Professional Documents
Culture Documents
A Discontinuous Galerkin Method With Splitting Applied To Visco-Elastic Flow by Martien A. Hulsen
A Discontinuous Galerkin Method With Splitting Applied To Visco-Elastic Flow by Martien A. Hulsen
Martien A. Hulsen
1 Introduction 1
3 Numerical methods 9
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.2 A DG-method for hyperbolic systems . . . . . . . . . . . . . . . . . . 9
3.2.1 Discretisation of the weak form . . . . . . . . . . . . . . . . . 9
3.2.2 Characteristic decomposition and splitting . . . . . . . . . . . 11
3.2.3 Initial and boundary conditions . . . . . . . . . . . . . . . . . 12
3.2.4 The hyperbolic visco-elastic system . . . . . . . . . . . . . . . 13
3.3 A DG-method for an incompressible visco-elastic system . . . . . . . 13
3.3.1 The incompressible limit . . . . . . . . . . . . . . . . . . . . . 13
3.3.2 Removing the irregularities for the incompressible limit . . . 15
3.3.3 Discretisation of the weak form . . . . . . . . . . . . . . . . . 16
3.3.4 A remark on boundary conditions . . . . . . . . . . . . . . . 19
3.4 A mixed method for viscous terms . . . . . . . . . . . . . . . . . . . 20
3.5 Numerical quadrature . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.6 Time integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.6.1 Explicit methods . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.6.2 Mixed explicit/implicit methods . . . . . . . . . . . . . . . . 24
3.7 Limit behaviour . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.7.1 No viscoelasticity . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.7.2 Creeping flow . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.7.3 Small relaxation times . . . . . . . . . . . . . . . . . . . . . . 27
3.7.4 Large relaxation times . . . . . . . . . . . . . . . . . . . . . . 27
3.8 Handling negative values of c2s . . . . . . . . . . . . . . . . . . . . . . 28
ii
E Forces and torques on a rigid body submerged in a fluid 52
Bibliography 54
iii
Chapter 1
Introduction
The simulation of viscoelastic fluid flow has become an active research area in
recent years (Crochet & Walters 1993). The majority of articles contain simulations
of creeping steady flow, motivated by the importance of industrial processing of
polymer melts (see, for example, Hulsen & van der Zanden 1991). For polymer
solutions however, it has become clear that inertia can have major influence on the
flow behaviour (Joseph 1990). At higher flow rates the flow becomes unsteady. In
order to study this behaviour time-dependent equations have to be solved.
Phelan et al. (1989) showed that allowing a weak compressibility into the equa-
tions of viscoelastic flow makes these equations hyperbolic (see also: Joseph 1990).
In that case the equations can formally be written in the following form
∂u X
d
∂u
˜+ Ai (u) ˜ + f (u) = 0 in (0, T ) × Ω, (1.1)
∂t i=1
¯ ˜ ∂xi ˜ ˜ ˜
∂u X ∂
d
˜+ F i (u) + f (u) = 0, (1.2)
∂t i=1
∂xi ˜ ˜ ˜ ˜ ˜
where u is the vector of conserved variables (specific mass, momentum and energy)
and F i˜, i = 1, . . . , d are the flux vectors with respect to all co-ordinate directions.
˜
Comparing (1.2) with (1.1) we see that1
∂F i
Ai (u) = ˜ , i = 1, . . . , d. (1.3)
¯ ˜ ∂u
˜
The majority of numerical methods for hyperbolic systems have been developed
with equation (1.2) in mind, including the possibility of shocks, i.e. discontinuous
solutions satisfying the conservation laws in integral form (Whitham 1974). Our
main application will be visco-elastic fluid flow. Although of course mass, momen-
tum and energy are still conserved, the full system including the stresses cannot be
written in conservation form (Hulsen 1986). Whether discontinuous solutions are
a physical reality for viscoelastic flows is far from a resolved problem and we will
assume that for the general non-conservative system (1.1) the solutions are smooth.
To solve the equations numerically, Phelan et al. (1989) apply a finite differ-
ence method that was originally developed for the Euler equations but does not
1 We have assumed that A (u), F (u), and f (u) are functions of u only. An explicit dependence
i i
¯ the
on x presents no difficulty in ˜ following
˜ ˜ ˜ as the dependence
as ˜long ˜ is smooth.
1
require a conservative form. Unfortunately their method is only a first-order up-
winding method, which is known to be much too diffusive. Higher-order methods
have been developed (see, for example, Hirsch 1990), but typically within the frame
of conservative systems. These schemes achieve to be of higher order by consid-
ering the variables in multiple neighbouring cells. A different approach towards
higher-order methods is followed by Cockburn & Shu (1989). They use the discon-
tinuous Galerkin method with higher-order polynomials. Since their work is still
within the framework of conservative hyperbolic systems, Hulsen (1992) has adapted
the method to non-conservative systems with smooth solutions. We should note,
however, that the equations for viscoelastic flow in the usual form are not fully
hyperbolic because
2
Chapter 2
The basic equations describing the flow of viscoelastic fluids consist of the basic
laws of continuum mechanics and the constitutive equation describing a particular
fluid. To simplify the basic equations, we assume that the flow is incompressible
and isothermal. The last assumption means that we do not have to consider the
energy equation.
3
The function f m is an isotropic tensor function
where
1
g m (bm ) = f m Gm (bm − 1) . (2.11)
Gm
For very fast deformation rates, kλm Lk 1, the tensors bm approach the Finger
deformation tensor. Equation (2.9) means that the stress-strain relation is given by
a neo-Hookean model2 . For small and moderate deformation rates the tensors bm
can be interpreted as a ‘reversible strain’ of mode m. In this report the formulation
in bm will be preferred over the τ m formulation.
1 Model having a Gordon-Schowalter derivative instead of the upper-convected derivative cannot
be written in the form (2.5). The same is true for models like Larson’s differential model (Larson
1988) and the FENE-type model of Chilcott & Rallison (1988).
2 Models having a different non-linear stress-strain law are, for example, Larson’s differential
model (Larson 1988) and the FENE-type model of Chilcott & Rallison (1988).
4
2.4 Eigenvalues and eigenvectors
We now try to write our equations into the form (1.1). Combining the equations of
sections 2.1 and 2.2 leads to the following system for (u, p, b1 , b2 , . . . , bM )
∂u 1 1X
+ u · grad u + grad p − Gm div bm − νs ∆u − f = 0, (2.12a)
∂t ρ ρ m
div u = 0, (2.12b)
∂bm
+ u · grad bm − L · bm − bm · LT + g m (bm ) = 0, m = 1, . . . , M. (2.12c)
∂t
where νs = ηs /ρ, the kinematic viscosity of the solvent. It is easy to see that we miss
a term like ∂p/∂t in order to write system (2.12) into the form (1.1). Therefore we
introduce the concept of weak compressibility (Phelan et al. 1989; Edwards & Beris
1990; Joseph 1990) where equation (2.12b) is replaced by an equation for weakly
compressible flow. The new system now becomes
∂u 1 1X
+ u · grad u + grad p − Gm div bm − νs ∆u − f = 0, (2.13a)
∂t ρ ρ m
∂p
+ u · grad p + κ div u = 0, (2.13b)
∂t
∂bm
+ u · grad bm − L · bm − bm · LT + bm div u + g m (bm ) = 0, m = 1, . . . , M.
∂t
(2.13c)
P
where κ m Gm is a very large number: the artificial compression modulus. As
suggested by Edwards & Beris (1990) we included an extra term bm div u in (2.13c),
which vanishes for incompressible flow.
In the following we take ηs = 0 and write (2.13) into the following form
∂u ∂u ∂u ∂u
˜ + A(u) ˜ + B (u) ˜ + C (u) ˜ = l.o.t., (2.14)
∂t ¯ ˜ ∂x ¯ ˜ ∂y ¯ ˜ ∂z
where l.o.t. represents the lower-order terms and the system vector u is defined by
˜
uT = (u, v, w, p, bT1 , . . . , bTM ), with bTm = (bxx , byy , bzz , bxy , bxz , byz )m . (2.15)
˜ ˜ ˜ ˜
The Cartesian co-ordinates are denoted by (x, y, z). Expressions for the coefficient
matrices A(u), B (u) and C (u) are given in Appendix B. Next we consider small
¯ ˜ around
perturbations ¯ ˜ a basic ¯ ˜solution u0 :
˜
u = u0 + δu. (2.16)
˜ ˜ ˜
Substituting this into (2.14) and discarding O(u2 ) terms, we arrive at
˜
∂δu 0 ∂δu 0 ∂δu 0 ∂δu
˜ + A(u ) ˜ + B (u ) ˜ + C (u ) ˜ = l.o.t., (2.17)
∂t ¯ ˜ ∂x ¯ ˜ ∂y ¯ ˜ ∂z
where the lower-order terms linear in δu have been moved to the l.o.t.-terms. In
the following we simply abbreviate A(u˜0 ) by A.
For computation of eigenvalues and ¯
¯ ˜eigenvectors we only consider the first-order
terms of (2.17) and write down a new system where the l.o.t.-terms have been
discarded:
∂δu ∂δu ∂δu ∂δu
˜ + A ˜ + B ˜ + C ˜ = 0, (2.18)
∂t ¯ ∂x ¯ ∂y ¯ ∂z ˜
5
Next we look for local plane wave solutions of the form
δu = U eik(n·x−ct) , (2.19)
˜ ˜
where k is the wavenumber, n is the unit vector in the wave direction and c is
the wave speed. With ‘local’ we mean that the coefficient matrices in (2.18) are
considered to be constant with respect to the wavelength 2π/k. Substitution of
(2.19) into (2.18) leads to the following eigenvalue problem: find U 6= 0 such that
˜ ˜
(K n − cI )U = 0, (2.20)
¯ ¯ ˜ ˜
with
K n = nx A + ny B + nz C , (2.21)
¯ ¯ ¯ ¯
where nx , ny and nz are the components of the vector n in the Cartesian (x, y, z)
system. The solution for c can be found from
det(K n − cI ) = 0. (2.22)
¯ ¯
Note, that the matrix K n , and thus the eigenvalues and eigenvectors, depend on
¯
the wave direction n. Developing the determinant leads to the following equation
(see Appendix B)
where λ = u · n − c and the shear and compression wave speeds cs and cc , are given
by
1X
c2s = Gm b(m)
nn , (2.24)
ρ m
1X
c2c = κ/ρ + Gm b(m)
nn , (2.25)
ρ m
(m)
with bnn = n · bm · n. From equation (2.23) we obtain 6M + 4 solutions for c
6
See Appendix B for an expression of the eigenvectors.
Any vector a can now be decomposed into the base of ri or li vectors
˜ ˜ ˜
X
m T
li a
a= ˜ ˜ ri , (2.32)
˜ i=1 lTi ri ˜
˜ ˜
Xm
aT r i
a= ˜ ˜ li . (2.33)
˜ i=1 lTi ri ˜
˜ ˜
This is called characteristic decomposition of the vector a with respect to the right
and left eigenvectors respectively. The right eigenvector ˜decomposition will be the
most important for us and the quantity wi = lTi a is called the characteristic variable
corresponding to eigenvalue ci . ˜ ˜
We have only considered the first-order terms of the system (2.17). If the lower-
order terms are taken into account, the system is still called hyperbolic but waves are
damped or amplified independent of the wave length. Whenever νs 6= 0, the type of
the system changes to mixed parabolic-hyperbolic type (van der Zanden & Hulsen
1988) and the wave analysis given above is not applicable anymore. However, when
νs is small only short wave lengths are strongly damped and the wave analysis is
still useful for the lower wave lengths (Joseph 1990).
∂δu ∂δu
˜ + K n ˜ = 0. (2.34)
∂t ¯ ∂n ˜
After multiplying this equation with lTi we get
˜
∂δwi ∂δwi
+ ci = 0, i = 1, . . . , nv , (2.35)
∂t ∂n
with δwi = lTi δu. These equations become decoupled in the characteristic system.
˜ ˜ is an incoming wave c < 0, an entry conditions has to be known
Whenever there i
for δwi . This is discussed further in Appendix C. The results are summarised in
the following.
Hereafter the (x, y, x) co-ordinates are transformed such that the positive x
directions corresponds to the direction of n. We divide the boundary into the
following subtypes:
7
a. v, w, bm , m = 1, . . . , M , for supercritical inflow,
b. bm , m = 1, . . . , M for subcritical inflow,
c. v or τxy and w or τxz for subcritical outflow,
d. none for supercritical outflow.
Other combinations of b.c. are possible but we will not discuss these. On a su-
percritical inflow boundary all convection and shear waves are going in, i.e. all
variables must be prescribed: tangential velocities and the complete fluid memory
bm , 1, . . . , M . On an subcritical inflow boundary we lose the possibility to prescribe
the tangential velocities3 u, w and only the fluid memory bm , 1, . . . , M can be pre-
scribed. In two-dimensional flows the b.c. are very similar except that the variables
w and τxz disappear.
For the parabolic-hyperbolic system (νs 6= 0) less restrictions exist. For example
the following b.c. are possible
• u or σ = −p + txx on the complete boundary,
• v or txy and w or txz on the complete boundary,
• bm , m = 1, . . . , M on an inflow boundary u < 0.
If νs is small these b.c. may lead to steep boundary layers problems in order to
adapt to the prescribed variables that are not allowed in the hyperbolic case νs = 0.
For either νs = 0 or νs 6= 0 it is assumed that proper initial conditions are:
specification at the initial time of
• u, fulfilling div u = 0,
• bm , m = 1, . . . , M ,
for all x ∈ Ω.
3 Characteristic variables of the shear wave consist of a combination of tangential velocities and
shear stresses. We could have chosen to prescribe v and w but this would mean that the fluid
memory bm , 1, . . . , M could not be fully prescribed and be constrained in a rather complicated
way.
8
Chapter 3
Numerical methods
3.1 Introduction
The discontinuous Galerkin (DG) method originates from Lesaint & Raviart (1974)
who apply this method to the neutron transport equation. Various names exist for
this method: discontinuous finite element method (Thomasset 1981), discontinuous
Galerkin method (Johnson 1987), Lesaint-Raviart method (Fortin & Fortin 1989),
upwinding by discontinuity (Pironneau 1989) and discontinuous Galerkin finite el-
ement method (Cockburn & Shu 1989).
The method has been introduced successfully into the field of viscoelastic flow
by Fortin & Fortin (1989) for upwinding the advection term in the constitutive
equation using quadratic elements. Others have used the DG method in various
forms for solving viscoelastic flow problems (see, for example, Fortin & Fortin 1990,
Basombrı́o et al. 1991, Fortin & Zine 1992, Fortin et al. 1992, Baaijens 1994a).
The DG method together with explicit Runge-Kutta methods for the time-
discretisation has been applied to conservative hyperbolic systems by Cockburn
and his coworkers (Cockburn & Shu 1989; Cockburn et al. 1989; Cockburn et al.
1990). They also introduce the concepts of characteristic decomposition and split-
ting of element boundary fluxes and local projection limiters for resolving sharp
gradients, such as shocks, without oscillations. Their system of equations is of the
conservative form (1.2). Hulsen (1992) has extended the method of splitting for DG
methods to non-conservative systems. In this report we will discuss the application
to visco-elastic fluid flows. Limiters are not considered here. Since sharp stress
gradients are possible in viscoelastic flows (stress boundary layers, sharp corners)
it may however appear to be necessary to introduce limiters for these problems too
(Baaijens 1994b; Baaijens 1995).
where V is a suitable function space for both u and v , for example (H 1 (Ω))nv .
We divide the domain Ω into finite elements ˜ Ω ˜= ∪ne Ω . On each element
e=1 e
Ωe we approximate u by uh , which consists of polynomials of the order `. The
approximation uh is ˜allowed
˜ to be discontinuous across element boundaries. We
˜ (e) (e)
will denote this space by Vh . Note that Vh ⊂ (L2 (Ω))nv .
(e) (e)
We approximate the space V by Vh . Note that Vh is only an approximation
9
of the space V on element level1 . If the system can be written in a conservative
form, it is possible to integrate partially on element level. The boundary integrals
can be used then to define weak element boundary conditions based on fluxes of
neighbouring elements (Cockburn & Shu 1989; Hulsen 1992). For non-conservative
systems this is not possible and we have to proceed differently.
If we substitute the approximations uh , v h into (3.1) the integral has to be split
into a sum over element integrals and a ˜sum˜ over element boundary integrals. The
latter terms appear, because on the element boundaries the normal component of
∂uh /∂xi , i = 1, . . . , d
˜
∂u X ∂uh
d
˜ = ni ˜ , (3.2)
∂n i=1
∂xi
is infinite2 . We find
Z ! Z
Xne Xd
T ∂uh ∂uh
vh ˜ + Ai (uh ) ˜ − f (uh ) dΩ + v Th ∆n dγ = 0, (3.3)
e=1 Ωe ˜ ∂t i=1
¯ ˜ ∂xi ˜ ˜ γh ˜ ˜
where γh = ∪ne=1
e
∂Ωe , which consists of all element boundaries3 , and ∆n is given by
˜
Z n+ X d Z n+ Xd
∂uh ∂uh
∆n = Ai (uh ) ˜ dn = Ai (uh )ni ˜ dn
˜ n −
i=1
¯ ˜ ∂x i n−
i=1
¯ ˜ ∂n
Z u+ X d Z u+
˜ ˜
= ni Ai (u) du = K n (u) du, (3.4)
u− i=1 ¯ ˜ ˜ u− ¯ ˜ ˜
˜ ˜
with n a coordinate in the direction of the unit normal vector4 n on γh ; the indices
− and + denote at the ‘backside’ and ‘frontside’ of the vector n and the matrix K n
is given by ¯
X
d
K n (u) = ni Ai (u). (3.5)
¯ ˜ i=1
¯ ˜
Xd Z n+ Xd
∆n = ni dF i (u) = ni F i (u+ ) − F i (u− ) . (3.6)
˜ i=1 n
− ˜ ˜ i=1
˜ ˜ ˜ ˜
10
but others are possible. Note that for the conservative system this choice would be
O([u]3 ) and exact for linear K n (u). We drop the term O([u]2 ) and obtain
˜ ¯ ˜ ˜
Z !
Xne
∂uh X
d
∂uh
v Th ˜ + Ai (uh ) ˜ + f (uh ) dΩ
e=1 Ωe ˜ ∂t i=1
¯ ˜ ∂xi ˜ ˜
Z
+ v Th H n (u− , u+ )[u]d γ = 0. (3.9)
γh ˜ ¯ ˜ ˜ ˜
We still have to choose v h on γh . This choice will be based on characteristic de-
composition of H n (u− , u˜+ ).
¯ ˜ ˜
3.2.2 Characteristic decomposition and splitting
We define for i = 1, 2, . . . , nv
eigenvalues of H n (u− , u+ ),
ci :
¯ ˜ ˜
ri : right eigenvectors of H n (u− , u+ ),
˜ ¯ ˜ ˜
li : left eigenvectors of H n (u− , u+ ),
˜ ¯ ˜ ˜
− +
where H n (u , u ) is given by (3.8). We decompose [u] into characteristic compo-
¯ ˜ ˜ to (2.32)
nents according ˜
Xnv T
li [u]
[u] = ˜ T ˜ ri . (3.10)
˜ l r ˜
i=1 ˜i ˜i
Substitution of (3.10) into (3.9) and using H n (u− , u+ )ri = ci ri we obtain for the
integral on the boundary ¯ ˜ ˜ ˜ ˜
Z Z X
nv T
li [u]
v Th H n (u− , u+ )[u] dγ = v Th ˜ T ˜ ci ri dγ
γh ˜ ¯ ˜ ˜ ˜ γh ˜ i=1 l i ri ˜
Z X ˜ ˜
nv
= vih lTi [u]ci dγ, (3.11)
γh i=1 ˜ ˜
where vih is the characteristic component with respect to the li base of v h : vih =
v Th ri . Now we take ˜ ˜
˜ ˜
(
−
vih , if ci < 0,
vih = + (3.12)
vih , if ci > 0.
Reorganising (3.9) with (3.11) and (3.12) into a sum over elements and considering
each element independently we obtain the method: find for e = 1, 2, . . . , ne , uh ∈
(e) ˜
Vh such that
Z
∂uh X
d
∂uh
v Th ˜ + Ai (uh ) ˜ − f (u) dΩ
Ωe ˜ ∂t i=1
¯ ˜ ∂xi ˜ ˜
X Z
lT [u] (e)
+ v Th ˜iT ˜ ci ri dγ = 0 for all v h ∈ Vh , (3.13)
(in) ∂Ωe
˜ l i ri ˜ ˜
˜ ˜
where
X
= sum over all i with ci < 0, (3.14)
(in)
11
and n is the outward6 normal on ∂Ωe , which is needed to define [u] and the positive
direction for K n and ci . ˜
¯
The resulting discretised system can be written in the form
M U̇ + S (U , t) = 0, (3.15)
¯ ˜ ˜ ˜ ˜
where M is the mass matrix and U is a vector containing all unknowns. The
coupling¯ between elements only exists
˜ through the boundary integral terms. This
means that the mass matrix M can be inverted on element level and the system
may be explicitly written as ¯
then proper boundary conditions at xp are given by (Cockburn & Shu 1989)
w1 (xp , t) wq+1 (xp , t)
.. ..
. = B (xp , t) . + g (xp , t), (3.19)
¯ ˜
wq (xp , t) wnv (xp , t)
wi = lTi u, i = 1, . . . , nv . (3.20)
˜ ˜
Implementation of (3.19) in our scheme (3.13) is easy and given by Cockburn
et al. (1989) for a one-dimensional system but equally applies to more-dimensional
systems as well. We define
In the change to the ‘per element’ formulation (3.13), we choose the outward normal and thus n
is opposite in the two elements.
12
and implement (3.19) as follows
+ +
w1 wq+1
.. .
. = B .. + g , (3.22)
+
¯ ˜
wq wn+v
wi+ = wi− , for i = q + 1, . . . , nv . (3.23)
Note that (3.23) means that the plus signs in the right-hand side of (3.22) may be
changed to a minus sign. From lTi [u] = [wi ] the boundary integral in (3.13) can be
˜ ˜
computed on the complete boundary.
13
Compression waves
cc
0
lT [u] (u + cc )(ρcc [u] + [p] − [τxx ])
0
c1 ˜1T ˜ r1 =
l 1 r1 ˜ 2ρc2c κ
˜ ˜
..
.
[p] − [τxx ]
c c [u] +
ρ
0
0
cc |u| 1 κ
−−−−→
,
(3.27)
2 ρcc
0
..
.
0
and
−cc
0
lT2 [u] (u − cc )(−ρcc [u] + [p] − [τxx ])
0
c2 ˜ T ˜ r 2 =
l 2 r2 ˜ 2ρc2c κ
˜ ˜
..
.
[p] − [τxx ]
−cc [u] + ρ
0
0
cc |u| 1 κ
−−−−→
.
(3.28)
2 ρcc
0
..
.
0
We conclude from the equations given above that in the limit the irregularity comes
from the cc [u] and κ/ρcc terms.
Shear waves
From (B.22) and (B.23) we find that the shear waves are independent of κ. Hence
there is no irregularity in the incompressible limit. The eigenvalues and eigenvectors
are given in Appendix B and are not recalled here.
14
Convection
The limit for κ → ∞ is regular and is easy to compute. We find from (B.41) that
0
0
0
Xnv
lTi [u] [τxx ]
cc |u|
ci ˜ T ˜ ri = u[u]conv −−−−→ u d1 , (3.29)
l r
i i ˜ ˜ ˜
i=7 ˜ ˜ d
˜2
..
.
dM
˜
where dm , m = 1, . . . , M are given by
˜
(m)
[bxx ]
(m)
(m) bxy [τxy ]
[byy ] −
ρc2s
(m)
b [τ ]
(m)
[bzz ] −
xz xz
ρcs2
dm = (m)
(m)
bxx [τxy ] .
(3.30)
˜ [bxy ] − 2
ρcs
(m)
(m) b [τ ]
[b ] −
xx xz
xz
ρcs2
(m) bxz [τxy ] bxy [τxz ]
(m) (m)
[byz ] − −
ρc2s ρc2s
15
If the velocity is discretised by RTk and the pressure by polynomials of the order
k, i.e. the pressure belongs to the space Pk , the combined space RTk × Pk is a
proper mixed discretisation. The same is true for the RT[k] × Qk combination for
quadrilaterals. In practice we work with the reduced spaces RTk1 and RT[k] 1
where
the divergence is a constant per element. With these elements the number of internal
degrees of freedom are reduced and only one pressure variable per element remains
in the discretised system. Using the penalty method this pressure variable can be
eliminated as well7 .
For extensive discussion of these and other mixed element spaces we refer to the
literature cited above and to Roberts & Thomas (1991). In Appendix D we will
discuss some properties of the RTk and RT[k] spaces.
X Z ∂u
h
vh · ρ +ρuh · grad uh + grad p − div τ − ρf dΩ
e Ωe ∂t
Z !
1
+ v h · n([p] − [τnn ]) dγ + b.i.t. = 0, (3.32a)
2 ∂Ωe
Z
− qh div uh dΩ = 0, (3.32b)
Ωe
Z ∂b
m
βh : + uh · grad bm − L · bm − bm · LT + g m (bm ) dΩ
Ωe ∂t
+ b.i.t. = 0, m = 1, . . . , M. (3.32c)
c 3 = c5 = u · n + cs
c4 = c6 = u · n − cs
cconv = u · n
16
and
lT3 [u] 1 [v] [τxy ]
˜T ˜ = − , (3.35)
l 3 r3 2 cs ρc2s
˜ ˜
lT4 [u] 1 [v] [τxy ]
˜T ˜ = − − , (3.36)
l 4 r4 2 cs ρc2s
˜ ˜
lT5 [u] 1 [w] [τxz ]
˜T ˜ = − , (3.37)
l 5 r5 2 cs ρc2s
˜T ˜
l6 [u] 1 [w] [τxz ]
˜T ˜ = − − . (3.38)
l 6 r6 2 cs ρc2s
˜ ˜
The matrix R = [r3 , r4 , r5 , r6 ] is given by
¯ ˜ ˜ ˜ ˜
0 0 0 0
ρcs −ρcs 0 0
0 0 ρcs −ρcs
0 0 0 0
xy
R = H xy
H xz H xz , (3.39)
¯ ¯ 1xy H ¯ 1
¯ 1
¯ 1
H 2 H2 xy
H2 xz
H2 xz
¯ ¯. ¯. ¯.
..
. .
. .
. ..
H xy H xy H xz H xz
¯M ¯M ¯M ¯M
where H xy xz
and H m , m = 1, . . . , M are defined by
¯m ¯
0 0
−2b(m) 0
xy
0 −2b(m)
xz
H m = (m) ,
xy
Hm = 0 .
xz
(3.40)
˜ −bxx ˜
(m)
0 −bxx
(m) (m)
−bxz −bxy
Furthermore, we have
0
0
0
0
[u]conv = d1 , (3.41)
˜ ˜
d2
˜
..
.
dM
˜
where dm , m = 1, . . . , M are given by
˜ (m)
[bxx ]
(m)
(m) bxy [τxy ]
[byy ] −
ρc2s
(m)
b [τ ]
(m)
[bzz ] −
xz xz
ρcs2
dm =
(m)
(m)
bxx [τxy ] .
(3.42)
˜ [bxy ] −
ρc2s
(m)
(m) bxx [τxz ]
[b ] −
xz
ρcs2
(m) bxz [τxy ] bxy [τxz ]
(m) (m)
[byz ] − −
ρc2s ρc2s
17
In order to obtain a symmetric (v, p) system we partially integrate the grad p term
in (3.32a) on element level and obtain for the pressure terms
XZ XZ 1
··· − p div v h dΩ + v h · n (p+ + p− ) dγ + · · · (3.43)
e Ωe e ∂Ωe 2
Since the normal velocities are continuous across element boundaries the internal
boundary integrals cancel out and the pressure terms become
XZ Z
··· − p div v h dΩ + v h · npb dγ + · · · (3.44)
e Ωe ∂Ω
where pb = 12 (p+ +p− ) is the pressure on the boundary. A similar partial integration
can be carried out for the div τ term. The terms involving the stress τ become
XZ XZ
···+ τ : (grad v h )T dΩ + n · τ · v h − (n · v h )n dγ
e Ωe e ∂Ωe
Z
b
− v h · nτnn dγ + · · · , (3.45)
∂Ω
where ∂Ωint
e denotes internal element boundaries. The weak system of equation now
becomes: find (uh , ph , b1 , . . . , bM ) ∈ Uh × Qh × (Bh )M such that for all elements
Ωe
X Z ∂u
h
Z
vh · ρ + ρuh · grad uh − div τ − ρf dΩ − ph div v h dΩ
e Ωe ∂t Ωe
Z ! Z
1
− v h · n[τnn ] dγ + b.i.t. − v h · n(−pb + τnn
b −
− τnn ) dγ = 0, (3.47a)
2 ∂Ωint
e ∂Ω
Z
− qh div uh dΩ = 0, (3.47b)
Ωe
Z ∂b
m
βh : + uh · grad bm − L · bm − bm · LT + g m (bm ) dΩ
Ωe ∂t
+ b.i.t. = 0, m = 1, . . . , M. (3.47c)
18
which means that of all pressure degrees of freedom, only one per element remains,
for example
R Z
Ωe ph dΩ
p̄h = , Ae = dΩ. (3.49)
Ae Ωe
Equation (3.47b) also reduces to one equation per element, which is a constraint on
the normal velocities on the boundary.
After numerical integration of the integrals we obtain a system of the following
form
where p is a small number, ∆t is the time step in the time integration scheme
(see section 3.6) and p̄rh is a reference pressure, for which we will usually take the
pressure at the previous time step. From (3.51) we find that
Z Z
∆tp (p̄h − p̄rh ) dΩ + div uh dΩ = 0, (3.52)
Ωe Ωe
which leads to
1
∆t(p − pr ) = Lu, (3.53)
˜ ˜ p Ae ¯ ˜
1
M u̇ + C u = Ru (u, b) + LT pr , (3.54a)
¯ ˜ ∆t ¯ ˜ ˜ ˜ ˜ ¯ ˜
ḃ = Rb (u, b), (3.54b)
˜ ˜ ˜ ˜
where C is defined by
¯
1
C= LT L. (3.55)
¯ p Ae ¯ ¯
19
3.4 A mixed method for viscous terms
The viscous term −νs ∆u in (2.12a) cannot be discretised with RTk or RT[k] el-
ements in the usual way. The reason is the discontinuity of tangential velocities
across element boundaries: div uh is integrable but grad uh is not. Therefore we
have to use a mixed method as described by Raviart (1981).
The mixed method will be illustrated for the Stokes problem:
grad p − ν∆u = f in Ω, (3.56a)
div u = 0 in Ω, (3.56b)
with boundary conditions
u = ū on ∂Ω. (3.57)
Using
∆u = grad(div u) − curl curl u, (3.58)
and introducing
ω = ν curl u, (3.59)
we can write (3.56) as follows:
grad p+ curl ω = f , (3.60a)
div u = 0, (3.60b)
curl u −ν −1 ω = 0. (3.60c)
The weak form of (3.60) is: find (u, p, ω) ∈ U × Q × Θ such that
Z Z Z Z
− p div v dx+ v · curl ω dx = − pv · n dγ + f · v dx, (3.61a)
Ω Ω Ω
Z ∂Ω
− q div u dx = 0, (3.61b)
Z Ω Z Z
u · curl θ dx − ν −1 ω · θ dx = − θ · (n × u) dγ. (3.61c)
Ω Ω ∂Ω
20
The structure of the discretised systems becomes
0 LT B u Ru
L ¯ ¯0 ¯0 p˜ = ˜0 (3.62)
B¯T ¯0 −A ¯ ω̃ R˜ω
¯ ¯ ¯ ˜ ˜
If we include viscous terms in the viscoelastic system (3.50) as described above for
the Stokes system, we get system of the following form
21
Table 3.1: Number of integration point ng for the volume integrals of triangular
elements. The co-ordinates and weights can be found elsewhere (Dunavant 1985).
k ng ng for A
¯
0 1 3
1 3 6
2 6 12
3 12 16
k1 = ∆tG(U n , tn ),
˜ ˜ ˜
k2 = ∆tG(U n + 12 k 1 , tn + 12 ∆t),
˜ ˜ ˜ ˜
k3 = ∆tG(U n + 12 k 2 , tn + 12 ∆t),
˜ ˜ ˜ ˜
k4 = ∆tG(U n + k 3 , tn + ∆t),
˜ ˜ ˜ ˜
U n+1 = U n + 16 (k 1 + 2k 2 + 2k 3 + k 4 ), (3.68)
˜ ˜ ˜ ˜ ˜ ˜
where ∆t is the time-step. These are not the only RK methods, many others are
possible, but if the system is linear all RK methods of the same order are identical.
We will use an RK-method of the same order in time as the order of the space
discretisation (= polynomial order +1), i.e. RK-order = k + 1.
Since we use the penalty method to eliminate the pressure vector, we do not
solve for u̇, but solve the vector d = un + ∆tu̇ and write (3.50) in the form
22
Table 3.2: Factors for the critical eigenvalue.
k k k
fwave frelax fvisc
k=0 2 1 12
k=1 6 1 60
k = 2 12.56 1 170.1
k = 3 19.50 1 380.2
with
k |u| + cx |v| + cy k 1 k 1 1
µ = fwave ( + ) + frelax + fvisc νs ( 2 + 2 ). (3.75)
hx hy λmin hx hy
where cx and cy are the (local) shear wave speeds in x and y direction respectively,
hx and hy are typical mesh sizes in x and y direction and λmin is the minimum
effective relaxation time of all modes:
λmin = min λeff
m, (3.76)
m
where λeff
m is given by
λm
λeff
m = . (3.77)
maximum eigenvalue of ∂g m (bm )/∂bm
˜ ˜ ˜
and g m is the ‘vector of components’ of the corresponding tensor g m in (2.10). The
˜ fk , fk
factors k
wave relax and fvisc depend on the polynomial order k and are given in the
table 3.2. The CFL condition is now given by
zreal
∆t ≤ , (3.78)
µcrit
where −zreal is the intersection point of the stability diagram of the RK method
with the negative real axis. The values of zreal for the four RK-methods are given
in table 3.3. We usually multiply the upper bound in (3.78) by a factor of 0.9 to be
on the save side. It is possible to adapt ∆t at the start of each time step.
23
Table 3.3: Intersection points of the stability diagram with the negative real axis
for RK-methods.
zreal
RK1 2
RK2 2
RK3 2.5127
RK4 2.7852
U̇ = Ge (U , t) + Gi (U , t), (3.79)
˜ ˜ ˜ ˜ ˜
the Ge term can be treated explicitly and the Gi term implicitly. The multi-step
˜ is given by
method ˜
PJ−1 X
J−1
γ0 U n+1 − k=0 αk U n−k
˜ ˜ = βk Ge (Un−k ) + Gi (U n+1 ). (3.80)
∆t ˜ ˜ ˜
k=0
where J is the order of the method. The coefficients can be found elsewhere
(Karniadakis et al. 1991; Hulsen 1996).
The method can be directly applied to the system (3.64), while treating the
penalty term and the viscous term implicitly. The left-hand side of the system
becomes
1
∆t (γ0 M + C ) B un+1
= ··· , (3.81)
¯ ¯ ¯ ˜
BT −A ω n+1
¯ ¯ ˜
γ0 bn+1 = · · · , (3.82)
˜
where the right-hand side depends on the solutions at former time steps. The matrix
system is solved by direct matrix methods.
The stability of the method has been analysed for a one-dimensional linear
equation by Hulsen (1996), because the analysis by Karniadakis et al. (1991) is
incorrect. By using (3.74) we can extend the stability analysis to our system. It is
however not possible to have an adaptive time step strategy for multi-step methods.
b. Creeping flow: ρ → 0.
24
c. Small relaxation times: λm → 0.
d. Large relaxation times: λm → ∞. This leads to an elastic neo-Hookian solid.
In order to study these limits we write out the following
lT [u] 1 [v] [τ ]
c3 ˜3T ˜ = (u + cs )
xy
−
l 3 r3 2 cs ρc2s
1 [v] [τxy ]
˜ ˜
[τxy ]
= u − u 2 + [v] − , (3.83)
2 cs ρcs ρcs
lT [u] 1 [v] [τ ]
c4 ˜4T ˜ = (u − cs ) −
xy
−
l 4 r4 2 cs ρc2s
1 [v] [τxy ]
˜ ˜
[τxy ]
= −u − u 2 + [v] + , (3.84)
2 cs ρcs ρcs
lT [u] 1 [w] [τ ]
c5 ˜5T ˜ = (u + cs )
xz
−
l 5 r5 2 cs ρc2s
1 [w] [τxz ]
˜ ˜
[τxz ]
= u − u 2 + [w] − , (3.85)
2 cs ρcs ρcs
lT [u] 1 [w] [τ ]
c6 ˜6T ˜ = (u − cs ) −
xz
−
l 6 r6 2 cs ρc2s
1 [w] [τxz ]
˜ ˜
[τxz ]
= −u − u 2 + [w] + , (3.86)
2 cs ρcs ρcs
and the split vectors
0
1 [τxy ]
2 ρu[v] − u + ρcs [v] − [τxy ]
lT [u] cs
c3 ˜3T ˜ r3 = 0 , (3.87)
l 3 r3 ˜
˜ ˜ 0
[v] [τxy ] [τxy ]
1
2 u c − u ρc2 + [v] − ρc
xy
H
s s s ˜
0
1 [τxy ]
2 ρu[v] + u − ρcs [v] − [τxy ]
lT [u] cs
c4 ˜4T ˜ r4 = 0 , (3.88)
l 4 r4 ˜
˜ ˜ 0
[v] [τxy ] [τxy ]
1
2 −u − u + [v] + H xy
cs ρc2s ρcs ˜
0
0
lT [u] 1 [τxz ]
ρu[w] − u + ρcs [w] − [τxz ]
c5 ˜5T ˜ r5 =2 c , (3.89)
l 5 r5 ˜ s
˜ ˜ 0
[w] [τxz ] [τxz ]
1
2 u c − u ρc2 + [w] − ρc
xz
H
s s s ˜
0
0
lT [u] 1 [τxz ]
ρu[w] + u − ρcs [w] − [τxz ]
c6 ˜6T ˜ r6 = 2 cs , (3.90)
l 6 r6 ˜
˜ ˜ 0
[w] [τxz ] [τxz ]
1
2 −u − u + [w] + H xz
cs ρc2s ρcs ˜
25
where
H xy
1 H xz
1
H 2
˜ xy H 2
˜ xz
H xy = ˜. , H xz = ˜. . (3.91)
˜ .. ˜ ..
H xy
M H xz
M
˜ ˜
The remaining split vector, corresponding to the convection, is u[u]conv as given by
(3.41) and (3.42). ˜
In supercritical flows, i.e. |u| > cs , all split vectors are ‘one-sided’ in the upstream
direction. This means that all vectors in (3.33) contribute to the same side. Then,
we find that
0
6 ρu[v] − [τxy ]
X lTi [u]
ci ˜ T ˜ ri + cconv [u]conv = ρu[w] − [τxz ] . (3.92)
l r
i=3 i i ˜ ˜ 0
˜ ˜
u[b] + [v]H xy + [w]H xz
˜ ˜ ˜
3.7.1 No viscoelasticity
In this limit we have Gm → 0. From (2.24) we obtain
P 1
cs ∼ ( m Gm ) 2 → 0. (3.93)
Furthermore, since [τxy ] and [τxz ] scale like Gm , we get
[τxy ] [τxz ]
[τxy ] → 0, [τxz ] → 0, → 0, → 0. (3.94)
ρcs ρcs
We distinguish two different cases:
• Zero normal velocity: u = 0, exactly, for example on a fixed wall or a plane
of symmetry. We always have subcritical flow: cs > |u| = 0. On an internal
boundary with u = 0 we find that
0
0
lT3 [u] lT5 [u] lT4 [u] lT6 [u]
c 3 ˜ T ˜ r 3 + c5 ˜ T ˜ r 5 = c4 ˜ T ˜ r 4 + c6 ˜ T ˜ r 6 → 0
l 3 r3 ˜ l 5 r5 ˜ l 4 r4 ˜ l 6 r6 ˜
0
˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜ 1 1
xy xz
2 [v]H + 2 [w]H
˜ ˜
(3.95)
On an external boundary with u = 0 we have only one incoming wave:
0
0
lT [u] lT [u]
c4 ˜4T ˜ r4 + c6 ˜6T ˜ r6 →
0
(3.96)
l 4 r4 ˜ l 6 r6 ˜ 0
˜ ˜ ˜ ˜
[v]H xy + [w]H xz
˜ ˜
where we have used ρcs [v] = [τxy ] and ρcs [w] = [τxz ] according the outgoing
boundary conditions as discussed in section 3.2.3.
• Non-zero normal velocity: u 6= 0. In this case we have supercritical flow in
the limit cs → 0. Hence, in this limit we can use equation (3.92) but now
with the results of (3.94):
0
X6 ρu[v]
lT [u]
ci ˜iT ˜ ri + cconv [u]conv →
ρu[w] .
(3.97)
i=3
l i ri ˜ ˜ 0
˜ ˜
u[b] + [v]H xy + [w]H xz
˜ ˜ ˜
26
We see that the limit Gm → 0 is regular and that the equations for bm are decoupled
from the (u, p) equations. The discretisation of the (u, p) system corresponds to the
incompressible Euler equation or to the Navier-Stokes equation if viscous terms are
taken into account (see section 3.4). The method for the Navier-Stokes equation is
identical to the method discussed by Raviart (1981, pages 108–109) and Thomasset
(1981, pages 104–105).
lT [u] lT [u]
c4 ˜4T ˜ r4 + c6 ˜6T ˜ r6 →
l 4 r4 ˜ l 6 r6 ˜
˜ ˜ ˜ ˜
0
− 12 [τxy ]
1
− 2 [τxz ]
. (3.99)
0
1 [τxy ] [τxy ] 1 [τxz ] [τxz ]
2 −u 2
+ [v] + H xy
+ 2 −u 2
+ [w] + H xz
ρcs ρcs ˜ ρcs ρcs ˜
Due to this singular behaviour it not possible with the split DG-method as used
here to describe the limit of creeping flows. A possibility to resolve the singular
behaviour is using mixed methods with [τxy ] = [τxz ] = 0, such as described by
Raviart (1981, pages 109–111).
27
which means that
b1 = b2 = · · · = bM = b = F · F T , (3.103)
X
M
τ = Gm (bm − 1) = G(b − 1), (3.104)
m=1
P
with G = m Gm .
Within the split DG-method it is not possible to continue, since we cannot compute
cs . If the numerical errors remain very local in space or time, for example a single
integration point or time step, we may possibly continue with our computations by
taking cs → 0. Note that in this limit we do not have Gm → 0. We distinguish two
different cases
28
The terms like [τxy ]/ρcs are singular. Therefore, we do not split these terms
and sum the split vectors for these terms so that they vanish. So we end up
with
0
− 12 [τxy ]
lT4 [u] lT6 [u]
c 4 ˜ T ˜ r 4 + c6 ˜ T ˜ r 6 → 1
− 2 [τxz ] . (3.110)
l 4 r4 ˜ l 6 r6 ˜
0
˜ ˜ ˜ ˜ 1
2 [v]H
xy
+ 12 [w]H xz
˜ ˜
On an external boundary with u = 0 the limit is regular because of the way
the boundary conditions are handled
0
0
lT [u] lT [u]
c4 ˜4T ˜ r4 + c6 ˜6T ˜ r6 → 0 ,
(3.111)
l 4 r4 ˜ l 6 r6 ˜ 0
˜ ˜ ˜ ˜
[v]H xy + [w]H xz
˜ ˜
where we have used ρcs [v] = [τxy ] and ρcs [w] = [τxz ] according the outgoing
boundary conditions.
• Non-zero normal velocity: u 6= 0. In this case we have supercritical flow in
the limit cs → 0. Hence, in this limit all split vectors are one-sided
0
6 ρu[v] − [τxy ]
X lTi [u]
ci ˜ T ˜ ri + cconv [u]conv = ρu[w] − [τxz ] . (3.112)
l r
i=3 i i ˜ ˜ 0
˜ ˜
u[b] + [v]H xy + [w]H xz
˜ ˜ ˜
There is a problem with this formulation: the splitting is discontinuous at a change
of sign for the normal velocity. This can lead to jumps in the solution. Therefore,
we split as follows
6
X lT [u]
ci ˜iT ˜ ri + cconv [u]conv =
i=3
l i ri ˜ ˜
˜ ˜
0 0 0
ρu[v] − 21 [τxy ] − 12 [τxy ]
ρu[w] + 1
− 2 [τxz ] 1
− 2 [τxz ] . (3.113)
+
0 0 0
1 xy 1 xz 1 xy 1 xz
u[b] 2 [v]H + 2 [w]H 2 [v]H + 2 [w]H
˜ ˜ ˜ ˜ ˜
The first terms is upwinded in the u-direction. The last two terms are distributed
on either side of an internal boundary. We keep (3.112) for the external boundaries.
Note that (3.110) is obtained for u = 0 on an internal boundary.
29
Chapter 4
In this chapter we will shortly discuss two flow examples: stability of planar Couette
flow and the flow around a sphere. Based on these examples we will make some
conclusions about the suitability of the DG method with splitting for computation
of viscoelastic flows.
The value of We crit can be increased by stretching the elements in the direction
of the flow (x), which has also been found by Keiller (1992) for a finite difference
method. For example using a computational area of 10H × H for a 3 × 3 mesh
increases We crit from 3.7 to 7.0. This can be used to good effect in channel type
flows as demonstrated by Draad (1996), who has used the methods developed in
this report to compute transition in two-dimensional channel flows.
1 In this chapter we denote the polynomial order by p, which is the notation used in the litera-
ture.
30
u = γ̇y
u
y H
x
Figure 4.1: Planar Couette base flow
We crit 3.5
2.5
2
2 3 4 5 6 7
N
Figure 4.2: Critical We for various resolutions N and two values of polynomial
order p. The elasticity number E = Re/We = 1.
31
R
R = 2a
a 30a
Figure 4.3: A sphere falling along the axis of a long cylindrical tube.
4.3 Conclusions
The DG method based on splitting, as developed in the previous chapter, has only an
acceptable stability for channel type flows, but there is no improvement compared to
other methods such as EVSS. For flows with complex geometries the method in its
32
current form is far inferior to existing methods and therefore not a good candidate
to compute general flows with viscoelastic fluids. For example, the standard DG
method is much better for complex flow geometries. At this moment it is unclear
why the splitted method fails for flows with complex geometries.
33
Appendix A
In this appendix we give some models that satisfy the neo-Hookean form (2.9) and
(2.10). For an overview of various models we refer to the books of Tanner (1985),
Bird et al. (1987) and Larson (1988) and the review article by Bird & Wiest (1995)
Giesekus
5
λm bm + bm − 1 + αm (bm − 1)2 = 0 (A.2)
Leonov 2D
5
1
λm bm + (b2m − 1) = 0 (A.3)
2
Leonov 3D
5
1 1
λm bm + b2m − 1 − (I1m − I2m )bm = 0 (A.4)
2 3
where
I1m = tr bm (A.5)
1 2
I2m = I1m − tr(b2m ) = tr b−1
m (A.6)
2
Phan-Thien/Tanner
5
λm bm + Y (tr bm )(bm − 1) = 0 (A.7)
where
(
1 + m (tr bm − 3) linear form
Y (tr bm ) = (A.8)
em (tr bm −3) exponential form
Modified Leonov 2D
5
1
λm bm + φ(2I1m )(b2m − 1) = 0 (A.9)
2
where
I1m = tr bm (A.10)
−1
2αm βm
φ(x) = 1 + arctan[ (x − 6)] (A.11)
π 4
34
Modified Leonov 3D
5
1 2 1
λm bm + φ(I1m + I2m ) bm − 1 − (I1m − I2m )bm = 0 (A.12)
2 3
where
I1m = tr bm (A.13)
1 2
I2m = I1m − tr(b2m ) = tr b−1
m (A.14)
2
−1
2αm βm
φ(x) = 1 + arctan[ (x − 6)] (A.15)
π 4
35
Appendix B
In this appendix we will derive the eigenvalues and eigenvectors of the system of
equations given by (2.18).
.. ..
. 1/ρ .
.. ..
x x x
uI . 0 . D1 D2 . . . DM
¯ .. .. ¯ ¯ ¯
. 0 .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
..
κ 0 0 .
A= , (B.1)
¯ . . . . . . . . . . .
..
E1 x
.
¯
.
..
E x2 uI
¯. ¯
. ..
. .
.
..
E xM
¯
.. ..
. 0 .
. ..
. y y y
vI . 1/ρ . D1 D2 . . . DM
¯ .. .. ¯ ¯ ¯
. 0 .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
..
0 κ 0 .
B= , (B.2)
¯ . . . . . . . . . .
y ..
E1 .
¯
.
..
E y2 vI
¯. ¯
.. .
..
.
..
E yM
¯
36
.. ..
. 0 .
. .
. . z z z
wI . 0 . D1 D2 . . . DM
¯ .. . ¯ ¯ ¯
. 1/ρ ..
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
..
0 0 κ .
C = , (B.3)
¯ . . . . . . . . . . .
.
..
E z1
¯
..
E2 z
. wI
¯ ¯
.. ..
. .
z
.
.
EM .
¯
1 0 0 0 0 0
Gm
Dxm =− 0 0 0 1 0 0 , (B.4)
¯ ρ
0 0 0 0 1 0
0 0 0 1 0 0
Gm
Dym =− 0 1 0 0 0 0 , (B.5)
¯ ρ
0 0 0 0 0 1
0 0 0 0 1 0
Gm
Dzm =− 0 0 0 0 0 1 , (B.6)
¯ ρ
0 0 1 0 0 0
−bxx 0 0
byy −2bxy 0
bzz 0 −2bxz
E xm =
0
, (B.7)
¯ −b xx 0
0 0 −bxx
byz −bxz −bxy m
−2bxy bxx 0
0 −byy 0
0 bzz −2byz
E ym =
−byy
, (B.8)
¯ 0 0
−byz bxz −bxy
0 0 −byy m
−2bxz 0 bxx
0 −2b byy
yz
0 0 −bzz
E zm
= . (B.9)
¯ −byz −bxz bxy
−bzz 0 0
0 −bzz 0 m
Next we want to develop the determinant equation (2.22). For this we need the
37
matrix K n given by (2.21), which leads to the matrix equation
¯
.. .
. nx /ρ ..
(u · n − c)I .. .. D D . . . D
. n /ρ . 1 2 M
¯ y ¯ ¯ ¯
.. ..
. nz /ρ .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
K n − cI = κnx κny κnz ... , (B.10)
¯ ¯
. . . . . . . . . . . . . . . .
E1 .
..
¯
E2
..
¯. (u · n − c)I
.
.
. ¯
..
EM .
¯
where
(m)
with bnn = n · bm · n.
4. Develop the 3 × 3 determinant. This leads to
where cc is given by
1X
c2c = κ/ρ + Gm b(m)
nn . (B.16)
ρ m
u · n + cc , u · n − cc , u · n + cs , u · n − cs , u · n + cs , u · n − cs , u · n, . . . , u · n .
| {z } | {z } | {z }
compression shear convection
(B.17)
38
The characteristic decomposition of a vector δu is given by
˜
Xm T
li δu
δu = ˜ ˜ ri ,
˜ i=1 lTi ri ˜
˜ ˜
X6
lTi δu
= ˜T ˜ ri + δuconv , (B.18)
l r ˜
i=1 ˜i ˜i
˜
where
Xm T
li δu
δuconv = ˜ T ˜ ri , (B.19)
˜ l r ˜
i=7 ˜i ˜i
39
(m) (m)
−bxx −bxx 0 0 0 0
(m) (m) (m) (m)
byy byy −2bxy −2bxy 0 0
(m) (m)
bzz (m)
bzz 0 0
(m)
−2bxz −2bxz
Hm =
0
. (B.25)
0
(m) (m)
¯ 0 −bxx −bxx 0
(m) (m)
0 0 0 0 −bxx −bxx
(m) (m) (m) (m) (m) (m)
byz byz −bxz −bxz −bxy −bxy
It is easily verified that the matrix LT R is a diagonal matrix. The diagonal com-
ponents are the inproducts lTi ri , i =¯1, .¯. . , 6, which are given by
˜ ˜
(
2ρc2c , for i = 1, 2
lTi ri = (B.26)
˜ ˜ 2ρc2s , for i = 3, 4, 5, 6
X
M
δw1 = ρcc δu + δp − Gm δb(m)
xx , (B.27)
m=1
X
M
δw2 = −ρcc δu + δp − Gm δb(m)
xx , (B.28)
m=1
X
M
δw3 = ρcs δv − Gm δb(m)
xy , (B.29)
m=1
X
M
δw4 = −ρcs δv − Gm δb(m)
xy , (B.30)
m=1
X
M
δw5 = ρcs δw − Gm δb(m)
xz , (B.31)
m=1
X
M
δw6 = −ρcs δw − Gm δb(m)
xz . (B.32)
m=1
0
0
0
P
κ(δp − G δb
(m)
)
δp − m m xx
ρc2c
δuconv = , (B.33)
˜ d1
˜
d2
˜.
..
dM
˜
40
where di , i = 1, . . . , M are given by
˜
(i) P (m)
(i) bxx (δp − m Gm δbxx )
δbxx +
ρc2c
(i) P (m) (i) P (m)
(i) byy (δp − m Gm δbxx ) bxy m Gm δbxy
δbyy − −
ρc2c ρc2s
(i) P (m) (i) P (m)
b (δp − G δb ) b G δb
δb
(i)
−
zz m m xx
−
xz m m xz
zz 2 2
di =
ρcc ρcs .
(i) P (m)
˜ (m) bxx m Gm δbxy
δbxy −
ρc2s
(i) P (m)
(i) b G δb
δb −
xx m m xz
xz
ρc 2
P (i) P
s
(i) P
b
(i)
(δp − G δb
(m)
) b G δb
(m)
b G δb
(m)
(i) yz m m xx xz m m xy xy m m xz
δbyz − − −
ρc2c ρc2s ρc2s
(B.34)
With help of (2.4) and (2.9) we can rewrite the equations (B.27)–(B.34) as follows
and
0
0
0
κ(δp − δτ )
δp − xx
ρc 2
δuconv = c , (B.41)
˜ d
1
d˜
2
˜.
..
dM
˜
where di , i = 1, . . . , M are given by
˜
(i)
(i) bxx (δp − δτxx )
δbxx +
ρc2c
(i) (i)
(i) byy (δp − δτxx ) bxy δτxy
δbyy − −
ρc2c ρc2s
(i) (i)
b (δp − δτ ) b δτ
(i)
δbzz −
zz xx
−
xz xz
2 2
di = ρc c ρc s . (B.42)
(i)
˜ (i) bxx δτxy
δbxy −
ρc2s
(i)
(i) b δτ
δb −
xx xz
xz
ρcs2
(i) b(i) yz (δp − δτxx )
(i)
bxz δτxy bxy δτxz
(i)
δbyz − − −
ρc2c ρc2s ρc2s
41
Appendix C
Before discussing the boundary conditions of the viscoelastic system we will first
consider a simple example.
On an interval [a, b] we have the prototype system
∂u ∂u 1 ∂τ
+U + = 0, ρ > 0, (C.1)
∂t ∂x ρ ∂x
∂p ∂τ ∂u
+U +G = 0, G > 0, (C.2)
∂t ∂x ∂x
which may also be written as
∂u ∂u U 1/ρ u
˜ + A ˜ = 0, with A =
G U
, u= . (C.3)
∂t ¯ ∂x ˜ ¯ ˜ τ
a. U < −c: supercritical flow. Both the variables w1 and w2 are transported in
negative x-direction.
42
If we consider boundary conditions at x = b, waves travelling in positive x-direction
are outgoing and waves travelling in negative x-direction are in-going. Characteris-
tic variables that are outgoing cannot be prescribed: they are determined from the
‘inside’. Characteristic variables that are in-going have to be prescribed directly or
should be derivable from the outgoing characteristic variables. For the three flow
conditions given above we have for the b.c. at x = b:
a. U < −c; supercritical inflow: Both the variables w1 and w2 are in-going and
should be fully prescribed. From (C.6) and (C.7) we find that this is equivalent
to prescribing both u and τ .
w2 = w1 − 2ρcu, (C.10)
w2 = −w1 + 2τ, (C.11)
but others are possible. The first equation can be used when we prescribe u = ū:
w2 = w1 − 2ρcū, (C.12)
Both boundary conditions are of the ‘reflecting’ type: the outgoing wave w1
comes back via w2 . Prescribing w2 directly, independent of w1 , is a non-reflecting
boundary condition. However, this type of b.c. will be not be discussed here.
Proper initial conditions for (C.6) and (C.7) are specification of w1 and w2 at the
initial time for all x ∈ (a, b). This is equivalent with specification of both u and τ
at the initial time for all x ∈ (a, b).
The results of the linear system can be used to analyse the boundary conditions
of the nonlinear hyperbolic visco-elastic system. When we consider (2.18) on a
boundary with outward normal n, the plane waves in the direction n are important.
Writing (2.18) locally on the boundary we have for the plane waves
∂δu ∂δu
˜ + K n ˜ = 0. (C.14)
∂t ¯ ∂n ˜
After multiplying this equation with lTi we get
˜
∂δwi ∂δwi
+ ci = 0, i = 1, . . . , nv , (C.15)
∂t ∂n
with δwi = lTi δu. These equations become decoupled in the characteristic system.
˜ ˜ systems can now be used to discuss b.c. similarly to the linear
These decoupled
system. The results are for the perturbed system in δu, but we assume that they
are valid for the non-linear system in u as well. ˜
The introduction P ˜
of weak compressibility is artificial. The coefficient κ is a very
large number: κ m Gm . We are really only interested in the limit κ → ∞.
Upon taking this limit we make the following observations
43
i. The wave speed cc becomes infinite and the set of eigenvalues and vectors cor-
responding to the compression: c1 , l1 , r1 , c2 , l2 , r2 , and the characteristic
variables δw1 and δw2 degenerate, as˜expected.
˜ ˜In this
˜ limit the pressure be-
comes a Lagrangian multiplier corresponding to the constraint div u = 0. The
wave analysis for the compression is not valid anymore. For finite, but large, κ
we conclude that we have subcritical flow (|u| < cc ) on the complete boundary.
This means that the characteristic variables δw1 and δw2 , as given by (B.35)
and (B.36), are going out and going in, respectively. Possible b.c. are: prescrib-
ing either δu, the normal velocity, or δ(−p + τxx ) = δσxx , the normal stress.
Since these types of b.c. do not degenerate in the limit κ → ∞ we assume that
they are valid in this limit as well.
ii. The eigenvalues, eigenvectors and characteristic variables corresponding to shear
waves remain regular in the incompressible limit. In fact, they are independent
of κ. Hence, it seems legitimate to assume that the b.c. for δw3 , . . . , δw6 , as
given by (B.37)–(B.40), are still valid for κ → ∞. This leads to the following
classification
a. supercritical inflow, u < −cs : δv, δτxy , δw, δτxz must all be prescribed.
b. sub-critical flow, −cs ≤ u < cs : prescribing either δv or δτxy and either δw
or δτxz are possible boundary conditions.
c. supercritical outflow, u ≥ cs : δv, δτxy , δw, δτxz cannot be prescribed.
iii. The eigenvalues, eigenvectors and characteristic variables corresponding to con-
vection remain regular in the incompressible limit. Using ρc2c ∼ κ for κ → ∞
we find from the equations (B.41) and (B.42) in the limit
0
0
0
δτxx
δuconv = d1 , (C.16)
˜ ˜
d
˜2
..
.
dM
˜
where di , i = 1, . . . , M are given by
˜ (m)
δbxx
(m)
(m) bxy δτxy
δbyy −
ρc2s
(m)
b δτ
(m)
δbzz −
xz xz
ρc 2
s
dm =
(m)
(m)
bxx δτxy .
(C.17)
˜ δbxy − 2
ρcs
(m)
(m) b δτ
δb −
xx xz
xz
ρcs2
(m) b(m) xz δτxy bxy δτxz
(m)
δbyz − −
ρc2s ρc2s
It is not difficult to see that δuconv only depends on δbm , m = 1, . . . , M , and
not on δu and δp. Furthermore, ˜ it is easy to verify that
X
M XM PM (m)
(m) Gm bxx
Gm dm (4) = Gm δbxy − m=1 2 δτxy = δτxy − δτxy = 0,
m=1
˜ m=1
ρcs
(C.18)
44
and
X
M X
M PM (m)
Gm bxx
Gm dm (5) = Gm δb(m) − m=1
δτxz = δτxz − δτxz = 0,
m=1
˜ m=1
xz
ρc2s
(C.19)
PM (m)
where we have used that ρc2s = m=1 Gm bxx . We conclude from (C.18)
and (C.19) that it is not possible to derive δτxy and δτxz from δuconv . These
shear stresses are not convected with speed u. This was expected ˜because they
are involved and determined by the shear waves. All other variables of δbm ,
m = 1, . . . , M are convected with speed u and have to be prescribed at an
inflow boundary (u < 0).
The results derived above can be summarised as follows. We have assumed that
the results for δu can be extended to u. We divide the boundary into the following
sub types: ˜ ˜
On a supercritical inflow boundary all convection and shear waves are going in,
i.e. all variables must be prescribed: tangential velocities and the complete fluid
memory bm , 1, . . . , M . On an subcritical inflow boundary we lose the possibility
to prescribe either the tangential velocities or the shear stresses. Prescribing shear
stresses is much easier, because they are determined from bm . This means that the
fluid memory bm , 1, . . . , M can then be fully prescribed.
45
Appendix D
Note that the space is defined per element K. The dimension of the space (per
element) is
(
(k + 1)(k + 3) for d = 2,
dim RTk (K) = 1 (D.3)
2 (k + 1)(k + 2)(k + 4) for d = 3.
A vector field q that is an element of the RTk space has the following properties
i. div q ∈ Pk (K).
ii. the normal component q · n is a polynomial of degree k on a line (d = 2) or a
surface (d = 3) with normal vector n (x · n = constant).
iii. the normal components q · n are continuous across element boundaries.
The properties ii. and iii. means that we need to define the following number of
normal components q · n on the boundaries of an element
(
3(k + 1) for d = 2,
nboun = (D.4)
2(k + 1)(k + 2) for d = 3.
Note that a triangle has three sides and a tetrahedron has four. From (D.3) and
(D.4) we find that the number of internal degrees of freedom is
(
k(k + 1) for d = 2,
ninternal = dim RTk (K) − nboun = 1 (D.5)
2 k(k + 1)(k + 2) for d = 3.
The internal degrees of freedom that can be used are of the form
Z
pm q dx, pm ∈ Pk (K). (D.6)
K
Note that indeed ninternal independent degrees of freedom, as given by (D.5), can
be defined. An RTk triangle has been depicted in figure D.1.
The RTk space can be used to solve the second-order elliptic problem
∆u + f = 0 in Ω, (D.7)
u=0 on ∂Ω, (D.8)
46
k+
ts
1q
ne n
·n
po
co
c om
mp
o
·n
ne
R
nt
1q
s
K pm q dx
k+
k + 1 q · n components
−g + grad u = 0, (D.9)
− div g + f = 0. (D.10)
Good1 approximation spaces for Q × U are: Qh = RTk (K) and Uh = Pk (K), where
Uh is discontinuous across element boundaries. According to the resemblance with
the velocity-pressure problem in fluid flow we can use identical discretisations for
that problem: RTk for the discretised velocity uh and Pk for the discretised pressure
ph .
where uh ∈ RTk (K). From property i. of the RTk space, i.e. div uh ∈ Pk (K), and
(D.13) we get
div uh = 0, (D.14)
i.e. the divergence is zero exactly and not only in a weak form. The subspace that
fulfills (D.14) is denoted by RTk0 :
47
The RTk0 space is not a practical space to work with, because
Z Z
div q dx = q · n ds = 0, (D.16)
K ∂K
which means that there is constraint on the normal components on the element
boundary. These are, in return, coupled with the normal components of neighbour
elements because of the continuity requirement iii. However, relaxing the constraint
(D.16) leads to a space that can be used in practice:
i.e. the divergence is a constant per element. The number of variables that can be
eliminated nelim is of course equal to the reduction of div uh from Pk (K) to P0 (K):
(
1
2 (k + 1)(k + 2) − 1 for d = 2,
nelim = dim Pk (K) − 1 = 1
(D.18)
6 (k + 1)(k + 2)(k + 3) − 1 for d = 3.
By substituting nelim independent function from Pk (K)\P0 (K) for φ into (D.21) we
see that nelim variables from the internal variables given in (D.6) can be expressed
in the normal velocities on the boundary and thus can be eliminated on element
level.
The dimension of RTk1 can be calculated from (D.3) and (D.18)
where the number of boundary degrees of freedom nboun are still given by (D.4). In
figure D.2 we have depicted the triangular (reference) elements for the space RTk1
that are used in this report.
48
k=0
k=1
k=2
R
+1 Internals: (yu − xv) dΩ
k=3
R R R
Internals: (yu − xv) dΩ, xyu dΩ, xyv dΩ
+3
49
D.3 The RT[k] , RT[k]
0
and RT[k]
1
spaces
The RT[k] space can be introduced as follows (Brezzi & Fortin 1991; Roberts &
Thomas 1991):
Note that the space is defined per element K. A vector field q that is an element
of the RTk space has the following properties
i. div q ∈ Qk (K).
The properties ii. and iii. means that we need to define 4(k + 1) normal components
q · n on the boundaries of an element.
Similar to the RTk0 and RTk1 spaces we can define RT[k] 0 1
and RT[k] spaces. In
1
the practical implementation we will use the RT[k] space.
where (ξ1 , ξ2 , ξ3 ) are co-ordinates in the reference space, also denoted by x̂. The
deformation of line elements is determined by the deformation gradient tensor F (x̂):
∂xi
dx = F (x̂) · dx̂ with Fij = , (D.27)
∂ξj
where dx̂ and dx are infinitesimal line elements in the reference space and real
space respectively. The transformation of a reference triangle has been depicted in
figure D.3. The transformation of the reference rectangle is similar.
The deformation of volume and area elements is given by
dV = JdV̂ , (D.28)
−T
da = JF · dâ, (D.29)
where J = det F . The vector quantities for RT elements transform such that the
component normal to a surface is preserved by the transformation except for the
stretching of the surface:
q̂ · dâ = q · da (D.30)
50
F (x̂) q
ξ2
(0, 1)
q̂
x̂ x
ξ1
(0, 0) (1, 0)
K̂ K
q = J −1 F · q̂. (D.31)
This is called the Piola transformation. By defining the usual operators in both the
reference and the real space, we find that
Z Z
φq · da = φ̂q̂ · dâ, (D.32)
Z ∂Ω Z∂ Ω̂
φ div q dx = ˆ dx̂,
φ̂divq̂ (D.33)
Ω Ω̂
Z Z
φq · grad() dx = ˆ
φ̂q̂ · grad() dx̂, (D.34)
Ω Ω̂
51
Appendix E
We consider a rigid body submerged in a fluid as depicted in figure E.1. The force
n
t2
t1
B
∂ B
Figure E.1: Rigid body B B
in a fluid. The boundary is denoted by ∂ . The vectors
n and t1 , t2 are the normal and tangential vectors respectively.
F and torque M P with respect to a point xP acting on the body B due to the
flow of the fluid is given by
Z
F = (−pn + 2ηs n · d + n · τ ) dS, (E.1a)
Z∂ B
MP = (x − xP ) × (−pn + 2ηs n · d + n · τ ) dS. (E.1b)
∂ B
The contribution of the viscous terms can be simplified as follows. Assume that the
body is rotating with a rate w0 . Consider a frame that is fixed to the body and
thus has the same rotation rate of w0 . In this frame we have on the wall
∂u0n ∂u0 ∂u01 ∂u0 ∂u02
2n · d = 2 n+( n + )t1 + ( n + )t2
∂n ∂t1 ∂n ∂t2 ∂n
∂u01 ∂u02
= t1 + t2
∂n ∂n
= w0 × n
= (w − w0 ) × n, (E.2)
where (n, t1 , t2 ) is a local Cartesian (not curved) system, w 0 is the vorticity in the
moving frame1 . We have used that u0 = 0 on the wall and div u0 = 0. The relation
1 Only quantities that are frame dependent are denoted with a prime.
52
between the vorticity in the stationary frame w and the moving frame is w0 is
w0 = w − w 0 . (E.3)
53
Bibliography
54
Fortin, A., & Fortin, M. 1990. A preconditioned generalised minimal residual al-
gorithm for the numerical solution of viscoelastic fluid flows. J. Non-Newtonian
Fluid Mech., 36, 277–288.
Fortin, A., & Zine, A. 1992. An improved GMRES method for solving viscoelas-
tic fluid flow problems. J. Non-Newtonian Fluid Mech., 42, 1–18.
Fortin, A., Zine, A., & Agassant, J.-F. 1992. Computing viscoelastic fluid
flow problems at low cost. J. Non-Newtonian Fluid Mech., 45, 209–229.
Fortin, M., & Fortin, A. 1989. A new approach for the fem simulation of
viscoelastic flows. J. Non-Newtonian Fluid Mech., 32, 295–310.
Gorodtsov, V.A., & Leonov, A.I. 1967. On a linear instability of a plane
parallel Couette flow of viscoelastic fluid. J. Appl. Math. Mech., 31, 310–319.
Hirsch, C. 1990. Numerical Computation of Internal and External Flows. Vol. 2.
New York: John Wiley.
Hulsen, M.A. 1986. Analysis of the equations for viscoelastic flow: type, boundary
conditions and discontinuities. Tech. rept. WTHD185. Fac. of Mech. Eng. &
Marine Tech., Delft University of Technology.
Hulsen, M.A. 1988a. Analysis and Numerical Simulation of the Flow of Viscoelas-
tic Fluids. Ph.D. thesis, Delft University of Technology, Delft, The Netherlands.
Hulsen, M.A. 1988b. Some properties and analytical expressions for plane flow of
Leonov and Giesekus models. J. Non-Newtonian Fluid Mech., 30, 85–92.
Hulsen, M.A. 1990. A sufficient condition for a positive definite configuration
tensor in differential models. J. Non-Newtonian Fluid Mech., 38, 93–100.
Hulsen, M.A. 1992. The Discontinuous Galerkin Method with Explicit Runge-
Kutta Time Integration For Hyperbolic and Parabolic Systems with Source
Terms. Tech. rept. MEMT19. Fac. of Mech. Eng. & Marine Tech., Delft Uni-
versity of Technology.
Hulsen, M.A. 1996. Stability of the implicit/explicit extension of the stiffly-stable
schemes of Gear. Tech. rept. MEAH138. Lab. for Aero & Hydrodynamics,
Delft University of Technology.
Hulsen, M.A., & van der Zanden, J. 1991. Numerical simulation of contraction
flows using a multi-mode Giesekus model. J. Non-Newtonian Fluid Mech., 38,
183–221.
Johnson, C. 1987. Numerical Solution of Partial Differential Equations by the
Finite Element Method. Cambridge: Cambridge University Press.
Joseph, D.D. 1990. Fluid Dynamics of Viscoelastic Liquids. New York: Springer
Verlag.
Joseph, D.D., Renardy, M., & Saut, J.C. 1985. Hyperbolicity and change of
type in the flow of viscoelastic fluids. Arch. Rat. Mech. Anal., 87, 213–251.
Karniadakis, G.E., Israeli, M., & Orszag, S.A. 1991. High-order splitting
methods for incompressible Navier-Stokes equations. J. Comput. Phys., 97,
414–443.
Keiller, R.A. 1992. Numerical instability of time-dependent flows. J. Non-
Newtonian Fluid Mech., 43, 229–246.
Larson, R.G. 1988. Constitutive Equations for Polymer Melts and Solutions.
Boston: Butterworths.
Lesaint, P., & Raviart, P.-A. 1974. On a finite element method for solving the
neutron transport equation. In: de Boor, C. (ed), Mathematical Aspects of
Finite Elements in Partial Differential Equations. New York: Academic Press.
55
Nedelec, J.C. 1980. Mixed finite elements in R3 . Numer. Math., 35, 315–341.
Phelan, F.R., Malone, M.F., & Winter, H.H. 1989. A purely hyperbolic
model for unsteady viscoelastic flow. J. Non-Newtonian Fluid Mech., 32, 197–
224.
Pironneau, O. 1989. Finite Element methods for Fluids. New York: John Wiley.
Ralston, A., & Rabinowitz, P. 1978. A First Course in Numerical Analysis.
London: McGraw-Hill.
Raviart, P.-A. 1981. Les Méthodes d’Éléments Finis en Mécanique des Fluides.
Collection de la Direction de Études et Recherches d’Électricité de France, no.
40. Paris: Editions Eyrolles.
Raviart, P.-A., & Thomas, J.-M. 1977. A mixed finite element method for 2-nd
order elliptic problems. In: Galligani, I., & Magenes, E. (eds), Mathe-
matical Aspects of the Finite Element Method. Lectures Notes in Mathematics
606. New York: Springer Verlag.
Renardy, M., & Renardy, Y. 1986. Linear stability of plane Couette flow of an
upper convected Maxwell fluid. J. Non-Newtonian Fluid Mech., 22, 23–33.
Roberts, J.E., & Thomas, J.-M. 1991. Mixed and Hybrid Methods. In: Ciar-
let, P.G., & Lions, J.L. (eds), Finite Element Methods (Part 1). Handbook
of Numerical Analysis, vol. 2. Amsterdam: North-Holland.
Szabó, B., & Babuška, I. 1991. Finite Element Analysis. New York: John Wiley.
Tanner, R.I. 1985. Engineering Rheology. Oxford: Clarendon Press.
Thomasset, F. 1981. Implementation of Finite Element Methods for Navier-Stokes
Equations. New York: Springer Verlag.
van der Zanden, J., & Hulsen, M.A. 1988. Mathematical and physical re-
quirements for successful computations with viscoelastic fluid models. J. Non-
Newtonian Fluid Mech., 29, 93–117.
van der Zanden, J., Kuiken, G.D.C., Segal, A., Lindhout, W.J., &
Hulsen, M.A. 1985. Numerical experiments and theoretical analysis of the
flow of an elastic liquid of the upper-convected Maxwell type in the presence
of geometrical discontinuities. Appl. Sci. Res., 42, 303–318.
Whitham, G.B. 1974. Linear and Nonlinear Waves. New York: John Wiley.
56