Comput. Methods Appl. Mech. Engrg.

193 (2004) 1–23

www.elsevier.com/locate/cma

An efficient solver for the fully coupled solution of

large-displacement fluid–structure interaction problems
Matthias Heil *

Department of Mathematics, University of Manchester, Oxford Road, Manchester M13 9PL, UK

Received 11 March 2003; accepted 19 September 2003

Abstract

This paper is concerned with the fully coupled (ÔmonolithicÕ) solution of large-displacement fluid–structure inter-
action problems by NewtonÕs method. We show that block-triangular approximations of the Jacobian matrix, obtained
by neglecting selected fluid–structure interaction blocks, provide good preconditioners for the solution of the linear
systems with GMRES. We present an efficient approximate implementation of the preconditioners, based on a Schur
complement approximation for the Navier–Stokes block and the use of multigrid approximations for the solution of the
computationally most expensive operations. The performance of the the preconditioners is examined in representative
steady and unsteady simulations which show that the GMRES iteration counts only display a mild dependence on the
Reynolds number and the mesh size. The final part of the paper demonstrates the importance of consistent stabilisation
for the accurate simulation of fluid–structure interaction problems.
Ó 2003 Elsevier B.V. All rights reserved.

Keywords: Fluid–structure interaction; Fully coupled solution; Finite elements; Preconditioning; Multigrid; Stabilisation

1. Introduction

Large-displacement fluid–structure interaction problems arise in many applications in which thin-walled

shell structures interact with fluid flows. Examples from engineering include flutter analyses in aeroelasticity
[1], the inflation of parachutes [2], the deformation of structures by explosions [3] and the design of sails [4].
Numerous applications exist in physiological fluid mechanics and include the flow of blood in the veins and
arteries and the flow of air in the pulmonary airways [5], the dynamics of heart valves [6] and many others.
The computational modelling of large-displacement fluid–structure interaction problems typically re-
quires the solution of the Navier–Stokes equations in a moving domain, coupled to the equations of large-
displacement shell theory. The coupling between the two domains often leads to highly nonlinear behaviour
and the development of robust and efficient solution techniques for such problems presents one of the great

*
Tel.: +44-161-275-5808; fax: +44-161-275-5819.
E-mail address: m.heil@maths.man.ac.uk (M. Heil).
URL: http://www.maths.man.ac.uk/~mheil/.

0045-7825/$ - see front matter Ó 2003 Elsevier B.V. All rights reserved.
doi:10.1016/j.cma.2003.09.006
2 M. Heil / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1–23

challenges in computational mechanics; see, e.g., the recent special volume of this journal [7]. Most com-
putational methods can be classified as partitioned or fully coupled (ÔmonolithicÕ) schemes. Partitioned
methods utilise separate solvers for the fluid and solid domains and attempt to obtain a coupled solution via
a fixed-point iteration. Starting with an initial guess for the wall shape, the fluid equations are solved and
the traction that the fluid exerts on the wall is evaluated. This new traction is then used to update the wall
shape and the iteration is continued until convergence. This approach is relatively easy to implement and is
convenient in situations in which specialised solvers for the fluid and solid problems are readily available;
indeed, if compatibility with existing commercial Ôblack boxÕ fluid and solid solvers is required, this method
presents the only feasible approach. However, partitioned methods also have serious drawbacks. Fixed-
point iterations tend to converge slowly (though Aitken extrapolation can lead to a certain improvement;
see, e.g. [8] and, more recently [9]) and, in the presence of strong fluid–solid coupling, the iteration can
diverge even if a good initial guess for the solution is available. When applied to time-dependent problems,
partitioned solvers lead to staggered time-integration schemes [1,10]. The implementation of such methods
as Ôweakly coupledÕ schemes is straightforward but it is often difficult to ensure their overall stability and
temporal accuracy. ÔStrong couplingÕ, which requires costly sub-iterations at every time-step, can improve
the temporal accuracy, but, of course, only if the sub-iterations converge.
In a fully coupled (ÔmonolithicÕ) approach, the fluid and solid equations are discretised and solved si-
multaneously. For highly transient fluid–structure interaction problems which involve compressible fluids
(e.g. the modelling of explosions [3]), the use of fully coupled but explicit time-stepping schemes can be
efficient. However, if the problem is steady and/or the fluid is incompressible then the solution of a large
system of coupled nonlinear algebraic equations is required. Since good initial guesses for the solutions tend
to be readily available (either from the solution at a previous time-step or from the solution at a previous
value of the control parameter during steady-state parameter studies), the solution of the nonlinear system
by NewtonÕs method is attractive since it provides a robust and rapidly converging scheme. Over the past
years we have successfully used this approach in a wide variety of large-displacement fluid–structure in-
teraction problems [11–15]. The main computational cost of this method arises from the repeated assembly
of the Jacobian matrix and the solutions of the associated linear systems for the Newton corrections.
In this paper, we develop an efficient preconditioning technique that allows the rapid iterative solution of
the linear systems that arise in the Newton method. The preconditioners are based on block-triangular
approximations of the Jacobian matrix and are obtained by neglecting selected fluid–structure interaction
blocks. The structure of the paper is as follows. Section 2 introduces the model problem that we used to
assess the efficiency of the numerical schemes. In Section 3 we present the coupled finite-element discret-
isation of the governing equations and discuss the stabilisation of the fluid equations on coarse meshes.
After presenting some representative simulations in Section 4, we show in Section 5 that the direct use of the
block-triangular approximate Jacobians in the Newton method causes a severe deterioration of its con-
vergence rate. We then demonstrate that the approximate Jacobians provide good preconditioners for the
solution of the linear systems with GMRES. Next, we reduce the cost of the preconditioning operation by
approximating the Navier–Stokes block with a pressure Schur-complement preconditioner and present an
efficient implementation based on multigrid approximations for the computationally most expensive op-
erations. Finally, we discuss the adaptive choice of the convergence tolerance for the GMRES iteration and
demonstrate the importance of consistent stabilisation in the numerical simulation of fluid–structure in-
teraction problems.

2. The model problem

We will assess the efficiency of the numerical schemes in the model problem sketched in Fig. 1. Viscous
fluid is driven through a two-dimensional channel in which part of one wall is replaced by a pre-stressed
 M. Heil / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1–23 3

pext

pentry pexit
x2 a
x1

L up L L down

Fig. 1. Sketch of the model problem considered in this study. Viscous fluid is driven through a channel in which part of one wall is
replaced by a pre-stressed elastic membrane.

elastic membrane. The membrane is loaded by the external pressure pext and by the fluid traction. This
system has been studied by many authors (see e.g. [5], for a recent review) and has features that are rep-
resentative of many other large-displacement fluid–structure interaction problems. Specifically, the system
has highly nonlinear flow characteristics and a propensity to develop large-displacement self-excited os-
cillations (see Section 4).
We model the flow by the unsteady Navier–Stokes equations


oui oui op o2 ui ouj
Re þ uj ¼ þ 2 and ¼ 0; ð1a; bÞ
ot oxj oxi oxj oxj

where we have scaled all lengths on the channel width a, the velocities on the average inflow velocity U and
the pressure on the viscous scale, lU =a. Throughout this paper, the summation convention is used and,
unless stated otherwise, all subscripts range from 1 to 2. The Reynolds number is Re ¼ qaU =l. We impose
steady, fully developed inflow at the upstream end of the channel, u ¼ ð6x2 ð1  x2 Þ; 0Þ at x1 ¼ Lup =a and
assume parallel, axially traction-free outflow at the downstream end x2 ¼ ðL þ Ldown Þ=a. This is equivalent
to setting pexit ¼ 0. No-slip conditions apply on all channel walls.
We model the elastic section of the wall as a pre-stressed, thin-walled elastic beam whose midplane we
parametrise by the nondimensional Lagrangian coordinate f 2 ½0; L=a. In its undeformed configuration,
the wall is located at rw ¼ ðf; 1ÞT and the displacement field vðf; tÞ displaces its material particles to their
new position Rw ¼ rw þ v. In the undeformed position, the elastic wall is subject to an initial stress
r
0 ¼ T0 =h0 (where h0 is the wall thickness and T0 the initial longitudinal tension) which we assume to be
much larger than the additional stress generated by the wallÕs deformation. This allows us to assume in-
crementally linear elastic behaviour which implies that r ¼ r0 þ c, where r ¼ r
=E is the dimensionless
second Piola–Kirchhoff stress and r0 ¼ r
0 =E the dimensionless pre-stress.  E is the incremental
 YoungÕs
2 2
modulus and c is the geometrically nonlinear extensional strain, c ¼ v1;f þ 12 ðv1;f Þ þ ðv2;f Þ , where a comma
denotes a derivative. The wall deformation is governed by the principle of virtual displacements [16],
Z L=a "
2
 #
1 h0 a
ðr0 þ cÞdc þ j dj  f  dRw D df ¼ 0; ð2Þ
0 12 a h0
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
   ffi 2 2
where f ¼ f
=E is the nondimensional traction acting on the wall, D ¼ 1 þ v1;f þ v2;f and
   .
j ¼ v2;ff 1 þ v1;f  v1;ff v2;f D ð3Þ

is the wall curvature. The first two terms in (2) represent the variation in the wallÕs strain energy due to its
extension and bending, respectively. The last term represents the virtual work done by the load f acting on
4 M. Heil / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1–23

the deformed wall. We assume that the elastic membrane is pinned at both ends and impose v ¼ 0 at
f ¼ 0; L=a.
Fluid and solid interact via the no-slip condition
oRw
u¼ on the wall; ð4Þ
ot
and via the (combined) traction on the wall


oui ouj
fi ¼ pext Ni þ Q pdij   Nj ; ð5Þ
oxj oxi
 T
where N ¼ v2;f ; 1 þ v1;f =D is the outer normal on the wall. The nondimensional parameter
lU
Q¼ ð6Þ
aE
represents the ratio of the viscous stresses to the elastic modulus of the wall and provides a measure of the
strength of fluid–structure interaction; if Q ¼ 0, the wall shape only depends on the transmural pressure
difference.

3. Discretisation

3.1. Solid equations

Carrying out the variations with respect to the displacements vi and their derivatives transforms equation
(2) into a variational equation of the form
Z L=a  
/ð0Þ i ð1Þ i ð2Þ i
i dv þ /i dv;f þ /i dv;ff df ¼ 0: ð7Þ
0

We discretise this equation by displacement-based finite elements. The /-terms contain up to second de-
rivatives of the displacements, therefore we need shape functions with continuous first derivatives across the
element boundaries. Isoparametric Hermite elements with nodal displacements and slopes as independent
degrees of freedom [17] were chosen such that the displacements vi were interpolated as
X
vi ¼ V ij wðSÞ
j ; ð8Þ
j

where the wðSÞ

j are piecewise cubic Hermite polynomials; see [18] for further details of the implementation.
This transforms (7) into
 Z L=a   
ð0Þ ðSÞ ð1Þ ðSÞ ð2Þ ðSÞ
/i wj þ /i wj;f þ /i wj;ff df dV ij ¼ 0: ð9Þ
0

The variations of those V ij that are not determined by the boundary conditions are arbitrary and the ex-
pressions multiplied by the corresponding dV ij have to vanish. This provides a system of nonlinear alge-
braic equations for the unknown V ij . These equations still contain the load terms f, which depend on the
fluid pressure and the velocity gradients at the wall (see (5)). The strong nonlinearity of the solid equations
can result in complicated load–displacement paths. In particular, multiple equilibrium configurations can
co-exist at a given external pressure, pext . Consequently, continuation techniques, such as KellerÕs arclength
method [19] are often required to compute strongly deformed equilibrium configurations. In the present
problem, we implemented a simple displacement-control technique by adding the equation
 M. Heil / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1–23 5

fctrl ¼ v2 ðfctrl Þ  Vctrl ¼ 0 ð10Þ

to Eq. (9). Upon choosing a control point (characterised by the value of its Lagrangian coordinate fctrl ) on
the elastic membrane, this equation prescribes the vertical wall displacement at that point and allows us to
treat the external pressure pext as an unknown in steady simulations.

3.2. Fluid equations

3.2.1. Galerkin discretisation

We discretise the Arbitrary Lagrangian–Eulerian (ALE) form of the Navier–Stokes equations (1a,b) on
a body-fitted moving mesh and perform the mesh update in response to changes in the wall position by a
modified version of Kistler and ScrivenÕs ÔMethod of SpinesÕ [20]. Briefly, we associate each nodal point j in
the fluid mesh with a fixed material point on the wall which we identify by its Lagrangian coordinate fðrefÞ
j .
As the wall deforms, each fluid node remains located at a fixed, pre-determined fraction xj 2 ½0; 1 along the
T
straight line from the spine origin R0 ðfÞ ¼ ðf; 0Þ to Rw ðfÞ such that the position vector to fluid node j is
given by
      
Rj ¼ R0 fðrefÞ
j þ xj Rw fðrefÞ
j  R0 fðrefÞ
j : ð11Þ

Fig. 2 illustrates the resulting meshes. Using standard isoparametric Taylor–Hood elements [21], the ve-
locities, the global coordinates and the pressure are represented by
X X X
ui ¼ U ij wðFÞ
j ; xi ¼ X ij wðFÞ
j and p ¼ P j wðPÞ
j ; ð12Þ
j j j

where the wðFÞ

j and wðPÞ
j are bi-quadratic and bi-linear shape functions in the local element coordinates,
ij
respectively. X are the nodal coordinates, specified by (11). We characterise the fluid meshes by hFEM , the
length of the element edges in the undeformed configuration.
ðFÞ
The residuals of the momentum equations (1a) are weighted by the velocity shape functions wl . We
perform the usual integration by parts and obtain
Z Z "
 #
oui   ðFÞ ðFÞ
ðFÞ ðMÞ oui ðFÞ owl oui owl
fil ¼ Re þ uj  uj wl  p þ dv ¼ 0; ð13Þ
Xf ot oxj oxi oxj oxj

Fig. 2. Illustration of the coarsest (hFEM ¼ 1=4) and finest (hFEM ¼ 1=16) fluid meshes used in the computations.
6 M. Heil / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1–23
RR
where P dv is the integral over the fluid domain Xf and
Xf

ðMÞ
X oX ij ðFÞ
uj ¼ w ð14Þ
j
ot j
is the mesh velocity. All time-derivatives were evaluated by a second-order backward Euler scheme (BDF2)
with constant time-step Dt. The fully implicit treatment of the momentum equations leads to a consistent
mass representation. Similarly, the continuity equation (1b) is weighted by the bilinear pressure shape
ðPÞ
functions wl which yields
Z Z
ðPÞ ouj ðPÞ
fl ¼ wl dv ¼ 0: ð15Þ
Xf oxj

3.2.2. Stabilisation by Petrov–Galerkin weighting

If the mesh Reynolds number ReM satisfies
1  
ReM ¼ Re hu  uðMÞ  < 1 ð16Þ
2
in all elements (where  h is a characteristic length of the element), the Galerkin discretisation is known to
yield smooth solutions for the velocity and pressure fields (see, e.g. [22]). For the cases considered in this
study, it is possible to use sufficiently fine meshes to avoid the occurrence of the spurious ÔwigglesÕ that
develop in the solution if the mesh Reynolds number becomes too large. However, the use of multigrid
methods for preconditioning purposes (introduced in Section 5.2.2) requires the solution of the fluid
equations on coarse meshes on which the criterion (16) is severely violated. The development of ÔwigglesÕ
ðFÞ
can be suppressed with a Petrov–Galerkin formulation in which the weighting functions wl for the mo-
mentum equations are replaced by
  ðFÞ
ðF;stabÞ ðFÞ ðMÞ owl
wl ¼ wl þ sstab uk  uk : ð17Þ
oxk
sstab is a stabilisation parameter, which should be (just) large enough to suppress the ÔwigglesÕ. Following the
ideas in [23–25] we set
( n   o1

h 2 u  uðMÞ  1 þ 1 þ 
h
if ReM > 1;
sstab ¼ ReM ðMÞ
Dtjuu j ð18Þ
0 otherwise:
At large ReM , this choice is consistent with the ÔoptimalÕ stabilisation for steady advection diffusion
problems (see, e.g. [26,27]) and it leads to small values of sstab in problems which are dominated by the
unsteady terms since sstab  Dt as Dt ! 0.
Consistent use of the weighting functions wiðF;stabÞ adds the terms
2 0 1 3
Z Z 6 B ou C
ðF;stabÞ 6 B i
 
ðMÞ oui C op o2 ui 7
7
  ðFÞ
ðMÞ owl
fil ¼ sstab 6 ReB þ uj  uj C þ  7 ku  u k dv ð19Þ
Xf 4 @ |{z}
ot oxj A oxi ox2j 5 oxk
|fflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflffl} |{z} |{z}
I II III IV

to the weak form (13) of the momentum equations. The second-order term IV is typically evaluated ele-
ment-by-element. For the purposes of stabilisation (i.e. the suppression of ÔwigglesÕ) it is sufficient to include
ðF;stabÞ
only the advective term II into fil . This is equivalent to the classical SUPG stabilisation of Hughes and
Brooks [28]. However, if the terms I, III and IV are neglected in (19), the weak form of the equations
becomes inconsistent in the sense that an exact solution of (1a,b) no longer satisfies the weak equations. The
choice (18) for sstab ensures that the stabilisation terms disappear as h and Dt ! 0. However, as long as
 M. Heil / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1–23 7

sstab > 0 anywhere in the domain, the inconsistency introduces additional errors into the solution. We will
demonstrate in Section 5.3 that, in fluid–structure interaction problems, these errors can be highly sig-
nificant.

3.3. The coupled solution

We solve the coupled system of nonlinear algebraic equations (9), (10), (13) and (15) by NewtonÕs
method. This requires the repeated assembly and solution of linear systems Jx ¼ r, with the following
block structure
2 30 1 0 1
S Csu Csp xs rs
4 Cus F G 5 @ xu A ¼  @ r u A : ð20Þ
Cps D 0 xp rp
Here rs , ru and rp represent the current residuals of the solid, momentum and continuity equations, re-
spectively and xs , xu and xp are the associated Newton corrections. The diagonal block S in the Jacobian
matrix has two contributions
S ¼ SðsolidÞ þ SðFSIÞ ; ð21Þ
ðsolidÞ
where S is the classical tangent stiffness matrix of the wall equations. The additional contribution,
SðFSIÞ , arises from the fluid–structure interaction: A change in a solid degree-of-freedom affects the velocity
gradients in (5) because the nodal positions in the fluid mesh are determined by the wall displacement field.
The diagonal block


1
F ¼ Re MþN þA ð22Þ
Dt
represents the fluid momentum matrix. Its three components M, N and A arise from the discretisation of
the time-derivative, the ALE advection operator and the viscous term in the Navier–Stokes equations,
respectively. The off-diagonal blocks G and D are the discrete gradient and divergence operators and
without stabilisation we have G ¼ DT . The term 1=Dt in the above equation indicates the dependence on the
time-step as Dt ! 0. The mass matrix M has two contributions
M¼c
M þ MðstabÞ ; ð23Þ
where c M represents the standard mass matrix corresponding to (13) and MðstabÞ arises from the differen-
tiation of the stabilisation term I in (19). Thus, consistent stabilisation makes M asymmetric. However,
MðstabÞ  Dt as Dt ! 0, so for small Dt, the mass matrix is dominated by the symmetric positive definite
contribution cM.
The fluid and solid variables are coupled through the interaction blocks C

. Their physical origin is as
follows: Csp and Csu represent the respective effects of the fluid pressure and the viscous fluid stresses on the
wall equations and arise through the load terms (5). The interaction blocks
1 ðtÞ
Cus ¼ CðsÞ
us þ C and Cps ð24Þ
Dt us
represent the effect of the wall displacement field on the discretised momentum and continuity equations,
respectively. In steady problems, this interaction occurs via the algebraic mesh-update (11): A change in
the wall displacement field changes the fluid mesh and thus the Jacobian of the isoparametric mapping in
the fluid elements. The term Dt1 CðtÞ
us arises from the fact that the time-derivative of the wall displacement
field affects the momentum equations by setting the fluid velocities along the wall via (4) and by setting the
mesh velocities uðMÞ via (14). For small time-steps, the Jacobian matrix therefore has large off-diagonal
entries.
8 M. Heil / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1–23

3.4. Implementation

The repeated assembly of the Jacobian matrix involves a significant amount of computational work and
needs to be optimised carefully. We assembled the blocks F, D and G by the standard finite element
assembly procedure, using the analytical expressions for the integrand and three point Gauss rules for the
numerical integration. The remaining entries in the Jacobian matrix were generated by finite differencing
equations (9), (10), (13) and (15). The number of elements used for the discretisation of the wall equations
(9) was matched to the number of elements in the fluid domain such that their respective nodal points
coincided. This resulted in a finer-than-necessary discretisation of the wall equations but ensured that every
solid degree of freedom only affected the position of a small number of fluid nodes. The (modest) increase in
the total number of degrees of freedom was more than compensated for by a dramatic increase in the
efficiency of the sparse mesh-update-algorithm which is used frequently during the finite difference oper-
ations. The stabilisation parameter sstab was computed element-wise by evaluating (18) at the element
centroids and its value was kept constant during the Newton iteration. h was determined by the procedure
proposed in Ref. [27] to ensure that it represents the element length in the direction of the ÔwindÕ u  uðMÞ .

4. Typical results

Figs. 3–5 present results from representative steady and unsteady simulations which will be used as test
cases for the numerical schemes presented below. All simulations were performed with the parameters
Lup =a ¼ 1, L=a ¼ 5, Ldown =a ¼ 10, Q ¼ 102 , r0 ¼ 103 , pexit ¼ 0, h0 =a ¼ 102 and fctrl ¼ 3:5.

4.1. Steady flows: multiple solutions

Fig. 3 illustrates the changes in the flow-field and the wall deformation when the external pressure is
varied while the flow rate and the fluid pressure at the downstream end of the rigid segment are held
constant. The viscous pressure drop in the downstream rigid segment causes the fluid pressure underneath
the membrane to exceed the fluid pressure pexit ¼ 0 at the outflow. At Re ¼ 500, an external pressure of
ðflushÞ
pext ¼ pext ¼ 1:524 is required to maintain the membrane in an approximately flush state (see Fig. 3a). An
increase in external pressure increases the compressive load on the wall and causes it to collapse more
strongly near its downstream end. At finite Reynolds number, the Bernoulli effect leads to a local reduction
in fluid pressure which augments the wall compression in the region of strongest collapse. The reduction in
channel width also increases the overall viscous flow resistance and requires an increase in upstream

Fig. 3. Streamlines and pressure contours for steady flow at Re ¼ 500. pext ¼ pext =E ¼ 1:524 (a) and pext ¼ 3:247 (b). The fluid pressure
is scaled on the viscous scale, p ¼ p
a=ðlU Þ.
 M. Heil / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1–23 9

Re=500
Re=250
-0.1
Re=0

-0.2
V ctrl

-0.3
A

-0.4

1.5 2 2.5 3 3.5

pext

Fig. 4. Pressure–displacement characteristics for steady flow at various Reynolds numbers. The wall deformation is characterised by
the vertical wall displacement Vctrl ¼ v2 ðf ¼ 3:5Þ at 70% of the membraneÕs length. At large Reynolds number, limit points develop in
the pressure–displacement curve.

0.4

0.3

0.2

0.1

0
V ctrl

-0.1

-0.2

-0.3

-0.4

-0.5

-0.6
0 20 40 60 80 100 120 140 160 180 200
time

Fig. 5. Time history of the vertical wall displacement Vctrl ðtÞ ¼ v2 ðf ¼ 3:5; tÞ for a self-excited oscillation at Re ¼ 500.


pext ¼ pext =E ¼ 2:5.

pressure to maintain the same flow rate. The resulting increase in fluid pressure in the upstream half of the
elastic segment causes it to bulge out (see Fig. 3b). The flow separates downstream of the point of strongest
collapse and a long region of reversed flow develops.
Fig. 4 illustrates the pressure/deformation characteristics of the system at three different Reynolds
number by plotting Vctrl ¼ v2 ðf ¼ 3:5Þ, the vertical wall displacement at 70% of the membraneÕs length, as a
10 M. Heil / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1–23

function of the external pressure. At zero Reynolds number, an increase in external pressure leads to an
approximately proportional increase Vctrl . At finite Reynolds number, the additional local pressure drop
generated by the Bernoulli effect increases the collapse at a given value of pext . At Re ¼ 500, this destabilising
effect is so strong that limit points develop in the Vctrl ðpext Þ curve. Hence, small changes in pext can lead to
large changes in the (steady) wall shapes. For instance, a small increase in the external pressure beyond
ðflushÞ
pext forces the system to jump (dynamically) from its ÔflushÕ state into a strongly collapsed equilibrium
configuration in which Vctrl ¼ 0:313 (marked as point A in Fig. 4). Upon reducing the external pressure,
the system evolves quasi-steadily along the curve from A to the limit point B where it has to undergo a
further discontinuous transition to a strongly inflated state with Vctrl > 0 (not shown in the diagram).

4.2. Unsteady flows––self-excited oscillations

As mentioned in the introduction, fluid-conveying elastic vessels are highly susceptible to the develop-
ment of large-amplitude self-excited oscillations. An example of such oscillations is shown in Figs. 5 and 6.
The simulation was started from the steady equilibrium configuration for Vctrl ¼ 0:35, pext ¼ 1:666. For
t P 0, the external pressure was set to pext ¼ 2:5 and the flow rate was held constant at the upstream end of
the channel. The time trace of the control displacement Vctrl ðtÞ ¼ v2 ðf ¼ 3:5; tÞ shown in Fig. 5 illustrates
that the system performs sustained large-amplitude oscillations with a fundamental period of T  25, with
superimposed higher harmonics which create an oscillation of great complexity.

Fig. 6. Instantaneous streamlines and pressure contours in an unsteady flow undergoing a large-amplitude self-excited oscillation.


t ¼ 2:0 (a), 8.0, 12.0, 14.0 (b–d). Re ¼ 500, pext ¼ pext =E ¼ 2:5. The fluid pressure is scaled on the viscous scale, p ¼ p
a=ðlUÞ.
 M. Heil / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1–23 11

Fig. 6 illustrates the interaction between the wall displacement and the fluid flow during the oscillation.
In Fig. 6(a) and (b), the wall collapses inwards and the size of the separation bubble at the downstream end
of the elastic segment increases. During the second half of the oscillation, when the downstream half of the
membrane moves outwards again, Fig. 6(c) and (d), the separation bubble detaches from the elastic wall,
and a large-amplitude wave pattern (a Ôvorticity waveÕ [29]) develops and travels downstream.

5. Solution strategies

For two-dimensional problems of moderate size, the direct solution of the linear system (20) is perfectly
feasible, both in terms of memory requirements and CPU times. In the simulations presented above, the
linear solves were performed with a frontal method (MA42 from the HSL2000 library [30]) and the Newton
iteration typically converged in 5–7 steps, reducing the maximum residual from Oð1Þ to below 108 . For
larger problems, particularly in 3D, the memory requirements of direct solvers tend to become prohibitive––
mainly because of the fill-in that occurs during the computation of the LU factors. In fluid–structure in-
teraction problems, a large amount of the fill-in is created by the coupling between the fluid and solid
variables. We will now explore alternative approaches which avoid the direct solution of the full system (20).

5.1. Direct solution with block-triangular approximate Jacobians

The additional fill-in associated with the fluid–structure interaction can be avoided by replacing the exact
Jacobian matrices by suitable approximations. We consider the three block-triangular approximations
2 3 2 3 2 3
S Csu Csp S 0 0 S 0 0
Jsup ¼ 4 0 F G 5 Jsub ¼ 4 Cus F G 5 Jdiag ¼ 4 0 F G 5; ð25Þ
0 D 0 Cps D 0 0 D 0
the use of which reduces the solution of the linear system (20) to the independent solution of two smaller
linear systems and an intermediate matrix–vector product. For instance, if we replace J by Jsub , the
Newton correction x can be computed via the sequence
solve Sxs ¼ rs for xs ; ð26Þ




rbu ru Cus
compute ¼ þ x; ð27Þ
rbp rp Cps s




F G xu rb T
solve ¼ u for ðxu ; xp Þ : ð28Þ
D 0 xp rbp
A similar sequence of steps is required if the sub-diagonal blocks are neglected. If all fluid–structure in-
teraction blocks are discarded, the intermediate matrix–vector product (27) can be omitted.
If we also approximate S in (26) by the tangent stiffness matrix SðsolidÞ , the two linear solves in (26) and
(28) involve the same matrices that would arise from the discretisation of the uncoupled fluid and solid
problems. The solution of the two linear systems can therefore utilise well-established solution techniques
for these sub-problems. We will exploit this in Section 5.2.2.
Fig. 7 compares the convergence history of the exact and inexact Newton methods for the steady
problem illustrated in Fig. 3. In all cases the number of Newton iterations required to reduce the maximum
residual below 108 tends to increase with the degree of the channel collapse. Fig. 7 shows the convergence
histories of the fastest and slowest converging Newton iteration over the whole range of the simulation (Vctrl
increasing from 0 to )0.4125). If the exact Jacobian is used, the Newton method displays the expected rapid
12 M. Heil / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1–23

100
Exact Jacobian ℑ
Approximate Jacobian ℑsub
10
-2
Approximate Jacobian ℑsup
Maximum residual Approximate Jacobian ℑdiag

10-4

10-6

-8
10

10-10

-12
10
0 10 20 30 40 50 60 70
Number of Newton iterations

Fig. 7. Convergence histories of the Newton iteration with the exact and approximate Jacobian matrices for a steady computation at
Re ¼ 500. Thin lines: the solution for Vctrl ¼ 0:0938 is used as an initial guess for Vctrl ¼ 0:1125. Thick lines: the solution for
Vctrl ¼ 0:3938 is used as an initial guess for Vctrl ¼ 0:4125.

convergence and the iteration count only increases very slightly (from 5 to 7) with the degree of collapse.
Neglecting the super- or sub-diagonal interaction blocks leads to a noticeable deterioration in the con-
vergence rate. Furthermore, the iteration count increases more rapidly with the degree of collapse (from
16 for Vctrl ¼ 0:1125 to 37 for Vctrl ¼ 0:4125). Neglecting all interaction blocks leads to a staggered
scheme, in which the update of the fluid and solid variables is performed independently. This leads to a
small saving in computational cost during each iteration because the matrix–vector product (27) can be
avoided. However, the convergence rate of the Newton iteration deteriorates dramatically and also
becomes extremely sensitive to the degree of collapse. Neglecting the block SðFSIÞ did not affect the con-
vergence rates in any of the four schemes. Overall, this behaviour is consistent with the observations in [31]
where only the case of neglecting the sub-diagonal block (in our notation) was considered.
When the approximate Jacobian matrices (25) were used in the unsteady simulation shown in Section
4.2, the Newton method failed to converge. Only the blocks Csu and SðFSIÞ could be neglected without
causing immediate divergence. However, in that case, no useful simplification of the block structure is
achieved.

5.2. Iterative solution by preconditioned Krylov subspace solvers

The numerical experiments presented in the previous section show that the use of approximate Jacobian
matrices in the Newton iteration can lead to a severe deterioration of its convergence rate. An alternative
approach is to use the approximate Jacobians as preconditioners in the iterative solution with a Krylov
subspace solver, such as GMRES. The iterative solution of the (right-) preconditioned version of the linear
system (20), JP1 Px ¼ r, involves matrix–vector products with the exact Jacobian matrix J and so-
lutions of linear systems with the preconditioning matrix P. If P is a good approximation to J, we can
expect rapid convergence of the ÔinnerÕ GMRES iteration to the exact solution of the original system (20),
while maintaining the quadratic convergence rate of the ÔouterÕ Newton iteration. Furthermore, the con-
vergence tolerance of the ÔinnerÕ iteration can be coupled to the size of the residual in the Newton iteration
to avoid Ôover-solvingÕ (see [32]). We will explore this in Section 5.2.3.
 M. Heil / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1–23 13

5.2.1. Preconditioning with the block-triangular approximate Jacobians

To illustrate the efficiency of the approximate Jacobians as preconditioners for the iterative solution of
the linear system (20), Fig. 8 shows representative convergence histories of the GMRES iterations in a
steady simulation at Re ¼ 500. The data was obtained from a computation in which the solution for
Vctrl ¼ 0:21 was used as the initial guess for Vctrl ¼ 0:28. In this particular case, the Newton iteration
converges in five steps and the lines in Fig. 8 show the convergence histories of the GMRES iterations for
the first and last Newton steps. During the first Newton steps, the GMRES convergence histories have an
initial stagnant period (which lasts for approximately 50% of the total number of iterations) during which
the scaled residual r ¼ jJx þ rj=jrj only decays very little. This is followed by a period of rapid convergence
during the second half of the iteration. For subsequent Newton steps, the initial stagnant period disappears.
Overall, the preconditioner Jsup produces slightly faster convergence rates than Jsub , especially during the
final Newton steps. The block diagonal preconditioner Jdiag requires approximately twice as many
GMRES iterations as the two block-triangular versions.
Table 1 provides a more comprehensive overview of the iteration counts for the three preconditioners
and presents the average number of GMRES iterations, NGMRES , for steady computations at different
Reynolds numbers. The average was taken over all Newton steps in a computation in which Vctrl was in-
creased from 0 to )0.35 in five steps. In all cases, the GMRES iteration was continued until
r < TOLGMRES ¼ 106 . Table 1 shows that an increase in Reynolds number leads to a slight increase in
NGMRES and confirms Jsup as the most efficient preconditioner. The data shown was obtained with
hFEM ¼ 1=8 but computations at different spatial resolutions displayed no evidence of any mesh-dependence
of the convergence rates.
Table 2 presents the average iterations counts for the three preconditioners in the unsteady simulation at
Re ¼ 500, shown in Section 4.2. Here NGMRES represents the average over all Newton iterations at 20 dif-
ferent time-steps during the systemÕs evolution. The data was obtained by restarting the simulation from the
four representative configurations shown in Fig. 6 and continuing the time-integration for five time-steps.
Selected computations were performed over several periods of the oscillation to confirm that the averages
shown in Table 2 are representative.

100
℘ = ℑdiag
10 -1
℘ = ℑsub

-2 ℘ = ℑsup
10
Scaled residual

10-3

10-4

-5
10

-6
10

10-7

2 4 6 8 10 12 14 16 18 20
Number of GMRES iterations

Fig. 8. Convergence histories of the preconditioned GMRES iteration for a steady computation at Re ¼ 500. The solution for
Vctrl ¼ 0:21 is used as the initial guess for Vctrl ¼ 0:28. Thick lines: first Newton step; thin lines: last (¼fifth) Newton step.
14 M. Heil / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1–23

Table 1
Average GMRES iteration counts for the steady simulations shown in Section 4.1
Re ¼ 0 Re ¼ 250 Re ¼ 500
P ¼ Jsup 3.2 5.4 6.2
P ¼ Jsub 3.8 6.2 7.0
P ¼ Jdiag 6.4 11.0 12.4
The approximate Jacobians Jsup , Jsub and Jdiag are used as right preconditioners for the iterative solution of the linear system (20)
(TOLGMRES ¼ 106 ).

Table 2
Average GMRES iteration counts for the unsteady simulation shown in Section 4.2
Dt ¼ 2  101 Dt ¼ 2  102 Dt ¼ 2  103
P ¼ Jsup 11.5 23.4 45.2
P ¼ Jsub 12.4 24.2 41.2
P ¼ Jdiag 22.7 46.2 87.2
The approximate Jacobians Jsup , Jsub and Jdiag are used as right preconditioners for the iterative solution of the linear system (20).
Re ¼ 500, TOLGMRES ¼ 106 .

Table 2 shows that, as in the steady case, the performance of Jsup and Jsub is very similar while the the
block-diagonal preconditioner Jdiag requires approximately twice as many iterations. The iteration count
increases slightly with a reduction in the time-step, NGMRES  Dt0:3 so that reducing the time-step by a factor
of 10 approximately doubles the number of iterations. This is documented in more detail in Fig. 9 which
compares representative convergence histories of the GMRES iterations (right-preconditioned with
P ¼ Jsup ) for different values of Dt. A reduction in Dt only leads to a slight reduction in the asymptotic
convergence rate of the GMRES iteration but significantly extends the length of the initial stagnant period.

100

-1
10

10-2
Scaled residual

10-3

-4
10

10-5
∆t=2 x10 -1
∆t=2 x10
-2
10-6
∆t=2 x10
-3

-7
10

10 20 30 40
Number of GMRES iterations

Fig. 9. Convergence histories of the GMRES iteration for different time-steps Dt. Jsup is used as the right preconditioner and the
results are for the unsteady simulation of Section 4.2 at t ¼ 14. Re ¼ 500. Thick lines: first Newton step; thin lines: last (¼fifth) Newton
step.
 M. Heil / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1–23 15

Furthermore, at small Dt the initial stagnant period is longer for the final rather than the initial Newton
steps.
The increase in NGMRES with a decrease in Dt presents a significant difference to the behaviour typically
observed in the iterative solution of pure fluid or solid problems. In uncoupled problems, a reduction in the
time-step tends to increase the diagonal dominance of the Jacobian matrix and this typically accelerates the
convergence of iterative solvers. In the present problem, the entries in the sub-diagonal block Cus grow like
1=Dt as Dt ! 0 and, in fact, tend to become the dominant entries in the matrix. Interestingly, the fact that
the block Cus is retained in Jsub does not significantly increase the efficiency of this preconditioner at small
Dt.

5.2.2. A pressure Schur complement preconditioner

The results shown in the previous section demonstrate that the approximate Jacobians present efficient
preconditioners for the iterative solution of the linear system (20). However, in their exact form, the ap-
plication of the preconditioners is very expensive and the computational cost is dominated by the solution
of the Navier–Stokes system (28). For the purposes of preconditioning, its solution does not have to be
computed exactly but can be replaced by a computationally cheaper approximation. For this purpose, we
replace the Navier–Stokes block in (25) by a global pressure Schur complement preconditioner [33]. This
results in preconditioners of the form
2 3 2 3 2 3
S Csu Csp S 0 0 S 0 0
Psup ¼ 4 0 F G 5 Psub ¼ 4 Cus F G 5 Pdiag ¼ 4 0 F G 5; ð29Þ
0 0 X Cps 0 X 0 0 X

where X is (an approximation of) the pressure Schur complement DF1 G. In the following we will use
ElmanÕs BFBt approximation [34]
1
X ¼ ðDGÞðDFGÞ ðDGÞ; ð30Þ
though other approximations could be used in its place. The action of the preconditioners defined by (29)
and (30) can then be obtained in a number of sub-steps. For instance, the solution of the system Psub x ¼ r
can be computed via
solve Sxs ¼ rs for xs ; ð31Þ




rbu ru Cus
compute ¼ þ x; ð32Þ
rbp rp Cps s

solve ðDGÞy ¼ rbp for y; ð33Þ

compute z ¼ ðDFGÞy; ð34Þ

solve ðDGÞxp ¼ z for xp ; ð35Þ

compute bu ¼  rbu  Gxp ; ð36Þ

solve Fxu ¼ bu for xu : ð37Þ

Similar sequences arise for Psup and Pdiag . The procedure can be viewed as an extension of the block
preconditioning strategy for ÔpureÕ fluids problems proposed in Ref. [35]. In this form, the application of the
preconditioner involves the solution of four linear systems and (up to) three matrix–vector products.
16 M. Heil / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1–23

The matrix S in (31) arises from the discretisation of a spatially one-dimensional problem and is
therefore much smaller than the matrices in (33), (35) and (37). The linear system (31) can therefore be
solved efficiently by a sparse direct solver; we used Demmel et al.Õs SuperLU solver [36]. The computa-
tionally most expensive part of the preconditioning operation is the solution of the three linear systems (33),
(35) and (37). We will now replace the exact solves for these systems by suitable approximations to improve
the overall efficiency of the algorithm.
In the absence of stabilisation, the matrix DG is symmetric and positive definite and plays a role similar
to a pressure Poisson operator (see [34]). This suggests that the systems (33) and (35) can be solved effi-
ciently by multigrid, [37], and that suitable approximations can be obtained by performing a fixed number
of multigrid cycles. We performed three V-cycles, using point Gauss–Seidel (with two pre- and two post-
smoothing steps on each level) as the smoother. The interpolation was based on the bilinear pressure in-
terpolation, full weighting was used for the injection and the linear system on the coarsest mesh was solved
with SuperLU.
At finite Reynolds number, the matrix F in (37) is nonsymmetric and has a complex spectrum. Con-
sequently, at large Reynolds number, multigrid with simple point Gauss–Seidel smoothing does not act as
an efficient solver. However, replacing the exact solution of (37) by a fixed number of multigrid V-cycles can
still provide an efficient preconditioner, provided a sufficient amount of stabilisation is added on each
multigrid level. The adaptive choice of the stabilisation parameter sstab via (18) ensures that sstab can be
relatively large on the coarser multigrid levels (on which the solutions only provide corrections for the finer
levels) but remains small enough not to affect the solution at the finest level, provided consistent stabili-
sation is used (see Section 5.3). The implementation of the multigrid solver was based on interpolation
consistent with the bi-quadratic variation of the velocity degrees of freedom and full weighting was used for
the injection. SuperLU was used as the direct solver on the coarsest level. ÔWith-the-windÕ numbering of the
equations is known to be beneficial for the efficiency of Gauss–Seidel smoothing but would be difficult and
costly to implement for the highly complex and time-varying flow fields encountered in the present problem
(see, e.g., Fig. 6). In each smoothing step, we performed four consecutive point Gauss–Seidel sweeps over
all variables, sweeping along the four principal directions of the structured mesh [38].
In this form, the computational work associated with a single application of the preconditioner scales
linearly with the number of unknowns in the problem. Hence, the ÔoptimalityÕ of the solver is directly re-
lated to the number of GMRES iterations required to reduce the scaled residual r below the convergence
tolerance.
Table 3 shows the average values for NGMRES for steady computations at various Reynolds numbers and
spatial resolutions. We only present data for Psup since it tended to be the most efficient preconditioner
throughout. The two numbers shown for each combination of Reynolds number and mesh size correspond
to the cases where the linear systems (33), (35) and (37) were solved exactly (by SuperLU) and approxi-
mately (by the fixed-cost multigrid approximations). The use of the multigrid approximations typically
increased the number of GMRES iterations by about 30% (interestingly, in one case it actually reduced
NGMRES ) but this was more than compensated for by a significant decrease in the overall CPU times. For

Table 3
Average GMRES iteration counts for the steady simulations shown in Section 4.1
Re ¼ 0 Re ¼ 250 Re ¼ 500
hFEM ¼ 1=4 14.4 (15.8) 41.2 (57.0) 46.8 (71.2)
hFEM ¼ 1=8 25.0 (25.8) 71.8 (86.8) 81.2 (104.0)
hFEM ¼ 1=12 35.6 (31.2) 91.0 (116.6) 111.4 (141.4)
hFEM ¼ 1=16 44.8 (46.4) 104.6 (130.2) 130.8 (173.0)
Psup is used as the right preconditioners for the iterative solution of the linear system (20), TOLGMRES ¼ 106 . The numbers in brackets
are for the case when the fixed-cost multigrid approximations for the systems (33), (35) and (37) are used.
 M. Heil / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1–23 17

instance, for the case Re ¼ 500 and hFEM ¼ 1=16, the 30% increase in NGMRES still allowed a reduction in the
total CPU time by a factor of 2.6.
The relative increase in NGMRES with an increase in Reynolds number from Re ¼ 250 to 500 is compa-
rable to that observed for the ÔexactÕ preconditioner Jsup in Table 1, whereas at smaller Reynolds numbers,
0:7
an increase Re has a more detrimental effect. The iteration count scales approximately like NGMRES  hFEM
which is comparable to the mesh dependence reported in Ref. [34], where the performance of the BFBt
preconditioner was examined in an Oseen problem.
Fig. 10 shows that the (significant) differences in the convergence histories at different stages of the
Newton iteration are similar to those observed for Jsup in Fig. 8. However, the deterioration in the as-
ymptotic convergence rate with a reduction in hFEM makes the presence of the initial stagnant period rel-
atively less important for simulations on fine meshes.
Table 4 presents the corresponding data for the unsteady simulations of Section 4.2 at Re ¼ 500.
The iteration counts are very similar to those presented in Table 2 for Jsup , and display very little
mesh-dependence. In fact, the iteration counts for Dt ¼ 2  102 are practically independent of hFEM . Even

0
10
hFEM = 1/4
-1
10 hFEM = 1/8
hFEM = 1/12
Scaled residual

10-2
hFEM = 1/16
-3
10

-4
10

10-5

10-6
100 200
Number of GMRES iterations

Fig. 10. Convergence histories of the preconditioned GMRES iteration for a steady computation at Re ¼ 500. Psup with fixed-cost
multigrid approximations for (33), (35) and (37) is used as the right preconditioner. The solution for Vctrl ¼ 0:21 is used as the initial
guess for Vctrl ¼ 0:28. Thick lines: first Newton step; thin lines: last (¼fifth) Newton step.

Table 4
Average GMRES iteration counts for the unsteady simulation of Section 4.2
Dt ¼ 2  101 Dt ¼ 2  102 Dt ¼ 2  103
hFEM ¼ 1=4 17.0 (17.6) 23.0 (24.2) 25.0 (28.0)
hFEM ¼ 1=8 17.9 (23.1) 24.2 (29.4) 46.0 (50.5)
hFEM ¼ 1=12 22.6 (28.3) 24.2 (30.4) 53.9 (64.6)
hFEM ¼ 1=16 28.2 (44.9) 23.9 (35.6) 55.8 (76.4)
Psup is used as the right preconditioners for the iterative solution of the linear system (20), TOLGMRES ¼ 106 . The numbers in brackets
are for the case when the fixed-cost multigrid approximations for the systems (33), (35) and (37) are used.
18 M. Heil / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1–23

0
10
∆t=2 x 10
-1

10-1 ∆t=2 x 10
-2

∆t=2 x 10 -3
-2
10
Scaled residual

10-3

10-4

10-5

10-6

10 20 30 40 50 60
Number of GMRES iterations

Fig. 11. Convergence histories of the GMRES iteration for different time-steps Dt. Psup with fixed-cost multigrid approximations for
(33), (35) and (37) is used as the right preconditioner and the results are for the unsteady simulation of Section 4.2 at t ¼ 14.
hFEM ¼ 1=12, Re ¼ 500. Thick lines: first Newton step; thin lines: last (¼fifth) Newton step.

the detailed convergence histories shown in Fig. 11 are very similar to those presented for Jsup in Fig. 9,
indicating that the approximate implementation of the preconditioner has little effect on its efficiency.

5.2.3. The choice of the convergence tolerance for the GMRES iteration
In order to allow a meaningful comparison between the various preconditioners, we used a fixed con-
vergence tolerance of TOLGMRES ¼ 106 for all studies in the previous sections. Dembo et al. [39] showed
that the adaptive choice
TOLGMRES ¼ OðjrjÞ ð38Þ
is sufficient to maintain the (asymptotically) quadratic convergence rate of the Newton iteration. (38) shows
that the linear system (20) needs to be solved more accurately as the Newton iteration approaches con-
vergence. This also implies that a natural lower limit for TOLGMRES is given by the convergence tolerance of
the Newton method itself.
In practice, the adaptive choice of TOLGMRES can be rather delicate and problem-dependent since the
analysis underlying the derivation of (38) only applies close to the converged solution. If the initial guess for
the Newton iteration has a large residual then the large value of TOLGMRES suggested by (38) can cause the
Newton method to diverge. Furthermore, any saving in computational work resulting from the reduction in
NGMRES must be compared to the potential additional work created by a possible small increase in the total
number of Newton steps. This is particularly important if, as in the present case, good (but costly to
construct) preconditioners already lead to small numbers of GMRES iterations. In such cases, even the cost
of constructing a single additional Jacobian matrix and the associated preconditioning operators can
outweigh the savings due to the reduction in NGMRES . Finally, the potential benefits resulting from the
adaptive adjustment of TOLGMRES depend strongly on the convergence characteristics of the GMRES it-
eration. If the GMRES iteration has a long initial stagnant period then an increase in TOLGMRES can only
lead to a small reduction in NGMRES .
We performed extensive numerical experiments to determine an ÔoptimalÕ strategy for the choice of
TOLGMRES and found that, for steady problems, the choice
 M. Heil / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1–23 19

8
< 104 if jrj > 104
TOLGMRES ¼ jrj if 106 6 jrj 6 104 ð39Þ
: 6
10 if jrj < 106

resulted in modest savings in CPU times (typically about 20–30% compared to a fixed tolerance of
TOLGMRES ¼ 106 ) while ensuring robustness in the sense that it only rarely resulted in costly additional
Newton steps. The benefits of the adaptive adjustment of TOLGMRES are limited by the fact that, in steady
problems, the GMRES iteration displays the initial stagnant period during the first few Newton steps. In
unsteady problems, the initial stagnant period is more pronounced during the final Newton steps but it
tends to persist for longer, particularly at small time-steps. Furthermore, the number of GMRES iterations
tends to be much smaller than in steady problems, and the adaptive adjustment of TOLGMRES in unsteady
problems only led to small savings in CPU time.

5.3. The importance of consistent stabilisation

The use of multigrid approximations for the solution of the linear systems (33), (35) and (37) presents a
key step in the efficient implementation of the preconditioning strategy. The stabilisation of the fluid
momentum equations, introduced in Section 3.2.2, is necessary to allow the solution of the linear system
(37) on the coarse meshes. The adaptive choice (18) for the stabilisation parameter sstab ensures that only
the minimum amount of stabilisation is applied on each multigrid level and that the stabilisation (whether
applied consistently or inconsistently) becomes insignificant as the element size hFEM and/or the time-step Dt
tend to zero. However, (18) shows that sstab  hFEM as hFEM ! 0 whereas, in the absence of stabilisation, the
finite element expansions (12) ensure the quadratic convergence of the solution under mesh refinement.
Thus, as long as ReM > 1 anywhere in the domain, any errors that are introduced by inconsistent stabili-
sation reduce the rate of convergence under mesh refinement. We will now demonstrate that in fluid–
structure interaction problems, the use of inconsistent stabilisation can introduce significant errors.
Fig. 12(a) and (c) present results of representative steady and unsteady simulations at different spatial
resolutions. (On the coarsest mesh, the flow is severely under-resolved and the Newton method diverges at
t ¼ 8.) The computations were performed with a direct solver and no stabilisation was applied. The figures
demonstrate that the finest mesh (hFEM ¼ 1=16) is adequate to resolve the results to within plotting accuracy.
Fig. 12(b) and (d) present the corresponding results (computed only on the finest mesh) with various forms
of stabilisation. Inconsistent stabilisation can be seen to introduce significant errors even though, in the
absence of stabilisation, the solution is fully converged on this mesh.
The explanation for the strong effect of inconsistent stabilisation in the present problem is that classical
SUPG stabilisation (implemented by including only term II into (19)) introduces artificial dissipation into
the problem. This implies that, for a given velocity field, a larger pressure drop is required to drive the flow.
In fluid–structure interaction problems, this can introduce large errors into the coupled solution since the
wall deformation depends strongly on the fluid pressure. To assess the significance of this effect in the
present problem, Fig. 13 shows the results of a steady computation in which the wall displacement field was
prescribed (a sinusoidal indentation with a maximum inwards displacement of 0.5; the resulting degree of
collapse is comparable to that in the most strongly collapsed configuration in Fig. 12). The lines show the
two velocity components u1 and u2 and the pressure p, plotted along the line x2 ¼ 0:4 (marked as the thick
dashed line in the inset). The figure shows that the different versions of stabilisation have little effect on the
velocity fields. However, the additional dissipation introduced by inconsistent stabilisation leads to a no-
ticeable increase in the pressure drop along the ÔmembraneÕ. With SUPG stabilisation, the fluid pressure at
the upstream end of the ÔmembraneÕ is Dp ¼ 14:6 higher than with consistent stabilisation (or without any
stabilisation at all). If the wall was elastic, this increase in fluid pressure would reduce the degree of collapse.
20 M. Heil / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1–23

0 0

-0.1 -0.1
no stabilisation
fully consistent stabilisation
stabilisation only with term II

V ctrl
V ctrl

hFEM = 1/4
-0.2 hFEM = 1/8 -0.2
hFEM = 1/12
hFEM = 1/16

-0.3 -0.3

1.3 1.4 1.5 1.6 1.7 1.4 1.5 1.6 1.7 1.8
(a) pext (b) pext

0.3 0.3

0.2 hFEM =1/4 0.2 no stabilisation

hFEM =1/8 fully consistent stabilisation
0.1 hFEM =1/12 0.1 stabilisation with terms II, III and IV
hFEM =1/16 stabilisation only with term II
0 0

-0.1 -0.1
V ctrl
V ctrl

-0.2 -0.2

-0.3 -0.3

-0.4 -0.4

-0.5 -0.5

-0.6 -0.6
0 10 20 30 40 0 10 20 30 40
(c) time (d) time

Fig. 12. Demonstration of the errors introduced by inconsistent stabilisation in steady ((a) and (b)) and unsteady ((c) and (d)) sim-
ulations at Re ¼ 500. (a) and (c) document the mesh convergence of the solutions in the absence of stabilisation; (b) and (d) dem-
onstrate the effect of inconsistent stabilisation on the finest mesh. The parameter values are the same as in Section 4.

Conversely, to maintain the same degree of collapse, the external pressure would have to be increased by an
amount comparable to Dp. The fluid pressures (nondimensionalised on the viscous scale) and the external
pressure (nondimensionalised on the elastic scale) are related by the parameter Q, via Dpext ¼ Q Dp (see (5)).
In the computations presented in Fig. 12, Q ¼ 102 , so the increase in external pressure required to ap-
proximately compensate for the rise in fluid pressure is Dpext ¼ 0:146 which is comparable to (in fact, nearly
twice as large as) the error in pext observed in Fig. 12(b) where Dpext ¼ 0:08 for Vctrl ¼ 0:35. This confirms
that the effect of artificial dissipation introduced by inconsistent stabilisation is large enough to explain the
differences observed in Fig. 12(b). In time-dependent simulations in which the external pressure is kept
constant, the increase in dissipation constantly extracts energy from the system. Over sufficiently long time-
scales, this can lead to a significant change in the systemÕs behaviour. For instance, in the simulation shown
in Fig. 12(d), the excessive dissipation causes the rapid decay of the oscillation.
Given that in the present problem, stabilisation is only required for the multigrid preconditioning op-
erations, it seems desirable to set sstab ¼ 0 on the finest mesh and use stabilisation (consistent or incon-
sistent) only for the auxiliary computations on the coarser meshes––following Gresho and LeeÕs maxim not
to ‘‘suppress the wiggles’’ [40]. This approach can be implemented in two ways: (i) Set sstab ¼ 0 for all
 M. Heil / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1–23 21

600
u1
2
1
400

x2
1.5 0
0 1 2 3 4 5
x1 200
u1 and u2

p
1 no stabilisation
stabilisation only with term II 0
p
fully consistent stabilisation

0.5 -200

0 -400
u2

0 1 2 3 4 5
x1

Fig. 13. The effect of inconsistent stabilisation on the velocities and pressures along the line x2 ¼ 0:4 (indicated by the dashed line in the
inset) in a steady simulation at Re ¼ 500. The wall shape is prescribed and the computation is performed on the finest mesh,
hFEM ¼ 1=16.

computations on the finest mesh and use (18) to determine the stabilisation parameter for the coarser
multigrid levels. (ii) Use (18) to determine the stabilisation parameter on all multigrid levels but set sstab ¼ 0
when computing the nonlinear residuals in the Newton method. Both methods ensure that the Newton
method can only converge to the solution for sstab ¼ 0. However, numerical experiments showed that in
case (i) the multigrid-preconditioned GMRES iteration fails to converge, indicating that the fine-grid
problem with sstab ¼ 0 is poorly approximated by the coarse grid problems with sstab 6¼ 0. With strategy (ii),
the GMRES iterations converged but the Newton iteration stagnated, indicating again that the stabilisation
terms still play an important role on the finest meshes. Finally, we mention that the above results are in-
dependent of the pressure nondimensionalisation––we observed exactly the same behaviour when the
pressure was scaled on the inertial scale. We conclude that the preconditioning strategy proposed in this
paper requires the use of consistent stabilisation on all multigrid levels.

6. Summary

We have developed an efficient preconditioning technique that allows the rapid iterative solution of
the linear systems that arise in the fully coupled solution of steady and unsteady large-displacement
fluid–structure interaction problems with NewtonÕs method. The preconditioners were derived from
block-triangular approximations of the Jacobian matrix. While the direct use of these matrices in the
Newton method leads to an unacceptable deterioration of its convergence rate, they provide efficient
preconditioners for the iterative solution of the linear system by GMRES. The block-triangular
structure of the approximate Jacobians makes it possible to perform the preconditioning operation in a
sequence of sub-steps. To improve the overall efficiency of the method, we developed approxi-
mate versions of the preconditioners, by replacing the Navier–Stokes blocks by a global pressure Schur
complement preconditioner, using ElmanÕs BFBt approximation [34] for the pressure Schur comple-
ment. Finally, we replaced the most computationally expensive operations by multigrid approximations.
22 M. Heil / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1–23

In this form, the work involved in a single application of the preconditioner scales linearly with number
of degrees of freedom in the problem. For steady problems, the mesh dependence of the iteration
counts was found to be comparable to that observed in applications of the BFBt preconditioner in
ÔpureÕ fluids problems and the iteration counts only displayed a weak dependence on the Reynolds
number. In time-dependent problems, the convergence rates were generally much faster and displayed
very little mesh dependence, though the iterations counts increased slightly with a reduction in the time-
step.

Acknowledgements

Financial support from the EPSRC is gratefully acknowledged. The author wishes to thank David
Silvester, Milan Mihajlovic and Andrew Hazel for many helpful and enjoyable discussions. The HSL
library routine MA42 (a frontal solver for sparse, unsymmetric systems [30]) and the the GMRES solver
from the CERFACS repository [41] were used in this work.

References

[1] C. Farhat, M. Lesoinne, Two efficient staggered algorithms for the serial and parallel solution of three-dimensional nonlinear
transient aeroelastic problems, Comput. Methods Appl. Mech. Engrg. 182 (2000) 499–515.
[2] K. Stein, R. Benney, T. Tezduyar, J. Potvin, Fluid–structure interactions of a cross parachute: numerical simulation, Comput.
Methods Appl. Mech. Engrg. 191 (2001) 673–687.
[3] F. Casadei, J.P. Halleux, A. Sala, F. Chille, Transient fluid–structure interaction algorithms for large industrial applications,
Comput. Methods Appl. Mech. Engrg. 190 (2001) 3081–3110.
[4] T. Charvet, F. Hauville, S. Huberson, Numerical simulation of the flow over sails in real sailing conditions, J. Wind Engrg. Indust.
Aerodynam. 63 (1996) 111–129.
[5] M. Heil, O.E. Jensen, Flows in deformable tubes and channels––theoretical models and biological applications, in: T.J. Pedley,
P.W. Carpenter (Eds.), Flow in Collapsible Tubes and Past Other Highly Compliant Boundaries, Kluwer, Dordrecht,
Netherlands, 2003, pp. 15–50.
[6] J.D. Hart, G.W.M. Peters, P.J.G. Schreurs, F.P.T. Baaijens, A three-dimensional computational analysis of fluid–structure
interaction in the aortic valve, J. Biomech. 36 (2003) 103–112.
[7] R. Ohayon, C. Felippa, Special issue: Advances in computational methods for fluid–structure interaction and coupled problems,
Comput. Methods Appl. Mech. Engrg. 190 (24–25) (2001).
[8] M. Heil, Stokes flow in an elastic tube––a large-displacement fluid–structure interaction problem, Int. J. Numer. Methods Fluids
28 (1998) 243–265.
[9] D.P. Mok, W.A. Wall, Partitioned analysis schemes for the transient interaction of incompressible flows and nonlinear flexible
structures, in: W.A. Wall, K.-U. Bletzinger, K. Schweizerhof (Eds.), Trends in Computational Structural Mechanics, CIMNE,
Barcelona, Spain, 2001.
[10] C.A. Felippa, K.C. Park, C. Farhat, Partitioned analysis of coupled mechanical systems, Comput. Methods Appl. Mech. Engrg.
190 (2001) 3247–3270.
[11] M. Heil, Finite Reynolds number effects in the propagation of an air finger into a liquid-filled flexible-walled channel, J. Fluid
Mech. 424 (2000) 21–44.
[12] M. Heil, J.P. White, Airway closure: surface-tension-driven non-axisymmetric instabilities of liquid-lined elastic rings, J. Fluid
Mech. 462 (2002) 79–109.
[13] A.L. Hazel, M. Heil, Three-dimensional airway reopening: the steady propagation of a semi-infinite bubble into a buckled elastic
tube, J. Fluid Mech. 478 (2003) 47–70.
[14] O.E. Jensen, M. Heil, High-frequency self-excited oscillations in a collapsible-channel flow, J. Fluid Mech. 481 (2003) 235–268.
[15] A.L. Hazel, M. Heil, Steady finite Reynolds number flow in three-dimensional collapsible tubes, J. Fluid Mech. 486 (2003) 79–103.
[16] G.A. Wempner, Mechanics of Solids with Applications to Thin Bodies, Sijthoff & Noordhoff, Alphen aan den Rijn, 1981.
[17] F.K. Bogner, R.L. Fox, L.A. Schmit, A cylindrical shell discrete element, AIAA J. 5 (1967) 645–750.
[18] M. Heil, T.J. Pedley, Large axisymmetric deformations of cylindrical shells conveying viscous flow, J. Fluids Struct. 9 (1995) 237–
256.
 M. Heil / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1–23 23

[19] H.E. Keller, Numerical solution of bifurcation and non-linear eigenvalue problems, in: P. Rabinowitz (Ed.), Applications of
Bifurcation Theory, Academic Press, New York, 1977, pp. 359–383.
[20] S.F. Kistler, L.E. Scriven, Coating flows, in: J. Pearson, S. Richardson (Eds.), Computational Analysis of Polymer Processing,
Applied Science Publishers, London, 1983.
[21] C. Taylor, P. Hood, A numerical solution of the Navier–Stokes equations using the finite element technique, Comput. Fluids 1
(1973) 73–100.
[22] T.J.R. Hughes (Ed.), Finite Element Methods for Convection Dominated Flows, vol. 34 of AMD, ASME, New York, 1979.
[23] T.E. Tezduyar, D.K. Ganjoo, Petrov–Galerkin formulations with weighting functions dependent on spatial and temporal
discretizations: applications to transient convection–diffusion problems, Comput. Methods Appl. Mech. Engrg. 59 (1986) 49–71.
[24] T.E. Tezduyar, Stabilized finite element formulations for incompressible flow computations, Adv. Appl. Mech. 28 (1992)
1–44.
[25] T.E. Tezduyar, Y. Osawa, Finite element stabilisation parameters computed from element matrices and vectors, Comput.
Methods Appl. Mech. Engrg. 190 (2000) 411–430.
[26] B. Fischer, A. Ramage, D.J. Silvester, A.J. Wathen, On parameter choice and iterative convergence for stabilised discretisations of
advection–diffusion problems, Comput. Methods Appl. Mech. Engrg. 179 (1999) 185–202.
[27] A. Ramage, A multigrid preconditioner for stabilised discretisations of advection–diffusion problems, J. Comput. Appl. Math. 101
(1999) 187–203.
[28] T.J.R. Hughes, A. Brooks, A multi-dimensional upwind scheme with no crosswind diffusion, in: T.J.R. Hughes (Ed.), Finite
Element Methods for Convection Dominated Flows, vol. 34 of AMD, ASME, New York, 1979, pp. 19–35.
[29] T.J. Pedley, K.D. Stephanoff, Flow along a channel with a time-dependent indentation in one wall: the generation of vorticity
waves, J. Fluid Mech. 160 (1985) 337–367.
[30] HSL 2000, A collection of Fortran codes for large scale scientific computation. Available from <http://www.numerical.rl.ac.uk>.
[31] O. Ghattas, X. Li, A variational finite-element method for stationary nonlinear fluid–solid interaction, J. Comput. Phys. 121
(1995) 347–356.
[32] C.T. Kelley, Iterative Methods for Linear and Nonlinear Equations, SIAM, Philadelphia, 1995.
[33] S. Turek, Efficient Solvers for Incompressible Flow Problems––An Algorithmic and Computational Approach, Springer, Berlin,
Heidelberg, New York, 1999.
[34] H.C. Elman, Preconditioning the steady-state Navier–Stokes equations with low viscosity, SIAM J. Sci. Comput. 29 (1999) 1299–
1316.
[35] H.C. Elman, D.J. Silvester, A.J. Wathen, Finite Elements and Fast Iterative Solvers, Oxford University Press, Oxford, to appear
in 2004.
[36] J.W. Demmel, S.C. Eisenstat, J.R. Gilbert, X.S. Li, J.W.H.R. Liu, A supernodal approach to sparse partial pivoting, SIAM J.
Matrix Anal. Appl. 20 (1999) 720–755.
[37] W.L. Briggs, V.E. Henson, S.F. McCormick, A Multigrid Tutorial, second ed., SIAM, Philadelphia, 2000.
[38] D.J. Silvester, personal communication, 2002.
[39] R.S. Dembo, S.C. Eisenstat, T. Steilhaug, Inexact Newton methods, SIAM J. Numer. Anal. 19 (1982) 400–408.
[40] P.M. Gresho, R.L. Lee, DonÕt suppress the wiggles––theyÕre telling you something, Comput. Fluids 9 (1981) 223–253.
[41] V. Frayss, L. Giraud, S.G.J. Langou, A set of GMRES routines for real and complex arithmetics on high performance computers,
CERFACS Technical Report TR/PA/03/3, Public domain software available on <www.cerfacs/algor/Softs>, 2002.