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.
Martien A. Hulsen
26th February 1996

MEAH-137
Copyright c 1996 Martien A. Hulsen

All rights reserved. No part of this publication may be reproduced,
stored in a retrieval system or transmitted by any means, electronic,
mechanical, photocopying, recording, or otherwise, without the prior
written permission of the author.
Laboratory for Aero and Hydrodynamics
Faculty of Mechanical Engineering and Marine Technology
Delft University of Technology
Abstract
In this report a numerical method for viscoelastic flow simulations is developed. It is
based on the discontinuous Galerkin (DG) method. In contrast to the standard DG
method, which is only applied to the convection terms, the DG method is applied to
the full hyperbolic system. In order to do so the characteristic values and directions
have to be determined locally and splitted according to their sign. The characteristic
values and directions have been determined analytically for multi-mode models with
an upper-convected derivative. The numerical method has been applied to stability
of a planar Couette flow and the flow around a sphere falling in a tube. The method
only seems to be useful in channel/pipe flows. In geometrically complex flows, such
as the flow around a sphere, the method is numerically stable only for very small
Weissenberg numbers.
Contents
1 Introduction 1
2 Governing equations and boundary conditions 3

2.1 Momentum equation and continuity equation . . . . . . . . . . . . . 3
2.2 Constitutive models . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.3 Alternative formulation of differential models . . . . . . . . . . . . . 4
2.4 Eigenvalues and eigenvectors . . . . . . . . . . . . . . . . . . . . . . 5
2.5 Boundary and initial conditions . . . . . . . . . . . . . . . . . . . . . 7
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
4 Some examples and conclusions 30

4.1 Stability of planar Couette flow . . . . . . . . . . . . . . . . . . . . . 30
4.2 Flow around a sphere falling a tube . . . . . . . . . . . . . . . . . . . 32
4.3 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
A Visco-elastic fluid models 34
B Eigenvalues and eigenvectors of the hyperbolic viscoelastic system 36
C Boundary conditions for the hyperbolic viscoelastic system 42
D Raviart-Thomas finite element spaces 46

D.1 The RTk spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
D.2 The RTk0 and RTk1 spaces . . . . . . . . . . . . . . . . . . . . . . . . 47
0 1
D.3 The RT[k] , RT[k] and RT[k] spaces . . . . . . . . . . . . . . . . . . . 50
D.4 Construction of the RTk space; curved elements . . . . . . . . . . . . 50
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 ˜ ˜ ˜
where Ω is a bounded subdomain of Rd , d = 1, 2 or 3, u = (u1 , u2 , . . . , unv )T ,

˜ A (u), i = 1, . . . , d
f (u) = (f1 , f2 , . . . , fnv )T is a source term and the matrices i
˜ ˜ Pd ¯ ˜
are such that any linear combination i=1 ξi Ai (u) has nv real eigenvalues and a
complete set of eigenvectors. A system having the ¯ latter
˜ property is called hyperbolic
(Whitham 1974). An important application of hyperbolic systems is in compressible
flow. The equations (conservation of mass, momentum and energy), are usually
written in the so-called conservation form
∂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
1. The flow is assumed to be incompressible, i.e. the compression wave speed is

assumed to be ∞. Only the introduction of a weak (artificial) compressibility
resolves this.
2. In many cases a solvent viscosity is present which introduces second order
terms (diffusion) in the momentum equation.
Both these problems need special attention and will be treated in this report.
As we will show in chapter 4 the method only seems to be useful in channel/pipe
flows. In more complex flows, such as the flow around a sphere, the method is nu-
merically stable only for very small Weissenberg numbers. Since the original DG
method (Fortin & Fortin 1989) is much more stable in these cases, further develop-
ments will be based on the original DG method. The main reason for documenting
the split DG method here that it has been extensively used for transition in channel
flows by Draad (1996).
2
Chapter 2
Governing equations and boundary conditions
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.
2.1 Momentum equation and continuity equation

The balance of linear momentum and the conservation of mass in a fixed bounded
space Ω become
∂u
ρ + ρu · grad u + grad p = ρf + div t, (2.1)
∂t
div u = 0, (2.2)
where ρ is the density, u the velocity vector, f a body force vector per unit mass,
p the pressure and t is the extra-stress tensor. The tensor t is symmetrical and
vanishes in equilibrium.
The extra-stress tensor t is determined by the deformation history of a material
particle and has to be specified by the constitutive equation (model) of the particular
fluid. Many models have been proposed in the literature. 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)
2.2 Constitutive models

All the models we will consider in the following have a viscous term and are of the
form
t = 2ηs d + τ , (2.3)
1 T T
where d = 2 (L + L ), with L = grad u, is Euler’s rate-of-deformation tensor and
ηs is the extra-viscosity, which may be zero. The stress tensor τ is either given
by differential equations (differential model) or by an integral equation (integral
model). We will not consider integral models in this report.
The differential models are given by superposition of sub-stresses τ m
X
M
τ = τ m, (2.4)
m=1
where M is the number of modes. The sub-stresses τ m , m = 1, . . . , M satisfy the

following differential equations
5
λm τ m + f m (τ m ) = 2ηm d, (2.5)
where λm and ηm are the relaxation time and viscosity parameter of mode m re-
5
spectively and (a) denotes the upper-convective derivative of a tensor a
5 ∂a
a= + u · grad a − L · a − a · LT . (2.6)
∂t
3
The function f m is an isotropic tensor function
f m (a) = f1m (I1 , I2 , I3 )1 + f2m (I1 , I2 , I3 )a + f3m (I1 , I2 , I3 )a2 , (2.7)
where fim , i = 1, 2, 3 are functions of the invariants I1 , I2 , I3 of a, given by

1 2
I1 = tr a, I2 = [I − tr(a2 )], I3 = det a. (2.8)
2 1
Examples of differential models are1 (Larson 1988)
Upper convected Maxwell: ηs = 0, f1m = f3m = 0, f2m = 1,
Oldroyd-B: ηs 6= 0, f1m = f3m = 0, f2m = 1,
Giesekus: f1m = 0, f2m = 1, f3m = constant,
Phan-Thien/Tanner: f1m = f3m = 0, f2m = Y (I1 ).
In Appendix A we have described these and other models in more detail.
The equations (2.1)–(2.5) form a system of partial differential equations in u, p
and τ m , m = 1, 2, . . . , M , which has to be supplemented by proper initial and
boundary conditions. Discussions on the type of the equations and whether the
system is a properly posed initial value problem can be found elsewhere (van der
Zanden et al. 1985; Joseph et al. 1985; Hulsen 1986; van der Zanden & Hulsen
1988; Hulsen 1988a; Joseph 1990). Hulsen (1988b, 1990) has derived some results
on constitutive models that are also relevant to this subject. Boundary and initial
conditions will be discussed further in section 2.5.
2.3 Alternative formulation of differential models

As described by Hulsen (1990) it is possible to identify configuration tensors bm in
the models given by (2.5) that satisfy the conditions
1. bm is positive definite,
2. bm = 1 in equilibrium.
The relation between bm and τ m is
ηm
τ m = Gm (bm − 1), Gm = . (2.9)
λm
The differential models given by (2.5) now become
5
λm bm + g m (bm ) = 0, (2.10)
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)
λ6M−2 (λ2 − c2s )2 (λ2 − c2c ) = 0, (2.23)
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
c = u · n, with multiplicity 6M − 2, (2.26)

c = u · n ± cs , with multiplicity 2, (2.27)
c = u · n ± cc , with multiplicity 1. (2.28)
(m)
Because bm is positive definite (see section 2.3), we have bnn > 0 and all the
solutions for the wave speed c are real for all n. This also means that system (2.18)
is hyperbolic. The eigenvalues given above correspond to convection, shear waves
and compression waves, respectively. In two dimensions the 3M + 3 solution vector
is given by u = (u, v, p, bT1 , . . . , bTm ) with bTm = (bxx , byy , bxy )m ; the expressions for
˜ are identical
the eigenvalues ˜ ˜
except ˜
the multiplicity of the convection, shear wave
and compression wave eigenvalues are 3M −1, 1, and 1, respectively. In the following
we denote the dimension of u by nv , which is 6M + 4 in 3D and 3M + 3 in 2D.
For each of the eigenvalues ˜ c = 1, . . . , n of the matrix K we define the left l
i v n i
and right eigenvectors ri as follows ¯ ˜
˜
K Tn li = ci li , (2.29)
¯ ˜ ˜
K n r i = ci r i . (2.30)
¯ ˜ ˜
The eigenvectors li and ri can be chosen such that (Hirsch 1990)
˜ ˜
lTi rj = 0 for i 6= j. (2.31)
˜ ˜
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).
2.5 Boundary and initial conditions

For the hyperbolic system (νs = 0) the characteristic components given in (2.32) can
be used to say something about the boundary conditions (b.c.) that are required.
When we consider (2.18) on a boundary with outward normal n, the incoming 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. (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:
a. u < −cs : supercritical inflow,
b. −cs ≤ u ≤ cs , u < 0: subcritical inflow,
c. −cs ≤ u ≤ cs , u ≥ 0: subcritical outflow,
d. u > cs , supercritical outflow.
Correct boundary conditions are as follows
• On the complete boundary, either u or σxx = −p + τxx can be prescribed.
• For the different boundary types we have to prescribe
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).
3.2 A DG-method for hyperbolic systems

Before applying the DG method to viscoelastic flow we briefly recall the method
introduced by Hulsen (1992) for non-conservative systems.
3.2.1 Discretisation of the weak form

The weak form of the system (1.1) is obtained by multiplication with a test function
v (x), and integration over the domain Ω
˜
Z
∂u X
d
∂u
vT ˜ + Ai (u) ˜ + f (u) dΩ = 0 for all v ∈ V, (3.1)
Ω˜ ∂t i=1
¯ ˜ ∂xi ˜ ˜ ˜
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
¯ ˜
Note that in (3.3) v h is still undefined on γh .

˜ )/∂x (i.e. the system is derivable from a conservative system)
If Ai (u) = ∂F i (u i
¯ ˜
we find that ˜ ˜
Xd Z n+ Xd

∆n = ni dF i (u) = ni F i (u+ ) − F i (u− ) . (3.6)
˜ i=1 n
− ˜ ˜ i=1
˜ ˜ ˜ ˜
For non-conservative systems ∆n depends on the path5 of u : u− → u+ . However,

˜ solution u is smooth, and thus
since we have assumed that the ˜ ˜ [u] →˜0 for h → 0,
˜
we can always approximate the integral by ˜ ˜
∆n = H n (u− , u+ )[u] + O([u]2 ), (3.7)

˜ ¯ ˜ ˜ ˜ ˜
where H n (a, a) = K n (a) and [u] = u+ − u− . An obvious choice is
¯ ˜ ˜ ¯ ˜ ˜ ˜ ˜
H n (a, b) = K n 12 (a + b) , (3.8)
¯ ˜ ˜ ¯ ˜ ˜
1V(e) (e) (e)
h is not a real approximating space of V since Vh is not a subspace of V . The space Vh
is an approximating space of (L2 (Ω))nv .
2 The tangential derivatives of u are finite, although they may jump across the element boundary.
˜
3 If an element boundary is internal to the domain Ω, i.e. it is connected to an element on both
sides, it is considered only once in the integral over γh .
4 The vector n is not unique. It may point to either side of γ .
h
5 We mean the path in u-space and not Rd .
˜
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 ¯
U̇ = G(U , t). (3.16)

˜ ˜ ˜
The structure of the system differs from standard finite elements because [u] on
the element boundary contains information from neighbouring elements. According˜
to Cockburn et al. (1990), the volume integrals and boundary integrals have to be
evaluated by numerical integration rules that integrate polynomials up to the order
2` and 2` + 1 exactly, respectively, where ` is the order of the polynomial used for
u.
˜
3.2.3 Initial and boundary conditions
System (1.1) should be supplemented with proper initial and boundary conditions.
The initial condition is given by
u(x, 0) = u0 (x), x ∈ Ω, (3.17)

˜ ˜
where u0 (x) is a function that makes sense with respect to the physical system
˜
investigated. For example if a component of u is the density, the corresponding
component of u(x) should be positive. For other ˜ systems the conditions may not
be so obvious. ˜
For t ∈ (0, T ) boundary conditions have to be given on ∂Ω. Consider the matrix
K n (u) on ∂Ω as given by (3.5) with n the outward normal. We define again ci , ri
¯ ˜l , i = 1, . . . , n as the eigenvalues, right and left eigenvectors of K (u). If at
and ˜
i v n
˜
point xp on the boundary ¯ ˜
c1 ≤ · · · ≤ cq < 0 < cq+1 ≤ · · · ≤ cnv , (3.18)
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)
where B is a q × (m − q) matrix, g a vector of length q and the characteristic

¯ w are given by
variables ˜
i
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
wi+ = lTi u+ , wi− = lTi u− , (3.21)

˜ ˜ ˜ ˜
6 In (3.9) n is inward to one element and outward to the other element sharing the same side.
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.
3.2.4 The hyperbolic visco-elastic system

With the results of Appendix B we can directly apply the method given by (3.13)
to the hyperbolic visco-elastic system (2.13). From (B.18) we find that the charac-
teristic decomposition of [u] is given by
˜
X6
lTi [u]
[u] = ˜ T ˜ ri + [u]conv , (3.24)
˜ l r ˜
i=1 ˜i ˜i
˜
where expressions for li , ri can be found in Appendix B. An expression for [u]conv

˜ by replacing δu by [u]. The eigenvalues are
can be obtained from ˜(B.33) ˜
˜ ˜
c1 = u · n + c c
c2 = u · n − cc
c3 = c5 = u · n + cs
c4 = c6 = u · n − cs
cconv = u · n
where cconv is the eigenvalue of [u]conv .
A disadvantage of the method ˜
P is the compressibility of the flow. The value of κ
needs to be very large (κ m Gm ) in order to approximate incompressible flow.
The discretised equation (3.16) is usually solved by an explicit time integration
method and a large value of κ then leads to severe limitations for the time step ∆t
due to the compression waves:
h
∆t < f (3.25)
cc
where f is a factor dependent on the order ` of the polynomial interpolation and h
is a typical size of the elements. To resolve this problem we will use special elements
that eliminate the compression waves all together. This will be the subject of the
next section.
3.3 A DG-method for an incompressible visco-elastic system

Application of the DG-method directly to the weak compressible viscoelastic system
leads to severe time step restrictions due to compression waves. Therefore we want
to develop a method for incompressible flow.
3.3.1 The incompressible limit

We want to take the limit κ → ∞ of the method (3.13) with [u] given by (3.24). To
find the irregularity in the limit we have to take the limit of ˜
lT [u]
ci ˜iT ˜ ri , for i = 1, . . . , nv , (3.26)
l i ri ˜
˜ ˜
on the element boundaries. We will discuss the three wave types separately.
13
Compression waves
In the limit κ → ∞ we have cc → ∞. This means that c1 = u + cc → ∞ and

c1 = u − cc → −∞. Hence, we always have an ‘inflow’ and and ‘outflow’ part. We
get from (B.22) and (B.23)
 
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
3.3.2 Removing the irregularities for the incompressible limit

In the previous subsection we observed two irregularity problems in the incompress-
ible limit. One of them is easily solved. In the limit the pressure is not hyperbolic
anymore, but becomes a Lagrange multiplier coupled with the div v = 0 constraint.
Therefore we divide the pressure equation in (3.13) by κ and take the limit. In fact,
we retain equation (2.12b) instead of (2.13b). This is reflected by the observation
that the fourth component in (3.29), u[τxx ], also vanishes after dividing by κ and
taking the limit.
Removing the second irregularity, corresponding to cc [u], is more difficult. When
taking the limit κ → ∞ we have cc → ∞. If we assume that the solutions remain
regular in the limit, it means that [u] = 0. If we put this constraint, i.e. the normal
velocity is continuous across the element boundaries, directly into the discretisation
(e)
space Vh , we see that the term cc [u] vanishes in (3.27) and (3.28). Discretisations
that fulfil this constraint are the finite element approximations of the space
H(div; Ω) = {q|q ∈ (L2 (Ω))d , div q ∈ L2 (Ω)}, (3.31)
where d is the dimension of the physical space, e.g. d = 2 or 3. Various type of

approximations are discussed in the book by Brezzi & Fortin (1991). In this report
we will use the discretisation spaces RTk for triangles and RT[k] for quadrilaterals,
introduced by Raviart & Thomas (1977). These spaces have been extended to
tetrahedral elements by Nedelec (1980).
(e)
The RTk and RT[k] spaces are like the Vh space defined in section 3.2.1, i.e.
vector variables q may be discontinuous across element boundaries, however with
one exception: the normal component q ·n must be continuous. This means that on
the element boundary the degrees of freedom consist of these normal components.
Depending on the polynomial order there are internal degrees of freedom as well.
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.
3.3.3 Discretisation of the weak form

The DG-method (3.13) with the restrictions to incompressible flow as described in
the previous subsections now becomes: find (uh , ph , b1 , . . . , bM ) ∈ Uh × Qh × (Bh )M
such that for all elements Ωe
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)
for all v h ∈ Uh ,, qh ∈ Qh , and βh ∈ Bh , where the spaces are given by: Uh =

RTk (K) or RT[k] , Qh = Pk (K) and Bh = (P` (K))6 . In (3.32) b.i.t. represents
boundary integral terms of the eigenvalues c ∈ {c3 , c4 , c5 , c6 , cconv }:
Z X
6
lT [u]
b.i.t. = v Th ci ˜iT ˜ ri + cconv [u]conv dγ, (3.33)
∂Ωe ˜ i=3
l i ri ˜ ˜ in
˜ ˜
where
( . . . )in = only terms with ci < 0 or cconv < 0. (3.34)
The eigenvalues are
c 3 = c5 = u · n + cs
c4 = c6 = u · n − cs
cconv = u · n
7 Other ways of dealing with the incompressibility constraint are:

• Introduction of a stream function (Raviart 1981; Thomasset 1981). This eliminates all
pressures.
• Introduction of Lagrange multipliers to impose the continuity of the normal velocities
(Brezzi & Fortin 1991; Roberts & Thomas 1991). With these so-called hybrid methods,
the velocities and pressures can be eliminated on element level.
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 we have defined the boundary normal stress by τnnb

= 12 (τnn
+ −
+ τnn ). Contrary
to div v h , computing grad v h for curved elements is not easy (see Appendix D).
Therefore, we integrate back partially and find for the stress terms:
Z Z !
X 1
··· − v h · div τ dΩ − v h · n[τnn ] dγ
e Ωe 2 ∂Ωint
e
Z
b −
− v h · n(τnn − τnn ) dγ + · · · (3.46)
∂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)
for all v h ∈ Uh ,, qh ∈ Qh , and βh ∈ Bh , where the spaces are given by: Uh =

RTk (K) or RT[k] , Qh = Pk (K) and Bh = (P` (K))6 . The polynomial order ` of the
approximation space Bh for bm can be taken independently from m. However, we
generally use ` = k.
If we reduce the space from RTk (K) to RTk1 (K) for uh and v h we see that the
pressure term in (3.47a) becomes
Z Z
− ph div v h dΩ = − div v h ph dΩ (3.48)
Ωe Ωe
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
M u̇ + LT p = Ru (u, b), (3.50a)

¯˜ ¯ ˜ ˜ ˜ ˜
Lu = 0, (3.50b)
¯˜ ˜
ḃ = Rb (u, b), (3.50c)
˜ ˜ ˜ ˜
where u, p and b are the vectors of unknowns of uh , ph and bm , respectively. The
˜ ˜ ˜
matrices M and L are constant. The coefficient matrices for ḃ have already been
¯ ¯ level. The vector of pressures p contains the
inverted on element ˜ averaged element
˜
pressures given by (3.49). These pressures can be eliminated by the penalty method
as follows. We penalise the pressure p̄h by
∆tp (p̄h − p̄rh ) + div uh = 0, (3.51)
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 ¯ ˜
and the system now becomes
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 ¯ ¯
3.3.4 A remark on boundary conditions

The boundary conditions corresponding to the shear waves and the convection as
discussed in section 2.5 and appendix C are quite naturally implemented by the
method for the full hyperbolic system given in section 3.2.3. However, the normal
velocity/pressure part does not have characteristics anymore and can be handled
by the traditional way of treating boundary conditions in finite elements.
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
for all (v, q, θ) ∈ U × Q × Θ, where U = H(div; Ω), Q = L2 (Ω) and Θ = (H 1 (Ω))3 .

Note that
1. for the boundary conditions (3.57) the pressure boundary integral in (3.61a)
vanishes.
2. in (3.61c) the boundary integral only involves the tangential velocity compo-
nents. This means that tangential velocities are prescribed weakly, contrary
to the normal velocity.
Good mixed methods are given by the following approximation spaces (Raviart
1981)
Uh = RTk (K),
Qh = Pk , discontinuous across elements,
Θh = Pk+1 , continuous across elements,
or
Uh = RT[k] (K),
Qh = Qk , discontinuous across elements,
Θh = Qk+1 , continuous across elements.
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
M u̇ + LT p + B ω = Ru (u, b), (3.63a)

¯ ˜ ¯ ˜ ¯˜ ˜ ˜ ˜
Lu = 0, (3.63b)
¯˜ ˜
B T u − Aω = Rω , (3.63c)
¯ ˜ ¯˜
ḃ = Rb (u, b). (3.63d)
˜ ˜ ˜ ˜
If we apply the penalty method to eliminate the pressure we get
1
M u̇ + C u + B ω = Ru (u, b) + LT pr , (3.64a)
¯ ˜ ∆t ¯ ˜ ¯ ˜ ˜ ˜ ˜ ¯ ˜
B T u − Aω = Rω , (3.64b)
¯ ˜ ¯˜
ḃ = Rb (u, b). (3.64c)
˜ ˜ ˜ ˜
Although the mixed method for viscous terms works well, there are some major
disadvantages compared to standard Galerkin methods for viscous flow:
a. Using traction boundary conditions is difficult.
b. The number of unknowns is larger, especially in 3D.
c. Even if the viscous terms are treated by explicit time integration a matrix system
must be solved for obtaining ω .
˜
d. It is very difficult to treat viscosities that are not constant, e.g. generalised New-
tonian models and temperature dependence. Implementation of other viscous
models, such as anisotropic fluids, is also difficult.
Unfortunately however, it is not possible to use standard Galerkin methods with
RT elements.
For computing forces and couples on submerged bodies we can use the equations
from appendix E.
3.5 Numerical quadrature

The integrals (3.47) and (3.61) are integrated numerically by Gaussian quadrature.
According to Cockburn et al. (1990) all volume integrals in the DG-method need
quadrature that integrates polynomials up to the order 2` exactly and boundary
integrals up to 2` + 1. We apply this rule of thumb with ` = k for the RTk space
as well. Therefore we apply it to all integrals in (3.47) and (3.61), except for the
integral that leads to the matrix A in (3.62). The latter integral is integrated with a
¯
quadrature rule that integrates polynomials up to 2(k + 1) exactly. The quadrature
rules that have been used are summarised in table 3.1. For the RT[k] space we use
(k + 1) × (k + 1) and (k + 2) × (k + 2) point Gauss-Legendre integration respectively.
3.6 Time integration

To solve the systems (3.50) and (3.63) we need a discretisation in time. We use
explicit methods (Runge-Kutta) and mixed explicit/implicit methods (Karniadakis
et al. 1991).
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
3.6.1 Explicit methods

Following Cockburn & Shu (1989) we can solve our systems using explicit Runge-
Kutta (RK) methods. In order to do so the systems (3.50) and (3.63) needs to be
written in the form (3.16). This can be done for both systems, at least in principle,
by elimination of the pressure vector p and, for (3.63), the vorticity vector ω . We
˜
will use the following standard RK methods (Ralston & Rabinowitz 1978)˜ for a
system of the form (3.16).
first-order
U n+1 = U n + ∆tG(U n , tn ) (=Euler Forward), (3.65)

˜ ˜ ˜ ˜
second-order
k 1 = ∆tG(U n , tn ),
˜ ˜ ˜
k 2 = ∆tG(U n + k 1 , tn + ∆t),
˜ ˜ ˜ ˜
U n+1 = U n + 12 (k 1 + k 2 ) (=Heun’s method), (3.66)
˜ ˜ ˜ ˜
third-order
k 1 = ∆tG(U n , tn ),
˜ ˜ ˜
k 2 = ∆tG(U n + 12 k 1 , tn + 12 ∆t),
˜ ˜ ˜ ˜
k 3 = ∆tG(U n − k 1 + 2k 2 , tn + ∆t),
˜ ˜ ˜ ˜ ˜
U n+1 = U n + 16 (k 1 + 4k 2 + k 3 ), (3.67)
˜ ˜ ˜ ˜ ˜
fourth-order
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
M d + ∆tLT p = ∆tRu (u, b) + M un , (3.69a)

¯˜ ¯ ˜ ˜ ˜ ˜ ¯˜
Ld = 0, (3.69b)
¯˜ ˜
ḃ = Rb (u, b). (3.69c)
˜ ˜ ˜ ˜
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
Now penalising div dh similar to (3.51), we get

p ∆t(p̄h − p̄rh ) + div dh = 0, (3.70)
and
1
∆t(p − pr ) = Ld, (3.71)
˜ ˜ p Ae ¯ ˜
and the system now becomes
(M + C )d = ∆tRu (u, b) + M un + ∆tLT pr , (3.72a)
¯ ¯ ˜ ˜ ˜ ˜ ¯˜ ¯ ˜
ḃ = Rb (u, b), (3.72b)
˜ ˜ ˜ ˜
where C is defined by
¯
1
C= LT L. (3.73)
¯ p Ae ¯ ¯
For the reference pressure p̄r we use the last computed value of p̄. The matrix
system is solved by direct matrix methods.
Since the RK methods are explicit, the time step ∆t has to satisfy a CFL
(Courant-Friedrich-Lewy) condition for stability. Following Hulsen (1992), we find
that a satisfactory estimate of the critical eigenvalue µcrit for 2D flows is
µcrit ≈ max µ, (3.74)
Ω
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
3.6.2 Mixed explicit/implicit methods

From (3.74) we see that the stability is determined by the largest eigenvalue. Some-
times this corresponds to a very fast transient behaviour that we’re not interested
in. The equations are called stiff. In particular the viscous terms can severely limit
the time step. To avoid this it is possible to apply implicit integration of the full
system, but this increases the computation time considerably, since all non-linear
terms need an iteration process. Furthermore, implementing this is not an easy task
for the DG-method, because the structure of the Newton matrices is different from
standard finite elements.
A different approach is given by Karniadakis et al. (1991), where some terms
are treated explicitly and others, typically the linear viscous terms, implicitly. If
the system (3.16) is written as
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.
3.7 Limit behaviour

In this section we consider the following limit behaviour
a. No visco-elasticity: Gm → 0. This leads to either the incompressible Euler or

Navier-Stokes equations.
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 
˜ ˜
˜ ˜ ˜
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).
3.7.2 Creeping flow

In this limit we have: ρ → 0. Since Gm is finite we find
ρc2s is finite, ρcs → 0, cs → ∞. (3.98)
In the limit we have subcritical flow (cs > |u|), hence on every (internal or external)
boundary the shear waves split to both sides. The splitting of the vectors according
to (3.87)–(3.90) is singular due to the terms [τxy ]/ρcs and [τxz ]/ρcs :
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).
3.7.3 Small relaxation times

In this limit we have: λm → 0. We expect Newtonian behaviour and take ηm finite
in the limit and find
ηm
Gm = → ∞, ρcs → ∞, cs → ∞, ρc2s → ∞. (3.100)
λm
In the limit we have subcritical flow (cs > |u|), hence on every (internal or external)
boundary the shear waves split to both sides. The splitting of the vectors according
to (3.87)–(3.90) is singular due to the terms ρcs [v] and ρcs [w]:
 
0
 1 ρu[v] − ρcs [v] − [τxy ] 
lT4 [u] lT6 [u]  12 
c 4 ˜ T ˜ r 4 + c6 ˜ T ˜ r 6 → 
 2 ρu[w] − ρcs [w] − [τxz ] 
 (3.101)
l 4 r4 ˜ l 6 r6 ˜  0 
˜ ˜ ˜ ˜ 1
2 [v]H
xy
+ 12 [w]H xz
˜ ˜
Due to this singular behaviour it not possible with the split DG-method as used
here to describe the limit of infinitely small relaxation times. A possibility to resolve
the singular behaviour is using standard finite elements with [v] = [w] = 0.
3.7.4 Large relaxation times

For large relaxation times λm → ∞ the behaviour becomes equal to an elastic neo-
Hookean solid. From equation (2.10) we find that the evolution equation for bm
becomes
5 1
bm = − g (bm ) → 0, (3.102)
λm m
27
which means that
b1 = b2 = · · · = bM = b = F · F T , (3.103)
where b is the (macroscopic) Finger deformation tensor and F the deformation

gradient. The distinction between modes vanishes and the stress-strain relation is
given by (2.4) and (2.9):
X
M
τ = Gm (bm − 1) = G(b − 1), (3.104)
m=1
P
with G = m Gm .
3.8 Handling negative values of c2s

Due to numerical errors it is possible that
1X
c2s = Gm b(m)
nn ≤ 0. (3.105)
ρ m
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
• 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
 − 21 [τxy ] 
 
lT3 [u]  0 

c3 ˜ T ˜ r 3 →  , (3.106)
0 
l 3 r3 ˜  
˜ ˜ 1 [τxy ] 
2 [v] − ρc H xy
s ˜
 
0
 − 21 [τxy ] 
 
lT4 [u]  0 

c4 ˜ T ˜ r 4 →  , (3.107)
0 
l 4 r4 ˜  
˜ ˜ 1 [τxy ] 
2 [v] + ρc H xy
s ˜
 
0
 0 
 
lT5 [u]  − 1
[τ ] 

c5 ˜ T ˜ r 5 →  2 xz , (3.108)
0 
l 5 r5 ˜  
˜ ˜ 1 [τxz ] 
2 [w] − ρc H xz
s ˜
 
0
 0 
 
lT6 [u]  − 1
[τ ] 

c6 ˜ T ˜ r 6 →  2 xz . (3.109)
0 
l 6 r6 ˜  
˜ ˜ 1 [τxz ] 
2 [w] + ρc H xz
s ˜
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 
˜ ˜
˜ ˜ ˜
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
Some examples and conclusions
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.
4.1 Stability of planar Couette flow

This test problem has been put forward by Brown et al. (1993). The base flow
and the dimensions have been depicted in figure 4.1. The viscoelastic model is the
upper-convected Maxwell model (UCM). The dimensionless numbers governing the
flow are the Weissenberg number We = λγ̇ and the Reynolds number Re = ρU H/η,
where U = γ̇H.
The time-dependent flow calculations are started from the base flow perturbed
with small (O(10−3 )) random disturbances. It is assumed that the damping of these
disturbances is governed by the linear stability. The base flow is linearly stable for
any We if Re = 0 (Gorodtsov & Leonov 1967) and most probably also for Re 6= 0
(Renardy & Renardy 1986). Hence any instabilities observed in computing the
initial value problem are of a numerical origin.
We have computed the initial value problem on a square domain H × H and
equal number of quadrilateral elements N × N in both space dimensions on an
equidistant mesh. We use periodical boundary conditions in x-direction. The time
discretisation is an explicit Runge-Kutta method.
We find a critical Weissenberg number We crit above which the flow becomes
unstable. In figure 4.2 we have depicted the results for We crit as a function of N and
for two values of the polynomial order1 p. The elasticity number E = Re/We = 1.
Note that the number of unknowns is proportional to approximately (N p)2 .
We observe that
• We crit is similar to the EVSS method (Brown et al. 1993).
• We crit is more or less independent of mesh size (h-refinement).
• p-refinement, i.e. increasing p, decreases We crit . This is unexpected and some-

what disappointing.
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
Critical Weissenberg number DG-split

4.5
p=1
p=5
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.
4.2 Flow around a sphere falling a tube

The flow around a sphere falling along the axis of a long cylindrical tube is a well-
known benchmark (see, for example, Bodart & Crochet 1994). In figure 4.3 we have
depicted the geometry for a sphere to tube ratio of 2:1.
We consider the steady flow of an Oldroyd-B model with ηs /(ηs + η1 ) = 0.1.
The dimensionless numbers characterising the flow are the Weissenberg number
We = λU/a and the Reynolds number Re = ρU a/(ηs + η1 ). In our method it is
not possible to use Re = 0 and therefore have used Re = 1 in the computations.
In order to compute the flow we use a frame that moves with the sphere. We
have used periodical boundary conditions with a period of 30a instead of the usual
uniform inflow and outflow boundary conditions. The mesh that is used has a
typical radial resolution of 0.03a on the surface of the sphere. The steady flow is
found as a limit of a time-dependent calculation.
We find that the flow becomes numerically unstable at relatively low values of
We, i.e. We crit = 0.1 for p = 1 and We crit = 0.2 for p = 5. For these values the
elasticity has only a small effect on the flow. We were not able to obtain higher
values. With other methods stable values for We have been obtained that are an
order of magnitude higher. For example, we have been able to obtain a We crit = 1.5
for p = 5 with the standard DG method of Fortin & Fortin (1989) using the same
mesh.
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
Visco-elastic fluid models
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)
Upper-convected Maxwell (ηs = 0) and Oldroyd-B (ηs 6= 0)

5
λm bm + bm − 1 = 0 (A.1)
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
Eigenvalues and eigenvectors of the hyperbolic

viscoelastic system
In this appendix we will derive the eigenvalues and eigenvectors of the system of
equations given by (2.18).
Writing (2.13) into a Cartesian (x, y, z) co-ordinate system we arrive at the

following expressions for the coefficient matrices A, B and C
¯ ¯ ¯
 
.. ..
 . 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 .
¯
where the matrices Dim and E im are given by

¯ ¯
 
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
Dm = nx Dxm + ny Dym + nz Dzm , (B.11)

¯ ¯x ¯y ¯z
E m = nx E m + ny E m + nz E m . (B.12)
¯ ¯ ¯ ¯
The strategy to develop the determinant of K n − cI is
¯ ¯
1. Multiply the first 3 rows of (B.10) with λ = u · n − c, which gives a factor of
λ−3 in front of the determinant.
2. Use the right-lower diagonal matrix in (B.10) to eliminate the upper-right
matrices λ(nx /ρ, ny /ρ, nz /ρ)T and λDm , m = 1, . . . , M .
¯
3. Develop the determinant from the lower-right corner which leads to the fol-
lowing 3 × 3 determinant
κ
λ6M−2 det (λ2 − c2s )I − nnT , (B.13)
¯ ρ ˜˜
where n = (nx , ny , nz )T and cs is given by

˜
1X
c2s = Gm b(m)
nn , (B.14)
ρ m
(m)
with bnn = n · bm · n.
4. Develop the 3 × 3 determinant. This leads to
λ6M−2 (λ2 − c2s )2 (λ2 − c2c ) = 0, (B.15)
where cc is given by
1X
c2c = κ/ρ + Gm b(m)
nn . (B.16)
ρ m
Next we want to discuss the eigenvectors and characteristic decomposition. We

order the 6M + 4 eigenvectors ci as follows
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
is the part corresponding to the convection. From (B.18) we find that

6
X lTi δu
δuconv = δu − ˜ T ˜ ri . (B.20)
˜ ˜ i=1 li ri ˜
˜ ˜
Since we are not interested in the individual eigenvectors of the convection, we
will only compute δuconv from (B.20). Furthermore, without loss of generality we
position the positive˜ x-axis in the direction n, i.e. we take K n = A, which makes
it more easy to compute the eigenvectors. The eigenvectors ¯li , ri , ¯i = 1, . . . , 6 are
written in m × 6 matrices: ˜ ˜
L = [ l1 , . . . , l6 ] and R = [r1 , . . . , r6 ]. (B.21)

¯ ˜ ˜ ¯ ˜ ˜
Substitution of the eigenvalues c1 , . . . , c6 into the eigensystem (2.22) leads to the
following matrices L and R
¯ ¯
 
ρcc −ρcc 0 0 0 0
 0 0 ρcs −ρcs 0 0 
 
 0 0 0 0 ρcs −ρcs 
 
 1 1 0 0 0 0 
 
L= F1 , (B.22)
¯  ¯ 
 F2 
 ¯. 
 
 .. 
FM
 ¯ 
cc −cc 0 0 0 0
0 0 cs −cs 0 0 
 
0 0 0 0 cs −cs 
 
κ κ 0 0 0 0 
 
R= H1 , (B.23)
¯  ¯ 
 H 
 ¯.
2 
 
 .. 
HM
¯
where F m and H m , m = 1, . . . , M are defined by
¯ ¯
 
−Gm −Gm 0 0 0 0
 0 0 0 0 0 0 
 
 0 0 0 0 0 0 
Fm = 0
, (B.24)
¯  0 −Gm −Gm 0 0 
 0 0 0 0 −Gm −Gm 
0 0 0 0 0 0
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
The characteristic variables δwi = lTi δu, i = 1, . . . , 6 of a vector δu are

˜ ˜ ˜
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
The vector δuconv can now be calculated from (B.20)

˜
 
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
δw1 = ρcc δu + δp − δτxx , (B.35)

δw2 = −ρcc δu + δp − δτxx , (B.36)
δw3 = ρcs δv − δτxy , (B.37)
δw4 = −ρcs δv − δτxy , (B.38)
δw5 = ρcs δw − δτxz , (B.39)
δw6 = −ρcs δw − δτxz , (B.40)
and
 
0
 0 
 
 0 
 
 κ(δp − δτ ) 
δp − xx 
 ρc 2 
δuconv =  c , (B.41)
˜  d 
 1 
 d˜ 
 2 
 ˜. 
 .. 
dM
˜
˜
 (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
Boundary conditions for the hyperbolic

viscoelastic system
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
The eigenvalues ci and eigenvectors li and ri of A are

˜ ˜ ¯

ρc c
c1 = U + c, l1 = , r1 = , (C.4)
˜ 1 ˜ G

−ρc −c
c2 = U − c, l2 = , r2 = , (C.5)
˜ 1 ˜ G
p
where c = G/ρ is the wave speed. The characteristic variables are
w1 = lT1 u = ρcu + τ, (C.6)

˜ ˜
w2 = lT2 u = −ρcu + τ. (C.7)
˜ ˜
Multiplying the system (C.3) by lTi , i = 1, 2 we obtain
˜
∂w1 ∂w1
+ c1 = 0, (C.8)
∂t ∂x
∂w2 ∂w2
+ c2 = 0. (C.9)
∂t ∂x
The variable w1 is transported with speed c1 = U + c and w2 with speed c2 = U − c.
We distinguish three ‘flow’-conditions:
a. U < −c: supercritical flow. Both the variables w1 and w2 are transported in
negative x-direction.
b. −c ≤ U < c: subcritical flow. The variables w1 and w2 are transported in

positive and negative x-direction, respectively.
c. U ≥ c: supercritical flow. Both the variables w1 and w2 are transported in

positive 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 .
b. −c ≤ U < c; sub-critical flow: The variable w1 is outgoing and cannot be

prescribed. The variable w2 is in-going and has to be prescribed directly or can
possibly depend w1 . From (C.6) and (C.7) we find for example
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)
and the second can be used when we prescribe τ = τ̄ :
w2 = −w1 + 2τ̄ . (C.13)
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.
c. U ≥ c; supercritical outflow. Both the variables w1 and w2 are outgoing and

cannot be prescribed. This means that neither u nor τ can be prescribed.
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
˜
˜  (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: ˜ ˜
a. u < −cs : supercritical inflow,

b. −cs ≤ u ≤ cs , u < 0: subcritical inflow,
c. −cs ≤ u ≤ cs , u ≥ 0: subcritical outflow,
d. u > cs , supercritical outflow.
Correct boundary conditions are as follows
• On the complete boundary, either u or σxx = −p + τxx can be prescribed.
• For the different boundary types we have to prescribe
a. supercritical inflow: v, w, bm , m = 1, . . . , M
b. subcritical inflow: there are various possibilities, for example:
1. full fluid memory: bm , m = 1, . . . , M
2. v, w and a subset of bm , m = 1, . . . , M such that δuconv is fully
specified. ˜
In the latter case τxy and τxz cannot be specified.
c. subcritical outflow: v or τxy and w or τxz
d. none.
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
Raviart-Thomas finite element spaces
D.1 The RTk spaces

The RTk space can be introduced as follows (Brezzi & Fortin 1991; Roberts &
Thomas 1991):
RTk (K) = (Pk (K))d + xPk (K), (D.1)
on a triangle or tetrahedron K ≡ Ωe and Pk (K) is given by
Pk (K) : the space of polynomials of degree ≤ k. (D.2)
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
Figure D.1: Degrees of freedom of an RTk triangle.
by a mixed method for the mixed formulation
−g + grad u = 0, (D.9)
− div g + f = 0. (D.10)
The weak formulation is: find (g, u) ∈ Q × U such that

Z Z
− g · h dx − u div h dx = 0 for all h ∈ Q, (D.11)
Ω Ω
Z Z
− v div g dx + vf dx = 0 for all v ∈ U. (D.12)
Ω Ω
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 .
D.2 The RTk0 and RTk1 spaces

When we use RTk × Pk for the (u, p) approximation in incompressible flows, the
discretised formulation of div u = 0 is
Z
qh div uh dx = 0 for all q ∈ Pk (K), (D.13)
Ω
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 :
RTk0 (K) = {q|q ∈ RTk (K), div q = 0}. (D.15)

1 By ‘good’ we mean the mixed space fulfills the LBB condition (Roberts & Thomas 1991).
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:
RTk1 (K) = {q|q ∈ RTk (K), div q ∈ P0 (K)}, (D.17)
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.
Reducing RTk to RTk1 can be achieved by elimination of internal degrees of freedom.

This can be seen as follows. For uh ∈ RTk1 (K) we have for the constant divergence
R
q
∂KR h
· n ds
div q h = . (D.19)
K dx
Using this and

Z Z Z
φ div q h dx = φq h · n ds − q h · grad φ dx, (D.20)
K ∂K K
for any smooth φ, we arrive at

Z Z R
φ dx
q h · grad φ dx = (φ − RK )q h · n ds. (D.21)
K ∂K K dx
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)
dim RTk1 (K) = dim RTk (K) − nelim

(
1
(k + 2)(k + 3) for d = 2,
= 21 (D.22)
6 (k + 1)(k + 2)(2k + 9) + 1 for d = 3,
and the number of internal degrees of freedom for RTk1 becomes

(
1
2 k(k − 1) for d = 2,
n1internal = dim RTk1(K) − nboun = 1
(D.23)
6 k(k − 1)(2k + 5) for d = 3,
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
Figure D.2: Triangular (reference) elements for the space RTk1 up to k = 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):
RT[k] (K) = Pk+1 (x) × Pk (y)ex + Pk (x) × Pk+1 (y)ey (D.24)
on a rectangle K ≡ Ωe and Pk (x) is given by
Pk (x) : the space of polynomials of degree ≤ k in one space dimension. (D.25)
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).
ii. the normal component q · n is a polynomial of degree k on the boundaries of

the element.
iii. the normal components q · n are continuous across element boundaries.
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.
D.4 Construction of the RTk space; curved elements

Introduction of the space RTk by means of (D.1) has advantages for theoretical
considerations but is not practical. The space RT[k] given by (D.24) is only for
rectangles. For a practical implementation of these elements and for constructing
elements with curved boundaries we work with a transformation of a reference
element (Raviart & Thomas 1977; Raviart 1981).
The transformation of a reference element to the real element is similar to the
deformation of bodies in continuum mechanics:
(ξ1 , ξ2 , ξ3 ) → x(ξ1 , ξ2 , ξ3 ), (D.26)
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
Figure D.3: Transformation of the reference element K̂ to the real element K.
With (D.29) it follows that
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
where φ is a scalar quantity. For curved elements the transformation F is not

constant and computing differential operations other than the ones above become
rather complicated. For example computing the velocity gradient L = (grad v)T is,
although straightforward, rather cumbersome. Contrary to standard finite elements
the transformation F is not isoparametric. We use the blending function method
(see, for example, Szabó & Babuška 1991) to generate curved elements.
51
Appendix E
Forces and torques on a rigid body submerged in a

fluid
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)
Substitution of (E.2) into (E.1) and using

Z
n dS = 0 (E.4)
∂ B
we obtain
Z
F = (−pn + ηs w × n + n · τ ) dS, (E.5a)
Z
∂ B
MP = (x − xP ) × (−pn + ηs (w − w0 ) × n + n · τ ) dS. (E.5b)
∂ B
Note that w0 cannot be eliminated for M P .
53
Bibliography
Baaijens, F.P.T. 1994a. Application of low-order discontinuous Galerkin methods

to the analysis of viscoelastic flows. J. Non-Newtonian Fluid Mech., 52, 37–57.
Baaijens, F.P.T. 1994b. Numerical experiments with a discontinuous Galerkin
method including monotonicity enforcement on the stick-slip problem. J. Non-
Newtonian Fluid Mech., 51, 141–159.
Baaijens, F.P.T. 1995. Numerical analysis of the non-isothermal viscoelastic
flow of a polymer melt around a cylinder. In: Dijksman, J.F., & Kuiken,
G.D.C. (eds), Numerical simulation of Non-Isothermal Flow of Viscoelastic
Liquids. Dordrecht: Kluwer.
Basombrı́o, F.G., Buscaglia, G.C., & Dari, E.A. 1991. Simulation of highly
elastic fluid flows without additional numerical diffusivity. J. Non-Newtonian
Fluid Mech., 39, 189–206.
Bird, R.B., & Wiest, J.M. 1995. Constitutive equations for polymeric liquids.
Annu. Rev. Fluid Mech., 27, 169–193.
Bird, R.B., Armstrong, R.C., & Hassager, O. 1987. Dynamics of Polymer
Liquids. 2 edn. Vol. 1. New York: John Wiley.
Bodart, C., & Crochet, M.J. 1994. The time-dependent flow of a viscoelastic
fluids around a sphere. J. Non-Newtonian Fluid Mech., 54, 303–329.
Brezzi, F., & Fortin, M. 1991. Mixed and Hybrid Finite Element Methods. New
York: Springer Verlag.
Brown, R.A., Szady, M.J., Northey, P.J., & Armstrong, R.C. 1993. On
the numerical stability of mixed finite-element methods for viscoelastic flows
governed by differential constitutive equations. Theoret. Comput. Fluid Dy-
namics, 59, 77–106.
Chilcott, M.D., & Rallison, J.M. 1988. Creeping flow of dilute polymer solu-
tions past cylinders and spheres. J. Non-Newtonian Fluid Mech., 29, 381–432.
Cockburn, B., & Shu, C.W. 1989. TVB Runge-Kutta local projection dis-
continuous Galerkin finite element method for conservation laws II: General
framework. Math. Comp., 52, 411–435.
Cockburn, B., Lin, S.Y., & Shu, C.W. 1989. TVB Runge-Kutta local projection
discontinuous Galerkin finite element method for conservation laws III: One-
dimensional systems. J. Comput. Phys., 84, 90–113.
Cockburn, B., Hou, S., & Shu, C.W. 1990. TVB Runge-Kutta local projection
discontinuous Galerkin finite element method for conservation laws IV: The
multi-dimensional case. Math. Comp., 54, 545–581.
Crochet, M.J., & Walters, K. 1993. Computational rheology: a new science.
Endeavour, New Series, 17, 64–77.
Draad, A.A. 1996. Laminar-Turbulent transition in pipe flow for Newtonian and
non-Newtonian fluids. Ph.D. thesis, Delft University of Technology, Delft, The
Netherlands. In preparation.
Dunavant, D.A. 1985. High degree efficient symmetrical Gaussian quadrature
rules for the triangle. Internat. J. Numer. Methods Engrg., 21, 1129–1148.
Edwards, B.J., & Beris, A.N. 1990. Remarks concerning compressible viscoelas-
tic fluid models. J. Non-Newtonian Fluid Mech., 36, 411–417.
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-
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-
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

A Discontinuous Galerkin Method With Splitting Applied To Visco-Elastic Flow by Martien A. Hulsen

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A Discontinuous Galerkin Method With Splitting Applied To Visco-Elastic Flow by Martien A. Hulsen

Uploaded by

Copyright:

Available Formats

A discontinuous Galerkin method with

splitting applied to visco-elastic flow.

26th February 1996

Copyright c 1996 Martien A. Hulsen

2 Governing equations and boundary conditions 3

4 Some examples and conclusions 30

A Visco-elastic fluid models 34

B Eigenvalues and eigenvectors of the hyperbolic viscoelastic system 36

C Boundary conditions for the hyperbolic viscoelastic system 42

D Raviart-Thomas finite element spaces 46

where Ω is a bounded subdomain of Rd , d = 1, 2 or 3, u = (u1 , u2 , . . . , unv )T ,

1. The flow is assumed to be incompressible, i.e. the compression wave speed is

Governing equations and boundary conditions

2.1 Momentum equation and continuity equation

2.2 Constitutive models

where M is the number of modes. The sub-stresses τ m , m = 1, . . . , M satisfy the

f m (a) = f1m (I1 , I2 , I3 )1 + f2m (I1 , I2 , I3 )a + f3m (I1 , I2 , I3 )a2 , (2.7)

where fim , i = 1, 2, 3 are functions of the invariants I1 , I2 , I3 of a, given by

2.3 Alternative formulation of differential models

λ6M−2 (λ2 − c2s )2 (λ2 − c2c ) = 0, (2.23)

c = u · n, with multiplicity 6M − 2, (2.26)

2.5 Boundary and initial conditions

a. u < −cs : supercritical inflow,

b. −cs ≤ u ≤ cs , u < 0: subcritical inflow,

c. −cs ≤ u ≤ cs , u ≥ 0: subcritical outflow,

d. u > cs , supercritical outflow.

Correct boundary conditions are as follows

• On the complete boundary, either u or σxx = −p + τxx can be prescribed.

• For the different boundary types we have to prescribe

3.2 A DG-method for hyperbolic systems

3.2.1 Discretisation of the weak form

Note that in (3.3) v h is still undefined on γh .

For non-conservative systems ∆n depends on the path5 of u : u− → u+ . However,

∆n = H n (u− , u+ )[u] + O([u]2 ), (3.7)

U̇ = G(U , t). (3.16)

u(x, 0) = u0 (x), x ∈ Ω, (3.17)

c1 ≤ · · · ≤ cq < 0 < cq+1 ≤ · · · ≤ cnv , (3.18)

where B is a q × (m − q) matrix, g a vector of length q and the characteristic

wi+ = lTi u+ , wi− = lTi u− , (3.21)

3.2.4 The hyperbolic visco-elastic system

where expressions for li , ri can be found in Appendix B. An expression for [u]conv

3.3 A DG-method for an incompressible visco-elastic system

3.3.1 The incompressible limit

In the limit κ → ∞ we have cc → ∞. This means that c1 = u + cc → ∞ and

3.3.2 Removing the irregularities for the incompressible limit

H(div; Ω) = {q|q ∈ (L2 (Ω))d , div q ∈ L2 (Ω)}, (3.31)

where d is the dimension of the physical space, e.g. d = 2 or 3. Various type of

3.3.3 Discretisation of the weak form

for all v h ∈ Uh ,, qh ∈ Qh , and βh ∈ Bh , where the spaces are given by: Uh =

( . . . )in = only terms with ci < 0 or cconv < 0. (3.34)

The eigenvalues are

7 Other ways of dealing with the incompressibility constraint are:

where we have defined the boundary normal stress by τnnb

for all v h ∈ Uh ,, qh ∈ Qh , and βh ∈ Bh , where the spaces are given by: Uh =

M u̇ + LT p = Ru (u, b), (3.50a)

∆tp (p̄h − p̄rh ) + div uh = 0, (3.51)

and the system now becomes

3.3.4 A remark on boundary conditions

for all (v, q, θ) ∈ U × Q × Θ, where U = H(div; Ω), Q = L2 (Ω) and Θ = (H 1 (Ω))3 .

M u̇ + LT p + B ω = Ru (u, b), (3.63a)

3.5 Numerical quadrature

3.6 Time integration

3.6.1 Explicit methods

∆tp (p̄h − p̄rh ) + div uh = 0, (3.51)