Professional Documents
Culture Documents
Introduction To Theoretical and Computational Fluid Dynamics - C. Pozrikidis PDF
Introduction To Theoretical and Computational Fluid Dynamics - C. Pozrikidis PDF
Introduction To Theoretical and Computational Fluid Dynamics - C. Pozrikidis PDF
Fluid Dynamics
This page intentionally left blank
Introduction to Theoretical and Computational
Fluid Dynamics
Second Edition
C. Pozrikidis
Oxford University Press, Inc., publishes works that further Oxford Universitys objective of
excellence in research, scholarship, and education.
Copyright
c 2011 by Oxford University Press, Inc.
www.oup.com
Oxford is a registered trademark of Oxford University Press. All rights reserved. No part of this
text may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means,
electronic, mechanical, photocopying, recording, or otherwise, without the prior permission of Oxford
University Press.
Scope
This book diers from a typical text on theoretical uid dynamics in that the discourse is carried
into the realm of numerical methods and into the discipline of computational uid dynamics (CFD).
Specic algorithms for computing incompressible ows under diverse conditions are developed, and
computer codes encapsulated in the public software library Fdlib are discussed in Appendix C.
This book also diers from a typical text on computational uid dynamics in that a full discussion
of the theory with minimal external references is provided, and no experience in computational uid
dynamics or knowledge of its terminology is assumed. Contemporary numerical methods and com-
putational schemes are developed and references for specialized and advanced topics are provided.
Content
The material covered in this text has been selected according to what constitutes essential knowledge
of theoretical and computational uid dynamics. This intent explains the absence of certain special-
ized and advanced topics, such as turbulent motion and non-Newtonian ow. Although asymptotic
and perturbation methods are discussed in several places, emphasis is placed on analytical and
numerical computation. The discussion makes extensive usage of the powerful concept of Greens
functions and integral representations.
Use as a text
This book is suitable as a text in an advanced undergraduate or introductory graduate course on uid
mechanics, Stokes ow, hydrodynamic stability, computational uid dynamics, vortex dynamics, or
a special topics course, as indicated in the Note to the Instructor. Each section is followed by a set of
problems that should be solved by hand and another set of problems that should be tackled with the
help of a computer. Both categories of problems are suitable for self-study, homework, and project
assignment. Some computer problems are coordinated so that a function or subroutine written for
one problem can be used as a module in a subsequent problem.
v
Preface to the Second Edition
The Second Edition considerably extends the contents of the First Edition to include contemporary
topics and some new and original material. Clarications, further explanations, detailed proofs, orig-
inal derivations, and solved problems have been added in numerous places. Chapter 1 on kinematics,
Chapter 8 on hydrodynamic stability, and Chapter 11 on vortex methods have been considerably
expanded. Numerous schematic depictions and graphs have been included as visual guides to il-
lustrate the results of theoretical derivations. Expanded appendices containing useful background
material have been added for easy reference. These additions underscore the intended purpose of
the Second Edition as a teaching, research, and reference resource.
FDLIB
The numerical methods presented in the text are implemented in computer codes contained in the
software library Fdlib, as discussed in Appendix C. The directories of Fdlib include a variety of
programs written in Fortran 77 (compatible with Fortran 90), Matlab, and C++. The codes
are suitable for self-study, classroom instruction, and fundamental or applied research. Appendix D
contains the User Guide of the eighth directory of Fdlib on hydrodynamic stability, complementing
Chapter 8.
Acknowledgments
I thankfully acknowledge the support of Todd Porteous and appreciate useful comments by Jerey
M. Davis and A. I. Hill on a draft of the Second Edition.
C. Pozrikidis
vi
Note to the Instructor
This book is suitable for teaching several general and special-topics courses in theoretical and com-
putational uid dynamics, applied mathematics, and scientic computing.
Some sections can be taught as a guided reading assignment at the instructors discretion.
Course on Stokes ow
Chapter 6 can be used in its entirety as a text in a course on theoretical and computational Stokes
ow. The course syllabus includes governing equations and fundamental properties of Stokes ow, lo-
cal solutions, particulate microhydrodynamics, singularity methods, boundary-integral formulations,
boundary-element methods, unsteady Stokes ow, and unsteady particle motion. The students are
assumed to have a basic undergraduate-level knowledge of uid mechanics. Some topics from previ-
ous chapters can be reviewed at the beginning of the course.
vii
viii
Notation
In the text, an italicized variable, such as a, is a scalar, and a bold-faced variable, such as a, is a
vector or matrix. Matrices are represented by upper case and bold faced symbols. Matrixvector
and matrixmatrix multiplication is indicated explicitly with a centered dot, such as AB. With this
convention, a vector, a, can be horizontal or vertical, as the need arises. It is perfectly acceptable to
formulate the product A a as well as the product a A, where A is an appropriate square matrix.
Index notation and other conventions are dened in Appendix A.
Polymorphism
Occasionally in the analysis, we run out of symbols. A bold faced variable may then be used to
represent a vector or a matrix with two or more indices. A mental note should be made that the
variable may have dierent meanings, depending on the current context. This practice is consistent
with the concept of polymorphism in object-oriented programming where a symbol or function may
represent dierent entities depending on the data type supplied in the input and requested in the
output. The language compiler is trained to pick up the appropriate structure.
v = [vx , vy , vz ].
Accordingly, v admits a dual interpretation as a physical entity that is independent of the chosen
coordinate system, and as a mathematical vector. In conceptual analysis, we refer to the physical
interpretation; in practical analysis and calculations, we invoke the mathematical interpretation.
To prevent confusion, the components of a vector in non-Cartesian coordinates should never be
collected into a vector. Similar restrictions apply to matrices representing physical entities that
qualify as Cartesian tensors.
ix
x
Units
In engineering, all physical variables have physical dimensions such as length, length over time, mass,
mass over length cubed. Units must be chosen consistently from a chosen system, such as the cgs
(cm, g, s) or the mks (m, kg, s) system. It is not appropriate to take the logarithm or exponential
of a length a, ln a or ea , because the units of the logarithm or exponential are not dened. Instead,
we must always write ln(a/b) or ea/b , where b is another length introduced so that the argument of
the logarithm or exponential is a dimensionless variable or number. The logarithm or exponential
are then dimensionless variables or numbers.
Computer languages
Basic computer programming skills are necessary for the thorough understanding of contemporary
applied sciences and engineering. Applications of computer programming in uid mechanics range
from preparing graphs and producing animations, to computing uid ows under a broad range of
conditions. Recommended general-purpose computer languages include Fortran, C, and C++.
Free compilers for these languages are available for many operating systems. Helpful tutorials can
be found on the Internet and a number of excellent texts are available for self-study.
However, there are some important dierences. The traction in a uid refers to an innitesimal
surface xed in space, whereas the traction in a solid refers to a material surface before or after
deformation. Consequently, there are several alternative, albeit inter-related, denitions for the
stress tensor in a solid. Another important dierence is that the velocity gradient tensor in uid
mechanics is the transpose of the deformation gradient tensor in solid mechanics. This dierence
is easy to identify in Cartesian coordinates, but is lurking, sometimes unnoticed, in curvilinear
coordinates. These important subtleties should be born in mind when referring to texts on continuum
mechanics.
Contents
Preface v
1 Kinematic structure of a ow 1
1.1 Fluid velocity and motion of uid parcels . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Subparcels and point particles . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.2 Velocity gradient . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.3 Dyadic base . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.4 Fundamental decomposition of the velocity gradient . . . . . . . . . . . . 4
1.1.5 Vorticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.1.6 Fluid parcel motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.1.7 Irrotational and rotational ows . . . . . . . . . . . . . . . . . . . . . . . . 8
1.1.8 Cartesian tensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2 Curvilinear coordinates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.2.1 Orthogonal curvilinear coordinates . . . . . . . . . . . . . . . . . . . . . . 13
1.2.2 Cylindrical polar coordinates . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.2.3 Spherical polar coordinates . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.2.4 Plane polar coordinates . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.2.5 Axisymmetric ow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.2.6 Swirling ow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.2.7 Nonorthogonal curvilinear coordinates . . . . . . . . . . . . . . . . . . . . 20
1.3 Lagrangian labels of point particles . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.3.1 The material derivative . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.3.2 Point-particle acceleration . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.3.3 Lagrangian mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.3.4 Deformation gradient . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
1.4 Properties of uid parcels and mass conservation . . . . . . . . . . . . . . . . . . . 32
1.4.1 Rate of change of parcel volume and Eulers theorem in kinematics . . . . 33
1.4.2 Reynolds transport theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 34
1.4.3 Mass conservation and the continuity equation . . . . . . . . . . . . . . . 35
xi
xii
4 Hydrostatics 274
4.1 Pressure distribution in rigid-body motion . . . . . . . . . . . . . . . . . . . . . . 275
4.1.1 Stationary and translating uids . . . . . . . . . . . . . . . . . . . . . . . 275
4.1.2 Rotating uids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
4.1.3 Compressible uids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
4.2 The LaplaceYoung equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
4.2.1 Meniscus inside a dihedral corner . . . . . . . . . . . . . . . . . . . . . . . 286
4.2.2 Capillary force . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288
4.2.3 Small deformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289
4.3 Two-dimensional interfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290
4.3.1 A semi-innite meniscus attached to an inclined plate . . . . . . . . . . . 291
4.3.2 Meniscus between two vertical plates . . . . . . . . . . . . . . . . . . . . . 295
4.3.3 Meniscus inside a closed container . . . . . . . . . . . . . . . . . . . . . . 297
4.4 Axisymmetric interfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298
4.4.1 Meniscus inside a capillary tube . . . . . . . . . . . . . . . . . . . . . . . . 299
4.4.2 Meniscus outside a circular tube . . . . . . . . . . . . . . . . . . . . . . . 301
4.4.3 A sessile drop resting on a at plate . . . . . . . . . . . . . . . . . . . . . 304
4.5 Three-dimensional interfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
4.6 Software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 308
7 Irrotational ow 522
7.1 Equations and computation of irrotational ow . . . . . . . . . . . . . . . . . . . . 523
7.1.1 Force and torque exerted on a boundary . . . . . . . . . . . . . . . . . . . 524
7.1.2 Viscous dissipation in irrotational ow . . . . . . . . . . . . . . . . . . . . 524
7.1.3 A radially expanding or contracting spherical bubble . . . . . . . . . . . . 525
7.1.4 Velocity variation around a streamline . . . . . . . . . . . . . . . . . . . . 526
7.1.5 Minimum of the pressure . . . . . . . . . . . . . . . . . . . . . . . . . . . . 527
7.2 Flow past or due to the motion of a three-dimensional body . . . . . . . . . . . . 528
7.2.1 Flow past a translating rigid body . . . . . . . . . . . . . . . . . . . . . . 528
7.2.2 Flow due to the motion and deformation of a body . . . . . . . . . . . . . 530
7.2.3 Kinetic energy of the ow due to the translation of a rigid body . . . . . . 533
7.3 Force and torque exerted on a three-dimensional body . . . . . . . . . . . . . . . . 534
7.3.1 Steady ow past a stationary body . . . . . . . . . . . . . . . . . . . . . . 534
7.3.2 Force and torque on a body near a point source or point-source dipole . . 536
7.3.3 Force and torque on a moving rigid body . . . . . . . . . . . . . . . . . . . 537
7.4 Force and torque on a translating rigid body in streaming ow . . . . . . . . . . . 539
7.4.1 Force on a translating body . . . . . . . . . . . . . . . . . . . . . . . . . . 539
7.4.2 Force on a stationary body in uniform unsteady ow . . . . . . . . . . . . 542
7.4.3 Force on a body translating in a uniform unsteady ow . . . . . . . . . . . 543
7.4.4 Torque on a rigid body translating in an innite uid . . . . . . . . . . . . 543
7.5 Flow past or due to the motion of a sphere . . . . . . . . . . . . . . . . . . . . . . 546
7.5.1 Flow due to a translating sphere . . . . . . . . . . . . . . . . . . . . . . . 546
7.5.2 Uniform ow past a sphere . . . . . . . . . . . . . . . . . . . . . . . . . . . 547
7.5.3 Unsteady motion of a sphere . . . . . . . . . . . . . . . . . . . . . . . . . . 548
7.5.4 A spherical bubble rising at high Reynolds numbers . . . . . . . . . . . . 548
7.5.5 Weisss theorem for arbitrary ow past a stationary sphere . . . . . . . . . 549
7.5.6 Butlers theorems for axisymmetric ow past a stationary sphere . . . . . 552
7.6 Flow past or due to the motion of a nonspherical body . . . . . . . . . . . . . . . 554
7.6.1 Translation of a prolate spheroid . . . . . . . . . . . . . . . . . . . . . . . 555
7.6.2 Translation of an oblate spheroid . . . . . . . . . . . . . . . . . . . . . . . 557
7.6.3 Singularity representations . . . . . . . . . . . . . . . . . . . . . . . . . . . 558
xxi
References 1200
Index 1225
This page intentionally left blank
Kinematic structure of a ow 1
The study of uid mechanics is divided broadly into two main themes, kinematics and dynamics.
Kinematics analyzes the motion of a uid with reference to the structure of the velocity eld, whereas
dynamics examines the forces developing in a uid as the result of the motion. Fundamental concepts
of kinematics and dynamics are combined with the fundamental principle of mass conservation and
Newtons second law of motion to derive a system of dierential equations governing the structure
of a steady ow and the evolution of an unsteady ow.
It is clear that the average parcel velocity depends The uid velocity is dened as the outer
on the parcel radius, . Taking the limit as tends to limit of the mean parcel velocity.
zero, we nd that the average velocity approaches a well-dened asymptotic limit, until becomes
comparable to the molecular size. At that stage, we observe strong oscillations that are manifes-
tations of random molecular motions. We dene the velocity of the uid, u, at position, x, and
time, t, as the apparent or outer limit of the velocity of the parcel as tends to zero, just before the
discrete nature of the uid becomes apparent. This denition is the cornerstone of the continuum
approximation in uid mechanics. Under normal conditions, the velocity, u(x, t), is an innitely
dierentiable function of position, x, and time, t. However, spatial discontinuities may arise under
extreme conditions in high-speed ows, or else emerge due to mathematical idealization. If a ow is
steady, u is independent of time, u/t = 0, and the velocity at a certain position in space remains
constant in time.
1
2 Introduction to Theoretical and Computational Fluid Dynamics
If we continue to subdivide a parcel into subparcels with decreasingly small size, we will
eventually encounter subparcels with innitesimal dimensions called point particles. Each point
particle occupies an innitesimal volume in space, but an innite collection of point particles that
belongs to a nite (noninnitesimal) parcel occupies a nite volume in space. By denition, a point
particle located at the position x moves with the local and current uid velocity, u(x, t).
uj (x, t) uj (x0 , t) + x
i Lij (x0 , t), (1.1.1)
where x = x x0 is the vector connecting the point x0 to the point x, and summation over the
repeated index i is implied on the right-hand side. We have introduced the velocity gradient or rate
of relative displacement matrix, L = u, with components
uj
Lij = . (1.1.2)
xi
Explicitly, the velocity gradient is given by
u uy uz
x
x x x
u
x uy uz
L u = . (1.1.3)
y y y
ux uy uz
z z z
In physical terms, the velocity gradient expresses the spatial variation of the velocity across a col-
lection of subparcels or point particles composing a parcel. In Section 1.1.8, we will show that L
satises a transformation rule that qualies it as a second-order Cartesian tensor.
In formal mathematics, the velocity gradient is the tensor product of the gradient vector
operator, , and the velocity vector eld, u. To simplify the notation, we have set
u u, (1.1.4)
1.1 Fluid velocity and motion of uid parcels 3
where denotes the tensor product of two vectors dened in Section A.4, Appendix A. Similar sim-
plied notation will be used for other tensor products involving the gradient operator; for example,
= is the two-dimensional operator of the second derivatives.
L(x0 , t).
u(x, t) u(x0 , t) + x (1.1.5)
Note the left-to-right vectormatrix multiplication in the second term on the right-hand side.
u = ui e i , (1.1.7)
where summation is implied over the repeated index, i. The velocity gradient can be expressed in
the dyadic form
L = Lij ei ej , (1.1.8)
where summation is implied over the repeated indices, i and j. The kl component of the matrix
ei ej is
Expression (1.1.8) is the algebraic counterpart of the matrix depiction shown in (1.1.3).
Projecting equation (1.1.8) from the right onto the unit vector eq , where q is a free index, and
using the orthogonality property of the unit vectors, we obtain
Formulating the inner product of this equation with the unit vector ep , where p is another free index,
and working in a similar fashion, we extract the components of the velocity gradient,
where the double dot product of two matrices with matching dimensions is dened in Section A.4,
Appendix A. This expression may also be regarded as a consequence of the orthogonality property
I = e1 e1 + e2 e2 + e3 e3 . (1.1.13)
In matrix notation underlying this dyadic expansion, all elements of I are zero, except for the three
diagonal elements that are equal to unity.
Without loss of generality, we may assume that the kth component of the unit vector ei is
equal to 1 if k = i or 0 if k = i, that is, [ei ]k = ik . The kl component of the matrix ei ej is
[ei ej ]kl = [ei ]k [ej ]l = ik jl . All elements of the matrix ei ej are zero, except for the element
located at the intersection of the ith row and jth column that is equal to unity, [ei ej ]ij = 1,
where summation is not implied over the indices i and j. For example,
0 1 0
e 1 e2 = 0 0 0 . (1.1.14)
0 0 0
This expression illustrates that the dyadic base provides us with a natural framework for representing
a two-index matrix.
1
L=+E+ I, (1.1.15)
3
where I is the identity matrix. We have introduced the vorticity tensor,
1 1 uj ui
(L LT ), ij = , (1.1.16)
2 2 xi xj
1 1 1 uj ui
1
E (L + LT ) I, Eij = + ij , (1.1.17)
2 3 2 xi xj 3
where the superscript T denotes the matrix transpose. Explicitly, the components of and E are
given in Table 1.1.1. In Section 1.1.8, we will demonstrate that and E obey transformation rules
that qualify them as second-order Cartesian tensors.
1.1 Fluid velocity and motion of uid parcels 5
uy ux uz ux
0
x y x z
1 ux uy uz uy
0
2 y x y z
ux uz uy uz
0
z x z y
ux 1 1 uy ux 1 uz ux
( + ) ( + )
x 3 2 x y 2 x z
1 ux uy uy 1 1 uz uy
E
2 ( y + x ) ( + )
y 3 2 y z
1 ux uz 1 uy uz uz 1
( + ) ( + )
2 z x 2 z y z 3
Table 1.1.1 Cartesian components of the vorticity tensor, , and rate-of-deformation tensor, E, where
u is the rate of expansion. The trace of the rate-of-deformation tensor, E, is zero by
construction.
1.1.5 Vorticity
Because the vorticity tensor is antisymmetric, it has only three independent components that can
be accommodated into a vector. To implement this simplication, we introduce the vorticity vector,
, and set
1
ij = ijk k , (1.1.18)
2
where ijk is the alternating tensor dened in Section A.4, Appendix A, and summation is implied
over the repeated index k. Explicitly,
0 z y
1
= z 0 x . (1.1.19)
2
y x 0
Explicitly,
u uy
u uz
u ux
z x y
= ex + ey + ez , (1.1.22)
y z z x x y
where ex , ey , and ez are unit vectors along the x, y, and z axes.
Observing that the vorticity of a two-dimensional ow in the xy plane is oriented along the z
axis, we write
Translation
The rst term on the right-hand side of (1.1.24) expresses rigid-body translation. Under the inuence
of this term, a uid parcel translates with the uid velocity evaluated at the designated particle
center, x0 .
uid parcels resembling slender needles initially aligned with the eigenvectors elongate or compress
in their respective directions while remaining mutually orthogonal (Problem 1.1.2). An ellipsoidal
uid parcel with three axes aligned with the eigenvectors deforms, increasing or decreasing its aspect
ratios, while maintaining its initial orientation. These observations suggest that the third term on
the right-hand side of (1.1.24) expresses deformation that preserves orientation.
To examine the change of volume of a parcel under the action of this term, we consider a
parcel in the shape of a rectangular parallelepiped whose edges are aligned with the eigenvectors
of E. Let the initial lengths of the edges be dx1 , dx2 , and dx3 . After a small time interval dt has
elapsed, the lengths of the sides have become (1 + 1 dt) dx1 , (1 + 2 dt) dx2 , and (1 + 3 dt) dx3 .
Formulating the product of these lengths, we obtain the new volume,
Multiplying the three factors, we nd that, to rst order in dt, the volume of the parcel has been
modied by the factor 1 + (1 + 2 + 3 )dt. However, because the sum of the eigenvalues of E is
equal to the trace of E, which is zero by construction, the volume of the parcel is preserved during
the deformation.
If the rate of strain is everywhere zero, the ow must necessarily express rigid-body motion,
including translation and rotation (Problem 1.1.3).
Expansion
The last term on the right-hand side of (1.1.24) represents isotropic expansion or contraction. Under
the action of this term, a small spherical parcel of radius a and volume V = 4 3
3 a centered at
the point x0 expands or contracts isotropically, undergoing neither translation, nor rotation, nor
deformation. To compute the rate of expansion, we note that, after a small time interval dt has
elapsed, the radius of the spherical parcel has become (1 + 13 dt) a and the parcel volume has been
changed by the dierential amount
4 1 3 4 3
d V = 1 + dt a3 a . (1.1.27)
3 3 3
Expanding the cubic power of the binomial, linearizing the resulting expression with respect to dt,
and rearranging, we nd that
1 d V
= . (1.1.28)
V dt
This expression justies calling the divergence of the velocity, u, the rate of expansion or
rate of dilatation of the uid.
Summary
We have found that a small uid parcel translates, rotates, expands, and deforms by rates that are
determined by the local velocity, vorticity, rate of expansion, and rate-of-strain tensor, as illustrated
in Figure 1.1.1. To rst order in dt, the sequence by which these motions occur is inconsequential.
8 Introduction to Theoretical and Computational Fluid Dynamics
Figure 1.1.1 Illustration of the expansion, rotation, and deformation of a small spherical uid parcel
during an innitesimal period of time in a three-dimensional ow.
A number of common ows consist of adjacent regions of nearly rotational and nearly irrota-
tional ow. For example, high-speed ow past a streamlined body is irrotational everywhere except
inside a thin layer lining the surface of the body and inside a slender wake behind the body, as
discussed in Chapter 8. The ow between two parallel streams that merge at dierent velocities is
irrotational everywhere except inside a shear layer along the interface. Other examples of irrotational
ows will be presented in later chapters.
x2
x
x 1
2
2
e2 1
x1
e3 e1
3
x
3
x3
Figure 1.1.2 Illustration of two coordinate systems with shared origin. The cosines of the angles, 1 ,
2 , and 3 , are the direction cosines of the x1 axis, subject to the convention that 0 i .
The bold arrows are Cartesian unit vectors.
respectively, by (e1 , e2 , e3 ) and (e1 , e2 , e3 ). The position of a point particle in physical space can be
expressed in the unprimed or primed coordinates and corresponding unit vectors as
where summation is implied over the repeated index, i. The primed coordinates are related to the
unprimed coordinates, and vice versa, by an orthogonal transformation.
Orthogonal transformations
It is useful to introduce a transformation matrix, A, with elements
Using the geometrical interpretation of the inner vector product, we nd that the rst row of A
contains the cosines of the angles subtended between each unit vector along the x1 , x2 , and x3 axes,
and the unit vector in the direction of the x1 axis,
called the direction cosines of e1 , as shown in Figure 1.1.1. The second and third rows of A contain
the corresponding direction cosines for the unit vectors along the x2 and x3 axes, e2 and e3 .
Projecting (1.1.29) on ep or ep , where p is a free index, and using the orthogonality of the
primed and unprimed unit vectors, we nd that the x = (x1 , x2 , x3 ) and x = (x1 , x2 , x3 ) coordinates
are related by the linear transformations
x = A x x = AT x , (1.1.32)
10 Introduction to Theoretical and Computational Fluid Dynamics
where the superscript T indicates the matrix transpose. In index notation, these equations read
xi = Aij xj , xi = xj Aji . (1.1.33)
Similar transformation rules can be written for the components of a physical vector, such as the
velocity vector, the vorticity vector, and the angular velocity vector.
Relations (1.1.32) demonstrate that the transformation matrix, A, is orthogonal, meaning that
its inverse, indicated by the superscript 1, is equal to its transpose, A1 = AT . The determinant of
A and the length of any vector represented by any column or row are equal to unity. The projection
of any column or row onto a dierent column or row is zero.
Rotation matrices
In practice, the primed axes can be generated from the unprimed axes by three sequential rotations
about the x1 , x2 , and x3 axes, by respective angles equal to 1 , 2 , and 3 . We may say that
the primed system is rotated with respect to the unprimed system, and vice versa. To develop the
pertinent transformations, we introduce the rotation matrices
1 0 0 cos 2 0 sin 2
R(1) = 0 cos 1 sin 1 , R(2) = 0 1 0 ,
0 sin 1 cos 1 sin 2 0 cos 2
(1.1.34)
cos 3 sin 3 0
R(3) = sin 3 cos 3 0 .
0 0 1
Each rotation matrix is orthogonal, meaning that its transpose is equal to its inverse. The deter-
minant of each rotation matrix and the length of any vector represented by a column or row are
equal to unity. The mutual projection of any two dierent columns or rows is zero. The orthogonal
transformation matrix of interest is
A = R(3) R(2) R(1) , (1.1.35)
where the order of multiplication is consequential.
Projecting (1.1.36) and (1.1.37) from the right onto a unit vector, em , noting that ej em = jm ,
where jm is the Kronecker delta, and using (1.1.29), we nd that
Projecting this expression onto a unit vector, ep , and working in a similar fashion, we obtain
Renaming the indices and working in a similar fashion, we derive the distinguishing properties of a
second-order Cartesian tensor,
Tij = Tkl Aki Alj , Tij = Aik Ajl Tkl . (1.1.40)
In vector notation,
T = AT T A, T = A T AT . (1.1.41)
The transformations (1.1.40) and (1.1.41) are special cases of similarity transformations encountered
in matrix calculus (Problem 1.1.5). Because the matrix A is orthogonal, its transpose can be replaced
by its matrix inverse.
Invariants
The characteristic polynomial of a Cartesian tensor, P() = det(T I), is independent of the
coordinate system where the tensor is evaluated. Consequently, the coecients and hence the
roots of the characteristic polynomial dening the eigenvalues of T are invariant under a change
of coordinates. To demonstrate the invariance of the characteristic polynomial, we recall that the
transformation matrix A is orthogonal and write
Tij ij = Tkl Aki Alj Aki Akj = Tkl Aki Alj Aki kl Alj = Aki ( Tkl kl ) Alj . (1.1.42)
T I = AT (T I) A. (1.1.43)
Taking the determinant of both sides, expanding the determinant of the product, and noting that
det(AT ) = 1/det(A), we nd that
3 3 tensors
In the case of 3 3 three tensors of central interest in uid mechanics, we recast the characteristic
polynomial into the form
det(T I) = 3 + I3 2 I2 + I1 , (1.1.45)
12 Introduction to Theoretical and Computational Fluid Dynamics
dened in terms of the roots of the characteristic polynomial, 1 , 2 , and 3 , which are the eigenval-
ues of T. In Chapter 3, we will see that the invariants (1.1.46) play an important role in developing
constitutive equations that provide us with expressions for the stresses developing inside a uid as
the result of the motion.
Kinematic tensors
The Cartesian components of the velocity transform like the components of the position vector, as
shown in (1.1.32),
Using the chain rule of dierentiation, we transform the velocity gradient matrix L introduced in
(1.1.2) as
uj xk uj uj ul
Lij = = = A ki = Aki A lj , (1.1.48)
xi xi xk xk xk
which is tantamount to
Comparing (1.1.49) with the rst equation in (1.1.40), we conclude that L is a second-order Cartesian
tensor. We note that the third invariant I3 dened in (1.1.46) is equal to the rate of expansion,
u, and conrm that the rate of dilatation of a uid parcel is independent of the coordinate
system where the ow is described.
The symmetric and antisymmetric parts of L also obey the transformation rules (1.1.40).
Thus, the rate-of-strain matrix, E, and vorticity matrix, , are both second-order Cartesian tensors.
It is a straightforward exercise to show that the sequence of matrices, L L, L L L, . . ., E E,
E E E, . . ., are all second-order Cartesian tensors.
Problems
u = u1 e1 + u2 e2 + u3 e3 , (1.2.1)
1 1 1
e1 + e2 + e3 , (1.2.2)
h1 1 h2 2 h3 3
where h1 , h2 , and h3 are metric coecients. The velocity gradient admits the dyadic representation
L u = Lij ei ej , (1.2.3)
14 Introduction to Theoretical and Computational Fluid Dynamics
x ex
e
x
0
Figure 1.2.1 Illustration of cylindrical polar coordinates, (x, , ), dened with respect to companion
Cartesian coordinates, (x, y, z).
where summation is implied over the repeated indices, i and j. To prevent misinterpretation, the
curvilinear components of the velocity gradient tensor, Lij , should not be collected into a matrix.
The rate of expansion is the trace of the velocity gradient tensor. We note that the trace of
the matrix ei ej is zero if i = j or unity if i = j, and obtain the expression
u = L11 + L22 + L33 . (1.2.4)
The components of the vorticity are given by
1 = L23 L32 , 2 = L31 L13 , 3 = L12 L21 . (1.2.5)
Other properties are discussed in Section A.8, Appendix A.
Cylindrical, spherical, and plane polar coordinates discussed in Sections A.9A.11, Appendix
A, are used extensively in theoretical analysis and engineering applications. Expressions for the
velocity gradient in the corresponding dyadic base are derived in this section.
Velocity gradient
To derive the components of the velocity gradient tensor in cylindrical polar coordinates, we express
the gradient operator and velocity eld as
1
= ex + e + e , u = ux ex + u e + u e , (1.2.8)
x
where ex , e , and e are the unit vectors dened in Figure 1.2.1, and ux , u , and u are the
corresponding velocity components. The cylindrical polar unit vectors are related to the Cartesian
unit vectors by
L = ex + e + e (ux ex + u e + u e ). (1.2.12)
x
Carrying out the dierentiation and using (1.2.11), we obtain
L u = L e e , (1.2.13)
where summation is implied over the repeated indices, and . The cylindrical polar components
of the velocity gradient tensor, L , are given in Table 1.2.1(a).
To generate the matrix e e , we write e = 0, cos , sin and e = 0, sin , cos ,
and formulate their tensor product,
0 0 0
e e = 0 cos sin cos2 . (1.2.14)
2
0 sin cos sin
(a) ux u u
Lxx = Lx = Lx =
x x x
ux u u
Lx = L = L =
1 ux 1 u u 1 u u
Lx = L = L = +
(b)
ur u u
Lrr = Lr = Lr =
r r r
1 ur u 1 u ur u u
Lr = L = + L =
r r r r r
1 ur u 1 u cot 1 u ur + u cot
Lr = L = u L = +
r sin r r sin r r sin r
(c)
ur u
Lrr = Lr =
r r
1 ur u 1 u ur
Lr = L = +
r r r r
Table 1.2.1 Components of the velocity gradient tensor in (a) cylindrical, (b) spherical, and (c) plane
polar coordinates.
1 (u ) 1 u
1 ux u
x = L L = , = Lx Lx = ,
x
u ux
= Lx Lx = . (1.2.16)
x
If a uid rotates as a rigid body around the x axis with angular velocity , the velocity components
are ux = 0, u = 0, and u = , and the only surviving vorticity component is x = 2.
e
er
r x
e x
0
Figure 1.2.2 Illustration of spherical polar coordinates, (r, , ), dened with respect to companion
Cartesian coordinates, (x, y, z).
and
der der de de
= e , = sin e , er , = cos e ,
d d d d
de
= cos e sin er , (1.2.22)
d
where Greek variables stand for r, , or . Using the representations (1.2.20), we derive the dyadic
expansion
L u = L e e , (1.2.23)
where summation is implied over the repeated indices, and . The spherical components of the
velocity gradient tensor, L , are given in Table 1.2.1(b). The individual matrices, e e , can be
formulated using expressions (1.2.21).
y e er
x
0
Figure 1.2.3 Illustration of plane polar coordinates, (r, ), dened with respect to companion Carte-
sian coordinates, (x, y).
where er and e are unit vectors in the r and directions, and ur and u are the corresponding
velocity components.
The plane polar unit vectors are related to the Cartesian unit vectors by
er = cos ex + sin ey , e = sin ex + cos ey . (1.2.29)
The inverse relations are
ex = cos er sin e , ey = sin er + cos e . (1.2.30)
All derivatives e / are zero, except for two derivatives,
er de
= e , = er , (1.2.31)
d
where Greek variables stand for r or . Using the representations (1.2.28), we derive the dyadic
expansion
L u = L e e , (1.2.32)
where summation is implied over the repeated indices and . The plane polar components of the
velocity gradient tensor, L , are given in Table 1.2.1(c). The individual matrices, e e , can be
formulated using expressions (1.2.29).
Figure 1.2.4 Illustration of nonorthogonal curvilinear coordinates. The solid lines represent covariant
coordinates, (, , ), and the dashed lines represent the associated contravariant coordinates.
1.2.5 Axisymmetric ow
The velocity of an axisymmetric ow lies in an azimuthal plane of constant azimuthal angle, ,
measured around the axis of revolution. An example is laminar streaming (uniform) ow past a
stationary sphere. Referring to cylindrical or spherical polar coordinates, we set u = 0 and obtain
u = ux ex + u e = ur er + u e , (1.2.35)
where the velocity components (ux , u ) and (ur , u ) depend on (x, ) in cylindrical polar coordinates
or (r, ) in spherical polar coordinates, but are independent of the azimuthal angle, . The vorticity
of an axisymmetric ow points in the azimuthal direction,
= e . (1.2.36)
1.2.6 Swirling ow
The velocity of a swirling ow points in the azimuthal direction,
u = u e . (1.2.37)
An example of a swirling ow is the laminar ow due to the slow rotation of a sphere in a uid
of innite expanse. The azimuthal velocity component, u , depends on (x, ) in cylindrical polar
coordinates or (r, ) in spherical polar coordinates, but is independent of the azimuthal angle, . A
swirling ow can be superposed on an axisymmetric ow. The velocity of an axisymmetric ow in
the presence of swirling motion is independent of the azimuthal angle, . The vorticity is oriented
in any arbitrary direction.
vector components, as discussed in Sections A.12A.17, Appendix A. The gradient operator and
velocity vector can be expressed in contravariant or covariant component form. Correspondingly,
the velocity gradient tensor can be expressed in its pure contravariant, pure covariant, or mixed
component form, in four combinations.
The physical meaning of these representations stems from the geometrical interpretation of the
covariant and contravariant base vectors, combined with the denition of the directional derivative
of the velocity as the projection from the left of the velocity gradient tensor onto a unit vector
pointing in a desired direction.
Problems
Eulerian framework
We can repeat the process for any appropriate kinematic or intensive thermodynamic variable, f ,
such as a spatial or temporal derivative of the velocity, the kinetic energy, the thermal energy, the
enthalpy, or the entropy per unit mass of a uid, and thus regard that variable as a function of x
and t, writing
This point of view establishes an Eulerian framework for describing the kinematic structure of a ow
and the physical or thermodynamic properties of a uid.
22 Introduction to Theoretical and Computational Fluid Dynamics
Point particles
It is physically appealing and often mathematically convenient to describe the state of a uid and
structure of a ow in terms of the state and motion of the constituent point particles. As a rst
step, we identify the point particles by assigning to each one of them an identication vector, ,
consisting of three dimensional or dimensionless scalar variables called Lagrangian labels that take
values in a subset of an appropriate set of real vectors. In practice, the triplet
= (1 , 2 , 3 ) (1.3.2)
can be identied with the Cartesian or some other curvilinear coordinates of the point particles at a
specied instant in time. At any instant, the lines of constant 1 , 2 , and 3 dene a right-handed
system of curvilinear coordinates in physical space, as discussed in Sections A.8A.17, Appendix A.
Lagrangian framework
The value of any kinematic, physical, or intensive thermodynamic variable at a particular location,
x, and at a certain time instant, t, can be regarded as a property of the point particle that happens
to be at that location at that particular instant. Thus, for any appropriate scalar, vector, or matrix
function f that can be attributed to a point particle in a meaningful fashion, we may write
where X(, t) is the position of the point particle labeled at time instant t, is a constant, and
F is an appropriate variable expressing a suitable property of the point particles. Equation (1.3.3)
suggests that, in order to obtain the value of the function f at a point x at time t, we may identify
the point particle located at x at time t, read its label , measure the variable F, and multiply
the value of F thus obtained by the coecient to produce f . Often f and F represent the same
physical variable and their distinction is based solely upon the choice of independent variables used
to describe a ow.
In this case, the variable f is the uid velocity, u, the variable F is the velocity of the point particle
labeled at time instant t, denoted by U(, t), and the coecient is equal to unity. Equation
(1.3.4) suggests that, to obtain the uid velocity u at a point x at time t, we may identify the point
particle that resides at x at time t, look up its label, , and measure its velocity, U.
In this case, the variable f is the vorticity, , the variable F is the angular velocity of the point
particle labeled at time instant t, denoted by (a, t), and the coecient is equal to 2.
1.3 Lagrangian labels of point particles 23
The expression between the two equal signs in (1.3.6) is the velocity gradient at the location of the
point particle labeled at time t, computed in terms of the relative velocity of neighboring point
particles, where X is the point-particle position. In this case, the coecient is equal to unity.
Chain rule
We have at our disposal two sets of independent variables that can be used to describe a ow,
including the Eulerian set, (x, t), and the Lagrangian set, (, t). A relationship between the partial
derivatives of a function f with respect to these two sets can be established by applying the chain
rule, obtaining
Df f (X(, t), t)
f
f
DX
i
= = + . (1.3.9)
Dt t t x xi t Dt
For simplicity, we drop the parentheses around the Eulerian partial derivatives on the right-hand
side and use (1.3.8) and (1.3.4) to obtain
Df f f f
= + ui = + u f, (1.3.10)
Dt t xi t
which relates the material derivative to temporal and spatial derivatives with respect to Eulerian
variables. Summation is implied over the repeated index i in the central expression.
If all point particles retain their value of f as they move about the domain of a ow, the
material derivative vanishes, Df /Dt = 0, and we say that the eld represented by f is convected by
the ow.
24 Introduction to Theoretical and Computational Fluid Dynamics
Convective derivative
The term u f in (1.3.10) can be written as |u| f /lu , where f /lu is the rate of change of
f with respect to arc length measured in the direction of the velocity, lu . If the eld f is steady,
f /t = 0, and if the point point particles retain their value of f as they move in the domain of
ow, Df /Dt = 0, then f /lu must be zero. In that case, the value of f does not change in the
direction of the velocity and is therefore constant along paths traveled by point particles, identied
as streamlines. However, the value of f is generally dierent along dierent streamlines.
The velocity distribution in a uid that rotates steadily as a rigid body around the origin with
angular velocity is u = x, and the associated vorticity is = 2. Setting u/t = 0, noting
that the velocity gradient tensor is equal to the vorticity tensor, , and using (1.1.18), we obtain
the jth component of the acceleration,
1
aj = ui ij = ui ijk k = jki k ui , (1.3.13)
2
which shows that
a = u = ( x). (1.3.14)
Invoking the geometrical interpretation of the cross product, we nd that the acceleration points
toward the axis of rotation and its magnitude is proportional to the distance from the axis of rotation.
(a)
Dux ux ux ux ux ux
ax = + u ux = + ux + uy + uz
Dt t t x y z
Duy uy uy uy uy uy
ay = + u uy = + ux + uy + uz
Dt t t x y z
Duz uz uz uz uz uz
az = + u uz = + ux + uy + uz
Dt t t x y z
(b)
ux ux ux u ux
ax = + ux + u +
t x
u u u u u u2
a = + ux + u +
t x
u u u u u u u
a = + ux + u + +
t x
(c)
ur ur u ur u ur u2 + u2
ar = + ur + +
t r r r sin r
u u u u u u ur u cot 2
a = + ur + + + u
t r r r sin r r
u u u u u u u
a = + ur + + + ur + u cot
t r r r sin r
(d)
ur ur u ur u2
ar = + ur +
t r r
u u u u u r u
a = + ur + +
t r r r
Table 1.3.1 Components of the point-particle acceleration in (a) Cartesian, (b) cylindrical polar, (c)
spherical polar, and (d) plane polar coordinates.
26 Introduction to Theoretical and Computational Fluid Dynamics
These expressions illustrate that the cylindrical polar components of the acceleration are not simply
equal to the material derivative of the corresponding polar components of the velocity. In the case
of a uid that rotates steadily as a rigid body around the x axis with angular velocity , so that
ux = 0, u = 0, and u = , we nd that ax = 0, a = u2 /, and a = 0. These expressions are
consistent with the corresponding Cartesian form (1.3.14).
Dur u2 + u2 Du ur u cot 2
ar = , a = + u ,
Dt r Dt r r
Du u
a = + ur + u cot . (1.3.16)
Dt r
We observe that the spherical polar components of the acceleration are not simply equal to the
material derivative of the corresponding polar components of the velocity.
Dur u2 Du ur u
ar = , a = + . (1.3.17)
Dt r Dt r
We observe once again that the plane polar components of the acceleration are not simply equal to
the material derivative of the corresponding polar components of the velocity.
as illustrated in Figure 1.3.1. The subscript t emphasizes that the mapping function C t may change
in time. Some authors unfortunately call the mapping function C t () a motion. We will assume
that C t is a dierentiable function of the three scalar components of .
It is important to realize that C t is time-dependent even when the velocity eld is steady,
and is constant only if a point particle labeled is stationary. If we identify the label with the
1.3 Lagrangian labels of point particles 27
2 x2
X
x1
1
3 x3
Figure 1.3.1 Lagrangian mapping of the parametric space, , to the physical space, x. The position
of a point particle in physical space is denoted by X.
Cartesian coordinates of the point particles at the origin of time, t = 0, then = X(t = 0) and
C 0 () = X(t = 0).
for any i, j, and k. When these conditions are fullled, the Lagrangian eld X() can be constructed
up to an arbitrary constant by integrating the dierential equations (1.3.20).
Lagrangian metric
The dierential vector (d1 , 0, 0) in labeling space is mapped to the following dierential vector in
physical space,
C t
dX(1) = d1 . (1.3.24)
1
Similar equations can be written for dX(2) and dX(3) ,
C t C t
dX(2) = d2 , dX(3) = d3 . (1.3.25)
2 3
The three vectors, dX(1) , dX(2) , and dX(3) , are not necessarily orthogonal.
Dyadic expansion
Let
ei be Cartesian unit vectors in the labeling space, , and ei be Cartesian unit vectors in physical
space, x, for i = 1, 2, 3. The Lagrange Jacobian tensor admits the bichromatic dyadic expansion
J = Jij
ei e j , (1.3.29)
where summation is implied over the repeated indices, i and j. Working as in Section 1.2 for the
velocity gradient, we may expand the Lagrange Jacobian tensor in a dierent base comprised of
Cartesian or curvilinear base vectors denoted by Greek indices,
J = J
e e . (1.3.30)
For example, e can be cylindrical polar unit vectors in labeling space and e can be spherical polar
unit vectors in physical space.
1.3 Lagrangian labels of point particles 29
J = Jij ei ej = J e e , (1.3.31)
where summation is implied over the repeated indices, i and j corresponding to Cartesian coordi-
nates, as well as over the repeated indices and corresponding to general curvilinear coordinates.
In this case, the transpose of the Lagrange Jacobian tensor is called the deformation gradient. It
can be shown that the relative deformation gradient obeys transformation rules that render it a
second-order Cartesian tensor (e.g., [365]). A distinguishing feature of the deformation gradient is
that it is dimensionless.
d = Ft ( ) d, (1.3.32)
where Ft ( ) is the relative deformation gradient, also simply called the deformation gradient and
denoted by F, with components
i
Fij = . (1.3.33)
j
d = F1 d. (1.3.34)
It is sometimes useful to introduce the displacement eld, v . Solving for and substituting
the result into (1.3.33), we obtain
Polar decomposition
The polar decomposition theorem guarantees that the deformation gradient tensor can be resolved
into the product of an orthogonal tensor representing rotation, R, and a symmetric and positive-
denite tensor representing deformation, U or V, so that
F = R U = V R, (1.3.37)
where U is the right stretch tensor and V is the left stretch tensor. A small material vector d
deforms under the action of U and then rotates under the action of R, or rotates under the action
of R and then deforms under the action of U. A practical way of carrying out the decomposition is
suggested by the forthcoming equations (1.3.39) and (1.3.48).
where the superscript T denotes the matrix transpose. This expression motivates introducing the
right CauchyGreen strain tensor, C, dened in terms of the deformation gradient as
C = FT F = U2 . (1.3.39)
By denition,
|d|2 = d C d. (1.3.40)
C = I + D + D T + DT D. (1.3.41)
The last quadratic term can be neglected in the case of small (linear) deformation.
Because C is symmetric and positive denite, it has three real and positive eigenvalues and
three corresponding orthogonal eigenvectors (Problem 1.3.6). If dj is the jth eigenvector of C with
corresponding eigenvalue s2j , then
where dj is the image of the eigenvector and sj is the stretch ratio, for j = 1, 2, 3.
C = I + 2E (1.3.45)
We see that the right GreenLagrange strain tensor determines the change in the squared length of
a material vector due to deformation.
B F FT = V 2 . (1.3.48)
By denition,
|d|2 = d B1 d. (1.3.49)
B = I + D + DT + D DT . (1.3.50)
The last quadratic term can be neglected in the case of small (linear) deformation.
Because B is symmetric and positive denite, it has three real and positive eigenvalues and
three corresponding orthogonal eigenvectors (Problem 1.3.6). The eigenvalues of B are the same as
those of C, but the corresponding eigenvectors are generally dierent. If dj is the jth eigenvector
of C corresponding to the eigenvalue s2j , then
d j = (FT )1 dj (1.3.51)
Constitutive equations
The CauchyGreen strain tensors nd important applications in developing constitutive equations
that relate the stresses developing in a uid to the deformation of uid parcels, as discussed in
Section 3.3.
32 Introduction to Theoretical and Computational Fluid Dynamics
Problems
(a) (b)
n
u n dt
u
dS
Figure 1.4.1 (a) Illustration of a convected uid parcel or stationary control volume in a ow. Point
particles over an innitesimal patch of a material surface, dS, move during a small period of time,
dt, spanning a cylindrical volume, dV = un dt dS, where un is the normal velocity. The dashed
line outlines the new parcel shape. (b) The surface integral
u n dS is the rate of change of
the parcel volume. In contrast, the surface integral u n dS is the rate of convective mass
transport outward from a control volume.
where dV () is an innitesimal volume in labeling space. The second integral shows that J plays
the role of a volume density distribution function.
Since point particles move with the uid velocity and the parcel shape changes only because
of normal motion across the instantaneous parcel conguration, we may also write
dVp
= u n dS, (1.4.3)
dt P arcel
where n is the unit normal vector pointing outward from the parcel and dS is a dierential surface
area, as illustrated in Figure 1.4.1. Both sides of (1.4.3) have units of length cubed divided by time.
The thin closed line in Figure 1.4.1(a) describes the boundary of a uid parcel at a certain time, and
the bold closed line describes the boundary after time dt. The volume between the two boundaries
is the change in the parcel volume after time dt. Further justication for (1.4.3) will be given in
Section 1.10.
Using the divergence theorem to convert the surface to a volume integral, and introducing the
rate of expansion, u, we obtain from (1.4.3)
dVp
= dV = J dV (). (1.4.4)
dt P arcel A
34 Introduction to Theoretical and Computational Fluid Dynamics
Comparing (1.4.2) with (1.4.4) and noting that the subset A is arbitrary to eliminate the integral
signs, we obtain Eulers theorem of kinematics stating that
1 DJ
= . (1.4.5)
J Dt
We have found that the rate of change of the Jacobian, J , following a point particle is proportional
to the rate of expansion of the uid.
A formal but more tedious method of deriving (1.4.5) involves taking the material derivative
of the determinant of the Jacobian matrix stated in (1.3.22), nding
X1 X2 X3
1 1 1
X X3
DJ D 1 X2
= det = J1 + J2 + J3 , (1.4.6)
Dt Dt 2 2 2
X X2 X3
1
3 3 3
where
U1 X2 X3 U1 Xj X2 X3
1 1 1
Xj 1 1 1
U X3
X3
X2 U1 Xj X2 = U1 J .
= det
1
J1 = det (1.4.7)
2 2 2
Xj 2 2 2
X1
U X2 X3 U1 Xj X2 X3
1
3 3 3 Xj 3 3 3
The determinants J2 and J3 are computed in a similar fashion. The corresponding matrices arise
from J by replacing X2 with U2 in the second column or X3 with U3 in the third column. The
results are then added to produce (1.4.5).
J
+ (J u) = 2 J , (1.4.8)
t
which can be regarded as a transport equation for J .
For example, P can be identied with the density or specic thermal energy of the uid. To derive
the last expression in (1.4.9), we have observed once again that the volume of integration in labeling
space is independent of time.
Expanding the derivative inside the last integral and using (1.4.5), we nd that
d DP
P dV = + P u dV. (1.4.10)
dt P arcel P arcel Dt
Expressing the material derivative inside the integrand in terms of Eulerian derivatives and rear-
ranging, we obtain
d P
P dV = + (Pu) dV. (1.4.11)
dt P arcel P arcel t
Finally, we apply the divergence theorem to the second term inside the last integral and derive the
mathematical statement of the Reynolds transport theorem,
d P
P dV = dV + P u n dS, (1.4.12)
dt P arcel P arcel t P arcel
where n is the normal unit vector pointing outward from the parcel and the last integral is computed
over the parcel surface. Applying (1.4.12) with P set to a constant, we recover (1.4.3).
The left-hand side of (1.4.12) expresses the rate of accumulation of P inside the parcel due,
for example, to diusion across the parcel surface. Introducing a corresponding physical law, such
as Ficks law of diusion or Newtons second law of motion, transforms (1.4.12) into a transport
equation expressing a balance over a xed control volume. This interpretation is the cornerstone of
transport phenomena (e.g., [36]).
where X is the point-particle position, Applying (1.4.9) and (1.4.11) with P = and requiring
that mass neither disappears nor is produced in the owwhich is tantamount to stipulating that
36 Introduction to Theoretical and Computational Fluid Dynamics
uid parcels retain their mass as they move about the domain of owwe set dmp /dt = 0 and
require that the integrands are identically zero. The result is the continuity equation expressing
mass conservation for a compressible or incompressible uid,
D
D D
dV (X) = 0, J = 0, + u = 0, + (u) = 0. (1.4.14)
Dt Dt Dt t
Mass conservation imposes a kinematic constraint, demanding that the structure of the velocity eld
be such that uid parcels do not tend to occupy the same volume in space, leaving behind empty
holes.
where Vc is a xed control volume that coincides with the instantaneous volume of a uid parcel, and
Bc is the boundary of the control volume. Equation (1.4.15) states that the rate of accumulation
of mass inside a stationary control volume is equal to the rate of mass transport into the control
volume through the boundaries.
Equation (1.4.15) can be produced by integrating the last equation in (1.4.14) over the control
volume and using the divergence theorem to obtain a surface integral. Working similarly with (1.4.8),
we obtain
J
dV = J u n dS + 2 J dV, (1.4.16)
Vc t Bc Vc
which can be regarded as a transport equation for J in Eulerian integral form, where u is
the rate of expansion.
ux uy uz
u + + = 0. (1.4.18)
x y z
A vector eld with vanishing divergence, satisfying (1.4.18), is called solenoidal. According to our
discussion in Section 1.1, uid parcels of an incompressible uid translate, rotate, and deform while
1.4 Properties of uid parcels and mass conservation 37
retaining their volume. The terms incompressible uid, incompressible ow, and incompressible
velocity eld are sometimes used interchangeably. However, strictly speaking, compressibility is
neither a kinematic property of the ow nor a structural property of the velocity eld, but rather a
physical property of the uid. Combining (1.4.18) with (1.4.5), we nd that
DJ
= 0, (1.4.19)
Dt
which states that the Jacobian is convected with the ow. The mapping function C t introduced
in (1.3.18) is then called isochoric, from the Greek word o, which means equal, and the word
o, which means space.
Following are three examples of solenoidal velocity elds describing the motion of an incom-
pressible uid:
1. Since the divergence of the curl of any continuous vector eld is identically zero, a velocity eld
that derives from a dierentiable vector eld, A, as u = A is solenoidal. The function A
is called the vector potential, as discussed in Section 2.6.
2. Consider a velocity eld that derives from the cross product of two arbitrary vector elds, A
and B, as u = A B. Straightforward dierentiation shows that u = B A A B.
We observe that, if A and B are both irrotational, u will be solenoidal. Since any irrotational
vector eld can be expressed as the gradient of a scalar function, as discussed in Section 2.1,
the velocity eld u = is solenoidal for any pair of dierentiable functions, and .
3. Consider a velocity eld that derives from the gradient of a scalar function, , called the
potential function, u = . This velocity eld is solenoidal, provided that satises Laplaces
equation, 2 = 0. In that case, is called a harmonic potential.
Polar coordinates
In cylindrical polar coordinates, (x, , ), the continuity equation for an incompressible uid takes
the form
ux 1 (u ) 1 u
u= + + = 0. (1.4.20)
x
In spherical polar coordinates, (r, , ), the continuity equation for an incompressible uid takes the
form
ur ur 1 u u 1 u
u= +2 + + cot + = 0. (1.4.21)
r r r r r sin
In plane polar coordinates, (r, ), the continuity equation for an incompressible uid takes the form
1 (rur ) 1 u
u= + = 0. (1.4.22)
r r r
Expressions for the divergence of the velocity in more general orthogonal or nonorthogonal coordi-
nates are given in Sections A.8A.17, Appendix A.
38 Introduction to Theoretical and Computational Fluid Dynamics
uj uj
ui ui = 0. (1.4.23)
xj xi xi
Integrating (1.4.23) over an arbitrary control volume that is bounded by a surface, D, and using the
divergence theorem, we obtain
uj uj
ui ui nj dS = 0, (1.4.24)
D xi xi
where n is the unit vector normal to D. Equation (1.4.24) places an integral constraint on the
mutual structure of any two incompressible ows over a common surface.
Pairs of extensiveintensive variables include momentum and velocity, thermal energy and
temperature, kinetic energy and square of the magnitude of the velocity. In developing dynamical
laws governing the behavior of uid parcels, it is useful to have expressions for the rate of change
of an extensive variable in terms of the rate of change of the corresponding intensive variable. Such
expressions can be derived using the continuity equation (1.4.14).
Momentum
An important extensive variable is the linear momentum of a uid parcel, dened as
Mp (X, t) U(X, t) dV (X) = (, t) U(, t) J dV (). (1.4.25)
P arcel A
Using the continuity equation (1.4.14), we express the rate of the change of the linear momentum
in terms of the acceleration of the point particles as
dMp DU D(J)
Angular momentum
The angular momentum of a uid parcel computed with respect to the origin is
Ap (X, t) X U(X, t) dV (X) = (, t) X U(, t) J dV (). (1.4.28)
P arcel A
Working as in the case of the linear momentum and recalling that DX/Dt = U, we nd that
dAp D(X U)
= dV (X), (1.4.29)
dt P arcel Dt
and then
dAp DU
= X dV (X), (1.4.30)
dt P arcel Dt
for compressible or incompressible uids.
Generalization
For an arbitrary scalar, vector, or tensor intensive eld, F, we nd that
d d
F dV = F J dV (), (1.4.31)
dt P arcel dt A
yielding
d DF
F dV = dV, (1.4.32)
dt P arcel P arcel Dt
for compressible or incompressible uids. Equation (1.4.27) emerges by setting F = U.
Problems
Velocity
By denition, the velocity of a point particle, U, is equal to the rate of change of its position in
space. If the x coordinate of a point particle has changed by an innitesimal displacement, dX, over
an innitesimal time period, dt, then, by denition, Ux = dX/dt. Writing the counterparts of this
equation for the y and z coordinates, we obtain
dX dY dZ
Ux = , Uy = , Uz = . (1.5.3)
dt dt dt
In vector notation,
dX
U= = Ux ex + Uy ey + Uz ez . (1.5.4)
dt
In the present context of isolated point-particle motion, the total derivative, d/dt, is the same as
the material derivative, D/Dt.
Acceleration
The acceleration vector, a, is dened as the rate of change of the velocity,
dU d2 X
a = ax ex + ay ey + az ez = = . (1.5.5)
dt dt2
Accordingly, the Cartesian components of the point-particle acceleration are
d2 X d2 Y d2 Z (1.5.6)
ax = , ay = , az = .
dt2 dt2 dt2
If the Cartesian coordinates of a point-particle are constant or change linearly in time, the acceler-
ation is zero.
Note that the rst unit vector, ex , is xed, while the second and third unit vectors, e and e ,
change with position in space.
Velocity
Taking the time derivative of (1.5.8) and using relations (1.5.9), we derive an expression for the
point particle velocity,
dX d dX dex d de
U = (X ex + e ) = ex + X + e + , (1.5.10)
dt dt dt dt dt dt
and then
dX dX d d
U = ex + e + e . (1.5.11)
dt dt dt dt
The cylindrical polar components of the velocity are then
dX d d
Ux = , U = , U = . (1.5.12)
dt dt dt
Since is a dimensionless function, all three right-hand sides have units of length divided by time.
Acceleration
Dierentiating expression (1.5.11) with respect to time and expanding the derivatives, we derive an
expression for the acceleration,
d2 X d dX d d
a= = ex + e + e
dt2 dt dt dt dt
d2 X d2 d de d d d2 d de
= 2
ex + 2
e + + e + 2
e + . (1.5.13)
dt dt dt dt dt dt dt dt dt
Now we substitute relations (1.5.9) and nd that
d2 X d2 X d2 d d d d d2 d d
a= 2
= 2
e x + 2
e + e + e + 2
e e . (1.5.14)
dt dt dt dt dt dt dt dt dt dt
Finally, we consolidate the terms on the right-hand side and obtain the cylindrical polar components
of the acceleration,
2
d2 X d2 d d2 d d 1 d 2 d
ax = , a = , a = + 2 = . (1.5.15)
dt2 dt2 dt dt2 dt dt dt dt
Note that a change in the azimuthal angle, , is accompanied by radial acceleration, a .
X = R er , (1.5.17)
where e are spherical polar unit vectors for = r, , . The rates of change of the unit vectors
following the motion of a point particle are given by the relations
der d d de d d
= sin e + e , = cos e er ,
dt dt dt dt dt dt
(1.5.18)
de d d
= cos e sin er .
dt dt dt
All three unit vectors change with position in space.
Velocity
Taking the time derivative of (1.5.17) and using relations (1.5.18), we derive an expression for the
point-particle velocity,
dX dR der dR d d
U= = er + R = er + R sin e + R e . (1.5.19)
dt dt dt dt dt dt
The spherical polar components of the velocity are
dR d d
Ur = , U = R , U = R sin . (1.5.20)
dt dt dt
Since the functions and are dimensionless, all three right-hand sides have units of length divided
by time.
Acceleration
Dierentiating expression (1.5.19) with respect to time, we obtain the point-particle acceleration, a.
Expanding the derivatives, we nd that
d2 X
a = A + B + C, (1.5.21)
dt2
where the term
d dR
d2 R dR der d2 R dR d dR d
A er = 2
er + = 2
er + sin e + e (1.5.22)
dt dt dt dt dt dt dt dt dt dt
corresponds to the rst term on the right-hand side of (1.5.19), the term
d d
dR d d2 d d
B R sin e = sin e + R 2 sin e + R cos e (1.5.23)
dt dt dt dt dt dt dt
d de dR d d2 d d
+R sin = sin e + R 2 sin e + R cos e
dt dt dt dt dt dt dt
d
2 d
2
R sin cos e R sin2 er (1.5.24)
dt dt
1.5 Point-particle motion 43
corresponds to the second term on the right-hand side of (1.5.19), and the term
d d
dR d d2 d de
C R e = e + R 2 e + R
dt dt dt dt dt dt dt
dR d d2 d d d
2
= e + R 2 e + R cos e R er (1.5.25)
dt dt dt dt dt dt
corresponds to the third term on the right-hand side of (1.5.19). Consolidating the various terms,
we derive the cylindrical polar components of the acceleration,
d2 R d
2 d
2
2
ar = R sin R ,
dt2 dt dt
d2 dR d d
2
a = R + 2 R sin cos , (1.5.26)
dt2 dt dt dt
d2 dR d d d
a = R sin + 2 sin + 2 R cos .
dt2 dt dt dt dt
Note that a change in the meridional angle, , or azimuthal angle, , is accompanied by radial
acceleration.
X = R er , (1.5.28)
where e are plane polar unit vectors for = r, . The rates of change of the unit vectors following
the motion of a particle are given by
der d de d
= e , = er . (1.5.29)
dt dt dt dt
Note that both unit vectors change with position in space.
Velocity
To derive the velocity components, we work as previously with cylindrical and spherical polar coor-
dinates, and nd that
dR d
Ur = , U = R . (1.5.30)
dt dt
Note that the right-hand sides have units of length divided by time.
44 Introduction to Theoretical and Computational Fluid Dynamics
Acceleration
The plane polar components of the particle acceleration are
d2 R d
2 d2 dR d 1 d 2 d
ar = R , a = R + 2 = R . (1.5.31)
dt2 dt dt2 dt dt R dt dt
Note that a change in the polar angle, , is accompanied by radial acceleration. If a particle moves
along a circular path of constant radius R centered at the origin, dR/dt = 0, the acceleration
components are
d
2 U2 d2
ar = R = , a = R . (1.5.32)
dt R dt2
A radial acceleration is necessary to follow the circular path.
x2 x
x
1
xnew
2 X
e2 1
x1
e3 e1
3
x3
Figure 1.5.1 Rotation of a point x around the x1 axis by angle . The new coordinates are found
using the projection matrix (1.5.36).
where a2 + b2 + c2 = 1. Carrying out the multiplications and simplifying, we derive the projection
matrix
2
a ab ac 0 c b
P = cos I + (1 cos ) ab b2 bc + sin c 0 a , (1.5.38)
2
ac bc c b a 0
where I is the identity matrix. The matrix in the second term on the right-hand side of (1.5.38)
is symmetric, whereas the matrix in the third term is skew-symmetric. This means that a rotation
vector, c, with the property that xnew = c x, cannot be found.
In practice, we may specify a point X on the x1 axis, and compute the direction cosines
X Y Z
a= , b= , c= , (1.5.39)
|X| |X| |X|
Problems
N
B
t
X
Figure 1.6.1 Illustration of a material vector with innitesimal length and designated beginning, A,
and end, B. The unit vector N is normal to the material vector.
D X u
= UB UA = Xi = X L, (1.6.1)
Dt xi
where L is the velocity gradient tensor introduced in (1.1.2) evaluated at the position of the material
vector, and summation is implied over the repeated index i.
Stretching
The rate of change of the length of the material vector, l |X|, is given by
D l D (X X)1/2 1 D (X X) X D X
= = 1/2
= . (1.6.2)
Dt Dt 2 (X X) Dt l Dt
1.6 Material vectors and material lines 47
Reorientation
To compute the rate of change of the unit vector, t, that is parallel to a material vector, X, at all
times, we write
D X D(t l) Dt D l
= = l +t , (1.6.5)
Dt Dt Dt Dt
and use (1.6.1) and (1.6.4) to obtain
Dt
= (t L) (I t t), (1.6.6)
Dt
where I is the identity matrix. The operator I t t on the right-hand side removes the component
of a vector in the direction of the unit vector, t, leaving only the normal component. Accordingly,
the right-hand side of (1.6.6) involves derivatives of the velocity components in a plane that is normal
to the material vector with respect to arc length measured in the direction of the material vector.
vector is aligned with the x axis, t = ex = 1, 0, 0 , we nd that
For example, if a material
Dt/Dt = 0, uy /x, uz /x , which shows that t tends to acquire components along the y and z
axes.
Mutual reorientation
The cosine of the angle subtended between two material unit vectors, t1 and t2 , is given by an
inner product, cos = t1 t2 . Taking the material derivative, we obtain
D cos Dt1 Dt2
= t2 + t1 . (1.6.7)
Dt Dt Dt
Substituting (1.6.6) and rearranging, we nd that
D cos
= (t1 t2 + t2 t1 ) : L cos (t1 t1 + t2 t2 ) : L. (1.6.8)
Dt
Because the tensors t1 t2 + t2 t1 and t1 t1 + t2 t2 are symmetric, the velocity gradient L can
be replaced by its symmetric constituent expressed by the vorticity tensor on the right-hand side.
Accordingly, the rst term on the right-hand side of (1.6.8) expresses the symmetric component of
48 Introduction to Theoretical and Computational Fluid Dynamics
(a) (b)
b
n
n
t
t
l
b l
Figure 1.6.2 (a) A closed or open (b) material line is parametrized by a Lagrangian label, . Shown
are the tangent unit vector, t, the normal unit vector, n, and the binormal unit vector, b.
the velocity gradient tensor in a dyadic base dened by the unit vectors t1 and t2 . If the two vectors
are initially perpendicular, cos = 0 and the second term on the right-hand side of (1.6.8) is zero.
Normal vector
It is of interest to consider the evolution of a unit vector, N, that is and remains normal to a material
vector, as shown in Figure 1.6.1. Taking the material derivative of the constraint N N = 1, we
nd that N DN/Dt = 0, which shows that the vector DN/Dt lacks a component in the direction
of N. Taking the material derivative of the constraint N X = 0 and using (1.6.1), we nd that
t DN/Dt = t L N, where t = X/l is the tangent unit vector. Combining these expressions,
we derive the evolution equation
DN
= (t t) L N + t N, (1.6.9)
Dt
where is an unspecied rate of rotation of N about t. In the case of two-dimensional or axisym-
metric ow where the vectors t and N are restricted to remain in the xy or an azimuthal plane,
= 0.
Next, we consider a material line consisting of a xed collection of point particles forming an open
or closed loop, as shown in Figure 1.6.2. To identify the point particles, we introduce a scalar label,
, taking values inside an appropriate set of real numbers, , and regard the position of the point
particles, X, as a function of and time, t, writing X(, t). The unit vector
1 X
t= (1.6.10)
h
is tangential to the material line, where h = |X/|. Because t is a unit vector, t t = 1, we have
t t/ = 12 (t t)/ = 0, which shows that the vector t/ is normal to t.
1.6 Material vectors and material lines 49
The arc length of an innitesimal section of the material line is dl = h d and the total arc
length of the material line is
L= h d. (1.6.11)
This expression shows that h is a scalar metric coecient for the arc length similar to the Jacobian
metric for the volume of a uid parcel, J . If we identify the label with the instantaneous arc
length along the material line, l, then h = 1.
b = t n. (1.6.13)
The three unit vectors, t, n, and b, dene three mutually orthogonal directions that can be used to
construct a right-handed, orthogonal, curvilinear system of axes so that
t = n b, n = b t, (1.6.14)
as shown in Figure 1.6.2. The plane containing the pair (t, n) is called the osculating plane, the
plane containing the pair (n, b) is called the normal plane, and the plane containing the pair (b, t),
is called the rectifying plane.
Dierentiating (1.6.13) with respect to arc length, l, expanding the derivative on the right-
hand side, and using (1.6.12), we nd that the vector db/dl is perpendicular to t. Since b is a unit
vector, d(b b)/dl = 2b (db/dl) = 0, which shows that db/dl is perpendicular to b. We conclude
that b must be parallel to n and write
b
= n, (1.6.15)
l
where is the torsion of the material line.
n
= t + b, (1.6.16)
l
involving the curvature and the torsion of the line.
50 Introduction to Theoretical and Computational Fluid Dynamics
Equations (1.6.12), (1.6.15), and (1.6.16) comprise the FrenetSerret relations. In the litera-
ture, the curvature, torsion, or both, may appear with opposite signs as a matter of convention.
Matrix formulation
The FrenetSerret relations can be collected into the matrix form
t 0 0 t
d
n = 0 n , (1.6.17)
dl
b 0 0 b
involving a singular skew-symmetric matrix on the right-hand side. One eigenvalue of this matrix is
zero and the other two eigenvalues are complex conjugates, given by = i(2 + 2 )1/2 .
The last integrand expresses the local rate of extension of the material line.
Dh U (t U) t
=t = U . (1.6.25)
Dt
Using the FrenetSerret relation (1.6.12) and setting the point-particle velocity equal to the uid
velocity, U = u, we obtain
Dh (u t) l
= + u n. (1.6.26)
Dt
The two terms on the right-hand side express the change in length of an innitesimal section of
the material line due to stretching along the line, and extension due to motion normal to the line.
These interpretations become more clear by considering the behavior of a circular material line that
exhibits tangential motion with vanishing normal velocity, or is expanding in the radial direction
while remaining in its plane (Problem 1.6.1).
Now substituting (1.6.26) into (1.6.24) and performing the integration, we obtain a revealing
expression for the rate of the change of the length of a material line,
dL end
= u t start + u n dl. (1.6.27)
dt Line
If a material line forms a closed loop, the rst term on the right-hand side vanishes and the tangential
motion does not contribute to the rate of change of the total arc length of the line.
Dt u
= (t L) (I t t) = (I t t) (1.6.28)
Dt l
or
Dt u u
= n n+ b b, (1.6.29)
Dt l l
52 Introduction to Theoretical and Computational Fluid Dynamics
+n = 2 = n . (1.6.31)
Dt Dt h Dt h Dt h Dt l Dt
Next, we project this equation onto n, recall that n n = 1 and thus n Dn/Dt = 0, and rearrange
to obtain an evolution equation for the curvature,
D Dh Dt
u Dt
= n = t n . (1.6.32)
Dt h Dt l Dt l l Dt
Taking the derivative of the rate of change Dt/Dt given in (1.6.28), we obtain
Dt
2 u u
= 2 (I t t) + (t n + n t). (1.6.33)
l Dt l l
Substituting this expression into (1.6.32) and simplifying, we derive the nal form
D u 2u
= 2 t 2 n. (1.6.34)
Dt l l
In the
case of an expanding circle of radius
a(t) in the xy plane, we set u = U cos , sin ,
t = sin , cos , and n = cos , sin , where U is the velocity of expansion and is the polar
angle. Substituting l = a and = 1/a, we obtain an expected equation, da/dt = U .
Dn u
1 2u
= n t 2
b b. (1.6.36)
Dt l l
If a line is and remains in the osculating plane containing t and n, the second term on the right-hand
side does not appear. Expression (1.6.36) can also be derived by substituting (1.6.34) into (1.6.31).
1.6 Material vectors and material lines 53
n+ = n . (1.6.38)
Dt Dt h Dt l Dt
Projecting this equation onto n, we obtain
D Dh Db
u Db
= n = t n . (1.6.39)
Dt h Dt l Dt l l Dt
Substituting (1.6.37), using the FrenetSerret relations, and simplifying, we obtain
D u 1 2u 1 1 3u
= ( t + b) + 2
( n + b) b. (1.6.40)
Dt l l l l3
The presence of a third derivative with respect to arc length is an interesting feature of this equation.
Spin vector
The evolution equations for the tangent, normal, and binormal unit vectors derived previously in
this section can be collected into the unied forms
Dt Dn Db
= s t, = s n, = s b, (1.6.41)
Dt Dt Dt
where
1 2u
u
u
s= 2
b t b n+ n b (1.6.42)
l l l
is the spin vector. Formulas (1.6.41) complement the Darboux relations (1.6.23).
yielding
D ds
t = 2 s t. (1.6.45)
Dt dl
Working similarly with the second and third equations in (1.6.43), we obtain the compatibility
condition
D ds
= 2 s , (1.6.46)
Dt dl
which can be regarded as an evolution equation for the Darboux vector.
U = ut t + un n + ub b. (1.6.47)
= t t + d1 d1 + d2 d2 (1.6.51)
The evolution equations for the rotated triad take the form
Dt D d1 D d2
= s t, = s d1 , = s d2 , (1.6.53)
Dt Dt Dt
1.6 Material vectors and material lines 55
t
d1
l
b d2
Figure 1.6.3 The orthonormal triad, (t, d1 , d2 ), arises from the FrenetSeret triad, (t, n, b), by ro-
tating the normal and binormal vectors, n and b, about the tangent vector, t, by angle .
where
u
u
s = st t b n+ n b (1.6.54)
l l
is the spin vector with an unspecied tangential component, st . Formulas (1.6.53) complement
the Darboux equations (1.6.23). Consistency between these formulas requires the compatibility
condition (1.6.46).
Problems
n t
n
n t
Using the label and , we eectively establish a mapping of the curved material surface in
the three-dimensional physical space to a certain area in the parametric (, ) plane. We may say
that (, ) are convected coordinates, meaning that the constituent point particles retain the values
of (, ) as they move in the domain of ow.
Since the lengths of the vectors t and t are equal to unity, t t = 1 and t t = 1. Taking
the or derivatives of these equations, we nd that
t t t t
t = 0, t = 0, t = 0, t = 0. (1.7.4)
It is evident from the denitions (1.7.1) that the unit tangent vectors satisfy the relation
(h t ) (h t ) 2X
= = . (1.7.5)
Expanding the rst two derivatives and projecting the resulting equation onto t or t , we obtain
h t h h t h
= h t + t t , = h t + t t . (1.7.6)
Orthogonal coordinates
In the case of orthogonal curvilinear coordinates, t t = 0, the second term on the right-hand
side of each equation in (1.7.6) does not appear. Since the vector t / lacks a component in the
direction of t and the vector t / lacks a component in the direction of t , we may write
t 1 h t 1 h
= t , = t . (1.7.8)
h h
is the surface metric, as shown in Figure 1.7.1. Combining (1.7.1) with (1.7.9), we nd that
h h
n= t t . (1.7.11)
hs
Since any tangential vector is perpendicular to the normal vector, we can write
X X
n = 0, n = 0. (1.7.12)
Dierentiating the rst equation with respect to and the second equation with respect to ,
expanding the derivatives and combining the results to eliminate the common term n 2 X/,
we obtain the useful relation
n n
t = t . (1.7.13)
l l
Surface area
The area of a dierential element of the material surface is dS = hs d d. The total area of the
material surface is
S= hs d d, (1.7.14)
where is the range of variation of and over the surface. Equation (1.7.14) conrms that hs is a
metric coecient associated with the surface coordinates, analogous to the Jacobian, J , associated
with the Lagrangian labels of three-dimensional uid parcels. To compute the rate of change of the
surface area of a material surface, we dierentiate (1.7.14) with respect to time and note that the
limits of integration are xed to obtain
dS Dhs 1 Dhs
= d d = dS, (1.7.15)
dt Dt S s Dt
h
where S denotes the surface. The last integrand, expressing the rate of expansion of an innitesimal
material patch, is identied with the rate of dilatation of the surface, as discussed in Section 1.7.3.
Orthogonal coordinates
If the surface coordinates (, ) are orthogonal, t t = 0 and |t t | = 1, the surface metric is
the product of the two surface coordinate metrics, hs = h h . Projecting equations (1.7.8) onto n,
we nd that
t t
n = 0, n = 0, (1.7.16)
yielding
n n
t = 0, t = 0. (1.7.17)
1.7 Material surfaces 59
N hs n. (1.7.18)
Using the denition of the normal unit vector, n, stated in (1.7.9), and the dynamical law (1.6.1)
for the rate of change of a material vector, we compute
DN D X
X X D X
= + (1.7.19)
Dt Dt Dt
and then
DN X
X X X
= L + L . (1.7.20)
Dt
In index notation,
DNi Xl Xk Xk Xl
= ijk Llj + ikj Llj , (1.7.21)
Dt
DNi X X
l k Xk Xl
Now using the rules of repeated multiplication of the alternating matrix discussed in Section A.4,
Appendix A, we obtain
DNi Xm Xn
= Llj (ip jl il jp ) pmn (1.7.24)
Dt
or
DNi Xm Xn Xm Xn
= Ljj imn Llj jmn . (1.7.25)
Dt
Switching back to vector notation, we express the nal result in the form
1 DN
= ( I L ) n = (n ) u, (1.7.26)
hs Dt
where = u is the rate of expansion [432]. Because the operator n involves tangential
derivatives, only the surface distribution of the velocity is required to evaluate the rate of change of
the material normal vector, in agreement with physical intuition. Consequently, the rate of change
60 Introduction to Theoretical and Computational Fluid Dynamics
of the material normal vector can be expressed solely in terms of the known instantaneous geometry
and motion of the surface and is independent of the ow o the surface [433].
Metric coecient
An evolution equation for the surface metric coecient, hs , can be derived by expressing (1.7.10) in
the form
X X X X
h2s = . (1.7.27)
Taking the material derivative, we obtain
Dhs X X D X X
2 hs =2 , (1.7.28)
Dt Dt
which can be restated as
Dhs DN
=n . (1.7.29)
Dt Dt
Using expression (1.7.26), we obtain the evolution equation
1 Dhs
= n L n, (1.7.30)
hs Dt
where = u is the rate of expansion
Surface divergence
Since the scalar n L n represents the normal derivative of the normal component of the velocity,
the right-hand side of (1.7.30) represents the divergence of the velocity in the tangential plane,
s u n L n = (P ) u = trace P L P , (1.7.31)
called the surface divergence of the velocity, where P = I n n is a tangential projection operator
and I is the identity matrix. Equation (1.7.30) becomes
1 Dhs
= s u. (1.7.32)
hs Dt
If a material
surface is inextensible, the surface divergence of the velocity is zero. For example, if
n = 1, 0, 0 , we obtain s u = uy /y + uz /z.
Taking the material derivative of (1.7.10) and carrying out straightforward dierentiations,
we nd that
Dhs U U
= h t h t n. (1.7.33)
Dt
Rearranging the triple scalar products, we obtain
Dhs U U
= h (t n) h (t n) . (1.7.34)
Dt
1.7 Material surfaces 61
Comparing this equation with (1.7.32), we derive an alternative expression for the surface divergence
of the velocity,
h h U U
s u = (t n) + (n t ) . (1.7.35)
hs l l
If the surface coordinates are orthogonal, t n = t and n t = t .
Normal vector
Combining (1.7.26) with (1.7.30), we derive an evolution equation for the unit vector normal to the
surface,
Dn 1 DN Dhs
= n = n ( s u) L n. (1.7.36)
Dt hs Dt Dt
Rearranging the expression inside the last parentheses, we obtain
Dn
= (n n) L n L n = P L n, (1.7.37)
Dt
where P = I n n is a tangential projection operator and I is the identity matrix (Problem 1.7.1).
We observe that n Dn/Dt = 0, as required. Expression (1.7.37) is consistent with the rst term on
the right-hand side of (1.6.9). Rearranging the last expression in (1.7.37), we conrm that the rate
of change of the normal vector, Dn/Dt, has only a tangential component,
Dn
= n (L n) n = w n, (1.7.38)
Dt
where w n (L n).
Density-weighted metric
Combining (1.7.26) and (1.7.30) with the continuity equation (1.2.6), we obtain the compact forms
1 D(N) 1 D(hs )
= L n, = n L n. (1.7.39)
hs Dt hs Dt
These equations nd useful applications in developing evolution equations for physical quantities
dened over a material surface.
h t n = h h U t n
h
(h U )
t
t t n + h2 U
= t n. (1.7.41)
h
62 Introduction to Theoretical and Computational Fluid Dynamics
+ h2 U
h t n = (t n). (1.7.42)
h h h
Working in a similar fashion with another term, we nd that
(U t )
t
t
h t n = h U t n = h U (t n). (1.7.43)
Combining these results, we obtain
(U t )
(U t )
h n t h n t
hs (h U ) t
t
= + U h2 (t n) h (t n) . (1.7.44)
h h h
Working in a similar fashion with the second tangential velocity, we obtain
(U t )
(U t )
h n t h n t
hs (h U ) 2 t
t
= + U h (t n) h (t n) . (1.7.45)
h h h
The term on the right-hand side of (1.7.33) involving the normal component of the velocity takes
the form
(U n)
(U n)
n n
n n
h n t h n t = Un n h t h t . (1.7.46)
Substituting (1.7.44), (1.7.45), and (1.7.46) into (1.7.33), we obtain four terms,
1 Dhs 1 (h U ) (h U )
= +
hs Dt h h
1
t
t
+U h2 (t n) h (t n)
hs h
1
t
t
+U h2 (t n) h (t n)
hs h
h n h
n
+Un (t n) (t n) , (1.7.47)
hs hs
with the understanding that the point-particle velocity is the uid velocity, U = u.
The last term on the right-hand side of (1.7.47) expresses dilatation due to normal motion associated
with expansion or contraction. In Section 1.8, we will see that the expression enclosed by the last
large parentheses is equal to twice the mean curvature of the surface, m , as shown in (1.8.19),
h h n n
2m = (t n) + (n t ) . (1.7.49)
hs l l
The expression on the right-hand side can be computed readily from a grid of surface marker points
by numerical dierentiation.
In summary, we have derived an expression that delineates the signicance of tangential and normal
motions,
1 Dhs
= s us + 2m u n. (1.7.50)
hs Dt
Only the rst term on the right-hand side appears over a stationary surface where the normal velocity
is zero. The evolution equation (1.7.50) for a material surface is a generalization of the evolution
equation (1.6.26) for a material line.
The preceding interpretations become more evident by assuming that the surface coordinates are
orthogonal, t t = 0. The rst term on the right-hand side of (1.7.47) is the standard expression
for the surface divergence of the velocity in orthogonal curvilinear coordinates,
1 (h U ) (h U )
s us = + . (1.7.51)
h h
The term enclosed by the rst square brackets on the right-hand side of (1.7.47) is zero. To show
this, we note that t n = t and t n = t , recall that t t / = 0 because t is a unit
vector, and nd that the terms enclosed by the rst pair of square brackets simplify into
t
t 1
t h t
h2 t + h t = h2 + h t = + h t , (1.7.52)
h h
which is zero in light of (1.7.6). Working in a similar fashion, we nd that the term enclosed by the
second square brackets on the right-hand side of (1.7.47) is also zero.
64 Introduction to Theoretical and Computational Fluid Dynamics
where and vary over the domain in parametric space, and N hs n is the material normal
vector. In various applications of uid mechanics and transport phenomena, q can be identied
with the velocity, the vorticity, the temperature gradient, or the gradient of the concentration of a
chemical species. When q is the velocity, Q is the volumetric ow rate. When q is the vorticity, Q
is the circulation around a loop bounding an open surface, as discussed in Section 1.12.2.
Taking the time derivative of (1.7.53), using (1.7.26), and introducing the rate of expansion,
= u, we obtain
dQ Dq
= + q q L n dS. (1.7.54)
dt S Dt
The second and third terms inside the integrand express the eect of surface dilatation. Expressing
the material derivative, Dq/Dt, in terms of Eulerian derivatives and using the vector identity (A.6.11)
in Appendix A, to write
(q u) = q u ( q) + u q q L, (1.7.55)
we recast (1.7.54) into the form
dQ q
= + (q u) + u ( q) n dS. (1.7.56)
dt S t
If the vector eld q is solenoidal, q = 0, the third term inside the integral does not appear. The
Zorawski condition states that, for the ow rate Q to remain constant in time, the expression inside
the tall parentheses on the right-hand side of (1.7.56) must be zero.
Problems
Orthogonal coordinates
If the coordinates and are orthogonal, the metric tensor is diagonal, g = g = 0, h2s = g g ,
and hs = h h .
66 Introduction to Theoretical and Computational Fluid Dynamics
dX dn f d 2 + 2 f d d + f d 2 , (1.8.9)
where
X n 2X X n 2X
f = = n, f = = n, (1.8.10)
2 2
and
1 X n X n
2X X n X n
f = + = n= = . (1.8.11)
2
Equations (1.7.13). were used to derive the last two expressions for f . The rate of change of the
tensors g and f following a point particle in a surface can be computed from their denition using
the evolution equations discussed in Section 1.7 (Problem 1.8.1).
Orthogonal coordinates
Using equations (1.7.17), we nd that, if the coordinates and are orthogonal, the o-diagonal
components of f vanish, f = f = 0.
1.8.3 Curvatures
The normal curvature of a surface at a point in the direction of the axis, denoted by K , is dened
as the curvature of the trace of the surface in a plane that contains the normal vector, n, and the
tangential unit vector, t , drawn with the heavy line in Figure 1.8.1. In practice, the trace of the
surface on the normal plane may not be available and the normal curvature must be extracted from
the curvature of another surface line that is tangential to t at a point, denoted by , such as the
dashed line in Figure 1.8.1. If l is the arc length along such a line, then, by denition,
t
n , (1.8.12)
l
where n is the unit vector normal to the dashed line in Figure 1.8.1.
Meusniers theorem
Meusniers theorem states the the normal curvature in the direction of the axis is given by
n t f f
K t = n = n n = = 2. (1.8.13)
l l g h
1.8 Dierential geometry of surfaces 67
n
n
Figure 1.8.1 The normal curvature of a surface at a point in the direction of the axis, denoted
by K , is dened as the curvature of the trace of the surface in a plane that contains the normal
vector, n, and tangential vector, t .
A similar equation can be written for the axis. To generalize these expressions, we consider a
tangent vector constructed as a linear combination of the two tangent vectors corresponding to the
surface coordinates, and ,
X X
= + , (1.8.14)
where is a free, positive or negative parameter. The normal curvature in the direction of is
given by
t n f + 2f + f 2
K = n n = n= t = , (1.8.15)
l l g + 2g + g 2
where t = /| | is a tangent unit vector and l is the corresponding arc length. Equation (1.8.13)
arises for = 0, and its counterpart for the axis arises in the limit as .
Mean curvature
Using (1.8.14), we nd that two tangent vectors corresponding to 1 and 2 are perpendicular if
they satisfy the relation
g + (1 + 2 ) g + 1 2 g = 0. (1.8.16)
Now using (1.8.15), we nd that the mean value of the directional normal curvatures in any two
perpendicular planes is independent of the plane orientation. Motivated by this observation, we
68 Introduction to Theoretical and Computational Fluid Dynamics
1 h h n n
m = (t n) (t n) . (1.8.19)
2 hs l l
Principal curvatures
The maximum and minimum directional normal curvatures, K , over all possible values of , are
called the principal curvatures. The corresponding values of are found by setting K / = 0 and
using the last expression in (1.8.15) to obtain a quadratic equation,
(f g g f ) 2 + (f g g f ) + f g g f = 0. (1.8.20)
The sum and the product of the two roots are given by
f g g f f g g f
1 + 2 = , 1 2 = . (1.8.21)
f g g f f g g f
Direct substitution shows that these equations satisfy (1.8.16), and this demonstrates that, if Kmax
is the maximum principal curvature corresponding to a particular orientation, then Kmin will be the
minimum principal curvature corresponding to the perpendicular orientation. The mean curvature
of the surface is
1
m = (Kmax + Kmin ). (1.8.22)
2
f f f + f 2
Kmax = , Kmin = , K = . (1.8.23)
g g g + g 2
Now let be the angle subtended between the direction corresponding to and the principal direc-
tion of the maximum curvature, varying in the range [0, 12 ]. Using the geometrical interpretation
of the inner vector product, we nd that
X X
X X X X
+ = + cos (1.8.24)
1.8 Dierential geometry of surfaces 69
1
3
0
4
2
Figure 1.8.2 Five points are arranged along two curvilinear axes, (, ) in a surface. If the axes are
orthogonal, knowledge of the position of these points is sucient for computing the mean curvature
of the surface at the intersection.
and
X X
X X X X
+ = + sin , (1.8.25)
yielding
Combining these expressions with (1.8.23), we derive Eulers theorem for the curvature, stating that
the normal curvature in an arbitrary direction is related to the principal curvatures by
Numerical methods
In numerical practice, the curvature of a surface can be computed by tracing two curvilinear axes
with a set of marker points whose position is described as X(, ), and then constructing parametric
representations for the and coordinate lines using methods of curve tting and function interpo-
lation, as discussed in Section B.4, Appendix B. The partial derivatives of X with respect to and
can be computed by numerical dierentiation, as discussed in Section B.5, Appendix B.
Assume that the positions of ve points along the and axes are given, Xi for i = 04,
as shown in Figure 1.8.2. Using centered dierences to approximate the tangential vectors at the
location of the central point, X0 , we obtain
X2 X 1 X4 X 3
t , t . (1.8.28)
|X2 X1 | |X4 X3 |
The numerical accuracy is of rst order with respect to and . If the points are evenly spaced
with respect to and , the accuracy becomes of second order with respect to and .
70 Introduction to Theoretical and Computational Fluid Dynamics
z (x)
n
n
Figure 1.8.3 Explicit description of a surface as z = q(x, y), showing the normal unit vector, n, and
the unit vector normal to the intersection of the surface with a plane that is normal to the y axis,
denoted by n(x) .
2 0
. (1.8.30)
l |X2 X1 | |X2 X0 | |X0 X1 |
The normal unit vector, n , follows by dividing the right-hand side of (1.8.30) by its length, and the
directional curvature is extracted from the formula
t
= n . (1.8.31)
l
Similar equations are used to compute n and . If the curvilinear axes are orthogonal, the results
can be substituted into (1.8.13) and then into (1.8.18) to obtain an approximation to the mean
curvature at the central point, X0 . In that case, but not more generally, knowledge of the position
of ve points is sucient for estimating the mean curvature.
Description as z = q(x, y)
Assume that a surface is described explicitly by the function z = q(x, y), as shown in Figure 1.8.3.
The unit vector normal to the intersection of the surface with a plane that is perpendicular to the
y axis is
1
n(x) = (qx ex + ez ), (1.8.32)
(1 + qx2 )1/2
1.8 Dierential geometry of surfaces 71
(a) (b)
y y t
n
>0
n x0
t
R
l <0
x x
Figure 1.8.4 (a) Illustration of a line in the xy plane showing the tangential and normal unit vectors,
t and n, and the sign of the curvature, . (b) Illustration of plane polar coordinates, (r, ), in the
xy plane centered at a point, x0 .
where qx = q/x, qy = q/y, and ex , ey , and ez are unit vectors along the x, y, and z axes.
Applying Meusnier formula (1.8.13), we nd that the corresponding normal curvature is
1 + qx2
1/2
Kx = x n(x) n = x , (1.8.34)
1 + qx2 + qy2
where x is the curvature of the planar intersection. We observe that Kx = x only when qy = 0,
corresponding to a cylindrical surface.
A similar expression can be written for Ky . The mean curvature of the surface is not necessarily
equal to the average of Kx and Ky .
dn
=t . (1.8.37)
dl
If the line is tangential to the x axis at a point, t = 1, 0 and = dnx /dx.
where ex and ey are unit vectors for the x and y axes, and a prime denotes a derivative with respect
to x. The curvature of the line is
dt dx dt 1
= n = n = (q ex + ey ) (ex + q ey ) . (1.8.39)
dl dl dx (1 + q 2 )3/2
q 1 1
q
= 2 3/2
=
=
. (1.8.40)
(1 + q ) q 1 + q 2 1 + q 2
The slope angle, , is dened by the equation tan = q , where /2 < < /2, as shown in
Figure 1.8.4(a). We note that 1 + q 2 = 1/ cos2 , and obtain
1 d cos
= . (1.8.41)
q dx
R er + R e R er + R e dl
n= t= = R2 + R2 , (1.8.42)
R2 + R2 R2 + R2 d
where er and e are unit vectors in the radial and polar directions, and a prime denotes a derivative
with respect to . The curvature of the line is
dt d dt 1
= n = n = 2 (R er + R e ) (R er + R e ) . (1.8.43)
dl dl d (R + R2 )3/2
1.8 Dierential geometry of surfaces 73
dy Y q
q = = , q = , (1.8.45)
dx X X
and carrying out the dierentiations, we obtain the formula
X Y Y X
= , (1.8.46)
(X2 + Y2 )3/2
where a subscript denotes a derivative with respect to . Formulas (1.8.40) arise by setting = x.
If increases in a direction that is opposite to that of t, a minus sign is introduced in front of the
fraction on the right-hand side of (1.8.46).
Often in practice, the functions X() and Y () are reconstructed numerically from data de-
scribing the location of marker points along the line using, for example, cubic spline interpolation,
as discussed in Appendix B. The derivatives of these functions are computed by numerical dier-
entiation, as discussed in Sections B.4 and B.5, Appendix B. In the simplest implementation, the
interpolating variable, , is identied with the arc length of the polygonal line connecting successive
marker points.
Dierentiating the expression Xl2 + Yl2 = 1 with respect to l and using the resulting equation to
simplify (1.8.47), we obtain
Xll Yll
= = . (1.8.48)
Yl Xl
These expressions allow us to reconstruct a curve in terms of the curvature, (l), by integrating the
ordinary dierential equations Xll = Yl and Yll = Xl subject to suitable boundary conditions.
74 Introduction to Theoretical and Computational Fluid Dynamics
This property allows us to write a general expression for the mean curvature with reference
to an arbitrary system of Cartesian coordinates whose axes are not necessarily tangential or normal
to the surface at a point,
nx ny nz
2m n = + + . (1.8.53)
x y z
If a surface is described implicitly by the equation F (x, y, z) = 0, the unit vector normal to
the surface is given by
1
n= F, (1.8.54)
|F |
and the mean curvature is given by
1
1 1
2m = F = 2 F F (F ) F, (1.8.55)
|F | |F | |F |3
where F is the matrix of second derivatives.
1.8 Dierential geometry of surfaces 75
2m = . (1.8.57)
x (1 + qx2 + qy2 )1/2 y (1 + qx2 + qy2 )1/2
Carrying out the dierentiations, we obtain
(1 + qy2 ) qxx 2qx qy qxy + (1 + qx2 ) qyy
2m = . (1.8.58)
(1 + qx2 + qy2 )3/2
This formula also arises from (1.8.55) by setting F (x, y, z) = z f (x, y). For a nearly at surface,
2m qxx qyy .
(a) (b)
n
A
t t
n
B D
D
n
t
t
C
C
b t
Figure 1.8.5 (a) A contour integral of the binormal vector, b, can be used to dene the mean
curvature. (b) The surface-average value of the mean curvature over an interfacial patch can be
computed from the position of four marker points.
for any arbitrary dierentiable vector function F, where C is a closed contour bounding a surface,
D, n is the unit vector normal to D pointing toward a designated side, t is the unit vector tangential
to C, and b t n is the binormal vector, as shown in Figure 1.8.5(a). Applying this identity with
F = n, we obtain
n t dl = n n (n) n dS. (1.8.65)
C D
If the surface is small, we may assume that the mean curvature and normal vector are constant and
obtain an approximation for the mean curvature,
1
m n n t dl, (1.8.67)
2S C
where S is the surface area of D. The counterpart of this expression for a section of a two-
dimensional surface with arc length l is
1
m n t, (1.8.68)
2l
where t is the dierence in the unit tangent vectors between the last and rst segment points.
1.8 Dierential geometry of surfaces 77
(a) (b)
x
y
n n
R2
x R2
Figure 1.8.6 The second principal curvature of an axisymmetric surface, 2 = 1/R2 , is the curvature
of the dashed line representing the trace of the surface in a plane that is normal to the surface and
also normal to a plane of constant azimuthal angle, .
Numerical methods
To illustrate the practical application of the method, we consider a rectangular surface element
dened by four points, AD, that lie at the intersections of a pair of lines and a pair of lines
representing surface curvilinear coordinates, as shown in Figure 1.8.5(b). Applying the trapezoidal
rule to approximate the contour integral in (1.8.67), we obtain
1 X
X
m n n +n (B A )
4S A B
X
X
+ n +n (C B )
B C
X
X
+ n +n (D C )
C D
X
X
+ n +n (A D ) . (1.8.69)
D A
The normal and tangent unit vectors on the right-hand side can be computed from the position of
surface marker points using standard methods of numerical dierentiation, as discussed in Section
B.5, Appendix B.
Description as = w(x)
The shape of an axisymmetric surface in a meridional plane of constant azimuthal angle can be
described by the function = w(x), as shown in Figure 1.8.6(a). The normal unit vector is
1
n= (e w ex ), (1.8.70)
1 + w2
where a prime denotes a derivative with respect to x. The mean curvature is given by the divergence
of the normal vector,
nx 1 (n ) w
1
2m = + = + . (1.8.71)
x 1 + w2 1 + w2
Carrying out the dierentiations, we obtain
w 1 1
2m = + , (1.8.72)
(1 + w2 )3/2 w 1 + w 2
which can be rearranged into the expression
1 1 + w2 ww
2m = . (1.8.73)
w (1 + w2 )3/2
The rst term on the right-hand side of (1.8.72) is the principal curvature in an azimuthal plane,
w
1 = . (1.8.74)
(1 + w2 )3/2
The second term on the right-hand side of (1.8.72) is the second principal curvature,
1
2 = , R2 = 1 + w2 = = , (1.8.75)
R2 sin n
where R2 is the second principal radius of curvature and the angle is dened in Figure 1.8.6(a).
We have found that the second principal radius of curvature, R2 , is the signed distance of the point
where the curvature is evaluated from the intersection of the extension of the normal vector with
the x axis. If n is negative, R2 is also negative, and vice versa.
Description as x = q()
The shape of an axisymmetric surface in an azimuthal plane can be described by the function
x = q(), as shown in Figure 1.8.6(b). The normal unit vector is
1
n=
(ex q e ), (1.8.76)
1+ q 2
where a prime denotes a derivative with respect to . The mean curvature is the divergence of the
normal vector,
nx 1 (n ) 1 q
2m = + =
. (1.8.77)
x 1 + q 2
1.8 Dierential geometry of surfaces 79
The rst term on the right-hand side of (1.8.78) is the principal curvature in an azimuthal
plane,
q 1 1
1 = =
. (1.8.79)
(1 + q 2 )3/2 q 1 + q 2
The second term is the second principal curvature,
1
2 = , R2 = 1 + q 2 = = , (1.8.80)
R2 q sin n
where R2 is the signed second principal radius of curvature, and the angle is dened in Figure
1.8.6(b).
Problems
Computer Problems
spheroid is specied in parametric form as X = a cos , Y = b sin cos , and Z = b sin sin , where
0 is a dimensionless parameter and is the azimuthal angle.
(a) Derive an expression for the mean curvature. Prepare a graph of the scaled mean curvature
am against for aspect ratios b/a = 0.1 (prolate spheroid), 1.0 (sphere), and 10 (oblate spheroid).
Verify that, when a = b, the mean curvature takes the uniform value 1/a = 1/b for a sphere.
(b) Repeat (a) using (1.8.28)(1.8.31) with suciently small increments and to approximate
the curvature.
Write a program that uses cubic spline interpolation with periodic end conditions to compute the
Cartesian coordinates and curvature of a point along the loop as a function of a suitably chosen
parameter that increases monotonically from 0 to 1 along the loop, from start to nish (Section
B.5, Appendix B). One plausible choice for is the polygonal arc length, dened as the length of
the polygonal line connecting successive nodes, divided by the value at the last node. The program
should return the coordinates (x, y) corresponding to a specied value of in the interval [0, 1], where
= 0 and 1 correspond to x = 1.1 and y = 1.0. Generate a table of 32 values of the quadruplet
(, x, y, ) at 32 evenly spaced intervals of between 0 and 1.
y n
t
Fluid 1
q
l
Fluid 2
0
x
Figure 1.9.1 Point particles along a two-dimensional interface are identied by a parameter, .
requires that
n
= q(0 ) q(), (1.9.2)
t
where the time derivative is taken keeping xed. Substituting the expression for n from the last
integral of (1.9.1) and transferring the derivative inside the integral as a material derivative, D/Dt,
we obtain
D l
Now we take the limit as tends to 0 and derive the dierential equation
D l
q
= . (1.9.4)
Dt
Expanding the material derivative on the left-hand side, we obtain
D l D l
q
+ = . (1.9.5)
Dt Dt
To compute the second material derivative on the left-hand side, we use the Pythagorean theorem
to write
l X
2 Y
2 1/2
= + , (1.9.6)
and then compute
D l
1 1 X D X
Y D Y
= 2 +2 . (1.9.7)
Dt 2 l/ Dt Dt
Interchanging the order of the material and the derivative inside the square brackets on the right-
hand side and setting DX/Dt = ux and DY /Dt = uy , we nd that
D l
1 X ux Y uy
1 X u
= + = . (1.9.8)
Dt l/ l/
82 Introduction to Theoretical and Computational Fluid Dynamics
Rearranging, we obtain
D l
l X u l u
= = t , (1.9.9)
Dt l l l
where t = X/l is the tangent unit vector shown in Figure 1.9.1. Substituting the nal expression
into (1.9.5), we obtain
D u q
+ t = , (1.9.10)
Dt l l
which is the targeted evolution equation for the surfactant concentration.
D
= + ux , (1.9.11)
Dt t t x
where the last expression involves Eulerian derivatives with respect to x and t. Similar expressions
can be written with reference to curvilinear axes.
Ficks law
The diusive ux of a surfactant along an interface can be described by Ficks law,
q = Ds , (1.9.12)
l
where Ds is the surfactant surface diusivity with unit of length squared divided by time. In practice,
the surfactant diusivity is typically small. Substituting this expression into (1.9.10), we derive a
convectiondiusion equation,
D u
+ t = Ds . (1.9.13)
Dt l l l
u = ut t + un n, (1.9.14)
where n is the normal unit vector, as shown in Figure 1.9.1. The tangential and normal velocities
are ut = ut and un = un. Noting that tn = 0, tt = 1, and nn = 1, and using the Frenet-Serret
relations dt/dl = n and dn/dl = t, we compute
u ut n ut
t = + un t = + un , (1.9.15)
l l l l
1.9 Interfacial surfactant transport 83
where is the interfacial curvature. Substituting this expression into (1.9.13), we obtain
D ut
+ + un = Ds . (1.9.16)
Dt l l l
The rst term inside the parentheses on the left-hand side expresses the eect of interfacial stretching,
and the second expresses the eect of interfacial expansion.
Stretching of a at interface
As an application, we consider a at interface along the x axis that is stretched uniformly under the
inuence of a tangential velocity eld, ux . Identifying the arc length l with x and setting = 0, we
nd that the transport equation (1.9.16) reduces to
D ux
+ = Ds . (1.9.17)
Dt x x x
The material derivative can be resolved into Eulerian derivatives with respect to x and t, yielding
ux
(ux )
+ ux + = Ds or + = Ds . (1.9.18)
t x x x x t x x x
This equation could have been derived directly by performing a surfactant molecular balance over a
stationary dierential control volume along the x axis, taking into consideration the convective and
diusive ux contributions.
d
+ = 0. (1.9.19)
dt
The solution reveals that, when is positive, the surfactant concentration decreases exponentially
due to dilution, (t) = (t = 0) exp(t).
D ur Ds 2
+ = 2 . (1.9.20)
Dt a a 2
Resolving the material derivative into Eulerian derivatives with respect to r and t, we obtain
ur Ds 2
+ = 2 . (1.9.21)
t a a 2
84 Introduction to Theoretical and Computational Fluid Dynamics
If the surfactant concentration is uniform at the initial instant, it will remain uniform at any time,
governed by the linear equation
d ur
+ = 0. (1.9.22)
dt a
In the case of expansion, ur > 0, the surfactant concentration decreases due to dilution, d/dt <
0. In the case of contraction, ur < 0, the surfactant concentration increases due to compaction,
d/dt > 0.
Interfacial markers
The material derivative expresses the rate of change of the surfactant concentration following the
motion of point particles residing inside an interface. In numerical practice, it may be expedient to
follow the motion of interfacial marker points that move with the normal component of the uid
velocity and with an arbitrary tangential velocity, vt . If vt = 0, the marker points move normal to
the interface at any instant. The velocity of a marker point is
v = un n + vt t. (1.9.23)
By denition,
D d
= + (ut vt ) , (1.9.24)
Dt dt l
where d/dt is the rate of change of the surfactant concentration following a marker point. Substi-
tuting this expression into the transport equation (1.9.16), we obtain
d ut
+ (ut vt ) + ( + un ) = Ds , (1.9.25)
dt l l l l
which can be restated as
d (ut )
+ vt + un = Ds . (1.9.26)
dt l l l l
The second term on the left-hand side represents the interfacial convective ux.
To derive an evolution equation for the surface surfactant concentration, we introduce cylin-
drical polar coordinates, (x, , ), and express the number of surfactant molecules inside a ring-like
material section of the interface conned between 0 and as
l(,t)
l
n(, t) = 2 ( , t) ( ) dl( ) = 2 ( , t) ( ) d . (1.9.27)
l(0 ,t) 0
1.9 Interfacial surfactant transport 85
y
n
Fluid 1
q
l t
Fluid 2
0
x
Figure 1.9.2 Point particles along the trace of an axisymmetric interface in an azimuthal plane are
identied by a parameter . The angle is subtended between the x axis and the straight line
dened by the extension of the normal vector.
Conservation of the total number of surfactant molecules inside the test section requires that
n
= 20 q(0 ) 2q(), (1.9.28)
t
where q is the ux of surfactant molecules along the interface by diusion, and the time derivative
is taken keeping xed. The counterpart of the balance equation (1.9.4) is
D l
(q)
(, t) (, t) = , (1.9.29)
Dt
D u u 1 (q)
+ (t + )= . (1.9.30)
Dt l l
In deriving this equation, we have set D/Dt = u , where is the distance of a point particle from
the axis of revolution.
where is the curvature of the interface in an azimuthal plane. Next, we introduce the angle
subtended between the x axis and the normal vector to the interface at a point, as shown in Figure
1.9.2. Substituting u = un sin ut cos , we obtain
D u cos sin
1 (q)
t
+ ut + ( + ) un = . (1.9.32)
Dt l l
The sum of the two terms inside the innermost parentheses on the left-hand side is twice the mean
curvature of the interface, 2m . The rst two terms inside the large parentheses can be consolidated
to yield the nal form
D 1 (u )
1 (q)
t
+ + 2 m un = . (1.9.33)
Dt l l
The rst term inside the large parentheses on the left-hand side expresses the eect of axisymmetric
interfacial stretching, and the second expresses the eect of interfacial expansion.
Marker points
An evolution equation for the rate of change of the surface surfactant concentration following in-
terfacial marker points that are not necessarily point particles can be derived working as in Section
1.9.1 for two-dimensional ow. The result is
d 1 (u )
1 (q)
t
+ (ut vt ) + + 2 m un = , (1.9.34)
dt dl l l
which can be restated as
d 1 (ut ) 1 (q)
+ vt + 2 m un = . (1.9.35)
dt l l l
When vt = 0, the marker points move normal to the interface and the third term on the right-hand
side does not appear.
where hs is the surface metric and is the xed support of the patch in the (, ) plane. A mass
balance requires that
dn
= b q dl, (1.9.37)
dt C
where C is the edge of the patch, b = t n is the binormal vector, t is the unit tangent vector, n
is the unit vector normal to the patch, and q is the tangential diusive surface ux. Substituting
1.9 Interfacial surfactant transport 87
the last integral in (1.9.36) in place of n into the left-hand side of (1.9.37), and transferring the time
derivative inside the integral as a material derivative, we obtain
D(hs )
d d = b q dS. (1.9.38)
Dt
Next, we expand the material derivative inside the integral, use (1.7.32) to evaluate the rate of
change of the surface metric, and apply the divergence theorem to convert the line integral on the
right-hand side into a surface integral, obtaining
D
+ s u dS = s q dS, (1.9.39)
P atch Dt P atch
where s (I n n) is the surface gradient and s u is the surface divergence of the velocity.
Because the size of the material patch is arbitrary, the integrands must balance to zero, yielding the
transport equation
D
+ s u = s q. (1.9.40)
Dt
D C
+ s us + n s un + un s n = s q. (1.9.41) t
Dt b
The inner product, n s un , is identically zero due to the orthogo- Illustration of an interfacial
nality of the surface gradient and normal vector. Setting the surface patch occupied by surfactant.
divergence s n equal to twice the mean curvature of the interface,
2m , we obtain the transport equation
D
+ s us + 2m un = s q. (1.9.42)
Dt
The rst term inside the parentheses on the left-hand side expresses the eect of interfacial stretching,
and the second expresses the eect of interfacial expansion. In the case of a stationary interface,
un = 0, only the tangential velocity eld aects the surfactant concentration.
Ficks law
The tangential diusive ux vector can be described by Ficks law,
q = Ds s , (1.9.43)
88 Introduction to Theoretical and Computational Fluid Dynamics
where Ds is the surfactant surface diusivity. Substituting this expression into (1.9.40), we derive a
convectiondiusion equation,
D
+ s u = s (Ds s ). (1.9.44)
Dt
When the diusivity is constant, the diusion term becomes Ds 2s , where 2s s s is the
surface Laplacian operator, sometimes also called the LaplaceBeltrami operator. Expression for this
operator in orthogonal and nonorthogonal curvilinear coordinates can be derived using the formula
presented in Appendix A.
Interfacial markers
In numerical practice, it may be expedient to follow the motion of interfacial marker points moving
with the normal component of the uid velocity, un n, and with an arbitrary tangential velocity, vs .
The marker-point velocity is
v = un n + vs . (1.9.45)
When vs = 0, a marker point moves with the uid velocity normal to the interface alone. Conse-
quently, if the interface is stationary, the marker point is also stationary. When vs = u un n, the
marker points are material point particles moving with the uid velocity. The rate of change of the
surfactant concentration following the marker points is related to the material derivative by
d D
= (us vs ) . (1.9.46)
dt Dt
Combining this expression with (1.9.42), we obtain
d
+ (us vs ) + (s us + 2m un ) = s q. (1.9.47)
dt
Rearranging, we derive the alternative form
d
+ s (us ) vs s + 2m un = s q. (1.9.48)
dt
If the point particles move normal to the interface, the third term on the left-hand side expressing
a convective contribution does not appear.
Problems
z
n
z
z=f(x, y, t)
y y
Figure 1.10.1 The shape of a material surface can be described in Eulerian form in global Cartesian
coordinates as z = f (x, y).
A function that describes the shape of a material line or surface in Eulerian form satises an
evolution equation that emerges by requiring that the motion of point particles on either side of the
line or surface is consistent with the stationary or evolving shape of the line or surface described by
the Eulerian form. This evolution equation can be regarded as a kinematic compatibility condition,
analogous to a kinematic boundary condition specifying the normal or tangential component of the
velocity on a rigid or deformable boundary.
f (x + ux t, y + uy t, t + t) = f (x, y, t) + uz t. (1.10.1)
90 Introduction to Theoretical and Computational Fluid Dynamics
Expanding the left-hand side of (1.10.1) in a Taylor series with respect to the rst two of its arguments
about the triplet (x, y, t), we nd that
f f
f (x, y, t + t) + ux t + uy t = f (x, y) + uz t. (1.10.2)
x y
Now dividing each term by t, taking the limit as t becomes innitesimal, and rearranging, we
derive the targeted evolution equation
f f f
+ ux + uy uz = 0, (1.10.3)
t x y
expressing kinematic compatibility. If the shape of the material surface is stationary, f /t = 0,
the remaining three terms on the left-hand side of (1.10.3) must balance to zero.
Level-set formulation
A material surface can be described implicitly by the equation
The function F (x, y, z, t) is negative above the material surface, positive below the material surface,
and zero over the material surface. Equation (1.10.3) can be restated in terms of the material
derivative of F as
DF F
+ u F = 0. (1.10.5)
Dt t
Introducing the upward normal unit vector, n = F/|F |, we obtain
1 F
= un , (1.10.6)
|F | t
where un u n is the normal velocity component. This form shows that un must be continuous
across a material surface. The tangential velocity component may undergo a discontinuity that is
inconsequential in the context of kinematics.
The implicit function theorem allows us to generalize these results and state that, if a material
surface is described in a certain parametric form as
F (x, t) = c, (1.10.7)
where c is a constant, then the evolution of the level-set function F (x, t) is governed by equation
(1.10.5) or (1.10.6). This is another way of saying that the scalar eld represented by the function
F (x, t) is convected by the ow. Stated dierently, a material surface is a convected level-set of the
function F (x, t) determined by the constant c.
F (r, , t) f (, t) r. (1.10.8)
1.10 Eulerian description of material lines and surfaces 91
vt = (I n n) u = n u n, (1.10.11)
where is an arbitrary time-dependent coecient allowed to vary over the material surface. The
projection operator I n n extracts the tangential component of a vector that it multiplies.
Equation (1.10.10) reveals that it is kinematically consistent to allow imaginary interfacial particles
to move with their own velocity,
v = un n + (I n n) u, (1.10.12)
that can be dierent than the uid velocity. When = 1, the interfacial particles represent marker
points devoid of physical interpretation.
It will be necessary to introduce the arc length along the material line, l , increasing in the
direction of the parameter . A point particle in an evolving material line moves tangentially and
normal to the reference line. If at time t the point particle is at a position corresponding to l , then
at time t + t the same point particle will be at a position corresponding to l + l , where
R 1
l = u t = u t, (1.10.14)
R+ 1 + R
92 Introduction to Theoretical and Computational Fluid Dynamics
R
un
x
xR u
R
n
R
R
t
l
Figure 1.10.2 An evolving material line can be described in Eulerian form with respect to a stationary
reference line in terms of the normal displacement, . The radius of curvature of the reference line
is denoted by R.
u is the velocity component tangential to the reference line evaluated at the position of the material
line, R is the signed radius of curvature of the reference line, and R = 1/R is the curvature of the
reference line. For the conguration illustrated in Figure 1.10.2, R and R are positive. Expression
(1.10.14) arises by approximating the line locally with a circular arc. Using the second FrenetSerret
relation, nR /l = R tR , we nd that
(nR ) xR (nR )
R = tR = , (1.10.15)
l l l
where tR is the unit vector tangent to the reference line. Now exercising geometrical reasoning, we
nd that
(l + l , t + t) = (l , t) + uR
n t, (1.10.16)
where uR
n is the velocity component normal to the reference line evaluated at the position of the
material line, as illustrated in Figure 1.10.2. Expanding the left-hand side in a Taylor series and
linearizing with respect to t, we derive the compatibility condition
u
+ uR
n = 0. (1.10.17)
t 1 + R l
In the case of a rectilinear reference line, R = 0, we recover the compatibility condition derived
earlier in this section.
Dierentiating expression (1.10.13) with respect to l and using the second FrenetSerret
relation, we obtain a vector that is tangential to the material line,
x R
= (1 + R ) tR + n . (1.10.18)
This expression shows that an arbitrary tangential velocity with components
v = (1 + R ), vnR = (1.10.19)
can be included on the left-hand side of (1.10.17), where () is an arbitrary coecient.
1.10 Eulerian description of material lines and surfaces 93
x(, , t) = xR (, ) + (, , t) nR (, ), (1.10.20)
where the function xR (, ) describes a steady reference surface, and comprise a system of
two surface curvilinear coordinates in the reference surface, nR (, ) is the unit vector normal to
the reference surface, and (, , t) is the normal displacement from the reference surface. This
representation is appropriate when the material surface is suciently close to the reference surface
so that (, , t) is a single-valued and continuous shape function.
Working as in Section 1.10.3 for a material line, we nd that, in the case of orthogonal surface
curvilinear coordinates, (, ), the shape function evolves according to the equation
u u
+ + uR
n = 0, (1.10.21)
t 1 + K l 1 + K l
where l and l is the arc length in the reference surface along the two surface curvilinear coordinates,
K and K are the corresponding principal curvatures, u and u are velocity components tangential
to the reference surface evaluated at the material surface, and uR n is the velocity component normal
to the reference surface. If the reference surface is a sphere of radius a, we may identify with the
meridional angle, , with the azimuthal angle, , and set l = a, l = (a sin ) , = 1/a, and
= 1/a.
Equation (1.10.21) can be expressed in coordinate-free vector notation in terms of the surface
gradient of the reference surface, R
s = P , and corresponding surface velocity, us = P u, as
R R R
s s F un = 0,
+ uR R R (1.10.22)
t
where PR = I nR nR is the tangential projection operator and I is the identity matrix. The
surface function, F(l, ), satises the dierential equation
F 1
= , (1.10.23)
l 1 + K l
where l is the arc length along the intersection of the material surface with a normal plane, corre-
sponding to the tangent unit vector t, K is the corresponding principal curvature, and is the angle
subtended between the intersection and a specied tangential direction. For example, = 0 may
correspond to the direction of maximum principal curvature. The pair (l, ) constitutes a system of
plane polar coordinates tangential to the material surface at a point of interest. By analogy with
(1.10.15), we write
(nR ) xR (nR )
K = t = , (1.10.24)
l l l
which demonstrates that the right-hand side of (1.10.23) is independent of .
94 Introduction to Theoretical and Computational Fluid Dynamics
Problems
To describe a streamline, we introduce a variable that increases monotonically along the streamline
in the direction of the velocity vector. One acceptable choice for is the time it takes for a point
particle to move along the streamline from a specied initial position as it is convected by the
frozen instantaneous velocity eld. If the ow is steady, is the real time, t. The streamline is
then described by an autonomous system of dierential equations, with no time dependence on the
right-hand side,
dx
= u(x, t = t0 ), (1.11.1)
d
dx dy dz
= = = d, (1.11.2)
ux uy uz
conrms that an innitesimal vector that is tangential to a streamline is parallel to the velocity.
1.11 Streamlines, stream tubes, path lines, and streak lines 95
Ideally, the time step should be adjusted according to both the magnitude of the velocity
and local curvature of the computed streamline, so that sharply turning streamlines are described
with sucient accuracy and the computation does not stall at regions of slow ow with simple
structure. However, implementing these conditions increases the complexity of the computer code.
Unless a high level of spatial resolution is required, the method of constant travel distance should
be employed.
for i = 1, 2, where n is the unit vector normal to D1 or D2 . Using the divergence theorem, we
compute
Q2 = Q1 u n dS + u dV, (1.11.4)
St Vt
where St is the surface of the stream tube extending between the two loops, and Vt is the volume
enclosed by the surfaces D1 , D2 , and St . Since the velocity is tangential to the stream tube and
therefore perpendicular to the normal vector on St , the surface integral over St is zero and (1.11.4)
simplies into
Q2 = Q 1 + u dV. (1.11.5)
Vt
96 Introduction to Theoretical and Computational Fluid Dynamics
n
St D2
n
Vt
n
D
1
Figure 1.11.1 Illustration of a stream tube and two closed loops wrapping around the stream tube.
This equation suggests that the volumetric ow rate may increase or decrease along a stream tube
according to whether the uid inside the stream tube undergoes expansion or contraction.
Incompressible uids
If the uid is incompressible, u = 0, the ow rate across any cross-section of a stream tube
is constant, Q1 = Q2 . Consequently, in the absence of singularities, a stream tube that carries a
nite amount of uid may not collapse into a nonsingular point where the uid velocity is nonzero.
If this occurred, the ow rate at the point of collapse would have to vanish, which contradicts the
assumption that the stream tube carries a nite amount of uid. A similar argument can be made to
show that a streamline, approximated as a stream tube with innitesimal cross-section, may not end
suddenly in a ow, but must meet another streamline or multiple streamlines at a stagnation point,
form a closed loop, extend to innity, or cross the boundaries of the ow. A third consequence of
(1.11.5) is that the distance between two adjacent streamlines in a two-dimensional incompressible
ow is inversely proportional to the local magnitude of the uid velocity. The faster the velocity,
the closer the streamlines.
FrenetSerret framework
As a preliminary, we introduce a system of orthogonal curvilinear coordinates constructed with
reference to a streamline. We begin by labeling point particles along the streamline by the arc
length, l, and observe that the unit vector t = dx/dl is tangential to the streamline and therefore
1.11 Streamlines, stream tubes, path lines, and streak lines 97
(a) (b)
n y
n y t
x
t
x R
l
b r
z
Figure 1.11.2 Illustration of a streamline and associated curvilinear axes constructed with reference
to the streamline in (a) three-dimensional, and (b) two-dimensional ow.
parallel to the velocity, as shown in Figure 1.11.2(a). The principal unit vector normal to the
streamline, n, is dened by the rst FrenetSerret relation
1 t
n= , (1.11.6)
l
where is the signed curvature of the streamline, as discussed in Section 1.6.3. The curvature of
the streamline shown in Figure 1.11.2(a) is positive, > 0. Next, we introduce the binormal unit
vector dened by the equation b = t n. The three unit vectors, t, n, and b, dene three mutually
orthogonal directions that can be used to construct a right-handed, orthogonal, curvilinear system
of axes, so that t = n b and n = b t, as discussed in Section 1.6.2. The second and third
FrenetSerret relations are
n b
= t + b, = n, (1.11.7)
l l
where is the torsion of the streamline.
uy u (u n) n uz u (u b) b
= n= u, = b= u. (1.11.8)
x l l l x l l l
Because the velocity is tangential to the streamline and thus perpendicular to the normal and
binormal vectors along the entire streamline, u n = 0, u b = 0, and the rst term after the second
equal sign of each equation in (1.11.8) is identically zero. Using the FrenetSerret formulas (1.11.7)
98 Introduction to Theoretical and Computational Fluid Dynamics
uy uz
= ux , = 0. (1.11.9)
x x
In Section 1.11.4, these expressions will be used to relate the vorticity to the structure of the velocity
eld around a streamline.
Vorticity in two-dimensional ow
Consider a two-dimensional ow in the xy plane, as illustrated in Figure 1.11.2(b). Using relations
(1.11.9), we nd that the z vorticity component at the origin of the local Cartesian axes is given by
uy ux
z = = ux n (u) t. (1.11.10)
x y
Introducing plane polar coordinates with origin at the center of curvature of a streamline at a point,
as shown in Figure 1.11.2(b), and noting that ux = u along the y axis pointing in the radial
direction, we obtain
u u
z = + , (1.11.11)
R r r=R
where R = 1/ is the radius of curvature of the streamline. This expression reveals that point
particles spin about the z axis due to the global motion of the uid associated with the curvature
of the streamline, but also due to velocity variations in the normal direction.
Point particles spin about the tangential vector due to the twisting of the streamline pattern,
which is possible only in a genuine three-dimensional ow.
Point particles spin about the normal vector due to velocity variations in the binormal direction,
which is also possible only in a genuine three-dimensional ow.
Point particles spin about the binormal vector due to velocity variations in the normal direction
and also due to the global uid motion of the uid associated with the curvature of the
streamline.
1.11 Streamlines, stream tubes, path lines, and streak lines 99
To derive the tangential vorticity component, we refer to Figure 1.11.2(a) and project the denition
of the vorticity, = u, onto the unit tangent vector, t. Rearranging the resulting expression
and introducing the normal and binormal vectors, we obtain
To derive the normal vorticity component, we write u = ut t and invoke once again the
denition of the vorticity, = u, to obtain
= (ut t) = ut t + ut t. (1.11.14)
n n = ut ( t) n + (ut t) n = ut ( t) (b t) + (t n) ut , (1.11.15)
and then
ut
n = ut b (t t) + b ut = ut b (t) t t t + . (1.11.16)
lb
Because the length of the tangent unit vector t is constant, (t) t = 12 (t t) = 0. Using (1.11.6),
we nd that b (t t) = b n = 0, and conclude that n = ut /lb . By denition, n = 0 in
two-dimensional or axisymmetric ow.
To derive the binormal vorticity component, we project (1.11.14) onto the binormal vector,
b, and work in a similar fashion to obtain the binormal vorticity components,
b b = ut ( t) b + (ut t) b = ut ( t) (t n) + (t b) ut , (1.11.17)
yielding
t ut ut
b = ut n (t t) n ut = ut n = ut . (1.11.18)
l ln ln
In the case of two-dimensional ow, the vorticity vector points in the binormal direction and (1.11.18)
reduces to (1.11.10).
dX
= u X(t), t , (1.11.19)
dt
100 Introduction to Theoretical and Computational Fluid Dynamics
where X is the position of the point particle along its path. Formal integration yields the position
of the point particle at time t,
t tt0
X(t; t0 ) = X0 (t0 ) + u(X( ; t0 ), ) d = X0 (t0 ) + u(X(t0 + ; t0 ), t0 + ) d, (1.11.20)
t0 0
which can be regarded as a parametric representation of the path line in time. To compute a path
line, we select an ejection location and time and integrate equation (1.11.19) in time using a standard
numerical method, such as a RungeKutta method discussed in Section B.8, Appendix B.
Streaklines
A streakline is the instantaneous chain of point particles that have been released from the same or
dierent locations at the same or dierent prior times in a ow. Streaklines can be produced in
the laboratory by ejecting a dye from a stationary or moving needle. Regarding the injection time,
t0 as a Lagrangian marker variable, we describe the shape of a streakline at a particular time t by
(1.11.20). When the point particles are injected at the same location, the term X0 (t0 ) after each
equal sign is constant.
Problems
Computer Problems
= 0. (1.12.1)
This property imposes restrictions on the structure of the vorticity eld, similar to those imposed
on the structure of the velocity eld for an incompressible uid.
The vortex tubes of a two-dimensional ow are cylindrical surfaces perpendicular to the plane
of the ow. The vortex lines of an axisymmetric ow with no swirling motion are concentric circles
and the vortex tubes form concentric axisymmetric surfaces. The vortex lines of an axisymmetric
ow with swirling motion are spiral lines.
102 Introduction to Theoretical and Computational Fluid Dynamics
D L2
L3
L
L1
Figure 1.12.1 Illustration of a vortex tube. The circulation around a loop that wraps around the tube
once is equal to the ow rate of the vorticity across a surface, D, bounded by the loop.
1.12.2 Circulation
The circulation around a closed loop residing in the domain of a ow, L, is dened as
C= u dX = u t dl, (1.12.2)
L L
where X is the position of point particles along the loop, l is the arc length along the loop, and
the tangent unit vector, t = dX/dl, is oriented in a specied direction. Using Stokes theorem, we
derive an alternative expression for the circulation,
C= ( u) n dS = n dS, (1.12.3)
D D
where D is an arbitrary surface bounded by the loop, as illustrated in Figure 1.12.1. The direction
of the normal unit vector, n, is chosen so that t and n constitute a right-handed system of axes.
Equation (1.12.3) states that the circulation around a loop is equal to the ow rate of the vorticity
across any surface that is bounded by the loop.
D1
n
D
B
D
2
n L
Figure 1.12.2 Illustration of the surface of an arbitrary body, DB , a loop in the ow, L, and two
surfaces bounded by the loop.
where the integration domain, Vt , is the volume of the vortex tube enclosed by D1 , D2 , and the
surface of the vortex tube. Since the vorticity eld is solenoidal, = 0, the right-hand side of
(1.12.5) is identically zero. We conclude that the circulation around any loop that wraps a vortex
tube once, denoted by and called the cyclic constant or strength of the vortex tube, is independent
of the location and shape of the loop around the vortex tube at any instant.
Similar arguments can be made to show that the circulation around a loop that lies on a
vortex tube but does not wrap around the vortex tube, such as the loop L3 shown in Figure 1.12.1,
is zero (Problem 1.12.1). The circulation around a closed loop that wraps m times around a vortex
tube is equal to m, where is the strength of the vortex tube.
where n is the normal unit vector oriented as shown in Figure 1.12.2. Because each integral on the
left-hand side of (1.12.6) is equal to the circulation around L, the integral on the right-hand side,
and therefore the ow of vorticity across DB , must necessarily vanish,
n dS = 0. (1.12.7)
DB
104 Introduction to Theoretical and Computational Fluid Dynamics
D dX 1
u = (dX L) u = dX (u u) = 0, (1.12.9)
L Dt L 2 L
where L = u is the velocity gradient tensor. Equation (1.12.10) identies the rate of change of
circulation around a material loop with the circulation of the acceleration eld, Du/Dt, around the
loop. If the acceleration eld is irrotational, Du/Dt can be expressed as the gradient of a potential
function, as discussed in Section 2.1, and the closed line integral is zero. The rate of change of
circulation then vanishes and the circulation around the loop is preserved during the motion.
An alternative evolution equation for the circulation can be derived in Eulerian form by
applying the general evolution equation (1.7.56) for q = . Recalling that the vorticity eld is
solenoidal, = 0, we nd that
dC
= + ( u) n dS, (1.12.11)
dt D t
Problems
where 3 is the three-dimensional delta function, t is the unit vector tangent to the line vortex, and
l is the arc length along the line vortex. The velocity eld induced by a line vortex will be discussed
in Section 2.11.
where is a tangential vector eld, called the strength of the vortex sheet. Since the vorticity eld
is solenoidal, = 0, the surface divergence of must vanish,
(I N N) = 0, (1.13.3)
106 Introduction to Theoretical and Computational Fluid Dynamics
(a) (b)
+
N
L P
n
n
D D
t e
Q t
E E
Vortex sheet +
Vortex sheet T
Figure 1.13.1 (a) Illustration of a three-dimensional vortex sheet. The loop L pierces the vortex sheet
at the points P and Q. (b) Closeup of the intersection between a surface, D, and a vortex sheet;
T is the intersection of D and E, and the unit vector n is normal to D.
where I is the identity matrix and N is the unit vector normal to the vortex sheet. To satisfy this
constraint, a vortex sheet must be a closed surface, terminate at the boundaries of the ow, or
extend to innity.
Using Stokes theorem, we nd that the circulation around a loop L that pierces a vortex
sheet at two points, P and Q, as shown in Figure 1.13.1(a), is given by
C= u t dl = n dS, (1.13.4)
L D
where D is an arbitrary surface bounded by L and n is the unit vector normal to D. Substituting
into the last integral the vorticity distribution (1.13.2), we obtain
The inner integral is over the arbitrary surface, D, and the outer integral is over the vortex sheet.
vortex sheet, E, and is perpendicular to t, as shown in Figure 1.13.1(b). If l is the arc length along
t and l is the arc length along e , then
is a dierential surface element over the vortex sheet, E. The dierential arc length dln in the
direction of the normal unit vector n corresponding to a given dl is
dln = n e dl . (1.13.8)
where is a small length. Using the properties of the three-dimensional delta function to simplify
the term inside the large parentheses, we nd that
n(x ) (x )
C=
dl(x ). (1.13.10)
T n(x ) e (x )
The strength of the vortex sheet, , lies in the plane containing t and e , which is perpendicular to
the plane containing e and n. Thus,
[(n e ) e ] = 0, (1.13.11)
which is equivalent to
(n e )(e ) = n (1.13.12)
Taking the limit as the loop, L, collapses onto the vortex sheet on both sides, while the point Q
tends to the point P , we nd that
(u+ u ) t = e , (1.13.14)
where t is the unit vector tangent to T . The superscripts plus and minus designate, respectively,
the velocity just above and below the vortex sheet.
Equation (1.13.14) reveals that the tangential component of the velocity undergoes a discon-
tinuity across a vortex sheet. Mass conservation requires that the normal component of the velocity
is continuous across the vortex sheet. Equation (1.13.14) then allows us to write
u+ u = N, (1.13.15)
108 Introduction to Theoretical and Computational Fluid Dynamics
where N is the unit vector normal to the vortex sheet, as shown in Figure 1.13.1(a). We conclude
that , N, and the dierence u+ u dene three mutually perpendicular directions, so that
= N (u+ u ). (1.13.16)
We have found that a vortex sheet represents a singular surface across which the tangential compo-
nent of the velocity changes from one value above to another value below. The dierence between
these two values is given by the right-hand side of (1.13.15) in terms of the strength of the vortex
sheet.
Principal velocity
The mean value of the velocity above and below a vortex sheet is called the principal velocity of the
vortex sheet, or more precisely, the principal value of the velocity of the vortex sheet, denoted by
1 +
uP V (u + u ). (1.13.17)
2
Combining (1.13.17) with (1.13.15), we obtain an expression for the velocity on either side of the
vortex sheet in terms of the strength of the vortex sheet and the principal velocity,
1
u = uP V + N. (1.13.18)
2
Given the strength of the vortex sheet, the principal velocity is much easier to evaluate than the
physical uid velocity on either side of the vortex sheet.
1.13.3 Two-dimensional ow
Next, we describe the structure of line vortices and vortex sheets in two-dimensional ow in the xy
plane. The vorticity is perpendicular to the xy plane, = z ez , and the vortex lines are innite
straight lines parallel to the z axis.
Point vortex
A rectilinear line vortex is called a point vortex. The location of a point vortex is identied by
its trace in the xy plane, x0 = (x0 , y0 ). The vorticity in the xy plane is described by the singular
distribution
z (x) = 2 (x x0 ), (1.13.19)
where 2 is the two-dimensional delta function and is the strength of the point vortex. More will
be said about point vortices in Chapter 11.
Vortex sheet
A two-dimensional vortex sheet is a cylindrical vortex sheet whose generators and strength, , are
oriented along the z axis, as illustrated in Figure 1.13.2 (a). We can set
= ez , (1.13.20)
1.13 Line vortices and vortex sheets 109
(a) (b)
y
y
+
+
_
A
B x
t
T
x
z
Figure 1.13.2 Illustration of (a) a two-dimensional vortex sheet with positive strength, , in the xy
plane and (b) an axisymmetric vortex sheet.
where is the strength of the vortex sheet and ez is the unit vector along the z axis. The vorticity
eld in the xy plane is represented by the singular distribution
z (x) = (x ) 2 (x x ) dl(x ), (1.13.21)
T
where 2 is the two-dimensional Dirac delta function and T is the trace of the vortex sheet in the
xy plane. Using (1.13.15), we nd that the discontinuity in the velocity across a two-dimensional
vortex sheet is given by
u+ u = t, (1.13.22)
where t is a unit vector tangent to T , as shown in Figure 1.13.2. In terms of the principal velocity
of the vortex sheet, uP V , the velocity at the upper and lower surface of the vortex sheet is given by
1 1
u+ = uP V + t, u = uP V t. (1.13.23)
2 2
When is positive, we obtain the local velocity prole shown in Figure 1.13.2(a).
The circulation around a loop that lies in the xy plane and pierces the vortex sheet at the
points A and B is given by
C= z (x) dA(x) = (x )2 (x x ) dl(x ) dA(x) = (x ) dl(x ), (1.13.24)
D D T TAB
where D is the area in the xy plane enclosed by the loop and TAB is the section of T between the
points A and B, as illustrated in Figure 1.13.2(a). Fixing the point A and regarding the circulation,
C, as a function of location of the point B along T , denoted by , we obtain
d
= . (1.13.25)
dl
110 Introduction to Theoretical and Computational Fluid Dynamics
Comparing (1.13.26) with (1.13.19) allows us to regard a cylindrical vortex sheet as a continuous
distribution of point vortices. More will be said about vortex sheets and their self-induced motion
in Chapter 11.
1.13.4 Axisymmetric ow
The vorticity vector in an axisymmetric ow without swirling motion points in the azimuthal direc-
tion and the vortex lines are concentric circles, as shown in Figure 1.13.2(b).
(x) = 2 (x x0 ), (1.13.27)
where 2 is the two-dimensional delta function in an azimuthal plane. More will be said about line
vortex rings and their self-induced motion in Chapter 11.
Vortex sheet
An axisymmetric vortex sheet can be identied by its trace in an azimuthal plane of constant
azimuthal angle, . The strength of the vortex sheet, , is oriented in the azimuthal direction, as
shown in Figure 1.13.2(b). The vorticity distribution in an azimuthal plane is given by (1.13.21),
where 2 is the two-dimensional delta function.
Problems
Describing a ow in terms of an auxiliary eld is motivated by two reasons. First, the number
of scalar ancillary elds necessary to describe the ow of an incompressible uid is less than the
dimensionality of the ow by one unit, and this allows for analytical and computational simpli-
cations. For example, we will see that a two-dimensional incompressible ow can be described in
terms of a single scalar ow, called the stream function. A three-dimensional incompressible ow
can be described in terms of a pair of scalar stream functions. The reduction in the number of scalar
functions with respect to the number of nonvanishing velocity components is explained by observing
that the latter may not be assigned independently, but must be coordinated so as to satisfy the con-
tinuity equation for incompressible uids, u = 0. Imposing additional constraints reduces further
the number of required scalar functions. Thus, a three-dimensional incompressible and irrotational
ow can be expressed in terms of a single scalar function, called the potential function.
In some cases, expressing the velocity eld in terms of an auxiliary eld allows us to gain
physical insights into how the uid parcel motion aects the global structure or evolution of a ow.
For example, representing the velocity eld in terms of the vorticity and rate of expansion illustrates
the eect of spinning and expansion or contraction of small uid parcels on the overall uid motion.
The auxiliary elds discussed in this chapter are introduced with reference to the kinematic
structure of a ow discussed in Chapter 1. Another class of auxiliary elds are dened with reference
to the dynamics of a ow and, in particular, with respect to the stresses developing in the uid as
a result of the motion. Examples of these dynamical elds will be discussed in Chapter 6.
111
112 Introduction to Theoretical and Computational Fluid Dynamics
u = (2.1.1)
is irrotational, where is a scalar function called the potential function or the scalar velocity
potential. We conclude that a potential ow is also an irrotational ow. Families of irrotational
ows can be produced by making dierent selections for . Next, we inquire whether the inverse
is also true, that is, whether an irrotational ow can be expressed as the gradient of a potential
function, as shown in (2.1.1).
n
x2
tB tA
Boundary
D
B A
L
t x1
where D is an arbitrary surface bounded by L, t is the unit vector tangential to L, and n is the
unit vector normal to D oriented according to the counterclockwise convention with respect to t, as
illustrated in Figure 2.1.1. One consequence of equation (2.1.2) is that the circulation around any
reducible loop that lies inside an irrotational ow is zero.
Next, we choose two points x1 and x2 on a reducible loop and decompose the line integral in
(2.1.2) into two parts,
x2 x2
u tA dl = u tB dl. (2.1.3)
x1 x1
The integral on the left-hand side is taken along path A with corresponding tangent vector tA = t,
while the integral on the right-hand side is taken along path B with corresponding tangent vector
tB = t, as shown in Figure 2.1.1. Equation (2.1.3) states that the circulation around any path
connecting two ordered points x1 and x2 on a reducible loop is the same. Consequently, a single-
valued scalar function of position can be introduced, , such that
x2
u t dl = (x2 ) (x1 ). (2.1.4)
x1
Taking the limit as the second point, x2 , tends to the rst point, x1 , and using a linearized Taylor
series expansion, we nd that
u t = t . (2.1.5)
Since the integration path, and thus the tangent unit vector t, is arbitrary, we conclude that u = ,
as shown in (2.1.1).
m=1 x1
m=2
A B t
x1
A
t
x2 B
x2
Figure 2.1.2 Illustration of two irreducible loops in a doubly connected domain of ow with number
of turns m = 1 and 2. The ow can be resolved into the ow due to a line vortex with circulation
inside the boundaries and a complementary irrotational ow described in terms by a single-valued
potential.
where is the lowest value of the circulation corresponding to a loop that performs a single turn,
called the cyclic constant of the ow around the boundary.
In the case of an irreducible loop, a surface D bounded by L must cross the boundaries of
the ow. Since D does not lie entirely inside the uid, Stokes theorem (2.1.2) cannot be applied.
Progress can be made by breaking up the circulation integral around the loop into two parts and
working as in (2.1.3) and (2.1.4), to nd that
A B = m, (2.1.7)
where A and B denote the change in the potential function from the beginning to the end of
the paths A and B illustrated in Figure 2.1.2.
u = v + , (2.1.8)
2.1 Irrotational ows and the velocity potential 115
where v represents a known irrotational ow whose cyclic constants around the boundaries are the
same as those of the ow of interest. For example, v could be identied with the ow due to a line
vortex lying outside the domain of ow in the interior of a toroidal boundary. The strength of the
line vortex is equal to the cyclic constant of the ow around the boundary, as illustrated in Figure
2.1.2. Because the velocity potential dened in (2.1.8) is a single-valued function of position, it
can be computed by standard analytical and numerical methods without any added considerations.
d
+ = t = t, (2.1.9)
dl Illustration of a
where is the strength of the vortex sheet and l is the arc length two-dimensional vortex sheet.
along the vortex sheet measured in the direction of the tangent unit
vector, t.
Projecting (2.1.9) onto t and integrating with respect to arc length, l, we nd that the jump
in the velocity potential across the vortex sheet is given by
+ = . (2.1.10)
We have assumed that + and have the same value at the designated origin of arc length along
the vortex sheet, l = 0. Equation (2.1.10) nds useful applications in computing the self-induced
motion of two-dimensional vortex sheets discussed in Section 11.5.
where is the strength of the vortex sheet and t is the tangent unit vector along the path. If t is
tangential to at every point, the integration path coincides with a vortex line and the integrand
in (2.1.11) is identically zero. We conclude that the jump of the velocity potential across a vortex
sheet is constant along vortex lines.
116 Introduction to Theoretical and Computational Fluid Dynamics
u = 2 , (2.1.12)
which can be regarded as a Poisson equation for , forced by the rate of expansion, u.
Using the Poisson inversion formula derived in Section 2.2, we obtain an expression for the
potential of a three-dimensional ow in terms of the rate of expansion,
1 1
(x) = (x ) dV (x ) + H(x), (2.1.13)
4 F low r
where r = |xx |, (x ) u(x ) is the rate of expansion, the gradient involves derivatives with
respect to x , and H(x) is a harmonic function determined by the boundary conditions, 2 H = 0.
If the domain of ow extends to innity, to ensure that the volume integral in (2.1.13) is nite, we
require that the rate of expansion decays at a rate that is faster than 1/d2 , where d is the distance
from the origin.
Two-dimensional ow
The counterpart of (2.1.13) for two-dimensional ow in the xy plane is
1 r
(x) = ln (x ) dA(x ) + H(x), (2.1.14)
2 F low a
Velocity eld
To obtain an expression for the velocity in terms of the rate of expansion, we take the gradient of
both sides of (2.1.13) and (2.1.14). Interchanging the gradient with the integral, we nd that
1 x
u(x) = (x) = (x ) dV (x ) + H(x) (2.1.15)
4 F low r 3
for two-dimensional ow, where x = x x . Later in this section, we will see that the integrals on
the right-hand sides of the last two equations can be interpreted as volume or areal distributions of
point sources. The densities of the distributions are equal to the local volumetric or areal rate of
expansion.
2.1 Irrotational ows and the velocity potential 117
2 = 0. (2.1.17)
Curvilinear coordinates
In cylindrical polar coordinates, (x, , ), Laplaces equation reads
2 1
1 2
2 = + + . (2.1.18)
x2 2 2
1 2
1
1 2
2 = 2
r + 2 sin + 2 . (2.1.19)
r r r r sin r sin 2
1
1 2
2 = r + 2 . (2.1.20)
r r r r 2
Expressions in more general orthogonal or nonorthogonal curvilinear coordinates are provided in
Sections A.8A.17, Appendix A.
where n is the unit vector normal to the boundaries pointing into the ow.
Equation (2.1.22) shows that, if the normal velocity component is zero over all boundaries,
the kinetic energy and therefore the velocity must be zero and the uid must be quiescent. In turn,
118 Introduction to Theoretical and Computational Fluid Dynamics
+
n
t
D
B
Figure 2.1.3 A doubly connected domain of ow containing a toroidal boundary is rendered singly
connected by introducing a ctitious boundary surface ending at the boundary.
the velocity potential must be constant throughout the domain of ow. When the velocity potential
has the same constant value over all boundaries, mass conservation requires that the right-hand side
of (2.1.22) is identically zero. Consequently, the uid must be quiescent and the potential must be
constant and equal to its boundary value throughout the domain of ow.
where is the potential on either side of D. The circulation around the loop L shown in Figure
2.1.3 is equal to the cyclic constant of the ow around the toroidal boundary,
= u t dl = t dl = dl = + . (2.1.24)
L L L l
Expression (2.1.25) shows that the kinetic energy is zero and the uid is quiescent under two
sets of conditions: (a) either the potential is constant or the normal component of the velocity is
zero at the boundaries, and (b) either the cyclic constant, , or the ow rate, Q, is zero.
2.1 Irrotational ows and the velocity potential 119
One important consequence of the integral representation of the kinetic energy is that, given bound-
ary conditions for the normal component of the velocity, an incompressible irrotational ow in a
singly connected domain is unique and the corresponding harmonic potential is determined uniquely
up to an arbitrary constant.
To prove uniqueness of solution, we consider two harmonic potentials representing two distinct
ows and note that their dierence is also an acceptable harmonic potential, that is, the dierence
potential describes an acceptable incompressible irrotational ow. If the two ows satisfy the same
boundary conditions for the normal component of the velocity, the dierence ow must vanish
and the corresponding harmonic potential must be constant. As a consequence, the two chosen
ows must be identical and the corresponding harmonic potentials may dier at most by a scalar
constant. Similar reasoning allows us to conclude that the boundary distribution of the potential
uniquely denes an irrotational ow in a singly connected domain.
In the case of ow in a doubly connected domain, we use (2.1.25) and nd that specifying (a)
boundary conditions either for the normal component of the velocity or for the potential, and (b)
either the value of the cyclic constant , or the ow rate across a surface ending at a toroidal
boundary Q, uniquely determines an irrotational ow. Similar conclusions can be drawn for a
multiply connected domain.
Kelvin demonstrated that, of all solenoidal velocity elds that satisfy prescribed boundary conditions
for the normal component of the velocity, the velocity eld corresponding to an irrotational ow has
the least amount of kinetic energy. Stated dierently, vorticity increases the kinetic energy of a uid.
Manipulating the last integral on the right-hand side and using the divergence theorem, we obtain
(v u) dV = (v u) n dS, (2.1.27)
B
which is zero in light of the equality of the normal component of the boundary velocity. We conclude
that the right-hand side of (2.1.26) is positive and thereby demonstrate that the kinetic energy of
a rotational ow with velocity v is greater than the kinetic energy of the corresponding irrotational
ow with velocity u.
120 Introduction to Theoretical and Computational Fluid Dynamics
Point source
Consider the solution of (2.1.17) in an innite three- or two-dimensional domain in the absence
interior boundaries, subject to a singular forcing term on the right-hand side,
2 3D = m 3 (x x0 ), 2 2D = m 2 (x x0 ), (2.1.28)
where 3 is the three-dimensional (3D) delta function, 2 is the two-dimensional (2D) delta function,
m is a constant, and x0 is an arbitrary point in the domain of ow. Using the method of Fourier
transforms, or else by employing a trial solution in the form of a power or logarithm of the distance
from the singular point, r = |x x0 | , we nd that
m 1 m r
3D = , 2D = ln , (2.1.29)
4 r 2 a
where a is an arbitrary length.
Since the potential satises Laplaces equation everywhere except at the point x0 , it describes
an acceptable incompressible and irrotational ow with velocity
m x x0 m x x0
u3D = 3D = , u2D = 2D = , (2.1.30)
4 r 3 2 r2
respectively, for three- or two-dimensional ow. The corresponding streamlines are straight radial
lines emanating from the singular point, x0 . It is a straightforward exercise to verify that the ow rate
across a spherical surface or circular contour centered at the point x0 is equal to m. Consequently,
the function 3D or 2D can be identied with the potential due to a point source with strength m
in an innite domain of ow. In Section 2.2.3, we will discuss the ow due to a point source in a
bounded domain of ow.
In light of (2.1.30), the integrals on the right-hand sides of (2.1.15) and (2.1.16) can be
interpreted as volume or areal distributions of point sources of mass forced by the rate of expansion.
In Section 2.2.2, we will see that the potential due to a point sink provides us with the free-space
Greens function of Laplaces equation
Point-source dipole
Now we consider two point sources with strengths of equal magnitude and opposite sign located
at the points x0 and x1 . Exploiting the linearity of Laplaces equation (2.1.17), we construct the
associated potential by linear superposition,
m 1 m 1
3D = + ,
4 |x x0 | 4 |x x1 |
(2.1.31)
2D m |x x0 | m |x x1 |
= ln ln ,
2 a 2 a
2.1 Irrotational ows and the velocity potential 121
respectively, for three- and two-dimensional ow. Placing the point x1 near x0 , expanding the
potential of the second point source in a Taylor series with respect to x1 about the location of the
rst point source, x0 , and retaining only the linear terms, we nd that
m 1
3D = (x1 x0 ) 0 ,
4 |x x0 |
(2.1.32)
m |x x0 |
2D = (x1 x0 ) 0 ln ,
2 a
where the derivatives of the gradient 0 are taken with respect to x0 .
1 x x0 1 x x0 m x1
3D = d, 2D = d, (2.1.33)
4 r 3 2 r 2
where r = |x x0 | is the distance of the eld point, x, from the singular A point source with
point, x0 . The corresponding velocity elds are given by strength m and a point
sink with strength m
1 1 x x merge into a dipole.
u3D = 3D = ( 3 I + 3 ) d,
4 r r5
(2.1.34)
2D 2D 1 1 x x
u = = ( 2 I + 2 ) d,
2 r r4
where I is the identity matrix and x = x x0 . The ow rate across a spherical surface, and therefore
across any closed surface, enclosing a three-dimensional point-source dipole is zero. Similarly, the
ow rate across any closed loop enclosing a two-dimensional point-source dipole is zero.
Problems
(a) (b)
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
y
y
0.2 0.2
0.4 0.4
0.6 0.6
0.8 0.8
1 1
1 0.5 0 0.5 1 1 0.5 0 0.5 1
x x
Figure 2.1.4 Streamline pattern due to (a) a three-dimensional and (b) a two-dimensional potential
dipole pointing in the positive direction of the x axis.
tangential component of the velocity is zero over the other surface. Does this imply that the velocity
eld must vanish throughout the whole domain of ow?
2.1.3 Innite ow
Consider a three-dimensional incompressible and irrotational ow in an innite domain where the
velocity vanishes at innity. In Section 2.3, we will show that the velocity potential must tend to
a constant value at innity, as shown in equation (2.3.18). Based on this observation, show that, if
the ow has no interior boundaries and no singular points, the velocity eld must necessarily vanish
throughout the whole domain of ow.
f 2 g = (f g) f g, (2.2.1)
which can be proven by straightforward dierentiation working in index notation. Interchanging the
roles of f and g, we obtain
g 2 f = (gf ) g f. (2.2.2)
f 2 g g 2 f = (f g gf ). (2.2.3)
If both functions f and g satisfy Laplaces equation, the left-hand side of (2.2.3) is zero, yielding a
reciprocal relation for harmonic functions,
(f g gf ) = 0. (2.2.4)
Integrating (2.2.4) over a control volume that is bounded by a singly or multiply connected surface,
D, and using the divergence theorem to convert the volume integral into a surface integral, we obtain
the integral form of the reciprocal theorem,
(f g g f ) n dS = 0, (2.2.5)
D
where n it the unit vector normal to D pointing either into the control volume or outward from
the control volume. Equation (2.2.5) imposes an integral constraint on the boundary values and
boundary distribution of the normal derivatives of any pair of nonsingular harmonic functions.
2 G(x, x0 ) + 3 (x x0 ) = 0, (2.2.6)
where 3 is the three-dimensional delta function, x is a eld point, and x0 is the location of the
Greens function, also called the singular point or the pole. When the domain of ow extends to
innity, the Greens function decays at least as fast as the inverse of the distance from the pole,
1/|x x0 |.
124 Introduction to Theoretical and Computational Fluid Dynamics
G(x, x0 ) = 0 (2.2.7)
when x lies on SB . Unless qualied, a Greens function of the rst kind is simply called a Greens
function.
The normal derivative of a Greens function of the second kind, also called a Neumann function,
is zero over a specied surface, SB ,
Physical interpretation
Comparing (2.2.6) with (2.1.28), we nd that, physically, a Greens function represents the steady
temperature eld due to a point source of heat with unit strength located at the point x0 , in the
presence of an isothermal or insulated boundary, SB . A Greens function of the rst kind represents
the temperature eld due to a point source of heat subject to the condition of zero boundary
temperature. A Greens function of the second kind represents the temperature eld due to a point
source of heat subject to the condition of zero boundary ux.
1
G(x, x0 ) = , (2.2.9)
4r
where r = |x x0 | is the distance of the eld point from the pole.
1
G(x, x0 ) = + H(x, x0 ). (2.2.10)
4r
The decomposition (2.2.10) shows that, as the observation point, x, approaches the singular point,
x0 , all Greens functions exhibit a common singular behavior. The precise form of H depends on the
geometry of the boundary associated with the Greens function, SB . In the absence of a boundary,
H is identically zero. For a limited class of simple boundary geometries, the complementary part,
H, can be found by the method of images. The construction involves introducing Greens functions
and their derivatives at strategically selected locations outside the domain of ow.
1 1
G(x, x0 ) = , (2.2.11)
4r 4rim
0 = (2w x0 , y0 , z0 )
xim (2.2.12)
is the image of the pole, x0 , with respect to the wall. The minus and plus signs apply, respectively,
for the Greens function of the rst or second kind.
1 1 a 1
G(x, x0 ) = , (2.2.13)
4r 4 |x0 xc | rinv
a2
xinv
0 = xc + (x0 xc ). (2.2.14)
|x0 xc |2
By construction, rinv /r = a/|x0 xc |, which demonstrates that the Greens function is zero when
the point x lies on the sphere.
The corresponding Greens function of the second kind representing the ow due to a point sink
of mass in the presence of an interior impermeable sphere will be derived in Section 7.5.5. A Greens
function of the second kind representing the ow inside a sphere cannot be found. The physical
reason is that uid cannot escape through the impermeable boundaries of an interior, completely
enclosed domain.
126 Introduction to Theoretical and Computational Fluid Dynamics
(a) (b)
n
n
D D
D
Figure 2.2.1 Illustration of (a) a nite control volume in a ow bounded by an interior and an exterior
closed surface and (b) an innite control volume bounded by an interior closed surface.
where the normal unit vector, n, points into the control volume, Vc . When the point x0 is located
precisely on the boundary, D, the improper but convergent integral on the left-hand side of (2.2.15)
is called a principal value integral (P V ). Using relations (2.2.15), we derive the identity
PV
1
n(x) G(x, x0 ) dS(x) = n(x) G(x, x0 ) dS(x) , (2.2.16)
D D 2
where plus or minus sign on the left-hand side applies, respectively, when the point x0 lies inside or
outside the control volume.
Next, we integrate (2.2.17) over a control volume, Vc , that is bounded by a surface associated with
the Greens function, SB . In the case of innite ow, Vc is also bounded by an outer surface with
2.2 The reciprocal theorem and Greens functions of Laplaces equation 127
large dimensions, S . Using the divergence theorem, we convert the volume integral on the right-
hand side to a surface integral, invoke the distinguishing properties of the delta function to write,
for example,
G(x, x1 ) 3 (x x2 ) dV (x) = G(x2 , x1 ), (2.2.18)
Vc
and nd that
G(x2 , x1 ) G(x1 , x2 ) = [G(x, x1 ) G(x, x2 ) G(x, x2 ) G(x, x1 )] n(x) dS(x). (2.2.19)
SB ,S
Since either the Greens function itself or its normal derivative is zero over SB , the integral over SB
on the right-hand side disappears. In the case of innite ow, we let the large surface S expand to
innity and nd that, because the integrand decays at a rate that is faster than inverse quadratic,
the corresponding integrals make vanishing contributions. We conclude that a Greens function of
the rst or second kind (Neumann function) satisfy the symmetry property
Physical interpretation
In physical terms, equation (2.2.20) states that the temperature or velocity potential at a point,
x2 , due to a point source of heat or point sink of mass located at another point, x1 , is equal to
the temperature or velocity potential at the point x1 due to a corresponding singularity located at
x2 . In Chapter 10, we will see that Greens functions appear as kernels in integral equations for
the boundary distribution of the harmonic potential or its normal derivatives, and the symmetry
property (2.2.20) has important implications on the properties of the solution.
One noteworthy consequence of (2.2.20) is that, when the pole of a Greens function of the
rst kind is placed at the boundary SB where the Greens function is required to vanish, the Greens
function is identically zero, that is, G(x, x0 ) = 0 when x0 is on SB for any x. This behavior can
be understood in physical terms by identifying the Greens function with the temperature eld
established when a point source of heat is placed on an isothermal body. Because the heat of the
point source is immediately absorbed by the body, a temperature eld is not established.
N
G(x, x1 , . . . , xN ) = G(x, xi ). (2.2.21)
i=1
Physically, this Greens function can be identied with the temperature eld due to a collection of
point sources of heat or with the velocity potential due to a collection of point sinks of mass.
128 Introduction to Theoretical and Computational Fluid Dynamics
A periodic Greens function represents the temperature eld or harmonic potential due to a
simply, doubly, or triply periodic array of point sources of heat or point sinks of mass. In certain
cases, a periodic Greens function cannot be computed simply by adding an innite number of
Greens functions with single poles, as the superposition produces divergent sums. Instead, the
solution must be found by solving the dening equation
2 G(x, x0 ) + 3 (x xi ) = 0, (2.2.22)
i
where the sum is computed over the periodic array. The solution can be found using Fourier series
expansions.
2 G(x, x0 ) + 2 (x x0 ) = 0, (2.2.30)
with respect to the wall. The minus or plus sign apply, respectively, for the Greens function of the
rst or second kind (Neumann function).
By construction, rinv /r = a/|x0 xc |, which shows that the Greens function is zero when the eld
point x lies on the circle.
130 Introduction to Theoretical and Computational Fluid Dynamics
The Greens function of the second kind representing the ow due to a point sink of mass in
the presence of an interior circular boundary will be derived in Section 7.8. The Greens function of
the second kind for ow in the interior of a circular boundary cannot be found.
Innite ow
An apparent complication is encountered in the case of innite two-dimensional ow. In the limit
as the contour C expands to innity, the Greens function may increase at a logarithmic rate, and
the integrals of the two terms in (2.2.19) over the large contour, C , which is the counterpart of the
large surface S in the three-dimensional ow, may not vanish. However, expanding the Greens
functions inside the integrand of (2.2.19) in a Taylor series with respect to x0 about the origin, we
nd that the sum of the two integrals makes a vanishing contribution and the combined integral
over C cancels out.
Problems
Vc
D
n
S
x0 n
D
n
D
n
Figure 2.3.1 Illustration of a control volume used to derive a boundary-integral representation of the
potential of an irrotational ow at a point, x0 .
By convention, the normal unit vector, n, points into the control volume.
132 Introduction to Theoretical and Computational Fluid Dynamics
By convention, the normal unit vector, n, points into the control volume. Considering the integrals
over S , we write dS = 2 d, where is the solid angle dened as the area of a sphere of unit radius
centered at x0 . Using (2.2.8), we obtain
1
G(x, x0 ) = n(x) + H(x, x0 ), (2.3.4)
42
where n = 1 (x x0 ) is the normal unit vector pointing into the control volume, as shown in
Figure 2.3.1. Taking the limit as tends to zero, using (2.2.8) and (2.3.4) and recalling that the
complementary component H is nonsingular throughout the domain of ow, we nd that
1
G(x, x0 ) n(x) (x) dS(x) = n(x) (x) 2 d(x) 0 (2.3.5)
S 4 S
and
1
(x) n(x) G(x, x0 ) dS(x) (x0 ) 2 d(x) = (x0 ). (2.3.6)
S 42 S
Substituting (2.3.5) and (2.3.6) into (2.3.3), we obtain the nal result
(x0 ) = G(x, x0 ) n(x) (x) dS(x) + (x) n(x) G(x, x0 ) dS(x), (2.3.7)
D D
which provides us with a boundary-integral representation of a harmonic function in terms of the
boundary values and the boundary distribution of its normal derivative, which is equal to the normal
component of the velocity. To compute the value of at a chosen point, x0 , inside a selected control
volume, we simply compute the two boundary integrals on the right-hand side of (2.3.7) by analytical
or numerical methods.
The integral representation (2.3.7) can be derived directly using the properties of the delta
function, following the procedure that led us from (2.2.17) to (2.2.19). However, the present deriva-
tion based on the exclusion of a small sphere centered at the evaluation point bypasses the use of
generalized functions.
Physical interpretation
The symmetry property (2.2.20) allows us to switch the order of the arguments of the Greens
function in the integral representation (2.3.7), obtaining
(x0 ) = G(x0 , x) n(x) (x) dS(x) + (x) n(x) G(x0 , x) dS(x). (2.3.8)
D D
2.3 Integral representation of three-dimensional potential ow 133
The two integrals on the right-hand side represent boundary distributions of Greens functions and
Greens function dipoles oriented normal to the boundaries of the control volume, amounting to
boundary distributions of point sources and point-source dipoles. Making an analogy with corre-
sponding results in the theory of electrostatics concerning the surface distribution of electric charges
and charge dipoles, we call the rst integral in (2.3.8) the single-layer harmonic potential and the sec-
ond integral the double-layer harmonic potential. The distribution densities of these potentials are
equal, respectively, to the normal derivative and to the boundary values of the harmonic potential.
The integrals on the right-hand side represent the velocity due to boundary distributions of Greens
functions and Greens function dipoles.
where r = |x x0 | is the distance of the evaluation point, x0 , from the integration point, x.
where Q is the ow rate across S or any other closed surface enclosed by S , dened as
Q= n(x) (x) dS(x) = n(x) (x) dS(x), (2.3.12)
S B
134 Introduction to Theoretical and Computational Fluid Dynamics
The normal unit vector, n, points inward over S . Concentrating on the ow rate, we write
2 2 d
Q= dS(x) = R d(x) = R d(x), (2.3.14)
S r r=R S r r=R dR S
where is the solid angle dened as the surface area of a sphere of unit radius centered at the point
x0 . Substituting the denition of , we obtain
Q
= . (2.3.15)
4R2 R
Integrating this equation with respect to R treating Q as a constant, we obtain
Q
(R, x0 ) = + c(x0 ), (2.3.16)
4R
where c is independent of R. Taking derivatives of the last equation with respect to x0 keeping R
xed, we nd that
0 c = 0 = 0 (x) d(x) = (x) d(x) = 0. (2.3.17)
S S
The value of zero emerges by letting R in the last integral tend to innity and invoking the original
assumption that the velocity decays at innity. We have thus shown that c is an absolute constant.
Substituting (2.3.15) into (2.3.11), we derive a simplied boundary integral representation lacking
the integral over the large surface,
1 1 1 (x)
(x0 ) = c n(x) (x) dS(x) + n(x) (x0 x) dS(x). (2.3.18)
4 B r 4 B r 3
This equation nds useful applications in computing an exterior ow.
Letting the point x0 in (2.3.18) tend to innity and recalling that r is the distance of the
evaluation point, x0 , from the integration point, x, over the interior boundary B, we deduce the
asymptotic behavior
Q
(x0 ) = c + , (2.3.19)
4|x0 |
where the dots represent decaying terms. This expression demonstrates that the potential of an
innite three-dimensional ow that decays at innity tends to a constant value far from the interior
boundaries.
Based on (2.3.19), we nd that the expression for the kinetic energy (2.1.22) becomes
1 1
K= u n dS + c Q, (2.3.20)
2 B 2
2.3 Integral representation of three-dimensional potential ow 135
where B is an interior boundary. Following the arguments of Section 2.1, we deduce that, given
the boundary distribution of the normal component of the velocity, a potential ow that decays
at innity is unique and the corresponding harmonic potential is determined uniquely up to an
arbitrary constant.
G(x, x0 ) = , (2.3.21)
4 r |x0 | rinv
where r = |x x0 |, rinv = |x xinv
0 |, and x0
inv
is the inverse of the point x0 with respect to the
2 2
sphere, located at xinv
0 = x0 a /|x0 | . The gradient of the Greens function is
1 x x0 a x xinv
0
G(x, x0 ) = 3
3 . (2.3.22)
4 r |x0 | rinv
Evaluating this expression at a point x on the sphere and projecting the result onto the inward
normal unit vector, n = a1 x, we obtain
1 1 2 a 1 2
which is identical to (2.3.8) except that the right-hand side includes a volume integral over the
control volume Vc enclosed by the boundary D, called the volume potential. Physically, the volume
potential represents a volume distribution of point sources.
Adopting the free-space Greens function, we obtain the counterpart of Greens third identity
for compressible ow,
1 1
(x0 ) = n(x) (x) dS(x)
4 D r
1 x0 x 1 1 2 (2.3.28)
+ n(x) (x) dS(x) (x) dS(x).
4 D r3 4 Vc r
The normal unit vector, n, points into the control volume, Vc .
Problem
stating that the mean value of a function that satises Laplaces equation over the surface of a
sphere is equal to value of the function at the center of the sphere. One interesting consequence
of the mean-value theorem is that, if the harmonic potential of an innite ow with no interior
boundaries decays at innity, it must vanish throughout the whole space.
Using (2.4.2), we nd that the mean value of a harmonic function over the volume of a sphere
of radius a is equal to value of the function at the center of the sphere,
a a
1 1 1
(x) dV (x) = (x) dS(x) dr = 4r 2 (x0 ) dr = (x0 ), (2.4.3)
V V 0 Sr V 0
4 3
where V = 3 a is the volume of the sphere and Sr denotes a sphere of radius r. Rearranging, we
obtain
1
(x0 ) = 4 3 (x) dV (x), (2.4.4)
3 a
stating that the mean value of a function that satises Laplaces equation over the volume of a
sphere is equal to value of the function at the center of the sphere.
Accordingly,
with the equality holding when x = x0 . However, since the positive scalar function |u(x)e0 | satises
Laplaces equation, a point x can be found such that
which contradicts the original assumption. One corollary of this result is that, if the velocity of an
innite irrotational ow with no interior boundaries vanishes at innity, it must vanish throughout
the whole space.
where the normal unit vector, n, points outward from the sphere. The rst integral on the right-hand
side is equal to 1. Rearranging, we obtain
1 1 1
G(x, x1 ) dS(x) = G(x0 , x1 ) + . (2.4.10)
4a2 4|x1 x0 | 4a
Biharmonic functions
Next, we derive a mean-value theorem for a function, , that satises the biharmonic equation,
4 = 0. Working as in the case of harmonic functions, we consider a spherical control volume of
radius a centered at a chosen point, x0 . Identifying D with the spherical boundary of the control
volume and substituting in (2.3.28) r = a, n = a1 (x0 x), and f = , we obtain
1 1 1
(x0 ) = n(x) (x) dS(x) + (x) dS(x) 2 (x) dS(x). (2.4.12)
4a 4a2 4a
Applying the divergence theorem, we nd that the rst surface integral on the right-hand side is
equal to the negative of the integral of the Laplacian 2 over the volume of the sphere. We note
2.5 Two-dimensional potential ow 139
that the function 2 satises Laplaces equation and use the mean-value theorem expressed by
(2.4.4) to replace its volume integral over the sphere with 4 3 2
3 a (x0 ), obtaining
1 1 1
(x0 ) = a2 2 (x0 ) + (x) dS(x) 2 (x) dS(x). (2.4.13)
3 4a2 4a
Finally, we use the mean-value theorem expressed by (2.4.4) for the harmonic function 2 to
simplify the last integral, and rearrange to derive the mean-value theorem expressed by
1 1
(x) dS(x) = (x0 ) + a2 2 (x0 ), (2.4.14)
4a2 6
where denotes a sphere. This equation relates the mean value of a biharmonic function over the
surface of a sphere to the value of the function and its Laplacian at the center of the sphere.
Working as in (2.4.4), we nd that the mean value of a biharmonic function over the volume
of a sphere is given by
1 1 2 2
4 3 (x) dV (x) = (x0 ) + a (x0 ). (2.4.15)
3 a 10
Problem
4 G + 3 (x x0 ) = 0, (2.4.16)
where 3 is the three-dimensional delta function. Conrm that the free-space Greens function is
G = r/8, where r = |x x0 |. Derive the counterpart of (2.4.10).
where the normal unit vector, n, points into the control area enclosed by C. Greens third identity
for a single-valued potential takes the form
1 r 1 (x)
(x0 ) = ln n(x) (x) dl(x) + n(x) (x0 x) dl(x), (2.5.2)
2 C a 2 C r 2
Q |x0 |
(x0 ) = ln + c + , (2.5.3)
2 a
where c is a constant, |x0 | is the distance from the origin assumed to be in the vicinity of the
boundary C, a is a constant length, and
Q= n dl (2.5.4)
C
is the ow rate across C (Problem 2.5.3). Now we consider the kinetic energy of the ow expressed
by the potential
Q |x0 |
ln . (2.5.5)
2 a
Repeating the procedure of Section 2.3 for three-dimensional ow, and nd that a two-dimensional
ow in an innite domain described by a single-valued harmonic potential is determined uniquely
by specifying the normal component of the velocity along the interior boundaries.
Compressible ow
The integral representation of the potential of a compressible ow stated in (2.3.27) is also valid
for two-dimensional ow, provided that the control volume is replaced by a control area, and the
boundary, D, is replaced by a closed contour, C, enclosing the control area.
where denotes the circle or the disk inside the circle, l is the arc length around the circle, and
dA = dx dy is the dierential area inside the circle.
2.5 Two-dimensional potential ow 141
Working as in the case of three-dimensional ow, we derive mean-value theorems for functions
that satisfy the biharmonic equation, 4 = 0, stating that
1 1
(x) dl(x) = (x0 ) + a2 2 (x0 ) (2.5.7)
2a 4
where denotes the circle, and
1 1
(x) dA(x) = (x0 ) + a2 2 (x0 ), (2.5.8)
a2 8
where denotes the disk inside the circle (Problem 2.5.2). If is a harmonic function, the last term
on the right-hand side of (2.5.7) or (2.5.8) does not appear.
1 x x0 x xinv
0
G(x, x0 ) = 2 . (2.5.10)
2 r2 rinv
Evaluating this expression at a point x on the circle and projecting the resulting equation onto the
inward normal unit vector, n = x/a, we obtain
1 1 2 1
2
n(x) G(x, x0 ) = (a x0 x) 2 (a x inv
0 x) . (2.5.11)
2a r 2 rinv
Recalling that for a point on the circle rinv /r = a/|x0 | and simplifying, we nd that
1 2 |x0 |2 a2 1
n(x) G(x, x0 ) = 2
(a x0 x) 2 (a2 2
x0 x) = (a2 |x0 |2 ). (2.5.12)
2ar a |x0 | 2ar2
Substituting this expression into the rst integral on the right-hand side of (2.5.1) and noting that
the second integral disappears because the Greens function is zero on the circular boundary, we nd
that the potential at a point x0 located in the interior or exterior of a circle of radius a centered at
the origin is given by the Poisson integral
a2 |x0 |2 (x)
(x0 ) = dl(x), (2.5.13)
2a |x x0 |2
where denotes the circle. The plus or minus sign corresponds, respectively, to interior or exterior
ow (e.g., [106], p. 242). Applying this equation for constant interior provides us with an
interesting identity.
142 Introduction to Theoretical and Computational Fluid Dynamics
Problems
u = A. (2.6.1)
A3 A2 A1 A3 A2 A1
u1 = , u2 = , u3 = . (2.6.2)
x2 x3 x3 x1 x1 x2
A1 = f1 (x1 ), (2.6.3)
where f1 is an arbitrary function. Integrating the second and third equations in (2.6.2) with respect
to x1 from an arbitrary position, x1 = a, we nd that
x1 x1
A2 (x) = u3 (x ) dx1 + f2 (x2 , x3 ), A3 (x) = u2 (x ) dx1 + f3 (x2 , x3 ), (2.6.4)
a a
where f2 , f3 are two arbitrary somewhat related functions of x2 and x3 . Substituting expressions
(2.6.4) into the rst equation in (2.6.2), we obtain
x1
u2 u3
f
3 f2
Since the velocity eld u is solenoidal, we may simplify the integrand in (2.6.5) to obtain
x1 f
u1 3 f2
f = u1 (a, x2 , x3 ) e1 , (2.6.8)
Stream functions
In Section 2.9, we will show that the vector potential of a two-dimensional or axisymmetric ow
can be described in terms of one scalar function, called, respectively, the Helmholtz two-dimensional
stream function or the Stokes axisymmetric stream function. Both are dened uniquely up to an
arbitrary scalar constant. The vector potential of a three-dimensional ow can be described in terms
of two scalar functions, called the stream functions. For example, the vector potential of a three-
dimensional ow can be expressed in the form A = , where and are two scalar stream
functions. It is interesting to note that A = , which shows that A ( A) = 0,
that is, the eld A is perpendicular to its curl. A vector eld that possesses this property is called
a complex lamellar eld ([14], p. 63). In contrast, a vector eld that is parallel to its own curl,
A ( A) = 0, is called a Beltrami eld.
Problems
u A = (A) u. (2.6.9)
= u. (2.7.1)
To carry out this inversion, we may use (2.6.1), (2.6.3), and (2.6.4), where u and A are replaced,
respectively, by and u. However, it is expedient to work in an alternative fashion by expressing
the velocity in terms of a vector potential, as shown in (2.6.1).
= A = ( A) 2 A. (2.7.2)
A = B + C, (2.7.3)
2 B = , (2.7.4)
C = ( B). (2.7.5)
(B + C) = , (2.7.6)
as required. The computation and signicance of the primary and complementary parts will be
discussed later in this section.
2.7 Representation of an incompressible ow in terms of the vorticity 145
^x
x x^
Figure 2.7.1 A rotating uid particle induces a rotary ow expressed by the BiotSavart integral.
Under these stipulations, we obtain a particular solution of (2.7.4) using the Poisson inversion
formula with the free-space Greens function,
1 (x )
B(x) = dV (x ), (2.7.7)
4 F low r
where r = |x x |. Substituting (2.7.7) into the right-hand side of (2.7.3), taking the curl of the
resulting expression, and switching the curl with the integral operator on the right-hand side, we
derive an expression for the velocity,
1 1
Ck
ui (x) = ijk k (x ) dV (x ) + . (2.7.8)
4 F low xj r xj
Carrying out the dierentiations under the integral sign and switching back to vector notation, we
derive an integral representation for the velocity,
1 1
u(x) = 3
(x ) x dV (x ) + C, (2.7.9)
4 F low r
where x = x x . The rst two terms on the right-hand side express the primary ow, and the last
term represents the complementary ow.
The volume integral on the right-hand side of (2.7.9) is similar to the BiotSavart integral
in electromagnetics expressing the magnetic eld due to an electrical current. By analogy, the
integral in (2.7.9) expresses the velocity eld induced by the rotation of point particles distributed
along the vortex lines, as shown in Figure 2.7.1. In formal mathematics, the BiotSavart integral
represents a volume distribution of singular fundamental solutions, called rotlets or vortons, with
vector distribution density equal to the vorticity.
146 Introduction to Theoretical and Computational Fluid Dynamics
B
n
Flow
n
B
Figure 2.7.2 Illustration of a ow domain enclosed by a collection of interior and exterior boundaries,
collectively denoted as B.
where B represents the boundaries of the ows and the normal unit vector n points into the ow,
as shown in Figure 2.7.2.
Now switching to vector notation and using the decomposition (2.7.3), we obtain the integral
representation
1 1 1 1
u(x) = (x ) dV (x ) + n(x ) (x ) dS(x ) + C. (2.7.13)
4 F low r 4 B r
The second term on the right-hand side involving the boundary integral disappears when the vortex
lines cross the boundaries at a right angle. The volume integral disappears when the curl of the
vorticity eld is irrotational, in which case all Cartesian components of the velocity satisfy Laplaces
equation, 2 u = 0. The curl of the vorticity of a two-dimensional or axisymmetric ow lies in the
plane of the ow or in an azimuthal plane and the volume integral produces respective motion in
these planes.
2.7 Representation of an incompressible ow in terms of the vorticity 147
Innite decaying ow
In the case of ow in an innite domain with no interior boundaries decaying far from the origin,
the representation (2.7.13) simplies into
1 1 1 1 2
u(x) = (x ) dV (x ) = u(x ) dV (x ). (2.7.14)
4 F low r 4 F low r
To derive the last expression, we have noted that 2 u = for any solenoidal velocity eld,
u = 0.
Next, we change the variable of dierentiation from x to x while simultaneously introducing a minus
sign, writing
1 1
and then
1 i (x )
1 i (x )
B(x) = + dV (x ). (2.7.17)
4 F low xi r r xi
Because the vorticity eld is solenoidal, the second term inside the integral is zero. Using the
divergence theorem, we nd that
1 1
B(x) = (x ) n(x ) dS(x ), (2.7.18)
4 B r
where the normal unit vector, n, points into the domain of ow, as shown in Figure 2.7.2. The
boundary integral vanishes when the vorticity vector is perpendicular to the normal vector or,
equivalently, the vortex lines are tangential to the boundaries of the ow. This is always true in
the case of two-dimensional or axisymmetric ow, but not necessarily true in the case of three-
dimensional ow (Problem 2.7.1). In the case of an unconned or partially unbounded ow, the
boundaries of the ow include the whole or part of a spherical surface of large radius R that lies
inside the uid. As the radius of the large surface tends to innity, the corresponding integral
vanishes provided that the vorticity decays faster than 1/d, where d is the distance from the origin.
Complementary velocity
When the boundary integral on the right-hand side of (2.7.18) vanishes, B = 0 and (2.7.5) shows
that the complementary velocity eld, w C, is irrotational. In that case, w can be expressed
148 Introduction to Theoretical and Computational Fluid Dynamics
as the gradient of a potential function, , so that w = . To ensure that the velocity eld u is
solenoidal, we require that satises Laplaces equation, 2 = 0. In the more general case where
the boundary integral on the right-hand side of (2.7.18) does not vanish, we write
w = w + , (2.7.19)
n=n (2.7.20)
over B, where n is the unit vector normal to B pointing into the ow. A nonsingular solenoidal
eld that satises (2.7.20) exists only if the ow rate of across the boundary B is zero, which
is always true: using (2.7.20) and recalling that the vorticity eld is solenoidal, = 0, we nd
that
n dS = n dS = dV = 0. (2.7.21)
B B F low
2 C = ( B + C). (2.7.22)
where r = |x x | (Problem 2.7.4). Expression (2.7.23) relates the complementary vector potential,
C, to the extended eld, , through a BiotSavart integral. We may then identify with the
extension of the vorticity eld outward from the domain of ow into the boundary. The extension
can be implemented arbitrarily, subject to the constraint imposed by (2.7.20).
Taking the curl of (2.7.23) and repeating the manipulations that led us from (2.7.10) to
(2.7.13), we derive two equivalent expressions for the complementary velocity eld,
1 1
w(x) = (x ) x dV (x ) (2.7.24)
4 VB r 3
2.7 Representation of an incompressible ow in terms of the vorticity 149
and
1 1 1 1
w(x) = (x ) dV (x ) n(x ) (x ) dS(x ), (2.7.25)
4 VB r 4 B r
where the normal unit vector, n, points into the ow. If the extended vorticity is irrotational, the
volume integral on the right-hand side of (2.7.25) vanishes, leaving an expression for the complemen-
tary ow in terms of a surface integral over the boundary, B, involving the tangential component of
alone. One way to ensure that is irrotational is to set = H, where the function H satises
Laplaces equation, 2 H = 0, and then compute H by solving Laplaces equation inside VB subject
to the Neumann boundary condition (2.7.20).
We recall that the normal unit vector, n, points into the ow.
150 Introduction to Theoretical and Computational Fluid Dynamics
1 uj (x )
2 uj (x )
1
= uj (x ) (2.7.32)
xj r xi xi xj r xj xi r
and use the continuity equation to simplify the rst integral on the right-hand side, obtaining
1 uj (x )
1
1
= uj (x ) u j (x ) . (2.7.33)
xj r xi xi xj r xj xi r
Substituting this expression into the volume integral of (2.7.31) and using the divergence theorem
to convert it to a surface integral, we obtain
1 1
The rst and third integrands can be combined to form a double cross product, resulting in an
expression for the complementary ow in terms of the tangential and normal components of the
boundary velocity,
1
1 1 1
The rst boundary integral on the right-hand side of (2.7.35) expresses the velocity eld due to a
vortex sheet with strength u n wrapping around the boundaries of the ow. Since the strength of
the vortex sheet is equal to the tangential component of the velocity, the purpose of the vortex sheet
is to annihilate the tangential component of the boundary velocity. The second integral expresses a
boundary distribution of point sources whose strength is equal to the normal velocity .
2.7.3 Two-dimensional ow
The counterpart of expression (2.7.7) for two-dimensional ow in the xy plane is
1 r
B(x) = ln z (x ) ez dA(x ), (2.7.36)
2 F low a
2.7 Representation of an incompressible ow in terms of the vorticity 151
where a is a constant length and ez is the unit vector along the z axis. Note that both the vorticity,
, and vector potential, B, are oriented along the z axis. Since the vortex lines do not cross the
boundaries of the ow, the divergence of B is zero and the velocity eld is given by
1 1
u(x) = 2
(x ) x dA(x ) + H(x), (2.7.37)
2 F low r
where x = x x and H is a two-dimensional harmonic function, 2 H = 0. An alternative integral
representation for the velocity is
1 r
u(x) = ln n(x ) (x ) dl(x )
2 C a (2.7.38)
1 r
ln (x ) dA(x ) + H(x),
2 F low a
where C is the boundary of the ow. In the case of innite ow with no interior boundaries, both
the boundary integral and the gradient H on the right-hand side of (2.7.38) vanish.
Problems
2.7.3 Complementary ow
(a) Conrm that (2.7.23) is a particular solution of (2.7.22).
(b) Derive (2.7.29) from (2.7.28).
u = B = ( B) 2 B. (2.7.40)
1 1 1
A(x) = B(x) = u(x ) dV (x ) = u(x ) dV (x ). (2.7.42)
4 F low r 4 F low r
1 1 1
A(x) = u(x ) dV (x ) dV (x ). (2.7.43)
4 F low r 4 F low r
Finally, we invoke the denition of the vorticity and use the divergence theorem to nd
1 1 1 1
A(x) = (x ) dV (x ) + n(x ) u(x ) dS(x ), (2.7.44)
4 F low r 4 B r
where the normal unit vector, n, points into the ow. Based on (2.7.44), we obtain a general
expression for the velocity eld,
1 1
1
x
u(x) = (x ) x dV (x ) + n(x ) u(x ) 3 dS(x ). (2.7.45)
4 F low r 3 4 B r
Note that the second integral on the right-hand side is a simplied version of the right-hand side of
(2.7.35). Show that
1 1
B(x) = u(x ) n(x ) dS(x ) = 0, (2.7.46)
4 B r
of either a potential function or a vector potential. In this section, we consider a ow that is neither
incompressible nor irrotational.
HodgeHelmholtz decomposition
If the domain of ow is innite, we require that the velocity, u, decays at a rate that is faster than
1/d, where d is the distance from the origin. This constraint ensures that the rate of expansion
and vorticity decay at a rate that is faster than 1/d2 . If the velocity does not decay at innity, we
subtract out the nondecaying far-eld component and consider the remaining decaying ow. Under
these assumptions, the velocity eld is subject to the fundamental theorem of vector analysis, also
known as the Hodge or Helmholtz decomposition theorem, stating that u can be resolved into two
constituents,
u = + w, (2.8.1)
w = 0. (2.8.2)
This last constraint allows us to express w in terms of a vector potential, A, and therefore recast
(2.8.1) into the form
u = + A. (2.8.3)
= u = w. (2.8.4)
u = 2 , (2.8.5)
which, along with (2.8.4), shows that the rate of expansion of the ow with velocity and the
vorticity of the ow with velocity w are identical to those of the ow u.
Integral representation
Combining (2.8.1) with (2.1.15), (2.7.9), and (2.7.19), we derive a representation for the velocity in
terms of the rate of expansion, the vorticity, and an unspecied irrotational and solenoidal velocity
eld described by a harmonic potential, H,
1 x 1 1
u(x) = 3
(x ) dV (x ) + 3
(x ) x dV (x ) + v(x) + H(x), (2.8.6)
4 F low r 4 F low r
where (x ) = u(x ) is the rate of expansion. The complementary rotational eld, v, can be
computed by extending the vortex lines outward from the domain of ow across the boundaries, as
154 Introduction to Theoretical and Computational Fluid Dynamics
discussed in Section 2.7. In the case of unbounded ow that decays at innity, the last two terms
on the right-hand side of (2.8.6) do not appear.
Specifying the distribution of the rate of expansion, u, and vorticity, , throughout the
domain of ow denes the rst three terms on the right-hand side of (2.8.6). Consequently, the
gradient, H, and thus the velocity, u, are dened uniquely by prescribing the boundary distribution
of the normal velocity component, u n. An important consequence is that it is not generally
permissible to arbitrarily specify the distribution of both the rate of expansion and vorticity in a
ow while requiring more than one scalar condition at the boundaries.
Problems
2.9.1 Two-dimensional ow
Examining the streamline pattern of a two-dimensional incompressible ow, we nd that the ow
rate across an arbitrary line, L, that begins at a point, A, on a particular streamline and ends
2.9 Stream functions for incompressible uids 155
(a) (b)
y
y
B B
2 2 n
L
A
L
1 x
n
1
A
x z
Figure 2.9.1 Illustration of two streamlines in (a) a two-dimensional or (b) axisymmetric ow, used
to introduce the stream function.
at another point, B, on another streamline is constant, independent of the precise location of the
points A and B along the two streamlines, as illustrated in Figure 2.9.1(a). Consequently, we may
assign to every streamline a numerical value of a function, , so that the dierence in the values
of corresponding to two dierent streamlines is equal to the instantaneous ow rate across any
line that begins at a point A on the rst streamline and ends at another point B on the second
streamline. Accordingly, we write
B
2 1 = u n dl, (2.9.1)
A
where the integral is computed along the line L depicted in Figure 2.9.1(a). The right-hand side of
(2.9.1) expresses the ow rate across L. Since the streamlines ll up the entire domain of a ow, we
may regard a eld function of position, x and y, and time t, called the stream function.
Next, we consider two adjacent streamlines and apply the trapezoidal rule to approximate the
integral in (2.9.1), obtaining the dierential relation
d = ux dy uy dx. (2.9.2)
Since the right-hand side of (2.9.2) is a complete dierential, we may write
ux = , uy = , (2.9.3)
y x
which can be recast into the compact vector form
u = ( ez ), (2.9.4)
where ez is the unit vector along the z axis. Equation (2.9.4) suggests that a vector potential of a
two-dimensional ow is
A = ez . (2.9.5)
156 Introduction to Theoretical and Computational Fluid Dynamics
It is evident from the denitions (2.9.3) that u = 0, which shows that the gradient is
perpendicular to the streamlines. Accordingly, the stream function is constant along the streamlines
of a steady or unsteady ow.
In summary, we have succeeded to express the velocity eld and vector potential of a two-
dimensional ow in terms of a single scalar function, the stream function, . Using the denitions
(2.9.3), we nd that the stream function of a certain ow is determined uniquely up to an arbitrary
constant.
Point source
The stream function associated with a two-dimensional point source of strength m located at a
point, x0 , introduced in (2.1.29), is given by
m
(x) = , (2.9.6)
2
where is the polar angle subtended between the vector x x0 and the x axis. This example makes
it clear that, when the domain of ow contains point sources or point sinks or is multiply connected
and the ow rate across a surface that encloses a boundary is nonzero, the stream function is a multi-
valued function of position. An example is the ow due to the radial expansion of a two-dimensional
bubble.
Poisson equation
Taking the curl of (2.9.4), we conrm that the vorticity is parallel to the z axis, = z ez . The
nonzero component of the vorticity is the negative of the Laplacian of the stream function,
z = 2 . (2.9.7)
Using the Poisson formula to invert (2.9.7), we derive an expression for the stream function in terms
of the vorticity,
1 r
(x) = ln z (x ) dA(x ) + H(x), (2.9.8)
2 F low a
where r = |x x |, a is a specied length, and H is a harmonic function in the xy plane. In the case
of innite ow with no interior boundaries, H is a constant usually set to zero. It is worth observing
that (2.9.5) and (2.9.8) are consistent with the more general form (2.7.36).
1 1
1 2
ur = , u = , z = r 2 . (2.9.9)
r r r r r r 2
Substituting expression (2.9.6), we conrm the radial direction of the velocity due to a point source.
2.9 Stream functions for incompressible uids 157
2.9.2 Axisymmetric ow
To describe an axisymmetric ow without swirling motion, we introduce cylindrical coordinates,
(x, , ), as illustrated in Figure 2.9.1(b). Working as in the case of two-dimensional ow, we
assign to every streamline a numerical value of a scalar function, (x, , t), called the Stokes stream
function, so that the dierence in the values between any two streamlines is proportional to the
instantaneous ow rate across an axisymmetric surface whose trace in a meridional plane begins at
a point A on one streamline and ends at another point B on the second streamline. By denition,
B
2 1 = u n dl. (2.9.10)
A
Multiplying the integral on the right-hand side by 2 produces the ow rate through an axisymmetric
surface whose trace in an azimuthal plane is the line L shown in Figure 2.9.1(b).
Next, we consider two streamlines that lie in the same azimuthal plane and are separated by
a small distance, and apply the trapezoidal rule to express (2.9.10) in the dierential form
d = ux d u dx, (2.9.11)
1 1
ux = , u = . (2.9.12)
x
Combining these equations, we derive the vector form
u= e , (2.9.13)
where e is the azimuthal unit vector. Thus,
A= e (2.9.14)
is the vector potential of an axisymmetric ow in the absence of swirling motion. Equations (2.9.12)
show that, u = 0, which demonstrates that the gradient of the stream function, , is perpen-
dicular to the streamlines of the ow. Accordingly, the Stokes stream function, , is constant along
the streamlines of an axisymmetric ow.
In summary, we have managed to express the velocity eld and vector potential of an axisym-
metric ow in terms of a single scalar function, the Stokes stream function, . It is clear from the
denitions (2.9.12) that is determined uniquely up to an arbitrary scalar constant. If the domain
of ow is multiply connected, can be a multi-valued function of position. An example is provided
by the ow due to the expansion of a toroidal bubble.
It is worth noting that the stream function of two-dimensional ow has units of velocity
multiplied by length, whereas the Stokes stream function of axisymmetric ow has units of velocity
multiplied by length squared.
158 Introduction to Theoretical and Computational Fluid Dynamics
Point source
As an example, the Stokes stream function associated with a three-dimensional point source with
strength m located at the point x0 , introduced in (2.1.29), is given by
m m x x0
(x) = cos = , (2.9.15)
4 4 |x x0 |
where is the polar angle subtended between the vector x x0 and the x axis.
Vorticity
Taking the curl of (2.9.13), we nd that the vorticity is oriented in the azimuthal direction, = e ,
where e is the corresponding unit vector. The azimuthal component of the vorticity is given by
1 2 1 2 2 1
= E = + , (2.9.16)
x2 2
where E 2 is a second-order linear dierential operator dened as
2 2 1
E2 2
+ 2
. (2.9.17)
x
The inverse relation, providing us with the stream function in terms of the vorticity, is discussed in
Section 2.12.
1 2 1 2 1 2 cot
= E = + , (2.9.19)
sin r2 r2 2 r 2
where the second-order dierential operator E 2 was dened in (2.9.17). In spherical polar coordi-
nates,
2 sin 1
2 1 2 cot
E2 2
+ 2
= 2
+ 2 2
2 , (2.9.20)
r r sin r r r
involving derivatives with respect to the radial distance, r, and meridional angle, .
2.9.3 Three-dimensional ow
Finally, we consider the most general case of a genuine three-dimensional ow. Inspecting the
streamline pattern, we identify two distinct families of stream surfaces, where each family lls the
2.9 Stream functions for incompressible uids 159
n
2 x
D
z 2
1
1
Figure 2.9.2 Illustration of two pairs of stream surfaces in a three-dimensional ow used to dene the
stream functions.
entire domain of ow, as illustrated in Figure 2.9.2. A stream surface consists of all streamlines
passing through a specied line.
Focusing on a stream tube that is conned between two pairs of stream surfaces, one pair in
each family, we note that the ow rate, Q, across any open surface that is bounded by the stream
tube, D, is constant. We then assign to the four stream surfaces the labels 1 , 2 , 1 , and 2 , as
shown in Figure 2.9.2, so that
Q= u n dS = (2 1 )(2 1 ). (2.9.21)
D
In this light, and emerge as eld functions of space and time, called the stream functions. Now
using Stokes theorem, we write
() n dS = t dl = (2 1 )(2 1 ) = Q, (2.9.22)
D C
where C is the contour enclosing D and t is the unit vector tangent to C pointing in the counter-
clockwise direction with respect to n. Comparing (2.9.22) with (2.9.21) and remembering that the
surface D is arbitrary, we write
u = () = = (). (2.9.23)
Since the gradients and are normal to the corresponding stream tubes, the expression
u = underscores that the intersection between two stream tubes is a streamline. Our
analysis has revealed that
A = , A = , (2.9.24)
where L = I 2 + and I is the identity matrix. Because each term on the right-hand side of
(2.9.25) represents a solenoidal vector eld, the vorticity eld is solenoidal for any choice of and
(Problem 2.9.2).
Problems
a3 1 a3
ur = U cos 1 3 , u = U cos 2 + 3 , (2.9.26)
r 2 r
where U and a are two constants. Verify that the velocity eld is solenoidal, compute the vorticity,
and discuss the physical interpretation of this ow.
2.10 Flow induced by vorticity 161
Computer Problem
2.9.6 Streamlines
Draw the streamlines of the ows described in (a) Problem 2.9.2 and (b) Problem 2.9.5.
or
1 1
u(x) = (x ) dV (x ), (2.10.3)
4 F low r
1
+ (x ) (x x0 ) 3
dV (x ) + . (2.10.4)
4 F low r x =x
0
Next, we examine the physical interpretation of the two leading terms on the right-hand side.
162 Introduction to Theoretical and Computational Fluid Dynamics
Rotlet or vorton
The rst term on the right-hand side of (2.10.4) represents the ow due to a singularity called a
rotlet or vorton, located at the point x0 . The strength of this singularity is equal to the integral of
the vorticity over the domain of ow. Recalling that the vorticity eld is solenoidal, we write
dV = ( x) dV = x ( n) dS, (2.10.5)
F low F low S
where S is a closed surface with large size enclosing the vortex region. Assuming that the vorticity
decays fast enough for the last integral in (2.10.5) tends to zero as S expands to innity, we nd
that the leading term on the right-hand side of (2.10.4) makes a zero contribution.
Point-source dipole
Concentrating on the second term on the right-hand side of (2.10.4), we change the variable of
dierentiation in the gradient inside the integrand from x to x, while simultaneously introducing a
minus sign, and obtain
x
1
u(x) = [ x ] (x ) dV (x ), (2.10.6)
4 F low |x|3
where x = x x0 and x = x x0 . In index notation,
1 2 1
where n is the normal unit vector over the boundaries pointing outward from the ow. Assuming
that the vorticity over the large surface S decays suciently fast so that the last integral in (2.10.8)
vanishes, we nd that the integral on the right-hand side of (2.10.7) is antisymmetric with respect
to the indices i and l, and write
1
xi l (x ) dV (x ) = mil mnk xn k (x ) dV (x ). (2.10.9)
F low 2 F low
Substituting the left-hand side of (2.10.9) for the integral on the right-hand side of (2.10.7), con-
tracting the repeated multiplications of the alternating matrix, noting that the function 1/|x x0 |
satises Laplaces equation at every point except x0 , and switching to vector notation, we nally
obtain
1 1
u(x) = d , (2.10.10)
4 |x x0 |
where
1
d= x (x ) dV (x ). (2.10.11)
2 F low
2.10 Flow induced by vorticity 163
Expression (2.10.10) shows that, far from the vortex, the ow is similar to that due to a point-source
dipole located at the point x0 whose strength is proportional to the moment of the vorticity. Cursory
inspection of the volume and surface integrals in the preceding expressions reveals that (2.10.10) is
valid, provided that the vorticity decays at least as fast as 1/|x x0 |4 ([24], p. 520).
Next, we use the divergence theorem to convert the volume integral of the second term inside the
integral on the right-hand side into an integral over a large surface, S . Assuming that the velocity
decays suciently fast for the integral to vanish, we obtain
1
K= A dV. (2.10.14)
2 F low
The vector potential, A, can be expressed in terms of the vorticity distribution using (2.10.1).
In the case of axisymmetric ow, we express the vector potential, A, in terms of the Stokes
stream function, , and write dV = 2dx d to obtain
K = (x, ) (x, ) dx d, (2.10.15)
F low
where is the azimuthal component of the vorticity. An analogous expression for two-dimensional
ow is discussed in Section 11.1.
A useful representation of the kinetic energy in terms of the velocity and vorticity emerges by
using the identity
1 1
[ u (u x) (u u) x ] + u u = u (x ) (2.10.16)
2 2
([24], p. 520). Solving for the last term on the left-hand side, substituting the result into the rst
integral of (2.10.12), and using the divergence theorem to simplify the integral, we nd that
K= u (x ) dV. (2.10.17)
F low
The velocity can be expressed in terms of the vorticity distribution using (2.10.2) or (2.10.3).
164 Introduction to Theoretical and Computational Fluid Dynamics
The irrotational ow induced by the vortex sheet can be described in terms of a velocity
potential, . In Section 10.6, we will see that can be represented in terms of a distribution of
point-source dipoles oriented normal to the vortex sheet. Combining equations (10.6.1) written for
the free-space Greens function and equation (10.6.8), we obtain
1 1
(x) = 3
x n(x ) (+ )(x ) dS(x ), (2.10.20)
4 Sheet r
where n is the unit vector normal to the vortex sheet pointing into the upper side corresponding to
the plus sign. Taking the gradient of (2.10.20), integrating by parts, and using the denition of ,
we recover (2.10.18).
Problems
2.10.1 Impulse
The impulse required to generate
the motion of a uid with uniform density is expressed by the
momentum integral P = u dV . Show that two ows with dierent vorticity distributions but
identical dipole strengths, d, require the same impulse.
2.10.2 Far ow
Derive the second-order term in the asymptotic expansion (2.10.5) and discuss its physical interpre-
tation. Comment on the asymptotic behavior of the ow when the coecient of the dipole vanishes.
2.10.3 Identities
Prove identities (2.10.8) and (2.10.16).
(a) (b)
n
x
D D Side
n
n Sph
t L t
Figure 2.11.1 (a) Illustration of a closed line vortex, L, and a surface bounded by the line vortex,
D. (b) The velocity potential at a point, x, is proportional to the solid angle subtended by the line
vortex, .
the ow induced by a line vortex, L, with strength , illustrated in Figure 2.11.1(a). To compute the
vector potential, A, and associated velocity eld, u, we substitute the vorticity distribution (1.13.1)
into the BiotSavart integrals (2.10.1) and (2.10.2), and nd that
1 1
A(x) = t(x ) dl(x ), u(x) = 3
t(x ) x dl(x ), (2.11.1)
4 L r 4 L r
where x = x x , r = |x|, and t is the unit tangential vector along the line vortex.
Using (2.10.11) and applying Stokes theorem, we nd that the associated coecient of the
dipole prevailing in the far eld is given by
d= (x x0 ) t(x) dl(x) = n(x) dS(x), (2.11.2)
2 L D
where x0 is an arbitrary point, D is an arbitrary closed surface bounded by the line vortex, and n
is the unit vector normal to D, as shown in Figure 2.11.1(a).
To derive an expression for the potential, we consider the closed line vortex depicted in Figure
2.11.1(a), express the velocity given in the second equation of (2.11.1) in index notation, and apply
Stokes theorem to convert the line integral along the line vortex into a surface integral over an
arbitrary surface D that is bounded by the line vortex, obtaining
xn xn
Expanding the product of the alternating tensors and recalling that 1/r satises Laplaces equation
in three dimensions, everywhere except at the point x , we obtain
1
The right-hand side of (2.11.5) expresses the velocity as the gradient of the velocity potential
1
x x
(x) = n(x ) dS(x ) = n(x ) dS(x ), (2.11.6)
4 D r 4 D r3
which is consistent with the far-eld expansion (2.10.10) for a more general ow.
Solid angle
A geometrical interpretation of the potential, , emerges by introducing a conical surface that
contains all rays emanating from a specied eld point, x, and passing through the line vortex,
as illustrated in Figure 2.11.1(b). We dene a control volume that is bounded by the surface D,
a section of a sphere with radius R centered at the point x and conned by the conical surface,
denoted by Sph, and the section of the conical surface contained between the spherical surface and
the line vortex, denoted by Side. Departing from the integral representation (2.11.6) and using the
divergence theorem, we nd that
x x
(x) = n(x ) dS(x ). (2.11.7)
4 Sph,Side r 3
Because the normal vector is perpendicular to the vector x x , the integral over the conical side
surface Side is identically zero. The normal vector over the spherical surface Sph is given by
n= (x x), (2.11.8)
R
where = 1 if the spherical surface is on the right or left side of D, and = 1 otherwise. For the
conguration depicted in Figure 2.11.1(b), we select = 1. Equation (2.11.7) then yields
x x x x
(x) = dS(x ) = dS(x ) (2.11.9)
4 Sph r3 R 4 R2 Sph
or
(x) = (x), (2.11.10)
4
where (x) is the solid angle subtended by the line vortex at the point x. For the conguration
depicted in Figure 2.11.1(b), the solid angle is positive.
2.11 Flow due to a line vortex 167
x e
t
x
l b
Figure 2.11.2 Closeup of a line vortex illustrating a local coordinate system with the x, y, and z axes
parallel to the tangential, normal, and binormal vectors. For the conguration shown, the curvature
of the line is positive by convention.
The solid angle, , and therefore the scalar potential , is a multi-valued function of position.
In the case of ow due to line vortex ring, changes from 2 to 2 as the point x crosses the
plane of the ring through the interior. This means that undergoes a corresponding discontinuity
equal to , in agreement with our discussion in Section 2.1 regarding the behavior of the potential
in multiply connected domains.
Next, we consider the velocity at a point x that lies in the normal plane containing n and b,
at the position x = e, where e is a unit vector that lies in the normal plane and is the distance
from the x axis. Using the integral representation for the velocity (2.11.1), we obtain
e x
u(x) = t(x ) dl(x ), (2.11.11)
4 L r3
To desingularize the last two integrals, we expand the position vector, x , and tangential vector
along the line vortex, t(x ), in a Laurent series with respect to arc length, l, measured from the
origin in the direction of t, and retain two leading terms to obtain
x
1 2 x
2 t
x = l+ l + , t(x ) = t0 + l + . (2.11.13)
l 0 2 l2 0 l 0
Considering the rst integral in (2.11.12), we truncate the integration domain at l = a and
use the expansion for the tangent vector shown in (2.11.15) to obtain
a a a
t(x ) 1 l
I(x) e 3
dl(x ) = e t0 3
dl(x ) c0 e n 0 3
dl(x ) + , (2.11.16)
a r a r a r
where l/. Evaluating the integrals in the limit as tends to zero and correspondingly a/
tends to innity, we nd that
e t0
I(x) = 2 + , (2.11.18)
2
where the three dots denote terms that increase slower that 1/ 2 as 0.
yielding
1 a
l2
J (x) = c0 b0 dl(x ) + , (2.11.20)
2 a r3
2.11 Flow due to a line vortex 169
Finally, we substitute (2.11.18) and (2.11.22) into (2.11.12) and obtain an expression for the
velocity near the line vortex, rst derived by Luigi Sante Da Rios in 1906 (see [339]),
a
u(x) = t0 e c0 b0 ln + . (2.11.23)
4 4
The rst term on the right-hand side of (2.11.23) describes the expected swirling motion around
the line vortex, which is similar to that around a point vortex in two-dimensional ow discussed in
Section 2.13.1. In the limit as /a tends to zero, the second term diverges, showing that the self-
induced velocity of a curved line vortex (c0 = 0 is innite. This singular behavior reects the severe
approximation involved in the mathematical fabrication of singular vortex structures. In Section
11.10.1, we will discuss the regularization of (2.11.16) accounting for the nite size of the vortex
core.
If the product of the curvature and the binormal vector, cb, is constant along the line vortex,
the line vortex translates as a rigid body. Examples include a circular vortex ring, an advancing
helical vortex advancing rotating about its axis, and a planar nearly rectilinear line vortex with
small amplitude sinusoidal undulations rotating as a rigid body about the centerline (Problems
2.11.2, 2.11.3). Numerical methods for computing the motion of line vortices based on the LIA are
discussed in Section 11.10.1.
Da Rios Equations
Da Rios (1906, see [339]) used the LIA to derive a coupled nonlinear system of ordinary dierential
equations governing the evolution of the curvature and torsion of a line vortex. To simplify the
notation, we introduce a scaled time, t t ln /(4), and recast (2.11.24) into the form
V X = c b, (2.11.25)
170 Introduction to Theoretical and Computational Fluid Dynamics
where a dot denotes a derivative with respect to t. The FrenetSerret relations derived in Section
1.6.2 state that
t = c n, b = n, n = c t + b, (2.11.26)
where is the torsion of the line vortex and a prime denotes a derivative with respect to arc length
along the line vortex, l.
Equations (1.6.34) and (1.6.40) provide us with evolution equations for the curvature and
torsion. In the present notation, these equations take the form
c = 2c V t V n,
1 c 1
= V ( t + cb) + V ( n + b) V b. (2.11.27)
c c c
Dierentiating (2.11.25) and using the Frenet-Serret relations, we nd that
V = c b + c n, V = c2 t + (2c + c ) n + b,
V b = (2c + c ) + ,
(2.11.28)
where we have dened
c 2 c . (2.11.29)
Substituting these expressions into (2.11.27) and simplifying, we obtain the Da Rios equations
c = 2c c (2.11.30)
and
c 1 1 2
= c c + 2
= c . (2.11.31)
c c 2 c
Equation (2.11.30) can expressed in the form
Dc2
= 2 (c2 ) . (2.11.32)
Dt
Integrating with respect to arc length along a closed line vortex provides us with a geometrical
conservation law [33].
Schrodinger equation
Hasimoto [171] discovered that the local induction approximation can be reformulated as a nonlinear
Schrodinger equation for a properly dened complex scalar function. To demonstrate this reduction,
we recall the evolution equations for the tangent, normal, and binormal unit vectors stated in
(1.6.28), (1.6.36), and (1.6.37). In the present notation, these equations read
1
t = V (V t) t, n = (V n) t ( V b ) b,
c
1
b = (V b t + (V b) n. (2.11.33)
c
2.11 Flow due to a line vortex 171
Next, we introduce the complex vector eld q n+i b, where i is the imaginary unit, i2 = 1,
as discussed in Section 1.6.3. The second and third equations in (2.11.34) can be unied into the
complex form
q = i (c + i c ) t + q , (2.11.36)
and then rearranged into an evolution equation for the function Q dened in (1.6.21),
Q = i ( t + Q ). (2.11.37)
Combining this expression with (1.6.20), we derive the evolution equation
2Q ( t)
= = i ( t + Q ) . (2.11.38)
t l t
Expanding the time and arc length derivatives, we obtain
t + t = i ( t + t Q Q ). (2.11.39)
Separating the tangential from the normal components, we obtain
= i ( + ), t = i ( t Q ). (2.11.40)
Making substitutions in the second equation, we obtain
(c b + c n) = i (c + ) n b, (2.11.41)
c c
which requires that
1 2 1
= c c = (c ) = ( ) , i = c2 ( + i ). (2.11.42)
2 2
Integrating the rst equation, we nd that
1
= + A(t) , (2.11.43)
2
where A(t) is a specied function of time, Substituting this expression into the rst equation of
(2.11.40) we derive a nonlinear Schrodinger equation,
1 2 1 2
= 2 + || + A(t) , (2.11.44)
i t l 2
which is known to admit solutions in the form of nonlinear traveling waves called solitons.
172 Introduction to Theoretical and Computational Fluid Dynamics
Problems
1
2 x e
(x , ) dA(x ),
u(x) = d (2.12.1)
4 F low 0 r3
where r = |x x |. Next, we substitute
set e = (0, sin , cos ), and compute the outer vector product, nding
2 cos +
1 1 d (x , ) dA(x ),
u(x) = 3
x cos (2.12.3)
4 F low r
0 x sin
2.12 Axisymmetric ow induced by vorticity 173
u (x) = 1 1
x cos d (x , ) dA(x ). (2.12.5)
4 F low r 3
u 0 x sin
Finally, we substitute
1
r2 = x2 + 2 + 2 2 cos = x2 + ( + )2 4 cos2 ( ), (2.12.6)
2
and integrate to nd that u = 0, as expected, and
! " ! "
ux 1 I31 (x, , ) + I30 (x, , )
(x, ) = (x , ) dA(x ), (2.12.7)
u 4 F low x I31 (x, , )
where
2
cosm
Inm (x, , ) = n/2 d. (2.12.8)
0 x2 + ( + )2 4 cos2 ( 12 )
Working in a similar fashion with the vector potential given in (2.10.1) in terms of the vorticity
distribution, and recalling the A = (/) e , we derive a corresponding representation for the
Stokes stream function,
(x, ) = I11 (x, , ) (x , ) dA(x ), (2.12.9)
4 F low
where x = x x .
where
/2
(2 cos2 1)m
Jnm (k) d, (2.12.11)
0 (1 k 2 cos2 )n/2
= /2, and
4
k2 (2.12.12)
x2 + ( + )2
174 Introduction to Theoretical and Computational Fluid Dynamics
(a) (b)
3.5 3
2
3
2.5
y
y
2
1
1.5 2
3
1 3 2 1 0 1 2 3
0 0.2 0.4 0.6 0.8 1 x
k
Figure 2.12.1 (a) Graphs of the complete elliptic integral of the rst kind, F (k) (solid line) and
second kind, E(k) (broken line), computed by an ecient iterative method. (b) Streamline pattern
in a plane passing through the axis of revolution of a line vortex ring with positive circulation;
lengths have been scaled by the ring radius.
The integrals Jnm can be expressed in terms of the complete elliptic integrals of the rst and
second kind, F and E, dened as
1 /2
d du
F (k) =
(2.12.13)
0 (1 ) (1 k2 2 )1/2
2 1/2
0 1 k 2 sin2 u
and
/2
E(k) 1 k 2 sin2 u du, (2.12.14)
0
where and u are dummy integration variables. Graphs of these integrals are shown in Figure
2.12.1(a). As k tends to unity, F (k) diverges as F (k) ln(4/ 1 k 2 ), whereas E(k) tends to unity.
2 k2 2
J10 (k) = F (k), J11 (k) = F (k) 2 E(k),
k2 k
(2.12.15)
E(k) 2 2 k2
J30 (k) = , J31 (k) = 2 F (k) + 2 E(k)
1 k2 k k (1 k 2 )
2.12 Axisymmetric ow induced by vorticity 175
(e.g., [150], p. 590). Substituting the expression for J11 (k) into (2.12.10) and then into (2.12.9), we
obtain the Stokes stream function
2 k2
1 2
(x, ) = F (k) E(k) (x , ) dA(x ). (2.12.16)
2 F low k k
Computation of the complete elliptic integrals
To evaluate the complete elliptic integrals of the rst and second kind, we may compute the sequence
2
1 (1 Kp1 )1/2
K0 = k, Kp = 2
(2.12.17)
1 + (1 Kp1 )1/2
P =1+ K1 1 + K2 . (2.12.19)
2 2
Alternative polynomial approximations are available (e.g., [2], Chapter 17). In Matlab, complete
elliptic integrals can be computed using the native function ellipke.
Velocity potential
The velocity potential due to a line vortex ring can be deduced from expressions (2.11.6) and (2.11.9)
for line vortices. Identifying the surface D in (2.11.6) with a circular disk of radius bounded by
the vortex ring, we obtain
(x, ) = ring (x), (2.12.22)
4
176 Introduction to Theoretical and Computational Fluid Dynamics
where ring is the angle subtended by the ring from the point x, given by
2
d
(x) = x 2 2 2 2 3/2 d . (2.12.23)
0 0 [ x + ( cos ) + sin ]
where the functions I30 and J30 are dened in (2.12.8), (2.12.10), and (2.12.11), and k 2 is dened
in (2.12.12). Finally, we obtain
E(k)
(x, , ) = 4x 2
d , (2.12.25)
0 1 k [ x 2 + ( + )2 ]3/2
where E(k) is the complete elliptic integral of the second kind. An alternative method of computing
is available [302].
or
2 k2
2
(x, ) = F (k) E(k) dA(x ), (2.12.27)
2 AV k k
where AV is the area occupied by the vortex in an azimuthal plane, dA = dx d is a corresponding
dierential area, x = x x , and the integrals Inm are dened in (2.12.8).
Radial velocity
The radial velocity component is found by dierentiating the stream function, nding
1
u (x, ) = =
I11 (x, , ) 2 dA(x ), (2.12.28)
x 4 AV x
where x = x x . Using the divergence theorem, we derive a simplied expression in terms of an
integral along the trace of the vortex contour in an azimuthal plane, C,
u (x, ) = I11 (x, , ) nx (x ) 2 dl(x ), (2.12.29)
4 C
where n is the normal unit vector pointing outward from the vortex contour.
2.12 Axisymmetric ow induced by vorticity 177
Axial velocity
To develop a corresponding expression for the axial velocity component, we introduce the velocity
potential function
(x, ) = ring (x, , ) dA(x ), (2.12.30)
4 AV
1
ux (x, ) = = ring (x, , ) dA(x ). (2.12.31)
x 4 AV x
To compute the right-hand side of (2.12.32), we introduce a branch cut so that the solid angle
becomes single-valued [302].
To circumvent introducing a branch cut, we resort to the integral representation of the Stokes
stream function and write
1 1
ux (x, ) = = I11 (x, , ) 2 dA(x ). (2.12.33)
4 AV
F G
Shari, Leonard & Fertziger [368] considered the right-hand side of (2.12.1) and expressed the curl
of the vorticity in terms of generalized functions. Their analysis implies that
Taken together, equations (2.12.32) and (2.12.37) provide us with a basis for the contour dynamics
formulation of axisymmetric vortex ow with linear vorticity distribution discussed in Section 11.9.3.
178 Introduction to Theoretical and Computational Fluid Dynamics
(a) (b)
3 3
2 2
1 1
0 0
y
y
1 1
2 2
3 3
3 2 1 0 1 2 3 3 2 1 0 1 2 3
x x
Figure 2.12.2 Streamline pattern associated with Hills spherical vortex (a) in a stationary frame of
reference and (b) in a frame of reference traveling with the vortex.
Hills vortex
Hills spherical vortex provides us with an important example of an axisymmetric vortex with linear
vorticity distribution, = . To describe the ow, we introduce cylindrical polar coordinates
with origin at the center of the vortex, (x, , ), and associated spherical polar coordinates, (r, , ).
In a frame of reference moving with the vortex, the Stokes stream function inside Hills vortex is
given by
2 2
int = (a r2 ), (2.12.38)
10
where a is the vortex radius. Outside the vortex, the stream function is
2 2 a3
ext = a 1 3 , (2.12.39)
15 r
where r is the distance from the origin, r2 = x2 + 2 , x = r cos , and = r sin . We may readily
verify that int = ext = 0 and (int /r) = (ext /r) at r = a, ensuring that the velocity
is continuous across the vortex contour. In fact, the exterior ow is potential ow past a sphere
discussed in Section 7.5.2. Cursory inspection of the exterior ow reveals that the spherical vortex
translates steadily along the x axis with velocity
2
V = a2 , (2.12.40)
15
while maintaining its spherical shape.
The streamline pattern in a stationary frame of reference and in a frame of reference translating
with the vortex is shown in Figure 2.12.2. Comparing the pattern shown in Figure 2.12.2(a) with
2.13 Two-dimensional ow induced by vorticity 179
that shown in Figure 2.12.1(b) for a line vortex ring, we note a similarity in the structure of the
exterior ow and conclude that the particular way in which the vorticity is distributed inside a
vortex plays a secondary role in determining the structure of the ow far from the vortex.
Problem
Computer Problems
where x = x x and the integration is performed over the domain of ow. Using (2.7.36), we nd
that the associated stream function is
1 x2 + y2
(x) = ln z (x ) dA(x ), (2.13.2)
4 a2
where a is a constant length. An alternative representation for the velocity originating from the
integral representation (2.7.38) is
1 x2 + y2
u(x) = ln z (x ) ez dA(x ), (2.13.3)
4 a2
where ez is the unit vector along the z axis.
y x x2 + y 2
ux (x) = , uy (x) = , (x) = ln , (2.13.4)
2 x2 + y 2 2 x2 + y 2 4 a2
where 2 is the two-dimensional delta function, is the strength of the point vortex, and x = x x0 .
The polar velocity component is
1
u (x) = , (2.13.5)
2 r
where r = |x x0 |. The streamlines are concentric circles centered at x0 , the velocity decays like
1/r, and the cyclic constant of the motion around the point vortex is equal to . The velocity eld
is irrotational everywhere except at the singular point, x0 . The associated multi-valued velocity
potential is
(x) = , (2.13.6)
2
where is the polar angle subtended between the vector x x0 and the x axis. We observe that the
harmonic potential increases by each time a complete turn is performed around the point vortex.
Complex-variable formulation
It is sometimes useful to introduce the complex variable z = x + iy, and the complex velocity
v = ux + iuy , where i is the imaginary unit, i2 = 1. The rst two equations in (2.13.4) can be
collected into the complex form
1
v (z) = ux i uy = , (2.13.7)
2i z z0
where an asterisk denotes the complex conjugate.
2.13 Two-dimensional ow induced by vorticity 181
1.5
0.5
2y/a
0
0.5
1.5
2
2 1.5 1 0.5 0 0.5 1 1.5 2
2x/a
Figure 2.13.1 Mesmerizing streamline pattern due to an array of evenly spaced point vortices with
positive (counterclockwise) circulation separated by distance a.
To overcome this diculty, we express the stream function corresponding to the velocity eld
induced by the individual point vortices as
r0 rm
0 (x, y) = ln , m (x, y) = ln (2.13.8)
2 a 2 |m| a
for m = 1, 2, . . ., where
rm [ (x xm )2 + (y ym )2 ]1/2 (2.13.9)
is the distance of the eld point, x = (x, y), from the location of the mth point vortex. The
denominators of the fractions in the arguments of the logarithms on the right-hand sides of (2.13.8)
have been chosen judiciously to facilitate forthcoming algebraic manipulations.
It is important to observe that, as m tends to , the fraction on the right-hand side of the
second equation in (2.13.8) tends to unity and its logarithm tends to vanish, thereby ensuring that
remote point vortices make decreasingly small contributions. If the denominators were not included,
remote point vortices would have made contributions that are proportional to the logarithm of the
distance between a point vortex from the point (x, y) where the stream function is evaluated.
182 Introduction to Theoretical and Computational Fluid Dynamics
Next, we express the stream function due to the innite array as the sum of a judiciously
selected constant expressed by the term after the rst equal sign in (2.13.10), and the individual
stream functions stated in (2.13.8), obtaining
(x, y) = ln( 2) + m (x, y). (2.13.10)
2 m=
where denotes the product. An identity allows us to compute the innite product on the right-hand
side of (2.13.12) in closed form, obtaining
2 r0 # rm
= {cosh[k(y y0 )] cos[k(x x0 )]}1/2 (2.13.13)
a m=1,2,...
|m| a
(e.g., [4], p. 197). Substituting the right-hand side of (2.13.13) into (2.13.12), we derive the desired
expression for the stream function
(x, y) = ln cosh[k(y y0 )] cos[k(x x0 )] , (2.13.14)
4
where k = 2/a is the wave number. Dierentiating the right-hand side of (2.13.14) with respect to
x or y, we obtain the corresponding velocity components,
sinh[k(y y0 )]
ux (x, y) = ,
2a cosh[k(y y0 )] cos[k(x x0 )]
(2.13.15)
sin[k(x x0 )]
uy (x, y) = .
2a cosh[k(y y0 )] cos[k(x x0 )]
As the wave number k tends to zero, expressions (2.13.14) and (2.13.15) reproduce the stream
function and velocity eld associated with a single point vortex. The streamline pattern due to the
periodic array exhibits a cats eye pattern, as illustrated in Figure 2.13.1.
Because of symmetry, the velocity at the location of one point vortex induced by all other
point vortices is zero and the array is stationary. Far above or below the array, the x component of
the velocity tends to the value /a or /a, while the y component decays at an exponential rate.
This behavior renders the innite array a useful model of the ow generated by the instability of a
shear layer separating two parallel streams that merge at dierent velocities. The KelvinHelmholtz
instability causes the shear layer to roll up into compact vortices represented by the point vortices
of the periodic array.
2.13 Two-dimensional ow induced by vorticity 183
ux (x) = 2 (x x 0 ) ,
2 r2 r4
(2.13.17)
1 x x x0
uy (x) = 2 (x x0 ) .
2 r2 r4
The streamline pattern of the ow due to a point-vortex dipole oriented along the y axis, with
= ey and > 0 is identical to that due to a potential dipole oriented along the x axis shown in
Figure 2.1.3(b). An alternative method of deriving the ow due to a point-vortex dipole employes
the properties of generalized delta functions. We set
(x) = (x x0 ), (2.13.18)
and then use (2.13.1) and (2.13.2) to derive (2.13.16) and (2.13.17).
where x = x x , d = dl is the dierential of the circulation along the vortex sheet, and C is the
trace of the vortex sheet in the xy plane. Comparing (2.13.19) with (2.13.2), we interpret a vortex
sheet as a continuous distribution of point vortices.
It is sometimes useful to introduce the complex variable z = x + iy and dene the complex
velocity v = ux + iuy , where i is the imaginary unit, i2 = 1. Equations (2.13.19) may then be
stated in the complex form
1 d(z )
v (z) = , (2.13.20)
2i C z z
where an asterisk denotes the complex conjugate.
184 Introduction to Theoretical and Computational Fluid Dynamics
where k = 2/a is the wave number and T is the trace of the vortex sheet in the xy plane inside
one period. As the scaled wave number ka tends to zero, (2.13.21) reduces to (2.13.19).
where ez is the unit vector along the z axis and the gradient involves derivatives with respect
to the integration point, x . Using Stokes theorem, we convert the area integral into a line integral
along the closed contour of the patch, C, obtaining
x2 + y 2
u(x) = ln t(x ) dl(x ), (2.13.23)
4 C a2
where t is the tangent unit vector pointing in the counterclockwise direction around C. An alterna-
tive way of deriving (2.13.23) departs from the identity
= 2 (x x ) t(x ) dl(x ), (2.13.24)
C
where 2 is the two-dimensional delta function in the xy plane. Substituting (2.13.24) into (2.13.3)
and using the properties of the delta function, we recover (2.13.23). If a ow contains a number of
disconnected vortex patches with dierent vorticity, the integral in (2.13.23) is computed along each
vortex contour and then multiplied by the corresponding values of the constant vorticity, .
(a) (b)
C3
C
U
CL 2
CR C2
1
CB
C1
Figure 2.13.2 Illustration of (a) a periodic vortex layer with constant vorticity and (b) a periodic
compound vortex layer consisting of two adjacent vortex layers with constant vorticity 1 and 2 .
where k = 2/a is the wave number, C is the contour of one arbitrary patch in the array, and t is
the tangent unit vector pointing in the counterclockwise direction.
As an application, we consider the velocity due to the periodic vortex layer illustrated in
Figure 2.13.2(a). We select one period of the vortex layer and identify C with the union of the
upper contour, CU , lower contour, CB , left contour, CL , and right contour, CR . The contributions
from the contours CL and CR to the integral in (2.13.25) cancel because the corresponding tangent
vectors point in opposite directions and the logarithmic function is periodic inside the integral.
Consequently, the contour C reduces to the union of CU and CB .
Problems
We will discuss constitutive equations relating the stresses developing on the surface of a
material uid parcel to the parcel deformation. In the discourse, we will concentrate on a special
but common class of incompressible uids, called Newtonian uids, whose response is described by
a linear constitutive equation. The equation of motion for an incompressible Newtonian uid takes
the form of a second-order partial dierential equation in space for the velocity, called the Navier
Stokes equation. Supplementing the NavierStokes equation with the continuity equation to ensure
mass conservation, and then introducing appropriate boundary and initial conditions, we obtain a
complete set of governing equations. Analytical, asymptotic, and numerical methods for solving
these equations under a broad range of conditions will be discussed in subsequent chapters.
The equation of motion can be regarded as a dynamical law for the evolution of the velocity
eld, providing us with an expression for the Eulerian time derivative, u/t, or material derivative,
Du/Dt. To derive a corresponding law for the evolution of the vorticity eld expressing the rate
of rotation of uid parcels, = u, we take the curl of the NavierStokes equation and derive
the vorticity transport equation. Inspecting the various terms in the vorticity transport equation
allows us to develop insights into the dynamics of rotational ows. Vortex dynamics provides us
with a natural framework for analyzing and computing ows dominated by the presence or motion
of compact vortex structures, including line vortices, vortex patches, and vortex sheets.
186
3.1 Forces acting in a uid, traction, and the stress tensor 187
(a) (b)
n y (y)
(y) f
n (x)
f
f f
n(x)
x
n
(z)
z n
(z)
f
Figure 3.1.1 (a) Illustration of the traction vector exerted on a small surface on the boundary of a
uid parcel or control volume, f , and normal unit vector, n. The traction vector has a normal com-
ponent and a tangential component; the normal component is the normal stress, and the tangential
component is the shear stress. (b) Illustration of three small triangular surfaces perpendicular to
the three Cartesian axes forming a tetrahedral control volume; n is the unit vector normal to the
slanted face of the control volume, and f is the corresponding traction.
dF
f . (3.1.1)
dS
The traction depends on the location, orientation, and designated side of the innitesimal surface
element. The location is determined by the position vector, x, and the orientation and side are
determined by the normal unit vector, n. The traction has units of force divided by squared length.
In terms of the traction, the surface force exerted on a uid parcel is
F= f dS. (3.1.2)
P arcel
188 Introduction to Theoretical and Computational Fluid Dynamics
The same expression provides us with the surface force exerted on a control volume occupied by
uid.
f N = (f n) n, f T = n (f n) = f (I n n), (3.1.3)
where I is the identity matrix. The projection matrix I n n extracts the tangential component
of a vector that it multiplies.
Body force
A long-range ambient force eld acting on the molecules of a uid parcel imparts to the parcel a
body force given by
B
F = b dV, (3.1.4)
P arcel
where b is the strength of the body force eld and is a companion physical constant that may
depend on time as well on position in the domain of ow. Examples include the gravitational or
an electromagnetic force eld. In the case of the gravitational force eld, b is the acceleration of
gravity, g, and is the uid density, . In the remainder of this book we will assume that the body
force eld is due to gravity alone.
The rst row of contains the components of f (x) , the second row contains the components of f (y) ,
and the third row contains the components of f (z) , so that
(x) (x) (x)
xx xy xz fx fy fz
= yx yy yz fx(y) fy(y) fz(y) . (3.1.6)
zx zy zz (z) (z) (z)
fx fy fz
The diagonal elements of are the normal stresses exerted on the three mutually orthogonal sides
of the control volume that are normal to the x, y, and z axes. The o-diagonal elements are the
tangential or shear stresses. Later in this section, we will show that satises a transformation rule
that qualies it as a second-order Cartesian tensor.
3.1 Forces acting in a uid, traction, and the stress tensor 189
Our point of departure is Newtons second law of motion for the uid parcel enclosed by the
tetrahedron, stating that the rate of change of momentum of the parcel is equal to the sum of the
surface and body forces exerted on the parcel. Using (3.1.2) and (3.1.4), we obtain
d
u dV = f dS + g dV. (3.1.7)
dt P arcel P arcel P arcel
Because the size of the tetrahedron is innitesimal, the traction exerted on each side is approximately
constant. Neglecting variations in the momentum and body force over the parcel volume, we recast
equation (3.1.7) into the algebraic form
D( u V )
= f (x) Sx + f (y) Sy + f (z) Sz + f S + g V, (3.1.8)
Dt
where D/Dt is the material derivative, V is the volume of the tetrahedron, S (x) , S (y) , S (z) ,
and S are the surface areas of the four sides of the tetrahedron, and f is the traction exerted on
the slanted side. Dividing each term in (3.1.8) by S and rearranging, we obtain
1 D( u V )
S (x) S (y) S (z)
g V = f (x) + f (y) + f (z) + f. (3.1.9)
S Dt S S S
In the limit as the size of the parcel tends to zero, the ratio V /S vanishes and the left-hand side
disappears.
Now we introduce the unit vector normal to the slanted side pointing outward from the
tetrahedron, n, and use the geometrical relations
f = n . (3.1.11)
fj = ni ij , (3.1.12)
where summation is implied over the repeated index, i. Equation (3.1.11) states that the traction,
f , is a linear function of the normal unit vector, n, with a matrix of proportionality that is equal to
the stress tensor, .
190 Introduction to Theoretical and Computational Fluid Dynamics
Newtons third law requires that the parcel exerts on the ambient uid a force of equal magnitude
in the opposite direction.
The total force exerted on the parcel is the sum of the surface force given in (3.1.13) and the
body force given in (3.1.4),
Ftotal = ( + g) dV. (3.1.15)
P arcel
If the divergence of the stress tensor balances the body force, the total force exerted on the parcel is
zero. Later in this chapter, we will see that this is true when the eect of uid inertia is negligibly
small.
where n is the unit vector normal to the boundary pointing into the uid.
is a tensor
We will show that the matrix of stresses, , is a Cartesian tensor, called the Cauchy stress tensor or
simply the stress tensor. Following the standard procedure outlined in Section 1.1.7, we introduce
3.1 Forces acting in a uid, traction, and the stress tensor 191
two Cartesian systems of axes, xi and xi , related by the linear transformation xi = Aij xj , where A
is an orthogonal matrix, meaning that its inverse is equal to its transpose. The stresses in the xi
system are denoted by and those in the xi system are denoted by . The traction exerted on a
small surface in the xi system is denoted by f , and the same traction exerted on the same surface
in the xi system is denoted by f . Next, we introduce the vector transformation
ni ij
= Ajl nk kl . (3.1.18)
ni ij
= Ajl Aik ni kl , (3.1.19)
which demonstrates that satises the distinguishing property of second-order Cartesian tensors
shown in (1.1.40) with in place of T.
One important consequence of tensorial nature of is the existence of three scalar stress
invariants. In particular, both the trace and the determinant of are independent of the choice of
the working Cartesian axes.
3.1.2 Torque
The torque, T, exerted on a uid parcel, computed with respect to a chosen point, x0 , consists of
the torque due to the surface force, the torque due to the body force, and the torque due to an
external torque eld with intensity c. Adding these contributions, we obtain the total torque
Ttotal = x (n ) dS + x ( g) dV + c dV, (3.1.20)
P arcel P arcel P arcel
where x = x x0 and is an appropriate physical constant associated with the torque eld, c.
For example, a torque eld arises in a suspension of magnetized or bipolar particles by applying an
electrical eld.
Using the divergence theorem, we convert the surface integral on the right-hand side of (3.1.20)
into a volume integral, obtaining in index notation
lk
Titotal = ijk jk + ijk xj + ijk xj gk + ci dV, (3.1.21)
P arcel xl
In the absence of an external torque eld, c = 0, the torque exerted on a parcel with respect
to a point x1 , denoted by Ttotal (x1 ), is related to the torque with respect to another point, x0 , by
When the rate of change of linear momentum of a uid parcel is negligible, the total force exerted
on the parcel vanishes and the torque is independent of the point with respect to which it is dened.
where x = x x0 and n is the unit vector normal to the boundary pointing into the uid. When
the surface force exerted on the boundary is zero, the torque is independent of the location of the
center of torque, x0 .
f () = e , (3.1.24)
where e is the unit vector in the direction of the coordinate line, and summation over the index
is implied on the right-hand side. The stress tensor itself is given by the dyadic decomposition
= e e , (3.1.25)
where the matrices e e provide us with a base of the three-dimensional tensor space, as discussed
in Section 1.1.
Conversely, the components of the stress tensor can be extracted from the stress tensor by
double-dot projection,
= : (e e ). (3.1.26)
The double dot product of two matrices is dened in Section A.4, Appendix A.
3.1 Forces acting in a uid, traction, and the stress tensor 193
(a) (b)
x
x
xx
x
x
x
z
(c) (d)
y r
r
y
rr r r rr
r
r
r
x
x
Figure 3.1.2 (a) Denition of the nine components of the stress tensor in orthogonal curvilinear
coordinates. Components of the stress tensor in (b) cylindrical, (c) spherical, and (d) plane polar
coordinates.
The components of the stress tensor in cylindrical, spherical, and plane polar coordinates are dened
in Figure 3.1.2(bd). Plane polar coordinates arise from spherical polar coordinates by setting = 0,
or from cylindrical polar coordinates by setting x = 0 and relabeling as r and as .
Figure 3.1.3 Denition of the nine components of the stress tensor in nonorthogonal curvilinear coor-
dinates. The solid lines represent covariant coordinates and the dashed lines represent the associated
contravariant coordinates.
ponents, in four combinations. The physical meaning of the pure and mixed representations stems
from the geometrical interpretation of the covariant and contravariant base vectors, combined with
the denition of the traction as the projection from the left of the stress tensor onto a unit vector
pointing in a specied direction.
Problems
f N = [ : (n n)] n. (3.1.27)
The double dot product of two matrices is dened in Section A.4, Appendix A.
ij = ik nk xj dS + xj dV , (3.1.29)
VP P arcel P arcel xk
in dyn/cm2 , where lengths are measured in cm. Evaluate the normal and shear component of the
traction (a) over the surface of a sphere centered at the origin, and (b) over the surface of a cylinder
that is coaxial with the x axis.
. (3.2.2)
Eulerian form
Expressing the material derivative, D/Dt, in terms of Eulerian derivatives, we obtain the Eulerian
form of the equation of motion,
u
+ u u = + g. (3.2.4)
t
The second term on the left-hand side can be regarded as a ctitious nonlinear inertial force. Using
the continuity equation,
+ ( u) = 0, (3.2.5)
t
196 Introduction to Theoretical and Computational Fluid Dynamics
Stressmomentum tensor
It is useful to introduce the stressmomentum tensor,
u u, (3.2.7)
and recast (3.2.6) into the compact form
(u)
= + g. (3.2.8)
t
If the ow is steady, the left-hand side of (3.2.8) is zero and the divergence of the stressmomentum
tensor balances the body force.
where n is the normal unit vector pointing into the control volume. Decomposing the stress
momentum tensor into its constituents and rearranging, we obtain
(u)
dV + (u) (u n) dS = n dS + g dV. (3.2.10)
Vc t D D Vc
Equation (3.2.10) states that the rate of change of momentum of the uid occupying a xed control
volume is balanced by the ow rate of momentum normal to the boundaries, the force exerted on
the boundaries, and the body force exerted on the uid residing inside the control volume. The
integral momentum balance allows us to develop approximate relations between global properties
of a steady or unsteady ow. In engineering analysis, we typically derive expressions for boundary
forces in terms of boundary velocities, subject to rational simplications for inlet and outlet velocity
proles. The formulation can be generalized to describe other transported elds, such as heat or the
concentration of a chemical species (e.g., [36]).
Combining this equation with the continuity equation (3.2.5) and rearranging the right-hand side,
we obtain
1 1
|u|2 + |u|2 u = ( u) : u + g u. (3.2.12)
t 2 2
The double dot product of two matrices is dened in Section A.4, Appendix A.
where f = n is the traction and n is the normal unit vector pointing into the control volume.
The four terms on the right-hand side of (3.2.13) represent, respectively, the rate of supply of kinetic
energy into the control volume by convection, the rate of working of the traction at the boundary of
the control volume, the rate of energy dissipation, and the rate of working against the body force.
We have found that the rate of dissipation of internal energy per unit volume of uid is given by the
double dot product of the stress tensor and velocity gradient tensor, : u.
where n is the normal unit vector pointing into the control volume. If the normal component of the
velocity obeys the no-penetration boundary condition for a translating body, u n = V n, we may
apply the divergence theorem to nd that WB = Vcv g V, where Vcv is the volume of the control
volume and V is a constant velocity. The last term in (3.2.13) may then be identied with the rate
of working necessary to elevate the uid inside the control volume with velocity V.
where X is the position of the point particles occupying the parcel. Taking the material derivative
of the kinetic energy integral, and recalling that, because of mass conservation, D( dV )/Dt = 0, we
express the rate of change of the kinetic energy in terms of the point-particle acceleration,
dK 1 D(u u) Du
= dV = u dV. (3.2.16)
dt 2 P arcel Dt P arcel Dt
Now we use the equation of motion to express the acceleration in terms of the stress tensor and
body force, obtaining
dK
= u ( ) dV + u g dV. (3.2.17)
dt P arcel P arcel
The last integral is equal to dPp /dT . Using the divergence theorem to convert the rst integral on
the right-hand side into a surface integral and rearranging, we derive an energy balance expressed
by the equation
d(K + P)
= u f dS : u dV, (3.2.19)
dt P arcel P arcel
where f = n is the traction and n is the normal unit vector pointing into the parcel.
The rst integral on the right-hand side of (3.2.19) is the rate of working of the traction on
the parcel surface. It then follows from the rst thermodynamic principle that the second integral
expresses the rate of change of internal energy, which is equal to the rate energy dissipation inside
the parcel, I, expended for increasing the temperature of the uid,
dI
= : u dV, (3.2.20)
dt P arcel
which is consistent with the third term on the right-hand side of (3.2.13).
The rst integrand is equal to u u, which is identically zero. Switching to index notation,
replacing the total torque with the right-hand side of (3.1.21) and rearranging, we obtain
Du
k jk
ijk xj gk ilk lk ci dV = 0. (3.2.23)
P arcel Dt xj
Since the volume of the parcel is arbitrary, we may discard the integral sign and use the equation of
motion (3.2.1) to simplify the integrand, nding
ilk lk + ci = 0. (3.2.24)
Multiplying (3.2.24) by imn and manipulating the product of the alternating tensors, we nd that
mn nm = mni ci , (3.2.25)
which shows that, in the absence of an external torque eld, c, the stress tensor must be symmetric,
ij = ji or = T , where the superscript T designates the matrix transpose. In that case, only
three of the six nondiagonal components of the stress tensor are independent, and the remaining
three nondiagonal components are equal to their transpose counterparts. The traction may then be
computed as
f n = n. (3.2.26)
In the remainder of this book, we will tacitly assume that the conditions for the stress tensor to be
symmetric are satised.
Principal directions
The symmetry of the stress tensor in the absence of a torque eld guarantees the existence of three
real eigenvalues and corresponding orthogonal eigenvectors. The traction exerted on an innitesimal
planar surface that is perpendicular to an eigenvector points in the normal direction, that is, it lacks
a shearing component. Setting the Cartesian axes parallel to the eigenvectors renders the stress
tensor diagonal. In the case of an isotropic uid, dened as a uid that has no favorable direction,
the eigenvectors of the stress tensor must coincide with those of the rate-of-deformation tensor, as
will be discussed in Section 3.3.
into the equation of motion provides us with three scalar component equations corresponding to the
chosen coordinates.
To derive the plane polar components of the hydrodynamic volume force, = , we express the
gradient and stress tensor in the corresponding forms
1
= er + e , = rr er er + r er e + r e er + e e , (3.2.27)
r r
and compute
1
= = er + e . (3.2.28)
r r
rr r 1 r
= er + e + er + e
r r r
(e e ) 1 (e e )
+ er + e , (3.2.29)
r r
where summation is implied over the repeated indices, and , standing for r or . Expanding the
derivatives of the products and grouping similar terms, we obtain
1 r
1
rr r
= + er + + e
r r r r
e e 1 e 1 e
+ er e + e r e + e e + e e . (3.2.30)
r r r r
Now we recall that all derivatives e / are zero, except for two derivatives,
er de
= e , = er , (3.2.31)
d
and nd that
1 r
1
1 er 1 e
rr r
= + er + + e + r e e + , (3.2.32)
r r r r r r
where summation is implied over the repeated index, . Simplifying, we derive the expressions given
in Table 3.2.1(c).
Expressions for the components of the hydrodynamic volume force in cylindrical and spherical
polar coordinates are collected in Table 3.2.1(a, b).
3.2 Cauchy equation of motion 201
(a) xx 1 (x ) 1 x
x = + +
x
x 1 ( ) 1
= + +
x
x 1 ( 2 ) 1
= + 2 +
x
(b)
1 (r2 rr ) 1 (r sin ) 1 r +
r = 2
+ +
r r r sin r sin r
1 (r2 r ) 1 1 r + 2 cot
= 2
+ + +
r r r r sin r
(c)
1 (rrr ) 1 r 1 (r2 r ) 1
r = + , = +
r r r r r2 r r
Table 3.2.1 (b) The x, , and components of the hydrodynamic volume force in cylindrical polar
coordinates. (b) The r, , and components of the hydrodynamic volume force in spherical
polar coordinates. (c) The r and components of the hydrodynamic volume force in plane polar
coordinates.
Velocity
Our rst task is to compute the velocity of a point particle in an inertial frame in terms of its
coordinates in a noninertial frame. We begin by introducing three unit vectors, e1 , e2 , and e3 ,
associated with the noninertial coordinates, y1 , y2 , and y3 , and describe the position of a point
202 Introduction to Theoretical and Computational Fluid Dynamics
Acceleration
The acceleration of a point particle in the inertial system is
Du D2 X
a(X) = . (3.2.38)
Dt Dt2
Taking the material derivative of (3.2.36), we obtain
dU D2 Yi DYi dei DYi d dei
a(X) = + ei + +( ei ) + Y + (Yi ). (3.2.39)
dt Dt2 Dt dt Dt dt dt
The second term on the right-hand side represents the acceleration of the point particles in the
noninertial frame,
D2 Yi
a(Y) = ei . (3.2.40)
Dt2
Using the second relation in (3.2.34) to simplify the third and last terms on the right-hand side of
(3.2.39), we nd that
dU d
a(X) = + a(Y) + 2 v + Y + ( Y). (3.2.41)
dt dt
The right-hand side involves position, velocity, and acceleration in the noninertial system alone.
3.2 Cauchy equation of motion 203
xy
y Centrifugal force
Coriolis force
Figure 3.2.1 Illustration of the Coriolis and centrifugal forces on the globe due to the rotation of the
earth.
Equation of motion
Substituting the right-hand side of (3.2.41) for the point-particle acceleration into the equation of
motion (3.2.3), and rearranging, we derive a generalized equation of motion in the noninertial frame,
Dv
= + g + fI, (3.2.42)
Dt
where v is the velocity eld in the noninertial frame,
dU d
f I + 2 v + ( y) + y (3.2.43)
dt dt
is a ctitious inertial force per unit volume of uid, and y is the position in the noninertial frame.
The ctitious inertial force consists of (a) the linear acceleration force, dU/dt; (b) the Coriolis
force, 2 v; (c) the centrifugal force, ( y); and (d) the angular-acceleration force,
(d/dt) y.
To understand the physical motivation for this nomenclature, we consider a uid in rigid-body
rotation. Describing the motion in a stationary frame of reference, we identify the term u2 / with
the centripetal point-particle acceleration force, which must be balanced by an opposing radial force.
Describing the motion in a noninertial frame of reference that rotates with the uid, we nd that
the Coriolis force is zero, and the centrifugal force is u2 /. The temptation to interpret the
centripetal acceleration force, in the inertial system as the centrifugal force is then apparent.
Problems
= pth I, (3.3.1)
where I is the identity matrix. The thermodynamic pressure, pth , is a function of the density and
temperature, and depends on the chemical composition of the uid in a manner that is determined
by an appropriate equation of state.
Ideal-gas law
In the case of an ideal gas, the thermodynamic pressure, pth , is related to the density, , by Clapey-
rons ideal gas law,
RT
pth = , (3.3.2)
M
where R = 8.314 103 kg m2 /(s2 kmole K) is the ideal gas constant, T is Kelvins absolute tempera-
ture, which is equal to the Celsius centigrade temperature reduced by 273 units, M is the molar mass,
dened as the mass of one mole comprised of a collection of NA molecules, and NA = 6.022 1026 is
the Avogadro number. The molar mass of an element is equal to the atomic weight of the element
listed in the periodic table expressed in grams.
but also on the history of the motion of all point particles comprising the uid, from inception of
the motion, up to the present time. Leaving aside physiochemical interactions that are independent
of the uid motion, we argue that the stress eld depends on the structure of the velocity eld at all
prior times. This reasoning leads us to introduce a constitutive equation for the stress tensor that
relates the stress at a point at a particular instant t = to the structure of the velocity eld at all
prior times,
The nonlinear functional operator G may involve derivatives or integrals of the velocity with the
respect to space and time. Coecients appearing in the specic functional form of G are regarded
as rheological properties of the uid.
p I + , (3.3.4)
is supported by the observation that rigid-body motion does not generate a deviatoric stress eld.
Replacing the velocity in the arguments of the operator in (3.3.3) with the deformation gradient F
introduced in (1.3.31), we obtain
If the deformation gradient is zero throughout the domain of ow, the uid executes rigid-body
motion.
It can be shown that a simple uid is necessarily isotropic, that is, it has no favorable or unfavorable
directions in space ([365], p. 67).
A purely viscous uid lacks memory, that is, it is inelastic. Enforcing the principle of material
objectivity, we nd that the functional form of the operator G in (3.3.8) must be such that the
antisymmetric part of F drops out and the stress tensor must be a function of the rate-of-deformation
tensor, E 12 (L + LT ) 13 I. We thus obtain a constitutive relation rst proposed by Stokes in
1845,
where fi are functions of the three invariants of the rate-of-deformation tensor introduced in (1.1.46).
Built in equation (3.3.10) is the assumption that the principal directions of the stress tensor are
identical to those of the rate-of-deformation tensor, as required by the stipulation that the uid is
isotropic.
3.3 Constitutive equations for the stress tensor 207
As an example, we consider unidirectional shear ow along the x axis with velocity varying
along the y axis, u = [ux (y), 0, 0]. The shear stress is determined from the constitutive equation
du n du
x x
xy = 0 , (3.3.11)
dy dy
where 0 and n are two physical constants. This scalar constitutive equation corresponds to a
generalized Newtonian uid, called the power-law uid. Setting n = 0 we obtain a Newtonian uid.
f0 = p, f1 = 2, (3.3.12)
where p is the reaction pressure, allowed to be a linear function of the rate of expansion, and the
coecient is a physical constant with dimensions of mass per time per length called the dynamic
viscosity or simply the viscosity of the uid. The viscosity is often measured in units of poise, which
is equal to 1 gr/(cm sec). In general, the viscosity of a gas increases, whereas the viscosity of a liquid
decreases as the temperature is raised. The viscosity of water and air at three temperatures is listed
in the rst column of Table 3.3.1. The constitutive equation for a Newtonian uid thus takes the
linear form
= p I + 2 E. (3.3.13)
Because the trace of the rate-of-deformation tensor E is zero, the trace of is equal to 3p, as
required.
Dilatational viscosity
The reaction pressure in a uid that has been left alone in a macroscopic state of rest for a long
time period of time reduces to the thermodynamic pressure, pth , determined by the local density
of the uid and temperature according to an assumed equation of state. Under ow conditions, we
note the assumed linearity of the stress tensor on the rate-of-deformation tensor and recall that the
uid is isotropic to write
p = pth , (3.3.14)
208 Introduction to Theoretical and Computational Fluid Dynamics
Table 3.3.1 The dynamic viscosity () and kinematic viscosity () of water and air at three temper-
atures; cp stands for centipoise, which is one hundredth of the viscosity unit poise: 1 cp = 0.01
g/(cm sec).
where u is the rate of expansion and is a physical constant with dimensions of mass per
time and length, called the dilatational or expansion viscosity [351]. The Newtonian constitutive
equation then takes the form
= pth I + I + 2 E. (3.3.15)
The coecient 3 is sometimes called the bulk viscosity. The constitutive equation (3.3.15) is known
to describe with high accuracy the stress distribution in a broad range of uids whose molecules
are small compared to the macroscopic dimensions of the ow and whose spatial conguration
is suciently simple. Substituting into (3.3.15) the denition of the rate-of-deformation tensor,
E 12 (L + LT ) 13 I, we obtain
= pth I + I + (L + LT ), (3.3.17)
where L is the velocity gradient tensor, the superscript T denotes the matrix transpose, and
2
= (3.3.18)
3
is the second coecient of viscosity.
Internal energy
Substituting the Newtonian constitutive equation (3.3.15) into (3.2.20), we nd that the rate of
production of internal energy inside a uid parcel is
dIp
= pth dV + 2 dV + 2 E : E dV. (3.3.19)
dt P arcel P arcel P arcel
The rst term on the right-hand side expresses reversible production of energy in the usual sense of
thermodynamics. The second term expresses irreversible dissipation of energy due to the expansion
or compression of the uid, which further justies calling the expansion viscosity. The third term
expresses irreversible dissipation of energy due to pure deformation.
3.3 Constitutive equations for the stress tensor 209
= p I + [u + (u)T ]. (3.3.20)
Explicitly, the nine components of the stress tensor are given by the matrix equation
ux uy ux uz ux
p + 2 ( + ) ( + )
x x y x z
xx xy xz uy
yx ux uy uy uz
yy yz = ( + ) p + 2 ( + ) . (3.3.21)
y x y y z
zx zy zz
ux uz uy uz uz
( + ) ( + ) p + 2
z x z y z
Polar coordinates
Explicit expressions for the components of the stress tensor in terms of the velocity and pressure
in cylindrical, spherical, and plane polar coordinates are given in Table 3.3.2. Note that the stress
components remain symmetric in orthogonal curvilinear coordinates. In the case of two-dimensional
ow expressing rigid-body rotation with angular velocity around the origin in the xy plane, we
substitute ur = 0 and u = r and nd that rr = p, r = 0, r = 0, = p, where p is the
pressure.
= p I. (3.3.22)
In real life, no uid can be truly inviscid and equation (3.3.22) must be regarded as a mathematical
idealization arising in the limit as the rate-of-deformation tensor, E, tends to become vanishingly
small. Under certain conditions, superuid helium behaves like an inviscid uid and may thus be
used in the laboratory to visualize ideal ows.
It is instructive to note the similarity in functional form between (3.3.22) and (3.3.1). However,
it is important to acknowledge that the pressure in (3.3.22) is a ow variable, whereas the pressure
in (3.3.1) is a thermodynamic variable determined by an appropriate equation of state.
210 Introduction to Theoretical and Computational Fluid Dynamics
(a)
ux u
x u
u
1 ux
xx = p + 2 , x = x = + , x = x = +
x x x
u u
1 u
= p + 2 , = = + ,
1 u u
= p + 2 +
(b)
ur u
1 u
r
rr = p + 2 , r = r = r +
r r r r
1 u
r u
1 u
ur
r = r = +r , = p + 2 +
r sin r r r r
sin u
1 u
= = +
r sin r sin
2 u
= p + ( + ur sin + u cos )
r sin
(c)
ur u
1 u
r
rr = p + 2 , r = r = r +
r r r r
1 u ur
= p + 2 +
r r
Table 3.3.2 Constitutive relations for the components of the stress tensor for an incompressible New-
tonian uid in (a) cylindrical, (b) spherical, and (c) plane polar coordinates. Note that the stress
components remain symmetric in orthogonal curvilinear coordinates.
Problems
The third term on the right-hand side expresses the spatial derivative of the ith velocity component
in the direction of the normal vector. When the uid is inviscid, we obtain a simplied form involving
the pressure alone, f = pn.
1
In terms of the vorticity tensor, = 2 (L LT ), we obtain
f = p n + 2 ( n u + n ). (3.4.3)
f = p n + 2 n u + n . (3.4.4)
The union of the second and third terms on the right-hand side expresses the traction due to the
deviatoric component of the stress tensor. If the velocity eld is irrotational, the last term on the
right-hand side vanishes and the part of the traction corresponding to the deviatoric component
of the stress tensor is proportional to the derivative of the velocity in the direction of the normal
vector, n u.
Introducing a local Cartesian coordinate system with two axes perpendicular to the normal unit
vector, n, we nd that the second term on the right-hand side of (3.4.5) expresses the viscous stress
associated with the normal derivative of the normal velocity component. Multiplying (3.4.1) with
the projection matrix I n n, we obtain the tangential component of the traction,
f T = 2 n E (I n n) = 2 n (n E) n. (3.4.6)
The structure of the ow must be such that the right-hand side of (3.4.6) is identically zero at a free
surface.
where n is the normal unit vector pointing outward from the parcel. Substituting (3.4.1) into
(3.1.20), we nd that, in the absence of an external torque eld, the torque with respect to a point,
x0 , exerted on the parcel is given by
Ttot = p (x x0 ) n dS + 2 (x x0 ) (E n) dS + (x x0 ) dV g, (3.4.8)
P arcel P arcel P arcel
where n is the normal unit vector pointing outward from the parcel.
p p g x, p I + 2E = + (g x) I. (3.4.9)
In hydrodynamic variables, the total force and torque with respect to a point x0 exerted on a uid
parcel are given by the surface integrals
Ftot = p n dS + 2 E n dS (3.4.10)
P arcel P arcel
and
Ttot = p (x x0 ) n dS + 2 (x x0 ) (E n) dS. (3.4.11)
P arcel P arcel
3.4 Force and energy dissipation in incompressible Newtonian uids 213
where n is the normal unit vector pointing into the uid. The rst term on the right-hand side
represents the form drag, and the second term represents the skin friction.
To obtain an expression for the torque with respect to a point, x0 , we substitute (3.4.1) into
(3.1.23) and obtain
T= p (x x0 ) n dS + 2 (x x0 ) (E n) dS. (3.4.13)
B B
In hydrostatics, or when viscous stresses are insignicant, the force and torque can be com-
puted from knowledge of the pressure distribution over the boundary.
where
2 E : E (3.4.15)
expresses the rate of viscous dissipation per unit volume of uid. Equation (3.4.14) is a special
version of (3.3.19) applicable to compressible Newtonian uids. Viscous forces dissipate energy,
converting it into thermal energy and thereby raising the temperature of the uid. When the rate
of viscous dissipation is negligible, the sum of the kinetic and potential energies remains constant in
time.
u ( ) = (u ) : u (3.4.16)
over the volume of a uid parcel, and using the divergence theorem, we obtain
u ( ) dV = u f dS : u dV, (3.4.17)
P arcel P arcel P arcel
214 Introduction to Theoretical and Computational Fluid Dynamics
where f = n is the boundary traction and n is the normal unit vector pointing into the parcel.
Substituting the Newtonian constitutive equation (3.3.20), we obtain
u ( ) dV = u f dS dV. (3.4.18)
P arcel P arcel P arcel
The two terms on the right-hand side of (3.4.18) represent, respectively, the rate of working of the
surface traction and the rate of viscous dissipation.
Working in a similar fashion with the deviatoric part of the stress tensor dened in (3.3.4),
we obtain
u ( ) dV = u f dS dV, (3.4.19)
P arcel P arcel P arcel
In the case of a uid with uniform viscosity, we use the Newtonian constitutive relation (3.3.20)
and the continuity equation to obtain
= 2 u = . (3.4.20)
Using this relation, we nd that, if the ow is irrotational or the vorticity is uniform throughout
the domain of ow, the rate of working of the deviatoric viscous traction is balanced by the rate of
viscous dissipation,
u f dS = dV. (3.4.21)
P arcel P arcel
where the normal unit vector, n, points into the control volume. The ve integrals in (3.4.22)
represent, respectively, the rate of accumulation of kinetic energy inside the control volume, the rate
of convection of kinetic energy into the control volume, the rate of working of surface forces, the
rate of viscous dissipation, change of potential energy associated with the body force.
When the ow is irrotational or the vorticity is uniform throughout the domain of ow, we
use (3.4.21) and obtain the simplied form
1
1
|u|2 dV = ( |u|2 ) u n dS + p u n dS + g u dV. (3.4.23)
Vc t 2 D 2 D Vc
This energy balance is also valid when the ow is not irrotational or has uniform vorticity, but the
uid can be considered to be inviscid.
3.5 The NavierStokes equation 215
Problem
= p + 2 (E) = p + 2 u + 2 E. (3.5.3)
The three Cartesian components of the NavierStokes equation for a uid with uniform viscosity
are displayed in Table 3.5.1.
When the uid density and acceleration of gravity are uniform throughout the domain of ow,
it is convenient to work with the hydrodynamic pressure, p p gx, and associated hydrodynamic
stress tensor indicated by a tilde introduced in (3.4.9), obtaining the equation of motion
Du
= = p + 2 (E), (3.5.4)
Dt
which is distinguished by the absence of the body force. In solving the NavierStokes equation,
the distinction between the regular and modied pressure becomes relevant only when boundary
conditions for the pressure or traction are imposed.
Kinematic viscosity
Dividing both sides of (3.5.2) by the density, we obtain the new form
u 1 2
+ u u = p + 2 u + E + g, (3.5.5)
t
216 Introduction to Theoretical and Computational Fluid Dynamics
ux ux ux ux p 2 ux 2 ux 2 ux
( + ux + uy + uz )= + ( + + ) + gx
t x y z x x2 y 2 z 2
uy uy uy uy p 2 uy 2 uy 2 uy
( + ux + uy + uz )= + ( + + ) + gy
t x y z y x2 y2 z 2
uz uz uz uz p 2 uz 2 uz 2 uz
( + ux + uy + uz )= + ( + + ) + gz
t x y z z x2 y2 z 2
Table 3.5.1 Eulerian form of the three Cartesian components of the NavierStokes equation for an
incompressible uid with uniform viscosity.
where = / is a new physical constant with dimensions of length squared over time analogous to
the molecular or thermal diusivity, called the kinematic viscosity of the uid. Sample values of the
kinematic viscosity for water and air are given in Table 3.3.1.
Du
= p + g, (3.5.6)
Dt
which shows that viscous forces are important only in regions where the curl of the vorticity is
signicant.
The equation of motion (3.5.6) reveals that irrotational ows and ows whose vorticity eld
is irrotational behave like inviscid ows. The dynamics of these ows is determined by a balance
between the inertial force due to the point-particle acceleration, the pressure force, and the body
force. Viscosity is important only insofar as to establish the vorticity distribution. Once this has
been achieved, viscosity plays no further role in the force balance. Flows with irrotational vorticity
elds include two-dimensional ows with constant vorticity and axisymmetric ows without swirling
3.5 The NavierStokes equation 217
(a) p p 2 ux 1 ux 1 2 ux
x = + 2 ux = + 2
+ + 2
x x x 2
p u 2 u
= + 2 u 2 2
p 2 u 1 (u ) 1 2 u 2 u
= + 2
+ + 2 2
2
x
1 p u 2 u
= + 2 u 2 + 2
1 p 2 u 1 (u ) 1 2 u 2 u
= + 2
+ + 2 2
+ 2
x
(b)
p ur 2 u 2 2 u
r = + 2 ur 2 2 2 2 u cot 2
r r r r r sin
1 p 2 ur u 2 cot u
= + 2 u + 2 2 2 2
r r r sin r sin
1 p 1 ur u
= + 2 u + 2 2 (u + 2 + 2 cos )
r sin r sin
1 2 f 1 f 1 2f
where: 2 f = r + sin +
r 2 r r r 2 sin r 2 sin2 2
(c)
p ur 2 u p 1 (rur ) 1 2 ur 2 u
r = + ( 2 ur 2 2 )= + + 2 2
r r r r r r r r 2 r
1 p u 2 ur 1 p 1 (ru ) 1 2 u 2 ur
= + (2 u 2 + 2 )= + + 2 2
+ 2
r r r r r r r r r
Table 3.5.2 Components of the hydrodynamic volume force for a Newtonian uid in (a) cylindrical, (b)
spherical, and (c) plane polar coordinates. The Laplacian operator 2 in spherical polar coordinates
is given in the fourth entry of (b).
motion where the azimuthal vorticity component increases linearly with distance from the axis of
revolution, = , where is a constant.
218 Introduction to Theoretical and Computational Fluid Dynamics
where F = p/ in the case of uniform-density uids. In the case of barotropic uids, the function F
is found by integrating the ordinary dierential equation
dF 1 dp
= . (3.5.8)
d d
For simplicity, in our discussion we consider uniform-density uids. Adaptations for barotropic uids
can be made by straightforward modications.
1
u u = (u u) u , (3.5.9)
2
we derive a new form of the NavierStokes equation,
u 1 p
+ u u + g x = u . (3.5.10)
t 2
Nonlinear terms appear in the second term on the left-hand side involving the square of the velocity
and in the rst term on the right-hand side involving the velocity and the vorticity. The term inside
the parentheses on the left-hand side of (3.5.10) is the Bernoulli function,
1 p
B(x, t) u u + g x. (3.5.11)
2
Physically, the Bernoulli function expresses the mass distribution density of the total energy con-
sisting of the kinetic energy, the internal energy due to the pressure, and the potential energy due
to the body force. In terms of the Bernoulli function, the NavierStokes equation takes the compact
form
u
+ B = u . (3.5.12)
t
Later in this section, we will see that, under certain conditions, the Bernoulli function is
constant along streamlines or even throughout the domain of ow. When this occurs, B is called
the Bernoulli constant.
3.5 The NavierStokes equation 219
Integrating (3.5.13) along a closed streamline and using the fundamental theorem of calculus
to set the integral of the left-hand side to zero, we obtain
t ( ) dl = 0, (3.5.14)
which shows that the circulation of the curl of the vorticity along a closed streamline is zero in a
steady ow.
Later in this chapter, we will discuss a class of steady two-dimensional ows where the vorticity
is constant along the streamlines and may thus be regarded as a function of the stream function,
z = f (), and a class of axisymmetric ows where the ratio / is constant along the streamlines
and may thus be regarded as a function of the Stokes stream function, / = f (). Under these
circumstances, the right-hand sides of equations (3.5.15) can be written as a gradient,
u = F (), (3.5.16)
220 Introduction to Theoretical and Computational Fluid Dynamics
dF
= f (). (3.5.17)
d
Equation (3.5.16) shows that, because the cross product u constitutes an irrotational vector
eld, the vortex force can be grouped with the gradient on the left-hand side of (3.5.13). In the case
of two-dimensional ow with constant vorticity, or axisymmetric ow where f () = , with being
constant, we nd that F = + c, where c is constant throughout the domain of ow.
ij (ui ij ) ui
u ( ) = ui = ij . (3.5.18)
xj xj xj
(ui ij ) ui uj
ui
u ( ) = + p ij . (3.5.19)
xj xj xi xj
Using the continuity equation to eliminate the term involving the pressure, we nd that
(ui ij ) u uj
ui
u ( ) =
i
+ . (3.5.20)
xj xj xi xj
Setting = and u = u , integrating over the volume of a uid parcel, and using the divergence
theorem, we recover (3.4.18).
Multiplying (3.5.20) by and (3.5.21) by , and subtracting corresponding sides of the resulting
equations, we obtain the dierential statement of the generalized Lorentz reciprocal identity
ui ij ui ij = u u , (3.5.22)
xj
which imposes a constraint on the mutual structure of the velocity and stress elds of two unre-
lated incompressible Newtonian ows. Equations (3.5.20), (3.5.21), and (3.5.22) also apply for the
hydrodynamic stress tensor dened in (3.4.9) (Problem 3.5.1).
3.5 The NavierStokes equation 221
Expressing the divergence of the stress tensors on the right-hand side of (3.5.22) in terms of
the point-particle acceleration and the body force using the equation of motion, we obtain
Du Du
ui ij ui ij = u u ( u u) g. (3.5.23)
xj Dt Dt
In terms of the hydrodynamic stress tensor indicated by a tilde, we obtain the simpler form
Du Du
ui ij ui ij = u u . (3.5.24)
xj Dt Dt
Integrating (3.5.24) over a chosen control volume, Vc , that is bounded by a surface, D, and using
the divergence theorem to convert the volume integral of the left-hand side into a surface integral
over D, we derive the corresponding integral form
Du Du
u f u f dV = u u dV, (3.5.25)
D Vc Dt Dt
where the normal unit vector, n, points into the control volume. In Section 6.8, we will see that
the reciprocal identities (3.5.24) and (3.5.25) nd extensive applications in the study of Stokes ow
where the eect of uid inertia is negligibly small.
Problems
3.5.2 Hydrostatics
Consider a body of uid with uniform density that is either stationary or translates as a rigid body.
Show that the general solution of the equation of motion is p = g x + c(t) or p = c(t), where c(t)
is an arbitrary function of time, and p is the hydrodynamic pressure. Discuss the computation of
c(t) for a ow of your choice.
3.5.3 Oseen ow
(a) Consider a steady streaming (uniform) ow with velocity V past a stationary body. Far from
the body, the ow can be decomposed into the incident ow and a disturbance ow v due to the
body, u = V + v. Introduce a similar decomposition for the pressure, substitute the decompositions
into the NavierStokes equation, and neglect quadratic terms in v to derive a linear equation.
(b) Repeat (a) for uniform ow past a semi-innite plate aligned with the ow. Specically, derive
a linear equation for the disturbance stream function applicable far from the plate.
When the density of the uid and acceleration of gravity are uniform throughout the domain
of a ow, it is convenient to introduce the hydrodynamic pressure incorporating the eect of the
body force, p p g x, and recast Eulers equation (3.6.1) into the form
Du
= p. (3.6.2)
Dt
This expression reveals that the point-particle acceleration eld, Du/Dt, is irrotational. As a conse-
quence, if a ow governed by the Euler equation is irrotational at the initial instant, it will remain
irrotational at all times. The permanence of irrotational ow for a uid with uniform density and
negligible viscous forces will be discussed in the context of vorticity dynamics in Section 3.11.
B = u . (3.6.5)
In a Beltrami ow where the velocity is parallel to the vorticity at every point, the right-hand side
of (3.6.5) is zero and the Bernoulli function is constant throughout the domain of ow.
Considering the more general case of a ow where velocity and vorticity are not necessarily
aligned, we apply (3.6.5) at a certain point on a chosen streamline and project both sides onto the
3.6 Euler and Bernoulli equations 223
tangent unit vector. Note that the right-hand side of (3.6.5) is normal to the velocity and therefore
also to the streamline, and integrating the resulting expression with respect to arc length along the
streamline, we nd that B is constant along the streamline. Symbolically, we write
B = F(streamline). (3.6.6)
1 B
= b, (3.6.8)
u ln
where u = |u| and ln is the arc length measured in the direction of n. Projecting (3.6.5) onto the
binormal unit vector, b, and working in a similar fashion, we nd that
1 B
= n, (3.6.9)
u lb
where lb is the arc length measured in the direction of b. If the vorticity is locally parallel to the
velocity at the chosen point, the right-hand sides of (3.6.8) and (3.6.9) are zero and B reaches a local
extremum at that point.
B = F () + c, (3.6.10)
where c is constant throughout the domain of ow. For example, in the case of two-dimensional ow
with uniform vorticity, z = , or axisymmetric ow with azimuthal vorticity = , where is a
constant, we nd that B = +c. Since the curl of the vorticity and thus the magnitude of viscous
forces are identically zero, these ows represent exact solutions of the NavierStokes equation. The
velocity eld can be computed working in the context of kinematics, and the pressure follows from
knowledge of the stream function using the derived expression for B. One example is the ow inside
Hills spherical vortex discussed in Section 2.12.2.
224 Introduction to Theoretical and Computational Fluid Dynamics
Irrotational ow
The right-hand side of (3.6.4) is zero in a steady or unsteady irrotational ow. Introducing the
velocity potential, , substituting u = in the temporal derivative on the left-hand side of (3.6.4),
and integrating the resulting equation in space, we derive Bernoullis equation
+ B = c(t), (3.6.11)
t
where c(t) is an unspecied time-dependent function. In practice, the value of c(t) is determined by
applying (3.6.11) at a point on a selected boundary of the ow and then introducing an appropriate
boundary condition for the velocity or pressure. Equation (3.6.11) can be interpreted as an evolution
equation for .
Steady irrotational ow
The Bernoulli equation for steady irrotational ow takes the simple form
B = c, (3.6.14)
where c is a constant. Since the velocity attains its maximum at the boundaries, the dynamic
pressure, p p g x attains a minimum at the boundaries.
we will see that the physical conditions for the vorticity to be uniform are that it is uniform at the
initial instant and the eect of viscosity is negligibly small. The key idea is to introduce the stream
function, , and write
+ B + = c(t), (3.6.17)
t
where c(t) is a time-dependent function.
v dU d
+ B = v
2 v ( y) y, (3.6.18)
t dt dt
where
is the vorticity in the noninertial frame,
1 p
B(x, t) vv+ gy (3.6.19)
2
is the corresponding Bernoulli function, y is the position vector in the noninertial frame, and the
rest of the symbols are dened in Section 3.2.6.
Consider a noninertial frame that translates with linear velocity U(t) with respect to an inertial
frame, and assume that the velocity eld in the noninertial frame, v, is irrotational. Expressing v
in terms of an unsteady velocity potential, , dened so that v = , and taking into account the
acceleration-reaction body force, we obtain the modied Bernoullis equation
1 p dU
+ vv+ + g y = c(t), (3.6.20)
t y 2 dt
where c(t) is an unspecied function of time. When U is constant or zero, we recover the standard
Bernoulli equation for unsteady irrotational ow.
226 Introduction to Theoretical and Computational Fluid Dynamics
In another application, we consider a noninertial frame whose axes rotate with constant an-
gular velocity about the origin of an inertial frame. We will assume that the ow is irrotational
in the inertial frame and appears to be steady in the noninertial frame. Eulers equation in the
noninertial frame (3.6.18) simplies into
B = v (
+ 2 ) ( y). (3.6.21)
Integrating in space, we derive a modied Bernoulli equation,
1
B | y|2 = F(streamline), (3.6.22)
2
which is inclusive of the Bernoulli equation (3.6.6) for an inertial frame. In the literature of geo-
physical uid dynamics, the term (g y + 12 | y|2 ) is sometimes called the geopotential.
Problems
specify either the three components of the velocity, or the three components of the traction, or a
combination of the velocity and the traction. If we specify the normal component of the velocity over
all boundaries of a domain whose volume remains constant in time, we must ensure that the total
volumetric ow rate into the domain of ow is zero, otherwise the continuity equation cannot be
satised. In certain computational procedures for solving the equations of incompressible Newtonian
ow, the boundary velocity is specied at discrete points and the total ow rate is not precisely zero
due to numerical error. This imperfection may provide an entry point for numerical instability.
n u = n V, (3.7.1)
where the boundary velocity V may be constant or vary over the boundary. An impermeable
boundary is not necessarily a rigid boundary, as it may represent, for instance, a uid interface or
the surface of a exible body.
Rigid-body motion
If a boundary moves as a rigid body, translating with velocity U and rotating about a point x0 with
angular velocity , the boundary velocity is V = U + (x x0 ) and the no-penetration condition
requires that
n u = n [ U + (x x0 ) ] (3.7.2)
Two-dimensional ow
In the case of two-dimensional ow in the xy plane past an impermeable boundary executing rigid-
body motion, translating while rotating about an axis that is parallel to the z axis and passes
through the point x0 , equation (3.7.2) takes the form
V = U + z ez (x x0 ), (3.7.3)
228 Introduction to Theoretical and Computational Fluid Dynamics
where ez is the unit vector along the z axis and z is the angular velocity of rotation around the z
axis. Next, we write n = (dy/dl, dx/dl), where dx and dy are dierential increments around the
boundary corresponding to the dierential arc length dl measured in the counterclockwise direction.
Expressing the velocity in terms of the stream function, , we obtain
dy dx
un= + = = U n + z [ez (x x0 )] n (3.7.4)
y dl x dl l
or
dy dx dx dy
= Ux Uy z (x x0 ) + (y y0 ) . (3.7.5)
l dl dl dl dl
Integrating with respect to arc length, we obtain the scalar boundary condition
1
= Ux y Uy x z |x x0 |2 + c(t), (3.7.6)
2
where c(t) is a time-dependent constant.
Equation (3.7.6) reveals that the stream function is constant over a stationary impermeable
surface in a steady or unsteady ow. The value of the stream function is determined by the ow rate
between the surface and the rest of the boundaries. In general, the value of the stream function over
dierent disconnected stationary boundaries may not be specied a priori, but must be computed
as part of the solution.
Axisymmetric ow
In the case of axisymmetric ow past a body that translates along its axis of revolution with velocity
U = U ex , we introduce the Stokes stream function, , and derive the scalar boundary condition
1
= U 2 + c(t), (3.7.7)
2
where is the distance from the axis of revolution and c(t) is a time-dependent constant. This ex-
pressions shows that the Stokes stream function takes a constant value over a stationary impermeable
axisymmetric surface.
The no-slip boundary condition requires that the tangential component of the uid velocity is
zero over a stationary solid boundary. Combining this condition with the no-penetration condition,
3.7 Governing equations and boundary conditions 229
we nd that the uid velocity over a stationary impermeable solid boundary is equal to the boundary
velocity. If a boundary executes rigid body motion, translating with linear velocity U and rotating
about a point x0 with angular velocity , then u = U + (x x0 ).
Physical origin
In the case of gases, the molecules of the uid are adsorbed on the solid surface over a time period that
is long enough for thermal equilibrium to be established, yielding a macroscopic no-slip condition on
a solid surface. An analogous explanation is possible in the case of liquids based on the formation
of short-lived bonds between liquid and solid molecules due to weak intermolecular forces. However,
because these physical mechanisms depend on the properties of the solid and uid molecules, the
no-slip condition may occasionally break down.
An alternative explanation is that the proper boundary condition on a solid surface is the
condition of vanishing tangential traction, and an apparent no-slip condition arises on a macroscopic
level due to inherent boundary irregularities. Detailed computations with model geometries lend
support to this explanation [340].
In rareed gases, the slip coecient, , can be rigorously related to the mean free path, f ,
dening the Knudsen number, Kn f /a by the Maxwell relation
1
= , (3.7.10)
Kn 2
where is the tangential momentum accommodation coecient (TMAC) expressing the fraction of
molecules that undergo diusive rather than specular reection (e.g., [72, 362]). In the limit 2,
we obtain the no-slip boundary condition, . In the limit 0, we obtain the perfect-slip
boundary condition, 0.
230 Introduction to Theoretical and Computational Fluid Dynamics
The slip boundary condition has been used with success to describe ow over the surface
of a porous medium where the slip velocity accounts for the presence of pores, and ow in the
neighborhood of a moving three-phase contact line where the slip velocity removes the singular
behavior of the traction in the immediate neighborhood of the contact line [114].
To implement the far-eld condition, we decompose the velocity eld into an unperturbed
component and a disturbance component, and require that the ratio of the magnitudes of the dis-
turbance velocity and the unperturbed velocity decays at innity. This condition does not necessarily
imply that the disturbance velocity vanishes at innity. For example, in the case of uniform ow
past a sphere, the disturbance velocity due to the sphere decays at innity; whereas in the case of
parabolic ow past a sphere, the disturbance velocity grows at a rate that is less than quadratic. In
the case of simple shear ow over an innite plane wall with periodic corrugations or protrusions,
we require that the disturbance ow due to the protrusions grows at a less-than-a-linear rate. The
solution reveals that the disturbance velocity tends to a constant value far from the wall, thereby
introducing an a priori unknown slip velocity.
Fluid 1
Fluid 2
Fluid 1
V
(d) (e)
A
n
n S
r R
Fluid 1
ns Vcl
Contact line
Fluid 2
Figure 3.7.1 Illustration of (a) a stationary interface meeting a stationary solid surface at a static
contact line, (b) a stationary interface meeting a moving solid surface at a dynamic contact line, (c)
a contact line moving over a stationary solid surface due to a spreading liquid drop. The contact
angle, , is measure on the side of the uid labeled 2. (d) Contact line around the surface of a
oating particle. (e) Typical dependence of the dynamic contact angle, , on the velocity of the
contact line on a stationary surface, Vcl .
The angle subtended between (a) the vector that is normal to the contact line and tangential
to an interface, and (b) the vector that is normal to the contact line and lies on the solid boundary
is called the contact angle, . The contact angle is measured from the side of a designated uid, as
shown in Figure 3.7.1. In the case of a liquid-gas interface, is measured by convention from the
side of the liquid. With reference to Figure 3.7.1(d), the contact angle measured on the side of the
uid labeled 2 is given by
cos = ns n, (3.7.11)
where ns is the unit vector normal to the surface and n is the unit vector normal to the interface
232 Introduction to Theoretical and Computational Fluid Dynamics
pointing into the uid labeled 1. The plane containing ns and n at a point is normal to the contact
line at that point.
A rational framework for predicting the static contact angle employs three interfacial tensions
applicable to each uid-uid or uid-solid pair at the contact line, denoted by 1 , 2 , and . A
tangential force balance yields the Young equation,
2 = 1 + cos . (3.7.12)
Similar equations can be written for a contact line that lies at a corner or cusp where two tensions
are not necessarily aligned.
Conversely, the velocity of a contact line can be regarded as a function of the dierence,
S , A , or R . From this viewpoint, contact angle motion is driven by deviations of the
contact angle from an appropriate threshold. In the case of a spreading drop illustrated in Figure
3.7.1(c), > A resulting in a positive (outward) contact line velocity. In the case of a heavy drop
moving down an inclined plane, the contact angle at the front of the drop is higher than A , while
the contact angle at the back of the drop is less than R , resulting in a forward bulging shape.
3.8 Interfacial conditions 233
Problems
where n is the normal unit vector pointing into uid 1 by convention, and (1) , (2) are the stress
tensors in the two uids evaluated on either side of the interface, as shown in Figure 3.8.1(a). The
discontinuity in the interfacial traction, f , is determined by the physiochemical properties of the
uids and molecular constitution of the interface, and is aected by the local temperature and local
concentration of surface-active substances, commonly called surfactants, populating the interface.
A laundry detergent or dishwashing liquid is a familiar household surfactant.
(a) (b)
n
Fluid 1
Fluid 1
n (1)
f
D
n
(2)
f
Fluid 2 C
Fluid 2
t
b
Figure 3.8.1 (a) Illustration of an interface between two immiscible uids labeled 1 and 2, showing the
traction on either side of the interface. The normal unit vector, n, points into uid 1 by convention.
(b) Surface tension pulls the edges of an interfacial patch of a three-dimensional interface in the
tangential plane.
The jump in the hydrodynamic traction is related to the jump in the physical traction by
f = f + (1 2 ) (g x) n, (3.8.3)
An uncontaminated interface between two immiscible uids exhibiting isotropic surface tension, ,
playing the role of surface pressure. In a macroscopic interpretation, the surface tension pulls the
edges of an interfacial patch in the tangential plane, as shown in Figure 3.8.1(b). The surface tension
of a clean interface between water and air at 20 C is = 73 dyn/cm. The surface tension of a clean
interface between glycerin and air at the same temperature is = 63 dyn/cm.
The surface tension generally decreases as the temperature is raised and becomes zero at a critical
threshold. If an interface is populated by a surfactant, the surface tension is determined by the local
surface surfactant concentration, . The negative of the derivative of the surface tension, d/d, is
sometimes called the Gibbs surface elasticity. Typically, the surface tension decreases as increases
and reaches a plateau at a saturation concentration where the interface is covered by a monolayer
of surfactant molecules, as discussed in Section 3.8.2.
To derive an expression for the jump in the interfacial traction, f , we neglect the mass of the
interfacial stratum and write a force balance over a small interfacial patch, D, that is bounded by
3.8 Interfacial conditions 235
the contour C,
f dS + t n dl = 0, (3.8.4)
D C
where n is the unit vector normal to D pointing into uid 1, t is the unit vector tangent to C,
and b = t n is the unit binormal vector, as shown in Figure 3.8.1(b). Next, we extend smoothly
the domain of denition of the surface tension and normal vector from the interface into the whole
three-dimensional space. The extension of the surface tension can be implemented without any
constraints. In practice, this can be done by solving a partial dierential equation with the Dirichlet
boundary condition over the interface. The extension of the normal vector can be implemented by
setting n = F/|F |, where the equation F (x, y, z, t) = 0 describes the location of the interface.
A variation of Stokes theorem expressed by equation (A.7.8), Appendix A, states that, for
any arbitrary vector function of position, q,
q t dl = n q (q) n dS. (3.8.5)
C D
Applying this identity for the extended vector eld, g = n, and combining the resulting expression
with (3.8.4), we obtain
f dS = n (n) [(n)] n dS. (3.8.6)
D D
1
Expanding the derivatives inside the integral and noting that (n) n = 2 (n n) = 0, we obtain
f dS = n [ ( n) n ] dS. (3.8.7)
D D
Now we take the limit as the surface patch shrinks to a point, rearrange the integrand on the right-
hand side, and discard the integral sign on account of the arbitrary integration domain to obtain
the desired expression
f = n n (I n n) , (3.8.8)
f = n n (n ) n. (3.8.9)
The divergence of the normal vector on the right-hand side of (3.8.8) is equal to twice the
mean curvature of the interface, n = 2 m . Thus,
f = 2m n (n ) n. (3.8.10)
By denition, the mean curvature is positive when the interface has a spherical shape with uid 2
residing in the interior and the normal vector pointing outward into the exterior uid labeled 1. The
computation of the mean curvature is discussed in Section 1.8.
236 Introduction to Theoretical and Computational Fluid Dynamics
where n is the unit vector tangent to the interface, r = dx/dl is the unit vector tangent to the
contact line, the integration is performed around the contact line, and l is the arc length around the
contact line. Using a vector identity, we obtain an alternative expression for the capillary torque,
The surface tension inside the integrals on the right-hand sides of (3.8.11) and (3.8.12) is allowed to
be a function of position.
We observe that, if the contact angle is constant along the contact line, the torque with respect to
the center of the spherical particle is identically zero irrespective of the shape of the contact line
[373]. This result underscores the importance of contact angle hysteresis in imparting a non-zero
capillary torque for any contact line shape.
For large variations in the surfactant surface concentration, we may use Langmuirs equation
of state derived on the assumption of second-order adsorption/desorption kinetics,
= c + RT ln 1 = c 1 + ln 1 , (3.8.20)
0
where is the surfactant concentration at maximum packing and = 0 / is the surface
coverage at the reference state (e.g., [48]). Applying this equation for = 0 , we obtain
0
c =
. (3.8.21)
1 + ln(1 )
238 Introduction to Theoretical and Computational Fluid Dynamics
In deriving Langmuirs equation of state, ideal behavior is assumed to neglect cohesive and repulsive
interactions between surfactant molecules in the interfacial monolayer.
where n is the normal unit vector pointing into uid 1. Rearranging, we obtain an expression for
the jump in the interfacial pressure in terms of the normal component of f and the viscous normal
stress,
where f is given by an appropriate interfacial constitutive equation. When the uids are either
stationary or inviscid, only the rst term on the right-hand side of (3.8.23) survives. When the
interface exhibits isotropic tension, we use (3.8.8) and derive Laplaces law, p = 2m .
If the uids are inviscid, the velocity is allowed to be discontinuous across the interface,
yielding the simplied form
Du(2) Du(1)
t f N = t (p2 p1 ) = 2 + 1 + (2 1 ) g t, (3.8.25)
Dt Dt
where f N = f n is the normal component of the jump in the traction. Equation (3.8.25) imposes
a constraint on the tangential components of the point-particle acceleration on either side of the
interface between two immiscible uids.
Du
= + g + q, (3.8.26)
Dt
where q is a singular forcing function expressing an interfacial distribution of point forces,
q(x) = 3 (x x ) f (x ) dS(x ), (3.8.27)
I
I stands for the interface, 3 is the three-dimensional delta function, and f is the jump in the
traction across the interface. The velocity is required to be continuous, but the physical properties
of the uids and stress tensor are allowed to undergo discontinuities across the interface.
where n is the normal unit vector pointing into uid 1, and L denotes the layer. In the limit as
tends to zero, the volume integrals of the point-particle acceleration and gravitational force vanish.
The surface integrals can be rearranged to yield
0= n+ + dS n+ dS f dS, (3.8.29)
I I I
240 Introduction to Theoretical and Computational Fluid Dynamics
where n+ is the normal unit vector pointing into uid 1, + is the stress on the upper side of the
interface, and is the stress on the lower side of the interface. It is now evident that (3.8.29), and
therefore (3.8.26), is consistent with the interfacial condition (3.8.1).
A formal way of deriving (3.8.26) involves replacing the step functions, inherent in the repre-
sentation of the physical properties of the uids, and the delta functions, inherent in the distribution
of the interfacial force, with smooth functions that change gradually over a thin interfacial layer of
thickness . As long as is noninnitesimal, the regular form of the equation of motion applies in
the bulk of the uids as well as inside the interfacial layer. Taking the limit as tends to zero, we
derive (3.8.26).
Two-dimensional ow
In the case of two-dimensional ow, the interfacial distribution (3.8.27) takes the form of a line
integral,
q(x) = f (x ) 2 (x x ) dl(x ), (3.8.30)
I
where l is the arc length along the interface and 2 is the two-dimensional delta function in the plane
of the ow.
Problems
Expanding the velocity in a Taylor series with respect to x, y, and z, and enforcing the no-slip
and no-penetration boundary conditions, we nd that all velocity components and their rst partial
derivatives with respect to x and z at the origin are zero,
u
u
u(0) = 0, = 0, = 0. (3.9.1)
x 0 z 0
The continuity equation, u = 0, requires that the rst partial derivative of the normal component
of the velocity with respect to y is also zero at the origin,
u
y
= 0. (3.9.2)
y 0
The corresponding normal derivatives of the tangential velocity, ux /y and uz /y, are not nec-
essarily zero. With reference to arbitrary Cartesian coordinates that are not necessarily tangential
to the boundary, equations (3.9.1) and (3.9.2) take the form
u = 0, (I n n) u = (n u) n = 0, n (u) n = 0, (3.9.3)
evaluated at the boundary.
Using the denition of the vorticity in conjunction with the kinematic constraint expressed by
the second equation in (3.9.3), we nd that the component of the vorticity normal to the boundary
is zero. At the origin of the local coordinate system, the tangential components of the vorticity are
uz 1 ux 1
x = = fz , z = = fx , (3.9.6)
y y
These expressions along with (3.9.3) suggest that, with reference to a general coordinate system,
n u = n u (I n n) = n. (3.9.7)
Substituting (3.9.5) and (3.9.8) into (3.4.11) and (3.4.12), we derive simplied expressions for the
hydrodynamic force and torque exerted on a rigid boundary in terms of the vorticity and normal
derivatives of the tangential components of the velocity.
u = n ( n). (3.9.9)
The symmetric part of the velocity gradient tensor is the rate-of-deformation tensor and the anti-
symmetric part is the vorticity tensor. In the case of a compressible uid, the term n n is added
to the left-hand side of (3.9.9), where u is the rate of expansion.
f (I n n) = (n f ) n = n. (3.9.10)
Equation (3.9.7) shows that the vorticity vector is perpendicular to the skin-friction vector,
[f (I n n)] = 0, [n f n] = 0. (3.9.11)
A line over a boundary whose tangent vector is parallel to the skin friction vector at each point is a
skin-friction line. A line over the boundary whose tangent vector is parallel to the vorticity at each
point is a boundary or surface vortex line. Equations (3.9.11) show that the skin friction lines are
orthogonal to the surface vorticity lines. Equation (3.9.5) suggests that a skin-friction line can be
described in parametric form by an autonomous ordinary dierential equation,
dx
= n (u) (I n n) = n, (3.9.12)
d
where is an arbitrary time-like parameter.
A nodal point belongs to an innite number of skin-friction lines all of which except for one,
the one labeled AA in Figure 3.9.1(a), are tangential to a single line labeled BB. A focal point
belongs to an innite set of skin-friction lines that spiral away from or into the focal point, as shown
in Figure 3.9.1(b). A saddle point is the point of intersection of two skin-friction lines, as shown
3.9 Traction, vorticity, and kinematics at boundaries and interfaces 243
(a)
B A
A
B
B B A A
B
A
B A
(b) (c)
Vortex line
Skinfriction line
Figure 3.9.1 Illustration of singular points of the skin-friction pattern on a solid boundary [238]: (a)
nodal points, (b) a focus, and (c) a saddle point.
in Figure 3.9.1(c). Topological constraints require that the number of nodal points and foci on a
boundary exceed the number of saddle points by two. An in-depth discussion of the topography and
topology of skin-friction lines is presented by Lighthill [238] and Tobak & Peake [405].
The relation between the skin-friction lines and the trajectories of point particles in the vicinity of
a boundary becomes evident by introducing a system of orthogonal surface curvilinear coordinates
over the boundary, (, ). The point particles move with velocity that is nearly tangential to the
boundary, except when they nd themselves in the neighborhood of a singular point where the
streamlines turn away from the boundary. Denoting the instantaneous distance of a point particle
away from the boundary by , we expand the velocity in a Taylor series with respect to distance in
the normal direction and nd that the rate of change of distance traveled by a point particle along
the curvilinear axes is
244 Introduction to Theoretical and Computational Fluid Dynamics
dl
= U(, , t) t = (n u) t = t ,
dt
dl
= U(, , t) t = (n u) t = t , (3.9.13)
dt
where is the vorticity and the velocity gradient and vorticity are evaluated at the surface, = 0.
Dividing these equations side by side, we derive an alternative expression of (3.9.12),
dl dl
= . (3.9.14)
n (u) t n (u) t
f n = (n ) n = 0. (3.9.15)
Expressing the stress tensor in terms of the pressure and the rate-of-deformation tensor using the
Newtonian constitutive equation, = pI + u + (u)T , we nd that
f n = (n u) n n [(u) n] = 0. (3.9.16)
(I n n) = (n ) n = [ (u) n n u ] n. (3.9.18)
f s (I n n) = 2 (n u) n = 2 [(u) n] n (3.9.19)
or
f s (I n n) = 2 (u n) (n) u] n. (3.9.20)
The rst term on the right-hand side of (3.9.20) contains derivatives of the normal component of the
velocity tangential to the free surface. The derivatives of the normal vector in the second term are
related to the normal curvatures of the free surface in the normal plane that contains the velocity
and its conjugate plane, as discussed in the next section for steady ow.
3.9 Traction, vorticity, and kinematics at boundaries and interfaces 245
u
z x
Figure 3.9.2 A local Cartesian system attached to a free surface with the x axis pointing in the
direction of the velocity vector.
Steady ow
The rst term on the right-hand side of (3.9.19) vanishes over a stationary free surface where the
normal velocity is zero, u n = 0. It is useful to introduce a local Cartesian system with origin at
a point on a free surface, the x axis pointing in the direction of the velocity
vector,
and the y axis
pointing normal to the free surface, as shown in Figure 3.9.2. Setting n = 0, 1, 0 , we nd that the
tangential component of the vorticity is given by
nx nx
t = 2 0, 1, 0 (n) [ux , 0, 0] = 2ux ( ex ez ), (3.9.21)
z x
where ex is the unit vector along the x axis and ez is the unit vector along the z axis. The derivative,
x = nx /x, is the principal curvature of the free surface in the xy plane.
3.9.3 Two-dimensional ow
Simplied expressions for the boundary traction and vorticity can be derived in the case of two-
dimensional ow in the xy plane.
Rigid boundaries
First, we consider the structure of a ow near a stationary rigid boundary with reference to the local
coordinate system illustrated in Figure 3.9.3(a). Using (3.9.3), we nd that the z component of the
vorticity and tangential component of the wall shear stress at the origin are given by
ux ux
z = , xy = z = . (3.9.22)
y y
Thus, xy = z , revealing an intimate connection between the wall vorticity and the shear stress
at a solid boundary.
(a) (b)
y y
ux
x x
Figure 3.9.3 Illustration of (a) a two-dimensional ow near a solid wall and (b) a stagnation point on
a solid wall.
point, both must be zero at the stagnation point. Conversely, the vanishing of the shear stress or
vorticity provides us with a criterion for the occurrence of a stagnation point.
To demonstrate further that the shear stress is zero at a stagnation point by considering a
point that lies on the dividing streamline near the stagnation point, at the position (dx, dy). Setting
the position vector, (dx, dy), parallel to the velocity at that point, and expressing the velocity in a
Taylor series about the stagnation point, we nd that
uy uy
dx + dy
dy uy x y
tan = = = , (3.9.23)
dx ux ux ux
dx + dy
x y
where is the angle that the dividing streamline forms with the x axis, and all partial derivatives
are evaluated at the stagnation point. Using the continuity equation to write uy /y = ux /x
and enforcing the no-slip boundary condition, we nd that all partial derivatives uy /y, ux /x,
uy /x, at the origin are zero. Since the slope of the dividing streamline is nite, the partial
derivative ux /y must also be zero, yielding the condition of vanishing vorticity and wall shear
stress at a stagnation point.
Now assuming that the wall is at, we seek to predict the slope of the dividing streamline in
terms of the structure of the velocity. Since the rst-order terms in the fraction in (3.9.23) vanish,
we must retain the second-order contributions, obtaining
2 uy 2 uy 2 uy
2
(dx)2 + 2 (dx) (dy) + (dy)2
dy uy x xy y2
tan = = = 2 (3.9.24)
dx ux ux 2 2 ux 2 ux
(dx) + 2 (dx) (dy) + (dy)2
x2 xy y2
3.9 Traction, vorticity, and kinematics at boundaries and interfaces 247
(a) (b)
y
y
x x
n
R
t
r
Figure 3.9.4 (a) Illustration of a two-dimensional ow under a free surface. The origin of the plane
polar coordinates, (r, ), is set at the center of curvature at a point on the surface. (b) Illustration
of a stagnation point at a stationary free surface; the dividing streamline crosses the free surface
at a right angle.
or
2 uy 2 uy 2 uy
+ 2 tan + tan2
dy uy x2 xy y 2
tan = = = 2 . (3.9.25)
dx ux ux 2 ux 2 ux 2
2
+2 tan + tan
x xy y2
Enforcing the no-slip boundary condition and using the continuity equation, we nd that
2 uy 2 uy 2 ux
= 0, = = 0. (3.9.26)
x2 xy x2
x x x
tan = 3 )/ = 3 / , (3.9.27)
xy y 2 x x
where fx is the boundary traction, p is the hydrodynamic pressure, and all variables are evaluated
at the stagnation point [289]. To derive the expression in the denominator on the right-hand side of
(3.9.27), we have applied the NavierStokes equation at the stagnation point and used the no-slip
boundary condition to set 2 ux /y 2 = p/x, where p is the hydrodynamic pressure. Equation
(3.9.27) applies also at a stagnation point on curved wall, provided that the derivatives with respect
to x are replaced by derivatives with respect to arc length, l [248].
are
uy ux uy ux
z = , xy = + . (3.9.28)
x y x y
Setting the shear stress to zero, as required by denition for a free surface, we obtain
uy
z = 2 = 2 t (u) n = 2 t (u n) 2 t (n) u, (3.9.29)
x
where t is the tangent unit vector. Using the FrenetSerret relation t n = n/l = t, we write
z = 2 t (u n) 2u t, (3.9.30)
where l is the arc length along the free surface and is the curvature of the free surface. In terms
of the stream function, , we obtain
2
z = 2 2 , (3.9.31)
l2 ln
where ln is the arc length in the direction of the normal unit vector, n.
ux
xy = u t + . (3.9.32)
y
Introducing plane polar coordinates with origin at the center of curvature of the free surface at an
arbitrary point, (r, ), as shown in Figure 3.9.4(a), we nd that
u u u
xy = = R , (3.9.33)
R r r r r=R
where R = 1/ is the radius of curvature and all expressions are evaluated at r = R. Setting the
shear stress to zero and combining (3.9.32) with (1.11.10), we derive a simple expression for the
vorticity over a stationary free surface in terms of the tangential velocity and curvature of the free
surface,
z = 2 u t. (3.9.34)
This expression demonstrates that the vorticity is zero at a stagnation point along a free-surface
where the velocity is zero, at an inection point on a free-surface where the curvature is zero, and
over a planar free surface.
It is reassuring to observe that (3.9.34) also arises from the general relation (3.9.20) for three-
dimensional ow by setting nx /z = 0 to obtain T = 2ux x ez .
3.9 Traction, vorticity, and kinematics at boundaries and interfaces 249
1
x
2
Figure 3.9.5 Illustration of a stagnation point at a stationary uid interface. When the shear stress
is continuous across the interface, the slopes of the dividing streamlines obey a refraction law.
(1)
ux /x
tan 1 = 2 (1)
. (3.9.35)
ux /y
Dividing corresponding sides of (3.9.35) with their counterparts for the second uid and noting that,
because the velocity is continuous across the interface, the derivative ux /x is shared by the two
uids, we obtain
(2)
tan 1 ux /y
= (1)
. (3.9.36)
tan 2 ux /y
Now using (3.9.32) and requiring that the velocity is continuous across the interface, we nd
that the jump in the tangential component of the traction across the interface at the stagnation
point is given by
(1) (2)
ux ux
(f t)(1) (f t)(2) = 1 2 . (3.9.37)
y y
250 Introduction to Theoretical and Computational Fluid Dynamics
Assuming that the shear stress is continuous across the interface, we set the left-hand side of (3.9.37)
to zero and combine the resulting equation with (3.9.36) to derive the remarkably simple formula
tan 1 1
= , (3.9.38)
tan 2 2
which can be regarded as a refraction law for the dividing streamline [248]. When the viscosities of
the two uids are matched, the dividing streamlines join smoothly at the stagnation point.
Problems
(u t) 2
f n = p + 2 2 u n = p 2 + 2 , (3.9.39)
l lln l
where l is the arc length along the line, t is the tangent unit vector, is the curvature of the line,
ln is the arc length measured in the direction of the normal vector, n, and is the stream function.
(b) Use equation (3.9.39) to derive a boundary condition for the pressure at a free surface in the
presence of surface tension.
(c) Use equation (3.9.39) to derive a boundary condition for the jump in pressure across a uid
interface in the presence of surface tension.
over a time period T , which means that the magnitude of the rst partial derivative of the velocity
with respect to time is comparable to the ratio U/T .
Characteristic scales
Typically, but not always, the characteristic length L is related to the size of the boundaries, the
characteristic velocity U is determined by the particular mechanism driving the ow, and the char-
acteristic time T is either imposed by external means or simply dened as T = L/U . In the case
of unidirectional ow through a channel or tube, U can be identied with the maximum velocity
across the channel or tube. In the case of uniform ow past a stationary body, U can be identied
with the velocity of the incident ow. In the case of forced oscillatory ow, T can be identied with
the period of oscillation.
Nondimensionalization
Next, we scale all terms in the NavierStokes equation using the aforementioned scales and compare
their relative magnitudes. To accomplish the rst task, we introduce the dimensionless variables
u x t pL
u , x , t , p . (3.10.1)
U L T U
Expressing the dimensional variables in terms of corresponding dimensionless variables and substi-
tuting the result into the NavierStokes equation, we obtain the dimensionless form
u u = p
+ 2 u + Re g ,
+ Re u (3.10.2)
t Fr2 g
where the gradient involves derivatives with respect to the dimensionless position vector, x, and
g = |g|. Since the magnitudes of the dimensionless variables and their derivatives in (3.10.2) are of
order unity, and since g/g is a unit vector expressing the direction of the body force, the relative
importance of the various terms is determined by the magnitude of their multiplicative factors, which
are the frequency parameter, , the Reynolds number, Re, and the Froude number, Fr, dened as
L2 UL U
, Re , Fr , (3.10.3)
T gL
where is the kinematic viscosity of the uid with dimensions of length squared divided by time.
The frequency parameter, , expresses the relative magnitudes of the inertial acceleration
force and the viscous force, or equivalently, the ratio of the characteristic diusion time, L2 /, to
the time scale of the ow, T . The Reynolds number expresses the relative magnitudes of the inertia
convective force and the viscous force, or equivalently, the ratio of the characteristic diusion time,
L2 /, to the convective time, L/U . The Froude number expresses the relative magnitudes of the
inertial convective force and the body force. The group
Re gL2
= (3.10.4)
Fr U
expresses the relative magnitudes of the body force and the viscous force.
252 Introduction to Theoretical and Computational Fluid Dynamics
In the absence of external forcing, T , is identied with the convective time scale, L/U ,
reduces to Re, and the dimensionless NavierStokes equation (3.10.2) involves only two independent
parameters, Re and Fr.
p + 2 u + g = 0, (3.10.5)
stating that pressure, viscous, and body forces balance at every instant. A ow that is governed by
the Stokes equation is called a Stokes or creeping ow.
The absence of a temporal derivative in the equation of motion does not necessarily mean
that that the ow is steady, but only implies that the forces exerted on uid parcels are at equilib-
rium. Consequently, the instantaneous structure of the ow depends only on the present boundary
conguration and boundary conditions, which means that the ow is in a quasi-steady state. Stated
dierently, the history of motion enters the physical description only insofar as to determine the
current boundary conguration.
Steady and quasi-steady Stokes ows are encountered in a variety of natural and engineering
applications. Examples include slurry transport, blood ow in capillary vessels, ow due to the mo-
tion of ciliated micro-organisms, ow past microscopic aerosol particles, ow due to the coalescence
of liquid drops, and ow in the mantle of the earth due to natural convection. In certain cases, the
Reynolds number is low due to small boundary dimensions as in the case of ow past a red blood
cell, which has an average diameter of 8m. In other applications, the Reynolds number is low due
to the small magnitude of the velocity or high kinematic viscosity. An example is the ow due to
the motion of an air bubble in a very viscous liquid such as honey or glycerin.
The linearity of the Stokes equation allows us to conduct extensive theoretical studies, analyze
the properties of the ow, and generate desired solutions by linear superposition using a variety of
analytical and computational methods. An extensive discussion of the properties and methods of
computing Stokes ow will be presented in Chapter 6.
Unsteady Stokes ow
When Re 1 but 1, the nonlinear inertial convective term on the left-hand side of (3.10.2) is
negligible compared to the rest of the terms and can be discarded. Reverting to dimensional variables,
we nd that the nonlinear component of the point-particle acceleration, u u, is insignicant, and
the rate of change of momentum of a point particle can be approximated with the Eulerian inertial
acceleration force, u/t. The motion of the uid is governed by the unsteady Stokes equation,
also called the linearized NavierStokes equation,
u
= p + 2 u + g. (3.10.6)
t
3.10 Scaling of the NavierStokes equation and dynamic similitude 253
Because of the presence of the acceleration term involving a time derivative on the left-hand side, the
instantaneous structure of the ow depends not only on the instantaneous boundary conguration
and boundary conditions, but also on the history of uid motion.
Physically, the unsteady Stokes equation describes ows characterized by sudden acceleration
or deceleration. Three examples are the ow occurring in hydrodynamic braking, the ow occurring
during the impact of a particle on a solid surface, and the initial stages of the ow due to a particle
settling from rest in an ambient uid. The linearity of the unsteady Stokes equation allows us to
compute solutions for oscillatory and general time-dependent motion using a variety of methods,
including Fourier and Laplace transforms in time and space, as well as construct solutions by linear
superposition. The properties and methods of computing unsteady Stokes ow are discussed in
Sections 6.156.18.
To allow for this possibility, we rescale the pressure by introducing a new dimensionless pres-
sure dened as
p 1
= p. (3.10.7)
U 2 Re
Working as previously, we obtain a new dimensionless form of the NavierStokes equation,
u = Re
+ Re u u + 2 u + Re g . (3.10.8)
t Fr2 |g|
Now considering the limits Re 1 and 1, we nd that the viscous term becomes small compared
to the rest of the terms and can be neglected, yielding Eulers equation,
u + 1 g ,
=
+ u u (3.10.9)
Re t Fr2 |g|
where /Re = L/(T U ). We have found that, at high Reynolds numbers, a laminar ow behaves like
an inviscid ow in the absence of small-scale turbulent motion. By neglecting the viscous term, we
have lowered the order of the NavierStokes equation from two to one, by one unit. This reduction
has important implications on the number of required boundary conditions, spatial structure of the
ow, and properties of the solution.
equation becomes
u = 0,
(3.10.10)
F (u, ) = 0, (3.10.11)
where is the nondimensional stress tensor. The function F may contain dimensionless constants
dened with respect to the physical properties of the uid, the characteristic scales U , L, and T ,
the body force, and other physical constants that depend on the nature of the boundary conditions,
such as the surface tension, , and the slip coecient. Three dimensionless numbers pertinent to an
interface between two immiscible uids are the Bond number (Bo), the Weber number (We), and
the capillary number (Ca), dened as
gL2 LU 2 U
Bo = , We = . Ca = . (3.10.12)
The Bond number expresses the signicance of the gravitational force relative to surface tension.
The Weber number expresses the signicance of the inertial force relative to surface tension. The
capillary number expresses the signicance of the viscous stresses relative to surface tension.
In the space of dimensionless variables, the ow is governed by the equation of motion (3.10.2)
and the continuity equation (3.10.10), and the solution satises the conditions stated in (3.10.11). It
is evident then that the structure of a steady ow and the evolution of an unsteady ow depend on
(a) the values of the dimensionless numbers , Re, and Fr, (b) the functional form of the boundary
conditions stated in (3.10.11), and (c) the values of the dimensionless numbers involved in (3.10.11).
Two ows in two dierent physical domains will have a similar structure provided that, in the
space of corresponding dimensionless variables, the boundary geometry, initial state, and boundary
conditions are the same. Similarity of structure means that the velocity or pressure eld of the
rst ow can be deduced from those of the second ow, and vice versa, by multiplication with an
appropriate factor.
Dynamic similitude can be exploited to study the ow of a particular uid in a certain domain
by studying the ow of another uid in a similar, larger or smaller, domain. This is achieved by
adjusting the properties of the second uid, ow domain, or both, to match the values of , Re, Fr,
and any other dimensionless numbers entering the boundary conditions. Miniaturization or scale-up
of a domain of ow is important in the study of large-scale ows, such as ow past aircraft, and
small-scale ows, such as ow past microorganisms and small biological cells and ow over surfaces
with small-scale roughness and imperfections.
Problems
dC 1
= p + 2 u + g dX, (3.11.1)
dt L
where X is the position of a point particle along the loop. Since the integral of the derivative of a
function is equal to the function itself, the integral of the last term on the right-hand side of (3.11.1)
makes a vanishing contribution. A similar reasoning reveals that, when the density of the uid is
uniform or the uid is barotropic, the integral of the pressure term is also zero. Setting the Laplacian
of the velocity equal to the negative of the curl of the vorticity, we obtain the simplied form
dC 2
= ( u) dX = ( ) dX. (3.11.2)
dt L L
If the vorticity eld is irrotational, = 0, the integral on the right-hand vanishes and the rate
of change of the circulation is zero.
256 Introduction to Theoretical and Computational Fluid Dynamics
where D is an arbitrary surface bounded by the loop and n is the unit vector normal to D. Equation
(3.11.3) expresses Kelvins circulation theorem, stating that the circulation around a closed material
loop in a ow with negligible viscous forces is preserved during the motion. As a consequence, the
circulation around a loop that wraps around a toroidal boundary, and thus the cyclic constant of
the motion around the boundary, remain constant in time.
Now we consider a reducible loop, take the limit as the loop shrinks to a point, and nd that,
when the conditions for Kelvins circulation theorem to apply are fullled, point particles maintain
their initial vorticity, which means that they keep spinning at a constant angular velocity as they
translate and deform in the domain of ow. Physically, the absence of a viscous torque exerted on
a uid parcel guarantees that the angular momentum of the parcel is preserved during the motion.
(u u) = u u, (3.11.4)
and assuming that the acceleration of gravity is constant throughout the domain of ow, we derive
the vorticity transport equation,
D 1 1
+ u = u + 2 p + 2 + 2 u + 2 E , (3.11.5)
Dt t
where E is the rate-of-deformation tensor. Projecting both sides of (3.11.5) onto half the moment
of inertia of a small uid parcel provides us with an angular momentum balance.
3.11 Evolution of circulation and vorticity dynamics 257
Now restricting our attention to uids whose viscosity is uniform throughout the domain of
ow, we obtain the simplied form
D 1
= + u = u + 2 p + 2 . (3.11.6)
Dt t
We could solve the NavierStokes for the pressure gradient and substitute the result into the right-
hand side of (3.11.6) to derive an expression in terms of the velocity and vorticity, but this is necessary
neither in theoretical analysis nor in numerical computation. The left-hand side of (3.11.6) expresses
the material derivative of the vorticity, which is equal to the rate of change of the vorticity following
a point particle, or half the rate of change of the angular velocity of the point particle.
To illustrate further the nature of the nonlinear term, u, we begin with the statement
= 0 and use a vector identity to write
= ( u) = u (u)T = 0, (3.11.7)
ui uj u uj
i
j = j = j Eji = j + (1 ) , (3.11.9)
xj xi xj xi
where is an arbitrary constant. If the vorticity vector happens to be an eigenvector of the rate-of-
deformation tensor, E, it will amplify or shrink in its direction, behaving like a material vector.
Baroclinic production
The second term on the right-hand side of (3.11.6), 12 p, expresses baroclinic generation of
vorticity due to the interaction between the pressure and density elds. To illustrate the physical
process underlying this coupling, we consider a vertically stratied uid that is set in motion by
the application of a horizontal pressure gradient. The heavier point particles at the top accelerate
slower than the lighter uid particles at the bottom. Consequently, a vertical uid column buckles
in the counterclockwise direction, thus generating vorticity with positive sign. When the density of
258 Introduction to Theoretical and Computational Fluid Dynamics
the uid is uniform or the pressure is a function of the density alone, the baroclinic production term
vanishes. In the second case, the gradients of the pressure and density are parallel and their cross
product is identically zero.
Diusion
The third term on the right-hand side of (3.11.6), 2 , expresses diusion of vorticity with dif-
fusivity that is equal to the kinematic viscosity of the uid, . Under the action of this term, the
Cartesian components of the vorticity vector diuse like passive scalars.
Eulerian form
An alternative form of the vorticity transport equation (3.11.6) emerges by restating the nonlinear
term in the equation of motion in terms of the right-hand side of (3.5.9) before taking the curl. Noting
that the curl of the gradient of a continuous function is identically zero, we derive the Eulerian form
of the vorticity transport equation,
1
+ ( u) = 2 p + 2 . (3.11.10)
t
The second term on the left-hand side incorporates the eects of convection, reorientation, and
vortex stretching.
where a prime over the gradient designates dierentiation with respect to the integration point, x .
The last expression shows that the second term on the right-hand side of (3.11.11) generates a sheet
of vortex dipoles causing a discontinuity in the velocity gradient across the interface.
3.11 Evolution of circulation and vorticity dynamics 259
Two-dimensional ow
The nature of the last term on the right-hand side of (3.11.11) can be illustrated best by considering
a two-dimensional ow in the xy plane and computing
wz = 2 (x x ) [f (x ) ez ] dl(x ), (3.11.14)
I
where 2 is the two-dimensional delta function in the plane of the ow, l is the arc length along the
interface, and ez is the unit vector along the z axis [261]. Decomposing the jump in the interfacial
traction into its normal and tangential components designated by the superscripts N and T , we
obtain
wz = 2 (x x ) [f T n + f N t](x ) dl(x ). (3.11.15)
I
We have assumed that the triplet (t, n, ez ) forms a right-handed coordinates so that t ez = n
and n ez = t.
The tangential component inside the integrand can be manipulated further by writing
f N (x ) t(x ) 2 (x x ) dl(x ) = t(x ) [f N (x ) 2 (x x )] dl(x )
I
I
When the interface is closed or periodic, the rst integral on the right-hand side vanishes and
(3.11.15) simplies into
wz = f (x ) n(x ) 2 (x x ) dl(x ) + 2 (x x ) d f N (x ).
T
(3.11.17)
I I
The rst term on the right-hand side expresses the generation of a sheet of vortex dipoles due to
the tangential component of the discontinuity in the interfacial traction. The second term expresses
the generation of a vortex sheet due to the normal component of the discontinuity in the interfacial
traction.
As an application, we consider an innite at interface parallel to the x axis and assume that
the normal component f N is zero along the entire interface. Expression (3.11.17) becomes
2 (x x )
wz = f T dx . (3.11.18)
I y
Problem
3.11.1 Two-dimensional ow
Write the counterpart of equation (3.11.5) for the z vorticity component in two-dimensional ow in
the xy plane.
260 Introduction to Theoretical and Computational Fluid Dynamics
+ u + = 0, (3.12.3)
t
which ensures that the vector eld inside the parentheses is and remains irrotational.
If u and are the velocity and vorticity of a generalized Beltrami ow, their negative pair, u
and , also satises (3.12.2), and the reversed ow is also a generalized Beltrami ow (Problem
3.12.9). Since the direction of the velocity along the streamlines is reversed when we change the sign
of the velocity, a generalized Beltrami ow is sometimes called a two-way ow.
Figure 3.12.1 Illustration of the Burgers columnar vortex surviving in the presence of an axisymmetric
straining ow.
which is the foundation of Helmholtz theorems discussed in Section 3.11.3. Referring to (1.6.1), we
nd that the vorticity vector behaves like a material vector in the following sense: if dl is a small
material vector aligned with the vorticity vector at an instant, then /|| = dl/|dl| at all times,
where is the vorticity at the location of the material vector.
According to equation (3.12.1), the rate of change of the vorticity vanishes throughout an irrotational
ow, and this seemingly suggests that the ow will remain irrotational at all times. However, this
erroneous deduction ignores the possibility that vorticity may enter the ow through the boundaries.
The responsible physical mechanism is analogous to that by which heat enters an initially isothermal
domain across the boundaries of a conductive medium due to a sudden change in the boundary
temperature.
The process by which vorticity enters a ow can be illustrated by considering the ow around
an impermeable boundary that is placed suddenly in an incident ow. To satisfy the no-penetration
boundary condition, we introduce an irrotational disturbance ow. However, we are still left with
a nite slip-velocity amounting to a boundary vortex sheet. Viscosity causes the singular vorticity
distribution associated with the vortex sheet to diuse into the uid, complementing the disturbance
irrotational ow, while the boundary vorticity is continuously adjusting in response to the developing
ow. If the uid is inviscid, vorticity cannot diuse into the ow, the vortex sheet adheres to the
boundary, and the ow remains irrotational.
+ 2 = 0. (3.12.9)
d d d
A nontrivial solution that is nite at the axis of revolution is
2
x = A exp , (3.12.10)
4
where A is a dimensionless constant. Inspecting the argument of the exponential term, we nd that
the diameter of the vortex is constant along the axis of revolution, and the size of the vortex is
comparable to the viscous length scale (4/)1/2 .
The azimuthal velocity prole is computed by integrating equation (3.12.8), nding the Gaus-
sian distribution
1 2
U () = 2A 1 exp . (3.12.11)
4
The azimuthal velocity is zero at the axis of revolution and decays like that due to a point vortex
far from the axis of revolution. Accordingly A = C/(4), where C is the circulation around a loop
with innite radius. The rate of extension, , determines the radius of the Burgers vortex; as tends
to innity, we obtain the ow due to a point vortex (see also Problem 3.12.1).
y U0
z
U0
Figure 3.12.2 A diusing vortex sheet is subjected to a two-dimensional straining ow that stretches
the vortex lines and allows for steady state to be established. For the velocity prole shown, z < 0.
where U (y) is the velocity prole across the vortex layer and is the rate of elongation of the
extensional ow.
Similarity solution
The absence of an intrinsic characteristic length scale suggests that the vortex layer may develop in
a self-similar fashion so that
z = f (), (3.12.13)
(t)
where (t) is the nominal thickness of the vortex layer to be computed as part of the solution,
y
(3.12.14)
(t)
is a dimensionless similarity variable, and f is a function with dimensions of length. Substituting
expressions (3.12.12) and (3.12.13) into the z component of the vorticity transport equation and
rearranging, we obtain
d f
+ = , (3.12.15)
dt f + f
where a prime denotes a derivative with respect to . We observe that the left-hand side of (3.12.15)
is a function of t alone, whereas the right-hand side is a function of both t and y. Each side
must be equal to the same constant, set for convenience equal to 2A2 , where A is a dimensionless
coecient. Solving the resulting ordinary dierential equations for and f subject to the initial
condition (0) = 0 and the stipulation that f is an even function of , we obtain
2
2 (t) = A2 1 e2t , f () = f (0) exp(A2 2 ). (3.12.16)
264 Introduction to Theoretical and Computational Fluid Dynamics
y 2
z = B exp , (3.12.17)
(1 e2t )1/2 2(1 e2t )
where
1
1/2
B= f (0) (3.12.18)
A 2
is a new dimensionless constant. To compute the velocity prole, U (y), we substitute (3.12.17) into
the left-hand side of the last equation in (3.12.12). Integrating with respect to y, we obtain an
expression in terms of the error function,
w
2 2
erfw e d. (3.12.19)
0
The constant B then follows as
2
1/2
B = U0 (3.12.20)
(Problem 3.12.2). At the initial instant, the velocity prole U (y, 0) undergoes a discontinuity and
the thickness of the vortex layer is zero. As time progresses, the thickness of the vortex layer tends
to the asymptotic value = A (2/)1/2 where diusion of vorticity away from the zx plane is
balanced by convection.
Unstretched layer
In the absence of straining ow, = 0, the vortex layer diuses to occupy the entire plane. To
describe the developing vorticity eld, we linearize the exponential terms in (3.12.17) with respect
to their arguments and obtain
1
1/2 y2
z (t) = U0 exp . (3.12.22)
t 4t
Integrating with respect to y, we derive the associated velocity prole
y
3.12.3 Axisymmetric ow
In the case of axisymmetric ow without swirling motion, the axial and radial components of the
vorticity transport equation (3.12.2) are trivially satised. The azimuthal component of the vor-
ticity transport equation provides us with a scalar evolution equation for the azimuthal vorticity
component,
D
= 2 E 2 ( ), (3.12.24)
Dt
where E 2 is a second-order operator dened in (2.9.17) and (2.9.20). The vortex stretching term
is inherent in the material derivative of / on the right-hand side of the ratio (3.12.24). If the
azimuthal vorticity is proportional to the radial distance at the initial instant, = , where is
a constant, the right-hand side of (3.12.24) vanishes and the vorticity remains proportional to the
radial distance at all times.
= 0. (3.12.25)
Dt
Accordingly, the strength of a line vortex ring is proportional to its radius. The underlying physical
mechanism can be traced to preservation of circulation around vortex tubes.
Projecting the vorticity transport equation onto the vorticity vector, we nd that the enstrophy of
a ow with uniform physical properties evolves according to the equation
dE
=2 ( ) : E dV 2 : dV + n ( ) dS, (3.12.27)
dt F low F low B
where B stands for the boundaries of the ow. The three terms on the right-hand side represent, re-
spectively, intensication of vorticity due to vortex stretching, the counterpart of viscous dissipation
for the vorticity, and surface diusion across the boundaries.
shown in (3.2.43). Let be v the velocity in the noninertial frame dened in (3.2.36), and
be
the corresponding vorticity. Including the inertial forces in the inertial form of the NavierStokes
equation, taking the curl of the resulting equation, and rearranging, we nd that the modied or
intrinsic vorticity,
+ 2, satises the inertial form of the vorticity transport equation,
D
= + v = v + 2 (3.12.28)
Dt t
(Problem 3.12.6). The simple form of (3.12.28) is exploited for the ecient computation of an
incompressible Newtonian ow in terms of the velocity and intrinsic vorticity, as discussed in Section
13.2 (e.g., [383]).
Problems
D Df
( f ) = . (3.12.29)
Dt Dt
Based on this equation, explain why, if the eld represented by f is convected by the ow, Df /Dt = 0,
the scalar f will also be convected by the ow.
vorticity transport equation (3.12.2) are trivially satised, while the z component yields a scalar
convectiondiusion equation for the z vorticity component,
Dz z
+ u z = 2 z , (3.13.1)
Dt t
where 2 is the two-dimensional Laplacian operator in the xy plane.
It is tempting to surmise that z behaves like a passive scalar. However, we must remember
that vorticity distribution has a direct inuence on the development of the velocity eld mediated
by the BiotSavart integral.
PrandtlBatchelor theorem
In Section 3.11.1, we showed that the circulation of the curl of the vorticity along a closed streamline
is zero in a steady ow,
t ( ) dl = 0. (3.13.3)
In the case of two-dimensional ow, this equation states that the contribution of the tangential
component of the viscous force to the total force exerted on the volume of a uid enclosed by a
streamline is zero. Since the z vorticity component is constant along a streamline in the absence
of viscous forces, it can be regarded as a function of the stream function, . The line integral in
(3.5.14) then becomes
z z dz
t ( ) dl = t (ex ey ) dl = t (ex ey ) dl, (3.13.4)
y x d y x
where ex and ey are unit vectors along the x and y axis. Invoking the denition of the stream
function, we nd that
dz dz
t ( ) dl = t (ex ux + ey uy ) dl = t u dl. (3.13.5)
d d
Setting the last expression to zero to satisfy (3.13.3), we nd that either the circulation around the
streamline is zero or the vorticity is constant in regions of recirculating ow.
268 Introduction to Theoretical and Computational Fluid Dynamics
Consider an innite ow that decays at innity in the absence of interior boundaries. Integrating
(3.13.1) over the whole domain of ow and using the continuity equation, we nd that
d
= (u z + 2 z ) dA = (z u + z ) dA. (3.13.7)
dt F low F low
Using the divergence theorem, we convert the last areal integral into a line integral over a large loop
enclosing the ow. Since the velocity is assumed to vanish at innity, the line integral is zero and
the total circulation of the ow remains constant in time. Physically, vorticity neither is produced
nor can escape in the absence of boundaries.
z z
= r . (3.13.8)
t r r r
Since the nonlinear term vanishes, the ow due to a diusing point vortex is a generalized Beltrami
ow.
The absence of an intrinsic length scale suggests that the solution is a function of the dimen-
sionless similarity variable r/(t)1/2 , so that
z = f (), (3.13.9)
t
where f is an a priori unknown function. Substituting this functional form into (3.13.8), we derive
a second-order linear ordinary dierential equation,
2 (f ) + 2 f + 2f = 0, (3.13.10)
where a prime denotes a derivative with respect to . Requiring that the derivatives of f are nite
at the origin and the total circulation of the ow is equal to at any time, we obtain the solution
r2
z = exp . (3.13.11)
4t 4t
3.13 Vorticity transport in two-dimensional ow 269
(a) (b)
2 0.4
0.35
1.5 0.3
0.25
z
U
1 0.2
0.15
0.5 0.1
0.05
0 0
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
r/a r/a
Figure 3.13.1 Proles of (a) the dimensionless vorticity, z z a2 /, and (b) dimensionless velocity,
U au /, around a diusing point vortex at a sequence of dimensionless times, t/a2 =
0.02, 0.04, . . ., where a is a reference length.
Integrating the denition z = (1/r)(ru )/r, we derive an expression for the polar velocity com-
ponent,
r2
u = 1 exp , (3.13.12)
2r 4t
describing the ow due to a diusing point vortex known as the Oseen vortex. Radial proles of the
dimensionless vorticity, z z a2 /, and reduced velocity, U au /, are shown in Figure 3.13.1
at a sequence of dimensionless times, t/a2 , where a is a reference length. As time progresses,
the vorticity diuses away from the point vortex and tends to occupy the whole plane. Viscosity
smears the vorticity distribution and reduces the point vortex into a vortex blob.
2 + f () = 0, (3.13.15)
t
which is fullled if the stream function evolves according to the equation
= f () = 2 . (3.13.16)
t
Specifying the function f and solving for provides us with families of two-dimensional generalized
Beltrami ows. Setting f equal to zero or a constant value, we obtain irrotational ows and ows
with constant vorticity. Other families of steady and unsteady generalized Beltrami ows have been
discovered [419, 420].
Taylor cellular ow
An interesting unsteady generalized Beltrami ow arises by stipulating that the function f () in-
troduced in (3.13.14) is linear in . Setting f () = 2 , where is a constant with dimensions of
inverse length, and integrating in time the dierential equation comprised of the rst two terms in
(3.13.16), we obtain
2 = 2 . (3.13.20)
describes an exponentially decaying cellular periodic ow with wave numbers in the x and y directions
equal to and , where 2 + 2 = 2 and A is an arbitrary constant [397]. The corresponding
3.13 Vorticity transport in two-dimensional ow 271
0.5
0.4
0.3
0.2
0.1
0
Y
0.1
0.2
0.3
0.4
0.5
0.5 0 0.5
X
Figure 3.13.2 Streamline pattern of the doubly periodic decaying Taylor cellular ow with square
cells.
pressure distribution is found from (3.13.18). The streamline pattern for = in the plane of
dimensionless axes X = x and Y = y is shown in Figure 3.13.2.
Inviscid ow
When the eect of viscosity is insignicant, the terms multiplied by the kinematic viscosity in
(3.13.16) vanish. Accordingly, any time-independent solution of (3.13.14) represents an acceptable
steady inviscid ow. Examples include the ow due to a point vortex, the ow due to an innite array
of point vortices, and any unidirectional shear ow with arbitrary velocity prole. The corresponding
pressure distributions are found from (3.13.18).
z = 2 = (x, y, t) c , (3.13.22)
where is a specied function and c is a specied constant with dimensions of inverse squared
length. The vorticity transport equation reduces to a linear partial dierential equation for ,
2
+ c2 = (2 + c ), (3.13.23)
t x y y x
whose solution can be found for a number of cases in analytical form (e.g., [419, 420]).
272 Introduction to Theoretical and Computational Fluid Dynamics
Kovasznay ow
As an example, consider a steady ow with = c U y, where U is a constant [212]. Expression
(3.13.22) becomes
z = 2 = c (U y ). (3.13.24)
The steady version of the vorticity transport equation (3.13.23) simplies into
U
= U y. (3.13.25)
c x
A solution is
2
=U y sin(ky) exp(kx) , (3.13.26)
k
where is an arbitrary dimensionless constant, k is an arbitrary wave number in the y direction,
and = c/(kU ) is a dimensionless number.
It remains to verify that (3.13.26) is consistent with the assumed functional form (3.13.24).
Substituting (3.13.26) into (3.13.24), we obtain a quadratic equation for ,
Re
2 1 = 0, (3.13.27)
2
where Re = 2U/k is a Reynolds number. Retaining the root with the negative value, we obtain
1 Re2
1/2 Re
= + 4 . (3.13.28)
2 4 2 2
Streamline patterns in the plane of dimensionless axes X = x/a and Y = y/a, are illustrated
in Figure 3.13.3 for = 1.0 and Re = 10 and 50, where a = 2/k is the separation of two successive
cells along the y axis. In both cases, a stagnation point is present at the origin. The ow can be
envisioned as being established in the wake of an innite array of cylinders arranged along the y
axis, subject to a uniform incident ow along the x axis. As the Reynolds number is raised, the
regions of recirculating ow become increasingly slender.
Problems
ux = U (1 eyV / ), uy = V. (3.13.29)
3.13 Vorticity transport in two-dimensional ow 273
(a) (b)
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
Y
Y
0.2 0.2
0.4 0.4
0.6 0.6
0.8 0.8
1 1
0.5 0 0.5 1 1.5 0.5 0 0.5 1 1.5
X X
Figure 3.13.3 Streamline pattern of the Kovasznay ow for = 1.0 and Reynolds number (a) Re =
10 or (b) 50.
Discuss the interpretation of the solution in terms of vorticity diusion toward and convection away
from the plate.
= U y (1 exU/ ) (3.13.30)
can be derived as an extended Beltrami ow. Discuss the physical interpretation of this ow.
Hydrostatics 4
When a uid is stationary or translates as a rigid body, the continuity equation is satised in a trivial
manner and the NavierStokes equation reduces into a rst-order dierential equation expressing a
balance between the pressure gradient and the body force. This simplied equation of hydrostatics
can be integrated by elementary methods, subject to appropriate boundary conditions, to yield the
pressure distribution inside the uid. The integration produces a scalar constant that is determined
by specifying the level of the pressure at an appropriate point over a boundary. The product of
the normal unit vector and the pressure may then be integrated over the surface of an immersed
or submerged body to produce the buoyancy force. The torque with respect to a chosen point can
be computed in a similar way. Details of this procedure and applications will be discussed in this
chapter for uids that are stationary or undergo steady or unsteady rigid-body motion.
In the case of two immiscible stationary uids in contact, the interfacial boundary conditions
discussed in Section 3.8 require that the normal component of the traction undergoes a discontinuity
that is balanced by surface tension. In hydrostatics, the normal component of the traction is equal
to the negative of the pressure and the tangential component is identically zero. Accordingly, the
interface must assume a shape that is compatible with the pressure distribution on either side and
conforms with boundary conditions for the contact angle at a three-phase contact line. Conversely,
the pressure distribution in the two uids cannot be computed independently, but must be found
simultaneously with the interfacial shape so that all boundary conditions are fullled. Depending on
the shape of the contact line, the magnitude of the contact angle, and the density dierence between
the two uids, a stationary interface may take a variety of interesting and sometimes unexpected
shapes (e.g., [52, 197]). For example, an interface inside a dihedral angle conned between two
planes may climb to innity, as discussed in Section 4.2.1. The computation of interfacial shapes
presents us with a challenging mathematical problem involving highly nonlinear ordinary and partial
dierential equations [126].
Our rst task in this chapter is to derive the pressure distribution in stationary, translat-
ing, and rotating uids. The equations governing interfacial hydrostatics are then discussed, and
numerical procedures for computing the shape of interfaces with two-dimensional (cylindrical) and
axisymmetric shapes are outlined. The computation of three-dimensional interfacial shapes and
available software are reviewed briey at the conclusion of this chapter as specialized topics worthy
of further investigation.
274
4.1 Pressure distribution in rigid-body motion 275
where n is the normal unit vector pointing into the uid. Substituting expression (4.1.2) and using
the divergence theorem to convert the surface integral into a volume integral over the volume of the
body, we derive Archimedes buoyancy force,
dU
F = VB g , (4.1.4)
dt
where VB is the volume occupied by the body.
(a) (b)
y
p x
atm
x
h p g
Fluid 1
a
Fluid 2
Figure 4.1.1 (a) Depiction of a liquid layer resting between an underlying heavier liquid and overlying
air, illustrating the hydrostatic pressure prole. (b) A spherical particle oats at the at surface of
a liquid at oating angle ; the dashed line represents the horizontal circular contact line.
is the volume centroid of the body. Placing the pivot point x0 at the volume centroid, xc , makes the
torque vanish. Using the divergence theorem, we derive a convenient representation for the volume
centroid in terms of a surface integral,
2
x 0 0
1 0 y 2 0 n dS(x),
xc = (4.1.7)
2VB Body
0 0 z2
gx = 0, gy = g, (4.1.8)
where g = |g| is the magnitude of the acceleration of gravity. Applying the general equation (4.1.2)
with U = 0 for each uid, we obtain the pressure distributions
p1 = 1 gy + P1 , p2 = 2 gy + P2 , (4.1.9)
To compute the constant P1 , we require that the pressure at the free surface of uid 1 is
equal to the atmospheric pressure, patm , and nd that P1 = patm . To compute the constant P2 ,
we apply both equations in (4.1.9) at the interface located at y = h, and subtract the resulting
expressions to nd that p2 p1 = gh + P2 patm , where = 2 1 . Requiring continuity of
4.1 Pressure distribution in rigid-body motion 277
the normal component of the traction or pressure across the interface, we obtain P2 = patm gh.
Substituting the values of P1 and P2 into (4.1.9), we obtain the explicit pressure distributions
The satisfaction of the dynamic condition p1 (h) = p2 (h) can be readily veried.
The function P (t) is evaluated by requiring that the pressure at the free surface is equal to the
atmospheric pressure, patm , at any instant, obtaining
P (t) = patm + a cos(t) g a2 cos(t) . (4.1.12)
The term inside the rst square brackets on the right-hand side is the vertical component of the
position vector in a frame of reference moving with the free surface. The term inside the second
square brackets is the sum of the vertical components of the gravitational acceleration and inertial
acceleration dU(t)/dt. It is then evident that expression (4.1.13) is consistent with the pressure
distribution that would have arisen if we worked in a frame of reference where the free surface
appears to be stationary, accounting for the ctitious inertial force due to the vertical vibration.
( x) = p + g. (4.1.14)
In cylindrical polar coordinates, (x, , ), with the x axis pointing in the direction of the angular
velocity vector, , equation (4.1.14) takes the form
2 e = p + g, (4.1.15)
278 Introduction to Theoretical and Computational Fluid Dynamics
where is the magnitude of and e is the unit vector in the radial () direction. Expressing the
gradient of the pressure in cylindrical coordinates and integrating with respect to , we nd that
1
p = g x + 2 2 + P (t), (4.1.16)
2
which can be restated as
1
p = g x + | x|2 + P (t), (4.1.17)
2
where P (t) is an inconsequential function of time. The additional pressure expressed by the second
term on the right-hand side is necessary in order to balance the radial centrifugal force due to the
rotation.
Rotating container
As an application, we consider the pressure distribution inside a container that rotates steadily
around the horizontal x axis. Setting the y axis in the direction of gravity pointing upward so that
the component of the acceleration of gravity are gx = 0, gy = g, and gz = 0, we nd the pressure
distribution
1
p = gy + 2 (y 2 + z 2 ) + P, (4.1.18)
2
which can be restated as
1 g
2 1 g2
p= 2 y 2 + z 2 2 + P. (4.1.19)
2 2
The expression shows that surfaces of constant pressure are concentric horizontal cylinders with axis
passing through the point z = 0 and y = g/2 . The minimum pressure occurs at the common axis.
Rotary oscillations
Now we assume that the uid rotates around the origin as a rigid body with time-dependent angular
velocity (t). Substituting the associated velocity eld u = (t) x in the NavierStokes equation,
4.1 Pressure distribution in rigid-body motion 279
we obtain the pressure distribution p = p1 + p2 , where p1 is given by the right-hand side of (4.1.17)
and p2 satises the equation
d
x = p2 . (4.1.21)
dt
Taking the curl of both sides of (4.1.21) and noting the the curl of the gradient of any twice dif-
ferentiable function is identically zero, we arrive at the solvability condition d/dt = 0, which
contradicts the original assumption of unsteady rotation. The physical implication is that rotary
oscillation requires a velocity eld other than rigid-body motion. For example, subjecting a glass
of water to rotary oscillation does not mean that the uid in the glass will engage in rigid-body
oscillatory rotation. In fact, at high frequencies, the motion of the uid will be conned inside a
thin boundary layer around the glass surface, and the main body of the uid outside the boundary
layer will be stationary.
Ideal gas
In the case of an ideal gas, the pressure is given by the ideal gas law, p = RT /M , as discussed
in Section 3.3. Solving for the density and substituting the resulting expression into (4.1.1) for a
stationary uid, U = 0, we obtain a rst-order dierential equation,
1 p M
p = ln = g, (4.1.22)
p p0 RT
where p0 is an unspecied reference pressure. When the temperature is constant, the solution is
p M
ln = g x. (4.1.23)
p0 RT
We have found that the pressure exhibits an exponential dependence on distance instead of the
linear dependence observed in an incompressible liquid.
elevation of y = 1 km = 1000 m is
28.97 9.80665
p = 1.0 exp 1000 atm = 0.892 atm. (4.1.25)
8.314 103 298
The corresponding density distribution is found by substituting the pressure distribution (4.1.24) in
the ideal gas law.
Problems
3 3 + 2 (2s 1) = 0, (4.1.26)
where cos , s Ws /(gVs ) is a dimensionless constant, Ws is the weight of the sphere, and
Vs is the volume of the sphere (e.g., [318]). If the sphere is made of a homogeneous material with
density B , then s = B / is the density ratio.
Computer Problem
Figure 4.2.1 A hydrostatic meniscus forming outside a vertical elliptical cylinder with aspect ratio 0.6,
for contact angle = /4 and scaled capillary length /c = 4.5036, where c is the major semi-axis
of the ellipse. The computer code that generated this shape is included in Directory men ell inside
Directory 03 hydrostat of the software library Fdlib (Appendix C).
where = 2 1 is the density dierence, P1 and P2 are two pressure functions of time attributed
to each uid, n is the normal unit vector pointing into uid 1 by convention, and m is the mean
curvature of the interface. Rearranging, we obtain the LaplaceYoung equation
2m = g x + , (4.2.2)
where (P2 P1 )/ is a new function of time with dimensions of inverse length.
Table 4.2.1 Curvature of a two-dimensional interface in several parametric forms. Fluid 1 lies above
uid 2.
where a is a properly chosen length. For example, a can be the radius of a circular tube surrounded
by an interface. In terms of the capillary length or Bond number, the LaplaceYoung equation
(4.2.2) takes the form
1 Bo
2m = eg x + or 2m = eg x + , (4.2.5)
2 a2
where eg g1 g is the unit vector pointing in the direction of gravity, the plus sign applies when
> 0, and the minus sign applies when < 0.
Table 4.2.2 Principal and mean curvatures of an axisymmetric interface in several parametric forms
with uid 1 lying in the outer space; 1 is the principal curvature in an azimuthal plane, and 2 is
the principal curvature in the conjugate plane.
Boundary conditions
The solution of the dierential equations that arise from the LaplaceYoung equation is subject
to a boundary condition that species either the contact angle at a three-phase contact line where
the interface meets a solid boundary or a third uid, or the shape of the contact line. The rst
Neumann-like boundary condition is employed when a solid boundary is perfectly smooth, whereas
the second Dirichlet-like boundary condition is employed when a solid surface exhibits appreciable
roughness (e.g., [115]). Problems where the shape of the contact line is specied are much easier to
solve than those where the contact angle is specied.
284 Introduction to Theoretical and Computational Fluid Dynamics
(a) (b)
y
R 3
2
x
1
x
1
Fluid 1
h a 2
Fluid 2
3
2 1 0 1 2
y
2a
Figure 4.2.2 (a) Illustration of a meniscus developing between two parallel vertical plates separated by
distance 2a for contact angle < /2. (b) When is greater than /2, the meniscus submerges
and the capillary rise h is negative. The meniscus shape depicted in (b) was produced by the
Fdlib code men 2d (Appendix C) [318].
The pressure proles in the upper or lower uids are described by (4.1.9). Equating the
pressures on either side of the at interface outside and far from the plates, located at y = h, we
obtain
P2 P1 gh
= , (4.2.6)
4.2 The LaplaceYoung equation 285
where = 2 1 > 0. Next, we substitute this expression for into the LaplaceYoung equation
(4.2.2), compute the radius of curvature from the expression R = a/ cos , introduce the approxi-
mation
1 cos
2m = , (4.2.7)
R a
evaluate the resulting equation at the midplane, x = 0, and nd that
2
h cos = cos . (4.2.8)
ga a
The sign of the capillary rise in (4.2.8) is determined by the contact angle, . When < /2, the
meniscus rises; when > /2, the meniscus submerges; when = /2 the meniscus remains at
at the level of the free surface outside the plates. The maximum possible elevation or submersion
height is 2 /a.
Equation (4.2.8) can be derived directly by performing a force balance on the liquid column
raised capillary action. Setting the weight of the column reduced by the buoyancy force, 2ah g,
equal to the vertical component of the capillary force exerted at the two contact lines, 2 cos , and
rearranging, we obtain precisely (4.2.8). A formal justication for this calculation will be given in
Section 4.2.3. A numerical procedure for computing the meniscus shape without any approximations
will be discussed in Section 4.3.2.
We may now return to (4.2.2) and establish the conditions under which the assumption that
the right-hand side is nearly constant is valid. Requiring that the variation in the elevation of the
interface, |R| (1 sin ), is smaller than |h|, we nd that
a
2
1 + sin , (4.2.9)
which shows that the plate separation must be suciently smaller than the capillary length, otherwise
gravitational eects are important along the length of the meniscus.
The predictions of equation (4.2.8) also apply when the plate separation, 2a, changes slowly
in the z direction normal to the xy plane. An example is provided by the meniscus forming between
two vertical plates with small-amplitude sinusoidal corrugations.
(a) (b)
x x
R2
g g
Fluid 1 Fluid 1 h a
Fluid 2 h a
Fluid 2
R2
Figure 4.2.3 Illustration of an axisymmetric meniscus developing inside a vertical circular tube of radius
a for (a) contact angle < /2 and (b) > /2. The signed length R2 is the second principal
radius of curvature.
Alternatively, expression (4.2.10) can be derived by setting the weight of the raised column reduced
by the buoyancy force, a2 hg, equal to the vertical component of the capillary force exerted
around the contact line, 2a cos . A numerical procedure for computing the meniscus shape
without any approximations will be discussed in Section 4.4.1.
Setting the x axis upward against the direction of gravity with origin at the level of the
undeformed interface far from the walls, we nd that the LaplaceYoung equation (4.2.2) reduces
into
1 g
x, (4.2.12)
R
where = 2 1 . Using elementary trigonometry, we write the geometrical condition
where is the contact angle, d is the meniscus thickness at the vertical bisecting plane, as shown in
Figure 4.2.4, and the length c is dened in Figure 4.2.4. Combining equations (4.2.12) and (4.2.13),
we obtain
1k 1k
d= R= , (4.2.14)
k k g x
4.2 The LaplaceYoung equation 287
x
g
Fluid 1
d q
R R
R
c
Fluid 2
Figure 4.2.4 Illustration of a meniscus inside a dihedral corner conned between two vertical plates
intersecting at angle 2.
where k sin / cos . The meniscus rises if d > 0, which is possible only when + < 12 . When
this condition is met, the meniscus takes a hyperbolic shape that diverges at the sharp corner. In
practice, the sharp corner has a nonzero curvature that causes the meniscus to rise up to a nite
capillary height. Expressions (4.2.13) and (4.2.14) validate a posteriori our assumption that the
curvature of the meniscus in a vertical plane is much smaller than the curvature of the interface in
a horizontal plane.
The area occupied by the liquid in a plane perpendicular to the x axis, denoted by A(x), scales
as A(x) d2 . The volume of liquid residing inside the tapering liquid tongue, denoted by V , scales
as V x0 d(x)2 dx, where x0 is a positive elevation marking the beginning of the local solution.
Since d 1/x, we nd that V 1/x0 d0 , where d0 is the meniscus thickness corresponding to x0 .
We conclude that a nite amount of liquid resides inside the tapering liquid tongue.
The shape of the meniscus can be described in plane polar coordinates with origin at the
corner, (r, ), by the function r = q(x, ), as shown in Figure 4.2.4. Using the law of cosines, we nd
that
Solving this quadratic equation for q and eliminating R and d in favor of x using (4.2.12) and
(4.2.14), we nd that the meniscus is described by the equation
cos k2 sin2 (4.2.16)
x= ,
g kr
288 Introduction to Theoretical and Computational Fluid Dynamics
n
t
D
Fluid 1
e
Fluid 2 Vc n n
M
n
n e
Figure 4.2.5 Illustration of a meniscus bounded by two closed contact angles drawn as heavy lines.
cos( + ) 1
xcl = . (4.2.17)
g sin r
We have found that the meniscus climbs up to innity when + < /2. For a right-angled
corner where = /4, such as that occurring inside a square tube, the contact line must be less
than /4. When + > /2, the meniscus does not rise to innity but takes a bounded shape [93].
In the absence of gravity, a solution does not exist when + > /2. This discontinuous behavior
underscores the important eect of boundary geometry in capillary hydrostatics.
Next, we identify a surface or a collection of surfaces, D, whose union with the meniscus, M , encloses
a nite control volume, Vc , as depicted in Figure 4.2.5. Using the divergence theorem to manipulate
the rst integral on the right-hand side, we obtain
2m n dS = (g x) n dS + n dS + Vc g, (4.2.19)
M D D
where the normal unit vector n over D points into the control volume, as shown in Figure 4.2.5.
Applying Stokes theorem expressed by equation (A.7.8), Appendix A, for G = n, and recalling that
4.2 The LaplaceYoung equation 289
2m = n, we obtain
2m n dS = n t dl, (4.2.20)
M C
where the line integral is computed around each contact line, C, t is the tangent unit vector, and l is
the arc length along the contact line. Substituting this expression into (4.2.19) and setting e nt,
we obtain the force balance
e dl = (g x) + n dS + Vc g. (4.2.21)
C D
The left-hand side expresses the capillary force exerted around the contact lines. The last term on
the right-hand side expresses the weight of the uid residing inside the control volume, reduced by
the buoyancy force. The rst term on the right-hand side expresses a surface pressure force.
As an example, we consider the meniscus subtended between two vertical circular cylinders
bounded by a horizontal at bottom, and identify D with union of the surface of the cylinders below
the contact lines and the the at bottom. If the contact lines are horizontal circles, the horizontal
component of the right-hand side of (4.2.21) is zero and the horizontal component of the capillary
force exerted on the outer cylinder is equal in magnitude and opposite in direction with that exerted
on the outer cylinder.
Problems
4.2.3 Catenoid
(a) Verify that the equation = f (x) = c1 [1 + f (x)2 ]1/2 , where c1 is a constant, describes an
axisymmetric interface with zero mean curvature. The surface is called a catenoid and its trace in
an azimuthal plane is called a catenary.
(b) Show that the catenoid is described by the equation = c1 cosh[(x c2 )/c2 ], where c2 is a new
constant.
Computer Problem
f 1 1
f
f
= 2 3/2
= 2 1/2
= 2 1/2
= 2 + , (4.3.1)
(1 + f ) f (1 + f ) (1 + f )
where = (g/g)1/2 is the capillary length, = 2 1 is the dierence in the densities of the
uids on either side of the interface, and is a constant with dimensions of inverse length to be
determined as part of the solution; uid 2 is assumed to lie underneath uid 1. Integrating (4.3.1)
4.3 Two-dimensional interfaces 291
(a) (b)
y y
h h
g g
(1) (1)
n n
cl
cl Fluid 1 t Fluid 1 t
x x
Fluid 2 Fluid 2
Figure 4.3.1 Illustration of a semi-innite interface attached to an inclined plate with (a) a monotonic
or (b) reentrant shape. Far from the plate, the interface becomes horizontal.
1 f2
| cos | = + f + . (4.3.3)
2 2
Rearranging (4.3.2), we derive a rst-order ordinary dierential equation,
df 2
2 1/2
= 1 . (4.3.4)
dx 2f + 2 f 2 /2
The plus or minus sign must be selected according to the expected interfacial shape. Performing a
second integration subject to a stipulated boundary condition provides us with specic shapes.
cl = + , (4.3.5)
292 Introduction to Theoretical and Computational Fluid Dynamics
as shown in Figure 4.3.1. Since both the inclination angle and the contact angle vary between
0 and , the angle cl ranges between 0 and 2. When cl = , the meniscus is perfectly at. This
observation provides us with a practical method of measuring the contact angle in the laboratory,
by varying the plate inclination angle, , until the interface appears to be entirely at; at that
point = . Requiring that the interface becomes at as x tends to innity, and therefore the
curvature and slope both tend to zero, and using expressions (4.3.1) and (4.3.3), we obtain
= 0, = 1. (4.3.6)
The constants P1 and P2 dened in (4.1.6) are equal to the pressure at the horizontal interface far
from the inclined plate.
Small deformation
If the slope of the interface is uniformly small along the entire meniscus, f 1 for 0 x < ,
equation (4.3.1) simplies into a linear dierential equation, f f /2 . A solution that decays as
x tends to innity is
f h ex/ , (4.3.7)
where h f (0) is the elevation of the interface at the contact line. Enforcing the contact angle
boundary condition f (0) = tan cl , we obtain
h
= tan cl , (4.3.8)
which is valid when cl .
Monotonic shapes
When the interface takes a monotonic shape, as illustrated in Figure 4.3.1(a), the slope angle cl
lies in the second or third quadrant, 12 cl 32 . Applying (4.3.3) at the contact line with = 0
and = 1, we nd that the capillary rise h f (0) is given by
h2
= 2 1 | cos cl | . (4.3.9)
2
Using standard trigonometric identities, we obtain
h cl
= 2 cos , (4.3.10)
2
Since 12 cl 32 , the maximum possible capillary height, occurring when cl = 12 or 32 , is
|h|max = 2. In the limit as cl tends to , we nd that h/ cl , in agreement with the
asymptotic solution (4.3.8).
Having obtained the elevation of the interface at the plate, we proceed to compute the inter-
facial shape. Equation (4.3.4) with = 0 and = 1 becomes
df 2
2 1/2
= 1 , (4.3.11)
dx 2 f 2 /2
where the plus or minus sign is selected according to the expected interfacial shape.
4.3 Two-dimensional interfaces 293
dX 2 F2
= . (4.3.14)
dF F 4 F2
Integrating with respect to F , we obtain
2 2
H
X= d, (4.3.15)
2
F 4
where the maximum possible value of |H| is 2. Evaluating the integral with the help of mathe-
matical tables, we nd that
X = [ (F ) (H) ], (4.3.16)
where
2+ 4 F2
(F ) = ln 4 F2 (4.3.17)
|F |
(e.g., [150], pp. 8185). To obtain the shape of the meniscus, we plot X against F in the range
0 F H if H > 0, or in the range H F 0 if H < 0.
Reentrant shapes
The preceding analysis assumes that the interface has a monotonic shape, which is true if cl lies
in the second and third quadrants, 12 cl 32 . Outside this range, the interface turns upon
itself, as shown in Figure 4.3.1(b). Since uid 2 lies above uid 1 beyond the turning point where
the meniscus is vertical, the sign of the curvature must be switched. The capillary rise is given by
equation (4.3.9) with the minus sign replaced by the plus sign on the right-hand side,
h2
2
= 2 1 + cos cl . (4.3.18)
294 Introduction to Theoretical and Computational Fluid Dynamics
(a)
1.5
1
F
0.5
0
1.5 1 0.5 0 0.5 1 1.5 2 2.5 3 3.5 4
(b) (c)
2 2
1.5 1.5
1 1
0.5 0.5
0 0
y
0.5 0.5
1 1
1.5 1.5
2 2
1 0 1 2 3 1 0 1 2 3
x x
Figure 4.3.2 (a) The shape of the meniscus for any value of the contact angle, , and arbitrary plate
inclination angle, , can be deduced from the master curve drawn with the solid line, representing
the function . The dashed and dotted lines represent approximate solutions for small interfacial
deformation. (b, c) Shape of a semi-innite meniscus attached to an inclined plate produced by
the Fdlib code men 2d plate (Appendix C). The plate inclination angles are dierent, but the
contact angle is the same in both cases.
Using standard trigonometric identities, we recover (4.3.10), which shows that the maximum possible
capillary height, achieved for a horizontal plate, cl = 0 or , is |h|max = 2. The solution (4.3.16)
is valid in the range 0 F H if H > 0, or in the range H F 0 if H < 0, with the maximum
value of the dimensionless capillary height |H| being 2.
Master curve
The shape of the interface in the complete range of cl can be deduced from the master curve drawn
with the solid line in Figure 4.3.2(a), representing the function (F ) dened in (4.3.17). To obtain
4.3 Two-dimensional interfaces 295
the shape of the meniscus for a particular plate inclination angle, , we identify the intersection of
the inclined plate with the master curve consistent with the specied contact angle, . When the
capillary height is negative, we work with the mirror image of the master curve. The dashed line in
Figure 4.3.2(a) represents the leading-order approximation for small deformation,
(F ) ln |F | + 2 (ln 2 1). (4.3.19)
Dividing the small-deformation solution (4.3.7) by and taking the logarithm of the resulting ex-
pression, we nd X ln |F |+ln |H| , which shows that (F ) ln |F |, represented by the dotted
line in Figure 4.3.2(a).
Numerical methods
As a practical alternative, the shape of the interface can be constructed numerically according to
the following steps:
1. Compute the slope angle cl from equation (4.3.5).
2. Compute the capillary rise h using the formulas
(1 + | cos(cl )|)1/2 if 0 < cl < 12 ,
1
1 h (1 | cos(cl )|)1/2 if 2 < cl < ,
= (4.3.20)
2
(1 | cos(cl )|)1/2 if < cl < 32 ,
3
(1 + | cos(cl )|)1/2 if 2
< cl < 2.
3. Integrate the dierential equation (4.3.11) from f = h to 0 with initial condition x(f = h) = 0.
If h is negative, we use a negative spatial step.
The method is implemented in the Matlab code men 2d plate included in the software library
Fdlib discussed in Appendix C. The graphics generated by the code for two plate inclination angles
and xed contact angle is shown in Figure 4.3.2(b, c).
Because of symmetry, the slope of the interface is zero midway between the plates. Using
(4.3.1) and (4.3.2), obtain = f (0) and = 1. Combing the expression for with (4.3.21), we
nd that
h
= f (0), (4.3.22)
2
296 Introduction to Theoretical and Computational Fluid Dynamics
which provides us with an expression for the capillary rise in terms of the a priori unknown curvature
of the meniscus at the centerline. Substituting the derived values of and into (4.3.3), we obtain
a relation between the slope of the interface and the elevation of the meniscus,
Dierential equations
Equation (4.3.4) provides us with a rst-order dierential equation describing the shape of the
interface in terms of the a priori unknown capillary rise, h,
df 22
2 1/2
= 2
1 . (4.3.24)
dx 2 f (f + 2h)
The boundary conditions require that f (0) = 0 and f (a) = cot , where is the contact angle.
Bond number
To recast equation (4.3.25) in dimensionless form, we introduce the dimensionless variables F = f /a,
X = x/a, and H = h/a. Substituting these expressions into (4.3.25), we obtain
F = Bo (F + H) (1 + F 2 )3/2 , (4.3.26)
where Bo (a/)2 = ga2 / is a Bond number and a prime denotes a derivative with respect to
X. The boundary conditions require that F (0) = 0, F (0) = 0, and F (1) = cot . It is now evident
that the shape of the meniscus is determined by the Bond number and contact angle, .
Numerical method
A standard procedure for solving equation (4.3.26) involves recasting it as a system of two rst-order
nonlinear ordinary dierential equations,
dF dG
= G, = Bo (F + H) (1 + F 2 )3/2 , (4.3.27)
dX dX
with boundary conditions F (0) = 0, G(0) = 0, and G(1) = cot . Having specied values for Bo and
, we compute the solution by iteration using a shooting method according to the following steps:
4.3 Two-dimensional interfaces 297
(a) (b)
y y
x
x
2a
Figure 4.3.3 Illustration of (a) the free surface of a liquid in a rectangular container and (b) a two-
dimensional liquid bridge supported by two horizontal plates.
1. Guess the capillary rise, H. A suitable guess provided by equation (4.2.8) is H = cos /Bo.
2. Integrate (4.3.27) from X = 0 to 1 with initial conditions F (0) = 0 and G(0) = 0.
3. Check whether the numerical solution satises the third boundary condition, G(1) = cot . If
not, repeat the computation with a new and improved value for H.
The improvement in the third step can be made using a method for solving nonlinear algebraic
equations discussed in Section B.3, Appendix B. A meniscus shape computed by this method is
shown in Figure 4.2.2(b).
Working as in the case of a meniscus between two parallel plates, we derive the counterpart
of equations (4.3.27),
dF dG
= G, = Bo (F + ) (1 + F 2 )3/2 , (4.3.28)
dX dX
where
2 2
= f (0) (4.3.29)
a a
is an a priori unknown dimensionless constant. The boundary conditions require that F (0) = 0,
G(0) = 0, and G(1) = cot . The solution can be found using the shooting method described in
Section 4.3.2, where guesses and corrections are made with respect to . Assuming that the meniscus
has a circular shape provides us with an educated guess, = cos /Bo.
298 Introduction to Theoretical and Computational Fluid Dynamics
Once the interfacial shape has been found, the dimensionless midplane elevation, H = h/a,
can be computed by requiring that the volume per unit width of the liquid has a specied value v,
yielding the integral constraint
1
v
H= 2 F (X) dX. (4.3.30)
2a 0
If H turns out to be zero or negative, the meniscus will touch or cross the bottom of the container
and the solution will become devoid of physical relevance. In that case, the uid will arrange itself
inside each corner of the container according to the specied value of the contact angle. The shape
of the meniscus must then be found using a dierent type of parametrization (Problem 4.3.1).
Problems
Computer Problems
Using the expression for the mean curvature shown in the second row of Table 4.2.2, we derive
a second-order nonlinear ordinary dierential equation,
1 1
f = 2
(f + h) (1 + f 2 )3/2 f (1 + f 2 ). (4.4.1)
The boundary conditions require that f (0) = 0, f (0) = 0, and f (a) = cot , where is contact
line. Note that, if the pair (f, h) is a solution for a specied contact angle, , then the pair (f, h)
is a solution for the reected contact angle, . Equation (4.4.1) diers from its two-dimensional
counterpart (4.3.25) by one term expressing the second principal curvature.
An apparent diculty arises when we attempt to evaluate (4.4.1) at the tube centerline, = 0,
as the second term on the right-hand side becomes indeterminate. However, using the regularity
condition f (0) = 0, we nd that the mean curvature of the interface at the centerline is equal to
f (0), which can be substituted into the LaplaceYoung equation to give
h
f (0) = . (4.4.2)
2
This expression can also be derived by applying the lHopital rule to evaluate the right-hand side of
(4.4.1).
Near the center of the cylinder, the interfacial slope is small, f 1. Linearizing (4.4.1), we
obtain the zeroth-order modied Bessel equation
1 f
f = (f + h) . (4.4.3)
2
An acceptable solution that remains nite at the origin is
f () h [ I0 (/) 1 ], (4.4.4)
Parametric description
Following standard practice, we resolve (4.4.1) into a system of two rst-order dierential equa-
tions, which we then solve using a shooting method. To ensure good performance, we introduce
a parametric representation in terms of the slope angle dened by the equation tan = f ,
where 0 + /2. The contact angle boundary condition is automatically satised by
300 Introduction to Theoretical and Computational Fluid Dynamics
this parametrization. To derive the equations governing the shape of the interface in parametric
form, we write
df df d d tan d 1 d
f = = = = (4.4.5)
d d d d d cos2 d
and
dx dx d d
= = tan . (4.4.6)
d d d d
Substituting (4.4.5) into the left-hand side of (4.4.1), replacing f on the right-hand side by tan ,
and simplifying, we obtain
d cos
= , (4.4.7)
d Q
where
x + h sin
Q= . (4.4.8)
2
Substituting (4.4.7) into the right-hand side of (4.4.6), we nd that
dx sin
= . (4.4.9)
d Q
Equations (4.4.7) and (4.4.9) provide us with the desired system of ordinary dierential equations
describing the shape of the meniscus in parametric form. To complete the denition of the problem,
we must supply three boundary conditions: two because we have a system of two rst-order ordinary
dierential equations, and one more because of the unspecied capillary height h. Fixing the origin
and the radial location of the contact line, we stipulate that
where cl = 12 .
The denominator Q dened in (4.4.8) becomes unspecied at the origin where = 0 and
= 0. Combining (4.4.2) with (4.4.5) and (4.4.6), we nd that, at the origin,
d
2 dx
=2 , = 0, (4.4.11)
d 0 h d 0
which is used to initialize the computation. The rst equation in (4.4.11) can be derived by applying
the lHopital rule to evaluate the right-hand side of the rst equation in (4.4.7), and then solving
for d/d (Problem 4.4.1).
where
sin
W = Bo (X + H) , (4.4.13)
Q
H = h/a, and Bo = (a/)2 = ga2 / is the Bond number. The boundary conditions require that
At the origin, = 0,
d
2 1 dX
= , = 0. (4.4.15)
d 0 Bo H d 0
Numerical method
Having specied the Bond number, Bo, and contact angle, , we compute the solution using a
shooting method as described in Section 4.3.2 for the analogous problem of a meniscus between two
plates according to the following steps:
1. Guess the reduced capillary height, H. An educated guess provided by equation (4.2.10) is
H = 2 cos /Bo.
2. Integrate (4.4.12) from = 0 to cl . To initialize the computation, we use (4.4.15).
3. Check whether the numerical solution satises the third boundary condition in (4.4.14). If
not, the computation is repeated with a new and improved value for H.
The improvement in the third step can be done using the secant method discussed in Section B.3,
Appendix B.
Far from the cylinder, the interfacial slope is small, f 1, and the LaplaceYoung equation
(4.4.16) reduces to the zeroth-order Bessel equation,
f f
f = + 2. (4.4.17)
302 Introduction to Theoretical and Computational Fluid Dynamics
(a) (b)
4
x
3.5
2.5
h/a
2
h
Fluid 1 1.5
1
Fluid 2 0.5
0
1 0.5 0 0.5 1 1.5 2
log(l/a)
Figure 4.4.1 (a) Illustration of an axisymmetric meniscus developing around a vertical circular cylin-
der. (b) Dependence of the capillary rise, h, on the capillary length, , for contact angle / = 0.02
(highest line), 0.125, 0.250, 0.375, and 0.480 (lowest line). The dotted lines show the capillary
height of a two-dimensional meniscus attached to a vertical at plate. The dashed lines represent
the predictions of an approximate asymptotic solution for high capillary lengths.
An acceptable solution that decays at innity is proportional to the modied Bessel function of zero
order, K0 ,
f () a K0 (/), (4.4.18)
K0 (/)
f () + f () 0, (4.4.19)
K1 (/)
The boundary-value problem can be solved numerically using a standard shooting method
combined with secant updates, as discussed in Section B.3, Appendix B. In the numerical method, the
solution domain is truncated at a large radial distance, max , where the Robin boundary condition
(4.4.19) is applied [320]. The computer code is available in the software library Fdlib [318] (see
Appendix C).
Graphs of the capillary rise around the contact line are shown in Figure 4.4.1(b) together with
the predictions of an asymptotic analysis discussed later in this section, represented by the dashed
lines. As /a becomes smaller, the curvature of the cylinder becomes decreasingly important and
the results reduce to those for a two-dimensional meniscus attached to a vertical at plate. Families
of interfacial shapes are presented in Figure 4.4.2 for two contact angles.
4.4 Axisymmetric interfaces 303
(a) (b)
5 5
4 4
3 3
x/a
x/a
2 2
1 1
0 0
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
/a /a
Figure 4.4.2 Interfacial shapes of an axisymmetric meniscus outside a circular cylinder parametrized
by the capillary length for contact angle (a) = /4 and (b) /50. The higher the capillary length,
the higher the interfacial prole.
Asymptotics
When the interfacial elevation is suciently smaller than the capillary length, f / 1, while the
slope f is not necessarily small, equation (4.4.16) takes the gravity-free form
f
f = (1 + f 2 ) , (4.4.20)
describing a zero-mean-curvature interface. A solution of this nonlinear equation that satises the
prescribed contact angle boundary condition at the cylinder surface, f ( = a) = cot , is
f () a cos ln , (4.4.21)
+ 2 cos2
where /a and is a dimensionless constant. As / tends to zero, the outer asymptotic
solution (4.4.18) yields
f () a ( ln + E ), (4.4.22)
where E = 0.577215665 . . . is Eulers constant. Matching this expression with the functional form
of (4.4.21) in the limit as tends to innity, we recover the coecient of the outer eld, , the
dimensionless coecient , and then an expression for the capillary rise, h f ( = a),
h 4/a
= cos , cos ln E (4.4.23)
1 + sin
(Derjaguin (1946) Dokl. Akad. Nauk USSR 51, 517). The predictions of this equation, represented
by the dashed lines in Figure 4.4.1(b), are in excellent agreement with the numerical solution. A
formal asymptotic expansion with respect to the small parameter a/ is available [242].
304 Introduction to Theoretical and Computational Fluid Dynamics
(a) (b)
x
g
d
Fluid 2 Fluid 2
Fluid 1
Fluid 1
d
g
x
Figure 4.4.3 Illustration of (a) an axisymmetric sessile liquid drop resting on a horizontal plane, and
(b) an axisymmetric pendant liquid drop hanging under a horizontal plate.
Evaluating (4.4.24) at the origin, we nd that = f (0), which shows that the constant is equal
to twice the mean curvature at the centerline.
Parametric description
To formulate the numerical problem, we introduce the slope angle dened by the equation cot =
f , ranging from zero at the centerline to the contact angle at the contact line, and regard x and
along the interface as functions of , as shown in Figure 4.4.3(a). The contact angle boundary
condition is satised automatically by this parametrization. Following the procedure that led us from
equation (4.4.1) to equations (4.4.7) and (4.4.9), we resolve (4.4.24) into a system of two rst-order
equations,
dx sin d cos
= , = , (4.4.25)
d Q d Q
where
sin x
Q= + 2 . (4.4.26)
The boundary conditions require that (0) = 0 and x(0) = 0. One more condition emerges by
4.4 Axisymmetric interfaces 305
In terms of the non-dimensional variables X = x/ and = /, equations (4.4.25) become
dX sin d cos
= , = , (4.4.30)
d W d W
with boundary conditions (0) = 0 and X(0) = 0, where
sin
W = + X , (4.4.31)
and . The volume constraint (4.4.27) becomes
0
2 dX = V , (4.4.32)
D
where D = d/ and V = V /3 . Equations (4.4.29) state that, at the origin,
dX
d
2
= 0, = . (4.4.33)
d 0 d 0
The shape of the drop depends on the reduced volume, V , and contact angle, . Given the
values of V and , the drop shape can be found using a shooting method according to the following
steps:
1. Guess the value of the constant . A good estimate can be obtained by assuming that the
drop has a spherical shape with radius a = [3V /(4)]1/3 . Since the constant is equal to the
mean curvature at the centerline, we may set = 1/a and compute = /a.
2. Solve an initial-value problem from by integrating (4.4.30) from = 0 to . To initialize the
computation, we use (4.4.33).
306 Introduction to Theoretical and Computational Fluid Dynamics
2 2
1.5 1.5
x/a
x/a
1 1
0.5 0.5
0 0
Figure 4.4.4 (Left) Shapes of a sessile drop for dimensionless volume V V /3 = 4/3 and contact
angle / = 0.1, 0.2, . . . , 0.9, and (right) corresponding shapes for a smaller volume, V = 4/34 .
The axes are scaled by the equivalent drop radius, a = (3V /4)1/3 .
3. Check whether (4.4.32) is fullled. If not, repeat the computation with a new and improved
value for .
Drop proles for volume V = 4/3 and 4/34 and contact angle / = 0.1, 0.2, . . . , 0.9 are shown in
Figure 4.4.4. When the reduced volume V is small, the drops take the shape of sections of sphere.
As the drop volume increases, the interface tends to atten at the top due to the action of gravity.
Problems
Computer Problems
x
Fluid 1
g
xc a
Fluid 2
Figure 4.4.5 Illustration of a spherical particle straddling the interface between two immiscible uids.
Numerical implementations and variational formulations are discussed by Brown [58], Mili-
nazzo & Shinbrot [266], Hornung & Mittelmann [191], and Li & Pozrikidis [237]. Finite-dierence
calculations of liquid bridges connecting three equal spheres are presented by Rynhart et al. [356].
Finite-element calculations of wetting modes of superhydrophobic surfaces are presented by Zheng,
Yu & Zhao [441]. Finite-dierence methods combined with conformal mapping have been imple-
mented in recent work [320, 319, 321].
308 Introduction to Theoretical and Computational Fluid Dynamics
Figure 4.6.1 A meniscus forming between two vertical circular cylinders whose centers are separated
by distance equal to the cylinder diameter, for contact angle = /4, and capillary length /a =
4.5036, where a is the cylinder radius.
4.6 Software
Directory 03 hydrostat of the uid mechanics library Fdlib contains a collection of programs that
compute two-dimensional, axisymmetric, and three-dimensional hydrostatic shapes, as discussed in
Appendix C.
The meniscus forming around a vertical elliptical cylinder is shown in Figure 4.2.1 for contact angle
= /4. The interfacial shape was computed by solving the YoungLaplace equation using a
nite-dierence method implemented in boundary-tted elliptic coordinates generated by conformal
mapping [320]. The computer code resides in Directory men ell inside Directory 03 hydrostat of
Fdlib. For clarity, the height of the cylinder shown in Figure 4.2.1 reects the elevation of the
contact line. We observe that the elliptical cross-section causes signicant oscillations in the capillary
elevation around the value corresponding to a circle. High elevation occurs on the broad side of the
cylinder, and low elevation occurs at the tip of the cylinder.
We conclude that high boundary curvature causes a local decline in the capillary height below
the value corresponding to a at plate. In the case of a meniscus attached to a wavy wall, the contact
line elevation is higher in the troughs than in the crests of the corrugations [179]. This behavior is
consistent with our analysis in Section 4.2.1 on the shape of a meniscus inside a dihedral corner.
The meniscus developing between two vertical circular cylinders with the same radius is shown in
Figure 4.6.1. The interfacial shape was generated using the computer code men cc in Directory
03 hydrostat of Fdlib. The YoungLaplace equation is solved by a nite-dierence method in
boundary-tted bipolar coordinates. In the conguration shown in Figure 4.6.1, the cylinder centers
4.6 Software 309
are separated by distance equal to the common cylinder diameters. The contact angle has a specied
value, = /4, around both contact lines. We observe that the rise of the meniscus is most
pronounced inside the gap between the cylinders.
Surface evolver
The surface evolver program is able to describe triangulated surfaces whose shape is determined
by surface tension and other types of surface energy, subject to various constraints. Further infor-
mation is given in the Internet site: http://www.susqu.edu/brakke/evolver/evolver.html.
Computer Problem
In this chapter, we discuss a family of ows whose solution can be found either analytically or
numerically by solving ordinary and elementary one-dimensional partial dierential equations. We
begin by considering unidirectional ows in channels and tubes where the nonlinear convection term
in the equation of motion is identically zero. The reason is that uid particles travel along straight
paths with constant velocity and vanishing acceleration. Swirling ows inside or outside a circular
tube and between concentric cylinders provides us with a family of ows with circular streamlines
where the nonlinear term due to the particle acceleration assumes a simple tractable form. Solutions
for unsteady ows will be derived by separation of variables and similarity solutions, when possible.
Following the discussion of unidirectional ows, we address a rare class of ows in entirely or
partially innite domains that can be computed by solving ordinary dierential equations without
any approximations. Reviews and discussion of further exact solutions can be found in articles by
Berker [31], Whitham [428], Lagestrom [218], Rott [355], and Wang [419, 420, 421]. These exact
solutions are complemented by approximate, asymptotic, and numerical solutions for low- or high-
Reynolds number ows discussed later in this book.
310
5.1 Steady unidirectional ows 311
In Chapters 7, 8, and 10, we discuss precisely and nearly irrotational ows where the vorticity
is conned inside slender wakes and boundary layers. The velocity eld in the bulk of the ow is
obtained by solving Laplaces equation for a harmonic velocity potential. Once the outer irrotational
ow is available, the ow inside boundary layers and wakes is computed by solving simplied versions
of the equation of motion originating from the boundary-layer approximation. In Chapter 11, we dis-
cuss numerical methods for computing inviscid or nearly inviscid ows containing regions or islands
of concentrated vorticity, including point vortices, vortex rings, vortex laments, vortex patches, and
vortex sheets. In the nal Chapters 12 and 13, we discuss nite-dierence methods for solving the
complete system of governing equations for NavierStokes ow without any approximations apart
from those involved in the implementation of the numerical method.
The pressure distribution is recovered by integrating (5.1.1) and the last two equations in
(5.1.2), yielding
p = x + (gy y + gz z) + p0 , (5.1.3)
312 Introduction to Theoretical and Computational Fluid Dynamics
Poisson equation
The x component of the equation of motion (5.1.2) provides us with a Poisson equation for the
streamwise velocity component with a constant forcing term on the right-hand side,
2 ux 2 ux + gx
2
+ 2
= . (5.1.4)
y z
The boundary conditions specify the distribution of ux along a channel or tube wall or the shear
stress n ux along a free surface, where is the two-dimensional gradient operating in the yz
plane and n is the unit vector normal to the free surface. In the case of multilayer ow, the shear
stress, n ux , is required to be continuous across the layer interface. Accordingly, the normal
derivative of the velocity n ux undergoes a jump determined by the viscosities of the two uids.
When a channel or tube is horizontal and the walls are stationary, we obtain pressure-driven
ow due to an externally imposed pressure gradient, p/x = .
When the walls are stationary and the pressure does not change in the direction of the ow,
= 0, we obtain gravity-driven ow.
When = gx , the pressure assumes the hydrostatic distribution and the ow is driven by
the longitudinal translation of the whole wall or a section of a wall.
In the third case, the Poisson equation (5.1.4) reduces to Laplaces equation, 2 ux = 0, whose
solution is independent of the viscosity, . Mixed cases of pressure-, gravity-, and boundary-driven
ow can be constructed by linear superposition.
d2 ux + gx
= . (5.1.5)
dy 2
Setting the origin of the Cartesian axes at the lower wall, integrating twice with respect to y, and
enforcing the no-slip boundary conditions u = V1 at y = 0 and u = V2 at y = h, we nd that the
5.1 Steady unidirectional ows 313
V2
x
0
V1
Figure 5.1.1 Velocity proles of ow through a two-dimensional channel with parallel walls separated
by distance h. The linear prole (dashed line) corresponds to Couette ow, the parabolic prole
(dotted line) corresponds to plane HagenPoiseuille ow, and the intermediate prole (dot-dashed
line) corresponds to ow with vanishing ow rate.
y + gx
ux (y) = V1 + (V2 V1 ) + y (h y). (5.1.6)
h 2
The rst two terms on the right-hand side represent a boundary-driven ow. The third term rep-
resents pressure- and gravity-driven ow. When = gx , we obtain Couette ow with a linear
velocity prole, also called simple shear ow [94]. When V1 = 0 and V2 = 0, we obtain plane
HagenPoiseuille pressure- or gravity-driven ow (e.g., [392]). Velocity proles corresponding to the
extreme cases of Couette ow, HagenPoiseuille ow, and a ow with vanishing ow rate are shown
in Figure 5.1.1.
The ow rate per unit width of the channel is found be integrating the velocity prole,
h
1 1 + gx 3
Q= ux dy = (V1 + V2 ) h + h . (5.1.7)
0 2 12
Note the dependence on h or h3 , respectively, in the rst or second term on the right-hand side.
The mean uid velocity is
Q 1 1 + gx 2
umean
x = (V1 + V2 ) + h . (5.1.8)
h 2 12
In the case of pressure- or gravity-driven ow, the maximum velocity occurring at the midplane,
umax
x = ux (y = 12 h) is related to the mean velocity by umax
x = 32 umean
x .
314 Introduction to Theoretical and Computational Fluid Dynamics
must recirculate inside the channel. Near the side walls, we Flow in a long horizontal channel
obtain a reversing ow. Far from the side walls, we obtain that is closed at both ends.
nearly unidirectional ow. Equation (5.1.7) reveals the spontaneous onset of a pressure eld with
pressure gradient
p
= 6 2 (V1 + V2 ), (5.1.9)
x h
inducing a back ow that counteracts the primary ow due to the belt motion. The velocity prole
is illustrated by the dot-dashed line in Figure 5.1.1.
One constraint arises by stipulating that the rate of displacement of the liquid by the slab is
equal to the upward ow rate through the gaps,
Q = V b, (5.1.10)
5.1 Steady unidirectional ows 315
1 (a b)3
V + ( + g) = 0. (5.1.11)
6 a+b
Note that Q is negative, as required for the uid to escape upward. A second constraint arises by
balancing the x components of the surface and body forces exerted on the slab. Requiring that the
pressure force at the top and bottom counterbalances the weight of the slab and the force due to
the shear stress along the sides, we nd that
du
x
(ptop pbot ) b + s gLb + L=0 (5.1.12)
dy y=0
or
V + g
( + g) b + (s ) gb + + h = 0, (5.1.13)
h 2
where s is the density of the slab material. After simplication, we obtain
1
V ( + g) (a2 b2 ) = gb(a b), (5.1.14)
2
where = s . Solving (5.1.11) and (5.1.15) for V and + g, we obtain
1 ga2 (1 )3 3 (1 + )
V = , + g = g , (5.1.15)
4 1 + + 2 2 1 + + 2
where = b/a, independent of the slab length, L. We observe that the settling velocity, V , vanishes
when = 0 because the weight of the slab is innitesimal, and also when = 1 because the sides
of the slab stick to the side plates. The maximum settling velocity occurs at the intermediate value
0.204. In the case of a neutrally buoyant slab, = 0, the uid is quiescent and the pressure
assumes the hydrostatic distribution, = g.
In the inclined Cartesian coordinates depicted in Figure 5.1.2, the x velocity component, ux ,
depends on the distance from the inclined plane alone, y. The components of the acceleration of the
gravity vector are g = g (sin , cos , 0), where is the plane inclination angle with respect to the
horizontal, dened such that = 0 corresponds to a horizontal plane and = 12 corresponds to a
vertical plane. Since the pressure outside and therefore inside the lm is independent of x, we set
= 0. The pressure distribution across the lm is
h p
atm
Figure 5.1.2 Illustration of a at liquid lm with thickness h owing down an inclined plane due to
gravity.
where patm is the ambient atmospheric pressure. Integrating the dierential equation (5.1.5) twice
subject to the no-slip boundary condition at the wall, ux = 0 at y = 0, and the condition of vanishing
shear stress at the free surface, ux /y = 0 at y = h, we derive a semi-parabolic velocity prole rst
deduced independently by Hopf in 1910 and Nusselt in 1916,
g
ux = sin y (2h y), (5.1.17)
2
where = / is the kinematic viscosity of the uid (e.g., [139]). The velocity at the free surface
and ow rate per unit width of the lm are
h
g 2 g 3
umax
x = h sin , Q = ux dy = h sin . (5.1.18)
2 0 3
The mean velocity of the liquid is
Q g 2 2
umean
x = h sin = umax . (5.1.19)
h 3 3 x
In Section 6.4.4, these equations will be used to describe unsteady, nearly unidirectional lm ow
down an inclined plane with a mildly sloped free surface prole.
Problem
5.1.1 Two-layer ow
(a) Consider the ow of two superposed layers of two dierent uids in a channel with parallel walls
separated by distance h. Derive the velocity prole across the two uids in terms of the physical
properties of the uids and ow rates for the mixed case of boundary- and pressure-driven ow.
(b) Repeat (a) for gravity driven ow of two layers owing down an inclined plate.
(a) (b)
x=0 a
x=L
Figure 5.2.1 Illustration of (a) pressure- and gravity-driven ows and (b) boundary-driven ow through
a tube with circular cross-section of radius a.
prole in tube ow depends on two spatial coordinates determining the position over the tube
cross-section in the yz plane.
xx x=L xx x=0 dS + x 2L + gx 2 L = 0. (5.2.7)
bottom,top
2L + gx 2 L = 0.
x
p x=L p x=0 dS + (5.2.8)
bottom,top d
Treating the pressure dierence inside the rst integral as a constant, integrating over the cross-
section, and solving for the rst derivative of the velocity, we obtain
dux + gx
= , (5.2.9)
d 2
where = (p)x=L (p)x=0 /L is the negative of the pressure gradient. Equation (5.2.9) is the
rst integral of (5.2.1). Integrating (5.2.9) once subject to the no-slip boundary condition ux = 0 at
= a, we recover the parabolic prole (5.2.2).
3
b b
a a
y y y
a a
1 2
a
Figure 5.2.2 Illustration of pressure- and gravity-driven ows through a tube with (a) elliptical, (b)
equilateral triangular, and (c) rectangular cross-section.
the wall roughness, as discussed in Section 9.8.6 in the context of hydrodynamic stability. As the
Reynolds number increases above this threshold, the streamlines becomes wavy, random motion
appears, and a turbulent ow is ultimately established.
Boundary-driven ow
Next, we consider ow in a horizontal circular tube driven by the translation of a sector of the tube
wall with aperture angle 2 conned between the azimuthal planes < < in the absence of
a pressure drop, as shown in Figure 5.2.1(b). To compute the velocity prole, we solve Laplaces
equation, 2 ux = 0, subject to the no-slip boundary condition ux = 0 at = a, except that ux = V
at = a and < < , where V is the boundary velocity. The solution can be found using the
Poisson integral formula (2.5.13), and is given by
a2 2 d
ux (x0 ) = V . (5.2.11)
a + 2a cos(0 )
2 2 2
Sparrow [382] derives the velocity prole in a isosceles triangular tube in terms of an innite
series. Other triangular proles will be discussed in Section 5.2.9.
5.2 Steady tube ows 321
is a dimensionless function and n = (n 12 ) . Integrating the velocity prole over the tube
cross-section, we derive the ow rate
4 + gx 3 a
Q= ab ( ), (5.2.24)
3 b
where
6 1
() = 1 tanh(n ). (5.2.25)
n=1 n5
R2
R1
V2
V1
g
x
Figure 5.2.3 Illustration of ow in the annular space conned two inclined innite concentric cylinders.
If the tube is closed at both ends, the pressure drop is such that the ow rate is zero. When the
clearance of the annular channel is small compared to the inner cylinder radius, h R2 R1 R1 ,
the curvature of the walls becomes insignicant and expressions (5.2.26) and (5.2.27) reduce to
(5.1.6) and (5.1.7) with y = R1 describing ow in a channel conned between two parallel walls
(Problem 5.2.8).
ux () = V + a 2 (a2 c2 ) , (5.2.28)
ln 4 ln
where = a/c > 1. The corresponding ow rate is
1 a2 c2 2 a2 c 2
Qgap = V c2 + (a c2 ) (a2 + c2 ). (5.2.29)
2 ln 8 ln
The axial force exerted on the object can be computed by summing the pressure force normal to
the two discoidal sides and the shear force exerted on the cylindrical side with surface area 4bc,
yielding
u
c a
Uc b x
b
V
Figure 5.2.4 Flow past or due to the motion of a tightly tting cylinder of radius c inside a cylindrical
tube of radius a > c.
Two modular cases are of interest. In the rst case, the cylinder moves with velocity V
inside a tube that is closed at both ends. Mass conservation requires that the ow rate through
the narrow annular gap balances the rate of displacement of the uid on either side of the cylinder,
Qgap = V c2 . Evaluating the ow rate using (5.2.27) with R1 = c, R2 = a, V1 = V , V2 = 0, and
gx = 0, and rearranging, we obtain the negative of the pressure gradient,
4V 2
V = . (5.2.33)
a2 2 + 1 + ( 2 + 1) ln
The second fraction on the right-hand side is identied with a positive dimensionless pressure gra-
dient coecient, cVp . The velocity prole arises by setting in (5.2.28) = V . Substituting (5.2.33)
into (5.2.32), we obtain the drag force
2 + 1
Fx = 4bV . (5.2.34)
2 + 1 + ( 2 + 1) ln
The associated drag coecient is
Fx 2 2 + 1
cV = , (5.2.35)
6cV 3 + 1 + ( 2 + 1) ln
2
In the second modular case, the cylinder is stationary and the uid is forced to move at a
constant ow rate, Qgap = 12 Uc a2 , where Uc is the centerline velocity of an equivalent Poiseuille
ow. Evaluating the ow rate using (5.2.27) with R1 = c, R2 = a, V1 = 0, V2 = 0, and gx = 0, and
rearranging, we obtain the negative of the pressure gradient,
4Uc 4
U = . (5.2.36)
a2 4 1 ( 2 1)2 / ln
324 Introduction to Theoretical and Computational Fluid Dynamics
r
y
a
b
Figure 5.2.5 Illustration of a circular tube of radius a indented with a circular sector of radius b.
The second fraction on the right-hand side can be identied with a positive dimensionless pressure
gradient coecient, cU p . The velocity prole arises by setting in (5.2.28) V = 0 and = U .
Substituting (5.2.36) into (5.2.32), we obtain the drag force
2
Fx = 4b Uc . (5.2.37)
2 + 1 + ( 2 + 1) ln
The associated drag coecient is
Fx 2 2
cU = . (5.2.38)
6cUc 3 2 + 1 + ( 2 + 1) ln
The negative of the pressure drop and drag force coecient in the general case arise by
superposition,
4
= V + U = (V cVp Uc cU
p ), Fx = FxV + FxU = 6c (V cVD Uc cU
D ). (5.2.39)
a2
In the case of a freely suspended particle, we set Fx = 0 and nd that V /Uc = cU /cV .
axes has been placed at the leftmost point of the unperturbed circular contour. The axial velocity,
ux , satises the Poisson equation
1 ux
1 2 ux + gx
2 ux = r + 2 = , (5.2.40)
r r r r 2
where r is the distance from the origin and is the corresponding polar angle. Using the law of
cosines, we nd that
where is the distance from the center of the outer circle. The boundary conditions specify that
ux = 0 at = a and r = 2a cos (outer circle), and ux = 0 at r = b (inner circle). The velocity
prole can be constructed by modifying the parabolic prole of the Poiseuille ow, obtaining
+ gx 2 b2 + gx a
ux = (a 2 )(1 2 ) = (2 cos 1)(r 2 b2 ). (5.2.42)
4 r 4 r
When b = 0, we recover Poiseuille ow through a circular tube.
The no-slip boundary condition requires that f = v around the tube contour. The problem
has been reduced to computing the harmonic function f subject to the Dirichlet boundary condition.
The theory of functions of a complex variable guarantees that f can be regarded as the real part of
an analytic function, G(w), where w = y + iz is the complex variable in the yz plane and i is the
imaginary unit, f = Real{G(w)}. To compute the solution, we introduce a function w = F() that
maps a disk of radius centered at the origin of the complex plane to the cross-section of the tube
in the physical w plane, as discussed in Section 7.11. The tube contour is described parametrically
by the functions
yc () = Real F(ei ) , zc () = Imag F(ei ) , (5.2.45)
326 Introduction to Theoretical and Computational Fluid Dynamics
where is the polar angle in the plane, 0 < 2. Writing = ei and using the Poisson
integral (2.5.13), we obtain
2 2 2 v[ yc ( ), zc ( ) ]
f (y, z) = d , (5.2.46)
2 0 2 + 2 2 cos( )
where
y(, ) = Real F(ei ) , z(, ) = Imag F(ei ) , (5.2.47)
and 0 . In general, the integral in (5.2.46) must be evaluated by numerical methods. The
computation of the required mapping function F() is discussed in Section 7.11.
+ gx a2 b2
ux (y, z) = F(y, z), (5.2.48)
2 a2 + b 2
where a and b are two specied lengths,
y2 z2
F(y, z) = 1 H(y, z), (5.2.49)
a2 b2
is a suitable function, and H(y, z) is a harmonic function in the yz plane satisfying
2H 2H
+ = 0. (5.2.50)
y 2 z 2
The tube contour where ux = 0 is described implicitly by the equation F(y, z) = 0. When H = 0,
we obtain an elliptical tube with semiaxes a and b. More generally, the tube shape must be found
by numerical methods.
To obtain tubes with polygonal proles, we dene the complex variable w = y + iz and
introduce the function
w
m
H(y, z) = Imag ei , (5.2.51)
c
where is a dimensionless coecient, i is the imaginary unit, i2 = 1, c is a specied length, m is
a specied integer determining the polygonal shape, and is a specied phase angle.
5.2 Steady tube ows 327
(a) (b)
0.6 0.6
0.4 0.4
0.2 0.2
z/
z/
0 0
0.2 0.2
0.4 0.4
0.4 0.2 0 0.2 0.4 0.6 0.4 0.2 0 0.2 0.4 0.6
y/ y/
(c) (d)
1 1
0.5 0.5
z/
z/
0 0
0.5 0.5
1 1
1 0.5 0 0.5 1 1 0.5 0 0.5 1
y/ y/
Figure 5.2.6 (a, b) Tube contours resembling triangles with rounded corners ranging between a circle
and a triangle. (c, d) Tube contours resembling squares or rectangles with rounded corners.
Rounded triangles
1
For a = b = c = 3 , m = 3, and = 0, we obtain
where y = y/, z = z/, and is a specied length [422]. When = 0, the tube contour is a circle
of radius 13 centered at the origin. Comparison with (5.2.18) reveals that, when = 2/33/2 , the
tube contour is an equilateral triangle with side length equal to . Figure 5.2.6(a) shows contours
describing equilateral triangles with rounded corners for intermediate values of .
F(y, z) = 1 y 2 + z 2 (y 4 6 y 2 z 2 + z 4 ). (5.2.53)
When = 0, the tube contour is a circle of radius . As increases, squares with rounded corners
appear, as shown in Figure 5.2.6(c).
Problems
(1)n z y
ux (y, z) = V 1 + 2 exp(n ) cos(n ) , (5.2.55)
n=1
n a a
where n = (n 12 ) (e.g., [318]). The origin has been set on the painted surface midway between
two plates; the z axis is perpendicular to the painted surface.
(b) The amount of liquid deposited on the surface per channel is
a
Q= [ V ux (y, z) ] dy dz. (5.2.56)
0 a
5.2 Steady tube ows 329
Show that
a
(1)n z y 1
Q = 2V exp(n ) cos(n ) dy dz = 4V a2 , (5.2.57)
n=1
n 0 a a a 3
n=1 n
yielding Q 1.085 V a2 . The dependence of the ow rate on the second power of a linear boundary
dimension, in this case, a, is typical of boundary-driven ow. Explain why the thickness of the lm
left behind the brush is h = Q/(2V a).
(c) Repeat (a) and (b) assuming that the brush planes are inclined at an angle with respect to
the painted surface. Discuss the relation between Q and .
g
(1)n cosh[n (1 + z/a)] y
ux = sin a2 y2 + 4a2 3
cos(n ) (5.2.58)
2 n=1
n cosh(2n ) a
and
g
tanh(2n )
where n =
2 (2n 1) and is the kinematic viscosity of the uid.
Computer Problem
dp u2 d 1 d (u )
(5.3.1)
= , = 0.
d d d
The rst equation states that the centrifugal force is balanced by a spontaneously developing radial
pressure gradient. Integrating the second equation subject to the boundary conditions u = 1 R1
at = R1 and u = 2 R2 at = R2 , we obtain the velocity prole
2 1 2 1 R12
u () = , (5.3.2)
1 1
5.3 Steadily rotating ows 331
2
where R1 /R2 < 1 is a geometrical dimensionless parameter. When 2 = 1 , the uid
rotates like a rigid body with azimuthal velocity u () = . When 2 = 1 , we obtain the
ow due to a point vortex with strength = 21 R12 = 22 R22 located at the axis, generating
irrotational ow. As the clearance of the channel becomes decreasingly small compared to the radii
of the cylinders, 1, the ow in the gap resembles plane Couette ow in a channel with parallel
walls (Problem 5.3.1).
The pressure distribution is found by substituting the velocity prole (5.3.2) into the rst
equation of (5.3.1). When the uid rotates as a rigid body, 2 = 1 , we obtain
1
p= 2 2 + g x + p 0 , (5.3.3)
2
where p0 is a constant. The rst term on the right-hand side is the pressure eld established in
response to the centrifugal force. When the ow resembles that due to a point vortex, we obtain
1 R2
p = U12 21 + g x + p0 , (5.3.4)
2
where U1 = 1 R1 .
The circular Couette ow device provides us with a simple method for computing the viscosity
of a uid in terms of the torque exerted on the inner or outer cylinder, given by
2 1
T 2 2 = 4 R12 . (5.3.5)
1
Note that the shear stress, , and thus the torque is zero in the case of rigid-body rotation.
The stability of the circular Couette ow will be discussed in Section 9.9. Linear stability
analysis shows that the laminar ow discussed in this section is established only when the pair
(1 , 2 ) falls inside a certain range. Wavy motion leading to cellular ow is established outside this
range.
5.3.2 Ekman ow
Consider a semi-innite body of uid residing in the lower half-space from z = 0 to and rotating
around the z axis with constant angular velocity . We refer to a noninertial frame of reference
that rotates with the uid with angular velocity, = ez , and assume that the vertical velocity
component, uz , is zero while the horizontal velocity components, ux and uy , depend only on z and
are independent of x and y. Simplifying the equation of motion in the rotating noninertial frame by
setting the left-hand side of (3.2.43) to zero, we obtain
0 = p + 2 u + g + [2 u ( x) ]. (5.3.6)
The rst term inside the square brackets on the right-hand side is the Coriolis force, and the second
term is the centrifugal force. The x, y, and z components of (5.3.6) are
p d2 ux
0= + + gx + (2uy x),
x dz 2
p d 2 uy p
0= + 2
+ gy + (2ux y), 0= + gz . (5.3.7)
y dz z
332 Introduction to Theoretical and Computational Fluid Dynamics
The solution is
1
p = g x + 2 (x2 + y 2 ) + p0 , (5.3.8)
2
where p0 is a constant reference pressure, and the velocity components ux and uy satisfy the coupled
ordinary dierential equations
d2 ux d2 uy
= 2 u y , = 2 ux . (5.3.9)
dz 2 dz 2
It is convenient to collect these equations into the compact complex form
d2 (ux + i uy )
= 2 i (ux + iuy ), (5.3.10)
dz 2
where i is the imaginary unit, i2 = 1, and = / is the kinematic viscosity of the uid. Stipulating
that ux = Ux and uy = Uy at z = 0, and that ux and uy decay as z tends to negative innity, we
obtain
|z|
ux + i uy = (Ux + i Uy ) exp (1 + i) , (5.3.11)
where (2/)1/2 is the Ekman layer thickness. We observe that the velocity components in a
plane that is normal to the axis of rotation decay exponentially with downward distance from the
surface of the liquid.
Problem
ux 2 ux
= , (5.4.3)
t y 2
known as the one-dimensional unsteady heat conduction equation, where = / is the kinematic
viscosity of the uid.
The velocity is required to decay far from the plate, as y . The no-slip boundary condition
at the plate requires that
ux (y = 0, t) = V cos(t), (5.4.4)
where is the angular frequency of the oscillations and V is the amplitude of the vibrations. It is
convenient to express this boundary condition in the form
ux (y = 0, t) = V Real eit , (5.4.5)
where i is the imaginary unit, i2 = 1. Motivated by the linearity of (5.4.3), we also set
ux (y, t) = V Real f (y) eit , (5.4.6)
where f (y) is a dimensionless complex function satisfying the boundary condition f (0) = 1 and the
far-eld condition f () = 0. Substituting (5.4.6) into (5.4.3), we derive a linear ordinary dierential
equation
if = f , (5.4.7)
where a prime denotes a derivative with respect to y. The solution is readily found to be
where y y/ and
%
2 (5.4.9)
=
is the Stokes boundary-layer thickness. Substituting (5.4.8) into (5.4.6), we obtain
ux (y, t) = V Real exp[ it (1 i) y ] , (5.4.10)
(a) (b)
5 5
4 4
3 3
y/
y/
2 2
1 1
0 0
1 0.5 0 0.5 1 1 0.5 0 0.5 1
ux/V ux/U
Figure 5.4.1 Velocity proles (a) due to the in-plane oscillation of a plate in a semi-innite viscous
uid and (b) in oscillatory ow over a stationary plate. Proles are shown at phase angles t/ = 0
(dashed line), 0.25 (dash-dotted line), 0.50 (dotted line), 0.75, 1.0, 1.25, 1.50, 1.75, and 2.0. The
y coordinate is scaled by the Stokes boundary layer thickness dened in (5.4.9).
for y 0. We have found that the velocity prole is an exponentially damped wave with wave
number 1/ and associated wavelength 2, propagating in the y direction with phase velocity
cp = = (2)1/2 . Since the amplitude of the velocity decays exponentially with distance from
the plate, the motion of the uid is negligible outside a wall layer of thickness . Velocity proles at
a sequence of time instants over one period of the oscillations are shown in Figure 5.4.1(a).
It is interesting to note that the phase shift between the wall shear stress and the velocity of the
plate is 3/4, independent of the angular frequency, . This is a unique feature of oscillatory ow
driven by the motion by a at surface in the absence of other boundaries. We will see later in this
section that, for other geometries, the phase shift depends on the frequency of oscillation.
p
= gx + sin(t), (5.4.13)
x
5.4 Unsteady unidirectional ows 335
where is a constant amplitude. The velocity prole is governed by the following simplied version
of the general equation (5.4.2),
ux 2 ux
= sin(t) + . (5.4.14)
t y 2
The solution is
ux (y, t) = v(y, t) cos(t), (5.4.15)
where the complementary velocity v is given by the right-hand side of (5.4.11) with V = /(),
yielding
ux (y, t) = cos(t y) ey cos(t) . (5.4.16)
The velocity is zero over the plate located at y = 0, and describes uniform oscillatory ow with
velocity ux (, t) = U cos(t) far from the plate, where U = /(). Physically, the ow consists
of an outer potential core and a Stokes boundary layer with thickness = (2/)1/2 adhering to the
plate. Velocity proles at a sequence of time instants over one period are shown in Figure 5.4.1(b).
ux (y = 0, t > 0) = V, ux (y , t) = 0. (5.4.17)
Because of the linearity of the governing equation and boundary conditions, we expect that the uid
velocity will be proportional to the plate velocity, V , and set
ux = V f (y, t, ), (5.4.18)
where f is a dimensionless function. The arguments of f (y, t, ) must combine into dimensionless
groups. In the absence of external length and time scales, we formulate a combination in terms of
the similarity variable
y
= , (5.4.19)
t
so that f (y, t, ) = F (), where F (), is an unknown dimensionlessfunction. This means that the
velocity as seen by an observer who nds
herself at the position y = t, and is thus traveling along
the y axis with velocity v = dy/dt = /4t, remains constant in time. The boundary and far-eld
conditions require that F (0) = 1 and F () = 0. Substituting into (5.4.3) the self-similar velocity
prole ux = V F (), we obtain
dF dF
= . (5.4.20)
d t y d y
336 Introduction to Theoretical and Computational Fluid Dynamics
(a) (b)
1 2
0.8
1.5
erfc(w)
0.6
y/a
1
erf(w)
0.4
0.5
0.2
0 0
0 0.5 1 1.5 2 2.5 3 0 0.2 0.4 0.6 0.8 1
w ux/V
Figure 5.4.2 (a) Graphs of the error function (solid line) and complementary error function (dashed
line) computed using algebraic approximations. (b) Velocity proles of the ow due to the sudden
motion of a plate in a semi-innite uid at dimensionless times t/a2 = 0.001, 0.005, 0.02, 0.05,
0.1, 0.2, 0.5, and 1.0, where a is an arbitrary length scale,
Carrying out the dierentiations, we derive a second-order nonlinear ordinary dierential equation
1 dF d2 F
= . (5.4.21)
2 d d 2
Integrating twice subject to the aforementioned boundary and far-eld conditions, we obtain the
velocity prole
/2
ux 2 2
= F () = erfc 1 erf 1 e d. (5.4.22)
V 2 2 0
The complementary error function, erfc(w), and the error function, erf(w), can be computed by
accurate polynomial approximations. Graphs are presented in Figure 5.4.2(a). As w tends to
innity, erf(w) tends to unity and correspondingly erfc(w) tends to zero. The converse is true as w
tends to zero.
A sequence of evolving velocity proles plotted with respect to y/a are shown in Figure 5.4.2(b)
at a sequence of dimensionless times t/a2 , where a is an arbitrary length. At the initial instant, the
velocity prole is proportional to the discontinuous Heaviside function, reecting the sudden onset
of a vortex sheet attached to the plate,
The dimensionless Heaviside function, H(t), is dened such that H(t) = 0 for t < 0 and H(t) = 1 for
t > 0. As time progresses, the vortex sheet diuses into the uid, transforming into a vortex layer
of growing thickness.
5.4 Unsteady unidirectional ows 337
A singularity appears at the initial instant, t = 0, when the plate starts moving. Physically, the
plate cannot be pushed impulsively with constant velocity, but must be accelerated gradually from
the initial to the nal value over a nite period of time. Setting the shear stress equal to V /,
where is the eective
thickness of the vorticity layer, we nd that increases in time like the square
root of time, t.
x
=V (t) = V 1 (t), (5.4.25)
t y=0 t y=0
In the case of arbitrary plate translation with arbitrary time-dependent velocity V (t) and V (0) = 0,
we obtain
t+t t+t F
where t is an innitesimal time interval. This expression allows us to construct the ow due to the
translation of the plate by generalization to arbitrary y. Assuming that the uid is quiescent for
t < 0, we obtain
t F
This solution also describes the temperature distribution in a semi-innite conductive medium due
to an arbitrary boundary temperature (e.g., ([66]). The right-hand side of (5.4.29) expresses a
distribution of Greens functions of the unsteady diusion equation (6.17.4) given in (6.17.5).
Oscillatory Couette ow
Assume that the upper wall is stationary while the lower wall oscillates in its plane along the x axis
with angular frequency . The lower wall velocity is ux (y = 0, t) = V cos(t), where V is a constant
amplitude. Working as in Section 5.4.1 for ow in a semi-innite uid due to an oscillating plate,
we derive the velocity prole
ux (y, t) = V Real f (y) eit , (5.4.30)
where
exp[(1 i)y] exp[(1 i)(2h y)]
f (y) = , (5.4.31)
1 exp[(1 i) 2h]
y y/, h h/, and (2/)1/2 is the Stokes boundary-layer thickness. The structure of the
ow can be regarded as a function of the dimensionless Womersley number dened with respect to
half the channel width,
%
h 1 h 1
Wo = = h. (5.4.32)
2 2 2
When the Womersley number is small, the ow evolves in quasi-steady fashion and the velocity prole
is nearly linear with respect to y at any time. As the Womersley number tends to innity, we recover
the results of Section 5.4.1 for ow due to an oscillating plate in a semi-innite uid. Comparing the
present ow with that due to the oscillating plate, we nd that the phase shift between the velocity
and shear stress at the lower wall depends on the angular frequency, . This dierence underscores
the hydrodynamic importance of the second plate or another nearby boundary.
Transient Couette ow
Next, we assume that the upper wall is stationary while the lower wall is suddenly set in motion
parallel to itself along the x axis with constant velocity V . Initially, the ow resembles that due to
5.4 Unsteady unidirectional ows 339
the motion of a at plate immersed in a semi-innite uid. As time progresses, we obtain Couette
ow with a linear velocity prole. To expedite the solution, we decompose the ow into the steady
Couette ow and a transient ow that decays at long times. Applying the method of separation of
variables in y and t, we nd that the velocity is given by a Fourier series [318],
ny
y 2 1 n2 2
ux (y, t) = V 1 exp( t) sin . (5.4.33)
h n=1 n h2 h
Working in an alternative fashion, we apply the Laplace transform method and obtain
y
2h y
2h + y
is a complex function, y y/, h h/, and (2/)1/2 is the Stokes boundary-layer thickness
(e.g., [318]). It is sometimes useful to regard the structure of the ow as a function of the dimen-
sionless Womersley number dened in (5.4.32). When the Womersley number is small, we obtain
quasi-steady plane HagenPoiseuille ow with parabolic velocity prole.
Considering the limit of high frequencies, we replace the hyperbolic cosine in the denominator
on the right-hand side of (5.4.37) with half the exponential of its argument, resolve the hyperbolic
cosine in the numerator into its two exponential constituents, and derive the velocity prole
ux (y, t) Real e(1i) y + e(1i) (hy) 1 exp(it) . (5.4.38)
340 Introduction to Theoretical and Computational Fluid Dynamics
Comparing (5.4.38) with (5.4.16), we nd that the ow consists of an irrotational core executing
rigid-body motion with velocity ux = cos(t), and two Stokes boundary layers, one attached
to each wall. Inspecting the precise form of the velocity prole at high frequencies, we nd that the
amplitude of the velocity may signicantly exceed that of the plug ow in the central core.
The velocity vanishes at the initial instant, t = 0. At long times, we recover the HagenPoiseuille
parabolic prole.
f 2f
=1+ 2. (5.4.41)
The boundary conditions require that f = 0 at = 0 and 1 at any time, . The initial condition
requires that f = 0 at = 0 for 0 1.
for i = 2, . . . , N , where fin denotes the value of f at the ith grid point at time = n (Chapter
12 and Section B.4, Appendix B). Rearranging, we obtain the linear equation
fi1
n+1
+ (1 + 2)fin+1 fi+1
n+1
= fin + (5.4.43)
for i = 2, . . . , N , where / 2 . The no-slip boundary condition at the lower and upper
walls requires that f1 = 0 and fN +1 = 0. Collecting the nodal values fin , into a vector, f n , for
i = 2, . . . , N , we derive a system of linear algebraic equations,
A f n+1 = f n + e, (5.4.44)
where A is a tridiagonal matrix originating from (5.4.43) and all entries of the vector e are equal
to unity. The algorithm involves solving the system (5.4.44) at successive time instants using the
Thomas algorithm discussed in Section B.1.4, Appendix B.
Problem
5.4.1 Flow due to the application of a constant shear stress on a planar surface
Show that the velocity eld due to the sudden application of a constant shear stress along the
planar boundary of a semi-innite uid residing in the upper half-plane, y 0, is
2 1
ux (y, t) = t exp( 2 ) erfc , (5.4.45)
4 2
where = y/(t)1/2 is a similarity variable. Discuss the asymptotic behavior at long times.
Computer Problems
Rectilinear ow
In the case of rectilinear unidirectional ow along the tube axis, the axial velocity component satises
the linear equation
ux p 1 ux
= + + gx . (5.5.1)
t x
Expanding the derivative on the right-hand side, we obtain
ux p 2 ux 1 ux
= + 2
+ + gx . (5.5.2)
t x
If the ow is due to the cylinder translation along its length, we obtain the counterpart of Couette
ow (Problem 5.5.1).
Swirling ow
In the case of swirling ow with circular streamlines, the pressure assumes the hydrostatic distribu-
tion and the azimuthal velocity component satises the linear equation
u 1 (u )
= (5.5.3)
t
or
u 2 u 1 u u
= + 2 . (5.5.4)
t 2
The last term inside the parentheses on the right-hand side distinguishes this equation from its
counterpart for rectilinear ow.
where J0 is the zeroth-order Bessel function, = /, a = a/, a is the tube radius, = (2/)1/2
is the Stokes boundary-layer thickness, and an asterisk denotes the complex conjugate (e.g., [178],
p. 226; [318]). The zeroth-order Bessel function of complex argument can be evaluated using the
denition
where is a real positive number. The Kelvin functions of zeroth order, ber0 () and ker0 (), can
be computed by polynomial approximations (e.g., [2], pp. 379, 384).
The structure of the ow can be regarded as a function of the dimensionless Womersley number
dened with respect to the tube radius,
%
a
Wo a = 2 = 2 a. (5.5.8)
When the Womersley number is small, we obtain quasi-steady Poiseuille ow with a nearly parabolic
velocity prole. As the Womersley number tends to innity, we obtain a compound ow consisting
of a plug-ow core and an axisymmetric boundary layer wrapping around the tube wall. At high
frequencies, the amplitude of the velocity may exhibit an overshooting, sometimes mysteriously
called the annular eect, similar to that discussed in Section 5.4.5 for oscillatory plane Poiseuille
ow.
Applying the method of separation of variables with respect to radial distance and time for
the transient ow, we derive the velocity prole
+ gx 2
1 J0 (n /a) 2 t
ux (, t) = a 2 8 a2 exp , (5.5.10)
4 3 J1 (n )
n=1 n
n 2
a
where J0 and J1 are the Bessel functions of zeroth and rst order, and n are the real positive zeros
of J0 (e.g., [318]). The rst ve zeros of J0 are
1
0.8
0.6
0.4
0.2
y/a 0
0.2
0.4
0.6
0.8
1
Figure 5.5.1 Transient ow in a circular tube of radius a due to the sudden application of a constant
pressure gradient or sudden tilting. Velocity proles are shown at dimensionless times t/a2 =
0.001, 0.005, 0.010, 0.020, 0.030, . . . . The dashed parabolic line corresponds to the Poiseuille ow
with centerline velocity U established at long times.
accurate to the fth decimal place (e.g., [2]). Velocity proles at a sequence of dimensionless times
t/a2 are shown in Figure 5.5.1. Note the presence of boundary layers near the wall and the
occurrence of plug ow at the core at short times. The steady parabolic velocity prole is established
when t a2 /.
Finite-dierence methods
To compute a numerical solution, it is convenient to introduce the dimensionless variables
ux t
f= , = , = , (5.5.12)
( + gx )a2 a2 a
f 2f 1 f
=1+ 2 + . (5.5.13)
The initial condition requires that f = 0 at = 0 for 0 1 and the no-slip boundary condition
requires that f = 0 at = 1 at any time.
for i = 2, . . . , N , where fin denotes the value of f at the ith grid point at dimensionless time = n
(Chapter 12 and Section B.5, Appendix B). Rearranging, we obtain a linear equation,
(i ) fi1
n+1
+ (1 + 2) fin+1 (i + ) fi+1
n+1
= fin + (5.5.15)
One more equation is needed to formulate a system of N equations for the N unknowns, fin
for i = 1, . . . , N . This last equation arises by applying the governing dierential equation (5.5.13)
at the tube centerline. Since the third term on the right-hand side becomes indeterminate as 0,
we use lHopitals rule and nd that, at = 0,
f 2f
=1+2 2. (5.5.16)
Applying the nite-dierence approximation provides us with the desired dierence equation
Now collecting the nodal values fin into a vector, f n , for i = 1, . . . , N , and appending to (5.5.17)
equations (5.4.43), we obtain a linear system of algebraic equations,
A f n+1 = f n + e, (5.5.18)
where A is a tridiagonal matrix and the vector e is lled with ones. The algorithm involves solving
system (5.5.18) at successive time instants using the Thomas algorithm discussed in Section B.1.4,
Appendix B.
Q 2
1 J0 (n /a)
2 t
ux = 2 1 + 2 exp( ) , (5.5.19)
a2 a2 2
n=1 n
J0 (n ) n 2
a
and
Q t
p 1
= 8 4 1 + exp n2 2 , (5.5.20)
x a 2 n=1 a
346 Introduction to Theoretical and Computational Fluid Dynamics
where n , for n = 1, 2, . . ., are the positive zeros of the second-order Bessel function, J2 . The rst
ve zeros of J2 are
accurate to the fth decimal place (e.g., [2]). The singular behavior of the pressure gradient at t = 0
is evident by the divergence of the sum in (5.5.20).
where = /a and n are the positive zeros of the Bessel function of the rst kind, J1 . The rst
ve zeros of J1 are
accurate to the fth decimal place (e.g., [2]). At short times, the circumferential ow near the
cylinder surface resembles that due to the impulsive translation of a at plate.
a2
u = f (, ), (5.5.24)
where = ( a)/(t)1/2 and = (t)1/2 /a, are dimensionless variables, and f is a dimensionless
function ([371], p. 199). Considering (5.5.3), we compute the derivatives
u a2 (u ) a
= (f + f ), = f , (5.5.25)
t 2t
5.5 Unsteady ows inside tubes and outside cylinders 347
and
1 (u )
a 1
= f + f , (5.5.26)
where a subscript denotes a derivative with respect to and a subscript denotes a derivative
with respect to . Substituting these expressions into (5.5.3) and simplifying, we obtain
1 a 1
( f + f ) = f + f . (5.5.27)
2
Problems
(a) (b)
y y
x
x
Figure 5.6.1 Illustration of stagnation-point ow (a) in the interior of a uid and (b) over a solid
wall. In both cases, the origin of the y axis is set at the stagnation point of the outer ow. The
angle denes the slope of the straight dividing streamline of the outer inviscid ow. The angle
denes the actual slope of the generally curved dividing streamline of the viscous ow.
Computer Problems
Reviews of exact solutions for stagnation-point ows obtained by solving ordinary dierential
equations can be found in References [42, 421]. In this section, we discuss several fundamental
congurations.
Outer ow
We begin constructing the solution by introducing the outer ow prevailing far from the stagnation
point, denoted by the superscript . The far-ow stream function, velocity, and vorticity in the
upper half-space, y > 0, are given by
1 2
= (xy sin + y cos ),
2
u
x = (x sin + y cos ), u
y = y sin , z = cos , (5.6.1)
where is a constant rate of elongation and is a free parameter. Setting the stream function to
zero, we nd that is related to the angle subtended between the straight dividing streamline
and the x axis by
as shown in Figure 5.6.1(a). Equations (5.6.1) describe oblique stagnation-point ow with constant
vorticity against a planar surface in the upper half space. The values = 0 and correspond,
respectively, to unidirectional simple shear ow toward the positive or negative direction of the x
axis with shear rate . The value = /2 corresponds to irrotational orthogonal stagnation-point
ow in the upper half-space with rate of extension .
The far-ow stream function, velocity components, and vorticity in the lower half-space, y < 0,
are given by
= xy sin , u
x = x sin , u
y = y sin , z = 0, (5.6.3)
These equations describe irrotational orthogonal stagnation-point ow with shear rate equal to sin
in the lower half space.
The outer ow in the upper and lower half-space constitutes an exact solution of the Navier
Stokes equation for a uid with uniform physical properties and zero or constant vorticity. The
velocity is continuous across the horizontal dividing streamline at y = 0. However, the derivatives
of the velocity undergo a discontinuity that renders the outer ow admissible only in the context of
ideal ow. When viscosity is present, a vortex layer develops around the dividing streamline along
the x axis rendering the derivatives of the velocity and therefore the stresses continuous throughout
the domain of ow.
350 Introduction to Theoretical and Computational Fluid Dynamics
Viscous ow
Motivated by the functional form of the stream function of the outer ow in the upper and lower
half-space, shown in (5.6.1) and (5.6.3), we express the stream function of the viscous ow in the
form
where F(y) and G(y) are two unknown functions satisfying the far-eld conditions
1 2
F() y sin , G(+) y sin , G() 0, (5.6.5)
2
and the symbol denotes the leading-order contributions. Straightforward dierentiation yields the
components of the velocity and the vorticity,
Substituting expressions (5.6.6) into the steady version of the two-dimensional vorticity transport
equation (3.13.1),
z z 2
z 2 z
ux + uy = 2
+ , (5.6.7)
x y x y 2
and factoring out the x variable, we derive two fourth-order ordinary dierential equations,
The rst equation involves the function F alone. A solution that satises the rst far-eld condition
(5.6.5) is the far ow itself,
Equation (5.6.6) then yields z = F . Substituting (5.6.9) into the second equation of (5.6.8), we
obtain an equation for g,
G + sin y G = 0. (5.6.10)
Integrating (5.6.10) twice subject to the aforementioned far-eld (5.6.5) and requiring that the
vorticity is continuous across the dividing streamline located at y = 0, we obtain an expression for
the negative of the vorticity,
&
cos + A erfc(y) for y > 0,
G = z = (5.6.11)
(A + cos ) erfc(y) for y < 0,
where = ( sin /)1/2 and A is a constant that must be found by specifying the rate of decay of
vorticity far from the x axis. Further integrations of (5.6.11) require two additional stipulations on
the behavior of the ow far from the stagnation point.
5.6 Stagnation-point and related ows 351
Working as in the case of the free stagnation-point ow, we express the stream function as
shown in (5.6.4), and nd that the functions F and G satisfy the dierential equations (5.6.8). To
satisfy the no-slip and no-penetration conditions on the wall, we require that
F(0) = 0, F (0) = 0, F (0) = 0. (5.6.12)
Setting without loss of generality the stream function to zero at the the plate located at y = 0, we
obtain
G(0) = 0. (5.6.13)
As y tends to innity, the function F must behave to leading order like F y sin , and the
function G must behave like G 12 y 2 cos .
To solve the third-order equation (5.6.16), we denote f = x1 , introduce the auxiliary functions
f = x2 and f = x3 , and resolve (5.6.16) in a system of three rst-order ordinary dierential
equations,
x1 x2
d
x2 = x3 , (5.6.17)
d
x3 x22 x1 x3 1
352 Introduction to Theoretical and Computational Fluid Dynamics
with boundary conditions x1 (0) = 0, x2 (0) = 0, and x2 () = 1. The solution can be computed
using a shooting method for boundary-value problems according to the following steps:
The new value of x3 (0) in Step 3 can be found using a method for solving the nonlinear algebraic
equation Q[x3 (0)] x2 (max ) 1.0 = 0, as discussed in Section B.3, Appendix B [317].
In practice, it is expedient to carry out the iterations using Newtons method based on varia-
tional equations. First, we dene the derivatives
dx1 dx2 dx3
x4 , x5 , x6 , (5.6.18)
dx3 (0) dx3 (0) dx3 (0)
expressing the sensitivity of the solution to the initial guess. Dierentiating equations (5.6.17) with
respect to x3 (0), we obtain
dx4 dx5 dx6
= x5 , = x6 , = 2 x2 x5 x4 x3 x1 x6 , (5.6.19)
d d d
subject to the initial conditions x4 ( = 0) = 0, x5 ( = 0) = 0, and x6 ( = 0) = 1 . The shooting
method involves the following steps:
f () = b + , (5.6.21)
G + FG F G = c, (5.6.22)
5.6 Stagnation-point and related ows 353
(a) (b)
1.5 3
2.5
1 2
1.5
0.5 1
0.5
0 0
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
Figure 5.6.2 Two-dimensional oblique stagnation-point ow against a at plate. (a) Graphs of the
functions f (solid line), f (dashed line), and f (dotted line). (b) Graphs of the functions g (solid
line), g (dashed line), and g (dotted line).
Letting y or tend to innity and using (5.6.21) and the far-eld condition G() 12 y 2 cos , we
obtain
where has been dened previously in (5.6.15). Substituting (5.6.25) and (5.6.24) into (5.6.23), we
obtain the dimensionless equation
where a prime denotes a derivative with respect to . The solution must satisfy the boundary
conditions g(0) = 0 and g (0) = 0, and a far-eld condition dictating that, as tends to innity, g
behaves like g 12 2 . Equation (5.6.16) suggests that g = bf is a particular solution of (5.6.26), and
354 Introduction to Theoretical and Computational Fluid Dynamics
this allows us to eliminate the forcing function on the right-hand side and obtain a homogeneous
equation.
The solution must satisfy the boundary conditions x7 (0) = 0 and x8 (0) = 0, and the far-eld
condition x9 () = 1. The solution can be found using the shooting method described previously
for the function F , where shooting is now done with respect to x9 (0). The functions x1 and x2
are known at discrete points from the numerical solution of (5.6.16). Because equation (5.6.26) is
linear, two shootings followed by linear interpolation are sucient. The numerical results, plotted
in Figure 5.6.2(b), show that x9 (0) 1.406544, to shown accuracy [108].
Vorticity
The vorticity can be computed in terms of the functions f and g as
sin
1/2
z = z x sin f () + [1 g () ] cos . (5.6.28)
The rst term on the right-hand side represents the constant vorticity of the incident ow. The
third term represents a vortex layer with uniform thickness attached to the wall. Far from the
origin, the second term on the right-hand side dominates and the ratio between the vorticity and
the x component of the incident ow is
sin
1/2
z
f (), (5.6.29)
ux
which reveals the presence of a vortex layer of constant thickness lining the wall. In physical terms,
diusion of vorticity from the wall is balanced by convection, thereby conning the vorticity gradient
near the wall.
The thickness of the vortex layer, , can be dened as the point where x3 () = 0.01. The
numerical solution shows that f (2.4) 0.01, and thus
1/2
= 2.4 . (5.6.30)
sin
As tends to zero, the thickness of the vortex layer understandably diverges to innity.
Pressure eld
Having computed the velocity eld, we substitute the result into the NavierStokes equation and
obtain a partial dierential equation for the pressure. Straightforward integration yields the pressure
5.6 Stagnation-point and related ows 355
V = x
eld
1 2 2 1
p = cx sin x2 sin [f () + f 2 ()] + g x + p0 , (5.6.31)
2 2
where the constant c was dened in (5.6.24) and p0 is a reference pressure. When = 0 or , the
pressure assumes the hydrostatic distribution, p = g x + p0 .
Stability
The behavior of perturbations introduced in the base ow derived in this section can be assessed by
performing a linear stability analysis, as discussed in Chapter 9. Conditions for stability subject to
suitable perturbations have been established [224].
Following our analysis in Section 5.6.2 for orthogonal stagnation-point ow against a at plate,
= 12 , we introduce the dimensionless variable = (/)1/2 y, and express the stream function in
the form (x, y) = ()1/2 x f (). The boundary conditions require that f (0) = 0 and f (0) = 1, and
the far-eld condition of vanishing x velocity requires that f () = 0. Making substitutions into the
vorticity transport equation, we nd that the dimensionless function f () satises the homogeneous
Hiemenz equation
f + f f f 2 = 0, (5.6.32)
where a prime denotes a derivative with respect to . The solution is given by
f () = 1 exp(), (5.6.33)
yielding Cranes [96] stream function
(x, y) = ()1/2 x (1 e ), (5.6.34)
356 Introduction to Theoretical and Computational Fluid Dynamics
(a) (b)
x
1.5
0.5
0
0 0.5 1 1.5 2 2.5 3
Figure 5.6.4 (a) Illustration of axisymmetric orthogonal stagnation-point ow. (b) Graphs of the
functions f (solid line), f (dashed line), and f (dotted line). The radial velocity prole is
represented by the dashed line.
which provides us with an exact solution of the NavierStokes equation with corresponding pressure
eld
1
p = f () + f 2 () + g x + p0 , (5.6.35)
2
where p0 is a reference pressure. The y velocity component tends to a constant value, v = ()1/2 ,
far from the stretching sheet. Thus, the stretching of the sheet causes a uniform ow against the
sheet to satisfy the mass balance.
The eect of wall suction or blowing at a constant velocity can be incorporated into the
similarity solution (e.g., [423]). In Section 8.3.2, Cranes solution will be generalized to situations
where the in-plane wall velocity exhibits an arbitrary power-law dependence on the x position,
subject to the boundary-layer approximation.
where f () is a dimensionless function. The axial and radial velocity components and the azimuthal
vorticity component are given by
ux = 2 ()1/2 f (), u = f (), = ()1/2 f (), (5.6.38)
where a prime denotes a derivative with respect to , The no-penetration and no-slip boundary
conditions require that f (0) = 0 and f (0) = 0, and the far-eld condition requires that f () = 1.
Substituting expressions (5.6.38) into the vorticity transport equation for axisymmetric ow,
and integrating once subject to the far-eld condition, we derive Homanns [188] ordinary dierential
equation
f + 2f f f 2 + 1 = 0, (5.6.39)
which diers from the Hiemenz equation (5.6.16) only by the value of one coecient. The solution
can be found using a shooting method similar to that used to solve (5.6.17), yielding f (0) = 1.3120,
as shown in Figure 5.6.4(b). Having computed the velocity eld, we substitute the result into the
NavierStokes equation and integrate in space to obtain the pressure eld
1
p = 2 2 2 f () + f 2 () + g x + p0 , (5.6.40)
2
where p0 is a reference pressure.
In Section 8.5.3, the formulation will be generalized to situations where the radial disk velocity
exhibits an arbitrary power-law dependence on the radial position, subject to the boundary-layer
approximation.
358 Introduction to Theoretical and Computational Fluid Dynamics
u
1 = 1 x1 , u
2 = 2 x2 , u
3 = (1 + 2 ) x3 , (5.6.43)
where 1 and 2 are two arbitrary shear rates with dimensions of inverse length. When 1 = 0
or 2 = 0, we obtain two-dimensional orthogonal stagnation-point ow in the x1 x3 or x2 x3 plane.
When 1 = 2 , we obtain axisymmetric orthogonal stagnation-point ow.
Our experience with two-dimensional and axisymmetric ows suggests expressing the solution
in the form
u1 = 1 x1 q (), u2 = 2 x2 w (), u3 = 1 q() + w() , (5.6.44)
The boundary value problem expressed by (5.6.47) can be solved using a shooting method
where guesses are made for q (0) and w (0). Numerical results show that q (0) = 1.233 and
w (0) = 0.570 when = 0.0; q (0) = 1.247 and w (0) = 0.805 when = 0.25; q (0) = 1.267 and
w (0) = 0.998 when = 0.50; q (0) = 1.288 and w (0) = 1.164 when = 0.75; q (0) = 1.312 and
w (0) = 1.312 when = 1.00.
5.6 Stagnation-point and related ows 359
Unsteady ow
The scaling of the velocity shown in (5.6.44) is useful for computing unsteady ow due to the start-up
of an outer stagnation-point ow. Substituting into the continuity equation and equation of motion
the functional forms
u1 = 1 x1 K1 (x3 , t), u2 = 2 x2 K2 (x3 , t), u3 = (1 )1/2 K3 (x3 , t),
1
p = (12 x21 + 22 x22 ) + (1 + 2 ) K4 (x3 , t) + g x + p0 , (5.6.48)
2
we obtain a system of four dierential equations for the functions Ki with respect to x3 and t, where
i = 14. The solution can be found by standard numerical methods (e.g., [371] p. 207). There is a
particular protocol of startup where the ow admits a similarity solution and the problem is reduced
to solving an ordinary dierential equation for a carefully crafted similarity variable [419].
Problems
where = (/)1/2 y is a scaled coordinate, i is the imaginary unit, and f , q are two unknown
complex functions (e.g., [353]; [428] p. 402). The no-slip and no-penetration conditions at the plate
require that f (0) = 0, f (0) = 0, q(0) = 0, and q (0) = 1. The far-eld condition requires that
f () or f () = 1, and q () = 0.
(a) Show that the function f satises equation (5.6.16), and is therefore identical to that for steady
stagnation-point ow, whereas the function q satises the linear equation
q + f q (f + i ) q = 0, (5.6.50)
where / is a dimensionless parameter. Demonstrate that the pressure eld is unaected by
the oscillation.
(b) Explain why the problem is equivalent to that of stagnation-point ow against a stationary at
plate, where the stagnation point oscillates about a mean position on the plate.
(c) Develop the mathematical formulation when the plate oscillates along the z axis ([428] p. 405).
Computer Problems
5.6.4 Stagnation-point ow
(a) Write a program that uses the shooting method to solve the system (5.6.17) and (5.6.27) for
two-dimensional stagnation-point ow and reproduce the graphs in Figure 5.6.2(a, b).
(b) Write a program that uses the shooting method to solve the system (5.6.39) for axisymmetric
orthogonal stagnation-point ow and reproduce the graphs in Figure 5.6.4(b).
(c) Write a computer program that uses the shooting method to solve equations (5.6.41) subject to
the stated boundary conditions. Compare the results with the Crane solution for the corresponding
two-dimensional ow.
(d) Write a computer program that uses the shooting method to solve equations (5.6.47) and prepare
graphs of the results for = 0.30, 0.60, and 0.90.
Von Karman [415] observed that the equation of motion can be reduced into a system of
ordinary dierential by introducing the transformations
where = x(/)1/2 is a scaled x coordinate and H, F , and G are three dimensionless functions.
The no-slip boundary condition at the disk surface requires that H(0) = 0, F (0) = 0, and G(0) = 1.
The continuity equation requires that H (0) = 0. The far-eld condition requires that F () = 0,
G() = 0, and H () = 0. Note that by demanding that H () = 0 or H() is a constant, we
allow for the onset of uniform axial ow toward the disk. Substituting expressions (5.7.1) into the
NavierStokes equation, we nd that the pressure must assume the functional form
p = Q() + g x + p0 , (5.7.2)
where Q is a dimensionless function and p0 is a constant. Note that the hydrodynamic pressure does
not exhibit the usual radial dependence due to the centripetal acceleration.
Combining the equation of motion with the continuity equation, we obtain a system of four
ordinary dierential equations for the functions H, F, G, and Q. Combining these equations further
5.7 Flow due to a rotating disk 361
(a) (b)
x 1
0.8
0.6
0.4
0.2
0
0 1 2 3 4 5
Figure 5.7.1 (a) Illustration of ow due to a rotating disk. (b) Graphs of the dimensionless functions
H (solid line), G (dashed line), F (dotted line), and Q (dot-dashed line).
to eliminate the functions Q and F , we derive two third-order coupled ordinary dierential equations
for the functions H and G,
1 2
H H H + H 2 G2 = 0, G G H + GH = 0, (5.7.3)
2
subject to the aforementioned boundary conditions. Once the solution has been found, the functions
F and Q arise from the expressions
1 1 2
F = H , Q= H H . (5.7.4)
2 2
An approximate solution of equations (5.7.3) is available [90]. To compute a numerical solution, we
recast these equations as a system of ve rst-order equations for H, H , H , G, and G . Denoting
H = x1 and G = x4 , we obtain
x1 x2
x2 x3
d
x3 = x1 x3 1 x22 + 2x24 , (5.7.5)
dt 2
x4 x5
x5 x1 x5 x4 x2
Since the initial values x3 (0) and x5 (0) are a priori unknown, we must solve a boundary-value
problem with two unknown end conditions. The solution can be found using a shooting method in
two variables according to the following steps:
362 Introduction to Theoretical and Computational Fluid Dynamics
Results computed using this method are plotted in Figure 5.7.1(b). The numerical solution shows
that H (0) = 1.0204 and G (0) = 0.6159, to shown accuracy.
The thickness of the boundary layer on the disk, , can be identied with the elevation of the
point where the function F () drops to a small value. The numerical solution shows that F () = 0.01
when
At the edge of the boundary layer, the axial velocity takes the value ux = 0.89 ()1/2 . The
associated ow feeds the boundary layer in order to sustain the radial motion of the uid away from
the centerline.
Torque
Neglecting end eects, we nd that the axial component of the torque exerted on a rotating disk of
radius a is given by
Tz = G (0) a4 (3 )1/2 . (5.7.8)
2
Stability
Observation reveals and stability analysis conrms that the steady ow over the disk is stable as
long as the radius of the disk, a, expressed by the Reynolds number, Re = a(/)1/2 , is lower than
about 285 [251]. Above this threshold, the ow develops non-axisymmetric waves yielding a spiral
vortex pattern.
Unsteady ow
The von Karman scaling of the velocity with respect to radial distance shown in (5.7.1) is also
useful in the case of unsteady ow due, for example, to the sudden rotation of the disk with constant
or variable angular velocity. Substituting into the continuity equation and equation of motion the
functional forms
we obtain a system of four partial dierential equations for the functions H, F, G, and Q with respect
to x and t whose solution can be found by numerical methods ([371] p. 204; [419]).
5.8 Flow inside a corner due to a point source 363
Problems
We begin the analysis by introducing plane polar coordinates with origin at the apex, (r, ), as
shown in Figure 5.8.1. Assuming that the ow is purely radial, we set u = 0 and use the continuity
equation to nd that the radial velocity component takes the functional form
A
ur = F (), (5.8.1)
r
where = /, F () is a dimensionless function, and A is a dimensional constant related to ow
rate, Q, by the equation
1
Q = A F () d. (5.8.2)
1
To satisfy the no-slip boundary condition at the two walls located at = , we require that
F (1) = 0. Since the velocity eld is purely radial, closed streamlines do not arise. However, it is
possible that the direction of the ow may reverse inside a certain portion of the channel. When
this occurs, the uid moves against the direction of the primary ow driven by the point source.
We are free to select the relative magnitudes of A and F . To remove this degree of freedom,
we specify that the maximum value of F is equal to unity, which is equivalent to requiring that
F = 0 when F = 1, where a prime denotes a derivative with respect to . Substituting (5.8.1) into
the two-dimensional vorticity transport equation, we obtain a third-order nonlinear equation for F ,
F + 42 F + 2 Re F F = 0, (5.8.3)
364 Introduction to Theoretical and Computational Fluid Dynamics
r x
Figure 5.8.1 Illustration of the JeeryHamel ow between two semi-innite planes intersecting at
angle 2, driven by a point source located at the apex.
F + 42 F + Re F 2 = , (5.8.4)
where is a constant. Substituting (5.8.1) into the radial component of the NavierStokes equa-
tion, using (5.8.4), and integrating once with respect to r, we obtain the hydrodynamic pressure
distribution,
2 1
p p g x = 2
A F () . (5.8.5)
r 4
Stokes ow
The third term on the left-hand side of (5.8.4) represents the eect of uid inertia. Neglecting this
term yields a linear equation whose solution is readily found by elementary analytical methods.
Solving for F and computing A in terms of Q from (5.8.2), we obtain
Q cos 2 cos 2
ur = . (5.8.6)
r sin 2 2 cos 2
A dierent way of deriving this solution based on the stream function is discussed in Section 6.2.5.
Inertial ow
To compute ow at nonzero Reynolds numbers, we multiply (5.8.4) by F and rearrange to obtain
1 2 1
(F ) + 22 (F 2 ) + Re (F 3 ) = F . (5.8.7)
2 3
One more integration subject to the condition F () = 0 when F = 1 yields the rst-order equation
2 1/2
F = (1 F )1/2 Re F (1 + F ) + 42 F + c , (5.8.8)
3
5.9 Flow due to a point force 365
where c is a new constant. Applying (5.8.8) at = 1 reveals that c is related to the magnitude of
the shear stress at the walls by
c = |F (1)|. (5.8.9)
When AF (1) < 0 or AF (1) < 0, we obtain a region of reversed ow adjacent to the walls. Given
the values of and Re, the problem is reduced to solving (5.8.8) and computing the value of the
constant c subject to the boundary conditions F (1) = 0. Once this is done, the results can be
substituted into (5.8.2) to yield the corresponding value of Q. Rosenhead [350] derived an exact
solution in terms of elliptic functions.
Numerical methods
To compute a numerical solution, we use a shooting method involving the following steps [190]:
Guess a value for c; integrate (5.8.8) by solving an initial value problem with the plus or minus
sign and initial condition F (1) = 0 from = 1 to = 1; examine whether F (1) = 0; if not,
repeat the computation with an improved estimate for c. The results show that the ow exhibits a
variety of features, especially at high Reynolds numbers, discussed in detail by a number of authors
[24, 136, 137, 428, 190]. Linear stability analysis shows that the ow becomes unstable above a
critical Reynolds number that decreases rapidly as the angle becomes wider [160].
Computer Problem
u u = p + 2 u + b (x x0 ), (5.9.1)
where b is a constant vector determining the magnitude and direction of the point force. The eect
of the distributed body force due to gravity has been absorbed into the hydrodynamic pressure,
366 Introduction to Theoretical and Computational Fluid Dynamics
p p g x. Physically, equation (5.9.1) describes the ow due to the steady gravitational settling
of a particle with innitesimal dimensions but nite weight in an ambient uid of innite expanse,
viewed in a frame of reference translating with the particle.
where F is a dimensionless function and = / is the kinematic viscosity of the uid. For future
convenience, cos has been used instead of the polar angle in the argument. The spherical polar
components of the velocity are
1 1
ur = = F , u = = F, (5.9.3)
r2 sin r r sin r r sin
and the azimuthal component of the vorticity is
= sin F , (5.9.4)
r
where a prime denotes a derivative with respect to cos . It is clear from these functional forms that
the streamlines are self-similar, that is, one streamline derives from another by a proper adjustment
of the radial scale.
Progress can be made by substituting (5.9.3) into the vorticity transport equation to derive a
fourth-order ordinary dierential equation for F. However, it is more expedient to work directly with
the generalized equation of motion (5.9.1), thus obtaining a simultaneous solution for the pressure.
Motivated again by the absence of an intrinsic length scale, we express the hydrodynamic pressure
in the functional form
p(r, ) = p0 + Q(cos ), (5.9.5)
r2
where Q is a dimensionless function and p0 is an unspecied constant. Substituting (5.9.3) and (5.9.5)
into the radial and meridional components of the NavierStokes equation for steady axisymmetric
ow, given by
ur u ur u2 p ur 2 u 2
ur + = + 2 ur 2 2 2 2 u cot ,
r r r r r r r
(5.9.6)
u u u ur u 1 p 2 ur u
ur + + = + 2 u + 2 2 2 ,
r r r r r r sin
where
1 2 f
1 f
1 2f
2 f = r + sin + , (5.9.7)
r 2 r r r 2 sin r2 sin2 2
5.9 Flow due to a point force 367
To evaluate the constants A and c, we note that both the positive and negative parts of the x
axis are streamlines and set F(1) = 0. Next, we require that the radial component of the velocity
and the vorticity are nite along the x axis and use (5.9.8) to nd that, as tends to 1, both
F (1) and the limit of (1 2 )1/2 F () must be bounded. Equation (5.9.10) then yields A = 0
and c = 0. Integrating (5.9.10), we nd
1 2
(1 2 ) F + 2F F = , (5.9.11)
2
where is a new integration constant. Applying (5.9.11) at the x axis corresponding to = 1 and
requiring a nonsingular behavior we obtain = 0. Integrating (5.9.11), we nally obtain the general
solution
sin2
F =2 (5.9.12)
d cos
where d is a new integration constant whose value must be adjusted so that the ow eld satises
(5.9.1) at the singular point, x0 [221, 385]. Substituting (5.9.12) into the rst equation of (5.9.8)
yields the function Q,
d cos 1
Q=4 . (5.9.13)
(d cos )2
To evaluate the constant d, we apply the macroscopic momentum balance (3.2.9) over a
spherical volume of uid of radius R centered at the point force. We note that the ow is steady
and use the properties of the delta function to nd
( u u) n dS = b, (5.9.14)
Sphere
where n is the normal vector pointing outward from the spherical surface. Using (5.9.12) and (5.9.13)
to evaluate the stress tensor and substituting the result in the x component of (5.9.14), we obtain
an implicit equation for d,
|b| 8 d d1
Re2 = 2
+ d2 ln + 2d, (5.9.15)
8 3 d 1 d+1
368 Introduction to Theoretical and Computational Fluid Dynamics
(a) (b)
5
4 4
3.5 3
2
3
1
2.5
0
y
2
Re
2
1
1.5
2
1 3
0.5 4
0 5
5 0 5
0 2 4 6 8 10 x
d
Figure 5.9.1 (a) The relationship between the eective Reynolds number, Re, dened in (5.9.15)
and the constant d. The dashed line represents the low-Re asymptotic given in (5.9.17), and the
dotted line represents the high-Re asymptotic given in (5.9.17). (b) Streamline pattern of the ow
due to a point source of momentum or point force for d = 1.1.
when Re is an eective Reynolds number of the ow. It is now evident that the structure of the
ow depends on the value of Re determining the intensity of the point force. In fact, this conclusion
could have been drawn on the basis of dimensional analysis at the outset. A graph of Re against
the constant d in its range of denition, d > 1, is shown with the solid line in Figure 5.9.1(a). The
streamline pattern for d = 1.1 is shown in Figure 5.9.1(b).
Low-Reynolds-number ow
To study the structure of ow for large values of d, we expand the right-hand side of (5.9.15) in
a Taylor series with respect to 1/d and nd that Re2 2/d, which conrms that Re is small,
as shown in Figure 5.9.1(a). Making appropriate approximations, we nd that the corresponding
stream function and modied pressure are given by
|b| |b| 1
= r sin2 , p = cos . (5.9.16)
8 4 r2
Evaluating the various terms of the NavierStokes equation shows that the nonlinear inertial terms
are smaller than the pressure and viscous terms by a factor of 1/d and the solution (5.9.15) satises
the equations of Stokes ow. The streamline pattern is symmetric about the midplane x = 0, as
illustrated in Figure 6.4.1(a).
High-Reynolds-number ow
To study the opposite limit as d tends to unity, we expand the right-hand side of (5.9.15) in a Taylor
5.9 Flow due to a point force 369
which conrms that Re is large. Making appropriate approximations, we nd that, as long as cos
is not too close to unity, that is, suciently far from the positive part of the x axis, the stream
function and hydrodynamic pressure are given by
1
= 2r (1 + cos ), p = , (5.9.18)
r 2 1 cos
independent of the magnitude of Re.
Problem
The study of uniform ow past a stationary particle or ow due to the translation of a particle
in an innite quiescent uid will lead us to examine the structure of the far eld where the Stokes-
ow approximation is no longer appropriate. In the far ow, the neglected inertial term u u
can be approximated with the linear form U u, where U is the uniform velocity of the incident
ow, yielding Oseens equation and corresponding Oseen ow. Computing solutions of the Oseen
equation is complicated by the quadratic coupling of U and u. Nevertheless, the preserved linearity
of the equation of motion with respect to u allows us to conduct some analytical studies and develop
ecient methods of numerical computation. In concluding this chapter, we discuss oscillatory and
more general time-dependent ow governed by the unsteady Stokes equation (3.10.6). Although
the new feature of time-dependent motion renders the analysis somewhat more involved, we are still
able to build a rm theoretical foundation and derive exact and approximate solutions using ecient
analytical and numerical methods.
+ g = 0, (6.1.1)
370
6.1 Equations and fundamental properties of Stokes ow 371
or
p + 2 u + g = 0, (6.1.2)
where u is the velocity, p is pressure, is the uid viscosity, is the uid density, and g is the
acceleration of gravity.
In terms of the hydrodynamic pressure excluding the eect of the body force, p p g x,
and associated stress tensor, = p I + 2E, the Stokes equation becomes
= 0, (6.1.3)
or
p + 2 u = 0, (6.1.4)
1
where E = 2
[u + (u)T ] is the rate-of-deformation tensor for an incompressible uid.
A ow that is governed by one of the four equivalent equations (6.1.1)(6.1.4) is called a Stokes
or creeping ow.
The total force exerted on an arbitrary uid parcel, including the hydrodynamic and the body force,
is given in (3.1.15), repeated below for convenience,
Ftot = ( + g) dV. (6.1.5)
P arcel
Referring to (6.1.1), we see that the hydrodynamic force balances the body force and the total force
exerted on the parcel is zero. The assumption of creeping ow guarantees that the rate of change of
momentum of a uid parcel is negligibly small.
The total torque with respect to a point, x0 , exerted on an arbitrary uid parcel was given in
(3.1.20) and (3.1.21), repeated below for convenience,
T tot
= x (n ) dS + x ( g) dV + c dV, (6.1.6)
P arcel P arcel P arcel
where x = x x0 and is an appropriate physical constant associated with the torque eld c.
Converting the surface integral to a volume integral and switching to index notation, we obtain
tot
lk
Ti = ijk jk + ijk xj + ijk xj gk + ci dV. (6.1.7)
P arcel xl
In the absence of an external torque eld, the stress tensor is symmetric and the total torque exerted
on any uid parcel is zero.
372 Introduction to Theoretical and Computational Fluid Dynamics
2 p = 0. (6.1.8)
The general properties of harmonic functions discussed in Section 2.1.5 ensure that, in the absence
of singular points, the pressure must attain extreme values at the boundaries of the ow or else
grow unbounded at innity. The hydrodynamic pressure, p, excluding hydrostatic variations, also
satises Laplaces equation, 2 p = 0.
4 u = 0. (6.1.9)
The mean-value theorems stated in (2.4.14) and (2.4.14) provide us with the identities
1 1
u(x) dS(x) = u(x0 ) + a2 2 u(x0 ) (6.1.10)
4a2 Sphere 6
and
1 1 2 2
4 3 u(x) dV (x) = u(x0 ) + a u(x0 ), (6.1.11)
3 a Sphere 10
relating the mean value of the velocity over the surface or volume of a sphere of radius a centered
at a point, x0 , to the value of the velocity and its Laplacian at the center of the sphere.
Irrotational ow
Equation (6.1.9) shows that any solenoidal velocity eld, u = , satises the Stokes equation
with a corresponding constant hydrodynamic pressure. However, since an irrotational ow alone
cannot be made to satisfy more than one scalar boundary condition, it is generally unable by itself
to represent an externally or internally bounded ow.
6.1 Equations and fundamental properties of Stokes ow 373
Any linear ow with constant vorticity satises the equations of Stokes ow with a corre-
sponding constant hydrodynamic pressure, including the simple shear ow u = [y, 0, 0], where is
a constant shear rate.
Mean-value theorems
The mean-value theorems for functions that satisfy Laplaces equation stated in (2.4.2) and (2.4.14)
provide us with the identities
1 1
(x0 ) = (x) dS(x) = 4 3 (x) dV (x), (6.1.13)
4a2 Sphere 3
a Sphere
stating that the mean value of the vorticity over the surface or volume of a sphere of radius a centered
at a point, x0 , is equal to the value of the vorticity at the center of the sphere.
We note that the outward normal vector over the surface of the sphere is n = a1 (x x0 ) and
use the divergence theorem to write
(x x0 ) u(x) dS(x) = a n(x) u(x) dS(x) = a (x) dV (x). (6.1.14)
Sphere Sphere Sphere
This equation relates the mean value of the angular moment of the velocity over the surface of a
sphere to the angular velocity of the uid at the center of the sphere.
6.1.4 Two-dimensional ow
The vorticity of a two-dimensional ow in the xy plane arises from the stream function, , as
z = 2 . Since the vorticity is a harmonic function, 2 z = 0, the stream function satises the
biharmonic equation,
4 = 0. (6.1.16)
The mean value of the stream function along a circle or over a circular disk are given by the mean-
value theorem stated in (2.5.7) and (2.5.8).
374 Introduction to Theoretical and Computational Fluid Dynamics
2 2 1 2 sin 1
2 1 2 cot
E2 2
+ 2
= 2
+ 2
= 2
+ 2 2
2 . (6.1.18)
x r r sin r r r
The steady vorticity transport equation for axisymmetric Stokes ow requires that the Stokes
stream function satises a fourth-order linear partial dierential equation,
E 2 (E 2 ) = E 4 = 0. (6.1.19)
Far from the axis of revolution, this equation reduces to the biharmonic equation corresponding to
two-dimensional ow in an azimuthal plane.
Swirling ow
To describe a swirling Stokes ow, such as that produced by the slow rotation of a prolate spheroid
around its axis of revolution, x, we introduce the swirl function, (x, ), dened by the equation
u = /, where is the distance from the axis of revolution. The vorticity points along the x axis,
indicated by the unit vector ex ,
1
= ex . (6.1.20)
Swirling motion does not generate a pressure eld in Stokes ow. Resorting to (6.1.9), we nd that
the swirl function satises a second-order partial dierential equation,
E 2 = 0, (6.1.21)
Using the concept of reversibility, we infer that the streamline pattern around an axisymmetric
body with fore-and-aft symmetry moving along its axis must also be axisymmetric and fore-and-aft
symmetric. The streamline pattern over a two-dimensional rectangular cavity must be symmetric
with respect to the midplane. A neutrally buoyant spherical particle convected in a parabolic ow
inside a cylindrical tube may not move toward the center of the tube or the wall, but must retain
the initial radial position.
where f = n is the hydrodynamic traction excluding the eect of the body force, and n is the unit
normal vector pointing into the control volume. Equation (6.1.22) states that the rate of working
of the hydrodynamic traction on the boundary is balanced by the rate of viscous dissipation inside
the control volume.
where F and T are the hydrodynamic force and torque with respect to x0 exerted on the body,
excluding contributions from the body force.
376 Introduction to Theoretical and Computational Fluid Dynamics
Equation (6.1.23) expresses a balance between the rate of supply of mechanical energy nec-
essary to sustain the motion of the body and the rate of viscous dissipation in the uid. Since the
right-hand side is nonpositive for any combination of V and , each projection V F and T
must be negative. Physically, these inequalities assert that the drag force exerted on a translating
body and the torque exerted on a rotating body always resist the motion of the body.
The rst integral on the right-hand side is nonnegative. We will show that the second integral on
the right-hand side is zero. Working toward that goal, we write
(ui ui )
(E E) : E dV = Eij dV, (6.1.25)
F low F low xj
where n is the unit vector normal to the boundaries, B, pointing into the uid. Since u and u
are identical over the boundary, the boundary integral on the right-hand side is zero. In fact, the
volume integral on the right-hand side is also zero. To show this, we use the continuity equation
and the Stokes equation to obtain
Eij 1 1 p
(ui ui ) dV = (ui ui ) 2 ui dV = (ui ui ) dV. (6.1.27)
F low x j 2 F low 2 F low xi
Pressurevorticity formulation
Consider a two-dimensional Stokes ow in the xy plane. Using the identity 2 u = along
with the Stokes equation, 2 u = p, we obtain
= p, (6.1.29)
where p is the hydrodynamic pressure excluding hydrostatic variations. Writing the two scalar
components of (6.1.29) in the x and y directions, we nd that the z vorticity component, denoted
by , and the hydrodynamic pressure, p, satisfy the CauchyRiemann equations,
1 p 1 p
= , = . (6.1.30)
x y y x
As a consequence, the complex function
p
f (z) = + i (6.1.31)
is an analytic function of the complex variable z = x + iy, where i is the imaginary unit. Assigning
to the complex function f (z) various forms provides us with dierent families of two-dimensional
Stokes ows.
where the subscripts R and I denote the real and imaginary parts. To derive an expression for the
stream function in terms of F , we note that, if H0 , H1 , H2 , and H3 are four harmonic functions and
c0 , c1 , c2 , and c3 are four arbitrary constants, then
satises the biharmonic equation, 4 = 0, and it is therefore an acceptable stream function. The
corresponding vorticity is
H1 H2 H3 H3
Equations (6.1.32) and (6.1.35) provide us with the vorticity, pressure, and stream function of
a two-dimensional Stokes ow in terms of a generating complex function, F , and a real function, H.
Having specied these functions, we obtain a family of ows with identical vorticity and pressure
elds parametrized by the constants c0 and c1 . For example, setting F equal to a constant produces
irrotational ows with vanishing vorticity and constant pressure.
In the particular case where c1 = 1/4, expression (6.1.35) takes the compact form
1 1 1
= c0 H (x x0 )FR (y y0 )FI = real (z z0 ) F (z) + G(z) , (6.1.36)
4 4 4
where G(z) is an arbitrary analytic function and an asterisk denotes the complex conjugate. Straight-
forward dierentiation yields the associated velocity eld,
i
ux + i uy = F + (z z0 ) F + G , (6.1.37)
4
where a prime denotes a derivative with respect to z.
An extensive discussion and further applications of the complex variable formulation of two-
dimensional Stokes ow is available elsewhere ([223], Chapter 7). The complex-variable formulation
is amenable to a boundary-integral representation that can be used to derive and numerically solve
boundary-integral equations [306].
The Stokes equation is satised for any dierentiable function . Using the continuity equation and
recalling that the pressure is a harmonic function, we nd that satises the biharmonic equation,
4 = 0. Expressing the three independent components of the stress tensor in terms of the pressure,
p, and stream function, , and using the Stokes equation to eliminate the pressure, we obtain
2 2 2 2 2 2
= 4 , = . (6.1.39)
x2 y 2 xy x2 y 2 xy
= + = + i (6.1.40)
z z x z y z 2 x y
and
2
1 2 2 2
= + 2i . (6.1.41)
z 2 z 4 x2 y2 xy
Similar expressions can be derived for . Combining these results, we nd that the complex function
= 2i satises the dierential equation
2
= 0. (6.1.42)
z 2 z
where 1 and 2 are two arbitrary analytic functions of z. Dierent selections for 1 and 2 produce
various types of two-dimensional Stokes ow. For example, setting
1 1
= ( ) 4 (6.1.44)
2 3
produces two-dimensional HagenPoiseuille ow in a two-dimensional channel conned between two
parallel walls located at y = a, where = z/a.
The formulation in terms of the Airy stress function and stream function is amenable to a
boundary-integral representation that can be used to derive integral equations [306].
6.1.11 Swirling ow
A swirling ow can be described by a single scalar function, (x, ), dened in cylindrical polar
coordinates, (x, , ), by the equations
1 1
x = , = (6.1.45)
2 2 x
380 Introduction to Theoretical and Computational Fluid Dynamics
([247], p. 325). All other stress components are identically zero. The Stokes equation is satised for
any function . For an incompressible Newtonian uid,
u u
x = , = , (6.1.46)
x
requiring that
x
= . (6.1.47)
x
Substituting expressions (6.1.45), we nd that satises the equation
2 1
3 (6.1.48)
+ = 0.
x2 3
The function cos 2/ 2 satises Laplaces equation, 2 = 0.
Problems
6.1.5 Swirling ow
(a) Conrm that the Stokes equation is satised for any choice of the function dened in (6.1.45).
(b) Show that is constant along a line of vanishing traction.
(c) Show that the derivative of normal to a line C rotating as a rigid body is zero. The azimuthal
velocity component is u = over C, where is the angular velocity of rotation.
(d) Conrm that the function cos 2/ 2 satises Laplaces equation, 2 = 0.
prole in the xy plane is described by the function y = yw (x) = L cos(kx), where k = 2/L is the
wave number. Show that the stream function is given by the asymptotic expansion
1 2
= y yLeky cos(kx) + , (6.1.49)
2
where the dots indicate quadratic and higher-order terms in . Compute the drag force exerted on
the wall over one period accurate to rst order in .
In this section, we derive a general solution and study two-dimensional Stokes ow in a wedge-
shaped domain bounded by two intersecting stationary, translating, or rotating planar walls repre-
senting rigid boundaries or free surfaces of vanishing shear stress. In Section 6.3, we will consider in
detail the structure of two-dimensional ow inside and around a corner with stationary walls driven
by an ambient ow.
To circumvent the computation of the pressure and thus reduce the number of unknowns, it is
convenient to describe the ow in terms of the stream function, . Since the boundary conditions
are applied at intersecting plane walls, it is appropriate to introduce plane polar coordinates with
origin at the vertex, (r, ), and separate the radial from the angular dependence by writing
q(r) = A r . (6.2.2)
The complex constant A is a measure of the intensity of the outer ow, and the complex exponent
determines the structure of the inner Stokes ow. It is understood that either the real or the
imaginary part of the complex stream function provides us with an acceptable solution.
382 Introduction to Theoretical and Computational Fluid Dynamics
d2 f
z = A r2 (2 f + ). (6.2.4)
d 2
Requiring that the vorticity is a harmonic function, 2 z = 0, provides us with a fourth-order
homogeneous ordinary dierential equation for f ,
where a prime denotes a derivative with respect to . Substituting an eigensolution of the form
f () = exp(), (6.2.6)
When has an imaginary component, the eigensolutions can be expressed in terms of trigonometric
functions. Solving for and substituting the result into (6.2.6), we obtain the general solution
B sin( ) + C sin[( 2) ] if = 0, 1, 2,
f () = B sin(2 ) + C + D if = 0, 2, (6.2.8)
B sin( ) + C sin( ) if = 1,
where B, C, and D are complex constants and and are real constants. A solution with = 0
was presented in (5.8.6) in the context of the JeeryHamel ow.
which can be substituted into (6.2.1) to yield the stream function in plane polar or Cartesian
coordinates,
(a) (b)
1.8 1.8
1.6 1.6
1.4 1.4
1.2 1.2
1 1
0.8 0.8
y
y
0.6 0.6
0.4 0.4
0.2 0.2
0 0
0.2 0.2
1 0.5 0 0.5 1 0 0.5 1 1.5
x x
Figure 6.2.1 Stokes ow near a stagnation point on a plane wall. The angle subtended between the
wall and the dividing streamline, measured from the positive x semiaxis, is (a) = /2 or (b) /4.
As tends to zero, we obtain unidirectional shear ow with parabolic velocity prole.
where F and G are two real constants determined by the intensity of the outer ow that is responsible
for the onset of the stagnation point. The Cartesian components of the velocity and pressure gradient
are
! "
2 3 cos
ux = Gy (3y cos 2x sin ), uy = Gy sin , p = 2G . (6.2.12)
sin
It is worth noting that the pressure gradient is constant and points into the same quadrant as
the velocity along the dividing streamline. Streamline patterns for = /2 and /4 are shown
in Figure 6.2.1. As the angle tends to zero, we obtain unidirectional shear ow with parabolic
velocity prole.
(a) (b)
1
0.9
0.8 0.8
0.7
0.6
0.6
0.5
0.4
y
y
0.4
0.3
0.2
0.2
0.1 0
0.1 0.2
0 0.2 0.4 0.6 0.8 1 0.2 0 0.2 0.4 0.6 0.8 1
x x
(c) (d)
1.8 1.8
1.6 1.6
1.4 1.4
1.2 1.2
1 1
0.8 0.8
y
0.6 0.6
0.4 0.4
0.2 0.2
0 0
0.2 0.2
1 0.5 0 0.5 1 1 0.5 0 0.5 1
x x
Figure 6.2.2 Streamline pattern of the ow due a scraper moving over a plane wall in a frame of
reference moving with the scraper. The inclined wall on the left side of each frame is stationary.
The scraping angle is (a) = /4, (b) /2, (c) 3/4, and (d) .
(a) (b)
1 1
0.8 0.8
0.6 0.6
0.4 0.4
y
y
0.2 0.2
0 0
0.2 0.2
0.4 0.2 0 0.2 0.4 0.6 0.4 0.2 0 0.2 0.4 0.6
x x
Figure 6.2.3 Streamline pattern of the ow due the translation of a section of a plane wall (a) close
to the moving section and (b) far from the moving section.
The case = describes ow driven by the motion of a semi-innite belt extending from
the origin to innity, as shown in Figure 6.2.2(d). A slight rearrangement of the general solution
(6.2.13) yields the simple form
() = V r sin = V y . (6.2.15)
In Section 6.2.3, this elementary ow will be used to construct a general solution for arbitrary
in-plane wall motion.
which describes the ow far from the moving section. Streamline patterns near and far from the
moving section are shown in Figure 6.2.3. The far ow in the rst quadrant resembles that due to
386 Introduction to Theoretical and Computational Fluid Dynamics
a point source with strength m = 2aV / located at the origin, in the presence of two intersecting
walls along the x and y axes where the no-slip and no-penetration conditions apply. The ow in the
second quadrant admits a similar interpretation.
Exploiting further the feasibility of superposition, we use the far-ow solution (6.2.17) to
derive the stream function of the ow above a plane wall subject to an arbitrary distribution of a
tangential in-plane velocity, V(x),
1 y2
(x, y) = V(x ) dx . (6.2.18)
(x x )2 + y2
Expression (6.2.16) arises by setting V equal to a constant V for |x| < a, and to zero otherwise.
When V is equal to a constant V over the entire x axis, we obtain = V y, describing uniform ow.
1
y2
ux (x, y) = V(x ) dx ,
y (x x )2 + y 2
(6.2.19)
1
y2 2
(x x )
uy (x, y) = 2 2
V(x ) dx = y 2 V(x ) dx .
x (x x ) + y [(x x )2 + y 2 ]2
In the limit y 0, the derivative inside the integral for the x velocity component tends to (xx ),
where is the one-dimensional delta function. When V(x) = x, we obtain orthogonal stagnation-
point ow with constant rate of elongation, .
6.2.5 Flow near a belt plunging into a pool or a plate withdrawn from a pool
Consider the ow due a belt plunging into a liquid pool or a plate withdrawn from a liquid pool
with velocity V at an angle , as shown in Figure 6.2.4. Over the belt or plate, located at = ,
we require the no-slip boundary condition / = V r. Motivated by this expression, we set the
exponent = 1. At the free surface of the pool, we require the no-penetration condition and the
condition of vanishing shear stress.
Using the general solution (6.2.8) and implementing the four boundary conditions f = 0,
f = V at = and f = 0, f = 0 at = 0, we derive the stream function [271]
2
= V r ( cos sin sin cos ). (6.2.20)
2 sin(2)
Streamline patterns for = /4 and /2 are shown in Figure 6.2.4. As in the case of the scraper
discussed in Section 6.2.2, the radial velocity component depends only on the polar angle, , and
is independent of radial distance, r. Because the shear stress diverges like 1/r near the corner, an
innite force is required to maintain the motion of the belt. In reality, the free surface will distort
near the contact point to prevent this unphysical singular behavior.
6.2 Stokes ow in a wedge-shaped domain 387
(a) (b)
0.5 0.5
0 0
0.5 0.5
y
y
1 1
1.5 1.5
2 2
0.5 0 0.5 1 1.5 2 0.5 0 0.5 1 1.5 2
x x
Figure 6.2.4 Flow due to (a) a belt plunging into a liquid pool at angle = /4 and (b) a plate
withdrawn from a liquid pool at angle = /2. In both cases, the shear stress is zero at the
horizontal free surface.
6.2.6 JeeryHamel ow
Consider Stokes ow between two intersecting planes located at = , due to a point source with
strength m located at the apex, as discussed in Section 5.8 for NavierStokes ow. Using the present
formulation, we derive the stream function
m sin 2 2 cos 2
() = , (6.2.21)
2 sin 2 2 cos 2
satisfying () () = m. This expression is consistent with the radial velocity given in (5.8.6).
Since the shear stress is zero at the midplane, = 0, the solution also describes the ow inside the
corner bounded by a stationary wall located at = and a free surface located at the midplane
= 0, where the ow is driven by a point source with strength 12 m located at the apex.
The denominator in (6.2.21) vanishes and the stream function diverges when = 257.45 due
to the neglected inertial forces [272]. This failure underscores the possible limitations of Stokes ow
(see also Problem 6.2.3).
Problems
6.2.2 Flow due to an imposed shear stress near a static contact line
Consider a ow driven by the application of a uniform shear stress, , along a at interface located
at = , in the presence of a stationary plane wall located at = 0. Derive the stream function
r2
= (1 cos 2 2 sin 2) sin 2)
8 cos 2 1 + sin 2
+(2 cos 2 sin 2)(cos 2 1) + 2(1 cos 2) . (6.2.22)
2 sin 2 2 cos 2
= r . (6.2.23)
2 sin 2 2 cos 2
Show that the denominator vanishes and the stream function diverges when = 257.45 . Discuss
the physical implication of this failure with reference to the neglected inertial forces [272].
(b) Generalize the results to arbitrary angular velocities of rotation.
6.3.1 Antisymmetric ow
Concentrating rst on the case of antisymmetric ow, illustrated in Figure 6.3.1, we require that the
velocity is zero at the walls and derive the boundary conditions = 0 and / = 0 at = .
Assuming that = 0, 1, 2, using (6.2.8), and stipulating that the velocity is antisymmetric with
respect to the midplane, / = 0 at = 0, we nd that
where A and B are two complex constants. When is complex, the complex conjugate of the right-
hand side of (6.3.1) is also added to extract the real part. Enforcing the boundary conditions at the
6.3 Stokes ow inside and around a corner 389
(a) (b)
1 0.5
0.8 0.4
0.6 0.3
0.4 0.2
0.2 0.1
0 0
y
y
0.2 0.1
0.4 0.2
0.6 0.3
0.8 0.4
1 0.5
1 0.5 0 0.5 1 0 0.2 0.4 0.6 0.8 1
x x
(c) (d)
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
y
0.1 0.1
0.2 0.2
0.3 0.3
0.4 0.4
0.5 0.5
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x x
Figure 6.3.1 Antisymmetric Stokes ow near two intersecting stationary plane walls for corner semi-
angle (a) = 135 , (b) 45 , (c) 30 , and (d) 10 .
(a) (b)
1 0.5
0.8 0.4
0.6 0.3
0.4 0.2
0.2 0.1
0 0
y
y
0.2 0.1
0.4 0.2
0.6 0.3
0.8 0.4
1 0.5
1 0.5 0 0.5 1 0 0.2 0.4 0.6 0.8 1
x x
Figure 6.3.2 Symmetric Stokes ow near two intersecting stationary planes for corner semi-angle
(a) = 135 and (b) 10 .
An obvious solution is = 1. However, since for this value the third rather than the rst expression
in (6.2.8) should have been chosen, this choice is disqualied. Combining (6.3.2) with (6.3.1) and
(6.2.1), we nd that the stream function is given by
Non-reversing ow
Dean & Montagnon [103] pointed out that real and positive solutions other than the trivial solution
exist in the range 73 < < 180 or 0.41 < a < ; the lower limit is a numerical approximation.
In practice, the nontrivial solutions can be computed by solving equation (6.3.4) using Newtons
method (Problem 6.3.1). We are mainly interested in the smallest solution plotted with the solid
line in the left column of Figure 6.3.3(a), describing the strongest possible velocity eld that is
expected to prevail in practice. When = 90 , we nd that = 2 corresponding to simple shear
ow. When = 180 , we nd that = 1.5 corresponding to ow around a at plate. The streamline
pattern for = 135 is shown in Figure 6.3.1(a).
Eddies
Moatt pointed out that, approximately when < 73 , equation (6.3.4) has complex solutions [271].
To study the structure of the ow, we write = R + iI and decompose (6.3.4) into its real and
imaginary parts indicated by the subscripts R and I, obtaining
sin[2(R 1)] cosh(2I ) = (1 R ) sin 2,
(6.3.6)
cos[2(R 1)] sinh(2I ) = I sin 2.
6.3 Stokes ow inside and around a corner 391
(a)
6 12
5 10
4 8
3 6
2 4
1 2
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
/ /
(b)
5 10
4 8
3 6
,
2 4
1 2
0 0
0 0.1 0.2 0.3 0.4 0 0.1 0.2 0.3 0.4
/ /
(c)
20 20
15 15
ln, ln
ln, ln
10 10
5 5
0 0
0 0.1 0.2 0.3 0.4 0 0.1 0.2 0.3 0.4
/ /
Figure 6.3.3 The left column describes antisymmetric ow and the right column describes symmetric
ow inside a corner conned between two intersecting planes. (a) Dependence of the real (solid
lines) and imaginary (broken lines) part of the exponent with the smallest real part on the
corner semi-angle, . The eigenvalue is real when 0.41 < < in antisymmetric ow or when
0.43 < < in symmetric ow. (b) Graphs of (solid lines) and (broken lines) dened in
(6.3.6) in the regime where eddies develop. (c) Graphs of the geometric ratio, (solid lines), and
decay ratio, (broken lines), dened in equations (6.3.14) and (6.3.15).
392 Introduction to Theoretical and Computational Fluid Dynamics
Since both the sine and cosine of must be negative, must lie in the range (2n1) < < (2n 12 ),
where n is an integer. The solution of the two nonlinear equations (6.3.8) can be computed using
Newtons method with suitable initial guesses for and . The solution branch with the smallest
real part, R , is shown in the left column of Figure 6.3.3(a, b). Combining (6.3.2) with (6.3.1) and
(6.2.1), we nd that the stream function for < 73 is given by
= C rR Real{ ei I ln r Q }, (6.3.9)
Thus,
Moatt observed that, as r tends to zero, the stream function changes sign an innite number
of times [271]. Physically, an innite sequence of self-similar eddies develop inside the corner. To
identify these eddies, we express the stream function in the form
The geometric ratio, , and amplication ratio, , are plotted on a linear-logarithmic scale with a
solid or broken line in the left column of Figure 6.3.3(c).
As the angle tends to zero, the geometrical ratio tends to unity, which means that the
eddies tend to become evenly spaced between two nearly parallel walls. As increases from zero,
takes values on the order of unity for small and moderate angles, revealing a slow spatial modulation
of the eddy size. As tends to the critical angle 73 , increases rapidly. Physically, the eddy
size decays quickly into the corner ow. The amplication ratio takes large values over an extended
range of angles, , and this reveals that the intensity of the eddy ow decays rapidly inside the corner.
The minimum value of achieved in the limit 0 is approximately 360. When = 45 , is
approximately 2000. Consequently, it takes one eddy for the magnitude of the velocity to decrease
by three orders of magnitude!
Channel ow
As tends to zero, the intersecting planes tend to become parallel yielding a two-dimensional channel
with parallel walls. The emerging ow can be identied with the ow far inside a deep cavity driven
by an overpassing uid. To derive an appropriate similarity solution, we may consider the limit as
tends to zero. However, it is more expedient to begin afresh working in Cartesian coordinates where
the walls are located at y = a.
Setting = Af (y) exp(k|x|), requiring that satises the biharmonic equation, and enforc-
ing the no-penetration condition = 0 at y = a, we obtain
where A and B are complex constants and k is the complex wave number,. Requiring the no-
slip condition /y = 0 at y = a provides us with an algebraic equation, sin 2ka = 2ka, which
admits only imaginary solutions. The solution with the smallest positive real part, 2ka = 4.21+2.26i,
describes a periodic eddy pattern with approximate wavelength L 4a/2.26 = 5.56a.
6.3.2 Symmetric ow
Next, we address the case of symmetric ow illustrated in Figure 6.3.2. Requiring that the velocity
vanishes at the walls located at = and the shear stress vanishes at the midplane = 0, we
derive the boundary conditions = 0 and / = 0 at = , and = 0 and 2 / 2 = 0 at
= 0. Working as in the case of the antisymmetric ow, we set
where A and B are two complex constants. Enforcing the boundary conditions, we derive an
eigenvalue problem expressed by
The obvious solutions = 1, 2 are disqualied for the reason that the second and third rather than
the rst expression in (6.2.8) should have been chosen.
394 Introduction to Theoretical and Computational Fluid Dynamics
Moatt pointed out that real and positive solutions exist in the range 78 < < 180 [271].
The solution branch with the smaller magnitude is plotted with the solid line in the right column
of Figure 6.3.3(a). When = 90 , we nd that = 2 corresponding to simple shear ow. When
= 180 , we nd that = 1.5 corresponding to ow along a semi-innite at plate. When
approximately < 78 = 0.43, equation (6.3.18) has complex solutions that can be computed
working in terms of and as in the case of antisymmetric ow. The solution branch with the
smallest real part R is shown in the right column of Figure 6.3.3(a, b). The corresponding geometric
ratio, , and amplication ratio, , are plotted on a linear-logarithmic scale in the right column of
Figure 6.3.3(c). Comparing these results with those shown in the left column of Figure 6.3.3 for
antisymmetric ow reveals that the size of the eddies decays less rapidly. However, the intensity of
the eddy motion decays at an appreciably higher rate.
Channel ow
As tends to zero, the intersecting planes tend to become parallel and the asymptotic solution
describes Stokes ow inside a liquid layer resting on a plane wall. To derive a similarity solution, it
is expedient to introduce Cartesian coordinates where the wall is located at y = 0 and the midplane
or free surface is located at y = a. Working as in the case of antisymmetric ow, we nd that the
wavelength of the periodic eddy pattern is L 2a/2.76 = 2.28 a [271].
We have considered solutions corresponding to with a positive real part, yielding a rapidly decaying
sequence of eddies. If 1 is a solution of (6.3.4) or (6.3.18) with positive real part, then 2 = 2 1
is also a solution with positive or negative real part corresponding to a ow whose velocity decays
like r1 +1 as r . In this light, the streamline patterns shown in Figures 6.3.1 and 6.3.2 can
be reinterpreted as those due to a disturbance near a corner, where the velocity decays far from the
corner.
Problem
which has only real solutions, precluding the occurrence of eddies. Conrm that, when > /2, the
smallest positive solution is = 2 /(2), corresponding to rotational ow. Conrm that, when
< /2, the smallest positive solution is = /(2), corresponding to irrotational ow.
(b) Repeat (a) for symmetric ow.
6.4 Nearly unidirectional ows 395
h
y
x
0
Figure 6.4.1 Illustration of ow past an obstacle in a HeleShaw cell conned between two parallel
plates separated by distance h.
Computer Problem
When the channel height, h, is small compared to the global dimensions of the channel and
size of the objects, pressure variations normal to the channel walls are negligibly small. The ow may
then be assumed to be locally unidirectional with a parabolic velocity prole described by (5.1.6).
If the walls are located at z = 0 and h, as shown in Figure 6.4.1, the nonzero velocity components
are
1 p 1 p
ux (x, y, z) = + gx z(h z), uy (x, y, z) = + gy z(h z) (6.4.1)
2 x 2 y
396 Introduction to Theoretical and Computational Fluid Dynamics
to leading-order approximation. Integrating these proles across the clearance of the channel, we
obtain the mean uid velocity
1 h h2
u(x, y) u dz = (p + g x), (6.4.2)
h 0 12
where the gradient operates with respect to x and y.
2 p = 0. (6.4.3)
These results suggest that ow in a HeleShaw cell with uniform gap past an obstacle is identical
to two-dimensional irrotational ow of an incompressible uid past a body with corresponding
geometry. Interestingly, highly viscous ow inside the HeleShaw cell provides us with a method of
visualizing irrotational ow normally occurring when the eect of viscosity is negligibly small.
Computing the ow in the HeleShaw is thus reduced to solving Laplaces equation (6.4.3).
Over an impermeable boundary, we require the no-penetration condition expressed as a Neumann
boundary condition for the pressure. Over the inlet and outlet, we typically impose a Dirichlet
boundary condition specifying the level of the pressure. Over an interface between two immisci-
ble uids, we require continuity of the normal velocity and specify that the pressure undergoes a
discontinuity determined by the interfacial curvature and surface tension.
The assumptions that lead us to the Laplace equation (6.4.3) cease to be valid in the vicinity
of a boundary where the ow becomes three-dimensional and the assumption of local unidirectional
ow is no longer appropriate. However, boundary eects in the HeleShaw cell are conned inside
thin uid layers whose thickness is of order of the plate separation, h, and are thus negligible
to a leading-order approximation. Neglecting inertial nonlinear eects due to the curvature of the
streamlines introduces an additional error that is small as long as the uid velocity is not exceedingly
high.
To develop a model for the ow inside the die, we neglect the the eect of gravity and assume
that the average velocity of the liquid through the channel is related to the pressure gradient by
equation (6.4.2) and the pressure satises Laplaces equation (6.4.3). The boundary conditions
6.4 Nearly unidirectional ows 397
(a) (b)
y Outlet
a h
n
l=b
l
x
0 w
Inlet
Figure 6.4.2 Flow in a coating die consisting of an inlet tube and a Hele-Shaw cell; (a) top view of
the die, and (b) cross-section of the inlet tube.
require that the pressure at the outlet is equal to the atmospheric pressure, p = pa at y = d, and
the uid cannot penetrate the side walls, p/y = 0 at x = 0 and w.
To complete the denition of the problem, we must provide a boundary condition for the
pressure at the inlet. A mass balance for the ow along the inlet tube requires that
dQ h3
= h n u = n p, (6.4.4)
dl 12
where Q is the ow rate and l is the arc length along the centerline of the inlet tube, 0 < l < b, and
n is the normal vector in the xy plane pointing into the HeleShaw cell. Evaluating the ow rate
along the inlet using Poiseuilles law, we set
a4 p
Q= . (6.4.5)
8 l
The inlet tube is allowed to be tapered, that is, the tube radius a can be a function of arc length along
the centerline, l. Combining (6.4.4) with (6.4.5), we derive the desired inlet boundary condition in
the form of a second-order partial dierential equation,
2 p 4 da p 2h3
+ + n p = 0, (6.4.6)
l2 a dl l 3a4
supplemented by the value of the pressure at the inlet, l = 0, and the condition that the ow rate
vanishes at the end of the inlet tube, p/l = 0 at l = b.
h0
hL
h(x)
L x
0
p(x)
pa pa
x
Figure 6.4.3 Lubrication ow between a horizontal and an inclined at surface (top), and the devel-
oping pressure eld (bottom). In this conguration, the pressure is the same on either end of the
lubrication zone.
so that the plate separation h is a function of x and y. Equations (6.4.1) and (6.4.2) are still valid
provided that the origin is set at the lower wall and the upper wall is located at z = h(x, y). Mass
conservation requires that (hu) = 0, yielding a Helmholtz equation,
Physically, this equation describes the steady distribution of temperature in a plate whose thermal
conductivity is proportional to h3 .
1 dp h3
Q= Vh . (6.4.8)
2 dx 12
Mass conservation requires that Q is a constant, independent of x. To compute the pressure dis-
tribution along the gap, we solve (6.4.8) for the unknown pressure gradient, dp/dx, and obtain the
6.4 Nearly unidirectional ows 399
Linear prole
To make the analysis more specic, we consider an inclined upper wall with linear prole, as shown
in Figure 6.4.3, and set h = h0 x, where is the constant slope and h0 the clearance of the
channel at the beginning of the lubrication zone. Equation (6.4.9) becomes
dp V Q
= 6 12 . (6.4.10)
dx (h0 x)2 (h0 x)3
To compute the ow rate, Q, we require that the integral of the right-hand side of (6.4.10) with
respect to x across the lubrication zone is pL p0 .
For illustration, we assume that the pressure is equal to the atmospheric pressure pa at both
ends of the lubrication zone, p0 = pL = pa . Setting the integral of the right-hand side of (6.4.10)
from x = 0 to L to zero, we obtain
h0 hL
Q=V , (6.4.11)
h0 + hL
where hL = h0 L is the clearance of the channel at the end of the lubrication zone. Substituting
this expression for Q into (6.4.10) and integrating with respect to x, and recover the pressure
distribution
x (L x)
p = pa + 6V . (6.4.12)
h0 + hL (h0 x)2
G()
Figure 6.4.4 Graph of the function G() expressing the hydrodynamic lift force exerted on the inclined
plane shown in Figure 6.4.3.
where x = x/L is the dimensionless position varying in the interval [0, 1]. The maximum pressure
occurs at the location
xmax = L (6.4.15)
2 1
and is given by
L
pmax = pa + 6V . (6.4.16)
( 1)(2 1) h20
As increases from unity to innity, the location of the maximum pressure is shifted from the left
end point, x = L, to the midpoint, x = 12 L.
The y component of the hydrodynamic force exerted on the sloped surface can be approximated
with the negative of the normal force exerted on the planar surface. The latter is found by integrating
the pressure over the planar surface, nding
L h
L2 0 h0 hL
Fy p dx = pa L + 6V 2 2 ln 2 . (6.4.17)
0 h0 hL h0 + hL
It is useful to recast this expression into the form
L2
Fy = pa L + 6V G(), (6.4.18)
h20
where
2
G() = 2 ln . (6.4.19)
1 2 1
6.4 Nearly unidirectional ows 401
x
Surface 1
(1)
V h
z
V (2)
y Surface 2
Figure 6.4.5 Lubrication ow between two solid but not necessarily rigid surfaces in close contact
moving with dierent velocities.
The second term on the right-hand side of (6.4.19) is the lubrication lifting or load force. A graph
of the function G() is shown in Figure 6.4.4.
When > 1, the plane wall moves toward the minimum gap and the lubrication force is
positive. Under these conditions, given V , the lift force will be able to balance the weight of
an overlying object whose lower surface is represented by the inclined plane, provided that is
suciently close to unity. Alternatively, given , the lift force will be able to balance the weight of
an overlying object provided that the velocity V is suciently large. When < 0, the wall moves
toward the maximum gap and the lubrication force pulls the object toward the moving plane, closing
the gap and choking the ow.
are the mean velocities of the uid along the gap, and the integration with respect to x is per-
formed over the lm thickness. Now adopting the approximations of lubrication ow, we use the
unidirectional channel ow solution (5.1.8) to write
1 (1) h2 p
uy (y, z, t) = (V + Vy(2) ) gy ),
2 y 12 y
1 h2 p
uz (y, z, t) = (Vz(1) + Vz(2) ) gz ). (6.4.23)
2 12 z
Finally, we combine (6.4.20), (6.4.21), and (6.4.23) to obtain the Reynolds lubrication equation
governing the distribution of the pressure inside the gap,
1 1
h3 p g = [h (V(1) + V(2) )] + (V(2) V(1) ) h Vx(2) + Vx(1) . (6.4.24)
12 2
The left-hand side of (6.4.24) expresses the eects of pressure- and gravity-driven ow. The right-
hand side represents the eects of Couette and squeezing ow.
where the function P(x) incorporates the eect of the ambient pressure and body force. The radial
velocity prole is found by integrating the lubrication equation of channel ow
p 2 u
= , (6.4.28)
x2
with boundary condition u = 0 at x = 0 and h, yielding
u = 3V x(h x). (6.4.29)
h3
The axial velocity, ux , arises by integrating the continuity equation using (6.4.29) and accounting
for the boundary conditions at the lower surface, yielding
x
6V V
ux = 3 (h ) d = 3 x2 3h 2x , (6.4.30)
h 0 h
The vertical component of the force exerted on the upper disk is found by integrating the
hydrodynamic normal stress over the surface of the upper disk from the origin up to the disk radius,
= a,
a a
3V 2
Fx = 2 p d = 2 3 + P(h) d, (6.4.31)
0 0 h
yielding
3 V a4
Fx = + a2 P(h). (6.4.32)
2 h3
This force is equal in magnitude and opposite in sign to the weight of a body whose lower surface
is represented by the upper disk, pushing against a at surface represented by the lower disk. The
magnitude of the force shown in (6.4.32) is equal to the force that one must apply in order to remove
a piece of adhesive tape from a solid surface, pulling it normal to the surface with velocity V . A thin
layer of an adhesive viscous liquid is assumed to separate the surface from the tape. A generalization
of (6.4.32) to non-Newtonian uids is available [231].
Further applications
Further applications of hydrodynamic lubrication are discussed in a comprehensive monograph by
Hamrock [164]. Lubrication theory nds important applications in describing the ow inside the
narrow gaps between two colliding particles or liquid drops where a squeezing ow is established.
y
p
a
g
x
Figure 6.4.6 Evolution of a thin liquid lm owing under the action of gravity down an inclined plane.
is small, h/x < 1, so that the ow is nearly unidirectional at every x station. Adapting the rst
equation in (5.1.17) for the velocity prole and equation (5.1.18) for the ow rate, we obtain
1 p h3 p
ux (x, y, t) = g sin y (2h y), Q(x, t) = g sin , (6.4.33)
2 x 3 x
where is the plane inclination angle. As part of the lubrication approximation, we assume that the
hydrodynamic pressure developing inside the lm due to the ow is independent of the transverse
position, y. Accounting for the pressure drop across the free surface due to surface tension and
adding the hydrostatic variation, we derive the pressure distribution
where is the curvature of the free surface in the xy plane and pa is the atmospheric pressure pre-
vailing above the lm. Substituting this equation in the expression for the ow rate and introducing
the approximation hxx , we obtain
h3
Q(x, t) = (g sin g cos hx + hxxx ), (6.4.35)
3
where the subscript x denotes a partial derivative with respect to x. Mass conservation requires that
the lm thickness and ow rate satisfy the equation
ht = Qx , (6.4.36)
where the subscript t denotes a partial derivative with respect to t. It will be noted that equation
(6.4.21) is a generalization of equation (6.4.36). Substituting (6.4.35) into (6.4.36), we derive a
fourth-order nonlinear partial dierential equation describing the evolution of the lm thickness,
1 3
ht + h (g sin g cos hx + hxxx ) = 0. (6.4.37)
3 x
To complete the denition of the problem, an initial lm prole and four periodicity or boundary
conditions must be provided.
6.4 Nearly unidirectional ows 405
If the lm surface travels without change in shape with phase velocity c, the lm thickness is
a function of the translating coordinate x ct alone. Substituting h(x, t) = () in the evolution
equation (6.4.37) and integrating once with respect to , we obtain
1 3 d
c= (g sin g cos + ) + , (6.4.38)
3
where d is a constant and is a shape function. This equation can be used to describe a propagating
gravity current down an inclined plane (e.g., [263]).
Dimensionless form
It is useful to introduce the dimensionless variables h = h/h, x = x/h, and t = g h sin t/(2), where
h is a reference length identied with a mean lm thickness, and = / is the kinematic viscosity
of the uid. Equation (6.4.37) becomes
2 3
ht + [ h (1 hx cot + hxxx ) ]x = 0, (6.4.39)
3
where = /(g h2 sin ) is an inverse Bond number determined by the physical properties of the
uid, the mean lm thickness, and the plane inclination angle.
In terms of the dimensionless variables h h/h, x x/h, and t = g ht/(2), we obtain the
dimensionless equation
2 3
ht + [ h (hx + hxxx ) ]x = 0, (6.4.41)
3
Numerical methods
To compute the evolution of a periodic lm from a specied initial state, we divide one period of
the lm with length L into N evenly spaced intervals separated by the grid points xi = (i 1)L/N ,
where i = 1, . . . , N + 1. Our objective is to compute the lm thickness at the grid points at a
sequence of time instants separated by a time interval, t.
406 Introduction to Theoretical and Computational Fluid Dynamics
(a) (b)
0
Regime A
y 2
V
4
h
6
x 8
Regime B
10
0 0.2 0.4 0.6 0.8 1 1.2
h
Figure 6.4.7 (a) Illustration of a at plate withdrawn with constant velocity, V , from a liquid pool.
(b) Distribution of the curvature along the lm surface far from the pool.
The value of h at the ith grid point at the time instant t = nt is denoted as hni , where
n = 0, 1, . . . , and h0i corresponds to the prescribed initial shape. Applying equation (6.4.40) at the
ith grid point and approximating the temporal derivative using a forward dierence and the spatial
derivatives using central dierences, we obtain
g t n
hn+1
i = hni Fi+1 Fin (6.4.42)
3 x
Consider a at plate withdrawn with constant velocity V at an angle from a liquid pool, as shown
in Figure 6.4.7(a). We are interested in computing the thickness of the liquid lm coated on the
plate far from the pool, denoted by h . The ow can be divided naturally into two regimes: (a)
Regime A prevails inside the lm far from the pool and (b) Regime B prevails near the neck of the
interface, as shown in Figure 6.4.7(a). The eect of gravity is insignicant in Regime A due to the
assumed small lm thickness. Hydrostatics dominates in Regime B due to the slow ow, and the
shape of the interface is determined by a balance between the gravitational and the capillary force.
6.4 Nearly unidirectional ows 407
Our plan is to separately consider the two regimes and then sensibly match the corresponding free
surface proles.
Film ow
Concentrating on Regime A, we introduce inclined Cartesian coordinates, (, ), as shown in Figure
6.4.7(a). Adapting the rst equation in (5.1.17) for the velocity prole and equation (5.1.18) for the
ow rate of lm ow, we obtain
1 p h3 p
u () = (2h ) + V, Q= + V h, (6.4.43)
2 3
where h() is the lm thickness and Q = V h is the constant ow rate. The pressure inside the
lm is determined by the normal interfacial force balance. Neglecting the eect of gravity and
approximating the curvature with the linear form h , we obtain
p() h + pa , (6.4.44)
where pa is the ambient pressure and the subscript denotes a derivative with respect to . Sub-
stituting this pressure distribution into the expression for the ow rate given in (6.4.43), setting
Q = V h , and rearranging, we obtain a third-order nonlinear dierential equation,
Ca h h
h = . (6.4.45)
3 h3
where Ca V / is a capillary number determining the relative magnitude of viscous and capillary
forces over the interface. The far-eld condition requires that h h as y . Linearizing the
right-hand side of (6.4.45) in the limit as and h h , we obtain
Ca h h
(h h ) . (6.4.46)
3 h3
The only acceptable solution of this third-order linear ordinary dierential equation that tends to
h as is
Ca
1/3
h h c exp , (6.4.47)
3 h
where c is a constant. Dierentiating this expression and eliminating the undesirable constant c, we
obtain
Ca
1/3 h h Ca
2/3 h h
h , h . (6.4.48)
3 h 3 h
Free-surface deformation
Next, we consider Regime B and introduce the solution for the semi-innite hydrostatic meniscus
attached to an inclined plate derived in Section 4.3.1 with zero apparent contact angle, = 0, and
408 Introduction to Theoretical and Computational Fluid Dynamics
arbitrary plate inclination angle, . Our analysis has shown that the curvature of the interface at
the apparent contact line is
hcl
cl = , (6.4.49)
2
where hcl is the capillary height at the apparent contact line, (/g)1/2 is the capillary length,
and is the density of the coated liquid. In the case of a vertical plate, = /2 and hcl / = 2.
Matching
To prevent a discontinuity in the pressure, we require that the curvature of the interface at the
beginning of Regime A is the same as that at the end of Regime B, yielding
hcl
h (6.4.50)
2
at the matching zone. The validity of this condition can be justied by formal matched asymptotic
expansions [430].
LandauLevich formula
It is convenient to introduce the dimensionless variables
h
h = , = Ca1/3 . (6.4.51)
h h
3
h = (h 1). (6.4.52)
h3
The primary and derived far-eld conditions (6.4.48) take the form
1 1
h 1, h (h 1), h (h 1). (6.4.53)
31/3 32/3
The matching condition (6.4.50) requires that
h 1
h hcl (6.4.54)
2 Ca2/3
Equation (6.4.52) can be integrated backward with respect to using a standard numerical
method, starting from an arbitrary and inconsequential initial position, , where h 1. The
necessary initial conditions for the numerical integration are provided by (6.4.53). A numerical
solution obtained for h = 1.01 is shown in Figure 6.4.7(b), where dimensionless variables indicated
by a hat are implied in the labels of the horizontal and vertical axes.
6.4 Nearly unidirectional ows 409
The numerical computations reveal that the second derivative tends to h 1.337578 as
. Substituting this result into (6.4.54) and rearranging, we obtain the LandauLevich
formula [222]
h
1.337578 Ca2/3 . (6.4.55)
hcl
The predictions of this equation are in good agreement with laboratory observations at small capillary
numbers.
In plane polar coordinates, (r, ), dened with respect to the xy axes such that x = r cos
and y = r sin , as shown in Figure 6.4.7(a), the stream function is
2 sin 2 2 cos 2
= V r ( cos sin + sin cos ) + V h . (6.4.56)
2( ) sin(2) sin 2 + 2 cos 2
Typical streamline patterns are shown in Figure 6.4.8 for two plate withdrawal angles.
Problems
Computer Problem
(a) (b)
0.5
0.5
0.5
y
y
1
1.5
1.5
2
0 0.5 1 1.5 2 0.5 0 0.5 1 1.5 2
x x
Figure 6.4.8 Streamline pattern in a liquid pool far from a plate withdrawn at inclination angle (a)
= /2 or (b) /4.
dimensionless amplitude. Run the program for = 0.50 and = 0, 0.5, and 1.0. Plot transient lm
proles at a sequence of times and discuss the behavior of the solution in each case.
The velocity and hydrodynamic pressure elds due to a point force are found by solving the
continuity equation, u = 0, and the singularly forced Stokes equation,
p + 2 u + b 3 (x x0 ) = 0, (6.5.1)
where x0 is the location of the point force, the constant b represents the direction and magnitude
of the point force, and 3 is the three-dimensional delta function. Equation (6.5.1) is the linearized
version of (5.9.1) in the absence of uid inertia. The solution of (6.5.1) must be found subject to a
boundary condition requiring that the velocity is zero over a stationary solid boundary of the ow
denoted by SB , that is, u = 0 when x is on SB .
6.5 Flow due to a point force 411
The vorticity, pressure, and stress elds due to the point force can be expressed in terms of a
vorticity tensor, , pressure vector, P, and stress tensor, T , as
1 1 1
i (x) = ij (x, x0 ) bj , p(x) = Pj (x, x0 ) bj , ik (x) = Tijk (x, x0 ) bj , (6.5.3)
8 8 8
where summation is implied over the repeated index, j. The stress tensor is given by
Gij (x, x0 ) Gkj (x, x0 )
Tijk (x, x0 ) = ik Pj (x, x0 ) + + . (6.5.4)
xk xi
We note that Tijk = Tkji , as required by the symmetry of the stress tensor, . If the domain of ow
is innite, , P, and T are required to decay far from the point force.
x0
n
Figure 6.5.1 Illustration of a point force and a closed surface, D, in the available domain of ow.
The point force may be located inside (as shown), outside, or precisely on D.
where n is the unit vector normal to D. This identity holds true independent of whether the singular
point x0 is located inside, precisely on, or outside D. Working in a similar fashion, we derive the
identities
Tijk (x, x0 ) ni (x) dS(x) = Tkji (x, x0 ) ni (x) dS(x) = jk (6.5.10)
D D
and
ilm Tmjk (x, x0 ) nk (x) dS(x) = ilj x0l , (6.5.11)
D
where n is unit normal vector pointing into the control volume, as shown in Figure 6.5.1, and x0l on
the right-hand side of (6.5.11) denotes the lth component of x0 . The coecient on the right-hand
sides is equal to 8, 4, or 0, depending on whether the point x0 is located inside, precisely on, or
outside a smooth surface, D. When the point x0 is on D, the surface integrals are improper but
convergent principal-value integrals.
which provides us with a relation between the velocity at the point x due to a point force placed
at the point x0 , and the velocity at the point x0 due to a point force placed at x. Among other
purposes, identity (6.5.12) serves as a useful check of accuracy in the computation of a Greens
function.
6.5 Flow due to a point force 413
3 (x x0 ) = , (6.5.13)
4 r
where r = |x x0 |. Recalling that the hydrodynamic pressure satises Laplaces equation and
balancing the dimensions of the pressure gradient with those of the delta function in (6.5.1), we set
1 1
p= b, (6.5.14)
4 r
and nd that (6.5.1) becomes
1 1
2 u = b I 2 , (6.5.15)
4 r
where I is the identity matrix. This form motivates expressing the velocity in terms of a generating
scalar function, H, as
1
u= ( I 2 H H ) b. (6.5.16)
The continuity equation is satised for any twice dierentiable function H. Substituting (6.5.16)
into (6.5.15) and discarding the arbitrary constant b, we obtain
1
( I 2 ) 2 H = 0, (6.5.17)
4r
which is satised by any solution of the Poisson equation
1
2 H = 0. (6.5.18)
4r
Taking the Laplacian of this equation, we nd that H is the Greens function of the biharmonic
equation,
4 H + 3 (x x0 ) = 0, (6.5.19)
Stokeslet
Substituting H = r/(8) into (6.5.16) and carrying out the dierentiations, we derive the ow due
to a point force in the standard form (6.5.2). The associated free-space Greens function, Gij , also
called the Stokeslet or the OseenBurgers tensor and denoted by Sij , is given by
ij xi xj
Sij (x) = + 3 , (6.5.20)
r r
414 Introduction to Theoretical and Computational Fluid Dynamics
(a) (b)
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
y
y
0.2 0.2
0.4 0.4
0.6 0.6
0.8 0.8
1 1
1 0.5 0 0.5 1 1 0.5 0 0.5 1
x x
Figure 6.5.2 Streamline pattern due to (a) a thee-dimensional and (b) a two-dimensional point force
pointing along the x axis.
where x = x x0 and r = |x|. The corresponding vorticity, pressure, and stress elds are given by
(6.5.6) with
xl xj xi xj xk
ij (x) = 2 ijl , Pj (x) = 2 , Tijk (x) = 6 . (6.5.21)
r3 r3 r5
The expression for Tijk is found by substituting the expressions for Sij and Pj into (6.5.4).
Stream function
The Stokes stream function due to a point force located at the origin and pointing along the x axis
is given by
bx
= r sin2 , (6.5.22)
8
where bx is the strength of the point force, r is the distance from the origin, and is the meridional
angle. The streamline pattern is shown in Figure 6.5.2(a). We observe that the point force causes
a forward motion of the uid throughout the domain of ow.
1 3 xi xj
fi (x) = ik (x) nk (x) = Tijk (x, x0 ) nk (x) bj = bj . (6.5.23)
8 4 a5
6.5 Flow due to a point force 415
F = b, (6.5.26)
independent of the radius of the sphere, a. Physically, in the absence of inertia, the net force exerted
on a volume of uid enclosed by two concentric spheres must vanish.
Fourier transforms
It is instructive to rederive the Stokeslet using the method of Fourier transforms. The three-
dimensional complex Fourier transform of a rapidly decaying scalar function f (x) dened in the
whole three-dimensional space is
1
f(k) = f (x) exp(i k x) dV (x), (6.5.27)
(2)3/2
where i is the imaginary unit and k is a real wave number vector. The inverse transform is
1
f (x) = f(k) exp(i k x) dV (k). (6.5.28)
(2)3/2
The Fourier transforms of the partial derivatives are
'
f
Taking the Fourier transform of the continuity equation (6.5.5) for the Stokeslet, we obtain
Taking the Fourier transform of (6.5.6) for a Stokeslet placed at the origin, x0 = 0, and using the
properties of the delta function, we nd that
4
i kl Pj (k) k k Slj (k) + lj = 0. (6.5.31)
2
416 Introduction to Theoretical and Computational Fluid Dynamics
Multiplying this equation by kl , summing over l, and using (6.5.30) to eliminate the Stokeslet, we
obtain
4
i k k Pj (k) + kj = 0. (6.5.32)
2
Rearranging, we obtain the Fourier transform of the pressure,
4 kj
Pj (k) = i . (6.5.33)
2 k
k
Substituting this expression into (6.5.31), we derive the Fourier transform of the Stokeslet,
4 1 ki kj
i kj 1 1
Pj (x) = 2 exp(i k x) dV (k) = 2 exp(i k x) dV (k) . (6.5.35)
kk xj kk
Expressing the wave number vector, k, in spherical polar coordinates, (r, , ), where the meridional
angle is measured from the orientation of k, we compute the integral
1
I exp(i k x) dV (k) = 2 exp(i kr cos ) sin d dk, (6.5.36)
kk 0 0
nding
sin(kr) 2 2
I = 4 dk = , (6.5.37)
0 kr r
where r = |x|. Substituting this expression into (6.5.35) reproduces the expression for the pressure
Greens function. The Stokeslet is derived working in a similar fashion (Problem 6.5.1(a)).
Blake [37] demonstrated that the Greens function can be constructed from a Stokeslet, S,
and a few image singularities, including a Stokeslet, a potential dipole, and a point-force dipole,
2
G(x, x0 ) = S(x x0 ) S(x xim
0 ) 2 h0 (x x0 ) + 2 h0 (x x0 ),
im im
(6.5.38)
where h0 = x0 w is the distance of the point force from the wall, and
0 = 2w x0 , y0 , z0
xim (6.5.39)
6.5 Flow due to a point force 417
x
x0
x
h0
z
y
x0im
Figure 6.5.3 Illustration of a point force located at a point, x0 , above an innite wall positioned at
x = w. The point xim0 is the image of the point force with respect to the wall.
is the image of the singular point, x0 , with respect to the wall. The tensors and contain,
respectively, potential dipoles and Stokeslet dipoles,
xi
ij x i xj
ij (x) = 3 = 3 + 3 5 = Dij (6.5.40)
xj r r r
and
S
j1 xi + i1 xj
i1
ij (x) = = x1 ij (x) , (6.5.41)
xj r3
where r = |x| and Dij is the potential dipole discussed in Section 6.6.1. The minus sign of applies
for j = 1, corresponding to the x direction, and the plus sign applies for j = 2, 3, corresponding to
the y and z directions [313].
Pressure eld
By analogy with (6.5.38), we express the pressure vector in the form
P(x, x0 ) = P S (x x0 ) P S (x xim
0 ) + 2h0 P (x x0 ),
im
(6.5.42)
where P S is the pressure vector associated with the Stokeslet given in (6.5.24), and
P
i (x) = 2Di1 . (6.5.43)
The minus sign applies for i = 1, corresponding to the x direction, and the plus sign applies for
i = 2 and 3, corresponding to the y and z directions. Because they are irrotational singularities, the
potential dipoles do not make a contribution to the pressure.
418 Introduction to Theoretical and Computational Fluid Dynamics
p + 2 u + b 2 (x x0 ) = 0, (6.5.44)
where b is the strength of the point force, 2 is the two-dimensional delta function and 2 is the
two-dimensional Laplacian operator in the xy plane. Working as in the case of three-dimensional
ow, we express the velocity, vorticity, pressure, and stress elds in the standard forms
1 1 1
ui (x) = Gij (x, x0 ) bj , i (x) = ij (x, x0 ) bj , p(x) = Pj (x, x0 ) bj ,
4 4 4
1
ik (x) = Tijk (x, x0 ) bj , (6.5.45)
4
involving the velocity, vorticity, pressure, and stress Greens functions, G, , P, and T , where
summation is implied over the repeated index, j.
4 H + 2 (x x0 ) = 0, (6.5.49)
given by
1 r
H= ln + 1 r2 . (6.5.50)
8 a
Substituting this expression into (6.5.16), we derive the velocity eld in the standard form (6.5.45),
where Gij is the free-space Greens function,
r xi xj
Sij = ij ln + 2 , (6.5.51)
a r
6.5 Flow due to a point force 419
also called the two-dimensional Stokeslet, x = x x0 and r = |x|. The associated vorticity, pressure,
and stress elds are given by (6.5.45) with
xl xj xi xj xk
ij (x) = 2 ijl , Pj (x) = 2 , Tijk (x) = 4 . (6.5.52)
r2 r2 r4
All these elds decay like 1/r with respect to the distance from the singular point, x0 . In contrast,
the corresponding eld for three-dimensional ow exhibit a faster, 1/r2 decay.
Stream function
The stream function associated with a point force is found by integrating the velocity,
1 r
= (1 ln ) (y ex x ey ) b, (6.5.53)
4 a
where ex and ey are unit vectors along the x and y axes. The streamline pattern due to a point
force pointing along the x axis is illustrated in Figure 6.5.2(b). Lengths have been normalized by a.
1 i |z z0 |
c0 = , H = y y0 , c1 = 0, F (z) = ln , (6.5.54)
4 2 a
and obtain
1 |z z0 |
= (y y0 )( ln + 1). (6.5.55)
4 a
The alternative choice
1 1 1 |z z0 |
c0 = , H = x x0 , c1 = , F (z) = ln , (6.5.56)
4 2 2 a
yields the stream function due to a point force of unit strength oriented along the y axis,
1 |z z0 |
= (x x0 )( ln + 1). (6.5.57)
4 a
These expressions are consistent with the compact form (6.5.53).
and third equations is replaced with 4. Equations (6.5.9)(6.5.11) are also valid for two-dimensional
Greens functions provided that the surface integral over a closed surface D is replaced by a line
integral along a smooth closed contour, C. The coecient on the right-hand sides of (6.5.10) and
(6.5.11) is equal to 4, 2, or 0, depending on whether the point x0 is located inside, on, or outside
a smooth contour C. When x0 is on C, the line integrals are improper principal-value integrals.
The Greens functions of Stokes ow can be classied in several categories according to the topology
of the domain of ow. First, we have the free-space Greens function for innite unbounded ow.
Second, we have the Greens functions for innite or semi-innite domains of ow bounded by solid
surfaces. Third, we have the Greens functions of interior ow in a completely conned domain.
Singly, doubly, and triply periodic Greens functions express the ow due to periodic arrays of point
forces. As the observation point, x, approaches the location of the point force, x0 , all Greens
functions exhibit a common singular behavior that is identical to that of the free-space Greens
function. The Greens functions for innite unbounded ow are required to decay at innity at a
rate that matches, or is lower than that of, the free-space Greens function.
Interior ow
The Greens functions of interior ow may carry a degree of freedom determining the ow rate or
pressure drop in an appropriate direction. For example, in the case of ow inside an innite tube,
the ow due to a periodic array of point forces oriented along the tube may generate either zero
axial ow rate or zero pressure drop. If a Greens function for one condition is available, the Greens
function for the other condition can be produced readily by adding an appropriate pressure-driven
ow, such as a Poiseuille ow.
A detailed discussion and explicit expressions of Greens functions for a variety of boundary geome-
tries can be found in References [306, 313]. Computer codes are available in the boundary-element
software library Bemlib [313] (see also Appendix C).
Problems
The linearity of the equations of Stokes ow allows us to generate solutions by superposing fun-
damental solutions that satisfy the governing equations but not necessarily the required boundary
conditions. The superposition is designed so that the ow expressed by discrete collections or con-
tinuous distributions of properly selected fundamental solutions satises the boundary conditions
in an exact or approximate sense. The fundamental solutions may become innite at a singular
point or else diverge at innity. In this section, we develop three classes of fundamental solutions
originating from the point source, the point force, and a quadratic ow. In Section 6.7, we will use
these fundamental solutions to study ow past particles and liquid drops.
In Section 6.1, we saw that any irrotational velocity eld satises the equations of Stokes ow with an
associated constant pressure eld set to zero. One example is the ow due to a point source located
at the point x0 , given in (2.1.30). Dierentiating the point source with respect to the position of the
singular point, x0 , we obtain a sequence of derivative vectorial or tensorial singularities expressing
irrotational ow. The rst three members of this chain are the potential dipole, the potential
quadruple, and the potential sextuple. The velocity and stress elds due to a point source with
scalar strength m, a potential dipole with vectorial strength d, and a potential quadruple with
tensorial strength q are shown in Table 6.6.1. The associated pressure elds are uniformly zero, and
the vorticity elds vanish throughout the domain of ow.
A second family of fundamental solutions originates from the point force. Dierentiating the
Stokeslet with respect to the singular point, x0 , we obtain derivative singularities representing mul-
tipoles of the point force. The velocity, pressure, and stress elds due to a point force with vectorial
strength b and a point-force dipole with tensorial strength p located at the point x0 in free space
are shown in Table 6.6.2. The point-force dipole in free space is also called the Stokeslet dipole.
Couplet or rotlet
The coecient of the Stokeslet dipole, p, can be resolved into a symmetric component, s = 12 (p+pT ),
and an antisymmetric component, r = 12 (p pT ), where the superscript T indicates the matrix
422 Introduction to Theoretical and Computational Fluid Dynamics
1
ui = m Mi ik = m TikM
4 4
xi ij xi xk
Mi = TikM = 2 6 5 = 2Dik
r3 r 3 r
1 D
ui = Dij dj ik = T dj
4 4 ijk
Mi ij xi xj
Dij = = 3 +3 5
x0j r r
M
D Tik jk xi + ki xj + ij xk xi xj xk
Tijk = =6 30 = 2Qijk
x0j r5 r7
1 Q
ui = Qijl qjl ik = T qjl
4 4 ijlk
Dij ij xl + jl xi + li xj xi xj xl
Qijl = = 3 + 15
x0l r5 r7
D
Q Tijk ij lk + il jk + ik jl jk xk + ik xj + jk xi
Tijlk = = 6 + 30 xl
x0l r5 r7
il xj xk + jl xi xk + kl xi xj xi xj xl xk
+30 210
r7 r9
Table 6.6.1 Velocity and stress elds at a point, x, due to irrotational singularities of three-dimensional
Stokes ow located at the point x0 in free space; r = |x| and x = x x0 . The associated
hydrodynamic pressure elds are constant.
1 D + D
ui = [Sijl ] sjl + [Sijl ] rjl , (6.6.1)
8
where the square brackets [ ]+ denote the symmetric part and the square brackets [ ] denote
the antisymmetric part with respect to the indices j and l. Cursory inspection yields the specic
6.6 Fundamental solutions of Stokes ow 423
1 1 S 1 S
ui = Sij bj p= P bj ik = T bj
8 8 j 8 ijk
ij xi xj xj xi xj xk
Sij = + 3 PjS = 2 = 2Mj S
Tijk = 6
r r r3 r5
1 1 SD 1 SD
ui = S D pjl p= P pjl ik = T pjl
8 ijl 8 jl 8 ijlk
D Sij 1 xi xj xl
Sijl = = 3 (ij xl il xj jl xi ) + 3
x0l r r5
SD
PjS xj
jl xj xl
Pjl = =2 = 2 3 + 6 5 = 2Djl
x0l x0l r3 r r
S
Tijk 6 xi xj xk xl
Tijlk
SD
= = 5 (il xj xk + jl xk xi + kl xi xj ) 30
x0l r r7
1 1 C
ui = Cim cm p=0 ik = T cm
8 8 imk
1 D xl C 1 SD ijm xk + kjm xi
Cim = jlm Sijl = iml 3 Timk = jlm Tijlk =3 xj
2 r 2 r5
Table 6.6.2 Velocity, pressure, and stress elds at the point x due to a three-dimensional point force
or point-force dipole located at the point x0 in free space; r = |x|, x = x x0 , M is the point
source, and D is the point-source dipole.
expressions
D + xi xi xj xl D 1
[Sijl ] = jl +3 , [Sijl ] = (ij xl il xj ), (6.6.2)
r3 r5 r3
D + D D
so that [Sijl ] + [Sijl ] = [Sijl ]. Exploiting the antisymmetry of r, we write
1
rjl = jlm cm , (6.6.3)
2
424 Introduction to Theoretical and Computational Fluid Dynamics
Expression (6.6.1) for the velocity due to the point-force dipole takes the form
1 D + 1 D 1 D +
ui = [Sijl ] sjl [Sijl ] jlm cm = [Sijl ] sjl + Cim cm , (6.6.5)
8 2 8
where
1 D 1 D xl
Cim jlm [Sijl ] = jlm Sijl = iml 3 (6.6.6)
2 2 r
is a new fundamental solution called the couplet or rotlet. The pressure eld associated with the
couplet is constant, and the stress eld is given in the third entry of Table 6.6.2.
Stresslet
Inspecting the symmetric component of the Stokeslet doublet given in (6.6.2), we recognize a point
source and a complementary fundamental solution called the stresslet, dened as
D + 1 D D xi xj xl
ijl [Sijl ] + jl Mi = (Sijl + Silj ) + jl Mi = 3 , (6.6.7)
2 r5
where Mi is the point source. The velocity, pressure, and stress eld due to a stresslet with strength
s is given in the rst entry of Table 6.6.3.
Point-force quadruple
Dierentiating the Stokeslet doublet, we obtain the Stokeslet quadruple shown in the second entry
of Table 6.6.3. Contracting the last two indices of the quadruple, we obtain the singularity
SQ
ij xi xj
Sijll = 2 3 + 3 5 . (6.6.8)
r r
The term inside the parentheses on the right-hand side is recognized as the potential dipole,
1 SQ 1
Dij = Sijll = 20 Sij , (6.6.9)
2 2
where S is the Stokeslet. This expression allows to restate all irrotational singularities, with the
exception of the point source, in terms of derivatives of the Laplacian of the Greens function.
1 1 1
ui = ijl sjl p= P sjl ik = T sjl
8 8 jl 8 ijlk
xi xj xl SD
PjS xj
jl xj xl
ijl = 3 Pjl = Pjl = =2 = 2 3 + 6 5 = 2Djl
r5 x0l x0l r 3 r r
ijl kjl 1 SD SD
Tijlk = ik Pjl + + = (Tijlk + Tiljk ) + jl TikM
xk xi 2
2 3 xi xj xk xl
= ik jl + 5 (ij xk xl + il xj xk + kj xi xl + kl xi xj ) 30
r3 r r7
1 1 SQ 1 SQ
ui = S Q tjlm p= P tjlm ik = T tjlm
8 ijlm 8 jlm 8 ijlmk
D
Q
Sijl 2 Sij 1
Sijlm = = = 3 (ij lm + il jm + jl im )
x0m x0l x0m r
3 xi xj xl xm
[ (ij xl + il xj + jl xi )xm + im xj xl + jm xi xl + lm xi xj ] + 15
r5 r7
SD
SQ
Pjl 2 PjS 6 xj xl xm
Pjlm = = = 5 (jl xm + jm xl + lm xj ) + 30
x0m x0l x0m r r7
SD
SQ
Tijlk 2 Tijk
S
Tijlmk = =
x0m x0l x0m
Table 6.6.3 Velocity, pressure, and stress elds at the point x due to a three-dimensional rotlet,
stresslet, or point-force quadruple located at the point x0 in free space; r = |x| and x = x x0 ,
D is the point-source dipole, and T M is the stress tensor due to a point source.
x1 exerted on a spherical surface that encloses a singularity is zero, except for the point force, the
point force dipole, and the couplet where
These expressions will nd applications in Section 6.7 where we discuss exact and approximate
singularity representations for several ows.
426 Introduction to Theoretical and Computational Fluid Dynamics
ui = Nij wj p = PjN wj N
ik = Tijk wj
ND Nij ND
PjN
Sijl = = ij 4xl + il xj + jl xi Pjl = = 10 jl
x0l x0l
N
ND
Tijk
Tijlk = = 3 ( 4ik jl + ij kl + kj il )
x0l
Table 6.6.4 Velocity, pressure, and stress eld at a point x due to three-dimensional interior rotational
singularities of Stokes ow located at the point x0 in free space; r = |x| and x = x x0 .
6.6.4 Interior ow
Families of fundamental solutions that have no singular points but diverge at innity can be con-
structed for interior ow [87]. To derive these singularities, we recall that the pressure is a harmonic
function and set
p = 10 (x x0 ) w, (6.6.11)
where w is a constant vector and the factor of ten has been introduced for future convenience.
Substituting this expression into the Stokes equation, we obtain a Poisson equation for the velocity,
2 u = 10 w, to be solved subject to the constraint u = 0 imposed by the continuity equation.
The solution is
u = N w, (6.6.12)
where N is a new singularity called the Stokeson, shown in the rst entry of Table 6.6.4.
The derivatives of the Stokeson with respect to the singular point, x0 , are legitimate funda-
mental solutions of interior Stokes ow. Dierentiating the Stokeson once, we obtain the Stokeson
dipole shown in the second entry of Table 6.6.4. The symmetric part of the Stokeson dipole is the
stresson, and the antisymmetric part is the roton. The stresson represents linear, purely straining
ow with vanishing vorticity and pressure, and the roton represents rigid-body rotation (Problem
6.6.3).
6.6 Fundamental solutions of Stokes ow 427
Point source
The velocity due to a point source located at a point x0 in a totally or partially innite domain of
ow arises from the pressure vector due to the point force at the same point in the same domain,
denoted by P(x, x0 ), as
1
u(x) = m M(x, x0 ), (6.6.13)
4
where m is the strength of the point source and
1
M(x, x0 ) = P(x0 , x). (6.6.14)
2
Note that the arguments x and x0 are switched on the left- and right-hand sides of (6.6.14). For
example, in the case of the free-space Greens function,
2 1
P(x0 , x) = (x0 x), M(x, x0 ) = (x x0 ), (6.6.15)
r3 r3
where r = |x x0 |.
where h = x w and
is the image of the point x with respect to the wall. In index notation,
Pi (x0 , x) = 2 MF
i (x0 x) 2 Mi (x0 x ) 4h Di1 (x0 x ),
S FS im FS im
(6.6.18)
where MF S is the point source in free space D F S is the point-source dipole in free space. Using
(6.6.14) and substituting the specic expressions for the singularities, we obtain
2
x0 2w + x rim + 3 (x0 xim )2
x
M(x, x0 ) = 3 + 3
1
y0 y + 2 x w 3 (x0 xim )(y0 y) , (6.6.19)
r rim 5
rim
z0 z 3 (x0 xim )(z0 z)
428 Introduction to Theoretical and Computational Fluid Dynamics
where
Next, we introduce the image of the point x0 with respect to the wall,
0 = (2w x0 , y0 , z0 ),
xim (6.6.21)
Combining the third with the fourth term on the right-hand side, we derive the nal expression
2 xim
1 x
M(x, x0 ) = MF S (x) + MF S (xim ) + 2 ( 3 + 3 5im ) yim + 2h0 D i1 . (6.6.25)
rim rim
zim
The rst two terms on the right-hand side represent point sources located, respectively, at x0 and
0 . The third term is a Stokeslet doublet, S ixx , and the fourth term is a potential dipole placed
D
xim
im
at x0 .
6.6.6 Two-dimensional ow
Fundamental solutions for two-dimensional ow are derived working as previously in this section
for three-dimensional ow. Examples of two-dimensional fundamental solutions are the Greens
functions discussed in Section 6.5, and the point source and its derivatives discussed in Section
2.1.6. Families of irrotational and rotational singularities are presented in Table 6.6.56.6.7. Some
of these singularities will be employed in Section 6.14 to compute streaming ow past a circular
cylinder.
Problems
1
ui = m Mi ik = m TikM
2 2
xi ij xi xk
Mi = TikM = 2 4 4 = 2Dik
r2 r2 r
1 D
ui = Dij dj ik = T dj
2 2 ijk
M
Mi ij xi xj D Tik jk xi + ki xj + ij xk xi xj xk
Dij = = 2 +2 4 Tijk = =4 4
16 = 2Qijk
x0j r r x0j r r6
1 Q
ui = Qijl qjl ik = T qjl
2 2 ijlk
Dij ij xl + jl xi + li xj xi xj xl
Qijl = = 2 +8
x0l r4 r6
D
Q
Tijk ij lk + il jk + ik jl jk xk + ik xj + jk xi
Tijlk = = 4 4
+ 16 xl
x0l r r6
il xj xk + jl xi xk + kl xi xj xi xj xl xk
+16 6
96
r r8
Table 6.6.5 Velocity and stress eld at the point x due to irrotational singularities of two-dimensional
Stokes ow located at the point x0 in free space; r = |x| and x = x x0 . The associated
hydrodynamic pressure elds are constant. Corresponding singularities of three-dimensional ow
are shown in Table 6.6.1.
1 1 S 1 S
ui = Sij bj p= P bj ik = T bj
4 4 j 4 ijk
r xi xj xj xi xj xk
Sij = ij ln + 2 PjS = 2 = 2Mj S
Tijk = 4
a r r2 r4
1 1 SD 1 SD
ui = S D pjl p= P pjl ik = T pjl
4 ijl 4 jl 4 ijlk
D Sij 1 xi xj xl
Sijl = = 2 (ij xl il xj jl xi ) + 2
x0l r r5
SD
PjS xj
jl xj xl
Pjl = =2 = 2 2 + 4 4 = 2Dij
x0l x0l r 2 r r
S
SD
Tijk 4 xi xj xk xl
Tijlk = = 4 (il xj xk + jl xk xi + kl xi xj ) 16
x0l r r6
1 1 C
ui = Cim cm p=0 ik = T cm
4 4 imk
xl 1 SD ijm xk + kjm xi
Cim = iml Timk
C
= jml Tijlk =2 xj
r2 2 r4
Table 6.6.6 Velocity, pressure, and stress eld at the point x due to two-dimensional rotational
singularities of Stokes ow located at the point x0 in free space; r = |x| and x = x x0 , a is a
chosen length, M is the two-dimensional point source, and D is the two-dimensional point-source
dipole.
1 1 1
ui = ijl sjl p= P sjl ik = T sjl
4 4 jl 4 ijlk
xi xj xl SD
PjS xj
jl xj xl
ijl = 2 Pjl = Pjl = =2 = 2 2 + 4 4 = 2Djl
r5 x0l x0l r 2 r r
ijl kjl 1 SD SD
Tijlk = ik Pjl + + = (Tijlk + Tilkk ) + jl TikM
xk xi 2
1 2 xi xj xk xl
= ik jl + 4 (ij xk xl + il xj xk + kj xi xl + kl xi xj ) 16
r2 r r6
1 1 SQ 1 SQ
ui = S Q tjlm p= P tjlm ik = T tjlm
4 ijlm 4 jlm 4 ijlmk
D
Q
Sijl 2 Sij 1
Sijlm = = = 2 (ij lm + il jm + jl im )
x0m x0l x0m r
2 xi xj xl xm
4 [ (ij xl + il xj + jl xi )xm + im xj xl + jm xi xl + lm xi xj ] + 8
r r6
SD
SQ
Pjl 2 PjS 2 xj xl xm
Pjlm = = = 4 (jl xm + jm xl + lm xj ) + 8
x0m x0l x0m r r6
SD
SQ
Tijlk 2 Tijk
S
Tijlmk = =
x0m x0l x0m
Table 6.6.7 Velocity, pressure, and stress eld at the point x due to two-dimensional rotational
singularities of Stokes ow located at the point x0 in free space; r = |x| and x = x x0 , D is the
point-source dipole, and T M is the stress tensor due to a point source.
6.7 Stokes ow past or due to the motion of particles and liquid drops
The fundamental solutions derived in Section 6.6 can be used as building blocks to produce the ow
past or due to the motion of solid particles and liquid drops. In Section 6.12, we will see that the
singularity representations allow us to develop generalized Faxen laws providing us with the force,
the torque, and higher moments of the traction exerted on a particle immersed in an arbitrary ow
in terms of the velocity and derivatives of the velocity of the ambient ow at selected locations.
Exact singularity representations will be developed in this section for ow due to the translation or
rotation of a sphere, arbitrary linear ow past a stationary sphere, ow due to the translation or
432 Introduction to Theoretical and Computational Fluid Dynamics
rotation of a prolate spheroid, ow due to the translation of a spherical drop, and linear ow past
a stationary sphere with the slip boundary condition over the surface of the sphere. Approximate
singularity representations can be constructed by numerical methods.
1 ij xi xj 1 ij xi xj
ui (x) = + 3 bj + 3 + 3 5 dj . (6.7.2)
8 r r 4 r r
Enforcing the no-slip and no-penetration boundary conditions, u = V at r = a, we obtain two
algebraic equations for the coecients of the singularities,
a2 b 2d = 8a3 V, a2 b + 6d = 0. (6.7.3)
The solution is
b = 6aV, d = a3 V. (6.7.4)
Substituting these expressions into (6.7.2) and grouping similar terms, we derive an explicit repre-
sentation of the velocity eld,
1 a a2 3 a a2 xi xj
ui (x) = 3 + 2 Vi + 1 2 Vj . (6.7.5)
4 r r 4 r r r2
Because of the point force singularity, the ow generated by the sphere decays like 1/r far from the
sphere.
Stream function
In spherical polar coordinates centered at the sphere with the x axis pointing in the direction of
translation indicated by the unit vector ex , the Stokes stream function is
1 a2
(r, ) = ar 3 2 sin2 V ex . (6.7.6)
4 r
Since the stream function diverges at innity, an unphysical innite amount of uid is convected
along the x due to the motion of the sphere. The streamline pattern in a stationary frame of reference
and in a frame of reference moving with the sphere are depicted in Figure 6.7.1.
6.7 Stokes ow past or due to the motion of particles and liquid drops 433
(a) (b)
3 3
2 2
1 1
0 0
y
y
1 1
2 2
3 3
3 2 1 0 1 2 3 3 2 1 0 1 2 3
x x
Figure 6.7.1 Streamline pattern of the ow due to a sphere translating along the x axis under con-
ditions of Stokes ow in (a) a stationary frame of reference and (b) a frame of reference moving
with the sphere.
Surface traction
The hydrodynamic traction exerted on the sphere is readily computed from the strength of the
singularities using Tables 6.6.1 and 6.6.2, and is found to be
1 S 1 D
fi = T bj nk + T dj nk . (6.7.7)
8 ijk 4 ijk
Substituting the expressions for the singularities, we nd that
3 xi xj xk 3 jk xi + ki xj + ij xk xi xj xk
fi = 5
bj nk + 5
5 dj nk . (6.7.8)
4 a 2 a a7
Setting nk = xk /a, observing that xk xk = a2 , and simplifying, we obtain
3 xi xj 3 2xi xj + ij a2 xi xj
fi = b j + 5 6 dj . (6.7.9)
4 a4 2 a6 a
Substituting expressions (6.7.4) and simplifying, we nd that
9 xi xj 3 3 xi xj + ij a2
fi = V j Vj , (6.7.10)
2 a3 2 a3
yielding the remarkably simple expression
3
fi = Vi . (6.7.11)
2 a
We have found that the traction is a constant vector oriented in the direction of translation. This
is a unique, remarkable yet fortuitous feature of the spherical geometry.
434 Introduction to Theoretical and Computational Fluid Dynamics
Stokes law
To compute the hydrodynamic force exerted on the sphere, we can either integrate the traction over
the surface of the sphere or use the properties of the singularities discussed in Section 6.6. The result
is the Stokes law,
F= f dS = 6a V. (6.7.12)
Sphere
The torque exerted on the sphere with respect to the center of the sphere is zero. The torque with
respect to any other point is nonzero.
As an application, we compute the velocity of a sphere that is settling or rising under the action of
gravity in an innite uid. Requiring that the hydrodynamic drag force exerted on the sphere given
in (6.7.12) cancels the buoyancy force and the weight of the sphere, we obtain
4 3
6a V + a (s ) g = 0, (6.7.13)
3
where s is the density of the sphere. Rearranging, we obtain
2 a2 (s )
V= g. (6.7.14)
9
A heavy sphere falls in the direction of gravity and a light sphere rises against the direction of
gravity.
1 1
ui (x) = u
i (x) + S D (x, x0 ) pjl + Qijl (x, x0 ) qjl . (6.7.15)
8 ijl 4
Substituting the specic expressions for the singularities and dening x = x x0 , we nd that
1 ij xl il xj jl xi xi xj xl
ui (x) = Aij xj + 3
+3 pjl
8 r r5
3 ij xl + jl xi + li xj xi xj xl
+ 5
+5 qjl . (6.7.16)
4 r r7
6.7 Stokes ow past or due to the motion of particles and liquid drops 435
Rearranging, we obtain
1 ij xl il xj 3 ij xl + li xj
ui (x) = Aij xj + pjl qjl
8 r3 4 r5
1 xi 3 xi 3 xi xj xl 15 xi xj xl
3
pll 5
qll + 5
pjl + qjl . (6.7.17)
8 r 4 r 8 r 4 r 7
Requiring that the velocity at the surface of the sphere is zero, we obtain
1 1 3 1 10
A+ 3
(p pT ) (q + qT ) = 0, pll = qll = 0, p= q, (6.7.18)
8 a 4 a5 a2
where pll is the trace of p and qll is the trace of q. Splitting the matrix A into a symmetric and an
antisymmetric part in the rst equation, we obtain
1 1 1 1 3 1
(A AT ) + (p pT ) = 0, (A + AT ) (q + qT ) = 0. (6.7.19)
2 8 a3 2 4 a5
Eliminating p in favor of q from the rst equation in (6.7.19) using the third equation in (6.7.18),
and simplifying the second equation, we nd that
2 5 2 5
q qT = a (A AT ), q + qT = a (A + AT ). (6.7.20)
5 3
Adding these equations and rearranging, we obtain
2 4
q= a5 (4A + AT ), p= a3 (4A + AT ). (6.7.21)
15 3
The symmetric and antisymmetric components of the coecient of the point-force dipole , satisfying
p = s + r, are
20
s= a3 E, r = 4a3 , (6.7.22)
3
where
1 1
E= (A + AT ), = (A AT ) (6.7.23)
2 2
are the rate-of-strain tensor and vorticity tensor of the linear ow.
Surface traction
The disturbance traction exerted on the sphere is given by
1 SD Q
fiD (x) = T (x, x0 ) pjl nk + T (x, x0 ) qjl nk = 3 Aij xj . (6.7.24)
8 ijlk 4 ijlk a
It is remarkable that the disturbance traction is proportional to the local velocity of the incident
ow.
436 Introduction to Theoretical and Computational Fluid Dynamics
Combining (6.7.24) with (6.7.28), we nd that the traction exerted on the sphere is
fi (x) = 3 ijk j xk or f (x) = 3 x. (6.7.32)
a a
It is remarkable that the traction is proportional to the velocity over the surface of the rotating
sphere.
6.7 Stokes ow past or due to the motion of particles and liquid drops 437
b
x
a
c c
z
Figure 6.7.1 The ow due to the rotation or translation of a prolate spheroid can be represented by
a singularity distribution over the focal length of the spheroid.
Translation
The ow due to a prolate spheroid translating with velocity V in an innite uid can be represented
by distributions of Stokeslets and potential dipoles over the focal length of the spheroid pointing in
the direction of translation, with constant or parabolic distribution densities, respectively [87]. The
induced velocity eld is
c
1 2
where X() = [, 0, 0] is a point along the centerline and jk is a dimensionless diagonal matrix with
nonzero components
e2 2e2
xx = , yy = zz = . (6.7.37)
1+e 1+e
2e + (1 + e2 ) ln 2e (1 3e2 ) ln
1e 1e
The force exerted on the spheroid is found by integrating the strength density of the point forces
over the focal length,
F = 16c V. (6.7.38)
In the limit e 0, all three coecients a11 , a22 , and a33 tend to 3/(8e), in agreement with our
earlier results for a sphere.
where S is the Stokeslet, D is the potential dipole, N is the Stokeson, and c is a constant representing
a uniform ow. Substituting the expressions for the singularities and requiring that the velocity is
continuous across the drop surface, we nd that
Requiring that the component of the velocity normal to the drop surface is zero in a frame of
reference moving with the drop, (uint V) n = 0, we obtain an additional condition,
c + a2 w = V. (6.7.42)
A fourth equation emerges by requiring that the shear stress is continuous across the interface. Using
the formulas given in the tables of Section 6.6, we nd that the shear stress exerted on the exterior
and interior sides of the interface are given by
1 xi xj ij xi xj
fkshear,ext = 3 4 bj + 6 4 3 6 dj (ik ni nk ) (6.7.43)
4 a a a
or
xi xj
fkshear,int = 3 a ij 3 wj (ik ni nk ). (6.7.44)
a
6.7 Stokes ow past or due to the motion of particles and liquid drops 439
Note that the projection operator I n n represented by the last term on the right-hand side
extracts the tangential component of the surface traction. Equating the right-hand sides of (6.7.43)
and (6.7.44), we obtain
d = 2a5 w. (6.7.45)
Solving the system of equations (6.7.41), (6.7.42), and (6.7.45) for the coecients of the fundamental
solutions, we obtain
3 + 2 1 2 + 3 1 1
b = 2a V, d = a3 V, c= V, w= V. (6.7.46)
+1 +1 2 +1 2a2 + 1
In the limit of high viscosity ratio, , the coecients b and d tend to corresponding values
for the solid sphere given in (6.7.4), c tends to V, and w vanishes, yielding a uniform interior ow.
Having computed the coecients of the singularities, we derive exact representations for the
velocity eld inside and outside the drop in a stationary frame of reference,
1 1 a a2
a3 a2
xi xj
uext = 3 + 2 + 2 Vi + 3 3 + 2 3 2 Vi (6.7.47)
i
4 +1 r r r r a2
and
1 1 a2
xi xj
uint
i = 2 + 3 2 2 Vi + 2 Vj . (6.7.48)
2 +1 r a
The streamline patterns for = 1 in a stationary frame of reference and in a frame of reference
moving with the drop is depicted in Figure 6.7.2.
Inside the drop, the azimuthal component of the vorticity, , increases linearly with distance
from the axis, , and the ow is identical to that inside Hills spherical vortex. As a consequence,
the interior ow satises the unsimplied NavierStokes equation with the nonlinear terms included
on the left-hand side. However, this is not true for the ow outside the drop.
Drag force
The hydrodynamic drag force exerted on the drop is found from the coecient of the point force,
3 + 2
F = b = 2a V. (6.7.49)
+1
This expression was derived independently by Hadamard and Rybczynski in 1911 by solving a
partial-dierential equation for the Stokes stream function. Equation (6.7.49) can be expressed in
terms of a drag coecient as
F 4 3 + 2
cD = , (6.7.50)
V 2 a2 Re + 1
where is the uid density, V is the magnitude of the drop velocity, and Re = 2aV / is the
Reynolds number based on the drop diameter, 2a. Inertial corrections to this formula are available,
as reviewed by Clift, Grace, & Weber ([89], Chapter 5).
440 Introduction to Theoretical and Computational Fluid Dynamics
(a) (b)
3 3
2 2
1 1
0 0
y
y
1 1
2 2
3 3
3 2 1 0 1 2 3 3 2 1 0 1 2 3
x x
Figure 6.7.2 Streamline pattern due to the translation of a spherical liquid drop in an innite uid
with the same viscosity, = 1, (a) in a stationary frame of reference and (b) in a frame of reference
moving with the drop.
Using the expressions given in tables of Section 6.6, we nd that the normal component of the
traction on either side of the interface is
3
(f n)ext = (a2 b + 4d) x g x + c1 , (f n)int = 6 w x d g x + c2 , (6.7.52)
4a5
where n is the outward unit normal vector and c1 , c2 are two undened constants. Subtracting these
6.7 Stokes ow past or due to the motion of particles and liquid drops 441
expressions, we obtain
3
(f n)ext (f n)int = (a2 b + 4d 8a2 w) x ( d ) g x + c1 c2 . (6.7.53)
4a5
Substituting (6.7.51) into (6.7.46) and then into (6.7.53), we nd that the jump in the normal
component of the traction is constant and equal to c1 c2 = 2/a. Since the singularity solution
satises all required boundary conditions, surface tension is not necessary for a settling or rising drop
to maintain the spherical shape. However, surface tension does aect the stability of the spherical
shape.
Stability
Linear stability analysis for small perturbations shows that the spherical drop is stable, except in
the complete absence of surface tension. In practice, we nd that, if the surface tension is suciently
high or a perturbation is suciently small, the drop is able to restore its spherical shape. Otherwise,
the interface either elongates into an oblate shape or obtains a toroidal shape. Physically, the back
side of the drop is convected outward or inward under the inuence of the local stagnation-point
ow. Stages in the evolution of a settling drop with a prolate or oblate initial shape are shown in
Figure 6.7.3 for = 1 in the absence or presence of surface tension. The evolution of the interface
was computed using the boundary-integral method for Stokes ow [305].
6.7.6 Flow past a spherical particle with the slip boundary condition
Consider an innite linear ow with velocity u = U + A x past a translating and rotating
spherical particle of radius a, where U is a constant velocity and A is the transpose of the velocity
gradient tensor, AT = u [249]. The continuity equation requires the trace of A to be zero. The
no-penetration condition and a slip boundary condition are enforced at the particle surface,
u = V + (x xc ) + uS , (6.7.54)
where V is the velocity of translation of the particle center, xc , and is the angular velocity of
rotation about xc . The rst two terms on the right-hand side of (6.7.54) describe rigid-body motion,
while the third term expresses a slip velocity given by the NavierMaxwellBasset slip formula
1 a
uS = f (I n n) = n f n, (6.7.55)
where f n is the traction, is the stress tensor, n is the unit normal vector pointing into
the uid, I n n is the tangential projection operator, is the dimensionless Basset particle slip
coecient ranging from zero for vanishing shear stress and perfect slip to innity for nonzero shear
stress and no-slip, and = a/ is the slip length, as discussed in Section 3.7.4.
For convenience, we set the particle center at the origin, xc = 0, and express the velocity eld
in terms of ve singularities,
ui = Ui + Aij xj + a Sij bj + a3 Dij dj + a3 Cij cj + a3 Sijl
D
pjl + a5 Qijl qjl , (6.7.56)
D
where Sij is the Stokeslet, Dij is the potential dipole, Cij is the rotlet, Sijl is the Stokeslet dipole,
and Qijl is the potential quadrapole.
442 Introduction to Theoretical and Computational Fluid Dynamics
(a) (b)
1
0 0
0
0
1
2 1 2
2
2
3 4 4
3
4
4
6 6
5
5
6
6
8 8
7
7
8 8
10 10
9 9
1 0 1 1 0 1 1 0 1 1 0 1
Figure 6.7.3 Stages in the evolution of a settling drop with an oblate or prolate initial shape (a) in
the absence and (b) in the presence of surface tension for unit viscosity ratio, = 1. Snapshots
are shown at equal time intervals. When the surface tension is suciently strong, the drop restores
its spherical shape.
and thus
1
b+d= (V U), 3 [p]+ + 9 q = E, (6.7.58)
2
where the brackets [ ]+ denote the symmetric part of the enclosed tensor, and E is the rate-of-
deformation tensor, E = [A]+ .
6.7 Stokes ow past or due to the motion of particles and liquid drops 443
The traction exerted on the surface of the sphere is fi = ik xk /a, and its tangential component is
| xi x m xk S
fi = fi fm 2
= (Aik + Aki + a Tijk bj + a3 Tijk
S
dj + a3 Tijk
C
cj + a3 Tijlk
SD
pjl + a5 Tijlk
Q
qjl )
a a
x i xm x k S
Amk + Akm + a Tmjk gj + a3 Tmjk
D
dj + a3 Tmjk
C
cj
a3
+a3 Tmjlk
SD
pjl + a5 Tmjlk
Q
qjl . (6.7.62)
where the square brackets [ ] denote the antisymmetric part of the enclosed tensor and the su-
perscript T denotes the matrix transpose. Combining the rst equation in (6.7.58) with the rst
equation in (6.7.65), we obtain
3 +2 1
b= (V U), d= (U V). (6.7.68)
4 +3 4 +3
The force exerted on the sphere is
+2
F = 8 a b = 6 a (V U). (6.7.69)
+3
As tends to zero, the Stokes-law coecient of six tends to four, indicating a substantial reduction
in the drag force. Combining the second equation in (6.7.58) with equations (6.7.66) and (6.7.67),
we nd that
5 +2 1 1
[p]+ = E, [p] = , q= E, (6.7.70)
6 +5 2 +3 6 +5
(4 2 + 20 + 15) A + ( 2 + 5 + 15)AT
p = [p]+ + [p] = . (6.7.71)
6 ( + 3)( + 5)
The coecient of the couplet inherent in the antisymmetric part of the Stokeslet dipole is
1 a3
cSD 3
m = mjl a [pjl ] = mjl a3 Ajl = , (6.7.72)
2 +3 2 +3 m
where is the vorticity of the ambient linear ow. The torque experienced by the sphere is
T = 8 (a3 c + cSD ) = 4a3 (2 ). (6.7.73)
+3
A torque-free particle rotates with angular velocity that is equal to half the vorticity of the incident
ow. When = 0, a rotating particle does not generate a ow and the torque vanishes.
velocity vanishes over the surface of the sphere ([61], p. 120). The approximate solution reproduces
precisely Stokes law (Problem 6.7.4(a)). The perfect agreement is explained by recalling that the
exact solution is composed by a Stokeslet and a potential dipole, and noting that average velocity
of the dipole is zero over the surface of a sphere. Burgers performed a similar computation for the
disturbance ow due to a sphere held stationary in a paraboloidal ow and found that the drag force
is still in perfect agreement with the exact solution which can be found readily using Faxen relations
discussed in Section 6.12 (Problem 6.7.4(b)).
Numerical methods
To develop a formal procedure for computing singularity representations, we introduce a positive
functional expressing the dierence between the specied boundary conditions and the corresponding
boundary values computed using the singularity representation. When the problem requires that
u = V over a boundary, D, we introduce a positive functional F(u V) that vanishes when u = V.
Other boundary conditions are implemented in a straightforward fashion.
The general strategy is to represent the velocity u in terms of a certain singularity distribution,
and then minimize the functional, F, with respect to the strength of the singularities and possibly
the location of their poles (e.g., [442]). When velocity boundary conditions are specied, a wise
choice is the least-squares collocation functional
M
F(u V) = |u(xi ) V(xi )|2 , (6.7.74)
i=1
where xi , is a collection of M collocation points over D. Minimizing (6.7.74) with respect to the
strength of the singularities yields a system of linear equations. When the force or torque acting
on the boundary D are specied, the strengths of the singularities must satisfy additional linear
constraints originating from (6.6.10).
The least-squares collocation method has found extensive applications in a variety of numerical
studies of Stokes ow involving suspended particles. In certain cases, instead of using singularity ex-
pansions, it is more expedient to use alternative expansions in terms of spherical harmonic functions
([220], 335; [208], Chapter 13). A generalized implementation allows the singularities to relocate
as part of the optimization (e.g., [442]).
Problems
Computer Problem
ui ij ui ij = u u , (6.8.1)
xj
where and are the uid viscosities. For simplicity, we omit the tilde above the hydrodynamic
stress, pressure, and traction. If both ows satisfy the Stokes equation, the right-hand side of (6.8.1)
is zero, yielding the Lorentz reciprocal identity [245]
( ui ij ui ij
) = 0, (6.8.2)
xj
which is the counterpart of Greens second identity for two scalar harmonic potentials.
Assuming that the two Stokes ows of interest are free of singular points inside a control
volume, Vc , bounded by a surface, D, we integrate (6.8.2) over the control volume and use the
6.8 The Lorentz reciprocal theorem and its applications 447
where f = n is the boundary traction and the unit normal vector n may point either into or out
of the control volume.
The reciprocal identities (6.8.2) and (6.8.3) allow us to obtain information about a certain
ow without explicitly solving the equations of Stokes ow, but merely by using information about
another ow. An alternative form of the reciprocal theorem involving the velocity and the pressure
is stated in Problem 6.8.1.
Force
To compute the force exerted on the particle, we turn to (6.8.3) and identify u with the velocity
eld generated when the particle translates with velocity V. Exploiting the linearity of the Stokes
equation, we express the corresponding traction exerted at the particle surface as
f t = Rt V, (6.8.4)
where Rt is a traction resistance matrix for translation indicated by the superscript t, not to be
confused with the matrix transpose.
Next, we select a control volume enclosed by the particle surface, denoted by P , and a surface
with large size enclosing the particle, denoted by S . Applying the reciprocal relation (6.8.3) for
the pair of ows ut and uD with equal uid viscosities, we obtain
ut f D dS = uD f t dS. (6.8.5)
S ,P S ,P
Letting the surface S tend to innity and noting that the velocity decays at least as fast as 1/d
and the traction decays at least as fast as 1/d2 , where d is the distance from a designated particle
center, we nd that the contribution of the surface integrals over S disappears. Equation (6.8.5)
then yields
V F =
D
uD f t dS, (6.8.6)
P
where FD is the disturbance force exerted on the particle. In the absence of inertia, the force exerted
on any uid volume by the ambient ow is zero and the disturbance force, FD , is equal to the total
448 Introduction to Theoretical and Computational Fluid Dynamics
This expression allows us to compute the hydrodynamic force exerted on a stationary particle in
terms of the incident velocity over the particle surface and the traction resistance matrix for trans-
lation, Rt [56].
Torque
Next, we consider the torque exerted on a rigid particle that is held stationary in an ambient ow.
Repeating the previous steps for the force, we introduce a rotary traction resistance matrix , Rr ,
dened by the equation
f r = Rr O, (6.8.8)
where f r is the traction established when the particle rotates about a specied point, x0 , with
angular velocity O. The analysis provides us with the counterpart of (6.8.7) for the torque with
respect to x0 ,
T= u Rr dS. (6.8.9)
P
This expression allows us to evaluate the hydrodynamic torque exerted on a stationary particle in
terms the incident velocity over the particle surface and the traction resistance matrix for rotation,
Rr [56].
Spherical particle
As an application, we use expressions (6.7.11) and (6.7.32) for the traction on a spherical particle of
radius a, and deduce the translational and rotary traction resistance matrices
t 3 1 r 3
Rij = ij , Rij = ijk xk , (6.8.10)
2 a a
where x = x x0 is the distance from the particle center, x0 . Substituting these expressions into
(6.8.7) and (6.8.9), we obtain
3
F= u dS, T=3 x u dS. (6.8.11)
2 a P a P
Because the matrix Rt is constant and diagonal, the force is proportional to the surface average of
the incident velocity, u .
inside a tube can be computed from knowledge of the traction exerted on the particle when it
translates or rotates in an otherwise stationary uid inside the tube.
uD = V + (x x0 ) u , (6.8.12)
and recalling that the disturbance force is equal to the total force, we obtain
VF +T =
t t
u f t dS + V F, (6.8.13)
P
where Ft and Tt are the force and torque with respect to the point x0 exerted on the particle when
it translates with velocity V. Working in a similar fashion in the case of particle rotation, we obtain
the corresponding equation
V Fr + Tr = u f R dS + O F, (6.8.14)
P
where Fr and Tr are the force and torque with respect to the point x0 exerted on the particle when
it rotates around the point x0 with angular velocity O.
Now we take advantage of the linearity of the equations of Stokes ow with respect to the
boundary conditions and write
Ft = X V, Tt = Z V, Fr = W O, Tr = Y O, (6.8.15)
where X, Z are resistance matrices for translation and W, Y are resistance matrices for rotation.
Using the reciprocal theorem, we may show that the resistance matrices X and Y are symmetric
450 Introduction to Theoretical and Computational Fluid Dynamics
and Z is the transpose of W, W = ZT (Problem 6.8.3). Using the denitions (6.8.15), we recast
(6.8.13) and (6.8.14) into the forms
1
VX+Z= u Rt dS F,
P
(6.8.16)
1
VZ +Y =
T
u R dS T.
r
P
Given the force, F, and torque, T, exerted on the particle and the resistance matrices for translation
and rotation, we obtain a linear system of equations for the particle translational and rotational
velocities, V and .
where f F and f T are the tractions exerted on the particle when it moves under the inuence of an
external force, F, or an external torque, T (Problem 6.8.4). An important consequence of (6.8.17)
is that the translational and angular velocities of a force-free and torque-free particle immersed in
an arbitrary ambient ow can be computed from the distribution of the incident velocity over the
particle surface and the traction exerted on the particle surface when it moves under the inuence
of an external force or torque.
Analogous symmetry properties exist for the Greens functions of potential ow, linear elastostatics
and, more generally, for the Greens function of linear, elliptic partial dierential equations involving
self-adjoint dierential operators. Physically, the symmetry property states that the velocity at a
point, x, due to a point force at another point, x0 , can be deduced from the velocity at the point
x0 due to a point force located at x.
To prove (6.8.18), we introduce the velocity elds induced by two point forces located at x1
and x2 with strengths b(1) and b(2) , given by
S
B
x1
x2
Figure 6.8.1 Illustration of two point forces near a boundary, SB , introduced to prove the symmetry
of the Greens functions of Stokes ow.
as illustrated in Figure 6.8.1. The corresponding stress elds satisfy the equations
(1) (2)
ij (1) ij (2)
= bi 3 (x x(1) ), = bi 3 (x x(2) ), (6.8.20)
xj xj
where 3 is the three-dimensional delta function. Substituting these expressions into the right-hand
side of (6.8.2) with equal uid viscosities, = , we obtain
Next, we integrate (6.8.21) over a control volume that is conned by a solid boundary where the
Greens functions are required to be zero, SB , two spherical surfaces of innitesimal radii enclosing
x1 and x2 , and, in the case of innite ow, a large surface enclosing SB as well as the points x1
and x2 . Using the divergence theorem, we convert the volume integral on the left-hand side of the
resulting equation into a surface integral over all surfaces enclosing the control volume.
The surface integral over SB vanishes because the Greens function and hence the velocities
u(1) and u(2) are zero over SB . Letting the radius of the large surface tend to innity and the radii
of the small spheres enclosing x1 and x2 to zero, we nd that the corresponding surface integrals
make insignicant contributions. As a result, the whole integral of the left-hand side of (6.8.21)
vanishes. Using the properties of the delta function to manipulate the right-hand side, we nally
arrive at the equation
(1) (2)
bk bi Gki (x1 , x2 ) Gik (x2 , x1 ) = 0. (6.8.22)
Property (6.8.18) follows by observing that the constants b(1) and b(2) are arbitrary.
Repeating this procedure with some modications, we can show that the two-dimensional
Greens functions also satisfy the symmetry property (6.8.18) (Problem 6.8.5).
452 Introduction to Theoretical and Computational Fluid Dynamics
D x0
x
D+
Figure 6.8.2 Illustration of a point force inside cylindrical tube with arbitrary cross-section. The
corresponding ow aords one degree of freedom determining the axial ow rate or pressure drop.
1 1
[ uP
i (x) Tikj (x, x0 ) Gik (x, x0 ) ij
P
(x) ] = upi (x) 3 (x x0 ), (6.8.23)
xj 8 8
where k is a free index indicating the direction of the point force. We will assume that the velocity
generated by the point force decays far upstream and downstream, and the pressure tends to a
constant value far from the point force, so that
8 8
Pj (x , x0 ) jx , Tijk (x , x0 ) jx ik , (6.8.24)
a2 a2
where are dimensionless constants and a is a specied length. These circumstances arise if the
tube is closed at both ends.
Next, we integrate (6.8.23) over a volume of uid inside the tube bounded by a far upstream
cross-sectional surface, D+ , and a far downstream cross-sectional surface, D . Using the divergence
theorem, we convert the volume integral into a surface integral. Recalling that uP and G are both
zero when the point x lies at the tube surface, and observing that G is virtually zero on D due
to the fast decay of the ow due to a point force, we nd that
p
i (x) Tikj (x, x0 ) nj (x) dS(x) = 8 ui (x0 ),
uP (6.8.25)
D
where n is the unit normal vector pointing into the control volume. Evaluating the integral using
(6.8.24), we obtain (+ ) Qp = a2 upi (x0 ), where
Qp = uP n dS (6.8.26)
D
6.8 The Lorentz reciprocal theorem and its applications 453
n
y
C
Figure 6.8.3 Illustration of a periodic array of two-dimensional point forces above a plane wall.
is the axial ow rate of the pressure-driven ow. Rearranging, we derive an expression for the
pressure drop due to a point force inside a closed tube,
a2 p
+ = u (x0 ). (6.8.27)
Qp i
This pressure drop induces a back ow that cancels the forward ow induced by the point force.
The functional dependence of the pressure drop on the position of the point force is precisely that
exhibited by the velocity prole of the corresponding pressure-driven ow.
where k is a free index indicating the direction of the point forces and 2 is the two-dimensional
delta function.
454 Introduction to Theoretical and Computational Fluid Dynamics
When the point forces are perpendicular to the wall, the induced velocity decays far from the
wall. When the point forces are parallel to the wall, the induced velocity tends to a constant value
far from the wall,
as y , where is a dimensionless constant. The corresponding pressure and stress elds decay
at an exponential rate.
Next, we integrate (6.8.29) over one period of the ow conned between the wall and a periodic
line far above the wall, C, and use the divergence theorem to convert the areal integral into a line
integral along the boundary of one period of the ow. Noting that the net contribution of the
periodic segments is zero and recalling that ussf and G are zero when x lies on the wall, we obtain
ssf
Gik (x, x0 ) ij (x) nj (x) dl(x) = 4 ussf
i (x0 ), (6.8.31)
C
where n is the unit normal vector pointing outward from the control volume. Substituting (6.8.30)
and evaluating the stress eld of the simple shear ow, we nd that
L = ussf
i (x0 ), (6.8.32)
The results are readily generalized to the ow induced by a doubly periodic array of three-
dimensional point forces arranged in a Bravais lattice parallel to an innite plane wall located at
x = 0. The counterpart of (6.8.30) is
x0
Gij (x, x0 ) 8 Jij , (6.8.34)
A
where x0 is the distance of the point forces from the wall, J is the identity matrix except that the
rst diagonal element is set to zero, and A is the area occupied by one periodic cell in the yx plane
parallel to the wall.
Problems
ui
ui
ui ik p + ui ik p + = 0. (6.8.35)
xk xk xk
6.9 Boundary-integral representation of Stokes ow 455
D
n
SB
Vc D n
n
S
x0
V
Figure 6.9.1 A control volume used to derive the boundary-integral representation of Stokes ow and
associated integral equations.
the point force outside the control volume. Noting that the function inside the large parentheses in
(6.9.2) is nonsingular throughout Vc , we integrate (6.9.2) over Vc and use the divergence theorem to
convert the volume integral over Vc into a surface integral over D, obtaining
[ Gij (x, x0 ) ik (x) ui (x) Tijk (x, x0 ) ] nk (x) dS(x) = 0. (6.9.3)
D
In equation (6.9.3) as well as in all subsequent equations, the unit normal vector, n, points into the
control volume, Vc .
Integral representation
Now we place the point x0 inside the control volume, Vc , and introduce a small spherical volume
V of radius centered at x0 . The function inside the square brackets in (6.9.2) is nonsingular
throughout the reduced volume, Vc V . Integrating (6.9.2) over Vc V and using the divergence
theorem to convert the volume integral into a surface integral, we obtain
[Gij (x, x0 ) ik (x) ui (x) Tijk (x, x0 )] nk (x) dS(x) = 0, (6.9.4)
D,S
where S is the spherical surface enclosing the volume V , as illustrated in Figure 6.9.1. Letting
the radius shrink to zero, we nd that, to leading order in , the tensors G and T reduce to the
Stokeslet and its associated stress tensor,
ij xi xj xi xj xk
Gij + 3 , Tijk 6 , (6.9.5)
5
where x = x x0 . Over S , the normal vector is n = 1 x and the dierential surface area is
dS = 2 d, where is the dierential solid angle. Substituting these expressions along with (6.9.5)
6.9 Boundary-integral representation of Stokes ow 457
As 0, u and over S tend to their respective values at the center of the small volume V , which
are u(x0 ) and (x0 ). Since x decreases linearly with as 0, the contribution of the stress term
inside the integral on the right-hand side of (6.9.6) decreases linearly with whereas the contribution
of the velocity term tends to a constant value. Thus, in the limit 0, equation (6.9.6) becomes
1
[Gij (x, x0 ) ik (x) ui (x) Tijk (x, x0 )] nk (x) dS(x) = 6ui (x0 ) 4 xi xj dS(x). (6.9.7)
D S
Equation (6.9.9) provides us with an integral representation of the velocity eld in terms of
boundary distributions of the Greens function, G, and associated stress tensor, T . The densities
of these distributions are proportional, respectively, to the boundary values of the traction and
velocity. Making an analogy with corresponding results in the theory of electrostatics, we term
the rst distribution involving the boundary traction the single-layer potential, and the second
distribution involving the boundary velocity the double-layer potential.
Physical interpretation
The symmetry property (6.5.12) allows us to switch the order of the indices of the Greens function
in the single-layer potential, provided that we also switch the order of the arguments, x and x0 ,
obtaining
1 1
uj (x0 ) = Gji (x0 , x)fi (x) dS(x) + ui (x) Tijk (x, x0 ) nk (x) dS(x). (6.9.10)
8 D 8 D
It is now evident that the single-layer potential represents a boundary distribution of point forces
with surface strength equal to f . The physical interpretation of the double-layer potential will be
discussed later in this section.
458 Introduction to Theoretical and Computational Fluid Dynamics
Consider a ow that is bounded by an internal or external solid stationary surface, B. Since the
velocity is zero over B, the corresponding double-layer integral does not appear. To also eliminate
the single-layer potential, we use a Greens function that vanishes over B, that is, G(x, x0 ) = 0
when x is on B. In this way, B is conveniently excluded from the domain of the boundary-integral
equation.
Integral identities
Useful identities can be derived by applying the boundary-integral equation for simple ows that
are known to be exact solutions of the equations of Stokes ow. In the case of rigid-body motion
with translational velocity V and angular velocity , the uid velocity is u = V + x, where
x = x xc and xc is the center of rotation. Setting f = pn, where p is the constant hydrodynamic
pressure, we derive the identities
Tijk (x, x0 ) nk (x) dS(x) = 8 ij (6.9.11)
D
and
ilm xm Tijk (x, x0 )nk (x) dS(x) = 8 jlm x0m , (6.9.12)
D
where D is the smooth boundary of an arbitrary control volume, x0m denotes the mth component
of x0 , and = 1 or 0 depending on whether the point x0 is located inside or outside D. These
equations reiterate identities (6.5.10) and (6.5.11). Similar identities can be written for linear or
parabolic ow.
6.9 Boundary-integral representation of Stokes ow 459
Now combining the two representations to eliminate the double-layer integral from the right-hand
side of (6.9.13) and adding the incident velocity eld u to both sides of the resulting equation, we
derive a simplied single-layer representation,
1
uj (x0 ) = u
j (x 0 ) fi (x) Gij (x, x0 ) dS(x). (6.9.15)
8 P
The disturbance velocity eld is represented by a surface distribution of point forces whose strength
density is the boundary traction.
Gij
+ (x0 , xc ) (xk xc,k ) fi (x) dS(x) + . (6.9.16)
xck P
The rst integral on the right-hand side of (6.9.16) is the force exerted on the particle, F. The
second integral is the rst moment of the boundary traction. Subsequent integrals express higher
moments of the boundary traction.
The rst term on the right-hand side of (6.9.16) represents the ow due to a point force. The
second term represents the ow due to a point-force dipole. Subsequent terms represent the ow
due to point-force quadruples and higher-order multipoles. The coecient of the point-force dipole,
q, can be decomposed into the coecient of the stresslet, s, and an antisymmetric component, r,
460 Introduction to Theoretical and Computational Fluid Dynamics
amounting to a torque mediating couplet, as discussed in Section 6.6.2. Equation (6.9.16) shows
that, far from a particle, the disturbance ow due to the particle is similar to that due to a point
force with strength F. When F = 0, the far ow is similar to that produced by a point-force
dipole. Consequently, in the case of ow in an innite domain, the disturbance ow decays as 1/d
or 1/d2 , respectively, for F = 0 or F = 0, where d is the distance from the particle center.
Translating sphere
Substituting into (6.9.15) the expression for the traction exerted on a sphere translating with velocity
V in an innite quiescent uid, given in (6.7.11), we obtain
3
uj (x0 ) = Vi Sij (x, x0 ) dS(x), (6.9.17)
16a Sphere
where Sij is the Stokeslet. Substituting into the left-hand side the singularity representation (6.7.1)
expressed in the form
3 1
uj (x0 ) = Vi a + a3 20 Sij (x0 , xc ), (6.9.18)
4 8
discarding the arbitrary velocity Vi , and rearranging, we obtain the identity
1 1
2
Sij (x, x0 ) dS(x) = 1 + a2 2c Sij (x0 , xc ), (6.9.19)
4a Sphere 6
where xc is the center of the sphere, the gradient c involves derivatives with respect to xc , and
the point x0 lies outside the sphere. This expression is consistent with the mean-value theorem for
biharmonic functions stated in (2.4.14).
For a point x0 that lies inside the sphere, the representation (6.9.17) yields uj (x0 ) = Vj , and
we nd that
16
Sij (x, x0 ) dS(x) = a ij , (6.9.20)
Sphere 3
independent of the precise position of x0 inside the sphere. This identity can be conrmed readily
when the point x0 is the center of the sphere.
Using the denition of the Greens function and the symmetry property of the velocity Greens
function tensor, we write
Pj (x, x0 )
= 2 Gjk (x0 , x). (6.9.22)
xk
This equation reveals that all derivatives of the pressure can be expressed in terms of the Laplacian
of the Greens function.
The vector P and tensor P express the pressure corresponding to the Greens function and asso-
ciated stress tensor expressing a stresslet.
where I denotes the interface. Since the point x0 is located in the exterior of the interface, we may
use the reciprocal identity (6.9.1) for the incident ow u to write
Gji (x0 , x) fi (x) dS(x) = ui (x) Tijk (x, x0 ) nk (x) dS(x). (6.9.27)
I I
Combining the last two representations to formulate the total velocity and traction, u = u + uD
and f = f + f D , and adding the incident velocity eld u to both sides of the resulting equation,
we obtain
1 1
uj (x0 ) = uj (x0 ) Gji (x0 , x) fi (x) dS(x) + ui (x) Tijk (x, x0 ) nk (x) dS(x). (6.9.28)
8 I + 8 I +
The notation I + emphasizes that the velocity and traction are evaluated on the outer side of the
interface indicated by the normal vector, n.
Gij
+ (x0 , xc ) [(xk xc,k ) fi (x) (uk ni + ui nk )(x)] dS(x) +
xck I+
1
Pj (xc , x0 ) u n dS + . (6.9.29)
8 I
The rst series on the right-hand side contains multipoles of the point force, while the second series
contains multipoles of the pressure. It is evident now that the pressure term P(xc , x0 ) represents
6.9 Boundary-integral representation of Stokes ow 463
the velocity eld at the point x0 due to a point source at the point xc , in agreement with our earlier
observations. Formula (6.9.22) suggests that all terms after the rst term in the second series on the
right-hand side of (6.9.29) can be incorporated into the rst series. Consequently, the disturbance
ow far from the interface can be represented in terms of an expansion of multipoles of the point
force supplemented by a point source.
where I denotes the interior side of the interface. Combining (6.9.28) with (6.9.30) to formulate
the jump in the traction across the interface, f f + f , and assuming that the velocity is
continuous across the interface, we derive the integral representation
1 1
uj (x0 ) = uj (x0 ) Gji (x0 , x) fi (x) dS(x) + ui (x) Tijk (x, x0 ) nk (x) dS(x),
8 I 8 I
(6.9.31)
where n is the unit normal vector pointing into the ambient uid. In the case of equal viscosities,
= 1, the double-layer integral does not appear.
Since the force on the exterior side of the interface is equal to the force on the interior side
of the interface to satisfy global equilibrium, and since the particle volume is xed, the multipole
expansion (6.9.29) simplies into
Gij
uj (x0 ) = u
j (x0 ) + (x0 , xc ) pik + , (6.9.32)
xck
where
pik = (xk xc,k ) fi (x) (1 ) (uk ni + ui nk )(x) dS(x). (6.9.33)
I+
In the case of equal uid viscosities, = 1, the second term inside the integral does not appear.
Working in a similar fashion, we nd that the velocity eld in the interior of a drop, capsule,
or cell is given by the right-hand side of (6.9.31), provided that all terms, including the rst term
representing the ambient ow, are divided by the viscosity ratio, ([306], Chapter 5).
6.9.7 Axisymmetric ow
In the case of axisymmetric ow, we express the Cartesian velocity and traction components in terms
of their radial and meridional counterparts, and perform the integration in the azimuthal direction,
464 Introduction to Theoretical and Computational Fluid Dynamics
, to reduce the surface integrals into line integrals along the boundary contour in an azimuthal
plane. The corresponding kernels of the single- and double-layer potential involve complete elliptical
integrals arising from integrals with respect to the azimuthal angle, . Further discussion and specic
expressions can be found elsewhere (e.g., [306, 313]).
6.9.8 Two-dimensional ow
To derive the boundary-integral representation of two-dimensional Stokes ow, we repeat our earlier
steps for three-dimensional ow. Applying the reciprocal theorem, we nd that
Gij (x, x0 ) fi (x) dl(x) = ui (x) Tijk (x, x0 ) nk (x) dl(x), (6.9.34)
C C
where the point x0 lies outside a selected domain of ow enclosed by the contour C, and l is the arc
length along C. For a point x0 located inside the selected domain of ow, we obtain the boundary-
integral representation
1 1
uj (x0 ) = Gij (x, x0 ) fi (x) dl(x) + ui (x) Tijk (x, x0 ) nk (x) dl(x), (6.9.35)
4 C 4 C
where the unit normal vector, n, points into the domain of ow. All properties, interpretations, and
simplications of the boundary-integral equation discussed earlier for three-dimensional ow also
apply to two-dimensional ow. Exceptions arise in the case of innite ow past a body and ow
produced by the motion of a body in an innite uid. In both cases, if the force exerted on the
body is nonzero, the disturbance velocity far from the body increases at a logarithmic rate and the
boundary integrals at innity may not be overlooked.
The pressure vector, P, and stress tensor, T , associated with a Greens function for an entirely
or partially innite domain of ow constitute two fundamental solutions of Stokes ow. Specically,
P(x, x0 ) represents the velocity at the point x0 due to a two-dimensional point source located at x.
Similarly, the ow given in (6.9.23) represents the velocity at the point x0 due to a two-dimensional
stresslet located at x. The corresponding pressure eld may be expressed in terms of a pressure
tensor P ST R , as shown in (6.9.23). In the case of ow in innite space,
ik xi xk
Pik
ST R
(x, x0 ) = 2 2
+4 4 . (6.9.36)
r r
The pressure eld of a two-dimensional Stokes ow can be expressed in terms of two distributions
corresponding to the single- and double-layer potentials,
1
p(x0 ) = Pi (x, x0 ) fi (x) dl(x) + ui (x) Pik
ST R
(x, x0 ) nk (x) dl(x). (6.9.37)
4 C 4 C
The vector P and tensor P ST R express the pressure eld corresponding to the Greens function and
associated stress tensor representing a stresslet.
6.10 Boundary-integral equation methods 465
Problems
1 ui
uj (x0 ) = Gji (x0 , x) p ni + nk (x) dS(x) (6.9.38)
8 D xk
1 Gji (x0 , x)
+ ui (x) Pj (x, x0 ) ni (x) + nk (x) nk (x) dS(x)
8 D xk
To derive the boundary-integral equation, we examine the behavior of the boundary integrals
as the eld point, x0 , approaches the boundary D of a selected domain of ow. Due to the weak
singularity of the Stokeslet, the single-layer integral remains continuous as the point x0 approaches
and then crosses D from either side. If the velocity and normal vector vary continuously over D, the
double-layer potential undergoes a discontinuity of magnitude 8u(x0 ) as the point x0 crosses D,
PV
lim ui (x) Tijk (x, x0 ) nk (x) dS(x) = ui (x) Tijk (x, x0 ) nk (x) dS(x) 4 uj (x0 ), (6.10.1)
x0 D D D
where P V designates the principal value of the double-layer potential dened as the improper double-
layer integral computed when the point x0 is located precisely on D. The plus sign applies when the
point x0 approaches D from the side of the selected volume of ow indicated by the normal vector,
466 Introduction to Theoretical and Computational Fluid Dynamics
The rst integral on the right-hand side is continuous due to the reduced singularity of the kernel.
A discontinuity is evident from identity (6.9.11).
Substituting (6.10.1) with the plus sign into (6.9.9) or with the minus sign into (6.9.3), we
nd that, when the eld point x0 is located on D,
P V
1 1
uj (x0 ) = Gij (x, x0 ) fi (x) dS(x) + ui (x) Tijk (x, x0 ) nk (x) dS(x). (6.10.3)
4 D 4 D
This integral equation imposes an integral constraint on the mutual distributions of the boundary
velocity and traction. Conversely, the integral equation allows us to compute the boundary velocity
from the boundary traction, and vice versa, as required.
In summary, equations (6.9.2), (6.9.10), and (6.10.3) apply, respectively, when the point x0 is
located outside, inside, and precisely on a smooth boundary of a selected control volume, D.
where
PV
IjD (x0 ) ui (x) Tijk (x, x0 ) nk (x) dS(x) (6.10.5)
D
Alternatively, prescribing the boundary traction f over D provides us with a Fredholm integral
equation of the second kind for the boundary velocity originating from (6.10.3),
P V
1 1 S
ui (x0 ) = ui (x) Tijk (x, x0 ) nk (x) dS(x) I (x0 ), (6.10.6)
4 D 4 j
where
IjD (x0 ) Gij (x, x0 ) fi (x) dS(x) (6.10.7)
D
Prescribing the velocity over a portion of D and the traction over the remaining portion of D
provides us with a Fredholm integral equation of mixed kind for the unknown boundary distributions.
Applications of the mixed formulation can be found in ows bounded by solid boundaries and free
surfaces or interfaces. Once the integral equations have been solved, the velocity, pressure, and stress
elds can be computed using the boundary-integral representations (6.9.10) and (6.9.25).
6.10.2 Two-dimensional ow
Repeating the preceding derivations with minor modications, we derive corresponding results for
two-dimensional ow. As the eld point x0 approaches and then crosses a boundary contour C
from either side, the single-layer potential remains continuous. If the velocity and normal vector
vary continuously over a smooth contour C, the double-layer potential undergoes a discontinuity of
magnitude 4u(x0 ),
PV
lim ui (x) Tijk (x, x0 ) nk (x) dl(x) = ui (x) Tijk (x, x0 ) nk (x) dl(x) 2 uj (x0 ), (6.10.8)
x0 C C C
where the plus sign applies when x0 approaches C from the side of the ow indicated by the normal
vector, n, and the minus sign otherwise (Problem 6.10.1). The principal value of the double-layer
integral is the value of the improper double-layer integral computed when the evaluation point x0
is located precisely on C.
V
1 2 3 K
N
N1 T
L R
i i 1
where S is the Stokeslet, T is the corresponding stress tensor, V is the lid velocity, and the point
x0 lies inside the cavity. The integral equation (6.10.9) takes the corresponding form
PV
Sij (x, x0 ) fi (x) dl(x) = 2uj (x0 ) V Txjk (x, x0 ) dl(x), (6.10.11)
T,L,B,R T
where the point x0 is located on T, L, B, or R. The problem has been reduced to solving an integral
equation of the rst kind for the boundary traction f inside the single-layer integral on the left-hand
side.
Boundary-element method
To solve the integral equation (6.10.11), we may discretize the boundaries of the ow into straight
elements and approximate the traction components with constant functions constant over each
element, as shown in Figure 6.10.1. This approximation allows us to recast (6.10.11) into the
6.10 Boundary-integral equation methods 469
discretized form
N
(l) (l)
K
(l)
fi Aij (x0 ) = 2uj (x0 ) V Bj (x0 ), (6.10.12)
l=1 l=1
where f (l) is the constant value of the traction over the lth element, N is the total number of segments,
and the sum on the right-hand side runs over the K lid elements. The inuence coecient, A(l) ,
and forcing vector, B(l) , are dened as
PV
(l) (l)
Aij Gij (x, x0 ) dl(x), Bj Txjy (x, x0 ) dl(x), (6.10.13)
El El
where El denotes the lth boundary element. Applying (6.10.12) at the midpoint of the mth element,
denoted by X(m) , we obtain a system of N linear equations for f (l) ,
N
(l) (l)
K
(l)
fi Aij (X(m) ) = 2uj (X(m) ) V Bj (X(m) ), (6.10.14)
l=1 l=1
When the mth element midpoint, X(m) , is located over the lid, the kernel Txjy is identically
(l)
zero and Bi (X(m) ) = 0 for l, m = 1, . . . , K. The associated inuence coecient A(l) (X(m) ) corre-
sponding to the lid segments can be computed exactly by noting that the o-diagonal components
(l) (l)
are zero, Axy (X(m) ) = 0 and Ayx (X(m) ) = 0, and evaluating the integrals
|X (m) x| |X (m) x|
A(l)
xx = ( ln + 1) dx, A (l)
yy = ln dx (6.10.15)
El a El a
C w = d, (6.10.17)
where C is an inuence coecient matrix and d is a constant vector. The rst set of N equations in
(6.10.17) correspond to the x components of (6.10.14), and the second set of N equations correspond
to the y components of (6.10.14) for X(1) , X(2) , . . ., X(N ) .
470 Introduction to Theoretical and Computational Fluid Dynamics
Regularization
The linear system (6.10.17) is nearly singular due to the unspecied level of the pressure inside the
ow. To prevent numerical diculties, we may specify the normal component of the traction at a
chosen element, discard one equation, and then solve a system of 2N 1 equations for the remaining
unknowns. More reliable methods of removing the singularity by eigenvalue deation are available
(e.g., [313]). In the present case, we may exploit the anticipated symmetry of the ow setting, for
example,
When an even number of elements is employed, implementing the symmetry condition reduces
the size of the nal system of equations by a factor of two. The simplied linear system contains
equations corresponding to N/2 collocation points distributed over the left or right half of the cavity.
Problem
(a) (b)
9
4 8
3.5 7
3 6
2.5 5
2 4
1.5 3
1 2
0.5 1
0 0
2 1.5 1 0.5 0 0.5 1 1.5 2 2 0 2
Figure 6.10.2 Stages in (a) the gravitational spreading of a liquid drop over a plane wall and (b)
the gravitational rise of a liquid drop away from a wall, in the absence of surface tension. The
viscosity of the drop is equal to that of the ambient uid. Interfacial proles are shown at equal
time intervals.
Computer Problem
y
b
x
a z a
Figure 6.11.1 In slender-body theory, the ow due to the axial or transverse translation of an elongated
circular cylinder is described by distributions of point forces and potential dipoles over the centerline.
In this section, we review the basic concepts of the slender-body theory and present a theo-
retical framework that supports the basic premise of the slender-body approximation. Our analysis
will conrm that the centerline singularity distributions are rational approximations of surface dis-
tributions arising from the boundary-integral representation, thus validating an otherwise heuristic
approach.
a
1 (x )2
ux (x, b) = Vx + d. (6.11.1)
8 a [(x )2 + b2 ]1/2 [(x )2 + b2 ]3/2
Introducing the dimensionless variable w = (x )/b and rearranging the integrand, we obtain
B
2 1
ux (x, b) = Vx dw, (6.11.2)
8 A (w 2 + 1)1/2 (w2 + 1)3/2
where
a+x ax
A= , B= . (6.11.3)
b b
Both A and B are positive when a < x < a. Performing the integration, we nd that
B A
u (x, b) = Vx . (6.11.6)
8 A2 1 B2 1
Asymptotics
As the aspect ratio b/a tends to zero, the linear functions A(x) and B(x) tend to innity and (6.11.4)
takes the asymptotic form
ux (x, b) Vx ln(4AB) 1 . (6.11.7)
4
The product AB = (a2 x2 )/b2 describes a parabolic distribution that vanishes at both cylinder
ends, x = a. Applying (6.11.7) at the midpoint of the cylinder and at a point located at a quarter
of the length of the rod away from the tips, we obtain
L a L 3
ux (0, b) = Vx ( 2 ln 1 ), ux ( , b) = Vx ( 2 ln + ln 1 ). (6.11.8)
4 b 2 4 b 4
To satisfy the no-slip boundary condition, we require that ux (x, b) = Vx . When the aspect ratio L/b
is large, the two velocities in (6.11.8) are not too far apart, and the choice
2
= 1 (6.11.9)
ln
L
b 2
provides us with a reasonable singularity representation. The drag force exerted on the cylinder is
given by
F = Vx L ex , (6.11.10)
where ex is the unit vector along the x axis. The radial surface velocity given in (6.11.6) is of order
b/L, which is small as long the aspect ratio is large, corroborating the consistency of the approximate
representation.
Surface distribution
To justify the ow representation in terms of a line distribution of point forces along the centerline,
we consider the underlying dimensionless velocity eld
a
1
vi (x) = Six x, X() d, (6.11.11)
8 a
where X() = [, 0, 0], and compare it with that induced by an equivalent uniform distribution of
point forces over the cylindrical surface,
1 1
vi (x) = Six x, X() dS, (6.11.12)
8 2b
474 Introduction to Theoretical and Computational Fluid Dynamics
(a) (b)
3 0.4
0.3
2.8
0.2
2.6
0.1
2.4 0
v
v
0.1
2.2
0.2
2
0.3
1.8 0.4
1 0.5 0 0.5 1 1 0.5 0 0.5 1
x/a x/a
(c) (d)
20 20
15
15 10
5
f
10 0
x
f
5 10
15
0 20
0 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2
s/a s/a
Figure 6.11.2 (a) Axial and (b) radial surface velocity due to a centerline distribution (solid line) and
an equivalent surface distribution (dashed line) of point forces, for a cylinder with aspect ratio
L/a = 10. (c) Axial and (d) radial components of the traction on a cylinder with aspect ratio
L/b = 10 computed by a boundary-element method (solid lines). The dashed line in (c) represents
the prediction of slender-body theory.
where X() = [, b cos , b sin ] and dS = b d d is a dierential surface area. The surface integral
in (6.11.12) was computed by numerical integration in terms of the axisymmetric Greens function
of Stokes ow [318]. Graphs of the axial and radial velocities, vx and v , computed from (6.11.11)
and (6.11.12) are shown in Figure 6.11.2(a, b) with the solid and broken line, respectively, for a
cylinder with aspect ratio L/a = 10. The good agreement justies replacing the surface distribution
with a line distribution along the centerline. Even better agreement is obtained for higher aspect
ratios, L/a.
6.11 Slender-body theory 475
1 y2
uy (x, b, ) = Vy 2 2 1/2
+ 2 2 3/2
d
8 a [(x ) + b ] [(x ) + b ]
a
2 1 y2
+Vy b + 3 d, (6.11.13)
4 a [(x )2 + b2 ]3/2 [(x )2 + b2 ]5/2
where y = b cos and is the azimuthal angle. Introducing the dimensionless variable w = (x)/b
and rearranging the integrand, we obtain
B
1 cos2
uy (x, b, ) = Vy 2 1/2
+ dw
8 A (w + 1) (w 2 + 1)3/2
B
1 cos2
G = + cos2 B + A . (6.11.17)
B2 + 1 A2 + 1 (B 2 + 1)3/2 (A2 + 1)3/2
476 Introduction to Theoretical and Computational Fluid Dynamics
Asymptotics
In the limit as the aspect ratio a/b and thus A and B tend to innity, we obtain
uy (x, b, ) = Vy [ ln(4AB) + 2 cos2 ] + Vy (1 + 2 cos2 ). (6.11.19)
8 2
To eliminate the dependence, we set
1
= (6.11.20)
4
and obtain
L
uy (x, b) Vy [ ln(4AB) + 1 ] Vy ( 2 ln + 1 ). (6.11.21)
8 8 b
Working as in the case of axial motion, we require that uy = Vy at x = 0 and nd that
4
= . (6.11.22)
ln + 12
L
b
Integrating the distributed point forces, we nd that the drag force exerted on the cylinder is given
by
F = Vy L ey , (6.11.23)
where ey is the unit vector along the y axis. It is worth noting that the ratio of the magnitude of
the drag force in transverse and axial motions is approximately equal to two.
It remains to conrm that the x and z velocity components are zero on the surface of the
cylinder. Working as for the y velocity component, we nd that
uz (x, b, ) = Vy sin 2 + Vy sin 2, (6.11.24)
8 4
which is zero in light of (6.11.22). The magnitude of the x velocity component at the surface of the
cylinder is of order b/L, which is small as long the aspect ratio is large.
Surface distribution
To justify the ow representation in terms of a line distribution of point forces and potential dipoles
along the centerline, we consider the underlying dimensionless velocity eld
a
1 1
(a) (b)
0.23 5
0.22
0
0.21
0.2
y
F
v
0.19
0.18
10
0.17
0.16 15
1 0.5 0 0.5 1 0 0.2 0.4 0.6 0.8 1 1.2
x/a l/L
Figure 6.11.3 (a) Velocity due to a centerline distribution of point forces and dipoles (solid line), and
an equivalent surface distribution of point forces (dashed line), for cylinder aspect ratio L/a = 10.
(b) Graphs of the dimensionless cylindrical polar components of the traction around the cylinder
plotted with respect to arc length; Fx (solid lines), F (dashed lines), and F (dotted lines).
Shown in Figure 6.11.3(a) is the y component of the surface velocity induced by the centerline
point force and potential dipole distributions (solid line) and by the corresponding surface distribu-
tion of point forces (dashed line) for a cylinder with aspect ratio L/a = 10. The good agreement
provides support for the basic premise of the slender-body representation.
where l is the arc length along the trace of the cylinder contour in a meridional plane, F (l) are
dimensionless functions, and = x, , . The y and z components of the traction are
Vy
The traction coecient functions F were computed by solving an integral equation using a
highly accurate boundary-element method. Graphs of the dimensionless traction Fourier coecients
are plotted in Figure 6.11.3(b) with respect to arc length for a cylinder with moderate aspect ratio,
L/b = 10. Singularities are observed at the sharp corners at either end. The coecients F (l)
and F (l) are nearly identical over the main body of the cylinder, meaning that the streamwise
traction coecient fy is nearly uniform around the cylinder cross-section and the traction fz is
nearly zero. The horizontal line in Figure 6.11.3(b) represents the predictions of the slender-body
theory, fy = Vy /(2b). The successful comparison provides further validation for the slender-body
approximation.
where S is the Stokeslet, D is the potential dipole, X() = [, 0, 0] is a point along the centerline,
and
Vx 0
1
b = Vy , d = b2 Vy (6.11.32)
4
Vz Vz
are the densities of point force and potential dipole distributions. The force exerted on the cylinder
is given by
F = L (Vx ex + Vy ey + Vz ez ), (6.11.33)
A settling needle
As an application, we consider a slender needle settling under the action of gravity at an angle with
respect to the vertical direction, as shown in Figure 6.11.4, where = 0 corresponds to a horizontal
needle and = /2 corresponds to a vertical needle. Reversibility of Stokes ow requires that the
6.11 Slender-body theory 479
g Vt
Vn Vd
Vs
Figure 6.11.4 Illustration of a slender needle settling under the action of gravity in an innite uid.
needle maintain its initial inclination. Balancing the drag force with the weight of the needle, W ,
we nd that
where Vt is the velocity of the needle tangential to the centerline and Vn is the velocity of the needle
normal to the centerline. Setting 2, we nd that the ratio between the horizontal drift velocity,
Vd = Vt cos Vn sin , and the vertical settling velocity, Vs = Vt sin + Vn cos , is
where u is the velocity of an imposed ow, l is the arc length along the centerline, and X(l) is the
position along the centerline [177, 239, 240]. Motivated by the second expression in (6.11.32), we
extract the normal component of the dipole and set the tangential component to zero to obtain
1
d = b2 b (I t t), (6.11.37)
4
where t is the unit tangent vector along the centerline and I n t is a normal projection operator.
Numerical methods
Given the velocity along the rod centerline, equation (6.11.36) provides us with an integral equation
of the rst kind for the point force strength density, b. A numerical solution can be computed
by dividing the centerline into straight segments and assuming that the force density b takes the
480 Introduction to Theoretical and Computational Fluid Dynamics
constant value b(k) over the kth segment (e.g., [177]). Enforcing (6.11.36) at the midpoint of each
segment provides us with a system of linear equations for the segment values, b(k) .
The inuence coecients are computed by integrating along the centerline numerically or
analytically, as discussed previously in this section for a circular cylinder. The element length
must be suciently larger than the rod radius, b, otherwise the basic premise of the slender-body
approximation is violated and numerical instabilities arise. To improve the performance of the
method, the formulation can be recast in terms of an integral equation of the second kind for the
Stokeslet density distribution [200].
Integral equations
The numerical procedure can be formalized in terms of the integral equation
1 2q 1 2q
u(l0 ) = u (l0 ) + (b t t)0 ( 2 ln 1) + [b (I t t)]0 ( 2 ln + 1)
4 b 8 b
1 1
+ S(l0 , l) b(l) dl + D(l0 , l) d(l) dl, (6.11.38)
8 |ll0 |>q 4 |ll0 |>q
where q is a cut-o length allowed to depend on l0 , and the subscript 0 indicates evaluation at l0
[156]. The second term on the right-hand side originates from the rst equation in (6.11.8) and the
third term onthe right-hand side originates from (6.11.21) with L = 2q. The second term disappears
when q = 2b e 0.82b. The last integral on the right-hand side of (6.11.38) can be neglected due
to the fast decay far from the evaluation point at l0 .
Note that the integrals are performed along the entire length of the centerline. The last integral
on the right-hand side of (6.11.39) can be neglected due to the fast decay away from the evaluation
point at l0 [177].
0.8
0.6
x/d
0.4
0.2
0
0 2 4 6 8
b /(d)
y
Figure 6.11.5 Critical evaluation of slender-body theory for shear ow past a rod attached to a
plane wall. The circles represent numerical results for the dimensionless centerline force density,
by /(L), obtained by an accurate boundary-element method for a cylinder with aspect ratio
b/d = 0.001, 0.005, and 0.01. The lines represent the results of a highly accurate boundary-element
method for b/d = 0.001 (dotted line), 0.005, and 0.01.
1 2q 1 2q
u(l0 ) = u (l0 ) + (b t t)0 ( 2 ln 1) + b (I t t) 0 ( 2 ln + 1)
4 b 8 b
1 1
+ (G S)(l0 , l) b(l) dl + G(l0 , l) b(l) dl. (6.11.40)
8 |ll0 |<q 8 |ll0 |>q
In the numerical method, the rod centerline is divided into N equal elements of length l = d/N .
Applying (6.11.40) at the midpoint of each element, setting 2q = l, and requiring u(l0 ) = 0, we
obtain a system of linear equations for the linear density, b. The integrals on the right-hand side
of (6.11.40) can be computed over the union of the elements using the GaussLegendre quadrature
(Section B.6, Appendix B). Numerical results obtained with 32 elements for three rod aspect ratios,
b/d, are represented by the circles in Figure 6.11.5. The predictions of the slender-body theory shown
in this gure are compared with highly accurate results obtained by a boundary-element method
for b/d = 0.001 (dotted line), 0.005, and 0.01 [309]. The excellent agreement over the entire length
of the rod except near the tip where the slender-body theory is expected to fail corroborates the
fundamental premises of the slender-body approximation.
Problem
where Gij is a Stokes ow Greens function for the velocity and Mkj is an integral, dierential, or
integro-dierential operator acting with respect to the point x0 to generate point forces and their
derivatives expressing multipoles of the point force inside or over the surface of the particle. To
satisfy the boundary condition u = V at the particle surface, we require that
when the point x lies at the particle surface. The traction on the particle surface is given by the
corresponding expression
where the superscript t stands for translation. Combining (6.12.3) with (6.8.4) and (6.8.7), we obtain
an expression for the force exerted on the particle,
Fk = u
i (x) Mkj [Tijl (x x0 )] nl (x) dS(x), (6.12.4)
P
The boundary integral representation (6.9.9) for the ambient ow, u , at a point x0 located
inside the volume occupied by the particle takes the form
1 1
uj (x0 ) = f (x) Gij (x, x0 ) dS(x) u (x) Tijk (x, x0 ) nk (x) dS(x), (6.12.5)
8 P i 8 P i
where the unit normal vector n points outward from the particle. Operating on both sides by Mkj ,
we obtain
1
Mkj [u
j (x 0 )] = f (x) Mkj [Gij (x, x0 )] dS(x)
8 P i
1
u (x) Mkj [Tijk (x, x0 )] nk (x) dS(x). (6.12.6)
8 P i
Incorporating the boundary condition (6.12.2) in the rst integral on the right-hand side of (6.12.6),
and noting that the force due to the ambient ow over a uid parcel occupied by the particle is zero,
we nd that the rst integral is identically zero. Combining the resulting equation with (6.12.4), we
derive the generalized Faxen relation
Fk = 8Mkj [uj (x0 )], (6.12.7)
providing us with the hydrodynamic force exerted on a stationary solid particle. This expression
allows us to compute the force exerted on the particle in terms of the velocity of the incident ow
alone.
Spherical particle
In Section 6.7.1, we found that the ow due to the translation of a spherical particle in an innite
uid can be presented in terms of a point force and a potential dipole placed at the particle center,
x0 . Introducing the representation (6.7.1) with coecients given in (6.7.4) and using (6.6.9) to
express the potential dipole, D, in terms of the Stokeslet, S, we obtain
3 1
ui (x) = Vk kj ( a + a3 20 )Sij (x, x0 ), (6.12.8)
4 8
where a is the particle radius, x0 is the particle center, and the Laplacian 20 operates with respect
to x0 . Faxens law for the force then follows from (6.12.7) as
F = 6a u (x0 ) + a3 2 u (x0 ). (6.12.9)
Thus, the force exerted on a spherical particle can be found from the incident velocity and its
Laplacian at the particle center. Comparing the Faxen relation (6.12.9) with the rst relation in
(6.8.11), we conrm the mean-value theorems stated in (6.1.10).
Prolate spheroid
Using the singularity representation shown in (6.7.36), we derive Faxen relation for the force on a
prolate spheroid,
c
1 2 2
Fk = 8kj u
j X() + b 1 2 2 u j X() d, (6.12.10)
c 4 c
484 Introduction to Theoretical and Computational Fluid Dynamics
where X() = [, 0, 0] is a point along the centerline. As the eccentricity tends to zero, c 0, we
recover the Faxen law for a sphere.
and working as previously for the force, we nd that the torque exerted on the particle is given by
the generalized Faxen relation
This expression allows us to compute the torque exerted on the particle in terms of the velocity of
the incident ow alone.
Spherical particle
In Section 6.7.3, we found that the ow due to the rotation of a spherical particle in an innite uid
can be represented in terms of a couplet alone. Combining (6.7.29) with (6.6.6), we obtain
1 Sij
ui (x) = k a3 klj (x, x0 ), (6.12.13)
2 x0j
where S is the Stokeslet and x0 is the particle center. Faxens law then follows from (6.12.12) as
T = 4a3 u (x0 ). (6.12.14)
We have found that the torque exerted on a spherical particle can be computed from the curl of the
velocity representing the local vorticity of the incident ow at the particle center. Comparing the
Faxen relation (6.12.14) with that shown in (6.8.11), we conrm the mean-value theorem expressed
by identity (6.1.11). Combining (6.12.14) with (6.7.31), we nd that a freely suspended sphere
bearing zero torque rotates at an angular velocity that is equal to half the vorticity of the ambient
ow.
Prolate spheroid
Based on the singularity representation (6.7.33), we nd that the Faxen law for the axial component
of the torque on a prolate spheroid is expressed by
c
2
Tx = 4c2 (1 2 ) u X() d, (6.12.15)
c c
where X() = [, 0, 0]. As the eccentricity tends to zero, c 0, we recover the Faxen law for a
spherical particle.
6.12 Generalized Faxen laws 485
Generalized Faxen relations can be developed for higher moments of the traction on a stationary
particle. The zeroth-order moment is the force and the antisymmetric part of the rst-order moment
is the torque.
The symmetric part of the rst-order moment is the coecient of the stresslet expressing the sym-
metric part of the point force dipole in the multipole expansion (6.9.16). To compute the coecient
of the stresslet for a particle that is held stationary in an arbitrary ow, we assume that the dis-
turbance ow generated when the particle is held stationary in a purely straining ow with velocity
u = E x admits the singularity representation
where E is a symmetric matrix with zero trace. The coecient of the stresslet arising when the
particle is held stationary in an arbitrary ambient ow with velocity u is
slk = 4 L0klj [uj (x0 )] + L0lkj [uj (x0 )] , (6.12.17)
Spherical particle
The singularity representation (6.7.15) with coecients given in (6.7.21), combined with the de-
nitions of the singularities discussed in Section 6.6, show that, in the case of a spherical particle of
radius a,
5 3
1 2 2
i = Ejl a
uD 1+ a 0 Sij (x, x0 ), (6.12.18)
6 x0l 10
where S is the Stokeslet. The Faxen relation for the coecient of the stresslet then follows from
(6.12.17) as
10 1 u u
a3 (1 + a2 2 )
j
sij = + i , (6.12.19)
3 10 xi xj
To derive the Faxen relations for a particular particle shape, we require singularity representations
for the ow due to the particle translation or rotation, as well as for the disturbance ow produced
when the particle is held stationary in a straining ambient ow. The singularity representations can
be computed using numerical and approximate methods, as discussed in Section 6.7.7.
486 Introduction to Theoretical and Computational Fluid Dynamics
where the subscript ext indicates the exterior particle surface. Now we substitute uD = u u into
the right-hand side, slightly modify the left-hand side, and rearrange to obtain
(u V) f dS + V F =
t D D
u f dS
t
u f t dS, (6.12.21)
Pext Pext Pext
Next, we describe the velocity of the exterior ow ut by the singularity representation (6.12.1)
and project the boundary-integral representation (6.12.6) onto V and obtain
1
Vk Mkj [uj (x0 )] = fi (x) Vk Mkj Gij (x, x0 ) dS(x)
8 Pext
1
ui (x) Vk Mkj Tijk (x, x0 ) nk (x) dS(x), (6.12.22)
8 Pext
where is the viscosity of the ambient uid and n is the unit normal vector pointing outward from
the particle.
or
(ut V) f dS + V F = u f t dS + 8V M[u(x0 )]. (6.12.25)
Pext Pext
Since the velocities ut V and u are tangential to the interface and the tangential components of f
and f t are continuous across the interface, the domain of integration of both integrals in (6.12.25) can
be switched from the exterior to the interior side of the interface. Applying the reciprocal theorem
for the disturbance ow ut V and for the interior ow u, we then nd that the rst two integrals
in (6.12.25) are exactly the same. Rearranging, we derive a generalized Faxen relation for the force
expressed by (6.12.7).
F = a 2 + a2 20 u (x0 ). (6.12.27)
+1 +1
As tends to innity, we recover expression (6.12.9) for a spherical particle.
Problems
uniformly negligible throughout the domain of ow. In the case of interior ow, this assumption can
be validated by making the Reynolds number, dened with respect to the size of the boundaries,
suciently small. Having obtained the solution of the equations of Stokes ow, we may proceed to
study the eect of uid inertia by expanding the velocity and pressure in regular perturbation series
with respect to the Reynolds number [412].
Exterior ow
In the case of exterior ow in an innite domain, the assumption that inertial forces are negligible
throughout the domain of ow may not be valid. As an example, we consider uniform (streaming)
ow with velocity U past a stationary three-dimensional body. The Stokes-ow solution reveals
that, when the force exerted on the body is nonzero, the perturbation ow decays like 1/d, which
means that the magnitude of the left-hand side of the NavierStokes equation decays like U/d2 ,
whereas the magnitude of the right-hand side decays like /d3 , where d is the distance from the
body. This means that inertial forces dominate and the Stokes-ow approximation ceases to be valid
at distances greater than /(U ) from the body.
These diculties can be resolved by admitting that the ow near the body is governed by
the equations of Stokes ow, whereas the ow far from the body is governed by the equations
of NavierStokes or Oseen ow incorporating a linearized convective contribution, as discussed in
Section 6.13.1. The method of matched asymptotic expansions may then be applied to derive a
uniformly valid solution.
6.13.1 Oseen ow
In the case of steady streaming (uniform) ow with velocity U past a stationary body, Oseen
proposed replacing the Stokes equation with the linearized NavierStokes equation
U u = p + 2 u, (6.13.1)
subject to the condition that u tends to U far from the body [287]. The left-hand side captures the
dominant contribution of the inertial forces far from the body. When the Reynolds number dened
with respect to the distance from the body is suciently low, the left-hand side of (6.13.1) is small
compared to the right-hand side and the Oseen equation (6.13.1) describes Stokes ow.
Axisymmetric ow
In the case of axisymmetric ow, it is convenient to align the x axis with the velocity of the outer
streaming ow so that U = U ex , introduce cylindrical polar coordinates, (x, , ), and describe the
ow in terms of the Stokes stream function, . A linear vorticity transport equation corresponding to
the Oseen equation (6.13.1) arises by simplifying the general vorticity transport equation (3.12.24),
obtaining
U
= 2 E 2 ( ), (6.13.4)
x
where = / is the kinematic viscosity of the uid,
1 2 1 2 2 1
= E = 2
+ 2
, (6.13.5)
x
is the azimuthal vorticity, and E 2 is a second-order linear dierential operator dened in (2.9.17),
repeated below for convenience,
2 2 1
E2 + . (6.13.6)
x2 2
Substituting into (6.13.4) expression (6.13.5) and rearranging, we obtain
U
(E 2 ) (E 2 ) = 0. (6.13.7)
x
When U = 0, this equation describes axisymmetric Stokes ow.
U u = p + 2 u + b 3 (x x0 ), (6.13.10)
where b is a constant vector and 3 is the three-dimensional delta function. The fundamental
solution of the Oseen equation, called the Oseenlet and denoted by O, can be derived working as
in Section 6.5 for the Stokeslet [166]. The pressure due to an Oseenlet is identical to that of the
Stokeslet. The velocity is given by the counterpart of expression (6.5.16),
1 1
u= O b = ( I 2 H H ) b, (6.13.11)
8
which ensures the satisfaction of the continuity equation. The generating function H satises the
linear dierential equation
1 1
U H + 2 H = 0, (6.13.12)
4r
where = / is the kinematic viscosity of the uid. The Laplacian, Q 2 H, is the Greens
function of the linear convectiondiusion equation,
1
U Q = 2 Q + 3 (x x0 ), (6.13.13)
given by
1
Q(x, x0 ) = (x, x0 ) exp U (x x0 ) , (6.13.14)
2
where (x, x0 ) is the Greens function of the Helmholtz equation
To solve equation (6.13.17) for H, we work in hindsight and assume that H is a function of
the dimensionless variable alone. The gradient is
|U| dH x
H = ( e), (6.13.19)
2 d r
and the Laplacian is
|U|
2 d2 H x x |U| dH x
2 H = H = 2
( e) ( e) + . (6.13.20)
2 d r r 2 d r
Simplifying the expression for the Laplacian, we obtain
|U|
2 d2 H 2 |U| dH 2 |U| 1 d2 H dH
2 H = (r e x) + = + . (6.13.21)
2 d 2 r 2 d r r d 2 d
Substituting the last expression into (6.13.17) and rearranging, we derive the ordinary dierential
equation
d dH
1
= e , (6.13.22)
d d 4 |U|
whose solution is
dH 1 1 e 1
1 e
= , H() = d. (6.13.23)
d 4 |U| 4 |U| 0
Expression (6.13.19) for the gradient now becomes
1 1 e x
H = ( e). (6.13.24)
8 r
1 2H 1 2H 2H
ux = (2 H 2
) bx = ( 2 + ) bx . (6.13.28)
x y z 2
Substituting the asymptotic form (6.13.27), we nd that, near the singular point, in limit as r tends
to zero, the streamwise velocity component behaves as
1 U
ux (Sxx ) bx , (6.13.29)
8 2
where Sxx is the Stokeslet. The second term on the right-hand side expresses an induced uniform
ow with velocity
1 U
ux = bx . (6.13.30)
8 2
In the case of streaming ow past a sphere of radius a, we set bx = 6aU from the Stokes ow
solution, and compute the induced velocity
3 aU 3
ux = Re, (6.13.31)
8 16
where Re = 2aU/ is the Reynolds number based on the sphere diameter, 2a. Using Stokess law,
we nd that the force exerted on the sphere is given by
3
Fx = 6a (U + ux ) = 6aU (1 + Re), (6.13.32)
16
yielding the drag coecient
F 12 3
cD 2 2
= (1 + Re). (6.13.33)
U a Re 16
The second term on the right-hand side expresses the rst inertial correction to Stokess law. The
procedure can be extended using the method of matched asymptotic expansions to derive further
corrections to the drag force [412].
we obtain
1
E2 = 0, (6.13.35)
2 2
where X = Ux x/ is a dimensionless variable and = /|Ux | is a length scale. Working in hindsight,
we introduce spherical polar coordinates, (r, , ), set
(a) (b)
5 5
4 4
3 3
2 2
1 1
0 0
y
y
1 1
2 2
3 3
4 4
5 5
5 0 5 5 0 5
x x
Figure 6.13.1 Approximate streamline pattern of Oseen ow (a) past a sphere and (b) due to the
translation of a sphere at Re = 2.
bx = 6aUx , dx = a3 Ux , (6.13.42)
and obtain
3 Re r
1 a
= + a2 U x (1 + cos ) 1 exp[ (1 cos ) ] + sin2 , (6.13.43)
Re 4 a 4 r
where Re = 2aUx / is the Reynolds number dened with respect to the sphere diameter, 2a.
Near the sphere, the ratio r/a is of order unity and the argument of the exponential term is small.
Expanding the exponential term in a Taylor series, we obtain the stream function for Stokes ow,
supplemented by a small correction whose magnitude is proportional to the Reynolds number. Unlike
in Stokes ow, the streamline pattern in Oseen ow is no longer symmetric with respect to the mid-
plane of the sphere, as shown in Figure 6.13.1(a) for Re = 2.
1
(x, x0 ) = K0 (cr), (6.13.50)
2
and K0 is the zeroth-order modied Bessel function of the second kind. Thus,
1
Q 2 H = K0 (cr) exp(c e x), (6.13.51)
2
where e = U/|U| is the unit vector in the direction of the streaming ow, x = x x0 , and r = |x|.
To simplify the analysis, we may assume that the streaming ow occurs along the x axis and
set U = U ex , e = ex , and e x = x x x0 . The solution of (6.13.51) is given by the indenite
integral
1
H= [K0 (cr) ecx + ln(cr) ] dx. (6.13.52)
4c
The two-dimensional Oseenlet may now be computed from the expression
O = 4 (I 2 H H) = 4 (Q I H). (6.13.53)
496 Introduction to Theoretical and Computational Fluid Dynamics
For a point force pointing in the direction of the outer streaming ow along the x axis, we obtain
1 x 1 y
Oxx = K0 (cr) ecx + K1 (cr) ec x , Oyx = K1 (cr) ec x . (6.13.54)
cr r cr r
For a point force pointing normal to the outer streaming ow, we obtain
1 y 1 x
Oyx = K1 (cr) ec x , Oyy = K0 (cr) ecx K1 (cr) ec x . (6.13.55)
cr r cr r
We note that Oxx + Oyy = 4Q, as required. These expressions can also be derived by applying
Fourier transforms, as discussed in Section 6.5 for the Stokeslet ([230], Problem 6.13.4).
where E = 0.577215665 is Eulers constant, S is the two-dimensional Stokeslet, and the dots
denote decaying terms. We observe a leading-order Stokes ow solution complemented by streaming
ow in the direction of the point force. In Section 6.14, these expressions will be used to describe
streaming ow past a circular cylinder.
Problems
Now we enforce the boundary condition u = V at the surface of the cylinder located at r = a,
and obtain a system of two equations for three unknowns, b, d, and c,
1 a 1
ln b d + c = V, a2 b + 4 d = 0, (6.14.3)
4 2a2
where V is the cylinder velocity. The solution is
1 1 a 1
b = 2 V, d = a2 V, c = [1 + (ln ) ] V, (6.14.4)
2 2 2
where is an indeterminate dimensionless constant. Substituting these expressions into (6.14.2),
we derive an explicit expression for the velocity eld,
r 1 1 a2 a2 xi xj
ui (x) = ( ln + 2
) + ( 1 2 ) 2 Vj + Vi . (6.14.5)
2 a 2 2 r 2 r r
498 Introduction to Theoretical and Computational Fluid Dynamics
Note that the length has disappeared after these substitutions. The force per unit length exerted
on the cylinder is
F = s = 2 V. (6.14.6)
Surface traction
The hydrodynamic traction exerted on the cylinder can be computed from the strength of the
singularities using the expressions given in Tables 6.5.3 and 6.5.4,
1 S 1 D
fi = T bj nk + T dj nk . (6.14.7)
4 ijk 2 ijk
We nd that
1 xi xj xk 2 jk xi + ki xj + ij xk xi xj xk
fi = bj nk + ( 4 ) dj nk . (6.14.8)
a4 a4 a6
Setting nk = xk /a, xk xk = a2 , and simplifying, we obtain
1 xi xj 2 2xi xj + ij a2 xi xj
fi = 3
bj + ( 4 5 ) dj . (6.14.9)
a a5 a
Substituting the expressions for b and d given in (6.14.4) and simplifying, we nd that
xi xj 2 xi xj + ij a2
fi = 2 V j ( ) Vj , (6.14.10)
a3 a3
yielding
fi = Vi . (6.14.11)
a
We have found that, as in the case of a translating sphere, the traction is a constant vector oriented
in the direction of translation. This is a unique and rather fortuitous feature of the circular geometry.
Integral representation
We have seen that the potential dipole can be expressed in terms of the Laplacian of the Stokeslet,
1
D(x, x0 ) = 20 S(x, x0 ). (6.14.12)
2
Using the second equation in (6.14.3) to write
a2
d= b, (6.14.13)
4
we recast the singularity representation (6.14.1) into the form
1 a2
ui (x) = Sij (x, x0 ) sj + 2 Sij (x, x0 ) sj + ci . (6.14.14)
4 16 0
6.14 Flow past a circular cylinder 499
Applying the mean-value theorem for biharmonic functions (2.5.7) for the Stokeslet, we obtain
1 1
Sij (x, ) dl() = Sij (x, x0 ) + a2 20 Sij (x, x0 ), (6.14.15)
2a Cylinder 4
where the point x lies outside the cylinder and the point lies at the surface of the cylinder.
Combining the last two equations, we nd that
ui (x) = Sij (x, ) dl() Vj + ci . (6.14.16)
4a Cylinder
The integral on the right-hand side represents a uniform distribution of two-dimensional point forces
with strength density fj = Vj /a around the cylinder contour. This interpretation is an essential
aspect of the slender-body theory discussed in Section 6.11.
1 r xi xj 1
uouter
i (x) ij ln + 2 bj + ( 2 ln 2 ) bi , (6.14.17)
4 r 4
where = /|V| and b = [bx , 0]. To satisfy the force balance, the strength of the point force, b,
must be the same as that occurring in the inner Stokes ow solution.
6.14.3 Matching
In Section 6.14.1, we saw that the inner Stokes ow consists of a point force, a potential dipole,
and a uniform ow. The outer limit of the inner ow computed from (6.14.2) by neglecting the
fast-decaying dipole is
1 r xi xj
uinner
i (x) (ij ln + 2 ) bj + ci . (6.14.18)
4 r
Comparing (6.14.17) with (6.14.18) and requiring functional consistency between the inner and outer
limits, we set
1
c= (2 ln 2 ) b. (6.14.19)
4
Substituting into this equation the expressions for b and c given in (6.14.4), we obtain
Re 1
1+ ln = (2 ln 2 ), (6.14.20)
2 2 2 2
500 Introduction to Theoretical and Computational Fluid Dynamics
where Re = D|V|/ is the Reynolds number dened with respect to the cylinder diameter, D = 2a.
Solving for the coecient , we obtain [219]
2 2
= . (6.14.21)
8 1 7.4
ln + ln
Re 2 Re
Substituting (6.14.21) into (6.14.6), we obtain an expression for the drag force,
2
F 2 V.
7.4 (6.14.22)
ln
Re
As Re tends to zero, the drag force diverges, and this underscores the ill-posedness of two-dimensional
streaming Stokes ow.
The analysis can formally extended using the method of matched asymptotic expansions to
derive further terms in the expression for the drag force [412].
Problems
Flows governed by the unsteady Stokes equation include those generated by hydrodynamic
braking and by the translational or rotational vibrations of rigid and liquid particles suspended in
6.15 Unsteady Stokes ow 501
a viscous uid, provided that the amplitude of the oscillation is small compared to the particle size.
Other examples include oscillatory ow past a suspended particle due to the passage of a sound wave,
and the ow due to the transient motion of a particle in a quiescent, accelerating, or decelerating
ambient uid. Unsteady Stokes ow occurs in the inner ear due to the small-amplitude vibrations
of the eardrum transmitted through the ossicular bone chain.
2 p = 0, (6.15.2)
as in the case of steady Stokes ow. Taking the curl of (6.15.1), we nd that the Cartesian compo-
nents of the vorticity evolve according to the unsteady heat conduction equation,
= 2 , (6.15.3)
t
where = / is the kinematic viscosity of the uid.
Any irrotational velocity eld described by a harmonic potential, u = , satises the equa-
tions of unsteady Stokes ow. The associated hydrodynamic pressure is given by the simplied
Bernoulli equation
p = + c(t), (6.15.4)
t
where c(t) is a time-dependent function. However, because a general potential ow is not generally
able to satisfy both the no-penetration and the no-slip conditions over a solid boundary, it must be
complemented by a rotational ow.
Unsteady Stokes ow shares many of the properties of steady Stokes ow discussed previously
in this chapter, including reversibility, the reciprocal identity, Faxens laws, and uniqueness subject
to specied boundary conditions for the velocity (e.g., [217, 306]).
where s is a complex variable. Taking the Laplace transform of the unsteady Stokes equation (6.15.1)
and the continuity equation, u = 0, and using the properties of the Laplace transform, we obtain
Brinkmans equation and the continuity equation in the Laplace domain,
s u = p + 2 u, u = 0. (6.15.6)
An additional term appears on the right-hand side of the Stokes equation in the presence of a
distributed or localized body force.
Having computed the solution subject to appropriate boundary conditions, we recover the
physical variables in the time domain in terms of the Bromwich integral in the complex s plane,
u +i u
p (x, t) = 1 p (x, s) est ds, (6.15.7)
2i i
where i is the imaginary unit and is a suciently large real positive number chosen so that, as we
move upward, all singularities of the Laplace transformed functions lie on the left of the integration
path.
6.15.3 Oscillatory ow
The linearity of the equations of unsteady Stokes ow requires that the velocity, vorticity, hydro-
dynamic pressure, and hydrodynamic stress of an oscillatory ow with angular frequency exhibit
identical harmonic dependencies. We may thus write
u u
p = p exp(it), (6.15.8)
where i is the imaginary unit and a bar denotes the amplitude of the underlying variable. Substituting
these expressions into (6.15.1) and into the continuity equation, u = 0, we obtain the equations
of oscillatory Stokes ow,
i u = p + 2 u, u = 0. (6.15.9)
It is worth noting the similarity with the Laplace transformed equations (6.15.6), subject to the
substitution s i.
where b is a constant amplitude. The induced velocity and pressure elds are computed by solving
the equation
i u = p + 2 u + b 3 (x x0 ), (6.15.11)
1
u= ( I 2 H H) b. (6.15.12)
1
2 H + i H = 0, (6.15.13)
4r
where r = |x x0 | is the distance from the point force. Taking the Laplacian of (6.15.13), we nd
that the Laplacian of the generating function, Q 2 H, is the Greens function of the Helmholtz
equation
2 Q + i Q + 3 (x x0 ) = 0, (6.15.14)
which is given by
1 R
Q= e , (6.15.15)
4r
where
1/2 1i r
Rr i = (6.15.16)
2
is a kinematic diusion length. Note that the real part of scaled distance R is positive to ensure
exponential decay.
To compute the function H, we introduce spherical polar coordinates with origin at the point
force, set
1 d 2 dH
1 R
Q 2 H = 2
r = e (6.15.18)
r dr dr 4r
and nd that
dH 1
= 2
1 (1 + R) eR . (6.15.19)
dr 4R
504 Introduction to Theoretical and Computational Fluid Dynamics
Integrating, we obtain
1
1/2 eR 1 1 1 + i eR 1
H= i = . (6.15.20)
4 R 4 2 R
Oscillatory Stokeslet
Substituting (6.15.20) into (6.15.12), we derive the velocity eld due to an oscillating point force in
the standard form
1
ui (x, x0 ) = Sij (x) bj , (6.15.21)
8
1 dH 1 1 dH d2 H
S = 8 Q I+ 2 2
x x , (6.15.23)
r dr r r dr dr
where I is the identity matrix. Making substitutions, we obtain
ij xi xj
Sij (x) = A(R) + 3 C(R), (6.15.24)
r r
where
1 1
2 3 3
6
A(R) = 2 1 + + 2 eR 2 , C(R) = 2 1 + + 2 eR + 2 , (6.15.25)
R R R R R R
are frequency-dependent functions. One may conrm that A(0) = C(0) = 1, which shows that, at
small frequencies or close to the point force, the unsteady Stokeslet reduces to the regular Stokeslet
for Stokes ow.
The vorticity, pressure, and stress elds associated with the oscillatory Stokeslet are given by
(6.15.8) with
1 1 1
i = ij bj , p= P j bj , ik = Tijk bj , (6.15.26)
8 8 8
where xl xj
ij = 2 ijk (R + 1) eR , Pj = 2 3 , (6.15.27)
r3 r
2 2 xi xj xk
Tijk = 3 (ij xk + kj xi ) (R + 1) eR C 3 ik xj (1 C) 2 5 C 2(R + 1)eR ,
r r r5
and the function C is given in (6.15.25). Note that Tijk is symmetric with respect to the indices i
andk.
6.15 Unsteady Stokes ow 505
1 ij xi xj
Performing a Taylor series expansion, we nd that K(0) = 0 and L(0) = 6, which is consistent with
equation (6.5.23) for Stokes ow. The force exerted on the spherical surface is found by integrating
the traction,
1 1
F= (3 K + L) b = [ 2(R + 1) eR + 1 ] b. (6.15.30)
6 3
We note that the force exerted on a small sphere with innitesimal radius (R = 0) is equal to b,
whereas the force exerted on a sphere with innite radius (R ) is equal to 13 b. The dierence
between these two values is equal to the rate of change of momentum of the uid surrounding the
point force.
1 1+i 1 1 1 3
H= 1 + R R2 R + . (6.15.31)
4 2 2 6 24
Simplifying, we obtain
1 1+i r 1 1 2
H= + 1 R+ R + . (6.15.32)
4 2 8 3 12
The constant term does not generate a ow; the linear term generates a steady Stokeslet, the
quadratic term generates a uniform ow; further terms generate a sequence of unclassied singular-
ities. It is interesting to recall that the Oseenlet also generates a uniform ow in addition to the
Stokeslet near the point force. The corresponding expansion for the oscillating Stokeslet is
S(x) = S (0) (x) + S (1) (x) + R2 S (2) (x) + R3 S (3) (x) + , (6.15.33)
All constituent tensors, S (n) , have dimensions of inverse length. As noted, the tensor S (1) represents
uniform streaming ow.
For example, applying the boundary-integral representation (6.15.36) for streaming oscillatory
ow with uniform velocity u = U and pressure p = i U x, and setting f = pn, we derive the
identity
1
ij = i xi nk (x) Skj (x, x0 ) dS(x) + Tijk (x, x0 ) nk (x) dS(x), (6.15.37)
8 D 8 D
where = / is the kinematic viscosity of the uid.
velocity is U = d, and the characteristic length is the particle size, L, where is the angular
frequency and d is the amplitude of the vibration. Requiring that U L/ Re L2 /T , we
obtain T L/U or d L, which shows that the equations of unsteady Stokes ow apply when
the amplitude of vibration is much smaller than the particle size. Enforcing the boundary condi-
tion of rigid-body motion at the mean position of the particle instead of the instantaneous position
introduces an error that is comparable to d/L, which is negligible as long as the amplitude of the
oscillation is small compared to the particle size. The boundary-integral formulation provides us
with the integral representation (6.15.36), where the integration domain, D, is the mean particle
surface.
Translational vibrations
In the case of translational vibrations with velocity V = V exp(it), we set u = V at the mean
position of the particle surface, P , and obtain the integral representation
1 1
uj (x0 ) = fi (x) Sij (x, x0 ) dS(x) + Vi Tijk (x, x0 ) nk (x) dS(x), (6.15.38)
8 P 8 P
for a point x0 lying in the ambient uid. Combining this representation with identity (6.15.37)
written with = 0 to eliminate the double-layer potential, we derive a single-layer representation,
1
uj (x0 ) = fi (x) + i (V x) ni Sij (x, x0 ) dS(x). (6.15.39)
8 P
Now we move the point x0 to the particle surface and derive an integral equation that can be solved
by numerical methods (e.g., [304, 243]). A similar reduction of the boundary-integral representation
to a single-layer potential is not possible in the case of rotational oscillations.
Low-frequency asymptotics
To study the behavior of the ow at low angular velocities in the general case of translational and
rotational vibrations, we introduce the dimensionless complex frequency
1/2 1i L
i L= , (6.15.40)
2
where L is a particle length scale. Next, we expand the velocity and traction in asymptotic series
with respect to ,
are the translational and rotational velocities about the center of rotation, xc . Substituting these
expansions along with (6.15.33) into (6.15.36), and collecting terms of zeroth and rst order in ,
we obtain the integral equations
(0) 1 (0) (0)
uj (x0 ) = f (x) Sij (x, x0 ) dS(x) nk (x) dS(x) (6.15.44)
8 P i
and
1 (0) 1 (1) (0)
F = fi (x) Sij (x, x0 ) dS(x). (6.15.45)
6a j 8 P
The double-layer potential has disappeared from (6.15.44) due to the boundary condition or rigid-
body motion.
The solution of the zeroth-order integral equation (6.15.44) can be expressed in the form
where Rt and Rr are, respectively, the steady translational and rotary surface traction resistance
matrices introduced in (6.8.4) and (6.8.8). The steady force and torque exerted on the particle can
be expressed in terms of the resistance matrices introduced in (6.8.15),
Comparing the last four equations, we conclude that the rst correction to the traction, force, and
torque exerted on the particle are given by
(1) t t
f R R
F = 1 X F(0) = X (X V + P ). (6.15.48)
6a 6a
T ZT ZT
We have found that the rst-order correction can be computed from the resistance matrices for
steady Stokes ow.
1
Q= ( ker0 i kei0 ), (6.15.51)
2
where r = |x x0 |, = /, = r/, and ker0 , kei0 are Kelvin functions (e.g., [2], p. 379). To
compute the solution, we express the Laplacian in plane polar coordinates,
1 d dH
2 H = r = Q, (6.15.52)
r dr dr
and integrate once to nd that
r
dH 1
= ( ker0 i kei0 ) d. (6.15.53)
dr 2r 0
Abramowitz & Stegun ([2], p. 380) provide us with the indenite integrals
ker0 () d = ker1 kei1 = kei0 (6.15.54)
2
and
kei0 () d = kei1 + ker1 = ker0 . (6.15.55)
2
dH
= ( kei0 + iker0 ). (6.15.56)
dr 2
The oscillatory Stokeslet is given by
dH x
S = 4(I 2 H H) = 4 I Q , (6.15.57)
dr r
which can be recast into the form
C()
S = A() I + x x, (6.15.58)
r2
where
1 dH 1 1
A() = 4 (G ) = 2 (ker0 kei0 ) 2 i (kei0 + ker0 ),
r dr
(6.15.59)
1 dH d2 H 2 2
C() = 4 ( 2
) = 2 (ker0 + kei0 ) + 2 i (kei0 + ker0 )
r dr dr
Oscillatory ow
The Stokes ow solution for oscillatory ow, is given by the real or imaginary part of the right-hand
side of (6.15.8). Selecting the real part, we introduce a regular asymptotic expansion for the velocity
with respect to Re,
1 (0) it
u(x, t) = (u e + u(0) eit ) + Re u(1) (x, t) + , (6.15.60)
2
where the superscript (0) denotes the Stokes ow solution, the superscript (1) denotes the rst
inertial correction, an asterisk designates the complex conjugate, and the three dots denote high-
order corrections with respect to Re, The nonlinear convection term on the left-hand side of the
equation of motion becomes
1
u u = F (u(0) ) + O(Re), (6.15.61)
4
where
F (u(0) ) u(0) u(0) e2it + u(0) u(0) e2it + u(0) u(0) + u(0) u(0) . (6.15.62)
Substituting expressions (6.15.60) and (6.15.62) into the NavierStokes equation, nondimensional-
izing all terms as discussed in Section 3.10, retaining terms of rst order in Re, and reverting to
physical variables, we obtain a forced unsteady Stokes equation
u(1) 1
= p(1) + 2 u(1) F , (6.15.63)
t 4
where p(1) is the rst inertial correction to the pressure. We observe that the nonlinear acceleration
term produces second harmonics of the fundamental frequency, , represented by the terms inside
the square brackets on the right-hand side of (6.15.62). The time-independent term enclosed by
the parentheses on the right-hand side of (6.15.62) is responsible for the onset of inertial steady
streaming superimposed on the underlying oscillatory ow.
Steady streaming
To describe the steady streaming ow, we integrate equation (6.15.63) over one period to eliminate
the time dependence, and derive a forced Stokes equation,
1
p(1) + 2 u(1) u(0) u(0) + u(0) u(0) = 0, (6.15.64)
4
where a bar designates a time average. The problem is reduced to solving the forced Stokes ow equa-
tion (6.15.64) together with the continuity equation, subject to appropriate boundary conditions.
6.16 Singularity methods for oscillatory ow 511
At high frequencies, the forcing term is conned inside a Stokes layer surrounding the boundaries of
the ow, and a solution can be found using the boundary-layer approximation (e.g., [24], p. 358).
Problems
Dierentiating the point source with respect to the singular point, x0 , we obtain a sequence
of derivative singularities expressing irrotational ow. The rst two are the potential dipole and the
512 Introduction to Theoretical and Computational Fluid Dynamics
potential quadruple. The velocity elds are shown in Table 6.6.1 and the corresponding pressure
elds are found by dierentiating the pressure of the point source.
The velocity and stress elds due to a symmetric Stokeslet quadruple with strength q are
given by
1 SQ SQ
ui = S qj , ik = T qj , (6.16.4)
4 ij 4 ijk
where
x + x xj
(6 + 6R + 3R2 + R3 ) + 2ik 5 (3 + 3R + R2 )
SQ ij k jk i
Tijk =
r5 r
xi xj xk
2 (15 + 15R + 6R + R ) eR .
2 3
(6.16.5)
r7
The traction exerted on a spherical surface of radius r centered at the singular point is
ij xi xj
Interior ow
The velocity, pressure, and stress elds due to an oscillating Stokeson with strength s = s exp(it)
is given by [303]
ui = Sij
N
sj , p = PjN sj , ik = Tijk
N
sj . (6.16.12)
The pressure vector is the same as that of the steady Stokeson, PjN = 10 xj . The velocity tensor is
given by
Sij
N
= ij 2r 2 Q(R) xi xj W(R), (6.16.13)
F = b iVp V, (6.16.15)
where Vp is the particle volume and V is the particle velocity of translation [303].
514 Introduction to Theoretical and Computational Fluid Dynamics
1
Q(R) = 5
2 R3 30(1 + R2 ) sinh R + 30R cosh R
4R
15
W(R) = 5
(3 + R3 ) sinh R 3R cosh R
R
15
I(R) = 5
R(6 + R2 ) cosh R 3(2 + R2 ) sinh R
R
15
J (R) = 10 5
R(18 + R2 ) cosh R (7R2 + 18) sinh R
R
Table 6.16.1 Functions related to the velocity eld and force exerted on a spherical surface due to an
oscillating Stokeson in interior ow.
e
c = 8a3 , (6.16.16)
1+
where 2 = ia2 / is a dimensionless frequency. Using (6.16.11), we nd that the torque exerted
on the sphere is
1 2 8 2
T = (1 + + ) e c = a3 3 + . (6.16.17)
3 3 1+
As 0, we recover our earlier results for Stokes ow.
Substituting the expressions for the singularities and enforcing the boundary condition u = V at
r = a, we nd that
1 2 1 1
b = 6a(1 + + ) V, q = 6a3 (e 1 2 ) V, (6.16.19)
2 3
where 2 = ia2 / is a dimensionless frequency. As 0, we obtain expressions (6.7.4) for
steady Stokes ow.
6.17 Unsteady Stokes ow due to a point force 515
Using (6.15.30) and (6.16.8) or (6.16.15) along with the coecients (6.16.19), we obtain the
force exerted on the sphere,
1 i 2
F = [2e ( + 1) + 1] b e (1 + ) q = b iVs V, (6.16.20)
3 3
where Vs = 4 3
3 a is the volume of the sphere. Making substituting, we recover the BoussinesqBasset
formula (e.g., [413])
1 2
F = 6a (1 + + )V. (6.16.21)
9
It is remarkable that the expression for the force is a binomial in the complex frequency parameter
. The three terms in the parentheses on the right-hand side of (6.16.21) represent, respectively,
the steady Stokes drag force, the unsteady BoussinesqBasset force, and an acceleration reaction
associated with the virtual mass in potential ow. The quadratic dependence is true only for the
spherical shape [303].
Problems
The Laplacian of the generating function, Q 2 H, is the Greens function of the unsteady
diusion equation, satisfying
1 Q
= 2 Q + 3 (x x0 ) (t t0 ), (6.17.4)
t
given by
r2
Q(r, t) = exp (6.17.5)
(4 t)3/2 4 t
(see also Section 12.2.1). Solving the Poisson equation 2 H = Q in spherical polar coordinates for
a radial eld,
1 2 H
2 H = 2 r = Q, (6.17.6)
r r r
we nd that
1
1/2 r 2
H(r, t) = 1 exp d. (6.17.7)
4r t 0 4 t
In terms of the error function,
w
2 2
erf(w) e d, (6.17.8)
0
we obtain
1
1/2 r
H(r, t) = r ( t)1/2 erf . (6.17.9)
4r t (4 t)1/2
Straightforward dierentiation according to (6.17.2) produces the evolving velocity eld.
6.17 Unsteady Stokes ow due to a point force 517
Long-time asymptotics
The behavior of the ow at long times can be found by expanding the integrand in (6.17.7) in a
Taylor series with respect to the argument of the exponential, nding
1
1/2 r 2 1 2
2
H(r, t) = + d. (6.17.10)
4r t 0 4 t 2 4 t
Performing the integration, we obtain
1
1/2 1 r2 1 r4
H(r, t) = + . (6.17.11)
4 t 3 4 t 10 (4 t)2
Substituting this expansion into (6.17.2), we obtain
1
1/2 4 1 r3
u(x, t) = I6 2
S (3) + b, (6.17.12)
4 t 3 4 t (4 t)
where I is the identity matrix and the tensor S (3) is dened in (6.15.34). This expression nds ap-
plications in the computation of the long-time decay of the angular velocity autocorrelation function
of small particles executing Brownian motion [183].
Laplace transform
The ow due to a localized impulse can also be derived by applying the Laplace transform to obtain
equations (6.15.6) with a forcing term in the equation of motion due to the impulse,
s u = p + 2 u + b 3 (x x0 ), u = 0. (6.17.13)
The solution is
1
ui (x, s) = Sij (x, s) bj , (6.17.14)
8
where x = x x0 and S is the unsteady Stokeslet tensor given in (6.15.24) with R = r s/.
Inversion yields the results produced earlier in this section using th method of Greens functions.
To compute the long time behavior of the ow, we use expansion (6.15.33) for the oscillatory
Stokeslet at small s, nding
1 (0) 4 1
and /( )1/2 is a dimensionless integration variable. Evaluating the integral for small values
of , we obtain
1 1 3
I= + + . (6.17.19)
2 48
Substituting this expansion into (6.17.17) and performing the integration, we obtain
r 1 1
H = 3/2
R+ R3 + , (6.17.20)
8 3 120
where R r/[(t tstart )]1/2 , which is consistent with the expression derived in Section 6.5.5 for
the Stokeslet.
In the case of a point force moving with constant velocity, V, the position of the point force
is x0 (t ) = x0 (t) V . Making substitutions, we obtain
ttstart (x x (t) V )2
0
Q(x, t) = exp dt. (6.17.22)
0 (4 )3/2 4
Letting the upper integration limit tend to innity and performing the integration, we obtain
1 |V|
Q = exp (r e x) , (6.17.23)
4r 2
where x = x x0 (t), r = |x|, and e = V/|V| is the unit vector in the direction of translation.
Expression (6.17.23) is consistent with the Laplacian of the generating function associated with the
Oseenlet, given in (6.13.17).
6.18 Unsteady particle motion 519
Problems
= 6a (1 + + 1 2 )V,
F (6.18.1)
9
where 2 = sa2 /. Taking the inverse Laplace transform, we obtain
t
dV
d 1 dV
F(t) = 6aV(t) 6a2 Vs , (6.18.2)
0 dt t 2 dt
where Vs = 4 3
3 a is the volume of the sphere. The three terms on the right-hand side of (6.18.2)
represent the Stokes drag force, the BoussinesqBasset viscous memory force, and an acceleration
reaction. The inverse square root kernel of the BoussinesqBasset memory integral reects the
diusion of vorticity away from the particle surface.
Equation (6.18.2) can be used to derive an integro-dierential equation describing the grav-
itational settling of a spherical particle released from rest. Using Newtons second law of motion
incorporating the hydrodynamic drag force, the buoyancy force, and the particle weight, we obtain
t
1 dV dV
d
(s + ) Vs = 6aV(t) 6a2 + (s ) Vs g, (6.18.3)
2 dt 0 dt t
where s is the density of the sphere. Equation (6.18.3) can be recast as a second-order ordinary
dierential equation in time for V(t), which can be integrated by standard numerical methods ([89],
520 Introduction to Theoretical and Computational Fluid Dynamics
p. 288; [437], p. 376). It should be pointed out that, because the boundary condition is applied at
a xed position, equations (6.18.1)(6.18.3) are strictly valid only for small travel distances, much
less than the particle size.
Equations similar to (6.18.2) and (6.18.3) have been developed for spherical bubbles and drops
and nonspherical rigid particles [89].
= 8 a3 3 +
T = 8 a3 3 +
1
s
, (6.18.4)
3 1+ 3 /a2 /a2 + s
where 2 = sa2 /. Assuming that the sphere has started rotating at t = 0 in a quiescent uid, we
obtain the torque by taking the inverse Laplace transform denoted by L1 . Resorting to tables of
the Laplace transform, we read
1
1 1
L1 = L1 = et erfc( t), (6.18.5)
s+ s s + s
T = 8 a3 3 +
1 1
1
s , (6.18.6)
3 /a2 s s /a2 + s
and nd that
t
8 d
1
T(t) = a3 3 (t) +
e(t ) erc (t ) d , (6.18.7)
3 0 dt (t )
where = /a2 is the inverse of the viscous time scale [303]. We note the presence of an instanta-
neous viscous torque together with a viscous history torque. There is no added mass term.
Problem
Computer Problem
Under certain conditions, the distance across which the vorticity penetrates the uid from
the boundaries is small compared to the overall size of the boundaries, and the bulk of the ow is
nearly irrotational. This occurs when the rate of diusion of vorticity into the ow is comparable
to the rate of convection of vorticity by tangential uid motion along the boundary. Balancing the
orders of magnitude of the rate of convection and diusion of vorticity yields the formal requirement
that the Reynolds number, Re, dened with respect to the typical size of the boundaries, should
be suciently high, Re 1. When this condition is met, vorticity is convected along thin layers
wrapping around the boundaries, and is then channeled into slender wakes or deposited into regions
of rotational ow, commonly called vortices or regions of separated ow. The point where a vortex
layer detaches from a boundary and enters the bulk of the ow denes the position where a boundary
layer is said to separate. The relevant physical mechanisms and the mathematical description of
these processes will be discussed in Chapter 8.
Examples of nearly irrotational ows include high-Reynolds-number ows past an aircraft wing
or streamlined ground vehicles. Under normal operating conditions, the vorticity is conned inside
thin boundary layers lining the wing or the surface of a vehicle, as well as inside wakes developing
behind these bodies. Another example is the ow generated by the propagation of surface waves in
the ocean, which is irrotational everywhere except inside a thin boundary layer containing a small
amount of vorticity along the free surface.
522
7.1 Equations and computation of irrotational ow 523
Prandtl proposed decoupling the irrotational ow far from the boundary layers from the ow
inside the boundary layers. In the rst step, a potential-ow problem is solved assuming that the
domain of irrotational ow extends all the way up to the boundaries and neglecting the presence
of wakes and regions of recirculating uid. In the second step, the potential-ow solution is used
to study the development and instability of boundary layers, subject to appropriate simplifying
assumptions. In this chapter and in Chapter 10, we discuss the properties and computation of the
outer potential ow. In Chapters 8 and 9, we discuss the structure and stability of the boundary-layer
ow.
u = , (7.1.1)
where u is the velocity. Taking the curl of (7.1.1) and remembering that the curl of the gradient
of a twice dierentiable function is identically zero, we nd that the vorticity vanishes and conrm
that a potential ow is also an irrotational ow. In Section 2.1, we saw that the converse is also
true, that is, any irrotational ow can be described in terms of a single- or multi-valued velocity
potential, . The advantages of introducing the potential become evident by observing that, instead
of computing the three components of the velocity, we only need to compute the scalar potential
function of an irrotational ow.
Incompressibility condition
The continuity equation for an incompressible uid requires that the velocity eld be solenoidal,
u = 0, and the potential be a harmonic function,
2 = 0. (7.1.2)
Laplaces equation (7.1.2) can be solved by a variety of analytical and numerical methods subject
to one scalar boundary condition over each boundary. Since the normal component of the velocity
at the edge of a boundary layer over an impermeable stationary body is small, we must enforce
the no-penetration condition, n = 0. The irrotational ow computed subject to this condition
exhibits a nonzero tangential velocity, which amounts to a surface of discontinuity identied as a
vortex sheet. This apparent jump is a macroscopic representation of a vortex layer or boundary
layer whose thickness is small compared to the macroscopic length scale of the ow.
Bernoulli equation
Since the velocity eld of an irrotational ow can be computed without reference to the equation of
motion, kinematics and dynamics are decoupled. The equation of motion provides us with a rst-
order linear partial dierential equation for the pressure, which can be integrated to yield Bernoullis
equation (3.9.22), repeated for convenience,
+ B = c(t), (7.1.3)
t
524 Introduction to Theoretical and Computational Fluid Dynamics
where
1 p
B uu+ gx (7.1.4)
2
is the Bernoulli function and c(t) is a specied function of time. We have assumed that the density
of the uid is uniform throughout the domain of ow.
It is worth emphasizing that an irrotational ow and the associated pressure eld computed
using Bernoullis equation constitute an exact solution of the full NavierStokes equation with the
viscous force included. However, with some fortuitous exceptions, the solution cannot satisfy more
than one scalar boundary condition. One exception is the two-dimensional ow generated by the
steady rotation of a circular cylinder around its center, which is identical to the irrotational circu-
latory ow due to a point vortex. Another exception is the ow generated by the radial expansion
or contraction of a spherical bubble which is identical to the irrotational ow due to a point source,
as discussed in Section 7.1.3.
where n is the normal unit vector pointing into the ow. Strictly speaking, these equations provide
us with the force and torque exerted on a uid surface around the edge of boundary layers, subject
to the assumption that viscous stresses make negligible contributions. These predictions must be
corrected taking into consideration the viscous stresses near the boundaries and the pressure drop
occurring across boundary layers, as discussed in Chapter 8.
Departing from (3.4.14) and (3.4.15), noting that the rate-of-deformation tensor is the same as
the velocity gradient tensor in an irrotational ow, using the continuity equation, and applying the
divergence theorem, we nd that the rate of viscous dissipation in an irrotational ow with uniform
viscosity is given by
dV = 2 () : () = n (u u) dS, (7.1.6)
F low F low B
where n is the normal unit vector pointing into the ow and B denotes the boundaries. The right-
hand side of (7.1.6) expresses the rate of viscous dissipation in terms of the derivatives of the square
of the magnitude of the velocity normal to the boundaries.
7.1 Equations and computation of irrotational ow 525
Using equation (3.4.21), we nd that the rate of viscous dissipation inside a uid parcel
balances the rate of working of the deviatoric viscous traction f = n,
dV = u f dS, (7.1.7)
P arcel P arcel
where is the deviatoric component of the stress tensor and n is the normal unit vector pointing
into the parcel. Further discussion of the relation between the drag force exerted on a boundary and
the rate of viscous dissipation is available [201].
1 1 d 2 a3 1 1 da3
2
p g (x xc ) p (t) = . (7.1.11)
3 r dt2 6 r 4 dt
To derive an evolution equation for the bubble radius in terms of the internal bubble pressure,
pB (t), we enforce the radial component of the interfacial boundary condition (3.8.8) with constant
surface tension, ,
rr + pB (t) = 2m , (7.1.12)
set m = 1/a, and obtain
u
r
pB (t) = p(r = a, t) 2 +2 . (7.1.13)
r r=a a
526 Introduction to Theoretical and Computational Fluid Dynamics
Note that the normal viscous stress expressed by the second term on the right-hand side makes a
nonzero contribution to the interfacial force balance. Rearranging, we obtain
da
p(r = a, t) = pB (t) 4 2 . (7.1.14)
a dt a
We will assume that the bubble is so small that hydrostatic variations can be neglected over the
bubble diameter, so that g x g xc over the bubble surface. Using (7.1.11) the evaluate the
pressure at the bubble surface, we derive a second-order nonlinear evolution equation governing the
bubble radius,
d2 a 3 da
2 4 da 1
known as the generalized Rayleigh equation. In general, the solution of (7.1.15) subject to a certain
initial condition and a specied ambient pressure eld or pressure inside the bubble must be found
by numerical methods [299].
When pB (t) = p (t), time does not appear explicitly on the right-hand side of (7.1.15). In
that case, it is possible to reduce the order of the equation by regarding da/dt a function of a, and
writing da/dt = f (a). Substituting this expression into (7.1.15), we obtain a rst-order dierential
equation,
df 3 4 1
a + f+ = 2 . (7.1.16)
da 2 a af
If viscous eects and surface tension are negligible, we obtain the exact solution f = da/dt = c/a3/2 ,
where c is a constant. Using the denition of f and integrating once with respect to time we obtain
the exact solution
5 1 da
2/5
a = a0 1 + t , (7.1.17)
2 a0 dt t=0
The collapse of a bubble near a wall has important consequences on mechanical damage due
to cavitation [38]. Observations and numerical simulations have revealed that the bubble no longer
has a spherical shape, but develops a dimple that drives a strong damaging jet of ambient uid
toward the wall.
u u
n (u) b = b (u) n, = 0, = u, (7.1.18)
lb ln
7.1 Equations and computation of irrotational ow 527
where u is the magnitude of the velocity. The third equation shows that, if a streamline has a
circular shape and the uid moves in the clockwise direction so that the curvature is positive,
u/ln is negative and the velocity on the inner side of the circle is greater than the velocity on the
outer side. The associated counterclockwise rotation of uid parcels balances the clockwise rotation
due to the global motion of the uid along the circular streamline. Generalizing this result, we nd
that the velocity of an irrotational ow around a sharp corner reaches a maximum at the salient
edge.
1
2 p = 2 (u u) = u : u. (7.1.19)
2
To derive the last expression, we have noted that the individual Cartesian velocity components
satisfy Laplaces equation in an irrotational ow, 2 ux = 2 uy = 2 uz = 0. Integrating (7.1.19)
over a control volume, Vc , bounded by a closed surface, D, and using the divergence theorem, we
obtain
n p dS = u : u dV < 0, (7.1.20)
D Vc
where the normal unit vector n points outward from the control volume. Next, we identify D with a
small surface that encloses a point where the hydrodynamic pressure allegedly reaches a minimum,
so that p is constant over this surface. By construction, n p is positive over the surface, which
contradicts inequality (7.1.20).
Problems
(a) (b)
n
n
B B V
V
n
SB
SB
n
Figure 7.2.1 Illustration of (a) ow past a translating rigid body and (b) ow due to the motion or
deformation of a body in the presence of a stationary boundary, SB .
An important area of study in the eld of incompressible potential ow concerns ow past a sta-
tionary or moving body and the ow due to the motion or deformation of a body in an eectively
inviscid uid. In this section, we discuss the general properties of these ows concentrating on the
behavior of the ow far from the body and kinetic energy of the uid, and in Section 7.3, we discuss
the developing forces and torques. Interestingly, we will nd that the behavior of the far ow, the
kinetic energy of the uid, and force exerted on the body are related by simple functional forms.
We begin by considering an ambient potential ow past a three-dimensional rigid body that translates
with velocity V in the possible presence of a stationary boundary, as illustrated in Figure 7.2.1(a).
It is benecial to decompose the harmonic potential into two constituents, = + d , where
describes the undisturbed ambient ow prevailing in the absence of the body and d is the
disturbance potential due to the body.
Boundary-integral representation
Next, we write the boundary integral representation (2.3.8) for the disturbance potential d at a
point in the ow, x0 . The boundary, D, consists of the surface of the body, B, the surface of an
impenetrable stationary boundary, SB , and a large surface enclosing B and SB , denoted by S . The
double-layer integral over SB vanishes because of the zero normal velocity. The single-layer integral
over SB can be made to vanish by using an appropriate Greens function. As the large surface
expands to innity, the integrals over S decay to zero. Enforcing the boundary no-penetration
7.2 Flow past or due to the motion of a three-dimensional body 529
Now we apply the integral relation (2.3.2) for a test ow with velocity V and cor-
responding potential V x . Identifying the control volume with the volume occupied by the
body and the boundary of the control volume with the surface of the body, we obtain
[V x (x)] n(x) G(x0 , x) dS(x) = G(x0 , x) n(x) [V (x)] dS(x). (7.2.3)
B B
Use of this identity is permissible because the point x0 lies in the exterior of the body. Combining
(7.2.2) with (7.2.3) to eliminate the single-layer potential, we derive a simplied representation
for the disturbance potential in terms of a double-layer potential. Adding to both sides of the
representation the potential of the incident ow, we obtain
(x0 ) = (x0 ) [V x (x)] n(x) G(x0 , x) dS(x). (7.2.4)
B
The modular cases of ow past a stationary body and ow due to the translation of a body arise by
setting, respectively, V equal to zero or equal to a constant.
Far-eld expansion
The representation (7.2.4) provides us with a convenient starting point for assessing the asymptotic
behavior of the ow from the body. Assuming that the point x0 lies far from the body, we select a
point xc inside or in the vicinity of the body, expand the Greens function dipole in a Taylor series
with respect to the point x about the point xc , and retain only the leading term to obtain
We have found that the coecient of the dipole can be computed from knowledge of the
velocity of the body and distribution of the velocity potential over the surface of the body. In
Section 7.3, we will see that the coecient of the dipole can be used to compute the force exerted
on a rigid body in accelerating motion.
530 Introduction to Theoretical and Computational Fluid Dynamics
The leading term on the right-hand side of (7.2.7) represents the ow due to a point source whose
strength is equal to the ow rate across the surface of the body. The second integral is the coecient
of the point-source dipole,
d (x) n(x) + (x xc ) n(x) (x) dS(x). (7.2.8)
B
Using the divergence theorem, we nd that the integral on the right-hand side of (7.2.8) remains
unchanged when the domain of integration, B, is replaced by any other closed surface enclosing B.
When the ow rate across a surface enclosing the body is zero, the integral is independent of the
location of the pivot point, xc .
As in the case of ow past a translating body discussed earlier in this section, the strength of the
dipole can be computed from the distribution of the velocity potential over the surface of the body.
Applying the divergence theorem, we nd that the second integral on the right-hand side of
(7.2.10) can be made to disappear by identifying the point xc with the volume centroid of the body
dened as
1
Xc x dV (x). (7.2.11)
VB Body
A representation of the centroid in terms of a surface integral is given in (4.1.7).
7.2 Flow past or due to the motion of a three-dimensional body 531
where the subscript on the square brackets in the second expression denotes the Cartesian compo-
nents.
If the point xc is placed at centroid of the body, Xc , dened in (7.2.11), the tensor is
identically zero and expression (7.2.16) simplies into
d = VB (V + V + ), (7.2.18)
532 Introduction to Theoretical and Computational Fluid Dynamics
which provides us with a method of computing the elements of the added-mass tensors and
from the coecient of the dipole corresponding to six independent modes of rigid-body motion.
We have noted that i decays fast enough so that the integral on the left-hand side of (7.2.19) over
a large closed surface is innitesimal. Interchanging the roles of i and j , we obtain an expression
with identical right-hand side, completing the proof.
The symmetry of ij implies that the component of the coecient of the dipole in a particular
direction due to translation in another direction is equal to the component of the dipole in the second
direction due to translation in the rst direction.
Kinetic energy of the ow due to the motion of a rigid body and grand added mass
Equation (2.1.22) provides us with an expression for the instantaneous kinetic energy of a potential
ow in terms of a boundary integral involving the boundary distribution of and the boundary
distribution of the normal velocity component. In the case of ow due to the motion of a rigid body,
possibly in the presence of stationary boundaries, we use the boundary conditions (7.2.9) to obtain
1 1 1
K= u n dS = V n dS (x xc ) n dS. (7.2.20)
2 B 2 B 2 B
Rearranging, we nd that
1 1
K = V n dS (x xc ) n dS. (7.2.21)
2 B 2 B
and substitute the decomposition (7.2.12) into the right-hand side of (7.2.21) to obtain the compact
quadratic form
1
K= VB Wi Aij Wj , (7.2.23)
2
where A is the 6 6 grand added-mass tensor dened as
1
Aij = Ni j dS, (7.2.24)
VB B
7.2 Flow past or due to the motion of a three-dimensional body 533
The tensors and dened in (7.2.17) comprise, respectively, the top and bottom 3 3 diagonal
blocks of A.
It is evident from the right-hand side of (7.2.25) that A depends on the instantaneous body
shape, orientation, and presence of other stationary objects, but is independent of the bodys linear
or angular acceleration. Physically, A is a measure of the sensitivity of the kinetic energy of the
uid to the velocity of translation and angular velocity of rotation, and may thus be regarded as an
inuence matrix for the kinetic energy.
Interchanging the role of the two potentials on the left-hand side produces the same right-hand side.
Since A is symmetric, it has only 21 independent components. Six of these components can be made
to vanish by making appropriate choices for the center of rotation xc and directions of the Cartesian
axes.
Later in this chapter, we will discuss a method of computing a potential ow in terms of discrete
or continuous distributions of dipoles inside the volume or over the surface of a body. Equations
(7.2.18) and (7.2.27) will provide us with the added-mass tensor and kinetic energy of uid in
terms of the total strength of the dipoles, thus circumventing the need for detailed computation.
534 Introduction to Theoretical and Computational Fluid Dynamics
Problems
Force
Substituting Bernoullis equation (7.1.3) into the rst equation of (7.1.5), we obtain the force exerted
on the body,
1
F= (u u) n dS VB g, (7.3.1)
2 B
7.3 Force and torque exerted on a three-dimensional body 535
n
n
n
E E
B
Vc
Figure 7.3.1 The force exerted on a stationary rigid body, B, immersed in a potential ow can be
computed in terms of surface integrals over the boundary E of a control volume that is partly
conned by the body.
where B stands for the surface of the body, n is the normal unit vector pointing into the uid, and
VB is the volume of the body.
To simplify the computation of the integral on the right-hand side of (7.3.1), we introduce
a control volume, Vc , that is enclosed by the surface of the body, B, and a closed boundary or
a collection of closed boundaries, denoted by E, as shown in Figure 7.3.1. Using the divergence
theorem, we write
(u u) dV = (u u) n dS + (u u) n dS. (7.3.2)
Vc B E
Since the ow is irrotational, the velocity gradient tensor is symmetric, u = (u)T , and
(u u) dV = 2 (u) u dV = 2 u u dV = 2 (uu) dV. (7.3.3)
Vc Vc Vc Vc
The continuity equation, u = 0, was used to derive the last expression. Applying once again the
divergence theorem and using the no-penetration boundary condition, u n = 0 over B, we obtain
(u u) dV = u (u u) dA, (7.3.4)
Vc E
where the normal unit vector n over E points outward from the control volume, as shown in Figure
7.3.1. Combining (7.3.2), (7.3.4), and (7.3.1), we derive an expression for the force as a surface
integral over E,
1
F= (u u) n 2(u n) u dS VB g. (7.3.5)
2 E
536 Introduction to Theoretical and Computational Fluid Dynamics
The integral on the right-hand side of (7.3.5) is independent of the shape of E. Identifying E with
B and using the no-penetration condition, u n = 0, we recover (7.3.1). The usefulness of the
generalized expression (7.3.5) will become evident when we discuss particular applications.
Torque
Substituting Bernoullis equation (7.1.3) into the second equation of (7.1.5) and rearranging, we nd
that the torque with respect to a point, xc , exerted on the body is given by
1
T= (x xc ) n (u u) dS VB (Xc xc ) g, (7.3.6)
2 B
where Xc is the center of volume of the body dened in (7.2.12). When the center of the torque,
xc , is identied with the volume centroid, the torque due to the body force disappears. Working as
previously for the force, we derive the generalized expression
1
T= (x xc ) [(u u) n 2(u n) u] dS VB (Xc xc ) g, (7.3.7)
2 E
where E is an arbitrary surface which, along with B, encloses a control volume, and the normal unit
vector n over E points outward from the control volume. Using the divergence theorem, we can
show that the integral on the right-hand side of (7.3.7) is independent of the shape of E. Identifying
E with B and using the no-penetration condition, u n = 0, we recover (7.3.6).
7.3.2 Force and torque on a body near a point source or point-source dipole
As an application, we consider the ow due to a point source with strength m located at a point, xs ,
in the vicinity of a body, in the absence of any other interior or exterior boundaries. Decomposing
the velocity eld into a singular component due to the point source and a nonsingular complementary
component due to the presence of the body, denoted by v, we obtain
m x xs
u= + v, (7.3.8)
4 r 3
where r = |x xs | is the distance from the point source.
Next, we identify E with a spherical surface of small radius centered at the point source,
denoted by S , and a large surface enclosing the body and the point source. As the large surface
expands to innity, the corresponding integrals in (7.3.5) make innitesimal contributions. Over
the surface of the small sphere enclosing the point force, the inward unit normal vector is given by
n = 1 (x xs ). Consequently,
m m
2 m m
u= n + v, uu= n v + v v, un= + v n. (7.3.9)
42 42 22 42
Substituting these expressions into (7.3.5), evaluating the velocity v at the location of the point
source, and simplifying, we obtain
1 m m m
F= [ 2
n (vs n) 2( 2
+ vs n)( 2
n + vs ) ] dS VB g, (7.3.10)
2 S 2 4 4
7.3 Force and torque exerted on a three-dimensional body 537
where vs v(xs ). Simplifying and performing the integration in the limit as 0, we derive the
remarkably simple expression
F = m v(xs ) VB g. (7.3.11)
Substituting (7.3.9) into (7.3.7) and working in a similar fashion, we obtain an expression for the
torque,
We have found that the force and torque can be computed in terms of the reected velocity, v, at
the position of the point source.
F = d v(xd ) VB g, (7.3.13)
Multiple singularities
If the ow is due to a collection of singularities including point sources and point-source dipoles, the
right-hand sides of (7.3.11), (7.3.12), (7.3.13), and (7.3.14) are summed over all singularities.
Force
Substituting Bernoullis equation (7.1.3) into the expression for the force (7.1.5), we nd that
1
F= p n dS = + u u n dS VB g. (7.3.15)
B B t 2
To assess the contribution of the six fundamental modes of rigid body motion, we express in terms
of the six fundamental potentials introduced in (7.2.12), Expanding the time derivative, we obtain
dVi di i i+3
= i + i+3 + Vi + i , (7.3.16)
t dt dt t t
538 Introduction to Theoretical and Computational Fluid Dynamics
for j = 16. The material point xc (t) denes the instantaneous position of the body, and the unit
vector e(t) denes the instantaneous orientation. Substituting (7.3.16) into (7.3.15), we nd that
dVi di
F= n i dS + n i+3 dS + FS (V, ), (7.3.18)
B dt B dt
where all terms that depend on the instantaneous position, orientation, linear and angular velocities
of the body, but are independent of the linear or angular acceleration, have been collected into the
term FS expressing the force exerted on a body in steady motion. The rst two terms on the right-
hand side of (7.3.18) express the acceleration reaction. In terms of the 3 3 added-mass tensors
and introduced in (7.2.17), we obtain
dV d
F = VB + + FS (V, ). (7.3.19)
dt dt
The rst term on the right-hand side of (7.3.19) is a generalized version of Newtons second law
of motion, where the coecient matrix VB plays the role of an eective or virtual mass. Since
the product VB is the mass of uid displaced by the body, it is natural to call the coecient of
virtual inertia or added mass.
Torque
Working in a similar fashion, we derive an expression for the torque with respect to a point, xc (t),
exerted on a moving rigid body,
dVi di
T= (x xc ) n i dS + (x xc ) n i+3 dS + TS (V, ). (7.3.20)
B dt B dt
The rst two terms on the right-hand side express the torque due to unsteady translation and
rotation. The third term is analogous to that on the right-hand side of (7.3.18).
dW
R = VB A + RS , (7.3.21)
dt
where the superscript S denotes the value of R in steady motion. Since the matrix A is symmetric,
the order of the vector-matrix multiplication in the rst term on the right-hand side of (7.3.21) is
inconsequential. The rst term on the right-hand side of (7.3.21) expresses a variation of Newtons
second law of motion where the matrix coecient VB A plays the role of an eective or virtual mass
7.4 Force and torque on a translating rigid body in streaming ow 539
n
n B
E
Vc
VB
Figure 7.4.1 Illustration of an arbitrary body, B, translating with velocity V in an innite ambient
uid. To compute the force and torque exerted on the body, we introduce a closed surface, E,
enclosing the body.
for the extended velocity vector. This observation provides justication for calling A the grand
coecient of virtual inertia or grand added mass.
Problems
F = VB + u u + Vi n dS VB g, (7.4.1)
dt B 2 t
540 Introduction to Theoretical and Computational Fluid Dynamics
where the material vector e is independent of time. The three terms on the right-hand side of (7.4.1)
represent, respectively, the acceleration reaction, the steady force, and the buoyancy force. Since
the body translates in a uid of innite expanse in the absence of exterior or interior boundaries,
the fundamental potentials i depend on the distance from the designated center of the body, xc ,
i x, xc (t), e = i x xc (t), e , (7.4.2)
We will show that the integral on the right-hand side of (7.4.4) is identically zero and the expression
for the force simplies into
dV
F = VB VB g. (7.4.5)
dt
where the normal unit vector n over E points outward from the control volume. Since the ow is
irrotational, u = (u)T and
(u u) dV = 2 (u) u dV = 2 u u dV = 2 (uu) dV. (7.4.7)
Vc Vc Vc Vc
To derive the last expression, we have used the continuity equation, u = 0. Applying the
divergence theorem once again and enforcing the no-penetration boundary condition, we obtain
(uu) dV = (V n) u dS + (u n) u dS. (7.4.8)
Vc B E
Using the irrotationality condition, u = (u)T , we establish that the integral over B on
the right-hand side of (7.4.10) remains unchanged when the domain of integration is replaced by
an arbitrary surface enclosing B (Problem 7.4.2(a)). Replacing B with E and combining the two
integrals on the right-hand side of (7.4.10), we obtain the compact form
dV 1
F = VB + u [(V u) n] + [u ( u V)] n dS VB g. (7.4.11)
dt E 2
Identifying E with B reproduces (7.4.4). Now we identify E with a large spherical surface and note
that, because in the far ow the velocity behaves like that due to a point-source dipole and thus
decays like 1/r 3 , the integral on the right-hand side (7.4.11) is identically zero, yielding (7.4.5).
We have thus arrived at the remarkable conclusion that the hydrodynamic force exerted on a
three-dimensional body translating steadily in a uid of innite expanse with no exterior or interior
boundaries is zero. It is important to emphasize that the presence of boundaries may cause the
development of a drag force, a lift force force, or both (Problem 7.4.1). Other exceptions arise in
the case of an innite cylindrical body with a smooth or corrugated surface.
DAlemberts paradox
The component of the steady force in the direction of translation can be computed in an alternative
fashion using an energy argument. Concentrating on the integral energy balance (3.4.23), repeated
here for convenience,
1 1
|u|2 dV = ( |u| 2
) u n dS + p u n dS + g u dV, (7.4.12)
Vc t 2 D 2 D Vc
we use the divergence theorem to convert the rst boundary integral on the right-hand side into
a volume integral. Combining the resulting expression with the integral on the left-hand side and
expressing the body-force term as shown in (3.2.14), we obtain
D 1
|u|2 dV = p u n dS (g x)(u n) dS, (7.4.13)
Vc Dt 2 D D
where D/Dt is the material derivative. Now we identify the control volume with the whole domain
of ow and recognize the left-hand side of (7.4.13) as the rate of change of the kinetic energy, dK/dt.
The boundary, D, consists of the surface of the body and a surface of large size extending to innity.
We note that, far from the body, the ow behaves like that due to a point-source dipole and eliminate
the corresponding integrals over the large surface. Enforcing the no-penetration condition over the
542 Introduction to Theoretical and Computational Fluid Dynamics
surface of the body and using the divergence theorem to manipulate the surface integral involving
the body force, we obtain
dK
= V (F + VB g). (7.4.14)
dt
The parentheses on the right-hand side enclose the hydrodynamic force, excluding the buoyancy
force.
Now working independently, we dierentiate (7.2.27) with respect to time and recall that the
virtual mass matrix is symmetric to express the rate of change of kinetic energy in the form
dK dVj
= VB ij Vi . (7.4.15)
dt dt
Combining (7.4.14) with (7.4.15) and using (7.2.18) to express the added mass in terms of the
coecient of the dipole, we obtain
dK dVj dV
= V (F + VB g) = VB ij Vi = (d VB V) , (7.4.16)
dt dt dt
which shows that the component of the hydrodynamic component of the force in the direction of
translation is nonzero only when the body accelerates, and vanishes when the body executes steady
motion (dAlemberts paradox). Physically, when a body translates with constant velocity in the
absence of exterior and interior boundaries, the whole ow pattern is convected with the body and
the kinetic energy remains constant in time, dK/dt = 0. Because the kinetic energy of the uid
remains constant in the absence of dissipation, work is not required to sustain the motion of the
uid.
The force exerted on a rigid body that is held stationary in an incident streaming (uniform) ow with
time-dependent velocity U(t) can be found using the results for ow due to the body translation.
In a frame of reference translating with velocity U(t), the far ow vanishes and the body appears
to translate with velocity V(t) = U(t) in an otherwise quiescent uid. Subtracting from the
acceleration reaction given in (7.4.5) the ctitious inertial force VB dU(t)/dt, we obtain
dV dU
F = VB + g + VB . (7.4.17)
dt dt
Rearranging, we nd that
dU
F = VB (I + ) VB g. (7.4.18)
dt
The term associated with the identity matrix I on the right-hand side expresses the force exerted
on the uid displaced by the body in accelerating streaming ow.
7.4 Force and torque on a translating rigid body in streaming ow 543
dV dU dV
B VB = VB (I + ) VB VB g + B VB g, (7.4.21)
dt dt dt
where B is the density of the body. Rearranging, we derive a linear ordinary dierential equation,
dV dU
(I + ) = (I + ) + ( 1) g, (7.4.22)
dt dt
where = B / is the density ratio. In Section 7.5 we will discuss applications of (7.4.22) to the
motion of a sphere.
where Xc is the volume centroid, x = x Xc , and summation is implied over the repeated index i
over 1, 2, 3. This expression is the counterpart of (7.4.4) for the force.
To simplify the computation of the integral of the quadratic term involving u u on the right-
hand side of (7.4.23), we introduce a control volume, Vc , conned between the body, B, and a closed
544 Introduction to Theoretical and Computational Fluid Dynamics
exterior surface , E, as shown in Figure 7.4.1. Using the divergence theorem stated in equation
(A.7.3), Appendix A, we restate the integral of the quadratic term as
(u u) x n dS = (u u) x n dS + [(u u) x] dV, (7.4.24)
B E Vc
where the normal unit vector n over E points outward from the control volume. Focusing on
the volume integral, we recall that, because the ow is irrotational, the velocity gradient tensor is
symmetric, u = (u)T , yielding
[(u u) x] dV = 2 [(u) u] x dV = 2 [u u] x dV (7.4.25)
Vc Vc Vc
and then
[(u u) x] dV = 2 (uu) x dV = 2 [u (u x)] dV. (7.4.26)
Vc Vc Vc
To derive the penultimate integral, we have used the continuity equation, u = 0. Applying the
divergence theorem once more and enforcing the no-penetration boundary condition, we obtain
[(u u) x] dV = 2 (u n) u x dS 2 (V n) u x dS. (7.4.27)
Vc E B
Identifying E with B reproduces (7.4.23). The integrand in the second integral on the right-hand
side of (7.4.29) can be simplied by writing (u V) n u (V n) = V (u n).
Next, we add and subtract one term from the second integral on the right-hand side of (7.4.29),
obtaining
dVi
T= x n i dS x [V (u n) + V n] dS
dt
B
B
1
+ V n dS + x [ (u u) n (u n) u ] dS VB (Xc xc ) g. (7.4.30)
B E 2
The second integral on the right-hand side of (7.4.30) remains unchanged when the surface of the
body, B, is replaced by a closed surface enclosing B (Problem 7.4.2(b)). Replacing B with a large
7.4 Force and torque on a translating rigid body in streaming ow 545
spherical surface and noting that, far from the body, the ow is similar to that due to a point-source
dipole and the velocity decays like 1/d3 , where d is the distance from the origin, we nd that the
second and fourth integrals on the right-hand side of (7.4.30) are identically zero.
Finally, we express the third integral on the right-hand side of (7.4.30) in terms of the coe-
cient of the dipole using (7.2.10), we nd that
dVi
T= (x n) i dS V d VB (Xc xc ) g, (7.4.31)
B dt
Steady ow
In the case of steady ow and in the absence of signicant gravitational eects, the hydrodynamic
torque is
T = V d, (7.4.32)
which suggests that the torque exerted on a steadily translating body vanishes only when the coef-
cient of the dipole is oriented in the direction of translation so that the cross product is zero.
Using (7.2.18) to express the coecient of the dipole, d, in terms of the added mass tensor ,
and remembering that is symmetric, we nd
T = VB V ( V). (7.4.33)
Problems
d = 2a3 V (7.5.1)
placed at the instantaneous center of the sphere, x0 . In this case, not only the far ow, but also the
entire ow is described by a potential dipole. Using (2.1.33) and (2.1.34), we nd that the harmonic
potential and velocity eld are given by
1 a3 1 a3 3
= V x, u= V + 2 (V x) x , (7.5.2)
2 r3 2 r 3 r
where r = |x| and x = x x0 . The streamline pattern shown in Figure 7.5.1(a) is identical to that
outside Hills spherical vortex illustrated in Figure 2.13.3(a).
1 a3
= V sin2 . (7.5.3)
2 r
The surface velocity is equal to V at the centerline located at = 0 and , and 12 V at the midplane
located at = 12 .
Fundamental potentials
Using (7.5.2) and noting that a sphere rotating about its center does not generate a ow, we nd
that the six fundamental potentials dened in (7.2.13) are given by
1 a3
i = xi , i+3 = 0 (7.5.5)
2 r
7.5 Flow past or due to the motion of a sphere 547
(a) (b)
3 3
2 2
1 1
0 0
y
y
1 1
2 2
3 3
3 2 1 0 1 2 3 3 2 1 0 1 2 3
x x
Figure 7.5.1 Streamline pattern of irrotational ow (a) due to a translating sphere and (b) past a
stationary sphere.
for i = 1, 2, 3, where the point xc has been identied with the center of the sphere, x0 . If we set
the point xc at a location other than the center of the sphere, the rst three potentials associated
with translation will be modied and the last three potentials associated with rotation will express
a nontrivial ow.
In spherical polar coordinates where the x axis is parallel to the far-eld velocity, U = U ex ,
the Stokes stream function is given by
1 a3
= U r 2 sin2 1 3 , (7.5.7)
2 r
where = r sin is the distance from the axis of revolution. The surface velocity is zero at the
front and back of the sphere corresponding to = 0 and , and is equal to 32 U at the midplane
corresponding to = 12 .
548 Introduction to Theoretical and Computational Fluid Dynamics
3 dU 1 dV
F= VB VB VB g, (7.5.8)
2 dt 2 dt
and the acceleration of the sphere is
dV 3 dU 1
= +2 g, (7.5.9)
dt 2 + 1 dt 2 + 1
where = s / and s is the density of the sphere. In the absence of the gravitational force, the
acceleration of a sphere that is heavier than the uid, > 1, is lower than that of the uid, while
the acceleration of a sphere that is lighter than the uid, < 1, is higher than that of the uid.
dV 1
=2 g. (7.5.10)
dt 2 + 1
One interesting consequence of this expression is that the acceleration of a sphere with negligible
mass, = 0, such as an air bubble, is twice the acceleration of gravity.
energy of the uid remains constant in time, and thus obtain the Levich equation
V (F + VB g) = dV (7.5.11)
F low
([235] p. 444; [259]). In Section 8.1.2, we will see that, at high Reynolds numbers, the rate of
dissipation inside the boundary layer around the surface of the bubble is negligible compared to
the viscous dissipation in the bulk of the uid. Consequently, the overall rate of dissipation can be
computed using (7.1.6). According to (7.1.7), the right-hand side of (7.5.11) is equal to the rate of
working of the deviatoric part of the boundary traction computed using the solution for irrotational
ow.
Substituting (7.5.2) into the last integral of (7.1.6) and carrying out the integration produces
the value 12a|V|2 , which shows that the hydrodynamic drag force exerted on the sphere is
D F + VB g = 12a V. (7.5.12)
It is interesting that the magnitude of the drag force is twice that predicted by Stokess law for
a solid sphere moving at low Reynolds numbers given in (6.7.12). The drag coecient for an air
bubble moving with velocity V in a viscous uid at high Reynolds numbers is then
D 24
cD = , (7.5.13)
V 2 a2 Re
where Re = 2aV / is the Reynolds number and is the kinematic viscosity of the uid.
Newtons second law of motion requires that the force F is balanced by the weight of the
bubble, F + B VB g = 0. Neglecting the density of the bubble, B 0, we set F = 0 and use
(7.5.12) to nd the terminal velocity
1 ga2
V = . (7.5.14)
9
It is worth emphasizing that this prediction arises from a relatively simple calculation using the
potential ow solution alone.
2
xinv () = x0 + (x x0 ) (7.5.16)
r2
550 Introduction to Theoretical and Computational Fluid Dynamics
xs
inv
xs
x0
x inv
x
Figure 7.5.2 Illustration of the inverse of a eld point, x, and inverse of a point source, xs , with
respect to a sphere of radius centered at x0 .
is the inverse of the eld point, x, with respect to a sphere of radius centered at x0 , and r = |xx0 |
is the distance of x from the center of the sphere [426]. The right-hand side of (7.5.15) involves the
potential of the incident ow along a line starting at the center of the sphere and ending at the
inverse of the evaluation point, x, with respect to the sphere, xinv (a), as shown in Figure 7.5.2.
2
xinv
s () = x0 + (xs x0 ), (7.5.19)
rs2
where rs = |xs x0 | is the distance of the point source from the center of the sphere. By denition
of the inverse points,
where
rsinv = |xinv
s () x0 |, rinv = |xinv () x0 | (7.5.21)
7.5 Flow past or due to the motion of a sphere 551
are the distances of the inverse points from the center of the sphere, x0 , as illustrated in Figure
7.5.2. Rearranging (7.5.20), we nd that
r rinv
= sinv , (7.5.22)
rs r
which demonstrates that the triangles (x0 , xinv , xs ) and (x0 , xinv
s , x) depicted in Figure 7.5.2 are
similar. Accordingly,
r |xinv
s () x|
= inv (7.5.23)
rs |x () xs |
It is convenient to write
xinv
s () = x0 + es , (7.5.25)
where es = (xs x0 )/rs is the unit vector pointing from the center of the sphere to the point source,
and = 2 /rs . Straightforward manipulation of the integral in (7.5.24) yields the more useful form
a2 /rs
m a 1 m d
d (x) = 2
+ . (7.5.26)
4 rs |x (x0 + a es /rs |) 4a 0 |x (x0 + es )|
The rst term on the right-hand side of (7.5.26) represents a point source with strength ma/rs
placed at the inverse point of the original point source with respect to the sphere. The second term
represents a continuous distribution of point sinks extending from the center of the sphere to the
inverse point of the point source. The net discharge of the point source and point sinks balances to
zero.
The force exerted on the sphere can be deduced immediately from the exact solution using
expression (7.3.11), where v(xs ) = d (xs ). Taking the gradient of (7.5.24), we nd that
m a a
1 m 1 1
v(xs ) = d es . (7.5.27)
4 rs |xs xinv
s (a)|
2 2 ars 0 |xs xinvs ()|
2
Next, we substitute
2 2
xs xinv
s () = (1 )(xs x0 ), |xs xinv
s ()| = (rs ), (7.5.28)
rs2 rs
and obtain
2
m ars rs a d 2
v(xs ) = es . (7.5.29)
4 (rs2 a2 )2 a 0 (rs2 2 )2
552 Introduction to Theoretical and Computational Fluid Dynamics
The integral on the right-hand side is equal to a2 /[rs2 (rs2 a2 )]. Making substitutions and simplifying,
we nd that
m a3
v(xs ) = es . (7.5.30)
4 rs (rs a2 )2
2
Butlers sphere theorem [63] states that the disturbance stream function due to the presence
of the sphere, denoted by the subscript d, is given by
r a2
d = ( , ). (7.5.32)
a r
It is reassuring to conrm that the stream function given in (7.5.7) for uniform ow past a sphere
is in agreement with (7.5.32).
A second theorem due to Butler states that, if all singularities of (r, ) lie inside the sphere,
and if decays like 1/r far from the sphere, then (7.5.32) provides us with the disturbance ow
in the interior of the sphere.
We can verify that + d = 0 at the surface of the sphere, ensuring that the no-penetration
condition is satised.
m x xs xs
m r cos xs xs
= + = + . (7.5.35)
4 |x xs | |xs | 4 (r2 + x2s 2rxs cos )1/2 |xs |
m r a2 cos rxs xs
d = + . (7.5.36)
4 a (a + r xs 2a rxs cos )
4 2 2 2 1/2 |xs |
It can be shown that the ow expressed by (7.5.36) is identical to that described by the potential
given in (7.5.18) [63].
Using this expression, we nd that the disturbance ow is generated by an image ring whose strength
and location are given by
a a2 a2
im = , xim
R = xR 2 ,
im
R = R 2 . (7.5.42)
rR rR rR
The image ring passes through the inverse point of the trace of the primary ring with respect to the
sphere in any azimuthal plane.
Computer Problems
(b) Demonstrate that the disturbance potential can be described by an image system consisting of a
line vortex ring of radius a2 /(2b) lying in the xz plane and centered at x = a2 /(2b) and z = 0, and
a vortex sheet subtended by the image ring [426].
Computer Problem
coordinates that conform with the body geometry, subject to the no-penetration boundary condi-
tion. The procedure is explained in detail in classical texts of uid mechanics (e.g., [220] Chapter
5; [268]). Two examples concerning the motion of a prolate or oblate spheroid will be discussed in
Sections 7.6.1 and 7.6.2.
Singularity methods
An interesting method of computing a potential ow past or due to the motion of a body involves
representing the harmonic potential in terms of discrete or continuous distributions of singular fun-
damental solutions, including the point source and the point-source dipole, and then computing the
strengths of the singularities, and possibly their optimal location, by analytical or numerical methods
in order to satisfy stipulated boundary conditions. One advantage of the singularity representation
is that it provides us with the kinetic energy of the uid, the added-mass tensor, and the accelera-
tion reaction directly from the strength of the singularities, thereby bypassing the computation of
the velocity eld and pressure distribution over the body and inside the ow. Applications will be
discussed in Sections 7.6.1 and 7.6.3.
Boundary-integral methods
Another powerful class of numerical methods for computing potential ow originates from the
boundary-integral representation discussed in Chapter 2, as well as from generalized single- or
double-layer representations discussed in Chapter 10. The numerical procedure involves apply-
ing the boundary-integral representation at the boundaries of the ow to derive Fredholm integral
equations of the rst or second kind for the boundary distributions of unknown functions, including
the harmonic potential and its normal derivative. The integral equations can be solved by highly
accurate boundary-element or panel methods. The engineering importance and popularity of the
boundary-integral methods justies their separate discussion in Chapter 10.
b
x
c
a
c
z
Figure 7.6.1 The potential ow due to the translation of a prolate spheroid can be represented by
potential dipoles distributed over the focal length of the spheroid pointing in the direction of
translation.
Using Maclaurin series expansions, we nd that, as the eccentricity tends to zero and the prolate
spheroid tends to a sphere, x , y , z 43 e3 .
The total coecient of the dipole is found by integrating the density of the parabolic distri-
bution, yielding
c
Vi 8 3 Vi
di = 2 (c2 2 ) d = c , (7.6.4)
i c 3 i
where summation is not implied over the repeated index i. Substituting expression (7.6.4) into
(7.2.27) and noting that the volume of the spheroid is VB = 4 2
3 ab , we obtain the kinetic energy of
the uid
4 3 Vx2 Vy2 V 2
2
K= c + + z ab2 (Vx2 + Vy2 + Vz2 ). (7.6.5)
3 x y z 3
Using (7.2.18) we nd that the nonzero elements of the diagonal added-mass tensor are
a
2 e3 a
2 e 3
xx = 2 1, yy = zz = 2 1. (7.6.6)
b x b y
As the eccentricity of the spheroid tends to zero, we recover the results of Section 7.5.1 for a sphere,
xx = yy = zz = 1/2.
where the dimensionless parameter ranges from unity to innity and the dimensionless varies
between 1 and 1 (e.g., [220], p. 141). It is convenient to set = cos , where the parameter
varies in the range [0, ] along the contour of the generating ellipse. The surfaces of constant
7.6 Flow past or due to the motion of a nonspherical body 557
or are confocal prolate spheroids and hyperboloids of two sheets. The surface of the spheroid
corresponds to = 0 = 1/e. As 0 1, the spheroid reduces to a slender needle.
The velocity potential and Stokes stream function of the axisymmetric ow due to a prolate
spheroid translating along the x axis are given by
Vx 1 + 1 Vx 2 1 + 1
= 2c 1 ln , = ln , (7.6.8)
x 2 1 x 2 1 2 1
1 2 1 + 1
Vy Vz
= ln 2 cos + sin , (7.6.9)
c 2 1 1 y z
where the dimensionless parameter varies from zero to innity and the dimensionless parameter
varies in the interval [1, 1]. We may set cos , where the angle varies from 0 to along the
contour of the generating ellipse (e.g., [220], p. 144).
Consider an oblate spheroid with minor semiaxis along the x axis equal to a and major semiaxis
normal to the x axis equal to b, where a < b, as shown in Figure 7.6.2. The focal length of the
spheroid is c = eb, where
a
2 1/2
c 2 = b 2 a2 , e 1 (7.6.11)
b
The dimensionless eccentricity, e, varies in the interval (0, 1] whose lower or upper limit corresponds
to a circular disk or a sphere. The surface of the spheroid corresponds to = 0 = a/c. The surfaces
of constant z or are, respectively, planetary ellipsoids and hyperboloids of revolution of one sheet
with common focal circle at x = 0 and = c.
The velocity potential and Stokes stream function of the axisymmetric ow due to axial
translation along the x axis are given by
Vx 1 Vx 2
a
c
z b
Figure 7.6.2 Illustration of the potential ow due to the translation of an oblate spheroid. Solutions
for the axisymmetric ow due to axial translation and three-dimensional ow due to transverse
translation are available.
where
1
x = 1 e2 arcsine. (7.6.13)
e
The velocity potential of the three-dimensional ow due to transverse translation normal to the x
axis is given by
1 2
V
y Vz
= arccot ( cos + sin ), (7.6.14)
c 2 + 1 y z
where
1 02 + 2
y = z = arccot0 . (7.6.15)
0 02 + 1
As the aspect ratio a/b tends to zero and the eccentricity tends to unity, we obtain the ow due
to the translation of a circular disk of innitesimal thickness (Problem 7.6.1). In the opposite limit
where a/b tends to unity and the eccentricity tends to zero, we recover the results of Section 7.5.1
for a sphere.
(a) (b)
y x y
r Vy
Vx
i
x
x
1 i N
z z
Figure 7.6.3 Representation of the ow due to the translation of an axisymmetric body in terms of
(a) point sources and sinks for axial motion and (b) point-source dipoles for transverse motion.
It is convenient to work in a frame of reference translating with the body along with the x
axis with velocity Vx . In this frame, the body appears to be stationary and the uid approaches
the body along the x axis with velocity Vx . Referring to spherical polar coordinates centered at
each singularity, and using the expression (2.9.15) for each singularity, we derive the Stokes stream
function
1
N
1 2
(x, ) = Vx mi cos i , (7.6.16)
2 4 i=1
where is the distance from the axis of symmetry, mi is the strength of the ith singularity, and i
is the angle subtended between the eld point x and the ith singularity located at the point xi on
the x axis, satisfying
as shown in Figure 7.6.3(a). The problem has been reduced to computing the strengths of the
singularities so that = 0 over the contour of the body in an azimuthal plane. Having obtained the
strengths of the singularities, we recover the coecient of the dipole, d = (dx , 0, 0), as
N
dx = xi mi . (7.6.18)
i=1
560 Introduction to Theoretical and Computational Fluid Dynamics
(a) (b)
2.5 4
2
3
1.5
2
1
1
0.5
y/d
y/d
0 0
0.5
1
1
2
1.5
3
2
2.5 4
2 1 0 1 2 4 2 0 2 4
x/d x/d
Figure 7.6.4 Rankine ovoids generated by introducing a point and a point sink separated by distance
2d in streaming ow for (a) = 0.2 and (b) 0.05. As decreases, the dividing stream surface
becomes a sphere.
Since the strengths of the singularities add up to zero, the result is independent of the chosen origin
of the x axis.
Rankine ovoids
Introducing one point source and one point sink with strengths of equal magnitude and opposite
sign, m, located at x = d, we nd that = 0 over a family of axisymmetric bodies parametrized
by the dimensionless number = Vx d2 /m, known as Rankine ovoids. Two examples for = 0.5 and
0.05 ar shown in Figure 7.6.4. As tends to zero, the two singularities merge to yield a point-source
dipole with strength
2
dx = 2md = Vx d3 . (7.6.19)
Referring to (7.5.1), we nd that, in this limit, the ovoids reduce to a sphere of radius a satisfying
the equation dx = 2a3 Vx . Solving for the radius of the sphere, we nd that (a/d)3 = 1/().
Numerical methods
Given a body of a certain shape, the strengths of the point sources or sinks, mi , can be computed
by pointwise collocation. The procedure involves requiring that = 0 at Nc selected points over
the contour of the body in an azimuthal plane, where Nc N , and then solving a linear system
of equations for the strength of the singularities, mi . Selecting Nc > N and then solving an
overdetermined system of linear equations by a least-squares or minimization method helps reduce
the sensitivity of the solution to the position of the collocation points.
7.6 Flow past or due to the motion of a nonspherical body 561
Using the expression for the potential due to a point source given in Section 2.1.6, we nd
that the harmonic potential corresponding to the distribution (7.6.20) is given by
b
1 1
(x) = Vx x () d, (7.6.21)
4 a |x |
where the point = (, 0, 0) lies along the x axis. In fact, it can be shown that the continuous distri-
bution (7.6.20) is equivalent to a series expansion in terms of Legendre polynomials with argument
cos , scaled over the interval [a, b]. The coecients of the series are related to the moments of the
distribution density function, () ([24], p. 460).
1
N
1 y
u(x) = Vy di ey + 3 (x x i ) , (7.6.22)
4 i=1
|x xi |3 |x xi |5
where xi = (xi , 0, 0) is the location of the ith singularity and ey is the unit vector along the y axis.
Evaluating (7.6.22) at the surface of the body, enforcing the no-penetration boundary condition,
u n = V n = Vy ny , and writing ny = (y/) n and nz = (z/) n , we obtain
N 1 (x xi )nx + n
di 3
+3 = 4, (7.6.23)
i=1
|x xi | n |x xi |5
In an advanced implementation of the singularity method, the locations of the singularities are not
specied at the outset but computed instead as part of the solution to minimize an appropriate
positive functional enforcing the boundary conditions (e.g., [442]). This modication allows us to
achieve a high degree of accuracy with a small number of singularities.
Problem
Computer Problems
f(x)
x
a a
z
Figure 7.7.1 Illustration of a slender axisymmetric body. The ow due to the translation of the body
can be represented in terms of singularities distributed over an interval inside the body.
The density of the distribution, , must be adjusted to satisfy the no-penetration condition at
the surface of the body. To achieve this, we write a mass balance over a xed control volume that
is conned between the instantaneous surface of the body and two planes that are perpendicular to
the x axis. Let A = f 2 be the cross-sectional area of the body and Q(x) be the instantaneous ow
rate in the exterior of the body through a plane that is perpendicular to the x axis. In the limit as
the distance between the two planes becomes innitesimal, we obtain
A Q f Q
= or 2f = , (7.7.2)
t x t x
Next, we introduce the instantaneous ow rate q(x) across the whole area of a plane that is perpen-
dicular to the x axis and use the point-source representation to write
q
= . (7.7.3)
x
564 Introduction to Theoretical and Computational Fluid Dynamics
Since, in a frame of reference moving with the body, the radius of the body is constant,
f f
+ Vx = 0. (7.7.4)
t x
Combining this equation with (7.7.2) and approximating Q with q, we nd that the density distri-
bution of the point sources is given by
A A
= = Vx . (7.7.5)
t x
Although the assumptions underlying these derivations fail near the end of the body where (7.7.2)
ceases to be valid, expression (7.7.5) provides us with a reasonable approximation of the ow over
the main portion of the body.
The coecient of the dipole, d, is parallel to the centerline. Using (7.7.5) and integrating by
parts under the stipulation f (a) = 0, we nd that
a a
d(f 2 )
dx = () d = Vx d = Vx VB , (7.7.6)
a a d
where VB is the volume of the body. Substituting (7.7.6) into (7.2.18) and (7.2.23), we nd that,
at this level of approximation, the added mass and kinetic energy of the uid are both zero. A
higher-order approximation is required to account for the presence of the body.
Prolate spheroids
To assess the accuracy of the slender-body theory, we compare the approximate solution (7.7.6) with
the exact solution for a prolate spheroid given in (7.6.4). As the aspect ratio b/a tends to zero, the
exact solution yields
4 b
2 b
2 b
dx = Vx a3 1 ln + . (7.7.7)
3 a a a
Neglecting the second term inside the square bracket yields precisely the right-hand side of (7.7.6).
The slender-body approximation incurs a relative error on the order (b/a)2 ln(b/a), which is small
for bodies with moderate and high aspect ratio. Consequently, slender-body theory provides us with
an ecient method for approximating the coecient of the dipole with the product of the volume
of the body and velocity of translation in axial motion.
Asymptotic expansions
Equation (7.7.6) can be derived formally working within the framework of asymptotic expansions.
We begin by considering the no-penetration condition, ux nx + u n = Vx nx , and introduce the
approximations nx f /x and n 1. Since ux and u have comparable magnitudes while nx
is small compared to n , we may approximate
f
u Vx . (7.7.8)
x
7.7 Slender-body theory 565
where = ( x)/ is an auxiliary variable. Next, we evaluate (7.7.9) at a point on the surface
of the body, = f (x), expand () in a Taylor series about x and retain only the leading constant
term. As f /a tends to zero, the limits of integration with respect to become innite yielding the
approximation
1 d 1 (x)
u (x) = (x) 2 + 1)3/2
= . (7.7.11)
4f (x) ( 2 f (x)
Working in a similar fashion, we nd that the axial component of the velocity at a point on
the surface of the body is given by
a
1 x
ux (x) = = () d. (7.7.12)
x 4 a [(x )2 + 2 ]3/2
Expanding () in a Taylor series about x and retaining only the constant and linear terms, we
obtain
1 d 4(a2 x2 )
ux (x) = ln , (7.7.13)
4 dx f 2 (x)
which can be used along with (7.7.11) and Bernoullis equation to evaluate the surface pressure.
In Section 7.8, we will see that streaming ow past a circular cylinder can be represented in
terms of a two-dimensional dipole oriented along the y axis with strength 2Vy f 2 = 2Vy A, where A
is the cross-sectional area of the body. Since the two-dimensional dipole emerges by integrating the
three-dimensional dipole along the x axis, it is appropriate to introduce a representation in terms
of a distribution of three-dimensional dipoles with linear distribution density 2Vy f 2 (x), obtaining
the potential
a
1 y
(x) = Vy f 2 () d, (7.7.14)
2 a |x |3
566 Introduction to Theoretical and Computational Fluid Dynamics
where the point = (, 0, 0) lies on the x axis. This representation can be derived more rigorously
working as in the case of axial motion in the context of asymptotic expansions (Problem 7.7.1(b)).
The coecient of the dipole,
a
dy 2Vy f 2 (x) dx = 2Vy VB , (7.7.15)
a
is equal to twice that for axial motion shown in (7.7.7). A comparison with the exact solution for
prolate spheroids conrms the consistency and accuracy of the slender-body representation (7.7.14)
for bodies with moderate and large aspect ratio (Problem 7.7.1(c)). The nonzero component of the
added-mass matrix and kinetic energy of the uid are yy = 1 and K = 12 VB Vy2 .
Problems
(a) (b)
F
V
Figure 7.8.1 (a) Illustration of irrotational ow past a translating two-dimensional rigid body with
nonzero circulation around the body. (b) Direction of the lift force, F, on a translating two-
dimensional body with positive circulation around the body.
The no-penetration boundary condition requires that u n = V n over the contour of the
body, B. Making substitutions and rearranging, we obtain
n d = n V n n v. (7.8.1)
The counterpart of the boundary-integral representation (7.2.2) for the single-valued disturbance
potential, d , is
d (x0 ) = G(x0 , x) n(x) V (x) v(x) dl(x)
B
+ d (x) n(x) G(x0 , x) dl(x), (7.8.2)
B
Using the counterpart of (7.2.3) for two-dimensional ow to simplify the single-layer potential
on the right-hand side of (7.8.2), and adding to both sides of the resulting equation the potential of
the incident ow, , we obtain
(x0 ) = (x0 ) + G(x0 , x) n(x) v(x) dl(x)
B
+ [ V x + (x) ] n(x) G(x0 , x) dl(x), (7.8.3)
B
Far-eld expansion
To assess the behavior of the ow far from the body, we follow a procedure similar to that outlined
in Section 7.2 for three-dimensional ow. The result is the asymptotic expansion
(x0 ) = (x0 ) + d c G(x0 , xc ) + + , (7.8.5)
2
where the gradient c involves derivatives with respect to an arbitrary point inside or in the vicinity
of the body, xc ,
d = AB V n(x) (x) (x xc ) v(x) n(x) dl(x) (7.8.6)
B
is the coecient of the dipole, AB is the cross-sectional area of the body in the xy plane. It will be
noted that d is dened in terms of the single-valued part of the velocity potential along the body
contour.
Far-eld expansion
To assess the behavior of the ow far from the body, we expand the Greens function and its dipole
in a Taylor series with respect to x about a chosen point, xc , note that
v n dl = 0, (7.8.8)
B
where
d= n(x) d (x) (x xc ) n(x) ud (x) dl(x), (7.8.10)
B
is the coecient of the dipole. We can demonstrate using the divergence theorem that the integral
on the right-hand side of (7.8.10) remains unchanged when the contour of integration, B, is replaced
by any other contour enclosing B.
7.8 Flow past or due to the motion of a two-dimensional body 569
The rst term on the right-hand side of (7.8.9) represents the ow due to a point source
associated with the change in the area occupied by the body. To leading order, far from the body,
the ow behaves like that due to a point source and a point vortex whose strength is equal to the
circulation around the body. In the absence of circulation and when the area occupied by the body
does not change in time, the far ow resembles that due to a two-dimensional potential dipole.
(x) = Vx 1 [x, xc , e(t)] + Vy 2 [x, xc , e(t)] + z 3 [x, xc , e(t)] + + 4 [x, xc , e(t)] , (7.8.14)
2
where i are four fundamental single-valued velocity potentials, xc is the chosen center of rotation,
the unit vector e describes the instantaneous body orientation, and is the polar angle around a
point vortex located inside the body. The square brackets contain the arguments of the fundamental
potentials, i .
Substituting (7.8.14) into (7.8.12) and identifying for convenience xc with the areal centroid
Xc to discard the last term on the right-hand side, we obtain
d = AB (I + ) V + z + + (x Xc ) n(x) v(x) dl(x), (7.8.15)
B
570 Introduction to Theoretical and Computational Fluid Dynamics
where
1 1 1
ij = ni j dl, i = ni 3 dl, i = ni 4 dl, (7.8.16)
AB B AB B AB B
for i = 1, 2. Working as in Section 7.2, we can show that the 22 added-mass matrix is symmetric.
If the circulation around the body is nonzero, the omitted term represented by the dots is innite;
otherwise, it is zero. The nite part of the kinetic energy represented by the rst term on the right-
hand side of (7.8.17) is amenable to the analysis of Section 7.2 for three-dimensional ow, subject
to straightforward changes in notation.
The three terms on the right-hand side represent, respectively, the acceleration reaction, a steady
force, and the buoyancy force. Using the divergence theorem, we nd that the integral on the right-
hand side of (7.8.18) remains unchanged when the contour of integration is replaced by any other
closed contour enclosing B. When the circulation around the body is nonzero, the velocity decays
like 1/d and the integral over a circle with large radius takes the KuttaJoukowski value ez V,
where d is the distance from the origin and ez is the unit vector along the z axis. Substituting this
expression into (7.8.18), we obtain the nal expression
dV
F = AB + ez V AB g, (7.8.19)
dt
which shows that the steady drag force is zero and the lift force is proportional to the circulation
around a translating two-dimensional body, independent of the body shape.
The direction of the lift force for positive circulation is illustrated in Figure 7.8.1(b). Physically,
the circulation accelerates the uid above the body and decelerates the uid below the body, thus
7.9 Computation of two-dimensional ow past a body 571
causing a pressure dierence that results in a lift force. The lifting action of an airfoil hinges on its
ability to generate a sucient amount of positive circulation, so that the lift force counterbalances
the aircraft weight. The generation of circulation relies on viscous eects inside boundary layers
developing around the airfoil, as discussed in Chapter 8.
The torque exerted on a two-dimensional body can be computed working as in Section 7.3 for
three-dimensional ow.
Problems
Singularity representation
The ow due to translation with a specied circulation around the cylinder, C, can be represented
by a point-source dipole with strength
d = 2a2 V, (7.9.1)
and a point vortex with strength = C, both placed at the center of the cylinder. The harmonic
potential and velocity eld are given by
a2 1 2 1
(x) = V x + , u(x) = a2 2 V + 4 (V x) x + e , (7.9.2)
r2 2 r r 2 r
572 Introduction to Theoretical and Computational Fluid Dynamics
where x = x xc , r = |xc | is the distance of the eld point, x, from the instantaneous center of the
cylinder, xc , is the polar angle around the center of the cylinder, and e is the corresponding unit
vector.
Fundamental potentials
Comparing the potential given in (7.9.2) with (7.8.14), we deduce that the four fundamental poten-
tials for translation, rotation, and circulation, are given by
a2 a2
1 = x, 2 = y, 3 = 0, 4 = 0. (7.9.3)
r2 r2
Since the velocity due to the point vortex is parallel to the surface of the cylinder, the last term on
the right-hand side of (7.8.15) is zero.
d = a2 (I + ) V. (7.9.4)
Remembering that d = 2a2 V according to (7.9.1), we obtain = I, which can be substituted into
(7.8.19) to yield the force exerted on the cylinder,
dV
F = a2 + ez V a2 g, (7.9.5)
dt
where the last term represents the buoyancy force.
a2 r
(x) = sin V ex ln . (7.9.6)
r 2 a
where V = V ex . This expression reveals that the structure of the ow is determined by the
dimensionless circulation parameter /(4aV ). Streamline patterns for several values of with
> 0 and V > 0 are shown in Figure 7.9.1. In all cases, the lift force points in the positive direction
of the y axis due to the high velocity, and thus low pressure, on the upper side of the cylinder
compared to that on the lower side.
(a) (b)
3 3
2 2
1 1
0 0
y
y
1 1
2 2
3 3
3 2 1 0 1 2 3 3 2 1 0 1 2 3
x x
(c) (d)
3 3
2 2
1 1
0 0
y
1 1
2 2
3 3
3 2 1 0 1 2 3 3 2 1 0 1 2 3
x x
Figure 7.9.1 Streamline patterns of the ow due to the translation of a circular cylinder in the positive
direction of the x axis with counterclockwise circulation corresponding to circulation parameter (a)
/(4aV ) = 0, (b) 0.25, (c) 0.5, and (d) 1.0.
cylinder discussed in Section 7.9.1. Working in a frame of reference where the ow far from the
cylinder appears to be stationary, we obtain
a2
1 2 1
(x) = 1 + 2 U x + , u(x) = U + a2 2
U 4 (U x) x + e , (7.9.7)
r 2 r r 2 r
where = C is the strength of a point vortex, x = x xc , r = |xc | is the distance of the eld
point x from the center of the cylinder, xc , is the polar angle measured around xc , and e is the
corresponding unit vector. In plane polar coordinates where the x axis is parallel to U, the stream
574 Introduction to Theoretical and Computational Fluid Dynamics
(a) (b)
3 3
2 2
1 1
0 0
y
y
1 1
2 2
3 3
3 2 1 0 1 2 3 3 2 1 0 1 2 3
x x
(c) (d)
3 3
2 2
1 1
0 0
y
1 1
2 2
3 3
3 2 1 0 1 2 3 3 2 1 0 1 2 3
x x
Figure 7.9.2 Streamline patterns of uniform ow past a stationary or rotating circular cylinder for
values of the circulation parameter (a) /(4aU ) = 0, (b) 0.5, (c) 1.0, and (d) 1.2.
function is
a2
r
(r, ) = U r sin ln , (7.9.8)
r 2 a
where U = U ex . Note that the stream function is zero when r = a, as required. Equation
(7.9.8) reveals that the structure of the ow depends on the dimensionless circulation parameter
/(4aU ). Streamline patterns for positive values of with < 0 and U > 0 are shown in
Figure 7.9.2.
7.9 Computation of two-dimensional ow past a body 575
2
N 1
Figure 7.9.3 The ow past a stationary two-dimensional body can be represented by a vortex sheet.
The tangential component of the velocity around the surface of the cylinder is readily found
by dierentiating the stream function (7.9.8) with respect to the radial distance, r, yielding
u (r = a) = = 2U sin + . (7.9.9)
r r=a 2a
We observe that u becomes zero when sin = , and this reveals the onset of a symmetric pair
of stagnation points on the surface of the cylinder when 0 < < 1. For larger values of , the
stagnation points move o the surface of the cylinder and then merge to yield a free stagnation
point inside the ow, as shown in Figure 7.9.2. It is evident from (7.9.9) that, when > 0, the
magnitude of the velocity at the top of the cylinder is higher than that at the bottom of the cylinder.
Bernoullis equation then shows that the surface pressure at the top is lower than that at the bottom,
which provides us with a physical explanation for the occurrence of a lift force toward the positive
direction of the y axis noted by Magnus in 1853.
The strength of the vortex sheet is equal to the tangential component of the uid velocity,
ut = u t, where t is the tangential unit vector pointing in the counterclockwise direction. This
point of view suggests expressing the stream function at an arbitrary point x0 in the ow as
L
1 |x0 x|
(x0 ) = (x0 ) ln ut (x) dl(x) + c, (7.9.10)
2 0 a
where is the stream function of an incident ow, L is the total arc length of the contour of the
body in the xy plane, c is an arbitrary constant, and a is a chosen length.
Applying (7.9.10) at a point at the surface of the body and enforcing the no-penetration
boundary condition, = 0, we obtain an integral equation of the rst kind for the tangential
576 Introduction to Theoretical and Computational Fluid Dynamics
boundary velocity,
1 L
|x0 x|
ln ut (x) dl(x) = (x0 ) + c. (7.9.11)
2 0 a
Numerical methods
A simple numerical method for solving the integral equation (7.9.11) involves tracing the contour of
the body with N marker points arranged in the counterclockwise direction, and then approximating
the contour of the body with a polygonal line connecting successive marker points. Assuming that
the value of ut is constant and equal to ui over the ith segment, we obtain the discrete version of
(7.9.11),
are inuence coecients. The tangential velocities ui can be computed by pointwise collocation.
The methodology involves applying (7.9.12) at the midpoint of each segment, yj = 12 (xj + xj+1 ),
to derive a system of N linear equations,
for j = 1, . . . , N , where summation is implied over the repeated index, i. To ensure that the
circulation around the body has a prescribed value, , we introduce the additional constraint
N
ui |xi xi+1 | = , (7.9.15)
i=1
and solve a system of the N + 1 linear equations comprised of (7.9.14) and (7.9.15) for N + 1
unknowns, ui and c.
The o-diagonal inuence coecients of the inuence matrix, Ai (yj ) for i = j, can be com-
puted using a standard integration method, such as the trapezoidal rule. or a Gaussian quadrature
(Section B.6, Appendix B). The diagonal inuence coecients corresponding to the host collocation
point, Aj (yj ), can be computed analytically as
1 |yj xj |
Aj (yj ) = |yj xj | (ln 1) (7.9.16)
a
for j = 1, . . . , N . An integral identity ensures that the choice of the length, a, is immaterial on the
accuracy of the computations.
7.9 Computation of two-dimensional ow past a body 577
where u u +
t ut is the discontinuity in the tangential velocity across the plate, the plus
superscript indicates evaluation at the upper side of the plate, and the minus superscript indicates
evaluation at the lower side of the plate. Equation (7.9.11) provides us with an integral equation,
a
1 |x0 x|
ln u(x) dx = (x0 ) + c, (7.9.18)
2 a a
where the constant c determines the circulation around the plate, .
In the case of uniform ow with velocity U = (Ux , Uy ), the stream function of the incident
ow is = Ux y Uy x and the solution of the integral equation (7.9.18) is given by
1
u(x) = 2Uy x + (7.9.23)
a2 x2
(Section 7.12.2). When the circulation assumes the Kutta value = 2aUy , the singularity at the
trailing edge, x = a, disappears and uid particles on either side of the plate leave the plate in the
tangential direction without making a sharp turn.
578 Introduction to Theoretical and Computational Fluid Dynamics
Slender-body theory
Slender-body theory provides us with useful approximate solutions for ow past an elongated body
with high aspect ratio. Consider the ow due to the axial translation of a slender two-dimensional
body whose contour is symmetric with respect to the x axis, in the absence of circulation around
the body. The upper and lower surfaces of the body are described by the function y = f (x) for
a < x < a. Working as in the case of ow due to the motion of an axisymmetric body discussed
in Section 7.7, we represent the ow due to translation with velocity Vx along the x axis in terms of
two-dimensional point sources distributed along the centerline. The induced harmonic potential is
Vx a |x |
(x) = ln () d, (7.9.24)
2 a a
where is the density of the distribution, a is a chosen length, and the point = (, 0) lies at the x
axis. The no-penetration condition requires that
ux nx + uy ny = Vx nx (7.9.25)
around the body contour. Introducing the approximations nx df /dx and ny 1 on the upper
surface of the body, and noting that ux and uy have comparable magnitudes while nx is small
compared to ny , we derive the approximation
df
uy (x, f ) Vx . (7.9.26)
dx
Next, we dierentiate (7.9.24) with respect to y to produce uy , obtaining
a (ax)/y
Vx () Vx ()
uy (x) = y d = d, (7.9.27)
2 a (x )2 + y 2 2 (a+x)/y 2 + 1
where AB is the area occupied by the body. Remarkably but fortuitously, this formula produces the
exact answer for a circular cylinder derived in Section 7.9.1.
Problems
Computer Problem
CauchyRiemann equations
The companion harmonic functions and are related through their denition by the Cauchy
Riemann equations,
ux = = , uy = = . (7.10.1)
x y y x
Since and is a pair of conjugate harmonic functions, they can be identied with the real or
imaginary part of an analytic function of a complex variable, z = x + iy,
w(z) + i , (7.10.2)
called the complex potential, where i is the imaginary unit, i2 = 1. A function of a complex
variable, z, is analytic at a point, z0 , if its rst derivative with respect to z is nite and independent
of the direction of dierentiation at every point in a neighborhood of z0 . An entire function is
analytic at each point in the complex plane. In the remainder of this chapter, z = x + iy will denote
the complex variable rather than the Cartesian coordinate normal to the xy plane.
Velocity
Dierentiating (7.10.2) and using the CauchyReimann equations (7.10.1), we obtain the two velocity
components in the complex form
dw
= ux i uy u , (7.10.3)
dz
where u ux + i uy is the complex velocity, and an asterisk denotes the complex conjugate. If
the velocity eld is continuous, the complex potential must be an analytic function of z everywhere
except at isolated singular points. Selecting dierent analytic functions for the complex potential,
w(z), allows us to construct diverse families of incompressible irrotational ows.
Uniform ow
The complex potential of uniform (streaming) ow with velocity U = (Ux , Uy ) in the xy plane is
given by
w = (Ux i Uy ) z = U z, (7.10.4)
7.10 Complex-variable formulation of two-dimensional ow 581
0.8
1
0.6
y
x 0.4
0.5
0.2
0
0
0 0.5 1 1.5 0 0.2 0.4 0.6 0.8 1
x x
0.5 0.5
0.5
y
0 0
y
0.5
0.5 0.5
1
1 1 1.5
1.5 1 0.5 0 0.5 1 1.5
1 0.5 0 0.5 1 1 0.5 0 0.5 1 x
x x
Figure 7.10.1 (a) Illustration of potential ow inside or around a corner with aperture angle . (be)
Streamline patterns of ow in a corner with aperture angle (b) = /4, (c) /2, (d) 3/4, and
(e) 2. (f) Streamline pattern of the ow due to a periodic array of point sources arranged along
the x axis; the x and y axes have been scaled with half the point source separation.
where U = Ux + iUy is a complex velocity. The potential function, stream function, and complex
velocity are given by
dw
= Ux x + Uy y, = Ux y Uy x, = u = U x i U y . (7.10.5)
dz
Flow inside and around a corner
The complex potential of irrotational ow inside or around a corner with aperture angle = /m,
as shown in Figure 7.10.1(a), is given by
w = Az m , (7.10.6)
where A is a real constant with appropriate dimensions. In plane polar coordinates, (r, ), dened
such that x = r cos , y = r sin , and z = r exp(i), one wall is located at = 0 and the second wall
is located at = . The potential function, stream function, and complex velocity are given by
dw
= Arm cos(m), = Arm sin(m), = u = Amz m1 . (7.10.7)
dz
582 Introduction to Theoretical and Computational Fluid Dynamics
Point source
m z z0 m |z z0 | m |z z0 | dw m z z0
w= ln , = ln , = arg , =
2 a 2 a 2 a dx 2 |z z0 |2
Point-source dipole
dx + i dy 1 1 dx (x x0 ) + dy (y y0 )
w= , =
2 z z0 2 |z z0 |2
1 dx (y y0 ) dy (x x0 ) dw dx + idy 1
= , =
2 |z z0 |2 dx 2 (z z0 )2
Point vortex
z z0 |z z0 | |z z0 | dw z z0
w= ln , = arg , = ln , =
2i a 2 a 2 a dx 2i |z z0 |2
Table 7.10.1 The complex potential, w, the corresponding real harmonic potential, , the stream
function, , and the complex velocity eld, dw/dz = ux iuy of elementary potential ows due
to singularities located at z0 . An asterisk designates the complex conjugate.
Streamline patterns for several angles aperture angles are shown in Figure 7.10.1(be). When = ,
corresponding to m = 1, we obtain uniform streaming ow along the x axis. When = 2,
corresponding to m = 1/2, we obtain ow around a semi-innite at plate represented by the
positive part of the x axis.
The complex potential and velocity due to a point source, a point-source dipole, and a point vortex
are given in Tables 7.10.1. The complex potential and velocity due to a periodic array of point
sources and a periodic array point vortices are given in 7.10.2. The streamline pattern of the ow
due to a periodic array of point sources is shown Figure 7.10.1(f). The streamline pattern due to a
periodic array of point vortices is shown in Figure 2.10.1.
7.10 Complex-variable formulation of two-dimensional ow 583
m k
w= ln sin (z z0 )
2 2
m
= ln cosh k(y y0 ) cos k(x x0 )
4
m k k k k
= arg sin (x x0 ) cosh (y y0 ) + i cos (x x0 ) sinh (y y0 )
2 2 2 2 2
dw m k
= cot (z z0 )
dz 2a 2
k
w= ln sin (z z0 )
2i 2
k k k k
= arg sin (x x0 ) cosh (y y0 ) + i cos (x x0 ) sinh (y y0 )
2 2 2 2 2
= ln cosh k(y y0 ) cos k(x x0 )
4
dw k
= cot (z z0 )
dz 2ai 2
Table 7.10.2 The complex potential and velocity eld due to a periodic arrangement of point sources
or point vortices along the x axis; a is the period, k = 2/a is the wave number, and z0 is the
location of one point source or point vortex in the array.
where y0im is the image of the point source with respect to the upper or lower wall. Since the two
images are separated by distance 2h, the choice is inconsequential. For example, of the lower wall
is located at y = yw , then the image point with respect to the lower wall is y0im = 2yw y0 and the
corresponding image point with respect to the upper wall is y0im = 2yw + 2h y0 . Setting m = 1
we obtain the Greens function of the second kind (Neumann function) of Laplaces equation whose
normal derivative vanishes over the lower and upper walls.
7.10.2 Computation of the complex potential from the potential or stream function
The theory of harmonic functions provides us with a method of computing the stream function from
the potential function, and vice versa, yielding the complex potential in terms of one its harmonic
components. The procedure involves integrating the CauchyRiemann equations along a path in
the complex plane, obtaining
y x
(x, y) = (x0 , y0 ) (x0 , y ) dy + (x , y0 ) dx ,
y0 x x0 y
y x
(x, y) = (x0 , y0 ) + (x0 , y ) dy (x , y0 ) dx , (7.10.10)
y0 x x0 y
where (x0 , y0 ) is an arbitrary point where the potential function and stream function are arbitrarily
assigned [106]. The integrals can be computed by analytical or numerical methods.
Integrating the series with respect to z and setting for convenience the integration constant
to zero, we obtain the expansion
m i z z0 bn
w(z) = U (z z0 ) + ln + b0 + , (7.10.11)
2 a n=1
(z z0 )n
7.10 Complex-variable formulation of two-dimensional ow 585
where a is a constant length and bn are constant complex coecients. Referring to Table 7.10.1, we
recognize the second term on the right-hand side of (7.10.11) as a point source and a point vortex
located at the point z0 . The strength of the point source, m, is equal to the ow rate across a closed
contour enclosing the body. The strength of the point vortex, , is equal to the cyclic constant of
the motion around the body. For example, in the case of ow due to the isotropic expansion of a
circular bubble centered at the point z0 , and all coecients bn in (7.10.11) are zero.
1 1
F = i p dz = i dz = i dw , (7.10.12)
Body 2 Body dz dz 2 Body dz
where p is the real hydrodynamic pressure, is the uid density, the path of integration is taken
in the counterclockwise direction, and an asterisk denotes the complex conjugate. Since the stream
function is constant and the complex potential is real around the body contour, we can replace dw
in the last integral with dw, obtaining
dw
2
1
F = i dz. (7.10.13)
2 Body dz
Assuming that the functions inside the last integrals in (7.10.13) and (7.10.14) are analytic through-
out the domain of ow, we use Cauchys integral theorem to replace the integrals along the contour of
the body with corresponding integrals along another closed contour enclosing the body. Introducing
the Laurent series of the complex potential stated in (7.10.11) allows us to compute the integrals in
terms of the coecients of the far-eld expansion.
Uniform ow
As an application, we consider uniform ow with velocity U past a stationary body. Substituting
into (7.10.13) the Laurent series (7.10.11) with m = 0, we obtain
2
1 1 nbn
F = i U + dz. (7.10.15)
2 Body 2i z z0 i=1 (z z0 )n+1
1 1
F = i U + U + dz, (7.10.16)
2 Body i z z0
586 Introduction to Theoretical and Computational Fluid Dynamics
where the three dots represent terms that decay at a quadratic or higher rate. Evaluating the integral
by the method of residues, we obtain the real components of the force,
Fx = Uy , Fy = Ux , (7.10.17)
which shows that uniform incident past a body combined with circulation around the body generates
a lift force that is independent of the body shape. These expressions are in agreement with our earlier
results in Section 7.8 obtained by dierent methods.
Working in a similar fashion, we nd that the hydrodynamic torque with respect to a point,
z0 , is given by
T = 2 Imag{ U b1 }. (7.10.18)
Because the complex constant b1 depends on the body shape, the result for the torque is less general
than that for the force.
Fx = Vy , Fy = Vx , (7.10.19)
in agreement with our previous results in Section 7.8. A lift force is established when Vx > 0 in the
presence of positive circulation, > 0.
a2 z zc
w(z) = V + ln , (7.10.20)
z zc 2i a
where is the circulation around the cylinder, zc is the instantaneous center of the cylinder, and
V = Vx +iVy is the complex velocity of translation. The rst term on the right-hand side of (7.10.20)
represents a point-source dipole and the second term represents a point vortex.
ow due to a point vortex. The complex potential (7.10.21) can be used to derive the complex
potential of uniform ow past a body with arbitrary cross-section using the method of conformal
mapping, as discussed in Section 7.11.
a2
w(z zc ) = w (z zc ) + w ( ), (7.10.22)
z zc
where w (z) w (z ) and a is the cylinder radius.
Uniform ow
As an example, we consider uniform ow past a circular cylinder with vanishing circulation around
the cylinder, and set w = U z. The restriction of vanishing circulation stems from the required
absence of singularities inside the cylinder. To apply the circle theorem, we write
w (z zc ) = U (z zc ) + U zc (7.10.23)
a2
w(z) = U z + U + U zc , (7.10.24)
z zc
which is consistent with (7.10.21). The inconsequential constant U zc on the right-hand side plays
no role on the velocity eld.
m z zs a (zs zc )
w(z) = ln + ln . (7.10.27)
2 a z zc a
588 Introduction to Theoretical and Computational Fluid Dynamics
Rearranging, we obtain
m z zs z zc (z zc )(zs zc )
w(z) = ln ln + ln 1 (7.10.28)
2 a a a2
and then
m z zs z zc a (zc zs )
w(z) = ln + ln + ln . (7.10.29)
2 z zc a (zs zc ) a
The last term on the right-hand side of (7.10.29) can be discarded as inconsequential. The resulting
expression shows that the disturbance ow due to the cylinder can be represented in terms of a
point sink with strength m located at the center of the cylinder and a point source with strength
m located at the inverse point of the primary point source with respect to the cylinder,
a2 a2
zsinv = zc + 2
(zs zc ) = zc + . (7.10.30)
|zs zc | (zs zc )
The real part of (7.10.29) with m = 1 is the two-dimensional Greens function of the second kind
or Neumann function, N (z, z0 ), in the exterior of a circular cylinder.
Problems
y
= F(z)
d z2
d z1 d
1
z
0 d2
x 0
Figure 7.11.1 Conformal mapping of the z plane to the plane using the mapping function = F (z).
The angle subtended between two innitesimal vectors is preserved after mapping.
= F (z), (7.11.1)
as shown in Figure 7.11.1. When the function F (z) is multivalued, we introduce an appropriate
branch cut in the z plane so as to render the mapping unique. The image of an open or closed line
in the z plane is another open or closed line in the plane with dierent orientation and shape.
z = f (). (7.11.2)
When the function f () is multi-valued, we introduce an appropriate branch cut in the plane so
as to render the inverse mapping unique.
590 Introduction to Theoretical and Computational Fluid Dynamics
Dierential vectors
Dierentiating F (z) and f () with respect to z or , we nd that the ratio of two corresponding
innitesimal dierential vectors in the z and planes, dz and d, are related by
d dz
= F (z) = |F (z)| exp i Arg{F (z)} , = f () = |f ()| exp i Arg{f ()} , (7.11.3)
dz d
where Arg denotes the argument of a complex number, dened as the polar angle around the origin
of the complex plane. Since the functions F (z) and f () are analytic, the derivatives in equations
(7.11.3) depend on z or but are independent of the orientation of corresponding dierential pairs,
dz and d.
The rst equation in (7.11.3) states that the length of the dierential vector d is equal to the
length of dz multiplied by the scalar factor |F (z)|. The direction of d is rotated with respect to
that of dz by an angle that is equal to the argument of F (z). A singular behavior is expected at a
critical point of the mapping (7.11.1) where F (z) becomes innite or, equivalently, f (z) is zero. A
similar interpretation of the dierentials pertains to the second equation in (7.11.3).
The CauchyRiemann equations associated with the inverse mapping function, f (), take the
corresponding forms
x y x y
= , = . (7.11.7)
2x 2x 2y 2y
+ 2 = 0, + 2 = 0. (7.11.8)
2 2
x x
= 0, (7.11.9)
where x is the real position vector in the z plane. This condition is the cornerstone of orthogonal
grid generation by conformal mapping, which involves producing the images of lines of constant
and constant and identifying their intersections as computational nodes (e.g., [317]).
= F (z) = az + b, (7.11.10)
multiplies the distance from the origin, |z|, by the scalar factor |a|, rotates a line connecting the
origin to a point z by an angle that is equal to the argument of a, and then shifts every point by
the complex number b. Circles in the z plane remain circles in the plane. The inverse mapping
function,
b
z = f () = , (7.11.11)
a
performs a similar transformation.
Mobius transformation
The Mobius transformation, also called the partial fractional transformation,
Az + B
= F (z) = , (7.11.12)
Cz + D
592 Introduction to Theoretical and Computational Fluid Dynamics
maps the whole complex plane onto itself, where AD are four complex constants. Circles and
straight lines in the z plane are mapped to circles or straight lines in the plane, and vice versa.
When C = 0, we obtain a linear transformation with shift. The inverse transformation,
D B
z = f () = , (7.11.13)
C + A
Exponential function
The exponential function,
2z
= F (z) = exp , (7.11.14)
b
where b and are two real and positive constants, maps the semi-innite horizontal strip x > 0 and
0 < y < b to the exterior a circle of radius centered at the origin, as shown in Figure 7.11.2(a).
The vertical side of the strip is mapped to the circle, and the two semi-innite horizontal sides are
mapped to a semi-innite section of the axis, > . The inverse mapping function,
b
z = f () = ln , (7.11.15)
2
becomes unique by introducing a branch cut along the positive axis, > , and specifying that
the argument of takes values in the range [0, 2).
The same exponential function (7.11.14) maps the rectangle conned between 0 < x < a and
0 < y < b to an annulus with inner radius and outer radius
2a
= exp , (7.11.16)
b
as shown in Figure 7.11.2(b). As the aspect ratio a/b tends to innity, the annulus reduces to the
exterior of a disk of radius .
b
z = f () = ln , (7.11.18)
2
becomes unique by introducing a branch cut along the positive axis, 0 < < , and specifying
that the argument of takes values in the range [0, 2).
7.11 Conformal mapping 593
(a)
y
b o
x
o
(b)
y
b o
x o
o
a
Figure 7.11.2 The exponential function maps (a) a semi-innite strip to the exterior of a circular
disk, and (b) a rectangle to an annulus.
Logarithmic function
The logarithmic function,
b z
= F (z) = ln , (7.11.19)
2 a
where a and b are two real and positive constants, maps the z plane to the innite horizontal strip
< < , and 0 < < b. The mapping becomes unique by introducing a branch cut along the
positive x axis. The inverse mapping function is
2
z = a exp . (7.11.20)
b
The same function (7.11.19) maps the upper half z plane to the innite horizontal strip < < ,
and 0 < < 12 b.
g g g g g g
= + , = + , (7.11.22)
x x x y y y
g = H g, (7.11.23)
Transformation metric
Using the CauchyRiemann equations (7.11.4), we derive an expression for the determinant of the
Jacobian matrix,
2
2
2
2
h2 det(H) = + = + = |F (z)|2 , (7.11.25)
x y x y
where h = |F (z)| is the metric of the transformation. The inverse of the Jacobian matrix is
1
H1 = HT , (7.11.26)
h2
H HT = h2 I, (7.11.27)
Gradient projection
Using (7.11.23), we nd that the projection of the gradients of two arbitrary functions, g1 and g2 ,
at corresponding points is given by
1 ) (H g
(g1 ) (g2 ) = (H g 2 ) = g
1 (HT H) g
2 = h2 (g
1 ) (g
2 ). (7.11.28)
Setting g1 = g2 = g, we nd that
|g| = h |g|. (7.11.29)
Expression (7.11.28) reveals that, if the vectors g1 and g2 are orthogonal, (g1 ) (g2 ) = 0, the
1 and g
vectors g 2 are also orthogonal, (g 1 ) (g
2 ) = 0.
7.11 Conformal mapping 595
Convectiondiusion in irrotational ow
As an application, we consider a steady temperature eld, T (x, y), governed by the steady convection
diusion equation in a two-dimensional irrotational ow with velocity u = ,
T T
ux + uy = u T = () (T ) = 2 T, (7.11.33)
x y
where is the thermal diusivity. Using the preceding relations for the gradient and Laplacian, we
nd that T (, ) satises the same equation in the plane,
(T
() ) =
2 T. (7.11.34)
Accordingly, a solution in the transformed plane can be used to produce a solution in the physical
plane, and vice versa, subject to appropriate boundary conditions [26].
These properties allow us to identify with the potential function and with the stream
function of a ow in the plane, so that
dW
= U i U , (7.11.37)
d
where U = (U , U ) is the velocity in the plane. To nd the relation between the magnitude and
direction of the velocity in the z plane, u, and the corresponding velocity in the plane, U, we use
the chain rule to write
dW dw dz dw
= = f (), (7.11.38)
d dz d dz
which shows that the ratio of the two complex velocities is given by
U i U
= f () = |f ()| exp i arg[f ()] . (7.11.39)
ux i uy
Equation (7.11.39) allows us to compute u from U, and vice versa, in terms of the mapping function
f (). Comparing (7.11.39) with the second equation in (7.11.3), we nd that dierential vectors and
velocities are amplied in inverse proportion so that the ow rates across corresponding dierential
line elements are the same in both planes.
In the case of a point source in the z plane, we choose a reference length, a, and write
m z z0 m f () f (0 )
W () = w[f ()] = w(z) = ln = ln . (7.11.40)
2 a 2 a
Expanding f () in a Taylor series about the singular point 0 , we nd we nd that
m 0 m 0
W () ln[f (0 ) ] ln , (7.11.41)
2 a 2 a
which shows that the ow in the plane contains a point source with identical strength placed at
the image point, 0 . An exception occurs when the point z0 happens to be a critical point of the
conformal mapping. Replacing m with i, we obtain a corresponding result for a point vortex with
strength .
y n
L
x
Figure 7.11.3 Mapping of a ow domain in the z plane to a corresponding domain in the plane.
An impermeable boundary in the z plane remains impermeable in the plane.
The Neumann boundary condition remains a Neumann boundary condition but undergoes
a quantitative transformation. To see this, we express the normal unit vector in the z plane as
n = (dy/dl, dx/dl) and apply the chain rule to obtain
dy dx dy dx
dy dx
n = = + , (7.11.43)
x dl y dl x dl y dl x dl y dl
where l is the arc length around the boundary, as shown in Figure 7.11.3. Now using the Cauchy
Riemann equations, we nd that
dy dx
dy dx
d d
n = + + = (7.11.44)
y dl x dl y dl y dl l l
and then
dL = |F (z)| N ,
n = N (7.11.45)
dl
where N is the normal unit vector pointing into the ow in the plane and L is the arc length
along the transformed boundary, as shown in Figure 7.11.3. Equation (7.11.45) ensures that an
impermeable boundary in the z plane where n = 0 remains impermeable in the plane, that
= 0.
is, N
b
x
a
Figure 7.11.4 The exterior of a circle of radius a in the z plane is mapped to the upper half-plane
using the linear fractional transformation (7.11.47). Corresponding points are shown with the same
symbol. Uniform ow past a cylinder with nonzero circulation around the cylinder in the z plane
transforms to ow due to a point-source dipole and a point vortex supplemented by their images
in the plane. Streamlines for zero circulation are shown.
z+a + ib
= F (z) = ib , z = f () = a , (7.11.47)
za ib
where b is a real positive constant, as shown in Figure 7.11.4. Innity is mapped to the rst singular
point = i b, and the center of the circle is mapped to the second singular point = i b.
Substituting into (7.11.46) the inverse transformation given by the second expression in
(7.11.47), we derive the complex potential of the ow in the plane,
+ ib ib + ib
W () = U a + Ua + ln . (7.11.48)
ib + ib 2i ib
Referring to Table 7.10.1, we nd that the ow in the plane is represented by a point-source dipole
with strength d = 4iabU placed at the rst singular point, = ib, and a point vortex with
strength placed at the second singular point, = ib. Both singularities are accompanied by their
images with respect to the wall to satisfy the no-penetration condition.
7.11 Conformal mapping 599
v
W () = ln . (7.11.50)
2i v
Substituting the conformal mapping function given by the rst expression in (7.11.47) and rearrang-
ing, we obtain the corresponding complex potential of the ow in the z plane,
(z + a)(zv a) (z a)(zv + a) z a
w(z) = ln + ln v . (7.11.51)
2i (z + a)(zv a) + (z a)(zv + a)
zv a
Simplifying the rst fraction and discarding the inconsequential constant last fraction, we obtain
zv z zv z
Discarding the constant last term, we nd that the ow consists of a primary point vortex in the z
plane and a reected point vortex with opposite strength located at the image of the point vortex
with respect to the cylinder, in agreement with (7.10.32).
A three-parameter family of functions = F (z) can be found that map a simply connected
domain, D, to another simply connected domain D . To remove the three degrees of freedom, we
generalize the denition of a point in the complex plane to include innity and choose one of the
following three sets of conditions:
600 Introduction to Theoretical and Computational Fluid Dynamics
1. Stipulate that an arbitrary point, z0 , inside D is mapped to the origin, = 0, and specify the
direction of an innitesimal vector starting at z0 whose image lies on the axis.
2. Stipulate that an arbitrary point, z0 , inside D is mapped to the origin = 0, and another
arbitrary point 1 at the boundary of D is mapped to an arbitrary point 1 at the boundary
of D .
3. Stipulate that three arbitrary points, z0 , z1 , and z2 , at the boundary of D are mapped in the
same order to three arbitrary corresponding points 0 , 1 , and 2 at the boundary of D .
As an example, the Mobius transformation (7.11.12) is dened by specifying the three ratios A/D,
B/D, and C/D.
Problems
F (z) F (z0 )
w(z) = ln . (7.11.53)
2i a F (z)F (z0 )/a
1 F (z) F (z0 )
w(z) = ln . (7.11.54)
2 a F (z)F (z0 )/a
x
c c a
Figure 7.12.1 Mapping the exterior of an ellipse with major and minor semiaxes a and b in the z
plane to the exterior of a circle of radius in the plane.
2
W () = U + U + ln . (7.12.1)
2i
To ensure that the far ows in the two complex planes behave in a similar manner so that uniform
ow far from the disk in the plane is uniform ow far from the body in the z plane, we require
that the mapping function, = F (z), and its inverse, z = f (), are linear functions far from the
body, as |z| or || tends to innity, so that their rst derivatives tend to a constant.
1 c2
z = f () = + , (7.12.3)
4
projects the exterior of a disk of radius = 12 (a + b) centered at the origin of the plane to the
exterior of the ellipse in the z plane. Since f () exhibits the required linear behavior at innity, it
is acceptable for the study of uniform ow.
602 Introduction to Theoretical and Computational Fluid Dynamics
Decomposing (7.12.3) into its real and imaginary parts, we obtain the explicit coordinate
transformations
1 c2 1 c2
x= 1+ , y = 1 , (7.12.4)
4 ||2 4 ||2
where c is the focal length and ||2 = = 2 + 2 . The forward transformation mapping the
exterior of the ellipse to the exterior of the disk is found by solving the quadratic equation (7.12.3)
for . We note that the root with the negative sign corresponds to a point inside the ellipse and
choose the root with the positive sign to obtain
1
= F (z) = z + z 2 c2 . (7.12.5)
2
The square root on the right-hand side becomes unique by introducing a branch cut along the x axis
between the two focal points, extending from c to c.
The complex potential of the ow in the z plane is found by substituting (7.12.5) into (7.12.1)
and setting = 12 (a + b) to obtain
1
1 (a + b)2 z + z 2 c2
w(z) = U z + z 2 c2 + U + ln . (7.12.6)
2 2 z + z 2 c2 2i a+b
When a = b, we recover the complex potential for ow past a circular cylinder of radius a.
1 a2
z = f () = + , (7.12.7)
4
mapping the exterior of a disk of radius = 12 a centered at the origin of the plane to the
whole complex z plane. The circular contour of the disk is mapped to the at plate. The inverse
transformation (7.12.5) with c = a becomes
1
= F (z) = z + z 2 a2 , (7.12.8)
2
where the branch cut of the square root coincides with the length of the plate. A dierent method
of deriving (7.12.7) is discussed in Problem 7.13.1.
Substituting (7.12.8) into (7.12.1) and setting = 12 a, or else setting in (7.12.6) b = 0, provides
us with the complex potential of the ow in the z plane,
1
1 a2 z + z 2 a2
w(z) = U z + z a + U 2 2 + ln . (7.12.9)
2 2 z + z 2 a2 2i a
7.12 Applications of conformal mapping to ow past a two-dimensional body 603
The tangential velocity on the upper or lower surfaces of the plate, indicated by a plus or
minus superscript, is
1
ux (x, y = 0)) = Ux Uy x + (7.12.12)
2 a x2
2
for a x a. We observe that the velocity diverges at both ends of the plate, x = a. The force
exerted on the plate follows from (7.10.17). Streamline patterns for several ow congurations are
shown in Figure 7.12.2.
KuttaJoukowski condition
The KuttaJoukowski condition requires that the circulation established around the plate is such
that the velocity is nite at the trailing edge and the two uid streams merge smoothly above and
below the airfoil (e.g., [99]). Physically, the action of viscosity during the unsteady startup process
is such that, when the nal potential ow state is established, viscous eects are signicant only
insofar as to ensure that the KuttaJoukowski condition is satised. Equation (7.12.12) shows that,
if the trailing edge is located at x = a, the circulation around a at plate necessary to satisfy the
KuttaJoukowski condition is
= 2aUy . (7.12.13)
Accordingly, when /(2aUy ) = 1, the streamlines merge smoothly at the trailing edge, as
shown in Figure 7.12.2(b).
The analytical computation of potential ow past an airfoil with a specied shape is generally
intractable. Early work in aerodynamics before the advent of high-speed computers concentrated
on families of airfoil shapes generated by mapping a circle using carefully crafted transformations.
One important family of shapes is generated by the Joukowski transformation,
2
z = f () = + , (7.12.14)
604 Introduction to Theoretical and Computational Fluid Dynamics
(a) (b)
2 2
1.5 1.5
1 1
0.5 0.5
y/a
y/a
0 0
0.5 0.5
1 1
1.5 1.5
2 2
2 1.5 1 0.5 0 0.5 1 1.5 2 2 1.5 1 0.5 0 0.5 1 1.5 2
x/a x/a
(c) (d)
2 2
1.5 1.5
1 1
0.5 0.5
0 0
y
0.5 0.5
1 1
1.5 1.5
2 2
2 1.5 1 0.5 0 0.5 1 1.5 2 2 1.5 1 0.5 0 0.5 1 1.5 2
x x
Figure 7.12.2 (a, b) Streaming ow past a at plate with incident velocity Ux = Uy and circulation
parameter (a) /(2aUy ) = 0 or (b) 1 (KuttaJoukowski value). (c, d) Streamline patterns
for incident velocity Ux = 0 and (c) = 0 or (d) 1 (KuttaJoukowski value).
where is a specied length, as shown in Figure 7.12.3. It will be noted that (7.12.14) includes as
special cases (7.12.7) and (7.12.3), corresponding to the at plate and the ellipse. An alternative
version of (7.12.14) is
z 2
2
= . (7.12.15)
z + 2 +
The critical points of the Joukowski transformation where df /d = 0 are located at = ,
corresponding to z = 2. A circle centered at the origin of the plane and passing through both
singular points is mapped to a at plate with half length equal to a = 2 in the z plane. A smooth
7.12 Applications of conformal mapping to ow past a two-dimensional body 605
x
2 2
Figure 7.12.3 Mapping with the Joukowski transformation. The bold circle in the plane transforms
into a plate in the xy plane, and the dashed circle transforms into an airfoil.
curve in the plane passing through the rst singular point, = , and enclosing the second
singular point, = , is mapped to a cusped curve in the z plane. The cusp is located at the image
of the rst singular point, as shown in Figure 7.12.3.
where a(n) are complex coecients determined by the body shape and orientation. The linear term
on the right-hand side of (7.12.16) satises the requirement that df /d tends to unity as tends to
innity. The constant a(0) determines the relative position of the body in the z plane and can be set
to zero. A nite set of subsequent coecients can be computed by stipulating that a collection of
points around the body contour are mapped to a corresponding collection of points on the circular
contour in the plane. However, the precise location of the image points must be computed as a
part of the solution.
606 Introduction to Theoretical and Computational Fluid Dynamics
Alternatively, nonstandard body shapes can be generated by keeping selected terms in the sum
on the right-hand side of (7.12.16). For example, to obtain rounded squares, we set all coecients
to zero except for the third coecient, a(3) < 0, according to the Lewis transformation. An innite
sum that yields a perfect square is available (e.g., [210]).
Numerical methods
To formalize a numerical method, we map the exterior of a smooth body to the exterior of the
disk of radius centered at the origin of the plane. The position around the circular contour is
= exp(i), where is the polar angle and i is the imaginary unit. Equation (7.12.16) provides
us with a Fourier expansion,
z() = ei + a(0) + a(n) eni . (7.12.17)
n=1
Separating the real part (R) from the imaginary part (I), we obtain
and
1
= x() cos + y() sin d, (7.12.20)
2 0
and
2
(n) 1
aR = x() cos(n) y() sin(n) d,
2 0
2
(n) 1
aI = x() sin(n) + y() cos(n) d (7.12.21)
2 0
1. Trace the contour of the body with N points distributed in the counterclockwise direction,
located at (xi , yi ), for i = 1, . . . , N .
2. Assign values for to the marker points, i .
3. Reconstruct the functions x() and y() by cubic-spline interpolation.
4. Compute and the coecients a(n) for n = 0, 1, . . . , nmax by performing the integration in
(7.12.20) and (7.12.21) analytically or using the trapezoidal rule implemented by a fast Fourier
transform, where nmax is a specied truncation level (e.g., [317]).
7.12 Applications of conformal mapping to ow past a two-dimensional body 607
Figure 7.12.4 Grids of nodes generated by mapping the exterior of a body described by a chain a
marker points, indicated by circles, to the exterior of a disk.
5. Compute the right-hand side of expressions (7.12.18) and (7.12.19), and formulate the dier-
ences between the specied and computed coordinates, xi xi x(i ) and yi yi y(i ).
6. Adjust the values i to either drive the squares of the distances |zi |2 to zero or minimize the
)N
positive functional i=1 |zi |2 , using, for example, Newtons method.
Grids of points generated using the minimization method for truncation level nmax = N/2 are
shown in Figure 7.12.4. The radial lines pass through a chain of marker points distributed around
the body. The dierence between the specied and reconstructed body contour diminishes rapidly
as more terms are retained in the expansion.
Another way of obtaining the mapping function is by mapping the exterior of a body of
interest to the interior of an auxiliary body using the inverse transformation = 1/z, where the
origin lies inside the body. At the second stage, we map the interior of the auxiliary body to the
upper half-plane or interior of a disk. However, even if the original body has a polygonal shape, the
auxiliary body will have a curved shape, which precludes the application of specialized methods for
polygons. Other methods of computing a function that maps the interior of a body to a disk involve
solving an integral equation over the body contour and developing a variational formulation (e.g.,
[65]).
608 Introduction to Theoretical and Computational Fluid Dynamics
= exp( 0 ), (7.12.22)
where 0 is a specied threshold. The left end of the strip, R = 0 , is mapped to the contour of the
body. An orthogonal grid can be generated by mapping lines of constant R or constant I inside
the strip to the physical z plane.
For example, substituting (7.12.22) into the inverse mapping function (7.12.3) for an ellipse
with major semiaxis a and minor semiaxis b, we obtain
It is convenient to set 0 = atanh(b/a) and obtain the simplied expression z = A cosh , where
c2
A = a cosh 0 b sinh 0 = cosh 0 , (7.12.24)
a
and c is the focal length of the ellipse. The real and imaginary parts of provide us with elliptic
coordinates.
Problems
A2 = yc21 a21 = yc22 a22 = (yc2 2d)2 a21 = (yc1 + 2d)2 a22 , (7.12.27)
z 2
q
= , (7.12.28)
z + 2 +
where 1 < q < 2. Demonstrate that a circle centered at the origin of the plane and passing through
both singular points, = transforms into two circular arcs passing through the points z = q.
Computer Problem
y
4
3
i
i
N
x
N 1 2
Figure 7.13.1 One version of the SchwarzChristoel transformation maps a polygon in the z plane
to the upper half plane.
polygon and all angles i are equal for a regular polygon. In all cases, the sum of the angles satises
the geometrical constraint
N
i = 2, (7.13.1)
i=1
where positive and negative angles may appear inside the sum. For example, in the case of a square,
N = 4 and i = 12 for i = 14. The SchwarzChristoel transformation provides us with the
inverse mapping function in the dierential form
dz #N
1
=c , (7.13.2)
d i=1
( /
i) i
where c is a complex constant and n are the images of the vertices of the polygon along the real
axis of the plane. The associated integral form is
#
N
1
z = f () = z0 + c d, (7.13.3)
0 i=1
( i )i /
where z0 and 0 are a pair of corresponding points and is an integration variable. The complex
powers in (7.13.2) and (7.13.3) are computed best by setting, for example,
i
( i )i / = exp ln( i ) , (7.13.4)
where the branch cut of the logarithmic function is the negative real axis. Given the polygon shape,
the problem is reduced to specifying or computing the images of the vertices, i , and the values of
the three constants, c, z0 , and 0 . Riemanns mapping theorem discussed in Section 7.11.10 allows
us to arbitrarily choose the images of three vertices, i , and compute the rest of the vertices so as
to map the polygon to the upper half-plane. It is permissible to set 1 = or N = , in which
case the corresponding factors in the product in (7.13.2) or (7.13.3) do not appear.
7.13 The SchwarzChristoel transformation 611
1
d 1
3
x 1 2 3
2
0 b a
0 1
2 3
Figure 7.13.2 Mapping a triangle in the z plane to the upper half plane. The dot-dashed line is
the image of the axis. Corresponding streamlines are shown as dashed lines.
Triangle
In the simplest application, we map the interior of a triangle with vertices at the points z1 = b + i d,
z2 = 0, and z3 = a, to the upper half plane, where a, b, and d are three real positive constants,
as illustrated in Figure 7.13.2. The Riemann mapping theorem allows us to arbitrarily select the
images of all three vertices. It is convenient to set 1 = , 2 = 0, and 3 = 1. Applying (7.13.3)
with z0 = 0 and 0 = 0, we nd that
d
z = f () = c 2 / ( 1)3 /
. (7.13.5)
0
Unfortunately, it is not possible to perform the integration analytically for an arbitrary triangular
shape. To evaluate the constant c, we require that f (1) = a and obtain
1
d c (1 2 /) (1 3 /)
a=c = . (7.13.6)
0
2 / ( 1)3 / (1)3 / (2 2 / 3 /)
The values of the gamma function, , can be read from tables or approximated by series expansions
(e.g., [2], p. 255). Substituting (7.13.6) into (7.13.5), we obtain
(2 2 / 3 /)
d
z = f () = a . (7.13.7)
(1 2 /) (1 3 /) 0 2 / (1 )3 /
The axis is mapped to the dot-dashed line connecting the second to the rst vertex in the xy
plane, as shown in Figure 7.13.2.
To compute the singular integral on the right-hand side of (7.13.7), denoted by I, we subtract
out the singularity at the origin by restating the integral as
d 1
I=J + =J + 12 / , (7.13.8)
0
2 / 1 2 /
612 Introduction to Theoretical and Computational Fluid Dynamics
where
1 1
J 1 d. (7.13.9)
0 2 / (1 )3 /
The integral on the right-hand side of (7.13.9) can be computed by standard numerical methods, as
discussed in Section B.6, Appendix B.
Uniform ow along the axis in the plane corresponds to ow due to a point-source dipole
placed at the vertex z1 in the z plane, and oriented perpendicular to the bisector of the corresponding
triangle angle. A streamline is illustrated with the dashed lines in Figure 7.13.2.
Semi-innite strip
As the elevation of the rst vertex, d, tends to innity while b lies between 0 and a, the triangle
becomes a vertical semi-innite strip with width a. Setting 2 = 3 = 12 , we obtain
21
d d
z = f () = c =c = i c [ arcsin(2 1) + ], (7.13.10)
0 1/2 ( 1)1/2 1 (2 1)1/2 2
1
where = 2 1 and = 2 ( + 1). Requiring that f (1) = a, we nd that c = ia/ and obtain he
mapping function
a
z = f () = [ arcsin(2 1) + ]. (7.13.11)
2
The forward mapping function follows by inversion,
1 z
= F (z) = 1 cos . (7.13.12)
2 a
Uniform ow along the axis in the plane corresponds to ow descending along the left vertical
side of the strip and leaving up along the right side of the strip in the z plane (Problem 7.13.4).
Rectangle
Next, we map the interior of a rectangle with vertices at the points
z1 = a + i b, z2 = a, z3 = a, z4 = a + i b, (7.13.13)
to the upper half plane, as shown in Figure 7.13.3. Motivated by the symmetry of the rectangle,
we specify that 2 = 1 and 3 = 1. Requiring that the origin of the z plane is mapped to the origin
of the plane, we set z0 = 0 and 0 = 0. Anticipating the symmetry of the solution, we set 1 =
and 4 = , where > 1 is a real constant to be computed as part of the solution. Substituting
these values along with i = /2 into (7.13.3) for i = 14, we obtain the transformation
d
z = f () = c . (7.13.14)
0 (2 1)1/2 (2 2 )1/2
7.13 The SchwarzChristoel transformation 613
4
b
1 4
1
2 3 x 1 2 3 4
3
a 0 a
2 1 0 1
Figure 7.13.3 Mapping a rectangle in the z plane to the upper half plane. Corresponding labels of
the four vertices are shown.
To compute the values of the two constants c and , we require that, as we move along the axis
from the origin up to 3 and then up to 4 , we nd ourselves at the corresponding vertices z3 and
z4 . The rst condition yields the real equation
1
d
c 2 1/2
= a, (7.13.15)
0 (1 ) (2 2 )1/2
which shows that c is real. The second condition yields the real equation
d
c 2 1/2
= b. (7.13.16)
1 ( 1) (2 2 )1/2
Combining these equations to eliminate c, we obtain a nonlinear equation,
d b 1 d
Q() 2 1)1/2 (2 2 )1/2
+ 2 )1/2 (2 2 )1/2
= 0. (7.13.17)
1 ( a 0 (1
The solution can be found using standard numerical methods, such as Newtons method discussed
in Section B.3, Appendix B.
Certain aspects of the computation require special attention. The integral in (7.13.15) can be
written as
1
d 1 1
= F , (7.13.18)
0 (1 )
2 1/2 (2 2 )1/2
where
1 /2
d du
F (k) =
(7.13.19)
0 (1 ) (1 k2 2 )1/2
2 1/2
0 1 k 2 sin2 u
is the complete elliptic integral of the rst kind discussed in Section 2.12, 0 < k < 1, and = sin u.
614 Introduction to Theoretical and Computational Fluid Dynamics
a
2
1
x x
1 0 1
a
2 1 2
Figure 7.13.4 Mapping of a semi-innite domain above a wall with (a) a forward or (b) backward
facing step to the upper half plane shown in (c).
To compute the singular integral on the left-hand side of (7.13.16), denoted by I, we subtract
out the singularities on either end of the integration domain by restating the integral into the form
1 d d
I=J + + , (7.13.20)
(2 1)1/2 1 ( 1)1/2
2
1 (2 2 )1/2
where
1 1 1 1
J + 2 d. (7.13.21)
1 ( 2 1)1/2 (2 2 )1/2 (2 1) 1/2 2
( 1)1/2 2
( ) 1/2
The J integral can be computed by standard numerical methods, as discussed in Section B.6,
Appendix B. The two improper integrals enclosed by the square brackets on the right-hand side of
(7.13.20) can be computed by elementary analytical methods.
Uniform ow along the axis in the plane corresponds to ow due to a horizontal point-
source dipole located at the midpoint of the upper side of the rectangle in the z plane, x = 0, y = 0,
as shown in Figure 7.13.3.
A third example is the semi-innite region above a wall with a forward facing step whose
boundary can be regarded as a generalized polygon with one vertex at innity, as shown in Figure
7.13.4(a). Mapping the vertex at innity in the z plane to the positive innity of the axis, specifying
7.13 The SchwarzChristoel transformation 615
and then
z = f () = c ( 1) ln( + 1) + i . (7.13.24)
2
The constant c is determined by requiring that f (1) = ia, yielding c = 2a/.
and then
z = f () = c ( 1) + ln( + 1) i . (7.13.27)
2
The constant c is determined by requiring that f (1) = ia, yielding c = 2a/.
For example, in the case of the generalized polygon with two collapsed sides and one vertex
at innity illustrated in Figure 7.13.5,
4 2
1 = , 2 = 2 3 = , 3 = , 4 = , 5 = . (7.13.28)
2 3 3 2
Note that the angle constraint (7.13.1) is satised.
616 Introduction to Theoretical and Computational Fluid Dynamics
/3
3
4 1
Figure 7.13.5 Illustration of a generalized polygon with two sides collapsed and one vertex at innity.
Channel-like domain
Consider an innite channel-like domain dened by N vertices, as shown in Figure 7.13.6. The left
aperture angle dened in this gure, L , is related to the vertex angles 1 and 2 by
1 + 2 = + L . (7.13.29)
For example, when 1 = 12 and 2 = 12 , L = 0. The right aperture angle, R , is related to the
corresponding vertex angles 1 and 2 by
1 + 2 = + R . (7.13.30)
N
1 + 2 + 1 + 2 + i = 2 (7.13.31)
i=1
requires that
N
L + R + i = 0, (7.13.32)
i=1
The SchwartzChristoel transformation mapping the channel to the upper half plane is
given by the dierential form
dz c #N
1
= 1+ / , (7.13.33)
d L
i=1
( /
i) i
where the left innity is mapped to the origin of the plane and the right innity is mapped to the
innity of the plane. For example, in the case of a channel conned between two parallel horizontal
walls, we set N = 2, L = 0, 1 = 0, and 2 = 0, yielding dz/d = 1/. Integrating, we obtain
z = c ln(/0 ), where c and 0 are two constants.
7.13 The SchwarzChristoel transformation 617
2
z plane
K+3 K+1
N
K+2
1 R
L
1
2 1
2
K
2
plane
K +1 N 1 2 K
plane
N K +1
h
1 2 K
Figure 7.13.6 A channel-like domain with two vertices at innity can be mapped to the upper half
plane and then to an innite strip in the plane.
The upper half plane can be subsequently mapped to a horizontal strip in the complex,
= R + i I plane, < R < , 0 < I < h, as shown in Figure 7.13.6. The mapping is
mediated by the logarithmic function and its inverse,
h
= ln , = a exp , (7.13.34)
a h
where a is a positive constant [129]. Making substitutions into (7.13.33), we nd that
dz
#N
( i ) i /
L
= exp exp 1 , (7.13.35)
d h i=1
h
618 Introduction to Theoretical and Computational Fluid Dynamics
# ( i ) i /
N
dz
= q exp (L R ) sinh , (7.13.37)
d h i=1 2h
where q is a new constant. We can separate the lower from the upper wall vertices to obtain
# ( i ) i / # ( i + ih) i /
K N
dz
= p exp (L R ) sinh cosh , (7.13.38)
d h i=1 2h 2h
i=K+1
where p is a new constant. In the case of a channel with symmetric extensions, L = R , the
exponential factor in front of the products on the right-hand side does not appear. In the case of a
channel with a at top wall, the second product on the right-hand side of (7.13.38) does not appear.
In the case of a channel whose upper and lower walls are repeated along the x axis with period
L, we obtain [130]
dz
# #
K
( i mL) i / #
N
( i + ih mL) i /
In practice, accurate results can be obtained even when the product is truncated at small values
of m. In the case of a channel with a at top wall, the second product on the right-hand side of
(7.13.39) does not appear.
Substituting the rst expression in (7.13.40) into the right-hand side of (7.13.3), we obtain
the required mapping function
# N
i / d
1+
z = f () = f [W( )] q( ) = z0 + 2ic i i . (7.13.41)
0 i=1 1 (1 )2
7.13 The SchwarzChristoel transformation 619
z plane
plane plane
y
i
i 1 2
3
4 x 1
1 2 3 4 N 1
3
1 N
4
N
2
Figure 7.13.7 Mapping a polygon in the z plane to a unit disk centered at the origin of the plane.
Rearranging and taking into account (7.13.1) to simplify the integrand, we obtain
N
#
1
z = q( ) = z0 + d d, (7.13.42)
0 i=1 ( i ) i
/
where d is a new constant. The image points, i = (i i)/(i + i) for i = 1, . . . , N , are distributed
around the unit circle in the plane. We can place 1 or N at innity to obtain, respectively, 1 = 1
or N = 1. Although the corresponding multiplier does not appear in (7.13.3), it must be included
inside the product in (7.13.42). It will be noted that the disk transformation (7.13.42) is identical
in form to the semi-innite plane transformation (7.13.3). The only dierence is that the images,
i , are distributed around the unit circle in the counterclockwise fashion instead of the real axis.
Regular polygons
In the case of a regular N -sided polygon centered at the origin of the z plane, z0 = 0, we set 0 = 0,
i = 2/N , and choose
i 1
i = exp 2i (7.13.43)
N
for i = 1, . . . , N , so that the vertices are distributed uniformly around the unit circle with 1 = 1.
The transformation (7.13.42) with z0 = 0 and 0 = 0 becomes
N
# i 1 2/N
z = q( ) = d exp 2i d. (7.13.44)
0 i=1
N
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
y/a
0 0
I
0.2 0.2
0.4 0.4
0.6 0.6
0.8 0.8
1
1 0.5 0 0.5 1 1 0.5 0 0.5 1
R x/a
Figure 7.13.8 Corresponding grids generated by mapping a disk of unit radius in the plane to a
heptagon of radius a in the z plane.
yielding
d
z = q( ) = A , (7.13.46)
0 (1 N )2/N
where A is a new constant computed by requiring q(1) = a, and a is the radius of the polygon.
Performing the integration with respect to , we obtain an inconvenient expression in terms of the
Gauss hypergeometric function. Although an approximate Taylor series expansion is available [225],
in practice, it is expedient to compute the mapping function by numerical integration. Corresponding
grids generated by mapping the disk to a heptagon are shown in Figure 7.13.8.
Fractal shapes
Polygonal shapes can be generated by distributing a specied number of images i around the unit
disk and specifying the corresponding angles i in the physical z plane, subject to the restriction
imposed by (7.13.1). Fractal contours can be generated in the limit as the number of vertices tends
to innity [344].
z plane plane
2
2 2 N
N 3
1 3
1
x
i
1
3
i 2
N
Figure 7.13.9 Mapping the exterior of a polygon to the interior or exterior of a unit disk in the
complex plane.
sequentially in the clockwise direction, and compute the vertex angles, i , where
N
i = 2. (7.13.47)
i=1
For the conguration shown in Figure 7.13.9, 1 > 0 and 2 < 0. The vertices of the exterior
auxiliary polygon are mapped to the origin of the plane, so)that the corresponding images are
zero, j = 0, and the sum of the corresponding angles satises j j = 2.
Applying (7.13.2) for the union of the vertices of the outer and inner polygons with the
aforementioned conventions, we obtain the transformation
# N d
1
z = q( ) = z0 + d i / 2
, (7.13.48)
0 i=1 ( i )
where z0 and 0 is an arbitrary pair of corresponding points and d is a complex constant. The image
points, i , are distributed in the counterclockwise direction around the unit disk centered at the
origin of the complex plane, as shown in Figure 7.13.10 ([276], p. 193).
Regular polygons
In the case of a regular N -sided polygon centered at the origin of the z plane, we set i = 2/N for
i = 1, . . . , N , and select the images shown in (7.13.43) so that the vertices are distributed uniformly
around the unit circle with 1 = 1. The transformation (7.13.48) reduces to
d
z = q( ) = z0 + d (N 1)2/N 2 . (7.13.49)
0
The power (N 1)2/N is computed best after the polynomial N 1 has been factorized into a
product of monomials involving the N th roots of unity.
622 Introduction to Theoretical and Computational Fluid Dynamics
Figure 7.13.10 Grids generated by mapping the exterior of the unit circle to the exterior of an equi-
lateral triangle, square, or regular pentagon.
([65], p. 153). The computation of the unknown parameters is analogous to that for the correspond-
ing mapping of the interior of a polygon previously discussed.
N
i = 0, (7.13.51)
i=1
where < i < . Positive angles correspond to counterclockwise rotation and negative angles
correspond to clockwise rotation.
7.13 The SchwarzChristoel transformation 623
2
y
N
1
1
i
x 3 i
o o o o
1 2 3 N
Figure 7.13.11 Mapping a semi-innite domain bounded by a periodic polygonal line to the upper
half-plane.
Modifying the SchwartzChristoel transformation (7.13.2) to account for all vertex images,
we obtain the inverse mapping function
dz #N #
1
i / #N
# i /
=c =c ( i ) ( i )2 k 2 2 ) . (7.13.52)
d i=1
( + k i ) i=1
k= k=1
sin(w) = w 1 (7.13.54)
k2
k=1
where d = c/. Requiring that the rst point, z1 , be mapped to the origin of the plane, 1 = 0,
and integrating, we nd that
#N
i i /
z = z1 + d sin d. (7.13.56)
0 i=1
624 Introduction to Theoretical and Computational Fluid Dynamics
dz N
i
d exp i i . (7.13.57)
d i=1
Because of (7.13.51), the argument of the exponent is insensitive to the denition of the origin of
the axis. Integrating along a periodic, we nd that
L N
i
d= exp i i , (7.13.58)
i=1
which can be used to evaluate the constant d once the positions i are available. Substituting this
expression into (7.13.57), we nd that dz/d L/ far from the wall. To compute the N 1
unknown images, 2 , . . . , N , we integrate along the axis from the rst up to the jth vertex, we
obtain
j #
N
i i /
zj = z1 + d sin d (7.13.59)
0 i=1
for j = 2, . . . , N . Requiring that this expression reproduces the vertices of the polygonal boundary
in the physical plane provides us with a system of nonlinear equations.
Triangular asperities
In the case of a wall with an innite sequence of triangular asperities, N = 2 and 2 = 1 , as shown
in Figure 7.13.12(a). Without loss of generality, we may assume that 1 < 0 so that the integrand
tends to zero at the origin. The transformation (7.13.56) with 1 = 0 becomes
L i1 (12 /) sin / 1 /
z = z1 + e d. (7.13.60)
0 sin (2 )/
Since 1 < 0, the exponent inside the integral is positive and a singularity does not appear at the
origin, = 0. The integral can be computed accurately using the numerical methods discussed
in Section B.6, Appendix B. Applying equation (7.13.59) for j = 2 provides us with an algebraic
equation for 2 , which can be recast into the form
L i (11 2 /) 2 sin / 1 /
F(2 ) z1 z2 + e d = 0. (7.13.61)
0 sin (2 )/
To compute the constant 2 , we solve the real equation FF = 0 using, for example, Newtons
method discussed in Section B.3, Appendix B. To evaluate the integral on the right-hand side,
denoted by I, we remove the singularity of the integrand at = 2 by writing
2 1 / sin / 1 /
sin / 2
I= d
0 sin (2 )/ (2 )/
sin / 1 / 2 1 1 /
2
+ d. (7.13.62)
/ 0 2
7.13 The SchwarzChristoel transformation 625
(a)
L
1
1 3
y
x
2
2
(b) (c)
1.2 1.2
1 1
0.8 0.8
0.6 0.6
y/L
y/L
0.4 0.4
0.2 0.2
0 0
Figure 7.13.12 (a) Mapping a semi-innite domain bounded by a periodic sequence of asperities to
the upper half-plane. (b, c) Grids generated by mapping a uniform Cartesian grid in the plane
to the z plane.
The rst integral on the right-hand side can be computed accurately using the numerical methods
discussed in Section B.6, Appendix B. The second integral on the right-hand side is equal to 2m /m,
where m = 1+1 / > 0. A grid of points generated by mapping a uniform Cartesian grid from the
to the z plane is shown in Figure 7.13.12(c). In the case of symmetric asperities where x2 = x1 + 12 L,
while y1 and y2 are arbitrary, we set 2 = 12 and obtain
L i1 /2 1 /
z = z1 + e tan d. (7.13.63)
0
Polygonal strip
Transformations mapping a periodic channel with polygonal walls to an innite strip are avail-
able [129, 130, 132, 133]. The mapping functions are useful in developing orthogonal grids for use
with nite-dierence methods, or studying scalar conduction in domains with involved and fractal
geometry [53].
626 Introduction to Theoretical and Computational Fluid Dynamics
dz #N
i
=c exp log( i ) . (7.13.64)
d i=1
An equivalent form is
dz #N 1 d
where is the slope angle dened in Figure 7.13.13. In the case of an N -sided polygonal boundary,
is a discontinuous function,
d
N
= i ( i ), (7.13.66)
d i=1
Smooth contours
In fact, the transformation (7.13.65) applies for polygonal as well as smooth shapes with nonsingular
slope functions (). A slight rearrangement yields
dz #N 1 d d
involving the known function d/d and the unknown function d/d, where is an arc-length like
parameter describing the physical contour [129, 133]. To compute the mapping function, we may
trace a smooth contour enclosing a domain of interest in the z plane with a number of marker points,
introduce the corresponding images, i , construct the function d/d by interpolation, and work as
in the case of a polygon to generate a system of nonlinear equations.
To map a smooth region to the unit disk in the plane, we use the transformation
dz #N 1 d d
corresponding to the polygon transformation (7.13.42), where d is a complex constant and the
integral is performed in the counterclockwise direction around the unit circle. Writing = exp(i),
we obtain
dz #N 1 2 d d
A weak logarithmic singularity occurs when the point is located at the unit circle. Working in a
similar fashion, we derive transformations that map the exterior of a smooth contour to the upper
half-plane or exterior of the unit disk.
Problem
Computer Problems
The plate aspect ratio, a/L, is determined implicitly by the dimensionless parameter 2 / ranging
in the interval (0, 1). Generate a grid of points similar to that shown in Figure 7.13.12(b) for several
values of 2 /.
Boundary-layer analysis 8
There is an important class of ows where the gradient of the vorticity or the vorticity itself vanishes
virtually everywhere, except inside thin layers or narrow columns of uid wrapping around or trailing
behind solid boundaries, free surfaces, and uid interfaces, loosely identied as boundary layers.
Inspecting the equation of motion, we nd that viscous forces are small and can be neglected outside
the boundary layers, but make important contributions and must be retained inside the boundary
layers.
Following Prandtls original idea, a large body of theoretical and laboratory work has con-
rmed the physical relevance and eciency of boundary-layer theory, opening a new era in the study
of viscous ow (e.g., [395]). Prandtl boundary layers have been identied over surfaces translating
at high speed or stretching in an otherwise quiescent uid in a broad range of applications. The
importance of boundary-layer ows in practical aerodynamics has spawned a large body of literature
discussed in comprehensive monographs and reviews [41, 70, 71, 260, 352, 363, 376].
629
630 Introduction to Theoretical and Computational Fluid Dynamics
The physical processes determining the behavior of boundary layers developing over solid
surfaces in unsteady ow become evident by considering the structure of a boundary layer developing
along a rigid body that is held stationary in an impulsively started incident ow. At the initial
instant, the velocity eld is irrotational everywhere, except inside a thin vortex layer lining the
body. Viewed at a distance, the vortex layer resembles a vortex sheet whose strength is equal to the
negative of the tangential velocity of the potential ow. As soon as the motion has been initiated, the
vortex sheet diuses into the ow. Locally around the boundary, the ow resembles that developing
over an innite at plate subject to an impulsively started uniform ow described by a similarity
solution, as discussed in Section 5.4.3. When the thickness of the vortex layer becomes signicant,
convection of vorticity in the tangential and normal directions becomes important and the leading-
order similarity solution is modied with the addition of a second-order term that is proportional
to time since startup [391]. Examination of the second-order term shows that the boundary shear
stress vanishes and back ow occurs at the point where the irrotational ow decelerates along the
boundary. The time of separation predicted by the second-order solution is in remarkably good
agreement with experimental observation.
More generally, changes in the conditions of an incident ow aect the distribution of the
tangential velocity and pressure gradient that drive the ow inside the boundary layer, and therefore
cause the vorticity to diuse across the boundaries and enter or exit the ow. The precise description
of the evolution of an unsteady boundary-layer ow presents us with signicant analytical and
computational challenges that have been tackled only for a limited number of ows [352].
A dierent class of boundary layers occurs at free surfaces and interfaces between two streams
of the same or dierent uids. In the case of interfacial ow, an irrotational incident ow is not
capable of satisfying both the kinematic boundary condition of continuity of velocity and the dy-
namic boundary condition specifying the discontinuity in the interfacial traction, and the role of the
boundary layer is to make an appropriate adjustment. An example is the boundary layer established
around the surface of a gas bubble rising through a liquid. Another example is the boundary layer
established around a liquid jet discharging into an ambient uid.
In this chapter, we introduce the fundamental concepts involved in the formulation of boundary-
layer theory and discuss analytical, approximate, and numerical methods for solving the governing
equations in various types of steady and unsteady ow. We will consider Prandtl boundary layers
developing over stationary surfaces in nonaccelerating and accelerating ow, and boundary layers
developing over moving or stretching surfaces in an otherwise quiescent uid. By way of applica-
tion, the theory will be used to predict the expansion of a two-dimensional or axisymmetric oil slick
oating on a pool of an otherwise quiescent uid.
y U
V
Figure 8.1.1 Illustration of the Prandtl boundary layer developing around a two-dimensional body.
Shown are the typical velocity, U , the typical length, L, and the boundary-layer thickness, .
Continuity equation
We begin the boundary-layer analysis by introducing Cartesian coordinates where the x axis is
tangential and the y axis is perpendicular to the surface of the body at a point, as shown in Figure
8.1.1. Next, we apply the continuity equation at a point in the vicinity of the origin,
u v
+ = 0, (8.1.1)
x y
where u = ux and v = uy are the components of the velocity along the x and y axes.
Let L be the typical dimension of the body, U be the typical magnitude of the velocity of
the outer irrotational ow, be the typical thickness of the boundary layer, dened as the region
around the body across which the velocity undergoes a rapid variation and the magnitude of the
viscous force is signicant, and V be the typical magnitude of the velocity component normal to
the body at the edge of the boundary layer. We expect that the magnitude of the derivative u/x
inside the boundary layer will be comparable to the ratio U/L, and the magnitude of the derivative
v/y will be comparable to the ratio V /. The continuity equation (8.1.1) requires that
U V
or V U , (8.1.2)
L L
stating that V scales with the boundary-layer thickness, . In fact, V is roughly proportional to the
boundary-layer thickness, .
632 Introduction to Theoretical and Computational Fluid Dynamics
Next, we turn to examining the two components of the equation of motion inside the boundary
layer in the vicinity of the origin written for the hydrodynamic pressure that excludes the eect of
the body force. For simplicity, the hydrodynamic pressure is denoted without a tilde, as p.
The scalings shown underneath equation (8.1.3), combined with the fundamental assumption
that L, have two important consequences. First, the penultimate viscous term on the right-hand
side is small compared to the last term and can be neglected, yielding the boundary-layer equation
u u u p 2u
+u +v = + 2. (8.1.5)
t x y x y
Second, the magnitude of the last viscous term must be comparable to the magnitude of the inertial
terms on the left-hand side, requiring that
U2 U
2. (8.1.6)
L
Rearranging, we nd that
L
1/2 L
= , (8.1.7)
U Re
where Re = U L/ is the Reynolds number.
The magnitudes of all nonlinear convective and viscous terms are of order . Unless the
magnitude of the time derivative on the left-hand side is of order unity, the dynamic pressure
gradient across the boundary layer must also be of order , p/y . This deduction allows us to
write to leading-order approximation
p
0. (8.1.10)
y
We conclude that nonhydrostatic pressure variations across the boundary layer are negligible, and
the dynamic pressure inside the boundary layer is primarily a function of position along the wall.
Boundary-layer equations
To compute the streamwise pressure gradient, we evaluate the x component of the Euler equation
at the edge of the boundary layer, obtaining
1 p U U
= +U , (8.1.11)
x t x
where U is the tangential component of the velocity of the outer ow. The boundary-layer equation
(8.1.5) then becomes
u u u U U 2u
+u +v = +U + 2, (8.1.12)
t x y t x y
where U is regarded as a known forcing function.
The continuity equation (8.1.1) and the boundary-layer equation (8.1.12) provide us with a
system of two second-order nonlinear partial-dierential equations for the x and y velocity compo-
nents, u and v, to be solved subject to the no-slip and no-penetration boundary conditions requiring
that u and v are zero along the boundary, and a far-eld condition requiring that, as y/ tends to
innity, u tends to the tangential component of the outer velocity, U . Because the boundary-layer
equations do not involve a second partial derivative of v with respect to y, a far-eld condition for v
is not required. The pressure plays the role of a forcing function computed by solving the equations
governing the outer irrotational ow.
which shows that the sign of the curvature of the velocity prole at the boundary is opposite to that
of the streamwise acceleration of the outer ow, dU/dx. Thus, the ow inside the boundary layer
in a decelerating outer ow, corresponding to dU/dx < 0, reverses direction, causing convection of
vorticity away from the boundary and the consequent formation of wakes and vortices inside the
bulk of the ow.
When dU/dx > 0, the pressure gradient is negative, dp/dx < 0, and the boundary layer is
subjected to a favorable pressure gradient. In the opposite case where dU/dx < 0, the pressure
gradient is positive, dp/dx > 0, and the boundary layer is subjected to an adverse pressure gradient.
Equation (8.1.13) shows that an adverse pressure gradient promotes ow separation.
p U2
= , (8.1.14)
R
where R is the radius of curvature of the boundary. We conclude that the dynamic pressure drop
across the boundary layer is of order , provided that R is not too small, that is, provided that the
boundary is not too sharply curved. For simplicity, in the remainder of this chapter we denote l and
, respectively, by x and y.
The parabolic nature of the boundary-layer equation (8.1.12) with respect to x implies that a
perturbation introduced at some point along the boundary layer modies the structure of the ow
downstream but leaves the upstream ow unchanged. The absence of the second partial derivative
with respect to x due to the boundary-layer approximation precludes a mechanism for upstream
signal propagation.
8.1 Boundary-layer theory 635
The absence of the Reynolds number in the governing equations (8.1.17) suggests that the
Reynolds number is signicant only insofar as to determine the physical thickness of the boundary
layer and magnitude of the y velocity component according to (8.1.12), and does not have an inuence
on the structure of the ow inside the boundary layer. However, the assumption that the Reynolds
number is suciently high is necessary for this asymptotic behavior to prevail.
which, according to (8.1.7), scales as U 3 /L and therefore remains nite as the viscosity tends to
vanish or the Reynolds number tends to innity. The rate of viscous dissipation inside the boundary
layer per unit surface area of the boundary is
2
u U2
dy , (8.1.19)
0 y
which, according to (8.1.7), scales as U 3 /Re1/2 . Since the rate of viscous dissipation per unit
volume of uid outside the boundary layer scales with U 3 /(LRe), the rate of viscous dissipation
inside the boundary layer makes a dominant contribution.
Flow separation
Boundary-layer analysis for laminar ow is based on two key assumptions: the Reynolds number is
suciently high, but not so high that the ow becomes turbulent, and the vorticity remains conned
inside boundary layers wrapping around the boundaries. The physical relevance of the second
assumption depends on the structure of the incident ow and boundary geometry. Streamlined
bodies allow laminar boundary layers to develop over a large surface area, whereas blu bodies cause
vorticity to concentrate inside compact regions forming steady or unsteady wakes. For example, the
alternating ejection of vortices of opposite sign into a wake is responsible for the von Karman vortex
street illustrated in Figure 8.1.2(a) (e.g., [187]). These limitations should be born in mind in carrying
out a boundary-layer analysis.
where F is the drag force per unit length exerted on the cylinder. In Figure 8.1.2(b), the drag
coecient is plotted against the Reynolds number, Re = U D/, on a log-log scale. Using (6.14.22),
we nd that, in the limit of vanishing Reynolds number, the drag force is given by the modied
Stokes law
4U
F . (8.1.21)
7.4
ln
Re
Correspondingly, the drag coecient is given by
8
cD . (8.1.22)
7.4
Re ln
Re
The predictions of this formula are represented by the dashed line in Figure 8.1.2(b). The apparent
change in the functional form of the drag coecient at a critical Reynolds number on the order of
8.1 Boundary-layer theory 637
(a) (b)
2.5
2
Vortex street 1.5
log10(cD)
1
0.5
1
2 0 2 4 6
log10(Re)
Figure 8.1.2 (a) Changes in the structure of streaming ow past a circular cylinder with increasing
Reynolds number showing boundary-layer separation and the development of a wake consisting of
a vortex street. (b) Laboratory data for the drag coecient on a circular cylinder that is held
stationary in streaming ow plotted against the Reynolds number dened with respect to the
cylinder diameter. The dashed line represents an analytical prediction for low-Reynolds-number
ow.
103 , shown in Figure 8.1.2(b), is due to the detachment of the boundary layer from the surface of
the cylinder at a certain point at the rear surface of the cylinder, in a process that is described as
ow separation, as shown in Figure 8.1.2(a). When the Reynolds number becomes on the order of
105 , the ow becomes fully turbulent and the boundary layer reattaches, causing a sudden decline
in the drag coecient.
stress, and a boundary layer must be established. An example is the boundary layer developing over
the surface of a spherical bubble rising through a liquid at high Reynolds numbers, as discussed in
Section 7.5.4 [259]. An important dierence between this boundary layer and the Prandtl boundary
layer developing over a solid surface is that, in the case of a free surface, the velocity does not have
to undergo a jump whose magnitude is comparable to the velocity of the incident ow.
Free surfaces
The tangential component of the vorticity at an arbitrary surface was given in (3.9.18) as
(I n n) = (n ) n = [ (u) n n u ] n, (8.1.23)
where f is the traction and n is the normal unit vector. Combining these expressions, we obtain
1 1
(I n n) = f n 2 (n u) n = f n + 2 [(u) n] n. (8.1.25)
Requiring that the tangential component of the traction vanishes at the free surface, we nd that the
tangential component of the vorticity is given by equation (3.9.19), repeated here for convenience,
f s (I n n) = 2 (n u) n = 2 [(u) n] n, (8.1.26)
It is useful to decompose the velocity eld into the outer component, u , and a complementary
velocity eld associated with the boundary layer, ubl , so that u = u + ubl . Assume that the outer
velocity eld u is irrotational. Applying (8.1.25) at the edge of the boundary layer where the
vorticity is zero, we nd that
1 1
(I n n) = f n 2 (n u ) n = f n + 2 [(u ) n] n = 0. (8.1.27)
f s (I n n) 2 (n u ) n 2 [(u ) n] n. (8.1.28)
For the vorticity inside the boundary layer to be of order U/L, the jump in the complementary
velocity ubl across the boundary layer must be of the same order of magnitude as the boundary
layer thickness, , which scales with Re1/2 , and is therefore small as long as the Reynolds number
is suciently high. The smallness of ubl allows us to linearize the equation of motion and vorticity
8.1 Boundary-layer theory 639
y U
U
U
Figure 8.1.3 Illustration of a boundary layer developing in the wake of a two-dimensional body.
transport equation with respect to u about u . Since the magnitude of ubl is much smaller than
that of u , the tendency for back ow in a decelerating ow along a free surface is much weaker
than along a solid wall, and the rate of viscous dissipation inside the boundary layer is comparable
to that inside the incident ow.
Fluid interfaces
To describe the ow around a uid interface, we introduce a boundary layer on either side of the
interface, demand that the velocity is continuous across the interface, and require that the two
boundary layers develop so as to satisfy the tangential force balance. The mathematical formulation
is discussed by Harper & Moore [169] with reference to the boundary layer developing around the
surface of a viscous drop rising through an otherwise quiescent ambient liquid at high Reynolds
numbers.
where c is a dimensional constant determined by the particular form of the velocity distribution at
the beginning of the wake.
Writing a momentum integral balance over a rectangular control area with two parallel sides
perpendicular to the x axis containing the object, we obtain an expression for the drag force per
unit width exerted on the object generating the wake,
2 2
F = (U u ) dy U (U u) dy. (8.1.32)
Substituting the velocity prole, performing the integration, and rearranging, we nd that
1 F
c= (8.1.33)
2 U 3/2 1/2
(Problem 8.1.4). Other free boundary layers are discussed in Problems 8.2.3, 8.5.5, and 8.5.6.
A variety of other specialized methods have been developed for computing the ow inside
boundary layers developing over arbitrarily shaped bodies that either move or are held stationary in
an arbitrary irrotational ow. A review of early work can be found in a comprehensive monograph
edited by Rosenhead [352], a concise and illuminating discussion is given in a textbook by White
([427], Chapter 4), and more recent developments are discussed by Cebeci & Bradshaw [70, 71].
Problems
+u = +U +u u , (8.1.34)
t x t x
with boundary conditions u = 0 at = 0 and far-eld condition that u tends to U as tends to
innity.
(b) Show that, in the case of steady ow, the equation of motion (8.1.12) can be restated as a
nonlinear parabolic equation,
2
= (U 2 )1/2 , (8.1.35)
x 2
u u 2u
u +v = 2, (8.2.1)
x y y
to be solved for the x and y velocity components, u = ux and v = uy , subject to the continuity
equation (8.1.1).
(a) (b)
6
5
y U
Streamline
4
y = g(x)
3
2
y =(x)
Boundary layer 1
x
0
0 0.5 1 1.5
Figure 8.2.1 (a) Illustration of a boundary layer developing along a semi-innite at plate that is held
parallel to a uniform incident stream. (b) Graphs of the Blasius self-similar streamwise velocity
prole u/U = f (solid line), its integral f (dashed line), and derivative f (dotted line).
in Section 8.2, L, is the streamwise distance, x. Equation (8.1.7) provides us with an expression for
the boundary-layer thickness in terms of the local Reynolds number, Rex U x/,
x
1/2 x
(x) = . (8.2.2)
U Re1/2
x
Recall that this scaling arose by balancing the magnitudes of the inertial and viscous forces inside
the boundary layer.
where the function f () is the indenite integral or antiderivative of the function F (), satisfying
df /d = F (). The principal advantage of using the stream function is that the continuity equation
is satised automatically and does not need to be considered in further analysis.
Dierentiating (8.2.5) with respect to y or x, and invoking the denition of the stream function,
we obtain expressions for the x velocity component,
d U
1/2
u(x, y) = = U (U x)1/2 f = U (U x)1/2 f , (8.2.6)
y dy x
and y velocity component,
[ x f ()] 1 U
1/2
v(x, y) = = (U )1/2 = f (U x)1/2 f , (8.2.7)
x x 2 x x
where a prime denotes a derivative with respect to . Simplifying, we obtain
1 U
1/2
u(x, y) = U f , v(x, y) = (f f ). (8.2.8)
2 x
Substituting these expressions into the boundary-layer equation (8.2.1), we derive the Blasius equa-
tion
1
f + f f = 0. (8.2.9)
2
An alternative form is
d ln f 1
= f. (8.2.10)
d 2
Enforcing the no-slip and no-penetration boundary conditions, we obtain
Since the ow in the boundary layer reduces to the outer uniform ow far from the plate,
f () = 1. (8.2.12)
Equations (8.2.11) and (8.2.12) provide us with boundary and far-eld conditions to be used in
solving the Blasius equation (8.2.9). Since boundary conditions are specied at both ends of the
computational domain, [0, ), we must solve a boundary-value problem involving a third-order,
nonlinear, ordinary dierential equation.
Rescaling
If f () = g() is a solution of (8.2.9), then f () = g() is also a solution of (8.2.9) for any value
of the constant . Requiring that the function g() satises the boundary conditions g(0) = 0,
g (0) = 0, and g (0) = , where is an arbitrary value, and setting = 1/[g ()]1/2 , ensures that
the function g() satises the boundary conditions (8.2.11) and (8.2.12), and therefore represents
644 Introduction to Theoretical and Computational Fluid Dynamics
the desired solution, f (). However, in practice, it is more expedient to compute the solution by a
shooting method, as discussed later in this section.
where a prime denotes a derivative with respect to the integration variable, . Integrating (8.2.10)
with respect to and using (8.2.13), we derive Weyls integral equation for f ,
f 1
ln = ( )2 f () d, (8.2.14)
f0 4 0
where f0 = f (0) ([352], p. 233).
which shows that the curvature of the streamwise velocity prole vanishes at the wall, in agreement
with equation (8.1.13).
Numerical solution
To solve the Blasius equation (8.2.9), we rename x1 = f , denote the rst and second derivatives of
the function f as x2 df /d and x3 dx2 /d = d2 f /d 2 , and decompose the third-order Blasius
equation into a system of three rst-order nonlinear equations,
dx1 dx2 dx3 1
= x2 , = x3 , = x1 x 3 . (8.2.16)
d d d 2
The accompanying boundary and far-eld conditions are
originating from (8.2.11) and (8.2.12). The solution can be computed using a shooting method
according to the following steps:
Numerical experimentation shows that integrating up to = 10 in the second step yields satisfactory
accuracy. The improvement in the third step can be made using a method for solving nonlinear
algebraic equations, as discussed in Section B.3, Appendix B.
Variational equations
In practice, it is expedient to carry out the iterations using Newtons method based on derived
variational equations (e.g., [317]). First, we dene the derivatives
dx1 dx2 dx3
x4 , x5 , x6 , (8.2.18)
dx3 (0) dx3 (0) dx3 (0)
expressing the sensitivity of the solution to the initial guess. Dierentiating equations (8.2.16) with
respect to x3 (0), we obtain
dx4 dx5 dx6 1
= x5 , = x6 , = (x4 x3 + x1 x6 ), (8.2.19)
d d d 2
subject to the initial conditions
x4 ( = 0) = 0, x5 ( = 0) = 0, x6 ( = 0) = 1. (8.2.20)
The shooting method involves the following steps: guess a value for x3 (0); integrate equations
(8.2.17) and (8.2.19) with the aforementioned initial conditions; replace the current initial value,
x3 (0), with an updated value,
x2 () 1
x3 (0) x3 (0) . (8.2.21)
x5 ()
The iterations converge quadratically for a reasonable initial guess.
u
f (0) 0.332
w
xy (x) = = U2 = U 2. (8.2.23)
y y=0 Rex Rex
w
We observe that xy takes an innite value at the leading edge, and decreases as the inverse square
root of the streamwise distance or local Reynolds number, Rex , along the plate. However, the
physical signicance of the singular behavior at the origin is undermined by the breakdown of the
assumptions that led us to the boundary-layer equations at the leading edge.
Drag force
Even though the shear stress is innite at the leading edge, the inverse-square-root singularity is
integrable and the drag force exerted on any nite section of the plate extending from the leading
edge up to an arbitrary point is nite. Using the similarity solution, we nd that the drag force
exerted on the plate over a length extending from the leading edge up to a certain distance x, is
given by
x x
3/2 1/2 d
D(x) w
xy () d = 0.332 U . (8.2.24)
0 0
0.664
D(x) = U 2 x. (8.2.25)
Rex
D(x) 1.328
cD (x) 1 2
= . (8.2.26)
2 U x Rex
The predictions of the last two equations for the drag force and drag coecient agree well with
laboratory measurements up to Rex 1.2 105 . At that point, the ow inside the boundary layer
develops a wavy pattern and ultimately becomes turbulent. Above the critical value of Rex , the
function cD (Rex ) jumps to a dierent branch with signicantly higher values. The transition from
steady laminar ow, presently considered, to unsteady turbulent ow will be discussed in Section
9.8.2 in the context of hydrodynamic stability.
Vorticity transport
Neglecting the velocity component along the y axis, we nd that the z component of the vorticity
inside the boundary layer is given by
u f () U 2 U
z (x, y) = = f () . (8.2.27)
y Rex (x)
We observe that the magnitude of the vorticity at a particular location, , decreases like the inverse
of the local boundary-layer thickness due to the broadening of the velocity prole.
8.2 Boundary layers in nonaccelerating ow 647
The streamwise rate of convection of vorticity across a plane that is perpendicular to the plate
is given by
u 1
u(x, y) z (x, y) dy u(x, y) dy = U 2 , (8.2.28)
0 0 y 2
independent of the downstream position, x. Thus, the ux of vorticity across the plate is zero and
viscous diusion of vorticity does not occur at the wall, in agreement with our earlier observation that
the gradient of the vorticity vanishes at the wall. Consequently, all convected vorticity is generated
at the leading edge where the boundary-layer approximation ceases to be valid. Viscous stresses at
the leading edge somehow generate the proper amount of vorticity necessary for the establishment
of the Blasius self-similar ow.
Displacement thickness
Because of the widening of the streamwise velocity prole, the streamlines inside the Blasius bound-
ary layer are deected upward, away from the plate, as shown in Figure 8.2.1(a). Consider a
streamline outside the boundary layer, described by the equation y = g(x), and write a mass bal-
ance over a control area that is enclosed by (a) the streamline, (b) a vertical plane at x = 0, (c) a
vertical plane located at a certain distance x, and (d) the at plate. Since the streamwise velocity
prole is at at the leading edge located at x = 0, we obtain
g(0) g(x)
U dy = u(x, y) dy. (8.2.29)
0 0
Taking the limit as the streamline under consideration moves farther from the plate, we nd that
where
u
(x) 1 dy (8.2.32)
0 U
is the displacement thickness. Using the numerical solution of the Blasius equation to evaluate the
integral on the right-hand side, we derive the exact expression
x
1/2 df
x
1/2
(x) = 1 d = 1.721 , (8.2.33)
U 0 d U
which shows that the displacement thickness, like the 99% boundary layer thickness, increases like
the square root of the streamwise position, x.
Physically, the displacement thickness represents the vertical displacement of the streamlines
far from from the plate with respect to their elevation at the leading edge. Laboratory experiments
648 Introduction to Theoretical and Computational Fluid Dynamics
have shown that the laminar boundary layer undergoes a transition from the laminar to the turbulent
state when the Reynolds number based on the displacement thickness reaches the approximate value
600, that is, when 600 /U . At that point, turbulent shear stresses become signicant and the
present analysis based on the assumption of laminar ow ceases to be valid.
Iterative improvement
The displacement thickness describes the surface of a ctitious impenetrable but slippery body
that is held stationary in the incident irrotational ow. An improved boundary-layer theory can
be developed by replacing the tangential velocity of the outer ow along the plate, U , with the
corresponding tangential velocity of the irrotational ow past the ctitious body. The irrotational
ow past the ctitious body must be computed after the displacement thickness has been established,
as discussed in this section. This iterative improvement provides us with a venue for describing the
ow in the framework of matched asymptotics (e.g., [277]).
Momentum thickness
It is illuminating to perform a momentum integral balance over the control area employed to dene
the displacement thickness [415]. Since the upper boundary of the control volume is a streamline, the
associated tangential ow does not contribute to the rate of momentum input. Assuming that the
normal stresses on the vertical sides are equal in magnitude and opposite in sign, which is justied by
the assumption that the pressure drop across the boundary layer is negligibly small, and neglecting
the traction along the top streamline, we obtain
g(0) g(x)
U ( U ) dy u ( u) dy D(x) = 0, (8.2.34)
0 0
where D(x) is the drag force exerted on the plate dened in (8.2.24). Now we make the two upper
limits of integration equal by recasting (8.2.34) into the form
g(x)
U 2 g(x) g(0) (U 2 u2 ) dy D(x) = 0. (8.2.35)
0
Finally, we take the limit as the streamline dening the top of the control area is moved far from
the plate, and use (8.2.30) to obtain the relation
where f () = df /d. The momentum thickness, like the displacement thickness and the 99%
boundary-layer thickness, increases as the square root of the streamwise position, x. Dierentiating
(8.2.36) with respect to x and using (8.2.24), we derive the relation
w
d xy
= , (8.2.39)
dx U 2
which provides us with a method of extracting the wall shear stress from the momentum thickness,
and vice versa, based on available information.
Shape factor
The ratio of the displacement to the momentum thicknesses is called the shape factor,
H . (8.2.40)
Substituting the right-hand sides of expressions (8.2.33) and (8.2.38) into (8.2.40), we nd that, for
the Blasius boundary layer over a at plate, H = 2.591. Inspecting the denitions of and
in equations (8.2.32) and (8.2.37), we nd that the shape factor is greater than unity as long as
the streamwise velocity u is less than U within a substantial portion of the boundary layer. The
satisfaction of this restriction is consistent with physical intuition. The smaller the value of H, the
more blunt the velocity prole across the boundary layer.
Relation between the wall shear stress and the momentum thickness
The momentum thickness is related to the wall shear stress, and vice versa, by the integral momentum
balance expressed by equation (8.2.36). Dierentiating (8.2.24) with respect to x, we nd that
dD(x) w
= xy (x). (8.2.41)
dx
Expressing the drag force in terms of the momentum thickness using (8.2.36), we obtain
d(x)
w
xy (x) = U 2 . (8.2.42)
dx
If the shear stress is known, the momentum thickness can be computed by integrating with respect
to x. If the momentum thickness is known, the shear stress can be computed by dierentiating with
respect to x.
Given the velocity prole across a boundary layer, the wall shear stress can be computed in
two ways: directly by dierentiation, and indirectly by evaluating the momentum thickness and then
dierentiating it with respect to streamwise position x to obtain the wall shear stress according to
equation (8.2.42). The indirect method is less sensitive to the structure of the velocity prole near
the wall. For the velocity prole that arises by solving the Blasius equation, the two aforementioned
methods are equivalent. To see this, we substitute (8.2.38) into (8.2.42) and obtain (8.2.23). We
conclude that the approximations that led us to the boundary-layer equations are identical to those
that led us to the simplied integral momentum balance (8.2.32).
650 Introduction to Theoretical and Computational Fluid Dynamics
Nothing can stop us from introducing a self-similar velocity prole with some reasonable form that is
either stipulated by physical intuition or suggested by laboratory observation. Our goal is to adjust
an unspecied function involved in this form so that the two methods of computing the wall shear
stress discussed previously in this section are equivalent. A reasonable velocity prole is
& y
u df () sin 2(x) for 0 y (x),
= = (8.2.43)
U d 1 for y (x),
where y/(x) and (x) is an unspecied function playing the role of a boundary-layer thickness,
similar to the 99 thickness introduced in (8.2.22). Note that the velocity distribution (8.2.43) con-
forms with the required boundary and far-eld conditions, f (0) = 0, f (0) = 0, and f () = 1, but
does not satisfy the Blasius equation; a prime denotes a derivative with respect to . Dierentiating
the prole (8.2.43) with respect to y, we obtain the wall shear stress
du
U
w
xy (x) = . (8.2.44)
dy y=0 2 (x)
The displacement and momentum thicknesses dened in (8.2.32) and (8.2.37) are found to be
2 2 1
(x) = (1 ) (x) = 0.363 (x), (x) = ( ) (x) = 0.137 (x), (8.2.45)
2
and the shape factor is H = 2.660. It is reassuring to observe that the shape factor is remarkably
close to that arising from the exact solution of the Blasius equation, H = 2.591.
Substituting the expressions for the momentum thickness and wall shear stress into (8.2.42),
we derive an ordinary dierential equation for (x),
U d(x)
= 0.137 U 2 . (8.2.46)
2 (x) dx
0.327 x
1/2 x
1/2
w
xy (x) = U 2, (x) = 1.4043 , (x) = 0.8544 . (8.2.48)
Rex U U
These expressions are in excellent agreement with their exact counterparts shown in (8.2.23), (8.2.33),
and (8.2.38). However, the agreement is fortuitous and thus atypical of the accuracy of the approx-
imate method (Problem 8.2.2).
8.2 Boundary layers in nonaccelerating ow 651
(a) (b)
6
y
5
x
1
V
0
0.5 0 0.5 1 1.5 2
Figure 8.2.2 (a) Illustration of the Sakiadis boundary layer developing over a semi-innite belt trans-
lating in an otherwise quiescent uid. (b) Graphs of the self-similar streamwise velocity prole
u/V = f (solid line), its integral f (dashed line), and derivative f (dotted line).
Sakiadis investigated the ow due to the translation of a semi-innite belt with velocity V along
the x axis normal to a vertical stationary wall, as illustrated in Figure 8.2.2(a) [360, 361]. At high
velocities, a boundary layer is established over the belt, while the uid is quiescent far from the
belt. The Sakiadis boundary layer is the mirror image of the Blasius boundary layer established
in streaming ow over a stationary plate discussed in Section 8.2.1, in that the wall and far-eld
velocities play opposite roles.
Similarity solution
Carrying out the boundary-layer analysis, we nd that the ow in the Sakiadis boundary layer is
governed by the Blasius ordinary dierential equation (8.2.9), where the belt velocity, V , is used in
place of the free-stream velocity, U , in the similarity variable
y V
1/2
= y. (8.2.49)
(x) x
The function f () satises the boundary conditions f (0) = 0 and f (0) = 1 and the far-eld condition
f () = 0. The solution can be found using the shooting method discussed in Section 8.2.1 for the
Blasius boundary layer. Numerical computations show that the far-eld condition is satised when
f (0) = 0.443748.
The prole of the streamwise velocity, u/V = f (), is drawn with the solid line in Figure
8.2.2(b) along with the proles of f () and f (), drawn with the dashed and dotted lines. As
in the case of the Blasius boundary layer, the curvature of the velocity prole is zero at the wall.
The 0.01% boundary-layer thickness, 0.01 , is dened as the elevation where u/V = 0.01. Using the
652 Introduction to Theoretical and Computational Fluid Dynamics
In fact, the velocity prole is reasonably well approximated with the complementary error
function, u/V erfc(/2). The corresponding numerical coecients on the right-hand sides of
(8.2.52) are 1.128 and 0.856 [361].
Writing a momentum integral balance over an innite control area with two parallel sides
perpendicular to the x axis, we nd that the momentum integral,
M = u2 dy, (8.2.55)
is constant, independent of x. Motivated by this observation and carrying out a dimensional analysis,
we express the stream function in the form
M
1/3
= x1/3 f (), (8.2.56)
8.2 Boundary layers in nonaccelerating ow 653
Figure 8.2.3 Illustration of a boundary layer developing due to the spreading of a two-dimensional or
axisymmetric jet discharging along the x axis into a quiescent ambient uid.
Making substitutions in the steady version of the Prandtl boundary-layer equation (8.1.12)
with U set to zero, we obtain a nonlinear homogeneous ordinary dierential equation,
1 1 2 1
f + f f + f = f + (f f ) = 0, (8.2.58)
3 3 3
accompanied by the boundary conditions f (0) = 0, f (0) = 0, and the far-eld condition f () = 0,
where a prime denotes a derivative with respect to . The integral constraint (8.2.55) requires that
f 2 d = 1. (8.2.59)
f () = 6 tanh(), (8.2.60)
where = 1/481/3 . The axial volumetric ow rate per unit width is given by
M x
1/3
Q(x) = u dy = 12 . (8.2.61)
We observe that Q(x) increases in the streamwise direction due to the entrainment of ambient uid.
654 Introduction to Theoretical and Computational Fluid Dynamics
Problems
Computer Problem
(a) (b)
x
Accelerating flow Decelerating flow
x
m>0 m<0
(c)
m<0
Decelerating flow
Accelerating flow
m>0
Figure 8.3.1 Examples of boundary layers in accelerating or decelerating ow. (a, b) Flow past a
wedge-shaped body and (c) streaming ow past a at plate at a nite angle of attack.
U (x) = xm , (8.3.1)
where is a positive dimensional coecient and m is a positive or negative exponent [121]. When
m = 0, we recover the Blasius ow over a semi-innite at plate at zero angle of attack. When
m = 1, we obtain orthogonal stagnation-point ow against a at plate. Intermediate values of
m correspond to accelerating symmetric ow past a wedge with semi-angle = m/(m + 1), as
illustrated in Figure 8.3.1(a).
Dierentiating the outer velocity described by (8.3.1), we obtain an expression the streamwise
acceleration of the outer ow
dU
= mxm1 , (8.3.2)
dx
which shows that the outer ow accelerates when m > 0 and decelerates when m < 0. In the rst
case, the continuity equation requires that the normal derivative of the velocity component of the
outer ow normal to the boundary, V , is negative, V /y < 0. Since V is zero at the wall, it must
be negative at the edge of the boundary layer. The motion of the uid normal to the boundary
contains the vorticity toward the boundary, thereby restricting the growth of the boundary layer.
656 Introduction to Theoretical and Computational Fluid Dynamics
Substituting (8.3.2) into the boundary-layer equation (8.1.12), we obtain the specic form
u u 2u
u +v = 2 mx2m1 + 2 . (8.3.3)
x y y
Working as in Section 8.2 for the Blasius boundary layer, we identify the characteristic length L
with the current streamwise position, x, and use (8.1.7) to write
x
1/2
1/2
(x) = x(1m)/2 . (8.3.4)
U (x)
The exponent (1 m)/2 determines the rate of spatial growth of the boundary layer in accelerating
or decelerating ow.
Similarity solution
Next, we postulate that the velocity prole across the boundary is self-similar, that is, u is a function
of the similarity variable
y U (x)
1/2
1/2
=y =y , (8.3.5)
(x) x x1m
and write
where F () is a dimensionless function. This self-similar velocity prole can be derived from the
stream function shown in (8.2.5), repeated here for convenience,
where U (x) is given in (8.3.1) and F () = f (). Dierentiating (8.3.7) and using the relations
m 1 y
1/2
1/2
= , = , (8.3.8)
x 2 x x1m y x1m
we obtain the velocity components
df df
u(x, y) = = U (x) = xm (8.3.9)
y d d
and
1 df
v(x, y) = = (xm1 )1/2 [(1 m) (1 + m) f ]. (8.3.10)
x 2 d
Further dierentiation provides us with expressions for the derivatives of the velocity involved in the
boundary-layer equation,
u df d2 f u d2 f 2u d3 f
2
= mxm1 +U , =U , = U . (8.3.11)
x d d 2 x y d 2 y y2 d 3 y
8.3 Boundary layers in accelerating or decelerating ows 657
Substituting these expressions into (8.3.3), we derive a third-order nonlinear inhomogeneous ordinary
dierential equation for the function f ,
1
f + (m + 1) f f m f 2 + m = 0. (8.3.12)
2
The boundary and far-eld conditions require that f (0) = 0, f (0) = 0, and f () = 1. When
m = 0, equations (8.3.9), (8.3.10), and (8.3.12) reduce to the Blasius equations (8.2.8) and (8.2.9).
When m = 1, we obtain equation (5.6.16), providing us with an exact solution to the equation of
motion describing orthogonal stagnation-point ow.
Velocity prole
Applying the FalkerSkan boundary-layer equation (8.3.12) at the wall, y = 0, and enforcing the
aforementioned boundary conditions, we nd that f (0) = m, which is positive when m < 0,
corresponding to decelerating ow. Thus, the curvature of the velocity prole at the wall is positive
in a decelerating ow. Noting that f must become negative at a suciently high value of
for the streamwise velocity f to tend to a constant value, reveals the presence of an inection
point. In Chapter 9, we will see that the inection point renders the boundary layer susceptible to
hydrodynamic instability mediated by the growth of small perturbations.
Numerical solution
Since boundary conditions are specied at both ends of the solution domain, [0, ), we encounter
a two-point boundary-value problem involving a third-order dierential equation. To solve the
FalknerSkan equation, we rename x1 = f , denote the rst and second derivatives of the function f
as x2 df /d and x3 dx2 /d = d2 f /d 2 , and decompose the third-order equation into a system
of three rst-order nonlinear equations,
The accompanying boundary and far-eld conditions are shown in (8.2.17). Working as in the case of
the Blasius equation, we implement Newtons method by dening the derivatives (8.2.18) satisfying
the initial conditions (8.2.20). Dierentiating equations (8.3.13) with respect to x3 (0), we obtain
The shooting method is implemented by guessing a value for x3 ( = 0), integrating equations (8.3.13)
and (8.3.14) with the aforementioned initial conditions, and replacing the guessed value, x3 ( = 0),
with an updated value computed using formula (8.2.21).
(a) (b)
2 5
1.8 4.5
1.6 4
1.4 3.5
1.2
x (0), ,
3
1 2.5
0.8
3
2
0.6
1.5
0.4
1
0.2
0.5
0
0
0 0.5 1 1.5 2 0 0.2 0.4 0.6 0.8 1
m u/U
Figure 8.3.2 (a) Graphs of the wall shear stress coecient x3 (0) = f (0) (solid line), coecient
pertaining to the displacement thickness (dashed line), and coecient pertaining to the momen-
tum thickness (dotted line). (b) Velocity proles across the FalknerSkan boundary layer: from
bottom to top, m = 1.8 (dotted line), 1.6, 1.4, . . . 0.4, 0.3, 0.2, 0.1, 0.0 (dashed line representing
the Blasius boundary layer), 0.05, 0.08, and 0.0904.
solution that appears to be most physical relevant is retained. Results for x3 (0) obtained using the
shooting method discussed previously in this chapter are shown with the solid line in Figure 8.3.2(a).
Velocity proles expressed by the derivative f () are shown in Figure 8.3.2(b) for several values of
m. The proles for m < 0, corresponding to decelerating ow, exhibit an inection point near the
wall.
The wall shear stress is found by dierentiating the streamwise velocity prole and evaluating the
resulting expression at the wall, nding
u
w
xy (x) = = f (0) 3/2 x(3m1)/2 . (8.3.15)
y y=0
We observe that, when m = 1/3, the shear stress is independent of the streamwise position, x;
when m > 1/3, the shear stress increases with the streamwise position; when m < 1/3, the shear
stress decreases with the streamwise position; when m = 1, corresponding to orthogonal stagnation-
point ow, the shear stress increases linearly with the streamwise position; when m = 0.0904,
corresponding to decelerating ow, f (0) = 0 and the shear stress, and thus the skin friction, vanish
uniformly along the wall. The rate of convection of vorticity across a plane that is perpendicular to
the wall is given by the right-hand side of (8.2.28). Because U depends on the streamwise position,
vorticity must diuse across the wall in order to satisfy the vorticity balance at every x station.
8.3 Boundary layers in accelerating or decelerating ows 659
The displacement and momentum thicknesses dened in equations (8.2.32) and (8.2.37) are
given by
x
1/2
1/2
u
(x) 1 dy = = x(1m)/2 (8.3.16)
0 U U
and
x
1/2
1/2
u u
(x) 1 dy = = x(1m)/2 , (8.3.17)
0 U U U
where
= (1 f ) d, = f (1 f ) d (8.3.18)
0 0
are dimensionless coecients dependent on m, plotted with the dashed and dotted lines in Figure
8.3.2(a). When m > 1, both and decrease in the streamwise direction due to the acceleration
of the incident ow. When m < 1, both and increase in the streamwise direction due to
the deceleration of the incident ow. When m = 1, corresponding to orthogonal stagnation-point
ow, the displacement and momentum thicknesses are constant along the wall. Physically, viscous
diusion of vorticity is balanced by convection against the at plate.
V (x) = xm , (8.3.19)
Similarity solution
As a preliminary, we introduce the similarity variable dened in (8.3.5), where the sheet velocity,
V (x), is used in place of the far-eld velocity, U (x),
y V
1/2
1/2
=y =y . (8.3.20)
(x) x x1m
Performing the boundary-layer analysis in the absence of outer ow, U = 0, we nd that the function
f () satises the homogeneous FalknerSkan equation
1
f + (m + 1) f f m f 2 = 0, (8.3.21)
2
660 Introduction to Theoretical and Computational Fluid Dynamics
(a) (b)
2.5 8
7
2
6
5
x (0), ,
1.5
1
3
2
0.5
1
0 0
0.5 0 0.5 1 1.5 2 0 0.2 0.4 0.6 0.8 1
m u/U
Figure 8.3.3 (a) Graphs of the negative of the wall shear stress coecient x3 (0) = f (0) (solid
line), coecient pertaining to the displacement thickness (dashed line), and coecient per-
taining to the momentum thickness (dotted line). (b) Velocity proles across the boundary layer
above a stretching sheet: from bottom to top, m = 1.8 (dotted line), 1.6, 1.4, 1, 2, 1.0 (dot-dashed
line), 0.8, . . ., 0.0 (dashed line representing the Sakiadis boundary layer), 0.10, 0.20, 0.28,
and 0.33.
where a prime denotes a derivative with respect to . The boundary and far-eld conditions require
that f (0) = 0, f (0) = 1, and f () = 0. When m = 0, we obtain the Blasius equation describing
the Sakiadis boundary layer over an inextensible belt. When m = 1, we obtain the homogeneous
Hiemenz equation (5.6.32), which is satised by Cranes exact solution of the unsimplied Navier
Stokes equation, f () = 1 e , as discussed in Section 5.6.3.
Solutions for other values of m computed by a shooting method are shown in Figure 8.3.3.
The wall shear stress tends to zero at a critical yet unphysical negative value of m. Applying (8.3.21)
at the wall, y = 0, and enforcing the aforementioned boundary conditions, we nd that f (0) = m,
which is negative when m < 0, corresponding to decelerating ow. Thus, the curvature of the
velocity prole at the wall is negative in a decelerating ow. Since f must become positive as
tends to innity for the streamwise velocity f to decay to zero, we infer the presence of an inection
point in the velocity prole. In Chapter 9, we will see that the inection point renders the boundary
layer susceptible to hydrodynamic instability mediated by the growth of small perturbations.
The displacement and momentum thicknesses dened in equations (8.2.51) are given by
x
1/2
1/2
u
(x) dy = = x(1m)/2 (8.3.22)
0 V V
8.3 Boundary layers in accelerating or decelerating ows 661
and
u
2 x
1/2
1/2
(x) dy = = x(1m)/2 , (8.3.23)
0 V V
where
= f d = f (), = f 2 d (8.3.24)
0 0
are dimensionless coecients plotted with the dashed and dotted lines in Figure 8.3.3(a) against
the exponent, m. When m > 1, both and decrease in the streamwise direction due to the
severe stretching of the sheet. When m < 1, both and increase in the streamwise direction
due to the deceleration of the sheet. When m = 1, corresponding to the Crane boundary layer, the
displacement and momentum thicknesses are constant over the sheet.
Similarity solution
We are interested in situations where the sheet velocity, V (x, t) = u(x, y = 0, t), allows for a similarity
solution such that
x y
(x, y, t) = t (x, y), x = , y = , (8.3.26)
t t
where the exponents , , and are determined as part of the solution and a tilde indicates a
similarity variable. The two velocity components are given by
u= = t , v= = t . (8.3.27)
y y x x
Substituting these expressions along with the expressions
u 2 2
u 2
= t1 ( ) x y , = t , (8.3.28)
t y x y y 2 x x y
and
u 2 2u 3
3
= t2 , = t , (8.3.29)
y y 2 y2 y 3
662 Introduction to Theoretical and Computational Fluid Dynamics
2 2 2 2
3
2+1
( ) x y + t+1
= t . (8.3.30)
y x y y 2 y x y x y 2 y 3
To eliminate the explicit time dependence on both sides of this equation, we require that
1 1
= , = , (8.3.31)
2 2
and derive a nonlinear partial dierential equation for x and y,
2 1 2 3
( 1) x y + = . (8.3.32)
y y x y 2 x y 2 y 3
The no-penetration boundary condition requires that (x, 0) = 0, and the far-eld condition requires
that / x 0 and / y 0 as y . The value of and the choice of further boundary
conditions depend on the specics of the problem under consideration, as discussed in Sections 8.3.4
and 8.3.5.
Equation (8.3.32) admits the solution (x, y) = xf (y) + , resulting in an ordinary dier-
ential equation,
1
f + ( y + f )f + f (1 f ) = 0, (8.3.33)
2
where is a constant and a prime denotes a derivative with respect to y. The no-penetration
condition requires that f (0) = 0, and the far-eld condition requires that f (y) decays to zero as y
tends to innity.
where q is a positive exponent, = y/x is a similarity variable, and the positive exponent is
determined as part of the solution. Substituting (8.3.34) into (8.3.32) and simplifying, we nd that
1
[(1 q + ) 1] f + ( ) f + xq1 [(q ) f 2 q f f ] = x2 f . (8.3.35)
2
This equation is physically acceptable under two complementary conditions. First, setting the power
of x on the right-hand side of (8.3.35) to zero, = 0, we obtain
1
f + y f [(1 q) 1] f = q xq1 ( f 2 f f ). (8.3.36)
2
When q > 1, the right-hand side vanishes as x 0, yielding a homogeneous integral equation,
1
f + y f [(1 q) 1] f = 0. (8.3.37)
2
8.3 Boundary layers in accelerating or decelerating ows 663
x a a
Figure 8.3.4 Illustration of an unsteady boundary layer extending between x = a(t) and a(t) due
to the spreading of a two-dimensional oil slick at the surface of the ocean.
Alternatively, we require that the power of x on the right-hand side of (8.3.35) matches that of the
last term on the left-hand side, q 1 = 2 or = 1 q, and obtain
1
f + q f f (2q 1) f 2 = x2 (2 1) f + f . (8.3.38)
2
The right-hand side vanishes as x 0, yielding the homogeneous FalknerSkan equation (8.3.21)
with m = 2q1. The properties of these equations have been investigated with respect to uniqueness
of solution [135, 296, 297].
Oil ow
The oil phase will be denoted by the superscript or subscript o. To leading order, the x and y
components of the equation of motion inside the oil lm take the form
o
xy po po
= , = o g, (8.3.39)
y x y
o
where xy is the shear stress. Integrating the second equation and requiring that the pressure at
the oilair interface, located at y = h(x, t), is equal to the atmospheric pressure, patm , we obtain
the oil pressure distribution
po = o g (y + h) + patm . (8.3.40)
Substituting this expression into the rst equation in (8.3.39), and integrating with respect to y
subject to the condition of zero shear stress at the oilair interface, we obtain the interfacial shear
664 Introduction to Theoretical and Computational Fluid Dynamics
Continuity of velocity and shear stress at the oilwater interface requires that
u
h
o
u(y = 0) = uo (y = 0) V (x), = xy (y = 0) = o gh , (8.3.44)
y y=0 x
where u is the x component of the water velocity.
Similarity solution
Progress can be made by assuming that the boundary layer develops according to the similarity
solution discussed in Section 8.3.3. The edges of the spreading lm are located at
x = a(t) = x t , (8.3.45)
where x is a constant to be determined as part of the solution.
Substituting the similarity solution into (8.3.43), stipulating that Ao (t) = qo tn , and rearrang-
ing, we obtain
x
Ao (t) = qo tn = t h(x, t) dx, (8.3.46)
0
where qo is a constant coecient and n is a specied exponent. This equation is satised only if
h(x, t) = tn h(x), (8.3.47)
where h(x) is an unknown function to be determined as part of the solution. Substituting this
expression along with the similarity solution into the interfacial shear stress condition given in the
second equation in (8.3.44), we obtain
2
dh
t+322n = o g h . (8.3.48)
y 2 y=0 dx
To eliminate the explicit time dependence on the left-hand side, we require that the exponent of
time, t, on the left-hand side is zero. Combining this condition with (8.5.36), we obtain
4n 1 4n + 3 1
= , = , = . (8.3.49)
8 8 2
8.3 Boundary layers in accelerating or decelerating ows 665
dh2 2
=2 , (8.3.50)
dx o g y 2 y=0
The similarity solution predicts that the x component of the water velocity is
u = t , (8.3.51)
y
and this motivates expressing the mean oil velocity in the corresponding form
x
uo = t x G(x) = t x G(x) = G(x), (8.3.52)
t
where G(x) is an unknown function. Combining this expression with (8.3.47), we nd that
Substituting this expression along with (8.3.47) into the mass balance equation (8.3.42), we obtain
an ordinary dierential equation,
d
x(G )h] + nh = 0. (8.3.54)
dx
When n = 0, this equation is satised by the constant function G(x) = . Assuming with good
reason that h is nearly constant except at the leading edge, we derive the more general, albeit
approximate, result
G(x) n. (8.3.55)
A mass balance over the oil lm provides us with an evolution equation for the lm thickness
h (us h)
+ = 0, (8.3.57)
t x
666 Introduction to Theoretical and Computational Fluid Dynamics
where us is the surface velocity parallel to the interface. This equation also describes the evolution of
the surface concentration of an insoluble surfactant, . In the preceding and forthcoming equations,
is substituted for h.
In the similarity solution, the lm thickness and surface tension are assumed to depend ex-
clusively on the scaled coordinate x, so that
3 dh d 0
x + h = 0, (8.3.61)
4 dx dx y
which can be restated as
d 0 3 3
y x h + h = 0, (8.3.62)
dx 4 4
where y0 ( / y)y=0 . Numerical solutions of the boundary-layer equations have been obtained
by nite-dierence methods [135, 297].
where f is an unknown function. Substituting this functional form into the boundary-layer equation,
we obtain a third-order partial dierential equation,
1 f f
f + (m + 1) f f mf 2 + m = l f f , (8.3.65)
2 l l
8.4 Integral momentum balance analysis 667
where m = (l/U ) dU/dl and a prime denotes a partial derivative with respect to . The boundary
and far-eld conditions require that f (l, 0) = 0, f (l, 0) = 0, and f (l, ) = 1. In the case of the
FalknerSkan boundary layer, the velocity prole is described by (8.3.1) with l = x and = y,
the coecient m is constant, the function f is independent of x and only depends on , and the
right-hand side of (8.3.65) is identically zero yielding the FalknerSkan ordinary dierential equation
(8.3.12).
Approximating the partial derivatives with respect to l in (8.3.65) with rst- or second-order
backward dierences, we obtain an ordinary dierential equation for f with respect to that can
be solved subject to the aforementioned boundary conditions using a standard shooting method. In
another implementation of the method, equation (8.3.65) is recast as a system of three rst-order
equations with respect to , which is then integrated using a standard nite-dierence method.
Problem
Computer Problems
y U
x
x1 x2
Figure 8.4.1 A rectangular control volume introduce to perform an integral momentum balance. The
balance provides us with a relation between the displacement thickness, the momentum thickness,
and the wall shear stress.
other types boundary-layer ows. In this section, we discuss the development and implementation
of the method to ow past a solid surface or due to in-plane boundary motion in an otherwise
quiescent uid. When appropriate, the method will be validated by comparison with exact or
numerical solutions of the boundary-layer equations derived by dierent methods.
Next, we consider a control area conned between two vertical planes located at x1 and x2 ,
the at surface located at y = 0, and a horizontal plane located at the elevation y = h, as shown
in Figure 8.4.1. Introducing the Newtonian constitutive equation for the stress tensor, neglecting
the normal viscous stresses expressed by the term 2 ux /x or 2 uy /y over the vertical and
top planes, and also neglecting gravitational eects, we nd that the x component of the integral
momentum balance becomes
x2 h h h
u
dy dx [u(u)]x=x1 ,y dy + [u(u)]x=x2 ,y dy
x1 0 t 0 0
x2 h h x2
+ [v(u)]x,y=h dx = (p)x=x1 ,y dy (p)x=x2 ,y dy w
xy (x) dx. (8.4.2)
x1 0 0 x1
Taking the limit as x1 tends to x2 , recalling that the pressure is constant across the boundary layer,
setting u(x, y = h) = U (x), and rearranging, we obtain an integro-dierential relation,
h p
h
u
dy = h u2 dy U (x) v x,y=h xyw
(x). (8.4.3)
o t x y=h x 0
8.4 Integral momentum balance analysis 669
To reduce the number of unknowns, we eliminate the y velocity at the top plane, v(x, y = h), in
favor of the velocity prole, u, using the continuity equation, setting
h
u
v x,y=h = (x, y ) dy . (8.4.4)
0 x
Moreover, we use the x component of Eulers equation (6.5.3) to evaluate the streamwise pressure
gradient, nding
p U U
= ( +U ). (8.4.5)
x t x
Substituting expressions (8.4.4) and (8.4.5) into (8.4.3) and rearranging, we obtain
h
(U u) h
dy = u(U u) dy
o t x 0
h h
U (U u) dy + U (U u) dy + xy
w
(x), (8.4.6)
x 0 x 0
which can be interpreted as an evolution law for the momentum decit expressed by the term
(U u).
Equation (8.4.8) can be derived from the continuity equation (8.1.1) and the boundary-layer
equation (8.1.12) in a way that bypasses the integral momentum balance [301]. Multiplying the
left-hand side of (8.1.1) by U u, adding the product to the right-hand side of (8.1.12), using the
continuity equation, integrating the result from y = 0 to , and enforcing the boundary conditions,
we obtain (8.4.8). This derivation ensures that no error is introduced by replacing the boundary-layer
equations with the integral momentum balance.
670 Introduction to Theoretical and Computational Fluid Dynamics
Substituting the expressions for the displacement and momentum thickness given in (8.3.16) and
(8.3.17) for m = 1, we nd that
w
xy = (2 + ) 3/2 x, (8.4.11)
which is consistent with (8.3.15) for m = 1, provided that f (0) = 2 + . Substituting expressions
(8.3.18), we obtain
f (0) = (2f 2 + f + 1) d. (8.4.12)
0
To conrm this identity, we recast the FalknerSkan equation (8.3.12) for m = 1 into the form
f + f (f 1) 2f 2 + f + 1 = 0. (8.4.13)
Integrating and using the boundary and far-eld conditions f (0) = 0, f (0) = 0, and f () = 1,
reproduces (8.4.12).
main idea is to introduce a meaningful boundary-layer thickness, (x), similar to the 99 boundary-
layer thickness, and then employ a sensible velocity prole across the boundary layer, u/U = F (),
where y/(x) is a dimensionless similarity variable. At the second stage, we compute (x) to
satisfy the integral momentum balance (8.4.8) for solving a dierential equation.
The implementation of the method in the case of ow over a at plate at zero angle of attack
where F () is a quarter of a period of a sinusoidal function, as shown in (8.2.43), was discussed in
Section 8.2. In this section, we illustrate the implementation for arbitrary steady ows.
Pohlhausen polynomials
Pohlhausen described the velocity prole across the boundary layer by a fourth-order polynomial,
&
u a(x) + b(x) 2 + c(x) 3 + d(x) 4 for 0 < < 1,
= F () = (8.4.14)
U 1 for > 1,
where a(x)d(x) are four position-dependent coecients to be computed as part of the solution.
The functional form (8.4.14) satises the no-slip boundary condition at the wall corresponding to
= 0. To compute the four coecients, we require four equations. First, we demand that the overall
velocity prole is continuous with smooth rst and second derivatives at the edge of the boundary
layer corresponding to = 1, and thus obtain three conditions,
F ( = 1) = 1, F ( = 1) = 0, F ( = 1) = 0, (8.4.15)
where a prime denotes a derivative with respect to . A fourth condition arises by applying the
boundary-layer equation (8.1.13) at the wall located at y = 0, and then using the no-slip and no-
penetration conditions to set the left-hand side to zero. Evaluating the streamwise pressure gradient
using (8.4.5) with the time derivative on the right-hand side set to zero, we obtain
2u
= U U , (8.4.16)
y 2 y=0
where U dU/dx. Next, we express the velocity in terms of the function F () introduced in
(8.4.14), and obtain
F ( = 0) = , (8.4.17)
where
2 (x)
(x) U (8.4.18)
is a dimensionless function expressing the ratio of the magnitude of the inertial acceleration forces
in the outer irrotational ow to the magnitude of the viscous forces developing inside the boundary
layer; if dU/dx = 0, then = 0. By denition, the eective boundary-layer thickness (x) is related
to (x) by
1/2
(x) . (8.4.19)
U
672 Introduction to Theoretical and Computational Fluid Dynamics
(a) (b)
1 0.15
0.9
0.8 0.1
0.7
0.6 0.05
0.5
0.4 0
0.3
0.2 0.05
0.1
0 0.1
0.2 0 0.2 0.4 0.6 0.8 1 1.2 0 0.5 1 1.5 2
F m
Figure 8.4.2 (a) Proles of the Pohlhausen polynomials for = 20, 12 (dashed ), 6, 0, 6, 12, and
15. (b) Exact (solid line) and approximate (dashed line) values of the HolsteinBohlen function
for the FalknerSkan boundary layer in accelerating or decelerating ow.
Requiring that the Pohlhausen prole (8.4.14) satises equations (8.4.15) and (8.4.17), we obtain
a=2+ , b= , c = 2 + , d=1 . (8.4.20)
6 2 2 6
Substituting these expressions into (8.4.14) and rearranging, we obtain the velocity prole in terms
of the parameter ,
&
u (2 2 2 + 3 ) + 16 (1 )3 for 0 < < 1,
= F () = (8.4.21)
U 1 for > 1.
A family of proles for = 20, 12 (dashed line), 6, 0, 6, 12, and 15, is shown in Figure
8.4.2(a). When > 12, corresponding to a strongly accelerating external ow according to (10.5.14),
the prole exhibits an overshooting, which undermines the physical relevance of the fourth-order
polynomial expansion. When = 12, the slope of the velocity prole is zero at the wall and the
ow is on the verge of reversal. At that point, the approximations that led us to the boundary-layer
equations cease to be valid, and the boundary layer is expected to separate from the wall producing
a region of recirculating uid attached to the wall.
The displacement thickness, momentum thickness, and wall shear stress can be computed in
terms of (x) and (x) using the proles (8.4.21), and are found to be
8.4 Integral momentum balance analysis 673
1 1
= 1 F () dy = (3 ),
0 10 12
1 1 5 2
= F () 1 F () dy = (37 ), (8.4.22)
0 315 3 144
w U F ()
U 1
xy = = (2 + ).
=0 6
Expressing (x) in terms of (x) using the denition (8.4.19), we obtain corresponding expressions
in terms of (x) and the known distribution U (x). Substituting expressions (8.4.22) into the
momentum integral balance (8.4.8) provides us with a rst-order nonlinear ordinary dierential
equation for (x) with respect to x. Having solved this equation, we recover the boundary-layer
thickness, (x), using the denition (8.4.19).
d 630
= , (8.4.23)
dx 37 U
which can be integrated from the leading edge up to a certain point, x, subject to the initial condition
(x = 0) = 0, to give
x
1/2
(x) = 5.836 . (8.4.24)
U
It is instructive to compare this expression with that shown in (8.2.47), corresponding to a sinusoidal
velocity prole, and note a signicant dierence in the numerical coecient on the right-hand side.
Substituting (8.4.24) into (8.4.22), we nd that
0.343 x
1/2 x
1/2
w
xy = U 2, = 1.751 , = 0.685 , (8.4.25)
Rex U U
which are reasonably close to the exact expressions given in (8.2.23), (8.2.33), and (8.2.38). The
predicted shape factor is H = 2.556.
HolsteinBohlen function
In the case of the Blasius boundary layer, we were able to derive an analytical solution for (x).
Under more general circumstances, the solution must be obtained by numerical methods. It is
convenient to introduce the dimensionless HolsteinBohlen function
2 (x) 2 (x)
(x) (x) = U (x), (8.4.26)
2 (x)
674 Introduction to Theoretical and Computational Fluid Dynamics
whose physical interpretation is similar to that of dened in (8.4.18). Using the expression for the
momentum thickness given in (8.4.22), we obtain a relationship between and ,
1 1 5 2 2
= 2
(37 ) . (8.4.27)
315 3 144
The value = 12 corresponds to = 0.15673 where the boundary layer is expected to separate,
as shown in Figure 8.4.2(a). In the case of the FalknerSkan boundary layer, we substitute into
(8.4.26) the velocity distribution U = xm and the analytical solution for the momentum thickness
given in (8.3.17), and nd that = m 2 , which is plotted with the solid line in Figure 8.4.2(b). As
the exponent m increases from zero, yielding a strongly accelerating ow, increases and tends to
a constant.
Numerical solution
To expedite the numerical solution, we multiply both sides of the momentum integral balance (8.4.8)
at steady state by , and rearrange to obtain
d
1 d2 T ()
=2 , (8.4.28)
dx U dx U
where
T () S() 2 + H() , (8.4.29)
H / is the dimensionless shape factor, and S() is a dimensionless shear function dened as
w
S() . (8.4.30)
U xy
Physically, the shear function expresses the ratio between the wall shear stress and the average value
of the shear stress across the boundary layer, and is thus another measure of the bluntness of the
boundary-layer velocity prole. Using expressions (8.4.25), we nd that
1
315 3 12 1 1 5 2
H= 1 5 , S= (2 + ) (37 ), (8.4.31)
10 37 3 144 2 315 6 3 144
In the case of the Blasius or FalknerSkan boundary layer, equation (8.4.28) is satised if
is a zero of the nonlinear algebraic equation
2m T () + (m 1) = 0. (8.4.32)
A physically acceptable solution branch is shown with the dashed line in Figure 8.4.2(b). When
m = 1, corresponding to orthogonal stagnation-point ow, we nd that = 0.0770356 corresponding
to = 7.0523231, indicated by the reclusive circle in Figure 8.4.2(b).
More generally, equation (8.4.28) must be solved by numerical methods. The numerical pro-
cedure involves the following steps:
8.4 Integral momentum balance analysis 675
The numerical integration typically begins at a stagnation point where the tangential velocity U
vanishes and the right-hand side of (8.4.28) is undened. The initial value of is found by solving
equation (8.4.32) with a proper value for the exponent m pertinent to the ow near the stagnation
point. Although the assumptions that led us to the boundary-layer equations cease to be valid at
the stagnation point where the local ow is not nearly unidirectional, starting the integration at a
stagnation point does not undermine the validity of the solution.
d
dT d
U
= 2 + . (8.4.34)
dx U d d U (U )2
Combining the left-hand side with the rst term inside the square brackets on the right-hand side,
we nd that
d
2 U
= T , (8.4.35)
dx U 1 2T (U )2
where T = dT /d. Evaluating the expression on the right-hand side using the denition of T (),
we extract the required initial value
d
U
= 0.0652 , (8.4.36)
dx U x=0 (U )2 x=0
boundary, l, measured in the direction of the tangential velocity of the incident ow, and begin the
integration at a stagnation point. A diculty arises at the critical point where the acceleration dU/dl
becomes zero or innite. However, the ambiguity can be removed by carrying out the integration
at that point using the FalknerSkan similarity solution with a proper value for the exponent m
(Problem 8.4.4).
U () = = 2U sin , (8.4.37)
r=a
where is the polar angle measured around the center of the cylinder in the counterclockwise
direction, as shown in Figure 8.4.3(a). The arc length around the cylinder measured from the front
stagnation point is l = a. The required derivatives of the velocity with respect to arc length are
dU 1 dU U d 2 U 1 d2 U U
= =2 cos , 2
= 2 2
= 2 2 sin . (8.4.38)
dl a d a dl a d a
Equation (8.4.36) yields
= 0, (8.4.39)
dl dU /dl l=0
Graphs of the solution are shown in Figure 8.4.3(be). The velocity prole across the bound-
ary layer at dierent stations around the cylinder can be inferred from the scaled Pohlhausen proles
shown in Figure 8.4.2 using the local value of . The shear function becomes negative when separa-
tion occurs, even though the average value of the shear stress across the boundary layer is positive.
At the front stagnation point, the values of all functions shown in Figure 8.4.3 are close to those
predicted by the FalknerSkan similarity solution for orthogonal stagnation-point ow with m = 1
and shear rate = 2U /a.
The numerical solution reveals that = 12 at the meridional angle = 109.5 . At that
point, the shear stress becomes zero and the boundary layer is expected to separate. Comparing this
result with the experimentally observed value = 80.5 , we nd a serious disagreement attributed
to the deviation of the actual outer ow from the idealized potential ow described by (7.9.9), due to
the presence of a wake. To improve the solution, we may describe the tangential velocity distribution
U by interpolation based on data collected in the laboratory. When this is done, the predictions of
the boundary-layer analysis are in excellent agreement with laboratory observation ([427], p. 323).
8.4 Integral momentum balance analysis 677
(a)
(b) (c)
1 5
4
0.5
, ,
0
/12,
0.5
1
1 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0 0.1 0.2 0.3 0.4 0.5 0.6
/ /
(d) (e)
2.5 4
2
3.5
1.5
S
3
xy,
2.5
0.5
0 2
0 0.1 0.2 0.3 0.4 0.5 0.6 0 0.1 0.2 0.3 0.4 0.5 0.6
/ /
Figure 8.4.3 (a) Illustration of the Prandtl boundary layer around a circular cylinder of radius a
held stationary in an incident streaming ow with velocity V . (b) Distribution of the dimensionless
1
parameters 12 (solid line) and (dashed line) computed by the von Karman-Pohlhausen method.
(c) Boundary-layer thickness (solid line), displacement thickness (dashed line), and momentum
thickness (dotted line); all are reduced by (a/V )1/2 . (d) Distribution of the wall shear stress
reduced by V /a (solid line) and shape factor S (dashed line). (e) Distribution of the shear
function, H.
678 Introduction to Theoretical and Computational Fluid Dynamics
Thwaites function
Thwaites proposed replacing the function T () dened in (8.4.29) by the linear function T () =
0.45 6.0, motivated by laboratory observations [402]. An analytical solution can be obtained,
0.45 x 5
2 (x) = 2 (x0 ) + U () d, (8.4.40)
U 6 x0
where x0 is an arbitrary point (Problem 8.4.3). The predictions of this equation are in excellent
agreement with experimental observations.
KarmanPohlhausen method
To implement the Karman Pohlhausen method for steady ow, we introduce the similarity variable
= y/(x), where (x) is an appropriate boundary-layer thickness, and dene the dimensionless
function
2
V (8.4.42)
expressing the ratio of the magnitude of the inertial acceleration forces to the magnitude of the
viscous forces developing inside the boundary layer, where V dV /dx. If V = 0, then = 0.
Conversely, the boundary-layer thickness (x) is related to the dimensionless parameter (x) by
1/2
. (8.4.43)
V
8.4 Integral momentum balance analysis 679
The velocity prole across the boundary layer can be represented by the complementary
Pohlhausen polynomial,
u
= Fc () 1 F (), (8.4.44)
V
satisfying the boundary conditions
Fc ( = 1) = 0, Fc ( = 1) = 0, Fc ( = 1) = 0, Fc ( = 0) = , (8.4.45)
where a prime denotes a derivative with respect to . The last condition arises by applying the
boundary-layer equation at the wall, located at y = 0, and then using the no-slip and no-penetration
boundary conditions in the absence of a pressure gradient to obtain
2u
= V V . (8.4.46)
y2 y=0
Explicitly, the velocity prole is given by
&
u 1 2 + 2 3 4 16 (1 )3 for 0 < < 1,
= Fc () = (8.4.47)
V 0 for > 1.
The displacement thickness, momentum thickness, and wall shear stress can be computed in terms
of (x) and (x) using the proles (8.4.21), and are found to be
11 1 2 V 1
= (3 ), = (23 + ), w
xy = (2 + ). (8.4.48)
10 12 126 12 72 6
Expressing (x) in terms of (x) using the denition (8.4.43), we obtain corresponding expressions
in terms of (x) and the boundary distribution of the acceleration, V (x). Substituting expressions
(8.4.48) into the momentum integral balance (8.4.41) provides us with a rst-order nonlinear ordi-
nary dierential equation for (x) with respect to x. Having solved this equation, we recover the
boundary-layer thickness, (x), using the denition (8.4.43).
0.2
0.2
0.4
0.6
0.8
Figure 8.4.4 Exact (solid line) and approximate (dashed line) values of the HolsteinBohlen parameter
for the boundary layer due to a stretching sheet.
HolsteinBohlen parameter
In the case of ow due to boundary stretching, the dimensionless HolsteinBohlen function is dened
as
2 (x) 2 (x)
(x) (x) = V , (8.4.52)
2 (x)
where V = dV /dx. Using the expression for the momentum thickness given in (8.4.48), we obtain
a relationship between and ,
11 1 2 2
= (23 + ) . (8.4.53)
1262 12 72
In the case of the generalized Crane boundary layer, we substitute the velocity distribution
V = xm and the analytical solution for the momentum thickness given in (8.3.23) into the denition
(8.4.52), and nd that = m 2 , which is plotted with the solid line in Figure 8.4.4. When m = 0,
we obtain = 0 due to the absence of acceleration. As the exponent m increases from zero,
increases and tends to a constant value in a strongly accelerating ow.
Numerical solution
To expedite the numerical solution, we multiply both sides of the momentum integral balance (8.4.41)
at steady state by , and rearrange to obtain
d
1 d2 S() 2
=2 , (8.4.54)
dx V dx V
8.4 Integral momentum balance analysis 681
where
w
S() (8.4.55)
V xy
is a dimensionless shear function. Using expressions (8.4.48) based on the Pohlhausen polynomials,
we nd that
1 11 1 2
S= (2 + ) (23 + ). (8.4.56)
126 6 12 72
Equation (8.4.54) can be solved using the numerical methods discussed in Section 8.4.2.
In the case of the Sakiadis or generalized Crane boundary layer, equation (8.4.54) is satised
for a constant satisfying the nonlinear algebraic equation
2m S() 2 + (m 1) = 0. (8.4.57)
A physically acceptable solution branch is shown with the dashed line in Figure 8.4.4. The approx-
imate solution is reasonably close to the exact solution plotted with the solid line.
where a tilde denotes a similarity variable. Substituting these expressions into the integral momen-
tum balance expressed by (8.4.41), and recalling that = , we obtain an ordinary dierential
equation,
x d(V ) d 2 dV u
+ + = . (8.4.59)
V V 2 dx dx V dx V 2 y y=0
A similarity solution in the xy domain can be sought in terms of the complementary Pohlhausen
polynomials dened in (8.4.47),
where = x/(x) is a similarity variable and (x) is the boundary-layer thickness. To evaluate
the parameter inherent in the Pohlhausen polynomials, we apply the boundary-layer equation at
the surface of the stretching sheet and nd that
2u
V V
= +V = t3/2 ( ) V x V + V V . (8.4.61)
y 2 y=0 t x
682 Introduction to Theoretical and Computational Fluid Dynamics
d d x dV 2 dV 1
x = + (2 + ) . (8.4.65)
dx V dx V V dx dx 6
The solution must be found by numerical methods according to the physics of the problem under
consideration.
dh2 u
2 V
=2 = (2 + ), (8.4.66)
dx o g y y=0 o g 6
where o is the oil density. To circumvent the integrable singularity at the leading edge, x = x , we
introduce the variable = (x x)1/2 and recast (8.4.66) into the form
dh2 4 V
= (2 + ). (8.4.67)
d o g 6
Integrating from the leading edge where = 0 up to the centerline, we obtain the lm thickness.
d( c ) 8 11 1
= 2 + (2 + ) , (8.4.68)
dx 3 4n x 2 6
8.4 Integral momentum balance analysis 683
where
3 + 4n
c= = . (8.4.69)
n 3 4n
d( c )2 2 8 1
= 2 ( c ) + H() c (2 + ) , (8.4.70)
dx x 3 4n 2 6
where
c 1 11 1 2 c
H() = (23 + ) (3 ). (8.4.71)
126 12 72 10 12
Using relations (8.4.48) and the expression for the boundary-layer thickness in terms of from
(8.4.63), we nd that
1 11 1 2 c 2
( c )2 = (23 + ) (3 ) , (8.4.72)
V 126 12 72 10 12
which can be used to compute the value of , given the the left-hand side, by numerically solving an
algebraic equation. Equation (8.4.70) accompanied by (8.4.72) can be integrated from the leading
edge toward the centerpoint using elementary numerical methods.
The unknown value x corresponding to the leading edge is determined by the oil slick volume
per unit width, Ao = qo tn ,
g 2 q 4
1/8
o
x = , (8.4.73)
where is a numerical coecient. In practice, we may guess a value for x , solve the preceding
equations, compute the volume of oil, and then rescale x to ensure a specied volume. When
n = 0, the numerical computations show that = 1.74. The distribution of the boundary-layer
thickness for n = 0 computed using the numerical method discussed in this section is shown in
Figure 8.4.5. We see that the lm thickness h is nearly constant, except near the leading edge. The
results are in excellent agreement with those obtained by other numerical methods [60, 75].
Problems
(a) (b)
4.5 0.4
4 0.35
3.5
0.3
3
0.25
, , *
2.5
0.2
h
2
0.15
1.5
0.1
1
0.5 0.05
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x x
Figure 8.4.5 (a) Boundary layer thickness (solid line), momentum thickness (dashed line), and
displacement thickness (dotted line), along a two-dimensional oil slick expanding due to gravity,
reduced by x and plotted against the dimensionless similarity distance x/x . (b) Corresponding
distribution of the lm thickness for a xed amount of oil, n = 0.
Computer Problem
The formulation of the boundary-layer theory and the exact or approximate solution of the boundary-
layer equations for axisymmetric and three-dimensional ows are carried out according to the prac-
tices and ideas discussed earlier in this chapter for two-dimensional ow, with some minor modi-
cations.
8.5 Axisymmetric and three-dimensional ows 685
(a) (b)
x
y v
u
l
w
x
Figure 8.5.1 Illustration of (a) a boundary layer in axisymmetric ow past a body and (b) axisymmetric
orthogonal stagnation-point ow against a at surface.
Continuity equation
The continuity equation for axisymmetric ow in the absence or presence of swirling motion takes
the form
1 (u) v
+ = 0. (8.5.1)
l
When the distance from the axis of symmetry, , is much larger than the typical boundary-layer
thickness around the body contour, the continuity equation (8.5.1) can be recast into the approximate
convenient form
(u) (v)
+ = 0. (8.5.2)
l
To satisfy this equation, we may introduce a stream function, (l, ), dened by the equations
1 1
u= , v= . (8.5.3)
l
In the case of a cylindrical body, l = x, = , u = ux , and v = u . In the case of ow over a plane
that is perpendicular to the x axis, as shown in Figure 8.5.1(b), l = and = x, u = u and v = ux .
686 Introduction to Theoretical and Computational Fluid Dynamics
Boundary-layer equations
The normal component of the equation of motion states that the pressure drop across the boundary
layer can be neglected to leading-order approximation. Consequently, the pressure distribution,
p(l, ), can be computed using Eulers equation for the incident ow,
1 p U U W 2 1 1 p W W U W
= +U , = +U + , (8.5.4)
l t l l t l l
where U is the meridional tangential velocity and W is the azimuthal swirling velocity of the outer ir-
rotational ow. The corresponding components of the equation of motion simplify into the boundary-
layer equations
u u u w 2 1 p 2u
+u +v = + 2 (8.5.5)
t l l l
and
w w w uw 1 1 p 2w
+u +v + = + . (8.5.6)
t l l 2
In the absence of swirling motion, equation (8.5.6) is trivially satised and equation (8.5.5)
reduces into the boundary-layer equation (8.1.12) for two-dimensional ow with a straightforward
change in notation,
u u u U U 1 p 2u
+u +v = +U + 2. (8.5.7)
t l t l l
Manglers transformation
Mangler invented a remarkable transformation that reduces the boundary-layer equations (8.5.1)
and (8.5.7) for axisymmetric ow without swirling motion to the simpler boundary-layer equations
(8.1.1) and (8.1.12) for two-dimensional ow written for the judiciously chosen modied variables
1 l
L d
x = 2 dl, y = , u = u, v = v+ u , (8.5.8)
L2 0 L dl
where L is a specied length [254]. To demonstrate the reduction of the continuity equation, we
write
x y 2 d 2 y d
= + 2 + = 2 + (8.5.9)
l l x l y L x L dl y L x dl y
and
x y
= + . (8.5.10)
x y L y
8.5 Axisymmetric and three-dimensional ows 687
u L L2 y d
+ v+ 2 u = 0. (8.5.12)
x y dl
Now invoking the last denition in (8.5.8), we obtain the continuity equation for the transformed
variables in the two-dimensional xy domain.
Manglers transformation reduces the problem of computing a boundary layer over an axisym-
metric body to the problem of computing a boundary layer over a body with modied geometry
in a ctitious two-dimensional ow. A practical drawback is that the pressure distribution of the
ctitious ow can be considerably more involved than that of the physical ow.
1 (U ) 1 () 1 U
w
sh
+ + (2 + ) = , (8.5.13)
U2 t l U l U 2
which is a slight modication of (8.4.8) (Problem 8.5.1). The numerator of the fraction on the right-
w
hand side, sh , is the wall shear stress. The displacement and momentum thicknesses are dened as
in the case of two-dimensional ow in terms of integrals with respect to normal arc length, .
Using the fourth-degree Pohlhausen polynomial and the HolsteinBohlen dimensionless pa-
rameter introduced in (8.4.26), we nd that the momentum integral balance for steady ow takes
the form
1 d 2
T ()
=2 , (8.5.14)
2 dl U U
where the function T () dened in (8.4.29), and a prime denotes a derivative with respect to tangen-
tial arc length, l (Problem 8.5.1). At the front stagnation point of an axisymmetric body, reduces
to l and the tangential velocity U behaves like l, where is the local shear rate. Making substitu-
tions in the left-hand side of (8.5.14) and requiring that d/dl is nite, we obtain the starting value
= 0.0571. Equation (8.5.6) is integrated using the methods discussed earlier for two-dimensional
ow. Boundary-layer separation occurs when = 0.15673, corresponding to = 12.
In the case of uniform ow past a stationary sphere, numerical integration of (8.5.14) shows
that the boundary layer separates at the meridional angle = 110 , which is substantially higher than
the measured value 83 (Problem 8.5.6). The discrepancy is attributed to the poor representation
of the outer irrotational ow due to the presence of a wake.
688 Introduction to Theoretical and Computational Fluid Dynamics
and introduce the dimensionless parameter m = (l/U )dU/dl. The counterpart of equation (8.3.65)
for axisymmetric ow is
m+1 l d f f
f + + f f mf 2 + m = l f f , (8.5.16)
2 dl l l
where a prime denotes a partial derivative with respect to . The solution can be computed using
a shooting method or a direct nite-dierence method as in the case of two-dimensional ow.
U () = m , (8.5.18)
where is a positive constant and m is a positive exponent. The ow inside the boundary layer is
governed by the continuity equation,
1 (u ) ux
+ = 0, (8.5.19)
x
and the boundary-layer equation,
u u 2 u
u + ux = . (8.5.20)
x x2
To carry out the boundary-layer analysis, we introduce the dimensionless similarity variable
U
1/2
= x, (8.5.21)
8.5 Axisymmetric and three-dimensional ows 689
where
V
1/2
= x (8.5.27)
is a dimensionless similarity variable, the function f () satises the homogeneous equation
1
f + (m + 1) + 1 f f mf 2 = 0, (8.5.28)
2
and a prime denotes a derivative with respect to . The boundary conditions require that f (0) = 0
and f (0) = 1, and the far-eld condition requires that f () = 0.
layers were discussed in Section 8.3.3. The ow inside the axisymmetric boundary layers is governed
by the continuity equation (8.5.1) and the counterpart of equation (8.3.25) for axisymmetric ow,
u u u 2u
+u +v = , (8.5.29)
t x x2
where u = u is the radial velocity and v = ux is the axial velocity. The ow will be described in
terms of a time-dependent Stokes stream function, (, x, t), dened such that
1 1
u u = , v ux = . (8.5.30)
x
The ordered pair, (, x), denes a system of two Cartesian coordinates similar to the (x, y) coordi-
nates in two-dimensional ow.
Similarity solution
We are interested in situations where the radial sheet velocity, u (, x = 0, t) V (, t), allows for a
similarity solution such that
x
(, x, t) = t (, x), = , x = , (8.5.31)
ta t
where the exponents , , and will be determined as part of the solution. The radial and axial
velocity components are given by
1 1 1 1
u= = t , v= = t . (8.5.32)
x x
Substituting these expressions along with the expressions
u 1 2 2
= t1 ( ) x (8.5.33)
t x x x2
and
u 1 2 1
u 1 2 2u 1 3
= t , = t2 , = t3 , (8.5.34)
x x x x2 x2 x3
into the boundary-layer equation (8.5.29), we obtain
2 2
( ) x
x x x2
1 2 1 2
3
2+1
+t2+1 = t . (8.5.35)
x x x x2 x3
To eliminate the explicit time dependence on both sides of this equation, we set
1 1
= , = 2 , (8.5.36)
2 2
8.5 Axisymmetric and three-dimensional ows 691
1 1 2 1 1 2 3
2 1 x + = . (8.5.37)
2 x x x x 2 x2 x3
The no-penetration boundary condition requires that (, 0) = 0, and the far-eld condition requires
that / 0 and / x 0 as x .
where is a constant and a prime denotes a derivative with respect to x. The no-penetration
condition requires that f (0) = 0 and the far-eld condition requires that f (x) decays to zero as x
tends to innity.
(, x) q f (), (8.5.39)
where q is a positive exponent, = x/ is a similarity variable, and the positive exponent will be
determined as part of the solution. Substituting (8.5.39) into (8.5.37) and simplifying, we obtain
1
[(2 q + ) 1] f + ( ) f + q2 [(q 1) f 2 q f f ] = 2 f . (8.5.40)
2
This equation is physically acceptable under two complementary conditions. First, setting the power
of on the right-hand side of (8.5.40) to zero, = 0, we obtain
1
f + x f [(2 q) 1] f = q2 [ (q 1)f 2 qf f ]. (8.5.41)
2
When q > 2, the right-hand side vanishes as 0, yielding a homogeneous integral equation,
1
f + x f [(2 q) 1] f = 0. (8.5.42)
2
Alternatively, we require that the power of on the right-hand side of (8.5.40) matches that of the
last term on the left-hand side, q 2 = 2 or = 2 q, and thus obtain
1
f + q f f (2q 3) f 2 = 2 (2 1) f + f . (8.5.43)
2
The right-hand side vanishes as 0, yielding the homogeneous FalknerSkan equation (8.5.28)
with m = 2q 3. The properties of the equations presented in this section have been investigated
with respect to uniqueness of solution [296].
692 Introduction to Theoretical and Computational Fluid Dynamics
= t (), w
x = = t2 . (8.5.45)
x x=0 x x=0
Substituting these expressions into the integral momentum balance (8.5.44) and recalling that =
2 , we obtain an ordinary dierential equation,
d(V ) d 2 dV 1
u
( ) + + + = . (8.5.46)
V V 2 d d V d V 2 x x=0
Given the radial velocity, V (), the solution can be found by sensible approximate methods.
v
x
Figure 8.5.2 Illustration of an unsteady boundary layer due to the axisymmetric spreading of an oil
slick of radius a(t) at the surface of the ocean.
originating from the third equation in (8.4.48), and rearranging, we obtain the ordinary dierential
equation
d d dV 2 dV 1
= + (2 + ) . (8.5.52)
d V d V V d d 6
The solution must be found by numerical methods according to the physics of the problem under
consideration.
Oil ow
To leading order, the and x components of the equation of motion inside the thin oil lm, denoted
by the subscript or subscript o, take the form
o
x po po
= , = o g, (8.5.53)
x x
o
where x is the shear stress. Integrating the second equation and requiring that the pressure at
the oilair interface, located at x = h(, t), is equal to the atmospheric pressure, patm , we obtain
the oil pressure distribution
po = o g (x + h) + patm . (8.5.54)
Substituting this expression into the rst equation of (8.5.53) and integrating with respect to x
subject to the condition of zero shear stress at the oil-air interface, we obtain
o h
x (x = 0) = o g h . (8.5.55)
694 Introduction to Theoretical and Computational Fluid Dynamics
A mass balance for the oil provides us with an evolution equation for the lm thickness,
h 1 ( uo h)
+ = 0, (8.5.56)
t
where uo () is the mean radial velocity across the oil lm. The total volume of oating oil is
a(t)
Vo (t) = 2 h(, t) d. (8.5.57)
0
Continuity of velocity and shear stress at the oilwater interface requires that
u
h
o
u(x = 0) = uo (x = 0) V (), = x (x = 0) = o gh , (8.5.58)
x x=0
where u is the radial () component of the water velocity.
Similarity solution
We assume that the boundary layer develops according to the similarity solution derived in Section
8.5.3. The radius of the spreading lm is
= a(t) = t , (8.5.59)
where is a constant to be determined as part of the solution.
Substituting the similarity solution into (8.5.57), stipulating that Vo (t) = qo tn , and rearrang-
ing, we obtain
Vo (t) = qo tn = 2 t2 h() d, (8.5.60)
0
where qo is a constant coecient and n is a specied exponent. This equation is satised only if
h(, t) = tn2 h(), (8.5.61)
where h() is an unknown function to be determined as part of the solution. Substituting this
expression along with the similarity solution into the interfacial shear stress condition given by the
second equation in (8.5.58), we obtain
1 2
h (8.5.62)
t+422n = o g h .
x2 x=0
To eliminate the explicit time dependence on the left-hand side, we require that the exponent of
time, t, on the left-hand side is zero. Combining this condition with (8.5.36), we obtain
2 3 + 4n 1
= n, = , = . (8.5.63)
3 12 2
The interfacial condition (8.5.62) then becomes
h2 1 2
=2 , (8.5.64)
o g x2 x=0
coupling the oil with the water ow.
8.5 Axisymmetric and three-dimensional ows 695
1
u = t , (8.5.65)
x
and this motivates setting
uo = t G() = t2 G(x) = G(), (8.5.66)
t
where G() is an unknown function. Combining this expression with (8.5.61), we nd that
Substituting this expression along with (8.5.61) into the balance equation (8.5.56), we derive the
ordinary dierential equation
1 d 2
h(G ) + nh = 0. (8.5.68)
d
When n = 0, this equation is satised by the constant function G() = . Assuming with good
reason that h is nearly constant except at the leading edge, we derive the more general albeit
approximate result G(x) n/2.
dh2 u
2 1
=2 = V (2 + ). (8.5.69)
d o g x x=0 o g 6
To circumvent the integrable singularity at the leading edge located at = , we recast this
equation into the form
dh2 4 1
= V (2 + ), (8.5.70)
d o g 6
where = ( )1/2 . Integrating from the leading edge where = 0 up to the centerline, we
obtain the lm thickness.
d( ) 1 1
= (2 ) 2 + (2 + ) 2 . (8.5.71)
d 6
696 Introduction to Theoretical and Computational Fluid Dynamics
(a) (b)
4.5 0.4
4 0.35
3.5
0.3
3
0.25
*
2.5
, ,
0.2
h
2
0.15
1.5
0.1
1
0.5 0.05
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Figure 8.5.3 (a) Boundary-layer thickness (solid line), momentum thickness (dashed line), and
displacement thickness (dotted line), over an axisymmetric oil slick expanding due to gravity,
reduced by and plotted against the dimensionless similarity distance / . (b) Corresponding
distribution of the lm thickness for a xed amount of oil, n = 0.
d( )2 2
= (2 ) 2 ( ) + H() (2 + ) ( ), (8.5.72)
d 6
where the function H is given in (8.4.71). Using relations (8.4.48) and the expression for the
boundary-layer thickness in terms of given in (8.5.50), we obtain equation (8.4.72), which can
be used to compute the value of given the left-hand side, by numerically solving an algebraic
equation. Equation (8.5.72) accompanied by (8.4.72) can be integrated from the leading edge up to
the centerline of the oil slick using elementary numerical methods.
The unknown value corresponding to the radius of the leading edge is determined from the
oil slick volume, Vo = qo tn ,
g 2 q 4
1/12
o
= , (8.5.73)
where is a numerical coecient. In practice, we may guess a value for , solve the boundary-layer
equations, compute the volume of oil, and then rescale to meet a specied volume. When n = 0,
the numerical computations show that = 1.14. The distribution of the boundary-layer thickness
computed using the numerical method discussed in this section is shown in Figure 8.5.3. The results
are in excellent agreement with those obtained by dierent methods [75].
of state, (h) [297]. The two-dimensional version of this problem was discussed in Section 8.3.5.
Balancing the induced Marangoni traction with the shear stress, we obtain
u
= . (8.5.74)
x x=0
A mass balance over the oil lm provides us with the evolution equation
h 1 (us h)
+ = 0, (8.5.75)
t
where us = u (, x = 0, t) is the radial surface velocity.
In the similarity solution, the lm thickness and the surface tension are assumed to depend
exclusively on the scaled radial coordinate , so that
h(, t) = h(), (, t) = (). (8.5.76)
The interfacial condition (8.5.74) requires that
1 2
d
t2 = . (8.5.77)
x2 x=0 d
To eliminate the explicit time dependence on the left-hand side, we require that 2 = 0.
Combining this condition with (8.5.36), we obtain
3 1
= 1, = , = . (8.5.78)
4 2
The evolution equation (8.5.75) for the lm thickness becomes
3 dh 1 d 0
+ h = 0, (8.5.79)
4 d d x
where x0 ( / x)x=0 . Numerical solutions of the boundary-layer equations can be obtained by
nite-dierence methods [297].
8.5.6 Three-dimensional ow
The computation of boundary layers in three-dimensional ow is complicated by the arbitrary orien-
tation of the tangential component of the velocity of the incident ow. The boundary-layer equations
arise from two tangential components of the equation of motion by discarding the second partial
derivatives of the tangential velocity components with respect to arc length in the tangential direc-
tions. The normal component of the equation of motion ensures that the pressure drop is negligible
across the boundary layer. Accordingly, the tangential pressure gradient can be set equal to that of
the incident irrotational ow.
where P = I n n is the tangential projection operator. The continuity equation retains its full
form for three-dimensional ow.
Problems
u +v = , (8.5.82)
x
where u = ux is the axial velocity and v = u is the radial velocity. The corresponding Stokes
8.6 Oscillatory boundary layers 699
Solve (8.5.84) subject to the integral constraint (8.5.81), to obtain the solution
4 2
f () = , (8.5.85)
4 + 2
where = 3M/(16) is a dimensionless constant. Show that the axial volumetric ow rate is
given by Q = 8x, independent of M .
Computer Problem
vibrations of rigid or deformable objects. For example, an oscillating boundary layer develops at the
bottom of the ocean due to the periodic ow induced by the propagation of free-surface or internal
gravity waves.
Role of vorticity
To illustrate the physical mechanism governing the structure of an oscillatory boundary layer, we
consider uniform pulsating ow with angular velocity past a stationary rigid body. Reversal of the
velocity of the outer ow over one period causes the sign of the boundary vorticity to alternate in
time. As a result, vorticity of positive and negative sign diuses across the boundary and enters the
ow, and the boundary layer consists of successive vortex layers that penetrate and tend to cancel
each other as they are convected by the incident ow. An example is the Stokes boundary layer
arising in oscillatory unidirectional streaming ow over an innite plate, as discussed in Section 5.4.1.
The Stokes boundary layer consists of traveling bands of vorticity with alternating sign penetrating
the ow by a distance that is comparable to the Stokes boundary-layer thickness, = (2/)1/2 ,
where is the angular frequency of the oscillations and is the kinematic viscosity of the uid.
Similar boundary layers occur in oscillatory ow through tubes, as discussed in Section 5.5.
u
t (n u ) n = (I n n) u
Ut cos(t) t, (8.6.1)
where n is the normal unit vector pointing into the ow, t is a tangent unit vector, and Ut is the
amplitude of the tangential velocity. Neglecting the curvature of the boundary, we approximate the
ow at a particular point on the boundary with the ow over a at plate due to an overpassing
oscillatory streaming ow with velocity Ut cos(t)t, as discussed in Section 5.4.2. Introducing a
local coordinate system with the y axis perpendicular to the boundary at a point and origin at that
point, and using (5.4.16), we nd that the velocity prole across the oscillatory boundary layer is
given by
u = Ut cos(t) cos(t y) ey t, (8.6.2)
Drag force
The force exerted on the boundary consists of the acceleration reaction associated with the outer
unsteady irrotational ow and a viscous drag due to the Stokes boundary layer. It might appear that
the local solution (8.6.2) alone is sucient for computing the viscous drag. However, this approach
leads us to the erroneous conclusion that the phase shift between the drag force and the velocity
at the edge of the boundary layer is equal to 3/4, independent of the boundary geometry ([24], p.
335). The viscous normal stresses due to the boundary curvature must be taken into consideration.
8.6 Oscillatory boundary layers 701
To perform this calculation, we work in a frame of reference where the body appears to
execute translational oscillations with velocity v = V cos(t) and the velocity decays far from the
body, where V = U. Noting that the average kinetic energy of the uid is constant over each
period of the oscillation, T = 2/, and integrating the energy balance over one period, we obtain
T
T
u f dS dt = dV dt, (8.6.3)
0 Body 0 F low
Substituting (8.6.2) into (3.4.15), we nd that the rate of dissipation inside a small volume
within the boundary layer is given by
u
2 3 y 2y/
= Ut2 cos2 (t
t
= )e , (8.6.4)
y 4
where ut is the tangential velocity. The average rate of viscous dissipation over one period of the
oscillation is thus given by
T
1
D dS dy dt = Ut2 dS. (8.6.5)
T 0 0 Body 2 Body
D = D cos(t ), (8.6.6)
where D is the amplitude and is the phase shift between the drag force and the velocity. The
average rate of working of the body due to its motion against the uid, expressed by the integral on
the left-hand side of (8.6.3), is
T
1 1
W V D cos(t) cos(t ) dt = cos V D. (8.6.7)
T 0 2
Our analysis in Sections 6.15 and 6.16 has indicated that, unless the boundary has sharp edges, in
which case the linearized equation of motion ceases to be valid, D is an analytic function of (i/)1/2
702 Introduction to Theoretical and Computational Fluid Dynamics
and = /4. To compute the damping force, we must evaluate the surface integral of the square of
the amplitude of the tangential velocity of the outer ow.
As an example, we use the solution for streaming irrotational ow past a sphere of radius a
to evaluate the integral in (8.6.8) and nd that
a2
1/2
UD= 6a|V|2 . (8.6.9)
In the case of streaming irrotational ow past a cylinder of radius a and length L, we nd that
a2
1/2
UD= 4L|V|2 . (8.6.10)
Problem
Computer Problem
Flows in industrial, laboratory, and natural settings are subjected to disturbances with small
or large amplitude due to equipment and building vibration, Brownian motion of microscopic sus-
pended particles, and other reasons. In some technological applications, perturbations with a suitable
amplitude and form are purposely introduced in order to initiate a certain desirable action, such
as enhance mixing or delay boundary-layer separation. It is then possible that natural or articial
disturbances may amplify in time or space leading to unsteady motion or to a new steady state.
The behavior of a disturbance depends on its specic form as well as on the structure of the
unperturbed ow. In the context of hydrodynamic stability, the unperturbed ow is a base ow.
Perturbations may exhibit dierent types of behavior depending on the Reynolds number and other
dimensionless numbers characterizing the base ow. In some cases, perturbations may grow while
being convected with the base ow, causing a convective instability. In other cases, perturbations
may spread out to contaminate the whole domain of ow, causing an absolute instability. More
involved types of behavior are possible [195]. For example, disturbances in a globally unstable ow
may initiate self-excited modes. Establishing criteria for the resilience and thus physical relevance
of a particular steady or unsteady base ow is the main directive of hydrodynamic stability analysis.
To assess the stability of a base ow, we may subject it to a broad range of perturbations
and observe their subsequent evolution. If all perturbations decay, the ow is stable. If certain
perturbations amplify, the ow is unstable and may not be realized in a physical setting unless an
703
704 Introduction to Theoretical and Computational Fluid Dynamics
external mechanism that controls the growth of the unstable perturbations is provided. In certain
cases, assessing the behavior of perturbations can be done by simple physical arguments. More
generally, the study of perturbations requires detailed analysis and numerical computation.
In principle, the behavior of perturbations can be studied by solving the equation of motion
together with the continuity equation subject to imposed boundary conditions and a suitable initial
condition. However, since the number of admissible disturbances is innumerable, it is futile to
attempt to exhaust all possibilities. One way to make progress is to assume that the magnitude of
the perturbation is small and remain small during an initial period of time, linearize the governing
equations with respect to all ow variables around the base state, and solve the governing equations
for a wide range of initial conditions using, for example, the method of Laplace transform. This
approach encapsulates the essence of linear stability analysis. However, even after linearization,
a general solution can be obtained by analytical methods only for a limited family of ows by
performing a normal-mode analysis that considers perturbations with exponential growth or decay
in time.
If linear stability analysis indicates that perturbations grow in time, the base ow hosting the
perturbations is unstable. However, since neglected nonlinear eects may promote the growth or
perturbations, the converse is true only if the magnitude of perturbations is and remains suciently
small in time. In certain cases, nonlinear eects slow down or even suppress the growth of unstable
perturbations and lead to a new steady or periodic state. Assessing the eect of nonlinearities can be
challenging. Progress can be made working under the auspices of weakly nonlinear stability theory
where a disturbance is described by a perturbation series and the analysis is carried out up to the
second or higher order with respect to a perturbation parameter. A full assessment of the nonlinear
eects requires the use of numerical methods, such as the nite-dierence methods discussed in
Chapter 13.
In this chapter, we introduce the basic concepts underlying the formulation of the linear
stability problem for internal, external, interfacial, and free-surface ows. In the case of interfacial
ows, we employ the method of domain perturbation where a uid is articially extended into another
uid or a boundary by analytical continuation. Having laid the theoretical foundation and derived
the governing equations, we present the normal-mode stability analysis and study the properties
of a fundamental class of viscous and inviscid ows. Solutions of the linearized equations will be
obtained by analytical and numerical methods for solving ordinary and linear partial dierential
equations involving an eigenvalue.
Further discussion of linear and nonlinear stability analysis can be found in classical texts
by Lin [241], Chandrasekhar [73], and Betchov & Criminale [34], and in other relevant reviews and
monographs [95, 112, 195, 257, 364].
u
= F u (u, p), (9.1.1)
t
where
1
F u (u, p) u u p + 2 u + g (9.1.2)
is a forcing function, is the uid density, and is the kinematic viscosity.
1
= F u = (u u) 2 p + 2 . (9.1.3)
t
Requiring that the left-hand side and the last term on the right-hand side are zero, we derive a
Poisson equation for the pressure,
2 p = (u u) = : (uu), (9.1.4)
subject to boundary conditions discussed in Section 13.3.3. Conversely, if the pressure eld develops
so that (9.1.4) is satised at all times, the gradient of the pressure on the right-hand side of (9.1.2)
will be such that F u = 0 at any time (Problem 9.1.1).
Now we take the partial derivative of (9.1.4) with respect to time, expand the derivatives
on the right-hand side, and use (9.1.1) to derive a Poisson equation for the rate of change of the
pressure,
p
2 = (F u u) (u F u ) = 2 : (uF u ). (9.1.5)
t
The solution can be found using the Poisson inversion formula, yielding a standard evolution equation
written in the symbolic form
p
= Fp u, F u (u, p) . (9.1.6)
t
The functional form of Fp depends on the instantaneous structure of the ow and required boundary
conditions.
706 Introduction to Theoretical and Computational Fluid Dynamics
Solenoidal projection
To illustrate further the role of the pressure in maintaining the velocity eld solenoidal, we consider
the right-hand side of (9.1.2) with the pressure term omitted, and dene the pressureless forcing
function
F u (u, 0) u u + 2 u + g. (9.1.7)
The Helmholtz decomposition theorem discussed in Section 2.8 allows us to express the function
F u (u, p = 0) as the sum of a solenoidal and an irrotational eld,
F u (u, 0) = A + , (9.1.8)
p = , (9.1.10)
which shows that the pressure gradient projects F u (u, 0) into the space of solenoidal functions,
thereby transforming it into a solenoidal eld, F u (u, p). In Section 13.5, we will see that this
interpretation is the foundation of a class of numerical methods for integrating the equation of
motion in time, called projection or pressure-correction methods.
F (, u) ( u) + 2 = u + u + 2 . (9.1.12)
One noteworthy feature of (9.1.12) is the absence of the pressure. This feature is exploited for the
expedient analytical or numerical computation of ows based on the vorticity transport equation,
as discussed in Chapters 11 and 13.
The left-hand side of each equation in (9.1.13) is zero in a steady ow, and the structure of the
velocity, pressure, and vorticity elds is governed by the equations of steady ow,
F u (uS , pS ) = 0, Fp (pS , uS ) = 0, F ( S , uS ) = 0, (9.1.14)
where the superscript S indicates a steady state.
Substituting expressions (9.1.15) into the right-hand sides of (9.1.2) and (9.1.12) and discard-
ing terms with quadratic dependence on , we obtain the linear forms
1
F u (u, p) uS uU uU uS pU + 2 uU (9.1.17)
and
F (, u) uS U uU S + U uS + S uU + 2 U . (9.1.18)
To derive these equations, we have used (9.1.14) to eliminate the terms involving the steady eld
alone on the right-hand sides.
Next, we substitute expressions (9.1.15), (9.1.17), and (9.1.18) into the evolution equations
(9.1.1) and (9.1.11), and thus derive a linear evolution equation for the unsteady components of the
velocity and vorticity,
uU 1
= uS uU uU uS pU + 2 uU (9.1.19)
t
and
U
= uS U uU S + U uS + S uU + 2 U . (9.1.20)
t
These two equations can be collected into the compact form
U
! U " ! " u
u A B 0
= pU , (9.1.21)
t U C 0 D
U
708 Introduction to Theoretical and Computational Fluid Dynamics
We observe that the precise form of A, B, C, and D depends on the structure of the base ow.
Substituting the right-hand side of (9.1.15) for the velocity and pressure into (9.1.4), and lin-
earizing the right-hand side of the resulting equation, we obtain a Poisson equation for the unsteady
component of the pressure,
We can work in a similar fashion with the evolution equation (9.1.6), recasting it into a form that is
similar to that shown in (9.1.21). However, this is not necessary for the purposes of our discourse.
Now we consider a base ow at steady state and identify the unsteady component with a
physical perturbation. Depending on the structure of the base ow and form of the perturbation,
a disturbance may behave in dierent ways. If the magnitude of the disturbance, dened in some
local or global sense, grows, remains constant, or decays asymptotically in time, the disturbance is
unstable, marginally stable, or stable. If every possible disturbance decays in time, the base ow is
linearly stable. If certain disturbances grow, the base ow is linearly unstable. An unstable ow can
be physically realized only if unstable disturbances are screened out by a control mechanism that
detects and suppresses the growth of perturbations.
To this end, we underline the limitations and advisory role of linear stability theory by pointing
out that a ow that is stable according to linear stability theory will not necessarily appear in
practice. The reason is that nonlinear interactions and small deviations from an assumed perfect
geometry of the domain of ow can be responsible for unstable behavior. For example, Poiseuille
ow in a tube is always stable according to linear theory but unstable in real life when the Reynolds
number is suciently large, as discussed in Section 9.8.6.
(a) (b)
l l
n n
x un un
X X
ut q ut
Figure 9.1.1 (a) Illustration of a material line or interface deviating from a reference shape correspond-
ing to a two-dimensional base ow. (b) A perturbed material line or interface can be described by
a chain of marker particles parametrized by the arc length along the unperturbed interface, l.
where X(l) is a steady streamline or interface regarded as a reference line, l is the arc length along
the reference line, n(l) is the unit vector normal to the reference line, and q(l, t) is the normal
displacement, as shown in Figure 9.1.1(a). Correspondingly, the velocity component normal to the
reference line evaluated at the position of the perturbed stream surface or interface can be expressed
in the form
where vn is the leading-order normal velocity component. Substituting the expressions = q and
un = vn into the general kinematic compatibility condition (1.10.17), and linearizing with respect
to , we obtain the evolution equation
q q
+ ut (l) vn = 0, (9.1.26)
t l
where ut is the velocity component of the base ow tangential to the reference line. Given the
tangential and normal velocities, the shape of an evolving interface can be computed by integrating
in time the convection equation (9.1.26) by analytical or numerical methods. Equation (9.1.26)
also describes the evolution of the trace of an axisymmetric interface in an azimuthal plane in
axisymmetric ow. Similar equations can be written for three-dimensional ow.
Numerical methods
In a numerical implementation, a two-dimensional or axisymmetric interface is typically described
by a chain of N marked points or nodes, as shown in Figure 9.1.1(b). The position of the ith marker
point corresponding to arc length li is
where qi (t) is the nodal value of the shape function q(t). Since tangential motion does not cause
a change in shape, we have stipulated that the marker points move normal to the unperturbed
stationary interface.
710 Introduction to Theoretical and Computational Fluid Dynamics
The shape function, q(l), and its derivative, q (l) dq/dl, can be constructed in terms of the
nodal values, qi , by cubic spline or another type interpolation, as discussed in Section B.4, Appendix
B. The numerical method provides us with the nodal derivatives,
q (l )
q (li ) =
i
qj , (9.1.28)
qj q=0
where summation is implied over the repeated index, j. Note the linear dependence of q (li ) on the
nodal displacements, qj . Linearizing the normal velocity at the ith node with respect to the position
of all nodes, and working in a similar fashion, we obtain an expression for the nodal velocities
v (l )
n i
vn (li ) = qj . (9.1.29)
qj q=0
Applying equation (9.1.26) at the ith node and substituting expressions (9.1.28) and (9.1.29), we
derive a system of linear ordinary dierential equations,
dqi
= Mij qj , (9.1.30)
dt
where summation is implied over the repeated index, j, and
v (l )
q (l )
n i i
Mij ut (li ) (9.1.31)
qj q=0 qj q=0
is a constant matrix. The general solution of the linear system (9.1.30) can be found in terms of the
eigenvalues, eigenvectors, and possibly generalized eigenvectors of the matrix M (e.g., [317]). If the
matrix M is diagonalizable, we obtain exponential solutions described as regular normal modes. As
an alternative, equation (9.1.30) can be integrated in time by numerical methods.
Discrete perturbations
In practice, the jth column of the matrix M introduced in (9.1.30) can be generated by a simple
numerical procedure, described as the method of discrete perturbations, according to the following
steps:
1. All marker points are positioned along the unperturbed interface at chosen arc length positions
li , where i = 1, . . . , N .
2. The jth marker point is displaced normal to the interface by a small distance, .
3. Cubic spline or another interpolation with appropriate boundary conditions is performed for
a function G(l) using the data set
4. The velocity component normal to the unperturbed contour is evaluated at all nodes. For
small , we compute
v (l )
1
n i
vn (li ). (9.1.34)
qj q=0
The value of must be such that the dierentiation error is larger than the numerical error in
evaluating the velocity.
5. The results of Steps 3 and 4 are used to compute Mij according to (9.1.31)
The computation is repeated for j = 1, . . . , N to generate all columns of M and the eigenvalues are
computed by standard numerical methods. This formulation is especially advantageous when the
normal velocity can be computed from a contour- or boundary-integral representation, as in the case
of Stokes ow, potential ow, or inviscid ow containing vortex patches. The main advantage is that
solving the equations of inviscid or viscous ow is entirely bypassed and the analysis is conducted
using an essentially geometrical approach [43]. Applications are discussed in Section 9.3.5.
g 3
ht + h (hx + hxxx ) = 0, (9.1.35)
3 g x
where is the lm density, is the lm viscosity, is the surface tension, g is the acceleration of
gravity, a subscript t denotes a derivative with respect to time, and a subscript x denotes a derivative
with respect to x. To derive a linearized evolution equation for the lm thickness, we set
h(x, t) = h + (x, t), (9.1.36)
where h is the mean lm thickness and (x, t) is a perturbation shape function. Substituting this
expression into (9.1.35) and linearizing with respect to , we obtain a fourth-order linear partial
dierential equation with respect to ,
g 3
t + h x + xxx = 0. (9.1.37)
3 g x
An initial condition and a boundary or periodicity condition for must be provided. The solution
can be computed by nite-dierence, nite-element, spectral, or other numerical methods.
Problems
zU zU
+ y = 0, (9.2.1)
t x
where zU = 2 and the superscript U denotes the unsteady component. Inverting the hyperbolic
operator on the left-hand side, we obtain the general solution
Fourier expansions
For simplicity, we consider perturbations that are initially symmetric with respect to the origin of
the x axis, x = 0, and express the stream function as a sine Fourier series in the y direction and a
cosine Fourier integral in the x direction. Antisymmetric perturbations can be treated in a similar
manner and a general perturbation can be expressed in terms of symmetric and antisymmetric modes
9.2 Initial-value problems 713
where bn () are real Fourier coecients with dimensions of velocity multiplied by squared length.
Without loss of generality, we may assume that bn are even functions, bn () = bn (); odd functions
are annihilated upon integration. Straightforward dierentiation yields the initial vorticity eld,
n
n
2
2
z (x, y) (x, y, t = 0) = sin (y + a) + 2 bn () cos x d. (9.2.4)
n=1
2a 2a
Fourier transform
The Fourier transform of a rapidly decaying function, f (x), is dened by the integral
f (k) f (x) eikx dx, (9.2.5)
where i is the imaginary unit. The Fourier transform of the derivatives are f (k) = ik f(k) and
f (k) = k 2 f(k). The inverse transform is
1
f (x) = f(k) eikx d. (9.2.6)
2
where
n
n
2
F(k, y, t) = sin (y + a) + 2 bn () cos[(x ty)] eikx dx d.
n=1
2a 2a
(9.2.8)
1 ity ity
= e exp[i( )x] dx + e exp[i( + )x] dx , (9.2.9)
2
714 Introduction to Theoretical and Computational Fluid Dynamics
yielding
where is the one-dimensional Dirac delta function. Substituting this expression into (9.2.8), using
the distinctive properties of the delta function, and recalling that bn (k) = bn (k), we nd that
(9.2.7) becomes
n
n
2
k2 + 2
(k, y, t) = 2 + k2 sin (y + a) bn (k) eikty . (9.2.11)
y n=1
2a 2a
where
n
n
2
p (k, y, t) = i + k2 bn (k) (n An n An ), (9.2.14)
n=1
2a
where n , n , and n are functions of k and t. Substituting (9.2.14) into (9.2.12) and matching the
functional forms on the left- and right-hand sides, we nd that
1
n = 2 . (9.2.15)
n
2a kt + k2
To simplify the notation, we dene the negative-order coecients, bn (k) bn (k), and obtain
n
2
p (k, y, t) = i
+ k2 bn (k) n An , (9.2.16)
n=
2a
where the prime indicates that the term n = 0 is excluded from the sum. Substituting the expression
for n , adding a judiciously selected homogeneous solution, and taking the inverse Fourier transform,
we derive the general solution,
2
n
+ k2
i 2a
(x, y, t) =
2 bn (k) (9.2.17)
2 n= n 2
2a kt + k
Bn eikty n sinh[k(a y)] + n sinh[k(a + y)] eik(xty) dk,
9.2 Initial-value problems 715
Bn = exp (y + a) . (9.2.18)
2a
To satisfy the no-penetration boundary condition, we require that the term enclosed by the large
square brackets is zero when y = a, yielding
Selecting the real part on the right-hand side of (9.2.16), we obtain ([437], p. 482)
2
n
+ k2
1 2a
(x, y, t) =
2 bn (k) n (x, y, t, k) dk, (9.2.20)
2 n= n
2a kt + k2
where
n
Evaluating the right-hand side of (9.2.20) at long times shows that the disturbance decays like 1/t2 ,
and this means that the base ow is stable (Eliassen, Hilland & Riis [119]; see [257, 112] ). Later
in this chapter, we will see that viscous forces render certain types of perturbations unstable in a
certain range of Reynolds numbers.
where s is a complex variable. Taking the Laplace transform of (9.1.21) provides us with a system
of linear, inhomogeneous, second-order system partial dierential equations in the spatial variables
for the Laplace transformed functions,
716 Introduction to Theoretical and Computational Fluid Dynamics
U
! "U ! " u ! "U
u A B 0 u
s = p + (x, t = 0+ ), (9.2.23)
C 0 D
where 0+ indicates the limit t 0 from positive times when the disturbance is present. Having
computed the solution subject to appropriate boundary conditions, we recover the physical variables
in the time domain in terms of the Bromwich integral in the complex s plane,
U
u +i u U
p (x, t) = 1 p (x, s) est ds, (9.2.24)
2i i
where i is the imaginary unit and is a suciently large real positive number chosen so that
all singularities of the Laplace transformed functions lie on the left of the integration path. The
integral in (9.2.24) can be evaluated using the method of residues by introducing a closed contour
that encloses all singularities of the integrand. In the absence of branch points of the Laplace-
transformed variables, the contour can be identied with the union of the vertical line s = and a
semi-circular contour of large radius.
(x x0 ) (y y0 ) (z z0 ) (t) b, (9.2.25)
where x0 is an arbitrary point in the domain of ow, b is a constant vector, and is the one-
dimensional delta function. The solution for the velocity is G(x, x0 , t) b, where G(x, x0 , t) is the
Greens function tensor for the velocity. The solution of the initial-value problem can be expressed
as (a) a volume integral of the Greens function over the domain of ow multiplied by an appropriate
density distribution function, q, determined by the initial condition, and (b) a convolution integral
in time, in the form
t
u(x, t) = G(x, x0 , t ) q(x0 ) dV (x0 ) d. (9.2.26)
0 F low
Unfortunately, because Greens functions are hard to compute, this approach nds limited applica-
tions in practice [195].
Problem
(b) Derive the counterpart of (9.2.20) for antisymmetric perturbations where the stream function at
the initial instant is given by a sine Fourier integral with respect to x.
Computer Problem
9.2.2 Perturbation ow
Plot the streamline patterns according to (9.2.3) and (9.2.20) at the initial instant and at a later
time of your choice for bn (k) = 0 and b1 (k) = a3 eka . Discuss the behavior of the perturbation.
A convenient set of fundamental modes that are analogous to the eigenvectors of a diagonal-
izable matrix are normal modes with exponential dependence on time. The unsteady component of
the ow takes the form
N M
u V
p (x, t) = (x, ) eit , (9.3.1)
where the superscript N M stands for normal mode, is a complex constant called the complex
growth rate, complex cyclic frequency, or complex angular velocity, i is the imaginary unit satisfying
i2 = 1, and V, , and are complex functions of x and . All dependent variables are assumed
to be complex, with the understanding that both the real and imaginary parts represent admissible
solutions.
Substituting (9.3.1) into the linearized equation of motion and vorticity transport equation
stated in (9.1.21), we obtain a system of linear homogeneous equations governing the spatial structure
of the normal modes,
1
(i uS + 2 ) V V uS = 0, (9.3.2)
and
(i uS + 2 ) V S + uS + S V = 0. (9.3.3)
V = 0. (9.3.4)
By introducing the normal modes, we have factored out the time dependence and derived a system
of partial dierential equations in space involving the a priori unknown complex growth rate, .
718 Introduction to Theoretical and Computational Fluid Dynamics
For example, in the case of viscous unidirectional ow in a channel, only a discrete spectrum
consisting of an innite number of discrete eigenvalues that form a complete set can be found. In the
case of inviscid unidirectional channel ow, only a continuous spectrum consisting of stable normal
modes can be found. The nature of the two components of the spectrum will be illustrated in Section
9.5.2 for inviscid Couette ow.
Completeness
Whether or not the normal modes provide us with a complete base of eigenfunctions capable of
describing an arbitrary disturbance by linear superposition depends on the topology of the domain
of ow and presence of singularities [217]. If the normal modes provide us with a complete base,
an arbitrary disturbance can be expressed as a linear combination of (a) the discrete normal modes
multiplied by appropriate complex coecients, and (b) distributions of the continuous normal modes
weighed by appropriate complex distribution density functions. It is important to keep in mind that,
even though the individual normal modes may grow or decay at an exponential rate, their super-
position may exhibit a dierent type of temporal behavior (e.g., [68]). Generalized normal modes
with nonexponential dependence on time, corresponding to the eigenvectors of nondiagonalizable
matrices, may arise in some ows. These generalized modes behave as tm exp(it), with an a
priori unknown algebraic exponent, m.
which shows that I is the temporal growth rate of the normal mode and R is the cyclic frequency
or angular velocity of the perturbation. If I is positive, the disturbance grows exponentially in time
and the base ow is linearly unstable; if I is negative, the disturbance decays and the normal mode
is linearly stable; if I is zero, the disturbance is neutrally stable.
associated eigensolution is a prime objective of the linear stability analysis. To establish the critical
conditions under which a ow is expected to be unstable, we examine the behavior of neutrally
stable perturbations with I = 0, with respect to the dimensionless numbers that characterize the
base ow, such as the Reynolds number or the Weber number.
Laplace transform
It is instructive to apply the method of Laplace transform for computing the evolution of normal
modes. Combining the normal-mode form (9.3.1) with the denition of the Laplace transform shown
in (9.2.22), we nd that the transformed variables, indicated by a caret, are given by
N M
u V
p (x, s, ) = 1 (x, ). (9.3.6)
s + i
Substituting the right-hand side of (9.3.6) into (9.2.23) and simplifying, we recover (9.3.2) and
(9.3.3), and thereby reconcile the inhomogeneous problem expressed by (9.2.23) with the eigenvalue
problem expressed by (9.3.2) and (9.3.3).
Equation (9.3.6) shows that = is is a simple pole of the ow variables in the Laplace-
transformed domain. The associated normal modes may then be used to evaluate the Bromwich
integral in (9.2.24) using the method of residues. Conversely, the poles of the Laplace-transformed
variables correspond to normal modes that fall within the discrete part of the spectrum. The
continuous part of the spectrum is associated with branch cuts of branch points in the complex s
plane.
9.3.1 Interfacial ow
In the case of interfacial ow, we substitute into the linearized kinematic compatibility condition
(9.1.26) the normal-mode expansions
dQ
i Q(l) + ut (l) Vn (l) = 0, (9.3.8)
dl
where Q(l) is a shape function and Vn (l) is the corresponding normal velocity. Rearranging and
observing that Vn (l) is an implicit function of the normal displacement, Q(l), we obtain an eigenvalue
problem governed by an ordinary dierential equation,
dQ
Vn (l) ut (l) = i Q(l). (9.3.9)
dl
The second term on the left-hand side does not appear in the absence of base ow. The form of the
shape function Q(l) can be deduced by physical intuition subject to volume and possibly contact
angle constraints, or else computed as part of the solution.
720 Introduction to Theoretical and Computational Fluid Dynamics
xi = Xi + Qi eit ni (9.3.10)
for i = 1, . . . , N , where Qi is the nodal value of the shape function Q, and ni is the corresponding
unperturbed normal unit vector. Because tangential motion does not cause a change in shape, we
have stipulated that the marker points move normal to the unperturbed stationary interface.
The shape function, Q(l), and its derivative, Q (l), can be constructed in terms of the nodal
values, Qi , by cubic spline or some other type of interpolation, yielding an expression for the deriva-
tive at the ith node,
Q
Qi = i
Qj , (9.3.11)
Qj Q=0
where summation is implied over the repeated index, j (e.g., [317]). Similarly, the normal velocity
at the ith node can be linearized with respect to the position of all nodes, yielding the corresponding
expression
V (l )
n i
Vn (li ) = Qj . (9.3.12)
Qj Q=0
Applying equation (9.3.9) at the ith node and substituting these expressions, we obtain an algebraic
eigenvalue problem expressed by the linear system
Mij Qj = i Qi , (9.3.13)
where
V (l )
Q
n i
Mij ut (li ) i
. (9.3.14)
Qj Q=0 Qj Q=0
The matrix M can be generated by the discrete perturbation method discussed in Section 9.1.7.
The method has been applied successfully to study the capillary instability of an innite quiescent
viscous cylindrical thread containing a periodic array of particles arranged along the centerline, and
the gravitational instability of a pendant drop [43, 322]. Two additional applications are discussed
in this section.
(a) (b)
1.5
a n
3 1
N1 2
l 0.5
Q/a
N 1
x
0
0.5
1
0 0.2 0.4 0.6 0.8 1
/
Figure 9.3.1 (a) Illustration of a stationary spherical viscous liquid drop suspended in an innite
quiescent ambient uid. (b) Graphs of the rst ve eigenfunctions.
unperturbed tangential velocity is identically zero, ut = 0, and the second term on the right-hand
side of (9.3.14) does not appear. Referring to (9.3.9), we set l = a and identify the normal velocity
with the radial velocity, un = ur , where a is the drop radius and is the polar angle.
The discrete perturbation method described in this section is implemented in the computer
code drop ax placed in directory 08 stab of the software library Fdlib (Appendix C). The interfacial
velocity is computed using the boundary-integral method for Stokes ow discussed in Section 6.9.
Solving the eigenvalue problem provides us with the dimensionless growth rate I I a/, where
is the surface tension. The computations reveal a double zero eigenvalue followed by a sequence of
negative eigenvalues, I = 0.46, 0.76, 1.04, 1.31, 1.57, 1.83, . . ., corresponding to decaying
modes. The rst ve eigenfunctions, Q(), normalized so that Q() = 1, are shown in Figure
9.3.1(b). The rst two eigenfunctions, corresponding to the zero eigenvalues, express inadmissible
radial expansion, Q = 1, and inconsequential translation, Q = cos . Further eigenfunctions
are precisely proportional to the Legendre polynomials, Ln (cos ), for n > 1, ensuring that the
deformation conserves the drop volume to rst order with respect to the dimensionless deformation
parameter, .
In the second application, we consider the instability of a spherical viscous liquid drop with radius
a and viscosity , settling with velocity V in the direction of gravity along the x axis in an innite
ambient uid with viscosity under conditions of Stokes ow, as illustrated in Figure 9.3.2(a).
Using the singularity solution given in (6.7.47) and (6.7.48), we nd that, in a frame of reference
translating with the drop, the interfacial velocity over the unperturbed spherical interface is
1 1 xi xj 1 1 xi xj
ui (r = a) = (2 + 1) Vi + 2 Vj Vi = V i + 2 Vj , (9.3.15)
2 +1 a 2 +1 a
722 Introduction to Theoretical and Computational Fluid Dynamics
(a) (b)
5
imaginary
a
0
l
x
5
10 8 6 4 2 0 2
V real
Figure 9.3.2 (a) Illustration of a settling viscous liquid drop observed in a frame of reference trans-
lating with the drop velocity, V . (b) Spectrum of dimensionless complex growth rates ia/ for
Ca = 0 (circles), 0.5 (squares), 1.0 (diamonds), 2.0 (plus signs), 3.0 (asterisks), and 5.0 (dots).
where x = x x0 is the distance from the drop center, x0 , is the viscosity ratio, and Vx = V ,
Vy = 0, Vz = 0 are the Cartesian components of the drop velocity. We can easily verify that the
normal velocity component is zero, as required, u n = 0. In a frame of reference moving with the
drop, the tangential surface velocity is
1 sin
ut (l) = u (r = a) = V , (9.3.16)
2 +1
where l = a and is the meridional angle. The maximum surface velocity occurs at the equator,
= 12 . Using expression (6.7.51) for the velocity of settling, we formulate the capillary number
V 3 3 + 2 a2 (d )g
Ca = (9.3.17)
2 +1
determining the signicance of the capillary pressure due to surface tension, , relative to the viscous
stresses due to the motion, where d is the drop uid density.
Referring to (9.3.9), we identify the normal velocity with the radial velocity, un = ur . The dis-
crete perturbation method is implemented in the computer code drop ax discussed in the preceding
section on drop relaxation. The spectrum of dimensionless complex growth rates, ia/, is shown
in Figure 9.3.2(b) for several capillary numbers, Ca. Zero eigenvalues representing inconsequential
translation appear for any capillary number. The eigenvalues are real at small capillary numbers but
become complex as the capillary number increases beyond a critical threshold. Since the real part
of ia/ is zero or negative independent of the capillary number, the interface is stable except in
the complete absence of surface tension corresponding to innite capillary number.
9.3 Normal-mode analysis 723
(a) (b)
1 1
0.8
0.6
0.5
0.4
Q/a
Q/a
0.2
0
0
0.2
0.5 0.4
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
/ /
(c) (d)
1 1.2
0.8 1
0.6 0.8
0.4 0.6
Q/a
Q/a
0.2 0.4
0 0.2
0.2 0
0.4 0.2
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
/ /
Figure 9.3.3 The rst three eigenfunctions representing the interfacial deformation of a moving vis-
cous drop for capillary number (a) Ca = 0.5, (b) 1.0, (c) 2.0, and (d) 3.0.
The rst three eigenfunctions corresponding to nonzero eigenvalues are shown in Figure 9.3.3
for Ca = 0.5, 1.0, 2.0, and 3.0. The corresponding eigenfunctions for zero capillary number, cor-
responding to a stationary drop relaxing after a normal-mode perturbation has been imposed, are
shown in Figure 9.3.1(b). In the case of a moving drop, the interfacial velocity eld causes the
normal-mode perturbations to concentrate at the back of the drop where a sharp peak arises at high
capillary numbers. The peak becomes innitely tall in the absence of surface tension due to the
uninhibited action of the local stagnation-point ow.
normal-mode perturbations with wavelength L, we substitute into the linearized equation (9.1.37)
the normal-mode form
where k = 2/L is the wave number and h is the mean lm thickness. Simplifying and rearranging,
we derive a purely imaginary growth rate,
h3 2
I = k (g + k 2 ). (9.3.19)
3
Since I < 0 under any conditions, normal-mode perturbations decay exponentially due to the
combined action of gravity and surface tension.
Problems
Fourier decomposition
A normal-mode disturbance can be expressed as a double complex Fourier integral with respect to x
and z in the horizontal plane where the velocity of the unperturbed base ow is constant. Because
the governing equations (9.4.2) and (9.4.3) are linear, each Fourier mode evolves independently and
can be studied in isolation. Motivated by this observation, we consider perturbations of the form
V F
(x, ) = G (y, ) exp[i(kx x + kz z)], (9.4.4)
Q
where F, G, and Q are complex functions of y to be determined as part of the solution, and kx , kz
are complex wave numbers in the x and z directions comprising the two-dimensional wave number
vector
k = (kx , 0, kz ). (9.4.5)
Substituting (9.4.4) into the continuity equation (9.3.4) and carrying out the dierentiations, we
obtain the equation
kx Fx + kz Fz iFy = 0, (9.4.6)
where a prime denotes a derivative with respect to y. Recalling that the vorticity is dened as the
curl of the velocity, we nd that the functions F and Q are related by
Substituting (9.4.4) into (9.4.3) and using (9.4.6) and (9.4.7), we obtain the vector equation
k x Fz 0
[ i ( U kx ) (kx2 + kz2 ) ] Q + Q i U kz Fy + U Fy 0 = 0. (9.4.8)
k z Fz 1
Returning to the three-dimensional problem, we introduce the unit wave number vector, k,
and its reciprocal vector, l, given by
1 1
k = (kz , 0, kz ), l = (kz , 0, kz ), (9.4.10)
k k
dened such that k l = 0. where k |k|. Assuming that the wave number is real and projecting
equation (9.4.8) onto the reciprocal unit vector, l, we obtain
k
k 2 k
i Uk k J + J + U Fy = 0. (9.4.11)
kx kx kx
We have introduced the component of the vorticity vector in the direction of the reciprocal vector,
1
J Q l = i k Fy (kx Fx + kz Fz ). (9.4.12)
k
Now we make the substitutions
k k
= , kx = k, kz = 0, = , (9.4.13)
kx kx
and
1
Fx = (kx Fx + kz Fz ), Fy = Fy , Fz = 0, Qz = J , (9.4.14)
k
9.4 Unidirectional ows 727
and observe that equations (9.4.11) and (9.4.12) reduce to equation (9.4.9) and the third equation in
(9.4.7) written for the tilde variables and physical parameters. This means that the study of three-
dimensional perturbations can be reduced to the study of two-dimensional perturbations with a
suitable change of the wavelength of the perturbation and kinematic viscosity of the uid determining
the Reynolds number of the base ow.
We conclude that, to assess the stability of a unidirectional ow with uniform physical proper-
ties, it is sucient to consider two-dimensional disturbances. Once the tilded variables corresponding
to an equivalent two-dimensional problem are available, the non-tilded variables corresponding to
the three-dimensional physical problem are recovered using relations (9.4.13) and (9.4.14). It should
be emphasized that Squires theorem may not be valid when the boundaries of the ow deform in
response to a perturbation, and are applicable only for spatially periodic modes corresponding to
temporal stability.
Inviscid ow
The rst denition in (9.4.13) shows that the growth rate in the equivalent two-dimensional problem
is higher than that in the physical three-dimensional problem. When the uid is inviscid, the two
problems occur at the same innite Reynolds number, yielding Squires theorem for inviscid ow: for
every unstable three-dimensional perturbation, we can nd another two-dimensional perturbation
with dierent wavelength that is more unstable.
where q(x) is a complex function playing the role of a stream function and ez is the unit vector
along the z axis. To conform with the periodicity of the ow, we set
where f (y) is a complex function with dimensions of velocity multiplied by length. To simplify the
notation, we have set k = kx . Substituting expression (9.4.16) into (9.4.15) and then into (9.4.9),
728 Introduction to Theoretical and Computational Fluid Dynamics
w f + k2 (9.4.18)
k
f 2k 2 f + k4 f = i (U c)(f k2 f ) U f , (9.4.19)
derived independently by Orr [283] in 1907 and by Sommerfeld [378] in 1908, known as the Orr
Sommerfeld equation, where c /k is the complex phase velocity.
Dimensionless form
Nondimensionalizing all variables using as characteristic length L, velocity V , and time L/V , as
discussed in Section 3.10, we recast the OrrSommerfeld equation into the dimensionless form
f 2k 2 f + 4 f = i Re k (U c)(f 2 f ) U f . (9.4.20)
f U c x VL
f = , U = , c = , x = , k = kL, Re = . (9.4.21)
VL L V L
A prime in (9.4.20) indicates a derivative with respect to the dimensionless transverse position, y.
(a) (b)
1.2
0.02 1
Unstable
0.8
growth rate
ka
0.6
Stable
0.02
0.4
0.04
1.5
6 0.2
1
4
0.5 2 4
x 10
ka 0 0 0.5 1 1.5 2 2.5 3 3.5 4
Re Re 4
x 10
Figure 9.4.1 (a) Dependence of the highest growth rate scaled by V /a on the Reynolds number,
Re, and scaled wave number, ka, for plane HagenPoiseuille ow. (b) Stability phase diagram
separating stable from unstable normal modes.
Plane Poiseuille ow
As an example, we consider plane HagenPoiseuille ow along the x axis in a channel conned
between two parallel plates located at y = a (e.g., [369]). The velocity of the undisturbed ow is
described by the parabolic prole
y2
U (y) = U0 1 2 , (9.4.22)
a
where U0 is the centerline velocity. The Reynolds number is dened as Re = U0 a/ where is the
kinematic viscosity of the uid. A graph of the highest growth rate representing the most unstable
mode, scaled by V /a, is shown in Figure 9.4.1(a), and a stability phase diagram in the (Re, ka) plane
is shown Figure 9.4.1(b). The complex phase velocity was computed using a nite-dierence method
discussed in Section 9.7. The neutral stability curve where cI = 0 separates stable from unstable
modes in Figure 9.4.1(b). The ow is stable below a critical Reynolds, Rec 5772 for any wave
number, as discussed in Section 9.8.5.
Considering the general case where k and are both complex, we introduce dimensionless
variables and express the solution of the eigenvalue problem associated with the OrrSommerfeld
equation in the functional form
F(
, k, Re) = 0, (9.4.23)
where = L/V is a dimensionless complex growth rate and k = kL is the dimensionless complex
wave number. Assuming that F is an analytic function of its arguments, we follow a solution branch
and write
F F F
dF = d
+ dk + dRe = 0. (9.4.24)
k Re
Next, we denote the real wave number of the temporal stability problem for neutral stability at
a particular Reynolds number by k0 , and the corresponding dimensionless real cyclic frequency of
the spatial stability problem for neutral stability at the same Reynolds number by 0 , where, by
denition, F(k0 , 0 , Re) = 0. If the Reynolds number is changed by a dierential amount, dRe,
while k0 is held constant, 0 will change by d
so that
F F
(d
)k + dRe = 0. (9.4.25)
Re
If the Reynolds number is changed by a dierential amount, dRe, while 0 is held constant, k0 will
change by dk so that
F F
(dk) + dRe = 0. (9.4.26)
k Re
Combining equations (9.4.25) and (9.4.26) to eliminate dRe, we obtain
F/ k
(d
)k = (dk) = (dk) cG dk, (9.4.27)
F/ k Re
where cG = ( / k)Re is the complex group velocity of the spatially growing waves under conditions
of neutral stability [141]. Equation (9.4.27) can be used to extract information on the temporal
stability problem using results from the spatial stability problem, and vice versa.
throughout the domain of ow and the right-hand side of the vorticity transport equation (9.4.17)
is identically zero. To simplify the functional form of (9.4.17), we introduce the new variable
2/3 k
k 2 + i (y + U0 c) , (9.4.28)
k
which is a linear function of y. Substituting (9.4.28) into (9.4.17), we obtain the Stokes equation
d2 w
+ w = 0, (9.4.29)
d 2
whose solution can be found in terms of contour integrals in the complex plane (e.g., [123]).
To expedite the study of the linear stability problem, the functional form (9.4.30) is sometimes
assumed at the outset. Dierentiating the stream function with respect to y or x, we derive the two
normal-mode disturbance velocity components,
! "N M ! "
ux f (y)
= exp[i(kx t)], (9.4.31)
uy i kf (y)
To obtain the corresponding normal-mode pressure, we substitute these expressions into the lin-
earized NavierStokes equation (9.4.2) and integrate the resulting rst-order dierential equation.
Problem
(U c)(f k2 f ) U f = 0, (9.5.1)
732 Introduction to Theoretical and Computational Fluid Dynamics
where a prime denotes a derivative with respect to y. Rearranging, we derive the standard form of
a second-order linear ordinary dierential equation with variable coecients,
U
f k2 + f = 0, (9.5.2)
U c
where k, or c, or both are allowed to be complex. For future reference, we introduce the variable
q f /(U c), and recast (9.5.2) into the form
The rst term on the left-hand side involves a linear self-adjoint operator, L(q) [(U c)2 q ] .
Riccati equation
To expedite the solution, it is sometimes useful to introduce the variable F f /f and rearrange
(9.5.2) into the Riccati equation,
U
F = F 2 + k2 + . (9.5.4)
U c
The reduction in the order of Rayleighs equation from two to one is oset by the occurrence of a
nonlinearity due to the quadratic term on the right-hand side.
Critical layer
The second term inside the parentheses on the left-hand side of the Rayleigh equation (9.5.2) becomes
singular at the critical layer where the the unperturbed uid velocity is equal to the phase velocity
of the perturbation, U = c. For this to occur, c must be real, which means that the ow is neutrally
stable. The associated disturbances fall within the continuous part of the spectrum (e.g., [67]).
We will see that the occurrence of this singularity has important implications on the eciency of
numerical methods used to compute normal modes.
(U c)(f k2 f ) = 0, (9.5.5)
Continuous spectrum
A second family of solutions of (9.5.5) corresponding to the continuous part of the spectrum emerges
by allowing the expression inside the second set of parentheses on the left-hand side to be singular
when U = c. The singularity occurs at the position y = (U U0 )/, which varies between
the lower and the upper boundary of the ow. Since c is real, the corresponding normal modes are
neutrally stable.
The eigensolutions can be expressed as f (y) = aG(y), where a is a reference length. The
function G, with dimensions of length, satises the dierential equation
G k2 G + (y ) = 0, (9.5.7)
where is the one-dimensional delta function with dimensions of inverse length. The solution of
(9.5.7) is the Greens function of the one-dimensional Helmholtz equation forced at the position
where the phase velocity becomes equal to the uid velocity of the base linear ow. Referring to
(9.4.32), we nd that the disturbance represents the ow due to a at vortex sheet with sinusoidal
strength situated along the x axis at the position y = . An acceptable solution of (9.5.7) must be
continuous with discontinuous derivatives at y = , satisfy the homogeneous equation G k2 G = 0
everywhere except at y = , respect the specied kinematic conditions at the boundaries of the ow,
and conform with the jump condition
G (+ ) G ( ) = 1, (9.5.8)
where the plus and minus subscripts indicate evaluation immediately above and below the critical
level y = (e.g., [67]).
Continuous spectrum
The general solution of (9.5.7), corresponding to the continuous part of spectrum , is parametrized
by the elevation = U/ where c = U , ranging from a to a. We nd that
&
1 sinh[k(a + )] sinh[k(a y)] for < y < a,
G(k, y, ) = (9.5.9)
k sinh(2ka) sinh[k(a )] sinh[k(a + y)] for a < y < .
It is a straightforward exercise to verify that (9.5.9) satises the four conditions stated at the end
of Section 9.5.1 (Problem 9.5.2).
734 Introduction to Theoretical and Computational Fluid Dynamics
where 0 (k, y) is the one-dimensional Fourier transform of the initial stream function with respect
to x. Comparing (9.5.11) with (9.5.10), we write
a
0 (k, y) = q(k, ) G(k, y, ) d. (9.5.12)
a
Operating on (9.5.12) with 2 /y 2 k 2 , switching the order of the second derivative and the integral
sign on the right-hand side, and using (9.5.7) and the distinguishing properties of the delta function,
we nd that
2
2
q(k, y) = k 0 (k, y), (9.5.13)
y 2
which provides us with a relation between q and the Fourier transform, 0 . Remembering that
z = 2 and taking the Laplacian of (9.5.11), we nd that aq(k, y) is the Fourier transform of
the initial vorticity distribution with respect to x.
Fourier expansion
Further progress can be made by expanding 0 (k, y) in a sine Fourier series with respect to y, similar
to that shown in (9.4.2),
n
where bn are complex coecients. Substituting (9.5.14) into the right-hand side of (9.5.13), we
obtain a sine Fourier series with respect to y for q(k, y),
n
2 n
Finally, we substitute (9.5.15) into the integrand of (9.5.10) and carry out the integration with
respect to to derive the general solution in terms of bn ,
n
2
(x, y, t) = bn (k) + k 2 eikx n (k, y, t) dk, (9.5.16)
n=1
2a
9.5 The Rayleigh equation for inviscid uids 735
where
a n
Rearranging, we obtain
a =a
1
F(k, y, , ) G(k, y, ) ei(+) d = ei(y+)
+ G (k, y, ) e i(+)
.
a 2 + k 2 =a
(9.5.20)
1 1
i(y+) i ia ia
F(k, y, , ) = e e sinh[k(a + y)]e + sinh[k(a y)]e .
2 + k 2 sinh(2ka)
(9.5.21)
Stream function
We return to (9.5.17), express the sine term in terms of complex exponentials, and compute
1
n (k, y, t) = Hn (k, y, t) Hn (k, y, r) , (9.5.24)
2i
736 Introduction to Theoretical and Computational Fluid Dynamics
where
1 n
Hn (k, y) =
2 exp[i (y + a) kty ] (9.5.25)
2a
n
2a kt + k2
1
sinh[k(a + y)] exp[i(n kta)] + sinh[k(a y)] exp(ikta) .
sinh(2ka)
Substituting (9.5.24) into (9.5.16) and dening the negative coecients bn (k) bn (k), we obtain
n 2
1 2a + k2
(x, y, t) = n 2 bn (k) Jn (k, y, t) dk, (9.5.26)
2i n= kt + k 2
2a
where a prime indicates that the term n = 0 is excluded from the sum, and
n
Jn (k, y) = exp i (y + a) + k(x ty) (9.5.27)
2a
1
We have conrmed that the set of the normal modes falling within the continuous part of the
spectrum comprises a complete set.
2 U
|f | dy = k2 + (U c
) |f |2 dy. (9.5.29)
a a |U c|2
Equating the real and imaginary parts of the left-hand and right-hand sides, we nd that
b b
2 U
|f | dy = k2 + (U c R ) |f |2 dy (9.5.30)
a a |U c|2
9.5 The Rayleigh equation for inviscid uids 737
and
b
U
2
cI k2 + |f | dy = 0. (9.5.31)
a |U c|2
The last equation requires that either cI = 0, corresponding to neutral stability, or the integral on
the left-hand side is zero, requiring that U changes sign at least once between y = a and b. We
conclude that, for a normal mode of a unidirectional shear ow to be unstable, the velocity prole
must exhibit at least one inection point [335]. This is a necessary but not sucient condition, that
is, a normal-mode disturbance of a unidirectional shear ow whose velocity prole has an inection
point is not necessarily unstable.
An important consequence of Rayleighs criterion is that the normal modes of innite simple
shear ow, Couette or Poiseuille channel ow, are stable. Curious though it may seem, viscous forces
are required to render perturbations unstable. Physically, viscous stresses sustain a perturbation for
a longer period of time, thereby giving them a better chance to grow.
where U0 is an arbitrary constant [128]. Consider a ow whose velocity prole has a single inection
point, and identify U0 with the velocity at the inection point. The sign of the product U (U U0 )
is constant throughout the integration domain. For the integral on the left-hand side of (9.5.32)
to be negative, the sign of U (U U0 ) must also be negative, otherwise the disturbance will be
neutrally stable. Consequently, for a normal mode of a unidirectional shear ow to be unstable, the
maximum of the absolute value of the vorticity of the base ow must occur at an inection point.
Combining Rayleighs and Fjrtofts theorems, we nd that the normal modes of the ow
shown in Figure 9.5.1(a) are stable, whereas those of the ow shown in Figure 9.5.1(b) may be stable
or unstable.
(a) (b)
Figure 9.5.1 Applications of Rayleighs theorem on the signicance of an inection point and Fjrtofts
theorem. Normal modes in ow (a) are stable, but those in ow (b) may be unstable.
eigenvalue problems (Problem 9.5.4) [317]. A more general but less accurate method of locating
eigenvalues is provided by Howards semi-circle theorem stating that c must fall inside a half-disk in
the upper half complex plane [192]. The center of the disk lies at the point 12 (Umax + Umin ) on the
real axis, and the disk radius is equal to 12 (Umax Umin ), where Umax and Umin are the maximum
and minimum values of U . One consequence of Howards semi-circle theorem is that there is always
a point where the real part of c is equal to U and the disturbance travels with the local uid velocity.
The region around that point is the critical layer.
To prove the semi-circle theorem, we multiply (9.5.3) by q and integrate the resulting equation
from y = a to b, where an asterisk indicates the complex conjugate. Enforcing the no-penetration
condition q(a) = 0 and q(b) = 0, we obtain
b
(U c)2 dy = 0, (9.5.33)
a
2 2 2
where |q | + k |q | is a non-negative function. Decomposing the integral into its real and
imaginary parts, we obtain
b b
(U 2 2cR U + c2R c2I ) dy = 0, cI (U cR ) dy = 0. (9.5.34)
a a
The second equation is satised for neutrally stable perturbations, cI = 0. Leaving this case aside,
we set the second integral to zero and nd that the phase velocity cR must take values between
Umax and Umin , which demonstrates that a disturbance cannot move faster than the uid [335].
Next, we work in two stages. First, we use the second equation in (9.5.34) to simplify the rst
equation, obtaining
b
(U 2 c2R c2I ) dy = 0. (9.5.35)
a
Problems
Computer Problem
U0 y
2b x
U0
Figure 9.6.1 Illustration of a vortex layer with thickness 2b separating two uniform streams owing
along the x axis with uniform velocities U0 . The horizontal lines represent the upper and lower
vortex contours in the unperturbed conguration. The sinusoidal lines represent the upper and
lower vortex in the perturbed conguration.
Base ow
Consider a homogeneous unbounded shear ow consisting of an innite a vortex layer with thickness
2b and uniform vorticity , as illustrated in Figure 9.6.1, considered by Rayleigh ([337], Vol. II,
p. 393). The vortex layer separates two uniform semi-innite streams translating along the x axis
with velocities U0 , where = U0 /b. In the unperturbed state, the velocity prole is at above
and below the vortex layer and varies linearly across the vortex layer. This piecewise linear velocity
prole can be regarded as an approximation of a continuous prole described, for example, by the
hyperbolic tangent function, U (y) = U0 tanh(y/b).
Rayleigh equation
Since U = 0 everywhere in the ow except at the edges of the vortex layer, the normal modes
are governed by the simplied Rayleigh equation (9.5.5). Leaving aside neutrally stable modes with
c = U corresponding to the continuous part of the spectrum, we obtain the Helmholtz equation,
f k 2 f = 0, describing irrotational perturbations, where k is the real wave number. The general
solution is
fn (y) = an eky + bn eky , (9.6.1)
where an and bn are constant coecients, and n = 1, 0, 1 indicate, respectively, the region below,
9.6 Instability of a uniform vortex layer 741
inside, and above the vortex layer. To ensure that the disturbance decays far from the vortex layer,
we set
a1 = 0, b1 = 0. (9.6.2)
The disturbance stream function and velocity of the normal modes are given by expressions (9.4.30)
and (9.4.31). Since the disturbance ow is irrotational throughout the domain of ow, we may
introduce the velocity potential given in (9.5.6), if required.
where is a small dimensionless number, A1 are two complex constants, the subscript 1 denotes
the lower vortex contour, and the subscript 1 denotes the upper vortex contour.
f1 (1 ) = f0 (1 ), f0 (1 ) = f1 (1 ). (9.6.4)
a0 + b0 = a1 , a0 + b0 = b1 , (9.6.6)
where dU/dx = inside the vortex layer and dU/dx = 0 outside the vortex layer. Continuity of
velocity then requires that
Working in a similar manner with the lower vortex boundary, we obtain that
f1 (b) = f0 (b) A1 . (9.6.10)
V A1 a0 + b0 = b1 , V A1 a0 + b0 = a1 , (9.6.11)
Kinematic compatibility
Since the boundaries of the vortex layer are material lines convected by the ow,
D
1 (x, t) y = 0, (9.6.12)
Dt
where D/Dt is the material derivative. Expanding all terms in Taylor series with respect to and
retaining only the linear contributions, we nd that
1 1
+ U (y = b) uN
y
M
(y = b) = 0. (9.6.13)
t x
Substituting (9.6.3) and (9.4.31) into (9.6.13), we obtain
a0 b0 = b1 1+ , a0 b0 = a1 1+ . (9.6.15)
k(c U0 ) k(c + U0 )
9.6 Instability of a uniform vortex layer 743
Finally we substitute (9.6.6) into (9.6.15) and derive a homogeneous system of two linear equations,
! " ! "
(2kc + 2kU0 + ) a0
= 0. (9.6.16)
(2kc 2kU0 ) b0
Setting the determinant of the matrix on the left-hand side to zero provides us with a quadratic
equation for the complex phase velocity whose solution is
U0 1/2
c= (1 2kb)2 e4kb . (9.6.17)
2kb
The quantity under the radical is negative when kb < 0.639 , and positive otherwise. In the rst
case, c is purely imaginary and the plus sign in (9.6.17) yields an unstable mode with growth rate
I = kc and vanishing phase velocity. The minus sign produces a companion stable normal mode.
When kb > 0.639 , c is real and the normal modes translate upstream or downstream with phase
velocity c, while maintaining their initial amplitude. The lack of dissipation in an inviscid uid
prevents the decay of the kinetic energy of a perturbation.
Flow instability
We have found that a normal-mode disturbance can be unstable only when the scaled wave number,
kb, is less than approximately 0.639, which means that the ratio between the wave length and
the layer thickness, L/(2b), is larger than approximately 4.92. Normal-mode disturbances with
shorter wavelengths travel along the vortex layer with phase velocity that depends on the scaled
wave number, kb. The signicance of these results on the behavior of general spatially periodic
perturbations that do not necessarily represent normal modes is discussed in Section 9.6.3.
(a) (b)
1 1
0.9 0.9
0.8 0.8
0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
kb kb
Figure 9.6.2 (a) Graphs of the dimensionless growth rate, 4bI /U0 (solid line), and scaled imaginary
part of the complex phase velocity, cI /U0 (dashed line), in the unstable regime of wave numbers.
The dotted line in the stable regime of wave numbers represents the scaled phase velocity, cR /U0 .
(b) The solid line represents the ratio of amplitudes of traveling waves along the upper and lower
vortex contours, |A1 /A1 |, in the regime of stable wave numbers. The dashed line represents the
phase shift / between growing waves on the upper and lower vortex contours in the regime
of unstable wave numbers. The dotted line represents the phase shift between the disturbance in
circulation and the wave along the upper vortex contour, /, in the regime of unstable wave
numbers.
where c c/U0 is the scaled complex phase velocity computed from (9.6.17). The ratio |A1 /A1 |
is equal to unity in the regime of unstable wave numbers. The solid in Figure 9.6.2(b) displays the
amplitude ratio |A1 /A1 | in the regime of stable wave numbers, corresponding to the plus sign in
equation (9.6.17). At high scaled wave numbers, kb, the lower contour is only slightly deformed with
respect to the upper contour. If the minus sign were chosen instead in equation (9.6.17), the ratio
|A1 /A1 | would be the inverse of that plotted in Figure 9.6.2(b), which means that the upper contour
would be only slightly deformed with respect to the lower contour. These observations demonstrate
that the two normal modes act in complementary ways to prevent spatial bias in a unidirectional
ow.
The dashed line in Figure 9.6.2(b) represents the phase shift of the sinusoidal waves on the
upper and lower vortex contours for the unstable normal mode in the regime of unstable wave
numbers, = arg(A1 /A1 ). As kb tends to zero, vanishes, indicating that the boundary
waves tend to grow in phase. As kb increases toward the threshold of neutral stability, increases
monotonically toward the maximum value of , indicating that boundary waves tend to grow out of
phase. The phase shift of the stable normal mode is the negative of that of the unstable mode shown
with the dashed line in Figure 9.6.2(b). The proles shown in Figure 9.6.3 illustrate the structure
of unstable and stable normal modes.
9.6 Instability of a uniform vortex layer 745
(a)
2 2
y/b
y/b
0 0
2 2
10 5 0 5 10 10 5 0 5 10
x/b x/b
(b)
3 3
2 2
1 1
y/b
y/b
0 0
1 1
2 2
3 3
8 6 4 2 0 2 4 6 8 8 6 4 2 0 2 4 6 8
x/b x/b
Figure 9.6.3 (a) Structure of an unstable (left) or stable (right) normal mode for kb = 0.48438. (b)
Illustration of two stable modes for kb = 0.73438.
and conrm that the perturbation causes a disturbance in the strength of the vortex layer. The
associated phase shift with respect to the displacement of the upper vortex contour is
A
1 A1
= arg . (9.6.20)
A1
A1 A1 2kb(1 + ) + 1 2 + 2kb(1 )c
= . (9.6.21)
A1 2kb(1 + c)
The phase shift is zero in the regime of stable wave numbers where the waves on the upper and
lower vortex contours are out of phase, = .
A graph of / corresponding to the plus sign in equation (9.6.17) is shown with the dotted
line in Figure 9.6.2(b) in the regime of unstable wave numbers. As kb tends to zero, tends to
1
2 . Physically, rotational uid tends to accumulate midway between the crests and troughs of
growing waves.
746 Introduction to Theoretical and Computational Fluid Dynamics
Vortex sheet
A surface of discontinuity, regarded as a vortex sheet, arises in the limit as kb tends to zero while
the upper and lower stream velocity, U0 = b, is held constant. Expanding (9.6.17) in a Taylor
series with respect to kb yields the complex growth rate
kc i kU0 , (9.6.22)
which shows that a vortex sheet is unstable for all wave numbers. More important, the growth rate
becomes innite as the wave length of a disturbance becomes innitesimally small. This singular
behavior undermines the physical relevance of a vortex sheet and imposes essential diculties in
computing its self-induced motion, as discussed in Section 11.5.
U = [ (x + b) (x b)]. (9.6.23)
0.2
0.2
0.4
0.6
y/L
0.8
1.2
1.4
1.6
1.8
1 0.5 0 0.5 1
x/L
Figure 9.6.4 Linear amplication and nonlinear evolution of a periodic disturbance on a vortex layer
computed by the method of contour dynamics. Proles of the vortex contours are shown at equal
time intervals from top to bottom.
The solution is required to be a continuous function, f (y). Integrating (9.6.24) with respect to y
over a small distance centered at the upper or lower vortex contour, located at y = b, using the
properties of the delta function, and rearranging, we nd that
f (b) f (b)
f (b+ ) f (b ) = , f (b+ ) f (b ) = , (9.6.25)
U0 c U0 + c
where the superscript indicate evaluation on the upper (+) or lower () side. These expressions
are consistent with conditions (9.6.9) and (9.6.10), subject to expressions (9.6.14).
Problems
Computer Problems
U f (k 2 U + U )f = c (f k2 f ), (9.7.1)
where a prime denotes a derivative with respect to y. Note that the right-hand side is independent
of the velocity prole of the base ow.
In the rst step of the numerical implementation, we introduce a one-dimensional uniform grid
of nodes located at yi for i = 0, . . . , N + 1, where y0 = A and yN +1 = B, as shown in Figure 9.7.1
To satisfy the no-penetration boundary condition, we require that f0 = 0 and fN +1 = 0. Because
9.7 Numerical solution of the Rayleigh and OrrSommerfeld equations 749
N+1
N
1
0
Figure 9.7.1 A nite-dierence grid for solving the Rayleigh or OrrSommerfeld equation determining
the linear stability of unidirectional ow.
the stream function takes the same value at the two walls, the disturbance ow does not generate
a net ow rate. Next, we approximate the second derivative, f , at the ith node with a central
dierence, and thereby replace the dierential equation (9.7.1) with the nite-dierence equation
fi1 2fi + fi+1 2 fi1 2fi + fi+1
U (yi ) 2
k U (yi ) + U (yi ) fi = c 2
k 2 fi , (9.7.2)
y y
where fi = f (yi ) is the value of the eigenfunction f (y) at the ith node. Denoting
Ui U (yi ), Ui U (yi ), h y, (9.7.3)
A f = c B f, (9.7.5)
where
2U1 h2 (k 2 U1 + U1 ) U1 0 0
U2
2U2 h2 (k 2 U2 + U2 ) U2 0
A .. .. .. .. ..
. . . . .
0 0 UN 2 UN h2 (k2 UN + UN )
(9.7.6)
is an N N tridiagonal matrix,
2 (kh)2 1 0 0 ... 0
1 2 (kh)2 1 0 ... 0
B .. .. .. .. .. .. (9.7.7)
. . . . . .
0 0 0 1 2 (kh)2
750 Introduction to Theoretical and Computational Fluid Dynamics
is another N N tridiagonal matrix. The structures of the matrices A and B depend on the
nite-dierence method chosen to approximate the derivatives. The structure of the matrix B is
independent of the velocity prole, U (y).
D f = c f. (9.7.8)
The complex growth rate, c, is an eigenvalue of the new matrix D B1 A. The eigenvalue
with the largest magnitude can be computed by the power method, and other eigenvalues can be
obtained by successive deation, as discussed in Section B.2, Appendix B. A good initial guess for
c is provided by Gerschgorins theorem discussed in Section B.2, Appendix B. A practical concern
is that the computation of the inverse matrix, B1 , necessary to obtain D, can be prohibitively
expensive or introduce signicant numerical round-o error.
E(c) f = 0, (9.7.9)
and the eigenvalues, c, are identied with the roots of the characteristic polynomial, det[E(c)] = 0.
Rearranging (9.7.5), we nd that
2 h2 (k 2 +
U1
U1 c ) 1 0 0 0
2 h2 (k 2 +
U2
1 U2 c
) 1 0 0
E .. .. .
.. .. .. ..
. . . . . .
2 h2 (k2 +
UN
0 0 0 1 UN c )
(9.7.10)
Because the matrix E is tridiagonal, its determinant can be computed using an ecient algorithm
discussed in Section B.2.1, Appendix B.
9.7 Numerical solution of the Rayleigh and OrrSommerfeld equations 751
Having specied the wave number, k, we compute the eigenvalues in three repetitive steps:
guess a complex value c; compute det(E) using a method for tridiagonal matrices; correct the value
of c to ensure that both the real and imaginary parts of det(E) are zero. The correction in the third
step can be made using Newtons method discussed in Section B.3, Appendix B, setting
d det[E(c)]
Since det[E(c)] is an analytic function of c, we may select a real or complex number with small
magnitude, , and approximate the derivative d det[E(c)]/dc with a nite dierence. Using, for
example, a forward dierence, we nd that
d det[E(c)] det[E(c + )] det[E(c)]
. (9.7.12)
dc
High accuracy in the computation of this derivative is not required.
It is instructive to compare the graphs presented in Figure 9.7.1 with corresponding graphs
presented in Figure 9.6.2(a) for a vortex layer. The comparison reveals that the detailed structure
of the velocity prolepiecewise linear versus hyperbolic tangentaects only mildly the stability of
a shear ow.
0.9
7
0.8 7
0.7
0.6
0.5
0.4 2
0.3
2
0.2
0.1
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
kb
Figure 9.7.2 Stability graph for inviscid shear ow with hyperbolic tangent velocity prole, U (y) =
U0 tanh(y/b). Graphs of the scaled imaginary part of the complex phase velocity, cI /U0 (dashed
lines), and dimensionless growth rate, 4bI /U0 (solid lines), in the unstable regime of wave numbers
for several channel half-width a/b = 2.0, 3.0, 4.0, 5.0, 6.0, and 7.0.
The no-penetration and no-slip boundary conditions require that f = 0 and f = 0 over a stationary
plane wall conning the ow. Applying (9.7.14) at the ith interior node and approximating the
second and fourth derivatives with central dierences, we obtain the dierence equation
fi2 4fi1 + 6fi 4fi+1 + fi+2 fi1 2fi + fi+1
i + [ U (yi ) i 2k ]
k y 4 y 2
fi1 2fi + fi+1
k 2 U (yi ) + U (yi ) i k3 fi = c k 2
fi , (9.7.15)
y 2
where fi f (yi ). In the absence of viscous eects, we set = 0 and obtain the corresponding
discretized Rayleigh equation (9.7.2).
Compiling the dierence equations at the nodes provides us with a system of linear equations
that is similar to (9.7.5), except that A is a complex pentadiagonal matrix due to the presence of the
fourth derivative. The matrix B is the same as that in the case of inviscid ow. The nal algebraic
system can be recast into the form E(c) f = 0, where E is a pentadiagonal matrix. Unfortunately,
the determinant of E may no longer be computed by a specialized algorithm, and must be found using
a general-purpose method such as Gauss elimination or LU decomposition discussed in Section B.1,
Appendix B. Results obtained by solving the generalized eigenvalue problem using a Matlab function
will be presented in Section 9.8.1. The computer code is available in program orr in Directory 08 stab
of Fdlib (Appendix C).
9.7 Numerical solution of the Rayleigh and OrrSommerfeld equations 753
where ak (c) are unknown constants, c it the complex phase velocity, and N is a specied truncation
level. Substituting (9.7.16) into the Rayleigh or OrrSommerfeld equation, multiplying the resulting
equation by a chosen set of weight functions, wk (y), and integrating the product with respect to y
over the solution domain, we obtain a homogeneous system of algebraic equations for the coecients
ak (c). In Galerkins method, the weighting functions, wk , are identied with the expansion functions,
gk (e.g., [318]). Requiring that the nal system of equations has a nontrivial solution provides us
with a generalized eigenvalue problem for c.
The choice of basis functions plays an important role in the accuracy and eciency of the
numerical method. In the case of plane Poiseuille ow between two plane walls separated by the
distance 2a, an appropriate choice for symmetric disturbances is gk = (1 y 2 ) y 2(k1) , where y = y/a
is a scaled position and the origin of the y axis has been placed midway between the two walls [143].
Analysis shows that expansions in Chebyshev polynomials yield higher accuracy than expansions in
other seemingly more relevant sets of orthogonal functions [284]. In the case of innite shear ow,
the arguments of the basis functions should exhibit exponential decay [274].
and k ik/.
754 Introduction to Theoretical and Computational Fluid Dynamics
Although seemingly innocuous, the procedure may suer from numerical instability leading
to unreliable results when the system (9.7.17) is integrated in the vicinity of walls, especially at high
Reynolds numbers. To bypass this diculty, we may perform forward and backward integration
and combine the solutions to eliminate spurious oscillations. Alternatives are ltering out numerical
instabilities, orthonormalizing the solution during the numerical integration, and performing parallel
shooting (e.g., [143]). Similar diculties are encountered when integrating the Rayleigh equation
for inviscid ow. The numerical computations proceed smoothly when integrating from a region
where the velocity prole of the base ow is nearly uniform to the main core of the shear ow, but
numerical instability is encountered when the integration continues into the region of uniform ow.
Physically, the perturbation decays exponentially into the region of vanishing shear rate.
The practical diculties associated with integrating the OrrSommerfeld or Rayleighs equa-
tion can be avoided by working with a modied nonlinear set of equations constructed according to
Riccatis method, or by using the compound matrix method briey discussed in the remainder of
this section.
Approximations may cause numerical instability that degrades the accuracy of the computations in
the case of unbounded ow. A remedy is to map the innite domain of ow onto a nite strip using
the transformation
y
= tanh , (9.7.22)
b
where b is a characteristic length scale comparable to the eective thickness of the shear ow and z is
a new dimensionless variable [264]. As y varies from to +, varies from 1 to 1. Substituting
(9.7.22) into Riccatis equation (9.5.4), we obtain
dF b U
= F 2 + k2 + , (9.7.23)
d 1 2 U c
Hyperbolic-tangent shear ow
dF k2 F 2 2
=b 2
. (9.7.24)
d 1 b c/U0
756 Introduction to Theoretical and Computational Fluid Dynamics
To compute dF/d at = 1, we use the lHopital rule to evaluate the rst term on the right-hand
side, nding
dF
dF
2 1
= kb , (9.7.25)
d =1 d =1 b 1 c/U0
which can be rearranged into
dF
2
= . (9.7.26)
d =1 b(1 kb)(1 c/U0 )
Viscous ow
Riccatis method reduces the OrrSommerfeld equation into a quadratic system of four rst-order
dierential equations for the four entries of the 2 2 Riccati matrix R dened by the equation
u = R v. The two-dimensional vectors u and v contain, respectively, the second and fourth,
and rst and third entries of the vector q dened in (9.7.17). The implementation of this powerful
method is discussed by Davey [101].
Computer Problems
where y = y/b, b is a specied length, and the parameter takes values in the interval [0, 1]. The
limiting values = 1 and 0 yield, respectively, a shear layer with hyperbolic tangent velocity prole
and a symmetric wake with Gaussian velocity prole. Assuming that the ow occurs in a bounded
domain, a y a, prepare and discuss a graph of the maximum growth rate against the wave
number kb for = 0, 0.5, 1.0, and a/b = 2.0, 3.0, 4.0.
(b) Repeat (a) for the velocity prole
U (y) = U0 tanh y + ( 1) sech2 y . (9.7.28)
The limiting values = 1 and 0 yield, respectively, a shear layer with hyperbolic tangent velocity
prole and the Bickley jet [416].
1.2
.1
10
0.
1
0.05
0.1
0.8 0.1 0.0
5 0
0
0.05
05
kb 0.6
0.
0.05
0 0.1
1.
.10
0.4 0 5 0.1
0.0
5
.0
0
0.1
0
0.2
0.05
0.05 0.05
5 10 15
Re
Figure 9.8.1 Stability graph for a viscous shear layer with hyperbolic tangent velocity prole, U =
U0 tanh(y/b). Contour plot of the dimensionless growth rate, I /(bU0 ), against the Reynolds
number, Re = U0 b/, and wave number, kb.
The fundamental signicance and practical importance of unidirectional and nearly unidirectional
shear ows has motivated numerous studies of their stability by analytical and numerical methods.
Drazin & Reid review channel, free shear layer, boundary layer, and jet ows ([112], p. 211). Huerre
& Monkewitz summarize in tabular form the stability characteristics of several families of external
shear ows and provide a comprehensive list of references ([195], Table 3). In this section, we review
the stability characteristics of a few representative ows.
A free shear layer is the diuse interface between two parallel uid streams merging with dierent
velocities. Consider a symmetric free shear layer with hyperbolic tangent velocity prole, U =
U0 tanh(y/b), where U0 is the magnitude of the velocity far above and below the shear layer and b
is half the shear layer thickness. A contour plot of the dimensionless growth rate, I /(bU0 ), with
respect to the Reynolds number, Re = U0 b/, and wave number, kb. is shown in Figure 9.8.1. The
graphs were produced using the nite-dierence method for the OrrSommerfeld equation discussed
in Section 9.7.2. The neutral stability curve corresponds to vanishing dimensionless growth rate
(zero contour line). In Section 9.7.1, we mentioned that, in the limit of innite Reynolds number,
the ow is unstable for disturbances with wave number kb < 1. The results in Figure 9.8.1 show
that, in fact, the ow is unstable at any nite Reynolds number, except in the theoretical limit of
Stokes ow. The range of unstable wave numbers broadens and the maximum growth rate increases
as the Reynolds number becomes higher [35]. We conclude the viscosity has a stabilizing inuence
on free shear ows.
758 Introduction to Theoretical and Computational Fluid Dynamics
dp/dx>0
Stable flow
dp/dx=0
dp/dx<0
Re
Rec 519
Figure 9.8.2 Stability graph of boundary layers with self-similar velocity proles. Schematic illus-
tration of regions of stable and unstable ow plotted with respect to the Reynolds number,
Re = U /, and wave number, k , where is the displacement thickness.
The velocity prole of a free shear layer represents an exact solution of the equations of steady
inviscid ow. However, because the velocity prole does not satisfy the steady version of the Navier
Stokes equation with the viscous force present, it is not an acceptable representation of a steady
unidirectional viscous ow. Including viscous eects causes the shear layer to spread out and the
vorticity to diuse away from the central region, as discussed in Chapter 5. Neglecting the temporal
or spatial evolution of a nearly steady or nearly unidirectional base ow may introduce signicant
error [195].
Schematic contour plots of the growth rate as a function of the Reynolds number and scaled
wave number are shown in Figure 9.8.2. We observe a family of loops separating a regime of stable
ow on the left from a regime of unstable ow on the right. The unstable normal modes are
called TollmienSchlichting waves. For the Blasius boundary layer corresponding to dp/dx = 0, the
critical Reynolds number for instability is Rec = 519, where Re = U / and is the displacement
thickness. When instability rst occurs at Rec , the wave number of the marginally stable mode is
k = 0.3, corresponding to wavelength L = 18 . In the opposite extreme case of a boundary
9.8 Instability of viscous unidirectional ows 759
(a) (b)
0.18 1
0.16 0.9
0.8
0.14 Even mode
Odd mode
0.7
0.12
0.6
0.1
bI/U0
0
c /U
0.5
R
0.08
0.4
Even mode
0.06
0.3
Odd mode
0.04
0.2
0.02 0.1
0 0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
kb kb
Figure 9.8.3 Normal modes of the inviscid Bickley jet with velocity prole U (y) = U0 sech2 (y/b); (a)
growth rate and (b) phase velocity of even (symmetric) or odd (antisymmetric) modes.
layer in orthogonal stagnation-point ow, Rec 1.4 104 . The Blasius boundary layer and other
boundary layers with favorable pressure gradient are stable in the limit of inviscid ow but unstable
at nite and suciently high Reynolds numbers. This behavior demonstrates once again that viscous
stresses may have a destabilizing inuence by sustaining the perturbations.
9.8.3 Jets
The stability properties of two-dimensional jets are generally similar to those of free shear layers
discussed in Section 9.8.1. Physically, the edges of a jet constitute shear layers susceptible to the
generic instability of free shear ows. The interaction of the two edges leads to multiple normal
modes corresponding to symmetric (sinuous) or antisymmetric (varicose) perturbations. In the case
of a circular jet, we obtain axisymmetric or spiral deformations.
Bickley jet
The Bickley jet is a prototypical two-dimensional jet with a symmetric velocity prole described by
where y = y/b and b is half the jet thickness. In the theoretical limit of inviscid ow, the jet exhibits
the sinuous and varicose unstable normal modes described in Figure 9.8.3. The results presented in
this gure were produced using the nite-dierence method discussed in Section 9.7.1. The wave
number, phase velocity, and eigenfunctions of neutrally stable waves are kb = 2, c = 23 U0 , and
f (y) = bU0 sech2 y for the symmetric mode, and kb = 1, c = 23 U0 , and f (y) = bU0 sechy tanh y for
the antisymmetric mode ([112], p. 233). Because the growth rate of the symmetric mode is higher
than that of the antisymmetric mode, the amplication of an arbitrary disturbance leads to the
formation of a staggered array of vortices in the arrangement of the von Karman vortex street, as
760 Introduction to Theoretical and Computational Fluid Dynamics
discussed in Section 11.3. Solving the OrrSommerfeld equation, we nd that the ow is unstable
when the Reynolds number, Re = bU0 /, exceeds the critical threshold Rec = 4. The wave number
at neutral stability corresponding to Rec is kb = 0.20.
Normal-mode analysis reveals that the plane Couette ow is stable at any Reynolds number, Re
a2 /, and for any wave number; is the wall shear rate or slope of the velocity prole and a is half
the channel width. Unstable behavior observed in practice at Reynolds numbers as low as Re = 350
is attributed to nonlinear eects due to the nite amplitude of the disturbances, wall roughness,
and deviation from unidirectional motion due to entrance eects. Transient growth may introduce
nonlinearity that is responsible for a secondary instability (e.g., [364]).
Normal-mode analysis reveals that the plane Poiseuille ow becomes unstable when the
Reynolds number, Re Ucl a/, exceeds the critical value Rec = 5772, where a is half the chan-
nel width, and Ucl is the centerline velocity. At lower Reynolds numbers, the ow is stable. The
wave number of the normal mode that rst becomes unstable at the critical Reynolds number is
ka 1.020. In practice, the ow becomes unstable when the Reynolds number exceeds the approx-
imate threshold 1500.
Poiseuille ow in a circular tube is stable against axisymmetric as well as more general three-
dimensional perturbations. In practice, the ow becomes unstable when the Reynolds number,
Re Ucl a/, exceeds a threshold as low as 1100, where a is the tube radius and Ucl is the centerline
velocity. The unstable behavior is attributed to the reasons stated previously in this section for
Couette ow.
Problem
Computer Problems
Von Karman provided an appealing physical interpretation of Rayleighs criterion [414]. Con-
sider a uid ring that is displaced in a way that preserves axisymmetry from an initial radial position,
1 , to a new radial position, 2 , due to the instability. Kelvins circulation theorem requires that
the circulation around the ring, C, is preserved after the perturbation. The old and new azimuthal
components of the velocity of the ring, (u )1 and (u )2 , are related by C1 = 21 (u )1 = 22 (u )2 .
The centrifugal force per unit volume due to the rotation of the uid at the unperturbed and per-
turbed states are
(u2 )1 (u2 )2 C12
F1 = , F2 = = . (9.9.5)
1 2 4 2 23
762 Introduction to Theoretical and Computational Fluid Dynamics
(a) (b)
Figure 9.9.1 Instability of rotating ows: (a) Taylor instability of a uid between two rotating
concentric cylinders and (b) Gortler instability of a high-Reynolds-number ow along a curved wall.
If the pressure eld remains unchanged, the radial pressure gradient at the new position has the
undisturbed value
dp
C22
= . (9.9.6)
d 2 4 2 23
Rayleighs criterion (9.9.1) arms that the ow is stable as long as the pressure gradient is able to
overcome the centrifugal acceleration, and thus push the uid ring back to its original position.
the wall as the stationary outer cylinder of a circular Couette ow allows us to make an analogy
between the instability of this ow and that of the ow between two concentric cylinders.
Problem
We begin in this section by consider the instability of the interface between two inviscid
uids that merge with dierent velocities forming a unidirectional base ow. Since the tangential
component of the uid velocity is allowed to undergo a discontinuity across the interface between
two inviscid uids, the interface can be regarded as a at vortex sheet whose self-induced motion
causes the growth or decay of perturbations. Conversely, the disturbance ow can be attributed
to the instability of an interfacial vortex sheet. The second interpretation allows us to study the
nonlinear stages of the motion using the vortex methods discussed in Chapter 11. In Sections 9.11
9.13, we discuss the instability of the interface between viscous uids due to the uid inertia, gravity,
viscosity stratication, or interfacial tension.
g Fluid 1
U
1
x
U2 Fluid 2
Figure 9.10.1 Illustration of a periodic vortex sheet representing the interface between two inviscid
uids in uniform motion along the x axis.
Rayleigh equation
Since U = 0 above and below the interface, Rayleighs equation (9.5.2) simplies into the Helmholtz
equation, f k 2 f = 0, inside each uid. Because the vorticity of the base ow vanishes above and
below the interface, the disturbance ow is irrotational. The general solution of the Helmholtz
equation is
where n = 1, 2 for the upper or lower uid, and An , Bn are four constants. Demanding that the
disturbance decays far from the interface, we set
A1 = 0, B2 = 0. (9.10.3)
Using (9.5.6), (9.10.2), and (9.10.3), we derive expressions for the normal-mode potential,
N
1
M
= i B1 eky exp[ik(x ct)], N
2
M
= i A2 eky exp[ik(x ct)]. (9.10.4)
Kinematic compatibility
The velocity on either side of the interface must be such that the motion of point particles adjacent to
the interface is consistent with the evolving shape of the interface expressed by (9.10.1), as discussed
in Section 1.10. Introducing the kinematic constraint D(y)/Dt = 0 on either side of the interface,
expanding all terms in Taylor series with respect to , and retaining only the linear contributions,
we nd that
+ Un uN
y
M
(y = 0) = 0, (9.10.5)
t x
where D/Dt is the material derivative and the subscript n of Un indicates that the velocity is
computed on the upper (n = 1) or lower (n = 2) side of the interface. Setting uy = /y, and
substituting into (9.10.5) expressions (9.10.1) and (9.10.4), we derive two equations,
A2 B1
A= = . (9.10.6)
c U2 c U1
It is clear that the phase velocity cannot be the upper or lower stream velocity.
9.10 Stability of a planar interface in potential ow 765
To express the pressure in terms of the velocity, we use Bernoullis equation. Linearizing the
quadratic terms and substituting the result into (9.10.7), we obtain
N M N M
N M N M
2
2 2 1 1
2 + U2 + g + 1 + U1 + g = , (9.10.8)
t x t x x2
evaluated at y = 0. Substituting the expression for the potential given in (9.10.4), we obtain
2 [ k(U2 c) A2 + gA ] + 1 [ k(U1 c) B1 + gA ] = k 2 A. (9.10.9)
Growth rates
To compute the complex phase velocity, we solve equations (9.10.6) for A2 and B1 in terms of A,
substitute the result into (9.10.9), and eliminate the arbitrary constant A to derive a quadratic
equation for c whose solution is
1
c= ( 1 U1 + 2 U2 D ) (9.10.10)
1 + 2
where
g
D= (1 + 2 ) + k (1 + 2 ) 1 2 U 2 (9.10.11)
k
is the discriminant, U = U2 U1 , and = 2 1 . If D is positive, c is real and the disturbance
travels with constant amplitude and phase velocity cR = c.
If D is negative, c is complex, one of the two solutions corresponding to the sign has a
positive imaginary part, and some disturbances are unstable. The phase velocity and growth rate
of unstable waves are
1 U1 + 2 U2 |D|
cR = , I = k . (9.10.12)
1 + 2 1 + 2
The critical wave numbers for neutral stability, kc , are found by setting the discriminant D to zero,
yielding a quadratic equation,
1 2
kc2 U 2 kc + g = 0. (9.10.13)
1 + 2
The two real roots enclose a nite band of unstable wave numbers.
766 Introduction to Theoretical and Computational Fluid Dynamics
g 1 + 2 k2
U 2 < 1+ . (9.10.14)
k 1 2 g
The critical velocity dierence under which the ow is stable to all disturbances, |Uc |, corresponds
to the minimum of the right-hand side of (9.10.14), regarded as a function of k, which occurs when
k = ( g/ )1/2 , yielding
+
2
1 2
|Uc |4 = 4 g . (9.10.15)
1 2
When the upper uid is heavier than the lower uid, < 0, the ow is unstable even in the absence
of ow.
1 N
2
= U2 U1 = U2 U1 + k (B1 + A2 ) exp[ik(x ct)]. (9.10.16)
x x y=0
1 U 2
kc1 = 0, kc 2 = . (9.10.18)
2
Intermediate wave numbers yield unstable normal modes with growth rate
1 k
1/2
I = k |U | 1 2 , (9.10.19)
2 U 2
9.10 Stability of a planar interface in potential ow 767
and phase velocity equal to the mean velocity of the unperturbed streams, cR = 12 (U1 + U2 ). The
phase shift between the disturbance in the strength and shape of the vortex sheet is equal to 12
for unstable wave numbers and zero for neutrally stable wave numbers. The maximum growth rate
corresponding to the most dangerous normal mode occurs when k = 23 kc2 . Large wave numbers
with small wavelengths are stabilized by the restraining action of surface tension.
Nonlinear evolution
The initial linear growth and subsequent nonlinear development of the KelvinHelmholtz instability
are illustrated in Figure 9.10.2(a) in the absence of surface tension. The evolution of the interface
depicted in this gure was computed using the point-vortex method for vortex sheets discussed
in Section 11.5. We observe that the growth of sinusoidal waves leads to the formation of a pe-
riodic array of spiral structures. Numerical evidence strongly suggests that the curvature of the
interface becomes discontinuous at a point immediately before the spiral structures develop (e.g.,
[215]). Since a geometrical singularity arises spontaneously from a smooth initial condition, the
problem of vortex-sheet motion is classied as ill-posed. Comparing the instability of the vortex
sheet illustrated in Figure 9.10.2(a) with the instability of a vortex layer illustrated in Figure 9.6.5
demonstrates the strong eect of the velocity discontinuity due to mathematical idealization involved
in the construction of a vortex sheet.
In the absence of surface tension, all wave numbers are unstable and the growth rate is given
by the simplied expression
I = Akg, (9.10.22)
768 Introduction to Theoretical and Computational Fluid Dynamics
(a) (b)
0
1
0.2
2
0
0.2
3
y/L
0.4 4
y/L
0.6
5
0.8
6
1
7
1.2
8
1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1 1 0 1 2
x/L x/L
Figure 9.10.2 Initial linear growth and subsequent nonlinear evolution of (a) the KelvinHelmholtz
instability and (b) the RayleighTaylor instability of a vortex sheet in the absence of surface tension.
Proles of the interface are shown at equal time intervals from top to bottom.
where
2 1
A= (9.10.23)
2 + 1
is the Atwood ratio. Because, the higher the wave number, the higher the growth rate of the
perturbation, a singularity may occur after a nite evolution time due to the amplication of small-
scale irregularities.
Nonlinear evolution
The initial growth and long-time evolution of the RayleighTaylor instability is illustrated in Figure
9.10.2(b) for A = 0.5 in the absence of surface tension. In these simulations, the evolution of the
interface was computed using a variation of the point-vortex method discussed in Section 11.5,
incorporating regularization. The development of a convoluted pattern with a secondary Kelvin
Helmholtz instability appearing along the sides of plunging sections of the interface is a striking
feature of the nonlinear motion.
Accelerating interfaces
Consider two adjacent uids accelerating with velocity V(t) normal to a at interface. The analysis
9.10 Stability of a planar interface in potential ow 769
of this section remains valid in a noninertial frame translating with the interface, provided that the
inertial acceleration force, dV/dt, is added to the gravitational force, g [399]. In this light, the
RayleighTaylor instability emerges as the instability of an accelerating interface between two uids
in the possible presence of a body force directed normal to the interface.
Wall eects
Container walls may have a strong stabilizing inuence on the interfacial instability by placing limits
on the maximum wave length (or minimum wave number) of normal modes that are allowed to enter
the physical system. As an example, we consider the instability of a horizontal interface inside a
vertical circular container of radius a that is partially lled with a liquid labeled 2 lying underneath
another liquid labeled 1. In the unperturbed state, the interface is at and the pressure distribution
in each uid is determined by hydrostatics.
Introducing cylindrical coordinates, (x, , ), with the x axis pointing upward against the
direction of gravity, we describe the position of the interface by the real or imaginary part of the
function x = (, , t), where
N
j
M
(, , t) = Bm,j ekx Jm (k) exp[i(m t)] (9.10.25)
for j = 1 (upper uid) or 2 (lower uid), where Bm,n are two complex constants. The plus or minus
sign in the rst exponential apply for the lower or upper uid, so that the perturbation decays far
from the interface (Problem 9.10.4).
m
= 0. (9.10.26)
d =a
When the cylinder radius is smaller than this critical value, the interface is stable in a conguration
where heavy uid lies above a light uid supported by interfacial tension.
770 Introduction to Theoretical and Computational Fluid Dynamics
where n is the unit vector normal to the free surface pointing into the ow. Using (9.10.4), we
evaluate the free-surface potential = Ac sin[k(x ct)]. Substituting this expression along with
the expression
u n = c nx c = c Ak sin[k(x ct)] (9.10.31)
x
into (9.10.30), we obtain
L
1 1
K c2 2 A2 sin2 [k(x ct)] dx = c2 2 A2 . (9.10.32)
2 0 2k
In the absence of surface tension, c2 = g/k, and the potential and kinetic energies are equal.
Standing waves
Superposing two waves with the same amplitude traveling in opposite directions, corresponding
to the two signs in (9.10.28), we obtain a standing gravitycapillary wave oscillating with angular
velocity kc. Whether a progressive or a standing wave arises in practice depends on the physical
mechanism that is responsible for initiating the disturbance and on the boundary conditions at the
point where the interface meets a container.
9.10 Stability of a planar interface in potential ow 771
In the unperturbed state, the interface is at and the two uids translate along the y axis
pointing upward against the direction of gravity with the same uniform seepage velocity, U . The
(0)
unperturbed velocity potentials, indicated by the superscript (0), are given by 1 = U y and
(0)
2 = U y. The interfacial velocity is V = U/n, where n < 1 is the porosity.
The horizontal interface is susceptible to the SamanTaylor instability [359]. To carry out the
linear stability analysis for two-dimensional perturbations in the xy plane, we describe the ow in a
frame of reference moving with the unperturbed interface. Working as in Section 9.10.1, we describe
the prole of the perturbed interface by the real or imaginary part of the function y = (x, t), where
is a small dimensionless coecient,
is the normal-mode wave form of the perturbation, A is the complex amplitude of the interfacial
wave, k = 2/L is the wave number, L is the wavelength, and c is the complex phase velocity. We
will describe the perturbations on either side of the interface separately, and then match the two
disturbance ows by requiring appropriate kinematic and dynamic conditions.
The normal-mode potentials are given in (9.10.4) in terms of the unknown constants B1 and
A2 . To compute the three unknowns, B1 , A2 , and c, we require three equations. Introducing the
kinematic constraint D(y )/Dt = 0 on either side of the interface, expanding all terms in Taylor
series with respect to , and retaining only the linear contributions, we nd that
n uN
y
M
(y = 0) = 0, (9.10.35)
t
where D/Dt is the material derivative. Making substitutions, we nd that B1 = A2 = ncA.
(0)
The driving potential in the ith uid on either side of the perturbed interface is i = i +
N
i
M
for i = 1, 2, evaluated at y = . Linearizing with respect to , we nd that
1 (y = ) U + N
1
M
(y = 0) = ( U + i nc ) ,
2 (y = ) U + N
2
M
(y = 0) = ( U i nc ) . (9.10.36)
772 Introduction to Theoretical and Computational Fluid Dynamics
From Darcys law, the pressure is related to the potential by pi = (i i + i gy) for i = 1, 2, where
2 2 /2 and 1 1 /1 . Substituting (9.10.36), we obtain an expression for the interfacial
pressure jump,
p2 p1 = U i (1 + 2 ) nc + g , (9.10.37)
k g + U + T k 2
I kc = . (9.10.38)
n 1 + 2
The ow is stable if I < 0, which requires that
g + T 2
g > (1 2 ) U T 2 or U< , (9.10.39)
1 2
and unstable otherwise. The second inequality shows that, when 1 = 2 and in the absence of
surface tension, the interface between a low-viscosity uid displacing a high-viscosity uid with the
same density is unstable.
Problems
2S 2S
2
+ = k 2 S, (9.10.40)
z x2
subject to periodicity or no-penetration conditions. Show that the growth rate satises (9.10.21).
9.11 Film leveling on a horizontal wall 773
y
Air
p
0
g Liquid h
x
Base state
In the unperturbed state, the lm is quiescent and the surfactant is distributed uniformly over the
774 Introduction to Theoretical and Computational Fluid Dynamics
where is the lm density, g is the acceleration of gravity, p0 is the ambient pressure, and the
superscript (0) denotes the unperturbed base state.
Normal-mode analysis
To carry out the normal-mode stability analysis for two-dimensional perturbations, we describe the
position of the interface by the real or imaginary part of the function
is the normal-mode wave form of the perturbation, A is the complex amplitude of the interfacial
wave, k = 2/L is the wave number, L is the wavelength, and c is the complex phase velocity.
Stream function
The stream function, , is dened by the equations u = /y and v = /x. In the linear
analysis, we set = (1) and introduce the normal-mode form
where y ky and the superscript (1) denotes the disturbance. The perturbation velocity components
(1) (1) (1) (1)
are ux = ux and uy = uy , where ux = (1) /y and uy = (1) /x.
The Reynolds number of the ow is so small that the motion of the uid is governed by
the equations of Stokes ow. Requiring that the stream function satises the biharmonic equation,
4 (1) = 0, we obtain
(y) = a1 ey + a2 y ey + a3 ey + a4 y ey , (9.11.5)
Wall conditions
The no-penetration and no-slip boundary conditions over the wall require that
(y = 0) = 0, (y = 0) = 0, (9.11.6)
a3 = a1 , 2a1 + a2 + a4 = 0. (9.11.7)
9.11 Film leveling on a horizontal wall 775
Kinematic compatibility
Kinematic compatibility requires that D[f (x, t) y]/Dt = 0, where D/Dt is the material derivative,
yielding
f f
+ ux uy = 0, (9.11.8)
t x
where the velocity is evaluated at the lm surface. Linearizing in the absence of a mean ow, we
obtain
(1)
+ =0 (9.11.9)
t x
evaluated at the undisturbed lm thickness, y = h. Substituting the normal-mode forms, we obtain
iAkc + ik(k) = 0, yielding
where 0 , 0 are uniform values corresponding to the at lm, and 1 , 1 are complex amplitudes.
Since the perturbations are small, we can write
1 1
= Ma , (9.11.12)
0 0
where Ma is the Marangoni number dened in Section 3.8.2.
Surfactant transport
The linearized form of the surfactant transport equation (1.9.13) is
(1)
(1) ux 2 (1) (9.11.13)
+ 0 = Ds ,
t x x2
where
Ds is the surface surfactant diusivity, and the velocity is evaluated at the unperturbed lm surface,
y = h. Expressing the velocity in terms of the stream function, we obtain
(1) 2 (1) 2 (1)
+ 0 = Ds . (9.11.15)
t xy x2
776 Introduction to Theoretical and Computational Fluid Dynamics
Substituting the normal-mode forms and rearranging, we nd that the complex amplitude of the
surfactant concentration is given by
1 (k)
=k . (9.11.16)
0 c + ikDs
Correspondingly, the complex amplitude of the surface tension is given by
1 Ma
= k (k). (9.11.17)
0 c + ikDs
Evaluating the derivative and remembering that a3 = a1 , we obtain
1 Ma
= k ek a1 (1 + q) + a2 (1 + k) + a4 (1 k) q , (9.11.18)
0 c + ikDs
where q = exp(2k).
(1)
(1)
x
+ = , (9.11.19)
y x y=h x
where
The term on the right-hand side of (9.11.19) represents the Marangoni traction. Expressing the
velocity in terms of the stream function and rearranging, we obtain
2 (1) (1)
1 (1)
= , (9.11.21)
y2 x 2 y=h x
Substituting the normal-mode forms, we nd that
d2
1
+ =i . (9.11.22)
dy 2 y=k k
Computing the derivatives, recalling that a3 = a1 , and rearranging, we nd that
1 k
(1 q) a1 + (k + 1) a2 + (k 1) q a4 = i e . (9.11.23)
2k
where 0 is the interfacial curvature in the unperturbed conguration. All terms in (9.11.24) are
evaluated at the unperturbed lm surface, y = h. The rst term on the right-hand side expresses the
contribution of the viscous normal stress at the interface. Because in the unperturbed conguration
the lm surface is at, 0 = 0, the last term on the right-hand side is identically zero. Consequently,
surface tension variations do not aect the normal force balance.
Dierentiating (9.11.24) with respect to x and using the tangential projection of the Stokes
equation to evaluate the pressure derivative on the left-hand side, we obtain the preferred pressure-
free form
(1)
p(1) 2 uy 3
= 2 u(1) = 2 + g 0 , (9.11.25)
x x
xy x x3
where all terms are evaluated at the unperturbed lm surface, y = h. Introducing the stream
function, we obtain
3 (1) 3 (1)
3
3 2 + = g 0 . (9.11.26)
x y y 3 x x3
where
1 g
+ 0 . (9.11.30)
2 k 2
Setting the determinant of the matrix M to zero provides us with a quadratic equation with two
purely imaginary roots for c, yielding negative growth rates, I ikcI .
778 Introduction to Theoretical and Computational Fluid Dynamics
Dimensionless numbers
A Bond number can be dened with respect to either the wavelength of the perturbation or the
layer thickness,
gL2 gh2
Bo , Bo , (9.11.31)
0 0
Corresponding dimensionless property groups can be introduced expressing the importance of the
surfactant diusivity,
0 L 0 h h
= , = = . (9.11.32)
Ds Ds L
Multiplying the third row of (9.11.29) by k/0 , we obtain the new matrix
2 1 1 0
i 1 k
1 q k + 1 ( k 1) q 2 e
M = 2 2 , (9.11.33)
c k(1 + q) + i (1 q)/k c k + i (c k + i ) q 0
Ma (1 + q) Ma (1 + k) Ma (1 k) q (c + ik/ ) ek
k 2 1
k 2 = 2
(Bo + 4 2 ) = (Bo + k2 ). (9.11.34)
0 8 2
Setting the determinant of the matrix M to zero provides us with a quadratic equation with two
purely imaginary roots for c, yielding negative growth rates, I ikcI .
Absence of surfactant
In the absence of surfactant, Ma = 0, we set 0 = and 1 = 0 and retain the rst three equations
in (9.11.29) involving the dimensionless group . After some algebra, we nd that the real part of
the phase velocity is zero, as expected by symmetry, and the dimensionless growth rate is given by
hI 1 1 sinh(2k) k
I = 2 . (9.11.35)
k k2 + cosh2 k
Since the fraction on the right-hand side is positive for any k, the growth rate is negative and the
lm is stable.
In Section 9.3.6, we derived an expression for the growth rate subject to the lubrication
approximation. In the present notation, equation (9.3.19) becomes
h3 2
I = i k (g + k2 ) = 2 k 2 . (9.11.36)
3 3h
This expression is the asymptotic form of (9.11.35) in the limit is k tends to zero (Problem 9.11.1(b)).
9.12 Film ow down an inclined plane 779
Computer code
Directory coat0 s of the software library Fdlib contains a code that computes the growth rates
based on the equations derived in this section (Section D.8, Appendix D.)
Problem
Computer Problem
9.12.1 Base ow
The base ow whose stability we wish to investigate is described by the Nusselt at-lm solution
discussed in Section 5.1.4, as illustrated in Figure 9.12.1. To simplify the notation, it is helpful
to nondimensionalize all variables using as characteristic velocity the unperturbed velocity at the
free surface, umax
x , characteristic length the unperturbed lm thickness, h, characteristic time the
ratio h/umax
x , and characteristic pressure and stress the combination umax
x /h. In dimensionless
780 Introduction to Theoretical and Computational Fluid Dynamics
y
h
Gas
Liquid
Figure 9.12.1 Periodic disturbances on a liquid lm owing down an inclined plane. The dashed line
represents the unperturbed lm surface.
where is the plane inclination angle, h is the unperturbed lm thickness, is the kinematic viscosity
of the uid, y = y/h is a dimensionless distance from the plane, and the superscript (0) denotes the
steady state. The dimensionless volumetric ow rate per unit width of the lm and mean velocity
of the uid are
2 2 2 2
Q(0) = Q(0) = , umean = umean = . (9.12.2)
gh3 sin 3 x
gh2 sin x 3
For future reference, we introduce the dimensionless stress tensor in the xy plane,
! "
(0) 2 (0) cot 1
= = 2 (1 y) . (9.12.3)
gh sin 1 cot
For convenience, we have assumed that the pressure takes the reference value of zero at the free
surface so that (0) = 0 at y = 1.
where
tumax 1
t = x
, =
, , x = x (9.12.5)
h x y h
9.12 Film ow down an inclined plane 781
are the dimensionless time, gradient operator, and position vector. The Reynolds number is
humax gh3 2
Re = x
= 2 sin = Fr2 , (9.12.6)
2 sin
where Fr umax
x / gh is the Froude number.
OrrSommerfeld equation
A normal-mode disturbance alters the velocity eld from the steady eld, u(0) (x), to an unsteady
eld,
where is a small dimensionless coecient and the superscript (1) denotes the perturbation. Nondi-
mensionalizing the normal-mode velocities given in (9.4.31) using the scales discussed in Section
9.12.1, we derive the dimensionless disturbance velocity
u(1)
x = f (y) exp[ik(x ct)], u(1)
y = ik f (y) exp[ik(x ct)], (9.12.8)
To complete the denition of the linear stability problem, we require four boundary conditions
for the eigenfunction f. Two boundary conditions arise by requiring that both velocity components
are zero at the plane, yielding
Two additional boundary conditions emerge by demanding that the shear stress is zero and the nor-
mal stress is consistent with the jump in interfacial traction due to surface tension at the free surface.
To prepare the ground for implementing these conditions, we digress to describe the kinematics of
the perturbed ow at the free surface.
Kinematic compatibility
In the normal-mode analysis, the scaled elevation of the deformed free surface is described by the
real or imaginary part of the function
is the normal-mode wave form of the perturbation, and A is a dimensionless complex constant.
Substituting (9.12.11) into the kinematic boundary condition
(9.12.13)
+ ux uy = 0,
t x
we obtain
ikA (c + ux ) exp[ik(x ct)] uy = 0, (9.12.14)
where the velocity components ux and uy are evaluated at the location of the perturbed free surface.
To compute the free surface velocity, we expand the velocity in a Taylor series around the undeformed
position,
u
where the dots represent terms with quadratic or higher-order dependence on . Substituting
(9.12.11) into (9.12.16) and the result into (9.12.14), and retaining terms of rst order with re-
spect to , we nd that A = f(1)/(c 1). Consequently, the deformed free surface is described by
the real or imaginary part of the function
f(1)
(x, t) = exp[ik(x ct)]. (9.12.17)
c 1
Free-surface conditions
At the free surface, the hydrodynamic traction undergoes a jump that is balanced by surface tension
or a more general interfacial force eld. For a clean interface with constant surface tension ,
1
n = n, (9.12.18)
Bo
where is the dimensionless stress tensor, n is the normal unit vector pointing into the ambient
gas, is the dimensionless curvature of the free surface in the xy plane, and
gh2
Bo = sin (9.12.19)
2
is a Bond number. It is sometimes convenient to introduce the inverse Bond number, , Weber
number, We, and a property group, S, that depends exclusively on the physical properties of the
uid, dened as
1 Re S 1 h(umax )2 2
1/3
x
= = = , We = , S = . (9.12.20)
Bo We Re2/3 sin1/3 g4
For water at room temperature, S is approximately equal to 4,280.
9.12 Film ow down an inclined plane 783
Next, we expand both sides of (9.12.18) in perturbation series with respect to . For this
purpose, we decompose the stress tensor and normal unit vector into their steady and unsteady or
disturbance components,
where
n(0) = 0, 1 , n(1) = , 0 . (9.12.22)
x
To compute the left-hand side of (9.12.18) in terms of ow variables at the undisturbed position of
the free surface, we apply once again the method of domain perturbation, writing
( n)y= = (y = 1) + + n
y y=1
(0)
= (0) (y = 1) + (1) (y = 1) + + (n(0) + n(1) )
y y=1
(0)
(0)
= (1) (y = 1) + n + . (9.12.23)
y y=1
Substituting the last expression into the left-hand side of (9.12.18), approximating = 2 / x2
on the right-hand side, and equating terms of same order in , we obtain
! " ! " ! "
(1) 0 1 2 0
(y = 1) = 2 + 2 , (9.12.24)
1 cot x 1
where is the inverse Bond number. Next, we decompose (9.12.24) into its two scalar constituents
expressing tangential and normal force balances,
(1) (1) 2
xy (y = 1) = 2, yy (y = 1) = 2 cot + . (9.12.25)
x2
Expressing the stress tensor in terms of its pressure and the viscous constituents, we obtain two
scalar linearized boundary conditions,
(1) (1) (1)
ux uy uy 2
+ = 2, p(1) = 2 + 2 cot 2 , (9.12.26)
y x y x
where all terms are evaluated at the unperturbed free surface position, y = 1.
(1) = Re
t p + u(0) u
(1) + u(1) u
(0) + (
2 u(1) ) t. (9.12.28)
t
Now dierentiating the second equation in (9.12.26) with respect to x, we obtain
(1)
p(1) 2 uy 3
=2 +2 cot 3 , (9.12.29)
x x y x x
evaluated at y = 1. Equating the right-hand sides of (9.12.28) and (9.12.29) and retaining only
terms of order , we obtain
u(1) ux
2 (1)
(1) (1)
x 2 uy 3
Re + + ux = 2 +2 cot 3 , (9.12.30)
t x x y x x
evaluated at y = 1. Finally, we substitute (9.12.8) along with (9.12.17) into (9.12.30) and obtain
the desired boundary condition
k
f (1) + k i Re (c 1) 3k f (1) i (2 cot + k 2 ) f(1) = 0, (9.12.31)
c 1
involving the Reynolds number, Re, and the inverse Bond number, .
Long-wave solution
Useful insights can be obtained by considering disturbances with long wavelength or small scaled
wave number, k [435]. It turns out from the numerical solution of the exact eigenvalue problem
that the fastest growing mode does have a long wavelength, and the approximate analysis for small
scaled wave numbers provides us with accurate predictions. Expanding the eigenfunction f and the
complex phase velocity c in Taylor series with respect to k, we obtain
f = f0 + k f1 + k 2 f2 + , c = c0 + kc1 + k2 c2 + . (9.12.32)
9.12 Film ow down an inclined plane 785
Substituting these expansions into the OrrSommerfeld equation and associated boundary condi-
tions, assuming that both Re and are of order unity, and collecting terms of same order in k, we
derive a sequence of eigenvalue problems.
f0 = 0, (9.12.33)
f0 = B y 2 , c0 = 2, (9.12.35)
where B is an arbitrary constant expressing the initial magnitude of the perturbation. Physically,
this solution tells us that long waves translate with a phase velocity that is equal to twice the velocity
of the liquid lm at the free surface. It is interesting to note that this behavior contradicts Rayleighs
theorem for inviscid ow discussed in Section 9.5.
Since k has been used as a perturbation variable, it appears neither in the zeroth-order equation
(9.12.33) nor in the rst-order equation (9.12.36). Substituting (9.12.35) into (9.12.36), we obtain
the simple form
A solution that satises the two boundary conditions at the wall is given by the fth-degree poly-
nomial
1
f1 = i B Re y 4 (y 5) + C y 3 , (9.12.39)
30
where C is a new constant. Substituting (9.12.39) into the third and fourth boundary conditions in
(9.12.37), we derive two linear homogeneous equations for B and C,
! 8
" ! "
c1 15 iRe 2 B
= 0. (9.12.40)
2i cot g C
786 Introduction to Theoretical and Computational Fluid Dynamics
Setting the determinant of the coecient matrix to zero to ensure a nontrivial solution, we obtain
C = 13 i B cos and
8 5
c1 = i (Re cot ). (9.12.41)
15 4
Approximate methods
Benjamin [28] developed an approximate method for solving the linear stability problem based on
an innite series expansion,
f(y) = An y n . (9.12.45)
n=0
Substituting (9.12.9) into the OrrSommerfeld equation, we obtain a recurrence relationship among
four groups of coecients, An , An2 , An4 , and An6 . Enforcing the boundary conditions provides
us with an algebraic eigenvalue problem with innite dimensions. In practice, the series (9.12.45)
and associated linear algebraic eigenvalue problem are truncated at a nite level. Benjamin noted
that truncating amounts to expressing the solution in a Taylor series with respect to k and Re, and
then truncating that series. For example, keeping 16 terms produces results that are accurate to
third order in k and Re. Benjamin solved the corresponding eigenvalue problem analytically and
derived an involved relationship between the critical wave number for neutral stability, the Reynolds
number, Re, and the Weber number, We.
Instead of implementing the long wave approximation in the nal linearized system and bound-
ary conditions, Benney [29] worked with an approximate form of the NavierStokes equation that
assumes long waves of nite amplitude at the outset. The governing equations are developed as dis-
cussed in Section 6.4. By carrying out a linear stability analysis of the resulting evolution equation
9.13 Capillary instability of a curved interface 787
for the lm thickness, Benney produced the second and third coecients of the complex growth rate
c shown in (9.12.32),
32 5
c2 = 2 Re (Re cot ), (9.12.46)
63 4
and
1 157 6 8 138904 2 1213952 3
Problems
9.12.2 Two-layer lm ow
Consider the ow of two superposed liquid lms down an inclined plane. Formulate and discuss the
equations and boundary conditions governing the temporal linear stability problem.
y
A
p
0
a
B
Normal-mode analysis
It is convenient to work in a frame of reference where the jet appears to be a stationary column
of uid and the ambient gas executes an inconsequential backward motion. The ow due to a
perturbation will be described in cylindrical coordinates coaxial with the undisturbed cylindrical
interface, (x, , ). In the case of axisymmetric perturbations, the jet radius is described by the real
part of the function
is the normal-mode wave form of the perturbation, A is the complex amplitude of the interfacial
wave, k = 2/L is the wave number, L is the wavelength, and c is the complex phase velocity.
where J0 is the zeroth-order Bessel function and I0 is the zeroth-order modied Bessel function. To
derive the last expression, we used the property of Bessel functions Jn (ix) = in In (x), where n is a
positive integer and x is a real argument. To compute the ratio B/A and complex phase velocity, c,
we require two constraints.
Kinematic compatibility
The rst constraint arises by substituting the expressions given into (9.13.2) and (9.13.3) in the
linearized kinematic condition at the interface stated in (1.10.3), /t (1) / = 0, obtaining
d dJ0 (ik)
ikcA = B =B = i kB J1 (ik), (9.13.5)
d d
evaluated at the unperturbed interface, = a, where J1 is the rst-order Bessel function. Simplify-
ing, we obtain
p p0 1 2
= 2 (a ) 2 . (9.13.7)
a x
The second term on the right-hand side is the curvature in a meridional plane and the second term
is the curvature in its conjugate orthogonal plane. The pressure in the liquid can be evaluated using
Bernoullis equation (3.6.11) for unsteady irrotational ow. Discarding the square of the velocity
and substituting into (9.13.7) the resulting linearized form, we obtain
(1) 2
= 2 + . (9.13.8)
t a x2
Substituting into (9.13.8) expressions (9.13.2) and (9.13.3), and evaluating the left-hand side at the
location of the unperturbed interface, = a, we nd that
ikcB I0 (ka) = A 1 (ka)2 . (9.13.9)
a2
790 Introduction to Theoretical and Computational Fluid Dynamics
I1 (ka)
c2 = 2
1 (ka)2 , (9.13.10)
ka I0 (ka)
The modied Bessel functions, I0 (ka) and I1 (ka), are positive for any scaled wavenumber, ka.
Accordingly, the sign of the right-hand side of (9.13.10) is determined by the magnitude of ka.
When ka > 1, the right-hand side of (9.13.10) is positive and the complex phase velocity, c,
takes two real values corresponding to stable normal modes. Axisymmetric waves with ka > 1 travel
undamped to the left or right with phase velocity
I (ka)
1/2
1 2
c= [1 (ka) ] . (9.13.11)
ka2 I0 (ka)
The permanence of the amplitude of the interfacial wave is a consequence of the absence of viscous
dissipation in an inviscid ow.
When ka < 1, the right-hand side of (9.13.10) is negative, and the complex phase velocity, c,
takes two complex conjugate values with zero real part. The negative imaginary part corresponds
to a stable mode, while the positive imaginary part corresponds to an unstable mode with growth
rate
k I (ka)
1/2
1 2
I = kcI = 1 (ka) . (9.13.12)
a2 I0 (ka)
The fastest growth occurs approximately when ka = 0.679, and the corresponding maximum growth
rate is
k
1/2
(I )max = 0.34 . (9.13.13)
a2
Physically, perturbations with large wavelength amplify due to a surface pressure generated by
surface tension.
We have found that, when the PlateauRayleigh criterion ka < 1 is fullled, the jet falls
prey to the capillary instability [334]. The growth of perturbations in the unstable regime can
be explained by considering the pressure distribution in the liquid column at the initial instant,
neglecting pressure variations due to the uid acceleration. Considering station A in Figure 9.13.1,
we note that the curvature of the interface in a meridional plane increases due to the interfacial
corrugation, whereas the curvature in a plane that is perpendicular to the x axis decreases due to
the increased jet radius; the inverse is true at station B. The linearized form of the expression for the
mean curvature on the right-hand side of (9.13.9) then shows that the pressure at station A will be
lower than the pressure at station B and the uid will be driven towards the crest when 0 < ka < 1,
thereby amplifying the interfacial corrugations.
9.13 Capillary instability of a curved interface 791
Nonlinear evolution
The initial amplication of interfacial corrugations due to the Rayleigh instability is followed by the
formation of liquid drops connected by axisymmetric bridges, eventually breaking up into smaller
satellite drops. Mansour & Lundgren [255] present numerical simulations based on the boundary-
integral method discussed in Chapter 10.
Three-dimensional corrugations
To study the behavior of non-axisymmetric disturbances, we describe the shape of the interface by
the real or imaginary part of the shape function
where m an integer representing the azimuthal wave number. Repeating the preceding analysis, we
nd that the growth rate is given by the generalized expression [334]
Im (ka)
c2 = 1 (ka)2 m2 , (9.13.15)
ka2 Im (ka)
where Im are mth-order modied Bessel functions. When m = 0, we recover (9.13.10). Since the
ratio of the modied Bessel functions on the right-hand side of (9.13.15) is positive for any integer
m > 0, the right-hand side is positive for any m 1 and the interface is stable to non-axisymmetric
perturbations. Thus, a cylindrical column of uid is expected to break up by developing axisym-
metric corrugations, in agreement with observation.
Normal-mode analysis
To carry out the normal-mode stability analysis for axisymmetric perturbations, we introduce cylin-
drical polar coordinates, (x, , ), where the x axis coincides with the axis of revolution of the
unperturbed interface. The radial position of the interface is described by the real or imaginary part
of the function
where a is the unperturbed thread radius, is a dimensionless coecient whose magnitude is much
less than unity,
(a) (b)
0.18
0.16
y
0.14
0.12
Fluid 2
0.1
s
a
0.08
Fluid 1
0.06
0.04
z
x 0.02
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
ka
Figure 9.13.2 (a) Illustration of an innite viscous thread suspended in a quiescent innite viscous
uid. (b) Dependence of the dimensionless growth rate, s I a1 /, on the scaled wave number,
ka, for viscosity ratio = 0 (dashed line) corresponding to a thread suspended in vacuum, 0.01,
0.05, 0.1, 0.2, 0.5, 1.0, 2.0, 5.0, 10.0, and 20.0 (lowest line).
is the normal-mode wave form of the perturbation, A is the complex amplitude of the interfacial
wave, k = 2/L is the wave number, L is the wavelength, and c is the complex phase velocity.
Normal-mode expansion
The stream function is expressed in the normal-mode form
(1)
j (x, , t) = j () exp[ik(x ct)], (9.13.23)
2
k j = j , (9.13.24)
d 2 d
where
d2 1 d
2
k 2 j = 0. (9.13.25)
d d
The general solution is
1 () = A1,1 I1 () + A2,1 I0 () (9.13.26)
for the thread uid labeled 1, and
2 () = B1,2 K1 () + B2,2 K0 () (9.13.27)
for the ambient uid labeled 2, where I0 , I1 , K0 , and K1 are Bessel functions, Ai,j , Bi,j are complex
constants, and k is the dimensionless radial position. The rst term on the right-hand side
of (9.13.26) or (9.13.27) satises (9.13.24) with j = 0. The second term satises (9.13.24) with j
being a nontrivial solution of (9.13.25).
Substituting the preceding expressions into (9.13.18) and using the properties of the Bessel
functions,
1
I0 (z) = I1 (z), I1 (z) = I0 (z)
I1 (z), K0 (z) = K1 (z),
z
1
K1 (z) = K0 (z) K1 (z), (9.13.28)
z
where a prime denotes a derivative with respect to z, we nd that
uxj = Uxj () exp[ik(x ct)], uj = Uj () exp[ik(x ct)], (9.13.29)
where
1 d1
Ux1 () = = A1,1 k I0 () + A2,1 2 I0 () + I1 () ,
d
ik
U1 () = 1 = ik A1,1 I1 () + A2,1 I0 () (9.13.30)
794 Introduction to Theoretical and Computational Fluid Dynamics
Pressure eld
The normal-mode pressure eld is
where P1 and P2 are the uniform unperturbed pressures, P1 P2 = /a is the unperturbed capillary
pressure, and is the surface tension. To compute the disturbance pressure eld, we consider the x
component of the Stokes equation for axisymmetric ow,
pj 2 ux j 1 uxj
= j + . (9.13.33)
x x2
Substituting (9.13.32) along with the rst equation into (9.13.29) and linearizing, we obtain
ij 1 d dUxj ij 1 d dUxj
j () = ij kUxj = ( ) k 2 Uxj (9.13.34)
k d d k d d
or
ij d2 Uxj 1 dUxj
j () = + k2 Uxj . (9.13.35)
k d2 d
Making substitutions, we nd that
Note that only one term of the general solution in each uid contributes to the pressure eld.
Kinematic compatibility
Kinematic compatibility requires that D[f (x, t) ]/Dt = 0 at the interface, where D/Dt is the
material derivative and the function f describes the position of the interface according to (9.13.16).
Carrying out the dierentiation, we obtain
f f
+ ux u = 0, (9.13.38)
t x
where the velocity is evaluated at the interface. Substituting the preceding expressions and lineariz-
ing, we obtain an expression for the interfacial amplitude,
U1 ( = a) 1
A= = A1,1 I1 (k) + A2,1 a I0 (k) . (9.13.39)
ick c
Because of the continuity condition (9.13.37), the outer uid velocity, U2 ( = a), could have been
used in this expression.
x1 x2
x1 x2 =a = 1 + 2 + = 0. (9.13.40)
x =a x =a
Substituting the normal-mode expansions, we obtain
dU
dU
x1 x2
1 + ikU1 2 + ikU2 = 0. (9.13.41)
d =a d =a
where
F1 = kI1 (k), F2 = I1 (k) + k I0 (k), F7 = k K1 (k), F8 = K1 (k) k K0 (k) . (9.13.45)
796 Introduction to Theoretical and Computational Fluid Dynamics
1 2 =a = p + 2 1 p + 2 2 = 2m , (9.13.46)
=a =a
where m is the mean curvature of the interface. In the linearized approximation,
1 1 A
2m = + 2 + xx = + 2 (1 k 2 ) exp[ik(x ct)]. (9.13.47)
a a a a
Using (9.13.39) to eliminate the interface amplitude, A, we obtain
1
2m = + 2(1)
m exp[ik(x ct)], (9.13.48)
a
where
1 k 2
2(1)
m = A1,1 I1 (k) + A2,1 a I0 (k) . (9.13.49)
ca2
Substituting the normal-mode expansions into (9.13.46), we obtain
U1
U2
1 k2
1 + 21 2 + 22 = 2
A1,1 I1 (k) + A2,1 a I0 (k) . (9.13.50)
=a =a ca
Using the expressions
dU1 1
where
1 k 2
=i . (9.13.55)
21 c k
If c is imaginary, as expected due to the absence of a mean base ow, the parameter is real.
9.13 Capillary instability of a curved interface 797
(a)
I1 (k) I0 (k) K1 (k) K0 (k)
k I0 (k) 2 I0 (k) + kI1 (k) k K0 (k) 2 K0 (k) + kK1 (k)
M=
k I1 (k) I1 (k) + k I0 (k) k K1 (k) [K1 (k) kK0 (k)]
kI0 (k) I1 (k) I1 (k) k I1 (k) I0 (k) [k K0 (k) + K1 (k)] k K1 (k)
(b)
I1 (k) k I0 (k) I1 (k)
k I0 (k) k [I0 (k) + k I1 (k)]
M =
k I1 (k) a2 I0 (k)
2
k I0 (k) I1 (k) I1 (k) (1 + k )I1 (k) kI0 (k) [kI0 (k) I1 (k)]
K1 (k) k K0 (k) K1 (k)
k K0 (k) k [K0 (k) + kK1 (k)]
k K1 (k) 2
k K0 (k)
2
[k K0 (k) + K1 (k)] [(1 + k ) K1 (k) + k K0 (k)]
(c)
I1 (k) k I0 (k) I1 (k)
I0 (k) I0 (k) + k I1 (k)
M =
I1 (k) k I0 (k)
k I0 (k) I1 (k) I1 (k) (k2 + 1) I1 (k) k I0 (k) [k I0 (k) I1 (k)]
K1 (k) k K0 (k) K1 (k)
K0 (k) K0 (k) + k K1 (k)
K1 (k) k K0 (k)
2
[k K0 (k) + K1 (k)] [(1 + k ) K1 (k) + k K0 (k)]
Table 9.13.1 Matrices arising from the linear stability analysis of a viscous liquid thread. Setting the
determinant of any one of these matrices to zero provides us with an algebraic equation for the
complex phase velocity.
k I0 (k) I1 (k) K1 (k) k K0 (k) K1 (k)
I1 (k) K1 (k) k K0 (k) K1 (k)
I1 (k) k I0 (k) I1 (k) k K0 (k) K1 (k)
I1 (k) k I0 (k) I1 (k) K1 (k)
Table 9.13.2 Determinants of matrices arising from the linear stability analysis of a viscous liquid
thread.
Switching the sign in the third column and eliminating the factor k in each term in the second row
and third rows, we obtain the new matrix M shown in Table 9.13.1(c). Expanding the determinant
of this matrix with respect to the last row and setting the resulting expression to zero, we derive
the equation
1
= k I0 (k) I1 (k) D1 (k 2 + 1) I1 (k) k I0 (k) D2
k K0 (k) + K1 (k) D3 + (1 + k 2 ) K1 (k) + k K0 (k) D4 , (9.13.56)
where = I1 (k) D1 k I0 (k) I1 (k) D2 , and D1 D4 are four determinants dened with respect
to the rst three rows of M and given in Table 9.13.2. Invoking the denition of from (9.13.55),
we nally derive a purely imaginary growth rate,
kc = i (1 k 2 ), (9.13.57)
21 a
where is computed from (9.13.56) [410]. A code that produces the growth rate is available in the
software library Fdlib (Section D.17, Appendix D, Directory thread0).
Graphs of the dimensionless growth rate s a1 I / plotted against the scaled wave number,
ka, are shown in Figure 9.13.2(b) for several viscosity ratios, = 2 /1 . The results demonstrate
that the thread is unstable for scaled wave numbers that are lower than the RayleighPlateau
threshold, ka < 1. When the ambient uid is inviscid, = 0, the growth rate tends to a nite limit
9.13 Capillary instability of a curved interface 799
(a) (b)
0.05
0.04 0.15
0.03
0.1
0.02
0.05
0.01
0 0
s
s
0.01
0.05
0.02
0.1
0.03
0.04 0.15
0.05
0 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2
ka ka
Figure 9.13.3 Eect of an insoluble surfactant on the instability of a viscous thread suspended in an in-
nite uid. The graphs demonstrate the dependence of the dimensionless growth rate s I a1 /
on the scaled wave number for = 0.001 (dashed lines), 0.05 (dotted lines), 0.1, 0.2, . . . , 0.5, and
negligible surfactant diusivity, = . The viscosity ratio is (a) = 1 and (b) 0.
as ka tends to zero. In contrast, when the ambient is viscous, the growth rate vanishes as ka tends
to zero. We conclude that the ambient uid viscosity has a profound eect on the evolution of long
waves.
Graphs of the dimensionless growth rate s a1 I /0 plotted against the scaled wave number,
ka, are shown in Figure 9.13.3 for viscosity ratio = 1 and 0. Two normal modes arise in the
presence of an insoluble surfactant. The rst normal mode is unstable for scaled wave numbers that
are lower than the RayleighPlateau threshold, ka < 1. The second normal mode is stable for all
wave numbers. The corresponding growth rates arise from the splitting of the contiguous lobe in
800 Introduction to Theoretical and Computational Fluid Dynamics
(a) (b)
Figure 9.13.4 Evolution of an annular layer coating the interior side of a circular tube subject to
an axisymmetric perturbation. The straight dotted lines show the initial position of the interface.
Frame (a) illustrates the initial amplication of sinusoidal interfacial corrugations and the forma-
tion of primary drops connected with axisymmetric bridges, as shown in frame (b). The bridges
eventually breakup to yield two alternating sequences of drops.
the absence of surfactant, = 0. Because the surfactant reduces the growth rate of the unstable
mode, it has a dampening action on the capillary instability. The eect is small when = 1, but
signicant when = 0.
Problems
Computer Problem
A related application concerns the ow of a viscous uid through an isotropic porous medium,
such as a brous matrix or ground rock. According to Darcys law, the macroscopic velocity of the
uid, U, dened as the average velocity of the uid over a volume that is small compared to the
global dimensions of the ow but large compared to the size of the bers or grains, is related to the
corresponding macroscopic hydrodynamic pressure, p, through the equation U = (/)p, where
is a physical constant called the permeability of the uid through the porous medium. It is evident
that the macroscopic pressure plays the role of a potential function, = (/) p. The continuity
equation requires that p is a harmonic function, 2 p = 0. The computation of a porous-medium
ow is then reduced to solving Laplaces equation subject to the Neumann boundary conditions over
the impermeable boundaries of the ow or to the Dirichlet boundary condition over boundaries that
are exposed to the atmosphere.
Motivated by the pervasiveness of Laplaces equation in the various branches of uid mechanics
and other physical and engineering sciences, we devote this chapter to discussing a powerful class
of numerical methods for computing potential ow in domains with arbitrary geometry, known as
802
10.1 The boundary-integral equation 803
n
x0
n +
+
Dint
Dext
Flow
Figure 10.1.1 Illustration of a ow domain bounded by an internal and an external surface whose
union comprises the boundary, D. In the case of exterior ow decaying at innity, Dext does not
appear.
where G(x0 , x) is a Greens function of Laplaces equation and n is the unit vector normal to D
pointing into the control volume. The Greens function satises the symmetry property G(x0 , x) =
G(x, x0 ). The rst integral on the right-hand side of (10.1.1), representing a distribution of point
sources, is the single-layer potential. The second integral, representing a distribution of point-source
dipoles oriented normal to D, is the double-layer potential.
Equation (10.1.1) is also valid at a point, x0 , located inside a partially or entirely innite
domain of ow that is bounded by an interior boundary D, provided that the velocity decays and
therefore the potential tends to the reference value of zero at innity.
Integral equations
The representation (10.1.1) allows us to compute the value of (x0 ) at any point x0 inside a selected
control volume in terms of the distribution of and its normal derivative, n, over the boundaries
of the control volume. In practice, physical arguments provide us with one scalar boundary condition
804 Introduction to Theoretical and Computational Fluid Dynamics
for the distribution of , n , or their linear combination, but not for both. It would appear
then that (10.1.1) is useful only when both boundary distributions are available. The basic idea of
the boundary-integral method is to let the point x0 approach the boundary D, and thereby reduce
(10.1.1) into a Fredholm integral equation for the unknown distribution.
Single-layer potential
Examining the singularity of the single-layer integrand in (10.1.1), we nd that, as the point x0
approaches and then crosses D, the single-layer potential varies in a continuous fashion.
Double-layer potential
The double-layer potential exhibits a jump across the integration domain, D. The discontinuity is
seen most clearly by writing
lim (x) n(x) G(x0 , x) dS(x) = J0 + 0 lim n(x) G(x0 , x) dS(x), (10.1.2)
x0 D D x0 D D
where 0 (x0 ),
J0 (x0 ) lim (x) (x0 ) n(x) G(x0 , x) dS(x), (10.1.3)
x0 D D
and the superscripts + or designate that the point x0 approaches D from within the control volume,
indicated by the normal vector, or from the exterior side, as shown in Figure 10.1.1. Examining the
singularity of the integrand, we nd that the function J0 (x0 ) varies continuously as x0 approaches
and then crosses D. To assess the behavior of the integral in the right-hand side of (10.1.2), we recall
that D represents the collection of all boundaries of an enclosed domain of ow and use identities
(2.2.15) to nd that
lim (x) n(x) G(x0 , x) dS(x) = J0 (x0 ) + (x0 ) (10.1.4)
x0 Dext D
for an interior boundary, where + = 0 and = 1. These equations demonstrate that, as the
point x0 crosses an exterior or interior boundary to enter the domain of ow, the double-layer integral
undergoes a discontinuity equal to (x0 ).
10.1 The boundary-integral equation 805
Combining (10.1.4) and (10.1.5) with (10.1.6) and (10.1.7), we obtain a unied relationship
between the limits of the double-layer potential and its principal value for an exterior or interior
boundary,
PV
1
lim (x) n(x) G(x0 , x) dS(x) = (x) n(x) G(x0 , x) dS(x) (x0 ). (10.1.8)
x0 D D D 2
The signicance of this equation hinges on the realization that the principal value of the double-
layer potential is much easier to compute than the limit of the double-layer potential as the point
x0 approaches the boundary from either side.
where the point x0 lies on D. It will be noted that (10.1.9) is identical to (10.1.1), except that the
right-hand side is multiplied by a factor of two and the principal value of the double-layer integral is
employed. Because of the symmetry property G(x2 , x1 ) = G(x1 , x2 ), the arguments of the Greens
function in the single-layer and double-layer potentials can be freely switched.
An alternative method of deriving (10.1.9) involves assuming that the point x0 lies on a smooth
boundary, D, and then repeating the analysis of Section 2.3, excluding from the control volume a
hemisphere centered at the point x0 . Generalizing this method, we nd that (10.1.9) is also valid
at a point x0 located at a boundary wedge or corner, provided that (x0 ) on the left-hand side is
multiplied by the factor /2, where is the solid angle subtended by the corner on the side of the
control volume (Problem 10.1.1). In the case of a smooth boundary, = 2.
806 Introduction to Theoretical and Computational Fluid Dynamics
where
I S (x0 ) G(x0 , x) n(x) (x) dS(x) (10.1.11)
D
is the known single-layer potential. Recasting (10.1.10) into the standard form of an integral equa-
tion, we obtain
PV
(x0 ) = (x) K(x0 , x) dS(x) 2 I S (x0 ), (10.1.12)
D
where
Regularity
To assess the singular nature of the kernel of the double-layer potential, we note the asymptotic
behavior (2.2.10) and nd that, as x x0 ,
1
1 x x0
K(x0 , x) 2 n(x) = n(x) . (10.1.14)
r 2 r3
When the normal unit vector varies continuously over D, as x x0 , the vectors x0 x and n tend
to become perpendicular and their dot product tends to vanish. Consequently, the singularity of the
kernel K(x0 , x) is of order 1/r2 , where is a positive number determined by the smoothness of
the boundary. If D is a smooth surface with nite curvature, = 0 and the singularity of the kernel
is of order 1/r. Since the order of the singularity is less than that dimensionality of the domain
of integration, which is equal to two, the kernel K(x0 , x) and the integral equation (10.1.12) are
weakly singular. Later in this chapter, we will see that this property has important implications on
the existence and uniqueness of solution, as well as on the feasibility of computing the solution by
iterative methods.
Uniqueness of solution
In Section 10.7, we will study the properties of the integral equation (10.1.12) and will nd that, when
D represents the interior boundary of an innite external ow, the integral equation has a unique
solution. However, when D is the exterior boundary of an internal ow in the absence of interior
boundaries, the integral equation has either no solution or an innite number of solutions that dier
10.1 The boundary-integral equation 807
by an arbitrary constant. The physical relevance of these results becomes evident by identifying
with a temperature eld and noting that a steady temperature distribution in a conned domain
will exist only if the ow rate of heat through the boundaries is zero. When this condition is met,
the temperature can be set to an arbitrary level.
where B is the surface of the body. Taking the limit as the point x0 approaches the surface of the
body and using (10.1.8), we derive an integral equation of the second kind for ,
PV
(x0 ) = 2 (x0 ) + 2 (x) n(x) G(x0 , x) dS(x). (10.1.16)
B
Substituting the free-space Greens function, G(x0 , x) = 1/(4r), where r = |x x0 |, we obtain the
specic form
1 PV
x x0
(x0 ) = 2 (x0 ) (x) n(x) dS(x). (10.1.17)
2 B r3
In Section 10.7, we will show that equation (10.1.17) has a unique solution for any ambient potential
representing an arbitrary irrotational incident ow.
As a specic application, we consider streaming ow with velocity U past a stationary sphere with
radius a centered at the origin of the chosen Cartesian coordinates. The exact solution for the
velocity potential was given in (7.5.6) as
1 a3
(x) = 1 + U x. (10.1.18)
2 |x|3
3
Substituting into (10.1.17) (x) = U x and the boundary distribution (x) = 2
U x, and
simplifying, we obtain the identity
3 PV
x x0
x0 = x
3
n(x) dS(x), (10.1.19)
2 Sphere r
where r = |x x0 |, n(x) = a1 x is the normal unit vector, the integration is performed over the
surface of the sphere, and the point x0 lies at the surface of the sphere (Problem 10.1.2).
808 Introduction to Theoretical and Computational Fluid Dynamics
n
C
l
Figure 10.1.2 In the case of axisymmetric ow, the boundary integrals can be reduced to line integrals
along the trace of the boundaries in an azimuthal plane.
1
with forcing function F = I D 2 , where
PV
I (x0 )
D
(x) n(x) G(x0 , x) dS(x) (10.1.21)
D
is the double-layer integral. The kernel of the integral equation (10.1.20) is the Greens function,
G(x0 , x). Inspecting the singular behavior shown in (2.2.10), we nd that, if the normal unit vector
varies continuously over the boundary, D, the integral equation is weakly singular. Further properties
of the solution will be discussed in Section 10.7.
n = n u = ux nx + u n . (10.1.23)
Applying the boundary-integral equation (10.1.9) with the free-space Greens function given by
G(x0 , x) = 1/(4|x x0 |), and expressing the dierential surface area as dS = d dl, we obtain
2
1 1
(x0 ) = d [n u](l) (l) dl
2 C 0 |x0 x|
P V 2
1 x0 x
+ n(x) d (l) (l) dl, (10.1.24)
2 C 0 |x x| 3
where l is the arc length along the trace of the boundary in the xy plane, C.
we obtain
I(x0 , l) = I10 (x, , 0 ),
(10.1.27)
K(x0 , l) = (xnx n )I10 (x, , 0 ) + 0 n I30 (x, , 0 ),
where x = x0 x and the integrals Inm are dened in (2.12.10) and (2.12.11). The integral equation
(10.1.24) takes the one-dimensional form
PV
1 1
(x0 ) = I(x0 , l)[n u](l) (l) dl + K(x0 , l) (l) (l) dl. (10.1.28)
2 C 2 C
Integral equations of the rst or second kind with respect to arc length, l, can be derived depending
on the imposed boundary conditions as discussed previously in this section.
Problems
where c = 1 or 0 when the point x0 is located inside or outside selected control volume, c = 12 when
x0 lies on a smooth patch of D, and c = /(2) when x0 lies at corner or wedge, where is the
solid angle subtended by the corner facing the control volume. In all cases, the normal unit vector,
n, points into the control volume.
PV
x x0 x x0 PV
x 3
n(x) dS(x) = (x x0 ) n(x) dS(x) x0 n(x) dS(x).
Sphere r Sphere r3 Sphere r
(10.1.30)
To compute the rst integral on the right-hand side, we introduce Cartesian coordinates with origin
at the center of the sphere and the x axis passing through the point x0 . Owing to symmetry, the y
and z components of the integral are identically zero. Carrying out the integration in terms of the
azimuthal angle, , show that the x component of the integral is equal to 4 3 x0 . Use (2.2.15) to
show that the second integral on the right-hand side of (10.1.30) is equal to 2. Combining these
results, conrm identity (10.1.19).
10.2 Two-dimensional ow
The results of Section 10.1 for three-dimensional ow can be adapted readily to two-dimensional
interior or exterior ow. The counterpart of the boundary-integral equation (10.1.9) is
PV
(x0 ) = 2 G(x0 , x) n(x) (x) dl(x) + 2 (x) n(x) G(x0 , x) dl(x), (10.2.1)
C C
where G(x,0 , x) is a two-dimensional Greens function, C is the boundary of a selected control area,
l is the arc length along C, and n is the unit vector pointing into the control area. In the remainder
of this section, we illustrate the application of (10.2.1) to several types of ow.
1
the free-space Greens function, G(x0 , x) = 2 ln(|x x0 |/a), we obtain an integral representation
for the single-valued potential ,
1 r 1 x x0
(x0 ) = (x0 ) ln n(x) v(x) dl(x) (x) n(x) dl(x), (10.2.2)
2 B a 2 B r2
where r = |x x0 |, a is a chosen length, and v is the velocity due to a point vortex whose strength
is equal to the circulation of the ow around the body, located at an arbitrary point inside the body.
Letting the point x0 approach the body, we derive an integral equation of the second kind for ,
1 PV x x0 1 r
(x0 ) = (x) n(x) 2
dl(x) + 2 (x 0 ) ln n(x) v(x) dl(x). (10.2.3)
B r B a
In Section 10.7, we will show that this integral equation admits a unique solution.
where W represents the four side and bottom walls, S represents the free surface, and a is a chosen
length. An ecient computational procedure for describing the evolution of the free surface subject
to an initial condition involves the following steps (e.g., [275]):
1. Trace the free surface with a collection of marker points and assign to the marker points initial
values for .
2. Solve the integral equation (10.2.4) for the normal velocity component over the free surface, S,
and for the potential over the walls, W . In this case, (10.2.4) is an integral equation of mixed
kind.
3. Dierentiate the potential along the free surface to compute the tangential component of the
uid velocity.
4. Advance the position of the marker points over a small time interval with the normal and
tangential uid velocities. To prevent clustering, it is preferable to move the marker points
with the normal velocity alone. This modication provides us with a kinematically consistent
description of the free surface.
5. At the new position, update the marker-point potential using Bernoullis equation. Applying
equation (7.1.3) at the free surface, we nd that, when the marker points move with the uid
velocity, the rate of change of the potential following their motion is given by
D 1 p
= + u = + |u|2 = |u|2 + g x + c(t). (10.2.5)
Dt t t 2
812 Introduction to Theoretical and Computational Fluid Dynamics
(a) (b)
1.2
y 1
0.8
y/L
0.6
W g 0.4
n
n n W 0.2
n
x 0
W a
0.6 0.4 0.2 0 0.2 0.4 0.6
x/L
Figure 10.2.1 (a) Illustration of liquid sloshing inside a two-dimensional tank. The evolution of the
free surface is computed by a boundary-integral method while the free-surface potential is updated
using Bernoullis equation. (b) Results of numerical simulations when the tank starts moving along
the x axis to the left and then it stops. The initial position of the free surface is shown as a dashed
line, and the interfacial marker points are shown as dots. Proles are shown at equal time intervals.
Note the deformation of the free surface and subsequent sloshing of the liquid.
If the marker points move with the normal velocity, we use the alternative evolution equation
d 1 p
= + (u n) n + |u|2 = (u n)2 |u|2 + g x + c(t). (10.2.6)
dt t 2
The time-dependent integration constant, c(t), in (10.2.5) and (10.2.6) can be set to zero
without any consequences on the motion of the uid. The pressure on the right-hand sides
derives from the dynamic boundary condition at the free surface. For example, in the case of
a free surface with surface tension , we set p = patm + , where is the curvature of the
free surface in the xy plane and patm is the ambient atmospheric pressure.
6. Return to Step 2 and repeat the cycle.
If the tank undergoes vertical or horizontal acceleration with velocity V(t), we work in a frame of
reference where the tank appears to be stationary. The acceleration of gravity, g, on the right-hand
side of (10.2.5) and (10.2.6) is then replaced by g dV/dt to account for the ctitious inertial
acceleration force.
S
P1 P
n 2
B
w
x
Figure 10.2.2 Flow due to the propagation of free-surface gravity waves over a at bottom.
constructed by the method of images. The image system consists of four periodic arrays of point
sinks with equal strengths. Using the expression for the potential due to a periodic array of point
sources shown in Table 7.8.1, we obtain
1
4
G(x, x0 ) = ln cosh[k(y ym )] cos[k(x xm )] , (10.2.7)
4 m=1
Using the Greens function (10.2.7) allows us to eliminate the double-layer potential over the walls
and provides us with a Fredholm integral equation of the rst kind for the normal component of the
velocity over the free surface alone.
Stages in the deformation of the free surface computed using the numerical method described
in this section in conjunction with the boundary-element method described in Section 10.3.1 are
shown in Figure 10.2.1(b) in the absence of surface tension. The tank starts moving along the x axis
with a constant acceleration for a certain time period, and then it stops. The numerical simulation
illustrates the initial deformation of the free surface and subsequent sloshing of the uid inside the
tank.
To simplify the boundary-integral formulation, we use a Greens function of the second kind
(Neumann function) that is repeated in the x direction with period a. The normal derivative of the
Greens function vanishes at the bottom, y = w. Using the method of images, we nd that
814 Introduction to Theoretical and Computational Fluid Dynamics
where k = 2/a is the wave number. Next, we identify our control area with one period of the ow
enclosed by the contours S, P1 , P2 , and B, as shown in Figure 10.2.2, and derive the integral equation
(10.2.1). The double-layer integral is identically zero over B because the normal derivative of the
Greens function vanishes when the point x is on B. The single-layer integral is also identically zero
over B because of the no-penetration boundary condition. The integrals over the vertical segments
P1 and P2 cancel each other because is periodic and the corresponding normal unit vectors point
into opposite directions. These simplications allow us to reduce the contour, C, to one period of
the free surface alone. The evolution of the ow can be computed using the numerical procedure for
the problem of uid sloshing discussed in Section 10.2.2.
Problems
approximated with a truncated polynomial expansion in terms of properly dened surface coordi-
nates over each element. The union of the local expansions provides us with an approximate solution
that is not necessarily continuous across the element edges or nodes.
The union of all element expansions contains M unknown coecients. When the unknown
function is approximated with a constant function over each element, M is equal to the number of
elements, NE . When higher-order expansions are employed, M is greater than NE . If the order of
the expansion of the unknown function is identical to that of the Cartesian coordinates of a point over
an element, we obtain an isoparametric representation. If it is higher, we obtain a superparametric
representation.
Global expansions
To facilitate the logistics of the computation, it is convenient to identify the coecients of the
expansion over each element with the values of the unknown function at corresponding element
nodes. The functions multiplying these coecients are local or element basis functions. Compiling
the local expansions and collecting the coecients corresponding to shared element nodes, we obtain
an expansion in terms of N global basis functions. It should be noted that, although supercially
innocuous, certain types of local expansions may introduce instabilities in the numerical solution,
as discussed later in this section.
Inuence coecients
The union of the local expansions of the unknown function is now substituted into the integral
equation, and the N coecients are extracted from the single- and double-layer potentials to yield
an algebraic equation,
F(c1 , c2 , . . . , cN ) = 0, (10.3.1)
where ci are the coecients of the global expansion. Since F is a linear function of ci , we can write
F(c1 , c2 , . . . , cN ) = A1 c1 + A2 c2 + + AN cN + B, (10.3.2)
where
The inuence coecients, Ai , are integrals of the single- or double-layer potential over selected
boundary elements. In certain cases, it is expedient to compute the inuence coecients by numerical
dierentiation,
selected collocation points over the boundary. In the method of weighted residuals, the discretized
integral equation is multiplied by a set of N chosen weighting functions and the product is integrated
over the boundary. The goal in both cases is to produce a system of N linear algebraic equations
for the N coecients of the global expansion.
Dierent choices for the weighting functions in the method of weighted residuals produce
dierent schemes. Identifying the weighting functions with the global basis functions, we obtain
Galerkins method. Identifying the weighting functions with delta functions whose singular points
are placed at selected locations over the boundary, we recover the collocation method. Because of
increased computational requirements, Galerkins method is suitable only for problems with notable
geometrical simplicity [17].
In solving the nal system of linear equations for the coecients of the local expansions, issues
of well-posedness, existence, and uniqueness of the solution of the integral equation arise. Ill-posed
integral equations and integral equations with no solution or multiple solutions lead to ill-conditioned
linear systems with sensitive, oscillatory, or nonconvergent solutions. Integral equations of the rst
kind may be susceptible to oscillations due to numerical error (e.g., [104]). Fortunately, experience
has shown that integral equations of the rst kind that arise from boundary-integral representations
are well-behaved.
n
M
xi
N
1 2 3 j j+1
Ej
where summation is implied over the repeated index, j. The inuence coecients Fj (x0 ) and Gj (x0 )
are dened as
1 P V x x0 1 r
Fj (x0 ) = 2
n(x) dl(x), G j (x 0 ) = ln dl(x). (10.3.7)
Ej r Ej a
The principal-value integral is relevant only when the evaluation point, x0 , lies on the element Ej .
i = Fj (xM M M
i ) j + 2 (xi ) + Gj (xi ) vj , (10.3.8)
where ij is Kroneckers delta. The solution can be found by the method of Gauss elimination
discussed in Section B.1.2, Appendix B.
Consider the diagonal coecients, Fi (xMi ), where summation is not implied over i. When the
elements are straight segments and the integration point, x, lies at the ith element, the distance x
xMi is perpendicular to the normal unit vector, the corresponding integrand in (10.3.7) is identically
zero, and Fi (xMi ) = 0. When the elements are curved segments such as circular arcs, we apply a
Taylor series expansion about the midpoint to nd that Fi (xM i ) 0 to leading-order approximation.
These simplications explain why the principal value of the double-layer integral is used instead of
the limit of the double-layer integral as the evaluation point approaches a boundary.
818 Introduction to Theoretical and Computational Fluid Dynamics
When i = j, the integral Gi (xM i ) exhibits a logarithmic singularity, which can be integrated
analytically over a straight segment to yield
1 li
i ) = li ln
Gi (xM 1 . (10.3.10)
2a
where li = |xi+1 xi | is the length of the ith element. A similar analytical integration is possible
when the boundary elements are circular arcs (Problem 10.3.1).
Problems
Computer Problems
Exterior ow
Considering an exterior ow that decays far from an interior boundary, D, we introduce the harmonic
potential in the interior of D so that = over D. Applying the reciprocal relation (2.3.2) for
at a point x0 in the exterior of D and subtracting (2.3.2) from (2.3.8), we obtain a generalized
representation in terms of a single-layer potential alone,
(x0 ) = G(x0 , x) q(x) dS(x), (10.4.1)
D
where q = n ( ) is the distribution density and n is the normal unit vector pointing into the
exterior of D. Working in a similar fashion, we introduce a harmonic potential in the interior of
D so that n = n over D, and derive the generalized double-layer representation
(x0 ) = (x) n(x) G(x0 , x) dS(x). (10.4.2)
D
where = is the distribution density and n is the normal unit vector pointing into the exterior of
D. The derivation of (10.4.1) and (10.4.2) hinges on the existence of the complementary potentials
and . In Section 10.7, we will see that it is always possible to nd the internal potential, , and thus
represent any exterior ow in terms of a single-layer potential alone. In contrast, a complementary
potential exists only when the ow rate of across the boundary D is zero,
n dS = 0. (10.4.3)
D
Consequently, only a limited class of exterior ows that satisfy (10.4.3) can be represented in terms
of a double-layer potential alone.
Interior ow
In the case of interior ow, we introduce a harmonic potential in the exterior of D so that =
over D, and and decays at innity. Repeating the preceding analysis, we derive the single-layer
representation (10.4.1). Working in a similar fashion, we introduce the harmonic potential in the
exterior of D so that n = n over D and decays at innity, and derive the double-layer
representation (10.4.2). Analysis undertaken in Section 10.7 guarantees the existence of both and
, and thus ensures that any nonsingular interior ow can be represented in terms of a single-layer
or double-layer potential alone.
820 Introduction to Theoretical and Computational Fluid Dynamics
Problems
The single-layer potential is the harmonic function due to a distribution of Greens functions over a
chosen surface, D,
(x0 ) = G(x0 , x) q(x) dS(x), (10.5.1)
D
where q is the surface density of the distribution, as shown in Figure 10.5.1. Physically, the single-
layer potential can be identied with the temperature eld due to a distribution of point sources of
heat or with the velocity potential due to a distribution of point sinks of mass with density q over
the surface D.
The single-layer potential is continuous throughout the domain of ow as well as across the
distribution domain, D. When G is a Greens function of the rst kind, vanishes over the associated
boundary SB . When G is a Greens function of the second kind (Neumann function), the normal
derivative of vanishes over SB . Using the general properties of the Greens functions we nd that,
far from D, decays at a rate that is equal to, or faster than 1/d, where d is the distance from the
origin. The precise rate of decay depends on the geometry of SB .
Integral identity
For simplicity, we assume that the distribution domain D is a closed surface with a continuously
varying normal unit vector, N, pointing outward from D, as shown in Figure 10.5.1. The integral
10.5 The single-layer potential 821
N S1
D
S2
N
N
N
S4
S
3
SB
Figure 10.5.1 Distribution domain of a hydrodynamic potential and illustration of several closed
surfaces in a ow.
Flow rates
Integrating the normal derivative of (10.5.1) over a surface that encloses the distribution domain,
D, but no other boundaries, such as the surface S1 illustrated in Figure 10.5.1, we nd that the
volumetric ow rate through the surface is
which shows that Q1 vanishes only when the total strength of the distribution dened as the integral
in (10.5.4) is zero. Working in a similar manner, we nd that the volumetric ow rate across a closed
822 Introduction to Theoretical and Computational Fluid Dynamics
uid surface that does not enclose any boundaries, such as the surface S2 illustrated in Figure 10.5.1,
is zero (Problem 10.5.1).
where the subscript 0 of the del operator designates dierentiation with respect to x0 . By denition,
when G is a Greens function of the second kind, the normal derivative of vanishes over the
associated boundary, SB .
Normal derivative
To investigate the behavior of the normal derivative across the distribution domain, D, we take the
limit of (10.5.5) as the point x0 approaches D, introduce the free-space Greens function G F S =
1/(4|x x0 |), note that 0 G F S = G F S , and write the identity
n(x0 ) 0 (x0 ) = J(x0 ) + N(x) 0 G F S (x0 , x) q(x0 ) dS(x), (10.5.6)
D
where
J(x0 ) N(x0 ) 0 G(x0 , x) q(x) N(x) 0 G F S (x0 , x) q(x0 ) dS(x). (10.5.7)
D
Because of the weak singularity of the integrand in (10.5.7), the integral is continuous across D.
Considering the second term on the right-hand side of (10.5.6), we use (10.5.2) to obtain
+
N(x0 ) 0 (x0 ) = J(x0 ), N(x0 ) 0 (x0 ) = J(x0 ) + q(x0 ), (10.5.8)
where the plus superscript denotes the exterior side of D indicated by the direction of N, and
the minus superscript denotes the interior side. Equations (10.5.8) demonstrate that the normal
derivative of the single-layer potential suers a discontinuity of magnitude q across the distribution
domain D,
+
N(x0 ) 0 (x0 ) N(x0 ) 0 (x0 ) = q(x0 ). (10.5.9)
Dirichlet problem
In the Dirichlet problem, the distribution of the single-layer potential is specied over D, and the
density of the distribution, q, is required. Applying (10.5.1) at a point x0 on D provides us with a
Fredholm integral equation of the rst kind for q. Numerical evidence suggests that, in general, the
solution of this equation is unique and the computational problem is tractable by standard numerical
methods, including the panel method discussed in Section 10.3.
Neumann problem
In the Neumann problem, the normal derivative of is specied over the external or internal side
of D, and the distribution density, q, is required. Using (10.5.11) to express the normal derivative
in terms of the principal value, we obtain a Fredholm integral equation of the second kind for q,
P V
q(x0 ) = 2 N(x0 ) 0 G(x0 , x) q(x) dS(x) 2 N(x0 ) 0 (x0 ) , (10.5.13)
D
where the plus superscript corresponds to the exterior side of D and the minus superscript corre-
sponds to the interior side of D. It is convenient to introduce the linear integral operator
P V
O q(x0 ) q(x0 ) 2 N(x0 ) 0 G(x0 , x) q(x) dS(x), (10.5.14)
D
In Section 10.7, we will nd that the integral equation (10.5.13) has a unique solution in
the case of exterior ow, but either no solution or an innite number of solutions in the case of
interior ow. In the case of interior ow, an innite number of solutions exist only when the
constraint (10.7.25) is fullled. Fortunately, the satisfaction of this constraint is guaranteed by
the continuity equation for any ow that does not contain point sources and point sinks with a
nonzero net discharge. We conclude that the single-layer representation is capable of representing
any external ow and any nonsingular internal ow.
According to our previous discussion, the integral equation (10.5.17) has a unique solution for any
incident ow.
Problems
Computer Problem
Jump condition
The double-layer potential is continuous throughout the domain of ow, but suers a discontinuity
across the distribution domain, D. To assess the discontinuity, we write
lim (x0 ) = lim I(x0 ) + q(x0 ) N(x) G(x0 , x) dS(x), (10.6.2)
x0 D x0 D D
where
I(x0 ) [ q(x) q(x0 )] N(x) G(x0 , x) dS(x). (10.6.3)
D
826 Introduction to Theoretical and Computational Fluid Dynamics
We note that I(x0 ) is continuous across D and use identity (10.5.2) to evaluate the second integral
on the right-hand side of (10.6.2), nding
lim (x0 ) + (x0 ) = I(x0 ), lim (x0 ) (x0 ) = I(x0 ) q(x0 ), (10.6.4)
x0 D+ x0 D
where D+ and D signify that x0 approaches D from the external side, indicated by the direction
of the normal unit vector, N, or from the internal side. When the point x0 is located precisely on
D, we obtain the principal value (PV) of the double-layer potential,
1
P V (x0 ) = I(x0 ) q(x0 ). (10.6.5)
2
Comparing the last three equations, we nd that
1 1
+ (x0 ) = P V (x0 ) + q(x0 ), (x0 ) = P V (x0 ) q(x0 ), (10.6.6)
2 2
and
1
P V (x0 ) = + (x0 ) + (x0 ) , (10.6.7)
2
which shows that the double-layer potential suers a discontinuity across the distribution domain,
+ (x0 ) (x0 ) = q(x0 ). (10.6.8)
It is instructive to observe the opposite signs on the right-hand sides of (10.6.6) for the double-layer
potential and (10.5.11) for the normal derivative of the single-layer potential.
Next, we select a point x0 in the exterior of D and note that both G(x0 , x) and q (x) are
nonsingular harmonic functions throughout the interior of D. Applying the reciprocal identity (2.2.1)
with q in place of f and G(x0 , x) in place of g, we restate (10.6.1) in terms of a single-layer potential
with distribution density N q ,
ext (x0 ) = G(x, x0 ) N(x) q (x) dS(x). (10.6.10)
D
where V is the interior of D. This identity demonstrates that the total strength of the single-layer
potential is zero and conrms that, far from D, ext decays like a Greens function dipole.
Taking the limit as the point x0 approaches D from either side and using (10.6.10) and
(10.6.12) in conjunction with (10.6.8), we obtain
+ (x0 ) (x0 ) = G(x, x0 ) N(x) [q q+ ](x) dS(x) = q(x0 ). (10.6.13)
D
To conrm this equation, we write the boundary-integral equation (10.1.9) for q+ with n = N and
for q with n = N. Adding the two equations, we obtain
q+ (x0 ) + q (x0 ) = 2 G(x0 , x) N(x) [q+ q ](x) dS(x)
D
PV
+2 [q+ q ](x) N(x) G(x0 , x) dS(x), (10.6.14)
D
where the evaluation point, x0 , lies on D. Identity (10.6.13) arises by substituting into this equation
the third equation in (10.6.9).
where x0 is on D. Using (10.6.8) to simplify the integrand of the double-layer potential on the right-
hand side, invoking the denition (10.6.1) to identify the double-layer potential with the principal
value of , and taking into account (10.6.7), we nally obtain
G(x0 , x) N(x) [+ ](x) dS(x) = 0, (10.6.16)
D
where x0 is on D. Since this equation holds true for any arbitrary double-layer potential, the
integrand must be identically zero and the normal derivative of the double-layer potential must be
continuous across D.
828 Introduction to Theoretical and Computational Fluid Dynamics
where P = I N N is the tangential projection operator. Equation (10.6.18) states that the
tangential component of the velocity undergoes a discontinuity across D. Referring to (1.13.15), we
nd that the ow can be regarded as though it were induced by a vortex sheet with strength
= N q (10.6.19)
situated over D, in the possible presence of an external or internal boundary, SB . We thus arrive
at the interesting conclusion that the ow due to a double-layer potential can be regarded as the
ow due to a vortex sheet whose strength derives from the tangential derivatives of the double-layer
distribution density, q.
where the subscript 0 indicates dierentiation with respect to x0 . One may readily verify by straight-
forward dierentiation that u = A = .
To derive the velocity eld, we use the BiotSavart integral in (2.10.18), and nd that
u(x0 ) = 0 (x0 ) = G(r) q(x) N(x) dS(x). (10.6.21)
D
where the plus sign indicates the outer side, the minus sign indicates the inner side, and
P V
P V (x0 ) = q(x) N(x) G(x, x0 ) dS(x) (10.6.23)
D
The properties of the integral equations (10.6.25) will be discussed extensively in Section 10.7.
The results will show that the exterior ow problem has a solution only when the ow rate of
the forcing function + across D is zero. When this condition is met, the integral equation has a
multiplicity of solutions that dier by an arbitrary constant. In contrast, the integral equation for
interior ow has a unique solution for any forcing function . Thus, the double-layer representation
is capable of representing any internal ow and a limited class of external ows.
The no-penetration boundary condition requires that N = 0 over the exterior side of B.
To derive an integral equation, we recall that the normal derivative of the double-layer po-
tential is continuous across B. Accordingly, the normal component of the velocity vanishes on the
interior side of B, the internal velocity eld computed from (10.6.26) is zero, and takes a constant
value, c, in the interior of B. Applying (10.6.26) at a point on the internal side of B and using
(10.6.6) provides us with an integral equation of the second kind for q,
P V
q(x0 ) = 2 q(x) N(x) G(x, x0 ) dS(x) + 2 (x0 ) c . (10.6.27)
B
In Section 10.7, we will show that (10.6.27) has a unique solution for any incident ow.
830 Introduction to Theoretical and Computational Fluid Dynamics
1. Describe the surface in terms of a collection of point particles moving with the uid velocity.
2. Assign to the point particles initial values for the potential, .
3. Dierentiate along the free surface to compute the tangential component of the uid velocity.
4. Solve the integral equation (10.6.22) for q for the interior problem using the free-space Greens
function.
5. Compute the vector potential over the free surface using (10.6.20) and then obtain the normal
component of the uid velocity using the equation N u = N ( A). The right-hand side
involves tangential derivatives of A.
6. Advect the marker points with the uid velocity. The tangential and normal components of
the velocity are available from Steps 3 and 5.
7. At the new position, update the potential using (10.2.5).
8. Return to Step 3 and repeat the computation for another step.
+ = (N q) N, (10.6.30)
10.6 The double-layer potential 831
which relates the tangential velocity to the strength of the vortex sheet, , given in (10.6.19).
Substituting = 0 into the left-hand side of (10.6.29) provides us with a vectorial integral
equation for the strength of the vortex sheet [213],
PV
1
(x0 ) N(x0 ) = 0 (x0 ) + (x) G(r) dS(x). (10.6.31)
2 B
In the case of two-dimensional or axisymmetric ow, we obtain an integral equation for the tangential
velocity (see Section 10.7).
10.6.5 Free-space ow
As a last topic of this section, we examine in more detail the case of the free-space Greens function,
G = 1/(4r), where r = |x x0 |. Exploiting the property G(r) = 0 G(r), we restate the
double-layer potential (10.6.1) as
(x0 ) = 0 q(x) N(x) G(x0 , x) dS(x), (10.6.32)
D
where the subscript 0 indicates dierentiation with respect to x0 . Taking the gradient of (10.6.32)
and using the identity
( A) = A + 2 A, (10.6.33)
Now we extend the domain of denition of q inward and outward from D according to (10.6.9),
select a point x0 in the exterior of D, and use the divergence theorem to write
q(x) N(x) G(x0 , x) dS(x) = [ q (x) G(x0 , x) ] dV (x). (10.6.35)
D V
and then
0 ext (x0 ) = 0 [ q (x) G(x0 , x) ] dV (x). (10.6.37)
V
A variation of Stokes theorem for a twice dierentiable vector function, F, states that
F dV = N F dS, (10.6.38)
V D
832 Introduction to Theoretical and Computational Fluid Dynamics
as discussed in Section A.7, Appendix A. Applying (10.6.38) for the volume integral in (10.6.38), we
obtain
0 ext (x0 ) = 0 N(x) q (x) G(x0 , x) dS(x), (10.6.39)
D
Working in a similar manner, we nd that the gradient of the double-layer potential at a point x0
in the interior of D is given by the corresponding expression
0 int (x0 ) = N(x) q+ (x) G(x0 , x) dS(x). (10.6.41)
D
Equations (10.6.40) and (10.6.41) provide us with the gradient of the double-layer potential in terms
of distributions of the Greens function dipoles. Because the corresponding distribution densities,
N(x) q+ (x) and N(x) q (x), are tangential to D, they can be determined from the
distribution of q over D independent of the extensions, q and q+ . To recover (10.6.21), we write
N q = N q+ = N q. (10.6.42)
Problems
Explain why the rst integral on the right-hand side is continuous across D, whereas the second
integral undergoes a discontinuity of magnitude N across D.
Computer Problem
is the known single-layer potential computed on the exterior (+) or interior () side of D. For
convenience, we repeat the integral equation (10.5.13) arising from the single-layer representation
with the Neumann boundary condition,
PV
q(x0 ) = 2 q(x) N(x0 ) 0 G(x0 , x) dS(x) 2 N(x0 ) 0 (x0 ), (10.7.4)
D
and the integral equation (10.6.22) arising from the double-layer representation with the Dirichlet
boundary condition,
PV
q(x0 ) = 2 q(x) N(x) G(x, x0 ) dS(x) 2 (x0 ), (10.7.5)
D
where the plus sign corresponds to the exterior problem and the minus sign corresponds to the
interior problem.
and
PV
(x0 ) = 2 (x) N(x0 ) 0 G(x0 , x) dS(x), (10.7.7)
D
where and are eigenfunctions and an asterisk denotes the complex conjugate. In the standard
terminology of integral equation theory, is a complex eigenvalue of the double-layer operator
PV
L (x0 ) 2 (x) N(x) G(x, x0 ) dS(x), (10.7.8)
D
Note that the adjoint operator L arises from L, and vice versa, by switching the arguments x0 and
x of the kernel. Using these denitions, equations (10.7.6) and (10.7.7) are written concisely as
Self-adjoint operators
It can be shown by direct substitution that any two complex functions, q1 and q2 , satisfy the
distinguishing property of adjoint operators,
L[q1 ], q2 = q1 , L [q2 ] , (10.7.11)
Spectrum
The homogeneous equations (10.7.10) have complex conjugate eigenvalues and the same number of
eigensolutions. The set of all eigenvalues is the spectrum of the double-layer operator or its adjoint.
The eigenvalues can be numerable or innumerable. However, an innite number of eigenvalues may
not accumulate at any point in the complex plane, except at innity. The spectral radius of L or
L is dened as the magnitude of the eigenvalue, , with the minimum norm.
Fredholm alternative
Fredholms alternative ensures that, if the coecient does not belong to the spectrum of L, then
the inhomogeneous equation
q = L[q] + F (10.7.13)
has a unique solution, where F is a forcing function. If belongs to the spectrum of L, the
inhomogeneous equation has either no solution or an innite number of solutions. An innite number
of solutions exist only if all eigenfunctions of the adjoint double-layer operator, , corresponding to
, are orthogonal to the forcing function F , that is, only when (F, ) = 0, where the projection (, )
is dened in (10.7.12). Similar results apply to the inhomogeneous equation involving the adjoint
operator dened in (10.7.9),
q = L [q] + F, (10.7.14)
Successive substitutions
The inhomogeneous equations (10.7.13) and (10.7.14) can be solved by the method of successive
substitutions, also called the method of xed-point iterations or successive approximations, only if
is less than the spectral radius of the double-layer operator. The method is implemented by guessing
a solution, computing the right-hand side of the integral equation, and then replacing the guessed
with the computed solution. In the extreme case when = 0, the solution is obtained after one
iteration. When the initial guess is set equal to zero, q = 0, we obtain a series of approximations
called the Neumann series of the integral equations.
Combining these expressions to eliminate the integral on the right-hand side, we obtain
1+
= , (10.7.20)
1
where
K+ /K . (10.7.21)
Since is real and positive, is real and lies outside the closed range [1, 1]. However, this deduction
may fail in the following three cases.
10.7 Investigation of integral equations of the second kind 837
Case = 1
When = 1, equation (10.7.19) shows that K+ = 0, which requires that is constant throughout V+ .
Since must vanish at innity, the value of this constant is necessarily zero. Since is continuous
across D, must also vanish throughout V and the normal derivative on the internal side of D
must be zero. Applying (10.7.17) on the internal side of D shows that = 0, which demonstrates
that = 1 cannot be an eigenvalue of the double-layer potential.
Case = 1
When = 1, equation (10.7.19) shows that K = 0, which requires that is constant throughout
V as well as on the interior side of D. Equation (10.7.17) yields the eigenfunction
= N + . (10.7.22)
The constraint (10.7.15) is satised and (10.7.22) constitutes a perfectly acceptable eigenfunction.
Writing the boundary-integral representation for dened in (10.7.16) allows us to identify with
the negative of the normal gradient of a potential function on the external side of D established
when is constant over D.
Although we are not be able to compute the eigensolution explicitly, except for one particular
case discussed in Problem 10.7.2, we can show that it must satisfy the integral constraint
dS = 0. (10.7.23)
D
Since is constant over D, equation (10.7.19) shows that, if (10.7.23) were not true, K+ would
have to vanish, which means that and thus would have to be zero. Equation (10.5.2) shows
that the corresponding eigensolution of (10.7.6) is constant, = c. One important consequence of
the last two equations is that the geometric multiplicity of the eigenvalue = 1 is exactly equal
to one. If this were not true, we would be able to nd a generalized or principal eigensolution
satisfying the inhomogeneous equation = LA [] + . According to Fredholms alternative, for
a solution of this equation to exist, must be orthogonal to . However, since is constant, we
obtain a contradiction with (10.7.23). We conclude that the algebraic and geometric multiplicities
of the eigenvalue = 1 are equal to one.
Summary
We have found that (a) the eigenvalues of the double-layer potential are real and lie outside the
semi-closed range (1, 1], and (b) = 1 is a marginal eigenvalue for any shape, D. The spectral
radius of the double-layer operator or its adjoint is equal to unity. Based on these results, we make
the following deductions regarding the properties of the integral equations for interior or exterior
ow:
838 Introduction to Theoretical and Computational Fluid Dynamics
Equations (10.7.2) and (10.7.4) for the exterior problem and equation (10.7.5) for the interior
problem have a unique solution for any boundary shape, D.
Equation (10.7.1) for the interior problem has either no solution or a multiplicity of solutions
that dier by an arbitrary constant. According to Fredholms alternative, multiple solutions
exist only if the projection of the forcing function onto the adjoint eigensolution , given in
(10.7.22), is zero. However, this is true for any ow that does not contain point sources and
point sinks. To see this, we denote the single-layer potential as I S and formulate its projection
on to obtain
, I S
= (x0 ) G(x0 , x) N(x) (x) dS(x) dS(x0 ) (10.7.24)
D
D
= N(x) (x) (x0 ) G(x0 , x) dS(x) dS(x0 ) = N(x) (x) c dS(x0 ) = 0,
D D D
This requires that the ow rate of the eld represented by the forcing function across D is
zero.
Equation (10.7.5) for the exterior problem has either no solution or a multiplicity of solu-
tions that dier by an arbitrary constant According to Fredholms alternative, an innity of
solutions exist only if the projection of the forcing function onto the adjoint eigensolution
corresponding to = 1 is zero,
+ dS = c N + dS = 0, (10.7.26)
D D
where c is the constant value of the potential over D dened in (10.7.16). The second expression
in (10.7.26) arises from the reciprocal theorem for harmonic functions discussed in Section 2.2.
None of the equations discussed can be solved by the method of successive substitutions.
where = (1 )/(1 + ) and r is the distance from the center of the sphere. The problem is
reduced to computing nontrivial interior and exterior harmonic potentials + and subject to
(10.7.27).
where Sn are surface spherical harmonics and bn , cn are constants (e.g., [220]). Enforcing the
matching conditions (10.7.28), we nd that
bn = cn , bn (n + 1) = ncn , (10.7.29)
which requires that = (n + 1)/n and b = 2n 1. Thus, the spectrum of the double-layer
operator for the sphere is the set of all negative integers.
Problems
The occurrence of multiple solutions and our inability to compute a solution by the method
of successive substitutions can be circumvented by replacing the original integral equations with
840 Introduction to Theoretical and Computational Fluid Dynamics
regularized integral equations that have a unique solution amenable to the method of successive
substitutions. The solution of the regularized equations or a modication of the solution also satisfy
the original integral equations.
where y and x are two arbitrary vectors and the parentheses denote the inner vector product.
Consequently, A and B have complex conjugate eigenvalues and the same number of eigenvectors and
generalized eigenvectors or principal vectors (e.g., [45, 318]). Each eigenvector of A corresponding
to a particular eigenvalue, 1 , is orthogonal to each eigenvector of B corresponding to a dierent
eigenvalue, 2 .
A u = u B = 1 u, B v = v A = 1 v. (10.8.2)
shares its eigenvalues with A, except that 1 has been replaced by zero, where w is an arbitrary
vector satisfying w u = 1 (e.g., [45], p. 361; [429], p. 596). Wielandts theorem also states that the
modied matrix
(1)
Bij Bij 1 vi zj (10.8.4)
shares its eigenvalues with B, except that 1 has been replaced by zero, where z is an arbitrary
vector satisfying z v = 1. The adjoint of B(1) , which is equal to its transpose,
(2)
Aij Aij 1 zi vj , (10.8.5)
shares its eigenvalues with the adjoint of B, which is the matrix A, except that 1 has been replaced
by zero.
The process of modifying a matrix to reduce the number of nonzero eigenvalues is called eigen-
value spectrum deation. Single deation eliminates one eigenvalue and leaves all other eigenvalues
and adjoint eigenvectors unaected, but alters the corresponding eigenvectors (Problem 10.8.1). To
implement eigenvalue deation, we must have available one eigenvalue and either the corresponding
eigenvector, u, or the adjoint eigenvector, v.
10.8 Regularization of integral equations of the second kind 841
which was shown to have a unique solution. Unfortunately, because the spectral radius of the adjoint
double-layer operator is equal to unity, the solution cannot be computed by the method of successive
substitutions.
Let us discretize the domain D into N boundary elements and approximate the unknown
function q(x) with a constant function over each element. In the numerical representation, equation
(10.8.6) takes the discrete form
x = A x + b, (10.8.7)
where A is an inuence matrix, the vector x contains the constant values of q over the boundary
elements, and the vector b contains the values of the forcing function at the midpoints of the
boundary elements. Instead of solving (10.8.7), we propose to solve the regularized equation
x = A(2) x + c, (10.8.8)
To ensure that the solution of (10.8.8) also satises (10.8.7), we require that
c = b + 1 z (v x). (10.8.9)
1
vx= v b. (10.8.10)
1 1
Substituting this expression into the last term on the right-hand side of (10.8.9), we obtain
1
c=b+ z (v b). (10.8.11)
1 1
x = A(2) x + I + z v b, (10.8.12)
1 1
where I is the identity matrix. The solution of (10.8.12) is identical to that of (10.8.7).
Identifying 1 = 1 with the eigenvalue of A with the maximum norm corresponding to the
eigenvalue of the adjoint double-layer operator with the minimum norm, = 1, and noting that
842 Introduction to Theoretical and Computational Fluid Dynamics
1 is not a eigenvalue of A(2) , we nd that the spectral radius of A(2) is less than unity. Equation
(10.8.12) may then be solved by the method of successive substitutions.
The generalized homogeneous equation of (10.8.13) arises by discarding the forcing function
expressed by the last two terms on the right-hand side, and multiplying the remaining terms on the
right-hand side by the eigenvalue , yielding
P V 1
(x0 ) = 2 (x) N(x0 ) 0 G(x0 , x) dS(x) + dS . (10.8.15)
D 2SD D
Integrating (10.8.15) over D and using (10.5.2), we nd that the last integral on the right-hand
side is zero, which allows us to identify the eigenvalues and eigenfunctions of (10.8.15) with those of
(10.7.7). However, since criterion (10.7.23) is violated, = 1 is no longer an eigenvalue and the
spectral radius of the integral operator on the right-hand side of (10.8.13) is higher than unity. This
guarantees that equation (10.8.13) can be solved by the desirable method of successive approxima-
tions.
where SDk is the surface area of the kth boundary. We observe that each surface undergoes an
individual deation.
10.8 Regularization of integral equations of the second kind 843
In Section 10.7, we demonstrated that this integral equation admits a unique solution. However,
because the spectral radius of the adjoint double-layer operator is equal to unity, the solution cannot
be found by the method of successive substitutions.
Working as in Section 10.8.2, we discretize the domain D into N boundary elements and
assume that q is constant over each element to obtain the linear algebraic system (10.8.7) with a
dierent inuence matrix, A, and known vector, b. Instead of solving this system, we propose to
solve the alternative system
x = A(1) x + c, (10.8.18)
where the matrix A(1) is dened in (10.8.3). To ensure that the solution of the altered system
(10.8.18) also satises the original system x = A x + b, we require that
c = b + 1 u (w x). (10.8.19)
y = A(1) y + b, (10.8.20)
where
y I 1 u w x. (10.8.21)
Projecting (10.8.21) onto w and recalling that wu = 1, we obtain wy = (11 ) wx. Substituting
this expression into (10.8.21) and rearranging, we obtain
1
x= I+ u w y. (10.8.22)
1 1
Instead of solving the original equation x = A x + b for x, we prefer to solve equation (10.8.20) for
y and then use (10.8.22) to recover x.
Identifying 1 = 1 with the eigenvalue of A with the maximum norm corresponding to the
eigenvalue = 1 of the double-layer operator with the minimum norm, and noting that 1 is not
a eigenvalue of A(1) , we nd that the spectral radius of A(1) is less than unity. Equation (10.8.20)
may then be solved by the method of successive substitutions.
and
1
q =Q Q dS. (10.8.24)
2SD D
The numerical procedure involves solving (10.8.23) for Q and then using (10.8.24) to recover q.
and
1
+ = dS, (10.8.26)
2SD D
where I+
S
is the single-layer potential on the right-hand side of (10.7.1).
and
1
+ = dS, (10.8.28)
2SDk Dk
where SDk is the surface area of the kth boundary. We observe that each surface undergoes an
individual deation.
10.8 Regularization of integral equations of the second kind 845
x = A(1) x + c, (10.8.29)
where the inuence matrix A is the same as that appearing in (10.8.18) but the vector c is dierent.
Working in the familiar way, we replace (10.8.29) with
y = A(1) y + b, (10.8.30)
where
y = (I + 1 u w) x. (10.8.31)
To perform eigenvalue deation, we must apply (10.8.31) with 1 = 1. However, inverting this
equation to recover x is prohibited by the singular nature of the matrix Iuw (Problem 10.8.3). We
conclude that (10.7.5) cannot be regularized for exterior ow. Similar diculties are encountered
when we attempt to deate equations (10.7.1) and (10.7.5) for interior ow.
Range completion
To complete the decient range of the double-layer operator in the case of exterior ow, we introduce
a composite representation,
(x) = q(x) N(x) G(x0 , x) dS(x) + (x0 ), (10.8.32)
D
where the harmonic function contributes a nite ow rate across D. Applying (10.8.32) at the
exterior side of D and using (10.6.6), we obtain an integral equation of the second kind for q,
PV
q(x0 ) = 2 q(x) N(x) G(x0 , x) dS(x) 2 (x0 ) + 2 + (x0 ). (10.8.33)
D
846 Introduction to Theoretical and Computational Fluid Dynamics
Equation (10.8.32) is the completed double-layer representation, and equation (10.8.33) is the asso-
ciated integral equation.
where c is a constant determined by the distribution density, q. To remove this dependency while
preserving linearity, we set
c=A q(x) dS(x), (10.8.35)
D
where A is a new constant independent of q. With these choices, the integral equation (10.8.33)
becomes
P V
q(x0 ) = 2 q(x) N(x) G(x0 , x) dS(x) 2A G(x0 , xs ) q(x) dS(x) + 2 + (x0 ). (10.8.36)
D D
To show that this equation has a unique solution, we consider the corresponding adjoint homogeneous
equation
PV
(x0 ) = 2 (x) N(x0 ) 0 G(x0 , x) dS(x) 2A G(x0 , xs ) (x) dS(x), (10.8.37)
D D
where is an eigenfunction. Integrating (10.8.37) over D and using (10.5.2), we nd that the last
integral is identically zero. Equation (10.8.37) then reduces to (10.7.7) with = 1, in violation
of (10.7.23). Consequently, the integral equation (10.8.37) has only the trivial null solution for any
value of A.
Problems
To ensure the linearity of the integral equation, we require that is a linear function of q, setting
= Aq, where A is a constant. Show that, with this choice, the solution of (10.8.33) is unique for
any positive A.
q = O[q] + F, (10.9.1)
where O is a compact linear integral operator dened over a specied line or surface, D, q is an
unknown density distribution, and F is a given forcing function. We proceed by discretizing D into
N boundary elements and approximate the integral operator O[q] with an integration quadrature
over each element to obtain
n=1,...,N
k=1,...,K
where {xn,k } is the union of the K quadrature base points over each element and A is an inuence
matrix (see Section B.6, Appendix B).
Nystrom method
In Nystroms method, equation (10.9.2) is enforced at the integration quadrature points over each
element to yield a system of N K linear algebraic equations for the unknown values, q(xn,k ) [15].
Jacobis method for linear systems is the discrete implementation of the method of successive sub-
stitutions involving three steps: guess the values q(xnk ), compute the right-hand side of (10.9.2) at
all quadrature base points, and replace the assumed with the computed values, q new (xnk ). Other
iterative schemes based on generalized conjugate gradient methods are discussed in Section B.1,
Appendix B.
848 Introduction to Theoretical and Computational Fluid Dynamics
To reduce the computational cost, it is helpful to compute the inuence matrix An,k once at
the outset and then store and recall its elements during the iterations. When N is on the order of
a few hundred and the unknown function is assumed to be a constant or linear function over each
element, the computer memory requirements are modest. When the number of boundary elements
is large or high-order approximations are employed, the size of the matrix An,k may exceed the
available memory, and this necessitates recomputing some or all of the elements, An,k , before each
iteration.
Element expansions
In a more general approach capturing the spirit of the spectral-element method, the unknown density,
q, is expanded into a polynomial or trigonometric series with respect to properly dened surface
variables over each element, and the coecients of the expansion are identied with the values of q
at corresponding element nodes. Initial values of q are assigned, the polynomial approximations are
evaluated over all elements, and the integral equation is enforced at the nodes. Finally, the double-
layer integral is computed and added to the forcing function F , thereby producing new values for
q at the nodes. The procedure is repeated until the values of q at all nodes change by less than
a preset minimum after one iteration. This method has the advantage of directly producing the
value of q at the nodes, and thus circumventing the need for further interpolation. This feature is
especially desirable when q represents a primary variable, such as the velocity.
Convergence
It is important to emphasize that the iterative procedures described in this section converge only
when the magnitude of the coecient is less than the spectral radius of the operator O, and the
integration domain, D, is a smooth Lyapunov surface. If D contains sharp edges or corners, the
iterations may diverge. In order to compute a solution, equation (10.9.2) must be restated in the
standard form of a linear system of equations and the solution must be computed by direct inversion.
In the context of parallel computation, we may distinguish two classes of problems according
to the topology of the ow domain. The rst class includes problems in domains that are bounded
by distinct closed surfaces representing, for example, the boundaries of a collection of suspended
particles. The second class includes problems in domains with contiguous but complex boundaries
representing, for example, the surface of a vehicle or aircraft.
M
q(x) = Om [q] + F (x). (10.9.3)
m=1
We then guess the distribution of q over each boundary, call it q (0) , and iterate individually on each
surface using the method of successive substitutions, where each boundary is assigned to a dierent
processor. Over the ith surface, we iterate based on the equation
M
q(x) = Oi [q] + G(x) where G(x) = Om [q] + F (x), (10.9.4)
m=1
and the prime after the sum indicates that the term m = i is excluded from the summation. These
local iterations are guaranteed to converge as long as || is less than the spectral radius of the
corresponding operators. After a number of local iterations have been carried out, the initial guess
q (0) is updated across all processors, a second global iteration is carried out, and the procedure is
repeated until a converged solution has been found. The frequency and protocol of communication
between the processors play an important role in determining the overall eciency of the method
(e.g., [102, 138]).
Computer Problem
The computational advantages of vortex methods stem from substantial simplications in the
vorticity transport equation that are not necessarily reected in the equation of motion. In the case
of two-dimensional ow, because of the absence of vortex stretching, the vorticity evolves under the
action of convection and viscous diusion alone. When viscous eects are insignicant, a region of
concentrated vorticity remains compact at all times and the vortices are simply convected by the
ow. These features allow us to reduce the computational domain by considering only those regions
of the ow where the vorticity takes substantial values, above a preset threshold. Another important
advantage of vortex methods is that computing the pressure is not required. The signicance of this
simplication will become evident in Chapter 13 where we discuss the subtleties involved in deriving
and implementing boundary conditions for the pressure from specied conditions for the velocity.
850
11.1 Invariants of vortex motion 851
3. Invert the denition of the vorticity, = u, to obtain the velocity eld at the new time
instant.
4. Return to Step 2 and repeat the computation for another time step.
To obtain the velocity eld from the vorticity eld in Step 2, we may use the BiotSavart integral
discussed in Sections 2.102.13. This approach is particularly eective in the case of inviscid or
slightly viscous ows considered in this chapter. Other methods discussed in Chapter 13 in the
context of the nite-dierence formulation involve expressing the velocity in terms of a solenoidal
vector potential, A, so that u = A and A = 0, and then solving a Poisson equation for A,
2 A = , or inverting Laplaces equation for the velocity 2 u = . These methods are
appropriate for viscous ows the vorticity is spread all over the domain of ow.
In this chapter, we outline and discuss the fundamental principles, governing equations, and
computational algorithms underlying a general class of vortex methods for two-dimensional, axisym-
metric, and three-dimensional ow. Further information on specic topics can be found in review
articles by Clements & Maull [88], Saman & Baker [358], Leonard [233, 234], Pullin [330], Puckett
[328], and in monographs by Newton [278] and Saman [357].
11.1.1 Three-dimensional ow
In Section 2.10, we saw that, if that the vorticity of a three-dimensional ow decays faster than
1/d3 , where d is the distance from the origin, the integral of the vorticity over the domain of ow is
zero. Thus, the total vorticity,
dV = 0 (11.1.1)
F low
(e.g., [24], Section 7.2). The angular impulse is proportional to the rst moment of the distributed
linear impulse expressed by the integrand dening P. To show that P and A remain constant in
time, we take the time derivative of expressions (11.1.2) and invoke the vorticity transport equation
for a homogeneous uid (Problem 11.1.1(a)).
852 Introduction to Theoretical and Computational Fluid Dynamics
Kinetic energy
In the absence of viscous dissipation, the total kinetic energy of the uid is conserved. Thus,
1 1
K= u u dV = A dV = u (x ) dV (11.1.3)
2 F low 2 F low F low
is a fourth invariant of the motion, where A is a solenoidal vector potential for the velocity dened
by the equation u = A (Section 2.6).
Helicity
Projecting Eulers equation onto the vorticity vector and then integrating the product over the
volume of the ow, we nd that the helicity,
H u dV, (11.1.4)
F low
is a fth invariant of the motion. Physically, the helicity is a measure of the net linkage of the vortex
lines [273].
Viscous eects
In the presence of viscous forces, the total vorticity dened in (11.1.1) remains zero and the linear
and angular impulse dened in (11.1.2) remain invariant during the motion. However, the kinetic
energy of the uid and the helicity change during the evolution of an unsteady viscous ow. It is
not surprising that the kinetic energy decreases due to viscous dissipation. Using the energy balance
(3.4.22) for innite ow, we nd that
dK
= 2 E : E dV = [ + 2 (u L) ] dV, (11.1.5)
dt F low F low
where L is the velocity gradient tensor and E is the rate-of-deformation tensor. Applying the
divergence theorem and simplifying, we nd that
dK
= dV (11.1.6)
dt F low
(Problem 11.1.1(b)). The integral on the right-hand side of (11.1.6) is the enstrophy of the ow.
Projecting the vorticity transport equation onto the vorticity vector, we derive an evolution
equation for the enstrophy,
d
dV = 2 ( ) : E dV 2 : dV, (11.1.7)
dt F low F low F low
where is the kinematic viscosity of the uid (Problem 11.1.1(b)). A more general evolution equation
for ow in a bounded domain is given in (3.12.27).
11.1 Invariants of vortex motion 853
11.1.2 Axisymmetric ow
In the case of axisymmetric ow with no swirling motion, the vorticity points in the azimuthal
direction, = e . In cylindrical polar coordinates, (x, , ), the expressions for the linear
momentum, angular momentum, kinetic energy, and helicity simplify into
P = 2 dA ex , A = 0, K = dA, H = 0, (11.1.8)
F low F low
where is the distance from the axis of revolution, dA = dx d is the dierential area in a plane
of constant azimuthal angle, , ex and e are coordinate unit vectors, and is the Stokes stream
function for axisymmetric ow; see equation (2.10.15). The helicity is identically zero because the
velocity is normal to the vorticity in an axisymmetric ow.
11.1.3 Two-dimensional ow
In the case of two-dimensional ow, we obtain a new set of invariants, some but not all of which are
related to the invariants of three-dimensional ow. In Section 3.13, we saw that the total circulation
of a two-dimensional ow is conserved. Accordingly,
C= z dA (11.1.9)
F low
is an invariant of the motion for inviscid or viscous uids, where dA = dx dy is a dierential area in
the xy plane of the ow and z is the nonvanishing vorticity component.
These invariants are related to the linear and angular impulse that must be expended in order to
generate the motion of the uid (e.g., [24], Section 7.2). Due to the constancy of C, the location of
the origin of the Cartesian axes is immaterial.
Kinetic energy
The kinetic energy of an innite two-dimensional ow with nonvanishing total circulation, C, is
innite. However, the quantity
1
W= z dA (11.1.11)
2 F low
is nite and remains constant during the motion, where is the stream function, assumed to decay
to zero at innity. Physically, W expresses the part of the kinetic energy of the uid that depends
on the particular way in which the vorticity is distributed in the ow (e.g., [24], Section 7.2).
854 Introduction to Theoretical and Computational Fluid Dynamics
Viscous eects
In the case of two-dimensional viscous ow with no interior boundaries, the total circulation, C, and
vorticity centroid, X , remain constant in time. The square of the dispersion length increases at a
constant rate according to the evolution equation
dD 2
= 4 (11.1.12)
dt
(e.g., [24], p. 536). This increase reects the tendency of the vorticity to diuse away from the
initial distribution and occupy all available space. Viscous dissipation causes W to monotonically
decrease during the evolution of an unsteady ow.
Problem
Point-vortex motion
The fundamental steps of vortex methods outlined in the Introduction of this chapter, combined
with the expressions for the induced velocity given in Section 2.13.1, provide us with a system of
coupled nonlinear ordinary dierential equations for the coordinates of the point vortices, Xi , in the
xy plane,
1 1
N N
dXi Yi Y j dYi Xi Xj
= j , = j , (11.2.1)
dt 2 j=1 |Xi Xj |2 dt 2 j=1 |Xi Xj |2
where j is the strength of the jth point vortex and the prime indicates that the term i = j
is excluded from the sum. An implicit supplement to this system is the condition of strength
invariance, di /dt = 0, for i = 1, . . . , N .
To compute the motion of a collection of point vortices from a given initial conguration, we
integrate equations (11.2.1) in time using a standard numerical method, such as the RungeKutta
method, as discussed in Section B.8, Appendix B.
11.2 Point vortices 855
To position of the point vortices in the complex plane is governed by the coupled dierential
equations
1
N
dZi 1
= , (11.2.2)
dt 2i j=1 Zi Zj
where Zi = Xi + iYi and i is the imaginary unit. Equations (11.2.1) are the real and imaginary parts
of the unifying complex form (11.2.2).
1
N
|x Xi |
(x) = i ln , (11.2.3)
2 i=1 a
where a is a chosen length. Expressing the vorticity in terms of the two-dimensional delta function
and using (11.2.3), we nd that the scalar invariants of the motion (11.1.10)(11.1.11) simplify into
N
1
N
1
N
C= i , X = i Xi , D2 = i Xi Xi ,
i=1
C i=1 C i=1
(11.2.4)
N N
|Xi Xj |
W= i j ln .
4 i=1 j=1 a
The availability of these invariants allows us to predict the trajectories of two point vortices, as
discussed in Section 11.2.1.
Hamiltonian formulation
Kirchho noted that the system of ordinary dierential equations (11.2.1) can be recast into the
Hamiltonian form
dXi W dYi W
i = , i = , (11.2.5)
dt Yi dt Xi
where summation over the repeated index i is not implied on the left-hand sides. This formulation
allows us to study the motion of point vortices in the context of Hamiltonian dynamics [278].
(a) (b)
3
y
1
3
3 2 1 0 1 2 3
x
Figure 11.2.1 (a) Two point vortices move along concentric circles centered at the centroid of vor-
ticity. (b) Streamline pattern due to a pair of point vortices with strengths of equal magnitude
and opposite sign, drawn in a frame of reference moving with the vortex pair. The x axis may be
regarded as an impenetrable wall.
while maintaining their initial separation, as illustrated in Figure 11.2.1(a). If 1 = 2 , the point
vortices rotate about the midpoint. If 1 2 , the second point vortex nearly orbits around the
nearly stationary rst point vortex.
1 2 N
Figure 11.2.2 A collection of N point vortices with identical strengths, , distributed at positions
corresponding to the zeros of the N -degree Hermite polynomial rotates as a whole with angular
velocity .
Vortex polygons
Havelock [172] considered the motion of N identical point vortices arranged at the vertices of a
regular polygon that is inscribed inside a circle of radius a. Assume that at an instant the point
vortices are located at the complex positions
i1
zi = xi + iyi = a exp 2i (11.2.8)
N 1
for i = 1, . . . , N , where i is the imaginary unit, so that z1 = a. Using Table 7.10.1, we nd that the
complex potential of the induced ow is
z N aN
w(z) = ln . (11.2.9)
2i aN
The complex velocity induced at the location of the rst point vortex, z1 = a, by all other point
vortices is
d z a z N 1 1 N 1
lim w ln = lim N N = . (11.2.10)
za dz 2i a 2i za z a N za 2i 2a
Accordingly, the vortex polygon rotates intact with angular velocity
= (N 1) . (11.2.11)
4a2
For N = 2, we recover formula (11.2.6) with 1 = 2 = and d = 2a. Linear stability analysis
reveals that the rotation is stable when N < 7, marginally stable when N = 7, and unstable when
N > 7 [172].
v i = xi . (11.2.15)
where c is a constant. To ensure that the coecients of the highest power of x on the right-hand
sides of the last two expressions are the same, we set c = 4N/ and nd that satises the
dierential equation
4 4
(x) x (x) + N (x) = 0. (11.2.19)
In terms of the dimensionless position,
%
2
x
x , (11.2.20)
equation (11.2.19) reduces to the Hermite equation
x) 2 x
( (
x) + 2N (
x) = 0, (11.2.21)
where a dot denotes a derivative with respect to x . The solution of the Hermite equation is the
N th-degree Hermite polynomial whose roots, i , are available in analytical or tabular form (e.g.,
[317]). The position of the point vortices in physical space is
1/2
xi i (11.2.22)
2
for i = 1, . . . , N , where is a free parameter whose sign matches the sign of .
11.2 Point vortices 859
where zi is the position of the ith point vortex. Using elementary calculus, we nd that
(zi ) 1 1 1 1
= + + + + + , (11.2.24)
2 (zi ) zi z1 zi zi1 zi zi+1 zi zN
where a prime denotes a derivative with respect to z (e.g., [317]). Note that one detrimental singular
term is missing from the sum. The complex velocity of the ith point vortex is given by
(zi )
ui i v i = , (11.2.25)
2i 2 (zi )
where ui = ux is the x velocity component and vi = uy is the y velocity component of the ith point
vortex.
If the point vortex assembly rotates intact as a cluster around the origin with angular velocity
, the velocity of the ith point vortex satises ui + i vi = i zi or
ui i vi = i zi (11.2.26)
for i = 1, . . . , N , where an asterisk denotes the complex conjugate. Substituting expression (11.2.25)
for the left-hand side, we nd that
(zi )
= i zi . (11.2.27)
2i 2 (zi )
This expression shows that the positions zi are zeros of the generally nonanalytic complex function
4
(z) (z) z (z), (11.2.28)
constituting a polynomial in z and z .
A rotating polygon
In the case of a regular polygon of N point vortices with radius a, we substitute (z) = z N aN
and nd that
4 N 1 4 2
(z) N (N 1) z N 2 z Nz = N z N 2 N 1 |z| . (11.2.29)
Requiring that (z) = 0 for z N = aN where |z| = a, we recover precisely Havelocks expression
(11.2.11).
860 Introduction to Theoretical and Computational Fluid Dynamics
(z) = (z M aM
1 )(z
M
aM
2 ), (11.2.30)
(z) = M z M 1 (2z M aM
1 a2 )
M
(11.2.31)
and
(z) = M z M 2 2(2M 1) z M (M 1) (aM M
1 + a2 ) . (11.2.32)
4
(3M 1) aM
1 (M 1) a2 =
M
|a1 |2 (aM
1 a2 ),
M
4
2 (M 1) a1
(3M 1) aM M
= |a2 |2 (aM
2 a1 ).
M
(11.2.34)
Dividing side by side to eliminate and rearranging, we obtain
3M 1 M
M ||2 ( + ||2 ) + 1 = 0, (11.2.35)
M 1
where = a2 /a1 is the radii ratio. Further rearrangement yields
3M 1 M
2 M +2 ( + 2 2 ) + 1 = 0, (11.2.36)
M 1
where ||/. If is real, in which case the polygons are in-phase, 2 = 1. If = || exp(i/M ),
in which the polygons are out of phase, 2 = exp(2i/M ).
Similar structures are obtained for higher values of M , as shown in Figure 11.2.3. The point
vortex clusters shown in this gure were generated using the program pvpoly2 in directory 09 vortex
of the software library Fdlib (Appendix C).
11.2 Point vortices 861
Figure 11.2.3 Equilibrium structures of two nested polygons of point vortices with identical strengths
rotating steadily in an innite uid. In-phase and out-of-phase polygons are observed for M > 4.
In the case of a regular polygon of N point vortices with radius a, we substitute (z) = z N aN
and nd that
4 N 1
(z) N (N 1) z N 2 z Nz + 2N 0 z N 2 . (11.2.40)
862 Introduction to Theoretical and Computational Fluid Dynamics
Simplifying, we obtain
4 2
(z) = N z N 2 N 1 + 20 |z| . (11.2.41)
In the case of two nested concentric M -sided regular polygons with radii |a1 | and |a2 |, we have
N = 2M and (z) = (z M aM 1 )(z
M
aM2 ). Working as previously , we obtain
(z) M z M 2 2(2M 1) z M (M 1) (aM M
1 + a2 )
4 2
|z| 20 (2z M aM
1 a2 ) .
M
(11.2.43)
The counterparts of equations (11.2.44) are
4
1 (M 1 + 20 ) a2 =
(3M 1 + 20 ) aM M
|a1 |2 (aM
1 a2 ),
M
4
2 (M 1 + 20 ) a1
(3M 1 + 20 ) aM M
= |a2 |2 (bM aM ). (11.2.44)
Dividing side by side to eliminate and rearranging, we obtain
3M 1 + 20 M
2 M +2 ( + 2 2 ) + 1 = 0, (11.2.45)
M 1 + 20
where = b/a is the radii ratio and ||/. If is real, the polygons are in-phase; if =
|| exp(i/M ), the polygons are out of phase.
Generalization
To generalize the preceding results, we consider a rotating assembly of point vortex consisting of
an M1 -sided polygon of point vortices of radius |a1 | and strength 1 , and a concentric M2 -sided
polygon of point vortices of radius |a2 | and strength 2 . It is convenient to introduce the polynomials
1 (z) = z M1 aM
1
1
and 2 (z) = z M2 aM 2
2 , and express the velocity of the ith point vortex in the
rst array as
and the velocity of the ith point vortex in the second array as
These expressions show that the position of the point vortices in the rst polygon are the zeros of
the function
4
1 (z) = 1 (z) 2 (z) + 1 (z) 2 2 (z) z 2 (z) , (11.2.48)
1
and the position of the point vortices in the second polygon are the zeros of the function
1 4
2 (z) = 2 (z) 1 (z) + 2 (z) 2 1 (z) z 1 (z) , (11.2.49)
1
(2 + )M M (2 + 1)M 1 2 2
2 M +2 + = 0, (11.2.53)
M 1 M 1
where a2 /a1 and ||/.
The analysis presented in this section can be extended to a higher number of nested point
vortex polygons with the same or dierent strengths [13].
The stream function of the ow due to a point vortex located at the point X in the presence
of an impermeable boundary is the Greens function of the rst kind of the two-dimensional Laplace
equation multiplied by the strength of the point vortex, , that is, (x, X) = G(x, X). In the
case of unbounded ow, we employ the free-space Greens function, G(x, X) = 1/(2) ln |x X|.
Specic expressions for several Greens functions generated by the method of images are given in
Section 2.2.8.
a2
Xinv = xc + (X xc ). (11.2.54)
|X xc |2
This construction shows that a single point vortex describes a circular path that is concentric with
the cylinder, moving around the center of the cylinder with polar velocity
|X xc |
u = . (11.2.55)
2 a |X xc |2
2
|X xc | c 1
u = 2 2
+ . (11.2.56)
2 a |X xc | 2 |X xc |
(a) (b)
3
0.8 2
0.6
0.4 1
0.2
0 0
y
y
0.2
0.4 1
0.6
0.8 2
3
1 0.5 0 0.5 1 3 2 1 0 1 2 3
x x
Figure 11.2.4 Streamline pattern due to a point vortex (a) inside and (b) outside a circular cylinder.
Figure 11.2.5 Trajectory of a point vortex in the rst quadrant near a right-angle corner, and illus-
tration of the three images.
(a) (b)
1.5 3
2.5
1 2
1.5
0.5 1
y
y
0.5
0 0
0.5
0.5
1 0.5 0 0.5 1
1
2 1 0 1 2
x
x
Figure 11.2.6 Streamline pattern due to a point vortex (a) between two parallel walls and (b) in a
semi-innite strip.
Assume that the point vortex is located at (X, Y ) and the walls are parallel to the x axis.
Using (2.13.14), we nd that the stream function of the induced ow is
cosh[k(x X)] cos[k(y Y )]
(x) = ln , (11.2.57)
4 cosh[k(x X)] cos[k(y Yim )]
where k = /h is a wave number and Yim is the y position of the image. Using (2.13.15), we nd
that the point vortex moves along the x axis parallel to the walls under the inuence of the ow
induced by the second array containing its image. The velocity of the point vortex is
sin(2kb)
Vx = , (11.2.58)
4h 1 cos(2kb)
where b is the distance of the point vortex from the lower wall. The streamline pattern in shown in
Figure 11.2.6(a). In the limit as kb tends to zero, we recover the earlier results for a point vortex in
a semi-innite uid above an innite plane wall.
The image system consists of the image system associated with the two parallel walls, and the
reection of the image system with respect to the third intersecting wall. The strengths of the
reected point vortices are the negatives of those of their images. The motion of the point vortex
can be computed by numerical methods.
1
N
dXi sinh[k(Yi Yj )]
= j ,
dt 2a j=1 cosh[k(Yi Yj )] cos[k(Xi Xj )]
(11.2.59)
1
N
dYi sin[k(Xi Xj ])
= j ,
dt 2a j=1 cosh[k(Yi Yj )] cos[k(Xi Xj )]
where k = 2/a is the wave number and the prime indicates that the term j = i is excluded from
the sum.
Problems
1 1 1
N N
dZi j dZi
= , = j cot k (Zi Zj ) , (11.2.60)
dt 2i Zi Z j dt 2ai 2
j=1 j=1
where i is the imaginary unit, i2 = 1, an asterisk denotes the complex conjugate, and the prime
indicates that the term j = i is excluded from the sum.
Figure 11.2.7 Two point vortices behind a cylinder in streaming ow emulate a stationary wake.
([220], p. 223). A pair of point vortices with opposite strength is symmetrically located above
and below the x axis. Verify that the point vortices will remain stationary provided that they lie
on the curve described by the equation 2ry = r 2 a2 , and the strengths of the point vortices are
= U r(1 a2 /r2 )(1 a4 /r 4 ), where r is the distance from the origin.
Computer Problems
compared to the point vortex separation, a, the motion can be described analytically in the context
of linear stability theory discussed in Chapter 9.
Xm = a (m + m ), Ym = a m (11.3.1)
for m = 0, 1, 2, . . ., where
are dimensionless complex displacements, 1 and 2 are dimensionless complex coecients with
small but comparable magnitudes, k = 2/L is the wave number of the perturbation, L is the
corresponding wavelength, and is the complex growth rate. If the imaginary part of is positive,
the periodic array is unstable.
The physical requirement that the separation between two consecutive point vortices is smaller
than the wavelength of the perturbation, L = k/2, imposes the geometrical constraint a/L =
ka/(2) < 1. When the ratio L/a is an integer, the array is perturbed in a commensurate and
spatially periodic fashion.
The motion of the point vortices is governed by the dierential equations (11.2.1). Linearizing
the terms inside the sums with respect to 1 and 2 , we obtain
Y m Yn 1 m n Xm Xn 1 m n + m n +
, (11.3.3)
|Xm Xn | 2 a (m n)2 |Xm Xn |2 a (m n)2 + 2(m n)(xm xn ) +
for any arbitrary integer pair, (m, n). Simplifying the second expansion, we obtain
Xm Xn 1 1 m n
+ . (11.3.4)
|Xm Xn |2 a m n (m n)2
The x position of each point vortex evolves according to the linearized ordinary dierential
equation
dXm m n
1e
i(nm)ka
= 2
= 2 exp[i(mka t)] , (11.3.5)
dt 2a n= (m n) 2a n=
(m n)2
where the prime indicates that the term m = n is excluded from the sum. Substituting into the
left-hand side the rst expressions in (11.3.1) and (11.3.2) and simplifying, we obtain
1 a i = 2 , (11.3.6)
2a
870 Introduction to Theoretical and Computational Fluid Dynamics
where
1 eilka 1 cos(lka) ka
= = = ka 1 , (11.3.7)
l2 l 2 2
l= l=
and the prime indicates that the singular term l = 0 is excluded from the sum.
Unfortunately, applying formula (11.2.1) for the y velocity component produces a divergent
sum. To circumvent this apparent but not essential diculty, we retain a large but nite number of
2N + 1 point vortices in the array and obtain
i(nm)ka
N
dYm 1 1e
= 1 exp[i(mka t)] . (11.3.8)
dt 2a mn n=
(m n)2
n=N
Because the rst sum on the right-hand side of (11.3.8) is independent of the position of the point
vortices, it can be discarded with no consequences on the ow. This manipulation is an example
of a mathematical renormalization. Substituting into the left-hand side the second expressions in
(11.3.1) and (11.3.2), and simplifying, we obtain
2 a i = 1 . (11.3.9)
2a
Combining equations (11.3.6) and (11.3.9), we nd that
a a 2
= i 1 , = 1. (11.3.10)
a2 L L 1
Since a/L < 1, the plus or minus sign can be selected according to the sign of the strength of
the point vortices so that the imaginary part of is positive, corresponding to instability. The
corresponding normal mode displacement causes the point vortices to move away from their original
location at an exponential rate (see also Problem 11.3.1).
An unstable and a stable normal mode displacement are illustrated in Figure 11.3.1 for negative
point vortex strength, < 0, corresponding, respectively, to 2 /1 = 1 and 1. In the case of the
unstable mode, the point vortices are displaced so that they concentrate near regions where the
interface slopes downward, causing a local counterclockwise rotation that promotes the instability.
The opposite is true in the case of the stable normal mode with 2 /1 = 1.
Vortex pairing
When the wavelength of the perturbation, L, is a multiple of the point vortex separation, a, the
ratio n L/a is an integer expressing the number of point vortices residing inside each period. The
expression for the complex growth rate in (11.3.10) becomes
n1
= i . (11.3.11)
a 2 n2
For xed point vortex strength, , the highest value of I corresponding to the most unstable
perturbation is obtained for the smallest possible integer, n = 2, associated with a pairing interaction
11.3 Rows of point vortices 871
Ustable Stable
Figure 11.3.1 An unstable and a stable normal mode displacement of an innite array of point vortices
with negative circulation.
between two adjacent point vortices. An unstable normal mode perturbation displaces the point
vortices upward and downward in succession, yielding a triangular wave.
Vortex sheet
An interesting interpretation of expression (11.3.11) for the growth rate arises by expressing the
strength of the point vortices in terms of the dierence in the velocity of the streams far above and
below the point-vortex array, U , the wavelength of the perturbation, L, and the number of point
vortices inside each period, n = L/a, nding = a U = L U/n. The growth rate is
U n 1 1 n1
= i = i kU , (11.3.12)
L n 2 n
where k = 2/L is the wave number. Increasing the number of point vortices inside each wave-
length, n, from two to innity, while keeping the velocity dierence constant and the wavelength
of the perturbation xed, doubles the growth rate. The limit n corresponds to a continuous
distribution of point vortices representing a vortex sheet.
Symmetric street
Using expressions (11.2.59), we nd that, when the point vortices are located above each other in
the arrangement of the symmetric vortex street shown in Figure 11.3.2(a), they translate parallel to
the x axis with velocity
sinh(kb) 1
Vx = = coth( kb), (11.3.13)
2a cosh(kb) 1 2a 2
where k = 2/a is the wave number. Linear stability analysis shows that the symmetric vortex
street is unstable and should not be expected to occur in practice (Problem 11.3.2).
872 Introduction to Theoretical and Computational Fluid Dynamics
(a) (b)
y y
a a
x b x
b
Figure 11.3.2 Illustration of (a) a symmetric and (b) an antisymmetric street of point vortices. The
strengths of the point vortices in the top row are equal in magnitude and opposite in sign that
those of the point vortices in the bottom row.
Antisymmetric street
When the vortices in the top row are displaced with respect to the point vortices in the bottom
row along the x axis by a distance that is equal to half their separation, forming the antisymmetric
vortex street shown in Figure 11.3.2(b), they translate along the x axis with velocity
sinh(kb) 1
Vx = = tanh( kb). (11.3.14)
2a cosh(kb) + 1 2a 2
Linear stability analysis reveals that the antisymmetric vortex street is unstable, except in the
special case where cosh(kb) = 2 corresponding to b/a = 0.281, where the ow is marginally stable
(Problem 11.3.2). However, nonlinearities promote the instability and the antisymmetric vortex
street is considered unstable.
Problem
Problem
(a) Write a program that computes the motion of the point vortices and plot proles of the evolving
array. The ordinary dierential equations governing the motion of the point vortices should be
integrated in time using the modied Euler method discussed in Section A.8, Appendix B. The
program should compute the ratio of the maximum displacement of the point vortices from the x
axis, b, to the initial amplitude, b0 .
(b) Consider an unstable normal mode perturbation where the initial position of the mth point
vortex is given by
z = (x X), (11.4.1)
as discussed in Section A.3, Appendix A. The distribution function has dimensions of inverse
squared length.
where r = |x X| is the distance from the center of the blob, is the blob radius, and f is a
dimensionless distribution function decaying far from the blob. As tends to zero, reduces to the
two-dimensional delta function. For the integral of over the entire xy plane to be equal to unity,
the function f must satisfy the integral constraint
1
wf (w) dw = . (11.4.4)
0 2 2
A possible choice for the core distribution function is the piecewise constant top-hat distribution
&
1/ 2 for w 1,
f (w) = (11.4.5)
0 for w 1,
describing the circular Rankine vortex with constant vorticity inside the core. Another choice is the
Gaussian distribution
1 w2
f (w) = e . (11.4.6)
2
Both choices satisfy the integral constraint (11.4.4).
Velocity eld
In plane polar coordinates centered at an axisymmetric vortex blob, the radial velocity is zero.
To compute the polar velocity, u , we resort to the denition of vorticity and derive the ordinary
dierential equation
1 d
(ru ) = 2 f (r/). (11.4.7)
r dr
Straightforward integration subject to the condition that the velocity is zero at the centerline yields
the velocity prole
u = F(r/), (11.4.8)
2r
where
r/
F(r/) = 2 2 w f (w ) dw (11.4.9)
0
is a dimensionless function. Equation (11.4.4) shows that, as the blob radius tends to zero for xed
r, the function F(r/) tends to unity. The Cartesian components of the velocity at a point x due
to a blob located at X are
y x
ux (x) = F(r/), uy (x) = F(r/), (11.4.10)
2 x2 + y 2 2 x2 + y 2
where x = x X is the distance from the centerpoint. This velocity diers from that due to a
point vortex only by the presence of the radial distribution function, F(r/). Far from the center of
the blob, F(r/) tends to unity and the blob induces a velocity eld similar to that due to a point
vortex.
11.4 Vortex blobs 875
For the top-hat distribution shown in (11.4.5), straightforward computation of the integral in
(11.4.8) yields
& 2
w for w 1,
F(w) = (11.4.11)
1 for w 1,
which shows that the uid inside a Rankine blob rotates like a rigid body, while the ow outside the
Rankine blob is identical to that due to a point vortex.
BiotSavart integral
It is instructive to rederive the velocity eld due to a vortex blob based on the BiotSavart integral
(2.1.26). Setting the origin at the center of a blob, we nd that the y component of the velocity at
a point x0 located on the x axis is given by
2
x0 r cos
uy (x0 ) = 2 (x) r dr d. (11.4.13)
2 0 0 x0 + r2 2x0 r cos
Substituting (11.4.3) into the integrand of (11.4.13), we obtain
2 x0 r cos
dXi 1
N
Yi Yj X X
i j
= j F ,
dt 2 j=1 |Xi Xj |2 j
dYi 1
N
X i Xj X X
i j
= j F , (11.4.16)
dt 2 j=1 |Xi Xj |2 j
where j is the strength of the jth blob and the prime indicates that the term i = j is excluded
from the sum. In order to preserve the total circulation, the strength of each blob remains constant
in time.
876 Introduction to Theoretical and Computational Fluid Dynamics
1
N
dXi sinh[k(Yi Yj )]
= j + Ui ,
dt 2a j=1 cosh[k(Yi Yj )] cos[k(Xi Xj )]
1
N
dYi sin[k(Xi Xj )]
= j + Vi , (11.4.17)
dt 2a j=1 cosh[k(Yi Yj )] cos[k(Xi Xj )]
where
1 Yi Yj 1 Xi Xj ma
N N
Ui = j 2 Gijm , Vi = j 2 Gijm ,
2 j=1 m= Rijm 2 j=1 m= Rijm
(11.4.18)
R
2
= (Xi Xj ma)2 + (Yi Yj )2 .
ijm
Gijm F 1, Rijm
j
The sums on the right-hand side of (11.4.17) express the velocity due to a periodic array of point
vortices. The sums in (11.4.18) express corrections due to the nite size of the blobs. The inner
innite sums over m converge rapidly and can be truncated after a sucient number of terms.
d d2
=2 or = , (11.4.19)
dt dt
where is the kinematic viscosity of the uid. We have found that the square of the radius or
a Gaussian blob, 2 , increases at a linear rate in time. Equations (11.4.16) and (11.4.19) provide
us with a system of ordinary dierential equations for the position of the centers and radii of a
collection of vortex blobs. To preserve the total circulation, which is an invariant of the motion, we
require that the strength of each blob remains constant during the motion. However, in practice,
the vorticity distribution of each blob becomes non-axisymmetric due to mutual interactions.
Problem
Fluid 1
t (1)
u
(2)
u
Fluid 2
Figure 11.5.1 Illustration of a two-dimensional vortex sheet described by a string of marker points
moving with the uid velocity normal to the vortex sheet, while executing an arbitrary tangential
motion.
Computer Problems
where u(i) is the uid velocity on the ith side of the vortex sheet, i = 1, 2, and t is the tangent
unit vector. Since nonzero viscosity causes the unphysical onset of innite shear stress and viscous
dissipation, the notion of a vortex sheet whose thickness is and remains innitesimal in time is
acceptable only in the context of inviscid ow.
Principal velocity
The principal velocity of the vortex sheet is dened as the mean value of the velocity of the uid on
878 Introduction to Theoretical and Computational Fluid Dynamics
1 (1)
uP V u + u(2) . (11.5.2)
2
Combining this denition with (11.5.1), we obtain
1 1
u(1) = uP V t, u(2) = uP V + t, (11.5.3)
2 2
subject to the conventions dened in Figure 11.5.1.
Marker points
To describe the motion of a vortex sheet, we mark its trace in the xy plane with a string of marker
points identied by a permanent label, , as depicted in Figure 11.5.1, and follow the evolution of the
vortex sheet by computing the trajectories of the marker points. Kinematic compatibility requires
that the component of the velocity of the marker points normal to the vortex sheet be equal to the
normal component of the uid velocity on either side of the vortex sheet. The tangential component
of the marker-point velocity can be arbitrary. If X(, t) is the position and V(, t) is the velocity of
a marker point,
X
= V(, t) = (u(i) n) n + Vt () t (11.5.4)
t
for i = 1, 2, where Vt () is an arbitrary tangential velocity allowed to vary in time and across the
marker points.
Using equations (11.5.3), we express the marker-point velocity in terms of the velocity of the
uid on either side of the vortex sheet, the principal velocity of the vortex sheet (PV), and the
strength of the vortex sheet, as
1
V(a, t) = (1 ) u(1) + u(2) = uP V + ( ) t, (11.5.5)
2
where is a dimensionless parameter related to the tangential marker-point velocity by
1
Vt = uP V t + ( ) . (11.5.6)
2
When = 0, 12 , 1 the marker points move, respectively, with the velocity of the uid on the rst
side of the vortex sheet, the principal velocity of the vortex sheet, and the velocity of the uid on
the second side of the vortex sheet. In the remainder of our discussion, will be assumed to be a
specied constant.
It will be useful to express the uid velocity on either side of the vortex sheet in terms of the
marker-point velocity, V, as
where 1 = and 2 = 1 . Equations (11.5.7) can be collected into the compact form
u(i) = V + i t (11.5.8)
for i = 1, 2.
BiotSavart integral
Substituting into (11.5.5) the BiotSavart integral (2.13.19), we obtain an integro-dierential equa-
tion for the motion of the marker points,
! "
X 1 PV
Y () + Y ( ) ( ) l 1
= V() = d + ( ) t + v, (11.5.9)
t 2 C X() X( ) |X() X( )|2 2
where C is the trace of the vortex sheet in the xy plane, P V denotes the principal-value integral,
and l is the arc length along the vortex sheet measured in the direction of the tangent unit vector,
t. The rst term on the right-hand side of (11.5.9) expresses the self-induced velocity of the vortex
sheet. The third term expresses a generally time-dependent external rotational or irrotational ow
with velocity v.
u(i) 1
+ u(i) u(i) = pi + g (11.5.10)
t i
for i = 1, 2, where pi is the pressure and i is the uid density. Slightly rearranging the left-hand
side, we obtain
u(i) 1
+ V u(i) + (u(i) V) u(i) = pi + g. (11.5.11)
t i
The rst two terms on the left-hand side express the derivative (u(i) /t) , representing the rate of
change of the uid velocity on either side of the vortex sheet following the marker points.
Dierentiating equation (11.5.8), relating the uid velocity to the marker-point velocity, we
obtain
u(i)
V
(t)
= + i . (11.5.12)
t t t
Substituting the right-hand side of (11.5.12) in place of the rst two terms in (11.5.11), and setting
u(i) V = i t, we nd that
V
(t)
(V + i t)
i + i i + = pi + i g, (11.5.13)
t t l
880 Introduction to Theoretical and Computational Fluid Dynamics
where l is the arc length along the vortex sheet. Next, we take the dierence of the equations that
arise by writing (11.5.13) for i = 1 and 2, nding
V
(t)
V ( (t)
+ + + = p + g, (11.5.14)
t t l l
= 2 2 1 1 = (1 ) 2 + 1 , ( = 2 2 2 1 = (1 )2 2 2 1 . (11.5.15)
2 1
Projecting (11.5.14) onto the tangent unit vector, t, to formulate the tangential derivatives of the
pressure dierence, and noting that, because t is a unit vector t t/l = 0 and t t/t = 0, we
nd that
V
V ( p
t + +t + = + g t. (11.5.16)
t t l l l
The jump in the pressure across the vortex sheet on the right-hand side of (11.5.17) depends on the
properties of the interface separating the two uids, as discussed in Section 3.8. For example, if the
vortex sheet represents the interface between two immiscible uids with surface tension ,
p = , (11.5.18)
where is the curvature of the vortex sheet in the xy plane, reckoned as positive when the interface
is downward convex. In the absence of surface tension, p = 0. Combining equation (11.5.17) with
the interfacial condition (11.5.18), we obtain the desired evolution equation for the strength of the
vortex sheet following the marker points involving the marker-point velocity distribution, V(), and
the tangential component of the marker-point acceleration, t (V/t) .
where
1
v
The time derivative inside the integral is taken keeping both and constant. Expanding the
derivative under the integral sign, we obtain
V
PV ! "
1 Y () + Y ( ) 1
t() = t() dl( ) + J () + G().
t 2 C X() X( ) |X() X( )|2 t
(11.5.21)
The function
! "
1 PV
Y () + Y ( ) 1 l
J () t() ( ) d (11.5.22)
2 C t X() X( ) |X() X( )|2
expresses the tangential component of the acceleration of the point particles if they were moving
with the principal velocity of the vortex sheet while preserving the current circulation. In practice,
this function can be computed by numerical time dierentiation. Substituting the right-hand side of
(11.5.21) into the second term on the right-hand side of (11.5.17), rearranging, and observing that
+ ( 12 ) = 12 (1 + 2 ), we obtain a linear Fredholm integral equation of the second kind for
(/t) ,
PV ! "
A Y () + Y ( ) 1
+ t() dl( ) = 2 F(), (11.5.23)
t C X() X( ) |X() X( )|2 t
where
2 1
A= (11.5.24)
1 + 2
is the Atwood ratio. The forcing function on the right-hand side of (11.5.23) is given by
v
V
1 p
F() = A J + t [ g] B t + A , (11.5.25)
t l l 1 + 2 l
where
(1 )2 + 1 (
(1 )2 2 2 1
B= = , A = = (11.5.26)
1 + 2 1 + 2 1 + 2 1 + 2
are dimensionless coecients dened in terms of the uid densities and the marker-point velocities.
Solving (11.5.28) for (/t) , substituting the result into the rst term on the left-hand side
of (11.5.17), multiplying the emerging equation by l/, and simplifying, we derive an integro-
dierential evolution equation for the circulation along the vortex sheet,
2 x V
1 ( 2
+ = p + g x . (11.5.29)
t t 2
The tangential acceleration in the second term on the left-hand side is computed from (11.5.21).
The right-hand side of (11.5.29) expresses the rate of production of vorticity along the vortex sheet.
1 ( = 1 , 1 1
= (1 + 2 ), B= , A = A. (11.5.30)
2 4 2 4
The forcing function dened in (11.5.25) takes the simpler form
v
1 1 V 1 p
F() = A J + t [ g]+ t . (11.5.31)
t 4 l 2 l 1 + 2 l
The motion of the vortex sheet is determined exclusively by the Atwood ratio, A. The evolution
equation (11.5.17) becomes
V V
1 2 p
+t = 2 A t + tg] . (11.5.32)
t l t 4 l 1 + 2 l
The rst two terms inside the square brackets on the right-hand side can be written in the compact
form
V
1
t + = t a, (11.5.33)
t 4 l
where
1 Du(1) Du(2)
a + (11.5.34)
2 Dt Dt
is the mean value of the uid acceleration on either side of the vortex sheet and D/Dt is the material
derivative. To show this, we write
Under the Boussinesq approximation, the factors and coecients dened in (11.5.15) and
(11.5.26) take the values
1 1
= , ( = (1 2) ,
B= , A = . (11.5.40)
2 2
The evolution equation for the strength of the vortex sheet reduces to
V
1 p
= t + (1 2) + 2 A g t, (11.5.41)
t l l l
884 Introduction to Theoretical and Computational Fluid Dynamics
and the evolution equation for the circulation along the vortex sheet reduces to
1 p
= ( ) 2 + 2 A g x + c(t). (11.5.42)
t 2
The simplications resulting from the Boussinesq approximation are reected in the absence of the
tangential marker-point acceleration on the right-hand sides of (11.5.41) and (11.5.42).
Problems
Xi
X i1/2 X i+1/2
Figure 11.6.1 Discretization of a vortex sheet into a collection of elements dened by end points and
midpoints. In the point-vortex method, the vortex sheet is replaced by a collection of point vortices
located at the midpoints.
Trapezoidal rule
In the point vortex method, the principal value of the BiotSavart integral (11.5.9) along a vortex
sheet is approximated with the trapezoidal rule. To implement the method, we divide the vortex
sheet into N elements whose end points are dened in terms of a chain of marker points, Xi1/2 ,
where i = 1, . . . , N + 1, as shown in Figure 11.6.1. The midpoint of the ith segment extending
between Xi1/2 and Xi+1/2 is denoted by Xi , the element length is denoted by li , the dierence in
the circulation across the end points is denoted by i , and the unit vector tangent to the segment
at the midpoint is denoted by ti . Using central dierences, we approximate the tangent unit vector
as
1
ti Xi+1/2 Xi1/2 . (11.6.1)
li
More advanced dierentiation methods based on cubic spline or trigonometric interpolation of the
coordinates of the marker points with respect to the arc length of the polygonal line connecting the
marker points oer higher accuracy.
where the prime indicates that the term i = j is excluded from the sum. Apart from the last two
terms, equation (11.6.2) is identical to (11.2.1) describing the motion of a collection of N point
886 Introduction to Theoretical and Computational Fluid Dynamics
vortices located at the element midpoints, Xi . The element circulation, j , plays the role of the
point vortex strength, j . Computing the evolution of the vortex sheet is thus reduced to computing
the motion of point vortices with generally evolving strength.
Having advanced the position of the midpoints, Xi , using (11.6.2), we reset the position of
the end points, Xi+1/2 , by linear or high-order interpolation. In the simplest approach, we set
Xi+1/2 = 12 (Xi + Xi+1 ). As an alternative, we advance the position of the end points according to
the velocity eld induced by the point vortices, setting
N ! "
dXi+1/2 1 Yi+1/2 + Yj j 1
= + ( ) i+1/2 t(Xi+1/2 ) + v(Xi+1/2 ).
dt 2 j=1 Xi+1/2 Xj |Xi+1/2 Xj |2 2
(11.6.3)
The tangential unit vector in the last term of (11.6.3) can be computed by numerical dierentiation
setting, for example, t(Xi+1/2 ) = (Xi+1 Xi )/|Xi+1 Xi |.
The acceleration of the point vortices involved in the rst term on the right-hand side is computed
by dierentiating the right-hand side of (11.6.2) with respect to time. To simplify the expressions,
we set = 12 and obtain
V
! "
1 Yi + Yj
N
dVi 1 dj
=
t a dt 2 j=1 Xi Xj |Xi Xj |2 dt
! "
1 d Yi + Yj
v
N
1
+ + (Xi ). (11.6.5)
2 j=1 dt Xi Xj |Xi Xj |2 j
t
Carrying out the dierentiation in the second sum on the right-hand side and simplifying, we nd
that
! " ! "
d Yi + Yj 1
1 2uxy + v (y 2 x2 )
= 4 , (11.6.6)
dt Xi Xj |Xi Xj |2 r u (y 2 x2 ) 2vxy
where
Substituting (11.6.6) into (11.6.5) and then into (11.6.4), we obtain a system of linear algebraic
equations for di /dt. The solution can be computed by Jacobi iterations, which involves assuming
11.6 The point-vortex method for vortex sheets 887
values for di /dt, substituting these values into the right-hand side of (11.6.5), and then computing
the right-hand side of (11.6.4) to produce new and improved values for di /dt, to be used in
subsequent iterations.
Boussinesq approximation
Adopting the Boussinesq approximation and setting in (11.6.4) = 12 , we obtain an explicit evolution
equation for the strength of the point vortices,
d i p Xi+1/2
= + 2Ag x , (11.6.8)
dt Xi1/2
where A is the Atwood ratio dened in (11.5.24). In the absence of surface tension, p = 0. We see
that the Boussinesq approximation bypasses solving a system of equations for the rate of the change
of the strength of the point vortices.
11.6.1 Regularization
In practice, unless a vortex sheet is suciently stretched during the evolution, the motion of the
point vortices representing the vortex sheet becomes disorganized and eventually chaotic. The origin
of this pathology can be traced back to the exponential amplication of the round-o error due to
the instability of an array of point vortices discussed in Section 11.3.
Repositioning
One way to regularize the motion is to redistribute the point vortices along the vortex sheet after
each time step or every few time steps, so that the point vortices become evenly spaced with respect
to the instantaneous arc length [127]. The strength of the point vortices at the new positions is
found from the strength of the point vortices at the old positions by interpolation with respect to
or with respect to arc length of the polygonal line connecting successive marker points.
1
xsi = (xi2 + 4 xi1 + 10 xi + 4 xi+1 xi+2 ), (11.6.9)
16
where the superscript s denotes a smoothed position [244].
replaced with a nearly singular integrand, and equation (11.5.7) is correspondingly modied. For
= 12 , the motion of the point vortices is governed by the equations
! "
1
N
dXi Yi + Yj j
= + v(Xi ), (11.6.10)
dt 2 j=1 Xi Xj |Xi Xj |2 + 2
where is a small numerical parameter with dimensions of length, included to eliminate the unstable
behavior due to the singular BiotSavart integral [216].
1
+( ) t + v, (11.6.11)
2
where k = 2/a is the wave number and C is the trace of the vortex sheet in the xy plane over one
period. The evolution equation for the strength of the vortex sheet remains unchanged.
where N is the number of point vortices inside one period and is a small dimensionless regularization
parameter. The evolution equation for the strength of the point vortices remain unchanged.
Stages in the evolution of a sinusoidally perturbed vortex sheet with negative strength im-
mersed in a homogeneous uid computed using (11.6.12) with N = 96 and = 0.125 or 0.250 are
shown in Figure 11.6.2. The vortex sheet rolls up into a periodic sequence of spirals whose turns
keep increasing in time. A higher number of tightly wound spiral turns develop as becomes smaller.
The numerical results suggest that, in the limit as tends to zero, a cusp develops spontaneously at
the center of the spiral after a nite period of time, immediately before the spiral starts forming its
turns [216].
Problems
(a) (b)
0 0
1 1
2 2
3 3
y/L
y/L
4 4
5 5
6 6
1 0 1 2 1 0 1 2
x/L x/L
Figure 11.6.2 Stages in the evolution of a sinusoidally perturbed vortex sheet with negative strength
computed using the regularized equation (11.6.12) with (a) = 0.250 or (b) 0.125. Proles are
shown at equal time intervals.
11.6.3 Smoothing
Discuss the mathematical action of the smoothing formula (11.6.9) with reference to a triangular
wave.
Computer Problem
periodic sinusoidal perturbation in shape with wavelength L and amplitude b0 . When viscous eects
are insignicant, the interface can be identied with a vortex sheet with vanishing initial strength.
Write a program that computes the motion of the vortex sheet using the point-vortex method,
subject to the Boussinesq approximation. The integration in time should be performed using the
modied Euler method discussed in Section B.8, Appendix B. The motion should be regularized using
a method of your choice. At the initial instant, assume that each point vortex is displaced along
the y axis in a sinusoidal fashion, while its strength remains zero. Run the program for b0 /L = 0.10
and N = 8, 16, and 32 point vortices inside each period separated by distance a = L/N . Discuss
the observed dierences in behavior.
N
z = i (t) 2 (x xi ), (11.7.1)
i=1
where 2 is the two-dimensional delta function. The motion of the point vortices is then computed
using the methods discussed in Section 11.2 [85]. A regularized version of method discretizes the
vorticity eld into vortex blobs, as discussed in Section 11.4 (e.g., [328]).
At the beginning of the computation, we must distribute the point vortices in a way that is consistent
with the initial state. In one implementation, a two-dimensional rectilinear or curvilinear grid
forming a two-dimensional array of cells is introduced and a point vortex is placed in the middle of
each cell. To ensure that the sum of the strengths of all point vortices is equal to the circulation
around a large loop enclosing the ow, the strength of each point vortex is set equal to the integral of
the vorticity or circulation of the uid inside the corresponding cell. Other methods of discretizing
the initial vorticity eld are available [78].
The simulations become prohibitively expensive when a large number of point vortices are employed.
Computing the velocity of N point vortices requires calculating N 2 mutual interactions, which can
be unaordable for N on the order of a few hundred or thousand. To circumvent this diculty, a
vortex-in-cell (VIC) method was proposed as a special implementation of the cloud-in-cell method
(CIC) [85]. The numerical scheme helps reduce the computational cost by circumventing the direct
computation of the mutual interactions.
11.7 Two-dimensional ow with distributed vorticity 891
Given the instantaneous position of the point vortices, the stream function is computed by
inverting the Poisson equation
2 = z . (11.7.2)
The velocity at the position of the point vortices arises by numerically dierentiating the stream
function.
where = 4n /A, A is the cell area, xn = (xn , yn ) is the position of the nth point vortex, and n
is the strength of the nth point vortex [85]. The contribution of the point vortex to all other grid
points is zero. Implementations of the vortex-in-cell method for vortex blobs are available [6].
Fast algorithms
A variety of computational techniques have been developed for reducing the O(N 2 ) operations of the
point-vortex method to O(N ln N ) or even O(N ) operations. Examples are tree codes and multipole
expansions [328]. One particular method involves successively discretizing the domain of ow into
smaller rectangles by subdividing a parent rectangle into four smaller rectangles until N rectangles
have been formed. Near-neighbors and well-separated rectangles are identied and, for the purpose
of computing velocities, point vortices that lie inside well-separated boxes are condensed into central
point vortices whose strength is equal to the sum of the strengths of the condensed point vortices.
Moore (1969) (see [265]) and Chorin [81] observed that, since point vortices and vortex blobs
in an inviscid uid maintain their original strength, viscous diusion can be emulated by allowing the
point vortices or vortex blobs to execute random motion in addition to the deterministic motion due
to mutual interactions. The diusion step displaces the position of the ith point vortex over one time
step by (xi , yi ), where x1 , y1 , . . . , xN , yN is a set of independent random displacement
whose probability density function forms a Gaussian distribution with zero mean and variance equal
to 2t, where is the kinematic viscosity. Chorin [81] applied the random vortex method to
simulate ow at high Reynolds number past a circular cylinder. Milinazzo & Saman [265] and
Roberts [347] showed that the casual application of the method may lead to signicant numerical
error due to slow convergence with respect to N . A large number of point vortices is required to
ensure statistical equilibrium.
Another way of emulating the eects of viscosity is to employ vortex blobs whose radius
increases in time due to diusion, as discussed in Section 11.4. The lack of random motion justies
calling these vortex methods deterministic (e.g., [328]).
Subtleties arise when we attempt to determine the motion of newly created blobs. An improved
version of the method involves discretizing the vortex sheet into elemental vortex patches using the
Prandtl boundary layer equations [83, 84]. Renements and improvements are available [328].
11.8 Two-dimensional vortex patches 893
Computer Problems
where x = x x , a is a chosen length, and l is the arc length around the contour measured in the
direction of the tangent unit vector, t.
The vorticity transport equation for two-dimensional ow in the absence of viscous forces
guarantees that the vorticity of the patch remains constant in time. Combining the contour integral
representation with the vorticity transport equation, we nd that the dynamics of the ow can be
described on geometrical grounds by following the evolution of the vortex contour alone.
(a) (b)
y
z
b
a x
Figure 11.8.1 (a) Illustration of a Rankine vortex of radius a. The uid executes rigid-body motion
inside the vortex and the velocity decays like that due to a point vortex outside the circular core.
(b) Illustration of Kirchhos elliptical vortex rotating with angular velocity z .
where r is the distance from the centerpoint. Outside a Rankine vortex of radius a, the ow is
identical to that due to a point vortex with strength = a2 situated at the center of the vortex,
and the polar velocity is
a2
u = . (11.8.3)
2 r
Note that the composite velocity is continuous, but the derivative of the velocity suers a disconti-
nuity across the vortex contour located at r = a.
r = a [1 + cos(m)], (11.8.4)
where 1 is a small dimensionless number, is the polar angle measured around the center of
the unperturbed vortex, and m > 1 is an integer determining the circumferential wave number. In
Cartesian coordinates, the vortex contour is described by the equations
x = xc () = a 1 + cos(m) cos , y = yc () = a 1 + cos(m) sin . (11.8.5)
Substituting these expressions into the contour integral representation (11.8.1), we derive an expres-
sion for the velocity induced by the deformed patch,
! " 2 2 . /
ux a 2 x xc () + y yc () sin + cos(m) cos
(x) = ln d, (11.8.6)
uy 4 0 a2 cos + cos(m) sin
Linearization
Next, we evaluate the velocity at a point on the vortex contour located at = 0 . As a preliminary,
we compute the squares of the x and y distances
2 2
(x)2 xc (0 ) xc () = a2 cos 0 cos + cos(m0 ) cos 0 cos(m) cos (11.8.7)
and
2 2
(y)2 yc (0 ) yc () = a2 sin 0 sin + cos(m0 ) sin 0 cos(m) sin . (11.8.8)
Normal velocity
The velocity component normal to the unperturbed circular contour arises by projecting the velocity
given in (11.8.12) onto the radial unit vector, n = [cos 0 , sin 0 ], nding
! " ! "
a 2 cos[m(w + 0 )] cos(w + 0 ) cos 0
un (0 ) = ln[1 cos w] d. (11.8.13)
4 0 cos[m(w + 0 )] sin(w + 0 ) sin 0
Simplifying the integrand, we obtain
a 2
un (0 ) = ln[1 cos w] cos[m(w + 0 )] sin w + m sin[m(w + 0 )] cos w dw, (11.8.14)
4 0
896 Introduction to Theoretical and Computational Fluid Dynamics
and then
2
a
un (0 ) = sin(m0 ) ln[1 cos w] sin(mw) cos w dw. (11.8.16)
4 0
Integrating by parts, we nd that the integral on the right-hand side is equal to 2 when m > 0,
and 2 when m < 0. Thus,
m a
un (0 ) = sin(m0 ). (11.8.17)
|m| 2
The calculations can be repeated with a sine instead of a cosine on the right-hand side of (11.8.4),
resulting in (11.8.17) with a negative cosine instead of a sine on the right-hand side.
Stability
Our analysis has indicated that, if
(, t) = a cos[m( ct)] (11.8.18)
is the normal displacement of the boundary velocity, then
m a
un (, t) = sin[m( ct)] (11.8.19)
|m| 2
is the velocity normal to the circular contour, where t stands for time and c is the phase velocity of
the contour wave. Kinematic compatibility requires that
u
+ un = 0, (11.8.20)
t a
where u = 12 a is the tangential velocity of the unperturbed vortex. Substituting (11.8.18) and
(11.8.19) into (11.8.20) and simplifying, we obtain
1
c + = 0. (11.8.21)
2 |m| 2
Rearranging, we derive an expression for the phase velocity rst deduced by Kelvin in 1880 [404],
1 1
c= 1 . (11.8.22)
2 |m|
Since c is real, boundary waves rotate with constant amplitude and the Rankine vortex is linearly
stable. The phase velocity increases from c = 14 when m = 2 for long waves to c = 12 as m =
for short waves.
We have analyzed the evolution of Kelvin waves around the Rankine vortex based on the
contour representation in a way that bypasses the computation of the pressure based on the Euler
or Bernoulli equation.
11.8 Two-dimensional vortex patches 897
where a and b are the semiaxes and is the natural parameter of the ellipse ranging in the interval
[0, 2). Without loss of generality, we assume that a b, as shown in Figure 11.8.1(b). Substituting
(11.8.23) into (2.13.23), we derive an integral representation for the induced velocity,
! " 2 ! "
ux (x a cos )2 + (y b sin )2 a sin
(x) = ln d. (11.8.24)
uy 4 0 a2 b cos
Contour velocity
The velocity at a point on the vortex contour corresponding to = 0 , located at x = a cos 0 and
y = b sin 0 , is
! " 2 ! "
ux R2 a sin
(0 ) = ln 2 d, (11.8.25)
uy 4 0 a b cos
where
Simplifying, we nd that
where
a2 b2
+ = + 0 , = 0 , = . (11.8.28)
a2 + b2
Substituting (11.8.27) into (11.8.25), we obtain
! " ! "
ux 2 a sin(+ 0 )
(0 ) = ln(1 cos + ) d+
uy 4 0 b cos(+ 0 )
2 ! "
a sin( + 0 )
+ ln(1 cos ) d . (11.8.29)
0 b cos( + 0 )
Simplifying, we obtain
! " ! "
ux a (A A1 ) sin 0
(0 ) = , (11.8.30)
uy 4 b (A + A1 ) cos 0
898 Introduction to Theoretical and Computational Fluid Dynamics
where
2
1 1 2 a2 b2
A() = ln(1 cos + ) cos + d+ = 2 = 2 , (11.8.31)
0 (a + b)2
and A1 A(1) = 2. Substituting the expressions for A and A1 into (11.8.30), we obtain
! " ! "
ux ab sin 0
(0 ) = . (11.8.32)
uy a+b cos 0
It is interesting that the tangential velocity at the tip of the vortex, 0 = 0, is the same as that at
the midpoint of the at side, 0 = 12 .
where = ab is the circulation around the vortex. We have shown that an elliptical vortex patch
rotates intact with angular velocity z given in (11.8.35). As the aspect ratio a/b increases while
the circulation around the vortex is held xed, the rotating elliptical vortex reduces to a rotating
vortex layer with elliptical distribution of circulation along the major axis. The limit a/b leads
us to a rotating vortex sheet (Problem 11.8.1).
Tangential velocity
Substituting (11.8.35) into (11.8.33), we obtain the velocity around the vortex contour in a frame of
reference rotating with the vortex,
! " ! "
vx a sin 0
(0 ) = z . (11.8.36)
vy b cos 0
where
1/2
J() = a2 sin2 + b2 cos2 (11.8.38)
is the metric of the arc length with respect to the native parameter around the elliptical shape.
Normal modes
To study normal mode perturbations with exponential dependence in time, we assume that the
shape function and normal velocity in the rotating frame behave as
q(, t) = Q() eit , vn (, t) = a Vn () eit , (11.8.44)
where is the complex growth rate, Q() is a complex dimensionless eigenfunction, and Vn () is
an accompanying dimensionless complex function. Substituting these expressions into (11.8.43), we
obtain
dQ
i Q() + z Vn () = 0. (11.8.45)
d
900 Introduction to Theoretical and Computational Fluid Dynamics
Given a linear relation between Vn and Q, we obtain an eigenvalue problem for the complex phase
velocity, . To ensure the satisfaction of the area-preserving condition (11.8.41), we now set
a
Q= (cos m + sin m), (11.8.46)
J
where m > 1 is the integer circumferential wave number and is an a priori unknown coecient
determining the phase of the normal mode with respect to the x axis. Substituting this expression
into (11.8.45), we nd that
dQc dQs
Substituting (11.8.52) into (11.8.49) and linearizing the logarithmic term, we obtain
! " 2 ! "
ux R2 R1 a sin + b ( cos )
(0 ) ln 2 + 2ab 2 d. (11.8.54)
uy 4 0 a R b cos + a ( sin )
Linearizing further, we nd that the velocity in a frame of reference rotating with the elliptical
vortex is given by
! " ! " 2 ! "
vx b sin 0 R2 a sin
(0 ) z ln 2 d
vy a cos 0 4 0 a b cos
2 ! " 2 ! " ! "
R1 a sin R2 b ( cos ) a sin 0
2ab d + ln 2 d + .
4 0 R2 b cos 4 0 a a ( sin ) z
b cos 0
(11.8.55)
The integrable logarithmic singularity in the second integral on the right-hand side of (11.8.56) can
be removed by expressing the integral in the form
2
R2
L(, 0 ) ln 2 L(0 , 0 ) ln[1 cos(0 0 )] d 2L(0 , 0 ) log 2. (11.8.58)
0 a
Because of the periodic integrands, the regularized integral can be computed eciently by the
trapezoidal rule. Machine accuracy can be achieved by using a large number of integration base
points.
Comparing (11.8.56) with the second expression in (11.8.44), we deduce the requisite dimen-
sionless function Vn () in relation to the dimensionless shape function Q() given in (11.8.46),
1 1 2 2 2
R1 2
R2
The procedure described in the paragraph following equation (11.8.48) may then be applied.
Love waves
Love [246] carried out a classical stability analysis and obtained the complex growth rate
z
2 a b
2m 1/2
= 2m 1 . (11.8.60)
2 a+b
902 Introduction to Theoretical and Computational Fluid Dynamics
C
i
t
N
1
2
Figure 11.8.2 Illustration of a vortex patch with constant vorticity, . The vortex contour, C, is
described by N marker points.
Perturbations with m = 2 are stable for any aspect ratio a/b; perturbations with m = 3 are stable
when a/b < 3; perturbations with m > 3 are stable below higher aspect ratio thresholds [270]. The
solution of the quadratic equation for the complex growth rate discussed in the paragraph following
equation (11.8.47) is in perfect agreement with Loves analytical predictions.
whose eigenvalues are equal to i. The corresponding eigenvectors are discrete representations
of the shape function Q(). The solution of the eigenvalue problem reproduces Loves eigenvalues
and corresponding eigenfunctions. Computer programs are available in directory kirchho inside
directory 09 vortex of the software library Fdlib (Appendix C).
for i = 1, . . . , N , where a is a constant length and t is the tangent unit vector pointing in the
counterclockwise direction around the vortex contour [113, 330]. To compute the integral in (11.8.62),
we interpolate for the shape of the contour C from the position of the marker points, as discussed
in Section B.4, Appendix B.
In the simplest implementation of the method, the vortex contour is approximated with a
polygonal line connecting a chain of marker points. To compute the contour integral on the right-
hand side of (11.8.62) over a segment that does not contain the ith marker point, Xi , we use a
standard numerical method, such as the trapezoidal rule, Simpsons rule, or a Gaussian quadrature,
as discussed in Section B.6, Appendix B. However, since the integrand in (11.8.62) exhibits a log-
arithmic singularity, these methods cannot be applied for the two segments that contain Xi as an
end point. The integration in these two cases can be done by elementary analytical methods [318].
Combining the trapezoidal rule for integration over the nonsingular segments with the ana-
lytical integration over the singular segments, we obtain
|Xi Xj |2
N
dXi |Xi Xj+1 |2
= (Xj+1 Xj ) ln 2
+ ln
dt 8 j=1 a a2
|Xj+1 Xj |2
(Xj+1 Xj ) ln 2
2 , (11.8.63)
4 j=i1,i
a
where the prime denotes that the terms j = i 1 and j = i are excluded from the sum. Equation
(11.8.63) provides us with a system of N nonlinear, coupled ordinary dierential equations for the
position of the marker points. The system can be integrated in time using Eulers method, a Runge
Kutta method, or another method, as discussed in Section B.8, Appendix B.
Multiple vortices
When the ow contains a collection of vortex patches with the same or dierent vorticity, the right-
hand side of (11.8.62) contains the sum of integrals over the contours of all patches multiplied by the
corresponding values of the vorticity, and equation (11.8.63) undergoes corresponding modications.
Figure 11.8.3 shows the evolution of two initially circular vortex patches with identical vorticity in
close proximity computed using a code encapsulated in the software library Fdlib (Appendix C).
We observe that the patches coalesce into a larger patch under the action of their mutually induced
velocity. Numerical investigation shows that the patches merge only if they are initially closer than
a critical distance.
Periodic arrangements
Using the results of Section 2.13.5, we nd that the motion of marker points distributed around the
contour of a vortex patch that is repeated periodically along the x axis with period a is governed
by the modied version of (11.8.62)
dXi
= ln cosh[k(Yi y)] cos[k(Xi x)] t(x) dl(x), (11.8.64)
dt 4 C
where C is the contour of one patch and k = 2/a is the wave number.
904 Introduction to Theoretical and Computational Fluid Dynamics
1 1 1
0 0 0
1 1 1
1 0 1 1 0 1 1 0 1
1 1 1
0 0 0
1 1 1
1 0 1 1 0 1 1 0 1
1 1 1
0 0 0
1 1 1
1 0 1 1 0 1 1 0 1
Figure 11.8.3 Evolution of two circular vortex patches with equal vorticity. Vortex contours are
shown at equal time intervals. The vortex patches are attracted and gradually coalesce into a
larger rotating elliptical vortex.
The contour of a patch can be approximated with a polygonal line connecting point particles.
To compute the improper integral over the adjacent segments of the ith point particle, Si1 and Si ,
we write
ln cosh[k(Yi y)] cos[k(Xi x)] t(x) dl(x) = I + J (11.8.65)
Sj
and
J ln k2 (Xi x)2 + k2 (Yi y)2 t(x) dl(x). (11.8.67)
Sj
The nonsingular integral in (11.8.66) can be computed with adequate accuracy using the trapezoidal
rule or a Gaussian quadrature. The singularity has been shifted to the integral in (11.8.67), which
can be computed by elementary methods.
11.9 Axisymmetric ow 905
Vortex layers
An example of an innite vortex patch is the innite vortex layer with constant vorticity discussed in
Section 9.6. When the boundaries of the vortex layer are parallel, the ow is unidirectional and the
velocity varies linearly with distance across the vortex layer and has two dierent constant values
above and below the vortex layer. Figure 9.6.5 shows stages in the instability of a periodically
perturbed vortex layer computed using the method of contour dynamics for periodic ow [323, 324].
Problem
Computer Problems
11.9 Axisymmetric ow
Vortex methods for axisymmetric vortex ow are extensions of those for two-dimensional ow dis-
cussed earlier in this chapter. However, the curvature of the vortex lines and the occurrence of vortex
stretching necessitates essential modications that demonstrate the subtlety of three-dimensional
vortex ows.
In contrast, an isolated rectilinear line vortex (point vortex) remains stationary. The formation,
motion, and stability of vortex rings has attracted a great deal of attention in the scientic and
popular literature [367].
If Vi is the self-induced velocity of the ith vortex ring that belongs to a collection of N coaxial
rings, the axial position, Xi , and radius, i , of the ring evolve according to the equations
dXi N
= Vi (i ) + j Ux (Xi Xj , i , j ),
dt j=1
(11.9.1)
di N
= j U (Xi Xj , i , j ),
dt j=1
where the prime indicates that the term i = j is excluded from the sum,
! " ! "
Ux 1 I31 (x, , ) + I30 (x, , )
(x, , ) = , (11.9.2)
U 4 x I31 (x, , )
and the integrals Inm are dened in (2.13.14). The vorticity transport equation requires that the
strength of each ring, i , remains constant in time.
Self-induced velocity
The divergence of the self-induced velocity of a line vortex ring with innitesimal core underscores
the importance of the core size. To compute the motion of a vortex ring, we may assume a sensible
vorticity distribution over the core, and thus transform the vortex ring into an axisymmetric vortex
blob. Kelvins formula for the self-induced velocity of a vortex ring whose vorticity is distributed
uniformly over a circular vortex core of radius a is
8 1
V = ln , (11.9.3)
4 a 4
where is the ring centerline radius, = a2 is the vortex ring strength, and = is the
uniform core vorticity ([220], p. 241). Hickss formula for the self-induced velocity of a vortex ring
whose vorticity is concentrated around a vortex sheet of radius a is identical to (11.9.3), except that
the factor 1/4 is replaced by 1/2 on the right-hand side [175]. For a ring with arbitrary core vorticity
distribution, we expect that the self-induced velocity is
8 1
V = ln + , (11.9.4)
4 a 2
where is a dimensionless constant taking the value of 1/4 for a Kelvin ring and the value of zero
for a Hicks ring (e.g., [125]).
As the radius of a ring, , changes during the motion, the azimuthal vorticity, , increases
or decreases by the same proportion. In response to this change, the radius of the core, a, adjusts
to preserve the ring strength and core volume, so that d(a2 )/dt = 0. Expanding the derivative, we
obtain an expression for the rate of change of the radius of the vortex core,
da 1 a d
= . (11.9.5)
dt 2 dt
11.9 Axisymmetric ow 907
The motion of the vortex rings is computed by simultaneously integrating in time equations (11.9.1)
and (11.9.5) for the axial position, radial position, and radius of the vortex core.
Bounded domains
When a vortex ring resides in a domain that is bounded by an impermeable axisymmetric surface,
the induced velocity must be complemented with that due to a potential ow that accounts for
the no-penetration boundary condition. When the boundary geometry is suciently simple, the
complementary ow can be expressed in terms of images of the vortex rings. The image of a vortex
ring with respect to a plane wall is another ring with opposite strength placed at the instantaneous
mirror-image position. The image of a ring with respect to a sphere was discussed in Section 7.5.
Contour dynamics
The numerical implementation of the contour dynamics method for axisymmetric ow is similar
to that for two-dimensional ow discussed in Section 11.8.3 (e.g., [302, 330]). Using (2.12.29) and
(2.12.37), we nd that the axial position, X, and radial position, , of point particles distributed
along the vortex contour evolve according to the equation
! " ! "
d X x I10 (x, , ) nx (x, ) + I11 (x, , ) n (x, )]
= dl(x), (11.9.6)
dt 4 C I11 (x, , ) nx (x)
where the integration is performed along the vortex contour, C, n is the normal unit vector pointing
outward from the vortex patch, x = X x, and the integrals Inm are dened in (2.13.14). Near the
evaluation point, X, the integrand in (11.9.6) behaves as shown in (11.8.62). The computation of
908 Introduction to Theoretical and Computational Fluid Dynamics
the singular integrals over the adjacent segments of a marker point is done by subtracting out the
singularity according to (11.8.66) (Problem 11.9.2).
Problems
Computer Problems
11.10 Three-dimensional ow
Vortex methods for three-dimensional ow are generalizations of those for two-dimensional and ax-
isymmetric ow discussed earlier in this chapter. However, the extensions are not always straightfor-
ward and the numerical implementation may lead to conceptual diculties and signicant numerical
error.
In Section 2.11.2, we saw that a curved line vortex with innitesimal cross-section moves with
an unphysical innite self-induced velocity. This singular behavior emphasizes that the centerline
curvature and the size of the vortex core plays a critical role in determining the motion of vortex
laments. Computing the self-induced motion of a line vortex with nite vortex core can be done
by several approximate methods. The point of departure in all cases is the BiotSavart integral
11.10 Three-dimensional ow 909
where x = x x , r = |x|, is the strength of the line vortex, and t is the tangent unit vector along
a line vortex, L. Since a line vortex moves with the uid velocity in the absence of viscous diusion,
the problem is reduced to solving an integro-dierential equation for the position of point particles
distributed along the line vortex, dX/dt = u(X).
grid labeled by discrete values of and . The strength of the vortex sheet, = n (u(1) u(2) ),
is tangential to the instantaneous position of the vortex sheet, where the superscripts (1) and (2)
designate evaluation at the upper or lower side of the vortex sheet and the normal unit vector n
points toward the upper side.
The marker points move normal to the vortex sheet with the common normal uid velocity
on either side of the vortex sheet, while executing an arbitrary tangential motion. The velocity of a
marker point can be expressed in terms of the principal velocity of the vortex sheet dened as the
principal value of the BiotSavart integral, as discussed in Section 11.5 for two-dimensional ow. To
compute the evolution of the strength of the vortex sheet, , we work as in Section 11.5, beginning
with Eulers equation on either side of vortex sheet written in the form of (11.5.12). The complexity
of the algebraic manipulations in the general case discourages a detailed derivation.
Equation (11.10.7) provides us with a basis for a numerical procedure according to the fol-
lowing steps: advance the position of the marker points with the principal velocity of the vortex
sheet; update the jump in the potential + according to (11.10.7); compute the surface gradi-
ent (+ ) over the vortex sheet; update the strength of the vortex sheet using the denition
= n (+ ) [64, 311].
11.10 Three-dimensional ow 911
N
(x) = i (t) 3 (x xi ), (11.10.8)
i=1
where 3 is the three-dimensional delta function. The ith term on the right-hand side of (11.10.8)
represents a vortex particle with strength i located at the point xi . The velocity eld arises by
taking the curl of the vector potential
1 i (t)
N
A(x) = , (11.10.9)
4 i=1 x xi (t)
computed according to (2.10.1). The vortex particles move with the uid velocity while their strength
evolves due to vortex stretching. Using equations (3.11.9) and (3.12.1), we obtain
dxi di
= u[xi (t), t], = i (t) [ u + (1 )(u)T ], (11.10.10)
dt dt
where is a free parameter. The choice = 0 helps preserve the total vorticity of the ow, which is
an invariant of the motion [431]. The alternative choice = 12 helps reduce the computational cost.
One diculty with the representation (11.10.8) is that, in general, the discretized vorticity
eld and vector potential are not solenoidal. Improvements can be made to rectify this deciency.
Vortex methods for three-dimensional ow are reviewed by Leonard [234], Kino & Ghonien [211],
Winckelmans & Leonard [431], and Puckett [328].
Computer Problem
It is helpful to keep in mind throughout this chapter that the particular way in which the
convectiondiusion equation enters a numerical procedure for computing the structure of a steady
ow or the evolution of an unsteady incompressible ow depends on the chosen computational
strategy. In certain cases, the convectiondiusion equation is integrated with reference to the
equation of motion. In other cases, the convectiondiusion equation is integrated with reference to
the vorticity transport equation. Examples in each category will be discussed in Chapter 13.
f
+ u(x, t, f ) f = 2 f (12.1.1)
t
in a specied one-, two-, or three-dimensional solution domain, subject to a given initial condition,
f (x, t = 0) = F(x), where F(x) is a prescribed function. The convection velocity, u, is a known
function of position, x, and time t, explicitly, as well as implicitly through its dependence on the
solution f . If u is constant, we obtain a linear convectiondiusion equation. If u depends on x and
t but is independent of f , we obtain a quasi-linear convectiondiusion equation.
912
12.1 Denitions and procedures 913
seemingly academic classication has important consequences on the eectiveness of the various
nite-dierence schemes.
To complete the statement of the computational problem, we must introduce a proper number
of boundary conditions. When the diusivity is nonzero, the convectiondiusion equation is a
second-order partial dierential equation and we must supply a number of boundary conditions that
matches the dimensionality of the unknown function f . Thus, if f is a three-dimensional vector, we
require three scalar conditions, one for each component of f over each boundary. Fewer boundary
conditions are needed when is zero.
Interpolation
Once a discrete nite-dierence solution is available, the values of the computed function between
grid points and time levels can be obtained by applying standard methods of function interpolation,
extrapolation, or approximation, as discussed in Sections B.4 and B.7, Appendix B.
Applying the nite-dierence equation sequentially at internal grid nodes provides us with a
system of linear or nonlinear algebraic equations that relate the values of the unknown functions at
914 Introduction to Theoretical and Computational Fluid Dynamics
the nodes. In certain cases, the solution domain is extended beyond the natural boundaries of the
physical problem and the nite-dierence equation is applied at boundary nodes for the purpose of
accurately implementing boundary conditions involving derivatives.
12.1.3 Consistency
The accuracy of a numerical computation based on a nite-dierence method depends on the control
parameters of the numerical method, including the grid spacing and time step. Assume that both
are reduced simultaneously but arbitrarily so that they may have dierent orders of magnitude. If
the nite-dierence equation approximates the partial-dierential equation with increasing accuracy
in some sensible fashion, then the nite-dierence method is consistent. However, if the grid spacing
and time step must be reduced simultaneously is special ways for the nite-dierence equation
to approximate the partial-dierential equation with increasing accuracy, then the nite-dierence
method is only conditionally consistent.
The consistency of a nite-dierence equation that arises by applying well established nite-
dierence formulas to approximate temporal and spatial derivatives in the partial dierential equa-
tion is guaranteed and does not need to be examined. By contrast, the consistency of a nite-
dierence equation that arises by heuristic or ad-hoc modications of well-established nite-dierence
approximations is subject to conrmation.
Undetermined coecients
Certain nite-dierence equations emerge by applying the dierential equation at a particular grid
point, and then replacing it with a combination of values of the unknown function at a group of
neighboring grid points at dierent times. The coecients that multiply the values of the function
are computed by imposing certain restrictions, including consistency with the dierential equation
and a desired degree of accuracy in approximating the partial derivatives.
does not grow unbounded in time, but either stays constant or decays at every point. It is not un-
reasonable to demand that the nite-dierence solution reproduces this behavior, that is, it provides
us with a numerical approximation that is free of oscillations. If it does, the nite-dierence method
is stable; otherwise the method is unstable. If the exact solution of the dierential equation grows
in time, the nite-dierence method is stable when it produces a numerical solution that grows at a
rate that not greater than that of the exact solution.
The stability of relatively simple nite-dierence methods for linear partial-dierential equa-
tions can be assessed by several methods, including the von Neumann stability method, the projec-
tion matrix method, and the discrete-perturbation method. The stability of involved nite-dierence
methods is harder to investigate. In practice, stability is often warranted by the absence of noticeable
spatial or temporal oscillations in the results of a computation.
The stability of nite-dierence methods for nonlinear dierential equations is typically ex-
amined by linearizing the dierential equation about a particular grid point and then studying the
performance of the nite-dierence method with reference to the linearized equation, as discussed
in Chapter 9 in the context of hydrodynamic stability. The local stability criteria obtained in this
manner provide us with an accurate characterization of the overall performance of the numerical
scheme.
12.1.5 Convergence
Stability imposes a modest restriction on the numerical method. Before the numerical results can
be claimed to bear any degree of physical relevance with respect to the physical problem governed
by the original partial-dierential equation, we must ensure that, as the size of the grid and time
step are made ner, the numerical solution converges to the exact solution.
f
+ u(f , x, t) f = 2 f . (12.1.2)
t
By contrast, the primary equation (12.1.1) expresses the nonconservative form. This terminology
emphasizes that the components of the matrix u f at the grid points telescope up to the bound-
aries, and thus conserve possible invariants of the solution in a nite-dierence discretization. The
conservative form is preferred over the nonconservative form due to improved accuracy and superior
916 Introduction to Theoretical and Computational Fluid Dynamics
numerical stability. In the case of incompressible uid ow, both the NavierStokes equation and the
vorticity transport equation can be recast into a conservative form similar to that shown in (12.1.2).
Problems
where h is a small interval. Derive relations among the coecients ac and AC so that the
nite-dierence approximations are consistent. This means that, in the limit as h tends to zero, the
approximations reproduce the exact values of the rst and second derivatives at x. Derive additional
relations among the coecients so that the discretization error is of second order in h and solve for
the coecients ac and AC.
(b) Compute the coecients ac and AC so that the error of the backward or forward dierence
approximations
is of second order in h.
(c) Compute the coecients ae so that the error of the central dierence approximation
is of fourth order in h.
(d) Approximating the rst derivative at x = 0 with a forward dierence, we obtain the discrete
boundary condition f2 f1 = 2x. Collect all nite-dierence equations into a linear system,
A f = b, where A is a tridiagonal matrix,
f = [ f1 , f2 , . . . , fN ] (12.1.7)
is the N -dimensional solution vector, and b is a known vector. Present the explicit forms of the
matrix A and vector b.
(e) The nite-dierence equation (12.1.6), is second-order accurate in x, whereas the corresponding
nite-dierence equation for the boundary condition at x = 0 derived in (d) is rst-order accurate in
x. This discrepancy somewhat compromises the overall accuracy of the nite-dierence method.
To obtain a fully consistent second-order method, we extend the solution domain beyond the natural
boundary at x = 0, introduce a ctitious exterior node, x0 = x, apply equation (12.1.6) for i = 1,
and approximate the derivative with a central dierence to obtain the discrete boundary condition
f2 f0 = 4x. Derive the corresponding linear system, B f = c, and present the explicit forms
of the tridiagonal coecient matrix B and known vector c.
Computer Problems
for < x < and t > t0 (e.g., [66], p. 53). At the initiation instant, t = t0 , the Greens function
is the one-dimensional delta function.
In terms of the Greens function, the solution of the unsteady diusion equation (12.2.1) is
constructed by superposition as
f (x, t) = F (x + ) G(, t) d, (12.2.4)
where the integration variable, , is the distance from the evaluation point, x. Substituting the
Greens function, we obtain the exact solution
1 2
f (x, t) = F (x + ) exp d (12.2.5)
4t 4t
for t > 0. The integral on the right-hand side can be computed numerically using the Gauss-Hermite
quadrature [317].
The goal of the nite-dierence method is to provide us with the values of the function, fin ,
at the grid points, xi , where i = 1, . . . , N + 1, at a sequence of successive time levels, tn , beginning
at the initial time level, t0 = 0, subject to the boundary conditions f1n = 0 and fN n
+1 = 0. By
denition,
fin f (xi , tn ). (12.2.6)
The initial condition species that fi0 = F (xi ).
12.2 One-dimensional diusion 919
(a) (b)
t
t
n +1
n
fi
n n
t x
1 n 1
x i1 i i+1
a b x
0
1 2 i N N+1
Figure 12.2.1 (a) Discretization of the space-time domain for solving the one-dimensional diusion
equation. The solution is advanced forward in time from the initial level, t = 0. (b) Computational
stencil of the forward-time, centered-space (FTCS) discretization.
fin+1 = fi1
n
+ (1 2) fin + fi+1
n
, (12.2.8)
where
t
(12.2.9)
x2
is a positive dimensionless constant called the diusion number. Equations (12.2.7) and (12.2.8) can
be applied at the internal grid points, i = 2, . . . , N , but not at the boundary points, i = 1, N + 1.
Equation (12.2.8) provides us with a straightforward algorithm for computing the value of f
at a grid point at the time level n + 1 in terms of the values of f at three grid points at the previous
time level, n. Since the algorithm does not require solving a system of algebraic equations, it is
classied as explicit. In summary, the FTCS discretization provides us with an explicit two-level
method with rst-order accuracy in time and second-order accuracy in space.
920 Introduction to Theoretical and Computational Fluid Dynamics
Consistency
To conrm the consistency of the dierence equation (12.2.8), we regard all discrete variables as
continuous functions of space and time, expand them in Taylor series about the point (xi , tn ), and
simplify to obtain the associated modied dierential equation
1 1
ft + ftt t + O(t2 ) = fxx + fxxxx x2 + O(x4 ), (12.2.10)
2 12
where subscripts denote partial derivatives with respect to corresponding variables, and all variables
are evaluated at xi and tn . Since, in the limit as t and x tend to zero independently, the
modied dierential equation (12.2.10) reduces to the partial dierential equation (12.2.1), the
FTCS discretization is consistent.
Dierentiating the governing equation (12.2.1) once with respect to t and twice with respect
to x, and combining the resulting equations, we derive a fourth-order dierential equation,
ftt = 2 fxxxx . (12.2.11)
Eliminating fxxxx on the right-hand side of (12.2.10) in favor of ftt and combining the resulting
expression with the second term on the left-hand side, we nd that, when = 1/6, the accuracy of
the FTCS discretization becomes of second order in both space and time.
Successive mapping
To formalize the action of the FTCS method, we collect the grid values of fin at the grid points
i = 2, . . . , N into a vector, f n , and use (12.2.8) in conjunction with the Dirichlet boundary conditions
f1n = 0 and fN n
+1 = 0 to obtain a linear system,
f n+1 = B f n , (12.2.12)
where B is a (N 1) (N 1) tridiagonal matrix with superdiagonal, diagonal, and subdiagonal
elements equal to , 1 2, and . This matrix form shows that the solution vector, f n+1 , arises by
projecting the vector f n onto the matrix B, and thereby establishes a relation between time stepping
and successive mapping in vector space.
Whether the vector f n will amplify or shrink during the successive mappings depends on
the spectral radius of the projection matrix, B, denoted by (B) and dened as the maximum
magnitudes of the eigenvalues of B. The theory of matrix calculus instructs us that, if (B) is
equal to unity, less than unity, or larger than unity, the length of f n will stay constant, decrease,
or increase during the successive mappings (e.g., [317]). Since, according to the exact solution, the
magnitude of f decays due to diusion at long times, we tolerate the rst behavior, welcome the
second behavior, and dismiss the third behavior as numerically unstable.
for m = 1, . . . , N 1 (e.g., [5] p. 57; [317]). This expression demonstrates that the spectral radius
of B is less than unity only when < 12 , and the FTCS method is conditionally stable.
If boundary conditions other than the homogeneous Dirichlet conditions f1n = 0 and fN n
+1 = 0
are specied at one end or both ends of the computational domain, the mapping matrix B will be
slightly altered. However, the performance of the nite-dierence method will still be determined by
the spectral radius, (B). As an example, we impose the Neumann boundary condition at the left
end, f /x = q at x = a, and retain the homogeneous Dirichlet condition at the right end, x = b.
To implement the left-end boundary condition with second-order accuracy, we extend the solution
domain beyond the physical boundary, x = a, introduce a ctitious node, x0 = x1 x, and use a
central dierence to obtain f2n f0n = 2qx. Having extended the solution domain, we may apply
the dierential equation at the rst node, x1 , and write (12.2.8) for i = 1.
To obtain the corresponding projection matrix, B, we collect the values of fin at the grid
points i = 1, . . . , N into a vector, f n , and combine (12.2.8) with the discretized Neumann boundary
condition to obtain a linear system, f n+1 = B f n + b, where B is an N N tridiagonal matrix with
superdiagonal, diagonal, and subdiagonal elements respectively equal to , 1 2, and , except
that the second entry of the rst row is equal to 2. All entries of the vector b are zero, except for
the rst entry that is equal to 2qx . The presence of the vector b does not aect the signicance
of the projection matrix B regarding the behavior of the numerical solution. When (B) is equal to
unity, less than unity, or larger than unity, the length of f n will stay constant, decrease, or increase
during the successive mappings. Unfortunately, the eigenvalues of B are not known in analytical
form.
The simple structure of the projection matrix associated with the FTCS method subject to Dirichlet
boundary condition at both ends allowed us to compute its eigenvalues and spectral radius exactly,
and thereby assess the stability of the numerical method without any approximations. Analytical
expressions for the eigenvalues of the projection matrix may not be available for more advanced
nite-dierence discretizations and more general boundary conditions. We may certainly compute
the eigenvalues using a numerical method, but this can be an arduous computational task.
Another way to assess the stability of the numerical method is by performing the von Neumann
stability analysis of the nite-dierence equation, neglecting the boundary conditions. The basic
idea is to examine the behavior of the numerical solution subject to a sinusoidal initial condition
with a prescribed wavelength, L. Motivated by the linearity of the governing equation, we separate
the temporal from the spatial dependencies by writing
where i is the imaginary unit, = 2x/L is the phase angle, and An is a constant coecient
dependent on the time level, n. Note that the superscript n is a time index, not an exponent.
922 Introduction to Theoretical and Computational Fluid Dynamics
The eciency of the von Neumann stability analysis is now apparent. One limitation of
the method is that, in its simple form described in this section, it does not incorporate the eect
of general boundary conditions, which may have a destabilizing inuence on the nite-dierence
method.
To examine the stability of the method, we substitute (12.2.14) into (12.2.17), set
An+1 An
= = G, (12.2.18)
An An1
and obtain a quadratic equation for the gain, G2 +G1 = 0, where = 2(1cos ) = 4 sin2 (/2)
is a real non-negative parameter. The solution is
1
G= ( 2 + 4)1/2 . (12.2.19)
2
12.2 One-dimensional diusion 923
Since the magnitude of the root corresponding to the minus sign is higher than unity, the CTCS
method is unconditionally unstable and thus of no practical value.
Performing the von Neumann stability analysis, we nd that the amplication factor satises
the quadratic equation
(1 + 2) G2 4 cos G 1 + 2 = 0. (12.2.22)
Examining the roots, we nd that |G| < 1 for any value of , which ensures that the Du FortFrankel
method is unconditionally stable.
Consistency
Since (12.2.20) was derived by an ad-hoc modication of the well-founded CTCS discretization, its
consistency must be examined by comparing the associated modied dierential equation (MDE)
with the given dierential equation (12.2.1). To derive the MDE, we regard all discrete variables
in (12.2.20) as continuous functions of space and time, expand them in Taylor series about (xi , tn ),
and simplify to obtain
t
2
ft = fxx ftt . (12.2.23)
x
We note that equation (12.2.23) is an accurate approximation of (12.2.1) only if the ratio (t/x)2
is suciently small, which requires that t is suciently smaller than x in appropriate units.
This restriction is milder than that arising from a stability constraint on the diusion number, .
We observe that (12.2.23) is classied as a hyperbolic dierential equation due to the presence of
the second derivative with respect to time, whereas (12.2.1) is classied as a parabolic dierential
equation, and conclude that adding a term with a wave-like character has a stabilizing inuence on
the numerical method.
Because of its advantages regarding numerical stability, the Du FortFrankel method enjoyed
widespread popularity in early numerical implementations. When using this method, care must be
taken so that the ratio (t/x)2 is suciently small, otherwise the results will not be physically
meaningful.
924 Introduction to Theoretical and Computational Fluid Dynamics
n +1 n +1 n +1
n +1/2
n n n
x x x
n 1 n 1 n 1
i1 i i+1 i1 i i+1 i1 i i+1
Figure 12.2.2 (a) Computational stencil of the BTCS discretization; the solution at three grid points
at the new time level is computed in terms of one current value. (b) Computational stencil of the
CrankNicolson discretization; the solution at three grid points at the new time level is computed
in terms of three current values. (c) Computational stencil of a three-level implicit method; the
solution at the new time level is computed in terms of values at two previous time levels.
Applying (12.2.1) at the point xi at the time instant tn+1 , and approximating the time deriva-
tive with a backward dierence and the space derivative with the central dierence, we obtain the
BTCS dierence equation
fi1
n+1
+ (1 + 2)fin+1 fi+1
n+1
= fin , (12.2.25)
where is the diusion number dened in (12.2.9). The presence of three unknown nodal values on
the left-hand side renders the BTCS method implicit.
Recasting (12.2.25) into a matrix form and implementing homogeneous Dirichlet boundary
conditions, we obtain a system of linear equations, A f n+1 = f n , where A is an (N 1) (N 1)
tridiagonal matrix with superdiagonal, diagonal, and subdiagonal elements equal to , 1 + 2,
and . Solving for f n+1 yields f n+1 = A1 f n , where A1 is the inverse of A. This expression
shows that stepping in time is tantamount to successively mapping the initial vector f 0 with the
projection matrix A1 . In practice, in order to compute the solution at the n + 1 time level, we solve
12.2 One-dimensional diusion 925
a system of linear algebraic equations, A f n+1 = f n , which renders the BTCS method implicit.
Since the coecient matrix A is tridiagonal and diagonally dominant, the linear system can be
solved eciently using the Thomas algorithm discussed in Section B.1.4, Appendix B.
To study the stability of the BTCS method, we consider the eigenvalues of the projection
matrix A1 = (C + I)1 , where C is a tridiagonal matrix with superdiagonal, diagonal, and
subdiagonal elements equal to 1, 2, and 1. Recalling that the eigenvalues of the inverse of a matrix
are equal to the inverse of the eigenvalues, and using (12.2.13), we obtain
1
m = , (12.2.26)
1 + 4 sin2 m
2N
where m = 1, . . . , N 1. Since the spectral radius of the projection matrix is less than unity, the
method is unconditionally stable. An independent way of arriving at this result is by performing
the von Neumann stability analysis, obtaining the amplication factor
1 1
G= = , (12.2.27)
1 + 2(1 cos ) 1 + 4 sin2 2
The low-order temporal accuracy of the BTCS method places a restriction on the size of the
time step for an accurate solution.
fi1 2fi n
fi1
n+1
+ 2 (1 + )fin+1 fi+1
n+1 n
= fi1 + 2 (1 )fin fi+1
n
. (12.2.29)
Deriving and examining the corresponding modied dierential equation shows that the Crank
Nicolson method is consistent and second-order accurate in both time and space.
Stability
Recasting (12.2.29) into a matrix form, we obtain a system of linear equations, A f n+1 = B f n ,
where A is a tridiagonal matrix with superdiagonal, diagonal, and subdiagonal elements equal to
926 Introduction to Theoretical and Computational Fluid Dynamics
, 2(1 + ), and , and B is another tridiagonal matrix with superdiagonal, diagonal, and
subdiagonal elements equal to , 2(1 ), and . Solving for f n+1 , we nd that
f n+1 = A1 B f n , (12.2.30)
which shows that stepping in time is equivalent to mapping successively with the projection matrix
P = A1 B. It can be shown that the spectral radius of the projection matrix is always less than
unity, thus ensuring that the CrankNicolson method is unconditionally stable (Problem 12.2.2).
Carrying out the von Neumann stability analysis, we derive the amplication factor
Because of its remarkable properties regarding accuracy and stability, the CrankNicolson
method has become a standard choice in practice.
Multiple substeps
It is instructive to regard the CrankNicolson method as the implementation of two substeps, where
each substep lasts for time interval 12 t. The rst substep is carried out using the explicit FTCS
method, while the second substep is carried out using the implicit BTCS method according to the
nite-dierence equations
n+1/2 n+1/2
fi fin n
fi1 2fin + fi+1
n
fin+1 fi n+1
fi1 2fin+1 + fi+1
n+1
1 = , 1 = . (12.2.32)
2 t
x2 2 t
x2
Adding these equations to eliminate the intermediate variables at time tn+1/2 reproduces (12.2.28).
The unconditional stability of the second substep prevails over the conditional stability of the rst
substep and renders the method overall unconditionally stable.
The particular choice = 12 provides us with a method that is second-order accurate in time
and capable of suppressing small-scale oscillations. These features render the dierence equation
(12.2.33) with = 12 preferable over the CrankNicolson method when the solution exhibits sharp
spatial variations. The choice = 12 (1+ 6
1
) provides us with a method that is second-order accurate
in time and fourth-order accurate in space.
12.2 One-dimensional diusion 927
Problems
1 2 sin2 (m/2N )
m (A1 B) = (12.2.34)
1 + 2 sin2 (m/2N )
for m = 1, . . . , N 1. Show that the spectral radius of the projection matrix is less than unity and
therefore the method is unconditionally stable.
f n 2fin + fi+1
= i1 i
2
i+1
+ (1 ) i1 2
, (12.2.35)
t x x
where is a positive parameter ranging in the interval [0, 1]. When = 0, we obtain the explicit
FTCS method; when = 1/2, we obtain the implicit CrankNicolson method; when = 1, we
obtain the implicit BTCS method.
(a) Show that the method is generally rst-order accurate in t and second-order accurate in x. When
= 12 (1 6
1
), the accuracy of the method increases to second order in t and fourth order in x.
(b) Provide an interpretation of the method in terms of a sequence of two elementary substeps.
(c) Perform the von Neumann stability analysis to derive the amplication factor
1 4(1 ) sin2
2
G= . (12.2.36)
1 + 4 sin2
2
Show that the method is unconditionally stable when 12 < < 1, and conditionally stable when
0 < < 12 . Derive the stability restriction 2 < 1/(1 2) ([343], p. 189). When = 12 , the critical
value of is shifted to innity and the method becomes unconditionally stable.
12.2.4 Dispersion
The dispersion equation in one dimension reads
f 3f
= , (12.2.37)
t x3
where the constant is the dispersion coecient. A solution describing the propagation of a si-
nusoidal wave is f (x, t) = exp[i k(x ct)], where k is a real wave number, c = k 2 is the phase
928 Introduction to Theoretical and Computational Fluid Dynamics
velocity of the traveling wave, and i is the imaginary unit. Since the phase velocity depends on the
wavelength, a wave packet disperses as each constituent wave travels with its own phase velocity.
Physically, the dispersion equation acts to propagate an arbitrary initial distribution while modifying
its shape.
(a) The FTCS discretization leads to an explicit ve-point two-level scheme described by the nite-
dierence equation
1 1
fin+1 = fi2
n n
+ fi1 + fin fi+1
n n
+ fi+2 , (12.2.38)
2 2
where = t/x3 . The numerical error is of order t and x2 . Carry out the von Neumann
stability analysis to derive the amplication factor
f 4f
= 4 , (12.2.42)
t x
where is a positive constant called the hyperdiusivity or fourth-order diusivity.
(a) The FTCS discretization leads to the explicit ve-point dierence equation
fin+1 = fi2
n n
+ 4fi1 + (1 6 ) fin + 4fi+1
n
fi+2
n
, (12.2.43)
where = t/x4 , with rst-order accuracy in time and second-order accuracy in space. Carry
out the von Neumann stability analysis, compute the amplication factor, and assess the stability
of the method.
(b) Repeat (a) for the implicit ve-point two-level BTCS method.
(c) Develop the counterpart of the CrankNicolson method.
12.3 Diusion in two and three dimensions 929
(i)
uj = rj sin (12.2.45)
K +1
Computer Problem
f 2f 2f f 2f 2f 2f
= ( + ), = ( + + ), (12.3.1)
t x2 y 2 t x2 y2 z 2
subject to a specied initial condition, f (x, t = 0) = F (x), where F (x) is a known function. For
simplicity, we assume that the solution domain extends over the whole two-dimensional plane or
three-dimensional space, and the function f decays to zero far from the origin.
where
t t
x = , y = (12.3.7)
x2 y 2
are the diusion numbers in the x and y directions. An alternative form of (12.3.6) is
1 n
n+1
fi,j n
= x (fi+1,j n
+ fi1,j )+2 1 fi,j n
+ (fi,j+1 n
+ fi,j1 ) , (12.3.8)
2x
where = (x/y)2 . As tends to zero or innity, we recover the FTCS formula for one-
dimensional diusion along the x or y axis.
Carrying out the von Neumann stability analysis, we derive the amplication factor
x y
G = 1 4 (x sin2 + y sin2 ), (12.3.9)
2 2
which shows that the FTCS method in two dimensions is stable provided that x + y < 12 , or
1
x < . (12.3.10)
2 (1 + )
This requirement imposes a strong constraint on the size of the time step and renders the explicit
discretization inecient. Similar diculties are encountered in three dimensions.
with rst-order accuracy in time and second-order accuracy in space. Rearranging, we derive the
implicit formula
x (fi1,j
n+1 n+1
+ fi+1,j n+1
) + (1 + 2x + 2y ) fi,j y (fi,j1
n+1 n+1
+ fi,j+1 n
) = fi,j . (12.3.12)
where n + 12 is an intermediate time level and the square brackets on the right-hand sides enclose
implicit discretizations. The rst substep is carried out using the implicit BTCS method for the x
direction, while the second substep is carried out using the implicit BTCS method for the y direction.
To eliminate the bias associated with this particular arrangement, the order is alternated after the
completion of each step. The overall method is second-order accurate in time and space.
Rearranging the dierence equations, we obtain a two-step implicit algorithm representing the
x and y sweeps,
n+1/2 n+1/2 n+1/2
x fi1,j + 2(1 + x ) fi,j x fi+1,j = y fi,j1
n
+ 2(1 y ) fi,j
n n
+ y fi,j+1 (12.3.15)
and
n+1/2 n+1/2 n+1/2
y fi,j1
n+1 n+1
+ 2(1 + y ) fi,j y fi,j+1
n+1
= x fi1,j + 2(1 x )fi,j + x fi+1,j . (12.3.16)
Completing each time step requires solving two systems of tridiagonal equations, which can be done
eciently using the Thomas algorithm described in Section B.1.4, Appendix B.
Carrying out the von Neumann stability analysis, we obtain the amplication factor
(1 2x sin2 2y sin2
x y
2
)(1 2
)
G= 2 x 2 y
. (12.3.17)
(1 + 2x sin 2 )(1 + 2y sin 2 )
Cursory examination reveals that |G| < 1 under any conditions, which ensures that the ADI method
is unconditionally stable. The second-order accuracy combined with the unconditional stability are
the reasons that the ADI method is a standard choice in practice.
(2 x 2x )(2 y 2y )fi,j
n+1
= (2 + x 2x )(2 + y 2y )fi,j
n
. (12.3.20)
The spatial decoupling is reected in the factorized nature of the dierence operators on either side.
(2 x 2x y 2y )fi,j
n+1
= (2 + x 2x + y 2y )fi,j
n
, (12.3.21)
(2 x 2x )(2 y 2y )fi,j
n+1
= (2 + x 2x )(2 + y 2y )fi,j
n
+ x y 2x 2y (fi,j
n+1
fi,j
n
). (12.3.22)
The ADI form (12.3.20) derives from the CrankNicolson form (12.3.22) by neglecting the last
term on the right-hand side. This simplication is permissible, for the neglected term is of fourth-
order in the spatial variables, whereas (12.3.20) is meant to be accurate only up to second order in
the spatial steps. We may say that the ADI method results from the approximate factorization of
the dierence operators on either side of the CrankNicolson equation (12.3.21).
The line GaussSeidel (LGS) scheme shown in the fourth entry of Table 12.3.1 arises from
the semi-implicit discretization of the Poisson equation. The relaxation parameter of the line
successive over-relaxation (LSOR) scheme shown in the last entry of Table 12.3.1 varies in the range
[0, 2]. The implicit LGS and LSOR schemes require solving tridiagonal systems of equations for each
grid line parallel to the x axis, which can be done eciently using the Thomas algorithm. Similar
equations can be written for the y axis.
To develop the ADI method, we introduce the parameter = 2/x , and recast the ADI equa-
tions (12.3.15) as shown in the rst entry of Table 12.3.2. Since the ADI method is unconditionally
934 Introduction to Theoretical and Computational Fluid Dynamics
Explicit FTCS
n+1
fi,j n
= fi,j n
+ x fi+1,j n
+ fi1,j 2 (1 + ) fi,j
n n
+ (fi,j+1 n
+ fi,j1 ) + x2 gi,j
fi1,j
n+1 n+1
+ 2 (1 + ) fi,j fi+1,j
n+1 n
= (fi,j1 n
+ fi,j+1 ) + x2 gi,j
fi1,j
n+1 n+1
+ 2 (1 + ) fi,j fi+1,j
n+1
= 2 (1 ) (1 + ) fi,j
n n
+ (fi,j1 n
+ fi,j+1 ) + x2 gi,j
Table 12.3.1 Iterative methods for solving the Poisson equation, 2 f + g = 0, on a uniform rect-
angular grid, where = (x/y)2 . All methods are second-order accurate in x and y. The
relaxation parameter in the PSOR and LSOR schemes ranges between 0 and 2.
stable, it might appear that the fastest approach to the steady state will be achieved using a large
value for t or a small value for . However, careful analysis shows that the minimum number of
iterations for a specied level of accuracy is achieved when a certain repetitive sequence of values
for is employed. In its generic implementation, the performance of the ADI method is comparable
to that of the line successive over-relaxation method (LSOR) ([5]; [184], p. 446).
To develop the SOR-ADI method shown in the second entry of Table 12.3.2, we set = 2 in
the rst ADI equation and = 2 in the second ADI equation so that the middle terms on the right-
hand sides disappear. A relaxation parameter, , is then introduced to accelerate the convergence
([186], Vol. II, p. 9). To demonstrate the consistency of the SOR-ADI scheme, we assume that the
solution depends only on x. The rst SOR-ADI equation yields
x2 gi,j .
n+1/2 n+1/2 n+1/2
fi1,j + 2(1 + )fi,j fi+1,j = 2(1 + )fi,j
n
+ (12.3.23)
2
12.3 Diusion in two and three dimensions 935
ADI
+ x2 gi,j ,
n+1/2 n+1/2 n+1/2
fi1,j + (2 + ) fi,j fi+1,j = fi,j1
n
+ ( 2 ) fi,j
n n
+ fi,j+1
+ fi+1,j + x2 gi,j
n+1/2 n+1/2 n+1/2
fi,j1
n+1 n+1
+ ( + 2 ) fi,j fi,j+1
n+1
= fi1,j + ( 2) fi,j
SOR-ADI
n+1/2 n+1/2 n+1/2
fi1,j + 2 (1 + ) fi,j fi+1,j
n
= fi,j1 + 2 (1 + ) (1 ) fi,j
n n
+ fi,j+1 + x2 gi,j ,
fi,j1
n+1 n+1
+ 2 (1 + ) fi,j fi,j+1
n+1
+ fi+1,j + x2 gi,j
n+1/2 n+1/2 n+1/2
= fi1,j + 2 (1 + ) (1 ) fi,j
Table 12.3.2 ADI methods for solving the Poisson equation, 2 f + g = 0, on a uniform Cartesian
grid, where = (x/y)2 . All methods are second-order accurate in x and y.
When = 1 + , we obtain
fi+1,j = x2 gi,j ,
n+1/2 n+1/2 n+1/2
fi1,j + 2fi,j (12.3.24)
which is the central dierence discretization of the ordinary dierential equation d2 f /dx2 + g = 0.
Douglas [111] developed an ADI method that is second-order accurate in both time and space
and unconditionally stable. The method proceeds in a predictorcorrection sense in three substeps.
The rst substep produces a predicted solution using the CrankNicolson method for the x direction,
while treating the y and z directions explicitly according to the dierence equation
1 1
(3 x 2x )fi,j,k
= (3 + x 2x + y 2y + z 2z )fi,j,k
n
, (12.3.25)
2 2
936 Introduction to Theoretical and Computational Fluid Dynamics
where an asterisk denotes the rst provisional solution. The second substep produces a predicted
solution using the x discretization of the rst substep and the CrankNicolson discretization for the
y direction, while treating the z direction explicitly according to the dierence equation
1 1 1 1
(3 y 2y )fi,j,k
= (3 + x 2x + y 2y + z 2z )fi,j,k
n
+ x 2x fi,j,k
, (12.3.26)
2 2 2 2
where a double asterisk denotes the second provisional solution. The third substep advances the
solution using the x and y discretizations of the second substep, while using the CrankNicolson
method for the z direction according to the dierence equation
1 1 1 1 1 1
(3 z 2z )fi,j,k
n+1
= (3 + x 2x + y 2y + z 2z ) fi,j,k
n
+ x 2x fi,j,k
+ y 2y fi,j,k
.
2 2 2 2 2 2
(12.3.27)
+ (x y x y + x z x z + y z y z ) x y z 2x 2y 2z fi,j,k
2 2 2 2 2 2 n
. (12.3.32)
4 8
The right-hand side can be rearranged to give
1 1 1
(3 x 2x )(3 y 2y )(3 z 2z )fi,j,k
n+1
2 2 2 (12.3.33)
1 1 1 1
= (3 + x 2x )(3 + y 2y )(3 + z 2z )fi,j,k
n
x y z 2x 2y 2z fi,j,k
n
,
2 2 2 4
which reveals a nearly factorized form. Performing the von Neumann stability analysis, we nd that
the method is unconditionally stable [317].
12.3 Diusion in two and three dimensions 937
f 2f f 2f
= 2, = 2, (12.3.34)
t x t y
which amounts to allowing diusion to operate sequentially in the two directions, each time neglecting
the other dimension. Each fractional step proceeds for the full time interval, t, and time is reset
to the initial value at the end of the rst fractional step. To preserve the spatial isotropy of the
Laplacian operator, the order of the equations in (12.3.34) can be switched after completing a full
time step. The fractional steps can be carried out using dierent methods for the one-dimensional
component diusion equations, as discussed in Section 12.2. The stability restrictions of the overall
method is the union of the restrictions pertaining to the individual fractional steps.
CrankNicolson method
Using the CrankNicolson method for each fractional step in two dimensions, we obtain the dierence
equations
1 1 1 1
(1 x 2x )fi,j
= (1 + x 2x )fi,j
n
, (1 y 2y )fi,j
n+1
= (1 + y 2y )fi,j
, (12.3.35)
2 2 2 2
where an asterisk denotes an intermediate solution. Eliminating the intermediate solution yields the
ADI equation in two dimensions. We conclude that the algorithm is second-order accurate in time
and space, while enjoying unconditional stability.
Problems
discretization corresponds to = 1. Perform the von Neumann stability analysis to derive the
amplication factor
1 4(1 ) x sin2 ( 12 x ) + y sin2 ( 12 y )
G= . (12.3.37)
1 + 4 x sin2 ( 12 x ) + y sin2 ( 12 y )
Show that, when > 12 , |G| < 1 for any values of x , y , x , and y , and the method is stable;
whereas when < 12 , the method is stable only if
1
x + y < . (12.3.38)
2(1 2)
1
Explain why the CrankNicolson method corresponding to = 2
is unconditionally stable.
Computer Problems
12.4 Convection
Having discussed the extreme case of pure diusion in one two and three dimensions, now we turn
our attention to the opposite extreme case of pure convection.
F(x)
characteristic
x
Figure 12.4.1 According to the linear convection equation, the initial distribution travels along the
x axis with velocity U . The dashed line is a characteristic line along which the solution remains
constant.
where U is a positive or negative constant convection velocity. For simplicity, we assume that the
solution domain extends over the whole x axis and stipulate the far-eld conditions f (, t) = 0.
If the solution domain were nite, we would have to specify one boundary condition at one end. An
initial condition stipulating that f (x, t = 0) = F (x) is provided, where F (x) is a known function.
One may readily verify that the exact solution is given by f (x, t) = F (x U t), which states
that the initial distribution, F (x), travels along the x axis with constant velocity U , as illustrated in
Figure 12.4.1. If the velocity U is negative, the motion occurs toward the negative direction of the
x axis. The value of the function f (x, t) remains constant along the characteristic lines x = + U t
in the xt plane, as shown with the dashed line in Figure 12.4.1, where is an arbitrary point.
As a preliminary, we dierentiate (12.4.1) twice, once with respect to t and the second time
with respect to x, and combine the resulting expressions to obtain the wave equation
2f 2f
2
= U2 , (12.4.2)
t x2
which is a prototypical second-order hyperbolic partial dierential equation. The hyperbolic nature
of the rst-order equation (12.4.1) thus becomes apparent. A function that satises the convection
equation (12.4.1) also satises the wave equation (12.4.2).
The numerical error due to the temporal and spatial discretizations is of order t and x2 . Since this
dierence equation emerged by applying standard nite-dierence approximations, its consistency
is guaranteed and need not be examined. Solving for fin+1 , we obtain the explicit formula
c n c n
fin+1 = f + fin fi+1 , (12.4.4)
2 i1 2
where
U t
c= (12.4.5)
x
is the dimensionless convection number, also called the Courant number.
The eigenvalues and thus the spectral radius of the (N 1)(N 1) matrix B can be computed
analytically and are given by
m
m = 1 + i c cos (12.4.6)
N
for m = 1, . . . , N 1, where i is the imaginary unit [317]. Since the magnitude of all eigenvalues
is greater than unity, the norm of the solution vector, f n , will amplify during the projections.
Accordingly, the numerical method is unconditionally unstable.
Another way of assessing stability is by performing the von Neumann analysis, obtaining the
amplication factor
An+1
G = 1 + i c sin , (12.4.7)
An
where = 2x/L is the phase angle. Since the magnitude of the right-hand side of (12.4.7) is
greater than unity for any value of c and , the solution vector, f n , will amplify for any wavelength,
L, and this conrms that the FTCS method is unconditionally unstable and must be abandoned. It
is instructive to recall that the FTCS method for the diusion equation was found to be conditionally
stable.
12.4 Convection 941
(a) (b)
t t
n +1 n +1
n n
x x
n 1 n 1
i1 i i+1 i1 i i+1
Figure 12.4.2 Computational stencil of the (a) FTBS and (b) FTFS discretization for solving the
convection equation.
Performing the von Neumann stability analysis, we derive the amplication factor
An+1
G = 1 c + c exp(i ), (12.4.12)
An
which reveals that G traces a circle of radius c centered at the point (1 c, 0) in the complex plane.
To guarantee stability, we require that |G| < 1, and obtain the stability constraint 0 < c < 1. When
the convection velocity U is negative, this constraint cannot be fullled and the FTBS method is
unstable.
942 Introduction to Theoretical and Computational Fluid Dynamics
Carrying out the von Neumann stability analysis, we derive the gain
An+1
G = 1 + c c exp(i ), (12.4.15)
An
which reveals that G is located on a circle of radius c centered at the point (1 + c, 0) in the complex
plane. To ensure stability, we require that |G| < 1 and obtain the stability constraint 1 < c < 0.
When the convection velocity U is positive, the method is unstable.
FTBS discretization
Considering rst the FTBS formula (12.4.9), we pretend that all discrete variables are continuous
functions of space and time and expand them in Taylor series about the doublet (xi , tn ) to obtain
1
fin + (ft )ni t + (ftt )ni t2 + O(t3 )
2
1
where ft = f /t, fx = f /x, and fxx = 2 f /x2 . Simplifying and rearranging, we nd that
1
(ft )ni t + c (fx )ni x = c (fxx )ni x2 (ftt )ni t2 + O(t3 ) + c O(x3 ). (12.4.17)
2
Next, we divide both sides by t and recall the denition c = U t/x to obtain
x2 c
(ft )ni + U (fx )ni = c (fxx )ni x2 (ftt )ni t + O(t2 ) + O(x3 ). (12.4.18)
t t
1
(ft )ni + U (fx )ni = U x (1 c) (fxx )ni + O(t2 ) + U O(x2 ). (12.4.19)
2
1
ft + U f x = U x (1 c) fxx . (12.4.20)
2
The right-hand side represents a small diusive term with numerical diusivity
1
num = U x (1 c), (12.4.21)
2
which is positive when 0 < c < 1. Negative numerical diusivity is tantamount to numerical
instability.
FTFS discretization
Working in a similar fashion with the FTFS method, we derive a modied dierential equation
involving a small diusion term involving the numerical diusivity
1
num = U x (1 + c), (12.4.22)
2
which is nonnegative when 1 < c < 0. Combining this result with the von Neumann stability
analysis, we conrm that positive numerical diusivity helps ensure numerical stability.
FTCS discretization
What went wrong with the FTCS discretization? Carrying out a consistency analysis, we derive a
modied dierential equation involving a small diusion term with numerical diusivity
1 1
num = c U x = U 2 t, (12.4.23)
2 2
(a) (b)
n n
t t
0 0
i x i x
Figure 12.4.3 Cone of inuence for the (a) FTBS and (b) FTFS method. The characteristic is drawn
as a heavy line.
According to the nite-dierence equations derived in the section, to compute the value fin , we use
information at all grid points residing inside a planar angle with apex at the point (xi , tn ), called
the numerical cone of inuence, as illustrated in Figure 12.4.3. The CFL condition, |c| < 1, requires
that the characteristic lies inside the numerical cone of inuence. If the numerical cone of inuence
does not include the characteristic that passes through a node, the numerical method will surely be
unstable. The CFL condition is thus necessary but not sucient for numerical stability.
1 1 x2
ft + U fx = t ftt + fxx . (12.4.27)
2 2 t
We conclude that the dierence equation is consistent only when x and t are reduced simulta-
neously so that the ratio x2 /t tends to zero and the last term on the right-hand side becomes
innitesimal. Keeping the ratio x/t constant yields a method that is rst-order accurate in time
and space.
Performing the von Neumann stability analysis, we derive the amplication factor
Requiring that |G| < 1, we nd that cos2 + c2 sin2 < 1, which reproduces the CFL condition,
|c| < 1.
To explain the conditional stability of Laxs method, we substitute ftt = U 2 fxx into the
modied dierential equation and group the rst with the second term on the right-hand side to
obtain the numerical diusivity
1 1
num = U x ( c), (12.4.29)
2 c
which is positive when |c| < 1. As the time step and thus c tends to zero, the numerical diusivity
becomes excessive, undermining the physical relevance of the calculations.
fin+1 = a1 fi1
n
+ a0 fin + a1 fi+1
n
, (12.4.30)
where a1 , a0 , and a1 are three constant coecients [228]. This formula encompasses all two-level
methods considered previously in this section.
To ensure consistency, we expand all variables in (12.4.30) in Taylor series about (xi , tn ) and
obtain
1 1
fin + (ft )ni t + (ftt )ni t2 = a1 [ fin (fx )ni x + (fxx )ni x2 ] + a0 fin
2 2
n 1
+a1 fi + (fx )i x + (fxx )i x2 .
n n
(12.4.31)
2
946 Introduction to Theoretical and Computational Fluid Dynamics
The solution is
1 1
a1 = c (c + 1), a0 = 1 c 2 , a1 = c (c 1). (12.4.34)
2 2
Substituting these values into (12.4.30), we derive the LaxWendro formula
1 1
fin+1 = n
c (c + 1) fi1 + (1 c2 ) fin + c (c 1) fi+1
n
. (12.4.35)
2 2
Rearranging, we nd that
1 1
fin+1 = fin + n
c (fi1 fi+1
n
) + c2 (fi1
n
2 fi+1
n n
+ fi+1 ). (12.4.36)
2 2
The rst two terms on the right-hand side implement the FTCS discretization. The last term cancels
a term originating from the second time derivative of the modied dierential equation.
To assess the numerical stability of the LaxWendro method, we carry out the von Neumann
stability analysis and nd that
which shows that the gain traces an ellipse with center at the point (1 c2 , 0) and semi-axes equal
to c2 and c in the complex plane. When |c| = 1, the ellipse reduces to the unit circle centered at
the origin. Geometrical arguments reveal that |G| < 1 when |c| < 1, and this demonstrates that the
method is stable when the CFL condition is fullled.
n +1 n +1 n +1
n +1/2
n n n
x x x
n 1 n 1 n 1
i1 i i+1 i1 i i+1 i1 i i+1
Figure 12.4.4 Computational stencil of (a) the explicit CTCS discretization involving three time levels,
(b) the implicit BTCS discretization involving two time levels. and (c) the implicit CrankNicolson
discretization involving three time levels.
The associated numerical error is on the order of t2 and x2 . The computational stencil is shown
in Figure 12.4.4(a). Rearranging, we derive the two-level explicit formula
fin+1 = fin1 + c fi1
n
c fi+1
n
. (12.4.39)
The solution at the rst time level, n = 1, must be computed using a two-level method with a
suciently small time step to prevent the onset of numerical instability.
To perform the von Neumann stability analysis, we substitute fin = An exp(i i) and nd
that
An+1 = An1 + 2 i An c sin . (12.4.40)
Dividing both sides by An , we obtain
An+1 An1
= + 2i c sin . (12.4.41)
An An
Next, we set An+1 /An = An /An1 = G, and derive a quadratic equation, G2 2i c sin G 1 = 0,
whose solution is
G = i 1 2 , (12.4.42)
where = c sin . To assess the magnitude of the
amplication factor, we consider two complemen-
tary possibilities. If 2 > 1, then G = i ( 2 1 ) and the magnitude of one of the roots is
greater than unity. If 2 < 1, then |G| = 1, which shows that the magnitude of the solution stays
constant during the successive mappings, in agreement with the exact solution. We note that 2 < 1
when |c| < 1 for any value of and conclude that the CTCS method is stable provided that the
CFL condition is fullled. Since the CTCS discretization for the unsteady diusion equation was
found to be unconditionally unstable, we infer that the presence of diusivity does not necessarily
promote numerical stability.
In practice, the eciency of the CTCS method is undermined by increased memory require-
ments associated with storage of information at three time levels. In addition, numerical error that
grows independently at every other time step, described as even-odd coupling, may arise.
948 Introduction to Theoretical and Computational Fluid Dynamics
The numerical error is of order t2 and x2 . The computational stencil is shown in Figure 12.4.4(c).
Rearranging, we derive the dierence equation
cfi1
n+1
+ 4fin+1 + cfi+1
n+1 n
= cfi1 + 4fin cfi+1
n
, (12.4.49)
yielding a tridiagonal system of algebraic equations at each time step. Carrying out the von Neumann
stability analysis, we nd that the amplication factor is the ratio of two conjugate numbers,
2 + i c sin
G= . (12.4.50)
2 i c sin
12.4.13 Summary
The restriction on the time step of conditionally stable explicit methods is not intolerable. The CFL
condition requires that the time step, t, be roughly smaller than x in appropriate dimensionless
units, which is not unreasonable. In contrast, the stability restriction for the diusion equation
discussed in Section 12.2 requires that the time step, t, be roughly smaller than x2 in appropri-
ate dimensionless units, which is typically intolerable. Although implicit methods allow us to use
larger time steps, the associated numerical error may erode the accuracy and therefore the physical
relevance of the solution. It is then not surprising that explicit methods are a standard choice in
practice.
The intensity of articial oscillations can be reduced by adding to the original dierential
equation explicit diusion or fourth-order diusion. In one dimension, these are implemented, re-
spectively, by the terms (x) 2 f /x2 and (x) 4 f /x4 . The positive coecients and can be
allowed to vary in space and time according to the structure of the solution. The optimal values of
these coecients depend on the particular problem under consideration and must be found by nu-
merical experimentation (e.g., [185], p. 207). A weak justication for introducing explicit numerical
diusion is that the approximations that lead us to the convection equation have neglected physical
diusion-like processes. Articially introducing these terms into the dierential equation may not
undermine the physical relevance of the solution.
950 Introduction to Theoretical and Computational Fluid Dynamics
The overall action of multistep methods for linear convection can be reduced to that of single-
step methods developed earlier in this section. Their discussion in the context of linear equations
serves as a point of reference for developing methods for nonlinear convection.
Richtmyers method
In Richtmyers method [342], a complete step is carried out in two substeps of equal duration 12 t.
The rst substep is executed using Laxs method based on formula (12.4.26), and the second substep
is executed using the CTCS method based on formula (12.4.39). The corresponding dierence
equations are
1 1
fin+1 = n
( + 1) fi2 + (1 2 ) fin + ( 1) fi+2
n
, (12.4.52)
2 2
where = 12 c = U t/(2x). Our analysis of the LaxWendro formula guarantees that Richtmyers
method is second-order accurate in time and space, and stable as long as || < 1 or |c| < 2.
Bursteins method
Bursteins method [47] diers from Richtmyers method in that the convection equation is applied
at the intermediate grid point i + 12 at the rst substep and at the regular grid point i at the second
substep. In both cases, the spatial step is 12 x. The counterparts of equations (12.4.51) are
MacCormack method
In MacCormacks method [250], a predictor step is made using the explicit FTFS method to produce
an approximation to fin+1 according to formula (12.4.14), denoted as fi ,
fi = (1 + c)fin cfi+1
n
. (12.4.54)
12.4 Convection 951
Rearranging, we obtain
fi = fin c ( fi+1
n
fin ). (12.4.55)
The corrector step uses the explicit forward time discretization and a hybrid forward/backward
space approximation involving the preliminary values according to the dierence equation
n
To deduce the action of the method, we use equation (12.4.55) to eliminate fi+1 on the right-
hand side of (12.4.57) in favor of fi , we obtain
1 n 1
fin+1 = fin ( f fi ) c (fi fi1
), (12.4.58)
2 i 2
and then
1 n
fin+1 = [ f + (1 c) fi + cfi1
) ]. (12.4.59)
2 i
Eliminating the intermediate variables denoted by an asterisk using equation (12.4.54), we recover the
LaxWendro formula (12.4.35). We conclude that MacCormacks method is second-order accurate
in time and space, and stable as long as the CFL condition is fullled, |c| < 1. In Section 12.4.16, we
will see that MacCormacks method is particularly eective for problems of nonlinear convection.
Flux-Corrected-Transport
We have seen that explicit numerical diusion may be necessary to enhance the performance of rst-
order methods. The idea behind ux-corrected transport is to use a predictorcorrector method
where the predictor step involves an articial dampening term [49, 50, 51]. The corrector step re-
moves excessive dissipation by introducing diusion with a negative diusivity, termed anti-diusion.
Laxs method
Laxs modication of the FTCS discretization yields the explicit two-level formula
1 1
fin+1 = (1 + cni1 ) fi1
n
+ (1 cni+1 ) fi+1
n
, (12.4.61)
2 2
where
cni1 = rfi1
n
, cni+1 = rfi+1
n
, (12.4.62)
and r = t/x. The method is rst-order accurate in time, second-order accurate in space, and
stable provided that |rfmax | < 1.
LaxWendro method
To achieve second-order accuracy, we discretize the conservative form as
1 1
fin+1 = fin + n
r (qi1 qi+1
n
) + r2 (A qi1
n
B qin + C qi+1
n
), (12.4.63)
2 4
where A, B, and C are adjustable coecients and r = t/x. The second term on the right-
hand side of (12.4.63) arises from the second term on the right-hand side of (12.4.36) by setting
c = 12 (fi1
n n
+ fi+1 ). Working as in the case of the linear convection, we nd that
n
A = fi1 + fin , n
B = fi1 + 2 fin + fi+1
n
, C = fin + fi+1
n
(12.4.64)
(e.g., [185], p. 207). The method is second-order accurate in space and time, and stable provided
that |rfmax | < 1.
BTCS discretization
The implicit BTCS discretization results in a quadratic system of algebraic equations that must be
solved by iteration,
n+1 2 n+1 2
r (fi1 ) + 4fin+1 + r (fi+1 ) = 4fin , (12.4.65)
where r = t/x. A suitable initial guess is provided by the converged solution at the previous
time step. In practice, increased computational cost renders this method uneconomical.
12.4 Convection 953
MacCormacks method
1 n
Substituting into the dierence formula (12.4.55) the expression c = 2
(fi+1 + fin ), we obtain a
tentative update denoted by an asterisk,
fi = fin c (qi+1
n
qin ). (12.4.66)
1
Setting in the dierence formula (12.4.59) c = 2 (fi + fi1
), we obtain the solution at the next time
level,
1 n
fin+1 = f + fi c (qi qi1
) . (12.4.67)
2 i
The bias in the solution due to the forward or backward dierencing is eliminated by alternating
the order after the completion of a time step. The method is stable as long as |rfmax | < 1.
f f f f
+U +V +W = 0. (12.4.68)
t x y z
The convection velocity, U = (U, V, W ), is assumed to be constant in time and space. The case
of two-dimensional convection arises by setting W = 0. An initial condition, f (x, t = 0) = F (x),
and suitable boundary conditions are provided. Assuming, for simplicity, that the solution domain
extends over the whole three-dimensional space or two-dimensional plane, we nd that the exact
solution is given by F (x Ut), which states that the initial distribution, F (x), travels with constant
velocity U. The characteristic lines along which the function f remains constant, are described by
the equation x Ut = , where is an arbitrary point in space.
Laxs method
Laxs method in two dimensions emerges by replacing (12.4.68) with the nite-dierence equation
1 n+1 1 n n
fi+1,j fi1,j
n n
fi,j+1 fi,j1
n
fi,j (fi+1,j + fi1,j
n n
+ fi,j+1 n
+ fi,j1 ) +U +V = 0.
t 4 2x 2y
(12.4.69)
1 1 1 1 1 1 1 1
n+1
fi,j = n
( + cx )fi1,j + ( cx ) fi+1,j
n n
+ ( + cy )fi,j1 + ( cy )fi,j+1
n
, (12.4.70)
2 2 2 2 2 2 2 2
954 Introduction to Theoretical and Computational Fluid Dynamics
where
U t V t
cx = , cy = (12.4.71)
x y
are convection numbers for the x and y directions. Performing the von Neumann stability analysis,
we nd that the method is stable provided that c2x + c2y < 12 . This condition imposes a strong
restriction on the time step that renders Laxs method inecient in practice. In three dimensions,
the stability constraint is c2x + c2y + c2z < 13 , which describes the interior of a sphere with radius 1/ 3
centered at the origin, where cz = W t/z and W is the convection velocity along the z axis.
Implicit methods
To achieve unconditional stability, we may resort to a fully implicit method. In three dimensions,
we obtain the dierence equation
1 n+1 n
1+ (cx x + cy y + cz z ) fi,j,k = fi,j,k , (12.4.72)
2
where
As a remedy, we may use an ADI partially implicit discretization, which requires solving a
one-dimensional equation in each substep (e.g., [110]; [182], Volume I, p. 442; [229]). The standard
ADI algorithm is unconditionally stable in two dimensions and unconditionally unstable in three
dimensions.
Operator splitting
Following the general blueprint of operator splitting and fractional step methods, we replace the
three-dimensional convection equation (12.4.68) with a system of three one-dimensional sequential
equations,
f f f f f f
+U = 0, +V = 0, +W = 0. (12.4.74)
t x t y t z
Each equation applies for the full time step, tn < t < tn + t.
Adopting the implicit BTCS discretization, we advance the solution in each step by solving
tridiagonal systems of equations,
1 n 1 1 n+1
(1 + cx x )fi,j,k = fi,j,k , (1 + cy y )fi,j,k = fi,j,k , (1 + cz x )fi,j,k = fi,j,k , (12.4.75)
2 2 2
12.4 Convection 955
where the starred and double starred variables denote the solution after the rst and second fractional
steps. Combining the three equations in (12.4.75), we derive an overall nite-dierence scheme
involving a factorized implicit operator in the left-hand side,
1 1 1 n+1 n
(1 + cx x )(1 + cy y )(1 + cz z )fi,j,k = fi,j,k . (12.4.76)
2 2 2
Comparing this formula with the fully implicit formula (12.4.72) shows that the fractional step
method implements the approximate factorization of the explicit BTCS scheme.
Problems
Computer Problems
12.5 Convectiondiusion
The methods developed previously in this chapter for the modular cases of pure diusion and pure
convection can be extended in a straightforward fashion to handle combined convectiondiusion.
In this section, we discuss methods for linear one-dimensional convectiondiusion governed by the
parabolic dierential equation
f f 2f
+U = 2, (12.5.1)
t x x
subject to an initial condition, f (x, t = 0) = F (x). Both the convection velocity, U , and diusivity,
, are assumed to be constant.
In previous sections, we found that the explicit FTCS scheme is conditionally stable for un-
steady diusion and unconditionally unstable for linear convection. Performing the von Neumann
stability analysis for convectiondiusion, we derive the amplication factor
G = 1 2 + 2 cos + i c sin , (12.5.6)
which shows that G traces an ellipse passing through the point (1, 0) in the complex plane. The
center of the ellipse is located at the point (1 2, 0), and the ellipse semi-axes are equal to 2 and
c. To guarantee stability, we require that the ellipse lies inside the unit disk.
First, we demand that the length of each semi-axis is less than unity,
1
< , c < 1. (12.5.7)
2
12.5 Convectiondiusion 957
A further restriction emerges by requiring that the curvature of the ellipse at the point (1, 0) is
higher than the curvature of the unit circle. Combining these restrictions, we obtain
It is instructive to recall that the FTCS discretization for pure convection leads to an unconditionally
unstable scheme. We conclude that the presence of diusion has a stabilizing action on the numerical
method.
The modied dierential equation corresponding to the dierence equation (12.5.10) is the
convectiondiusion equation with eective diusivity
1 2 1 c2
ef f = U t = 1 . (12.5.9)
2 2
Inequality (12.5.8) states that a necessary condition for numerical stability is that the positive
physical diusivity overtakes the negative numerical diusivity in absolute value.
The stability restriction (12.5.8) requires an excessively small time step, and this renders the
FTCS discretization uneconomical. Laxs modication leads to an unconditionally unstable method.
Regrettably, the explicit FTCS scheme and its variations are of limited interest.
fin+1 = (c + ) fi1
n
+ (1 c 2) fin + fi+1
n
. (12.5.13)
958 Introduction to Theoretical and Computational Fluid Dynamics
Carrying out a consistency analysis, we nd that the corresponding modied dierential equation is
the convectiondiusion equation with eective diusivity
1
ef f = 1 + (1 c) Rec . (12.5.14)
2
Since Rec vanishes as x tends to zero, the method is consistent.
Performing the von Neumann stability analysis, we derive the amplication factor
which shows that G traces an ellipse passing through the point (1, 0) in the complex plane. The
center of the ellipse is located at (1 c 2 , 0), and the semi-axes of the ellipse are c + 2 and c.
To guarantee stability, we demand that the ellipse lies inside the unit disk and derive the stability
constraint
When U is negative, we use a forward dierence for the convective spatial derivative and a centered
dierence for the diusive derivative. Working in a similar fashion, we nd that the method is stable
if c2 < |c| + 2 < 1. In practice, the numerical diusivity associated with upwind dierencing can
be signicant. This undesirable feature combined with the rst-order accuracy and the conditional
stability renders the method inferior to its alternatives.
Higher-order methods
To improve the accuracy and reduce the numerical diusivity of the rst-order upwind method,
we may approximate the convective derivative using a third-order backward dierence involving
four points, while keeping the central dierence for the second spatial derivative [232]. When U is
positive, the nite-dierence equation is
The discretization error is of order t and x2 . Rearranging, we derive the Leonard dierence
equation
1 n 1 1
fin+1 = c fi2 n
+ (c + ) fi1 + (1 c 2) fin + ( c + ) fi+1
n
. (12.5.20)
6 2 3
The corresponding modied dierential equation is the convectiondiusion equation with an ef-
fective diusivity identical to that of the FTCS scheme given in (12.5.9). Carrying out the von
Neumann stability analysis, we derive the amplication factor
c c c
G = exp(2i) + (c + ) exp(i) + 1 2 + ( + ) exp(i). (12.5.21)
6 2 3
The stability criteria are substantially milder than those of the rst-order upwind method.
To carry out the von Neumann stability analysis, we substitute fin = An exp(i i) and set
G = An+1 /An = An /An1 to nd
c + 2 1 2 1 c + 2
G= exp(i) + + exp(i). (12.5.24)
1 + 2 1 + 2 G 2 + 1
Rearranging, we obtain the quadratic equation
(1 + 2) G2 2 (2 cos + i c sin ) G 1 + 2 = 0. (12.5.25)
Detailed consideration reveals that |G| < 1 when |c| < 1, and this shows that the Du FortFrankel
method is stable as long as the CFL condition is fullled.
The absence of a stability restriction on the diusion number, , allows us to use large time
steps. However, a suciently small time step must be used to ensure that the solution is accurate
and consistent with the convectiondiusion equation.
960 Introduction to Theoretical and Computational Fluid Dynamics
BTCS
Implementing a backward dierence for the time derivative and centered dierences for the convective
and diusive spatial derivatives, we derive the fully implicit BTCS dierence formula
(c + 2) fi1
n+1
+ 2 (1 + 2) fin+1 + (c 2) fi+1
n+1
= 2fin . (12.5.27)
Since |G| < 1 for any values of and c, the method is unconditionally stable. However, the physical
restriction Rec < 2 must be satised for the results to be physically meaningful under any conditions.
CrankNicolson method
To improve the temporal accuracy of the BTCS method, we implement the fully implicit Crank
Nicolson method according to the dierence equation
(c + 2)fi1
n+1
+ 4 (1 + )fin+1 + (c 2)fi+1
n+1 n
= (c + 2)fi1 + 4(1 )fin (c 2)fi+1
n
.
(12.5.30)
It can be shown that |G| < 1 for any value of and c, and thus the method is unconditionally stable.
The physical restriction Rec < 2 still applies.
12.5 Convectiondiusion 961
b1 + b0 + b1 = a1 + a0 + a1 = 1. (12.5.33)
The last equality represents an arbitrary normalization that removes one degree of freedom. For
example, in the case of the BTCS discretization, a1 = 0, a0 = 1, a1 = 0, and
1 1
b1 = (c + 2) b0 = 1 + 2, b1 = (c 2). (12.5.34)
2 2
In the case of the CrankNicolson discretization,
1 1
a1 = (c + 2), a0 = 1 , a1 = (c 2),
4 4
1 1
b1 = (c + 2), b0 = 1 + , b1 = (c 2). (12.5.35)
4 4
Carrying out the von Neumann stability analysis for the general case, we derive the amplication
factor
1 (a1 + a1 )(1 cos ) i (a1 a1 ) sin
G= (12.5.36)
1 (b1 + b1 )(1 cos ) i (b1 b1 ) sin
([294], p. 39). Formulas (12.5.28) and (12.5.31) are special cases of (12.5.36).
fi = (1 + c) fin cfi+1
n n
+ (fi1 2fin + fi+1
n
), (12.5.38)
962 Introduction to Theoretical and Computational Fluid Dynamics
f f f 2f
+U = 0, = 2. (12.5.40)
t x t x
Both equations apply for a full time step, tn < t < tn + t, with the understanding that the time is
reset to the initial value, tn , at the end of the rst fractional step. The rst step takes us from the
current time level, f n , to f , and the second step takes us from f to the new time level, f n+1 .
For example, applying the FTCS discretization, we obtain the dierence equations
1 n 1 n
fi = cf + fin c fi+1 ,
fin+1 = fi1 + (1 2) fi + fi+1
. (12.5.41)
2 i1 2
The stability restrictions are the same as those for the FTCS method applied to the undivided
convectiondiusion equation.
Hopscotch method
The hopscotch method is named after a childrens game ([269], p. 77; [149]). The explicit FTCS
discretization is used to advance the solution at the odd-numbered grid points, x2i+1 , and then the
implicit BTCS discretization is used to advance the solution at the even-numbered grid points, x2i ,
where i is an integer. The order is reversed after the completion of each step. Since the solution at
every other grid point at the new time level is known, the implicit step does not have to be done
through matrix inversion and the method is eectively explicit. The hopscotch method is rst-order
accurate in time, second-order accurate in space, and stable as long as |c| < 1. There is no stability
restriction on the diusion number .
The explicit FTCS update of the ith node from tn to tn+1 yields
1 1
fin+1 = ( c + ) fi1
n
+ (1 2) fin ( c ) fi+1
n
. (12.5.42)
2 2
12.5 Convectiondiusion 963
The implicit BTCS update of the same node from tn1 to tn yields
1 1
n
( c + ) fi1 (1 + 2) fin ( c ) fi+1
n
= fin1 . (12.5.43)
2 2
Combining these equations, we derive the dierence equation
fin+1 = 2fin fin1 , (12.5.44)
which implements linear extrapolation. This formula is applicable only in the FTCS step.
Burgers equation
A prototypical equation for studying the performance of numerical methods is the Burgers convection
diusion equation whose convective and conservative forms are
f f 2f f q 2f
+f = 2, + = 2, (12.5.45)
t x x t x x
where q = 12 f 2 is the quadratic ux. The general solution in an unbounded domain extending over
the entire x axis can be found analytically using the ColeHopf transformation [30],
2 du
f = . (12.5.46)
u dx
Direct substitution shows that the function u satises the linear unsteady diusion
u 2u
= 2, (12.5.47)
t x
whose exact solution is available in integral form. Note that the transformation fails as tends to
zero, yielding the inviscid form. An exact solution is
cosh(x/L)
f (x, t) = , (12.5.48)
L sinh(x/L) + exp(t/L2 )
where L is a specied length [30]. Note that a discontinuity occurs when the denominator becomes
zero. As time progresses, the discontinuity progresses along the x axis and settles at the origin,
x = 0.
FTCS method
The fully explicit FTCS method is consistent, rst-order accurate in time, and second-order accurate
in space. Stability requires that
1 c2x c2y c2
x + y + z < , + + z < 2. (12.5.52)
2 x x z
In two dimensions, the sums on the left-hand sides are over x and y [180].
Upwind dierencing
Adopting a rst-order upwind dierence for the convective spatial derivative and the central dier-
ence for the diusive derivative yields a conditionally stable method. The stability constraint in two
dimensions is 4x + |cx | + |cy | < 1 ([294], p. 66).
Hopscotch method
The hopscotch method is implemented as discussed for one dimension in Section 12.6 [149]. In two
dimensions, we rst use the explicit FTCS method to advance the solution at the grid points xi,j ,
where i + j is an odd integer, and then use the implicit BTCS method to advance the solution at
the grid points xi,j , where i + j is an even integer. The order is reversed after the completion of a
time step. After the rst time step, the FTCS dierence equation is replaced with the equivalent
equation
n+1
fi,j n
= 2fi,j fi,j
n1
. (12.5.53)
The method is rst-order accurate in time, second-order accurate in space, and stable as long as
|cx | < 1 and |cy | < 1.
and
n+1
fi,j
n+1/2
fi,j
n+1/2 n+1/2
fi+1,j fi1,j f n+1 f n+1
i,j+1 i,j1
1 +U +V
2 t
2x 2y
n+1/2 n+1/2
fi1,j 2fi,j
n+1/2
+ fi+1,j f n+1 2f n+1 + f n+1
i,j1 i,j i,j+1
= + , (12.5.55)
x2 y 2
where the square brackets enclose implicit discretizations. The ADI method is unconditionally stable
and second-order accurate in time and space ([294], p. 66).
U = a1 U n+1 + (1 a1 a2 ) U n + a2 U n1 ,
V = b1 V n+1 + (1 b1 b2 ) V n + b2 V n1 , (12.5.56)
Fractional steps
To implement the fractional step method in two or three dimensions, we consider convectiondiusion
in each direction individually through a sequence of one-dimensional steps of equal duration, t. In
three dimensions, the fractional steps are carried out according to the one-dimensional convection
diusion equations
f f 2f f f 2f f f 2f
+ ux = 2, + uy = 2, + uz = 2. (12.5.58)
t x x t y y t z z
Each step applies for tn < t < tn + t, and time is reset to the initial value, tn , after completion of
the rst and second fractional steps. Each step can be carried out using an unconditionally stable
implicit method. When Dirichlet or Neumann boundary conditions are prescribed, the discretization
results in tridiagonal systems of equations.
Problems
Computer Problems
f f 3f
+ f + 3 = 0, (12.5.59)
t x x
where and are two positive constants.
(a) Conrm that an exact solution in an unbounded domain expressing the propagation of a solitary
wave is
3c
f= 2
, (12.5.60)
cosh (x ct d)
where c 0 and d are two arbitrary constants and = c/(4) [98].
(b) Develop an explicit nite-dierence method and compute the evolution of the solution from the
initial state described by (12.5.60). Discuss the reliability and accuracy of your computations.
Finite-dierence methods
for incompressible Newtonian ow
13
Having discussed nite-dierence methods for solving the convectiondiusion equation in its general
form, we proceed to develop corresponding methods for solving the equations of steady and unsteady
incompressible Newtonian ow. The set of governing equations includes the NavierStokes equation
and the continuity equation. The primary unknowns are the velocity and the pressure. However, we
recall that an arbitrary rotational ow can be described, and therefore computed, in terms of the
secondary variables discussed in Chapter 2, including the vorticity, the stream functions, and the
vector potential. Numerical methods based on these secondary elds will be outlined.
967
968 Introduction to Theoretical and Computational Fluid Dynamics
directly available, but must be derived from the equation of motion subject to specied boundary
conditions for the velocity or traction. We will see that the accurate implementation of derived
pressure boundary conditions requires special attention.
We will begin this chapter by discussing a class of methods for computing the structure of a
steady ow and the evolution of an unsteady ow based on the vorticity transport equation. The
numerical procedure involves computing the evolution of the vorticity eld, while simultaneously
obtaining the simultaneous evolution of the velocity eld by inverting the denition of the vorticity,
= u, subject to the continuity equation. One advantage of this approach is that the pressure
eld does not need to considered. One disadvantage is the need to derive boundary conditions for
the vorticity.
Methods based on the vorticity transport equation can be regarded as generalizations of the
vortex methods for inviscid or weakly viscous ows discussed in Chapter 11. What distinguishes
vortex methods from other methods based on vorticity transport is that the velocity eld is obtained
from the vorticity eld eciently using the BiotSavart integral or a related contour integral. In
the case of viscous ow, because the support of the vorticity is not necessarily compact, and it is
more expedient to recover the velocity from the vorticity eld by nite-dierence or other domain
discretization methods.
A variety of nite-dierence procedures are available for solving the equations of steady and
unsteady incompressible Newtonian ow, and a choice must be made according to the tolerated level
of programming complexity and available computational resources. In this chapter, we outline the
fundamental principles of several alternative formulations and discuss the basic steps involved their
numerical implementation. Extensions and discussion of particular issues and specialized methods
can be found in the references cited as well as in general reviews and monographs on nite-dierence
methods in uid dynamics dynamics (e.g., [9, 69, 155, 185, 286, 294, 346]). Numerical methods for
free-surface and interfacial ow are reviewed in Reference [131].
z = 2 , (13.1.1)
where 2 is the Laplacian operator in the xy plane. The computation proceeds according to the
two fundamental steps of the vortex methods discussed in Chapter 11.
13.1 Vorticitystream function formulation for two-dimensional ow 969
In the rst step, the rate of change of the vorticity is computed using the simplied vorticity
transport equation for two-dimensional ow written in the vorticitystream function form
z z z
+ = 2 z , (13.1.2)
t y x x y
where is the kinematic viscosity of the uid. The sum of the second and third terms on the
left-hand side of (13.1.2) is sometimes designated as the Jacobian, J (z , ). In the second step, the
simultaneous evolution of the stream function is obtained by solving the Poisson equation (13.1.1)
for in terms of z . Boundary conditions are requiring in both the integration of (13.1.2) and the
inversion of (13.1.1). It is instructive to note that the absence of an explicit evolution equation for
the pressure in the original system of governing equations is reected in the absence of an explicit
evolution equation for the stream function.
Pressure eld
One important feature of the vorticitystream function formulation is that computing the pressure
is not required. If the instantaneous pressure eld is desired, it can be computed a posteriori by
solving a Poisson equation that emerges by taking the divergence of the NavierStokes equation and
using the continuity equation to obtain
2 2 2
2
2 p = 2 . (13.1.3)
x2 y2 xy
Boundary conditions for the pressure arise by applying the equation of motion at the boundaries,
enforcing the specied boundary conditions, and then projecting the result onto the normal or
tangential unit vector, as discussed in Section 13.3 in the context of the velocitypressure formulation.
where c is an arbitrary constant set for simplicity to zero. The no-slip boundary condition requires
that the tangential component of the velocity is zero over the bottom, left, and right walls, whereas
the tangential velocity at the upper wall is equal to the wall velocity, V (t). In terms of the stream
function, we obtain the equivalent statement
= 0 at the bottom, =0 at the sides, =V at the lid. (13.1.5)
y x y
Given these boundary conditions for the velocity, we derive simplied expressions for the boundary
values of the vorticity in terms of the stream function. Beginning with (13.1.1) and noting that, for
970 Introduction to Theoretical and Computational Fluid Dynamics
y V
Ny+1
by
Ny
3
2
ay 1
1 2 3 i Nx Nx +1
x
ax bx
2 2
z = at the top and bottom walls, z = at the side walls. (13.1.6)
y 2 x2
Steady ow
Two distinct but somewhat related approaches are available for computing a steady ow. The rst
class of methods involves solving the equations of steady ow using an iterative scheme. The second
class of methods involves computing the solution of a ctitious transient ow problem governed by
a modied set of dierential equations from a given initial condition up to the steady state. The
solution of the modied problem at steady state satises the equations of steady two-dimensional
incompressible Newtonian ow.
13.1 Vorticitystream function formulation for two-dimensional ow 971
2 = z , (13.1.7)
and
1 z z
2 z = ux + uy . (13.1.8)
x y
In the special case of Stokes ow, the right-hand side of (13.1.8) vanishes, yielding Laplaces equation
for the vorticity, 2 z = 0. Equation (13.1.7) then shows that the stream function satises an
inhomogeneous biharmonic equation, 4 = z . The computational algorithm in the general case
of nonzero Reynolds-number ow involves the following steps:
Step 2: Solve the Poisson equation (13.1.7) for the stream function.
For boundary conditions, we have the choice between the Dirichlet boundary condition that spec-
ies the boundary distribution of the stream function, and the Neumann boundary condition that
species the boundary distribution of the normal derivative of the stream function, which is equal
to the tangential component of the velocity. A combination of the Dirichlet and Neumann boundary
conditions at dierent boundaries can also be employed. If we use the Neumann condition over all
boundaries, the Poisson equation will have a solution only if the following compatibility condition
is fullled,
n dl = z dx dy, (13.1.9)
W alls F low
where n is the normal unit vector pointing into the ow. Even when (13.1.9) is fullled by a fortuitous
guess of the vorticity distribution in Step 1, the singular nature of the linear system of equations
that arises from the nite-dierence discretization of (13.1.7) introduces additional complications.
For these reasons, the Dirichlet condition expressed by (13.1.4) is preferred around all boundaries.
Step 3: Compute the right-hand side of (13.1.8) and the boundary values of the vorticity using the
specied boundary conditions for the velocity.
Step 5: Check whether the computed vorticity eld agrees with that guessed in Step 1 at all grid points.
If it does not agree within a specied tolerance, replace the guessed with the computed vorticity and
return to Step 2.
Implementation
The details of the numerical implementation will be discussed with reference to ow in a cavity
driven by a moving lid, as illustrated in Figure 13.1.1.
972 Introduction to Theoretical and Computational Fluid Dynamics
Step 1: Assign initial values for the stream function to all internal and boundary (Nx + 1) (Ny + 1)
grid points.
A simple choice consistent with the no-penetration boundary condition is to set the initial stream
function to zero, corresponding to a quiescent uid.
Step 2: Assign values for the vorticity to all internal Nx Ny grid points.
A simple choice is to set all initial vorticity gird values to zero, corresponding to a quiescent uid.
Note that this choice disregards the motion of the uid due to the translation of the upper wall.
Step 3: Solve the Poisson equation (13.1.7) for subject to the Dirichlet boundary condition (13.1.4)
at all four walls.
Since the vorticity on the right-hand side is only an approximation to the exact solution, high accu-
racy is not necessary. To reduce the computational eort, we solve the Poisson equation iteratively
and carry out only a small number of iterations. To perform the iterations, we introduce a ctitious
unsteady diusionreaction problem governed by the equation
= (2 + z ), (13.1.10)
t
where is a diusivity. The solution of this equation at steady state satises (13.1.7). For simplicity,
we denote = z . Implementing the forward time centered space (FTCS) discretization discussed
in Section 12.3.7, we nd that
(l+1) (l) (l)
ij = ij + x Rij , (13.1.11)
where x = t/x2 is the diusion number in the x direction, the superscript (l) denotes the lth
iteration level,
for l = 1, 2, . . . , where is a relaxation factor used to control the iterations. When the iterations
are executed for the rst time, the initial guess (0) is set equal to that assigned in Step 1. When
the iterations converge, the solution is second-order accurate with respect to x and y.
Step 4: Use the simplied expressions (13.1.6) to compute the vorticity at the boundary grid points
by one-sided nite dierences.
To compute the vorticity at a grid point along the upper wall, we expand the stream function in a
Taylor series with respect to y about a grid point that lies at the lid and evaluate the series at the
13.1 Vorticitystream function formulation for two-dimensional ow 973
i,Ny +1 i,Ny V
i,Ny +1 = 2 2 , (13.1.15)
y 2 y
which is rst-order accurate in y. Working in a similar fashion for the bottom and side walls, we
derive the corresponding expressions
i,1 i,2 1,j 2,j Nx +1,j Nx ,j
i,1 = 2 , 1,j = 2 , Nx +1,j = 2 . (13.1.16)
y 2 y 2 y 2
To increase the accuracy of the method to second order, we expand the stream function in
a Taylor series about a grid point at the upper wall, evaluate the series at two layers immediately
below the upper wall, and retain terms up to third order to nd
1 2
1 3
3
i,Ny i,Ny +1 y + y 2 + y (13.1.17)
y i,Ny +1 2 y2 i,Ny +1 6 y 3 i,Ny +1
and
2
4 3
Step 5: Dierentiate the stream function to compute the velocity at the internal grid points subject
to the boundary values (13.1.4).
Step 6: Dierentiate the vorticity to compute the right-hand side of (13.1.8) at the internal grid
points subject to the boundary values computed from (13.1.15) and (13.1.16) or from (13.1.19) and
(13.1.20).
974 Introduction to Theoretical and Computational Fluid Dynamics
Step 7: Solve the Poisson equation (13.1.8) for the vorticity subject to the Dirichlet boundary condi-
tions for the vorticity derived in Step 4.
This is done by iteration using, for example, the forward timecentered space (FTCS) algorithm
discussed in Step 3, carrying out only a small number of iterations. Grid values of the forcing
function on the right-hand side at the internal nodes are available from Step 6.
Step 8: If the vorticity computed in Step 7 does not agree with that previously available within a
specied tolerance, we return to Step 2 and repeat the computations with the new values of the
vorticity.
This outer iteration is terminated when the absolute value of the dierence in the vorticity between
two successive iterations at each grid point becomes less than a present threshold, , or when the
sum of the absolute values of the dierences in the vorticity over all internal Nx Ny grid points
becomes less than Nx Ny .
One noteworthy feature of the algorithm is that the corner grid points do not enter the
computation, which means that the numerical scheme is oblivious to the velocity discontinuity at the
two upper corner points. The local jump may cause local oscillations and slow down the convergence
of the iterations, but does not have a deleterious eect on the overall numerical method.
Structure of the ow
Vorticity and stream function contour plots for ow in a cavity with aspect ratio Lx /Ly = 2, and
Reynolds number Re = V Lx / = 1 and 100 are shown in Figure 13.1.2. In the notation of Figure
13.1.1, Lx = bx ax and Ly = by ay . Stream function contours are both streamlines and particle
paths in a two-dimensional ow.
When Re = 1, we obtain a nearly creeping ow whose streamlines are symmetric with respect
to the midplane due to reversibility of Stokes ow, as discussed in Section 6.1.7. Small regions of
recirculating ow are present at the bottom two corners, requiring increased spatial resolution. The
vorticity is singular at the upper two cavity corners due to the discontinuous boundary velocity. A
local analysis shows that the shear stress diverges at these corners, and an innite force is required
to slide the lid as a result of the sharp-corner idealization, as discussed in Section 6.2. However,
these physical singularities do not deter the performance of the numerical method.
As the Reynolds number increases, the center of the central eddy is shifted toward the upper
right corner due to the uid inertia and the ow becomes unsymmetric. To compute ow at even
higher Reynolds numbers, it is helpful to perform parameter continuation where the initial guesses
for the stream function and vorticity are identied with the corresponding converged values at a
lower Reynolds number. In fact, the Reynolds number can be increased gradually up to a targeted
value in the course of the iterations.
The computational procedure discussed in this section can be improved in several ways (e.g., [159]).
In one improvement, instead of iterating on the Poisson equation for the vorticity in Step 7, we
13.1 Vorticitystream function formulation for two-dimensional ow 975
(a)
20
20
10
0 0
10
20 20
30
40
40 40
40
30 30
20 20 20 20
10 10 10 10
0 0 x 0 0 x
y y
(b)
1.2 1.2
1 1
0.8 0.8
0.6 0.6
y
0.4 0.4
0.2 0.2
0 0
0.2 0.2
0 0.5 1 1.5 2 0 0.5 1 1.5 2
x x
Figure 13.1.2 (a ) Vorticity and (b) stream function contour plots for ow in a rectangular cavity
with aspect ratio Lx /Ly = 2, at Reynolds number Re = V Lx / = 1 (left) and 100 (right).
iterate on the full diusionconvection equation (13.1.8). This requires that the nonlinear term on
the right-hand side is recomputed after each iteration using the updated vorticity. The iterations
can be carried out using an explicit or implicit nite-dierence method for the convectiondiusion
equation in two dimensions discussed in Section 12.7. Because the ow near the center of the cavity
is dominated by convection at high Reynolds numbers, using upwind dierencing helps improve
numerical stability. The rate of convergence of the inner and outer iterations depends on the details
of the particular implementation [198].
The method for computing the boundary values of the vorticity described in Step 3 has been
the subject of criticism [155]. It has been argued that it is not appropriate to explicitly impose the
boundary values of the vorticity. Instead, the boundary distribution of the vorticity should arise
976 Introduction to Theoretical and Computational Fluid Dynamics
implicitly as part of the solution using the natural boundary conditions for the velocity or traction.
This issue will be discussed further in Section 13.2 in the more general context of three-dimensional
ow.
This observation suggests reformulating the problem by retaining the vorticity transport equa-
tion (13.1.1) and replacing (13.1.2) with an evolution equation equation for the stream function ex-
pressed by (13.1.10), where the diusivity is a free parameter of the numerical method. The idea is
to compute the evolution of the ow from a certain initial condition subject to (13.1.1) and (13.1.10)
until a steady state has been established. The steady-state solution will satisfy the original equations
(13.1.7) and (13.1.8). An advantage of this approach is that all governing equations are parabolic
in time. Time-marching schemes similar to those developed in Chapter 12 for convectiondiusion
may then be employed.
For the problem of ow in a cavity illustrated in Figure 13.1.1, the method of modied dy-
namics is implemented according to the following steps:
1. Assign initial values to the stream function and vorticity at all internal and boundary grid
points. A simple choice is to set them both to zero.
2. Dierentiate the stream function to compute the two components of the velocity at all internal
grid points.
3. Compute the vorticity at the boundary grid points.
4. Advance the vorticity at all internal grid points on the basis of (13.1.2) using, for example, the
ADI method for the convectiondiusion equation described in Section 12.7, while keeping the
vorticity at the boundary grid points constant.
5. Advance the stream function at all internal grid points on the basis of (13.1.10) using, for
example, the ADI method for the convectiondiusion equation described in Section 12.7.
6. Return to Step 3 and repeat the computations for another time step.
13.1.5 Unsteady ow
To compute the evolution of an unsteady ow, we combine features of the direct method with
features of the method of modied dynamics for steady ow. The algorithm involves computing the
evolution of the vorticity eld using (13.1.2), while simultaneously describing the evolution of the
velocity eld in terms of the stream function using (13.1.1). To simplify the notation, we denote
u = ux , v = uy , and = z . A simple strategy for computing the evolution of ow in a cavity when
the lid starts translating suddenly with constant velocity V involves the following steps:
13.1 Vorticitystream function formulation for two-dimensional ow 977
1. At the initial instant, we set the stream function and velocity equal to zero at all internal and
boundary grid points. Then we set the x velocity component at the grid points at the upper
wall equal to the lid velocity, V .
2. Now we dierentiate the velocity to compute the vorticity. For the internal grid points, we
use central dierences. For grid points that lie on the lid, we use the denition of the vorticity
and apply second-order backward dierence to nd that
3V + 4 ui,Ny ui,Ny 1
i,Ny +1 = . (13.1.21)
2y
For grid points that lie at the bottom and side walls, we use the corresponding second-order
dierence formulas
4 ui,2 + ui,3 4 v2,j v3,j 4 vNx ,j + vNx 1,j
i,1 = , 1,j = , Nx +1,j = . (13.1.22)
2y 2x 2x
3. Next, we integrate (13.1.2) to obtain the vorticity at the next time level at all internal grid
points subject to the boundary conditions (13.1.21) and (13.1.22) using an explicit method,
such as the FTCS method. At high Reynolds numbers, we use upwind dierences.
4. In the fourth step, we solve the Poisson equation (13.1.1) for the stream function at the next
time level subject to the Dirichlet boundary condition (13.1.4).
5. Now we dierentiate the stream function to compute the velocity at the next time level at all
internal grid points.
6. Having advanced the ow by one time step, we return to Step 2 and repeat the computation
for another time step.
To improve the temporal accuracy and promote the numerical stability, we may update the
vorticity using an implicit or semi-implicit method, such as the ADI method.
ADI method
In a simple implementation of the ADI method, the convection velocity is held constant and equal
to initial value at the beginning of the rst substep during both substeps. In a more advanced
implementation, the Poisson equation is solved for an intermediate stream function after completing
the rst substep, and the convection velocity is set equal to an intermediate velocity computed by
dierentiating the intermediate stream function at the second substep. Since the convection velocity
is held constant and equal to its value at the beginning of a step or substep, the overall accuracy of
both methods is rst-order in time.
To achieve second-order accuracy, we use the ADI method with time-dependent velocities
described in (12.5.56) and (12.5.57). Collecting the values of the vorticity at all grid points into a
vector, , we obtain two ADI equations written in the symbolic form
where A, B, C, and D are tridiagonal matrices dependent on their stated arguments. Equations
(13.1.23) replace the explicit FTCS equation in Step 3. Steps 35 combine into the following inner
iterative loop: (a) Guess the velocities un+1 and solve the two tridiagonal systems (13.1.23) with
boundary conditions (13.1.21) and (13.1.22) for n+1/2 and n+1 ; (b) execute Steps 4 and 5; (c)
solve (13.1.23) with the computed values of un+1 or with a weighted average of old and new values.
If the boundary velocity changes in time, we solve the rst tridiagonal system in (13.1.23)
subject to the boundary conditions n+1/2 = 12 ( n + n+1 ), where n+1 has been approximated
from the previous inner iteration. To accelerate the convergence, we may replace the boundary
conditions for n+1 during the inner iterations with a weighted average of old and new values.
Further details on the implementation of the method are given in Reference [294] (p. 197).
Problems
13.1.2 Axisymmetric ow
Write the counterparts of equations (13.1.1)(13.1.3) for axisymmetric ow in terms of the Stokes
stream function.
Computer Problems
+ ( u) = 2 . (13.2.1)
t
Taking the divergence of (13.2.1) and recalling that the divergence of the curl of any twice dier-
entiable vector function is identically zero, we nd that the divergence of the vorticity, ,
evolves according to the unsteady heat conduction equation,
= 2 . (13.2.2)
t
Accordingly, the computed vorticity eld will be solenoidal, as required, provided that it is solenoidal
at the initial instant and the boundary distribution of remains zero at all times (Problem 13.2.1).
It has been argued that it is not entirely appropriate to impose an explicit local boundary
condition for the vorticity [155]. Instead, the boundary distribution of the vorticity should emerge
as part of the solution in terms of the specied boundary conditions for the velocity or traction.
Computational experiments have shown that computing instead of imposing the boundary values of
the vorticity promotes the stability of the numerical method at the cost of increased programming
complexity and computational eort.
Quartapelle & Valz-Gris [332] replaced the boundary conditions for the vorticity with an
integral constraint. In the case of two-dimensional ow with homogeneous boundary conditions for
the velocity, the constraint requires that the vorticity is orthogonal to any nonsingular harmonic
function dened in the domain of ow, that is, the integral of the vorticity multiplied by any
nonsingular harmonic function over the area of the ow is zero. The implementation of this method
is discussed by Quartapelle [331] and Anderson [8] in the context of Chorins vortex sheet method
(Section 11.6).
980 Introduction to Theoretical and Computational Fluid Dynamics
Noninertial frames
When the ow is described in a noninertial frame that translates with velocity V(t) while rotating
with angular velocity (t) about a chosen point, we work with the modied vorticity W = + 2,
which evolves according to the standard vorticity transport equation (13.2.1) for an inertial frame, as
discussed in Section 3.12.5 [383]. Since the acceleration of the frame of reference enters the solution
only through the boundary condition, using the modied vorticity simplies the implementation and
reduces the computational cost.
u = + A, (13.2.3)
as discussed in Section 2.8 [181]. In the case of ow past a solid boundary, the scalar potential
is computed by solving Laplaces equation subject to the no-penetration boundary condition,
n = n u = 0. The solution is unique up to an arbitrary but physically irrelevant constant. To
compute the vector potential, we write
= A = ( A) 2 A, (13.2.4)
2 A = . (13.2.5)
Taking the divergence of (13.2.5) and remembering that the vorticity eld is solenoidal, we nd
that the divergence A satises Laplaces equation, which means that the computed A will be
solenoidal provided that A = 0 over all boundaries.
n (A n) = 0. (13.2.6)
To validate this stipulation in light of the no-slip boundary condition, and also derive a boundary
condition for the normal component of A, it is convenient to introduce a local coordinate system
13.2 Vorticity transport equation for three-dimensional ow 981
n y
Figure 13.2.1 Illustration of a local coordinate system on a solid boundary used to derive a boundary
condition for the normal component of the vector potential.
where the x and z axes are tangential to the boundary and the y axis is normal to the boundary
at a point, as shown in Figure 13.2.1. The boundary condition (13.2.6) species that Ax = 0 and
Az = 0 at the origin. The normal component of the velocity is
Ax Az A A
( A)y = = tx tz (13.2.7)
z x z x
evaluated at the origin, where tx and tz are unit tangent vectors along the x and z axes. Rearranging,
we obtain
(A tx ) (A tz ) tx tz
( A)y = A( ). (13.2.8)
z x z x
Now implementing the boundary condition (13.2.6), we obtain
tx tz n n
( A)y = Ay n ( ) = Ay n (tx tz ), (13.2.9)
z x z x
which is zero in light of (1.7.13), consistent with the requirement that n ( A) = 0, originating
from the no-penetration boundary condition.
x
A = n (A) n + + A + , (13.2.11)
x z x z
yielding
A = n (A) n + 2m A n = 0, (13.2.12)
982 Introduction to Theoretical and Computational Fluid Dynamics
or
n (A) n = 2m A n, (13.2.13)
where m is the mean curvature of the boundary. In the case of a at boundary, n (A) n = 0.
Boundary conditions for multiply connected domains are available [341]. The numerical imple-
mentation of the vector potentialvorticity formulation for three-dimensional ow has been discussed
on several occasions (e.g., [11, 20, 252]). In the case of two-dimensional or axisymmetric ow, all
components of A are set to zero, except for the z or azimuthal components that are identied, respec-
tively, with the stream function or with the Stokes stream function divided by the distance from the
axis of revolution, . The vector potentialvorticity formulation then reduces to the vorticitystream
function formulation discussed in Section 13.1.
To validate the method, we must show that the curl of the velocity computed by solving
(13.2.14) is identical to the vorticity, , provided that is solenoidal and its boundary values are
computed from the denition = u [100, 407]. Working toward this goal, we use the vector
identity ( u) = ( u) 2 u in conjunction with (13.2.14). Dening ( u)
and rearranging, we obtain
= ( u). (13.2.15)
Taking the curl of both sides, we nd that
= ( ) 2 = 0. (13.2.16)
2
Since and thus is solenoidal, = 0, we can be sure that = 0. Because the boundary
values of are zero, must vanish identically and the curl of the computed velocity, u, must
be equal to the prescribed vorticity throughout the domain of ow.
To show that the velocity computed by solving (13.2.14) is solenoidal, we invoke once again
the identity ( u) = (u) 2 u. Using (13.2.14), we nd that ( u) = (u)+,
which shows that ( u) = 0. Integrating in space, we nd that u is a constant equal to zero
due to mass conservation throughout the domain of ow [100, 407].
In another version of the velocityvorticity formulation, the velocity eld is computed directly by
solving the CauchyRiemann system of equations = u and u = 0 for the velocity, subject
to the no-penetration boundary condition [142, 288].
Problem
We begin by recasting the NavierStokes equation as an evolution equation for the velocity
involving the pressure,
u 1
= N(u) p + L(u), (13.3.1)
t
where the terms N(u) and L(u) are associated, respectively, with the nonlinear inertial force and
the linear viscous force,
Because the uid is incompressible, the velocity eld is solenoidal and the term ( u) in the
denition of L(u) is identically zero. However, because numerical discretization destroys the exact
equality, keeping this term helps ensure numerical stability and minimize the numerical error.
The equation of motion is accompanied by the continuity equation for incompressible uids,
u = 0, (13.3.3)
stating that the velocity eld is and must remain solenoidal during the evolution.
The continuity equation requires that the right-hand side vanishes at all times. Accordingly, the
pressure must satisfy the pressure Poisson equation (PPE)
One may argue that, since the divergence and the linear operator L commute, the second term on
the right-hand side of (13.3.5) vanishes, yielding the simplied form
2 p = N(u). (13.3.6)
For reasons that will become evident in our further analysis, equation (13.3.5) is called the consistent
PPE and equation (13.3.6) is called the simplied PPE [153].
Substituting (13.3.6) into (13.3.4) and interchanging the divergence with the Laplacian, we
obtain an unsteady diusion equation for the rate of expansion,
= L(). (13.3.8)
t
The general properties of the unsteady diusion equation in a bounded domain show that the rate
of expansion will vanish at all times provided that (a) the initial velocity eld is solenoidal and
(b) the rate of expansion or its normal derivative are zero over all boundaries of the ow at all
times. When these conditions are met, it is permissible to replace the continuity equation with the
simplied pressure Poisson equation (13.3.5). The second condition underlines the importance of
accurately satisfying mass conservation at the grid points near the boundaries. When the initial
rate of expansion is not zero but the boundary distribution is kept zero during the evolution, the
magnitude of the divergence of the velocity will keep decreasing and eventually vanish during the
evolution.
at all times. The derivation of proper boundary conditions for the pressure is discussed in two
illuminating articles by Orszag, Israeli & Deville [285] and Gresho & Sani [153].
2u
n p = n, (13.3.10)
2
where n is the normal unit vector directed into the ow and is the arc length normal to the wall
measured in the direction of the uid. When the Reynolds number is suciently high, the right-hand
side of (13.3.10) is small and can be set to zero in order to simplify the implementation.
Implicit implementations
In a certain class of nite-dierence methods discussed in Section 13.4, the Neumann condition
(13.3.9) is not explicitly enforced. Instead, the numerical method employs an alternative boundary
condition that emerges by enforcing the continuity equation at grid nodes adjacent to the boundaries.
The consistent implementation of this constraint is equivalent to the Neumann boundary condition
(13.3.9). Other implementations have been proposed. For example, Quartapelle & Napolitano [333]
replaced the boundary condition for the pressure with an integral constraint involving the projection
of the boundary distribution of the pressure onto a nonsingular solution of the vectorial Helmholtz
equation dened within the available domain of ow.
2 p = w, (13.3.11)
986 Introduction to Theoretical and Computational Fluid Dynamics
Integrating the pressure Poisson equation over the domain of ow and using the divergence
theorem, we nd that a solution will exist only when a compatibility condition is fullled,
2
p dV = n p dS (13.3.13)
F low B
or
w dV = q dS, (13.3.14)
F low B
where B stands for the boundaries of the ow and n is the normal unit vector pointing into the ow.
The two-dimensional counterpart of (13.3.14) is
w dA = q dl, (13.3.15)
F low B
where l is the arc length around the boundary, B. When the compatibility condition is fullled, the
pressure can be determined only up to an arbitrary constant. When the compatibility condition is
not fullled, a solution for the pressure cannot be found.
The solvability condition (13.3.17) is the discrete implementation of the compatibility con-
dition (13.3.14). In this light, the left-hand side of (13.3.17) is recognized as the implementation
of a numerical integration quadrature pertinent to the volume or surface integrals on the left- and
right-hand sides of (13.3.14). The particular nature of this quadrature depends on the structure of
the matrix A, which is determined by the method used to discretize the Laplacian in the pressure
Poisson equation. This observation reveals an intimate relation between the numerical dierentiation
matrix embedded in A and the singular eigenvector of its transpose of this matrix with reference to
numerical integration.
13.3 Velocitypressure formulation 987
Regularization
In an alternative and more general approach, the right-hand side of the Poisson equation, 2 p = w,
is modied by an appropriate amount so as to yield a singular system with an innite number of
solutions [57, 144, 145, 173]. This is done by replacing w with w f , where f is a specied function
independent of w, and is a small dimensionless number to be found as part of the solution.
The pressure is computed by solving the altered pressure Poisson equation 2 p = w f . The
compatibility condition is satised if
= w dV + q dS / f dV. (13.3.18)
F low B F low
1. Solve the rst N 1 equations of the system (13.3.16) for the rst N 1 unknowns, set the
last unknown to zero, N = 0, and call the solution (1) .
(1)
2. Compute the residual of the last equation R(1) AN i i bN .
3. Solve the rst N 1 equations of the system (13.3.16) with q = 0 for the rst N 1 unknowns,
set the last unknown to zero, wN = 0, and call the solution (2) .
(2)
4. Compute the residual of the last equation R(2) AN i i bN .
(1)
5. Set = R (1)
/R (2)
and compute the nal solution = (2) .
The nal solution satises all N equations of the perturbed linear system A = b(w + f, q).
988 Introduction to Theoretical and Computational Fluid Dynamics
To formalize and optimize the method, we replace the linear system (13.3.16) with the per-
turbed system
A = b c, (13.3.19)
where c is a suitable vector normalized so that c c = 1, and the constant on the right-hand
side is adjusted to ensure the satisfaction of the solvability condition When the adjoint eigenvector
v is available, we may enforce the solvability condition by setting = (v b)/(v c). When the
linear system (13.3.16) is solved by iteration, the regularization embodied in (13.3.19) with a uniform
vector c can be implemented by shifting all components of the solution vector w by the same amount
after each iteration, so that one arbitrarily chosen component is anchored at a xed value.
The best way of regularizing the linear system (13.3.16) is by projecting the right-hand side
onto the orthogonal complement of the adjoint eigenvector of the matrix A, denoted by v and
satisfying v A = 0, thereby obtaining the regularized system
A = (I v v) b, (13.3.20)
13.3.7 Assessment
Comparing the formulation in primitive variables to the formulations based on the vorticity trans-
port equation discussed in the preceding sections of this chapter, we identify relative strengths and
weaknesses. One weakness is the need to derive boundary conditions for the pressure, and one
strength is the ease in extending the implementation to multi-uid and interfacial ows.
13.4 Implementation in primitive variables 989
Problem
Computer Problem
The nite-dierence method is based on the explicit forward-time discretization of the equation
of motion (1.4.32), yielding the velocity at the next time level, n + 1, from the velocity and pressure
at the current time level, n,
1
( u)n+1 ( u)n 1
= N(un ) + L( un ) 2 pn . (13.4.2)
t
Requiring that the divergence of the velocity is zero at the n + 1 time level, we obtain a Poisson
equation for the pressure,
2 pn = ( u)n + N(un ) + L( un ), (13.4.3)
t
which is a modication of the consistent pressure Poisson equation. Although small, the rst and
third terms on the right-hand side of (13.4.3) should be retained to prevent the onset of numerical
instability. The computational algorithm involves the following steps:
990 Introduction to Theoretical and Computational Fluid Dynamics
1. Specify an initial solenoidal velocity eld that satises the prescribed boundary conditions.
2. Compute the pressure eld by solving the Poisson equation (13.4.3), subject to the Neumann
boundary condition (13.3.9). In practice, this is done using an alternative but equivalent set
of boundary conditions to be discussed shortly.
3. Use (13.4.1) to advance the velocity eld by one step in time, subject to the prescribed bound-
ary conditions. The spatial derivatives on the right-hand side of (13.4.1) are computed using
central dierences. The size of time step is kept suciently small to suppress the growth of
numerical oscillations.
4. Return to Step 2 and repeat the computation for another time step.
The method can be implemented on a staggered grid that decouples velocity from pressure nodes,
or on a nonstaggered grid where all variables are dened at the same nodes.
Discrete values of the pressure are assigned to the primary nodes, (i, j), dened by the in-
tersection of the solid lines, shown as circles. Discrete values of the x component of the velocity
are dened at the intersection of horizontal primary grid lines and vertical secondary grid lines,
(i , j), shown as horizontal arrows. Discrete values of the y component of the velocity are dened at
the intersection of vertical primary grid lines and horizontal secondary grid lines, (i, j ), shown as
vertical arrows.
A distinguishing feature of the staggered-grid method is that the unknown functions and gov-
erning equations are dened or enforced at dierent nodes. This decoupling simplies the numerical
implementation and promotes the stability of the numerical method. We will see that distributing
the pressure nodes in the interior of the ow excluding the boundaries bypasses the derivation of
explicit boundary conditions for the pressure.
1 1
ui +1/2,j = (ui ,j + ui +1,j ), ui ,j+1/2 = (ui ,j + ui ,j+1 ).
2 2
1 1
vi+1/2,j = (vi,j + vi+1,j ), vi,j +1/2 = (vi,j + vi,j+1 ). (13.4.4)
2 2
13.4 Implementation in primitive variables 991
N y +2
V
by N y+1
N y+1
Ny
Ny
3
3
j 2
j 2
1
ay
(i,j)1 2 3 Nx N x+1
1
i
(i,j) 1 2 3 Nx N+1 Nx+2
x
i
x
ax bx
Figure 13.4.1 A staggered grid for computing ow in a rectangular domain representing a cavity. The
primary grid is drawn with heavy lines and the secondary grid is drawn with broken lines. Note
that the secondary grid conforms with the boundaries of the ow. The pressure is dened at the
circles, the x and y velocity components are dened at the horizontal or vertical arrows.
Unphysical x velocity components are required at the horizontal lines j = 1 and j = Ny + 2 for the
computation of the second y derivative of u near the top and bottom boundaries. Corresponding
unphysical y velocity components are required at the vertical levels i = 1 and Nx + 2 for the
computation of the second x derivative of v near the left and right boundaries. These exterior
velocities are computed by extrapolation to satisfy the boundary conditions at the physical levels
i = 1, i = Nx + 1, j = 1, and j = Ny + 1.
For example, approximating u with a parabolic function near the lid located at y = by , and
enforcing the no-slip boundary condition u = V , we obtain
u = V + A (y by )2 + B (y by ), (13.4.5)
where V is the lid velocity, and A, B are unknown coecients. Applying this expression at three
neighboring grid levels, we obtain
9 2 3 1 2 1
ui ,Ny = V + A h B h, ui ,Ny +1 = V + A h B h,
4 2 4 2
1 2 1
ui ,Ny +2 = V + A h + B h, (13.4.6)
4 2
992 Introduction to Theoretical and Computational Fluid Dynamics
uni,2 uni1,2
n+1
u u
n
vi,2 vi,1 i,2 i1,2 vi,2
n
+ + (13.4.16)
t x y x y
u u
i,2 i1,2 vi,2
n pi+1,2 2 pi,2 + pi1,2 pi,3 pi,2
n
+ + 2
+ = 0.
x y x y 2
The v velocity at the (i, 1) node is available from the prescribed boundary conditions. Straightfor-
ward modications are necessary for the two corner nodes, i = 2 and Nx + 1. Working in a similar
fashion for the last horizontal layer corresponding to j = Ny + 1 for i = 3, . . . , Nx , we obtain
To implement an explicit Neumann boundary condition for the pressure, we introduce the
exterior pressure nodes pi,1 . Using the continuity equation, we nd that v/y vanishes at the
bottom wall. Applying (13.3.10) at the lower wall located at y = ay , and approximating the
derivatives using central dierences, we obtain
p
pi,2 pi,1 2u
vi,2
y
n p = = 2 . (13.4.18)
y y=ay y y y=ay y
994 Introduction to Theoretical and Computational Fluid Dynamics
Rearranging, we nd that
vi,2
pi,1 = pi,2 2 , (13.4.19)
y 2
where all variables are evaluated at the nth time level. Working in a similar fashion with the upper,
left, and right walls, we nd that
vi,Ny +1 u2,j
pi,Ny +2 = pi,Ny +1 + 2 , p1,j = p2,j 2 ,
y x
(13.4.20)
uN +1,j
pNx +2,j = pNx +1,j + 2 x .
x
Equations (13.4.19) and (13.4.20) provide us with an alternative set of boundary conditions for the
Poisson equation (13.4.15), which is now also applied at pressure nodes adjacent to the boundaries,
i = 2, Nx + 1 and j = 2, Ny + 1.
It appears that, by using a staggered grid, we have circumvented the derivation of explicit
boundary conditions for the pressure. In fact, equation (13.4.16) and its counterparts for the other
three walls amount to the Neumann pressure boundary condition stated in(13.3.9) [153]. To see
this, we subtract (13.4.16) from (13.4.15) written for j = 2, and obtain
n
vi,j1 n
vi,j1 pni,j pni,j1
i,j1 ui,j1 ) +
(un+1 n
+ = 0. (13.4.21)
t y y y 2
As y tends to zero, the rst and second terms on the left-hand side vanish and the remaining terms
reproduce the Neumann boundary condition (13.4.19).
Assessment
Bypassing the derivation of the boundary conditions for the pressure is a signicant advantage of
the staggered-grid implementation. Unfortunately, the numerical formulation becomes considerably
more involved and prohibitively expensive for grids dened in curvilinear coordinates.
(13.4.1), we obtain the counterpart of equation (13.4.16). Working in a similar fashion, we derive
corresponding equations for the upper, left, and right walls.
It appears that we have again circumvented the explicit derivation of boundary conditions
for the pressure. However, it can be shown that (13.4.22) and its counterparts for the three other
walls amount to the Neumann boundary conditions (13.3.9) [153]. This is true even when rst-order
one-sided dierences are used to approximate the second partial derivatives. Unfortunately, solving
the Poisson equation by this method is susceptible to numerical instability due to the decoupling
of the pressure at adjacent nodes. A small numerical incompressibility must be tolerated in the
nite-dierence solution.
t
t
un+1 = un + N(un+1 ) + N(un ) + [ L(un+1 ) + L(un ) ] pn+1/2 . (13.4.23)
2
Since the method is implicit in the nonlinear terms, carrying out each time step requires solving a
system of nonlinear algebraic equations for the velocity at the n + 1 time level and for the pressure
at the n + 12 time level. In practice, this is done by iteration ([294], p. 167).
Problems
Computer Problem
Since the pressure p has not been updated in the convectiondiusion step, it loses its physical
meaning and must be regarded as an auxiliary function whose main purpose is to project the interme-
diate velocity, u , onto the subspace of solenoidal functions. To indicate this subtle dierentiation,
we replace the pressure p in (13.5.2) with a projection function, , writing
u 1
= . (13.5.3)
t
In the case of a constant density uid, the right-hand side of this equation is irrotational. The
relation between the projection function and the pressure will be discussed later in this section.
boundary conditions for u should be chosen to minimize its divergence while ensuring that the
stipulated boundary conditions for the physical velocity are observed at the end of a complete
time step. The derivation of boundary conditions for u and has been the subject of extensive
investigation [285].
A consistent set of boundary conditions include the Dirichlet boundary condition for the
normal and tangential velocity components of the intermediate velocity,
n
1 t +t
u n = Vn+1 n, u (I n n) = Vn+1 + dt (I n n), (13.5.4)
tn
where V is the prescribed boundary velocity, I is the identity matrix, and I n n is the tangential
projection operator. Correspondingly, the projection function satises the homogeneous Neumann
condition
tn +t
n dt = 0. (13.5.5)
tn
Straightforward substitution demonstrates that (13.5.4) and (13.5.5) ensure the required boundary
condition at the end of a complete step, un+1 = Vn+1 . In expressions (13.5.4) and (13.5.5), the
projection function is regarded as a continuous function of time over the interval where equation
(13.5.3) applies. In numerical practice, the continuous evolution is replaced by an instantaneous
projection.
Optimal and simplied boundary conditions with varying degrees of accuracy and ease of
implementation are available [154]. In the simplest scheme, the intermediate velocity, u , satises
the boundary conditions stipulated at the n + 1 time level, and the projection function satises
the homogeneous Neumann boundary condition n = 0. Improved boundary conditions are
discussed in Section 13.5.4.
= N + 2 , (13.5.6)
t
and
= 2 , (13.5.7)
t
where u is the rate of expansion. The convectiondiusion equation (13.5.6) receives the
presumably vanishing rate of expansion at the tn time level, and advances it over a time interval t
to an intermediate nonzero eld, . The projection equation (13.5.7) receives and advances it
over the same time interval t to the nal eld, n+1 , which is required to vanish throughout the
domain of ow and over the boundaries.
998 Introduction to Theoretical and Computational Fluid Dynamics
u u 2u u u 2u u u 2u
+ ux = , + uy = , + uz = . (13.5.8)
t x x2 t y y 2 t z z 2
Each component equation applies for the full time interval, t, and time is reset to the initial value,
tn , at the end of each fractional step. The three intermediate velocity elds, denoted by
are not generally solenoidal. In the case of two-dimensional ow, only the rst two steps in (13.5.8)
are present and the intermediate velocity elds are denoted by
The three steps in (13.5.8) can be carried out using the implicit BTCS method or the CrankNicolson
method with rst- or second-order accuracy in time, respectively. Both methods are unconditionally
stable and require the easy task of solving tridiagonal systems of equations. To express the con-
stituent convectiondiusion equations in conservative form, we may set the convection velocity in
all three steps equal to the presumed solenoidal velocity at the beginning of the step corresponding
to the time level, tn . Boundary conditions for the intermediate velocities will be derived.
Approximating the time derivative in the projection step in (13.5.3) with a dierence equation
and rearranging, we obtain
t
un+1 = u n+1 . (13.5.11)
To compute n+1 , we take the divergence of (13.5.11) and require that un+1 = 0 to derive a
Poisson equation,
2 n+1 = u , (13.5.12)
t
accompanied by the boundary condition
n n+1 = 0. (13.5.13)
As in the case of the Poisson equation for the pressure discussed in Section 13.3, the explicit imple-
mentation of this boundary condition can be circumvented by using a staggered grid and requiring
that un+1 = 0 at pressure nodes adjacent to the boundaries, or by using a non-staggered grid
and requiring that un+1 = 0 at nodes located at the boundaries.
Given an initial velocity eld along with a prescribed boundary condition, u = V, we compute
the evolution of the ow according to the following steps:
13.5 Operator splitting, projection, and pressure-correction methods 999
1. Assign the initial velocity to velocity nodes and provide an estimate for the projection function,
n+1 , at the pressure nodes.
2. Advance the velocity eld sequentially according to equations (13.5.8) with boundary condi-
tions u n = V n for the normal velocity component and
t
u (I n n) = (V + n+1 ) (I n n) (13.5.14)
for the tangential velocity component, where I is the identity matrix and I n n is the
tangential projection operator.
3. Solve the Poisson equation (13.5.12) with the Neumann boundary condition (13.5.13). Since
the ow rate of u and thus of across the boundaries of the ow vanishes because of the
imposed boundary conditions on u , the compatibility condition is automatically satised and
the Poisson equation has an innite number of solutions.
4. Compute the velocity un+1 at all internal and boundary grid points according to (13.5.11). If
the tangential boundary velocity is not equal to the specied velocity, producing a nonzero slip
velocity, return to Step 2 and repeat the computations with the current boundary distribution
of . Otherwise, proceed to Step 5.
5. Reset the time to tn+1 , return to Step 2, and repeat the computations for another time step.
Consider the familiar problem of two-dimensional ow in a cavity driven by a moving lid, illustrated
in Figure 13.1.1. The convectiondiusion equations in the x and y directions can be integrated
in time using an implicit method, such as the CrankNicolson method. When integrating in the x
direction, we use the side wall boundary conditions
t n+1
t n+1
where u = ux and v = uy . Boundary conditions over the upper and lower walls are not required.
When integrating in the y direction, we use the lower and upper wall boundary conditions
t n+1
t n+1
and
n
= Nn+ + 1 ) 2 n + 2 , (13.5.18)
t
where the numerical parameter varies in the interval (0, 1]. The term Nn+ is assumed to be
available with O(tm ) accuracy in terms of the extrapolated or interpolated velocity of the physical
velocity eld evaluated at time tn + t. When m = 2, we obtain second-order accuracy.
Now turning to the projection step, we replace the corresponding evolution equations (13.5.3)
and (13.5.7) with the dierence equations
un+1 u
= n+1 , (13.5.19)
t
and
n+1
= 2 n+1 . (13.5.20)
t
Solving (13.5.20) for , we obtain
t 2 n+1
= n+1 + . (13.5.21)
Now using (13.5.19) to eliminate the intermediate velocity u from (13.5.17), we obtain the
equation
un+1 un
= Nn+ (n+1 ) + (1 ) 2 un + 2 un+1 . (13.5.23)
t
Comparing this expression with the equation of motion discretized according to the generalized
CrankNicolson method shows that second-order temporal accuracy for the velocity is achieved
when = 12 , the pressure is evaluated according to
1 1 1
pn+ 2 = n+1 + O(t2 ) = n+1 t 2 n+1 + O(t2 ), (13.5.24)
2 2
and the nonlinear term is evaluated with second-order accuracy by extrapolation or interpolation, so
that m = 2 [59, 206]. Equation (13.5.24) provides us with a relation between the projection function
and the pressure, worthy of further investigation.
13.5 Operator splitting, projection, and pressure-correction methods 1001
We can infer the relation between the projection function and the pressure by taking the
divergence of the equation of motion and enforcing the incompressibility condition to derive the
pressure Poisson equation. Applying the Poisson equation at time tn+ , we obtain
2 pn+ = Nn+
E , (13.5.29)
where the subscript E denotes the exact value. Subtracting (13.5.27) from (13.5.29) and rearranging,
we nd that
2 pn+ n+1 + t 2 n+1 + Nn+ Nn+ = (Rn+1 R+ n
). (13.5.30)
E
t
By denition, the expression enclosed by the square brackets on the left-hand side is of O(tm ).
2 un+
E = (1 ) 2 un + 2 un+1 + O(tl ), (13.5.33)
where l is the temporal order of the physical velocity. Using (13.5.17) and the Laplacian of (13.5.19)
to eliminate the Laplacian of the velocity, we nally obtain
u un u n+
B
n = n + O(tl ) + O(tm ) . (13.5.34)
t t
In light of the rst boundary condition in (13.5.4), the right-hand side of (13.5.34) vanishes up to
order tm or tl . Consequently, the harmonic function is constant within this precision, and
equation (13.5.31) yields
or
t 2 = 0. (13.5.37)
Balancing the magnitude of the two terms on the left-hand side, we nd that is supported by
boundary layers of thickness (t)1/2 . Since the normal derivative of at the boundary is
zero, the corresponding normal derivative of is comparable to that of the pressure according to
(13.5.36). Consequently, is of order inside the boundary layer. Correspondingly, the second
term on the right-hand side of (13.5.36) makes a leading-order contribution of order inside the
boundary layer and decays exponentially outside the boundary layer [117, 154, 393]. As tends
to zero, yielding an explicit discretization, the thickness of the boundary layer vanishes and the
numerical method expectedly fails.
Under more general conditions of unsteady NavierStokes ow and in the presence of a dis-
tributed body force, the right-hand side of (13.5.25) is signicant throughout the domain of ow
and is of order t in the interior of the ow and varies by an amount of order t1/2 across the
boundary layers.
In the fully discrete time-space implementation, the right-hand side of (13.5.30) does not
generally vanish to machine accuracy, and is of order t hm , where the exponent m is determined
13.5 Operator splitting, projection, and pressure-correction methods 1003
by the spatial discretization. Accordingly, the right-hand sides of (13.5.35), (13.5.36), and (13.5.24)
include corrections that may lower the nominal order of the numerical method.
un+1 un 1 1
= n+1 + t 2 n+1
t 2
1 1
+ 3 N(un ) N(un1 ) + L(un+1 ) + L(un ) , (13.5.40)
2 2
which makes a clear distinction between p and . Setting
1
pn+1/2 n+1 + t 2 n+1 (13.5.41)
2
shows that (13.5.39) is second-order accurate in time.
To carry out the convectiondiusion step, we recast (13.5.38) into the form
1 1
(1 t 2 )(u un ) = t 3 N(un ) N(un1 ) + t 2 (un ), (13.5.42)
2 2
and then approximately factorize the operator on the left-hand side,
1 1 2 1 2 1 2
(1 t 2 ) (1 t 2 )(1 t 2 )(1 t 2 ). (13.5.43)
2 2 x 2 y 2 z
To compute the dierence u un , we solve three tridiagonal systems of equations using the Thomas
algorithm discussed in Section B.1.4, Appendix B (Problem 13.5.2). Boundary conditions for u
arise as discussed in Section 13.5.2. The Poisson equation (13.5.12) can be solved on a staggered
grid to avoid the explicit implementation of boundary conditions for the projection function, .
un+1 un 1 1 1
= N(un+1/2 ) ( + ) + L(un+1 ) + L(un ) + t L(), (13.5.46)
t 2 2
which implements a second-order discretization of the equation of motion provided that pn+1/2 =
+ and the last term on the right-hand side is zero. The algorithm can be implemented according
to the following steps [443]:
1. Given the pressure, pn1/2 , and the velocity, un , estimate un+1/2 by extrapolation and set
= pn1/2 .
2. Calculate the nonlinear term, N(un+1/2) ), compute u from (13.5.44), and call the solution
u,k , where k is an inner iteration number.
3. Introduce the discrete form of the equation of motion
un+1,k un 1 1
= N(un+1/2 ) pn+1/2,k + L(un+1,k ) + L(un ) , (13.5.47)
t 2 2
take its divergence, and require that un+1,k is solenoidal to obtain a Poisson equation for
pn+1/2,k . Solve the Poisson equation subject to appropriate boundary conditions and use
(13.5.47) to compute un+1,k .
4. Set = pn+1/2,k , return to Step 2 and repeat the computations with un+1 = un+1,k , where
un+1/2 is computed by interpolation or extrapolation, increasing k by one.
Problems
Computer Problem
The main idea behind the methods of modied dynamics is to amend either the continuity
equation or the equation of motion so as to render all governing equations parabolic in time. In
certain cases, the error introduced by modifying the original equations is mild and the transient
or steady solution obtained by solving the modied equations describes the physical evolution with
acceptable accuracy. In other cases, the transient evolution is purely ctitious and thus physically
irrelevant, and the temporal evolution is signicant only insofar as to provide us with a venue that
leads us to a steady state. The modied equations are designed so that the solution obtained
at steady state satises the equations of steady incompressible Newtonian ow within a specied
tolerance.
To expedite the approach to steady state, it is desirable to use larger times steps. However, a
semi-implicit or fully-implicit method must be employed to ensure numerical stability. Chorin [79]
implemented a variation of the explicit CTCS method according to the Du FortFrankel scheme.
The nite-dierence equation corresponding to the x component of the equation of motion at the
(i, j) grid point is
1 pn
un+1
ij = un1
ij + 2t + Fx , (13.6.2)
x
where
un un+1
and u = ux . The nonlinear term, N, is discretized using centered dierences in its conservative form.
In Chapter 12, we saw that the modied dierential equation corresponding to the Du FortFrankel
scheme is not consistent with the original dierential equation. However, since the dierence is
a small term involving second partial derivatives with respect to time, the modied and original
equations agree at steady state. The numerical scheme can be extended in a straightforward fashion
to three-dimensional ow where the Du FortFrankel scheme is applied to one out of the three second
spatial derivatives. Chorin [79] nds that, given boundary conditions for the velocity,
and provided
that the ow remains subsonic with respect to the articial speed of sound 1/ , the method is
stable as long as the maximum Courant number is less than 2(/n)1/2 /(1 + 51/2 ), where n = 2 or 3
for two- or three-dimensional ow.
pn+1
ij = p n
ij + , (13.6.4)
x y
13.6 Methods of modied dynamics or false transients 1007
where u = ux and v = uy . One notable feature of the articial compressibility method implemented
on a staggered grid is that boundary conditions for the pressure are not required. When steady state
has been reached, the pressure satises the pressure Poisson equation with Neumann boundary
conditions that arise by projecting the equation of motion onto the unit vector normal to the
boundaries.
= p N(u) + , (13.6.5)
t t
where is a positive constant. At steady state, equation (13.6.5) reduces to the familiar Poisson
equation for the pressure. The transient solution lacks physical meaning and equation (13.6.5) is
signicant only insofar as to provide us with a route toward the steady state.
The computations proceed by integrating (13.6.7) forward in time from a given initial state subject
to the specied velocity boundary conditions. The penalty-function formulation has found extensive
applications in numerical procedures based on nite-element methods [196].
Problem
Continuity
If the graph of a function does not exhibit sudden jumps inside a certain interval, a x b, the
function is continuous in that interval.
Intermediate-value theorem
Assume that a function, f (x), is continuous in an interval, a x b, and denote the minimum
value of the function as min inmum[f (x)], and the maximum value of the function as max
supreme[f (x)]. The intermediate-value theorem asserts, and geometrical intuition conrms, that,
for any number c such that min c max, there is at least one point, , in the interval a b,
such that f () = c, as illustrated in Figure A.1.1(a).
1008
A.1 Functions of one variable 1009
(a) (b)
y y
y=f(x)
max
c
min
x x
a b a b
Figure A.1.1 Graphical interpretation of (a) the intermediate-value theorem and (b) the mean-value
theorem.
Let w(x) be a nonnegative function whose denite integral is nite in an interval, a x b. If the
function f (x) is continuous in that interval, then at least one point, , can be found in the interval
a b, such that
b b
f (x) w(x) dx = f () w(x) dx. (A.1.2)
a a
The value f () is a weighted average of the function f (x). If w(x) is constant, the value f () is the
mean value of the function f (x).
If (a) a point, x, is close to another point, x0 , (b) the function f (x) is continuous in a neighborhood
of x0 , and (c) the graph of the function does not become vertical at x0 , it is reasonable to expect
that the dierence between f (x) and f (x0 ) will be proportional to the dierence between x and
x0 . The Taylor series provides us with an innite polynomial expression for the dierence in the
function values, f (x) f (x0 ), in terms of the corresponding dierence in the arguments, x x0 .
If the rst N + 1 derivatives of a function f (x) are continuous at every point between and
including the points x and x0 , then
N
1 (m)
f (x) = f (x0 ) + f (x0 ) (x x0 )m + RN (x, x0 ), (A.1.3)
m=1
m!
where the superscript (m) denotes the mth derivative and m! = 1 2 m is the factorial, subject
to the convention that 0 ! = 1. The remainder is given by
1
RN (x, x0 ) = f (N +1) () (x x0 )N +1 , (A.1.4)
(N + 1) !
1010 Introduction to Theoretical and Computational Fluid Dynamics
It is important to note that the remainder does not necessarily vanish in the limit N , but
may oscillate or grow unbounded. This means that the innite Taylor series without the remainder
does not necessarily converge to the function, f (x). The radius of convergence is determined by the
location of the pole of the complex function f (z) nearest x0 .
Problem
1
m
N
f (x, y) = f (x0 , y0 ) + (x x0 ) + (y y0 ) f (x, y) + RN (x, y, x0 , y0 ), (A.2.1)
m=1
m! x y x0 ,y0
where
1
N +1
RN (x, y, x0 , y0 ) = (x x0 ) + (y y0 ) f (x, y) (A.2.2)
(N + 1) ! x y x ,y
is the remainder. If x < x0 , x x x0 ; whereas, if x0 < x, x0 x x. Similar relations apply
for y . The operators on the right-hand sides of (A.2.1) and (A.2.2) are dened in terms of the
binomial expansion as
m
m m mf
1
Rational 1x
= 1 + x + x2 + x 3 + for 1 < x < 1
1
1+x = 1 x + x2 x3 + for 1 < x < 1
1 1 1
Exponential ex = 1 + x + 2 x2 + 6 x3 + + i! xi + for any real x
1 1 1
sinh x = x + 6
x3 + 5!
x5 + 7!
x7 + for any real x
1 1 1
cosh x = 1 + 2
x2 + 4!
x4 + 6!
x6 + for any real x
1 2 17
tanh x = x 3 x3 + 15 x5 315 x7 + for any real x
1 1
ax = 1 + x ln a + 2 (x ln a)2 + + i! (x ln a)i + for any real x
1 1 1
Logarithmic ln(1 + x) = x 2 x2 + 3 x3 4 x4 + . . . for 1 < x < 1
1 1 1
Trigonometric sin x = x 6 x3 5! x5 + 7! x7 + for any real x
1 1 1
cos x = 1 2 x2 + 4! x4 6! x6 + for any real x
1 2 17
tan x = x + 3
x3 + 15
x5 + 315
x7 + for any real x
1 3 5
Trigonometric arcsin x = x + 6
x3 + 40
x5 + 112
x7 + for any real x
1 3 5
arccos x =
2 x 6 x3 40 x5 112 x7 + for any real x
Table A.1.1 Maclaurin series expansions of common functions. A comprehensive compilation can be
found at the CRC Standard Mathematical Tables [444].
The rst set of large parentheses on the right-hand side of (A.2.3) denote the combinatorial,
#
l
m m! mp+1
= , (A.2.4)
k k ! (m k) ! p=1 p
where l is the minimum of k and m k. The combinatorial expresses the number of possible
combinations by which k objects can be chosen from a set of m identical objects, leaving behind
m k objects.
As the truncation level, N , tends to innity, we obtain an innite series polynomial represen-
tation. However, the remainder does not necessarily vanish, and the Taylor series does necessarily
converge as N . The Maclaurin series arises by setting x0 = 0 and y0 = 0.
1012 Introduction to Theoretical and Computational Fluid Dynamics
Problems
This property reveals that the delta function with argument of length has units of inverse
length.
3. The integral of the product of an arbitrary function, f (x), and the delta function over an
interval I that contains the point x0 is equal to value of the function at the singular point,
(x x0 ) f (x) dx = f (x0 ). (A.3.2)
I
The integral of the product of an arbitrary function, f (x), and the delta function over an
interval I that does not contain the singular point x0 is zero.
4. If a function, f (x), undergoes a discontinuity from a left value f to a right value f+ at the
point x0 , the derivative of the function at that point can be expressed in terms of the delta
function as
f (x0 ) = (f+ f ) (x x0 ) + , (A.3.3)
where the dots denote nonsingular terms.
A.3 Dirac delta function 1013
G 2
0
-1 -0.5 0 0.5 1
X
Figure A.3.1 A family of dimensionless test functions, G gL, described by (A.3.4), plotted against
the dimensionless distance X (x x0 )/L for = 1 (dashed line), 2, 5, 10, 20, 50, 100, and 500
(dotted line). As the dimensionless parameter tends to innity, we recover Diracs delta function
in one dimension.
In formal mathematics, (x x0 ) can be obtained from the family of Gaussian test functions,
1/2 x x
2
0
g(x x0 ) = exp , (A.3.4)
L2 L
in the limit as the dimensionless parameter tends to innity, where L is an arbitrary length. The
mean value of the Gaussian distribution is x0 and the variance is 2 = L2 /(2).
This property illustrates that the delta function in two dimensions has units of inverse squared
length.
3. The integral of the product of an arbitrary function, f (x, y), and the delta function over an
area D that contains the singular point, (x0 , y0 ), is equal to the value of the function at the
singular point,
2 (x x0 , y y0 ) f (x, y) dx dy = f (x0 , y0 ). (A.3.8)
D
The integral of the product of an arbitrary function, f (x, y), and the delta function over an
area D that does not contain the singular point, (x0 , y0 ), is equal to zero.
4. We may write
2 (x x0 , y y0 ) = 1 (x x0 )1 (y y0 ), (A.3.9)
r2
q(r) = 2
exp 2 , (A.3.10)
L L
in the limit as the dimensionless parameter tends to innity, where r |x x0 | and L is an
arbitrary length.
Graphs of the dimensionless function Q qL2 plotted against the dimensionless distance
X (x x0 )/L are shown in Figure A.3.2 for y = y0 , and = 1 (dashed line), 2, 5, 10, 20, 50, 100,
and 500 (dotted line). The maximum height of each graph is inversely proportional to the square of
its width, so that the area underneath each graph in the xy plane is equal to unity,
2
q(r) dx dy = q(r) r dr d = 2 q(r) r dr. (A.3.11)
0 0 0
independent of L.
A.4 Index notation 1015
Q 2
0
-1 -0.5 0 0.5 1
X
Figure A.3.2 A family of dimensionless test functions, Q qL2 , described by (A.3.10), plotted
against the dimensionless distance X (x x0 )/L for y = y0 and = 1 (dashed line), 2, 5, 10,
20, 50, 100, and 500 (dotted line). As the dimensionless parameter tends to innity, we recover
Diracs delta function in two dimensions.
Problems
where N in the last expression is the maximum value of the index i. The third expression denes
the inner product of a pair of vectors, u and v.
Free indices
The vector or matrix nature of an expression is determined by the number of free indices, that is,
the indices that appear only once in the expression. For example, Aij uj is a vector, whereas ui uj
and ui Aijl are two-dimensional matrices.
Kroneckers delta
Kroneckers delta, ij , represents the identity or unit matrix: ij = 1 if i = j, and ij = 0 if i = j.
Using this denition, we nd that
where the index i in the last expression ranges from 1 to N . If xi is a set of N independent variables,
then
xi xi
= ij , = N. (A.4.3)
xj xi
Alternating tensor
The alternating tensor ijk , where all three indices takes the values 1, 2, 3, is dened such that (a)
ijk = 0 if all three indices or any two indices have the same value, (b) ijk = 1 if the indices are
arranged in cyclic order, 123, 312, or 231, and (c) ijk = 1 otherwise. Thus, 132 = 1, 122 = 0,
ijj = 0, and ijk = jki = kij = jik . Formally, we dene
1
ijk = (i j)(j k)(k i). (A.4.4)
2
Two useful properties of the alternating tensor are
Antisymmetric matrix
A three-dimensional antisymmetric (skew-symmetric) matrix, Aij , has only three independent com-
ponents that can be represented in terms of a three-dimensional vector, v, as
Multiplying this denition by ijl and using (A.4.5), we obtain the inverse relation
1
vk = ijk Aij . (A.4.7)
2
For example, in uid mechanics the matrix A can be the vorticity tensor, , and the vector v can
be the vorticity vector, .
In uid mechanics, we avoid using the superscript T by indicating the one-index-product with
a centered dot. For example, v = A u is equivalent to vi = Aij uj , where summation is implied
over the repeated index, j. Similarly, w = u A is equivalent to wj = ui Aij , where summation is
implied over the repeated index, i. The scalar
s = (A u) v = v (A u) = (v A) u = u (v A) u (A.4.8)
is dened as s = vi Aij uj , where summation is implied over the repeated indices, i and j.
s u v = ui vi . (A.4.9)
A=uv (A.4.10)
Problem
Inner product
In Section A.4, we dened the inner product of a pair of vectors, a and b, as the scalar s ab = ai bi ,
where summation is implied over the repeated index, i. If the inner product is zero, the two vectors
are orthogonal. In two or three dimensions, the inner product of two vectors is equal to the product
of the lengths of the vectors, |a| and |b|, and the cosine of the angle subtended between the vectors
in their plane,
a b = |a| |b| cos . (A.5.1)
A similar interpretation applies in two dimensions.
Outer product
The outer or vector product of an ordered pair of three-dimensional vectors, a and b, is a new vector,
c a b, given by the determinant of a matrix,
e1 e2 e3
c a1 a2 a3 = e1 (a2 b3 a3 b2 ) + e2 (a3 b1 a1 b3 ) + e3 (a1 b2 a2 b1 ). (A.5.2)
b1 b2 b3
A.5 Three-dimensional Cartesian vectors 1019
In index notation,
ci = ijk aj bk . (A.5.3)
a b = b a. (A.5.4)
The outer product, c, is oriented normal to the plane of a and b. The direction of the outer product
is determined by the right-handed rule applied to the ordered triplet a, b, c.
The magnitude of the outer product is equal to the product of the lengths of the two vectors,
|a| and |b|, and the sine of the angle subtended between these vectors in their plane,
This expression demonstrates that, if the outer product is a null vector, the vectors a and b are
parallel, and vice versa.
V [a, b, c] (a b) c = (c a) b = (b c) a. (A.5.6)
Notice two cyclic permutations. The triple scalar product can be computed as the determinant of a
matrix,
a1 a2 a3
Since the determinant of a matrix is equal to the determinant of the matrix transpose,
a1 a2 a3 a1 b 1 c1
aa ab ac
If the vectors a, b, c are mutually orthogonal, the o-diagonal elements of the last matrix are zero.
d a (b c). (A.5.10)
1020 Introduction to Theoretical and Computational Fluid Dynamics
Writing
and then
a (b c) = b (a c) c (a b). (A.5.13)
a (b c) + b (c a) + c (a b) = 0. (A.5.14)
Notice the cyclic permutation of the vector triplets on the left-hand side.
e (b e) = b e (e b) (A.5.15)
e (b e) = b (I e e), (A.5.16)
where I is the identity matrix and denotes the tensor vector product.
Similarly, we dene the right cross product as a new matrix, C = A u, with elements
Problem
f f f
f = ex + ey + ez , (A.6.2)
x y z
where ei are unit vectors in the directions of the subscripted axes. Physically, the gradient points
into the direction where the function f increases most rapidly at a point. To prove this, we consider
the directional derivative of the function f , as discussed at the end of this section.
fi fx fy fz
f = = + + . (A.6.3)
xi x y z
If f = 0 at every point, the vector eld f is called solenoidal. The continuity equation requires
that the velocity eld of an incompressible uid is solenoidal throughout the domain of a ow.
f = trace(L). (A.6.5)
In fact, the velocity gradient is the tensor product of the gradient vector operator and the vector
eld of interest, f . To simplify the notation, we have denoted
f f . (A.6.6)
1022 Introduction to Theoretical and Computational Fluid Dynamics
f = det /x /y /z . (A.6.7)
fx fy fz
Explicitly,
f
z fy
f
x fz
f
y fx
f = ex + ey + ez . (A.6.8)
y z z x x y
In the case of the two-dimensional vector eld, f , lying in the xy plane, only the z component of the
curl survives.
2f 2f 2f
(f ) 2 f = + + . (A.6.9)
x2 y 2 z 2
The Laplacian is equal to the divergence of the gradient.
Vector identities
Let f be a scalar function and u, v be two vector functions. The following identities can be shown
working in index notation:
(f u) = f u + u f, (A.6.10)
(u v) = u v + v u + u ( v) + v ( u) , (A.6.11)
(u v) = v u u v, (A.6.12)
(u v) = u v v u + v u u v, (A.6.13)
(f ) = 0, (A.6.14)
( u) = 0, (A.6.15)
( u) = ( u) 2 u. (A.6.16)
Divergence of a matrix
The divergence of a two-dimensional matrix function of position, Q(x), is a vector, v = Q, with
components
Qij
vj = , (A.6.17)
xi
where summation is implied over the repeated index, i.
A.7 Divergence and Stokes theorems 1023
Directional derivatives
The rate of change of a scalar function, f , with respect to arc length, lt , in the direction of a unit
vector t, is given by
f f f f
t f = = tx + ty + tz . (A.6.18)
lt x y z
Using the geometrical interpretation of the inner vector product, we nd that t f is maximum
when t is parallel to f at a point.
f f f f
t f = = tx + ty + tz . (A.6.19)
lt x y z
Note that, by convention, the gradient f is multiplied from the left by t to produce a directional
derivative.
Problem
A.6.1 Curl
Compute the curl of the vector function v(x) = f (r) x, where r = |x|, x is the position vector, and
f (r) is a dierentiable function.
Making the three sequential choices f = (f, 0, 0), f = (0, f, 0), and f = (0, 0, f ), where f is
a dierentiable scalar function, we derive the vector form of the divergence theorem,
f dV = f n dS. (A.7.2)
Vc D
1024 Introduction to Theoretical and Computational Fluid Dynamics
Figure A.7.1 Illustration of (a) a volume used to derive the divergence theorem in space and (b) an
area used to derive the divergence theorem in the xy plane. (c) Illustration of a closed loop in
space, C, bounded by a surface, D, used to derive Stokes theorem.
The integral on the right-hand side involves the tangential component of the vector eld, q.
where l is the arc length along C and n is the unit vector normal to C pointing outward from Ac .
Making the sequential choices f = (f, 0) and f = (0, f ), where f is a dierentiable scalar
function, we obtain the vector form of the divergence theorem,
f dA = f n dl. (A.7.5)
Ac C
The particular choices f = x or f = y yield the area of Ac in terms of a line integral of the x or y
component of the normal vector.
A.7 Divergence and Stokes theorems 1025
where P = I n n is the tangential projection operator, l is the arc length along C, and t n = b
is the tangent unit vector shown in Figure A.7.1(c). The integrand on the left-hand side of (A.7.6) is
the surface divergence of the tangential component of f . If f is normal to D, the surface divergence
is identically zero.
Stokes theorem
Let C be an arbitrary closed loop with tangent unit vector t and D be an arbitrary surface bounded
by C, as illustrated in Figure A.7.1(c). Stokes theorem states that the circulation of a dierentiable
vector function f along C is equal to the ow rate of the curl of f across D,
f t dl = ( f ) n dS, (A.7.7)
C D
where l is the arc length along C and n is the unit vector normal to D oriented according to the
right-handed rule with respect to t and with reference to a designated side of D. As we look at
a designated side of D, the normal vector points toward us, while the tangent vector describes a
counterclockwise path.
Problems
A.7.2 Curl
0
Show that f (r) x t dl = 0, where r = |x|, x is the position vector, f (r) is a dierentiable function,
and t is the unit vector tangent to a closed contour.
1026 Introduction to Theoretical and Computational Fluid Dynamics
e
2
x
e3
e1
1
(ei ei ) ei
= 2 ei =0 (A.8.3)
j j
A.8 Orthogonal curvilinear coordinates 1027
for any j, where summation is not implied over the repeated index, i. This equation demonstrates
that the vector ei /j lacks a component in the direction of ei .
Orthonormality
If the base unit vectors form a local right-handed system of orthogonal axes, so that
ep eq = pqm em , (A.8.4)
then the variables 1 , 2 , and 3 comprise a system of orthogonal curvilinear coordinates, where
summation is implied over the repeated index, m. In that case, the inner product of any two dierent
base vectors gi and gj is zero, gi gj = 0 for i = j. Orthonormality of the base unit vectors requires
that
ei ej = ij , (A.8.5)
Metric tensor
An innitesimal vector in space can be described as
dx = g1 d1 + g2 d2 + g3 d3 , (A.8.6)
where gi are the local base vectors and di are appropriate dierential increments for i = 1, 2, 3.
Unless the base vectors are constant, equation (A.8.6) cannot be integrated readily to produce
an explicit expression for the position vector, x(1 , 2 , 3 ). Exceptions arise in the case of polar
coordinates discussed in Sections A.9A.11.
dx dx = gij di dj , (A.8.7)
where summation is implied over the repeated indices, i and j. We have introduced the diagonal
metric tensor g with diagonal components
where hi = |gi | are three positive metric coecients, and summation is not implied over the repeated
index, i. With these denitions, the base unit vectors are
g1 g2 g3
e1 , e2 , e3 . (A.8.9)
h1 h2 h3
The dierential arc length along a line over which 2 and 3 are constant but 1 changes, is
dl1 = h1 d1 . Similar relations can be written for the other two axes,
where dV (a) d1 d2 d3 . We see that the product h1 h2 h3 serves as a volume metric coecient
for the curvilinear coordinates. The size of a dierential surface element that lies in the surface over
which 1 , 2 , or 3 is constant is given, respectively, by
Vector components
The component of a vector function of position, f (x), in the direction of the unit vector, ei , is given
by the scalar product fi = f ei . By denition,
f = f1 e1 + f2 e2 + f3 e3 . (A.8.23)
To prevent misinterpretation the curvilinear components, fi , should not be collected into a vector.
Gradient operator
The denitions (A.8.24) motivate introducing the gradient operator
1 1 1
e1 + e2 + e3 = e1 + e2 + e3 . (A.8.25)
h1 1 h2 2 h3 3 l1 l2 l3
The unit vectors and dierentiation operators should not be transposed in these expressions.
1030 Introduction to Theoretical and Computational Fluid Dynamics
L f = e1 + e2 + e3 (f1 e1 + f2 e2 + f3 e3 ). (A.8.26)
h1 1 h2 2 h3 3
The expression enclosed by the rst set of parentheses on the right-hand side originates from the
gradient operator. Carrying out the dierentiation, we nd that
1 f1 f1 e1 1 f1 f1 e1
L= e1 e1 + e1 + e2 e1 + e2
h1 1 h1 1 h2 2 h2 2
1 f1 f1 e1
+ e 3 e1 + e3 + , (A.8.27)
h3 3 h3 3
where the three dots represent similar terms involving f2 and f3 . Using (A.8.17), (A.8.19), and
(A.8.20) to evaluate the derivative of the unit vectors, we obtain
1 f1 f1 h1 f1 h1 1 f1
L= e1 e1 e 1 e2 e1 e3 + e2 e1
h1 1 h1 h2 2 h1 h3 3 h2 2
f1 h2 1 f1 f1 h3
+ e 2 e2 + e3 e1 + e3 e3 + , (A.8.28)
h1 h2 1 h3 3 h1 h3 1
where the three dots represent similar terms involving f2 and f3 . Collecting similar terms, we derive
the dyadic expansion
L = Lij ei ej , (A.8.29)
where Lij are the components of L in the chosen curvilinear coordinates, and summation is implied
over the repeated indices, i and j. The calculations yield the diagonal components
1 f1 f2 h1 f3 h2
1 f1 h2 f2 f3 h1
L11 = + + , L22 = + + ,
h1 1 h2 2 h3 3 h2 h1 1 2 h3 3
(A.8.30)
1 f1 h3 f2 h3 f3
L33 = + + , (A.8.31)
h3 h1 1 h2 2 3
and the o-diagonal components
1 f2 f1 h1
1 f1 f2 h2
L12 = , L21 = ,
h1 1 h2 2 h2 2 h1 1
1 f3 f1 h1
1 f1 f3 h3
L23 = , L32 = .
h2 2 h3 3 h3 3 h2 2
The trace of L is the divergence of f , whereas the curl of f is related to the skew-symmetric part of
L, as discussed later in this section.
A.8 Orthogonal curvilinear coordinates 1031
f e1 + e2 + e3 (f1 e1 + f2 e2 + f3 e3 ). (A.8.33)
h1 1 h2 2 h3 3
Since the base unit vectors are orthonormal, the trace of the matrix ei ej is zero if i = j, or unity
if i = j. Using (A.8.27) and recalling that, because ei is a unit vector, ei ei /j = 0, where
summation is not implied over the repeated index, i, we nd that
1 f1 1 e1 1 e1
f = + f1 e2 + e3 + , (A.8.34)
h1 1 h2 2 h3 3
which also follows from (A.8.33). The three dots on the left-hand side represent similar terms
involving f2 and f3 . A slight rearrangement yields
1 f1 1 (h1 e1 ) 1 (h1 e1 )
f = + f1 e2 + e3 + . (A.8.35)
h1 1 h1 h2 2 h1 h3 3
Now using the relations (A.8.11) and similar relations and simplifying, we obtain
1
f = (h2 h3 f1 ) + (h3 h1 f2 ) + (h1 h2 f3 ) . (A.8.36)
h1 h2 h3 1 2 3
We recall that the products hi hj for i = j are surface metric coecients corresponding to coordinate
surfaces.
f = e1 + e2 + e3 (f1 e1 + f2 e2 + f3 e3 ). (A.8.38)
h1 1 h2 2 h3 3
Expanding the derivatives, we nd that
1 e1 e2 f2 e3 f3
= f1 e 1 + f2 e1 + e3 + f3 e1 e2
h1 1 1 1 1 1
1 e1 f1 e2 e3 f3
+ f1 e 2 e3 + f2 e2 + f3 e2 + e1 (A.8.39)
h2 2 2 2 2 2
1 e1 f1 e2 f2 e3
+ f1 e3 + e2 + f2 e3 e1 + f3 e3 .
h3 3 3 3 3 3
1032 Introduction to Theoretical and Computational Fluid Dynamics
1 = f1 e3 + f2 e3 + f3 e3 +
h2 2 2 2 2
1 e1 e2 f2 e3
f1 e2 + f2 e2 + + f3 e2 . (A.8.41)
h3 3 3 3 3
Using relations (A.8.3) and (A.8.15), we obtain
1 e2 f3
1 f2 e3
1 = f2 e3 + + f3 e2 , (A.8.42)
h2 2 2 h3 3 3
which can be rearranged into
1 f3 e3 f2 e2
1 = h3 f3 h2 e2 h2 + f2 h3 e3 . (A.8.43)
h2 h3 2 3 3 2
Using relations (A.8.20), we obtain
1 f3 h3 f2 h2
1 (h3 f3 ) (h2 f2 )
1 = h3 + f3 h2 f2 = . (A.8.44)
h2 h3 2 2 3 3 h2 h3 2 3
Similar relations can be written for 2 and 3 . Compiling these results, we nd that the curl is
given by the determinant of a matrix,
1 h1 e1 h2 e2 h3 e3
f = det /1 /2 /3 . (A.8.45)
h1 h2 h3
h 1 f1 h 2 f2 h3 f3
The determinant is computed according to the usual rules of the outer vector product, treating the
partial derivative operator triplet as a regular vector.
Problems
x ex
e
x
0
Figure A.9.1 Cylindrical polar coordinates, (x, , ), dened with respect to companion Cartesian
coordinates, (x, y, z).
x is the projection of the position vector onto the straight (rectilinear) x axis passing through
a designated origin, taking values in the range (, +).
is the distance of a point of interest from the x axis, taking values in the range [0, ).
is the azimuthal angle measured around the x axis, taking values in the range [0, 2).
The value = 0 corresponds to the rst and second quadrants, and the value = corresponds to
the third and fourth quadrants of the xy plane. An x half-plane of constant azimuthal angle is
called an azimuthal plane. In astronomy, land, and sea navigation, the yz plane lies in the horizon
and the y axis points to the north. A star may be located at a point x in the sky.
Using elementary trigonometry, we derive relations between the Cartesian and associated polar
cylindrical coordinates,
and the inverse relations between the polar cylindrical and Cartesian coordinates,
y
= y2 + z 2 , = arccos . (A.9.2)
In computing the inverse cosine function, arccos, care must be taken so that is a continuous
function of y and .
1034 Introduction to Theoretical and Computational Fluid Dynamics
Unit vectors
We consider a point in space and dene three vectors of unit length, denoted by ex , e , and e ,
pointing, respectively, in the direction of the x axis, normal to the x axis, and in the direction of
the azimuthal angle , as depicted in Figure A.9.1. Note that the orientation of the unit vectors
e and e changes with position in space, whereas the orientation of ex is xed and independent of
position in space. The position vector is
x = x ex + e . (A.9.3)
The dependence of the position vector on the azimuthal angle, , is mediated through the unit
vector e on the right-hand side. The absence of e from the right-hand side of (A.9.3) is justied
by observing that the distance from the origin, expressed by the position vector x, is perpendicular
to the third unit vector, e .
Vector components
Any vector, v, can be resolved as
v = vx ex + v e + v e , (A.9.4)
e de
= e , = e , (A.9.7)
d
where Greek variables stand for x, , and .
Metric coecients
The cylindrical polar coordinates, (x, , ), comprise a set of orthogonal curvilinear coordinates
(1 , 2 , 3 ), as discussed in Section A.8. The associated metric coecients are
hx = 1, h = 1, h = . (A.9.8)
A.10 Spherical polar coordinates 1035
Expressions (A.9.7) arise by substituting these metric coecients into the general relations (A.8.17),
(A.8.19), and (A.8.20).
Dierential operators
The gradient a scalar function, f (x), is
f f 1 f
f = ex + e + e . (A.9.9)
x
The Laplacian of a scalar function, f (x), is
2f 1 f
1 2f
2 f = 2
+ + 2 . (A.9.10)
x 2
The divergence of a vector function, f (x), is
fx 1 (f ) 1 f
f = + + . (A.9.11)
x
The curl of a vector eld, f (x), is
1 (f ) f
1 f f
f fx
x
f = ex + e + e . (A.9.12)
x x
The Laplacian of a vector function, f (x), is
f 2 f
f 2 f
2 f = ex 2 fx + e 2 f 2 2 + e 2 f 2 + 2 . (A.9.13)
The derivative of a vector function, f (x), with respect to arc length, l, measured in the
direction of a unit vector, t, is
f t f
t f
t f = = ex (t fx ) + e t f + e t f + , (A.9.14)
l
where the cylindrical polar coordinates of the gradient of the individual components, f , are
computed from (A.9.9).
Problem
e
er
r x
e x
0
Figure A.10.1 Spherical polar coordinates, (r, , ), dened with respect to companion Cartesian
coordinates, (x, y, z), and cylindrical polar coordinates, (x, , ).
r is the distance from the designated origin taking values in the range [0, ).
is the meridional angle subtended between the x axis, the origin, and the chosen point, taking
values in the range [0, ].
is the azimuthal angle measured around the x axis, taking values in the range [0, 2).
The value = 0 corresponds to the rst and second quadrants, and the value = corresponds to
the third and fourth quadrants of the xy plane. An r half-plane of constant azimuthal angle is
called an azimuthal plane. In astronomy, land and sea navigation, the yz plane is the horizon and
the y axis points to the north. A star may be located at the point x in the sky.
Using elementary trigonometry, we derive relations between the Cartesian, cylindrical, and
spherical polar coordinates,
and
Unit vectors
We consider a point in space and dene three vectors of unit length, denoted by er , e , and e ,
pointing, respectively, in the radial, meridional, and azimuthal direction, as illustrated in Figure
A.10.1. Note that the orientation of all three unit vectors changes with position in space. In
contrast, the orientation of the Cartesian unit vectors, ex , ey , and ez , is xed. The position vector
is proportional to the radial unit vector,
x = r er . (A.10.4)
The dependence on and is mediated through the unit vector er . The absence of e and e from
the right-hand side of (A.10.4) is explained by observing that the distance from the origin, expressed
by the position vector x, is perpendicular to the unit vectors e and e .
Vector components
A vector, v, can be resolved in the component form
v = vr er + v e + v e , (A.10.5)
der der de de
= e , = sin e , = er , = cos e ,
d d d d
de
= sin er cos e , (A.10.8)
d
where Greek variables stand for r, , and .
1038 Introduction to Theoretical and Computational Fluid Dynamics
Metric coecients
The spherical polar coordinates, (r, , ), comprise a set of orthogonal curvilinear coordinates iden-
tied with the triplet (1 , 2 , 3 ), as discussed in Section A.8. The associated metric coecients are
given by
hr = 1, h = r, h = r sin . (A.10.9)
Expressions (A.10.8) arise by substituting these metric coecients into the general relations (A.8.17),
(A.8.19), and (A.8.20).
Dierential operators
The gradient a scalar function, f , is given by
f 1 f 1 f
f = er + e + e . (A.10.10)
r r r sin
The Laplacian of a scalar function, f , is given by
1 2 f
1 f
1 2f
2 f = 2 r + 2 sin + 2 2 . (A.10.11)
r r r r sin r sin 2
The divergence of a vector function, f , is given by
1 (r2 fr ) 1 (sin f ) f
f = 2 + + , (A.10.12)
r r r sin
yielding
fr ur 1 f f 1 f
f = +2 + + cot + . (A.10.13)
r r r r r sin
The curl of a vector eld, f , is
1 (sin f ) f
1 1 fr (rf )
f = er + e
r sin r sin r
(A.10.14)
1 (rf ) fr
+e .
r r
The Laplacian of a vector function, f , is given by
2 f = qr er + q e + q e , (A.10.15)
where
fr 2 f 2 2 f
qr = 2 fr 2 2 2 f cot 2 ,
r2 r r r sin
2 fr f 2 cot f
q = 2 f + 2 2 2 , (A.10.16)
r 2 r sin r sin
1 fr f
q = 2 f + 2 f + 2 + 2 cos ,
r2
sin
and the Laplacian of the individual vector components is computed using (A.10.11).
A.11 Plane polar coordinates 1039
e
y er
x
0
Figure A.11.1 Plane polar coordinates, (r, ), in the xy plane dened with respect to companion
Cartesian coordinates, (x, y).
The derivative of a vector function, f , with respect to arc length, l, measured in the direction
of a unit vector, t, is given by
f t f + t f t fr cot
t f = = er t fr + e t f + t f
l r r r (A.10.17)
t fr t f
+e t f + + cot .
r r
The gradients of the individual spherical polar components, fr , f , and f , are computed
according to (A.10.10).
Problem
Unit vectors
We consider a point in the xy plane and dene two vectors of unit length, denoted by er and e ,
pointing in the radial or polar direction, as depicted in Figure A.11.1. Note that the orientation of
these unit vectors changes with position in the xy plane, whereas the orientation of the Cartesian
unit vectors ex and ey is xed. The position vector is proportional to the radial unit vector,
x = r er . (A.11.3)
The dependence of the position vector on the polar angle, , is mediated through the unit vector er .
Vector components
A vector in the xy plane, v, can be expressed in the form
v = vr er + v e , (A.11.4)
where the coecients vr and v are the plane polar components of v.
Metric coecients
The plane polar coordinates, (r, ), comprise a set of orthogonal curvilinear coordinates (1 , 2 ).
The associated metric coecients are
hr = 1, h = r. (A.11.8)
A.11 Plane polar coordinates 1041
Expressions (A.11.7) arise by substituting these metric coecients into the general relations (A.8.17),
(A.8.19), and (A.8.20).
Dierential operators
The dierential vector operators in plane polar coordinates, (r, ), derive from those in cylindrical
polar coordinates by discarding the dependence on x and renaming as r and as . The gradient
of a scalar function, f , is given by
f 1 f
f = er + e . (A.11.9)
r r
1 f
1 2f
2 f = r + 2 . (A.11.10)
r r r r 2
1 (rfr ) 1 f
f = + . (A.11.11)
r r r
1 (rf ) fr
f = ez , (A.11.12)
r r
where ez is the unit vector normal to the xy plane. The Laplacian of a vector function, f , is
fr 2 f
f 2 fr
2 f = er 2 fr 2 2 + e 2 f 2 + 2 . (A.11.13)
r r r r
The derivative of a vector function, f , with respect to arc length, l, measured in the direction
of a unit vector, t, is given by
f t f
t fr
t f = = er t fr + e t f + , (A.11.14)
l r r
where the plane polar coordinates of the gradient of the individual components, fr and f , are
computed from (A.11.9).
Problem
2
g x
g2 2
dx
g
x 1
g1
x1
x
x
1
Figure A.12.1 Nonorthogonal curvilinear coordinates in a plane; (g1 , g2 ) are the covariant base vec-
tors, and (g1 , g2 ) are the contravariant base vectors.
Base vectors
We begin the discussion of nonorthogonal coordinates by introducing two continuous intersecting
families of generally curvilinear axes in the xy plane, (x1 , x2 ), as shown in Figure A.12.1. The
position vector, x, is regarded as a function of x1 and x2 , signied by writing x(x1 , x2 ). Next, we
introduce a corresponding pair of tangent base vectors,
x x
g1 = , g2 = , (A.12.1)
x1 x2
where the derivative with respect to x1 is taken keeping x2 constant, and vice versa. The two base
vectors, g1 and g2 , are unit vectors only if the coordinates x1 and x2 measure physical arc length.
The theory of dierential calculus allows us to express an innitesimal vector in the xy plane at a
point, x, as
In general, equation (A.12.2) cannot be integrated readily to produce an explicit expression for the
position vector, x(x1 , x2 ).
The ordered doublet (x1 , x2 ) constitutes a pair of nonorthogonal curvilinear coordinates with
associated base vectors (g1 , g2 ). If the lines of constant x1 and x2 are straight, we obtain rectilinear
coordinates. If the angle subtended between the two vectors g1 and g2 is equal to 90 at every point
in the plane, we obtain orthogonal curvilinear or rectilinear coordinates.
A.12 Nonorthogonal coordinates 1043
Table A.12.1 Denitions of covariant and contravariant coordinates, base vectors, and vector com-
ponents in two dimensions. Analogous denitions are made in three dimensions.
Metric tensor
Using the denitions of the covariant and contravariant metric tensor, we nd that
dx dx = gij dxi dxj = g ij dxi dxj = dxi dxi , (A.12.16)
where summation is implied over the repeated indices, i and j. Combining (A.12.5) with (A.12.14)
and juxtaposing the two coordinates, we obtain
gij dxj = dxi , g ij dxj = dxi , (A.12.17)
where summation is implied over the repeated index, j. These expressions demonstrate that the
covariant metric tensor is the inverse of the contravariant metric tensor, and vice versa,
[gij ] = [g ij ]1 , [g ij ] = [gij ]1 , (A.12.18)
where the superscript 1 denotes the matrix inverse.
A.12 Nonorthogonal coordinates 1045
Because of the biorthonormality of the covariant and contravariant base vectors, the trace of
the matrices gi gj and gj gj satisfy the equations
trace(gi gj ) = ij , trace(gi gj ) = ij , (A.12.19)
where ij is Kroneckers delta.
Three-dimensional coordinates
All notions and equations discussed previously in this section in two dimensions can be extended
in a straightforward fashion to three dimensions. A system of three-dimensional nonorthogonal
coordinates is shown in Figure A.12.2. Subscripts and superscripts now vary from 1 to 3, and the
size of the metric tensor is 3 3. Given the covariant base vectors, the contravariant base vectors
can be computed using the formula
1
gi = ijk gj gk , gj gk = J jki gi , (A.12.21)
J
where ijk is the alternating tensor, and summation is implied over the repeated indices j and k.
We have introduced the volumetric Jacobian associated with the covariant base,
J g1 (g2 g3 ) = g, (A.12.22)
where
1
g det(gij ) = . (A.12.23)
det(g ij )
1046 Introduction to Theoretical and Computational Fluid Dynamics
g1 = J g2 g3 , g2 = J g3 g1 , g3 = J g 1 g 2 . (A.12.26)
Problems
a = a1 g1 + a2 g2 + a3 g3 = a1 g1 + a2 g2 + a3 g3 . (A.13.1)
We may write
a = ai g i = a i g i , (A.13.2)
where summation is implied over the repeated index, i. Using the biorthonormality of the base
vectors, we nd that
a1 = a g1 , a2 = a g 2 , a3 = a g 2 ,
(A.13.3)
a1 = a g1 , a2 = a g2 , a3 = a g3 .
Taking the inner product of (A.13.1) with each of the covariant vectors, g1 , g2 , and g2 , we
nd that
Taking the inner product of (A.13.1) with each one of the contravariant vectors, g1 , g2 , and g2 , we
nd that
a1 = a1 g 11 + a2 g 21 + a3 g 31 , a2 = a1 g 12 + a2 g 22 + a3 g 32
a3 = a1 g 13 + a2 g 23 + a3 g 33 . (A.13.5)
We may then write
ai = aj gji , ai = aj g ji , (A.13.6)
where summation is implied over the repeated index, j. Expressions (A.13.6) are used to raise or
lower the indices.
We thus write
f f f f 1 f f
f = g1 + g2 + g3 , f = 1
g + 2 g2 + 3 g3 , (A.13.13)
x1 x2 x3 x x x
and observe that the derivatives with respect to contravariant coordinates, xi , produce covariant
components. To signify this rule, sometimes the gradient is regrettably called a covariant vector.
Convective derivative
Using the biorthonormality relation between the covariant and contravariant base vectors, we nd
that the convective derivative in two dimensions can be computed either from the expression
f f f f
u f = (u1 g1 + u2 g2 ) (g1 + g 2 2 ) = u1 1 + u2 2 , (A.13.14)
x1 x x x
or from the expression
f f f f
u f = (u1 g1 + u2 g2 ) (g1 + g2 ) = u1 + u2 , (A.13.15)
x1 x2 x1 x2
where u represents the velocity. Similar expressions can be written in three dimensions, yielding
f f
u f = ui = ui , (A.13.16)
xi xi
where summation is implied over the repeated index, i.
Gradient operator
Motivated by (A.13.13), we introduce the gradient operator,
= gk = gk , (A.13.17)
xk xk
where summation is implied over the repeated index, k. Explicitly in three dimensions,
= g1 + g2 + g3 = g 1 1 + g2 + g3 . (A.13.18)
x1 x2 x3 x x2 x3
The base vectors and derivative operators may not be transposed.
Numerical computation
As an application, we consider the computation of the gradient of a scalar function, f , at a point
in a plane. We are given the values of the function at nodes distributed along two rectilinear axes
passing through that point, as shown in Figure A.13.1. For convenience, we denote x1 = and
x2 = . A suitable algorithm involves the following steps:
1. Compute the covariant base vectors by nite-dierence approximations
x xR xL x xT xB
g , g , (A.13.19)
R L T B
where L, R, T, B stand for left, right, top, and bottom nodes relative to the intersection.
A.14 Christoel symbols of the second kind 1049
Figure A.13.1 Illustration of two curvilinear coordinates supporting data passing through the central
point.
2. Compute the corresponding contravariant base vectors, g , g , using the formulas discussed
in this section.
3. Compute the covariant derivatives by nite-dierence approximations,
f fR fL f fT fB
, . (A.13.20)
R L T B
Problem
Using the biorthonormality of the base vectors once more, we nd that (gi gk )/xj = 0. Expanding
the derivative, we obtain
gk gi
gi = j gk = kij , (A.14.3)
xj x
which shows that
gk
= kij gi . (A.14.4)
xj
gi x gj
kij = gk = i j g k = gk , (A.14.5)
xj x xi
which reveals the symmetry property
gmi (gm gi ) gi gm
j
= j
= gm j
+ gi = kij gm gk + kmj gi gk , (A.14.7)
x x x xj
yielding
gmi
= kij gmk + kmj gik , (A.14.8)
xj
where summation is implied over the repeated index, k. Multiplying the last equation by g mp , where
p is a free index, summing over m, and noting that gmk g mp = kp , we obtain
gmi
g mp = pij + kmj gik g mp . (A.14.9)
xj
Switching the indices i and j, we obtain
gmj
g mp = pij + kmi gjk g mp . (A.14.10)
xi
Combining the last three equations, we nd that the Christoel symbols of the second kind can be
computed in terms of the metric tensor from the expression
1 km gmi gmj gij
kij = g + , (A.14.11)
2 xj xi xm
where summation is implied over the repeated index, m.
A.14 Christoel symbols of the second kind 1051
M (f )T , (A.14.12)
is a matrix dened such that its projection from the right on gi or gi provides us with the rate of
change of f in the respective directions,
f f
M gi = , M gi = . (A.14.13)
xi xi
By analogy with (A.13.13) for the gradient of a scalar, we introduce the covariant and contravariant
expansions
f f f
M= g1 + g2 + g3 ,
x1 x2 x3
(A.14.14)
f f f
M= g1 + g2 + g3 .
x1 x2 x3
Expressing the vector eld f in terms of the covariant or contravariant base vectors and
carrying out the dierentiations, we obtain four combinations. The rst combination is
(f i gi ) f i
M= g j
= + i
kj f k
gi gj . (A.14.15)
xj xj
The second combination is
(fi gi ) f
i
M= g j
= ij k g g .
k
f i j (A.14.16)
xj xj
The third combination is
(f i gi ) f i gi
M= gj = gi gj + f i gj . (A.14.17)
xj xj xj
Summation over the repeated indices i and j is implied in all four combinations. In solid mechanics,
continuum mechanics, and applied mathematics, the right gradient M = (f )T is simply called the
gradient of the vector eld f .
In uid mechanics, the gradient of the vector eld f is identied with the left gradient,
L = f = MT . (A.14.19)
1052 Introduction to Theoretical and Computational Fluid Dynamics
L= + i
kj f k
gj gi f,ji gj gi , (A.14.20)
xj
where f,ji is the rst covariant derivative. Similar representations can be derived from expressions
(A.14.16)(A.14.18).
As an example, using (A.14.20), we nd that the convective derivative of the velocity on the
left-hand side on the equation of motion is
ui
i
j u
u u = (um gm ) + i
kj u k
g j
g i = u + i
kj u k
gi = uj ui,j gi . (A.14.21)
xj xj
The term uj ui,j is the ith covariant component of the convective derivative.
f i f
i
f = trace(M) = + i
ki f k
= k
ij fk g ij . (A.14.22)
xi xj
Making substitutions, we obtain
f i 1 gim k
f = i
+ g mi f . (A.14.23)
x 2 xk
Further manipulation gives
1 i
f = f g , (A.14.24)
g xi
where g = det(gij ).
2 f = g g , (A.14.27)
g xi xk
where g = det(gij ).
f1 2 1 f1 3 1 1
1 g1 2 g1 3 g1
= (g g ) + (g g ) + f g + g + g + , (A.14.30)
x2 x3 x1 x2 x3
where the dots indicate similar terms involving f 2 and f 3 . Using (A.12.25) and (A.14.4), we obtain
1 f1 f1
= 2 g3 + g2 f 1 1i1 g1 + 1i2 g2 + 1i3 g3 gi + , (A.14.31)
J x x 3
where summation is implied over the repeated index, i. Using (A.12.25), we nd that the term
enclosed by the square brackets on the right-hand side of (A.14.31) is identically zero, yielding the
simple expression
1 fk
= ijk gk . (A.14.32)
J xj
f i
f i
g j gi
2 f = (gm gj ) m j
+ ikj f k gi + j
+ ikj f k gm . (A.14.34)
x x x xm
1054 Introduction to Theoretical and Computational Fluid Dynamics
ij contravariant components
ij covariant components
ij mixed right-covariant components
i j mixed left-covariant components
Table A.15.1 Possible components of a two-index tensor in contravariant, covariant, and mixed forms.
2 f = g kj f,jk
i
gi , (A.14.35)
where
2f i i f
l
i f
l
l f
i ijl l
i
f,jk = + jl + lk jk + jl jk ml f
+ imk m m i (A.14.36)
xk xj xk xj xl xk
is the covariant second derivative. This expression can be used to evaluate the viscous force on the
right-hand side of the NavierStokes equation (e.g., [417]).
Problem
= ij gi gj , = i j g i gj ,
(A.15.1)
= ij gi gj , = ij gi gj .
The scalar coecients representing the pure contravariant, the pure covariant, and the mixed com-
ponents are named as shown in Table A.15.1.
Repeating our previous analysis for vectors, we nd that the various tensor components derive
from one another by the relations
where summation is implied over the repeated indices, k and l. Cursory inspection reveals obvious
rules for raising and lowering the indices.
A.16 Rectilinear coordinates with constant base vectors 1055
ij ik j
k g
i
i g
j
= gk ij g i
g j
= g g + ij (g ) g j
+ (g k
g ) . (A.15.3)
xk xk xk xk
Computing the derivatives of the base vectors using (A.14.4), we obtain
ij ik j
= k
g g ij g mk imk gj + g ik jmk gm , (A.15.4)
x
which can be recast into the form
ij
= m
ik mj m
jk im g ik gj . (A.15.5)
xk
Working in a similar fashion, we nd the alternative forms
ij
= + iim mj + jim im gj ,
xi
i
j
= + i
im j m
m i
ij m g ,
j (A.15.6)
xi
j
j
= i
m
ik m
j
+ km i
m
g ik gj .
xk
The right-hand sides of the last four equations are vectors expressed in covariant or contravariant
form.
Problem
where the square brackets enclose the Cartesian coordinates, as illustrated in Figure A.16.1. The
covariant metric tensor is
! "
1 1
gij = . (A.16.2)
1 2
1056 Introduction to Theoretical and Computational Fluid Dynamics
x2 2
x
x
g
2
g2
g
1 x1
g1
x1
We can easily conrm that g ij is the inverse of gij and vice versa.
The traces of g1 g1 and g2 g2 are unity, whereas the traces of g1 g2 and g2 g1 are zero.
The covariant, contravariant, and Cartesian axes share a common origin. The contravariant
and covariant coordinates of the position vector with Cartesian coordinates x = [x, y] are
x1 = x g1 = x y, x2 = x g2 = y, x1 = x g1 = x, x2 = x g2 = x + y. (A.16.6)
Conversely,
x = x1 + x2 = x1 , y = x2 = x1 + x2 . (A.16.7)
If we know that contravariant or covariant coordinates, we can easily compute the Cartesian coor-
dinates.
A.16 Rectilinear coordinates with constant base vectors 1057
Gradient of a scalar
Now we consider a scalar eld,
f (x, y) = 2x + y = 2x1 + 3x2 = x1 + x2 , (A.16.8)
where x and y are Cartesian coordinates. In Cartesian coordinates, the gradient of this eld is
f = [2, 1]. Straightforward dierentiation yields
f f f f
= 2, = 3, = 1, = 1, (A.16.9)
x1 x2 x1 x2
which is consistent with (A.13.13).
Vector eld
Next, we consider a vector eld,
! "
2x + y
u= = (x + 3y) g1 + (x 2y) g2 = (2x + y) g1 + (3x y) g2 . (A.16.10)
x 2y
The contravariant and covariant components are
u1 = x + 3y = x1 + 4x2 , u2 = x 2y = x1 x2 ,
u1 = 2x + y = x1 + x2 , u2 = 3x y = 4x1 x2 . (A.16.11)
Using (A.14.22), we nd that u = 0.
Problems
y
^
^
x
x
^
Figure A.17.1 Illustration of helical coordinates, (x, , ), in relation to companion Cartesian and
cylindrical polar coordinates, (x, , ).
= , = + x, x = x, (A.17.1)
where 2/L is the helical wave number and L is the helical pitch. In the limit of innite
pitch, 0, the helical coordinates reduce to cylindrical polar coordinates. In problems with
helical symmetry, the partial derivative of a variable of interest, f , with respect to x is zero, so that
f (, ).
A vector eld, u, can be resolved into physical components corresponding to polar cylindrical
and associated helical coordinates,
u = u e + u e + ux ex = u e + u e + ux ex , (A.17.7)
where
1 1 1 1 1
e = g1 = g1 , e = g2 = g2 , ex = g3 = g3 (A.17.8)
g11 g22 g33 1 + 2 2
are position-dependent unit vectors. Projecting (A.17.7) onto e , e , or ex , we nd that
1
u = u , u = u + ux = u + ux , ux = ux . (A.17.9)
1 + 2 2 1 + 2 2
If the vector eld u is helically symmetric, the cylindrical polar as well as the helical components of
u are independent of x, and only depend on and .
The contravariant components of u corresponding to the helical coordinates, dened such that
u = u1 g1 + u2 g2 + u3 g3 , are given by
u u u ux ux
u1 = = u , u2 = = , u3 = = = ux . (A.17.10)
g11 g22 g33 1 + 2 2
In the case of helical symmetry, the contravariant components are independent of x, and only depend
on and .
Problem
bi
xi = (B.1.1)
Ai,i
for i = 1, . . . , N , where summation is not implied over the repeated index, i. For clarity, a comma
was inserted between the two matrix indices.
When the matrix A is lower triangular, we use the forward substitution algorithm
b1 1
i1
x1 = , xi = (bi Ai,j xj ) (B.1.2)
A1,1 Ai,i j=1
for i = 2, . . . , N .
When the matrix A is upper triangular, we use the backward substitution algorithm
bN 1 N
xN = , xi = (bi Ai,j xj ) (B.1.3)
AN,N Ai,i j=i+1
1060
B.1 Linear algebraic equations 1061
Gauss elimination is the simplest and most popular direct method. The basic idea is to solve
the rst equation of the given system, A x = b, for the rst unknown, x1 , and use the expression
thus obtained to eliminate x1 from all subsequent equations. We then retain the rst equation as is,
and replace all subsequent equations with their descendants that do not contain x1 . At the second
stage, we solve the second equation for the second unknown, x2 , and use the expression thus obtained
to eliminate x2 from all subsequent equations. We then retain the rst and second equations, and
replace all subsequent equations with their descendants that do not contain x1 or x2 . Continuing in
this fashion, we arrive at the last equation, which contains only the last unknown, xN .
Having completed the elimination, we compute the unknowns by the method of backward
substitution. First, we solve the last equation for xN , which thus becomes a known. Second, we
solve the penultimate equation for xN 1 , which also becomes a known. Continuing in the backward
direction, we scan the reduced system until we have computed all unknowns.
Pivoting
Immediately before the mth equation has been solved for the mth unknown in the process of elimi-
nation, where m = 1, 2, . . . , N 1, the linear system takes the form
(m) (m) (m)
A1,1 A1,2 A1,N b1
(m)
(m) (m) x1
0 A2,2 A2,N b(m)
x2 2
0 0 (m)
x3 b3
0 Am1,m1 Am1,m Am1,N
(m) (m) (m)
0
.. = .
.
,
(B.1.4)
(m) (m) . .
0 0 Am,m Am,N (m)
xN 1
bN 1
0 0
(m) (m)
xN (m)
0 0 AN,m AN,N bN
(m) (m)
where Ai,j are intermediate coecients and bi are intermediate right-hand sides. The rst equa-
tion of (B.1.4) is identical to the rst equation of the given linear system for any value of m.
Subsequent equations are dierent, except at the rst step corresponding to m = 1.
(m)
A diculty arises when the diagonal element, Am,m , is nearly or precisely zero, since we are
unable to solve the mth equation for xm , as required. However, the failure of the method does
not imply that the system does not have a unique solution. To circumvent this diculty, we simply
rearrange the equations or relabel the unknowns so as to bring the mth unknown to the mth equation
using the method of pivoting. If there is no way we can make this happen, the matrix A is singular
and the linear system has either no solution or an innite number of solutions.
In the method of row pivoting, potential diculties are bypassed by switching the mth equation
in the system (B.1.4) with the subsequent kth equation, where k > m; the value of k is chosen so
1062 Introduction to Theoretical and Computational Fluid Dynamics
(m) (m)
that |Ak,m | is the maximum value of the elements in the mth column below the diagonal, Ai,m , for
(m)
i m. If Ai,m= 0 for all i m, the matrix A is singular and the system under consideration does
not have a unique solution.
Setting up
Formulate the N (N + 1) partitioned augmented matrix
C(1) A b , (B.1.5)
1. Find the location of the element with the maximum norm in the rst column of C(1) .
(1)
This is done by searching for the maximum norm of the elements |Ci,1 |, for i = 1, . . . , N .
(1)
Assume that this is equal to |Ck,1 |, corresponding to the kth row.
2. Skip this step if k = 1. Otherwise, switch the rst row with the kth row of C(1) ; repeat
for the matrix L.
(1) (1)
3. Compute the rst column of L below the diagonal by setting Li,1 = Ci,1 /C1,1 , for i =
2, . . . , N .
4. Subtract from the ith row of C(1) the rst row multiplied by Li,1 , for i = 2, . . . , N , to
obtain a new augmented matrix,
C(2) A(2) |b(2) . (B.1.6)
Second pass
1. Find the location of the element with the maximum norm in the second column of C(2) ,
below the diagonal. This is done by searching for the maximum norm of the elements
(2) (2)
|Ci,2 |, for i = 2, . . . , N . Assume that this is equal to |Ck,2 |, corresponding to the kth row.
2. Skip this step if k = 2. Otherwise, switch the second row with the kth row of C(2) ; repeat
for the matrix L.
(2) (2)
3. Compute the second column of L below the diagonal, by setting Li,2 = Ci,2 /C2,2 , for
i = 3, . . . , N .
4. Subtract from the ith row of C(2) the second row multiplied by Li,2 , for i = 3, . . . , N , to
obtain a new augmented matrix,
C(3) A(3) |b(3) . (B.1.7)
...
B.1 Linear algebraic equations 1063
mth pass
1. Find the location of the element with the maximum norm in the mth column of C(m) ,
below the diagonal. This is done by searching for the maximum norm of the elements
(m) (m)
|Ci,m |, for i = m, . . . , N . Assume that this is equal to |Ck,m |, corresponding to the kth
row.
2. Skip this step if k = m. Otherwise, switch the mth row with the kth row of C(m) ; repeat
for the matrix L.
(m) (m)
3. Compute the mth column of L below the diagonal, by setting Li,m = Ci,m /Cm,m , for
i = m + 1, . . . , N .
4. Subtract from the ith row of C(m) the mth row multiplied by Li,m , for i = m + 1, . . . , N ,
to obtain a new augmented matrix,
C(m+1) A(m+1) |b(m+1) . (B.1.8)
...
(N 1) pass
At the end of the N 1 pass, corresponding to m = N 1, the augmented matrix, C(N ) , has
the form
C(N ) = A(N ) b(N ) , (B.1.9)
U x = b(N ) (B.1.10)
to extract the solution of the original system of equations, A x = b. This is done by solving
the last equation in (B.1.10) for the last unknown, xN ; once this is available, we solve the
penultimate equation for xN 1 ; continuing backward in this fashion, we nally compute x1 .
Complete the matrix L (optional)
Set the diagonal elements of the matrix L equal to 1.
LU decomposition
It can be shown by straightforward algebraic manipulations that the matrices L and U provide us
with the LU decomposition of the matrix A. Specically,
L U = Amod , (B.1.11)
where the matrix Amod is identical to A, except that the rows may have been switched due to
pivoting. If pivoting is disabled, Amod = A.
1064 Introduction to Theoretical and Computational Fluid Dynamics
Determinant
The determinant of the matrix A follows from the relation
where the plus sign applies when an even number of row interchanges have been made due to
pivoting, and the minus sign otherwise. In the absence of pivoting, the plus sign is chosen.
Other algorithms for performing the LU decomposition of a matrix directly without reference
to a particular system of linear equations are available. Examples are Crouts and Doolittles algo-
rithms [18, 318]. Once the L and U matrices have been obtained, the solution of the linear system
A x = b is found in two stages: rst, we solve the lower triangular system L y = b for y by
forward substitution; second, we solve the upper triangular system U x = y for x by backward
substitution.
D x = s, (B.1.13)
The Thomas algorithm proceeds in two stages. At the rst stage, the tridiagonal system (B.1.13) is
transformed into a bidiagonal system,
D x = y, (B.1.15)
At the second stage, the bidiagonal system (B.1.15) is solved by backward substitution. The al-
gorithm involves solving the last equation for the last unknown, xN , and then moving upward to
compute the rest of the unknowns in a sequential fashion. The combined algorithm is listed in Table
B.1.1.
B.1 Linear algebraic equations 1065
Reduction to bidiagonal:
! " ! "
d1 1 b1
=
y1 a1 s1
Do i = 1, 2, . . . , N 1
! " ! "
di+1 1 bi+1
=
yi+1 ai+1 ci+1 di si+1 ci+1 yi
End Do
Backward substitution:
xN = y N
Do i = N 1, N 2, . . . , 1 (step=1)
xi = yi di xi+1
End Do
Table B.1.1 The Thomas algorithm for solving a tridiagonal system of linear equations.
M x = N x + b, (B.1.17)
x = P x + c, (B.1.18)
where
P = M1 N, c = M1 b. (B.1.19)
The algorithm involves computing the sequence of vectors x(k) based on the equation
or on the equation
beginning with a certain initial vector, x(0) . If the spectral radius of the projection matrix P is less
than unity, that is, the magnitude of all eigenvalues of P is less than one, the sequence of vectors will
converge to the xed point of the mapping, denoted by X, that satises the equation X = P X + c
and therefore the given equation A X = b. In that case, the sequence x(k) will contain increasingly
accurate successive approximations to the solution.
Jacobis method
In Jacobis method, the iteration matrix, P, and constant vector, c, are constructed by solving the
individual scalar equations of the given system, A x = b, for the diagonal unknowns, obtaining the
iteration formula
1
N
(k+1) (k)
xi = bi Ai,j xj , (B.1.22)
Ai,i j=1
where the prime denotes that the term i = j is excluded from the sum. The matrix M dened in
(B.1.17) is the diagonal component of A, and the matrix N is the negative of A with the diagonal
elements set equal to zero. The components of the projection matrix are Pi,j = Ai,j /Ai,i if i = j,
Pi,j = 0 if i = j, and ci = bi /Ai,i , where summation is not implied over the repeated index, i.
A sucient but not necessary condition for the successive substitutions to converge is that
the matrix A is diagonally dominant,
N
|Ai,i | > |Ai,j | (B.1.23)
j=1
for any i, where summation is not implied the repeated index, i, and the prime denotes that the
term i = j is excluded from the sum.
GaussSeidel method
The GaussSeidel method is a variation of Jacobis method with the new feature that the components
of x(k+1) replace the corresponding components of x(k) as soon as the former are available. The
eective algorithm is
1
i1 N
(k+1) (k+1) (k)
xi = bi Ai,j xj Ai,j xj . (B.1.24)
Ai,i j=1 j=i+1
Let D be a diagonal matrix, L be a lower triangular, and U be an upper triangular matrix, con-
taining, respectively, the diagonal, lower triangular, and upper triangular components of A. The
GaussSeidel mapping takes the form
(D + L) x(k+1) = U x(k) + b, (B.1.25)
where
M = D + L, N = U. (B.1.26)
A sucient but not necessary condition for the successive substitutions to converge is that the matrix
A is symmetric and positive denite.
B.1 Linear algebraic equations 1067
Successive over-relaxation
The successive over-relaxation method (SOR) is constructed by restating the GaussSeidel iteration
algorithm in the residual correction form x(k+1) = x(k) + r(k) , where
i1 N
(k+1) (k) (k+1) (k)
xi = (1 ) xi + bi Ai,j xj Ai,j xj . (B.1.29)
Ai,i
j=1 j=i+1
When = 1, we obtain the GaussSeidel algorithm. A necessary condition for the successive
substitutions to converge is that 0 < < 2. Ideally, should take the value that minimizes the
spectral radius of the underlying projection matrix, P. The optimal value can be constructed during
the iterations.
r F = A x + b, (B.1.32)
To compute the travel distance, we note that, as we travel in the direction of r(0) , the vector x is
described by x = x(0) + r(0) , where is a scalar parameter. The value of the quadratic functional
along the travel path is a quadratic function of ,
1 (0)
F(x) = (x + r(0) ) A (x(0) + r(0) ) b (x(0) + r(0) ). (B.1.35)
2
We want to stop traveling when F has reached a minimum, F/ = 0. Taking the partial derivative
of the right-hand side of (B.1.35) with respect to and setting the resulting expression to zero, we
nd the optimal value
r(0) r(0)
1 = , (B.1.36)
r(0) (A r(0) )
p(k) p(k)
k = , (B.1.39)
p(k) (A p(k) )
until the minimum has been reached within a specied tolerance. While the method is guaranteed
to converge, the rate of convergence can be slow.
Directional search
In this method, the minimization problem is solved by selecting a set of search directions expressed
by the N -dimensional vectors p(1) , p(2) , p(3) , . . . , and then stepwise advancing an initial guess, x(0) ,
by setting
for k = 1, . . ., where k are appropriate coecients determining the length of each step. The evolved
solution at the end of the kth step is
k
x(k) = x(0) + i p(i) , (B.1.41)
i=1
B.2 Matrix eigenvalues 1069
for k > 0. Without loss of generality, we can begin the search from the origin, x(0) = 0. Our goal is
to compute the search directions so that the exact solution, X, is found exactly after N steps,
X = 1 p(1) + 2 p(2) + + N p(N ) , (B.1.42)
where N is the system size.
As we travel in the search direction p(k) , the vector x is described by x = x(k1) + k p(k) ,
and the quadratic functional along the path is given by
1 (k1)
F(x) = (x + k p(k) ) A (x(k1) + k p(k) ) b (x(k1) + k p(k) ), (B.1.43)
2
which is a quadratic function of k . Setting F/k = 0 to identify the optimal stopping point, we
nd that
p(k) r(k1)
k = , (B.1.44)
p(k) (A p(k) )
which is consistent with (B.1.39).
We have considerable freedom in selecting the search directions. In the steepest descend search
discussed earlier in this section, we have stipulated that p(k) = r(k1) . However, when the graph
of the functional exhibits narrow valleys, the search directions tend to align and successive points
bounce o opposite sides. Better methods for selecting the search directions are available.
In the method of conjugate gradients, the search vectors are computed so that they are A-
conjugate with each other, that is, p(i) A p(j) = 0 for i = j. The direction p(k) is aligned as much
as possible with p(k1) , subject to the A-conjugation constraint. The basic algorithm is shown in
Table B.1.1.
x(0) = 0
Do k = 1, . . . , N
If k = 1
p(1) = r(0) = b
Else If k > 1
r(k1) r(k1)
k = (k2) (k2)
r r
p(k) = r(k1) + k p(k1)
End if
r(k1) r(k1)
k = (k)
p (A p(k) )
x = x(k1) + k p(k)
(k)
Table B.1.1 A conjugate gradients algorithm for solving a linear system, A x = b, with a real,
symmetric, and positive denite-coecient matrix, A.
Gerschgorins theorem locates the eigenvalues of a general N N matrix, A, inside the union
of N disks in the complex plane. The ith disk is centered at the diagonal element, Ai,i , and the
corresponding radius is equal to the minimum of
N
N
|Ai,j |, |Ai,j |, (B.2.1)
i=1 j=1
where the prime denotes that the term i = j is excluded from the sum. Other methods for locating
eigenvalues are available [429]. Three general classes of methods for computing eigenvalues are
available.
This can be done using general-purpose numerical methods for solving nonlinear algebraic discussed
in Section B.3. To compute P (), we may perform the LU decomposition of the matrix A I for
a trial value of using, for example, the method of Gauss elimination discussed in Section B.2.1.2.
B.2 Matrix eigenvalues 1071
Choose x(0)
for k = 0, 1, ...
s = |x(k) |
x(k) 1s x(k)
x(k+1) = A x(k)
(k+1) = x(k) x(k+1)
end
If the matrix A is real and the dominant eigenvector is complex, a complex starting vector x(0) must
be provided.
Suppose that the eigenvalue of A with the maximum norm, 1 , is available. To obtain a
second eigenvalue, we may apply the power method to the singular matrix B = A 1 I. Note that
at least one eigenvalue of B is zero. Since the eigenvalues of B are shifted with respect to those
of A by 1 , the dominant eigenvalue has been moved to the origin and does not play a role in the
iterations. The power method produces the eigenvalue of B with the maximum norm, B . The
corresponding eigenvalue of A is A = B + 1 .
To compute the eigenvalue of A with the minimum norm, we apply the power method to the
inverse matrix, A1 , and thus obtain the corresponding eigenvalue with the maximum norm, A1 .
The eigenvalue of A with the minimum norm is A = 1/A1 . The success of this method hinges
on our ability to compute the inverse A1 with adequate precision.
To compute the eigenvalue of A that is closest to a specied complex number, c, we apply the
power method to the matrix B = (A c I)1 , and obtain the eigenvalue of B with the maximum
norm, B . The theory of eigenvalue shifting ensures that A = c + 1/B is the eigenvalue of A
closest to c. To avoid the demanding and possibly precarious computation of the inverse matrix,
1072 Introduction to Theoretical and Computational Fluid Dynamics
(A c I)1 , we carry out the iterations by selecting a starting vector, x(0) , and then solve a system
of linear equations,
In practice, we may perform the LU decomposition of the coecient matrix A c I and carry out
iterations by forward and backward substitution, as discussed in Section B.1. The method is reliable
even when c is close to an eigenvalue, rendering the matrix A c I nearly singular.
where w is a real vector with unit length, w w = 1. We want to compute the vector w so that the
elements in the rst column of the matrix
B=HAH (B.2.7)
are zero, except for the rst element, B1,1 , that is equal to 1 . Analysis shows that this is true when
1
1/2 1 (1)
(1)
w1 = (1 u1 ) , wi = u (B.2.8)
2 2w1 i
for i = 2, . . . , N (e.g., [317]). The eigenvalues of the bottom (N 1) (N 1) diagonal block of B
are also eigenvalues of A. Applying the power method to the reduced matrix B yields an additional
eigenvalue. The deation can be repeated until the sequentially deated matrix has been reduced
to a scalar.
Jacobis method
Jacobis method seeks to reduce the norm of the o-diagonal elements of a symmetric matrix, A, by
performing consecutive similarity transformations described as plane rotations by an angle, . The
algorithm involves the following steps:
1. Compute
Ai,i Aj,j
w cot 2 = . (B.2.9)
2Ai,j
B.3 Nonlinear algebraic equations 1073
sign(w)
tan = . (B.2.10)
|w| + w2 + 1
3. Set
1
cos = 2
. (B.2.11)
tan + 1
4. Compute
sin
r= . (B.2.12)
1 + cos
5. Compute the elements of B corresponding to the amended elements of A from the equations
and
Bi,k = Bk,i = Ai,k + (Aj,k rAi,k ) sin , Bj,k = Bk,j = Aj,k (Ai,k + rAj,k ) sin ,
(B.2.14)
for k = 1, . . . , N and k = i, j.
The algorithm involves sweeping the o-diagonal elements of the upper or lower triangular block of
the evolving matrix according to a certain protocol.
As the similarity transformations continue, the transformed matrix tends to become diagonal
at a quadratic rate. This means that, after a full sweep has been completed, the sum of the squares
of the norms of the o-diagonal elements has been roughly squared. In practice, only a few sweeps
are necessary to reduce the magnitude of the o-diagonal elements below a negligible threshold.
QR decomposition
The method of QR decomposition can be applied to an arbitrary matrix that is not necessarily
symmetric. The algorithm involves decomposing a matrix A as A = QR, where Q is an orthogonal
matrix and R is an upper triangular matrix whose diagonal elements are equal to unity. Setting
P = Q, we obtain B = QT A Q = R Q. As the transformations are repeated, the evolving matrix
tends to become precisely or nearly upper triangular.
fi (x1 , x2 , . . . , xN ) = 0, (B.3.1)
1074 Introduction to Theoretical and Computational Fluid Dynamics
Fixed-point iterations
To compute the solution of the vector equation f (x) = 0, we recast it into the form x = g(x), and
perform xed-point iterations based on the formula
beginning with a certain initial guess, x(0) , where the superscript (k) is an iteration index. It is
possible that the sequence x(k) will converge to the xed point of the mapping function g, denoted
by X, which, by denition, satises the equation X = g(X) or f (X) = 0. A sucient condition for
the xed-point iterations to converge is that
N
|Gi,j (X)| 1, (B.3.3)
i=1
for any j. The Jacobian matrix, G, with elements Gi,j = gi /xj , is the transpose of the gradient
of g, G (g)T . A necessary condition for the iterations to converge is that the spectral radius of
G evaluated at the a priori unknown xed point is less than unity.
where F1 is the inverse of F. The algorithm involves solving a linear system of equations for a
correction vector e,
and then setting x(k+1) = x(k) + e. The iterations are terminated when |e| falls below a specied
threshold.
f (x(k) )
x(k+1) = x(k) , (B.3.6)
f (x(k) )
Newtons method is guaranteed to converge so long as the initial guess is suciently close
to a root. The rate of convergence is quadratic: each time we carry out one iteration, the error is
raised to the second power and then multiplied by a constant coecient. In the case of a multiple
B.3 Nonlinear algebraic equations 1075
root, the rate of convergence is linear: each time we carry out one iteration, the error is multiplied
by a constant factor that is less than unity. In the case of a double root, the constant factor is
equal to one half. However, if necessary, second-order convergence can be rectied by a judicious
modication of the basic algorithm.
Secant method
The secant method is a modication of Newtons method designed to circumvent the computation
of derivatives. In the case of one equation, N = 1, the algorithm is
x(k) x(k1)
x(k+1) = x(k) f (k) . (B.3.7)
f (k) f (k1)
Two guesses are required to initialize the iterations. Each time we carry out one iteration, the error
is raised approximately to the 1.6 power.
Quasi-Newtons methods are modications of Newtons method where the Jacobian matrix is
either kept constant or constructed during the iterations.
The problem has been reduced to solving a system of two nonlinear equations,
bN (r, s) = 0, bN +1 (r, s) = 0, (B.3.9)
where the coecients bN and bN +1 are polynomial functions of r and s. The solution is found using
Newtons method for two equations. To evaluate the coecients bN and bN +1 and their partial
derivatives with respect to r and s, we construct the sequence
b1 = a1 , b2 = a2 + r b1 , b3 = a3 + r b2 + s b1 , ...,
bN = aN + r bN 1 + s bN 2 , bN +1 = aN +1 + r bN + s bN 1 , (B.3.10)
and then the sequence
c1 = 0, c2 = b 1 , c3 = b 2 + r c 1 , c4 = b3 + r c3 + s c2 , ...,
cN = bN 1 + r cN 1 + s cN 2 , cN +1 = bN + r cN + s cN 1 . (B.3.11)
1076 Introduction to Theoretical and Computational Fluid Dynamics
Deation
Once a root of the equation f (x) = 0 has been found, denoted by X, it can be screened out by
considering the deated equation f (x) = 0, where f (x) = f (x)/|x X|m and m is the multiplicity
of X.
Initial guess
A successful estimate of the location of a root can be made by examining whether any terms of the
function f take small or large values for small or large values of the individual scalar components
of the vector of independent variables, x. If they do, we either ignore these terms or discard the
rest of the terms, compute the solution of the simplied system, and check a posteriori whether the
approximations that lead us to the approximate solution are justied.
Parameter continuation
A successful initial guess can be found by the method of parameter continuation or embedding.
The idea is to perturb the equation f (x) = 0 to formulate a modied equation, q(x, ) = 0, where
q(x, = 1) = f (x) and the equation q(x, = 0) = 0 is easy to solve. The procedure involves solving
a sequence of equations, such as
where the initial guess in solving each equation is the converged solution of the previous equation.
In uid mechanics, the parameter can be identied with the Reynolds number.
f (N +1) ()
e(x) = N +1 (x), (B.4.2)
(N + 1)!
where
Vandermonde matrix
To compute the interpolating polynomial, we may enforce the interpolation constraint f (xi ) =
PN (xi ) at the N + 1 data points, and thereby derive a system of N + 1 equations for the N + 1
unknown coecients, ai . The matrix of the linear system multiplying the unknown vector is the
transpose of the Vandermonde matrix. The determinant of this matrix can be shown to be equal to
the product of xi xj with i > j, and this ensures that the solution of the linear system is unique
as long as the data points are distinct. Unfortunately, the Vandermonde matrix is nearly singular
for moderate and high values of N , placing limits on the accuracy and practicality of the numerical
method.
Lagrange interpolation
In Lagranges method, the interpolating polynomial is constructed in an expedient manner that
circumvents the explicit computation of the polynomial coecients, by setting
N +1
PN (x) = f (xi ) lN,i (x), (B.4.4)
i=1
where
(x x1 )(x x2 ) (x xi1 )(x xi+1 ) (x xN +1 )
lN,i (x) = (B.4.5)
(xi x1 )(xi x2 ) (xi xi1 )(xi xi+1 ) (xi xN +1 )
are N th-degree Lagrange polynomials dened in terms of the data points. An alternative compact
representation is
1 N +1 (x)
lN,i (x) = , (B.4.6)
N +1 (xi ) x xi
where the polynomial N +1 is dened in (B.4.3). The denominators of the fractions dening lN,i (x)
are constant, whereas the numerators are N th-degree polynomials in x.
Linear polynomial interpolation employs two consecutive data points and yields the rst-degree
polynomial
(i) f (xi+1 ) f (xi )
P1 (x) = f (xi ) + (x xi ) , (B.4.7)
xi+1 xi
1078 Introduction to Theoretical and Computational Fluid Dynamics
or
(i) x xi+1 x xi
P1 (x) = f (xi ) + f (xi+1 ) , (B.4.8)
xi xi+1 xi+1 xi
Quadratic interpolation employs three consecutive data points and yields the second-degree
polynomial
(i)
P2 (x) = f (xi ) + [ (ai (x xi ) + bi ](x xi ), (B.4.9)
for xi1 < x < xi+1 . The coecients ai and bi are computed sequentially as
1 f (x ) f (x ) f (x ) f (x )
i+1 i i1 i
ai = ,
xi+1 xi1 xi+1 xi xi1 xi
(B.4.10)
f (xi+1 ) f (xi )
bi = ai (xi+1 xi ).
xi+1 xi
Cubic-spline interpolation
Cubic-spline interpolation ts a third-degree polynomial over each interval between two consecutive
data points and matches the rst and second derivatives of adjacent polynomials at the data points.
Denoting the ith cubic polynomial by
(i)
P3 (x) = ai (x xi )3 + bi (x xi )2 + ci (x xi ) + yi , (B.4.11)
we nd that
1 bi+1 bi yi+1 yi 1
ai = , ci = hi (2bi + bi+1 ), (B.4.12)
3 hi hi 3
where hi = hi+1 hi and
1 d2 P3
(N )
bN +1 = 3 aN hN + bN . (B.4.13)
2 dx2 x=xN +1
i+2
hi bi + 2 (hi + hi+1 ) bi+1 + hi+1 bi+2 = 3 , (B.4.14)
hi+1 hi
Trigonometric interpolation
Consider a real function, f (x), dened in the interval [a, b]. Trigonometric or Fourier approximation
and interpolation represents the function in that interval with a truncated complete Fourier series,
1 M M
FM (x) = a0 + ap cos(pk
x) + bp sin(pk
x), (B.4.18)
2 p=1 p=1
where x = x a, k = 2/L is the wave number, L = b a is the length of the interval, M is a chosen
truncation level, and ap , bp are real Fourier coecients. Although x can be replaced by x without
loss of generality, this complicates the forthcoming algebraic expressions. Outside the interval [a, b],
the Fourier series yields the periodic repetition of the portion of f in [a, b].
where cp are complex Fourier coecients and i is the imaginary unit, i2 = 1. To ensure that the
right-hand side of (B.4.19) is real, we require that
cp = cp , (B.4.20)
where an asterisk denotes the complex conjugate. Using the Euler decomposition of the complex
exponential,
exp(ipk x) i sin(pk
x) = cos(pk x), (B.4.21)
1080 Introduction to Theoretical and Computational Fluid Dynamics
b L
exp[i (p q)k
x] dx = exp[i (p q)k
x] d
x = L pq
a 0
L L if p = q = 0
1
cos(pk
x) cos(qk
x) d
x= 2 L if p = q = 0
0
0 otherwise
L & 1
2 L if p = q = 0,
sin(pk
x) sin(qk
x) d
x=
0 0 otherwise.
L
cos(pk
x) sin(qk
x) d
x=0
0
Table B.4.1 Orthogonality of trigonometric functions in the interval [a, b]; L = b a, k = 2/L, p
= x a.
and q are two integers, and x
To compute the Fourier coecients, we use of the Fourier orthogonality properties shown in
Table B.4.1. Multiplying both sides of (B.4.19) by exp(i qx), where q is an integer, using the rst
orthogonality property in Table B.4.1, and then relabeling q p, we nd that
1 b
cp = f (x) exp(i pk
x) dx, (B.4.23)
L a
and thus
b b
2 2
ap = f (x) cos(pk
x) dx, bp = f (x) sin(pk
x) dx. (B.4.24)
L a L a
It can be shown that, when the coecients are computed in this fashion, in the limit as M tends
to innity, the Fourier representation (B.4.18) becomes exact. The integrals in (B.4.24) can be
computed by numerical integration, as discussed in Section B.6.
Let us divide the interval L into N subintervals separated by N +1 data points xi = a+(i1)h,
where i = 1, . . . , N + 1, h = L/N , and xN +1 = b. Using the trapezoidal rule to evaluate the complex
B.4 Function interpolation 1081
N &
N if p = sN
exp(ipk
xj ) =
0 otherwise
j=1
N &
N if p q = sN
exp[ijk(xp xq )] =
0 otherwise
j=1
N N if p q = sN = 0
1
cos(pk
xj ) cos(qk
xj ) = 2 N if p q = 0
j=1 0 otherwise
N & 1
2
N if p q = sN
sin(pk
xj ) sin(qk
xj ) =
0 otherwise
j=1
N
cos(pk
xj ) sin(qk
xj ) = 0
j=1
1 1 1
cp = f (x1 ) + p f (x2 ) + 2p f (x3 ) + + (N 1)p f (xN ) + f (xN +1 ) , (B.4.25)
N 2 2
where
2i
exp(ikh) = exp( ). (B.4.26)
N
Using the identities shown in Table B.4.2, we nd that, when the coecients are computed in this
fashion, the truncated Fourier series interpolates through the interior data,
for j = 2, . . . , N . The value of the truncated Fourier series at the rst and last data is the end point
average,
1
FM (x1 ) = FM (xN +1 ) = [ f (x1 ) + f (xN +1 ) ]. (B.4.28)
2
Thus, the truncated Fourier series interpolates through the mean of the rst and last points.
1082 Introduction to Theoretical and Computational Fluid Dynamics
If the interpolated function f (x) is repeated along the x axis with period L, f (x1 ) = f (xN +1 ),
formula (B.4.25) takes the simpler form
1 (j1)p
N
cp = f (xj ). (B.4.29)
N j=1
In that case,
FM (xj ) = f (xj ) (B.4.30)
for all j = 1, . . . , N + 1, that is, the truncated Fourier series interpolates through all N + 1 data. In
practice, when N is large, the sum on the right-hand side of (B.4.29) is computed most eciently
by the method of fast Fourier transform (FFT) requiring N log2 N operations.
Consider the computation of the derivatives of a function at the ith datum point, xi . If we
use an equal number of data points on either side of xi to construct the interpolating polynomial, we
obtain central dierences. Otherwise, we obtain spatially biased, forward or backward dierences.
Table B.5.1 summarizes nite-dierence formulas for computing the derivatives of a function of one
variable, x, at the point xi using the values of the function at a collection of evenly spaced data
points separated by distance x = h (e.g., [185]).
Finite-dierence formulas for computing the Laplacian of a function of two variables, x and y
at a grid point, (xi , yi ), can be derived working in a similar fashion. Consider a uniform Cartesian
grid with spacings x and y, and dene = (x/y)2 . The ve-point formula yields
0 0
1
(2 f )i,j = 1 2 (1 + ) 1 : F, (B.5.1)
x2
0 0
with accuracy of O(x2 ) and O(y 2 ), where
fi1,j+1 fi,j+1 fi+1,j+1
F = fi1,j fi,j fi+1,j . (B.5.2)
fi1,j1 fi,j1 fi+1,j1
The nine-point formula with x = y yields
1
1 4 1
2 f 4 20 4 : F, (B.5.3)
i,j 6x2
1 4 1
with accuracy of O(x2 ).
B.5 Numerical dierentiation 1083
Table B.5.1 Finite-dierence formulas for the derivatives of a function, f (x), at the grid point, xi ,
in terms of values of the function at a set of evenly spaced neighboring points separated by the
distance x = h. (Continuing.)
1084 Introduction to Theoretical and Computational Fluid Dynamics
A typical method of numerical integration involves approximating the integrand, f (x), with a global
interpolating polynomial or with a set of local interpolating polynomials dened in [a, b], as discussed
in Section B.4, and then integrating the interpolating polynomials over their domain of denition.
Trapezoidal rule
Approximating f (x) with a straight line between two consecutive data points, described by a rst-
degree local interpolating polynomial, we obtain the trapezoidal rule. When the base points are
distributed evenly with constant separation h = (b a)/N , we obtain
1 1
Itr (h) = h ( f1 + f2 + + fN + fN +1 ), (B.6.2)
2 2
where x1 = a and xN +1 = b. It can be shown that the leading-order numerical error is
1
Etr (h) = (b a) f () h2 , (B.6.3)
12
where the point lies somewhere inside the integration domain.
B.6 Numerical integration 1085
Romberg integration uses the results of two computations with dierent interval sizes h to
improve the accuracy to fourth order in h. For example,
1
Simpsons rule
Approximating f (x) with a parabola subtended across triplets of consecutive data points, described
by a second-degree local interpolating polynomial, we obtain Simpsons one-third rule. Assuming
that the number of intervals N is even, we obtain
1
Ismp (h) = h f1 + 4f2 + 2f3 + 4f4 + + 2fN 1 + 4fN + fN +1 . (B.6.5)
3
The numerical error is
1 1
ESmp (h) = (b a) f () h4 (f f1 ) h4 , (B.6.6)
180 180 N +1
where the point lies somewhere inside the integration domain. Romberg integration improves the
accuracy to sixth order in h by setting
1
Gauss quadratures
Gauss quadratures require that the data points are distributed in special ways over the integration
domain. Global polynomial interpolation followed by analytical integration leads to the numerical
approximation
1
NQ
IGauss = (b a) f (xi ) wi , (B.6.8)
2 i=1
where NQ is the chosen number of quadrature base points and wi are proper weights. The position
of the base points, xi , is determined by the zeros of a properly selected NQ -degree orthogonal
polynomial.
The GaussLegendre quadrature is designed for functions that are free of singularities inside
the entire integration domain, including the boundaries. The base points where the integrand is
evaluated are located at xi = 12 [a + b + ti (b a)], where ti are the zeros of the NQ -degree Legendre
polynomial dened in the interval [1, 1]. For each chosen NQ , the weights wi add up to two, which
is necessary for (B.6.8) is to be valid when f is a constant function. Polynomials of degree 2NQ 1
or less are integrated exactly. Values of zi and wi for several values of NQ are given in texts on
numerical methods (e.g., [317]).
1086 Introduction to Theoretical and Computational Fluid Dynamics
Other Gauss quadratures applicable to singular integrals and semi-innite or innite integra-
tion domains are available (e.g., [317]).
Two-dimensional integrals
The choice of a successful numerical method for computing the integral of a function of two variables
f (x, y) over a two-dimensional domain depends on the smoothness of the function and geometry of
the integration domain.
When the function f (x, y) is smooth and the domain is a rectangular area conned within
a < x < b and c < yd, the successive application of two GaussLegendre quadratures for integration
in each direction yields the compound quadrature
NQx Qy N
1
f (x, y) dx dy (b a)(c a) f (xi , yj ) wi wj , (B.6.9)
rectangle 4 i=1 i=1
where NQx and NQy are two independent integers determining the order of the quadrature, and xi ,
yj correspond to the Gauss-Legendre base points.
To compute the integral of a smooth function over the surface of a triangle with vertices at
the points v1 , v2 , and v3 , we use the quadrature
NQ
f (x, y) dx dy A f (xi ) wi , (B.6.10)
triangle i=1
xi = i v1 + i v2 + i v3 (B.6.11)
is the position of the base points over the triangle. The coecients i , i , and i , and corresponding
weights, wi , are are given in texts on numerical methods [317].
In the least-squares method, the coecients of the polynomial are found by minimizing the
integral of the squared dierence, |f (x) PN (x)|2 , multiplied by a chosen weighting function, w(x),
over [a, b]. Unfortunately, when N is higher than about ve, this approach results in a nearly singular
system of algebraic equations.
B.8 Ordinary dierential equations 1087
Orthogonal polynomials
In an alternative approach, we introduce a family of orthogonal polynomials, pi (x), for i = 0, 1, . . .,
where pi (x) is an ith degree polynomial with associated weighting function w(x) dened in the
interval [c, d]. By construction,
d
pi , pj pi (x) pj (x) w(x) dx = Dij , (B.7.1)
c
where D is a diagonal matrix. Two popular families of orthogonal polynomials are the Legendre
polynomials and the Chebyshev polynomials, both dened in the interval [1, 1]. An important
property of the Chebyshev polynomials is that their magnitude is less than unity for any value of x
in the interval [1, 1].
Next, we introduce the scaled variable t = c + (d c)(x b)/(a b), where a < b and c < d. As
x increases from a to b, the variable t increases from c to d. Finally, we express the approximating
polynomial in terms of a weighted sum of orthogonal polynomials of a chosen class as
N
Pn (t) = ai pi (t), (B.7.2)
i=0
and compute the coecients ai using the orthogonality condition for pi (t) stated in (B.7.1),
d
1
ai = f (t) pi (t) w(t) dt, (B.7.3)
Dii c
where summation is not implied over the repeated index, i. The integral in (B.7.3) can be computed
by numerical integration in terms of the values of f at data points, as discussed in Section B.6. To
relate t to x, we use the inverse transformation, x = a + (t c)(b a)/(d c).
In the following discussion, the superscript (k) denotes the kth time level, tk , and t = tk+1 tk
is the time step.
1088 Introduction to Theoretical and Computational Fluid Dynamics
Euler method
In Eulers method, the solution is advanced over a small time interval, t, by moving in the phase
space by a small distance with the initial phase velocity, so that
where the superscript tmp denotes a temporary solution, and and are positive numerical
parameters satisfying 2 = 1.
Setting = 12 and = 1 yields the standard version of the RK2 method, also called the
modied Euler method.
1
Setting = 1 and = 2 yields the midpoint RK2 method.
3 2
Setting = 4 and = 3 yields Heuns method.
Heuns method minimizes the magnitude of the numerical error. In all cases, each complete time
step introduces a numerical error of order t3 .
1
6. Set: f f inal = 6 f (k) + 4 f tmp1 + f tmp2
C.1 Installation
The source code of Fdlib, containing programs and data les, can be obtained from the Fdlib In-
ternet site stated above. The directories have been archived using the tar Unix facility into the
compressed FDLIB *.tgz le, where the asterisk denotes the version number encoded as year and
month (yy mm). To unravel the directories (folders) in a Unix system, issue the Unix command:
tar xzf FDLIB *.tgz. This will generate the FDLIB * directory containing nested subdirectories.
To unravel the directories in another operating system, double-click on the archived tar le and
follow the instructions of the invoked application.
To compile an application named nea krini, navigate to the subdirectory where the application
resides and type: make nea krini. To compile the Fortran 77 programs using a Fortran 90
compiler, make appropriate compiler call substitutions in the makeles. To remove the object les,
output les, and executable of an application named polihni, navigate to the subdirectory where the
application resides and issue the command: make clean.
1090
C FDLIB software library 1091
Subject Directory
Table C.1 The software library Fdlib is arranged in thirteen directories according to physical or
numerical classication.
CFDLAB
A subset of Fdlib has been combined with the X11 graphics library vogle into the integrated
application Cfdlab that visualizes the results of simulations and performs interactive animation.
Vogle is written in C but oers Fortran 77 and C++ interfaces. The source code of Cfdlab can be
downloaded from the Internet site: http://dehesa.freeshell.org/CFDLAB.
BEMLIB
A subset of Fdlib containing boundary-element and related codes have been arranged in the li-
brary Bemlib [313]. The source code of Bemlib can be downloaded from the Internet site:
http://dehesa.freeshell.org/BEMLIB.
Figure C.1 Streamlines of potential ow exiting a two-dimensional channel generated using code strml
in Directory 04 various of Fdlib.
1092 Introduction to Theoretical and Computational Fluid Dynamics
01 num meth
This directory contains a suite of general-purpose programs on general numerical methods and
dierential equations accompanying Reference [317].
Subdirectory Topic
Figure C.2 Finite-element solution of Laplaces equation with the Dirichlet boundary condition gen-
erated using code lapl3 d in Directory of 15 fem inside Directory 01 num meth of Fdlib.
C FDLIB software library 1093
02 grids
This directory contains programs that perform grid generation, adaptive discretization, parametriza-
tion, representation, and meshing of planar lines, three-dimensional lines, and three-dimensional
surfaces.
Subdirectory Topic
Figure C.3 Triangulation of a disk, square, square with a square hole and square with a circular hole
into six-node triangular elements generated using codes trgl6 disk, trgl6 sqr, trgl6 ss, and trgl6 sc
in Directory of trgl at inside Directory 02 grids of Fdlib.
1094 Introduction to Theoretical and Computational Fluid Dynamics
03 hydrostat
Subdirectory Topic
1.5
0.5
x/a
0.5
1.5
2
2 1 0 1 2
y/a
Figure C.4 Equilibrium position of a oating sphere generated using code sphere in Directory
03 hydrostat of Fdlib.
C FDLIB software library 1095
04 various
This directory contains miscellaneous codes that compute the structure and kinematics of various
ows.
Subdirectory Topic
05 lub
This directory contains codes that solve problems involving lubrication ows.
Subdirectory Topic
0.04
0.03
0.02
u
0.01
0
40
40
20 30
20
0 10
0
y 20 10
x
Figure C.5 Velocity proles of unidirectional pressure-driven ow through a rectangular tube com-
puted by a boundary-element method implemented in the code ow 1d in Directory 04 various of
Fdlib.
Figure C.6 Evolution of two lms owing down an inclined plane computed using code lms in
Directory 05 lub of Fdlib.
C FDLIB software library 1097
06 stokes
This directory contains codes that compute viscous ows at vanishing Reynolds number.
Subdirectory Topic
Figure C.7 Color-coded (or grayscale) plot of the shear stress over the surface of a particle attached to
a wall in simple shear ow computed by a boundary-element method implemented in code bump 3d
included in Directory 06 stokes of Fdlib.
1098 Introduction to Theoretical and Computational Fluid Dynamics
07 ptf
This directory contains codes that solve problems involving potential ow.
Subdirectory Topic
Figure C.8 Distribution of the pressure coecient around the NACA 23012 airfoil computed by a
panel method implemented in code airf 2d lvp included in Directory 07 ptf of Fdlib.
C FDLIB software library 1099
08 stab
This directory contains codes that perform the linear stability analysis of miscellaneous ows. The
fundamental concepts underlying the linear stability analysis are discussed in Chapter 9. The User
Guide of this Directory can be found in Appendix D.
Subdirectory Topic
09 vortex
Subdirectory Topic
Figure C.9 Evolution of a vortex patch computed by the method of contour dynamics implemented
in code vp 2d in Directory 09 vortex of Fdlib.
C FDLIB software library 1101
10 bl
Subdirectory Topic
11 fdm
This directory contains codes that solve problems using nite-dierence methods.
Subdirectory Topic
12 bem
This directory contains codes that produce solutions to Laplaces equation by boundary-element
methods.
Subdirectory Topic
13 turbo
Subdirectory Topic
0.5
velocity
0.5
0 200 400 600 800 1000 1200 1400 1600 1800 2000
time
Figure C.10 A time series of two velocity components in a turbulent ow contained in the le keller.dat
included in Directory stats inside Directory 13 turbo of Fdlib.
User Guide of directory 08 stab
of FDLIB on hydrodynamic stability
D
This appendix contains the User Guide of the eighth directory of the software library Fdlib on
hydrodynamic stability (see Appendix C). The individual subdirectories contain programs that per-
form the linear stability analysis of several families of inviscid and viscous ows at zero or nonzero
Reynolds numbers. The fundamental concepts underlying the linear stability analysis are discussed
in Chapter 9. The subdirectories of the main directory 08 stab are listed in Table D.1.
Subdirectory Topic
1103
1104 Introduction to Theoretical and Computational Fluid Dynamics
Subdirectory Topic
Subdirectory Topic
Table D.1 (Continued.) Contents of the Fdlib directory 08 stab on hydrodynamic stability.
The User Guide of the individual directories is presented in this appendix, including the
denition of the base state, the formulation of the linear stability problem, and the derivation of an
algebraic eigenvalue problem determining the complex growth rate. To facilitate the study of the
individual programs, some parts of the analysis and a few gures are repeated across the individual
directories. This User Guide is intended to be used in concert with Chapter 9 on the fundamentals
of hydrodynamic stability.
Fluid 1 (inner)
Fluid 2 (outer)
ae
a
ai
Figure D.1.1 Illustration of two annular layers conned between an inner cylindrical tube of radius ai
and an outer cylindrical tube of radius ae . The interface is occupied by an insoluble surfactant.
1106 Introduction to Theoretical and Computational Fluid Dynamics
Base state
We consider two stationary annular layers separated by a cylindrical interface of radius a, placed
between an inner cylindrical tube of radius ai and an outer cylindrical tube of radius ae , as illustrated
in Figure D.1.1. The inner uid is labeled 1 and the outer uid is labeled 2. In the unperturbed
conguration, the uids are quiescent and the surfactant is distributed uniformly over the interface.
Several special cases can be recognized:
When the inner cylinder is absent and the radius of the outer cylinder is innite, we obtain a
thread suspended in an innite ambient uid.
When the radius of the outer cylinder is innite, we obtain an annular layer coated on the
exterior surface of the inner cylinder.
Normal-mode analysis
To carry out the normal-mode stability analysis for axisymmetric perturbations, we introduce cylin-
drical polar coordinates, (x, , ), where the x axis coincides with the axis of revolution of the
interface. The radial position of the interface is described by the real or imaginary part of the
function
is the normal-mode wave form of the perturbation, A is the complex amplitude of the interfacial
wave, k = 2/L is the wave number, L is the wavelength, and c is the complex phase velocity.
where
2 2 1
E2 2
+ 2
(D.1.5)
x
is a second-order dierential operator. It is convenient to work with the vorticity transport equation
for axisymmetric ow,
D 1
j
j = 2 E 2 ( j ), (D.1.6)
Dt
where D/Dt is the material derivative and j is the kinematic viscosity of the jth uid. Substituting
expression (D.1.4) into (D.1.6), linearizing with respect to and rearranging, we obtain
1 2 (1)
E2 E j = 0 (D.1.7)
j t
for j = 1, 2.
To carry out the normal-mode analysis, we express the perturbation stream function in the
form
(1)
j (x, , t) = j () exp[ik(x ct)], (D.1.8)
2 2
k k j = 0, (D.1.9)
d 2 d d 2 d j
where
k
kj2 k 2 i c . (D.1.10)
j
for j = 1, 2, where k, j kj , I1 and K1 are Bessel functions, and Ai,j , Bi,j are complex
coecients to be found as part of the solution. The rst and third terms on the right-hand side
of (D.1.11) correspond to the rst operator, while the second and fourth terms correspond to the
second operator on the left-hand side of (D.1.9).
Substituting into (D.1.3) the preceding expressions and using the properties of the Bessel
functions
1
I0 (z) = I1 (z), I1 (z) = I0 (z)
I1 (z), K0 (z) = K1 (z),
z
1
K1 (z) = K0 (z) K1 (z), (D.1.12)
z
1108 Introduction to Theoretical and Computational Fluid Dynamics
we nd that
uxj = Uxj () exp[ik(x ct)], uj = Uj () exp[ik(x ct)], (D.1.13)
where
1 dj
Uxj () = = A1,j k I0 () + A2,j kj I0 (j ) B1,j k K0 () B2,j kj K0 (j ), (D.1.14)
d
and
k
Uj () = i j = ik A1,j I1 () + A2,j I1 (j ) + B1,j K1 () + B2,j K1 (j ) . (D.1.15)
Pressure eld
The normal-mode expression for the pressure is
pj = Pj + j () exp[ik(x ct)], (D.1.16)
where P1 and P2 are the unperturbed pressures, P1 P2 = 0 /a is the unperturbed capillary pressure,
0 is the unperturbed surface tension, and j are eigenfunctions. To compute the disturbance
pressure, we consider the x component of the NavierStokes equation for axisymmetric ow,
Duxj pj 2u 1 uxj
xj
j = + j + ( ) . (D.1.17)
Dt x x2
Substituting (D.1.16) together with the rst equation into (D.1.13) and linearizing, we obtain
j 1 d dUxj
j 1 d dUxj
Kinematic compatibility
Kinematic compatibility requires that D[f (x, t) ]/Dt = 0 at the interface, where D/Dt is the
material derivative and the function f describes the position of the interface according to (D.1.1).
Carrying out the dierentiation, we obtain
f f
+ uxj uj = 0, (D.1.23)
t x
where the velocity is evaluated at the interface. Substituting the preceding expressions and lineariz-
ing, we nd that
i 1
A= U ( = a) = [ A1,j I1 (k) + A2,j I1 (kj ) + B1,j K1 (k) + B2,j K1 (kj ) ]. (D.1.24)
ck j c
Surfactant transport
The interface is occupied by an insoluble surfactant with surface concentration . For a normal-mode
disturbance, = 0 + (1) , where 0 is the unperturbed surfactant concentration,
(1) 1 exp[ik(x ct)] (D.1.28)
is the normal-mode perturbation, and 1 is the complex amplitude of the perturbation. Linearizing
the surfactant transport equation in the absence of base ow, we derive the evolution equation
(1) ux 0 2 (1)
+ 0 = u + Ds , (D.1.29)
t x a x2
where the velocity is evaluated at the unperturbed position, = a. Substituting the normal-mode
forms and rearranging, we nd that
1 1 U
k
= 2
ik Uxj + j = i 2
A1,j [k I0 (k) I1 (k)]
0 s + Ds k =a a (s + Ds k )
+A2,j [kj I0 (kj ) I1 (kj )] B1,j [k K0 (k) + K1 (k)] B2,j [kj K0 (kj ) + K1 (kj )] , (D.1.30)
1110 Introduction to Theoretical and Computational Fluid Dynamics
where s ikc. If c is imaginary, as expected in the absence of a base ow, s = kcI is the real
growth rate.
where 0 is the surface tension in the unperturbed state corresponding to the surfactant concentration
0 , and 1 is the complex amplitude of the perturbation. In the linearized approximation, the surface
tension depends linearly on the surfactant concentration,
1 0
= Ma , (D.1.32)
1 0
where Ma is the Marangoni number. Substituting (D.1.30) and rearranging, we obtain
1 k
= i Ma 2
A1,j [k I0 (k) I1 (k)] + A2,j [kj I0 (kj ) I1 (kj )]
0 a (s + Ds k )
where
and
k
Ma 0 (D.1.36)
a (s + Ds k 2 )
is a dimensionless number.
x1 x2
ik 1 = 1 + ik U1 2 + ik U2 . (D.1.38)
d =a d =a
D.1 FDLIB User Guide: 08 stab/ann2l 1111
1 2 =a = p1 + 2 1 p 2 + 2 2 = 2m , (D.1.43)
=a =a
where m is the mean curvature. In the linearized approximation,
1 1 A
2m = + 2 + xx = + 2 (1 k 2 ) exp[ik(x ct)]. (D.1.44)
a a a a
Using (D.1.24) with j = 1 to eliminate the interfacial amplitude, A, we obtain
1
2m = + 2 (1)
m exp[ik(x ct)], (D.1.45)
a
where
1 k 2
2(1)
m = A 1,1 I 1 ( k) + A2,1 I 1 ( k 1 ) + B 1,1 K1 ( k) + B2,1 K 1 ( k 1 ) . (D.1.46)
ca2
Next, we substitute into (D.1.43) the normal-mode expansions and obtain
U1
U2
1 + 2 1 2 + 2 2
=a =a
2
1 1 k
= + 0 2
A1,1 I1 (k) + A2,1 I1 (k1 ) + B1,1 K1 (k) + B2,1 K1 (k1 ) . (D.1.47)
a ca
1112 Introduction to Theoretical and Computational Fluid Dynamics
= i k A1,j k I0 () I1 () + A2,j kj I0 (j ) I1 (j )
d j
1
1
B1,j k K0 () + K1 () B2,j kj K0 (j ) + K1 (j ) , (D.1.48)
j
together with the last expression in (D.1.20) for the pressure, we obtain
1,1 A1,1 + 2,1 A2,1 + 1,1 B1,1 + 2,1 B2,1 + 1,2 A1,2 + 2,2 A2,2 + 1,2 B1,2 + 2,2 B2,2
1 1
= i = (1 A1,1 + 2 A2,1 3 B1,1 4 B2,1 ), (D.1.49)
a a
where
1
k 3 1 k2
0
1,1 = s 1 I0 (k) + 2 1 k 2 I0 (k) I1 (k) I1 (k),
k s k 2
1
k 3 1 k 2
0
2,1 = 2 1 k k1 I0 (k1 ) I1 (k1 ) I1 (k1 ),
k1 s k2
1
k 3 1 k 2
0
1,1 = s 1 K0 (k) 2 1 k2 K0 (k) + K1 (k) K1 (k), (D.1.50)
k s k2
1
k 3 1 k 2
0
2,1 = 2 1 k k1 K0 (k1 ) + K1 (k1 ) K1 (k1 ),
k1 s k2
and
1
1
where
1 2 3 4
G1 = 1,1 , G2 = 2,1 , G3 = 1,1 + , G4 = 2,1 + ,
a a a a
G5 = 1,2 , G6 = 2,2 , G7 = 1,2 , G8 = 2,2 . (D.1.53)
I1 (k) I1 (k1 ) K1 (k) K1 (k1 ) I1 (k)
k I (k) k K0 (k) k1 K0 (k1 ) k I0 (k)
0 k1 I0 (k1 )
I1 (ki ) I1 (ki1 ) K1 (ki ) K1 (ki1 ) 0
k K0 (ki ) k1 K0 (ki1 )
M = k I0 (ki ) k1 I0 (ki1 ) 0
0 0 0 0 I1 (ke )
0 0 0 0 k I0 (ke )
F1 F2 F3 F4 F5
G1 G2 G3 G4 G5
I1 (k2 ) K1 (k) K1 (k2 )
k2 I0 (k2 ) k K0 (k) k2 K0 (k2 )
0 0 0
0 0 0
I1 (ke2 ) K1 (ke ) K1 (ke2 )
k K0 (ke ) k2 K0 (ke2 )
k2 I0 (ke2 )
F6 F7 F8
G6 G7 G8
Table D.1.1 Setting the determinant of the matrix displayed to zero provides us with an algebraic
equation for the complex growth rate.
and the coecient matrix, M, is shown in Table D.1.1. For a nontrivial solution to exist, the deter-
minant of the coecient matrix must be zero. In the numerical method implemented in the code,
the roots of the secular equation det(M) = 0 are found by Newtons method with a suitable initial
guess. The determinant of the matrix is calculated by LU decomposition using Gauss elimination.
A1,2 and A2,2 to zero and obtain a simplied expression for the exterior stream function,
j () = B1,j K1 () + B2,j K1 (j ) , (D.1.56)
Suspended thread
In the simplest conguration, both the internal and external cylinders are absent, yielding an innite
thread suspended in an innite ambient uid. The stream functions for the internal and external
ows are described, respectively, by equations (D.1.55) and (D.1.56). The linear system M q = 0
undergoes analogous simplications. When the interface is devoid of surfactants, the coecient
matrix reduces to that derived by Tomotika [410], as discussed in Section 9.13.2.
Program les:
1. ann2l
Computes the growth rate.
2. matrix
Formulates the matrix M.
3. bess I01K0
Computes the Bessel functions.
4. gel
Computes the determinant of a matrix by LU decomposition using Gauss elimination.
Input les:
1. ann2l.dat
Problem selection and specication of input parameters.
Output les:
1. ann2l.out
Growth rates.
Fluid 1 (inner)
Fluid 2 (outer)
ae
a
ai
Figure D.2.1 Illustration of two concentric annular layers conned between an inner cylindrical tube
of radius ai , and an outer cylindrical tube of radius ae .
Base state
We consider the instability of the interface between two annular layers separated by a cylindrical
interface of radius a, conned between an inner cylindrical tube of radius ai and an outer cylindrical
tube of radius ae , as illustrated in Figure D.2.1. The inner uid is labeled 1 and the outer uid is
labeled 2. The viscosity of the inner uid is 1 and the viscosity of outer uid is 2 = 1 , where is
the viscosity ratio. The interface is occupied by an insoluble surfactant that diuses and is convected
over the interface, but does not enter the bulk of the uids. In the unperturbed conguration, the
uids are quiescent and the surfactant is distributed uniformly over the cylindrical interface. Several
special cases can be recognized:
When the inner cylinder is absent and the radius of the outer cylinder is innite, we obtain a
thread suspended in an innite ambient uid.
When the inner cylinder is absent, we obtain an annular layer coated on the interior surface
of the outer cylinder.
When the radius of the outer cylinder is innite, we obtain an annular layer coated on the
exterior surface of the inner cylinder.
Normal-mode analysis
To carry out the normal-mode stability analysis for axisymmetric perturbations, we introduce cylin-
drical polar coordinates, (x, , ), where the x axis coincides with the axis of revolution of the
unperturbed interface. Next, we describe the radial position of the interface in terms of the real or
imaginary part of the function
is the normal-mode wave form of the perturbation, A is the complex amplitude of the interfacial
wave, k = 2/L is the wave number, L is the wavelength, and c is the complex phase velocity.
2
k 2 j = j , (D.2.9)
d d
D.2 FDLIB User Guide: 08 stab/ann2l0 1117
where
d2 1 d
2
k j = 0. (D.2.10)
d 2 d
The general solution is
j () = A1,j I1 () + A2,j I0 () + B1,j K1 () + B2,j K0 () , (D.2.11)
where I0 , I1 , K0 , and K1 are Bessel functions, Ai,j , Bi,j are complex constants, and k. The
rst and third terms on the right-hand side of (D.2.11) satisfy (D.2.9) with = 0. The second and
fourth terms satisfy (D.2.9) with being a nontrivial solution of (D.2.10).
Substituting the preceding expressions into (D.2.3) and using the properties of the Bessel
functions
1
I0 (z) = I1 (z), I1 (z) = I0 (z)
I1 (z), K0 (z) = K1 (z),
z
1
K1 (z) = K0 (z) K1 (z), (D.2.12)
z
we nd that
where
1 dj
Uxj () = = A1,j k I0 () + A2,j [2 I0 () + I1 ()]
d
B1,j k K0 () + B2,j [2 K0 () K1 ()], (D.2.14)
and
k
Uj () = i j = ik A1,j I1 () + A2,j I0 () + B1,j K1 () + B2,j K0 () . (D.2.15)
Pressure eld
The normal-mode expression for the pressure is
where P1 and P2 are the uniform unperturbed pressures, the pressure jump P1 P2 = 0 /a is
the unperturbed capillary pressure, 0 is the unperturbed surface tension, and j () are pressure
eigenfunctions. To compute the disturbance pressure, we consider the x component of the Stokes
equation for axisymmetric ow,
pj 2u 1 uxj
xj
= j 2
+ . (D.2.17)
x x
1118 Introduction to Theoretical and Computational Fluid Dynamics
Substituting (D.2.16) together with the rst equation in (D.2.13) and linearizing, we obtain
j 1 d dUxj j 1 d dUxj
2
1 dUxj 2
j () = i 2
+ k U xj = i 2 j k A2,j I0 () + B2,j K0 () . (D.2.19)
k d d
Note that only two terms contribute to the pressure eld.
A1,1 k I0 (k) + A2,1 [2 I0 (k) + k I1 (k)] B1,1 k K0 (k) + B2,1 [2 K0 (k) k K1 (k)]
= A1,2 k I0 (k) + A2,2 [2 I0 (k) + k I1 (k)] B1,2 k K0 (k) + B2,2 [2 K0 (k) k K1 (k)] (D.2.20)
and
A1,1 I1 (k) + A2,1 a I0 (k) + B1,1 K1 (k) + B2,1 a K0 (k)
= A1,2 I1 (k) + A2,2 a I0 (k) + B1,2 K1 (k) + B2,2 a K0 (k), (D.2.21)
Kinematic compatibility
Kinematic compatibility requires that D[f (x, t) ]/Dt = 0 evaluated at the interface, where D/Dt
is the material derivative and the function f describes the position of the interface according to
(D.2.1). Carrying out the dierentiation, we obtain
f f
+ ux u = 0, (D.2.22)
t x
where the velocities ux and u are evaluated at the interface. Substituting the preceding expressions
and linearizing, we nd that
i 1
A= Uj ( = a) = [ A1,j I1 (k) + A2,j a I0 (k) + B1,j K1 (k) + B2,j a K0 (k) ]. (D.2.23)
ck c
where ke = kae .
Surfactant transport
The interface is occupied by an insoluble surfactant with surface concentration . For a normal-mode
perturbation, = 0 + (1) , where 0 is the unperturbed surfactant concentration,
(1) 1 exp[ik(x ct)] (D.2.27)
is the normal-mode surfactant perturbation, and 1 is the complex amplitude of the perturbation.
Linearizing the surfactant transport equation in the absence of a base ow, we derive the evolution
equation
(1) ux 0 2 (1)
+ 0 = u + Ds , (D.2.28)
t x a x2
where the velocity is evaluated at the unperturbed position = a. Substituting the normal-mode
forms and rearranging, we nd that
1 1 1
k
= 2
ik Uxj + Uj = i 2
A1,j [k I0 (k) I1 (k)]
0 s + Ds k a =a a (s + Ds k )
+A2,j a [I0 (k) + k I1 (k)] B1,j [k K0 (k) + K1 (k)] + B2,j a [K0 (k) k K1 (k)] , (D.2.29)
where s ikc. If c is imaginary, as expected in the absence of a base ow, s = kcI is the real
growth rate.
Selecting j = 1, corresponding to the inner uid, we recast this equation into the form
1 = i ( 1 A1,1 + 2 A2,1 + 3 B1,1 + 4 B2,1 ), (D.2.33)
where
1 = [k I0 (k) I1 (k)], 2 = [I0 (k) + k I1 (k)], 3 = [k K0 (k) + K1 (k)],
a a
4 = [K0 (k) k K1 (k)], (D.2.34)
and
0
Ma (D.2.35)
s + Ds k2
is a dimensionless number pertinent to the surfactant.
x1 x2
1 + ik U1 2 + ik U2 = ik1 . (D.2.37)
d =a d =a
21 k A1,1 k I1 (k) + A2,1 [I1 (k) + k I0 (k)] + B1,1 k K1 (k) B2,1 [K1 (k) k K0 (k)]
22 k A1,2 k I1 (k) + A2,2 [I1 (k) + k I0 (k)] + B1,2 k K1 (k) B2,2 [K1 (k) k K0 (k)]
= ik 1 = k 1 A1,1 + 2 A2,1 + 3 B1,1 + 4 B2,1 . (D.2.39)
Consolidating the various terms, we obtain a linear algebraic equation,
F1 A1,1 + F2 A2,1 + F3 B1,1 + F4 B2,1 + F5 A1,2 + F6 A2,2 + F7 B1,2 + F8 B2,2 = 0, (D.2.40)
where
F1 = 21 k I1 (k) + 1 , F2 = 21 [I1 (k) + k I0 (k)] + 2 ,
F3 = 21 k K1 (k) + 3 , F4 = 21 [K1 (k) + k K0 (k)] + 4 ,
(D.2.41)
F5 = 22 k I1 (k), F6 = 22 [I1 (k) + k I0 (k)],
F7 = 22 k K1 (k), F8 = 22 [K1 (k) + k K0 (k)].
D.2 FDLIB User Guide: 08 stab/ann2l0 1121
1 2 =a = p1 + 2 1 p2 + 2 2 = 2 m , (D.2.42)
=a =a
where m is the mean curvature. In the linearized approximation,
1 1 A
2m = + 2 + xx = + 2 (1 k2 ) exp[ik(x ct)]. (D.2.43)
a a a a
Using (D.2.23) with j = 1 to eliminate the interfacial amplitude, A, we obtain
1
2m = + 2(1)
m exp[ik(x ct)], (D.2.44)
a
where
1 k 2
2(1)
m = 2
A1,1 I1 (k) + A2,1 a I0 (k) + B1,1 K1 (k) + B2,1 a K0 (k) . (D.2.45)
ca
Substituting into (D.2.42) the normal-mode expansions, we obtain
U1
U2
1
1 + 2 1 2 + 2 2 = (D.2.46)
=a =a a
2
1 k
+0 2
A1,1 I1 (k) + A2,1 a I0 (k) + B1,1 K1 (k) + B2,1 a K0 (k) .
ca
Substituting into the left-hand side the expression
dUj 1
= i k A1,j k I0 () I1 () + A2,j I0 () + I1 ()
d
1
B1,j k K0 () + K1 () + B2,j K0 () K1 () (D.2.47)
together with the last expression in (D.2.19) for the pressure, we nd that
1,1 A1,1 + 2,1 A2,1 + 1,1 B1,1 + 2,1 B2,1 + 1,2 A1,2 + 2,2 A2,2
1 1
+1,2 B1,2 + 2,2 B2,2 = i = 1 A1,1 + 2 A2,1 + 3 B1,1 + 4 B2,1 , (D.2.48)
a a
where
1 0 k 3 1 k 2
1,1 = 21 k 2 I0 (k) I1 (k) I1 (k),
k s k2
0 k 3 1 k 2
2,1 = 21 k k I1 (k) a I0 (k),
s k2
1 0 k 3 1 k 2
1,1 = 21 k 2 K0 (k) + K1 (k) K1 (k), (D.2.49)
k s k2
0 k 3 1 k2
2,1 = 21 k k K1 (k) a K0 (k),
s k 2
1122 Introduction to Theoretical and Computational Fluid Dynamics
and
1
1,2 = 2 2 k 2 I0 (k) I1 (k) , 2,2 = 2 2 k k I1 (k),
k
2
1
1,2 = 2 2 k K0 (k) + K1 (k) , 2,2 = 2 2 k k K1 (k). (D.2.50)
k
Consolidating the various terms, we derive a linear algebraic equation,
where
1 2 3 4
G1 = 1,1 , G2 = 2,1 , G3 = 1,1 , G4 = 2,1 ,
a a a a
G5 = 1,2 , G6 = 2,2 , G7 = 1,2 , G8 = 2,2 . (D.2.52)
and the coecient matrix, M, is given in Table D.2.1 [214]. When both the inner and outer cylinders
are absent and the interface is devoid of surfactants, the coecient matrix is in agreement with that
derived by Tomotika [410], as discussed in Section 9.13.2.
For a nontrivial solution to exist, the determinant of the matrix M must be zero, det(M) = 0.
To compute the growth rate, s = ikc, we eliminate the denominators from all entries of the matrix
M by multiplying corresponding rows by them, thereby obtaining a new matrix, Q. Setting the
determinant of Q to zero, we obtain a quadratic equation for the shifted growth rate,
where sr is an arbitrary reference value. In the numerical method, the binomial coecients, a, b, c,
are computed from the values P2 (sr ), P2 (sr ), P2 (sr + ), using nite-dierence formula with zero
error, where is an arbitrary oset. The growth rate is then computed using the quadratic formula,
yielding two normal modes. In the absence of surfactant, a = 0, yielding one normal mode.
Program les:
1. ann2l0
Solves the secular equation. The determinant of the matrix is calculated by LU decomposition
using Gauss elimination.
2. ann2l0 dr
Driver for ann2l0.
D.2 FDLIB User Guide: 08 stab/ann2l0 1123
I1 (k) a I0 (k) K1 (k) a K1 (k)
k I0 (k) 2 I0 (k) + k I1 (k) k K0 (k) 2 K0 (k) k K1 (k)
I (k ) ai I0 (ki ) K1 (kai ) ai K0 (ki )
1 i
k I (k ) k K0 (ki ) 2 K0 (ki ) ki K1 (ki )
M=
0 i 2 I0 (ki ) + ki I1 (ki )
0 0 0 0
0 0 0 0
F1 F2 F3 F4
G1 G2 G3 G4
I1 (k) a I0 (k) K1 (k) a K0 (k)
k I0 (k) 2 I0 (k) k I1 (k) k K0 (k) 2 K0 (k) + k K1 (k)
0 0 0 0
0 0 0 0
I1 (ke ) ae I0 (ke ) K1 (ke ) ae K0 (ke )
k K0 (ke ) 2 K0 (ke ) ke K1 (ke )
k I0 (ke ) 2 I0 (ke ) + ke I1 (ke )
F5 F6 F7 F8
G5 G6 G7 G8
Table D.2.1 Setting the determinant of the matrix displayed to zero provides us with a quadratic
equation for the complex growth rate.
3. matrix
Formulates the secular matrix, M.
4. bess I01K0
Computes the Bessel functions.
5. gel
Computes the determinant of a matrix by LU decomposition using Gauss elimination.
Input les:
1. ann2l0.dat
Problem selection and specication of input parameters.
Output les:
1. ann2l0.out
Growth rates.
1124 Introduction to Theoretical and Computational Fluid Dynamics
Fluid 1 (inner)
Fluid 2 (outer)
ae
a
ai
Figure D.3.1 Illustration of two annular layers separated by an elastic interface between an inner
cylindrical tube of radius ai and an outer cylindrical tube of radius ae .
Base state
We consider two stationary annular layers separated by a cylindrical interface of radius a, conned
between an inner cylindrical tube of radius ai and an outer cylindrical tube of radius ae , as illustrated
in Figure D.3.1. The inner uid is labeled 1 and the outer uid is labeled 2. The viscosity of the
inner uid is 1 and the viscosity of outer uid is 2 = 1 , where is the viscosity ratio. Several
special cases can be recognized:
When the inner cylinder is absent and the radius of the outer cylinder is innite, we obtain a
thread suspended in an innite ambient uid.
When the inner cylinder is absent, we obtain a core-annular arrangement consisting of an
annular layer coated on the interior surface of the inner cylinder.
When the radius of the outer cylinder is innite, we obtain an annular layer coated on the
exterior surface of the inner cylinder.
Normal-mode analysis
To carry out the normal-mode stability analysis for axisymmetric perturbations, we introduce cylin-
drical polar coordinates, (x, , ), where the x axis coincides with the axis of revolution of the
D.3 FDLIB User Guide: 08 stab/ann2lel 1125
interface. The radial position of the interface is described in terms of the real or imaginary part of
the right-hand side of the equation
is the normal-mode wave form of the perturbation, A is the complex amplitude of the interfacial
wave, k = 2/L is the wave number, L is the wavelength, and c is the complex phase velocity.
To carry out the normal-mode analysis, we express the stream function in the form
(1)
j (x, , t) = j () exp[ik(x ct)], (D.3.8)
1126 Introduction to Theoretical and Computational Fluid Dynamics
2 2
k k j = 0, (D.3.9)
d 2 d d 2 d j
where
k
kj2 k 2 i c . (D.3.10)
j
The general solution is
j () = A1,j I1 () + A2,j I1 (j ) + B1,j K1 () + B2,j K1 (j ) (D.3.11)
for j = 1, 2, where k, j kj , I1 and K1 are Bessel functions, and Ai,j , Bi,j are complex
constants. The rst and third terms on the right-hand side of (D.3.11) correspond to the rst
operator, while the second and fourth terms correspond to the second operator on the left-hand side
of (D.3.9).
Substituting the preceding expressions into (D.3.3) and using the properties of the Bessel
functions
1
I0 (z) = I1 (z), I1 (z) = I0 (z) I1 (z), K0 (z) = K1 (z),
z
1
K1 (z) = K0 (z) K1 (z), (D.3.12)
z
we nd that
where
1 dj
Uxj () = = A1,j k I0 () + A2,j kj I0 (j ) B1,j k K0 () B2,j kj K0 (j ) (D.3.14)
d
and
k
Uj () = i j = ik A1,j I1 () + A2,j I1 (j ) + B1,j K1 () + B2,j K1 (j ) . (D.3.15)
Pressure eld
The normal-mode expression for the pressure is
where P1 and P2 are the uniform unperturbed pressures, P1 P2 = 0 /a is the unperturbed capillary
pressure, and j () are pressure eigenfunctions. To compute the disturbance pressure, we resort to
the x component of the NavierStokes equation for axisymmetric ow,
Duxj pj 2u 1 uxj
xj
j = + j 2
+ ( ) . (D.3.17)
Dt x x
D.3 FDLIB User Guide: 08 stab/ann2lel 1127
Substituting (D.3.16) together with the rst equation in (D.3.13) and linearizing, we obtain
j 1 d dUxj j 1 d dUxj
2
j () = i + k U xj . (D.3.19)
k d 2 d j
and
Kinematic compatibility
Kinematic compatibility requires D[f (x, t) ]/Dt = 0 at the interface, where D/Dt is the material
derivative, and the function f describes the position of the interface according to (D.3.1). Carrying
out the dierentiation, we nd that
f f
+ ux u = 0, (D.3.23)
t x
where the velocity is evaluated at the interface. Substituting the preceding expressions and lineariz-
ing, we obtain
1 1
A=i U ( = a) = A1,j I1 (k) + A2,j I1 (kj ) + B1,j K1 (k) + B2,j K1 (kj ) . (D.3.24)
kc j c
Fluid 2
l t
n
Fluid 1
z
and
where (1) and (2) are the hydrodynamic stress tensors for the inner and outer uid, n is the unit
vector normal to the interface pointing into the inner uid labeled 1, as shown in Figure D.3.2, m
is the mean curvature of the interface, and f els is an elastic component.
It will be necessary to introduce the unit vector tangent to the interface in an azimuthal plane,
t, and the arc length l in the direction of t, as shown in Figure D.3.2. A force balance over a small
section of the interface shows that the jump in the hydrodynamic traction due to the elastic tensions
is given by
1
l
f els = (l l + ) n + (l ) t, (D.3.29)
l l
D.3 FDLIB User Guide: 08 stab/ann2lel 1129
where l and are the principal curvatures of the interface in an azimuthal and its conjugate plane,
l is the principal elastic tension in the direction of the tangential vector t, and is the principal
elastic tension in the azimuthal direction (e.g., [306], pp. 152153). All functions in equations
(D.3.28) and (D.3.29) are evaluated at the position of the perturbed interface. Consistent with the
normal-mode analysis, we express the principal elastic tensions as
Using (D.3.1) and standard expressions for the normal unit vector and curvatures, we derive
linearized forms for the normal unit vector,
principal curvatures,
1 A
l = Ak 2 exp[ik(x ct)], = + 2 exp[ik(x ct)], (D.3.32)
a a
and mean curvature,
1 1 A
2m = l + = + 2 + xx = + 2 (1 k 2 ) exp[ik(x ct)], (D.3.33)
a a a a
where ex and e are unit vectors in the axial and radial directions.
Linearizing the tangential and normal components of the interfacial force balance expressed
by (D.3.28), we obtain
u
x1 u1
u
x1 u1
x1 x2 =a
= 1 + 2 + = il exp[ik(x ct)],
x =a x =a
(D.3.34)
and
u1
u2
1 2 =a
= p1 + 2 1 p2 + 2 2 , (D.3.35)
=a =a
yielding
A
1 2 =a
= 2 (1 k2 ) exp[ik(x ct)]. (D.3.36)
a a
where the subscript R denotes the resting state. Assuming that the interface behaves like a thin
elastic sheet of an incompressible material that obeys the Mooney constitutive law (e.g., [152, 236,
262]), we write
2 2
l = E (2l + ), = E (l + 2 ), (D.3.38)
3 3
we nd that
2 2
l = E (2l + ), = E (l + 2 ). (D.3.40)
3 3
A
= . (D.3.41)
a
An evolution equation for l arises from the equation of motion of a material vector lying along the
trace of the interface in an azimuthal plane,
1 Dl
= t L t, (D.3.42)
l Dt
where D/Dt is the material derivative and L is the velocity gradient tensor. Linearizing, we obtain
1 j
l = . (D.3.43)
ca =a
Interfacial displacement
Setting the right-hand side of the rst equation in (D.3.37) equal to the right-hand side of the rst
equation in (D.3.39) and substituting (D.3.43), we obtain
l j
Integrating with respect to lR , we nd that the axial displacement of a material membrane point
which at the unstressed state was located at x, is given by
j
X1 = i , 1 = A. (D.3.48)
kca =a
The ratio X1 /1 determines the direction of displacement of the individual material points
corresponding to a normal mode.
Program les:
1. ann2lel
Computes the roots of the secular equation det(M) = 0 using Newtons method. The deter-
minant of the matrix is calculated by LU decomposition using Gauss elimination.
2. matrix
Formulates the matrix M.
3. bess I01K0
Computes the Bessel functions.
4. gel
Computes the determinant of a matrix by LU decomposition through Gauss elimination.
Input les:
1. ann2lel.dat
Problem selection and specication of input parameters.
Output les:
1. ann2lel.out
Growth rates.
1132 Introduction to Theoretical and Computational Fluid Dynamics
I1 (k) I1 (k1 ) K1 (k) K1 (k1 ) I1 (k)
k I (k) k K0 (k) k1 K0 (k1 ) k I0 (k)
0 k1 I0 (k1 )
I1 (ki ) I1 (ki1 ) K1 (ki ) K1 (ki1 ) 0
k K0 (ki ) k1 K0 (ki2 )
M = k I0 (ki ) k1 I0 (ki1 ) 0
0 0 0 0 I1 (ke )
0 0 0 0 k I0 (ke )
F1 F2 F3 F4 F5
G1 G2 G3 G4 G5
I1 (k2 ) K1 (k) K1 (k2 )
k2 I0 (k2 ) k K0 (k) k2 K0 (k2 )
0 0 0
0 0 0
I1 (ke2 ) K1 (ke ) K1 (ke2 )
k K0 (ke ) k2 K0 (ke2 )
k2 I0 (ke2 )
F6 F7 F8
G6 G7 G8
Table D.3.1 Setting the determinant of the matrix displayed to zero provides us with an algebraic
equation for the complex growth rate.
Base state
We consider the instability of an elastic interface separating two annular layers conned between
two concentric cylinders, as shown in Figure D.4.1. The inner uid is labeled 1 and the outer uid
is labeled 2. The viscosity of the inner uid is 1 and the viscosity of outer uid is 2 = 1 , where
is the viscosity ratio. In the unperturbed conguration, the uids are quiescent. Several special
cases can be recognized:
When the inner cylinder is absent and the radius of the outer cylinder is innite, we obtain a
thread suspended in an innite ambient uid.
When the inner cylinder is absent, we obtain an annular layer coated on the interior surface
of the outer cylinder.
D.4 FDLIB User Guide: 08 stab/ann2lel0 1133
Fluid 1 (inner)
Fluid 2 (outer)
ae
a
ai
Figure D.4.1 Illustration of two annular layers separated by an elastic interface between an inner
cylindrical tube with radius ai and an outer cylindrical tube with radius ae .
When the radius of the outer cylinder is innite, we obtain an annular layer coated on the
exterior surface of the inner cylinder.
Normal-mode analysis
To carry out the normal-mode stability analysis for axisymmetric perturbations, we introduce cylin-
drical polar coordinates, (x, , ), where the x axis coincides with the axis of revolution of the
unperturbed interface. Next, we describe the radial position of the interface in terms of the real or
imaginary part of the function
is the normal-mode wave form of the perturbation, A is the complex amplitude of the interfacial
wave, k = 2/L is the wave number, L is the wavelength, and c is the complex phase velocity.
2
k j = j , (D.4.9)
d 2 d
where
d2 1 d
2
k j = 0. (D.4.10)
d 2 d
The general solution is
j () = A1,j I1 () + A2,j I0 () + B1,j K1 () + B2,j K0 () , (D.4.11)
where I0 , I1 , K0 , and K1 are Bessel functions, Ai,j , Bi,j are complex constants, and k. The
rst and third terms on the right-hand side of (D.4.11) satisfy (D.4.9) with = 0. The second and
fourth terms satisfy (D.4.9) with being a nontrivial solution of (D.4.10).
Substituting the preceding expressions into (D.4.3) and using the properties of the Bessel
functions
1
I0 (z) = I1 (z), I1 (z) = I0 (z) I1 (z), K0 (z) = K1 (z),
z
1
K1 (z) = K0 (z) K1 (z), (D.4.12)
z
D.4 FDLIB User Guide: 08 stab/ann2lel0 1135
we nd that
uxj = Uxj () exp[ik(x ct)], uj = Uj () exp[ik(x ct)], (D.4.13)
where
1 dj
Uxj () = = A1,j k I0 () + A2,j [ 2 I0 () + I1 () ]
d
B1,j k K0 () + B2,j [ 2 K0 () K1 () ] (D.4.14)
and
k
Uj () = i j = ik A1,j I1 () + A2,j I0 () + B1,j K1 () + B2,j K0 () . (D.4.15)
Pressure eld
The normal-mode expression for the pressure is
pj = Pj + j () exp[ik(x ct)], (D.4.16)
where P1 and P2 are the uniform unperturbed pressures, P1 P2 = /a is the unperturbed capillary
pressure, is the surface tension, and j () are pressure eigenfunctions. To compute the disturbance
pressure eld, we consider the x component of the Stokes equation for axisymmetric ow,
pj 2u
xj 1 uxj
= j + . (D.4.17)
x x2
Substituting (D.4.16) together with the rst equation in (D.4.13) and linearizing, we obtain
j 1 d dUxj
j 1 d dUxj
j () = ij kUxj i = i k 2 Uxj , (D.4.18)
k d d k d d
yielding
j d Uxj
2
1 dUxj 2
j () = i 2
+ k Uxj = i 2 j k A2,j I0 () + B2,j K0 () . (D.4.19)
k d d
Note that only two terms contribute to the pressure eld.
A1,1 k I0 (k) + A2,1 [2 I0 (k) + k I1 (k)] B1,1 k K0 (k) + B2,1 [2 K0 (k) k K1 (k)]
= A1,2 k I0 (k) + A2,2 [2 I0 (k) + k I1 (k)] B1,2 k K0 (k) + B2,2 [2 K0 (k) k K1 (k)] (D.4.20)
and
A1,1 I1 (k) + A2,1 a I0 (k) + B1,1 K1 (k) + B2,1 a K0 (k)
= A1,2 I1 (k) + A2,2 a I0 (k) + B1,2 K1 (k) + B2,2 a K0 (k), (D.4.21)
Kinematic compatibility
Kinematic compatibility requires D[f (x, t) ]/Dt = 0 at the interface, where D/Dt is the material
derivative and the function f describes the position of the interface according to (D.4.1). Carrying
out the dierentiation, we nd that
f f
+ ux u = 0, (D.4.22)
t x
where the velocity is evaluated at the interface. Substituting the preceding expressions and lineariz-
ing, we nd
1 1
A=i U ( = a) = [ A1,j I1 (k) + A2,j a I0 (k) + B1,j K1 (k) + B2,j a K0 (k) ]. (D.4.23)
ck j c
A1,1 k I0 (ki ) + A2,1 [2 I0 (ki ) + ki I1 (ki )] B1,1 k K0 (ki ) + B2,1 [2 K0 (ki ) ki K1 (ki )] = 0,
A1,1 I1 (ki ) + A2,1 ai I0 (ki ) + B1,1 K1 (ki ) + B2,1 ai K0 (ki ) = 0, (D.4.25)
and
A1,2 k I0 (ke ) + A2,2 [2 I0 (ke ) + ke I1 (ke )] B1,2 k K0 (ke ) + B2,2 [2 K0 (ke ) ke K1 (ke )] = 0,
A1,2 I1 (ke ) + A2,2 ae I0 (ke ) + B1,2 K1 (ke ) + B2,2 ae K0 (ke ) = 0, (D.4.26)
where (1) and (2) are the hydrodynamic stress tensors in the inner and outer uid, n is the unit
vector normal to the interface pointing into the inner uid labeled 1, as shown in Figure D.4.2, is
the surface tension, m is the mean curvature of the interface, and f els is an elastic component.
It will be necessary to introduce the unit vector tangent to the interface in an azimuthal plane,
t, and the arc length l in the direction of t, as shown in Figure D.4.2. A force balance over a small
D.4 FDLIB User Guide: 08 stab/ann2lel0 1137
Fluid 2
l t
n
Fluid 1
z
section of the interface shows that the jump in the hydrodynamic traction due to the elastic tensions
is given by
l 1
f els = (l l + ) n + (l ) t, (D.4.28)
l l
where l and are the principal curvatures of the interface in a meridional and its conjugate
plane, l is the principal elastic tension in the direction of t, and is the principal elastic tension
in the azimuthal direction of increasing angle (e.g., [306], pp. 152153). All functions in equations
(D.4.27) and (D.4.28) are evaluated at the position of the perturbed interface.
Consistent with the normal-mode analysis, we express the principal elastic tensions as
l = l exp[ik(x ct)], = exp[ik(x ct)], (D.4.29)
where l and are complex coecients to be determined as part of the solution.
Using (D.4.1) and standard expressions for the normal unit vector and curvatures, we derive
linearized forms for the normal unit vector,
n = e + Aik exp[ik(x ct)] ex , (D.4.30)
principal curvatures,
1 A
l = Ak 2 exp[ik(x ct)], = + 2 exp[ik(x ct)], (D.4.31)
a a
and mean curvature,
1
1 A
2m = l + = + 2 + xx = + 2 (1 k 2 ) exp[ik(x ct)], (D.4.32)
a a a a
where ex and e are unit vectors in the axial and radial directions.
1138 Introduction to Theoretical and Computational Fluid Dynamics
Linearizing the tangential and normal components of the interfacial force balance expressed
by (D.4.27), we nd that
u
x1 u1
u
x2 u2
x1 x2 =a
= 1 + 2 + = i l exp[ik(x ct)]
x =a x =a
(D.4.33)
and
u1
u2
1 2 =a
= p1 + 2 1 p2 + 2 2 , (D.4.34)
=a =a
yielding
1
1 2 =a
= A 2 (1 k 2 ) exp[ik(x ct)]. (D.4.35)
a a
Both sides are evaluated at the location of the unperturbed interface, = a.
we nd that
2 2
l = E (2l + ), = E (l + 2 ). (D.4.39)
3 3
An evolution equation for l arises from the equation of motion of a material vector lying along the
trace of the interface in an azimuthal plane,
1 Dl
= t L t, (D.4.41)
l Dt
where D/Dt is the material derivative and L is the velocity gradient tensor. Linearizing, we nd
1 j
l = , (D.4.42)
ca =a
where j = 1 or 2.
Interfacial displacement
Setting the right-hand side of the rst equation in (D.4.36) equal to the right-hand side of the rst
equation in (D.4.38) and substituting (D.4.42), we nd
l j
We anticipate that, because of the absence of a base ow, the phase velocity c will be purely imaginary
and select the real parts of the last two equations to obtain the normal-mode parametrization
where
1 j
X1 = i , 1 = A. (D.4.47)
kca =a
The ratio X1 /1 determines the direction of displacement of the individual material points
corresponding to the normal modes.
I1 (k) a I0 (k) K1 (k) a K1 (k)
k I0 (k) 2 I0 (k) + k I1 (k) k K0 (k) 2 K0 (k) k K1 (k)
I (k ) ai I0 (ki ) K1 (ki ) ai K0 (ki )
1 i
k I (k ) k K0 (ki ) 2 K0 (ki ) ki K1 (ki )
M=
0 i 2 I0 (ki ) + ki I1 (ki )
0 0 0 0
0 0 0 0
F1 F2 F3 F4
G1 G2 G3 G4
I1 (k) a I0 (k) K1 (k) a K0 (k)
k I0 (k) 2 I0 (k) k I1 (k) k K0 (k) 2 K0 (k) + k K1 (k)
0 0 0 0
0 0 0 0
I1 (ke ) ae I0 (ke ) K1 (ke ) ae K0 (ke )
k K0 (ke ) 2 K0 (ke ) ke K1 (ke )
k I0 (ke ) 2 I0 (ke ) + ke I1 (ke )
F5 F6 F7 F8
G5 G6 G7 G8
Table D.4.1 Setting the determinant of the matrix displayed to zero provides us with an algebraic
equation for the complex growth rate.
and the coecient matrix, M, is given in Table D.4.1 [310]. The functions F1 F8 in the seventh
row and the functions G1 G8 in the eighth row are dened in the computer code. For a nontrivial
solution to exist, the determinant of the coecient matrix must be zero, det(M) = 0. Expanding
the determinant, we obtain a quadratic equation for the growth rate.
Program les:
1. ann2lel0
Computes the roots of the secular equation det(M) = 0. The determinant of the matrix is
calculated by LU decomposition using Gauss elimination.
2. matrix
Formulates the matrix M.
3. bess I01K0
Computes the Bessel functions.
4. gel
Computes the determinant of a matrix by LU decomposition using Gauss elimination.
D.5 FDLIB User Guide: 08 stab/ann2lvs0 1141
Fluid 1 (inner)
Fluid 2 (outer)
ae
a
ai
Figure D.5.1 Illustration of two annular layers separated by a viscous interface conned between an
inner cylindrical tube with radius ai and an outer cylindrical tube with radius ae .
Input les:
1. ann2lel.dat
Problem selection and specication of input parameters.
Output les:
1. ann2lel.out
Growth rates.
Base state
We consider the capillary instability of an innite, cylindrical, viscous interface separating two
concentric annular layers conned between two concentric cylinders, as shown in Figure D.5.1. The
inner uid is labeled 1 and the outer uid is labeled 2. The viscosity of the inner uid is 1 and the
viscosity of outer uid is 2 = 1 , where is the viscosity ratio. In the unperturbed conguration,
the uids are quiescent. Several special cases can be recognized:
When the inner cylinder is absent and the radius of the outer cylinder is innite, we obtain a
thread suspended in an innite ambient uid.
1142 Introduction to Theoretical and Computational Fluid Dynamics
When the inner cylinder is absent, we obtain an annular layer coated on the interior surface
of the outer cylinder.
When the radius of the outer cylinder is innite, we obtain an annular layer coated on the
exterior surface of the inner cylinder.
Normal-mode analysis
To carry out the normal-mode stability analysis for axisymmetric perturbations, we introduce cylin-
drical polar coordinates, (x, , ), where the x axis coincides with the axis of revolution of the
unperturbed interface. Next, we describe the radial position of the interface in terms of the real or
imaginary part of the function
is the normal-mode wave form of the perturbation, A is the complex amplitude of the interfacial
wave, i is the imaginary unit satisfying i2 = 1, k = 2/L is the wave number, L is the wavelength,
and c is the complex phase velocity.
2
k j = j , (D.5.9)
d2 d
where
d2 1 d
2
k j = 0. (D.5.10)
d 2 d
The general solution is
j () = A1,j I1 () + A2,j I0 () + B1,j K1 () + B2,j K0 () , (D.5.11)
where I0 , I1 , K0 , and K1 are Bessel functions, Ai,j , Bi,j are complex constants, and k. The
rst and third terms on the right-hand side of (D.5.11) satisfy (D.5.9) with j = 0. The second and
fourth terms satisfy (D.5.9) with j being a nontrivial solution of (D.5.10).
Substituting the preceding expressions in (D.5.3) and using the following properties of the
Bessel functions,
1
I0 (z) = I1 (z), I1 (z) = I0 (z)
I1 (z), K0 (z) = K1 (z),
z
1
K1 (z) = K0 (z) K1 (z), (D.5.12)
z
we nd that
where
1 dj
Uxj () = = A1,j k I0 () + A2,j [ 2 I0 () + I1 () ]
d
B1,j k K0 () + B2,j [ 2 K0 () K1 () ], (D.5.14)
and
k
Uj () = i j = ik A1,j I1 () + A2,j I0 () + B1,j K1 () + B2,j K0 () . (D.5.15)
1144 Introduction to Theoretical and Computational Fluid Dynamics
Pressure eld
The normal-mode pressure eld is
where P1 and P2 are the unperturbed uniform pressures, P1 P2 = /a is the unperturbed capillary
pressure, is the surface tension, and j () are pressure eigenfunctions. To compute the disturbance
pressure eld, we consider the x component of the Stokes equation for axisymmetric ow,
pj 2u
xj 1 uxj
= j + . (D.5.17)
x x2
Substituting (D.5.16) along with the rst equation into (D.5.13) and linearizing, we obtain
j 1 d dUxj
j 1 d dUxj
j () = ij k Uxj i = i k 2 Uxj (D.5.18)
k d d k d d
or
j d Uxj
2
1 dUxj
2
j () = i + k U x . (D.5.19)
k d 2 d j
A1,1 k I0 (k) + A2,1 [2 I0 (k) + k I1 (k)] B1,1 k K0 (k) + B2,1 [2 K0 (k) k K1 (k)]
= A1,2 k I0 (k) + A2,2 [2 I0 (k) + k I1 (k)] B1,2 k K0 (k) + B2,2 [2 K0 (k) k K1 (k)], (D.5.21)
and
Kinematic compatibility
Kinematic compatibility requires that D[f (x, t) ]/Dt = 0 at the interface, where D/Dt is the
material derivative and the function f describes the position of the interface according to (D.5.1).
Carrying out the dierentiation, we obtain
f f
+ ux u = 0, (D.5.23)
t x
D.5 FDLIB User Guide: 08 stab/ann2lvs0 1145
where the velocity is evaluated at the interface. Substituting the preceding expressions and lineariz-
ing, we nd that
1 1
A=i Uj (a) = A1,j I1 (k) + A2,j a I0 (k) + B1,j K1 (k) + B2,j a K0 (k) . (D.5.24)
kc c
A1,1 k I0 (ki ) + A2,1 [2 I0 (ki ) + ki I1 (ki )] B1,1 k K0 (ki ) + B2,1 [2 K0 (ki ) ki K1 (ki )] = 0,
A1,1 I1 (ki ) + A2,1 ai I0 (ki ) + B1,1 K1 (ki ) + B2,1 ai K0 (ki ) = 0, (D.5.26)
and
A1,2 k I0 (ke ) + A2,2 [2 I0 (ke ) + ke I1 (ke )] B1,2 k K0 (ke ) + B2,2 [2 K0 (ke ) ke K1 (ke )] = 0,
A1,2 I1 (ke ) + A2,2 ae I0 (ke ) + B1,2 K1 (ke ) + B2,2 ae K0 (ke ) = 0, (D.5.27)
where (1) and (2) are the hydrodynamic stress tensors in the inner and outer uid, n is the unit
vector normal to the interface pointing into the inner uid labeled 1, as shown in Figure D.5.2, m
is the mean curvature of the interface, and f v is a viscous component.
A force balance over a small section of the interface shows that the jump in the hydrodynamic
traction due to the interfacial tensions is given by
l 1
f v = (l l + ) n + (l ) t, (D.5.29)
l l
where l and are the principal curvatures of the interface in a meridional and its conjugate plane,
l and are the principal viscous tensions referring to orthogonal curvilinear axes corresponding
to the unit tangential vector t and to the azimuthal angle, , and l is the arc length measured in
the direction of t, as shown in Figure D.5.2 (e.g., [306], pp. 152153). All functions in equations
(D.5.28) and (D.5.29) are evaluated at the position of the perturbed interface.
1146 Introduction to Theoretical and Computational Fluid Dynamics
Fluid 2
l t
n
Fluid 1
z
Consistent with the normal-mode analysis, we express the principal viscous tensions as
l = l exp[ik(x ct)], = exp[ik(x ct)], (D.5.30)
where l and are complex coecients to be determined as part of the solution.
Using (D.5.1) and standard expressions for the normal and tangent unit vectors and curvatures,
we derive linearized forms for the normal and tangent unit vectors,
n = e + Aik exp[ik(x ct)]ex , t = ex + Aik exp[ik(x ct)]e (D.5.31)
principal curvatures,
1 A
l = Ak 2 exp[ik(x ct)], = + 2 exp[ik(x ct)], (D.5.32)
a a
and mean curvature,
1
1 A
2m = l + = + 2 + xx = + 2 (1 k 2 ) exp[ik(x ct)], (D.5.33)
a a a a
where ex and e are the unit vectors in the axial and radial directions.
To evaluate the viscous interfacial tensions, we introduce the principal extension ratios
l
l , , (D.5.34)
lR R
where the subscript R signies the resting state.
Es = (t u t) t t + (e u e ) e e , v = l t t + e e , (D.5.38)
where
l = (e s ) + 2s t u t, = (e s ) + 2s e u e , (D.5.39)
and e is the unit vector in the azimuthal direction. To interpret these expressions, we introduce
the principal extension ratios
u lR u 1 Dl ut
t u t = t= t= = + l un ,
l l lR l Dt l
1 u 1 D u
e u e = e = = , (D.5.40)
Dt
where D/Dt is the material derivative, ut u t is the tangential velocity, and un u n is the
normal velocity (e.g., [107]). The radial velocity can be expressed in terms of the tangential and
normal velocities as
x
u = ut un , (D.5.41)
l l
yielding
u ut un x ut
e u e = = = + un . (D.5.42)
l l x
The rate of surface dilatation is then
1 (ut )
= t u t + e u e = + 2m un . (D.5.43)
l
The rst term on the right-hand side expresses dilatation due to in-plane axisymmetric motion, and
the second term expresses dilatation due to normal motion.
where
1
1 = k 2 [(e + s ) I0 (k) (e s ) I1 (k)],
k
2 = k [(e + s )[2 I0 (k) + k I1 (k)] (e s ) I0 (k) ],
1
3 = k 2 [(e + s ) K0 (k) (e s ) K1 (k) ], (D.5.46)
k
4 = k [ (e + s )[2 K0 (k) k K1 (k)] (e s ) K0 (k) ],
and
where
1
1 = k 2 [(e s ) I0 (k) (e + s ) I1 (k)],
k
2 = k [(e s )[2 I0 (k) + k I1 (k)] (e + s ) I0 (k) ],
1
3 = k 2 [(e s ) K0 (k) (e + s ) K1 (k) ], (D.5.48)
k
4 = k [ (e s )[2 K0 (k) k K1 (k)] (e + s ) K0 (k) ].
x1 x2
1 + ik U1 2 + ik U2 = ikl . (D.5.50)
d =a d =a
dUxj ()
= A1,j k2 I1 () + A2,j k [2 I1 () + I0 ()]
d
+B1,j k 2 K1 () B2,j k [2 K1 () K0 ()] (D.5.51)
D.5 FDLIB User Guide: 08 stab/ann2lvs0 1149
21 k A1,1 k I1 (k) + A2,1 [I1 (k) + k I0 (k)] + B1,1 k K1 (k) B2,1 [K1 (k) k K0 (k)]
22 k A1,2 k I1 (k) + A2,2 [I1 (k) + k I0 (k)] + B1,2 k K1 (k) B2,2 [K1 (k) k K0 (k)]
= ik l = k 1 A1,1 + 2 A2,1 + 3 B1,1 + 4 B2,1 . (D.5.52)
where
F1 = 2 1 k I1 (k) + 1 , F2 = 2 1 [I1 (k) + k I0 (k)] + 2 ,
F3 = 2 1 k K1 (k) + 3 , F4 = 2 1 [K1 (k) + kK0 (k)] + 4 ,
(D.5.54)
F5 = 2 2 k I1 (k), F6 = 2 2 [I1 (k) + k I0 (k)],
F7 = 2 2 k K1 (k), F8 = 2 2 [K1 (k) + k K0 (k)].
1,1 A1,1 + 2,1 A2,1 + 1,1 B1,1 + 2,1 B2,1 + 1,2 A1,2 + 2,2 A2,2 + 1,2 B1,2 + 2,2 B2,2
1
= i = 1 A1,1 + 2 A2,1 + 3 B1,1 + 4 B2,1 , (D.5.59)
a a
1150 Introduction to Theoretical and Computational Fluid Dynamics
where
1 k 3 1 k 2
1,1 = 2 1 k 2 I0 (k) I1 (k) I1 (k),
k s k 2
k3 1 k 2
2,1 = 2 1 k k I1 (k) a I0 (k),
s k2
1 k 3 1 k2
1,1 = 2 1 k 2 K0 (k) + K1 (k) K1 (k), (D.5.60)
k s k2
k3 1 k 2
2,1 = 2 1 k k K1 (k) a K0 (k),
s k2
and
1
1,2 = 2 2 k 2 I0 (k) I1 (k) , 2,2 = 2 2 k k I1 (k),
k
1
1,2 = 2 2 k 2 K0 (k) + K1 (k) , 2,2 = 2 2 k k K1 (k). (D.5.61)
k
Consolidating the various terms, we obtain a linear algebraic equation,
G1 A1,1 + G2 A2,1 + G3 B1,1 + G4 B2,1 + G5 A1,2 + G6 A2,2 + G7 B1,2 + G8 B2,2 = 0, (D.5.62)
where
1 2 3 4
G1 = 1,1 , G2 = 2,1 , G3 = 1,1 , G4 = 2,1 ,
a a a a (D.5.63)
G5 = 1,2 , G6 = 2,2 , G7 = 1,2 , G8 = 2,2 .
I1 (k) a I0 (k) K1 (k) a K1 (k)
k K0 (k) 2 K0 (k) k K1 (k)
k I0 (k) 2 I0 (k) + k I1 (k)
I1 (ki ) ai I0 (ki ) K1 (ki ) a K1 (k)
k I0 (k) 2 I0 (k) + k I1 (k) k K0 (k) 2 K0 (k) k K1 (k)
I1 (ki ) ai I0 (ki ) K1 (ki ) ai K0 (ki )
M=
k I0 (ki ) 2 I0 (ki ) + ki I1 (ki ) k K0 (ki ) 2 K0 (ki ) ki K1 (ki )
0 0 0 0
0 0 0 0
F1 F2 F3 F4
G1 G2 G3 G4
I1 (k) a I0 (k) K1 (k) a K0 (k)
k I0 (k) 2 I0 (k) k I1 (k) k K0 (k) 2 K0 (k) + k K1 (k)
0 0 0 0
0 0 0 0
I1 (ke ) ae I0 (ke ) K1 (ke ) ae K0 (ke )
k K0 (ke ) 2 K0 (ke ) ke K1 (ke )
k I0 (ke ) 2 I0 (ke ) + ke I1 (ke )
F5 F6 F7 F8
G5 G6 G7 G8
Table D.5.1 Setting the determinant of the matrix displayed to zero provides us with an algebraic
equation for the complex growth rate.
Suspended thread
In the simplest conguration, both the internal and external cylinder are absent, yielding an innite
thread suspended in an innite ambient uid. The stream functions for the internal and external
ow are described, respectively, by equations (D.5.65) and (D.5.66). The linear system M q = 0
undergoes analogous simplications. In the absence of interfacial viscosity, the coecient matrix
reduces to that derived by Tomotika [410].
1152 Introduction to Theoretical and Computational Fluid Dynamics
Program les:
1. ann2lvs0
Computes the roots of the secular equation. The determinant of the matrix is calculated by
LU decomposition using Gauss elimination.
2. matrix
Formulates the matrix M.
3. bess I01K0
Computes the Bessel functions.
4. gel
Computes the determinant of a matrix by LU decomposition using Gauss elimination.
Input les:
1. ann2lvs.dat
Problem selection and specication of input parameters.
Output les:
1. ann2lvs.out
Growth rates.
Base state
In the unperturbed conguration, the interface is at and the uids execute unidirectional motion
parallel to the channel walls. In the inclined system of coordinates depicted in Figure D.6.1, the
unperturbed interface is located at y = 0, the lower wall is located at y = h1 , and the upper wall
is located at y = h2 = 2h h1 , where 2h is the channel width. For convenience, the x and y velocity
components are denoted by u = ux and v = uy .
(0) + j gx 2 (0)
uj = y + j y + u I , vj =0 (D.6.1)
2j
for j = 1, 2, corresponding to the lower or upper uid, where the superscript (0) denotes the
unperturbed base state, subject to the following denitions: is the negative of the axial pressure
gradient; gx = g sin is the x component of the acceleration of gravity; g is the magnitude of the
D.6 FDLIB User Guide: 08 stab/chan2l0 1153
g 2h
h2
h1 Fluid 2
Fluid 1
Figure D.6.1 Two-layer ow in an inclined channel conned between two parallel plane walls inclined
at angle .
(0)
acceleration of gravity; j are the uid densities; j are the uid viscosities; uI uj (y = 0) is the
common interfacial velocity. The coecients
u(0)
u(0)
x1 x2
1 , 2 , (D.6.2)
y y=0 y y=0
are the interfacial shear rates of the unperturbed ow. Continuity of shear stress across the interface
requires that 1 1 = 2 2 or
1 = 2 , (D.6.3)
U h r2 r2 1
1 = + 1 gx , 2 = (D.6.5)
h1 + r 1 r2 (1 + r)( + r)
1154 Introduction to Theoretical and Computational Fluid Dynamics
and
r U1 + U 2 h2 1 + r 2r
uI = + + 1 g x , (D.6.6)
+r 1 1 + r (1 + r)( + r)
where U = U2 U1 is the dierence in the wall velocities, r = h2 /h1 is the layer thickness ratio,
and = 2 /1 is the density ratio. When the densities of the two liquids are matched, = 1 and
1 = 2 = , the two terms enclosed by the large parentheses on the right-hand side of (D.6.5) and
(D.6.6) combine into an eective negative pressure gradient, + gx .
Normal-mode analysis
A normal-mode perturbation displaces the interface to a position given by the real or imaginary
part of the function
is the normal-mode wave form of the perturbation, A is the complex amplitude of the interfacial
wave, k = 2/L is the wave number, L is the wavelength, and c is the complex phase velocity.
Stream function
The stream function, , is dened by the equations u = /y and v = /x. In the linear
(0) (1)
analysis, we set j = j + j and introduce the normal-mode form
(1)
j (x, y, t) = j (y) exp[ik(x ct)] (D.6.9)
for j = 1, 2, where y = ky and the superscript (1) denotes the disturbance. The perturbation
velocity components are u(1) = (1) /y and v (1) = (1) /x.
The Reynolds number of the ow is so small that the motion of the uid is governed by
the equations of Stokes ow. Requiring that the stream function satises the biharmonic equation,
(1)
4 j = 0, we nd that
j (y) = a1j cosh y + a2j y cosh y + a3j sinh y + a4j y sinh y, (D.6.10)
Pressure eld
(0) (1)
The linearized pressure eld in the jth uid is pj = pj + pj . The disturbance pressure eld is
expressed by the normal-mode form
(1)
pj (x, y, t) = j qj (y) exp[ik(x ct)] (D.6.11)
D.6 FDLIB User Guide: 08 stab/chan2l0 1155
for j = 1, 2, where qj (y) are pressure eigenfunctions. Substituting this expression into the x compo-
nent of the Stokes equation, we nd that
(1) i 3 j 3 j
pj (y) = 2
+ , (D.6.12)
k x y y 3
yielding
d d3 j
qj (y) = i k2
j
. (D.6.13)
dy dy 3
Kinematic compatibility
Kinematic compatibility requires that D[f (x, t) y]/Dt = 0, where D/Dt is the material derivative
and the function f (x, t) describe the shape of the interface according to (D.6.7). In the linear
approximation,
+ uI v (1) (y = 0) = 0, (D.6.15)
t x
where v (1) = (1) /x. Substituting the preceding expressions, we nd that ikcA + uI ikA +
ika11 = 0, yielding a11 = (c uI ) A or
A = a11 , (D.6.16)
where 1/(c uI ).
1 2
1 (x, y) + = 2 (x, y) + , (D.6.17)
y y=0 y y=0
yielding
d
d
1 2
1 A + k = 2 A + k . (D.6.18)
dy y=0 dy y=0
1 2
(1 2 ) a11 + k k = 0, (D.6.19)
dy y=0 dy y=0
1156 Introduction to Theoretical and Computational Fluid Dynamics
yielding
1 (y = k1 ) = 0, 1 (y = k1 ) = 0, 2 (y = k2 ) = 0, 2 (y = k2 ) = 0, (D.6.21)
and
! "
cosh k2 k2 cosh k2 sinh k2 k2 sinh k2
w2 = 0, (D.6.23)
sinh k2
cosh k2 + k2 sinh k2 cosh k2 sinh k2 + k2 cosh k2
where w1 = a11 , a21 , a31 , a41 and w2 = a12 , a22 , a32 , a42 .
(1) u(1) v
(1)
1 2
1 + = 2 + 2 + gx A, (D.6.24)
y x y=0 y x y=0
where = 1 2 . Expressing the velocity in terms of the stream function and using (D.6.16) to
eliminate the interfacial amplitude, A, we nd that
2 (1) 2 1
(1) 2 (1) 2 2
(1)
gx
1 2
= + a11 , (D.6.25)
y 2 x2 y=0 y2 x2 y=0 1
yielding
d2 1
d2
gx
2
( 2
+ 1 = 2
+ 2 + a11 . (D.6.26)
dy y=0 dy y=0 1 k 2
Substituting the preceding expressions for j and simplifying, we nd that
where
gx
Bx . (D.6.28)
21 k 2
D.6 FDLIB User Guide: 08 stab/chan2l0 1157
(1) v
(1)
2
(1) (1)
p1 + 21 p2 + 22 2 gy (x) = 2 , (D.6.30)
y y=0 y y=0 x
where is the surface tension. Substituting the preceding expressions, we obtain
d1 d2
where
1 gy
+ = By + (D.6.35)
21 k2 21
is a property group with dimensions of velocity, and
gy
By . (D.6.36)
21 k 2
Finally, we substitute (D.6.29), (D.6.34), and (D.6.37) into the linear system (D.6.22) and obtain
the equivalent system
!
C11 C12 k1 cosh k1 sinh k1 (1 ) k1 cosh k1
C21 + C22 cosh k1 + k1 sinh k1 cosh k1 + (1 ) k1 sinh k1
"
k1 sinh k1
w2 = 0, (D.6.38)
(sinh k1 + k1 cosh k1 )
1158 Introduction to Theoretical and Computational Fluid Dynamics
where
Setting the determinant of Q to zero provides us with a secular equation for the computation of the
complex phase velocity, c.
Performing the Laplace expansion with respect to the rst column and rearranging, we obtain
Separating the complex phase velocity into its real and imaginary parts, we obtain the real
phase velocity
N
cR = u I + (D.6.42)
D
where
and
cosh k2 + k2 sinh k2 cosh k2
M11 = det( k1 cosh k1 sinh k1 + ( 1) k1 cosh k1
cosh k1 + k1 sinh k1 cosh k1 ( 1) k1 sinh k1
sinh k2 + k2 cosh k2
k1 sinh k1 )
(sinh k1 + k1 cosh k1 )
k2 cosh k2 sinh k2
M21 = det( k1 cosh k1 sinh k1 + ( 1) k1 cosh k1
cosh k1 + k1 sinh k1 cosh k1 ( 1) k1 sinh k1
k2 sinh k2
k1 sinh k1 )
(sinh k1 + k1 cosh k1 )
k2 cosh k2 sinh k2
M31
= det( cosh k2 + k2 sinh k2 cosh k2
cosh k1 + k1 sinh k1 cosh k1 ( 1) k1 sinh k1
k2 sinh k2
sinh k2 + k2 cosh k2 )
(sinh k1 + k1 cosh k1 )
k2 cosh k2 sinh k2
M41 = det( cosh k2 + k2 sinh k2 cosh k2
k1 cosh k1 sinh k1 + ( 1) k1 cosh k1
k2 sinh k2
sinh k2 + k2 cosh k2 )
k1 sinh k1
Table D.6.1 Matrices involved in the secular equation determining hydrodynamic stability. The ver-
tical lines separate the three matrix columns.
1160 Introduction to Theoretical and Computational Fluid Dynamics
Program les:
1. chan2l0
Evaluates the complex phase velocity.
2. chan2l0 dr
Driver for the subroutine chan2l0
3. det 33c
Determinant of a 3 3 complex matrix.
4. det 44c
Determinant of a 4 4 complex matrix.
Input le:
1. chan2l0.dat
Specication of input parameters.
Output le:
1. chan2l0.out
Recording of the computed phase velocity and growth rate.
Base state
In the inclined system of coordinates depicted in Figure D.7.1, the unperturbed interface is located
at y = 0, the lower wall is located at y = h1 and the upper wall is located at y = h2 = 2h h1 ,
where 2h is the channel width. The surfactant is distributed uniformly over the interface. For
convenience, the x and y velocity components are denoted by u = ux and v = uy .
g 2h
h2
h1 Fluid 2
Fluid 1
Figure D.7.1 Two-layer ow in an inclined channel conned between two parallel walls. The interface
is occupied by an insoluble surfactant.
for j = 1, 2, corresponding to the lower and upper uid, where the superscript (0) denotes the
unperturbed base state, subject to the following denitions: is the negative of the axial pressure
gradient; gx = g sin is the x component of the acceleration of gravity; g is the magnitude of the
(0)
acceleration of gravity; j are the uid densities; j are the uid viscosities; uI uj (y = 0) is the
common interfacial velocity. The coecients
u(0)
u(0)
x1 x2
1 , 2 , (D.7.2)
y y=0 y y=0
are the interfacial shear rates of the unperturbed ow. Continuity of shear stress across the interface
requires that 1 1 = 2 2 or
1 = 2 , (D.7.3)
and
r U1 + U 2 h2 1+r 2r
uI = + + 1 gx , (D.7.6)
+r 1 1 + r (1 + r)( + r)
where U = U2 U1 is the dierence in the wall velocities, r = h2 /h1 is the layer thickness ratio,
and = 2 /1 is the density ratio. When the densities of the two liquids are matched, = 1 and
1 = 2 = , the two terms enclosed by the large parentheses on the right-hand side of (D.7.5) and
(D.7.6) combine into an eective negative pressure gradient, + gx .
Normal-mode analysis
A normal-mode perturbation displaces the interface to a position given by the real or imaginary
part of the function
is the normal-mode wave form of the perturbation, A is the complex amplitude of the interfacial
wave, k = 2/L is the wave number, L is the wavelength, and c is the complex phase velocity.
Stream function
The stream function, , is dened by the equations u = /y and v = /x. In the linear
(0) (1)
analysis, we set j = j + j and introduce the normal-mode form
(1)
j (x, y, t) = j (y) exp[ik(x ct)], (D.7.9)
where y = ky and the superscript (1) denotes the perturbation. The perturbation velocity compo-
nents are u(1) = (1) /y and v (1) = (1) /x.
The Reynolds number of the ow is so small that the motion of the uid is governed by
the equations of Stokes ow. Requiring that the stream function satises the biharmonic equation,
(1)
4 j = 0, we nd that
j (y) = a1j cosh y + a2j y cosh y + a3j sinh y + a4j y sinh y, (D.7.10)
Pressure eld
(0) (1)
The linearized pressure eld in the jth uid is pj = pj + pj . The disturbance pressure eld is
expressed by the normal-mode form
(1)
pj (x, y, t) = j qj (y) exp[ik(x ct)] (D.7.11)
D.7 FDLIB User Guide: 08 stab/chan2l0 s 1163
for j = 1, 2, where qj (y) are pressure eigenfunctions. Substituting this expression into the x compo-
nent of the Stokes equation, we nd that
(1) i 3 j 3 j
pj (y) = 2
+ , (D.7.12)
k x y y 3
yielding
d d3 j
qj (y) = i k2
j
. (D.7.13)
dy dy 3
Kinematic compatibility
Kinematic compatibility requires that D[f (x, t) y]/Dt = 0, where D/Dt is the material derivative.
In the linear approximation,
+ uI v (1) = 0, (D.7.15)
t x
where the velocity is evaluated at y = 0. Substituting the preceding expressions, we nd that
ikcA + uI ikA + ika11 = 0, yielding a11 = (c uI ) A or
A = a11 , (D.7.16)
where 1/(c uI ) is the inverse of the complex phase velocity shifted by the real interfacial
velocity.
1 2
1 (x, y) + = 2 (x, y) + , (D.7.17)
y y=0 y y=0
yielding
d
d
1 2
1 A + k = 2 A + k . (D.7.18)
dy y=0 dy y=0
1 2
(1 2 ) a11 + k k = 0, (D.7.19)
dy y=0 dy y=0
1164 Introduction to Theoretical and Computational Fluid Dynamics
yielding
Wall conditions
The no-slip and no-penetration conditions at the lower and upper wall require that
1 (y = k1 ) = 0, 1 (y = k1 ) = 0, 2 (y = k2 ) = 0, 2 (y = k2 ) = 0, (D.7.21)
and
! "
cosh k2 k2 cosh k2 sinh k2 k2 sinh k2
w2 = 0, (D.7.23)
sinh k2
cosh k2 + k2 sinh k2 cosh k2 sinh k2 + k2 cosh k2
where w1 = a11 , a21 , a31 , a41 and w2 = a12 , a22 , a32 , a42 .
where 0 , 0 are uniform values corresponding to the at interface, and 1 , 1 are complex ampli-
tudes. Since the perturbations are small, we can adopt the linear constitutive equation
1 1
= Ma , (D.7.25)
0 0
where Ma is the Marangoni number.
Surfactant transport
Surfactant transport over the interface is governed by the equation
D u
+ t = Ds , (D.7.26)
Dt l l l
where D/Dt is the material derivative. Regarding as a function of x and t, we obtain
u u
x
+u + t + = Ds . (D.7.27)
t x x y x l l l
D.7 FDLIB User Guide: 08 stab/chan2l0 s 1165
(1) u(1) v
(1)
(1)
1 2
1 + = 2 + 2 + gx A + , (D.7.31)
y x y=0 y x y=0 x
where (1) (x, t) = 1 exp[ik(xct)] and = 1 2 . Expressing the velocity in terms of the stream
function and using (D.7.16) to eliminate A in favor of a11 , we obtain
2 (1) 2 1
2 (1) 2 2
(1) (1)
1 2 gx 1 (1)
= + a11 + , (D.7.32)
y 2 x 2 y=0 y 2 x 2 y=0 1 1 x
yielding
d2
d2
gx 1
1 2
+ 1 = + 2 + a11 + i . (D.7.33)
dy 2 y=0 dy 2 y=0 1 k 2 k1
Substituting the preceding expressions for j and simplifying, we obtain
1
(1 Bx ) a11 + a41 (a12 + a42 ) = i , (D.7.34)
21 k
where
gx
Bx . (D.7.35)
21 k 2
We recall that a11 = a12 and 1 /0 = Ma 1 /0 , and use (D.7.30) to obtain
where
0 i
Ma (D.7.37)
21 1 + Ds k i
v1
(1) v
(1)
2
(1) (1)
p1 + 21 p2 + 22 2 gy (x) = 0 2 . (D.7.38)
y y=0 y y=0 x
d1 d2
1 q1 + 2 q2 2ik2 1 + 2ik 2 2 gy A = 0 A k 2 , (D.7.39)
dy dy
where
1 gy 0
2
+ 0 = By + (D.7.43)
21 k 21
gy
By . (D.7.44)
21 k 2
Finally, we substitute (D.7.36), (D.7.42), and (D.7.45) into the linear system (D.7.22) and derive
the equivalent system A w2 = 0, where
!
C11 (C12 C13 ) k1 ( sinh k1 cosh k1 )
A=
C21 + (C22 + C23 ) (1 k1 ) cosh k1 + (k1 ) sinh k1
"
(1 ) k1 cosh k1 + ( k1 ) sinh k1 k1 sinh k1
, (D.7.46)
(1 k1 ) cosh k1 + [(1 ) k1 ] sinh k1 (sinh k1 + k1 cosh k1 )
and
C13 = 2 k1 sinh k1 ,
C21 = sinh k1 + (1 )(sinh k1 + k1 cosh k1 ),
C22 = (2 1 ) (cosh k1 + k1 sinh k1 ) Bx (sinh k1 + k1 cosh k1 ) + i k1 sinh k1 ,
In the absence of surfactant, = 0. Setting the determinant of Q to zero provides us with a secular
equation for the complex phase velocity, c.
Numerical method
The growth rates of the two normal modes are identied with the roots of a quadratic equation with
complex coecients,
The factor 1 + Ds ki on the left-hand side involving the surface surfactant diusivity arises from
the denition of the parameter . In the numerical method, the binomial coecients are obtained
1168 Introduction to Theoretical and Computational Fluid Dynamics
Program les:
1. chan2l0 s
Evaluates the complex phase velocity.
2. chan2l0 s dr
Driver for the subroutine chan2l0 s
3. det 33c
Determinant of a 3 3 complex matrix.
4. det 44c
Determinant of a 4 4 complex matrix.
5. quadc
Computes the roots of a quadratic equation with complex coecients.
Input le:
1. chan2l0 s.dat
Specication of input parameters.
Output le:
1. chan2l0 s.out
Recording of the computed phase velocity and growth rate.
Program les:
1. coat0 s
Evaluates the growth rate.
2. cramer 33c
Solves a system of three complex equations.
D.9 FDLIB User Guide: 08 stab/drop ax 1169
3. det 33c
Computes the determinant of a 3 3 complex matrix.
4. det 44c
Computes the determinant of a 4 4 complex matrix.
5. matrix
Evaluates the secular matrix.
Input le:
1. coat0 s.dat
Specication of input parameters.
Output le:
1. coat0 s.out
Recording of the computed phase velocity and growth rate.
Program les:
1. drop ax
Evaluates the rate of relaxation by discrete perturbations.
2. drop slp splines
Evaluates the single-layer potential on cubic splines.
3. ell int
Evaluates complete elliptic integrals.
4. gauss leg
GaussLegendre base points and weights.
5. sgf ax fs
Free-space Greens function of axisymmetric Stokes ow.
6. splc clm
Cubic-spline interpolation with clamped boundary conditions.
7. splc geo
Geometry of interpolated interface.
8. thomas
Thomas algorithm for tridiagonal systems.
1170 Introduction to Theoretical and Computational Fluid Dynamics
Program les:
1. det 33c
Determinant of a complex 3 3 matrix.
2. film0
Evaluates the complex phase velocity.
3. film0 dr
Driver for film0.
4. matrix
Evaluates the secular matrix.
5. cramer 33c
Solution of a complex 3 3 linear system.
Input le:
1. film0.dat
Specication of input parameters.
Output le:
1. film0.out
Recording of the computed growth rate.
Base state
The ow is described in the inclined coordinates dened in Figure D.11.1. The unperturbed cong-
uration is described by the at-lm Nusselt solution designated by the superscript (0). The velocity,
stream function, and pressure are given by
g g 1
u(0)
x = sin y(2h y), u(0)
y = 0, (0) = sin y2 (h y),
2 2 3
p(0) = g cos (h y) + pa , (D.11.1)
D.11 FDLIB User Guide: 08 stab/film0 s 1171
Air
Liquid
Figure D.11.1 Instability of a liquid lm owing down an inclined plane. The surface of the lm is
occupied by an insoluble surfactant.
where g = |g| is the magnitude of the acceleration of gravity, is the plane inclination angle, and
pa is the ambient atmospheric pressure. The free-surface velocity is
gh2
us = u(0)
x (y = h) = sin . (D.11.2)
2
Normal-mode analysis
To carry out the normal-mode stability analysis for two-dimensional perturbations, we describe the
position of the lm surface by the function
y = f (x, t) = h + (x, t), (D.11.3)
where is a dimensionless coecient, whose magnitude is small compared to unity,
(x, t) = A exp[ik(x ct)] (D.11.4)
is the normal-mode wave form of the perturbation, A is the complex amplitude of the interfacial
wave, k = 2/L is the wave number, L is the wavelength, and c is the complex phase velocity.
Stream function
The stream function, , is dened by the equations u = /y and v = /x. In the linear
analysis, we set = (0) + (1) and introduce the normal-mode form
(1) (x, y, t) = (y) exp[ik(x ct)], (D.11.5)
where the superscript (1) denotes the perturbation and y ky.
The Reynolds number of the ow is so small that the motion of the uid is governed by
the equations of Stokes ow. Requiring that the stream function satises the biharmonic equation,
(1)
4 j = 0, we obtain
(y) = a1 ey + a2 y ey + a3 ey + a4 y ey , (D.11.6)
where a1 a4 are four complex coecients.
1172 Introduction to Theoretical and Computational Fluid Dynamics
Wall conditions
The no-penetration and no-slip boundary conditions over the wall require that (0) = 0 and (0) =
0. Making substitutions, we obtain
a3 = a1 , 2a1 + a2 + a4 = 0. (D.11.7)
Kinematic compatibility
Kinematic compatibility requires that D[f (x, t) y]/Dt = 0, where D/Dt is the material derivative,
yielding
f f
+ ux uy = 0, (D.11.8)
t x
evaluated at the lm surface. Linearizing, we obtain
(1)
+ us + = 0. (D.11.9)
t x x
Substituting the preceding expressions, we nd that iAkc + us ikA + ik(k) = 0, where k = kh.
Rearranging, we obtain
A = (k), (D.11.10)
where = 1/(c us ) is the inverse of the complex phase velocity shifted by the real interfacial
velocity.
where 0 , 0 are uniform values corresponding to the at lm, and 1 , 1 are complex amplitudes.
Since the perturbations are small, we can write
1 1
= Ma , (D.11.12)
0 0
where Ma is the Marangoni number.
Surfactant transport
The linearized form of the surfactant transport equation is
(0)
2 (1)
+ u(0) (0) x
+ + = D s , (D.11.13)
t x
x x y dx x2
D.11 FDLIB User Guide: 08 stab/film0 s 1173
evaluated at the unperturbed interface, y = h. The second term inside the parentheses on the left-
(0) (0) (0)
hand side arises from the derivative ux /l (ux /y)(df /dx) = (ux /y)(d/dx). However,
because the shear stress and thus slope of the Nusselt velocity prole vanishes at the lm surface,
this term does not make a contribution. In terms of the stream function,
where a prime denotes a derivative with respect to k. Correspondingly, the complex amplitude of
the surface tension is
1 Ma
=k (k). (D.11.16)
0 c us + ikDs
Evaluating the derivative and remembering that a3 = a1 , we obtain
1 Ma
= k ek a1 (1 + q) + a2 (1 + k) + a4 (1 k) q , (D.11.17)
0 c us + ikDs
where q = exp(2k).
(1)
x us (1)
+ = 2 2 + . (D.11.18)
y x y=h h x
The last term on the right-hand side represents the Marangoni traction. Expressing the velocity in
terms of the stream function and rearranging, we obtain
2 (1) (1)
us 1 (1)
=2 2 + , (D.11.19)
y2 x 2 y=h h x
Substituting the normal-mode forms and using (D.11.10), we obtain
d2
h
k 2 + = 2 us (k) + i k 1 . (D.11.20)
dy 2 y=k
Computing the derivatives, recalling that a3 = a1 , and rearranging, we nd that
1q 1 k
k (1 q) a1 + (k + 1) a2 + (k 1) q a4 = a1 + a2 + q a4 + i h e , (D.11.21)
k 2
Dierentiating (D.11.22) with respect to x and using the tangential projection of the equation
of motion to evaluate the pressure derivative on the left-hand side, we obtain the preferred pressure-
free form
(1)
p(1) 2 uy 3
= 2 u(1)
x = 2 + g cos 0 3 , (D.11.23)
x xy x x
where all terms are evaluated at the unperturbed lm surface, y = h. Introducing the stream
function, we obtain
3 (1) 3 (1)
3
3 2 + = g cos 0 . (D.11.24)
x y y 3 x x3
Substituting the normal-mode forms and using (D.11.10), we obtain
g
3 (k) + (k) = i ( 2 cos + 0 ) (k), (D.11.25)
k
where a prime denotes a derivative with respect to y. Evaluating the derivatives, remembering that
a3 = a1 , and setting = 1/(us ), we nd that
1 g
[ (1 + q) a1 + a2 k a4 k q ] = i ( 2 cos + 0 ) a1 (1 q) + a2 k + a4 k q , (D.11.26)
2us k
where q = exp(2k).
where q = exp(2k),
1 2 2 us L
2 1 gL2
(gh2 cos + 0 k2 ) = cot + , Ca = sin , (D.11.28)
2us Ca 0 h 2 0
are a dimensionless group and a capillary number dened with respect to the wavelength of the
perturbation. Setting the determinant of M to zero provides us with a third-order algebraic equation
for the reduced and shifted complex phase velocity, . One trivial root is = 0 corresponding to
c = us . The other two roots can be computed in terms of the coecients of the remainder binomial
using the quadratic formula. In practice, the coecients are extracted by solving a system of complex
linear equations for three trial values of . Thus, in the presence of surfactant, the ow admits two
normal modes.
Absence of surfactant
In the absence of surfactant, Ma = 0, we set 1 = 0 and retain the rst three equations in (D.11.27).
Yih [435] derived analytical expressions for the phase velocity cR and rate of decay of surface waves
sI = kcI , where the subscripts R and I denote the real and imaginary part. The dimensionless
phase velocity and dimensionless growth rate are given by
cR 1 sI h cI k sinh(2k) 2k
=1+ 2
, s = = . (D.11.29)
us cosh k + k 2 us us 2k cosh2 k + k 2
Since the fraction on the right-hand side of the expression for the growth rate is positive for any k,
the growth rate is negative and the ow is stable.
Program les:
1. cramer 33c
Solves a system of three complex equations.
2. det 33c
Computes the determinant of a 3 3 complex matrix.
3. det 44c
Computes the determinant of a 4 4 complex matrix.
4. film0 s
Evaluates the growth rate.
5. matrix
Evaluates the secular matrix.
Input le:
1. film0 s.dat
Specication of input parameters.
Output le:
1. film0 s.out
Recording of the computed phase velocity and growth rate.
1176 Introduction to Theoretical and Computational Fluid Dynamics
Fluid 2
x
g
L
Fluid 1
Figure D.12.1 Instability of the interface between two semi-innite uids in Stokes ow.
Base state
The lower uid is labeled 1 and the upper uid is labeled 2. The viscosity of the lower uid is 1 and
the viscosity of the upper uid is 2 = 1 , where is the viscosity ratio. The origin of the y axis is
set at the location of the unperturbed interface. For convenience, the x and y velocity components
are denoted by u = ux and v = uy .
In the unperturbed conguration, the interface is at, the uids are quiescent, and the pressure
assumes the hydrostatic distribution
(0)
pj (y) = j gy + p0 , (D.12.1)
where 1, 2 for the lower or upper uid, j are the uid densities, g is the acceleration of gravity, p0
is an unspecied reference pressure, and the superscript (0) denotes the unperturbed base state.
Normal-mode analysis
A normal-mode perturbation displaces the interface to a position given by the real or imaginary
part of the function
is the normal-mode wave form of the perturbation, A is the complex amplitude of the interfacial
wave, k = 2/L is the wave number, L is the wavelength, and c is the complex phase velocity of the
perturbation.
D.12 FDLIB User Guide: 08 stab/if0 1177
Stream function
The stream function, , is dened by the equations u = /y and v = /x. In the absence of
(1)
base ow, we set j = j and introduce the normal-mode form
(1)
j (x, y, t) = j (y) exp[ik(x ct)], (D.12.4)
where j = 1, 2 for the lower or upper uid, and y = ky. The perturbation velocity components are
u(1) = (1) /y and v (1) = (1) /x.
The Reynolds number of the ow is so small that the motion of the uid is governed by
the equations of Stokes ow. Requiring that the stream function satises the biharmonic equation,
(1)
4 j = 0, we obtain
Pressure eld
(0) (1)
The linearized pressure eld in the jth uid is pj = pj + pj . The perturbation pressure is
described by the normal-mode form
(1)
pj (x, t) = j qj (y) exp[ik(x ct)] (D.12.6)
for j = 1, 2, where qj (y) are pressure eigenfunctions. Substituting this expression into the x compo-
nent of the Stokes equation, we obtain
i 3 j 3 j
(1) (1)
(1)
pj = + , (D.12.7)
k x2 y y 3
yielding
d d3 j
qj (y) = i k2
j
. (D.12.8)
dy dy 3
yielding a1 = c2 .
1 2
= , (D.12.10)
y y=0 y y=0
1178 Introduction to Theoretical and Computational Fluid Dynamics
yielding
a1 + b1 = c2 + d2 . (D.12.11)
Kinematic compatibility
Kinematic compatibility requires that D[f (x, t) y]/Dt = 0 at the interface, where D/Dt is the
material derivative. In the linearized approximation,
v (1) (y = 0) = 0. (D.12.12)
t
Substituting the preceding expressions, we nd that ikcA + ika1 = 0 or
a1 = cA. (D.12.13)
1 2
1 + = 2 + (D.12.14)
y x y=0 y x y=0
or
2 2 1
2 2 2
1 2
= , (D.12.15)
y 2 x2 y=0 y 2 x2 y=0
where = 2 /1 is the viscosity ratio. Substituting the preceding expressions for the stream function
and simplifying, we nd that
a1 + b1 = (c2 d2 ). (D.12.16)
d2 = c2 = a1 , b1 = a1 . (D.12.17)
(1) v
(1)
2
(1) (1)
p1 + 21 p2 + 22 2 + g = 2 , (D.12.18)
y y=0 y y=0 x
where is the surface tension and = 1 2 . Substituting the preceding expressions, we obtain
d1
1 q1 + 2 q2 2ik 2 (1 2 ) + gA = A k 2 , (D.12.19)
dy
D.13 FDLIB User Guide: 08 stab/ifsf0 s 1179
q1 (y = 0) = 2i k 2 b1 , q2 (y = 0) = 2i k 2 d2 , (D.12.20)
we obtain
i g
a1 = 2
+ A. (D.12.22)
21 (1 + ) k
Growth rate
Substituting expression (D.12.13) into (D.12.22), we obtain an imaginary phase velocity,
i g
c= 2
+ . (D.12.23)
21 (1 + ) k
1 g
I kcI = + k . (D.12.24)
21 (1 + ) k
When the uids are unstably stratied, 2 > 1 or < 0, the rst term inside the large parentheses
on the right-hand side is responsible for the RayleighTaylor instability. The second term expresses
the stabilizing inuence of surface tension.
Program les:
1. if0
Evaluates expression (D.12.24).
2. if0 dr
Driver for if0.
Input le:
1. if0.dat
Specication of input parameters.
Output le:
1. if0.out
Recording of the growth rate.
1180 Introduction to Theoretical and Computational Fluid Dynamics
Fluid 2
L
Fluid 1
Figure D.13.1 Instability of the interface between two semi-innite uids in simple shear ow. The
interface is occupied by an insoluble surfactant.
This directory contains a code that performs the linear stability analysis of the horizontal interface
between two semi-innite uids undergoing simple shear ow, as illustrated in Figure D.13.1. The
interface is occupied by an insoluble surfactant and the uid motion occurs under conditions of
Stokes ow.
Base state
The lower uid is labeled 1 and the upper uid is labeled 2. The viscosity of the lower uid is 1
and the viscosity of the upper uid is 2 = 1 , where is the viscosity ratio. The origin of the
y axis is set at the unperturbed interface. For convenience, the x and y velocity components are
denoted as u = ux and v = uy .
In the unperturbed conguration, the interface is at and the uids undergo simple shear ow
with a linear velocity prole parallel to the interface. The velocity proles in the lower and upper
uids are given by
(0) (0)
u1 = y, u2 = y, (D.13.1)
where is the shear rate in the upper uid and the superscript (0) denotes the base state. The
corresponding pressure distributions are
(0)
pj (y) = j g y + p0 (D.13.2)
for j = 1, 2, where j are the uid densities, g is the gravitational acceleration, and p0 is an
inconsequential interfacial pressure.
D.13 FDLIB User Guide: 08 stab/ifsf0 s 1181
Normal-mode analysis
A normal-mode perturbation displaces the interface to a position given by the real or imaginary
part of the function
is the normal-mode wave form of the perturbation, A is the complex amplitude of the interfacial
wave, k = 2/L is the wave number, L is the wavelength, and c is the complex phase velocity of the
perturbation.
Stream function
The stream function, , is dened by the equations u = /y and v = /x. In the linear
(0) (1)
analysis, we set j = j + j and introduce the normal-mode form
(1)
j (x, y, t) = j (y) exp[ik(x ct)] (D.13.5)
for j = 1, 2, where y = ky and the superscript (1) denotes the perturbation. The perturbation
velocity components are u(1) = (1) /y and v (1) = (1) /x.
The Reynolds number of the ow is so small that the motion of the uid is governed by
the equations of Stokes ow. Requiring that the stream function satises the biharmonic equation,
(1)
4 j = 0, and decays far from the interface, we obtain
Pressure eld
(0) (1)
The linearized pressure eld in the jth uid is pj = pj + pj . The perturbation pressure is
described by the normal-mode form
(1)
pj (x, y, t) = j qj (y) exp[ik(x ct)] (D.13.7)
for j = 1, 2, where qj (y) are pressure eigenfunctions. Substituting this expression into the x compo-
nent of the Stokes equation, we obtain
i 3 j 3 j
(1) (1)
(1)
pj = + , (D.13.8)
k x2 y y 3
yielding
d d3 j
qj (y) = i k 2
j
. (D.13.9)
dy dy 3
1182 Introduction to Theoretical and Computational Fluid Dynamics
yielding a1 = c2 .
Kinematic compatibility
Kinematic compatibility requires that D[f (x, t) y]/Dt = 0 evaluated at the interface, where D/Dt
is the material derivative. In the linearized approximation,
(1)
vj = 0, (D.13.11)
t
evaluated at y = 0 for either uid, j = 1, 2. Substituting the preceding expressions for the lower
uid velocity, we nd that ikcA + ika1 = 0 or
A = a1 , (D.13.12)
1 2
+ = + , (D.13.13)
y y=0 y y=0
yielding
d
d
1 2
A + k = A + k . (D.13.14)
dy y=0 dy y=0
1 2
(1 ) a1 + k =k , (D.13.15)
dy y=0 dy y=0
yielding
1 (1 ) a1 + b1 + c2 d2 = 0, (D.13.16)
Surfactant transport
Interfacial surfactant transport is governed by the convectiondiusion equation
D u
+ t = Ds , (D.13.20)
Dt l l l
where D/Dt is the material derivative, t is the unit tangent vector pointing in the direction of
increasing arc length, l, and Ds is the surface surfactant diusivity. Regarding as a function of x
and t, we obtain
u u
x
+ ux + t + = Ds . (D.13.21)
t x x y x l l l
The unit tangential vector can be linearized as
t ex + ey , (D.13.22)
x
where ex and ey are unit vectors along the x and y axes. To rst order in , x/l = 1. The
linearized surfactant transport equation becomes
(1) (1) u(1) u(0)
2 (1)
+ u(0) + 0 + = Ds , (D.13.23)
t x x y x x2
where (1) (x, t) = 1 exp[ik(x ct)], all terms are evaluated at the unperturbed position, y = 0, and
the velocity on the left-hand side is evaluated on either side of the interface. Choosing the upper
uid velocity, we obtain
1 ikc + 0 [ ik 2 (a1 + b1 ) + Aik ] = Ds k 2 1 . (D.13.24)
Substituting A = a1 and rearranging, we obtain
1 a1 + b1 + a1
= k, (D.13.25)
0 1 + iDs k
where we recall that = /k.
1184 Introduction to Theoretical and Computational Fluid Dynamics
(1) u(1) v
(1)
(1)
1 2
1 + = 2 + 2 + , (D.13.26)
y x y=0 y x y=0 x
where (1) (x, t) = 1 exp[ik(x ct)]. Expressing the velocity in terms of the stream function, we nd
that
2 (1) 2 1
2 (1) 2 2
(1) (1)
1 2 1 (1)
= + , (D.13.27)
y 2 x 2 y=0 y 2 x 2 y=0 1 x
yielding
d2
d2
1
1 2
+ 1 = + 2 +i . (D.13.28)
dy 2 y=0 dy 2 y=0 k1
Substituting the preceding expressions for j , simplifying, and using (D.13.19), we obtain an alge-
braic equation,
1 1 0
a1 + b1 (c2 d2 ) = i = i Ma . (D.13.29)
21 k 21 k 0
0 a1 + b1 + a1 /k
a1 + b1 (c2 d2 ) = i Ma , (D.13.30)
21 1 + iDs k
which can be rearranged into
a1 + b1 (c2 d2 ) = [ a1 (1 + ) + b1 ] , (D.13.31)
where
0
i Ma (D.13.32)
21 1 + Ds k i
is a dimensionless group. In the absence of surfactant, = 0. Setting a1 = c2 and rearranging, we
nd that
(1 ) b1 = [ 1 + (1 + ) ] c2 d2 . (D.13.33)
(1) v
(1)
2
(1) (1)
p1 + 21 p2 + 22 2 + g = 0 2 , (D.13.35)
y y=0 y y=0 x
where = 1 2 . Substituting the preceding expressions, we nd that
d1 d2
1 + 2 + i 2
+ 0 c2 = 0. (D.13.39)
2 k
c = i Ma + kDs . (D.13.42)
21 (1 + )
Since the growth rate I = kcI is always negative, this mode is stable. Since c2 = 0, the interface is
at and the perturbation ow is driven by in-plane interfacial motion.
1186 Introduction to Theoretical and Computational Fluid Dynamics
Program les:
1. ifsf0 s
Evaluates the analytical expressions for the phase velocity and growth rates.
2. ifsf0 s dr
Driver for ifsf0 s.
Input le:
1. ifsf0 s.dat
Specication of input parameters.
Output le:
1. ifsf0 s.out
Recording of the growth rate.
Base state
The layer is designated as uid 1 and the upper uid is designated as uid 2. The viscosity of the
layer uid is 1 and the viscosity of the upper uid is 2 = 1 , where is the viscosity ratio. The
origin of the y axis is set at the location of the unperturbed interface. For convenience, the x and y
velocity components are denoted by u = ux and v = uy .
In the unperturbed conguration, the interface is at, the uids are quiescent, and the pressure
distribution assumes the hydrostatic prole
(0)
pj (y) = j gy + p0 , (D.14.1)
where j = 1 or 2 for the lm or upper uid, j are the uid densities, g is the acceleration of gravity,
p0 is an unspecied reference pressure, and the superscript (0) denotes the unperturbed base state.
Normal-mode analysis
A normal-mode perturbation displaces the interface to a position described by the real or imaginary
part of the function
y
Fluid 2
x
h
g Fluid 1
Figure D.14.1 Gravitational (RayleighTaylor) instability and leveling of a liquid layer resting on a
horizontal wall underneath a semi-innite uid.
is the normal-mode wave form of the perturbation, A is the complex amplitude of the interfacial
wave, k = 2/L is the wave number, L is the wavelength, and c is the complex phase velocity.
Stream function
The stream function, , is dened by the equations u = /y and v = /x. In the absence of
(1)
base ow, we set j = j and introduce the normal-mode form
(1)
j (x, y, t) = j (y) exp[ik(x ct)], (D.14.4)
where j = 1 or 2 for the lm or upper uid, y = ky, and the superscript (1) denotes the perturbation.
The perturbation velocity components are u(1) = (1) /y and v (1) = (1) /x.
The Reynolds number of the ow is so small that the motion of the uid is governed by
the equations of Stokes ow. Requiring that the stream function satises the biharmonic equation,
(1)
4 j = 0, we obtain
where a1 , b1 , c1 , d1 , a2 , and b2 are six complex coecients. Note that the disturbance ow decays
exponentially with distance from the interface into the overlying uid.
Pressure eld
(0) (1)
The linearized pressure eld in the jth uid is pj = pj + pj . The perturbation pressure eld is
described by the normal-mode form
(1)
pj (x, y, t) = j qj (y) exp[ik(x ct)] (D.14.6)
for j = 1, 2, where qj (y) are pressure eigenfunctions. Substituting this expression into the x compo-
nent of the Stokes equation, we obtain
i 3 j 3 j
(1) (1)
(1)
pj = + , (D.14.7)
k x2 y y 3
yielding
1188 Introduction to Theoretical and Computational Fluid Dynamics
d d3 j
qj (y) = i k2
j
. (D.14.8)
dy dy 3
Wall conditions
The no-slip and no-penetration conditions at the wall, located at y = h, require that
d1
1 (y = k) = 0, (y = k) = 0, (D.14.9)
dy
a1 ek b1 k ek + c1 ey d1 k ek = 0,
(D.14.10)
a1 ek + b1 (1 y) ek c1 ey + d1 (1 + y) ek = 0.
Continuity of y velocity at the interface
In the linear approximation, continuity of y velocity at the interface requires that
(1) (1)
1 (x, y = 0, t) = 2 (x, y = 0, t), (D.14.11)
yielding
a1 + c1 = c2 . (D.14.12)
Kinematic compatibility
Kinematic compatibility requires that D[f (x, t) y]/Dt = 0 evaluated at the interface, where D/Dt
is the material derivative. In the linear approximation,
v (1) (y = 0) = 0. (D.14.13)
t
Substituting the preceding expressions and choosing the upper uid interfacial velocity, we nd that
ikc A + ikc2 = 0. Simplifying, we obtain
c2 = c A. (D.14.14)
1 2
= , (D.14.15)
y y=0 y y=0
yielding
a1 + b1 c1 + d1 = c2 + d2 . (D.14.16)
D.14 FDLIB User Guide: 08 stab/layer0 1189
(1) u(1) v
(1)
1 2
1 + = 2 + 2 (D.14.17)
y x y=0 y x y=0
or
2 (1) 2 1
(1) 2 (1) 2 2
(1)
1 2
= , (D.14.18)
y 2 x2 y=0 y2 x2 y=0
where = 2 /1 is the viscosity ratio. Substituting the preceding expressions and simplifying, we
nd that
a1 + b1 + c1 d1 = (c2 d2 ). (D.14.19)
(1) v
(1)
2
(1) (1)
p1 + 21 p2 + 22 2 + g = , (D.14.20)
y y=0 y y=0 x2
where is the surface tension and = 1 2 . Substituting the preceding expressions, we nd
that
d1
1 q1 + 2 q2 2ik2 (1 2 ) + g A = A k 2 , (D.14.21)
dy
evaluated at y = 0. Substituting
q1 (y = 0) = 2i k 2 (b1 + d1 ), q2 (y = 0) = 2i k2 d2 , (D.14.22)
we obtain
2i k 2 c 1 b1 + d1 d2 (1 )(c2 + d2 ) + g c2 = k2 c2 . (D.14.23)
where
i g
1 2
+ (D.14.25)
21 c k
is a dimensionless group. Rearranging the denition of , we derive an expression for the growth
rate,
1 g
I kcI = 2
+ . (D.14.26)
21 ( 1 + )k k
1190 Introduction to Theoretical and Computational Fluid Dynamics
and q exp(2k). Setting the determinant of the matrix M to zero provides us with a secular
equation for the computation of . Once is available, the growth rate follows from (D.14.26).
After some algebra, we derive the dimensionless growth rate
1 g
I = 2
+ F(k, ), (D.14.28)
21 (1 + )k k
where
1
2 sinh(2k) k + (sinh2 k k 2 )
F(k, ) = (1 + ) . (D.14.29)
(1 2 ) k 2 + (cosh k + sinh k)2
As k , the function F(k, ) tends to unity, yielding the growth rate of perturbations at the
interface between two semi-innite uids. In the limit 0, the function F(k, ) reduces the
second fraction on the right-hand side of (9.11.35) for a liquid lm underneath a constant-pressure
ambient gas.
When the uids are unstably stratied, 2 > 1 and < 0, the rst term inside the
parentheses on the right-hand side of (D.14.28) is responsible for the RayleighTaylor instability.
The second term inside the parentheses expresses the stabilizing inuence of the surface tension.
Program les:
1. layer0
Computes the growth rate by two methods.
2. layer0 dr
Driver for layer0
3. crout
Crout decomposition for computing the determinant.
Input le:
1. layer0.dat
Specication of input parameters.
D.15 FDLIB User Guide: 08 stab/layersf0 s 1191
y
L
Fluid 2
h Fluid 1 x
Figure D.15.1 Instability of a liquid layer coated on a horizontal wall underneath a semi-innite uid
undergoing simple shear ow. The interface is occupied by an insoluble surfactant.
Output le:
1. layer0.out
Recording of the growth rate.
Base state
The liquid layer is designated as uid 1 and the overlying semi-innite uid is designated as uid
2. The viscosity of the layer is 1 and the viscosity of the upper uid is 2 = 1 , where is the
viscosity ratio. The origin of the y axis is set at the wall. For convenience, the x and y velocity
components are denoted by u = ux and v = uy .
In the unperturbed conguration, the interface is at and the uids undergo unidirectional
ow parallel to the wall. The velocity proles inside the layer and upper uid are given by
(0) (0)
u1 = y, u2 = y + ( 1)h , (D.15.1)
where is the shear rate in the upper uid, h is the layer thickness, and the superscript (0) denotes
the unperturbed unidirectional ow. These expressions ensure continuity of velocity and shear stress
at the interface for any viscosity ratio. The corresponding pressure distributions are given by
(0)
pj (y) = j g (y h) + p0 (D.15.2)
Normal-mode analysis
A normal-mode perturbation displaces the interface to a position given by the real or imaginary
part of the function
is the normal-mode wave form of the perturbation, A is the complex amplitude of the interfacial
wave, k = 2/L is the wave number, L is the wavelength, and c is the complex phase velocity.
Stream function
The stream function, , is dened by the equations u = /y and v = /x. In the linear
(1)
analysis, we set j = (0) + j and introduce the normal-mode form
(1)
j (x, y, t) = j (y) exp[ik(x ct)], (D.15.5)
where the superscript (1) denotes the perturbation, y k(y h), and j = 1, 2 for the liquid
lm or semi-innite uid. The perturbation velocity components are u(1) = (1) /y and v (1) =
(1) /x.
The Reynolds number of the ow is so small that the motion of the uid is governed by
the equations of Stokes ow. Requiring that the stream function satises the biharmonic equation,
(1)
4 j = 0, we obtain
j (y) = a1j cosh y + a2j y cosh y + a3j sinh y + a4j y sinh y, (D.15.6)
where aij are eight complex coecients for i = 14 and j = 1, 2. To ensure that the perturbation
velocity decays in the upper uid far from the wall, we set
Pressure eld
(0) (1)
The linearized pressure eld in the jth uid is pj = pj + pj . The disturbance pressure eld takes
the normal-mode form
(1)
pj (x, y, t) = j qj (y) exp[ik(x ct)] (D.15.8)
for j = 1, 2, where qj (y) are pressure eigenfunctions. Substituting this expression into the x compo-
nent of the Stokes equation, we obtain
i 3 j 3 j
(1) (1)
(1)
pj = + , (D.15.9)
k x2 y y 3
D.15 FDLIB User Guide: 08 stab/layersf0 s 1193
yielding
d d3 j
qj (y) = i k2
j
. (D.15.10)
dy dy 3
yielding a11 = a12 . Combining this equation with (D.15.7), we obtain a32 = a12 = a11 .
Kinematic compatibility
Kinematic compatibility requires that D[f (x, t) y]/Dt = 0, where D/Dt is the material derivative.
In the linear approximation,
(1)
+ uI vj (y = h) = 0 (D.15.12)
t x
for j = 1, 2, where uI = h is the unperturbed interfacial velocity. Substituting the preceding
expressions, we obtain ikcA + uI ikA + ika11 = 0, yielding a11 = (c uI )A and
A = a11 , (D.15.13)
where 1/(c uI ) is the inverse of the complex phase velocity shifted by the real interfacial
velocity.
1 2
(x, t) + = (x, t) + , (D.15.14)
y y=h y y=h
yielding
d
d
1 2
A + k = A + k , (D.15.15)
dy y=k dy y=k
1 2
(1 ) a11 + k k = 0, (D.15.16)
dy y=h dy y=h
yielding
Wall conditions
The no-slip and no-penetration conditions at the horizontal wall require that
1 (y = k) = 0, 1 (y = k) = 0, (D.15.19)
where 0 , 0 are uniform values corresponding to the at interface, and 1 , 1 are complex ampli-
tudes. Since the perturbations are small, we can write
1 1
= Ma , (D.15.22)
0 0
where Ma is the Marangoni number.
Surfactant transport
Surfactant transport over the interface is governed by the equation
D u
+ t = Ds , (D.15.23)
Dt l l l
where D/Dt is the material derivative, l is the arc length along the interface measured in the direction
of the unit tangent vector, t, and Ds is the surface surfactant diusivity. Regarding as a function
of x and t, we obtain
u u
x
+ ux + t + = Ds . (D.15.24)
t x x y x l l l
The unit tangential vector can be linearized as
t ex + ey , (D.15.25)
x
where ex and ey are unit vectors along the x and y axes. To leading order, x/l = 1. The emerging
linearized form of the surfactant transport equation is
(1) (1) u(1) u(0)
2 (1)
+ u(0) + 0 + = Ds , (D.15.26)
t x x y x x2
D.15 FDLIB User Guide: 08 stab/layersf0 s 1195
where all terms are evaluated at the unperturbed position, y = h, and the velocity on the left-hand
side is evaluated on either side of the interface, and (1) (x, t) = 1 exp[ik(x ct)]. Substituting the
preceding expressions for the upper uid velocity, we obtain
(1) (1)
1 2 (1)
1 + = 2 + 2 + , (D.15.28)
y x y=h y x y=h x
where (1) (x, t) = 1 exp[ik(x ct)]. Expressing the velocity in terms of the stream function and
rearranging, we obtain
2 (1) 2 1
(1) 2 (1) 2 2
(1)
1 (1)
1 2
= + . (D.15.29)
y 2 x 2 y=h y 2 x 2 y=h 1 x
yielding
where
0
i Ma (D.15.33)
21 1 + Ds k i
(1) v
(1)
2
(1) (1)
p1 + 21 p2 + 22 2 + g = 0 2 , (D.15.34)
y y=h y y=h x
1196 Introduction to Theoretical and Computational Fluid Dynamics
d1 d2
1 q1 + 2 q2 2ik2 1 + 2ik 2 2 + g A = 0 Ak2 , (D.15.35)
dy dy
where all terms are evaluated at y = 0. Substituting further qj (0) = 2i k 2 a2j , we obtain
2ik2 (1 a21 2 a22 ) 2ik 2 1 (a21 + a31 ) + 2ik 2 2 (a22 + a32 ) + g A = 0 Ak2 , (D.15.36)
where
1 g 0
2
+ 0 = (Bo + 1) (D.15.39)
21 k 21
Setting the determinant of the matrix M to zero provides us with a quadratic equation for
two possible growth rates,
P2 () = (1 + Ds k i) det(M ) = a 2 + b + c = 0. (D.15.42)
D.16 FDLIB User Guide: 08 stab/orr 1197
The factor 1 + Ds k i on the left-hand side involving the surface surfactant diusivity arises from
the denition of the dimensionless parameter . In the numerical method, the binomial coecients
are computed by an exact nite-dierence based on the equations
P2 (1) = a + b + c, P2 (1) = a b + c, P2 (0) = c. (D.15.43)
The roots are found analytically using the quadratic formula.
As the dimensionless wave number kh increases, the eect of the wall becomes decreasingly
important, yielding shear ow past an interface separating two semi-innite uids. Our analysis in
directory ifsf0 s has revealed two normal modes with complex phase velocities given in (D.13.42)
and (D.13.42),
1 g 0
c = uI i + , c = uI i Ma + kDs . (D.15.44)
21 (1 + ) k2 21 (1 + )
Both modes are stable with negative growth rates and phase velocity equal to the undisturbed
interfacial velocity.
Program les:
1. layersf0 s
Evaluates the complex phase velocity.
2. layersf0 s dr
Driver for the subroutine layersf0 s
3. det 33c
Determinant of a 3 3 complex matrix.
4. det 44c
Determinant of a 4 4 complex matrix.
5. quadc
Computes the roots of a quadratic equation with complex coecients.
Input le:
1. layersf0 s.dat
Specication of input parameters.
Output le:
1. layersf0 s.out
Recording of the growth rate.
Program les:
1. thread0
Evaluates the growth rate from analytical expression derived in Section 9.13.2.
2. bess I01K01
Computes the Bessel functions.
3. thread0 dr
Driver calling the main function.
Input les:
1. thread0.dat
Specication of input parameters.
Output les:
1. thread0.out
Growth rates.
Program les:
1. thread1
Evaluates the phase velocity and growth rate from analytical expressions.
2. bess I01K01
Computes Bessel functions.
3. thread1 dr
Driver calling the main function.
D.21 Directory vl
This directory contains a code that computes the phase velocity and growth rate of periodic distur-
bances on a vortex layer from analytical expressions derived in Section 9.6.
D.22 Directory vs
This directory contains a code that produces the phase velocity and growth rate of periodic distur-
bances on a vortex sheet separating two streams merging with dierent velocities from analytical
expressions derived in Section 9.10.
References
[1] Abdallah, S. (1987) Numerical solutions for the incompressible NavierStokes equations in
primitive variables using a non-staggered grid II. J. Comp. Phys. 70, 193202.
[2] Abramowitz, M. & Stegun, I. A. (1972) Handbook of Mathematical Functions. Dover.
[3] Adamson, A. W. (1990) Physical Chemistry of Surfaces. Fifth Edition. Wiley.
[4] Ahlfors, L. A. (1979) Complex Analysis. McGrawHill.
[5] Ames, W. F. (1977) Numerical Methods for Partial Dierential Equations. Academic.
[6] Anderson, C. R. (1986) A method of local corrections for computing the velocity eld due to
a distribution of vortex blobs. J. Comp. Phys. 62, 111123.
[7] Andersson, C. R. (2009) Modied SchwarzChristoel mappings using approximate curve
factors. J. Comp. Appl. Math. 233, 11171127.
[8] Anderson, C. R. (1989) Vorticity boundary conditions and boundary vorticity generation for
two-dimensional viscous incompressible ows. J. Comp. Phys. 80, 7297.
[9] Anderson, P. A., Tannehill, J. C. & Pletcher, R. H. (1984) Computational Fluid
Dynamics and Heat Transfer. Taylor & Francis.
[10] Anshus, B. E. & Goren, S. L. (1966) A method of getting approximate solutions to the
Orr-Sommerfeld equation for ow on a vertical wall. AIChE J. 12, 10041008.
[11] Aragbesola, Y. A. S. & Burley, D. M. (1977) The vector and scalar potential method for
the numerical solution of two- and three-dimensional NavierStokes equations. J. Comp. Phys.
24, 398415.
[12] Aref, H. (1983) Integrable, chaotic, and turbulent vortex motion in two-dimensional ows.
Annu. Rev. Fluid Mech. 15, 345389.
[13] Aref, H., Newton, P. K., Stremler, M. A., Tokieda, T. & Vainchtein, D. L. (1983)
Vortex Crystals. Arch. Appl. Mech. 39, 281.
[14] Aris, R. (1962) Vectors, Tensors, and the Basic Equations of Fluid Mechanics. Prentice-Hall.
[15] Atkinson, K. E. (1973) Iterative variants of the Nystrom method for the numerical solution
of integral equations. Numer. Math. 22, 1731.
1200
References 1201
[16] Atkinson, K. E. (1976) A Survey of Numerical Methods for the Solution of Fredholm Integral
Equations of the Second Kind. SIAM.
[17] Atkinson, K. E. (1980) The numerical solution of Laplaces equation in three dimensions - II.
In Numerical Treatment of Integral Equations, J. Albrecht & L. Collatz (Eds.), 123, Birkhauser,
Basel, Switzerland.
[19] Atkinson, K. E. (1990) A Survey of Boundary Integral Methods for the Numerical Solution of
Laplaces Equation in Three Dimensions. In Numerical Solution of Integral Methods, Goldberg,
M. A. (Ed.), Plenum Press.
[20] Aziz, K. & Hellums, J. D. (1967) Numerical solution of the three-dimensional equations of
motion for laminar natural convection. Phys. Fluids 10, 314324.
[21] Babu, V. & Korpela, S. (1994) Numerical solution of the incompressible, three-dimensional
NavierStokes equations. Comput. & Fluids 23, 675691.
[22] Baker, G. R. (1983) Generalized vortex methods for free-surface ows. In: Waves in Fluid
Interfaces. Academic Press.
[23] Batchelor, G. K. (1954) The skin friction on innite cylinders moving parallel to their length.
Quart. J. Mech. Appl. Math. 7, 179192.
[25] Batchelor, G. K. (1970) Slender-body theory for particles of arbitrary cross-section in Stokes
ow. J. Fluid Mech. 44, 419440 .
[26] Bazant, M. Z. (2004) Conformal mapping of some non-harmonic functions in transport theory.
Proc. R. Soc. Lond. A 460, 14331452.
[27] Bell, J., Colella, P. & Glaz, H. (1989) A second-order projection method for the incom-
pressible NavierStokes equations. J. Comp. Phys. 85, 257283.
[28] Benjamin, T. B. (1957) Wave formation in laminar ow down an inclined plane. J. Fluid
Mech. 2, 554574.
[29] Benney, D. J. (1966) Long waves in liquid lms. J. Math. Phys. 45, 150155.
[31] Berker, R. (1963) Integration des equations du mouvement dun uide visqueux incompress-
ible. In Handbuch der Physik, Flugge, S. (Ed.), 8(2), 1384, Springer, Berlin.
[32] Bertsekas, D. P. & Tsitsiklis, J. N. (1989) Parallel and Distributed Computation. Prentice-
Hall.
1202 Introduction to Theoretical and Computational Fluid Dynamics
[33] Betchov, R. (1965) On the curvature and torsion of an isolated vortex lament. J. Fluid
Mech. 22, 471479.
[34] Betchov, R. & Criminale, W. O. (1967) Stability of Parallel Flows. Academic Press.
[35] Betchov, R. & Szewczyk, A. (1963) Stability of a shear layer between parallel streams.
Phys. Fluids 6, 13911396.
[36] Bird, R. B., Stewart, W. E. & Lightfoot, E. N. (2006) Transport Phenomena. Second
Edition. Wiley.
[37] Blake, J. R. (1971) A note for the image system for a Stokeslet in a no-slip boundary. Proc.
Camb. Phil. Soc. 70, 303310.
[38] Blake, J. R. & Gibson, D. C. (1987) Cavitation bubbles near boundaries. Annu. Rev. Fluid
Mech. 19, 99123.
[39] Blasius, H. (1908) Grenzschichten in Flussigkeiten mit kleiner reibung. Mathematik und
Physik 56, 1-37. Z. Math. Phys. 56, 137.
[40] Blasius, H. (1910) Funktionentheoretische methoden in der hydrodynakik. Z. Math. Phys. 58,
90110.
[41] Blottner, F. G. (1970) Finite dierence methods of solution of the boundary layer equations.
AIAA J. 8, 193205.
[44] Boas, M. L. (2005) Mathematical Methods in the Physical Sciences. Third Edition. Wiley.
[45] Bodewig, E. (1959) Matrix Calculus. North-Holland.
[46] Bogy, D. B. (1979) Drop formation in a circular liquid jet. Annu. Rev. Fluid Mech. 11,
207228.
[47] Burstein, S. Z. (1967) Finite-dierence calculations for hydrodynamic ows containing dis-
continuities. J. Comp. Phys. 2, 198222.
[48] Gupta N. R. & Borhan, A. (2003) Capsule motion and deformation in tube and channel ow.
In Modeling and Simulation of Capsules and Biological Cells, Pozrikidis C. (Ed.), Chapman &
Hall /CRC.
[49] Boris, J. P. & Book, D. L. (1973) Flux-corrected transport. I. SHASTA, A uid transport
algorithm that works. J. Comp. Phys. 11, 3869.
[50] Boris, J. P., Book, D. L. & Hain, K. (1975) Flux-corrected transport II: Generalizations
of the method. J. Comp. Phys. 18, 248283.
References 1203
[51] Boris, J. P. & Book, D. L. (1976) Flux-corrected transport. III. Minimal-error FCT algo-
rithms. J. Comp. Phys. 20, 397431.
[52] Boucher, E. A. (1980) Capillary phenomena: properties of systems with uid/uid interfaces.
Rep. Prog. Phys. 43, 497545.
[53] Brady, M. & Pozrikidis, C. (1993) Diusive transport across irregular and fractal walls.
Proc. R. Soc. Lond. A 442, 571583.
[54] Brenner, H. (1963) The Stokes resistance of an arbitrary particle. Chem. Eng. Sci. 18, 125.
[55] Brenner, H. (1964) The Stokes resistance of an arbitrary particleII: An extension. Chem.
Eng. Sci. 19, 599629.
[56] Brenner, H. (1964) The Stokes resistance of an arbitrary particleIV: Arbitrary elds of ow.
Chem. Eng. Sci. 19, 703727.
[57] Briley, R. W. (1974) Numerical method for predicting three-dimensional steady viscous ow
in ducts. J. Comp. Phys. 14, 828.
[58] Brown, R. A. (1979) Finite-element methods for the calculation of capillary surfaces. J. Comp.
Phys. 33, 217235.
[59] Brown, D. L., Cortez, R. & Minion, M. L. (2001) Accurate projection methods for the
incompressible NavierStokes equations. J. Comp. Phys. 168, 464499.
[60] Buckmaster, J. (1973) Viscous-gravity spreading of an oil slick. J. Fluid Mech. 59, 481491.
[61] Burgers, J. M. (1938) On the motion of small particles of elongated form, suspended in
a viscous liquid. Second Report on Viscosity and Plasticity. Kon. Ned. Akad. Wet. Verhand.
(Eerste Sectie) 16: 113.
[62] Burgers, J. M. (1948) A mathematical model illustrating the theory of turbulence. Adv. Appl.
Mech. 1, 171199.
[63] Butler, S. F. J (1953) A note on Stokess stream function for motion with a spherical bound-
ary, Proc. Camb. Phil. Soc. 49, 169174.
[64] Caflisch, R. E., Li, X. & Shelley, M. J. (1993) The collapse of an axi-symmetric, swirling
vortex sheet. Nonlinearity 6, 843867.
[65] Carrier, G. F., Krook, M. & Pearson, C. E. (1983) Functions of a Complex Variable:
Theory and Technique. Hod Books.
[66] Carslaw, H. S. & Jaeger, J. C. (1959) Conduction of Heat in Solids. Second Edition.
Oxford University Press.
[67] Case, K. M. (1960) Stability of inviscid plane Couette ow. Phys. Fluids 3, 143148.
[68] Case, K. M. (1960) Stability of an idealized atmosphere I. Discussion of results. Phys. Fluids
3, 149154.
1204 Introduction to Theoretical and Computational Fluid Dynamics
[69] Cebeci, T. (1982) Numerical and Physical Aspects of Aerodynamic Flows. Springer-Verlag.
[70] Cebeci, T. & Bradshaw, P. (1977) Momentum Transfer in Boundary Layers. Hemisphere.
[71] Cebeci, T. & Bradshaw, P. (1984) Physical and Computational Aspects of Convective Heat
Transfer SpringerVerlag.
[72] Cercignani, C. (2000) Rareed Gas Dynamics. Cambridge University Press.
[73] Chandrasekhar, S. (1961) Hydrodynamic and Hydromagnetic Stability. Dover.
[74] Chang, H.-C. (1994) Wave evolution on a falling lm. Annu. Rev. Fluid Mech. 26, 103136.
[75] Chebbi, R. (2001) Viscousgravity spreading of oil on water. AIChE J. 47, 288294.
[76] Chien, J. C. (1976) A general nite-dierence formulation with applications to NavierStokes
equations. J. Comp. Phys. 20, 268278.
[77] Chin, R. W., Abernathy, F. H. & Bertschy, J. R. (1986) Gravity and shear wave stability
of free surface ows. Part 1. Numerical calculations. J. Fluid Mech. 168, 501513.
[78] Chiu, C. & Nicolaides, R. A. (1988) Convergence of a higher-order vortex method for
two-dimensional Eulers equations. Math. Comp. 51, 507534.
[79] Chorin, A. J. (1967) A numerical method for solving incompressible viscous ow problems.
J. Comp. Phys. 2, 1226.
[80] Chorin, A. J. (1968) Numerical solution of the NavierStokes equations. Math. Comp. 22,
745762.
[81] Chorin, A. J. (1973) Numerical study of slightly viscous ow. J. Fluid Mech. 57, 785796.
[82] Chorin, A. J. & Bernard, P. S. (1973) Discretization of a vortex sheet, with an example
of roll-up. J. Comp. Phys. 13, 423429.
[83] Chorin, A. J. (1978) Vortex sheet approximation of boundary layers. J. Comp. Phys. 27,
428442.
[84] Chorin, A. J. (1980) Vortex models and boundary layer instability. SIAM J. Sci. Stat. Comput.
1, 121.
[85] Christiansen, J. P. (1973) Numerical simulation of hydrodynamics by the method of point
vortices. J. Comp. Phys. 13, 363379.
[86] Chwang, A. T. & Wu, T. Y.-T. (1974) Hydromechanics of low-Reynolds-number ow. Part
1, Rotation of axisymmetric prolate bodies. J. Fluid Mech. 63, 607622.
[87] Chwang, A. T. & Wu, T. Y.-T. (1975) Hydromechanics of low-Reynolds-number ow. Part
2, Singularity methods for Stokes ows. J. Fluid Mech. 67, 787815.
[88] Clements, R. R. & Maull, D. J. (1975) The representation of sheets of vorticity by discrete
vortices. Prog. Aerospace Sci. 16, 129146.
References 1205
[89] Clift, R., Grace, J. R. & Weber, M. E. (1978) Bubbles, Drops, and Particles. Academic
Press.
[90] Cochran, W. G. (1934) The ow due to a rotating disk. Proc. Camb. Phil. Soc. 30, 365375.
[91] Colton, D. & Kress, R. (1983) Integral Equation Methods in Scattering Theory. Wiley.
[92] Concus, P. (1968) Static menisci in a vertical right circular cylinder. J. Fluid Mech. 34,
481495.
[93] Concus, P. & Finn, R. (1974) On capillary free surfaces in a gravitational eld. Acta Math.
132, 207223.
[94] Couette, M. (1890) Etudes sur le frottement des liquides. Ann. Chim. Phys. 21, 433510.
[95] Craik, A. D. D. (1985) Wave Interactions and Fluid Flows. Cambridge University Press.
[96] Crane, L. J. (1970) Flow past a stretching plate. ZAMP 21, 645647.
[97] Crank, J. & Nicolson, P. (1947) A practical method for numerical evaluation of solutions
of partial dierential equations of the heat-conduction type. Proc. Camb. Phil. Soc. 43, 5067.
[98] Greig, I. S. & Morris, J. L. (1976) A hopscotch method for the Korteweg-de-Vries equation.
J. Comp. Phys. 20, 6480.
[99] Crighton, D. G. (1985) The Kutta condition in unsteady ow. Annu. Rev. Fluid Mech. 17,
411445.
[101] Davey, A. (1977) On the numerical solution of dicult eigenvalue problems. J. Comput.
Phys. 24, 331338.
[102] David, T. & Blyth, G. (1992) Parallel algorithms for panel methods. Int. J. Num. Meth.
Fluids 14, 95108.
[103] Dean, W. R. & Montagnon, P. E. (1949) On the steady motion of viscous liquid in a
corner. Proc. Camb. Phil. Soc. 45, 389394.
[104] Delves, L. M. & Mohamed, J. L. (1985) Computational Methods for Integral Equations.
Cambridge University Press.
[105] Dennis, S. C. R., Ingham, D. B., & Cook, R. N. (1979) Finite-dierence methods for
calculating steady incompressible ows in three dimensions. J. Comp. Phys. 33, 325339.
[107] Diaz, A., Barthes-Biesel, D. & Pelekasis, N. (2001) Eect of membrane viscosity on
the dynamic response of an axisymmetric capsule. Phys Fluids 13, 38353838.
1206 Introduction to Theoretical and Computational Fluid Dynamics
[108] Dorrepaal, J. M. (1986) An exact solution of the NavierStokes equation which describes
non-orthogonal stagnation-point ow in two dimensions. J. Fluid Mech. 163, 141147.
[109] Douglas, J. Jr. (1955) On the numerical integration of 2 u/x2 + 2 u/y 2 = u/t by
implicit methods. J. Soc. Indust. Appl. Math. 3, 4265.
[110] Douglas, J. Jr. & Gunn, J. E. (1964) A general formulation of alternating direction
methods. I. Parabolic and hyperbolic problems. Numerische Mathematik 6, 428453.
[111] Douglas, J. Jr. (1962) Alternating direction methods for three space variables. Numerische
Mathematik 4, 4163.
[112] Drazin, P. G. & Reid, W. H. (2004) Hydrodynamic Stability. Second Edition. Cambridge
University Press.
[113] Dritschel, D. G. (1989) Contour dynamics and contour surgery: numerical algorithms for
extended, high-resolution modelling of vortex dynamics in two-dimensional, inviscid, incom-
pressible ows. Comp. Phys. Rep. 10, 77146.
[114] Dussan V., E. B. (1979) On the spreading of liquids on solid surfaces: static and dynamic
contact lines Ann. Rev. Fluid Mech. 11, 371400.
[115] Dussan V., E. B. (1985) On the ability of drops or bubbles to stick to non-horizontal surfaces
of solids. Part 2. Small drops or bubbles having contact angles of arbitrary size. J. Fluid Mech.
151, 120.
[116] Du Fort, E. C. & Frankel, S. P. (1953) Stability conditions in the numerical treatment
of parabolic dierential equations. Mathematical Tables and Other Aids to Computation 7,
135152.
[117] E, W. & Liu, J.-G. (1995) Projection method I: Convergence and numerical boundary layers,
SIAM J. Num. Anal. 32 10171057.
[118] Edwards, D. A., Brenner, H. & Wasan D. T. (1991) Interfacial Transport Processes and
Rheology. Butterworth-Heinemann.
[119] Eliassen, A., Hilland, E. & Riis, E. (1953) Two-dimensional perturbation of a ow with
constant shear of a stratied uid. Publ. 1, Inst. Weather & Climate Res., Norwegian Acad.
Sci. Letters.
[120] Evans, H. (1968) Laminar Boundary Layers. Addison-Wesley.
[121] Falkner, V. M. & Skan, S. W. (1931) Solutions of the boundary layer equations. Phil.
Mag. 12, 865896.
[122] Faxen, H. (1924) Der widerstand gegen die bewegung einer starren kugel in einer zahen
ussigkeit, die zwischen zwei parallelen, ebener Wanden eingeschlossen ist. Ark. Mat. Astr.
Fys. 18(29), 3.
[123] Feldman, S. (1957) On the hydrodynamic stability of two viscous incompressible uids in
parallel uniform shearing motion. J. Fluid Mech. 2, 343370.
References 1207
[125] Fetter, A. L. (1974) Translational velocity of a classical vortex ring. Phys. Rev. A 10,
17241727.
[127] Fink, P. T. & Soh, W. K. (1974) Calculation of vortex sheets in unsteady ow and appli-
cations in ship hydrodynamics. Proc. 10th Symp. Naval Hydrodynamics, 463491.
[128] Fjrtoft, R. (1950) Application of integral theorems in deriving criteria of stability for
laminar ows and for the baroclinic circular vortex. Geophys. Publ. 17(5).
[131] Floryan, J. M. & Rasmussen, H. (1989) Numerical methods for viscous ows with moving
boundaries. Appl. Mech. Rev. 42, 323341.
[132] Floryan, J. M. & Zemach, C. (1988) Quadrature rules for singular integrals with applica-
tion to SchwartzChristoel mappings. J. Comp. Phys. 75, 1530.
[133] Floryan, J. M. & Zemach, C. (1993) SchwarzChristoel methods for conformal mapping
of regions with a periodic boundary. J. Comp. Appl. Math. 46, 77102.
[134] De Foy, B. & Dawes, W. (2000) Unstructured pressure-correction solver based on a con-
sistent discretization of the Poisson equation. Int. J. Num. Meth. Fluids 34, 463478.
[135] Foda, M. & Cox, R. G. (1980) The spreading of thin liquid lms on a water-air interface.
J. Fluid Mech. 101, 3351.
[136] Fraenkel, L. E. (1962) Laminar ow in symmetrical channels with slightly curved walls
I. On the Jeery-Hamel solutions for ow between plane walls. Proc. R. Soc. Lond. A 267,
119138.
[137] Fraenkel, L. E. (1962) Laminar ow in symmetrical channels with slightly curved walls II.
An asymptotic series for the stream function. Proc. R. Soc. Lond. A 272, 406428.
[139] Fulford, G. D. (1964) The ow of liquids in thin lms. Adv. Chem. Eng. 5, 151236.
[141] Gaster, M. (1962) A note on the relation between temporally-increasing and spatially-
increasing disturbances in hydrodynamic stability. J. Fluid Mech. 14, 222224.
1208 Introduction to Theoretical and Computational Fluid Dynamics
[142] Gatski, T. B., Grosch, C. E. & Rose, M. E. (1989) The numerical solution of the Navier
Stokes equations for three-dimensional, unsteady, incompressible ows by compact schemes. J.
Comp. Phys. 82, 298329.
[144] Ghia, U., Ghia, K. N. & Studerus, C. J. (1976) A study of three-dimensional laminar
incompressible ow in ducts. AIAA Paper 76-424.
[145] Ghia, K. N., Hankey, W. L. & Hodge, J. K. (1977) Study of incompressible Navier
Stokes equations in primitive variables using implicit numerical technique. AIAA Paper 77-648,
156165.
[146] Goldshtik, M. A. (1990) Viscous-ow paradoxes. Annu. Rev. Fluid Mech. 22, 441472.
[147] Golter, H. (1940) On the three-dimensional instability of laminar boundary layers on concave
walls. Tech. Mem. Nat. Adv. Comm. Aero. Wash. No 1375.
[148] Golter, H. (1957) A new series for the calculation of steady laminar boundary layer ows.
J. Math. Mech. 6, 166.
[149] Gourlay, A. R. (1970) Hopscotch: A fast second-order partial dierential equation solver.
J. Inst. Maths. Applics. 6, 375390.
[150] Gradshteyn, I. S. & Ryzhik, I. M. (1980) Table of Integrals, Series, and Products. Aca-
demic Press.
[151] Gray, J. & Hancock, G. J. (1955) The propulsion of sea-urchin spermatozoa. J. Exp. Biol.
32, 802814.
[152] Green, A. E. & Adkins, J. E. (1960) Large Elastic Deformations and Non-linear Continuum
Mechanics. Oxford University Press.
[153] Gresho, P. M. & Sani, R. L. (1987) On pressure boundary conditions for the incompressible
NavierStokes equations. Int. J. Num. Meth. Fluids 7, 11111145.
[154] Gresho, P. M. (1990) On the theory of semi-implicit projection methods for viscous incom-
pressible ow and its implementation via a nite-element method that also introduces a nearly
consistent mass matrix. Part 1: Theory. Int. J. Num. Meth. Fluids 11, 587620.
[155] Gresho, P. M. (1991) Incompressible uid dynamics: some fundamental formulation issues.
Annu. Rev. Fluid Mech. 23, 413453.
[156] Gueron, S. & Liron, N. (1992) Ciliary motion modeling, and dynamic multicilia interac-
tions. Biophys. J. 63, 10451058.
[157] Guj, G. & Stella, F. (1988) Numerical solutions of high-Re recirculating ows in vorticity-
velocity form. Int. J. Num. Meth. Fluids. 8, 405416.
References 1209
[158] Guj, G. & Stella, F. (1993) A vorticityvelocity method for the numerical solution of 3D
incompressible ows. J. Comp. Phys. 106, 286298.
[160] Hamadiche, M., Scott, J. & Jeandel, D. (1994) Temporal stability of Jeery-Hamel ow.
J. Fluid Mech. 268, 7188.
[161] Hamel, G. (1916) Spiralformige bewegungen zaher Flussigkeiten. Jahresb. Deutsch. Math.-
Verein. 25, 3460.
[162] Hama, F. R. (1962) Progressive deformation of a curved vortex lament by its own induction.
Phys. Fluids 5, 11561162.
[163] Hama, F. R. (1963) Progressive deformation of a perturbed line vortex lament by its own
induction. Phys. Fluids 6, 526534.
[165] Hancock, G. J. (1953) The self-propulsion of microscopic organisms through liquids. Proc.
R. Soc. Lond. A 217, 96121.
[166] Happel, J. & Brenner, H. (1973) Low Reynolds Number Hydrodynamics. Martinus Nijho.
[167] Hardin, J. C. (1982) The velocity eld induced by a helical vortex lament. Phys. Fluids 25,
19491952.
[169] Harper, J. F. & Moore, D. W. (1968) The motion of a spherical liquid drop at high
Reynolds number J. Fluid Mech. 32, 367391.
[170] Hartree, D. R. (1937) On an equation occurring in Falkner and Skans approximate treat-
ment of the equations of the boundary layer. Proc. Camb. Phil. Soc. 33, 223239.
[171] Hasimoto, H. (1972) A soliton on a vortex lament. J. Fluid Mech. 51, 477485.
[172] Havelock, T. H. (1931) The stability of motion of rectilinear vortices in ring formation.
Phil. Mag. Ser. 7. 11(70), Suppl. 617633.
[173] Henshaw, W. D. (1994) A fourth-order accurate method for the incompressible NavierStokes
equations on overlapping grids. J. Comp. Phys. 113, 1325.
[174] Hess, J. L. (1990) Panel methods in computational uid dynamics. Annu. Rev. Fluid Mech.
22, 255274.
[175] Hicks, W. M. (1923) On the mutual threading of vortex rings. Proc. R. Soc. Lond. A 102,
111131.
1210 Introduction to Theoretical and Computational Fluid Dynamics
[176] Hiemenz, K. (1911) Die Grenzschicht an einem in den gleichformigen ussigkeitsstrom einge-
tauchten geraden kreiszylinder. Dinglers Polyt. J. 326(21), 321410.
[177] Higdon, J. J. L. (1979) A hydrodynamic analysis of agellar propulsion. J. Fluid Mech. 90,
685711.
[179] Hill, A. I. & Pozrikidis, C. (2011) On the shape of a hydrostatic meniscus attached to a
corrugated plate or wavy cylinder J. Colloid Interf. Sci. 356, 763774.
[180] Hindmarsh, A. C., Gresho, P. M. & Griffiths, D. F. (1984) The stability of explicit
Euler time-integration for certain nite dierence approximations of the multi-dimensional
advection-diusion equation. Int. J. Num. Meth. Fluids 4, 853897.
[181] Hirasaki, G. J. & Hellums, J. D. (1970) Boundary conditions on the vector and scalar
potentials in viscous three-dimensional hydrodynamics. Quart. Appl. Math. 28, 293296.
[182] Hirsch, C. (1991) Numerical Computation of Internal and External Flows. Vols. I and II.
Wiley.
[183] Hocquart, R. & Hinch, E. J. (1983) The long-time tail of the angular-velocity autocorre-
lation function for a rigid Brownian particle of arbitrary centrally symmetric shape. J. Fluid
Mech. 137, 217220.
[184] Hoffman, J. D. (1992) Numerical Methods for Engineers and Scientists. McGrawHill.
[185] Hoffmann, K. A. & Chiang, S. T. (1993) Computational Fluid Dynamics for Engineers.
Vol I. Engineering Education System.
[186] Hoffmann, K. A. & Chiang, S. T. (1993) Computational Fluid Dynamics for Engineers.
Vol II. Engineering Education System.
[187] Homann, F. (1936) Einu groer zahigkeit bei stromung um zylinder. Forschg. Ing.-Wes. 7,
110.
[188] Homann, F. (1936) Der einu groer zahigkeit bei der stromung um den zylinder und um
die kugel. Z. Angew. Math. Mech. 16(3), 153164.
[189] Hooper, A. (1989) The stability of two superposed viscous uids in a channel. Phys. Fluids
A 1, 11331142.
[190] Hooper, A., Duffy, B. R. & Moffatt, H. K. (1982) Flow of uid of non-uniform viscosity
in converging and diverging channels. J. Fluid Mech. 117, 283304.
[191] Hornung, U. & Mittelmann, H. D. (1990) A nite element method for capillary surfaces
with volume constraints. J. Comp. Phys. 87, 126136.
[192] Howard, L. N. (1961) Note on a paper by John W. Miles. J. Fluid Mech. 10, 509512.
References 1211
[193] Howarth, L. (1951) The boundary layer in three-dimensional ow. Part II: The ow near a
stagnation point. Phil. Mag. Ser. 7 42, 14331440.
[194] Howell, L. H. (1993) Numerical conformal mapping of circular arc polygons. J. Comp. Appl.
Math. 46, 728.
[195] Huerre, P. & Monkewitz, P. A. (1990) Local and global instabilities in spatially devel-
oping ows. Annu. Rev. Fluid Mech. 22, 473537.
[196] Hughes, T. J. R., Liu, W. K. & Brooks, A. (1979) Finite element analysis of incompress-
ible viscous ows by the penalty function formulation. J. Comp. Phys. 30, 160.
[197] Isenberg, C. (1992) The Science of Soap Films and Soap Bubbles. Dover.
[198] Israeli, M. (1972) On the evaluation of iteration parameters for the boundary vorticity. Stud.
Appl. Math. 51, 6771.
[199] Jeffery, G. B. (1915) The two-dimensional steady motion of a viscous uid. Phil. Mag. 29,
455465.
[200] Johnson, R. E. (1980) An improved slender-body theory for Stokes ow. J. Fluid Mech. 99,
411431.
[201] Joseph, D. D., Liao, T. Y. & Hu, H. H. (1993) Drag and moment in viscous potential
ow. Eur. J. Mech. B/Fluids 12, 97106.
[202] Kaneda, Y. (1990) On the three-dimensional motion of an innitely thin vortex sheet in an
ideal uid. Phys. Fluids A 2, 18171826.
[203] Kaneko, A. & Honji, H. (1979) Double structures of steady streaming in the oscillatory
viscous ow over a wavy wall. J. Fluid Mech. 93, 727736.
[205] Kerner, W. (1989) Large-scale complex eigenvalue problems. J. Comput. Phys. 85, 185.
[206] Kim, J. & Moin, P. (1985) Application of a fractional-step method to incompressible Navier
Stokes equations. J. Comp. Phys. 59, 308323.
[207] Kim, S. (1985) A note on Faxen laws for nonspherical particles. Int. J. Multiph. Flow 11,
713719.
[208] Kim, S. & Karrila, S. J. (1991) Microhydrodynamics: Principles and Selected Applications.
Butterworth.
[209] Kim, S. & Lu, S.-Y. (1987) The functional similarity between Faxen relations and singularity
solutions for uid-uid, uid-solid and solid-solid dispersions. Int. J. Multiph. Flow 13, 837844.
[210] Kim, S. K. & Troesch, A. W. (1989) Streaming ows generated by high-frequency small-
amplitude oscillations of arbitrarily shaped cylinders. Phys. Fluids A 1, 975985.
1212 Introduction to Theoretical and Computational Fluid Dynamics
[212] Kovasznay, L. I. G. (1948) Laminar ow behind a two-dimensional grid. Proc. Camb. Phil.
Soc. 44, 5862.
[213] Kozlov, S. V., Lifanov, I. K. & Mikhailov, A. A. (1991) A new approach to mathemat-
ical modelling of ow of ideal uid around bodies. Sov. J. Num. Anal. Mod. 6, 209222.
[214] Kwak, S. & Pozrikidis, C. (2001) Eect of surfactants on the instability of a liquid thread
of annular layer, Part I: Quiescent uids. Int. J. Multiph. Flow 27, 137.
[215] Krasny, R. (1986) A study of singularity formation in a vortex sheet by the point-vortex
approximation. J. Fluid Mech. 167, 6593.
[216] Krasny, R. (1986) Desingularization of periodic vortex sheet roll-up. J. Comp. Phys. 65,
6593.
[217] Ladyzhenskaya, O. A. (1975) Mathematical analysis of NavierStokes equations for incom-
pressible liquids. Annu. Rev. Fluid Mech. 7, 249272.
[218] Lagestrom, P. A. (1964) Laminar Flow Theory. In Theory of Laminar Flows, Moore, F. K.
(Ed.), Princeton University Press.
[219] Lamb, H. (1911) On the uniform motion of a sphere through a viscous uid. Phil. Mag. 21,
112121.
[220] Lamb, H. (1932) Hydrodynamics. Dover.
[221] Landau, L. (1944) An exact solution of NavierStokes equations. Doklady Acad. Nauk. SSSR
43(7), 299301.
[222] Landau, L. & Levich, V. G. (1942) Dragging of a liquid by a moving plate. Acta Physic-
ochim. (USSR) 17, 4254.
[223] Langlois, W. E. (1964) Slow Viscous Flow. MacMillan.
[224] Lasseigne, D. G. & Jackson, T. L. (1992) Stability of a nonorthogonal stagnation ow to
three-dimensional disturbances. Theor. Comput. Fluid Dyn. 3, 207218.
[227] Lax, P. D. & Richtmyer, R. D. (1956) Survey of the stability of linear nite dierence
equations. Part I. An equivalence theorem. Comm. Pure Appl. Math. 9, 267293.
[228] Lax, P. D. & Wendroff, B. (1960) Systems of conservation laws. Comm. Pure Appl. Math.
13, 217237.
References 1213
[229] Lees, M. (1962) Alternating direction methods for hyperbolic dierential equations. J. Soc.
Indust. Appl. Math. 10, 610616.
[230] Lee, S. H. & Leal, L. G. (1986) Low-Reynolds-number ow past cylindrical bodies of
arbitrary cross-sectional shape. J. Fluid Mech. 164, 401427.
[231] Leider, P. J. & Bird, R. B. (1974) Squeezing ow between parallel disks. I. Theoretical
analysis. Ind. Eng. Che. Fundam. 13, 336341.
[232] Leonard, B. P., Leschziner, M. A. & McGuirk, J. (1978) Third-order nite-dierence
method for steady two-dimensional convection. Numerical Methods in Laminar and Turbulent
Flow, Proc. First Int. Conf. Swansea, 807819, Pentech.
[233] Leonard, A. (1980) Vortex methods for ow simulation. J. Comp. Phys. 37, 289335.
[234] Leonard, A. (1985) Computing three-dimensional incompressible ows with vortex elements.
Annu. Rev. Fluid Mech. 17, 523559.
[235] Levich, V. G. (1962) Physicochemical Hydrodynamics. Prentice-Hall.
[236] Li, X. Z., Barthes-Biesel, D. & Helmy, A. (1988) Large deformation and burst of a
capsule freely suspended in an elongational ow. J. Fluid Mech. 187, 179196.
[237] Li, X. & Pozrikidis, C. (1996) Shear ow over a liquid drop adhering to a solid surface. J.
Fluid Mech. 307, 167190.
[238] Lighthill, M. J. (1963) Attachment and separation in three-dimensional ow. In Laminar
Boundary Layer Theory, Rosenhead, L. (Ed.), Chap. II, Sec. 2.6, pp. 7282, Oxford University
Press.
[239] Lighthill, M. J. (1975) Mathematical Biouiddynamics. SIAM.
[240] Lighthill, M. J. (1976) Flagellar hydrodynamics, SIAM Rev. 18, 161230.
[241] Lin, C. C. (1955) The Theory of Hydrodynamic Stability. Cambridge University Press.
[242] Lo, L. L. (1983) The meniscus on a needlea lesson in matching. J. Fluid Mech. 132, 6578.
[243] Loewenberg, M. (1994) Axisymmetric unsteady Stokes ow past an oscillating nite-length
cylinder. J. Fluid Mech. 265, 265288.
[244] Longuet-Higgins, M. S. & Cokelet, E. D. (1976) The deformation of steep surface waves
on water. I. A numerical method of computation. Proc. R. Soc. Lond. A 350, 126.
[245] Lorentz, H. A. (1907) Ein allgemeiner Satz, die bewegung einer reibenden ussigkeit betr-
eend, nebst einigen anwendungen desselben. Abhand. Theor. Phys. (Leipzig) 1, 2342. Trans-
lated into English: Lorentz, H. A. Collected Papers Vol. IV, 714, Martinus Nijho.
[246] Love, A. E. H. (1893) On the stability of certain vortex motions. Proc. Lond. Math. Soc. 25,
1842.
[247] Love, A. E. H. (1944) A Treatise on the Mathematical Theory of Elasticity. Dover.
1214 Introduction to Theoretical and Computational Fluid Dynamics
[248] Lugt, H. J. (1987) Local ow properties at a viscous free surface. Phys. Fluids 30, 36473652.
[249] Luo, H. & Pozrikidis, C. (2008) Eect of surface slip on Stokes ow past a spherical particle
in innite uid and near a plane wall. J. Eng. Math. 62, 121.
[250] MacCormack, R. W. (1969) The eect of viscosity in hypervelocity impact cratering. AIAA
Paper 69-354.
[251] Malik, M. R. (1986) The neutral curve for stationary disturbances in rotating-disk ow. J.
Fluid Mech. 164, 275287.
[253] Mager, A. (1964) Three-dimensional laminar boundary layers. In: Theory of Laminar Flows,
Moore F. K. (Ed.), Princeton University Press.
[255] Mansour, N. N. & Lundgren, T. S. (1990) Satellite formation in capillary jet breakup.
Phys. Fluids A 2, 11411144.
[256] Mansour, M. L. & Hamed, A. (1990) Implicit solution of the incompressible NavierStokes
equations on a non-staggered grid. J. Comp. Phys. 86, 147167.
[258] Maxey, M. R. & Riley, J. J. (1983) Equation of motion for a small rigid sphere in a
nonuniform ow. Phys. Fluids 26, 883889.
[259] Moore, D. W. (1963) The boundary layer on a spherical gas bubble. J. Fluid Mech. 16,
161-176.
[261] McCraken, M. F. & Peskin, C. S. (1980) A vortex method for blood ow through heart
valves. J. Comp. Phys. 35, 183205.
[263] Mei, C. C. (1966) Nonlinear gravity waves in a thin sheet of viscous uid. J. Math. Phys. 45,
266288.
[264] Michalke, A. (1964) On the inviscid instability of the hyperbolic-tangent velocity prole. J.
Fluid Mech. 19, 543556.
[265] Milinazzo, F. & Saffman, P. G. (1977) The calculation of large Reynolds number two-
dimensional ow using discrete vortices with random walks. J. Comp. Phys. 23, 380392.
References 1215
[266] Milinazzo, F. & Shinbrot, M. (1988) A numerical study of a drop on a vertical wall. J.
Colloid Interf. Sci. 121, 254264.
[267] Milne-Thomson, L. M. (1940) Hydrodynamic images. Proc. Camb. Phil. Soc. 36, 246247.
[269] Mitchell, A. R. & Griffiths, D. F. (1980) The Finite-Dierence Method in Partial Dif-
ferential Equations. Wiley.
[270] Mitchell, T. B. & Rossi L. F. (2008) The evolution of Kirchho elliptic vortices. Phys.
Fluids 449, 054103.
[271] Moffatt, H. K. (1964) Viscous and resistive eddies near a sharp corner. J. Fluid Mech. 18,
118.
[272] Moffatt, H. K. & Duffy, B. R. (1980) Local similarity solutions and their limitations. J.
Fluid Mech. 96, 299313.
[273] Moffatt, H. K. & Tsinober, A. (1992) Helicity in laminar and turbulent ow. Annu. Rev.
Fluid Mech. 24, 281312.
[274] Murdock, J. N. & Stewartson, K. (1977) Spectra of the Orr-Sommerfeld equation. Phys.
Fluids 20, 14041411.
[275] Nakayama, T. (1990) A computational method for simulating transient motions of an in-
compressible inviscid uid with a free surface. Int. J. Num. Meth. Fluids 10, 683695.
[279] Ng, B. S. & Reid, W. H. (1979) An initial value method for eigenvalue problems using
compound matrices. J. Comp. Phys. 30, 125136.
[281] Newhouse, L. A. & Pozrikidis, C. (1992) The capillary instability of annular layers and
liquid threads. J. Fluid Mech. 242, 193209.
[282] Norbury, J. (1973) A family of steady vortex rings. J. Fluid Mech. 57, 417431.
[283] Orr, W. M. F. (1907) The stability or instability of the steady motions of a perfect liquid
and of a viscous liquid; Part I: A perfect liquid; Part II: A viscous liquid. Proc. R. Irish Acad.
A 27, 9138.
[284] Orszag, S. A. (1971) Accurate solution of the Orr-Sommerfeld stability equation. J. Fluid
Mech. 50, 689703.
1216 Introduction to Theoretical and Computational Fluid Dynamics
[285] Orszag, S. A., Israeli, M. & Deville, M. O. (1986) Boundary conditions for incompress-
ible ows. J. Sci. Comp. 1, 75111.
[286] Orszag, S. A. & Israeli, M. (1974) Numerical simulation of viscous incompressible ows.
Annu. Rev. Fluid Mech. 5, 281318.
[288] Osswald, G. A., Ghia, K. N. & Ghia, U. (1987) A direct algorithm for solution of incom-
pressible three-dimensional unsteady NavierStokes equations. AIAA Paper 871139, 408421.
[289] Oswatitsch, K. (1958) Die Ablosungsbedingung von Grenzschichten. In: IUTAM Symposium
on Boundary Layer Research, Freiburg, 1957, Gortler, H. (Ed.), 357367, SpringerBerlin.
[291] Peaceman, D. W. & Rachford, H. H. Jr. (1955) The numerical solution of parabolic and
elliptic dierential equations. J. Soc. Indust. Appl. Math. 3, 2841.
[292] Peregrine, D. H. (1981) The fascination of uid mechanics. J. Fluid Mech. 106, 5980.
[293] Petrila, T. & Trif, D. (2005) Introduction to Computational Fluid Dynamics. Springer.
[294] Peyret, R. & Taylor, T. D. (1983) Computational Methods for Fluid Flow. Springer
Verlag.
[295] Phan-Thien, N. & Kim, S. (1994) Microstrutures in Elastic Media. Oxford University Press.
[296] Phillips, W. R. C. (1996) On a class of unsteady boundary layers of nite extent. J. Fluid
Mech. 319, 151170.
[297] Phillips, W. R. C. (1997) On the spreading radius of surface tension driven oil on deep
water. Appl. Sci. Res. 57, 6780.
[298] Pierson, F. W. & Whitaker, S. (1977) Some theoretical and experimental observations of
the wave structure of falling liquid lms. Ind. Eng. Chem. Fund. 16, 401408.
[299] Plesset, M. S. & Prosperetti, A. (1977) Bubble dynamics and cavitation. Annu. Rev.
Fluid Mech. 9, 145185.
[300] Pogorzelski, W. (1966) Integral Equations and their Applications. Pergamon Press.
[301] Pohlhausen, E. (1921) Zur naherungsweisen integration der dierentialgleichung der lam-
inaren grenzschicht. Z. Angew. Math. Mech. 1, 252268.
[302] Pozrikidis, C. (1986) The nonlinear instability of Hills spherical vortex. J. Fluid Mech. 168,
337367.
[303] Pozrikidis, C. (1989) A singularity method for unsteady linearized ow. Phys. Fluids A 1,
15081520.
References 1217
[304] Pozrikidis, C. (1989) A study of linearized oscillatory ow past particles by the boundary
integral method. J. Fluid Mech. 202, 1741.
[305] Pozrikidis, C. (1990) The instability of moving viscous drops. J. Fluid Mech. 210, 121.
[306] Pozrikidis, C. (1992) Boundary Integral and Singularity Methods for Linearized Viscous
Flow. Cambridge University Press, Cambridge.
[307] Pozrikidis, C. (1993) Unsteady viscous ow over irregular boundaries. J. Fluid Mech. 255,
1134.
[308] Pozrikidis, C. (1995) A bibliographical note on the unsteady force on a spherical drop. Phys.
Fluids 6, 3209.
[309] Pozrikidis, C. (1997) Shear ow over a protuberance on a plane wall. J. Eng. Math. 31,
2942.
[310] Pozrikidis, C. (2000) Instability of two annular layers or a liquid thread bounded by an
elastic membrane. J. Fluid Mech. 405, 211241.
[311] Pozrikidis, C. (2001) Theoretical and computational aspects of the motion of three-
dimensional vortex sheets. J. Fluid Mech. 425, 335366.
[313] Pozrikidis, C. (2002) A Practical Guide to Boundary-Element Methods with the Software
Library BEMLIB. Francis & Taylor/CRC, Boca Raton.
[314] Pozrikidis, C. (2003) Eect of surfactants on lm ow down a periodic wall. J. Fluid Mech.
496, 105127.
[315] Pozrikidis, C. (2003) Numerical simulation of the ow-induced deformation of red blood
cells. Annals Biomed. Eng. 31, 11941205.
[316] Pozrikidis, C. (2003) On the relation between the pressure and the projection function for
the numerical computation of incompressible ow. Eur. J. Mech. B/Fluids 22, 105121.
[317] Pozrikidis, C. (2008) Numerical Computation in Science and Engineering. Second Edition.
Oxford University Press.
[318] Pozrikidis, C. (2009) Fluid Dynamics: Theory, Computation, and Numerical Simulation.
Second Edition. Springer.
[319] Pozrikidis, C. (2009) Shape of hexagonal hydrostatic menisci. Int. J. Num. Meth. Fluids
DOI 10.1002/d.2200.
[321] Pozrikidis, C. (2010) Hydrostatic meniscus between two eccentric circular cylinders. J. Col-
loid Interf. Sci. 349, 366373.
[322] Pozrikidis, C. (2011) Stability of sessile and pendant liquid drops. J. Eng. Math. DOI
10.1007/s10665-011-9459-3.
[323] Pozrikidis, C. & Higdon, J. J. L. (1985) Nonlinear Kelvin-Helmholtz instability of a nite
vortex layer. J. Fluid Mech. 157, 225263.
[324] Pozrikidis, C. & Higdon, J. J. L. (1987) Instability of compound vortex layers and wakes.
Phys. Fluids 30, 29652975.
[325] Pozrikidis, C. & Hill, A. I. (2011) Surfactant-induced instability of a sheared liquid layer
IMA J. Appl. Math. doi: 10.1093/imamat/hxq067.
[326] Pozrikidis, C. & Juneja, V. (2010) Eect of surface viscosity on the capillary instability
of an annular layer or viscous thread. IMA J. Appl. Math. 75, 932950.
[327] Prandtl, L. (1904) Uber Flussigkeitsbewegung bei sehr kleiner Reibung, Verh. III. Int. Math.
Kongr., Heidelburg, 484491. (Translated into English in NACA Tech. Mem. 452.)
[328] Puckett, E. G. (1993) Vortex methods: An introduction and survey of selected research
topics. In Incompressible Computational Fluids Dynamics: Trends and Advances. Cambridge
University Press.
[329] Pullin, D. I. (1981) The nonlinear behaviour of a constant vorticity layer at a wall. J. Fluid
Mech. 108, 401421.
[330] Pullin, D. I. (1992) Contour dynamics methods. Annu. Rev. Fluid Mech. 24, 89115.
[331] Quartapelle, L. (1981) Vorticity conditioning in the computation of two-dimensional viscous
ows. J. Comp. Phys. 40, 453477.
[332] Quartapelle, L. & Valz-Gris, F. (1981) Projection conditions on the vorticity in viscous
incompressible ows. Int. J. Num. Meth. Fluids. 1, 129144.
[333] Quartapelle, L. & Napolitano, M. (1986) Integral conditions for the pressure in the
computation of incompressible viscous ows. J. Comp. Phys. 62, 340348.
[334] Rayleigh, L. (1879) On the instability of jets. Proc. Lond. Math. Soc. 10, 413.
[335] Rayleigh, L. (1880) On the stability or instability of certain uid motions. Proc. Lond. Math.
Soc. 11, 5770.
[336] Rayleigh, L. (1883) Investigation of the character of the equilibrium of an incompressible
heavy uid of variable density. Proc. Lond. Math. Soc. 14, 170177.
[337] Rayleigh, L. (1894) The Theory of Sound. Macmillan (Reprinted by Dover 1945).
[338] Rayleigh, L. (1916) On the dynamics of revolving uids. Proc. R. Soc. Lond. A 93, 148154.
[339] Ricca, R. L. (1991) Rediscovery of Da Rios equations. Nature 352, 561562.
References 1219
[340] Richardson, S. (1973) On the no-slip boundary condition. J. Fluid Mech. 59, 707719.
[342] Richtmyer, R. D. (1963) A survey of dierence methods for non-steady uid dynamics.
NCAR Technical Notes 63-2.
[343] Richtmyer, R. D. & Morton, K. W. (1967) Dierence Methods for Initial-Value Problems.
Interscience.
[344] Riera, G., Carrasco, H. & Preiss R. (2008) The SchwarzChristoel conformal mapping
for Polygons with innitely many sides. Int. J. Maths. Math. Sci., No 350326.
[345] Riley, N. (1992) Acoustic streaming about a cylinder in orthogonal beams. J. Fluid Mech.
242, 387394.
[347] Roberts, S. G. (1985) Accuracy of the random vortex methods for a problem with non-
smooth initial conditions. J. Comp. Phys. 58, 2943.
[348] Rosenhead, L. (1930) The spread of vorticity in the wake behind a cylinder. Proc. R. Soc.
Lond. A 127, 590612.
[349] Rosenhead, L. (1931) The formation of vortices from a surface of discontinuity. Proc. R.
Soc. Lond. A 134, 170192.
[350] Rosenhead, L. (1940) The steady two-dimensional radial ow of viscous uid between two
inclined plane walls. Proc. R. Soc. Lond. A 175, 436467.
[351] Rosenhead, L. (1954) A discussion on the rst and second viscosities of uids. Proc. R. Soc.
Lond. A 226, 165.
[353] Rott, N. (1956) Unsteady viscous ow in the vicinity of a stagnation point. Quart. Apl.
Math. 13, 444451.
[354] Rott, N. (1958) On the viscous core of a line vortex. ZAMP 9, 543553.
[356] Rynhart, P. R., McLachlan, R., Jones, J. R. & McKibbin, R. (2003) Solution of the
Young-Laplace equation for three particles. Res. Lett. Math. Sci. 5, 119127.
[358] Saffman, P. G. & Baker, G. R. (1979) Vortex interactions. Annu. Rev. Fluid Mech. 11,
95122.
1220 Introduction to Theoretical and Computational Fluid Dynamics
[359] Saffman, P. G. & Taylor, G. I. (1958) The penetration of a uid into a porous medium
or HeleShaw cell containing a more viscous liquid. Proc. R. Soc. Lond. A 245, 312329.
[361] Sakiadis, B. C. (1961) Boundary-layer behavior on continuous solid surfaces: II The bound-
ary layer on a continuous at surface. AIChE J. 7, 221225.
[362] Schaaf, S. A. & Chambre. P. L. (1961) Flow of Rareed Gases. Princeton University
Press.
[364] Schmid, P. J. & Henningson, S. (2001) Stability and Transition in Shear Flows. Springer
Verlag.
[367] Shariff, K. & Leonard, A. (1992) Vortex rings. Annu. Rev. Fluid Mech. 24, 235279.
[368] Shariff, K., Leonard, A. & Ferziger, J. H. (1989) Dynamics of a class of vortex rings.
NASA TM-102257.
[369] Shen, S. F. (1954) Calculated amplied oscillations in plane Poisseuille and Blasius ows. J.
Aero. Sc. 21, 6264.
[370] Shercliff, J. A. (1975) A note on vorticity. J. Inst. Maths. Applics. 16, 259262.
[372] Shtern, V. & Hussain, F. (1993) Azimuthal instability of divergent ows. J. Fluid Mech.
256, 535560.
[373] Singh, P. & Hesla, T. I. (2004) The interfacial torque on a partially submerged sphere. J.
Colloid Interf. Sci. 280, 542543.
[376] Smith, F. T. (1982) On the high Reynolds number theory of laminar ows. IMA J. Appl.
Math. 28, 207281.
[377] Sod, G. A. (1985) Numerical Methods in Fluid Fynamics. Initial and Initial Boundary-Value
Problems. Cambridge University Press.
References 1221
[378] Sommerfeld, A. (1908) Ein Beitrag zur hydrodynamischen Erklarung der turbulenten
Flussigkeitsbewegung. Proc. Fourth Int. Cong. Math. Rome, 116124.
[379] Sotiropoulos, F. & Abdallah, S. (1990) Coupled fully implicit solution procedure for the
steady incompressible NavierStokes equation. J. Comp. Phys. 87, 328348.
[380] Sotiropoulos, F. & Abdallah, S. (1991) The discrete continuity equation in primitive
variable solutions of incompressible ow. J. Comp. Phys. 95, 212227.
[381] Sotiropoulos, F. & Abdallah, S. (1992) A primitive variable method for the solution of
three-dimensional incompressible viscous ow. J. Comp. Phys. 103, 336349.
[382] Sparrow, E. M. (1962) Laminar ow in isosceles triangular ducts. AIChE J. 8, 599604.
[383] Speziale, C. G. (1987) On the advantages of the vorticityvelocity formulation of the equa-
tions of uid dynamics. J. Comp. Phys. 73, 476480.
[384] Squires, H. B. (1933) On the stability for three-dimensional disturbances of viscous ow
between parallel walls. Proc. R. Soc. Lond. A 142, 621628.
[385] Squires, H. B. (1951) The round laminar jet. Quart. J. Mech. Appl. Math. 4, 321329.
[386] Stakgold, I. (1967) Boundary Value Problems of Mathematical Physics. Volume I. Macmil-
lan.
[387] Stakgold, I. (1968) Boundary Value Problems of Mathematical Physics. Volume II. Macmil-
lan.
[388] Stewartson, K. (1954) Further solutions of the Falkner-Skan equation. Proc. Camb. Phil.
Soc. 50, 454465.
[389] Struik, D. J. (1961) Lectures on Classical Dierential Geometry. Dover.
[390] Stuart, J. T. (1959) The viscous ow near a stagnation point when the external ow has
uniform vorticity. J. Aero/Space Sc. 26, 310311.
[391] Stuart, J. T. (1963) Unsteady boundary layers. In Laminar Boundary Layers, Rosenhead,
L. (Ed.), Oxford University Press.
[392] Sutera, S. P. & Skalak, R. (1993) The history of Poiseuilles law. Annu. Rev. Fluid. Mech.
25, 119.
[393] Strikwerda, J. C. & Lee, Y. S. (1999) The accuracy of the fractional step method, SIAM
J. Numer. Anal. 37, 3747.
[394] Tafti, D. (1995) Alternate formulations for the pressure equation Lplacian in a collocated
grid for solving the unsteady incompressible NavierStokes equations. J. Comp. Phys. 116,
143153.
[395] Tani, I. (1977) History of boundary-layer theory. Annu. Rev. Fluid Mech. 9, 87111.
[396] Tanner, R. I. (1988) Engineering Rheology. Oxford University Press.
1222 Introduction to Theoretical and Computational Fluid Dynamics
[397] Taylor, G. I. (1923) On the decay of vortices in a viscous uid. Phil. Mag. 46, 671674.
[398] Taylor, G. I. (1923) Stability of a viscous liquid contained contained between two rotating
cylinders. Phil. Trans. R. Soc. Lond. A 223, 289343.
[399] Taylor, G. I. (1950) The instability of liquid surfaces when accelerated in a direction per-
pendicular to their planes. Proc. R. Soc. Lond. A 201, 192196.
[400] Taylor, G. I. (1960) Deposition of a viscous uid on a plane surface. J. Fluid Mech. 9,
218224.
[401] Temam, R. (1968) Une methode dapproximation de la solution des equations de Navier
Stokes. Bull. Soc. Math. France 96, 115152.
[402] Thwaites, B. (1949) Approximate calculation of the laminar boundary layer. Aeronaut.
Quart. 1, 245280.
[403] Thomas, L. H. (1953) The stability of plane Poseuille ow. Phys. Rev. 91, 780783.
[404] Thomson, W. (1880) On the vibrations of a columnar vortex. Phil. Mag. 10, 155168.
[405] Tobak, M. & Peake, D. J. (1982) Topology of three-dimensional separated ows. Ann.
Rev. Fluid Mech. 14, 6185.
[406] Trefethen, L. N. (1980) Numerical computation of the Schwarz-Christoel transformation.
SIAM J. Sci. Stat. Comput. 1, 82102.
[407] Trujillo, J. R. (1994) Spectral Element Vorticity-Velocity Algorithm for the Incompressible
NavierStokes Equations. Doctoral dissertation, Princeton University.
[408] Truesdell, C. (1974) The meaning of viscometry in uid dynamics. Annu. Rev. Fluid. Mech.
6, 111146.
[409] Tryggvason, G. (1988) Numerical simulations of the Rayleigh-Taylor instability. J. Comp.
Phys. 75, 253283.
[410] Tomotika, S. (1935) On the instability of a cylindrical thread of a viscous liquid surrounded
by another viscous liquid, Proc. R. Soc. Lond. A 150, 322337.
[411] Tung, T. T. & Laurence, R. L. (1975) A coordinate frame for helical ows. Polym. Eng.
Sci. 15(6), 401405.
[412] Van Dyke, M. D. (1975) Perturbation Methods in Fluid Mechanics. Academic Press.
[413] Vojir, D. R. & Michaelides, E. E. (1994) Eect of the history term on the motion of rigid
spheres in a viscous uid. Int. J. Multiph. Flow 20, 547556.
[414] Von Karman, T. (1934) Some aspects of the turbulence problem. Proc. Fourth Int. Congr.
Appl. Mech., Cambridge, England.
[415] Von Karman, T. (1921) Uber laminare und turbulente Reibung. Z. Angew. Math. Mech. 1,
233252.
References 1223
[417] Wang, C. Y. (1981) On the low-Reynolds-number ow in a helical pipe. J. Fluid Mech. 108,
185194.
[418] Wang, C. Y. (1984) The three-dimensional ow due to a stretching at surface. Phys. Fluids
27, 19151917.
[419] Wang, C. Y. (1989) Exact solutions of the unsteady NavierStokes equations. Appl. Mech.
Rev. 42, S269S282.
[420] Wang, C. Y. (1991) Exact solutions of the steady-state NavierStokes equations. Annu. Rev.
Fluid Mech. 23, 159177.
[421] Wang, C. Y. (2008) Similarity stagnation point solutions of the NavierStokes equations-
review and extension. Eur. J. Mech. B/Fluids 27, 678683.
[422] Wang, C. Y. (2010) The rounded triangular cross sectionExact solutions for torsion, ow
and heat transfer. Z. Angew. Math. Mech. 90, 522527.
[423] Wang, C. Y. (2011) Review of similarity stretching exact solutions of the NavierStokes
equations. Eur. J. Mech. B/Fluids 30, 475479.
[424] Warming, R. F. & Hyett, B. J. (1974) The modied equation approach to the stability
and accuracy analysis of nite-dierence methods. J. Comp. Phys. 14, 159179.
[425] Watson, J. (1959) The two-dimensional laminar ow near the stagnation point of a cylinder
which has an arbitrary transverse motion. Quart. J. Mech. 12, 175190.
[428] Whitham, G. B. (1963) The NavierStokes equations of motion. In Laminar Boundary Lay-
ers, Rosenhead, L. (Ed.), Oxford University Press.
[429] Wilkinson, J. H. (1965) The Algebraic Eigenvalue Problem. Oxford University Press.
[430] Wilson, S. D. R. (1982) The drag-out problem in lm coating theory. J. Eng. Math. 16,
209221.
[432] Wu, J. Z. & Wu, J. M. (1996) Vorticity dynamics on boundaries. Adv. Appl. Mech. 32,
119275.
1224 Introduction to Theoretical and Computational Fluid Dynamics
[433] Wu, J.-Z., Yang, Y.-T, Luo, Y.-B. & Pozrikidis, C. (2005) Fluid kinematics on a
deformable surface. J. Fluid Mech. 541, 371381.
[438] Yon, S. & Pozrikidis, C. (1998) A nite-volume / boundary-element method for interfa-
cial ow in the presence of surfactants, with applications to shear ow past a viscous drop.
Computers & Fluids 27, 879902.
[439] Yu, Q. & Hu, G.-H. (1997) Development of a helical coordinate system and its application to
analysis of polymer ow in screw extruders, Part I. The balance equations in a helical coordinate
system. J. Non-Newt. Fluid Mech. 69, 155167.
[440] Zabusky, N. J., Hughes, M. H. & Roberts, K. V. (1979) Contour dynamics for the Euler
equations in two dimensions. J. Comp. Phys. 30, 96106.
[441] Zheng, Q.-S., Yu, Y. & Zhao, Z.-H. (2005) Eects of hydraulic pressure on the stability
and transition of wetting modes of superhydrophobic surfaces. Langmuir 21, 1220712212.
[442] Zhou, H. & Pozrikidis, C. (1995) Adaptive singularity method for Stokes ow past particles.
J. Comp. Phys. 117, 7889.
[443] Zhu, J. & Sethian, J. (1992) Projection methods coupled to level set interface techniques.
J. Comp. Phys. 102, 128138.
[444] Zwillinger, D. (2002) CRC Standard Mathematical Tables and Formulae. 31st Edition.
Chapman & Hall/CRC.
Index
1225
1226 Introduction to Theoretical and Computational Fluid Dynamics