Mathematical and numerical models for

fluid-structure interaction

Fabio Nobile
MOX, Dipartimento di Matematica, Politecnico di Milano
via Bonardi 9, 20133 Milano, fabio.nobile@polimi.it

International Summer School on

Mathematical Modeling and Computation
LSEC, Beijing, China, August 2-14, 2010
Outline

Introduction

Fluid models – Navier-Stokes equations

Derivation of the Navier-Stokes equations
Finite Element approximation

Structural Models and their finite element approximation

Coupled fluid-structure problem

Analysis of a linearized fluid structure model

Linearized fluid models – Added mass technique

Fluid equations in moving domains – ALE formulation

Algorithms for fluid-structure interaction with large displacements

Practical Informations
I Teacher: Fabio Nobile, MOX, Politecnico di Milano, Italy
url: http://mox.polimi.it/ nobile
email: fabio.nobile@polimi.it
I Teaching Assistant: Christian Vergara, Università degli Studi di
Bergamo, Italy
url: http://mox.polimi.it/ vergara
email: christian.vergara@unibg.it
I The course includes hand-on sessions. We will use the free software
FreeFem++ (a library for finite elements)
url: http://www.freefem.org
I students are invited to solve by their own during the afternoons the
exercices proposed.
I Reference material:
I Chapters 3, 8 and 9 of the book: Cardiovascular Mathematics:
Modeling and Simulation of the Circulatory System. L. Formaggia,
A. Quarteroni, A. Veneziani Eds, Springer, 2009.
I slides of the lectures
Introduction
Introduction

When do we have fluid-structure interaction?

Whenever there is a significant exchange of energy between a moving

fluid and a solid structure.

Examples:
I The effect of the wind on civil structures such as bridges, suspended
cables, tall structures like cooling towers or skyscrapers
I The action of the air on aeronautic structures (aeroelasticity, flutter)
I Aero-acoustics
I Vibrations in vessels. Water-hammer effect
I Biomechanics: blood flow in large arteries
I Effect on dams of water movement in a reservoir; sloshing of a fluid
in a tank.
I ...
FSI in aeroelasticity
A Piper PA-30 Twin Commanche,
known as NASA 808, was used at
the NASA Dryden Flight Research
Center as a rugged workhorse in a
variety of research projects associ-
ated with both general aviation and
military projects. The movie shows
a tail flutter test done in 1966.

Glider’s wings instabilities

FSI in civil engineering
The first Tacoma Narrows Bridge
opened to traffic on July 1, 1940. At
the time of its construction the bridge
was the third-longest suspension bridge
in the world. Its main span collapsed
into the Tacoma Narrows four months
later on November 7, 1940, due to
aeroelastic flutter caused by a 67 kilo-
metres per hour wind. The collapse
of the bridge was recorded on 16mm
film by Barney Elliott, owner of a local
camera shop.

car abandoned on the bridge ...

FSI in civil engineering

Galloping of a suspended cable

The vortex shedding frequency shifts
to the natural frequency of oscillation
of the cylinder giving a resonance phe-
nomenon.
This effect is called Vortex Induced Vi-
bration and the range of fluid velocities
for which this happens is named lock-
in region.

Simulation at Re=100, done in

FreeFem++ by S. Beretta, (Master
Thesis, Aerospace Eng., Politecnico di
Milano, 2007).
FSI in haemodynamics
Blood flow interacting with the arterial wall in the aorta
The fluid-structure interaction in the
cardiovascular system is the mecha-
nism that generates pressure waves
traveling from the heart to peripheral
vessels and guarantees a constan pres-
sure at capillary level.
Simulation done in LifeV by P.
Crosetto (Ecole Polytechnique Fed-
erale de Lausanne, 2010)

Idealized 2D simulation of an artificial aortic

valve.
Simulation done in FreeFem by L. Cassani
(Master thesis. Aerospace. Eng., Politec-
nico di Milano, 2010)
FSI in boat dynamics
Simulation of an Olympic Rowing Boat
The performance of a rowing boat is
strictly related (beside the ability of the
rowers) to the surface waves generated
during the motion (energy dissipated
to generate the waves + increased wa-
ter resistence)

Simulations done by A. Mola and N.

Parolini (MOX, Politecnico di Milano)
Which model? Which numerical technique?
I Several models available
I Several numerical tecniques depending on the application

We can at least define two categories of problems

I Small deformation regimes (aeroelsaticity, acoustic vibrations, ... )
I The focus is on small vibrations of the structure
I The fluid domain can be kept fixed
I Often a simplified fluid dynamics may be used (potential flow;
linearization around a steady state flow)
I Large deformations/displacements regimes (solid bodies / particles
moving in a fluid; flutter / galloping; soft structures)
I The focus is on both fluid and structure;
I The fluid domain changes considerably in time. Several techniques
can be used: Arbitrary Lagrangian Eulerian (ALE) formulation,
Fictitious Domains, Immersed Boundary Method, ...
I Lot of energy exchange between fluid and structure

In this course we will address both cases.

Derivation of the Navier-Stokes equations
Derivation of the Navier Stokes equations
We consider a domain Ωf occupied by a fluid moving at velocity u(x, t),
x ∈ Ωf , t > 0 (Eulerian frame of reference)
Ωf

Trajectories of fluid particles:

x(t) : ẋ(t) = u(x(t), t)

Given a scalar function f (x, t) (could be density, temperature, etc.), we

call material derivative its time variation along particle trajectories

Df d
Material derivative: (x, t) = f (x(t), t)
Dt dt

Applying the chain rule we have

3
Df d ∂f X ∂f ∂xi ∂f
= f (x(t), t) = + = + (u · ∇)f
Dt dt ∂t ∂xi |{z}
∂t ∂t
i=1
=ui
Derivation of the Navier Stokes equations
Consider now an arbitrary volume Vt transported by the fluid and a scalar
quantity f (x, t) defined in Ωf
Ωf

Vt
n

Reynolds transport formula

Z Z Z
d ∂f
f dx = dx + f u · n dA
dt Vt Vt ∂t ∂Vt
Z  
∂f
= + div(f u) dx
Vt ∂t
Conservation of mass

Principle of mass conservation: given an arbitrary volume Vt moving with

the fluid, the mass contained in it will not change in time.
We denote by %f (x, t) the fluid density. Then
Z Z  
d by Reynolds thm ∂%f
%f dx = 0, =⇒ + div(%f u) dx = 0
dt Vt Vt ∂t

Given the arbitrariness of the volume Vt , we conclude:

Mass conservation equation (or continuity eq.)
∂%f
+ div(%f u) = 0, in Ωf , t > 0
∂t
Incompressibility assumption
I Many fluids at low velocity can be considered incompressible.
I The incompressibility assumption is acceptable for Mach  1, where
the Mach number is the ration between the fluid velocity and the
speed of sound.

A flow is incompressible if any arbitrary subdomain Vt transported by the

flow does not change its volume.
Z
d by Reynolds thm
1 dx = 0, =⇒ div u = 0
dt Vt

Then, the continuity equation becomes

∂%f
+ div(%f u) =
∂%f
+
%f  :+0 u · ∇% = D%f = 0
divu

f
∂t ∂t Dt
i.e. the density remains constant along particle trajectories.
I A fluid initially homogeneous (%f |t=0 = %0 const) under an
incompressible flow, remains homogeneous at all times, i.e.
%f (x, t) = %0 , ∀x ∈ Ωf , t > 0.
Conservation of momentum
Principle of momentum conservation (Newton’s second law): the
variation of the momentum equals the forces applied to the system
Vt

f
Z Z Z t
d n
%f u dx = f dx + t dA
dt Vt Vt ∂Vt

where
I f are the external volume forces per unit volume (e.g. gravity∗%f )
I t are the internal surface tractions due to the interaction with the
surrounding fluid particles.
The Cauchy postulate asserts that t = t(x, n, t) depends only on the
unit outward normal vector n to Vt in x and not on the particular
volume Vt considered.
Moreover, the Cauchy theorem asserts that t is necessarily a linear
function of n, i.e. there exists a tensor σ f , (called the Cauchy stress
tensor), such that t = σ f · n.
Finally, σ f has to be symmetric if no concentrated couples exist in
the fluid.
Conservation of momentum
By applying the Reynols transport theorem + the divergence theorem +
arbitrariness of Vt , we get
Momentum conservation equation (first form)
∂%f u
+ div(%f u ⊗ u) = f + div(σ f ), in Ωf , t > 0
∂t

Observe that
∂%f u ∂%f ∂u
+ div(%f u ⊗ u) = u + %f + %f u · ∇u + u div(%f u)
∂t ∂t
 ∂t   
∂%f ∂u
=u + div(%f u) +%f + u · ∇u
∂t ∂t
| {z }
=0(continuity eq.)

Momentum conservation equation (second form)

 
∂u
%f + u · ∇u − div(σ f ) = f, in Ωf , t > 0
∂t
Constitutive relations
To characterize a particular fluid, the Cauchy stress tensor has to be
related to the fluid motion.
Stokesian fluid
I For a fluid at rest σ f = −pI (p= fluid pressure)
I σ f is a continuous function of ∇u
I σ f is isotropic =⇒ Q T σ f (∇u)Q = σ f (Q t ∇uQ) for all
orthogonal matricies Q.

From the previous assumptions it follows that σ f has to be a function of

T
the strain rate tensor D(u) = ∇u+∇
2
u

Newtonian fluid: is a Stokesian fluid for which the relation

σ f = σ f (D(u)) is linear. Then, necessarily

σ f (u, p) = −pI + 2µD(u) + λ(div u)I

with µ: dynamic viscosity and λ: second coefficient of viscosity.

Navier-Stokes equations for incompressible, homogeneous,
Newtonian fluids
  
∂u
%f + u · ∇u − div σ f (u, p) = f


∂t





div u = 0
σ f (u, p) = −pI + 2µD(u)






%f = %0 const.


Possible boudary conditions:

• imposed velocity profile: u = g on ΓD
• imposed traction (pressure and/or shear) σ f · n = d on ΓN

Example: flow in a channel

Γwall

Γin
u=0
Γout
I u = g on Γin (inflow velocity profile)
σ.n=0
u=g
I u = 0 on Γwall (perfect adherence to wall)
u=0
Γwall
I σ f · n = 0 on Γout (free stress condition)
Potential flows and Bernoulli law
We consider now an inviscid, incompressible fluid subject to conservative
forces f = ∇V .
If the flow is initially irrotational, it will stay so for all times (Kelvin
theorem)
Assume curl u = 0. Then, for a simply connected domain Ωf , there exists
a potential φ such that u = ∇φ.
I Continuity eq.: div u = 0 =⇒  ∆φ = 0
∂∇φ
Momentum eq.: %f ∂t + u · ∇u + ∇p = ∇V
I

identity u · ∇u = 12 ∇|u|2 :0
using the vector  − u ×
curl
u 

=⇒ ∇ %f ∂φ 1 2
∂t + %f 2 |u| + p − V = 0

∂φ 1
Bernoulli law %f + %f |u|2 + p − V = const
∂t 2

Potential equations
(
∆φ = 0  
∂φ
p = c + V − %f ∂t + 12 |∇φ|2
Effects of viscosity
When considering the flow over a wall, the viscosity forces the fluid
particles to adhere to the wall −→ boundary layers effects. Let
I U: far field velocity; L: characteristic length of the wall
I Re = %f UL/µ: Reynolds number; measures the relative importance
of inertia effects (U ∗ L) versus viscous effects (µ/%f ).
Flow over a flat wall (∂x p = 0)
q
Lx
I boundary layer of thickness z = Re
(increasing along the wall – Blasius
solution) z
y
slope ~
drag forces

x
I shear forces (drag) on the wall
(σ f · n) · ex ≈ µ ∂u
∂z
x

Flow over a curved wall

I The pressure gradient along the wall
causes the fluid to separate from the
wall. This is at the basis of the well
known Von Karman vortex shedding
phenomenon
Turbulence
The Navier-Stokes equations become
unstable when the transport term (u · ∇u)
becomes dominant with respect to the
viscous term (2µD(u)), i.e. when Re  1.

In such a situation, energy is tranferred from large eddies (large vorticous

structures) to smaller ones up to a characteristic scale, so calles
Kolmogorov scale where they are dissipated by viscous forces.

Kolmogorov scale lK ∼ L ∗ Re −3/4

A direct numerical simulation (DNS) aims at simulating all relevant scales

up to the Kolmogorov scale. Therefore, the mesh size has to be h ≈ lK .
 3
L
3D simulazions #Dofs ∼ ∼ Re 9/4
h
Unfeasible for nowadays computers for Re ∼ 104 ÷ 106 .
Kinetic energy decay
It is instructive to look at the fluid kinetic energy E = 21 %f |u|2 and
perform a Fourier analysis with respect to space wavenumbers:
Z Z
Ê (k) = E (x)e ix·k dx dA, with S(k) = {|k| = k}
log E(k) S(k) R3

inertial region

E(k)
=c
k ε 2/3 −
k 5/3

production dissipation
of energy tranfer of energy of energy
−1 −1 log(k)
L lk
(integral scale) (Kolmogorov scale)

The decay Ê (k) ∼ 2/3 k −5/3 is one of the main results of Kolmogorov
theory of turbulence, valid for homogeneous isotropic turbulence
Reynolds Averaged Navier Stokes (RANS) models
Since a Direct Numerical Simuation is most of the times impossible, the
idea behind RANS models is to “formally” average over many repetitions
of the experiment/simulation so as to filter out the small fluctuations at
the inertia and dissipation scale.
Split the fluid motion in u = ū + u0 , p = p̄ + p 0 , with
I ū: mean velocity; p̄: mean pressure (mean = ensamble average)
0 0
I u , p : fluctuations around the mean

By applying the ensamble average operator (·) to the Navier-Stokes

equations for Newtonian incompressible fluids we obtain:


∂ū
%f + div(%f u ⊗ u) − 2µ div D(ū) + ∇p̄ = f̄

Averaged N-S ∂t
div ū = 0

Problem: div(%f u ⊗ u) 6= div(%f ū ⊗ ū)

We need further equations to close the system.
Reynolds Averaged Navier Stokes (RANS) models
Define the
Reynolds stress tensor

τ R = %f (ū ⊗ ū − u ⊗ u) = −%f u0 ⊗ u0

Then

div(%f u ⊗ u) = div(%f u ⊗ u) − div(%f ū ⊗ ū) + div(%f ū ⊗ ū)

| {z } | {z }
=− div τ R =%f ū·∇ū
R
= − div τ + %f ū · ∇ū

RANS equations
  
% ∂ū + ū · ∇ū − div 2µD(ū) + τ R
+ ∇p̄ = f̄
f
∂t
div ū = 0


Turbulence models aim at modeling τ R as a function of ∇ū.

k − ε turbulence model
Introduces two new variables
I k: turbulence kinetic energy (per unit mass) ∼ 12 |u0 |2
I ε: turbulence kinetic energy dissipation rate ∼ 2 %µf D(u0 ) : D(u0 )

2
Reynolds stresses τ R = 2µT D(ū) − %f kI
3
turbulent viscosity µT = Cµ k 2 /ε

Equations for k and :

  
Dk R µT
%f = τ : ∇ū − ε + div (µ + )∇k


Dt σk
2

Dε ε R ε µT
%f = C1 τ : ∇ū − C2 + div (µ + )∇ε


Dt k k σ

with constants (C1 , C2 , Cµ , σk , σ ) to be tuned.

(typical values: C1 = 1.44, C2 = 1.92, Cµ = 0.09, σk = 1, σ = 1.3)
 Numerical approximation of the
Navier-Stokes equations by Finite Elements
Finite elements approximation of the Navier-Stokes
equations

We will consider only flows in laminar regimes (no turbulence models)

Navier-Stokes equations

∂u

 %f + %f u · ∇u − div σ f (u, p) = f f , in Ω, t > 0,
∂t




div u = 0,
 in Ω, t > 0,
 u = g, on ΓD , t > 0,

σ f · n = d, on ΓN , t > 0,





u = u ,
0 in Ω, t = 0
with σ f (u, p) = 2µD(u) − pI
Weak formulation
Let us define the following functional spaces:

V = [H 1 (Ω)]d V0 = [HΓ1D (Ω)]d ≡ {v ∈ V, v = 0 on ΓD }

Q = L2 (Ω)

Remark: if ΓD = ∂Ω, the pressure is defined only up to a constant. The

pressure space should be Q = L2 (Ω) \ R.

We multiply the continuity equation by a test function q ∈ Q and the

momentum equation by a test function v ∈ V0 and integrate over the
domain.
Integration by parts of the stress term:
Z Z Z
− div σ f · v = σ f : ∇v − (σ f · n) · v
Ω Ω ∂Ω

Z Z Z *0 Z


= 2µD(u) : ∇v − pI : ∇v − (σ · n) · v −
f  (σ f · n) · v
Ω Ω | {z } Γ
D ΓN | {z }
=p div v  =d
Weak formulation

Navier Stokes equations – weak formulation

Find (u(t), p(t)) ∈ V × Q, u(0) = u0 , u(t) = g(t) on ΓD such that
Z   Z Z Z Z
∂u
%f + u · ∇u · v + 2µD(u) : ∇v − p div v = f ·v+ d·v
Ω ∂t Ω Ω Ω ΓN
Z
q div u = 0
Ω

∀(v, q) ∈ V0 × Q

Weak solutions exist for all time ([Leray ’34], [Hopf ’51]). Uniqueness is
an open issue in 3D.
Energy estimate

Assume zero forcing terms (f = g = d = 0) and take v = u in the weak

formulation

1 ∂|u|2
Z Z Z
∂u d 1
• %f ·u= %f = %f |u|2
Ω ∂t Ω 2 ∂t dt Ω 2
| {z }
kinetic energy Ek
Z Z
1
Z
 :0 Z 1
• %f (u · ∇u) · u = %f u · ∇|u|2 = −  %f
(div
 u)|u|2 + %f |u|2 u · n
Ω Ω 2  Ω Γ 2
| N {z }
Z Z flux of Ek through ΓN

• 2µD(u) : ∇u = 2µD(u) : D(u) > 0

Ω
|Ω {z }
Z dissipation rate
• p div u = 0 by incompressibility constraint
Ω
Energy estimate
Putting everything together, we have
Z Z
∂Ek
+ 2µ D(u) : D(u) = − Ek u · n
∂t Ω ΓN

Fully Dirichlet problem (ΓD = ∂Ω)

In this case we have
Z
∂Ek
+ 2µ D(u) : D(u) = 0 =⇒ Ek (T ) ≤ Ek (0)
∂t Ω

i.e. the kinetic energy decreases due to viscous dissipation.

R boundary conditions (|ΓN | > 0)

Mixed Dirichlet / Neumann
In this case, the term − Ω Ek u · n is positive on the inflow boundary
u · n < 0 (pumps energy in the system) and negative on outflow
boundaries u · n > 0 (radiates energy outside the domain).
It is a good practice, therefore, to impose Neumann boundary conditions
only on outflow boundaries so that the system has a descresing energy.
Finite element approximation
Introduce finite dimensional spaces of finite element type

Vh ⊂ V; Vh0 = Vh ∩ V0 ; Qh ⊂ Q

Observe that pressure functions need not be continuous.

Moreover let us denote by uh0 ∈ Vh and gh ∈ Vh (ΓD ) suitable
approximations of the initial datum and Dirichlet boundary datum

Finite Element approximation (continuous in time)

Find (uh (t), ph (t)) ∈ Vh × Qh , uh (0) = uh0 , uh (t) = gh (t) on ΓD such that
Z   Z Z
∂uh
%f + uh · ∇uh · v + 2µD(uh ) : ∇v − ph div v
Ω ∂t Ω Ω
Z Z
= f ·v+ d·v
Ω ΓN
Z
q div uh = 0
Ω

∀(v, q) ∈ Vh0 × Qh
Algebraic formulation (b)
φj
Nu
I {φi }i=1 : basis of Vh0 ψl
φi

(b) N b
I {φ } u : basis of Vh \ Vh0 (shape functions
j j=1
corresponding to boundary nodes)
Np
I {ψl }l=1 : basis of Qh
Expand the solution (uh (t), ph (t)) on the finite element basis
b
Nu Nu
(b)
X X
uh (x, t) = ui (t)φi (x) + gi (t)φj (x)
i=1 j=1
| {z }
=gh known term
Np
X
ph (x, t) = pi (t)ψl (x)
l=1

Vectors of unknown dofs (nodal values if Lagrange basis functions are

used)
U(t) = [u1 (t), . . . , uNu (t)]T , P(t) = [p1 (t), . . . , pNp (t)]T
Algebraic system
Replacing in the Galerkin formulation the expansions of uh and ph on the
finite element basis and testing with shape functions φi and ψl we are led
to the following system of ODEs

M dU + AU + N(U)U + B T P = F (U),

u
dt t>0
BU = Fp ,


with
Z Z
Mij = %f φj φi mass matrix Aij = 2µD(φj ) : ∇φi , stiffness matrix
Ω Ω
Z
N(U)ij = uh · ∇φj · φi , !! depends on U; non linear term
Z Ω
Bli = − ψl div φi , divergence matrix
Ω
Z Z Z „ « Z
∂gh
(Fu (U))i = f · φi + d · φi − %f + uh · ∇gh · φi − 2µD(gh ) : ∇φi
Ω ΓN Ω ∂t Ω
Z
(Fp )l = div gh ψl
Ω
A simpler problem – Stokes equation
Let us consider for the moment the steady state Stokes problem


 −2µ div D(u) + ∇p = f, in Ω,

div u = 0, in Ω,


 u = g, on ΓD ,
(2µD(u) − pI) · n = d, on ΓN ,


Weak formulation: find (u, p) ∈ [H 1 ]d × L2 , u = g on ΓD , such that

(
a(u, v) + b(v, p) = F(v), ∀v ∈ [HΓ1D ]d
b(u, q) = 0, ∀q ∈ L2

with
Z Z
a(u, v) = 2µD(u) : ∇v, b(v, p) = − p div v,
Z Ω Z Ω

F(v) = f ·v+ d·v

Ω ΓN
Stokes problem
The stokes problem is well posed (admits a unique solution (u, p) which
depends continuously on the data).
In particular, the bilinear form b(·, ·) satisfies the important property
b(v, q)
inf2 sup ≥β>0
q∈L v∈H 1
Γ
kvkH 1 kqkL2
D

which implies that ∀q ∈ L2 , ∃vq ∈ HΓ1D : b(vq , q) > 0.

Finite elements approximation: find (uh , ph ) ∈ Vh × Qh , uh = gh on ΓD ,
such that
(
a(uh , vh ) + b(vh , ph ) = F(vh ), ∀vh ∈ Vh,0
(*)
b(uh , qh ) = 0, ∀q ∈ Qh

Corresponding algebraic formulation

(
AU + B T P = Fu BT
    
A U F
=⇒ = u
BU = Fp B 0 P Fp

The algebraic system has the typical structure of a saddle point problem
Spurious pressure modes
If these exists a pressure ph∗ such that

b(vh , ph∗ ) = 0, ∀vh ∈ Vh ,

this pressure will not be seen by the first equation of (*). It follows that,
if (uh , ph ) is a solution of (*), then also (uh , ph + ph∗ ) is a solution to the
system and we loose uniqueness of the pressure.
Such a pressure ph∗ is called a spurious pressure mode
A necessary and sufficient condition to avoid the presence of spurious
pressure modes is that the finite elements spaces Vh × Qh satisfy the
b(vh , qh )
inf-sup condition: inf sup ≥ βh > 0
qh ∈Qh vh ∈Vh0 kvh kH 1 kqh kL2

Observe that the inf-sup condition is satified by the continuous spaces

V0 × Q but not necessarily by the finite element spacese Vh,0 ⊂ V0 and
Qh ⊂ Q.
At the algebraic level, the inf-sup condition is equivalent to the request
that Ker(B T ) = ∅
Spurious pressure modes
Spaces that satisfy the inf-sup condition are called compatible.
Assume the spaces (Vh,0 , Qh ) are compatible with a constant βh
independent of h. Then, the optimality of the Galerkin projection holds:
 
kuh −uex kH 1 +kph −pex kL2 ≤ C inf kuex − vh kH 1 + inf kpex − qh kL2
vh ∈Vh qh ∈Qh

Examples of spaces that satisfy the inf-sup condition

P2 / P1 Pbubble
1 / P1 Q2 / Q1

Examples of spaces that do not satisfy the inf-sup condition

P1 / P1 P1 / P0 Q1 / Q1
Examples of spurious modes
u=1

Couette flow:

σ.n=0

σ.n=0
exact solution: u1 = y ; u2 = p = 0
u=0

Pressure Pressure
IsoValue IsoValue
-9.94024 -3.7491
-8.52021 -3.21351
-7.57352 -2.85645
-6.62683 -2.4994
-5.68015 -2.14234
-4.73346 -1.78528
-3.78677 -1.42822
-2.84008 -1.07116
-1.8934 -0.714103
-0.94671 -0.357045
-2.32106e-05 1.34456e-05
0.946664 0.357072
1.89335 0.71413
2.84004 1.07119
3.78673 1.42825
4.73341 1.78531
5.6801 2.14236
6.62679 2.49942
7.57347 2.85648
9.94019 3.74913

Pressure computed with P1 /P0 Pressure computed with P1 /P1

The issue of spurious pressure modes appears in the full Navier-Stokes

equations as well. It is related to the incompressibility contraint.
Back to the Navier-Stokes equations: Temporal
discretization
A simple time integration scheme: Implicit Euler with semi-implicit
treatment of the convective term:
n n−1
% u − u

f + un−1 · ∇un − div(2µD(un )) + ∇p n = f n
∆t in Ω, n = 1, 2, . . .
div un = 0


u0 = u0 , un = gn on ΓD − p n n + 2µD(un ) · n = dn on ΓN

In algebraic form this leads to a linear system to solve at every time step,
of the form
1
" 1 #" # " n #
∆t M + A + N(U
n−1
) B T Un Fu + ∆t MUn−1
=
B 0 Pn Fnp

Observe the structure of the matrix, equal to the one for the Stokes
problem.
Temporal discretization – second order scheme
As an example of second order scheme we consider a second order
Backward differentiation (BDF2) with semi-implicit treatment of the
convective term:

Approximation of the time derivative:

∂u  3un − 4un−1 + un−2

n≈
∂t t 2∆t

Extrapolation of convective field

u · ∇ut n ≈ (2un−1 − un−2 ) · ∇un



Resulting system:
M
" 3M #" # " #
2∆t + A + N(2U
n−1
− Un−2 ) B T Un Fnu + 2∆t (4U
n−1
− Un−2 )
=
B 0 Pn Fnp
Solution of the linear system

After temporal discretization (Implicit Euler, BDF2, ...) we are led to the
linear system
C B T Un
    
F
= u
B 0 Pn Fp
C = αM + A + N(U∗ ) with α, U∗ depending on the time marching
scheme chosen.

Solution by a direct solver is quite unfeasible. In 3D problems we have 3

velocity components + the pressure −→ huge sparse linear system
Direct solvers will fail because of memory requirements for fine meshes.

Iterative methods are more suited, therefore.

Pressure matrix equation
Let us formally eliminate the unknown Un :

1st eq. C U n + B T Pn = F u Un = C −1 Fu − B T Pn


=⇒
−1 T n −1
2nd eq. BUn = Fp =⇒ |BC {z B } P = BC Fu − Fp
| {z }
Σ χ

Pressure matrix equation ΣPn = χ

The matrix Σ = BC −1 B T can not be formed explicitely (it would imply

the computation of C −1 ).
The pressure system can be solved by a preconditioned iterative method
(e.g. GMRES) with a given preconditioner P. Steps needed at each
iteration:
I matrix vector multiplication w = Σy. It implies
I z = BT y
I C t = z (solution of a velocity problem)
I w = Bt
I computation of preconditioned residual
Preconditioners
A simple preconditioner has been proposed by Cahouet-Chabard (’88)
 −1
1 −1
PCC = µMp−1 + K
∆t p

with
Z
Mp : mass pressure matrix (Mp )ij = ψj ψi
ZΩ
Kp : stiffness pressure matrix (Kp )ij = ∇ψj · ∇ψi + εψj ψi , ε1
Ω

Preconditioning step. Computing preconditioned residual PCC z = r

implies
I Mp z1 = r (solve a system with mass matrix – well conditioned)
I Kp z2 = r (solve a system with stiffness matrix – ill conditioned)
I z = µz1 + z2 /∆t
Preconditioners
I An alternative approach to solving the pressure system consists in
solving the full system

C B T Un
    
F
= u
B 0 Pn Fp

with an iterative method (GMRES) with block preconditioner

 
Ĉ 0
P=
B Σ̂

where Ĉ and Σ̂ are good preconditioners of C and Σ, respectively.

I Leads to a similar sequence of operations as the previous approach.
I For Σ̂ one can use the Cahouet-Chabard preconditioner. Better
preconditioners have also been proposed, based on the solution of a
transport diffusion problem on the pressure space, that are more
robust with respect to the Reynolds number (see e.g.
[Elman-Silvester-Wathen ’05]).
Projection methods
The idea of projection methods is to avoid the costly solution of the
saddle point problem and to split the computation of velocity and
pressure at each time step.
Fractional step approach (assume homogeneous Dirichlet boundary cond.)

∂u
NS eqs. %f + L1 u + L2 u =f
∂t |{z} |{z}
viscous+transport terms incompress. constraint

Split in

ũn − un−1
%f + L1 ũn = f n

Step I: ∆t (transport-diffusion eq.)
un | = 0
∂Ω

un − ũn

n
%f ∆t + ∇p = 0


(L2 -proj. on divergence
Step II: div un = 0

 free functions)
 n
u |∂Ω · n = 0
Projection methods
I Step II corresponds to the L2 -projection of ũn onto the divergence
free space Hdiv ,0 = {v ∈ L2 (Ω), div v = 0, v · n = 0}. Indeed
Z
ũn − un
Z Z * 0Z
 *0

·v = − ∇p n ·v =  p n n 

∀v ∈ Hdiv ,0 , %f div v− p v · n = 0.
Ω ∆t Ω Ω ∂Ω
Z Z
n n
=⇒ ũ · v = u · v, ∀v ∈ Hdiv ,0
Ω Ω

I Step II can be equivalently written as

(
%f
∆p n = ∆t div ũn
n
∂n p = 0
un = ũn − ∆t∇p n /%f
Indeed, 1) take divergence of the equation:
un − ũn
„ «
%f
div %f + ∇p n = 0 =⇒ ∆p n = div ũn
∆t ∆t
2) take normal component of the equation on the boundary:
un − ũn
„ «
n · %f + ∇p n = 0 =⇒ ∂n p n = 0
∆t
Chorin-Temam projection algorithm
  n n

% ũ − u + u∗,n · ∇ũn − div(2µD(ũn )) = f n in Ω
f
Step I ∆t
 n
ũ = 0 on ∂Ω
(
%f
∆p n = ∆t div ũn in Ω
Step II
∂n p n = 0 on ∂Ω
Step III un = ũn − ∆t∇p n /%f
where u∗,n is a suitable extrapolation of the velocity at t n .
Remarks:
I The boundary condition ∂n p n = 0 is unnatural. It is a consequence
of the fact that in the projection step we can impose only un · n = 0
on the boundary instead of un = 0. This induces artificial boundary
layers on the pressure. √
I The Chorin-Temam algorithm induces a splitting error O( ∆t). It
can be shown that
1
kun − uex (t n )kH 1 + kp n − pex (t n )kL2 ≤ C ∆t 2
kun − uex (t n )kL2 ≤ C ∆t
Incremental Chorin-Temam algorithm
The classical Chorin-Temam method has the inconvenience that it does
not converge to the correct steady state solution. This comes from the
fact that in Step I the pressure does not appear at all.
Easy remedy: add to the first step the term ∇p n−1 and compute in the
second step a pressure correction δp n = p n − p n−1
Incremental Chorin-Temam scheme

  n n−1

% ũ − u + u ∗,n
· ∇ũn
− div(2µD(ũn )) = f n − ∇p n−1
f
Step I ∆t
 n
ũ = 0 on ∂Ω
(
%f
∆δp n = ∆t div ũn in Ω
Step II n
∂n δp = 0 on ∂Ω
∆t
Step III un = ũn − ∇δp n
%f
Step IV p n = p n−1 + δp n
Further improvements
I The incremental Chorin-Temam scheme can be written also for a
BDF2 discretization of the time derivative. In this case, the scheme
reads:
 n
3ũ − 4un−1 + un−2

Step I %f + (2un−1 − un−2 ) · ∇ũn
2∆t
− div(2µD(ũn )) = f n − ∇p n−1
3 %f
Step II ∆δp n = div ũn
2 ∆t
I For this scheme it has been proved (see e.g. [Guermond ’99]) that

kun − uex (t n )kH 1 + kp n − pex (t n )kL2 ≤ C ∆t

kun − uex (t n )kL2 ≤ C ∆t 2

I Further improvements can be achieved by correcting the pressure in

Step IV as
p n = p n−1 + δp n −µ div ũn
Some Books on the numerical approximation of the
Navier-Stokes equations with finite elements

[CM93] A.J. Chorin, J.E. Marsden. A mathematical introduction to fluid

mechanics. Third edition. Springer-Verlag, 1993
[ESW05] H. Elman, D. Silvester, A. Wathen. Finite elements and fast iterative
solvers: with applications in incompressible fluid dynamics. Oxford University
Press, New York, 2005
[GS00] P.M. Gresho, R.L. Sani. Incompressible Flow and the Finite Element
Method. (Two volumes), Wiley, 2000.
[L08] W. Layton. Introduction to the numerical analysis of incompressible
viscous flows. Society for Industrial and Applied Mathematics (SIAM),
Philadelphia, PA, 2008.
[Q09] A. Quarteroni. Numerical models for differential problems.
Springer-Verlag Italia, Milan, 2009.
Structural models and their finite element
approximation
Structural models – kinematics
Dynamics of structures is more conveniently described in Lagrangian
coordinates, i.e. equations are written in the reference configuration Ωs0
(e.g. the initial configuration).

Lagrangian map: x = x(ξ, t) = Lt (ξ)

I ξ: coordinate of material point

in reference configuration s −ξ s
Ωt
Ω0 η= x
I x: coordinate of material point x
in deformed configuration
ξ
I η(ξ, t) = x − ξ displacement of
material point

Kinematics
∂x
velocity : u(ξ, t) = (ξ, t) = ẋ(ξ, t) = η̇(ξ, t)
∂t
∂2x
acceleration : a(ξ, t) = 2 (ξ, t) = ẍ(ξ, t) = η̈(ξ, t)
∂t
Eulerian versus Lagrangian

Lagrangian frame (mostly used for solids)

Kinematics is described in the reference configuration
position x = x(ξ, t)
velocity u = u(ξ, t) = ẋ(ξ, t)
acceleration a = a(ξ, t) = ẍ(ξ, t)

Eulerian frame (mostly used for fluids)

Kinematics is described in the current (deformed) configuration
position (trajectories) x(ξ, t) : ẋ = u(x, t), x(ξ, 0) = ξ
velocity u = u(x, t)
acceleration a = a(x, t) = Du ∂u
Dt (x, t) = ∂t + u · ∇x u
Measures of strain
Let us introduce the deformation gradient tensor
∂xi
F = ∇ξ x = I + ∇ξ η Fij =
∂ξj
Given an infinitesimal material line segment dξ in Ωs0 , it will be
transformed through the Lagrangian map into dx = Fdξ.

I Right Cauchy-Green tensor C = FT F

Measures the change in kdξk due to the motion.
kdxk2 dxT ·dx dξ T FT Fdξ dξ T dξ
= = = C = vT Cv = λ2 (v)
kdξk2 kdξk2 kdξk2 kdξk kdξk
with v = dξ/kdξk unitary vector.

x=x(ξ,t)

The change in length of dξ

depends on its orientation. v||dξ||
λ(v)||dξ||
s s
Ω0 Ωt
Measures of strain

I Left Cauchy-Green tensor B = FFT

Its inverse measures the change in kdxk due to the motion

dξ = F−1 dx, =⇒ kdξk2 = dxT F−T F−1 dx = dxT B−1 dx

I Green strain tensor E = 12 (C − I)

measures the relative elongation of dξ

kdxk2 − kdξk2 dξ T Cdξ − dξ T dξ

2
= = vT (C − I)v = 2vT Ev
kdξk kdξk2

with v = dξ/kdξk unitary vector.

Deformation of volume elements

x=x(ξ,t)
dV0 dVt Consider an infinitesimal volume
dx3 dV0 = det(dξ 1 , dξ 2 , dξ 3 )
dξ 3 dξ 2 dx2
dx1
dξ 1

After deformation

dVt = det(dx1 , dx2 , dx3 )

= det(Fdξ 1 , Fdξ 2 , Fdξ 3 )
= det(F) det(dξ 1 , dξ 2 , dξ 3 )
= det(F)dV0

dVt
Hence J = det(F) = dV0 measures the change of volumes.
Deformation of surface elements – Nanson’s formula
x=x(ξ,t)
n
Consider an infinitesimal surface element
n0 dx
dA0 with normal n0 and an infinitesimal
dξ dA vector dξ
dA0 infinitesimal volume dV0 = dξ T n0 dA0
After deformation

dVt = dxT ndA = JdV0 = dξ T (Jn0 dA0 )= dxT (JF−T n0 dA0 )

From the arbitrariness of dx we obtain

Nanson’s formula ndA = JF−T n0 dA0

A direct consequence of Nanson’s formula is that divξ (JF−T ) = 0.

Indeed for any volume Vt = Lt (V0 )
Z Z Z Z
0= div(I)dx = ndA = JF−T n0 dA0 = divξ (JF−T )
Vt ∂Vt ∂V0 V0
Equations of motion
The equations of motion are derived in the same way as for fluids.
However, now we want to recast them in the reference configuration Ωs0 .
Conservation of mass:
Z
d
%s (x, t) dx = 0
dt Vt

where %s = %s (x, t) is the material density in the current configuration.

Let %0s be the material density in the reference configuration (density at
initial time). Then, for all t > 0 we have
Z Z
%s dx = %0s dξ =⇒ %s = J −1 %0s
Vt V0

Conservation of momentum:
Z Z Z
d
%s u dx = f dx + tdA
dt Vt Vt ∂Vt

We introduce as usual the Cauchy stress tensor σ s : t = σ s n

Equations of motion
Equations in the reference configuration:
Z Z Z Z
d d
I %s u dx = J%s u dξ = %0s a dξ = %0s η̈ dξ
dt Vt dt V0 |{z} V0 V0
%0s
Z Z
I f dx = J(ξ)f(x(ξ, t)) dξ
Vt V0 | {z }
f 0 (ξ)
I Z Z Z Z
Nanson’s f. −T
tdA = σ s ndA = Jσ s F n0 dA0 = divξ σ 0s dξ
∂Vt ∂Vt ∂V0 | {z } V0
σ 0s

momentum equation %0s η̈ − divξ σ 0s = f 0

where we have introduced the nominal stress tensor (also called first
Piola-Kirchhoff stress tensor) σ 0s = Jσ s F−T . Observe that the
divergence divξ is taken with respect to the reference coordinate ξ.

Balance of moments of forces (couples): leads to σ s = σ T

s
(symmetry of the Cauchy stress tensor).
Nominal stress tensor

σ 0s = Jσ s F−T

I σ 0s is called nominal stress tensor. It has the property that

σ 0s n0 dA0 = σ s ndA

so it effectively represents the “stress tensor” in the reference

configuration
I σ 0s is not symmetric. The symmetry of the Cauchy stress tensor σ s
implies
σ 0s FT = F(σ 0s )T
I One can also define the so called Second Piola Kirchhoff stress
tensor S = F−1 σ 0s = JF−1 σ s FT , which is symmetric. However, it
does not have a direct physical interpretation.
Constitutive relations

Differently than for fluids, an elastic body presents internal stresses as a

consequence of deformation gradients (instead of gradients of
deformation rates).
Cauchy elastic material
I σ 0s is a continuous function of F =⇒ σ 0s = σ 0s (F)
I Frame indifference =⇒ σ 0s (QF) = σ 0s (F)QT for all Q orthogonal.
Implies that σ 0s can be expressed as a funtion of C = FT F only:
σ 0s = σ 0s (C)
I Incompressible materials: one adds the constraint det F = 1 and a
corresponding pressure p (Lagrange multiplier)

σ s = σ̃ s (η) − pI, =⇒ f0 − pJF−T

σ 0s = σ s
Constitutive relations – Green elastic materials
For a Green elastic material (also called hyperelastic) the stress tensor is
derived from a strain energy funtion W
 
∂W ∂W
σ 0s = =⇒ (σ 0s )ij =
∂F ∂Fij

For incompressible materials σ 0s = ∂W

∂F − pJF−T .

I By frame indifference, W (QF) = W (F) for all Q orthogonal. This

implies that the strain energy W can be written as a function of C
∂W ∂W ∂C ∂W
W = W (C), therefore σ 0s = = = 2F
∂F ∂C ∂F ∂C

Observe that in this way σ 0s authomatically satisfies the symmetry

requirement σ 0s FT = F(σ 0s )T since ∂W
∂C is a symmetric matrix.
I Isotropic material: W (FQ) = W (F) for all Q orthogonal.
This relation, together with the frame indifference implies that W
depends only on the eigenvalues of C (or the three principal
invariants of C).
Principal invariants of a matrix
Given a matrix A ∈ R3×3 , its characteristic polynomial

p(A) = det(λI − A) = λ3 − I1 (A)λ2 + I2 (A)λ − I3 (A)

is invariant for changes of coordinate system.

The three coefficients I1 , I2 and I3 are called the principal invariants of
the matrix A. Denoting with λ1,2,3 the three eigenvalues of A, we have

I1 (A)= tr A = λ1 + λ2 + λ3
1
(tr A)2 − tr A2

I2 (A)= = λ1 λ2 + λ1 λ3 + λ2 λ3
2
I3 (A)= det A = λ1 λ2 λ3

It is also useful to compute the derivatives of the invariants with respect

to changes in the matrix:

∂Ik (A) ∂Ik (A) Ik (A + δA) − Ik (A)

∈ R3×3 , : δA = lim
∂A ∂A →0 
Derivatives of the principal invariants of a matrix
∂I1 (A) ∂I2 (A)
• =I • = (tr A)I − AT
∂A ∂A
∂I3 (A)
• = (det A)A−T = Cof A
∂A
Hints on the proof:
∂I1 (A) tr (A+δA)−tr A
1. ∂A
: δA = lim→0 
= tr(δA) = I : δA
−1
∂I3 (A)
3. ∂A
: δA = lim→0 det(A+δA)−det(A)

= lim→0 det(A)(det(I+A

δA)−1

observe now that det(I + A−1 δA) = 1 +  tr(A−1 δA) + O(2 )

2. left as an exercice

Isotropic hyperelastic materials

I Strain energy function: W = W (I1 (C), I2 (C), I3 (C))
I Nominal stress tensor:
 
0 ∂W ∂W ∂I1 ∂W ∂I2 ∂W ∂I3
σ s = 2F = 2F + +
∂C ∂I1 ∂C ∂I2 ∂C ∂I3 ∂C
 
∂W ∂W ∂W 2 −1
= 2F I+ (tr(C)I − C) + J C
∂I1 ∂I2 ∂I3
Examples of constitutive relations

Incompressible materials (J = 1)
µ
I Neo-Hookean: W (C) = 2 (I1 (C) − 3)

∂W
σ 0s = 2F − pJF−T = µF − p Cof F
∂C
µ1 µ2
I Mooney-Rivlin: W (C) = 2 (I1 (C) − 3) − 2 (I2 (C) − 3)

∂W
σ 0s = 2F − pJF−T = µ1 + µ2 tr(FT F) F − µ2 FFT F − p Cof F


∂C

Compressible materials
I Saint-Venant Kirchhoff: W (C) = λ2 tr(E)2 + µ tr(E2 )
with E = 12 (C − I) and (λ, µ) Lamé constants
∂W ∂W
σ 0s = 2F =F = F (λ(tr E)I + 2µE)
∂C ∂E
Examples of constitutive relations
Quasi incompressible materials
We split the deformation gradient in F = Fiso Fvol where
1
I Fiso = J − 3 F: isochoric deformation, det Fiso = 1
1
I Fvol = J 3 I: volumetric deformation

2 2
Correspondingly: C = FT F = J 3 FT
iso Fiso = J Ciso , with det Ciso = 1.
3

Then, the strain energy function takes the form

W (C) = Wiso (Ciso ) + Wvol (J)

and the nominal stress tensor

∂W ∂Wvol ∂J ∂Wiso ∂Wvol ∂Wiso ∂Ciso
σ 0s = = +2F = Cof F + 2F
∂F ∂J ∂F ∂C ∂J ∂Ciso | ∂C
{z }
4th order tensor

Example: quasi incompressible neo-Hookean material

µ K
(J − 1)2 + (log J)2


W (C) = (I1 (Ciso ) − 3) +
2 4
Equations of motion – hyperelastic materials
Compressible materials
∂W
%0s η̈ − divξ σ 0s (η) = f 0 , σ 0s = 2F
∂C

Incompressible materials
(
%0s η̈ − divξ σ 0s (η, p) = f 0 , σ 0s = 2F ∂W
∂C − p Cof F
J=1

Possible boundary conditions

I imposed displacement: η = g on ΓD
I imposed nominal stress: σ 0 n0 = d on ΓN
s

Example: incompressible Neo-Hookean material

 0 0 0
%s η̈ − divξ [µF(η) − p Cof F(η)] = f (η) in Ωs

in Ω0s

J(η) = 1


 η=g on ΓD
 0
σ s n0 = d on ΓN
Weak formulation and Energy estimate – compressible case
Multiply the equation by a test function φ in a suitable functional space
V, vanishing on ΓD , and integrate by parts

∂2η
Z Z Z Z
%0s 2 · φ dξ + σ 0s (η) : ∇ξ φ dξ = 0
f · φ dξ + d · φdA0
Ωs0 ∂t Ωs0 Ωs0 ΓN

Consider a hyperelastic material with σ 0s = ∂W

∂F and zero forcing terms
(f 0 = d = g = 0). The following energy conservation principle holds:
Energy conservation
Z Z
%s 2 %s 2
kη̇(T )kL2 + W (η(T )) = kη̇ 0 kL2 + W (η 0 )
|2 {z } | Ωs0 {z }
2 Ωs0
kinetic energy
elastic energy

Proof: Take φ = η̇
%0 d %0 d
Z Z
%0s η̈ · η̇ dξ = s (η̇)2 dξ = s kη̇k2L2
Ωs0 2 dt Ωs 2 dt
Z Z 0 Z
0 ∂W ∂W
σ s (η) : ∇ξ η̇ dξ = : Ḟ dξ = dξ
Ω s Ωs ∂F Ωs ∂t
0 0 0
Finite element approximation – compressible case
Consider a finite element space Vh ⊂ V (e.g. piecewise continuous
polynomials) defined on a suitable triangulation Th of Ωs0 .
Finite element formulation: find η h (t) ∈ Vh , η h = gh on ΓD , such that

Z 2 Z Z Z
∂ηh
%0s · φh + σ 0s (η h ) : ∇ξ φh = f0s · φh + d · φh dA0
Ωs0 ∂t 2 Ωs0 Ωs0 ΓN

for all φh ∈ Vh , vanishing on ΓD

Introduce a basis {φi }N s

of Vh and expand the displacement η h on the
PNi=1
s
basis: η h (t, ξ) = i=1 ηi (t)φi (ξ).
Then, we obtain a system of non-linear ODEs
Ms d̈s + Ks (ds ) = Fs , t>0 (1)
with
I Solution vector: ds (t) = [η 1 (t), . . . , η Ns (t)]T
Mass matrix: (Ms )ij = Ωs %0s φi φj
R
I
0

Stiffness (nonlinear) term: (Ks (ds ))i = Ωs σ 0s (η h ) : ∇ξ φi

R
I
0
Finite element approximation – incompressible case
∂W
Nominal stress tensor: σ 0s = σ
f0 − p Cof F, with σ
s
f0 =
s ∂F .
Let V and Q be suitable spaces for the displacement η and the pressure
p, and Vh ⊂ V and Qh ⊂ Q suitable finite element subspaces.
Finite element form: find (η h (t), ph (t)) ∈ Vh × Qh , η h = gh on ΓD , s.t.
∂ 2 ηh
Z Z h i

 %0s · φh + f0 (η ) − ph Cof F(η ) : ∇ξ φ
σ s h h h


 Ωs ∂t 2 Ωs0
 0
 Z Z
= f0s · φh + d · φh dA0
 Ω s ΓN

 Z 0

 (J(η h ) − 1)qh = 0


Ωs0

for all qh ∈ Qh and φh ∈ Vh , vanishing on ΓD .

(
Ms d̈s + Ks (ds ) + G (ds )ps = Fs
Algebraic nonlinear system:
D(ds ) = Fp
PNp
I ph (t) is expanded
R on a basis of Qh : ph (ξ, t) = l=1 pl (t)ψl (ξ)
I [Gs (ds )]il = Ωs Cof F(η h ) : ∇ξ φi ψj
R 0
I [Ds (ds )]l = Ωs J(η h )ψl
0
Steady state case – Newton method
Let us consider the steady state continuous compressible problem (or its
finite element approximation equivalently)

Z Z Z
∂W
σ 0s (η) : ∇ξ φ dξ = 0
f · φ dξ + d · φdA0 , with σ 0s =
Ωs0 Ωs0 ΓN ∂F

To solve numerically this problems we can use Newton’s algorithm.

Define
Z Z Z
L(η; φ) = σ 0s (η) : ∇ξ φ dξ − f 0 · φ dξ − d · φdA0
Ωs0 Ωs0 ΓN

and
L(η + δη; φ) − L(η; φ)
L0 (η; δη, φ) = lim
→0 

Newton’s method
given η 0 , compute for k = 1, . . . until convergence

L0 (η k ; δη, φ) = −L(η k , φ), η k+1 = η k + δη

Steady state case – Newton method
Observe that F = ∇ξ η =⇒ δF = ∇ξ (δη)

∂σ 0s  ∂ 2 W 
Z  Z 
L0 (η k ; δη, φ) = ( δF) : ∇ ξ φ = (  ∇ξ (δη)) : ∇ξ φ
Ωs0 ∂F ηk Ωs0 ∂F∂F η k

∂ 2 W  ∂δηl ∂φi
XZ 
=
Ωs0 ∂Fij ∂Flk η k ∂ξk ∂ξj

ijlk

Tangent problem

∂ 2 W 
Z  Z Z Z
0
( ∇ ξ (δη)) : ∇ ξ φ = f · φ + d · φ − σ 0s (η k ) : ∇ξ φ
Ωs0 ∂F∂F η k

Ωs0 ΓN Ωs0

Algebraic equivalent Ks0 (dks )δds = Fs − Ks (dks )


∂2W 
where (Ks0 (dks ))ij = Ωs ( ∂F∂F
R
 ∇ξ φj ) : ∇ξ φi
0 ηk
Steady state case – Newton method (incompressible case)
Let us consider now the incompressible case
Z Z Z Z
f0 (η) : ∇ξ φ − s


 σ s p Cof F(η) : ∇ ξ φ = f 0 · φ + d · φ dA0
ZΩs0 Ωs0 Ωs0 ΓN

 (J(η) − 1)q = 0


Ωs0

which we write as L(η, φ) + G(η, p; φ) − D(η; q) = 0

with L defined as before and
Z Z
G(η, p; φ) = − p Cof F(η) : ∇ξ φ, D(η; q) = − (J(η) − 1)q
Ωs0 Ωs0

The Newton method then reads: given (η 0 , p 0 ), compute for k = 1, . . .

until convergence

0 k 0 k k 0 k k 0 k
L (η ; δη, φ) + Gη (η , p ; δη, φ) + Gp (η , p ; δp, φ) − D (η ; δη, q)

= −L(η k , φ) − G(η k , p k ; φ) + D(η k ; q)

 k+1
η = η k + δη, p k+1 = p k + δp
Steady state case – Newton method (incompressible case)
Observe that
Z Z
∂J
D0 (η k ; δη, q) = − q δF = − q Cof F(η k ) : ∇ξ δη = Gp0 (η k , p k ; q, δη)
Ωs0 ∂F Ωs0

Therefore, the tangent problem can be written in abstract form as

Tangent problem
(
ak (δη, φ) + b k (φ, δp) = −L(η k ; φ) − G(η k , p k ; φ)
b k (δη, q) = −D(η k ; q)

with
∂ 2 W 
Z  Z  
∂ Cof F
ak (δη, φ) = ( ∇ ξ (δη)) : ∇ ξ φ − p k
∇ ξ (δη) : ∇ξ φ
Ωs0 ∂F∂F ηk Ωs0 ∂F
Z
k
b (φ, δp) = − δp Cof F(η k ) : ∇ξ φ
Ωs0

We see from here that the tangent problem has the usual structure of a
saddle point problem (Stokes) with a constraint that resembles a
“divergence” constraint. At the finite element level, we have to take
compatible finite element spaces (i.e. that satisfy the inf-sup condition).
Time discretization - Newmark scheme

We can use any ODE solver to solve the system (1). A popular scheme is
the following Newmark scheme

Define dns ≈ ds (t n ), ḋns ≈ ḋs (t n ), d̈ns ≈ d̈s (t n ), and two parameters

γ ∈ [0, 1] and β ∈ [0, 12 ].

Ms d̈ns + Ks (dns ) = Fns , (eq. collocated at time t n )

ḋns = ḋn−1
s + ∆t(γ d̈ns + (1 − γ)d̈n−1
s ), (Taylor expansion for ḋs )
2
∆t
dns = dn−1
s + ∆t ḋn−1
s + (2β d̈ns + (1 − 2β)d̈n−1
s ), (Taylor expansion for ds )
2

I Leads to a nonlinear system in (dns , ḋns , d̈ns ) at each time step.

1
I For linear second order systems it is unconditionally stable for β ≥ 4
and second order accurate for γ = 12 .
Time discretization - Newmark scheme

In the Newmark scheme, we can use the expansions for dns and ḋns to
express the acceleration d̈ns in terms of the displacement dns .

1 1 1 − 2β n−1
d̈ns = dn − ζ n , with ζ n = (dn−1 + ∆t ḋn−1 )+ d̈s
β∆t 2 s β∆t 2 s s
2β

In this way, the algebraic system can be written in the displacement only
1
Ms dns + Ks (dns ) = Fns + Ms ζ n
β∆t 2

At each time step we have to solve a nonlinear system in the

displacement vector dns . This can be done with Newton method as
illustrated in the steady state case.
Time discretization - Mid Point scheme
Another popular scheme is the following:
The equation is colocated at t n−1/2 . Central finite differences are used to
approximate both the velocity and the acceleration:
n n−1
  n n−1

Ms ḋs − ḋs + Ks ds + ds
 n− 1
= Fs 2
∆t 2

n
 ds − ds n−1 n n−1
 ḋ + ḋs
 = s
∆t 2
1
I Very similar to the Newmark scheme with the choice γ = 2 and
β = 14 .
I Second order accurate
I No numerical dissipation for linear problems.
Also in this case we can elimitare the velocity ḋns in the first equation and
n− 1
solve the problem in displacement only (most convenient variable ds 2 )
 4 2
n− 12 n− 12 n− 12
   
n−1 n−1

 ∆t 2 M s d s + K s ds = F s + M s 2ds + ∆t ḋs
∆t 2

n n− 12 n−1
 d s = 2d s − ds
1
4 n− 2

 n 4 n−1
ḋs = ∆t ds − ∆t ds − ḋn−1 s
Infinitesimal elasticity
We consider now small perturbations δη  1 around the reference
configuration (η = 0) and assume that the reference configuration is
stress free.
 
I reference state: η = 0, F = C = I, σ 0s  = 2F ∂W 
F=I
=0
∂C F=I
I perturbed state δη: define the linearized strain tensor
∇ξ δη+∇T
ξ δη
D(δη) = 2 . Then
δF = ∇ξ δη
= (FT δF + δFT F) = ∇ξ δη + ∇T
 
δC η=0 η=0 ξ δη = 2D(δη)

*0
∂ 2 W  ∂ 2 W 
   
∂W  
δσ 0s 

η=0
= 2δF  
  + 2F δC= 4 D(δη)
∂C η=0 ∂C∂C η=0 ∂C∂C η=0


It can be shown that for any isotropic material

∂ 2 W 

λ
4 = δij δlk + µ(δil δjk + δik δjl )
∂Cij ∂Clk η=0
 2

for suitable constants λ and µ (called Lamé constants)

Infinitesimal elasticity
Linearization of equations of motion:
!
∂2η
Z Z Z Z
∂W 0
δ %0s 2 · φ + 2F(η) : ∇ξ φ = δf · φ + δd · φ
Ωs0 ∂t s
Ω0 ∂C Ωs0 ΓN

Leads to
! Z
2
∂2W
Z Z Z
0 ∂ δη 0
%s · φ + 4 D(δη) : ∇ ξ φ = δf · φ + δd · φ
Ωs0 ∂t 2 Ωs0 ∂C∂C Ωs0 ΓN

Corresponds to the well known

Linear elasticity equations

∂ 2 δη
%s − div σ s (δη) = f, in Ωs0 ,
∂t 2
σ s (δη) = λ(div δη)I + 2µD(δη)

valid for small displacements.

Infinitesimal elasticity – incompressible case
In the incompressible case, we have σ 0s = 2F ∂W
∂C − p Cof F. Observe that
in the stress free configuration we must have
 
∂W  ∂W  1
σ 0s η=0 = 2

− pI = 0, =⇒ = pI
∂C η=0 ∂C η=0 2

We recall moreover the following formula:

∂ Cof F F→I
δF = J tr(F−1 δF)F−T − JF−T δFT F−T −→ (tr δF)I − δFT
∂F
Therefore
∂ 2 W 
 
∂W  ∂ Cof F
δσ 0s η=0

= 2δF + 2F δC − δp Cof F − p δF
∂C η=0
 ∂C∂C η=0
 ∂F
∂ 2 W 

= p∇ξ δη + 4 D(δη) − δpI − p(divξ δη − ∇Tξ δη)
∂C∂C η=0
!
∂ 2 W 

= 2p + 4 D(δη) − δpI − p divξ δη
∂C∂C η=0
Infinitesimal elasticity – incompressible case

On the other hand, the linearization of the incompressibility constraint

leads to
∂J F→I
0 = δ(J − 1) = : δF = JF−T : δF −→ tr δF = divξ δη = 0
∂F

Linear incompressible elasticity equations


∂ 2 δη
%s − div σ s (δη, δp) = f, in Ωs0 ,

∂t 2
div δη = 0, in Ωs
0

σ s (δη, δp) = 2(µ + p)D(η) − δpI

which, in the steady state case, corresponds to a Stokes problem.

Some Books on non-linear elasticity and its numerical
approximation

[O84] R.W. Ogden. Nonlinear elastic deformations. Halsted Press [John Wiley
& Sons, Inc.], New York, 1984
[BW08] J. Bonet, R.D. Wood. Nonlinear continuum mechanics for finite
element analysis. Second edition. Cambridge University Press, 2008.
[BLM00] T. Belytschko, W.K. Liu, B. Moran. Nonlinear finite elements for
continua and structures. John Wiley & Sons, Ltd., Chichester, 2000.
[ZT00] O.C. Zienkiewicz, R.L. Taylor. The finite element method. Vol. 2.
Solid mechanics. Fifth edition. Butterworth-Heinemann, Oxford, 2000
Coupled fluid-structure problem
Eulerian versus Lagrangian description
We consider now an incompressible fluid interacting with an elastic
structure featuring possibly large deformations.
As we have seen,
I Fluid equations are typically written in Eulerian form
I Structure equations are typically written in Lagrangian form on the
reference domain Ωs0 .

WARNING: The fluid equations are defined in a moving domain Ωft

Current Configuration Reference configuration

(at time t)
Ω st Ω s0
η

Γt
Γ0
A) Ω ft Ω f0
(u,p)
x2 ζ2
x1 ζ1
Fluid structure coupling conditions
At the common interface Γt (evolving in time), we impose
I Continuity of velocity (kinematic condition)
I Continuity of the normal stress (dynamic condition)
Warning: the continuity of stresses on Γt involves the physical stresses,
i.e. the Cauchy stress tensors.

∂η
(cont. velocity) u(t, x) = (t, ξ), with x = Lt (ξ)
∂t
(cont. normal stress) σ f (u, p)nf = −σ s (η)ns

1
Remember the relation σ s (η) = J(η) σ 0s (η)FT (η) between the Cauchy
stress tensor and the nominal stress tensor.
n nf

Ω st
η

Normal convention: nf = −ns

Γt
ns Ω ft
The coupled fluid-structure problem


∂u
% f
 + %f div(u ⊗ u) − div σ f (u, p) = f f
∂t in Ωft (η),

div u = 0

∂2η
%0s − divξ σ 0s (η) = f0s , in Ωs0 ,
∂t 2

∂η
u(t, x) = (t, ξ), with x = Lt (ξ) on Γt
∂t

σ f (u, p)nf = −σ s (η)ns . on Γt

Sources of non-linearity

The coupled fluid-structure problem is highly non-linear.

Sources on non-linearity are:

I Convective term div(u ⊗ u) of the Navier-Stokes equations

I Non-linearity in the structure model when using non-linear elasticity
I The fluid domain is itself an unknown
Global weak formulation

Fluid eqs. Multiply by (v, q) ∈ [HΓ1D (Ωft )]3 × L2 (Ωft )

Z   Z Z
∂u
%f + div(u ⊗ u) ·v+σ f : ∇v+div uq = f f ·v+ (σ f · nf ) · v
Ωft ∂t Ωft Γt

Structure eq. Multiply by φ chosen in a proper functional space

∂2η
Z Z Z
%0s · φ + σ 0s (η) : ∇ξ φ = f0s · φ + (σ 0s (η) · ns0 ) · φ
Ωs0 ∂t 2 Ωs0 Γ0
Global weak formulation - cont.

•If we take matching test functions at the interface: v ◦ Lt (ξ) = φ(ξ)

and thanks to the coupling condition (continuity of stresses), the
interface terms perfectly cancel.

Fluid-Structure functional space

V ≡ {(v, q, φ) : v ◦ Lt = φ on Γ0 }

Z  
∂u
%f + div(u ⊗ u) · v + σ f (u, p) : ∇v + div uq+
Ωft ∂t
∂2η
Z Z Z
%0s 2 · φ + σ 0s (η) : ∇ξ φ = ff · v + f0s · φ
Ω0s ∂t f
Ωt s
Ω0

∂η
+ coupling condition u ◦ Lt = ∂t on Γ0
Energy inequality

•Taking as test functions (v, q, φ) = (u, p, η̇) we can derive an Energy

inequality for hyperelastic materials

Fluid Structure Energy (kinetic + elastic)

%f %0 ∂η
EFS (t) ≡ ku(t)k2L2 (Ωf ) + s k (t)k2L2 (Ωs0 ) + W (η(t))
2 t 2 ∂t

Then, for an isolated system (no external forces)

Z T Z
EFS (T ) + 2µ D(u) : D(u) dΩ dt ≤ EFS (0)
0 Ωft

RT R
The term 2µ 0 Ωf D(u) : D(u) dΩ dt represents the energy dissipated
t
by the viscosity of the fluid
Energy Inequality

Key points in deriving an energy inequality:

I Perfect balance of work at the interface
Z Z
∂η
(σ f · nf ) · u = − (σ 0s (η) · ns0 ) ·
Γt Γ0 ∂t
I No kinetic flux through the interface
Z Z Z
∂u %f d %f
(time der.) %f ·u= |u|2 − |u2 |w · n
Ωft ∂t 2 dt Ωft 2 Γt
Z Z
%f
(convective term) %f div(u ⊗ u) · u = |u2 |u · n
Ωft 2 Γt

where w is the velocity at which the Rinterface moves.

Since w = u = η̇, the kinetic flux %2f Γt |u2 |(u − w) · n vanishes.
Recovering the fluid and structure subproblems
I Take as a test function (v, q, 0) (implies in particular that
v ∈ H10 (Ωft )). Fluid equations with Dirichlet boundary conditions
(R
% ∂u + div(u ⊗ u) · v + σ f : ∇v + div uq = Ωf f f · v

R
Ωf f ∂t
t t
∂η
u= ∂t ◦ (Lt )−1 (Dirichlet boundary cond.)

I Take as a test function (Ext(φ), 0, φ), where Ext(φ) is a suitable

extension of φ in the fluid domain:
Ext(φ) ∈ H1 (Ωft ), Ext(φ) ◦ Lt (ξ) = φ(ξ) on Γ0 . Then Structure
equation with Neumann boundary conditions

∂2η
Z Z
%0s 2 · φ + σ 0s (η) : ∇ξ φ = fs0 · φ + < Resf (u, p), Ext(φ) >
Ω0s ∂t s
Ω0

with Resf (u, p): residual of the fluid momentum equation

Z   
f ∂u
< Resf (u, p), v >= f − %f + div(u ⊗ u) · v − σ f : ∇v
Ωft ∂t
Observe that the term < Resf (u, p), Ext(φ) > represents the stress of
the fluid on the structure. Indeed:

< Resf (u, p), Ext(φ) >

Z » „ «–
∂u
= f f − %f + div(u ⊗ u) · Ext(φ) − σ f : ∇ Ext(φ)
Ωft ∂t
2 3
Z „ « Z
∂u
6 7
6 f
= 6f − %f + div(u ⊗ u) + div σ f 7 · Ext(φ) − (σ f nf ) · Ext(φ)
7
Ωft 4 ∂t 5 Γt
| {z }
=0
Z
=− (σ f nf ) · φ
Γt

Hence, step (2) corresponds to a structure problem with Neumann

1
boundary conditions at the interface: det(F) σ 0s FT ns = −σ f nf
Other splitting are also possible, leading e.g. to Fluid+Neumann b.cs /
Structure + Dirichlet b.cs.
Analysis of a linearized fluid structure model
Stokes + linear elasticity
We consider now a linearized fluid structure model around the rest state.
Then, the linearized Navier-Stokes equations reduce to the Stokes
problem and the equations for a non-linear hyperelastic isotropic solid
reduce to the linear elasticity equations.
Stokes
(
%f ∂u
∂t − div σ f (uf , pf ) = ff
f
in Ωf
div uf = 0
σ f (uf , pf ) = 2µf D(uf ) − pf I

Linear elasticity
∂ 2 ηs
%s − div σ s (η s ) = fs in Ωs
∂t 2
σ s (η s ) = 2µs D(η s ) + λs (div η s )I

Coupling conditions

uf = η̇ s , σ f (uf , pf ) · nf = −σ s (η s ) · ns on Γ
Boundary and initial conditions
s
ΓD
f
ΓD
nf
ns solid
fluid
Ωs
Ωf
Γ
s
ΓN
f
ΓN

Boundary conditions:


 uf = ġf on ΓD,f

σ · n = d
f f f on ΓN,f


 η s = g s on ΓD,s
σ s · ns = ds on ΓN,s


Initial conditions
(
uf |t=0 = uf ,0 in Ωf
η s |t=0 = η s,0 , η̇ s |t=0 = η̇ s,0 on Ωs
Linear elasticity in mixed form
It is actually more convenient to write a mixed formulation for the
elasticity equations as well, by introducing a pressure

ps = −λs div η s

which we can differentiate in time, thus obtaining

1 ∂ps
div η̇ s = −
λs ∂t
and the elasticity equation becomes

∂ 2 ηs
%s − div σ s (η s , ps ) = fs


∂t 2 in Ωs
1 ∂ps
+ div η̇ s = 0



λs ∂t
σ s (η s , ps ) = 2µs D(η s ) − ps I
Global weak formulation
Let Ωfs = Ωf ∪ Ωs . Similarly, ΓD = ΓD,f ∪ ΓD,s , and ΓN = ΓN,f ∪ ΓN,s .
Moreover we define the following functional spaces:
Vf = [H 1 (Ωf )]d , Vf ,0 = {v ∈ Vf , v = 0 on ΓD,f }
1 d
Vs = [H (Ωs )] , Vs,0 = {v ∈ Vs , v = 0 on ΓD,s }
1 d
Vfs = [H (Ωfs )] , Vfs,0 = {v ∈ Vfs , v = 0 on ΓD }
Qf = L2 (Ωf ), Qs = L2 (Ωs ), Qfs = L2 (Ωfs )
To derive a global weak formulation, we multiply the momentum
equations of the fluid and the solid by the same test function v ∈ Vfs and
integrate over the respective domains.
Stokes
Z Z Z Z
∂uf
%f · v + σ f : ∇v = ff · v + (σ f nf ) · v + df · v
Ωf ∂t Ωf Γ ΓN,f

Linear elasticity
∂ 2 ηs
Z Z Z Z
%s · v + σ s : ∇v = f s · v + ( σ s ns ) · v + ds · v
Ωs ∂t 2 Ωs Γ | {z } ΓN,s
=−σ f nf
Global weak formulation
Similarly, we multiply both the continuity equation of the fluid and the
equation defining the pressure for the solid by the same test function
q ∈ Qfs and sum them together.
global weak formulation
Find (uf (t), pf (t), η s (t), ps (t)) ∈ Vf × Qf × Vs × Qs , with uf = ġf on
ΓD,f , η s = gs on ΓD,s such that
Z  
∂uf
% · v + 2µ D(u ) : ∇v − p div v

f f f f

∂t



 Ωf Z  2

∂ ηs


· ∇v −


 + % s v + 2µ s D(η s ) : p s div v



 ZΩs ∂t 2 Z Z Z

= ff · v + fs · v + df · v + ds · v ∀v ∈ Vfs



Ωf Ωs ΓN,f ΓN,s
Z Z  
 ∂η s 1 ∂ps
∀q ∈ Qfs


 div uf q + div q+ q =0

 Ωf

 Ωs ∂t λs ∂t

uf = ∂ηs

on Γ



 ∂t

u (0) = u , η (0) = η , ∂ηs (0) = η̇

f f ,0 s s,0 ∂t s,0
Global weak formulation
We define now the following global functions:
(
uf in Ωf
velocity : u = ∂ηs
∂t in Ωs
(R t
0 f
u (s) ds + Ext(η s,0 ) in Ωf
displacement : η(t) =
η s (t) in Ωs
where Ext(η s,0 ) is a suitable extension of η s,0 in the fluid domain
(
pf in Ωf
pressure : p =
ps in Ωs
( (
uf ,0 in Ωf Ext(η s,0 ) in Ωf
initial conditions : u0 = η0 =
η̇ s,0 in Ωs η s,0 in Ωs
( ( (
ff in Ωf df in ΓN,f ġf in ΓD,f
forcing terms : f = d= ġ =
fs in Ωs ds in ΓN,s ġs in ΓD,s
Global weak formulation
... and the following bilinear forms: ∀v, w ∈ Vfs , q ∈ Qfs
Z Z
m(w, v) = %f w · v + %s w · v
Ωf Ωs
Z Z
af (w, v) = 2µf D(w) : ∇v, as (w, v) = 2µs D(w) : ∇v
Ωf Ωs
Z Z Z
b(v, q) = − div vq − div vq = − div vq
Ωf Ωs Ωfs

They satisfy the following properties:

I m(·, ·) is continuous and coercive in Qfs

m(v, v) ≥ Cm,1 kvk2L2 (Ωfs ) , m(w, v) ≤ Cm,2 kwk2L2 (Ωfs ) kvk2L2 (Ωfs )
I aj (w, v), with j = {f , s} are continuous and coercive in Vj,0 , thanks
to the Korn inequality
aj (v, v) ≥ Caj ,1 kvk2H 1 (Ωj ) , aj (w, v) ≤ Caj ,2 kwk2H 1 (Ωj ) kvk2H 1 (Ωj )
I b(v, q) satisfies the usual inf-sup condition between the spaces Vfs
and Qfs .
Global weak formulation

global weak formulation

For t > 0, find u(t), η(t), p(t), with u = ġ on ΓD , u(0) = u0 and
η(0) = η 0 , such that

∂u


 m( , v) + af (u, v) + as (η, v) + b(v, p)


 ∂t Z Z

= f ·v+ d·v ∀v ∈ Vfs,0




 Ωfs ΓN
Z
1 ∂p

 q − b(q, u) = 0 ∀q ∈ Qfs
λs Ωs ∂t





 ∂η = u



∂t
Finite element approximation (continuous in time)
We define finite element subspaces

Vhfs ⊂ Vfs fs
Vh,0 = Vhfs ∩ Vfs,0 Qhfs ⊂ Qfs

that satisfy the inf-sup condition.

Then, the finite element approximation simply reads
Finite element formulation
For t > 0, find (uh (t), η h (t), ph (t)) ∈ Vhfs × Vhfs × Qhfs , with uh = ġh on
ΓD , uh (0) = u0h and η h (0) = η 0h , such that
 ∂u
m( ∂t , vh ) + af (uh , vh ) + as (ηRh , vh ) + b(vRh , ph )
h

fs



 = Ωfs f · vh + ΓN d · vh ∀vh ∈ Vh,0
1
R ∂ph


 λ s
q − b(qh , uh ) = 0
Ωs ∂t h
∀qh ∈ Qhfs

 ∂ηh

∂t = uh

Observe that one should not expect the pressure to be continuous at the
interface. Therefore, it is better to work with a discontinuous pressure
space (or at least a space of pressures discontinuous at the interface).
Algebraic formulation
I let {φj }N fs
j=1 be a basis of Vh
u

N
I p
let {ψk }k=1 be a basis of Qhfs
Expansion of the solution:
X X X
uh (t) = uj (t)φj , η h (t) = ηj (t)φj , ph (t) = pk (t)ψk
j j k

Algebraic vectors and matrices:

U = [u1 , . . . , uNu ]T , Ξ = [η1 , . . . , ηNu ]T P = [p1 , . . . , pNp ]T
(Mfs )ij = m(φj , φi ), (Af ,s )ij = af ,s (φj , φi ), Bjl = b(φj , ψl )
algebraic system
 dU

 Mfs + Af U + As Ξ + B T P = Fu


 dt
 dΞ
−U=0

 dt
 1
 dP
Ms − BU = Fp


%s λ s dt
R
with (Ms )ij = Ωs
%s φj · φi
Algebraic formulation

algebraic system
BT
        
Mfs 0 0 U Af As U Fu
 0 d
I 0  Ξ +  −I 0 0  Ξ =  0 
0 0 %s1λs Ms dt P −B 0 0 P Fp

I Observe that the matrix Ms is singular (lives only in Ωs ). No

evolution equation is written for the pressure in Ωf .
I However, in Ωf the pressure is defined through the incompressibility
constraint in the velocity.
Well posedness analysis (sketch)
[Du-Gunzburger-Hou-Lee 2003, 2004]

We first prove that the discretized system admits a unique solution.

Observe that the matrix
 
Mfs 0 0
 0 I 0 
0 0 %s1λs Ms

multipying the time derivative is singular (Ms is singular), so the

concusion is not immediate.

I Define the space Wh = {vh ∈ Vfs , div vh qh = 0, ∀qh ∈ Qhfs }.

R
h Ωf
Clearly, Wh ⊂ Vnfs and uh ∈ Wh
I Let {φ̃j } be a basis of Wh and expand the solution uh on such basis:
P
uh (t) = j ũj (t)φ̃j .
Similarly, define the vector Ũ = [ũ1 , . . . , ũÑu ]T and the matrices M̃fs ,
Ãf , Ãs and B̃ using the basis functions φ̃j .
 I Split the basis of Qhfs as {ψk } = {ψks } ∪ {ψkf } where
{ψkf } = {ψk : supp(ψk ) ∩ Ωs = ∅}. Define then Qhf = span{ψkf }
and Qhs = span{ψks }.
I Then, the pressure vector P, the matrix B̃ and the matrix Ms can
also be split into
     
Pf B̃f 0 0
P= , B̃ = , Ms =
Ps B̃s 0 M̃s

accordingR to the splitRin the basis. Observe that

(B̃f )jl = Ωfs φ̃j ψlf = Ωf φ̃j ψlf = 0 since φ̃j ∈ Wh , so that
B̃ T P = B̃fT Pf + B̃sT Ps = B̃sT Ps .

The algebraic system becomes then:


d Ũ T
M̃fs dt + Ãf Ũ + Ãs Ξ̃ + B̃s Ps = F̃u

d Ξ̃
 dt − Ũ = 0
1 dPs
λs M̃s dt − B̃s Ũ = Fp

which now admits a unique solution (the matrix multiplying the time
derivative vectors is now non singular). From this, we infer the existence
and uniqueness of uh , η h and ph |Ωs .
I It remains to show that there exists a unique pressure ph |Ωf also in
the fluid domain. We go back to the original basis {φj } of Vhfs . The
matrix B can be still split according to the split in the basis of Qh in
B T = [BfT |BsT ] where now BfT is not zero anymore.
I We show that the bilinear form b(·, ·) satisfies an inf-sup condition
between the spaces Qhf and Vhfs (rememer that b satisfies already the
inf-sup cond, between Qh and Vhfs ). Indeed,
b(vh , qh ) b(vh , qh )
inf sup ≥ inf sup ≥β>0
qh ∈Qhf v ∈Vfs
h
kvh kH 1 (Ωfs ) kqh kL2 (Ωf ) qh ∈Qh
vh ∈V fs kvh k H 1 (Ω ) kqh kL2 (Ω )
fs fs
h h

It follows that Ker (BfT ) = ∅.

I Multiply the first equation of the system by Bf . Then
 
T dU T
Bf Bf Pf = Bf Fu − Mfs − Af U − As Ξ − Bs Ps
dt

which provides a unique Pf since Bf BfT is non singular. This give

the existence and uniqueness of ph |Ωf as well.
Well posedness analysis – a priori estimates
We consider the simpler case f = d = g = 0.
Energy estimate. We take vh = uh and qh = ph in the finite element
formulation and remember that ∂η ∂t = uh .
h

 
I m( ∂uh , uh ) = 1 d % f kuh k2
2 + % s kuh k 2
2
∂t 2 dt L (Ωf ) L (Ωs )

I af (uh , uh ) = 2µf kD(uh )k2L2 (Ωf )

I as (η h , uh ) = as (η h , ∂η d 2
∂t ) = µs dt kD(η h )kL2 (Ωs )
h

R ∂ph
I b(uh , ph ) = 1 1 d 2
λs Ωs ∂t ph = 2λs dt kph kL2 (Ωs )
from which we deduce that
 
%f kuh (t)k2L2 (Ωf ) + %s kuh (t)k2L2 (Ωs ) + 2µs kD(η h (t))k2L2 (Ωs )
Z t
1
+ kph (t)k2L2 (Ωs ) + 4µf kD(uh (s))k2L2 (Ωf ) ds = C (u0h , η 0h )
λs 0

where C (u0h , η 0h ) is a constant depending only on the initial data, but

not on h.
Well posedness analysis – a priori estimates
Estimate on ∂t uh . We differentiate the equations in time:
2
(
∂ph
m( ∂∂tu2h , vh ) + af ( ∂u
∂t , vh ) + as (uh , vh ) + b(vh , ∂t ) = 0
h

2
1 ∂ p ∂uh
R
λs Ωs ∂t 2 qh − b(qh , ∂t ) = 0
h

Taking vh = ∂f uh and qh = ∂t ph and proceeding as before, we get a

bound for k∂t uh (t)kL2 (Ωfs ) ≤ C , provided the data of the problem are
sufficinetly smooth.

Estimate for the pressure ph |Ωf . It remains to provide a bound for the
pressure ph |Ωf . We provide actually a bound for ph |Ωfs . From the inf-sup
condition we have
1 b(vh , ph (t))
kph (t)kL2 (Ωfs ) ≤ sup
β vh ∈Vfs kvh kH 1 (Ωfs )
h

1 −m( ∂u
∂t
h
(t), vh ) − af (uh (t), vh ) − as (η h (t), vh )
= sup
β vh ∈Vfs kvh kH 1 (Ωfs )
h
` ´
≤ C k∂t uh (t)kL2 (Ωfs ) + kuh (t)kH 1 (Ωf ) + kη h (t)kH 1 (Ωs )
Well posedness analysis – passage to the limit

I From the previous a prioir estimates conclude that

∗
I uh * u in L∞ (0, T ; L2 (Ωfs ))
I uh |Ωf * u in L2 (0, T ; H 1 (Ωf ))
∗
I η h * u in L∞ (0, T ; H 1 (Ωs ))
I ph * p in L2 (0, T ; L2 (Ωfs ))
∗
I ph |Ωs * p in L∞ (0, T ; L2 (Ωs ))

I These convergences are enough to pass to the limit in the equations.

It follows then that the limit (u, η, p) is a solution of the continuous
problem.
I From the a priori bounds one deduces also the uniqueness of the
solution.
I Finally, by similar arguments the optimality of the finite element
approximation can be established.
Temporal discretization – Mid Point

We discretize in time with the mid-point algorithm, i.e. all equations are
colocated at t n+1/2 and the time derivatives are discretized with a
centered finite difference.
Mip-point discretization

un+1 − unh un+1 + unh η n+1 + η nh


h h h
m( , v ) + a ( , v ) + a ( , vh )

h f h s

∆t 2 2




p n+1 + phn

 Z Z
+b(vh , h f n+1/2 · vh + dn+1/2 · vh fs
∀vh ∈ Vh,0


 )=

 2 Ωfs ΓN

phn+1 − phn un+1 + unh

Z
1
qh − b(qh , h

)=0 ∀qh ∈ Qhfs


λ ∆t 2



 s Ωs

 n+1 n n+1 n
 h − η h = uh + uh
η



∆t 2
Temporal discretization – Mid Point
In the previous scheme we can eliminate the third equation and solve the
n+ 1 un+1 +unh n+ 1 p n+1 +p n
system in the variables uh 2 = 2 and ph 2 = h 2 h .

n+ 1 n+ 1 n+ 1 n+ 1

2

 ∆t m(uh 2 , vh ) + af (uh 2 , vh ) + ∆t2 as (uh 2 , vh ) + b(vh , ph 2 )


1 1 2
= Ωfs f n+ 2 · vh + ΓN dn+ 2 · vh + ∆t m(unh , vh ) − as (η nh , vh )
R R


n+ 12 n+ 1

 2 R
qh − b(qh , uh 2 ) = λs2∆t Ωs phn qh
R
λs ∆t Ωs ph

n+ 21 n+ 12 n+ 12
un+1 = 2uh − unh , p n+1 = 2ph − phn , η n+1
h = ∆tuh + η nh

Algebraic system
 
2 ∆t
# " n+ 1 # n+ 1
BT
"
∆t Mfs + Af + 2 As U 2 Fu 2 + 2
∆t Mfs U −
n
As Ξn
2 1
=  
−B λs %s ∆t Ms Pn+ 2 2
λs %s ∆t Ms P
n
Stability analysis
Consider the homogeneous case (f = d = g = 0) for simplicity. Take
un+1 +un p n+1 −p n
vh = h 2 h and qh = h 2 h . and sum the first two equations.
Then
un+1 +unh phn+1 −phn
I The term b( h , 2 ) cancels out.
2
un+1 −u n
u n+1
+u n
1
m(un+1 n+1
 n n

h , uh ) − m(uh , uh )
I m( h h
, h 2 h ) = 2∆t
∆t
η n+1 n n+1
h +η h uh +uh
n
η n+1 +η n η n+1 −η n
I as ( 2 , 2 ) = as ( h 2 h , h ∆t h ) =
1 n+1 n+1
 n n

2∆t as (η h , η h ) − as (η h , η h )
R phn+1 −phn phn+1 +phn h i
I 1
λs Ωs ∆t 2 = 2λs1∆t kphn+1 k2L2 (Ωs ) − kphn k2L2 (Ωs )

Therefore
 
1 1 n+1 2
m(un+1
h , un+1
h ) + as (η n+1
h , η n+1
h ) + kp k L2 (Ωs )
2∆t λs h
un+1
h + unh un+1 + unh
+ af ( , h )
 2 2 
1 1
= m(unh , unh ) + as (η nh , η nh ) + kphn k2L2 (Ωs )
2∆t λs
Stability analysis

From which an unconditional stability can be established

1 n 2
m(unh , unh ) + as (η nh , η nh ) +
kp k 2
λs h L (Ωs )
n
X ui + ui−1 ui + ui−1
+ 2∆t af ( h h
, h h
)
2 2
i=1
= m(uh,0 , uh,0 ) + as (η h,0 , η h,0 ) + λs k div η h,0 k2L2 (Ωs )

I A better estimate for the pressure can be recovered from the inf-sup
condition.
I It can also be proved that the scheme is second order accurate in
time.
Some References

[DGeal04] Q. Du, M.D. Gunzburger, L.S. Hou, J. Lee. Semidiscrete finite

element approximations of a linear fluid-structure interaction problem. SIAM J.
Numer. Anal. 42(1):1–29, 2004.
[DGeal03] Q. Du, M.D. Gunzburger, L.S. Hou, J. Lee. Analysis of a linear
fluid-structure interaction problem. Discrete Contin. Dyn. Syst. 9(3)633–650,
2003.
[LeTM01] P. Le Tallec and J. Mouro. Fluid structure interaction with large
structural displacements. Comput. Methods Appl. Mech. Engrg.,
190:3039–3067, 2001.
[FQV09] Cardiovascular Mathematics. Modeling and simulation of the
circulatory system (Chapter 9). L. Formaggia, A. Quarteroni, A. Veneziani
(Eds.) Springer, 2009.
[N01] F. Nobile. Numerical approximation of fluid-structure interaction
problems with application to haemodynamics, PhD Thesis n.2458, École
Polytechnique Fédérale de Lausanne, 2001.
 Linearized fluid models:
the added mass technique
Linearized fluid model
Consider the Navier-Stokes equations for an inviscid incompressible fluid
(Euler equations):
 ∂u
%f ∂t + %f u · ∇u + ∇p = 0



div u = 0


+ b.cs and i.cs

Assume small perturbations around the rest state (ū, p̄) = (0, 0). Then
u · ∇u is negligible
Linearized inviscid fluid model

% ∂u + ∇p = 0,

f
∂t in Ωf , t > 0
div u = 0


u · n = g on ΓD , t > 0, p = d on ΓN , t > 0
f
u|t=0 = u0 , in Ω
Equation for the pressure

I We take the divergence of the momentum equation and use the

incompressibility constraint:
 
∂u
div %f + ∇p = 0 =⇒ ∆p = 0, in Ωf
∂t
I On ΓD take the momentum equation in the normal direction:
 
∂u ∂p ∂g
n · %f + ∇p = 0 =⇒ = −%f , on ΓD
∂t ∂n ∂t

Pressure equation

∆p = 0
 in Ωf
∂p
 ∂n = −%f ġ , on ΓD
p = d, on ΓN

Alternative formulation: potential of fluid velocity
I Assume that the fluid velocity is initially irrotational, i.e. curl u0 = 0.
Then the velocity field u(t) will be irrotational at all time.
I For a simply connected domain Ωf , we can introduce (up to a
constant) a potential φ : Ωf × R+ → R such that u(x, t) = ∇φ(x, t)
I Then, the momentum equation can be written as
∂∇φ ∂φ 1
%f + ∇p = 0, =⇒ = − p + c(t)
∂t ∂t %f
that is, the fluid potential is a primitive of the pressure (up to an
arbitrary constant which can be set to zero)

Equation for the potential


∆φ = 0
 in Ωf
∂φ
∂n = g, on ΓD
Rt
1

φ = − %f D(t), on ΓN , with D(t) = d(τ ) dτ.

0

I In what follows, we will use the pressure equation only, although the
potential equation is probably more common.
A toy fluid-structure problem
Piston interacting with an inviscid incompressible fluid
y

η
(−L,H) Γwall u.n=0 (0,H)
Γin Piston
Γp
p=g(t) Fluid .
u.n= η
(−L,0) Γwall u.n=0 (0,0) x

Simplifying assumptions
I Small displacements: the fluid domain is considered fixed.
I Fluid equations linearized around the rest state.




 ρf ∂t u + ∇p = 0
div u = 0


 Z
Fluid: u · n = 0, on Γwall Piston: m η̈ + c η̇ + kη = pdσ
 Γp
p = g (t), on Γin





u · n = η̇, on Γ
p
 

 ∆p = 0 in Ω,

∂ p = 0 on Γ
n wall
Pressure equation:


 p = g (t) on Γ in
∂n p = −ρf η̈ on Γp


Solution: p(x, y , t) = g (t) − %f (x + L)η̈

R
Force acting on the piston: F = Γp p dσ = g (t)H − %f LH η̈.

Replacing in the Piston equation, we get the effective piston dynamics

(m + %f LH ) η̈ + c η̇ + kη = g (t)H
| {z }
added mass

Remarks:
I The effective mass of the piston includes the mass of the fluid
(ma = %f LH) that has to be displaced.
I The presence of the fluid alters (descreases) the natural frequencies
of oscillation of the elastically supported piston
r r
k k
freq. of oscillation: in air: ω = , in the fluid: ω̃ =
m m + ma
More complex structure models

u.n=0

p=0
Σ

p=0
Ωf
Ωs
nf ns
u.n=0

Fluid linearized around the rest state

(ū, p̄) = (0, 0).

(p̄ = −%f gz if gravity is considered. See later)

More complex structure models

I Structure model: linear elasticity (small perturbations around

equilibrium state η = 0)
 2
 %s ∂ η2 − div σ s (η) = 0, in Ωs
 ∂t


σ s (η) · ns = −pns , on Σ



+ other b.cs and i.cs


with σ s (η) = λ div(η)I + 2µD(η)

I Fluid model: inviscid incompressible flow linearized around rest state


 ∆p = 0, in Ωf ,

∂ p = 0,
n on ΓD


 p = 0, on ΓN
∂n p = −ρf η̈ · nf , on Σ

The added mass operator
We introduce first the Neumann-to-Dirichlet operator
Sf : H −1/2 (Σ) → H 1/2 (Σ):

∆pλ = 0,


in Ωf ,
∂ p = 0, on ΓD
n λ
Sf λ = pλ |Σ , where pλ solves
pλ = 0,
 on ΓN

∂n pλ = λ, on Σ


Added Mass operator: MA = %f Sf

Then the traction on the structure interface is

σ s · ns = −p|Σ ns = −Sf (−%f η̈ · nf )ns = −MA (η̈ · ns )ns

Effective structure equation (weak form)

Z Z Z
%s η̈ · φ + σ s (η) : ∇φ = − MA (η̈ · ns )φ · ns
Ωs Ωs
|Σ {z }
fluid added mass
Properties of the added mass operator
I MA is symmetric: < MA λ, µ >=< λ, MA µ >

Z Z Z * 0Z

< MA λ, µ >= %f pλ µ = %f pλ ∂n pµ = − %
f pλ ∆pµ +
 %f ∇pλ ·∇pµ
Σ Σ
Ωf Ωf

Z
:

 0Z
= − %f ∆p pm u +
λ %f ∂n pλ pµ =< λ, MA µ >
Ωf Σ

I MA is positive:
Z
< MA λ, λ >= %f ∇pλ · ∇pλ = %f k∇pλ k2L2 (Ωf ) > 0, ∀pλ 6= const.
Ωf

I MA : L2 (Σ) → L2 (Σ) is compact:

c
∀λ ∈ L2 (Σ), pλ ∈ H 1 (Ωf ), pλ |Σ ∈ H 1/2 (Σ) ,→ L2 (Σ)

Therefore, MA has a sequence of positive eigenvalues

n→∞
λ1 > λ2 > . . . > λn > . . . > 0, with λ1 < +∞ and λn −→ 0 and a
2
corresponding sequence of L -orthogonal eigenfunctions.
Algebraic formulation of the fluid structure problem

I Let Vh,s and Vh,f be finite element spaces for the structure
displacement and fluid pressure, respectively (need not be
conforming at the common interface Σ)
I {φj } and {ψl }: basis fuctions for Vh,s and Vh,f , respectively.
Z Z Z
Structure equation: %s η̈ · φ + σ s (η) : ∇φ = − pφ · ns
Ωs Ωs Σ
 R
(Ms )ij = R Ωs %s φj φi

=⇒ Ms Ü + Ks U = −G P with (Ks )ij = Ωs σ s (φj ) : ∇φi
 R
Gil = Σ ψl φi · ns

Z Z  Z 
Fluid equation: ∇p · ∇ψ = − %f (η̈ · nf )ψ = %f (η̈ · ns )ψ
Ωf Σ Σ
Z
=⇒ Kf P = %f G T Ü with (Kf )ij = ∇ψj · ∇ψi
Ωf
Algebraic formulation of the fluid structure problem
The coupled fluid-structure problem at algebraic level reads:
       
Ms 0 Ü Ks G U 0
+ =
−%f G T 0 P̈ 0 Kf P 0

I Observe that the matrix multiplying the second order derivatives is

singular.
I Both block matrices are non-symmetric. However, it is possible to
reformulate the problem using the potential of the fluid displacement
instead of the pressure to obtain a symmetric system.

Elimination of the pressure

I From the second equation we have: P = %f Kf−1 G Ü.
I Replacing in the first equation we obtain:

Ms Ü + Ks U = −%f GKf−1 G T Ü
| {z }
MA
Added mass matrix

Effective structure dynamics: (Ms + MA )Ü + Ks U = 0

I The matrix MA = %f GKf−1 G T is called the added mass matrix. It is

the algebraic counterpart of the added mass operator MA .
I It can be shown that MA is symmetric and positive definite.
I The practical computation of the added mass matrix is quite
expensive as it implies the computation of Kf−1 . Other
reformulations of the pressure Poisson problem as boundary integral
problems could be more convenient.
Frequency analysis

I The previous model can also be used to compute the natural

frequencies of oscillation of the fluid-structure system.
I Assume U(t) = U cos(ωt) and P(t) = P cos(ωt). Then, the
(approximated) natural frequencies of oscillation are the generalized
eigenvalues of the algebraic system
     
Ks G U 2 Ms 0 U
=ω
0 Kf P −%f G T 0 P

This eigenvalue problem is non-symmetric, hence difficult to solve.

I Using the added mass matrix MA , we can solve, instead, the
symmetric eigenvalue problem

Ks U = ω 2 (Ms + MA )U
Modal reduction
Assume that the structure displacement can be described as a linear
combination of few modes only
m
X
η(x, y ) = αj (t)η j (x)
j=1

This can be the case for instance for

I Rigid body: in this case the motion is described by 6 dofs in three
dimensions (m = 6)

η 1 = e1 η 2 = e2 η 3 = e3
η 4 = y e1 − xe2 η 5 = ze1 − xe3 η 6 = ze2 − y e3

I Analysis of the first oscillation modes of a deformable structure;

(η j , λj ) are the first few eigenmodes/eigenvalues satisfying
Z Z
σ s (η j ) : ∇φ = λj %s η j · φ
Ωs Ωs
Start withP
weak formulation of the structure equation with
m
η(x, t) = j=1 αj (t)η j (x) and φ = η i , i = 1, . . . , m:

m
X Z m
X Z Z
α̈j (t) %s η j · η i + αj (t) σ s (η j ) : ∇η i = − pη i · ns
Ωs Ω
j=1 | {z } j=1 | s {z } | Σ {z }
(M
fs )ij (K
fs )ij Fei

To compute the right hand side, we solve m Poisson problems


∆pl = 0, in Ωf



∂ p = 0, and compute
n l on ΓD Z


 pl = 0, on Γ N (Mf )
A il = pl η i · ns
∂n pl = −%f η l · nf , on Σ Σ


Pm Pm f
Then p = l=1 α̈l pl and the r.h.s. becomes: Fei = l=1 (M A )il α̈l

Reduced structure model (M

fs + M
fA )α̈ + K
fs α = 0
Gravity effects – free surface problems
We consider now a fluid subject to gravity and with a free surface.
Γa’
ηz z (
z=0
Γa ū = 0
rest state:
g p̄ = −%f gz + pa
pa = atmospheric pressure
z=−H

Boundary conditions on the deformed (unknown) free surface Γ0a :

p = pa

Perturbation analysis: u = ū + δu, p = p̄ + δp

Linearization of the momentum equation:
∂u ∂δu
%f + u · ∇u + ∇p = −%f g ez =⇒ %f + ∇δp = 0
∂t ∂t
Linearization of the free surface condition on the reference surface Γa :
∀x ∈ Γa :

∂p ∂p̄ ∂δp  : h.o.t

p(x) = p(x + ηz (x)) − (x)ηz (x) = pa − (x)ηz (x) −  (x)ηz (x)
∂z ∂z ∂z
= pa + %f g ηz (x) = p̄(x) + %f g ηz (x)

=⇒ δp(x) = %f g ηz (x)
On the other hand, from the linearized momentum equation

∂δp ∂uz ∂ 2 ηz 1 ∂ 2 δp
(x) = −%f = −%f (x) = − (x)
∂z ∂t ∂t 2 g ∂t 2

Linearized inviscid fluid model with gravity and free surface



 ∆δp = 0, in Ωf


δp = 0,

on ΓD


 ∂n δp = 0, on ΓN

∂ δp = − 1 ∂ 2 δp ,

on Γa
z g ∂t 2
The linearized problem for the pressure has a non trivial dependence on
time. The added mass operator can not be straightforwardly derived in
this context.  
iωt
Frequency analysis: δp(x, t) = < δp(x)e
c

Pressure equation – frequecy domain


∆δp = 0, in Ωf

 c
2
ω

c = δp,
∂z δp g
c on Γa

+ other b.cs

Equivalently, this problem can be stated using the potential of the fluid
velocity δu = ∇φ. Linearized momentum equation (frequency domain):
∂δu  
%f + ∇δp = 0 =⇒ ∇ iω%f φb + δp
c =0
∂t
Potential equation – frequecy domain

∆φ = 0, in Ωf

 b
ω2 b c = −iω%f φb + const
δp
 ∂z φ = g φ, on Γa
b

+ other b.cs
Fluid-structure with gravity and free surface
Γa p=pa

I Fluid linearized around the rest

p=−ρ gz

p=−ρ gz
Σ Ωf state (ū, p̄) = (0, −%f gz)
f

Ωs

f
u.n=0
nf ns I Frequency analysis

For easiness of presentation we write (η, p) instead of (b

η , δp).
c

Fluid-structure problem


 %s ω 2 η + div σ s (η) = 0, in Ωs

σ s (η) · ns = −(p̄ + p)ns ,

 on Σ


∆p = 0, in Ωf

 ∂n p = −%f ω 2 η · ns , on Σ
ω2

∂z p = g p, on Γa




+ other b.cs

Global weak formulation

Z Z Z
2
−ω %s η · φ + σ s (η) : ∇φ + pφ · ns
Ωs Ωs Σ
ω2
Z Z Z Z
∇p · ∇ψ + ω 2 %f ψη · ns − pψ = − p̄φ · ns
Ωf Σ g Γa Σ

Algebraic formulation
       
Ks G U 2 Ms 0 U G P̄
=ω −
0 Kf P −%f G T S P 0
with Sij = Γa g1 ψj ψi
R
Added mass matrix for free surface flows
We eliminate P from the second equation

P = −ω 2 %f (Kf − ω 2 S)−1 G T U

Problem for U only:

Ks U = ω 2 (Ms + MA (ω))U , with MA (ω) = %f G (Kf − ω 2 S)−1 G T

Remarks
I The added mass depends on the angular frequency ω! This
technique is not very suited for time dependent problems
I The matrix (Kf − ω 2 S) might be singular! This will correspond to
resonance modes of the superficial waves
A better formulation consists in introducing the variable µ = ∂p 

I
∂z Γa
(proportional to the vertical displacement of the fluid at the free
surface) and solving the coupled problem in (η, µ).
Some References

[MO95] H. J-P. Morand, R. Ohayon, Fluid structure interaction. Wiley & Sons,
1995.
[N77] J.N. Newman. Marine Hydrodynamics. The MIT Press, 1977.
[M83] C.C. Mei. The Applied Dynamics of Ocean Surface Waves. John Wiley
& Sons Inc, 1983.
[M10] A. Mola. Multi-physics and Multilevel Fidelity Modeling and Analysis of
Olympic Rowing Boat Dynamics. PhD Thesis, o Virginia Polytechnic Institute
and State University, 2010.
[FMMM09] L. Formaggia, E. Miglio, A. Mola, A. Montano. A model for the
dynamics of rowing boats. Int. J. Numer. Methods Fluids 61(2):119–143,
2009.

[FMMP08] L. Formaggia, E. Miglio, A. Mola, N. Parolini. Fluid-structure

interaction problems in free surface flows: application to boat dynamics. Int. J.
Numer. Methods Fluids. 56(8)965–978, 2008.
ALE formulation for fluids in moving domains
How to deal with moving domains

The fluid equations are defined on a moving domain. How can we treat
them numerically?
Idea Use a moving mesh, that follows the deformation of the domain.

T hs Γt,h
w

f
T h,0 T f
t,h

At

The mesh follows the moving boundary and is fixed on the artificial
(lateral) sections.
Eulerian / Lagrangian / ALE frame of reference
ΓLw
t Γw
t

ΓL
in
Ω ft ΩL t
out
ΓL Γ
in Ω ft ΩA t Γ
t
out
t t t

ΓL t
w
Γw t

(Lagrangian) (ALE)
Lt At

Γ0
w

Γ0 Ω0 Γ0
in f out
u u

Γ0w

Neither an Eulerian nor a Lagrangian frame are suited to our case. We

want to follow the interface in a Lagrangian fashion, while keeping an
“Eulerian-type” description of the remaining part of the boundary.
A possible answer: Arbitrary Lagrangian Eulerian (ALE) frame:
Introduce a (fixed) frame of reference which is mapped at every time to
the desired physical domain. The equations of motion are then recast in
such reference frame.
ALE formulation (assuming known the domain
deformation)
The moving domain is recast at each time t to a fixed configuration Ωf0
through the ALE mapping At :

At x=x (t, ξ )
Ω0

               

               

               

At : Ωf0 −→ Ωft ,
               

               

               

               

ξ
               

                

Ωt
               

x(ξ, t) = At (ξ)
∂At
I domain velocity w(x, t) = ◦ A−1
t (x)
 ∂t

∂u 
 ∂u 

I ALE derivative = + w · ∇x u
∂t ξ ∂t x

∂JAt 
I Euler expansion = JAt div w, JAt = det(∇At )
∂t ξ
I ALE Transport formula
Z Z "  #
d ∂u 
u(x, t) = + u div w , ∀Vt ⊂ Ωft
dt Vt Vt ∂t ξ
Navier Stokes equations in ALE form

 
%f ∂u  + %f ((u − w) · ∇)u − div σf (u, p) = f
 
∂t ξ in Ωt

div u = 0

Hybrid formulation: the spatial derivative terms are computed on the

deformed configuration Ωft , whereas the time derivative is computed on
the reference configuration Ωf0 .

Using the Euler formula, the momentum equation can be equivalently

written in conservative form

%f ∂JAt u 
+ %f div((u − w) ⊗ u) − div σf (u, p) = f
JAt ∂t ξ
NS-ALE weak formulation
Consider a test function v̂ ∈ [H01 (Ωf0 )]d defined on the reference
configuration, and let us “map” it onto the deformed configuration:
v = v̂ ◦ (At )−1 .
We multiply the momentum eq. by v and integrate by parts
Non conservative formulation

Z  ! Z
∂u 
%f + (u − w) · ∇u · v + σ f : ∇v + div uq = f ·v
Ωft ∂t ξ Ωft

Again, we can write an alternative conservative formulation by using the

ALE transport formula
Conservative formulation
Z Z Z
d
%f u · v + %f div((u − w) ⊗ u) · v + σ f : ∇v + div uq = f ·v
dt Ωft Ωft Ωft
Finite Element ALE formulation
We introduce a triangulation Th,0 of the reference domain Ωf0 and choose
finite element spaces (Vh,0 , Qh,0 ) on Th,0 for velocity and pressure resp.
T hs Γt,h
w

f f
T h,0 T t,h

At

At time t, let Th,t be the image of Th,0 through the mapping At .

Similarly, we define finite element spaces (Vh,t , Qh,t ) on Th,t as the map
of (Vh,0 , Qh,0 ) through At .
Remark. Assume At is a piecewise linear function on Th,0 . Then, Th,t is
a triangulation of Ωft .
Moreover, if Vh,0 (resp Qh,0 ) is the space of piecewise polynomials of
degree p on Th,0 , then Vh,t (resp Qh,t ) is the space of piecewise
polynomials of degree p on Th,t
Finite Element ALE formulation – non conservative
find (uh , ph ) ∈ (Vh,t , Qh,t ), for each t > 0, such that

Z  ! Z
∂uh 
%f + (uh − w) · ∇uh · vh + σ f : ∇vh + div uh qh = f · vh
Ωft ∂t ξ Ωft

for all (vh , qh ) ∈ (Vh,t , Qh,t )

Basis of Vh,0 : {φ̂i (ξ)}N −1

i=1 ⇒ Basis of Vh,t : {φi (t, x) = φ̂i (ξ) ◦ (At ) }
u

Finite element solution (velocity):

Nf  Nf
X ∂uh  X dui (t)
u(t, x) = ui (t)φi (t, x), = φi (t, x)
∂t ξ  dt
i=1 i=1
Remark. ui are nodal values associated to moving nodes!
Algebraic formulation of the momentum equation
d
M(t) U(t) + A(t)U(t) + N(t, U − W)U(t) +B T (t)P(t) = F(t)
| dt
{z } | {z } | {z }
transport
stiffness
ALE time der.
Finite Element ALE formulation – conservative
find (uh , ph ) ∈ (Vh,t , Qh,t ), for each t > 0, such that

Z Z Z
d
%f uh · vh + %f div((uh −wh )⊗uh )·vh +σ f : ∇vh +div uh qh = f ·vh
dt Ωft Ωft Ωft

for all (vh , qh ) ∈ (Vh,t , Qh,t )

Algebraic formulation of the momentum equation

d
(M(t)U(t)) + A(t)U(t) + Nc (t, U − W)U(t) +B T (t)P(t) = F(t)
dt
| {z } | {z } | {z }
transport
stiffness
ALE time der.
R

with Nc (U)ij = Ω
div (uh − wh ) ⊗ φj · φi
I The two formulations (conservative and non conservative) are
equivalent at the time-continuous level. But they lead to different
time discretization schemes.
I Observe that all matrices depend on time!
ALE time discretization

The previous system of ODEs can be discretized by any ODE

discretization method. Let us consider an Implicit Euler with
semi-implicit treatment of the convective term

Non Conservative
Un − Un−1
Mn + An Un + N n (Un−1 − Wn )Un + (B n )T Pn = Fn
∆t

Conservative
M n Un − M n−1 Un−1
+ An Un + Ncn (Un−1 − Wn )Un + (B n )T Pn = Fn
∆t

Warning. All matrices change in time (because of the domain movement)

and have to be re-evaluated at each time step.
How to construct the ALE mapping in practice
T hs Γt,h
w

f
T h,0 T f
t,h

At

Assume the displacement η n of the boundary is known at time t n . We

can solve for instance a Laplace equation for the displacement of the
internal nodes
(
∆xn = 0, in Ωf0
x = ξ + η , on ∂Ωf0
n n

PNf
In practice, we solve it by the finite element method: xn = n
i=1 xi φi ,
and Xn = [x1n , . . . , xNn f ]T satisfying

Km Xn = Gn

Remark. the computed mapping xn (ξ) = At n (ξ) will be a piecewise

polynomial (for instance piecewise linear)
Given the ALE mapping at discrete times At n−1 , At n , At n+1 , ....
I Define a time continuous mapping At by piecewise polynomial
interpolation in time. E.g. linear interpolation

tn − t t − t n−1
At = At n−1 + At n , t ∈ [t n−1 , t n ]
∆t ∆t
I Compute the domain velocity w = ∂A ∂t .
t

E.g. linear interpolation w = (At n − At n−1 )/∆t

Remark on the Laplace problem

I simple to implement
I leads often to ‘”good quality” deformed meshes
I However, it does not guarantee that the mapping At n is invertible!

A more robust approach solves the elasticity equations (linear or

non-linear) to compute the mapping (the domain is thought as a
deforming elastic body - see e.g. [T.Tezduyar et al. ]).
Geometric Conservation Laws
The geometric conservation laws (GCL) have been advocated by
researchers working in fluid-structure problems using finite volume
methods: a constant solution of the original problem must be reproduced
exactly by the numerical scheme.
(see e.g. [Thomas-Lombard, AIAA ’79], [Lesoinne-Farhat, CMAME ’96],
[Formaggia-N., CMAME ’04])

The non conservative formulation always satisfies the GCL (according to

the definition above)

However, if we put uh = 1 in the conservative ALE-FE formulation of the

momentum equation we have
Z Z
d
vh − vh div wh = 0, ∀vh ∈ Vh,t
dt Ωft Ωft

This is always true at continuous level.

But what happens when we discretise in time?
GCL satisfying time advancing scheme

If we use a one-step scheme

Z Z Z
n+1 
vh = INTttn −

vh − vh div wh
Ωfn+1 Ωfn Ωft

A scheme satisfies the (discrete) GCL if the previous relation holds for
the numerical time integration rule used to integrate the ALE transport
n+1
term, here indicated by INTttn .
GCL satisfying time advancing scheme

Let At be a polynomial of degree s in time on each time slab (t n , t n+1 ).

Then, the GCL is fulfilled if the numerical time integration formula used
to compute the ALE transport term has a degree of exactness
r ≥ sd − 1, where d is the space dimension.

d s r Possible formula
2 1 1 Mid-point / Trapezoidal
2 2 2 Two-point Gauss
3 1 2 Two-point Gauss
3 2 3 Two-point Gauss

Remark: Only the integral containing the ALE transport term has to be
treated in this special way.
A conservative GCL-satisfying scheme in 2D
Let us consider
I 2D Conservative ALE formulation

I Implicit Euler discretization (one step scheme)

I a linear in time ALE mapping (constant velocity in [t n−1 , t n ])

According to the above criterion, to satisfy the GCL we have to integrate

the ALE transport term with a mid-point rule → integrate on the
intermediate configuration Ωfn+1/2 .
We integrate all the terms on the intermediate configuration
Z Z Z
1 1 n− 21
%f unh ·vh − %f un−1
h ·vh + div((uhn−1 −wh )⊗unh )·vh
∆t Ωfn ∆t Ωfn−1 Ωf
n− 1
2
Z Z
1
+ σ nf : ∇vh + div unh qh = f n− 2 · vh
Ωf 1 Ωf 1
n− n−
2 2
In matrix form
M n Un − M n−1 Un−1 1 n− 1 1 1 1
+An− 2 Un +Nc 2 (Un−1 −Wn− 2 )Un +(B n− 2 )T Pn = Fn− 2
∆t
A stability result R[Formaggia-N. EWJNM ’99]
Notation: kuk2n = Ωfn
u2
The previous scheme satisfies the following unconditional stability result
in the case of homogeneous Dirichlet conditions on all the boundary

n n
1
X X
%f kunh k2n + 2µ∆t kD(uih )k2i− 1 ≤ %f kuh0 k20 + C ∆t kf i− 2 k2i− 1
2 2
i=1 i=1

Conversely, the non-conservative formulation is only conditionally stable

with a limitation on ∆t that depends on the mesh velocity.
Some comments on the GCL
I If conservation properties are important, then GCL should be used.
In particular, they should be used in high speed compressible fluid
dynamics
I The result above shows that GCL also enhance stability. However, in
many problems with low speed incompressible fluid dynamics, also
non-satisfying GCL schemes seem to work well.
I The result above is one of the VERY FEW unconditional stability
results available for ALE schemes.
A stability result [Formaggia-N. EWJNM ’99]
Sketch of the Proof The critical terms are the time derivative and the ALE
transport terms. Take vh = unh :

Z Z
%f %f n− 12
kunh k2n − un−1
h · unh + %f div((un−1
h − wh ) ⊗ unh ) · unh
∆t ∆t Ωfn−1 Ωf 1
n−
2
| {z }
%f A
Z
1
+ 2µkD(un )k2n− 1 = f n− 2 · unh
2
Ωf
n− 1
2

Z
1 1 1 1 n− 21
A≥ kunh k2n − kunh k2n−1 − kun−1 k2n−1 − div(wh )(unh )2
∆t 2∆t 2∆t h 2 Ωf 1
n−
2
1 1
= kunh k2n − kun−1 k2n−1
2∆t 2∆t h
from which the result follows.
The last equality holds thanks to the GCL assumption:
Z tn Z Z
d 1 n− 1
kunh k2n − kunh k2n−1 = (unh )2 = div(wh 2 )(unh )2
t n−1 dt Ωft 2 Ωf
n− 1
2
An illustrative example [Formaggia-N. CMAME ’04]

We solve a scalar parabolic problem

(
∂u
∂t − 0.01∆u = 0, in Ωt , t > 0
u=0 on ∂Ωt , t > 0

in an expanding square Ωt = [0, 2 − cos(20πt)]2 .

The L2 energy is uniformly decreasing in time due to diffusion:

ku(t)kL2 (Ωt ) ≤ ku(0)kL2 (Ω0 ) , ∀t > 0

We solve the problem with different time integration schemes: Implicit

Euler (IE), second order backward difference BDF2, Crank Nicholson
(CN).
When working in fixed domains, the discrete energy is strictly decreasing
in all cases: ku n kL2 (Ω) ≤ ku0 kL2 (Ω) , ∀n = 1, 2, . . .
But what happens with moving domains?
An illustrative example [Formaggia-N. CMAME ’04]
An illustrative example [Formaggia-N. CMAME ’04]
Some References
[D83] J. Donea, An arbitrary Lagrangian-Eulerian finite element method for
transient fluid-structure interactions Comput. Methods Appl. Mech. Engrg.,
33:689–723:1982.
[FGG01] C. Farhat, P. Geuzainne, C. Grandmont, The discrete geometric
conservation law and the nonlinear stability of ALE schemes for the solution of
flow problems on moving grids, J. Comput. Physics 174:669–694, 2001.
[FN99] L. Formaggia, F. Nobile, A stability analysis for the Arbitrary
Lagrangian Eulerian formulation with finite elements, East-West J. Num.
Math., 7:105–132, 1999.
[FN04] L. Formaggia, F. Nobile, Stability analysis of second-order time
accurate schemes for ALE-FEM, Comput. Methods Appl. Mech. Engrg.,
193:4097–4116, 2004.
[EGP09] S. Étienne, A. Garon, D. Pelletier, Perspective on the geometric
conservation law and finite element methods for ALE simulations of
incompressible flow, J. Comput. Physics, 228:2313–2333, 2009.
[JLBL06] J-F. Gerbeau, C. Le Bris, T. Lelièvre, Mathematical Methods for the
Magnetohydrodynamics of Liquid Metals (Ch. 5), Oxford Univ. Press, 2006
Some References

[HLZ81] T.J.R. Hughes, W.K. Liu, T.K. Zimmermann, Lagrangian-Eulerian

finite element formulation for incompressible viscous flows, Comput. Methods
Appl. Mech. Eng. 29(3):329–349, 1981.
[LF96] M. Lesoinne, C. Farhat, Geometric conservation laws for flow problems
with moving boundaries and deformable meshes, and their impact in aeroelastic
computations, Comput. Methods Appl. Mech. Engrg. 134:71–90, 1996.
[N00] B. Nkonga, On the conservative and accurate CFD approximations for
moving meshes and moving boundaries, Comput. Methods Appl. Mech. Eng.,
190:1801–1825, 2000.

[TL79] P.D. Thomas, C.K. Lombard, Geometric conservation law and its
application to flow computations on moving grids, AIAA 17:1030–1037, 1979.
Algorithms for fluid-structure interaction
with large displacements
FSI problem - a three field formulation

Fluid problem FlALE (wn ; un , p n ) = 0 in Ωfn

Structure problem St(η n , η̇ n ) = 0; in Ωs0
ALE mapping Mesh(An , Ȧn ) = 0 in Ωf0

Coupling conditions

kinematic cond. un ◦ An = η̇ n on Γ0
n n f n s
dynamic cond. σ f (u , p )n = −σ s (η )n on Γt
n n n n
geometric cond. A = ξ + η , Ȧ = η̇ on Γ0
wn = Ȧn ◦ (An )−1 , Ωtn = An (Ωf0 )

Fully coupled non-linear problem in the unknowns (un , p n , An , Ȧn , η n , η̇ n )

Partitioned versus Monolithic solvers

I We refer to Monolithic schemes when we try to solve the

non-linear problem all at once. We assemble and solve a “huge”
linear system in the unknowns un , p n , An , Ȧn , η n , η̇ n .

I We refer to Partitioned strategies when we solve iteratively the

fluid and structure subproblems (never form the full matrix). This
allows one to couple two existing codes, one solving the fluid
equations and the other the structure problem.

I There are other options in between: for instance one can “formally”
write the monolithic problem and then use a block preconditioner to
solve it. In this way one recovers partitioned procedures (see e.g.
[Heil, CMAME ’04])
Partitioned fluid-structure algorithms
A simple loosely coupled partitioned FSI algorithm
Given the solution (un−1 , p n−1 , An−1 , Ȧn−1 , η n−1 , η̇ n−1 ) at time
step t n−1 :
1. Extrapolate the displacement and velocity of the interface: e.g.
˜ n = η̇ n−1 ,
η̇ ˜ n , on Γ0
η̃ n = η n−1 + ∆t η̇
(
Mesh(An , Ȧn ) = 0 in Ωf0
2. Compute the ALE mapping:
˜n
An = ξ + η̃ n , Ȧn = η̇ on Γ0
3. Solve the fluid problem with Dirichlet boundary conditions
(
FlALE (wn ; un , p n ) = 0, in Ωfn = An (Ωf0 )
˜ n ◦ (An )−1 ,
un = η̇ on Γn = An (Γ0 )

4. Solve the structure equation with Neumann boundary conditions

(
St(η n , η̇ n ) = 0, in Ωs0
σ s (η n )ns = −σ f (un , p n )nf , on Γn

5. Go to next time step

A staggered algorithm by Farhat and Lesoinne (’98)

A popular algorithm proposed by C. Farhat and M. Lesoinne, CMAME

1998, uses staggered grids:

I The fluid equation is discretized with a second order backward

difference scheme (BDF2) and colocated at t n+1/2
I The ALE mapping is piecewise linear in [t n−1/2 , t n+1/2 ]
I The structure equation is discretized by the mid-point scheme and
colocated at t n+1 .

un−1/2 pn−1/2 un+1/2 pn+1/2

ηn−1/2 σ n+1/2
f

.
ηn−1 η n−1 ηn
.
ηn
.
ηn+1 η n+1
Given the solution (un−1/2 , p n−1/2 , An−1/2 , η n , η̇ n ):
1. Extrapolate the interface displacement and velocity at t n+1/2 :
∆t n
η n+1/2 = η n + η̇ , η̇ n+1/2 = η̇ n , on Γ0
2
(
Mesh(An+1/2 ) = 0
2. Compute the ALE mapping at t n+1/2 :
An+1/2 = ξ + η n+1/2

t − t n−1/2 t n+1/2 − t
and A(t) = An+1/2 +An−1/2 , t ∈ [t n−1/2 , t n+1/2 ]
∆t ∆t
3. Solve the fluid problem with BDF2 at t n+1/2
(
GCL
FlALE (w; un+1/2 , p n+1/2 ) = 0, in Ωfn+1/2 = An+1/2 (Ωf0 )
un+1/2 = η̇ n+1/2 ◦ (An+1/2 )−1 , on Γn = An (Γ0 )

4. Solve the structure equation with Mid-Point at t n+1

(
St(η n+1 , η̇ n+1 ) = 0, in Ωs0
n
+η n+1
σs ( η 2 )ns = −σ f (un+1/2 , p n+1/2 )nf , on Γn+1/2

5. Go to next time step

The staggered temporal grid combined with the Mid-Point discretization
of the structure is particularly nice:
I The Mid-Point discretization satisfies:

η n+1 − η n η̇ n+1 + η̇ n
=
∆t 2
I The ALE mapping at the interface satisfies at t n
η n+1/2 + η n−1/2
 

n 1 n ∆t n n−1 ∆t n−1
A  = = η + η̇ + η + η̇ = ηn
Γ0 2 2 2 2
I Analogously, the mesh velocity in [t n−1/2 , t n+1/2 ] satisfies
 An+1/2 − An−1/2  η n+1/2 − η n−1/2
Ȧ(t) =  =

Γ0 ∆t Γ0 ∆t
 
1 ∆t ∆t
= ηn + η̇ n − η n−1 − η̇ n−1 = η̇ n
∆t 2 2
I The ALE mapping is perfectly consistent with the structure
displacement and velocity.
I Moreover, interpreting the fluid velocity as piecewise constant in
time, we have u(t)|Γ0 = un+1/2 = η̇ n = Ȧ(t)|Γ0 in [t n−1/2 , t n+1/2 ]
The fluid velocity matches at the same time the structure and
the mesh velocities
On partitioned procedures
Partitioned procedures may have serious stability problems in certain
applications.
Dangerous situations appear for incompressible fluids when the mass of
the structure is small compared to the mass of the fluid
The instability is due to a non perfect balance of energy transfer at the
interface. For instance, in the algorithm by Farhat & Lesoinne, we have:
Work done in [t n , t n+1 ] by
Z
n+ 1 1 1 1
fluid Wf 2 = (σ f (un+ 2 , p n+ 2 )nf ) · ∆tun+ 2
Γn+ 1
2

η n+1 + η n s η̇ n+1 + η̇ n
Z
n+ 1
structure Ws 2 = (σ s ( )n ) · ∆t
Γn+ 1 2 2
2

n+ 12 n+ 12
Ideally, one would like to have Wf + Ws = 0.
n+ 1 n+ 1
Energy unbalancing: ∆W = Wf 2 + Ws 2
η̇ n+1 − η̇ n
Z
1 1
∆W = −∆t 2 (σ f (un+ 2 , p n+ 2 )nf ) · = O(∆t 2 )
Γn+1/2 2∆t
A fully implicit partitioned FSI algorithm

To have a perfect energy balance, we need to satisfy exactly at each time

step the kinematic and dynamic coupling conditions.

A possible remedy: fixed point approach: subiterate at each time step,

eventually adding a relaxation step
I if kη n − η̃ n k > tol
set η̃ n ← ωη n + (1 − ω)η̃ n
repeat points 2-4 until convergence

I The fixed point algorithm is stable, in general.

I However, whenever the partitioned procedure is unstable, the fixed
point may need a very small relaxation parameter ω and the
convergence may be very slow.
A fully implicit partitioned FSI algorithm - Dir. Neumann
Given the solution (un−1 , p n−1 , An−1 , η n−1 , η̇ n−1 ) at time step t n−1 , set
η̇ n0 = η̇ n−1 and η n0 = η n−1 + ∆t η̇ n and for k = 1, 2, . . .
1. Compute the ALE mapping:
(
Mesh(Ank , Ȧnk ) = 0
Ank = ξ + η nk−1 , Ȧnk = η̇ nk−1 , on Γ0

2. Solve the fluid problem with Dirichlet boundary conditions

(
FlALE (wkn ; unk , pkn ) = 0, in Ωfn,k = Ank (Ωf0 )
unk = η̇ nk−1 ◦ (Ank )−1 , on Γn,k = Ank (Γ0 )

3. Solve the structure equation with Neumann boundary conditions

(
˜ n ) = 0,
St(η̃ nk , η̇ in Ωs0
k
n s
σ s (η̃ k )n = −σ f (unk , pkn )nn , on Γn,k

4. Relaxation step: If kη̃ nk − η nk−1 k > tol, set η nk = ω η̃ nk + (1 − ω)η nk−1

(and similarly for η̇ nk ) and go to step 1.
Formulation as an interface problem
I Consider the fluid structure problem discretized in time and assume
the solution up to time step t n−1 is known.
n
I Let (λn , λ̇ ) be a suitable displacement / velocity of the
fluid-structure interface at time t n .
Dirichlet-to-Neumann (non-linear) fluid map
n
Let us define the operator F that, given (λn , λ̇ )
I computes the ALE mapping (An , Ȧn )
n
I solves the fluid equations in ALE form with prescribed velocity λ̇ on
n n
the interface (Dirichlet problem) and gets (u , p )
I computes the normal stress on the interface σ f (un , p n ) · nf
n
F(λn , λ̇ ) = σ f (un , p n ) · nf

Dirichlet-to-Neumann (non-linear) structure map

n
Similarly, define the operator S that, given (λn , λ̇ )
I solves the structure problem with prescribed displacement λn on the
interface and gets (η n , η̇ n )
I computes the normal stress on the interface σ s (η n ) · ns
n
S(λn , λ̇ ) = σ s (η n ) · ns
Formulation as an interface problem
Then, the position/velocity of the interface at time step t n satisfies
n n
interface problem: F(λn , λ̇ ) + S(λn , λ̇ ) = 0

Equivalently, the previous interface problem can be written as

n n
fixed point problem: (λn , λ̇ ) = −S −1 F(λn , λ̇ )
n
i.e. the solution (λn , λ̇ ) is the fixed-point of the operator −S −1 F.
The Dirichlet-Neumann implicit algorithm presented before (without
relaxation), corresponds to
Fixed-point algorithm
n
Given (λn0 , λ̇0 ), compute for k = 1, 2, . . . , until convergence
n n
(λnk , λ̇k ) = −S −1 F(λnk−1 , λ̇k−1 )
n
Let us denote ζ n = (λn , λ̇ ) and Φ = −S −1 F. Then the fixed point
algorithm can be written as
ζ nk = Φ(ζ nk−1 )
Fixed-point algorithms on the interface problem
Constant relaxation
As the simple fixed-point iterations might not converge, it is useful to
add a relaxation step
n
ζ̃ k = Φ(ζ nk−1 )
n
ζ nk = ω ζ̃ k + (1 − ω)ζ nk−1
The relaxation parameter ω has to be tuned for each FSI problem
Variable relaxation (Aitken’s extrapolation)
A better strategy consists in using the so called Aitken’s extrapolation
(see [KW08] for application to FSI)

n n
ζ̃ k = Φ(ζ nk−1 ), rk = ζ̃ k − ζ nk−1
(rk−1 , rk − rk−1 )
ωk = −ωk−1
krk − rk−1 k2
n
ζ nk = ωk ζ̃ k + (1 − ωk )ζ nk−1
In the update of ωk one has to use a suitable inner product (·, ·) and
norm k · k2 on the space of the variable ζ nk . At the discrete level, ζ nk is a
vector and one can use the euclidean inner product and norm.
Load transfer between fluid and structure
At the discrete level, there is a proper way of exchanging information.
Conforming finite element spaces at the interface: let {φbi }N i=1 be the
b

basis functions (fluid and structure) corresponding to the interface nodes;

Nb
X Nb
X
uh |Γ = ui φbi |Γ , η h |Γ = ηi φbi |Γ
i=1 i=1
b b
and U = [u1 , . . . , uNb ], d = [η1 , . . . , ηNb ] are the vectors of dofs.
I Fluid Dirichlet b.c. uh = η̇ h
=⇒ nodal values ui = η̇i =⇒ Ub = ḋb
I Structure Neumann b.c. σ s · ns = −σ f · nf
Z
Fs,i = −σ f (uh , ph )nf ·φbi = < Res(uh , ph ), φbi > =⇒ Fbs = −Fbf
Γn | {z }
−Ff ,i
LOAD

i
111
000
000
111
000
111
000
111
000
111
Γ
000
111
NS
Non-conforming finite element spaces at the interface
Nb
I {φb } f : basis functions (fluid) corresponding to the interface
i i=1
PNb
nodes. uh |Γ = i=1f ui φbi |Γ
N
I {ψ b } bs basis functions (structure) corresponding to the interface
j j=1
PNbs
nodes. η h |Γ = j=1 ηj ψ bj |Γ

Fluid Dirichlet b.cs: unh = Πh η̇ nh where Πh is a suitable operator (can be

PNb
interpolation, L2 projection, etc.). Let Cij such that Πh ψjb = i=1f Cij φbi |Γ
N bs
X
=⇒ ui = Cij η̇j =⇒ Ub = C ḋb
j=1

Structure Neumann b.cs.: σ s ns = −Π∗h (σ f nf ) (with Π∗h adjoint of Πh )

Z Nbs
X
Fs,j = − (σ f ·nf )·Πh ψ bj = Cij < Res(uh , ph ), φbi >, =⇒ Fbs = −C T Fbf
Γ i=1
| {z }
−Ff ,i
This coupling strategy allows to balance energy at the discrete level.
Indeed:
I work done by the fluid at the interface (per unit time)
Z Z
Whf = (σ f nf ) · uh = (σ f nf ) · Πh η̇ h
Γ Γ

I Work done by the structure at the interface (per unit time)

Z Z Z
∗
s
Wh = (σ s ns )·η̇ h = − Πh (σ f nf )·η̇ h = − (σ f nf )·Πh η̇ h = −Whf
Γ Γ Γ

At the algebraic level this reads

Whf = Fbf · Ub = Fbf · C db

Whs = Fbs · ḋb = −(C T Fbf ) · ḋb = −Whf

Some References on partitioned FSI algorithms
[DDFQ06] S. Deparis, M. Discacciati, G. Fourestey, A. Quarteroni.
Fluid-structure algorithm based on Steklov-Poincaré operators. Comput.
Methods Appl. Mech. Engrg., 195, 5797-5812, 2006.
[FLLeT98] C. Farhat, M. Lesoinne, P. LeTallec. Load and motion transfer
algorithms for fluid/structure interaction problems with non-matching discrete
interfaces: momentum and energy conservation, optimal discretization and
application to aeroelasticity. Comput. Methods Appl. Mech. Engrg.
157(1-2):95–114, 1998.
[FL00] 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. 182:499–515, 2000.
[KW08] U. Kütter, W.A. Wall, Fixes-point fluid-structure interaction solvers
with dynamic relaxation, Comput. Mech. 43:61–72, 2009.
[LeTM01] P. Le Tallec and J. Mouro. Fluid structure interaction with large
structural displacements. Comput. Methods Appl. Mech. Engrg.,
190:3039–3067, 2001.
[PF95] S. Piperno, C. Farhat. Partitioned procedures for the transient solution
of coupled aeroelastic problems – Part II: energy transfer analysis and
three-dimensional applications. Comput. Methods Appl. Mech. Engrg.
124(1-2):79–112, 1995.
Instabilities of partitioned algorithms: added
mass effect
Mathematical explanation of instability: Added mass effect
Let us consider a toy linear FSI problem
η
Γwall u.n=0
Γin
Γp
p=g(t) Fluid .
u.n= η
Γwall u.n=0 Piston

Simplifying assumptions
I Inviscid, incompressible fluid.
I Small displacements: the fluid domain is considered fixed.
I Fluid equations linearized around the rest state.


ρf ∂t u + ∇p = 0
 Z
Fluid: div u = 0 Piston: m η̈ + c η̇ + kη = pdσ
 Γp
u · n = η̇, on Γp

Modified structure equation R R
we write the fluid pressure on the piston as Γp p = Γp pg − MA η̈, where
pg accounts for the pressure on the left boundary and MA is the usual
added mass operator

−∆pλ = 0 in Ω,

−∆p = 0 in Ω,


 g Z 
∂ p = 0 on Γ ,
n λ wall
∂n pg = 0 on Γwall ∪ Γp MA λ = pλ ,
 Γp pλ = 0 on Γin ,
pg = g on Γin .
 

∂n pλ = %f λ on Γp


Remember that for this problem (see Lecture 4)

Z
pg = g (t)H, MA = %f LH
Γp

Then the equation for the dynamics of the piston modifies as:
Z
(m + MA )η̈ + c η̇ + kη = pg dσ
Γp
Instability of explicit algorithms
Prototype of explicit algorithm: Leap-frog for the piston and Implicit
Euler for the fluid (LF-IE).
Given η n , η n−1 , un−1 , compute
un − un−1

ρ + ∇p n = 0

f
 
∆p n = 0


 ∆t 
div un = 0
 
η n − 2η n−1 + η n−2

div
→ n
∂n p = −ρf
n η n − η n−1 ∆t 2
u · n = , on Γp

 

 n
p = g (t n )


 ∆t
p n = g (t n ), on Γ

in

η n+1 − 2η n + η n−1 η n+1 − η n−1

Z
m 2
+c + kη n = p n dσ
∆t 2∆t Γp

Equivalent discretization of the modified structure equation:

η n+1 − 2η n + η n−1 η n+1 − η n−1
m + c + kη n
∆t 2 2∆t
η n − 2η n−1 + η n−2
Z
= −MA + pgn dσ
∆t 2 Γp
 I At the level of the modified structure equation, explicit FSI coupling
can be interpreted as treating explicitely the fluid added mass
I This can be catastrophic for stability if the fluid added mass is
greater than the structure mass !

Proposition 1 [Causin, Gerbeau, Nobile 2004]

The explicit Leap-Frog/Implicit Euler algorithm is unconditionally unstable
if
m < MA

Proof. Compute the characteristic polynomial χ(s) ∈ P3 of the 3 step

difference equation:
„ « „ «
“ m c ” 3 2m MA 2 m c 2MA MA
χ(s) = + s + − + k + s + − − s+ 2
∆t 2 2∆t ∆t 2 ∆t 2 ∆t 2 2∆t ∆t 2 ∆t
It follows that

χ(−∞) = −∞, χ(−1) = k + 4(MA − m)/∆t 2

Hence, if MA ≥ m, then χ(−1) ≥ 0

=⇒ ∃s ∗ ≤ −1 s.t. χ(s ∗ ) = 0, ∀∆t!!!!

Implicit algorithms
Prototype of implicit algorithm: Implicit Euler for the fluid and first order
BDF for the structure (BDF+IE)

un+1 − un


ρ f + ∇p n+1 = 0 (

 ∆t div ∆p n+1 = 0
n+1 →
div u =0 η n+1 −2η n +η n−1
n+1
− ηn ∂n p n+1 = −ρf ∆t 2
un+1 = η



n
∆t

η n+1 − 2η n + η n−1 η n+1 − η n

Z
m + c + kη n+1 = p n+1 dσ
∆t 2 ∆t Γp

Equivalent discretization of the modified stucture equation:

η n+1 − 2η n + η n−1 η n+1 − η n

Z
(m + MA ) + c + kη n+1 = pgn+1 dσ
∆t 2 ∆t Γp

I Implicit discretization −→ Stable for any ∆t !

Dirichlet-Neumann iterations

Let us solve at each time step the fully implicit problem by

Dirichlet-Neumann subiterations:

given an initial guess η0n+1 , we solve for each k = 1, 2, . . .

i. ∆pkn+1 = 0 in Ω
n+1
ηk−1 − 2η n + η n−1
∂n pkn+1 = −ρf on Γp
∆t 2
pkn+1 = g (t n+1 ) on Γin
η̃kn+1 n
− 2η + η n−1
η̃kn+1
−η n Z
ii. m +c + k η̃kn+1 = pkn+1 dσ
∆t 2 ∆t Γp

iii. ηkn+1 = ω η̃kn+1 + (1 − ω)ηk−1

n+1

I equivalent to a fixed point algorithm on η n+1 .

Proposition 2
The Dirichlet-Neumann iterative algorithm converges iff
2+
0<ω< where  = (c∆t + k∆t 2 )/m
1 + MA /m + 

I In the limit ∆t → 0, whenever the explicit algorithm diverges

(MA > m), the D-N iterative method needs a relaxation parameter
strictly smaller than 1 to converge.
I In this toy problem we can find a value ωopt for which the algorithm
converges in 1 iteration.
I In more complex situations this is not the case and even with an
optimal relaxation parameter, the convergence is very slow if
MA  m.
Other partitioned procedures: Robin-Robin
Idea: Iterate between fluid and structure solvers and use linear
combinations of the Dirichlet and Neumann coupling conditions.
([Badia-N.-Vergara, JCP 2008])
Solve iteratively for k = 1, 2, . . .

FlALE (wkn ; unk , pkn ) = 0
(R)
α un + σ n · nf = α η̇ n − σ n s
f k f ,k f k−1 s,k−1 · n


n
St(η nk , η̇ k ) = 0

(R) n n s n n f
αs η̇ k + σ s,k · n = αs uk − σ f ,k · n


with αs 6= αf
I The coefficient αs (resp. αf ) should incorporate (in a single
constant!) as much of the fluid (resp. structure) dynamics as
possible.
I This is not always feasible, but sometimes it works well.
Choice of αs
We have seen that for an inviscid, incompressible fluid, linearized around
the rest state, the effect of the fluid on the structure can be expressed by
the added mass operator.

p Γ ≈ −MA η̈ · nf

wall

Heuristic argument

η̇ n − η̇ n−1 s
(σ ns · ns ) · ns = −(σ nf ·nf )·ns ≈ −p n ≈ MA η̈ n ·nf ≈ −MA ·n
∆t
which resembles a Generalized Robing boundary condition

∆t −1 MA η̇ n + σ ns · ns = · · ·
| {z }
αs

Problem: MA is not a number, it’s an operator, in general !

Possible Remedies: Take αs ≈ λmax (MA ) – maximum eigenvalue of MA
– or give a rough estimate of the fluid added mass on the structure.
Choice of αf – I
Let’s as consider a simple case first: piston example
η
Γwall u.n=0
Γin
Γp
p=g(t) Fluid .
u.n= η
Γwall u.n=0 Piston
R
On Γp we have m η̈ + c η̇ + kη = Γp
pdσ
Consider, for instance a backward difference discretization:
η n − η n−1
Z
m δ− η̇ n + c η̇ n + kη n = p n dσ, η̇ n = δ− η n =
Γp ∆t
m Z
 m
=⇒ + c + ∆tk η̇ n − p n dσ = η̇ n−1 − kη n−1
∆t Γp ∆t
which resembles a Robin boundary condition
m 
+ c + ∆tk un + σ nf · nf = · · ·
| ∆t {z }
αf

I piston’s inertia, damping and stiffness effects can all be included in

the constant αf .
Choice of αf – II
A more complex situation: let us consider a full 3D elasticity model
The structure equation in weak form reads
∂2η
Z Z
0
%s 2 · φ + σ s (η) : ∇0 φ = − (σ f (u, p) · nf ) · φ
Ωs0 ∂t Γwall

Assume the structure is thin (shell-like structure with thickness hs ). Then

one could consider a shell model, instead
∂2η
Z Z
hs %s 2 · φ + E(η; φ) =− (σ f (u, p) · nf ) · φ
Γ0 ∂t | {z } Γwall
membrane and flexural terms

At least the mass term can be included in the Robin boundary conditions:
∆t −1 hs ρs un + σ f (un , p n ) = · · ·
| {z }
αf

In some cases, other terms could be included: e.g. for a linear membrane
type solid one could have a term proportional to the local curvature of
Γwall (see [N.-Vergara, SISC 2008])
(
hs ρs ∆tEhs “ 2 ” E , ν : elastic constants
αf = + 4ρ1 − 2(1 − ν)ρ2 ,
∆t (1 − ν )
2
ρ1 , ρ2 : mean and Gaussian curvatures
Some comments on Robin-Robin partitioned algorithms

I The choice of αs and αf is highly problem dependent

I In certain cases a good choice can lead to a significant improvement
in the convergence properties
I We have presented Robin-Robin procedures in the context of implicit
coupling algorithms (within subiterations). However, they may be
used also for explicit coupling: at each time step solve only one fluid
problem with Robin b.cs and one structure problem with Robin b.cs.
I In our experience (hemodynamics applications) Robin conditions on
the fluid are very effective, whereas Robin conditions on the
structure do not improve much the convergence (αs = 0 is often a
good choice).
Monolithic approach
Global weak formulation
We recall the global weak formulation for the fluid structure problem,
with the fluid in ALE formulation:
−1
Find (u, p, η) s.t. u = ∂η
∂t ◦ Lt on Γt and

Z  !
∂u 
%f + (u − w) · ∇u · v + σ f (u, p) : ∇v + div uq+
Ωft ∂t ξ
∂2η
Z Z Z
0 f
%s 2 ·φ+σ s (η) : ∇ξ φ = f ·v+ f0s ·φ, ∀(v, q, φ) ∈ VFS
Ωs0 ∂t Ωft Ωs0

with VFS ≡ {(v, q, φ) : v ◦ Lt = φ on Γ0 }.

complemented with a problem for the ALE mapping, e.g. a Laplace eq.:
find At s.t. At = η(t) on Γ0 and
Z
∇ξ At · ∇ξ ψ = 0, ∀ψ ∈ VA
Ωf0

We recall that the continuity of stresses is already embedded into the

global weak formulation.
Time discretization
We discretize now in time the global problem. To make the presentation
simple we consider the Implicit Euler scheme for the fluid and the
Mid-Point scheme for the structure.
Find (un , p n , η n , η̇ n , An , Ȧn ) such that

un − ũn−1
Z  
%f + (un − Ȧn ) · ∇un · v + σ f (un , p n ) : ∇v + div un q
An (Ωf0 ) ∆t
η̇ n − η̇ n−1 η n + η n−1
Z
+ %s · φ + σ 0s ( ) : ∇ξ φ
Ωs0 ∆t 2
Z
+ ∇ξ An · ∇ξ ψ
Ωf0
Z Z
f ,n s,n−1/2
= f ·v+ f0 · φ, ∀(v, q, φ) ∈ VFS , ψ ∈ VA
Ωft Ωs0
2 2
η̇ n = (η n − η n−1 ) − η̇ n−1 , Ȧn = (An − An−1 ) − Ȧn−1
∆t ∆t
un ◦ An = η̇ n , An = η n , on Γ0

where ũn−1 = (un−1 ◦ An−1 ) ◦ (An )−1 .

Time discretization
In the previous system we can eliminate the variables η̇ n and Ȧn and
write it in abstract form as
(
Lf (An , un ; p n , v, q) + Ls (η n ; φ) + LA (An , ψ) = F(v, φ)
un ◦ An = η̇ n , An = η n , on Γ0

with
un − ũn−1
Z „ «
Lf (An , un ; p n , v, q) = %f + (un − Ȧn ) · ∇un · v
An (Ωf0 ) ∆t
Z
+ σ f (un , p n ) : ∇v + div un q
An (Ωf0 )
n n−1
η n + η n−1
Z
η̇ − η̇
Ls (η n ; φ) = %s · φ + σ 0s ( ) : ∇ξ φ
Ωs0 ∆t 2
Z
LA (An , ψ) = ∇ξ A n · ∇ ξ ψ
Ωf0

Observe that the fluid operator Lf is linear in p n , non-linear in un and

depends also non-linearly on the ALE mapping through the mesh velocity
Ȧn and through the integration domain An (Ωf0 ).
Newton algorithm
We can now write a global Newton algorithm: given (un0 , p0n , η n0 , An0 )
solve for k = 0, 1, 2, . . . until convergence



 (Lf )0u (Ank , unk , pkn ; δu, v, q) + (Lf )0A (Ank , unk , pkn ; δA, v, q)
+Lf (Ank , unk ; δp, v, q) + (Ls )0η (η nk ; δη, φ) + LA (δA, ψ)




= F(v, φ) − Lf (Ank , unk ; pkn , v, q) − Ls (η nk ; φ) − LA (Ank , ψ)
2
δη − (unk ◦ Ank − η̇ nk )
 n
δu ◦ Ak = ∆t on Γ0




δA = δη − (An − η n ),

on Γ0
k k

unk+1 = unk + δu, pk+1

n
= pkn + δp, η nk+1 = η nk + δη, Ank+1 = Ank + δA
I The tangent operator contains, in particular, the term
(Lf )0A (Ank , unk , pkn ; δA, v, q) that gives the derivative of the fluid
operator with respect to the domain. This is called shape derivative
I Such term is computed by recasting first the integral on Ωf0 and then
differentiating all the terms appearing inside the integrand due to the
ALE mapping (similar procedure as for solids in Lagrangian frame)
I By doing so, the perturbation on the fluid velocity is considered
always in the reference configuration, i.e. δu = δu(ξ). This is why
δ(un ◦ An ) = δu ◦ An
Newton algorithm – algebraic form
We consider now a finite element formulation with conforming finite
elements at the fluid structure interface.
We group the degrees of freedom as
I (Ui , P): pressure dofs and velocity dofs on internal nodes only

I (Ub ): velocity dofs on the FS interface

I (Ai ), (Ab ): mesh displacement dofs on internal (resp. boundary)

nodes
I (di ), (db ): structure displacement dofs on internal (resp. boundary)
nodes

At algebraic level the Newton method leads to the following system

2 (ii) (ib) (ii) (ib)
32 3 2 R(i) 3
(Lf )0u,p (Lf )0u,p (Lf )0A (Lf )0A 0 0 δ(Ui , P) f
0 I 0 0 2
− ∆t I 0 6R(b) 7
7 6 δUb 7
6 76
6 6 f 7
(i) (ib) 7 6 (i) 7
0 0 LA LA 0 0 7 6 δAi 7 R 7
6 76
6 A7
7=6
6
0 0 0 I −I 0 7 6 δAb 7 6R(b)
6 76 7
6 A7
6 7
6 (bi) 0 (bb) (bi) (bb) (bb) (bi)
(Lf )0u,p (Lf )0A (Lf )0A (Ls )0η (Ls )0η 5 4 δdb 5 4 R(i)
7
4(Lf )u,p s
5
0 0 0 0
(ib)
(Ls )0η
(ii)
(Ls )0η δdi (i)
Rs

The 5th line corresponds to the equality of stresses in a weak sense.

Newton algorithm – algebraic form
I At each time step we have to solve a fully coupled, linearized fluid
structure problem.
I A direct solution is most of the times unfeasible. One has to use
iterative methods.
I Preconditioning of this global system is an issue.
I Block preconditioners go back to the direction of partitioned
schemes. For instance the lower block triangular preconditioner
(ii) (ib) (ii) (ib)
2 3
(Lf )0u,p (Lf )0u,p (Lf )0A (Lf )0A 0 0
6 0 7
2
− ∆t
6 0 I 0 0  *I 0 7
6  7
(i) (ib)
0 0 LA LA 0 0
6 7
6 7
6 0 7
0 0 0 I −I 0 7
6 >
 7
6 
6 (bi) 0 (bb) 0 (bi) (bb) 0 (bb) (bi 0 7
4(Lf )u,p (Lf )u,p (Lf )0A (Lf )A (Ls )0η (Ls )η 5
(ib) (ii
0 0 0 0 (Ls )0η (Ls )0η

corresponds to a Dirichlet-Neumann partitioned procedure applied

to the linearized fluid structure problem at each Newton iteration.
Quasi Newton algorithms

I Alternatively in the Newton algorithm one could replace the exact

tangent matrix with an approximated one. In this case, the
algorithm will not converge quadratically any more.
I For instance, since the terms (Lf )0A are the most difficult to
compute, one could drop them.
*0
2 3
(ii) (ib) (ii)  (ib) :0 3 2 R(i) 3
6 (Lf )0u,p (Lf )0u,p 0 0
  2
(L
f )A (L
 f )A 0 0 7 δ(Ui , P) f
 2 6R(b) 7
0 I 0 0 − ∆t I 0 7 6 δUb 7
6 76
6 f
7 6 (i) 7
6
(i) (ib)
7 6 δAi 7 RA 7
6 76 7
6 0 0 LA LA 0 0 7=6
6 (b)
δA
6 76 7
6 0 0 0 I −I 0 76 b 7
7 6RA 7
:0  : 0 (bb) δdb 5 4 R(i)
6 74 6 7
6 (bi) 0 (bb) 0 (bi)0 (bb) 0 (bi
)A (Ls )0η (Ls )0η 5 s
7
4(Lf )u,p (Lf )u,p (L
f )A (Lf 
5
 
(ib) (ii
δd i (i)
Rs
0 0 0 0 (Ls )0η (L )0
s η
Some References on monolithic FSI algorithms + added
mass effect
[BNV08] S. Badia, F. Nobile and C. Vergara, Fluid-structure partitioned
procedures based on Robin transmission conditions , J. Comput. Physics, 2008,
vol. 227/14, pp. 7027-7051
[BQQ08] S. Badia, A. Quaini, and A. Quarteroni. Modular vs. non-modular
preconditioners for fluid-structure systems with large added-mass effect.
Comput. Methods Appl. Mech. Engrg., in press, 2008.
[CGN05] P. Causin, J.F. Gerbeau, and F. Nobile. Added-mass effect in the
design of partitioned algorithms for fluid-structure problems. Comput. Methods
Appl. Mech. Engrg., 194(42-44):4506–4527, 2005.
[FM05] M.A. Fernández and M. Moubachir. A Newton method using exact
Jacobians for solving fluid-structure coupling. Computers & Structures,
83(2-3):127–142, 2005.
[H04] M. Heil. An efficient solver for the fully coupled solution of
large-displacement fluid-structure interaction problems. Comput. Methods
Appl. Mech. Engrg. 193(1-2):1–23, 2004.
[MS02] H.G. Matthies, J. Steindorf, Partioned but strongly coupled iteration
schemes for non-linear fluid-structure interaction, Computers and Structures,
80 (2002), 1991-1999.