Journal of Non-Newtonian Fluid Mechanics, 32 (1989) 295-310 295

Elsevier Science Publishers B.V., Amsterdam - Printed in The Netherlands

A NEW APPROACH FOR THE FEM SIMULATION

OF VISCOELASTIC FLOWS

MICHEL FORTIN
Dkpartement de Mathtkztiques, Universitk Luval, Q&bec GIK 7P4 (Canada)

and

ANDRk FORTIN
DCpartement de Mathkmatiques Appliqukes, &oIe Polytechnique de Mont&al,
Mont&al H3C 3A7 (Canada)
(Received June 6, 1988; in revised form January 27,1989)

Summary

We present a finite element method for the simulation of viscoelastic

flow. A decoupled approach is employed for the computation of the velocity
field and of the components of the extra-stress tensor. The high Weissenberg
number problem is addressed by using together two different types of
upwinding namely the streamline upwinding technique and the
Lesaint-Raviart method. Numerical results are presented for the case of the
4 to 1 plane contraction.

1. Introduction

The numerical simulation of the flow of viscoelastic fluids is still an

outstanding challenge for researchers. In the past few years intensive re-
search has pointed out many difficulties inherent to this type of simulation.
The most important problem is certainly the loss of convergence of almost
all numerical schemes for high values of the Weissenberg (Deborah) number.
It is more or less accepted that most of these difficulties are purely
numerical (improper boundary conditions, hyperbolic nature of the equa-
tions, inability of the classical FEM technique to deal with singularities,
etc.). Another possible cause is certainly due to the rheological models that
characterize the viscoelastic fluids. Although these models can describe

0377-0257/89/$03.50 0 1989 Elsevier Science Publishers B.V.

296

adequately most of the behaviour of such fluids, they get into trouble in the
vicinity of singularities due to complex boundaries (see Keunings [l] for a
complete review of the numerical simulation of viscoelastic flow). Finally,
the formidable amount of computation involved in many such simulations is
also a major drawback since one has to obtain approximations of velocities,
pressure and stress tensor at each node of very fine meshes. This results in
the resolution of very large coupled non-linear systems of equations that
must be performed on supercomputers.
It has recently been shown by Marchal and Crochet [2] that a way to
overcome the high Weissenberg problem is to use an upwinding technique
that takes into account the hyperbolic nature of the constitutive equation.
They used the Streamline-Upwind method introduced by Hughes-Brooks
[3] and they obtained numerical solutions at very high Weissenberg num-
bers. This method appears to be equivalent to a modification of the
constitutive equation and its accuracy is still to be proven.
In this paper, we present an alternative way for solving viscoelastic fluid
flows at high Weissenberg numbers. This is done by using two upwinding
techniques. The first was introduced by Lesaint and Raviart [4] for the
neutron transport equation and allows the decoupling of the computation of
the stress tensor separately on each element. We then use a streamline-up-
winding technique at element level to control internal oscillations that may
occur since our scheme is not TVD (Total Variation Decreasing) due to the
quadratic approximation used (see [5]).
Another important feature of the proposed method is that the number of
global degrees of freedom is strongly reduced (velocities and pressure only).
The stress tensor is computed on an element by element basis which further
reduces the computational cost. The upper convected Oldroyd-B model is
considered as the constitutive equation for the purpose of comparison only.
We will not discuss the appropriateness of this model with respect to other
existing models since we believe that it contains in itself most of the
numerical difficulties encountered in this type of simulation. Numerical
results for the flow in a 4: 1 plane contraction will be presented and
compared with existing numerical results. It will be demonstrated that this
technique allows us the computation of viscoelastic fluid flows at Weissen-
berg numbers higher then the critical value obtained with classical methods.

2. Basic equations and finite element formulation

The numerical simulation of a viscoelastic fluid flow is very difficult

because of the nature of the fluid itself. Indeed, the behaviour of viscoelastic
fluid cannot be predicted by the usual Newtonian stress-rate-of-strain rela-
tion that characterizes Newtonian fluids since they possess a “memory” of
 297

past deformations which is not the case for Newtonian fluids. Consequently,
models that can take into account those memory effects must be introduced
and solved together with the equations of change, introducing non-linearities
present even for creeping flows. We now recall briefly the basic equations.
The equations of change consist of the stress equations of motion

pDu/Dt-v.a=f, (2-l)

where u = -p6 + T is the Cauchy stress tensor, u is the velocity field and p
is the pressure, to which we add the incompressibility contraint

V *u=o. (2.2)

Moreover, a constitutive equation must be introduced relating the extra-stress

tensor T to the velocity field u. As mentioned previously, we consider the
upper convective Oldroyd-B model which yields the following relations:

T=T+T,,

T+X(DT/D~- (VU-T) -(T+u)~)] =~o(vu(vu)tj, (2.3)

T, = ql(vu + (VU)') and (Vu)ij= ui,jy

where n1 is the solvent viscosity (set to 0.11 in all applications) and

q, + q1 = 1. Finally, the system (2.1)-(2.3) must be completed by a set of
initial and boundary conditions depending on the problem at hand. It is
worth mentioning that for most problems the boundary conditions are not
obvious and care must be exercised in order to insure reliability of the
numerical solution. For example many authors believe that the no-slip
condition usually applied on solid walls can be a source of numerical
difficulties but the proper way to impose slip conditions near geometric
singularities is still unknown (see [6]).
This system is of hyperbolic nature and it is well known that for such
systems, classical methods often yield numerical solutions presenting spuri-
ous oscillations. It is also well known that a cure to this phenomenon is to
introduce an upwinding technique that kills these oscillations by introducing
a certain amount of numerical diffusion. However, this approach usually
results in a loss of accuracy in the numerical solution. Great care must be
taken in choosing the upwinding technique to avoid smooth and stable but
physically unrealistic solutions. Let us now describe how the Lesaint-Raviart
technique can be applied in the context of the numerical simulation of
viscoelastic flows. It is , however, necessary to consider first some simpler
problems.
298

3. The Lesaint-Raviart upwinding technique

3.1. Discretization of a 1 -D transport equation

We present in this section the method of Lesaint and Raviart for the
discretization of 1-D transport problems. Although this is not new, the
method is non standard and is worth some developments. Let us consider
the differential equation

T’(x) =f (x, 0, O<x<l,

T(0) = a. (3-l)
In order to apply the finite element method we introduce a partition of the
interval [OJ] into elements [xi, xi+,], 1 < i < n. On each element the
function T will be approximated by a polynomial of degree k, i.e. T E Pk[xi,
x~+~]. It must be pointed out that the approximation can be discontinuous
between elements. We thus let r- and T:.’ denote respectively the left- and
right-hand side approximations of T at node xi. The variational formulation
of the differential equation is obtained by multiplying (3.1) by a piecewise
discontinuous test function u and integrating over the interval [OJ]. Since u
is discontinuous, we obtain on each element [xi, xitl],
X,+1
T’u dx = x’+‘f(x, T)u dx, UE V, (3.2)
J x, J x,

where

V= {u]uEP,[xj, xi+i], l<i<n}.

Equation (3.2) can be solved separately on each element by inverting a linear

system of order k + 1. However, as it stands (3.2) takes no account of the
initial condition and consequently, is meaningless. It is thus necessary to
modify (3.2) so as to include the effect of the initial condition in the first
element and that the information can be transmitted from one element to
the next. This is done formally by integrating by parts the left-hand side of
(3.2), in taking the convention that

T(x;) = ?‘- (T,- = a), (3.3)

We easily obtain
X,+1
T’u dx= - x’+‘Tu’ dx + q;iu(xi+i) - 7J-u(xi). (3.4)
/ x, /x,
 299

A second integration by part yields, this time by taking

T(xi) = lJ+ and T(x,+~) = q;,
X,+1
T’v dx = ++‘T’v dx + [ Zj v(xi), P-5)
/ Xl /Xl

where [q] = q.’ - T- is the jump in T across an element boundary. The

jump of the variable T is an attempt to take into account the Dirac masses
that should appear in the derivative of discontinuous functions. Replacing
(3.5) in (3.2) we must now solve

j-x”‘T’v dx + [q] v(xi) = /“‘+‘f(x, T)v dx, v E V, (3.6)

x, x,
for each i. The additional term in (3.6) with respect to (3.2) is an upwinding
term that will transmit the information from one element to the next. It can
easily be seen that (3.6) is an approximation of (3.1) in the distribution
sense. Moreover, if T:.- is known, then (3.6) can be easily solved by using a
properly chosen quadrature formula for the right-hand side. It can be shown
that (3.6) provides an order k approximation of (3.1) of implicit
Runge-Kutta type which is strongly A-stable. This property is particularly
important for the resolution of stiff differential equations. This will certainly
be a valuable asset for approximating flows near geometric singularities.

3.2. Discretization of a 2-D transport equation

The two-dimensional case is slightly more complicated. The numerical

scheme must also take into account the information coming from the
upwind direction. This is not as simple as in the one-dimensional case but
the generalization is still possible (Fig. 1). Let us consider the typical
equation (Thomasset [7])
u.vT+a,T=f ina,
P-7)
T=g on I+,

r_i.;
Fig. 1. 2-D domain and its boundary.
300

>cl ale-

U
aK-
bK+

f
Fig. 2. 2-D element K and parts of CIK.

where I? is the boundary of D and if n denotes the outward normal to I

then I-={xEIY, u.n(x)<O}.
Suppose that we have a partition of Q into triangles or quadrilaterals.
Hereagain we will approximate T by piecewise polynomials of degree k on
each element K, (P,(K)). The weak form of (3.7) is again obtained by
multiplying by a test function u and integrating over 52. This yields the
following relation

JK(
uqT)u dx+
$K
a,Tv dx =
JK
fu dx, u E V. (3.8)

In this equation I’ is the set of the functions whose restriction to each

element K belong to Pk. We denote by X- (resp. UC’) the inflow (resp.
outflow) boundary of element K, i.e. u. n < 0 (resp. u. n > 0). Moreover,
since we are using a discontinuous approximation of T, there are two
different ways to compute the restriction to X of T. If K is the current
element we denote by T’ (internal) and Te (external) those two approxima-
tions (see Fig. 2). The one-dimensional case is now generalized by supposing
that

T= T’ on aK_,
T= T’ on aK+, (3.9)
[T] = T”- T’ on llK,

and we obtain after two integrations by parts of the first term of the
left-hand side of (3.8)

ds+k,,TU dx=IKfU dx, (3.10)

/ K(
u.vT)u dx+i,_u.n[T]u

for every u E V and every K. Recall that (3.9) is used only in the first
integration by parts. If Te is supposed to be known then (3.10) can be
solved on each element. It is noteworthy that this assumption imposes a
particular ordering of the elements to insure that T” is always known when
solving (3.10) on a particular element. Fortunately, Lesaint and Raviart have
 301

shown (see [4]) that this ordering always exists for flows without recircula-
tion. For flows with recirculation, (3.10) cannot be solved on an element by
element basis. There is, however, the possibility that consists in solving
(3.10) element by element by using the best possible ordering and by
sweeping the entire domain many times until convergence. This can be seen
as a block relaxation iterative method for solving the global linear system
(3.10). In all cases it can be shown that (3.10) has a unique solution.
We close this section by introducing a generalization of (3.10) that allows
partial upwinding. This is done by replacing (3.9) by

T=aT”+(l-a)T’ on SC-,
(3.9)’
T= aT’+ (1 - a)Te on aK+,

where cr (0 < (Y< 1) is a heuristically determined upwinding parameter. The

boundary integral appearing in (3.10) should now be replaced by the
following sum:

ai,u.n[T]o ds+ (1 -a)/ wn[T]u ds. (3.11)

aK+

When (Y equals 1, we have a fully upwinded scheme. For (Y less then 1, we

have a less upwinded scheme, cx= 0.5 corresponding to a “centered” discre-
tization. It is important to note that when (Y differs from 1, a boundary
condition must be imposed on I?+ and that is impossible. This difficulty can
be easily overcome by imposing CY= 1 in elements for which CM+ n If + 4.
The parameter (Y allows for the possibility of adding a smaller amount of
numerical (artificial) diffusion but to our knowledge there is no way to
choose its value in an optimal fashion with respect to accuracy and stability.
In all simulations we have chosen (Y= 1.
This transport equation has been studied by Lesaint and Raviart [4] and
convergence has been shown to be of 0( hk) in L2-norm when polynomials
of degree k are used. This result is not optimal but sufficient for our
purpose. It must be recalled (Leveque-Goodman [5]) that for k 2 2 such a
scheme cannot be total variation decreasing (TVD) so that oscillations are
likely to occur near singularities. As we shall see in the numerical results,
this will be the case in viscoelastic fluid flow simulations.

3.3. Application to viscoelastic fluid jlows

In order to apply the method of Lesaint and Raviart to the simulation of

viscoelastic flows, it is necessary to go back to the basic equations (2.1)-(2.3)
and their variational formulation. By multiplying eqns. (2.1) (2.2) and (2.3)
302

respectively by test functions D, CJand cp, and with integrating by parts when
necessary, we obtain

(Du/Dt, U) - (p, v .I+ + (T:vu) = (f, 0) nE v, (3.12)

(q,v-u)=O qEQ, (3.13)

(7 : ‘p) + A( (Dr/Dt : cp) - ((VU’T) + (T. (vu)‘): q)) (3.14)

=nrJ((Vu+ (vu)‘): (P) CpE w,

(r&=%((Vu+(V+cp) cPEW,

where ( , ) stands for the integration over D (a and : denote the usual scalar
and tensor product). The spaces V and Q are respectively (H’(n))* and
L*(G) and W is defined by

W= (cpl(~;~EL*(fJ), l<i, jG2, (pi2=(p2i].

In eqn. (3.12), no continuity requirement is needed on T (and thus on r

and 7,) and consequently the stress tensor can be approximated by discon-
tinuous functions (each component of r and TVbelongs to L*(a)). This is
very important as it will allow us to compute T locally on each element by
using the technique of the previous section.
In the present computation we used a complete biquadratic QP (see Fig.
3) approximation of the velocities and a piecewise linear approximation
(discontinuous at element interfaces) for the pressure. If o is a velocity test
function, then u is continuous in a but vv is discontinuous at the interface
of adjacent elements. Moreover, vv is an incomplete biquadratic poly-
nomial Q2@). For consistency, we will approximate T by piecewise discon-
tinuous polynomials in Q$‘) on each element. With this choice we have
formally a 0( h*) approximation. Moreover when A = 0 we are insured of a
consistent approximation of a standard Stokes (Newtonian) problem. The
first equation, eqn. (3.14) can now be written

(7: ~),+h{(aT/h: (P)K+ (WVT: C&- ((VWT)

+(T.h’d: ~)K)=Io((VII+(VU)~):(P)~, cp~Q$'), (3.15)

umu ” P T

Fig. 3. Biquadratic approximation (u, T) and linear approximation (p).

 303

where ( , )K now stands for the integration over K. Equation (3.15) is

hyperbolic and the technique of Section 3.2 can be applied. Putting

7=~&+(1--)7~ onX_,
~=cxT~+(~-cY)? onX+, (3.16)
[ 71 = Te - 4 on aK,

and replacing the term (u - VT : CJJ)~ in (3.14) by the sum

(u - VT: c& + a/,,_~

-n [ T] : cpds + (I- dj-,+u.n [Tl : Cpds, (3.17)

we obtain

(7: (P)K+ +T/at: ‘#‘)K+ (U-VT: ‘&-- ((vU*T) + (Ta(VU)t):(P)K

+f&u.“[T] :cp ds+ (I -a)l,,+u.“[r] : cp ds

= 9,,((vu + (vu)‘) : & cpE Qf', (3.18)

If we now suppose that u and 7e are known for a given element K, then T
can be computed on this element by solving the linear system resulting from
(3.18). Recall that an appropriate ordering of the elements is necessary to
insure that 7e is known. To prevent this difficulty, we have chosen to solve
(3.18) using the block relaxation strategy described in Section 3.2.
Using T = T + T, in (3.12) and rearranging, we easily obtain

(Du/Dt, V) -(p, v .u) +vI((vu+ (vu)‘:v~)

=(f,V)-(T:VV) DE v, (3.19)
(&Vu)=0 qE Q- (3.20)
If we now suppose that T is known, (3.12) can be viewed as a standard
Stokes problem that can be solved by a Uzawa’s algorithm or by a penalty
technique (see [g]). Special care must be taken with the total derivative
Du/Dt = &~/at + (u - v)u (3.21)
appearing in (3.19). It is generally accepted that for this kind of simulation
the convective term does not play an important role and can be neglected.
This will not be the case if the viscosity coefficient is variable (for instance if
there is a dependence on the velocity gradient). Moreover, the time deriva-
tives au/at and &r/at are not essential for the simulation of stationary
flows. We have realized however that our algorithm fails to converge when
those two terms were left out. The reason for this is not quite clear but is
304

probably due to the particular algorithm (Picard type) that was used (see
algorithm 3.1 below). Consequently, a transient algorithm is used even in the
simulation of a steady flow. The two time derivatives were discretized by a
backward Euler scheme yielding

arGh(L+l) = (rG+i -%)/At

This leads to the following algorithm that has been succesfully used in our
simulations.

Algorithm 3. I

Step I: Obtain a divergence-free velocity field (solving for example a New-

tonian flow).
For each time step do:
Step 2: Compute r by solving (3.18) on each element;
Step 3: Compute a new velocity field by solving (3.19)-(3.20) using in the
right-hand side the value of r obtained in Step 2;
Step 4: Check for convergence and go back to Step 2 if necessary.

The following comments are in order to emphasize the features of our

method. First, Step 2 requires the solution of a large number of small linear
systems (24 x 24) on each element that could be easily parallelized, although
we did not do it. The Stokes problem in Step 3 involves a symmetric matrix
that is computed once and for all unless the time step is changed. In this
global system, only the velocities and pressure are to be computed what
results in an important reduction in the size of the matrix involved. More-
over, the linear part of the pressure on each element is eliminated by a
special condensation method (see [9]). We now present some numerical tests
performed with this algorithm.

4. Numerical results

When using a numerical method on a problem for which error estimates

are not available and at first sight hard to get except on overly simplified
models, numerical experiments are the only way to check the procedure. In
order to cross-check our results with those already published in the litera-
ture, in particular those of Marchal and Crochet [2], we have tested our
scheme on a classical bench-mark problem, the 4 to 1 contraction. Figure 4
presents the geometry and boundary conditions for this problem.
Note that this problem has a singular solution due to the presence of a
geometric singularity near the reentrant corner. As it has been shown in [6],
 305

. 3
(0,4) )4
(8,3).
2
2 Y

(O,O) (8,O)
I .

2 L x
Fig. 4. The 4 to 1 contraction, geometry and boundary conditions. (1) Imposed velocity and
extra-stress components, (2) No-slip walls (u = u = 0), (3) symmetry line (V = 0), and (4) fully
developed profile (V = 0).

this singularity is stronger in the viscoelastic case than in the Newtonian

case.
Our tests can be summarized in a few statements that we shall try to
support by numerical examples:
(a) The upwinding imbedded in the Lesaint-Raviart method is not sufficient
to avoid trouble with singularities at least when large degree polynomials are
used.
(b) Introducing at element level an additional upwinding cures this problem.
(c) Mesh refinement should be used with care. Transition from a fine to a
coarse grid can cause spurious oscillations.
This being said, we shall try to assess these statements by having a close look
at our results.

4.1 The 4 to 1 contraction

This problem has been extensively used as a test case for the high
Weissenberg (or Deborah) number problem. We limited ourselves to the
case of a plane contraction, for lack of time reasons to adapt our code to the
axisymmetric case. Although the results of Marchal and Crochet indicate
that no spectacular phenomena should be expected for this problem, it is
nevertheless a good test of the possibility to get results at fairly high
Deborah numbers. In the present case, we define De = Xr, as in [2], where
I, is the shearstress at the outlet. We present results up to De = 12. Since
we are using an essentialy unsteady procedure, steady solutions are obtained
as a limit in time, that takes a longer and longer simulation as relaxation
time increases. Our present computing equipment (an engineering worksta-
tion) prevents us from getting results for larger Deborah numbers.
306

2 (b)

Fig. 5. Meshes A, B and C.

We tested three different meshes (Fig. 5). Mesh A is a relatively coarse,

almost uniform grid while mesh B is a refinement (by a factor of 2) of the
former. Mesh C is non-uniform and has been refined near the singularity.
Let us first consider what happens when the Deborah number increases
on mesh C. For De = 0.8 (Fig. 6a) a section of rrr through the singularity
(on the line y = 3 of Fig. 4) shows a sharp peak, not very different, although
higher, from what is observed in the Newtonian case. For De = 1.6, the peak
has grown and small oscillations begik to appear (Fig. 6b). Finally, at
De = 2.0, the result is unacceptable (Fig. 6~). Appearance of oscillations is
linked with a spurious change of type (cf. [lo]), there is a small region near
the singularity where the determinant of (7 - AZ) changes sign. This is a
clear indication that the upwinding technique is not strong enough. This
 307

150.00 - (a)

134.M

119.00 -

103.M - 103.50

88.00 - 88.W ,
72.M - 72.50 -I

57.w - 57.0 4

41.50 - 41.50 ,
26.00 - 26.00

1o.m - L

7.00 7.50 8.00 8.50 9.00 9.50 IO.0010.50Il.00II.50L2Ck7 7.00 7.50 8.00 8.50 9.00 9.50 10.0010.50IlW 11.5012.M)
150.00
(c)

1
134.M

119.00

103.50

88.03

72.50 -

57.w -

41.50 -

26.00 , 1

7.00 7.50 8.00 8.M 9.00 9.M 10.00LO.50Il.00L1.5012.00

Fig. 6. Section of 711 through the singularity (on the line y = 3 of Fig. 4) for mesh C. (a)
De = 0.8, (b) De =1.6, and (c) De = 2.0.

should not be a surprise since a second order scheme cannot be TVD for a
transport equation of the type we have for i.
A curve to this had thus to be found and it was quite natural to look for it
in an extra upwinding term. What we finally adopted was to add to eqn.
(3.18) a streamline-upwinding term. More precisely we added in eqn. (3.18)
the term

k(u . VT, u . vcp),. (4.1)

Note that this extra term does not destroy the element-wise decoupling
properties of the scheme. This technique is certainly not optimal; for large k
it forces r to become linear in the streamline direction. This can be
considered as overkill with respect to what we need; an extra fourth-order
term would likely be more appropriate. Nevertheless, it is clear from the
308
260.M - 26o.cm
(a) (bl
23350 - 23330

207.w - 2u7.00 -

180.50 - 180.50 -

154.M 154.00 -

127.50 -

101.00 -

‘7450 -

48.00 -

21.50 -

-5.w

3.Ol 3.50 4.00 4.30 3.a, 5.50 6.a) 6.30 7.00 7.50 8.a) 3.M 3.30 4.00 .Wl S.CO 5.54 6.a) 6.50 7.00 7.30 8.M

2w~00 - (C) ,

233.50 -

207.00 -

180.50 -

154.00

Fig. 7. Section of 711 through the singularity (on the line y = 3 of Fig. 4) at De = 8.0. (a)
Mesh A, (b) mesh B, and (c) mesh C.

results that any extra dissipation, even a small one, can have a strong effect.
Taking k = 0.001 yields oscillation free stresses up to De = 8. It would also
be interesting to check a piecewise bilinear approximation for r that would
be naturally free of internal oscillations. It might well be that the apparent
loss of accuracy would be compensated for by a better stability. However,
this would demand for a large enough viscous term in the Oldroyd model.
It is then interesting to compare our results on the three meshes for
De = 8. On the regular mesh, we get smooth results. As can be expected, the
singularity is higher and sharper on the finer grid, although not as much as
on mesh C (Fig. 7). A remarkable phenomenon appears on the non-uniform
mesh at the transition zone: one observes a few oscillations that are
obviously mesh dependent for they are not visible on the other two meshes.
This may be related with the well known fact that in a wave propagation
 309

3.w 3.50 4.cY.l 4.50 ml 5.50 ml 6.50 1.M) 7.50 a.00

Fig. 8. Section of T,, through the singularity (on the line y = 3 of Fig. 4) at De =12.0 (mesh
J9.

problem, a change of mesh is equivalent to a change in the physical

properties. Moreover, elements in the transition zone have large angles
which may be a source of numerical errors. Finally, we present in Fig. 8 the
results at De = 12. The singularity is stronger but the general behaviour is
similar as for De = 8.

5. Conclusions

We have presented a decoupled approach for the numerical simulation of

viscoelastic fluid flows. The features of our method are the computation of
the extra-stress tensor on an element-by-element basis and the introduction
of two types of upwinding techniques to take into account the hyperbolic
nature of the constitutive equations. This method is similar to the method of
characteristics used in [ll-121.
Although the method seems to give satisfactory results, there is still room
for further improvement. The most important drawback of our method is
the fact that algorithm 3.1 is a transient procedure even in the case of steady
flows. This can be improved by a better iterative method (more Newton like)
than Picard’s scheme. It is, however, important to preserve the decoupling
between the computation of the velocity and the extra-stress and this is the
reason why a Newton method cannot be employed. A last point remains to
be addressed, namely the nature of the internal (streamline) upwinding
method. In our numerical tests, the value of the parameter k in (4.1) was
made large enough to suppress the oscillations. The value of k was then
decreased as long as the solutions were free of oscillations. This procedure is
310

certainly not optimal and should be improved. Finally the use of different
approximations of the extra-stress tensor (linear for example) could possibly
preclude the need of internal upwinding.

References

1 R. Keunings, Simulation of viscoelastic fluid flow, in: C.L. Tucker III (ed.), Fundamen-
tals of Computer Modeling for Polymer Processing, Carl Hanser Verlag, Munich, 1987.
2 J.M. Marchal and M.J. Crochet, A new finite element for calculating viscoelastic flow, J.
Non-Newtonian Fluid Mech., 26 (1987) 77-134.
3 T.J.R. Hughes and A. Brooks, A theoretical framework for Petrov-Galerkin methods with
discontinuous weighting functions: Application to the Streamline Upwind Procedure, in:
R.H. Gallagher, D.H. Norrie, D.H. Oden and J.T. Zienkiewicz, (Eds.), Finite Element in
Fluids, Vol. 4. Wiley, New York, NY, 1982.
4 P. Lesaint and P.A. Raviart, On a finite element method for solving the neutron transport
equations, in: C. de Boor (Ed.), Mathematical Aspects of Finite Elements in Partial
Differential Equations. Academic Press, New York, NY, 1974, pp. 89-123.
5 J.B. Goodman and R.J. Leveque, On the accuracy of a stable scheme for two dimensional
conservation laws, Sot. Ind. Appl. Math. J. Numer. Anal. 25(2) (1988) 268-284.
6 G.G. Lipscomb, R. Keunings and M.M. Denn, Implications of boundary singularities in
complex geometries, J. Non-Newtonian Fluid Mech., 24 (1987) 85-96.
7 F. Thomasset, Implementation of Finite Element Methods for Navier-Stokes Equations.
Springer-Verlag, Berlin, 1981.
8 M. Fortin and A. Fortin, A generalisation of Uzawa’s algorithm for the solution of the
Navier-Stokes equations, Commun. Appl. Numer. Meth., 1 (1985) 205-208.
9 M. Fortin and A. Fortin, Experiments with several elements for viscous incompressible
flows, Int. J. Numer. Mech. Fluid., 5 (1985) 911-928.
10 D.D. Joseph and J.C. Saut, Change of type and loss of evolution in the flow of viscoelastic
fluids. J. Non-Newtonian Fluid Mech., 20 (1986) 117-141.
11 M. Fortin and D. Esselaoui, A finite element procedure for viscoelastic flows, Int. J.
Numer. Meth. Fluid., 7 (10) (1987) 1035-1052.
12 P.A. Tanguy and M. Gadbois, submitted to Int. J. Numer. Meth. Fluid.