Journal of Non-Newtonian Fluid Mechanics, 42 (1992) 1-18 1

Elsevier Science Publishers B.V., Amsterdam

An improved GMRES method for solving viscoelastic fluid

flow problems

Andre Fortin * and Abdelmalek Zine

De’partement de MathLmatiques Appliqukes l&ole Polytechnique de Montkal C.P. 6079,
Succursale A Montrkal H3C 3A7 (Canada)
(Received May 15, 1991; in revised form July 15, 1991)

Abstract

We present a modified generalised minimal residual (GMRES) method

for the solution of viscoelastic fluid flow problems. Belonging to a class of
so-called Krylov-Newton methods, it allows the solution of non-linear
systems without computation of a potentially large Jacobian matrix. For
viscoelastic fluids, GMRES can be used to decouple the computation of
the velocity and pressure from that of the stress tensor. This substantially
reduces the storage requirement with respect to a full Newton method.
Our method has shown to be much more efficient than the Picard algo-
rithm. Numerical results are also presented.
Keywords: GMRES; numerical simulation; viscoelastic fluid

1. Introduction

In a previous paper, Fortin and Fortin [l] proposed a preconditioned

generalised minimal residual method for solving viscoelastic fluid flow
problems. Although promising, the method did not allow the computation
of solutions at large values of the Deborah number in test problems such as
the 4 to 1 contraction or the stick-slip problem. A possible explanation is

Correspondence to: A. Fortin,Department de Mathematiques Appliquees Ecole Polytech-

nique de Montreal C.P. 6079, Succursale A Montreal H3C 3A7, Canada. (See also present
address.)
* Present address (until May 1992): Ecole des Mines de Paris, CEMEF, 06565 Valbonne
CCdex, France.

0377-0257/92/$05.00 0 1992 - Elsevier Science Publishers B.V. All rights reserved

2

that at large values of the Deborah number, the velocity components, the
pressure and the stress components can be of quite different orders of
magnitude, resulting in ill-conditioned non-linear systems. For example, in
the vicinity of singular points like reentrant corners, the stress components
can reach extremely high values while the velocities and pressure do not
vary dramatically.
In this paper, we propose several improvements to the previous method.
First, we apply GMRES to the stress components only, considering the
velocity and pressure as a function of the stress (through a Stokes problem)
as described in section 3.
The second modification is the use of the following change of the
variable [2,3]
s = 7 - 2Q(u) (1)
in the Oldroyd-B model (see section 2 for a definition of the symbols). This
allows for computation at larger values of the Deborah number. An
interesting explanation was proposed by M. Fortin [4]. In the original work
of Oldroyd [5], the tensor S was used first while the stress tensor T was
introduced through eqn. (1). However the use of T in place of S is
mathematically justified only in the absence of singularities such as those
appearing in most applications. Consequently, the correct formulation of
similar viscoelastic models should probably use the tensor S instead of T.
We also decided to use the streamline upwind (SU) method for the
treatment of the constitutive equation in order to make sure that the loss of
convergence was not due to the Lesaint-Raviart method used in Ref. 1.
Some other minor modifications are also introduced and will be outlined
in the next two sections. Finally, numerical results are presented in section
4 for the 4 to 1 contraction and the stick-slip problem. Convergence with
mesh refinement will also be addressed in both cases.

2. Description of the problem and variational formulation

2.1 Basic equations

There exists many models to describe the flow of viscoelastic fluids. We

have chosen the Oldroyd-B model mainly for the purpose of comparison
with existing data. However, the described methodology can also be ap-
plied to other possibly more relevant models.
The problem thus consists in finding a velocity field u, a pressure p and
a StreSS tensor T solution of the following system of equations, Written in its
most standard form: Conservation of momentum:
-v(2?&+(U)) +vp=v-7. (2)
 3

Conservation of mass:
v.u=o. (3)
Constitutive equation (Oldroyd-B model):
T+h[(u~V)T-vu*T-T~(VU)T] =2q”+(u), (4)
where qs and 17, are constant viscosities and A is a constant relaxation
time. + is the rate of deformation tensor defined by:
1 aui &lj
(+(">)ij=~ aX, + aX,
( 1
and
(VU),,= aUi/aXj.

In the above system, we have neglected the inertia term in the momen-
tum equation, making the usual creeping flow hypothesis. As mentioned in
the introduction, we use the change of variable (1) and rewrite the above
system as
-v* (2vj(u)) +vp =v*s, (5)
v*u=o, (6)
S+h[(u*V)S-vu+Sqvu)‘]

= -277,h[(u*V)j(u) -vu+(u) -$(u)*(VU)=], (7)

where
rl =rlv+rls.
This system of equations must be complemented by a proper set of
boundary conditions. Typically, the velocity is fixed on a part I1 of the
boundary and the stress tensor is required only on the inflow part ri, of the
boundary defined by
ri, = {XE r I u(x) ‘12 -CO},
where n is the outward unit normal to I’. This is a consequence of the
hyperbolic nature of the constitutive equation (7). These boundary condi-
tions will be described in details for each application in section 4.

2.2 Variational formulation

The finite element discretisation of the system (5-7) requires careful

attention. Among the many difficulties is the choice of the polynomial
approximations for the different variables u, p and S. Since the early days
of numerical simulations of viscoelastic fluid flow problems, it has been
clear that some choices were better than others.
r-rmI_,:i
Fig. 1. Mixed element of Marchal-Crochet.

Our choice is based on the analysis of Fortin and Pierre [6], who made a
mathematical analysis of the Stokes problem for the velocity-pressure-
stress formulation. They concluded that the polynomial approximations of
the different variables should satisfy a generalised Brezzi-Babuska condi-
tion. Their analysis was then applied to a discretisation introduced by
Marchal and Crochet [7] which is illustrated in Fig. 1. The velocity is
approximated by biquadratic polynomials while the pressure is only bilin-
ear. The stress components are then approximated by first decomposing
the element into 16 subelements on which bilinear (continuous) polynomi-
als are used. This leads to continuous Co approximations of all three
variables.
In this paper a slight modification of this element will be used in which
the pressure is piecewise linear and discontinuous at interelement bound-
ary resulting in a C- ’ approximation (see Fig. 2). This particular choice is
motivated by a better satisfaction of the incompressibility condition [8]
when using discontinuous pressure. This element was also used in Refs. 13
and 14.
The variational (Galerkin) formulation can now be written as: Find
u E (H’(Ln>)2, p E L2(fl) and S E L,(a) such that
u =ur on l?r,
S = Si, on Iin,
and

/n[2~;/(u):~(~)-pVu] dx= -p:Vv dx+j,/,,W ds VVEI/,

(8)

(qV.u dx=O &JEL~(SZ),

‘cl

j-(s+n[u3s-Vu~S-S~(Vu)T]):~ dn

= -29,~/~((u47)j(u) -Vu.+(u) -;J(u)+‘u)~):+ dx v4 E W,

(10)
~/---I-J; : : []
6 c : -

Fig. 2. Element used in the present study.

where V is the subspace of (H’(n)>* of functions vanishing on Ii, L,(R) is

the space of symmetric tensors whose components are square integrable (in
L*(n)> and W is the subspace of L,(R) of symmetric tensors vanishing on
Ii,. We also recall that H’(IR) is the space of square integrable functions
whose first order derivatives are also square integrable. In eqn. (8), the
vector T denotes a traction force usually imposed on the outflow boundary.
T is defined by:
T=(-pI+2yj(u)+S)*n,
where I is the identity matrix. In both our test problems we will impose
T=O.
It is worth noticing that the term

involves second order derivatives of the velocities. To avoid this difficulty,

an integration by parts is performed yielding

+(u):(wV)q5 dx+/(u*+@):d ds. (11)

n l-
The above boundary integral vanishes on solid boundaries (since u . IZ=
0) and also on the portion ri, of the boundary where the stresses are
imposed (since the test functions 4 vanish on Ii,>.
Our last remark of this section concerns the numerical treatment of the
constitutive equation (10). It is well known that the standard Galerkin
method applied to hyperbolic equations results in spurious oscillations and
ultimately in a loss of convergence. Following Marchal and Crochet [7], we
use a streamline upwind method consisting in modifying the test function 4
but only on the term

&d’)S:+ dx,

which becomes

(12)
6

See ref. 2 for a complete description and for the evaluation of k. This is
often referred to as the inconsistent streamline upwind method. The
modification of the test function introduces a certain amount of numerical
diffusion but was shown by Johnson [9] to be convergent with mesh
refinement.
Combining equations (g-12), we obtain the variational formulation:

/o[277j++i(v)-pV.v] dx= -/-S:Vv dx+l,/,,T.v ds VVEI/,

(13)

qV.u dx = 0 Vq eL2(LR), (14)

/n

S,(~+h[u.V~-V~.S-S.(Vu)=]}:~ dx

+ k(u4’S) :(+.V4) dx=2~,r[/nt(u):(~.V)O

/R

Vu~j(u)+j(r+(V~~):c$ dx-/r(zcn)j(u):c$ ds]. (15)

3. The iterative solver

The aim of this section is to present the iterative method for the solution
of non linear systems of equations and how it can be applied to the system
(13-15).
The method is based on the Newton algorithm for the treatment of
non-linearities and on GMRES for the solution of the associated linear
systems. This general approach if also referred to as a Krylov-Newton
method [lo]. After some preliminaries, we describe the GMRES method
and its application to the solution of viscoelastic fluid flow problems.

3.1 The Newton method

The Newton method is certainly one of the most popular algorithm for
the solution of non-linear system of equations of the form
F(X) = 0. (16)
because of its robustness and quadratic rate of convergence. The algorithm
proceeds as follows:
1. X0 given arbitrarily.
2. for n 2 1, solve
J,SX= -F(X,).
 3. x,+1 =X,+6X.
4. check for convergence and go back to 2 if necessary
The stopping criterion is often based on a combination of the norm of
the residual and the norm of the correction 6X, i.e.
II F(X,+,) II <em
and IlaX II
(17)
< w,
where eps is a prescribed tolerance and II II is a norm on R”.
The main drawback of this algorithm is the computation of the Jacobian
matrix J, and the resolution of the linear systems in step 2. In the case of
viscoelastic fluid flow problems, the size of this matrix can easily reach
30000 equations, even for two-dimensional applications on rather coarse
meshes. This is clearly a serious handicap in view of three dimensional
applications. Hence there is a need for a cost effective alternative.
The idea is quite simple. The Newton method will be used without
computing the Jacobian matrix. The price to be paid for this is the
computation of two smaller matrices, one for the solution of the Stokes
problem ((13) and (14)) and the other for the constitutive equation (15).
The number of unknowns is not reduced but the storage requirements for
the two matrices is much smaller than that of the Jacobian matrix. More-
over, the matrix associated with the Stokes problem is computed and
factorised once and for all. We are left with only a periodic update of the
other matrix.

3.2 The GMRES method

The Generalised Minimal RESidual method is an iterative solver for

possibly non-symmetric linear systems [lo]. We give now a very brief
description of the method.
To solve the system of equations
Ax =b,
we construct a Krylov space K of dimension k much smaller than the
dimension 12of the non linear system. This Krylov space is of the form
K = span{r,, AYE, A2r0,. . . , Ak-‘r,},
where r0 is the initial residual b -Ax,. We then look for a correction
6x E K to x0 minimising the residual
II A@,, + 6~) -b II
or equivalently
II A&x - rg II (18)
8

The above least square problem is then solved through a QR algorithm

which involves a modified Gram-Schmidt orthogonalization procedure.
Once the minimisation has been performed, the process can be restarted
with the new residual until convergence.
In the particular case where the matrix A is the Jacobian of a non-linear
system, the Krylov space can be constructed using the approximation
Jr = F(” + h4 -F(x) (h = 10-6),
h
(19)
avoiding the explicit construction of J. We refer to Saad and Schultz [ll]
for a complete description of the algorithm.
To apply the above method to the solution of viscoelastic fluid flows, all
we have to do at this stage is to define properly the function F appearing
in (16). This is the purpose of the next section.

3.3 Application to viscoelastic fluids

The choice of F must be done in order to obtain a non-linear system

that is as well conditioned as is possible. Here, the variable of interest will
be the tensor S. The velocity u and the pressure p will be considered as
functions of S through the solution of the Stokes problem ((13) and (14)).
This means that for a given tensor S, the resolution of the Stokes problem
will provide a velocity field and a pressure denoted u(S) and p(S)
respectively.
We now write eqn. (15) in matrix form:

where the subscripts u(S) expresses the dependence of the matrix E and
the right-hand side b on the velocity field. It is also a reminder of the
non-linearity of eqn. (20). We can now define the function F as
F(S) = h,(,, - E,,,,S.
However this choice is not the best since no form of preconditioning is
added. This can result in a break down of convergence at very low Deborah
number. We have chosen to use the following definition of F(S):

(21)
where u0 is the best approximation of the velocity field at hand, Usually u0
is chosen as the velocity field obtained at a lower value of the Deborah
number.
A few more remarks are necessary to describe the method. Firstly, all
the Stokes problems are solved by an Uzawa algorithm combined with a
 9

condensation method that eliminates the velocity degrees of freedom

associated with the center of each element and the pressure gradients. This
is achieved by using the continuity equation at the element level, thanks to
the discontinuous approximation of the pressure [12]. This method greatly
reduces the number of unknowns resulting in a very efficient Stokes solver.
It is also worth mentioning that the associated matrix is symmetric and can
be computed and. factorised once and for all.
On the other hand, the matrix EUo in eqn. (21) is not symmetric and has
to be updated periodically to enhance the preconditioning. This matrix is
much larger than the one required in the Stokes problem. This is mainly
due to the discretisation employed for the tensor S.
Finally, the choice of the norm in eqn (17) must be performed according
to the variational formulation of the problem. The usual Euclidian norm is
not the best choice in this case. Since we are looking for a tensor S in
L,(LR), it seems appropriate to use the norm in that space. The same norm
is used in the Gram-Schmidt method to orthogonalise the basis of the
Krylov space. Apparently, this results also in some form of preconditioning
of the non-linear system.

4. Numerical results

We will consider two different but classical problems: the stick-slip and
the 4 to 1 contraction problems. Figures 3 and 4 provide the geometry and
boundary conditions for these problems. Since convergence with mesh

,v---- y--- -1
j4

2 (0.0) 3

3. Symmetry axis

4. v=o

Fig. 3. The stick-slip problem.

 2. u=o,tJ=o

3. Symmetry axis

4. v=o

Fig. 4. The 4 to 1 contraction problem.

refinement is often a delicate issue in computing viscoelastic fluid flows, we

present in Figs. 5 and 6 a close-up of the different meshes that were used
for both problems. In these last two figures, the quadrilaterals used in the
discretisation are decomposed into eight triangles since our post-processer
cannot handle quadrilaterals. Tables 1 and 2 provide the different features
of these meshes in terms of the number of elements NE, number of
velocity unknowns NEQU and stress tensor unknowns NEQS. We have
also indicated the area A of the element adjacent to singular points
(reentrant corner for the contraction and the point where the boundary
condition changes from Dirichlet to Neumann for the stick-slip).
In all the following results, the convergence criterion was

IIF(S) IIL;(n) < 1c

where F(S) is defined by eqn. (21). Moreover, the dimension k of the
Krylov space was fixed to 30.

4.1 The stick-slip problem

Let us recall that the difficulty in this problem comes from the transition
from a Dirichlet boundary condition (stick) to a Neumann boundary
condition (slip). The point where this transition occurs is singular even in
the Newtonian case and it is extremely difficult to approximate the stresses
in the vicinity of that point.
Although the results are encouraging, we have not been able to over-
come the so-called high Weissenberg problem. Table 3 provides the maxi-
 11

-.2

- -1 -. 5 L3 .5 1 1.5 2
I I I I I L

1-.2
1.1

-.l
-.2
-. 5
-1 -. 5 8 .5 1 1.5 2

-1 -.5 0 .5 1 1.5 2
I I I I 1 I 1
Fig. 5. Meshes for the stick-slip problem.
 2
I

CONTRl

1.5
I

CONTR2
Fig. 6. Meshes for the contraction problem.

TABLE 1
Stick-slip problem
Mesh NE NEQU NEQS A
Slip1 210 1639 10584 1.44 E-2
Slip2 336 2639 16704 6.63 E-3
Slip3 384 3023 19008 2.46 E-3
 13

TABLE 2
Contraction

Mesh NE NEQU NEQS A

Contrl 250 1882 12360 7.72 E-3
Contd 326 2474 16080 2.19 E-3

TABLE 3
Stick-slip

Mesh Limit Deborah number

Slip1 9
Slip2 16
Slip3 10

mum value of the Deborah number for which we could get converged
solutions. This number is defined, as usual, as

De =A+,, (22)
where 9, is the wall shear stress in the upstream part of the domain. The
maximum Deborah number increases for the first two meshes up to 16 and
then it drops to 10 for the last one. We believe that this poor performance
with the last mesh is due to a very bad transition in the size of the elements
in the neighbourhood of the singular point. A more appropriate mesh with
a similar refinement near the singularity would require a large number of
elements ( = 600). Our computing facilities prevent however such a calcula-

-I. 8. I. 2. 3. 4. 5. 6. 1.
Fig. 7. Stick-slip: development of u1 along the line y = 0 (Mesh Slip2).
14

De= Is
leae. -

533.-
De = 12

688. -

-28. -18. 8. le. 20, 36. 46. 58.

Fig. 8. Stick-slip: development of S,, along the line y = 0 (Mesh Slip2).

tion with the present method. This can be clearly seen in Fig. 5. This is an
indication that great care must be exercised when refining a mesh. It also
shows that adaptive remeshing strategies could be of valuable help in such
cases.
Let us now take a look to the results. Figures 7-10 present the develop-
ment of ul, Sil, S,, and S, along the line y = 0, passing through the
singular point, for Deborah numbers in the range 0 to 15 obtained with
mesh SLIP2. Figures 11 and 12 show the development of u1 and S,, on the
same axis with the three different meshes at De = 8. These results are
similar to those obtained by Marchal and Crochet [7]. This should certainly
not be a surprise since we are using very similar discretizations and meshes.

6.4

6. - !
1.. -

2. - s-u
De = 6 12 15
6. \Y
-2. -

Fig. 9. Stick-slip: development of S,, along the line y = 0 (Mesh Slip2).

 15

.4

.3

.2

.l
s-u
, De=6Pl5
8.
L
-20. -18. 18. 20. 30. 40. 54
Fig. 10. Stick-slip: development of S,, along the line y = 0 (Mesh Slip2).

.5-

.5-

.4-

.3-

0. - I I
8. 2. 4. 6.
Fig. 11. Stick-slip: development of uI along the line y = 0 at De = 8 for the different
meshes.

Ea.

Era.

4m.

m.

-20. -18. 0. 18. 28. 38. 48. 59.

Fig. 12. Stick-slip: development of S,, along the line y = 0 at De = 8 for the different
meshes.
16

TABLE 4
Contraction
Mesh Limit Deborah number
Contrl 9
Contd 12

4.2 The 4 to 1 contraction problem

This is the most classical benchmark problem. Here again, a singularity

is present at reentrent corner causing very high stresses in its vicinity.

288.
1
S.U De= 2 6 11

158.

188.

5%.

t?. 5 I I

1. -3. -2. -1.

Fig. 13. Contraction: development of S,, along the line y = 0 (Mesh Contr2).

coNTR2De=8
120.-

laa. -
cm De-8
Ea. -

m. - I
1. -
I
28.-

8. sv I i I 1
-1. -3. -2. -1. 6. I. 2. 3. k.
Fig. 14. Contraction: development of S,, at De = 8 along the line y = 0 for the different
 17

1.1’
l.-
.9-
.8-
.l-

.5-

.4-
S.U De= 2 6 11
.3-
, 1 I 1 I 1
-28. -18. 8. 18. ?a. 3. 40. !
Fig. 15. Contraction: development of u1 on the symmetry axis y = 1 (Mesh Contr2).

Here again, the Deborah number is defined by eqn. 22 where i/, is now
computed in the downstream tube.
Table 4 provides the maximum Deborah numbers achieved with our
method. Here again, we get convergence with mesh size. As in the previous
case, it is interesting to look carefully at the evolution of the different
variables in the neighborhood of the singularity (on the line y = 0). Figure
13 shows the development of S,, along this line for Deborah numbers
between 2 and 11 while Fig. 14 presents the results obtained with our two
meshes for the same variable at De = 8. Finally, Fig. 15 shows the well
known overshoot of the first component of the velocity on the symmetry
axis. Comparison with existing results is also excellent in that case.

5. Conclusion

We have presented a decoupled method for the computation of the flow

of viscoelastic fluids. Although we have not reached the level of reliability
of the full Newton method, we can certainly be optimistic concerning the
development of efficient decoupled methods. One should remember that
this is probably the only realistic alternative for tackling three dimensional
applications.
The GMRES method is shown to be much more efficient than the
Picard method. For example, in the 4 to 1 contraction, starting with a
converged solution at De = 2.5, the Picard method converged at De = 2.75
in more than 2100 iterations. GMRES took only 7 iterations to converge
but from a computational point of view, each GMRES iteration is roughly
equivalent to k Picard iterations where k is the dimension of the Krylov
space. Since k was fixed to 30, we can conclude that GMRES was 10 times
18

faster. Finally, GMRES allowed computations at Deborah numbers roughly

twice as high as with the Picard method.
At this time we do not fully understand why our method loses conver-
gence at some value of the Deborah number. Since we know that a solution
exists for higher values (see Ref. 7), we can only presume that the GMRES
method fails to solve the ill-conditioned linear systems associated with high
Deborah numbers. The finite difference (19) can also be a source of error
and more accurate approximations should be also considered.
GMRES is not the only possible solver for non-symmetric linear systems.
Consequently, other iterative solvers should also receive attention and
could possibly replace the GMRES module in our method. These alterna-
tives are however left for future work.

References

1 A. Fortin, and M. Fortin, J. Non-Newtonian Fluid Mech., 36 (1990) 227-288.

2 T.W. Patten, P.W. Chang and B.A. Finlayson, Comput. Fluids, 7 (1979) 267-293.
3 X.L. Luo and R.I. Tanner, J. Non-Newtonian Fluid Mech., 31 (1989) 143-162.
4 M. Fortin, Private communication.
5 J.G. Oldroyd, Proc. Ry. Sot., London Ser. A., 200 (1950) 523-541.
6 M. Fortin and R. Pierre, Comput. Meth. Appl. Mech. Eng., 73 (1989) 341-350.
7 J.M. Marchal and M.J. Crochet, J. Non-Newtonian Fluid Mech., 26 (1987) 77-114.
8 D. Pelletier A. Fortin and R. Camarero, Int. J. Numer. Meth. Fluids, 9 (1989) 99-112.
9 C. Johnson, Numerical Solution of Partial Differential Equations by the Finite Element
Method, Cambridge University Press, 1990.
10 P.N. Brown and Y. Saad, SIAM J. Sci. Stat. Comput., 27 (1990).
11 Y. Saad and M.H. Schultz, SIAM J. Sci. Stat. Comput., 7 (1986) 856-869.
12 M. Fortin and A. Fortin, Int. J. Numer. Meth. Fluids, 5 (1985) 911-928.
13 M.J. Crochet, V. Delvaux and J.M. Marchal, “Numiform 89”, Thomson et al. (Eds.),
Balkema (1989) 17-22.
14 V. Delvaux and M.J. Crochet, Rheol. Acta, 29 (1990) l-10.