Professional Documents
Culture Documents
-
An introduction to numerical coupled simulation
Lecture Notes
May 6, 2008
Preface
These lecture notes have been written specifically for the lecture wb1417 Fluid-Structure Inter-
action at the TU Delft, The Netherlands. There are based on earlier lecture notes of Prof. D.J.
Rixen used in previous years for the course, and on lecture notes written by Prof. H. Bijl, A.
van Zuijlen and A. de Boer from the Faculty of Aerospace building on their expertise and the
knowhow gathered through several years of research and on published literature.
These lecture notes should be considered as the basis study material for the course. More
details can be found in the documents listed in the reference.
We did our best to compile a set of accurate and clear lecture notes. Nevertheless the notes
will be constantly improved and corrected over the years. Therefore feedback from students is
always appreciated!
We hope that you will have as much excitement studying this topic highly relevant to the
modern mechanical engineer (hence including aerospace engineers) as we had discovering the in
and outs during our research and teaching activities.
Prof. H. Bijl
Prof. D.J. Rixen
Contents
1 Introduction 1
5 Moving Meshes 67
5.1 Mesh deformation methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.1.1 Complete regeneration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.1.2 Transfinite interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.1.3 Master-slave coupling with transparent block boundaries . . . . . . . . . . 69
5.1.4 Spring analogy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.1.5 Least squares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.1.6 Solid body elasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.1.7 Laplacian smoothing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.1.8 Biharmonic operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.1.9 Radial basis function interpolation . . . . . . . . . . . . . . . . . . . . . . 73
5.2 Mesh quality metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
5.3.1 Example 1: Rotation and translation . . . . . . . . . . . . . . . . . . . . . 79
5.3.2 Example 2: Rigid body Rotation . . . . . . . . . . . . . . . . . . . . . . . 80
5.3.3 Example 3: Flow around airfoil . . . . . . . . . . . . . . . . . . . . . . . . 80
5.4 Importance of smooth mesh deformation for higher order time-integration . . . . 81
List of Figures
3.21 Unstable solution with the Forward Euler scheme for ∆t = 0.4. . . . . . . . . . . 47
3.22 Stability region for the Backward Euler scheme. . . . . . . . . . . . . . . . . . . . 48
3.23 Stability region for the Leapfrog scheme. . . . . . . . . . . . . . . . . . . . . . . . 48
Introduction
Fluid-structure interaction is a vast field. On one side of the spectrum of the problem handled,
the fluid and structure problems are fully linearized for small oscillations: vibro-acoustics. On
the other extreme of the spectrum, we have the general problems of transient motion of fluid-
structure interaction with large modifications of the fluid domain due to structural motion and
where the fluid behaves non-linearly, with viscous and thermal effects.
The problems of the latter domain can be handle nowadays thanks to the tremendous in-
crease of computing power and to significant advances in numerical methods obtained during
the last decade. The problems treated by advanced fluid-structure interaction include for in-
stance the sloshing of tanks in launchers, limit cycle oscillations of wings, instabilities in the
wind interaction with cable stayed bridges (think about the Tacoma or the Erasmus bridge),
bio-fluids in deformable vessels, sound production in instruments ... (see Figure 1.1).
Being able to model fluid-structure interactions with high accuracy is important mainly for
2 reasons:
• Predict the performance of fluid-structure systems. For instance evaluate the effect of
surgical interventions to the functioning of the heart, predict the actual lift of a wing
taking into account its elasticity ....
Advanced simulation in fluid-structure interaction becomes more and more important for
several reasons. One is that modern structures are optimized and are usually significantly more
flexible and have non-conventional (hence not well-known) configurations (modern airliners, X29
experimental aircraft with forward swept wings, new bridge designs). Another reason is that
one often wants to simulate problems which are otherwise very difficult to observe (bio-flows,
micro-fluidics in microsystems)
Classically, fluid-structure interaction problems have been analyzed using analytical or semi-
analytical methods. Those methods are very cumbersome/complex to use and based on strong
assumptions (inviscid, velocities away form the transonic regime, harmonic motion ....).
Here we will discuss basics of modern numerical methods to couple Computational Fluid
Dynamics (mainly Finite Volumes) and structural analysis (Finite Elements). Let us consider
the example of a wing section deforming and moving in a fluid flow at transonic speed (Figure
1.3). In is seen that next to the challenge of computing the flow and the structure dynamics,
important issues have to be treated such as the motion of the mesh to follow the moving
structure, the non-matching of the fluid and structural meshes, and the time-integration of the
2 CHAPTER 1. INTRODUCTION
coupled problem. These issues will be shortly discussed in these lecture notes. But first, we will
give a short summary of the fluid-mechanics and the finite volume method.
Remark These lecture notes have been widely inspired by the research done at the Center for
Aerospace Structures, University of Colorado at Boulder. Some relevant publications supporting
this text can be found in the bibliography at the end of the manuscript.
Chapter 2
As an example, let us consider the Eulerian field for the temperature TEulerian in a steady-state
flow taken as linearly variant in a one dimensional domain:
x
TEulerian = T0 + (TL − T0 )
L
If the particles in the entire domain have constant velocity in time, we can write uLagrangian = st
6 CHAPTER 2. FLUID MECHANICS AND ISSUES IN INTERACTION PROBLEMS
where s is the constant velocity. The Lagrangian description of the temperature is then
X + st
TLagrangian = T0 + (TL − T0 )
L
Let us note that (2.1) can also be written as
aLagrangian (X, t) = aEulerian (x = X + uEulerian (x, t), t) (2.2)
where uEulerian (x, t) is the Eulerian displacement field, namely the displacement undergone
between time 0 and t by a particle being at the spatial location x at time t. This is nearly never
used in practice since one is usually not interested in displacements when taking the Eulerian
approach (like in fluid), but relation (2.2) will be useful later in the theoretical developments.
In what follows we will shortly recall how the fluid mechanics equations can be written,
considering the problem to be continuous, namely that we are interested in the macro-behavior
of the flow, not in the very displacement of each of its particle. Obviously the Eulerian approach
will be chosen (the subscript ⋆Eulerian being dropped for simplicity).
If we want to write an equation similar to (2.3) for an infinitesimal volume dV , we can pass
the time-derivative operator under the integral sign, remembering that the elementary volume
dV itself depends on time since it follows the deformation of the particle collection due to the
flow. Hence Z
D (ρdV )
=0 (2.5)
V Dt
or
D (ρdV )
=0 (2.6)
Dt
which also writes
Dρ DdV
dV + ρ =0 (2.7)
Dt Dt
The time variation of an elementary volume dV enclosing a given set of fluid particles can be
expressed in terms of the velocity of the particles. Indeed, considering a finite volume of given
particles, Z Z
DVol D DdV
= dV = (2.8)
Dt Dt Vol Vol Dt
and on the other hand, the change of volume can be evaluated from the normal displacement
of the volume faces, namely
Z Z
DVol
= v n dS = ∇ v dV (2.9)
Dt Surf Vol
2.1.1 Conservation of mass 7
so that
Dρ
+ ρ∇ v = 0 (2.12)
Dt
which indicates that the variation of mass density for a given collection of particles is due to
the difference of velocity on its surfaces, the latter creating a variation of volume.
∂
Let us define ∂t the time derivative of a property considering the property for a point fixed
D ∂
in space during dt. The substantial derivative Dt is related to the time derivative ∂t by1
D⋆ △ ∂⋆
=
Dt ∂t given particle
X ∂xi 3
∂⋆ ∂⋆
= + (2.13)
∂t given space position x ∂t given particle ∂xi
i=1
indicating that the change of variation in time of a property for a given elementary volume of
fluid particles is equal to the time variation at a location fixed in space plus the space variation
of the property times the speed of the particles. Defining vi the velocity of the particles, we
write
3
D⋆ ∂⋆ X ∂⋆
= + vi
Dt ∂t ∂xi
i=1
∂⋆
= + v ∇⋆ (2.14)
∂t
1 D⋆
This relation can also be derived from (2.2) by considering that the substantial derivative Dt
corresponds to
deriving in time while always considering a given particle, namely keeping X constant:
D⋆ △ ∂⋆ ∂ ⋆ (x = X + u(x, t), t)
= =
Dt ∂t X ∂t X
∂ ⋆ (x, t) ∂ ⋆ (x, t) ∂ (X + u(x, t))
= +
∂t ∂x ∂t X
∂ ⋆ (x, t) ∂ ⋆ (x, t)
= + v
∂t ∂x
8 CHAPTER 2. FLUID MECHANICS AND ISSUES IN INTERACTION PROBLEMS
The mass conservation relation for a given collection of particles (2.12) can be transformed
using (2.14) as
∂ρ
+ v ∇ρ + ρ∇ v = 0
∂t
and thus
∂ρ
+ ∇ (ρv) = 0 (2.15)
∂t
This relation indicates that, for an observer fixed in space, the change of density at that location
is related to the mass flow in and out of the elementary volume fixed in space. Integration of
this last equation over a volume fixed in space,
Z
∂ρ
+ ∇ (ρv) dV = 0 (2.16)
Vf ix ∂t
Since Vf ix does not depend on time and using Gauss’s integration formula,
Z Z
∂
ρ dV + ρv n dS = 0 (2.17)
∂t Vf ix Sf ix
where n is the vector defining the external normal to the surface. This relation is the integral
counterpart of (2.15) and expresses the fact that for a finite volume in space, the variation of
the fluid mass it contains is compensated by the net mass flow through its surface boundary.
Remarks
• The different forms of mass conservation laws derived above can be classified as conservative/non-
conservative, integral/local:
Integral conservative form (fixed finite volume)
Z Z
∂
ρ dV + ρv n dS = 0 (2.17)
∂t Vf ix Sf ix
• Using the mass conservation (2.15), the relation (2.14) between substantial and fixed space
position time-derivative can be written as,
D⋆ ∂⋆ ∂ρ
ρ = ρ + ρv ∇ ⋆ + ⋆ + ∇ (ρv)
Dt ∂t ∂t
∂ρ⋆
= + ∇ (ρv⋆) (2.18)
∂t
2.1.2 Conservation of momentum 9
Dρinc
=0 (2.19)
Dt
∂ρinc
=0 (2.20)
∂t
∇ρinc = 0 (2.21)
Hence,
∇v =0 (2.22)
where p is the hydrodynamic pressure, τ the tensor of viscous stresses and f the external
body-forces (gravity for instance). Componentwise and using (2.18),
X ∂τri 3
∂ρvi ∂p
+ ∇ (ρvvi ) = − + + ρfi (2.26)
∂t ∂xi ∂xr
r=1
i = 1, 2, 3
∂(ρv)
+ ∇ (ρvv) = −∇p + ∇ τ + ρf (2.27)
∂t
Note the outer product vv. This last form is a local conservative form and expresses that in an
infinitesimal fixed volume, the rate of change of momentum is equal to the applied forces minus
the balance of momentum entering and exiting the fixed volume due to the flow (convective
term). This is also understood by taking the integral over a fixed volume of (2.27):
Z Z Z Z
∂
ρv dV + ρv(v n) dS = (−pn + τ n) dS + ρf dV (2.28)
∂t Vf ix Sf ix Sf ix Vf ix
10 CHAPTER 2. FLUID MECHANICS AND ISSUES IN INTERACTION PROBLEMS
vv
E =e+ (2.30)
2
e is the internal energy representing the atomic/molecular energy (vibration, spin, electronic
...).2
The conservative form of (2.29) is found by using (2.18):
∂(ρE)
+ ∇ (ρEv) = ∇ (−pv + τ v) + ρf v − ∇ q (2.31)
∂t
Noting that multiplying the local non-conservative momentum equation (2.24) by v, the
rate of change of kinetic energy for a particle collection writes
Dv
ρ v = (−∇p + ∇ τ ) v + ρf v (2.32)
Dt
Combining this relation with the energy conservation equation (2.29) and the definition (2.30),3
De ∂vi
ρ = −∇ q − p∇ v + τij (2.33)
Dt ∂xj
The last term corresponds to the dissipation due to the rate of mechanical energy expended in
the deformation of the fluid due to viscosity.
• p
The fluid system is fully determined if one takes into account the following equations.
2
Here we assume that only the internal energy and the kinetic energy are significant for the total energy E.
For instance we neglect potential energies.
3
From here on we use Einstein’s summation convention
2.1.4 Additional relations 11
Navier-Stokes law
For so-called Newtonian fluids (most of gases and liquids), viscous stresses at a point are pro-
portional to the rate of strain of the fluid (or, equivalently, to the space variation of the velocity
field). In general one can then write (Navier’s law)
∂vi ∂vj ∂vk
τij = τji = µ + + δij µ′ (2.34)
∂xj ∂xi ∂xk
where µ and µ′ are respectively the coefficient of viscosity (dynamic viscosity) and the second
coefficient of viscosity. For the direct stresses one has
∂vi ∂vk
τij = 2µ + µ′ when i=j
∂xj ∂xk
The coefficient of bulk viscosity κ corresponds to the viscosity observed when an average hy-
drodynamic compression is applied, namely
∂vk
τkk = 3κ
∂xk
and therefore κ = 2/3µ + µ′ . In general the viscosity associated to bulk hydrodynamic compres-
sion in neglected (might be important though for shock waves or damping in acoustics). Hence,
if κ = 0, one finds Stokes’ law.
2
µ′ = − µ (2.35)
3
and
∂vi ∂vj 2 ∂vk
τij = τji = µ + − δij µ (2.36)
∂xj ∂xi 3 ∂xk
Introducing this relation in the momentum equation (2.27), one obtains the so-called Navier-
Stokes equations.
Let us note that if one assumes the viscosity µ to be constant and the fluid to be incom-
pressible so that the velocity is divergence free, the viscous term in the Navier-Stokes equations
simplify to
∇ τ = µ∇ (∇ v) [constant µ and incompressible] (2.37)
This corresponds to a Laplace type of problem. Note that the viscous terms in the Navier-
Stokes equations are very similar to the stress terms in the linear equilibrium equation of solid
mechanics.
Fourier’s law
The heat flux is related to the space variation of temperature T like in diffusion laws. Fourier’s
law states that heat flux due to conduction is given by
q = −k∇T (2.38)
Equations of state
The 2 relations needed to fully determine the fluid behavior are relations between the thermo-
dynamic variables ρ, p, T, e: the equations of state. Choosing ρ and e as independent variables,
the equations of state must provide
p = p(e, ρ) (2.39)
T = T (e, ρ) (2.40)
ρ = constant (2.41)
e = cT (2.42)
p = ρRT (2.43)
e = cv T (2.44)
where R is the gas constant and cv the specific heat at constant volume.
In that case the energy equation in the Euler equation needs no to be solved since the first 4
relation together with the equation of state (2.47) are sufficient to determine ρ and v.
Note that if in the Navier-Stokes or Euler equations v, ρ, p and E are not function of time,
the flow is said to be steady state. This does not mean that the fluid does not move, but that
for a given position in space, the state of the fluid is unchanged. The inertia forces are then
coming only from the convective terms.
When the boundary condition (2.49) or (2.50) must be enforced on a moving boundary
(u̇ 6= 0), we have to remember that the fluid velocity v is the velocity at a given position
in space, whereas the velocity v in a structure is usually defined as the Lagrangian velocity,
meaning that one must enforce
v(x0 + u(x0 , t), t) = u̇(x0 , t) (2.52)
This expression indicates that the boundary condition for the fluid must be enforced for a
location in space that varies in time. This is one of the difficulties in fluid-structure interaction.
Several ways to treat the issue have been proposed. Two important ones are:
• Assuming that the body displacement is small compared to the flow disturbance it gen-
erates, one can state that x0 + u(x0 , t) ≃ x0 . Hence, this approach consists in assuming
that the boundary of the fluid does not change, but that the boundary velocity is not
zero:
v(x0 , t) = u̇(x0 , t) (2.53)
This approximation is known as the transpiration approach and has been often used in
the past to derive closed form solution of fluid flow around vibrating structures.
• If the body displacements are not small, another methods consists in re-writing the fluid
equations no longer for a point fixed in space as in the Eulerian representation, but for
a point in the fluid that has a given displacement, namely the one of a reference frame
following the moving body. This will be explained in chapter 2.3.
14 CHAPTER 2. FLUID MECHANICS AND ISSUES IN INTERACTION PROBLEMS
In this equation, the set of components in W will now be considered as unknowns and the terms
E, F are then non-linear functions of W . The first term represents the time variation of W
at a fixed point whereas the second and third terms represent a convection contribution. The
term J is the source term.
Integrating the local conservative form of the Navier-Stokes equations over a volume fixed
in space Z Z Z
∂
W dV + [E(W ), F (W )] n dS = J dV (2.56)
∂t Vf ix Sf ix Vint
Often the viscous terms are put on the right hand side of the equations so that these relations
are also put in the form
Z Z h i Z
∂
W dV + Ẽ(W ), F̃ (W ) n dS = R(W ) dV (2.57)
∂t Vf ix Sf ix Vint
where Ẽ and F̃ are similar to the definitions (2.55) except that the viscous forces are moved
to R. In this way one can solve for the viscous terms by Finite Element methods whereas the
time variation and convective terms are solved by Finite Volume.
The relations (2.57) express that the time variation of W included in the volume Vf ix
is balanced by the flow of W through its boundary surface Sf ix and, if present, source and
viscous contributions. Hence let us assume that the fluid domain is partitioned into cells. Joining
2.2. FINITE VOLUME DISCRETIZATION ON FIXED MESHES 15
the center of these cells, one defines a finite volume centered on the vertex i for which the
conservation relations (2.57) can be approximated by
d
(Vi Wi ) + Φi (Wj ) = Ri (Wj ) (2.58)
dt
In this relation
• Vi represents the volume of the finite volume cell i.
• Wi is the discretized value of W in the center of the cell, representing an average over
the cell.
• Φi arises from a discretization of the convective term through the edges of the cell i
(function of the discretized unknowns in the cell and in neighboring cells). Several ap-
proximations of the convective flux trough the edges can be defined based on elementary
solutions for one-dimensional flows from the cell center with state Wi to the neighboring
cell centers with state Wj . The construction of this flux discretization is an essential issue
(refer to specific C.F.D. courses or references).
• Ri (Wj ) represent the discretized contribution of the source and viscous terms, often
discretized using Finite Elements.
Solving the non-linear set of time-differential equations (2.58) for the discretized unknowns
Wi , one finds the solution of the fluid problem. Note that the Navier-Stokes equation are first
order non-linear differential equations and their mathematical type is very different from the
equations of motion in a solid. This can be understood by remembering that the fluid solution
includes shocks in supersonic regimes and the acceleration terms are written in terms of first
time-derivative of velocity unknowns and convective terms. Hence time-integrators must be
carefully chosen.
We observe that this technique is no longer applicable in the case the cells of the Finite
Volume discretization move for instance to follow the motion of bordering flexible structures.
Indeed in that case the integration volume in (2.56) is no longer constant in time and its
location varies so that the time derivative can not be brought outside of the volume integral. In
the following chapter we will indicate how the Finite Volume method can be adapted to take
account of moving meshes.
∂⋆ ∂⋆
= + ẋ ∇⋆ (2.60)
∂t moving frame ∂t given space location
where ẋ is the speed of the moving reference frame in the fixed one and ∇ is the space derivative
with respect to the fixed frame. Let us note that for the particular case where the frame moves
∂ D
exactly as the fluid particles, ∂t moving frame
= Dt and ẋ = v as defined in the previous chapter.
So (2.60) is equivalent to (2.14) in that case. Hence (2.60) is a generalization of (2.14) for a
point moving with arbitrary motion.
Next we note that in (2.60)
Combining (2.60) and (2.61), the first term of (2.59) can be written as
Z Z !
∂W ∂W
dV = − ẋ ∇W dV
Vmov ∂t Vmov ∂t moving frame
Z !
∂W
= − ∇ (ẋW ) + W ∇ ẋ dV (2.62)
Vmov ∂t moving frame
∂dV
= ∇ ẋ dV (2.63)
∂t moving frame
stating that the relative change of elementary volume for a volume attached to a moving frame
is equal to the divergence of the velocity of the motion. Finally combining this last equation
with (2.62) , one obtains
Z Z Z
∂W ∂(W dV )
dV = − ∇ (ẋW ) dV
Vmov ∂t Vmov ∂t moving frame Vmov
and finally
Z Z Z
∂W ∂
dV = W dV − W ẋ n dS (2.64)
Vmov ∂t ∂t Vmov Smov
where ẋ is the speed of the mesh. This last relation can be interpreted as described in figure
2.2:
The integral of the time variation of W on an instantaneous volume = the time variation
of the integral over the volume − flux of W across the boundaries due to the edge motion.
18 CHAPTER 2. FLUID MECHANICS AND ISSUES IN INTERACTION PROBLEMS
This conservative form on a moving cell is similar to the basic Finite Volume relations (2.56).
It shows that the Finite Volume method can still be applied on a moving mesh if one consider
an additional flux term corresponding to the flux of W through the boundary of the cell due
to the motion of the cell edges. Similarly to (2.58) the discretized Finite Volume equation can
be schematically written as
d
(Vi (t)Wi ) + Φci (Wj , x, ẋ) = Ri (Wj , x) (2.66)
dt
where the superscript c in the discretized convective term indicates that it includes the convec-
tion term associated to the mesh motion.
Let us further note that the procedure described here to write the fluid dynamics equation
on a moving reference volume can also be used for Finite Element discretization. Since in this
approach the observation frame is not fixed in space nor moves with the material but has an
arbitrary motion in order to follow specific boundary conditions, this approach is known as the
Arbitrary Lagrangian Eulerian formulation, often used for instance in solid mechanics when
simulating casting procedures for instance. In the world of fluid mechanics this approach is also
known as the dynamic mesh formulation.
and the Finite Volume equations discretized in time are obtained by applying some quadrature
formula to approximate the time integral. Since, unlike in classical Finite Volume schemes, the
2.3.3 Time-integration of F.V. on moving mesh 19
Figure 2.3: Different mesh configurations between two time-steps tn and tn+1
cell is moving, one has to specify on which configuration the flux and dissipation terms should
be evaluated. This is illustrated in Figure 2.3.
If we use a backward difference formula, we can write the discretized time equations for a
Finite Volume in the following general form
αn+1 V (xn+1
i )Win+1 + αn V (xni )Win + αn−1 V (xn−1 i )Win−1 (2.68)
X
+∆t βklm Φci Wjk , ẋlj , xm
j − γ R
k,l,m i Wj
k
, ẋ l
j , x m
j =0
k,l,m
where αn+1 , αn , αn−1 are known coefficients related to a given integration scheme and ∆t =
tn+1 − tn is the time-step. The discretized convective and dissipation terms are obtained as a
weighted sum of fluxes evaluated for different instances of W , ẋ, x, the indices k, l, m respectively
corresponding to possible choices of instances to evaluate the unknowns, the mesh velocity and
mesh position when computing the discretized fluxes and dissipation.
In particular, one specific method of the general expression (2.68) is the four-point rule
written as (omitting the dissipation term for simplicity)
4
X
αn+1 V (xn+1
i )Win+1 +αn V (xni )Win +αn−1 V (xn−1
i )Win−1 +∆t ωk Φci Wjn+1 , ẋsj k , xsj k =0
k=1
(2.69)
Hence the fluxes are computed with the unknowns at time tn+1 considered on four different
configurations of the cell between tn and tn+1 . The weight coefficients ωk must be chosen to
have, in some sense, the best approximation of the actual convective flow on the moving mesh.
One way to define the weighting coefficients ωk is to impose that the integration scheme
satisfies the so-called Discrete Geometric Conservation Law or DGCL requiring that the scheme
should be able to exactly integrate a uniform flow for any mesh motion.4 Substituting Wi (t) in
(2.67) by a constant uniform field W ∗ , we have
Z tn+1 Z tn+1 Z tn+1
d ∗ c ∗
(Vi (t)W ) dt + Φi (W , x, ẋ) dt = Ri (W ∗ , x) dt
tn dt tn tn
Noting that convective fluxes not associated to mesh motion are null and that the dissipation
terms are null for a uniform flow,
Z tn+1 Z
n+1 n n−1
αn+1 V (xi ) + αn V (xi ) + αn−1 V (xi ) = ẋ n dS dt (2.70)
tn Smov
4
Note that this is conceptually similar to the patch test for finite elements in solid mechanics.
20 CHAPTER 2. FLUID MECHANICS AND ISSUES IN INTERACTION PROBLEMS
This condition, for a determined set of time-integration coefficients αn−1 , αn , αn+1 , entirely
defines the weight coefficients ωk . It can be shown that to satisfy the DGCL, the coefficients
and the evaluation configurations must be chosen such as illustrated in Figure 2.4.
xn+1 − xn
ẋsk =
∆t
αn+1 αn−1 ∆tn
ω1 = ω2 = ω3 = ω4 = −
2 2∆tn−1
1 1 1 1
δ1 = 1− √ δ2 = 1+ √
2 3 2 3
For every integration scheme (thus not just for the four-point rule), one should choose
weighting coefficients so that the DGCL is satisfied since it has been shown that if the DGCL
is satisfied, integration schemes that are time accurate of order p on non-moving meshes will be
at least of order p − 1 on moving meshes.
where the first set of equation are the Finite Volume discretized equation, M , C, K are the
structural mass, damping and stiffness matrices, Fapplied , Ff luid are the external and fluid forces
applied to the structure and where K̃ is the stiffness matrix of the pseudo-structure, xo being
the mesh points in the fluid domain that have to follow the boundary points defined by xb .
2.4.1 Matching of meshes 21
The solution has to satisfy the boundary conditions on the fluid-structure interface bound-
ary:
where nb is the normal on the fluid-structure boundary interface and u is the structural dis-
placement field on the boundary.
The fluid is time-integrated using the techniques shortly described in section 2.3.3 while the
structure is typically integrated using Newmark’s integration scheme (mid-point/trapezoı̈dal
rule).
In this section we will briefly introduced three specific issues relevant to the implementation
of simulation solving these coupled fluid-structure equations:
• matching the fluid and structural mesh
• defining the mesh motion in the fluid grid (the pseudo-structure).
• time-integration procedures for the coupled problem.
These issues will be further discussed in later chapters.
Let us first answer the compatibility question: which displacement uF to impose on the fluid
nodes, given a structural displacement uS ?. In Figure 2.6 (left) it is shown that, using the shape
22 CHAPTER 2. FLUID MECHANICS AND ISSUES IN INTERACTION PROBLEMS
functions in the structural finite elements, one can reconstruct the displacement at the very
location of the fluid node and then enforce
X X
uF,i = Nj (χi )uS,j = ci,j uS,j (2.78)
j=dof of matching element j
where we defined cij the value of the shape function associated to the structural node j at the
location χi of the fluid vertex i. Obviously when all (translational) degrees of freedom of the
structural elements are unity, the displacement inside the element is also unity and thus we
have the property
X
cij = 1 (2.79)
j=dof of matching element
Now let us answer the equilibrium question: how to apply the fluid forces (pressure and
viscous) to the structure when there are only known at discrete locations in the fluid?. A first
way would be to interpolate the pressure across the fluid nodes and apply the interpolated
pressure to the structural boundary. Such an approach does however not guaranty that the
energy is properly transferred through the interface.
So let us call fS,j the structural forces that one should apply to the structural degrees of
freedom j in order to represent the fluid forces. The work produced by those forces over a small
structural displacement in the entire structure is equal to
X
δWS = fS,j δuS,j
jǫS
On the other hand, the work produced by the fluid forces on the fluid cell which vertices have
been moved according to (2.78) is
Z
δWF = (−pnb + τ nb ) δuF dS
SF
X XX
= fF,i δuF,i = fF,i (cij δuS,j )
iǫF iǫF j
Remembering that cij are the value of the structural shape functions pertaining to node j
evaluated at the position of the fluid vertices i, this last relation means that the fluid forces
should be considered as concentrated forces applied on the structural element at the location
of the fluid vertices. This is illustrated in Figure 2.6 (right).
A first idea is to associate a lineal spring to each edge of a cell (Figure 2.7). The spring
stiffnesses on the edges can be chosen as
1
kedge = (2.81)
lengthedge
in order to prevent the edge to have zero length. Nevertheless, as illustrated in Figure 2.7, this
approach breaks down for large mesh deformation since the edges can snap through.
Another approach consists in defining torsional springs between edges and choose the torsion
stiffness to be inversely proportional to the corresponding cell area (see Figure 2.8 for 2D and
3D implementation). Choosing the torsional stiffness at node i of triangle ijk to be
length2edge ij length2edge ik
kiijk = (2.82)
Area2ijk
24 CHAPTER 2. FLUID MECHANICS AND ISSUES IN INTERACTION PROBLEMS
one ensures that nodes will not collide and that the area will not go to zero (hence no snap
through).
• Solving the structural problem (often linear or smoothly non-linear) is usual much eas-
ier than solving the non-linear convective fluid equations. Hence mixing the fields in a
monolithic scheme usually prohibits to solve the structural part at low cost.
• The time-scale of the phenomena appearing in the fluid and in the structure are often
very different. Moreover explicit time-integration schemes are often used in CFD whereas
implicit schemes are much more efficient for the structure where smooth dynamics is
usually taking place. Hence the time-step required to solve the fluid is usually significantly
smaller than the required structural time-step. If a monolithic scheme is adopted, both
fields have to be integrated with the smallest time-step, thereby incurring unnecessary
computing cost.
Therefore, unless the structure is described as a simple model (e.g. a modal model), it is
often preferred to solve the fluid and structure separately and to iterate from one field to the
other: staggered schemes.
One simple staggering procedure is defined in Figure 2.9: at a given time step tn , one starts
from a known structural displacement un and assumes that at tn+1 the fluid mesh has to
2.4.3 Time integration of coupled models 25
Figure 2.9: Basic staggering procedure between the structure, the fluid and the pseudo-structure
of the fluid mesh
5
Note that if the time-step for computing the new fluid state is smaller than the structural time-step, several
intermediate states will be computed between W n and W n+1 .
26 CHAPTER 2. FLUID MECHANICS AND ISSUES IN INTERACTION PROBLEMS
Most real life problems involve time-dependent phenomena. Examples are vortex shedding be-
hind an airfoil at high angle of attack, turbulent flows and flutter behaviour. These phenomena
can be described by the full Navier-Stokes equations in differential form
where u = [u, v, w]T is the velocity vector, ν the kinematic viscosity, ρ the density, p the
pressure and F = [Fx , Fy , Fz ]T additional source terms. Due to the ∂t term time-dependence is
introduced. For 1D the Navier Stokes equations can be written in non-conservative form as
mass ∂t ρ + ∂x ρu = 0, (3.5)
momentum ∂t u + u∂x u − ν∂x2 u 1
= − (∂x p − F ),
ρ (3.6)
| {z } | {z } | {z }
Advection Dif f usion Other
The momentum equation can be split to three parts. The first term describes advection phonom-
ena, such as running waves. The second term describes diffusion phenomena and the last term
describes other phenomena introduced for example by pressure differences and gravity forces.
Because the full Navier Stokes equations are rather difficult to solve we start with the introduc-
tion of two simpel model problems. These model problems will be used in the remainder of the
notes to investigate properties of time-discretization methods.
ut + aux = 0, ∀ x ∈ Ω, (3.7)
(initial) u(x, 0) = f (x), (3.8)
(boundary) u(x, t) = g(x, t) ∀ x ∈ ∂Ω. (3.9)
28 CHAPTER 3. TIME DEPENDENT PROBLEMS
Here we use the short-hand notation ut = ∂t u and ux = ∂x u and Ω is the domain of computation
with ∂Ω its boundaries. If periodic boundary conditions are used and the initial form of u at
t = 0 is a known function f (x) we can obtain an analytical solution for this equation at t > 0
u(x, t) = f (x − at). (3.10)
An example of the time evolution of the solution is shown in Figure 3.1 where we start with
2
the initial condition u(x, 0) = e−30(x−0.5) and a = 1. You can clearly see that the distortion is
moving with speed at to the right (when a = −1 it moves with the same speed but then to the
left).
1 1
t=0 t=0
0.9 t = 0.1 0.9 t = 0.01
t = 0.2 t = 0.02
0.8 0.8
0.7 0.7
0.6 0.6
0.5 0.5
u
u
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x x
Figure 3.1: Solution to the advection equa- Figure 3.2: Solution to the diffusion equation
2 2
tion with u(x, 0) = e−30(x−0.5) and a = 1 at with u(x, 0) = e−30(x−0.5) and ν = 1 at t =
t = 0, t = 0.1 and t = 0.2. 0, t = 0.01 and t = 0.02.
The coefficients am are just the Fourier sine expansions of the initial solution u(x, 0) = f (x)
and therefore given by
Z 1
am = 2 f (x) sin(mπx) dx. (3.15)
0
3.1. TIME INTEGRATION 29
This final result may be regarded as an exact analytical solution, but it is more like a numer-
ical approximation. First the coefficients am can only be found exactly for special and simple
functions f (x) and more generally require numerical integration. Secondly, we can only sum a
finite number of terms of the infinite series. However generally a few terms of the series will be
quite sufficient, as the series converges extremely rapidly.
The time evolution of the solution with the same initial condition as for the linear advection
equation is shown in Figure 3.1 where we chose ν = 1. You can clearly see that the bump is
smoothed out over the domain, while the area under the solution remains constant (no total
heat or mass loss).
When we apply Forward Euler to the heat equation (3.11) with a central difference scheme in
space we obtain
and more complicated strategies are required to solve for un+1 than with explicit methods.
The simplest example is Backward Euler (or implicit Euler), where forward difference dis-
cretization in time is used and other terms are evaluated at the new time level tn+1 . Equation
(3.16) is then approximated by
un+1 − un
= F (un+1 ) or un+1 = un + ∆tF (un+1 ). (3.24)
∆t
When we apply Forward Euler to the heat equation (3.11) with a central difference scheme in
space we obtain
−ναun+1 n+1
i+1 + (1 + 2να)ui − ναun+1 n
i−1 = ui . (3.26)
To obtain the solution at the new time level a system of equations has to be solved.
θ-method Another well known example is the θ-method. In this method terms at the current
and the next time level are combined:
un+1 − un
= (1 − θ)F (un ) + θF (un+1 ), (3.27)
∆t
or
un+1 = un + ∆t (1 − θ)F (un ) + θF (un+1 ) , (3.28)
with 0 ≤ θ ≤ 1. As long as θ is larger than zero the method is implicit and for θ = 0 the method
is explicit. For the special values θ = 0 and θ = 1 we obtain:
• θ = 0: explicit Euler.
• θ = 1: implicit Euler.
3.1.3 Higher order time discretization 31
When we apply the θ-method to the heat equation (3.11) with a central difference scheme in
space we obtain
−ναθun+1 n+1
i+1 + (1 + 2ναθ)ui − ναθun+1 n n n n
i−1 = ui + να(1 − θ)(ui+1 − 2ui + ui−1 ), (3.30)
Truncation error
The order of accuracy of a scheme can be derived by looking at the truncation error of the
discretization. The truncation error is obtained by substituting the exact solution into the
numerical scheme and using Taylor series expansion to compare it with the original differential
equation. We give here an example for the derivation of the order of the forward Euler method
applied to the heat equation (3.21), with with ν = 1. Substituting the exact solution u(x, t) in
this equation gives the truncation error
u(x, t + ∆t) − u(x, t) u(x + ∆x, t) − 2u(x, t) + u(x − ∆x, t)
TFE = − . (3.31)
∆t (∆x)2
Substituting Taylor series expansions of the form
Since the exact solution satisfies the original differential equation ut − uxx = 0, the truncation
error simplifies to
2
TFE = 12 ∆t utt − 1
12 (∆x) uxxxx + O(∆t)2 + O(∆x)4 . (3.36)
We can see that when ∆t and ∆x go to zero the truncation error goes to zero and therefore the
discretization scheme is consistent. Consistency is one of the major requirements for a numerical
scheme. The term 12 ∆t utt is the leading error term in ∆t. When we divide ∆t by two, this term
32 CHAPTER 3. TIME DEPENDENT PROBLEMS
also becomes 2 times smaller and therefore the method is called first order accurate in time.
This means that the method converges linearly with ∆t. If we look at the error in space we see
1
that the term 12 (∆x)2 uxxxx is the leading error term. Here the error becomes 4 times smaller
when ∆x is divided by 2 and therefore the method is called second order accurate in space
which means that it converges quadratically with ∆x.
For the backward Euler method (3.25) the truncation error looks slightly different:
u(x, t + ∆t) − u(x, t) u(x + ∆x, t + ∆t) − 2u(x, t + ∆t) + u(x − ∆x, t + ∆t)
TBE = − . (3.37)
∆t (∆x)2
This is equivalent to
by substituting t = t − ∆t, which is easier to evaluate because there are no terms with both
∆t and ∆x anymore. Substituting Taylor expansions gives the following expression for the
truncation error
or
2
TBE = − 12 ∆t utt − 1
12 (∆x) uxxxx + O(∆t)2 + O(∆x)4 . (3.40)
This method is also consistent and first order accurate in time and second order accurate in
space.
The idea of higher order methods is to construct a scheme in which all lower order terms
cancel to make the method converge faster to the exact solution. A simple example is given
by the θ-method (3.29). The θ-method is actually a linear combination of the forward and
backward Euler method:
The truncation error of the θ-method is therefore also a linear combination of the truncation
errors of the forward and backward Euler method
If we substitute the truncation errors of the forward (3.36) and backward Euler (3.40) method
this becomes
1 2
2
Tθ = (1 − θ) 2 ∆t utt − 1
12 (∆x) uxxxx + θ − 12 ∆t utt − 1
12 (∆x) uxxxx + O(∆t)2 + O(∆x)4
(3.43)
2
= ( 12 − θ)∆t utt − 1
12 (∆x) uxxxx + O(∆t)2 + O(∆x)4 (3.44)
When θ 6= 12 the method is only first order accurate in time. However, when we choose θ = 12
the ∆t term disappears and the method becomes second order accurate in time. The θ-method
with the special choice θ = 12 is known as the Crank-Nicolson scheme, or the trapezoidal rule.
3.1.3 Higher order time discretization 33
Multi-step methods
In multi-step methods not only the solution at tn is used to obtain the new solution at tn+1 but
also previous solutions at tn−1 , tn−2 , etc. With a clever combination of time levels you can make
a scheme of arbitrary order of accuracy (just write out the truncation error with the Taylor
series and make sure that all the lower order terms cancel).
The simplest example is the three level method (Leapfrog method) given by
un+1 − un−1
+ F (un ) = 0. (3.45)
2∆t
However, this method has severe stability restrictions as we will show in the last section.
Other examples are obtained with the implicit backward differencing formula:
" K
#
1 X
(αk un−k
i ) + F (un+1 ) = 0 (3.46)
∆t
k=−1
3un+1
i − 4uni + un−1
i
un+1 n+1
i+1 − 2ui + un+1
i−1
− 2
= 0. (3.47)
2∆t (∆x)
One of the disadvantages of the multi-step methods is the start-up problem. In order to calculate
the solution at t1 you need not only the initial solution at t0 , but also the solution at t−1 or
earlier and these solutions are not known. This can be fixed by using an appropriate lower order
multi-step method in the first time-steps.
Multi-stage methods
Multi-stage methods do not use more previous time levels, but the solution at intermediate time
levels (stages) tn+α , 0 < α < 1 to obtain higher order accuracy in time. A linear combination
of these solutions at intermediate stages is used to obtain a higher order accurate solution at
tn+1 . The computation of the solution at the intermediate time levels can be done either in an
implicit or explicit way.
The intermediate stage solutions u(k) in Explicit Runge-Kutta (ERK) schemes are calculated
using the following formula
k−1
u(k) − un X
= akj F (u(j) ). (3.48)
∆t
j=1
where the sum is taken over all previous stages. The update to the new time level is then
obtained by
m
un+1 − un X
= bj F (u(j) ). (3.49)
∆t
j=1
34 CHAPTER 3. TIME DEPENDENT PROBLEMS
An example is the explicit RK3 method (third order accurate in time) which proceeds as follows:
u(1) = un , (3.50)
(2) n (1)
u = u + ∆t F (u ), (3.51)
h i
u(3) = un + ∆t 14 F (u(1) ) + 14 F (u(2) ) , (3.52)
h i
un+1 n 1 (1) 1 (2) 2
= u + ∆t 6 F (u ) + 6 F (u ) + 3 F (u ) . (3)
(3.53)
In Implicit Runge-Kutta (IRK) schemes a sum is taken over all stages including the current
stage to obtain the intermediate stage solutions
k
u(k) − un X
= akj F (u(j) ). (3.54)
∆t
j=1
This means that you have to solve a system of equations at each intermediate stage. The update
to the new time level (3.49) is then equal to that of the explicit schemes.
An example is the implicit RK2 method (third order accurate in time) which proceeds as
follows:
√
u(1) = un + γ∆t F (u(1) ), γ = 12 + 16 3 (3.55)
h i
u(2) = un + ∆t (1 − 2γ)F (u(1) ) + γF (u(2) ) , (3.56)
h i
un+1 = un + 12 ∆t F (u(1) ) + F (u(2) ) . (3.57)
Contrary to multi-step methods, RK-methods are self-starting: for the solution at t1 they only
need information at t0 . You pay for this with additional work: you have to perform an evaluation
or solve a system of equations at each intermediate stage. Multi-stage methods can also be made
of arbitrary order: to obtain higher order methods you have to evaluate more stages, which means
higher computational costs. However by creating higher order methods you hope to be able to
take larger time steps and obtain the same accuracy, which leads to lower computational costs.
The next example gives an idea of the trade-off between more stage evaluations and higher
accuracy.
Example: Flow past a cylinder As an example for the performance of higher order methods
we look at viscous flow past a cilinder. The configuration of the problem and a snapshot of the
velocity distribution is shown in Figure 3.3. The computation is performed with implicit time
integration methods of different order of accuracy (BDF2: second order, IRK3: third order,
IRK4: fourth order). The error in time is computed using the L2 -norm of the density. The
’temporally exact’ solution is obtained with a high order scheme and a very small time step.
The loglog-plot of the error of the methods against the time step is shown in Figure 3.4. The
slope of the line represents the order of the method. You can see that the higher the order of
the method, the more accurate the solution becomes at a given time step and the steeper the
slope. The meaning of Figure 3.4 becomes more clear if you look at the solutions obtained with
a first order method and a third order method for different values of the time step. Examples are
shown in Figures 3.6 and 3.7, respectively. You can see that the third order method converges
much faster to the exact solution than the first order method and for the same value of the time
step it is more accurate.
But what about the efficiency of the methods? We know that higher order methods are
more expensive. Therefore we also plotted the amount of work required by the method (how
3.1.3 Higher order time discretization 35
−1
BDF2
−1 IRK3
−2 IRK4
−2
−3
−3
log(error)
−4
log(error)
−4
−5
−5
−6
−6
−7
−7 BDF2
IRK3
IRK4
−8 −8
−1.6 −1.4 −1.2 −1 −0.8 −0.6 −0.4 −0.2 0 0 0.5 1 1.5 2 2.5
log(∆ t) log(work)
Figure 3.4: Error vs. timestep for BDF2, Figure 3.5: Error vs. work for BDF2, IRK3
IRK3 and IRK4 to obtain the order of the and IRK4 to obtain the efficiency of the
methods. methods.
0.4 ∆t = 1 0.4 ∆t = 1
∆t = 1/2 ∆t = 1/2
∆t = 1/4 ∆t = 1/4
∆t = 1/8 ∆t = 1/8
∆t = 1/16 exact
0.2 ∆t = 1/32 0.2
∆t = 1/64
∆t = 1/128
displacement
displacement
∆t = 1/256
∆t = 1/512
0 exact 0
-0.2 -0.2
-0.4 -0.4
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Time Time
Figure 3.6: Solutions obtained with a first or- Figure 3.7: Solutions obtained with a third
der method with different timesteps. order method with different timesteps.
many times the system of equations has to be solved) against the error in a loglog-scale. The
36 CHAPTER 3. TIME DEPENDENT PROBLEMS
results are shown in Figure 3.5. The closer a line is to the lower right corner, the more efficient
the method is. We can see that higher order methods can indeed be more efficient, especially
when the required accuracy is higher.
3.2 Stability
There are several forms of instability you can encounter when solving partial differential equa-
tions. Physical instability as for instance transition from laminar to turbulent flow or mathe-
matical instability. However, in this section we consider the numerical instability resulting from
discretizing the partial differential equation. A system is numerically stable if the error in the
time dependent problem remains bounded for t → ∞. This does not necessarily imply that the
solution itself remains bounded. Take for example the solution to the differential equation
where a stable scheme is used to discretize the equation, then both the exact solution, uex = ect ,
and the approximate solution, uh , go to infinity for t → ∞, so the equation is mathematically
instable. But if the scheme is numerically stable then the error uex − uh remains bounded. In
this way small errors (due to roundoff errors or in the initial/boundary conditions) lead to
accordingly small errors in the numerical solution.
An example of a stable and unstable solution in time is shown in Figure 3.8.
1.5
0.5
0
u
−0.5
−1
−1.5 exact
stable
unstable
−2
0 1 2 3 4 5
t
Figure 3.8: Stable (bold line) and unstable (dashed line) solution in time.
The main reason why we are interested in the stability analysis of a scheme is stated in the
famous Lax Theorem:
Theorem 3.2.1 (Lax Theorem) For a well-posed initial value problem and a consistent dis-
cretization scheme, stability is a necessary and sufficient condition for convergence.
This fundamental theorem shows that in order to establish convergence of a numerical scheme
i.e. to show that the approximate solution will converge to the exact solution for ∆t → 0,
∆x → 0 two requirements have to be met:
To obtain a stable scheme the spatial discretisation must be stable, otherwise no time-
integration method will lead to an overall stable algorithm. There are various ways to analyze
the stability of a numerical scheme, but in the next sections we restrict ourselves to the von
Neumann analysis and Matrix method. Also the concept of phase and diffusion error will be
introduced.
φ ≡ κj ∆x, (3.61)
and covers the domain (−π, π) in steps of π/N . The region around φ = 0 corresponds to the
low frequencies while the region close to φ = π is associated with the high-frequency range of
the spectrum.
Since we deal with linear schemes we can look at each individual harmonic separately. The
basic procedure is now to impose a spatial harmonic as an initial value on the mesh and ask
the question: Will its amplitude grow or decay in time?
So impose
To obtain numerical stability the amplitude of any harmonic E n must not grow in time. There-
fore there must yield
E n+1
|G| ≡ ≤ 1. (3.63)
En
For this reason the quantity G, defined by
E n+1
G= , (3.64)
En
38 CHAPTER 3. TIME DEPENDENT PROBLEMS
is called the amplification factor . The problem then becomes to solve for the G’s produced by
any given method. If equation (3.63) is satisfied for all possible combinations of parameters the
method is called unconditionally stable. If the method is stable only for certain combinations of
parameters the method is called conditionally stable and if it never satisfies (3.63) the method
is unconditionally unstable. The procedure can best be explained by some examples.
Example 1: Forward Euler Consider the finite-difference approximation of the heat equa-
tion (3.11) discretized by forward Euler in time and central discretization in space:
∆t
un+1
i = uni + να(uni+1 − 2uni + uni−1 ), α= . (3.65)
(∆x)2
Substituting (3.62) into (3.65) and dividing by eIiφ gives the relation
The discretization is only stable when να ≤ 12 or ∆t ≤ (∆x)2 /(2ν), because then |G| ≤ 1 for
all values of φ. This means that the method is only conditionally stable.
Example 2: Backward Euler We now discretize the heat equaton (3.11) by backward Euler
in time and central discretization in space.
−ναun+1 n+1
i+1 + (1 + 2να)ui − ναun+1 n
i−1 = ui , (3.68)
Substituting (3.62) into (3.65) and dividing by eIiφ gives the relation
where ω = ω(κ) is a complex function of the real number κ, representing the numerical dispersion
relation. The function Ê(κ) is obtained from the Fourier decomposition of the initial solution,
since for u(x, 0) = f (x) at t = 0 we have
Z L
1
Ê(κ) = f (x) e−Iκx dx. (3.72)
2L −L
3.2.2 Phase and diffusion error 39
where the exact dispersion relation ω̃ = ω̃(κ) can be obtained from the differential equation
as a solution of the equations in eigenspace. The approximate relation between ω and κ is the
numerical dispersion relation and can be derived from the amplification factor G.
From the definition of G (3.64) together with (3.71) we can derive
E n = Gn E 0 = Gn Ê = e−Iωn∆t Ê (3.75)
G = e−Iω∆t . (3.76)
will allow us to investigate the nature of the numerical errors. Since ω is a complex function the
amplification factor can be separated into an amplitude |G| and a phase Φ. With
ω = ξ + Iη (3.78)
we have
where
where Re(G) is the real part of G and Im(G) the imaginary part. The phase angle is actually
the angle between the imaginary and real part of G, so
Im(G) −1 Im(G)
tan(Φ) = ⇒ Φ = tan . (3.83)
Re(G) Re(G)
A similar decomposition can be performed for the exact solution leading to the exact amplitude
and phase
1 1.2
exact exact
0.9 FE art. diss.
1
0.8
0.7 0.8
0.6
0.6
0.5
u
u
0.4
0.4
0.3
0.2
0.2
0
0.1
0 −0.2
0 0.5 1 1.5 2 0 0.5 1 1.5 2
x x
Figure 3.9: Artificial dissipation with For- Figure 3.10: Artificial dispersion with Crank-
ward Euler α = 12 . Nicolson α = 5.
The error in amplitude also called the diffusion or dissipation error or artificial dissipation, is
defined by the ratio of the computed amplitude to the exact amplitude:
|G|
ǫD = . (3.85)
eη̃∆t
An example of artificial dissipation is shown for the linear advection equation (3.7) discretized
with forward Euler in Figure 3.9. You can clearly see that the amplitude is lower than it should
be. Another error that you can encounter is the error in phase, the so called phase or dispersion
error or artificial dispersion. This means that the solution runs faster or slower than it should.
An example is given in Figure 3.10 for the Crank-Nicolson scheme where you can see that the
wave is running behind the exact solution. This error can be defined as the difference
ǫφ = Φ − Φ̃ (3.86)
suitable for pure parabolic problems, such as the heat equation (3.11), where Φ̃ = 0. There is a
leading phase error when Φ > Φ̃ and a lagging phase error when Φ < Φ̃.
For convection dominated problems, such as the linear advection equation, the definition
ǫφ = Φ/Φ̃ (3.87)
is better adapted. There is a leading phase error when ǫφ > 1 and a lagging phase error when
ǫφ < 1. For the linear advection equation (3.7) the exact phase is given by
∆t
Φ̃ = aκ∆t = αφ, with α=a , (3.88)
∆x
because the solution is a single wave propagating with velocity a.
and
1 1
GBE = −Iφ
= (3.90)
1 + α(1 − e ) (1 + α) − α[cos(φ) − I sin(φ)]
∆t
respectively, with α = a ∆x . The error in the amplitude is measured by
ǫD = |G| (3.91)
The dispersion error (in this case equal to the amplification factor) of the two methods is
plotted against the phase angle φ in Figures 3.11 and 3.12 for different values of α. You can
2 1
εD
1 0.6
0.8 α = 0.5
0.5
0.6 α = 0.75 α = 0.75
0.4
0.4 α = 0.25 α=1
α = 0.5 0.3 α = 1.25
0.2
α = 1.5
0 0.2
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
φ/π φ/π
Figure 3.11: Error in amplitude ǫD of For- Figure 3.12: Error in amplitude ǫD of Back-
ward Euler for different values of α. ward Euler for different values of α.
see that forward Euler is unstable when α > 1 because then the diffusion error is larger than
1. For backward Euler you can see that the larger the value for α (so the larger the time
step) the smaller the diffussion error and the more the solution is damped, especially for the
high wavenumbers (large value of φ). This means that high-frequency errors (and solutions) are
damped more than low-frequency errors (and solutions).
The exact phase for this problem is given in (3.88). The phase error of the two methods is
plotted against the phase angle φ in Figures 3.11 and 3.12 for different values of α. You can
see that backward Euler has a lagging phase error and the phase error of forward Euler is more
dependent on the value of α and the frequency of the wave.
The effects of artificial dissipation and dispersion The effects of artificial dissipation and
dispersion can be described as good, bad and ugly. In accordance with the effects of the natural
dissipation, the numerical dissipation, as a result of the amplification factor |G| < 1, will reduce
the amplitude of the waves. This is good, when time-stepping is used to obtain steady state
solutions, because it ensures stability and reduces the error, resulting in faster convergence.
This is bad especially for unsteady computations, since it also reduces the amplitude of the
dependent flow variables. Adding too much numerical dissipation will give unphysical results.
The dispersion is plain ugly. In most cases, moderate dispersion can be lived with, since the
main features of the flow will be fine. However, it might cause instabilities or simply totally
wrong results, when coupled with some other physical effects.
42 CHAPTER 3. TIME DEPENDENT PROBLEMS
3 1
α = 0.25 α = 0.25
α = 0.5 0.9 α = 0.5
2.5 α = 0.75 α = 0.75
α=1 0.8 α=1
α = 1.25 α = 1.25
α = 1.5 0.7 α = 1.5
2
0.6
φ
φ
1.5 0.5
ε
ε
0.4
1
0.3
0.2
0.5
0.1
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
φ/π φ/π
Figure 3.13: Error in phase ǫφ of Forward Figure 3.14: Error in phase ǫφ of Backward
Euler for different values of α. Euler for different values of α.
∂t u = L(u) (3.92)
where L is the spatial linear differential operator, with initial and boundary conditions
After spatial discretisation this can be written as a linear matrix system as follows
∂t u = Su + r (3.95)
where u is the vector with values ui , S the matrix with spatial discretisation including boundary
conditions and r contains boundary values. In the next sections we will consider stability for
this system in space and time separately.
Stability in space
To obtain a stable scheme the spatial discretisation must be stable, otherwise no time-integration
method will lead to an overall stable algorithm. To perform stability analysis in space you have
to look at the eigenvalues λj of matrix S. The spatial discretization is stable when the real part
of all eigenvalues is smaller or equal to zero, Re(λj ) ≤ 0 for all j. Only for simple matrices an
analytic expression is possible for all the eigenvalues. In the other cases you have to obtain them
numerically.
The location of the eigenvalues depend on the problem to be solved:
• For diffusion-dominated flows, the λ eigenvalues tend to lie along the negative real axis.
3.2.3 Matrix method 43
• For periodic convection-dominated flows, the λ eigenvalues tend to lie along the imaginary
axis.
The location of the eigenvalues is important to decide what kind of time-integration method
can be used.
∂u ui − ui−1
≈ . (3.97)
∂x ∆x
This results in the following semi-discrete matrix system
1 0 0 ··· 0 0 g(t)
−1 0 1 · · · 0
0 0
a 0 −1 0 · · · 0
0
a
0
∂t u = . .. .. . . .. u+
.. .. , (3.98)
2∆x .. . . . . . 2∆x .
0 0 0 ··· 0 1 0
0 0 0 ··· −2 2 0
| {z } | {z }
S r
where we will use the values a = 0.5 and ∆x = 0.02. For this case the eigenvalues of matrix S
are plotted in Figure 3.15. You can see that for all the eigenvalues the real part is smaller or
equal to zero, which means that the discretization is stable. When we interchange the boundary
conditions, Dirichlet for the outflow and first order for the inflow the semi-discrete matrix system
has the following form
2 −2 0 · · · 0 0 0
−1 0 1 · · · 0 0 0
a 0 −1 0 · · · 0 0 a
du 0
= .. .. .. . . .. .. u + . . (3.99)
dt 2∆x . . . . . . 2∆x ..
0 0 0 ··· 0 1 0
0 0 0 ··· 0 1 g(t)
| {z } | {z }
S Q
44 CHAPTER 3. TIME DEPENDENT PROBLEMS
30 30
20 20
Im(λ) 10 10
Im(λ)
0 0
−10 −10
−20 −20
−30 −30
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
Re(λ) Re(λ)
Figure 3.15: Eigenvalues for the stable dis- Figure 3.16: Eigenvalues for the unstable dis-
cretization. cretization.
The eigenvalues of matrix S are plotted in Figure 3.16. The discretization is now clearly unstable,
because all the eigenvalues have a non-negative real part.
The effect of the (un)stability of the space discretization will be shown by discretizing the
system with the unconditionally stable (see the next section) backward Euler scheme. We will
use ∆t = 0.01 and g(t) = 0.1. The results for the stable and unstable space discretization
are shown in Figures 3.17 and 3.18 respectively. The small wiggles in the left of the domain
u u
ex ex
BE BE
1 1
0.8 0.8
0.6 0.6
u
0.4 0.4
0.2 0.2
0 0
−0.2 −0.2
0 0.5 1 1.5 2 0 0.5 1 1.5 2
x x
Figure 3.17: Solution for the stable discretiza- Figure 3.18: Solution for the unstable dis-
tion. cretization.
are resulting from the odd-even decoupling caused by the central space discretization: for the
evaluation of the solution in the ’even’ node ui you only take into account the values of the
’odd’ nodes ui−1 and ui+1 and vice versa, so the solutions between odd and even nodes can
diverge (this is beyond the scope of this lecture). In the left figure you can clearly see that the
boundary value g(t) = 0.1 at the left boundary is ’flowing’ into the domain. However, imposing
this same value on the right boundary makes the system unstable which is visible through the
large wiggles which enter the domain.
Equations in eigenspace To be able to look at the stability of the time discretization method
separately we want to define an equation which is representative for all space discretizations.
3.2.3 Matrix method 45
For this reason we will transform the equations to eigenspace. We start again with the linear
matrix system (3.95) and assume that matrix S has a complete eigensystem (it has a complete
set of linearly independent eigenvectors). The eigenvector matrix X is obtained by filling the
columns with these eigenvectors. This matrix has the property that
X −1 SX = Λ, (3.100)
X −1 ∂t u = X −1 SX · X −1 u + X −1 r. (3.101)
We now introduce the new variable w = X −1 u and q = X −1 r, then we can rewrite (3.101) as
w = Λw + q, (3.102)
which is a system of ordinary decoupled differential equations where each equation can be solved
separately. As a result each equation has the linear form
wt = λw + q (3.103)
with w and q elements of w and q, respectively, and λ the corresponding eigenvalue of S. When
a solution is expressed in terms of w, it is said to be in eigenspace. The numerical solution to
a set of linear ODE’s is entirely equivalent to the solution obtained when the equations are
transformed to eigenspace, solved there in their uncoupled form, and then transformed back
to real space. (However, in most practicle computations the generation of the equations in
eigenspace is too much work.) The importance of this concept resides in its message that we
can analyze time-marching methods by applying them to a single, uncoupled equation and our
conclusions will apply in general.
The goal in our analysis is to study typical behaviour of general situations, not particu-
lar problems. Therefore we replace r with its corresponding Fourier term which leads to the
representative equation
wt = λw + aeµt , (3.104)
which can be used to evaluate various time-marching methods. In such evaluations the parame-
ters λ and µ must be allowed to take the worst possible combination of values that might occur.
In the next section we will not consider the term aeµt coming from the boundary conditions
resulting in the evaluation of the simple linear equation of the form
ut = λu. (3.105)
We will use this linear equation to evaluate the performance of several time integration methods.
Stability in time
When the linear equation (3.105) is discretized with a time integration method we can perform
von Neumann stability analysis to obtain the operational form of this equation which has the
general form
P (G) = 0. (3.106)
46 CHAPTER 3. TIME DEPENDENT PROBLEMS
The term P (G) is a polynomial in the amplification factor G referred to as the characteristic
polynomial.
The m roots σm of the characteristic polynomial P (G) say something about the stability
of a time-marching method. A time-marching method is stable if and only if all the roots are
smaller than one, because then errors will not grow in time. Since each eigenvalue λk has a
characteristic polynomial we must have
k
|σm |≤1 for all m and k. (3.107)
Example 1: Forward Euler As a first example take the forward Euler time-discretization.
Discretizing (3.105) results in
G − (1 + λ∆t), (3.110)
σ = 1 + (∆tλ). (3.111)
The stability region is shown in Figure 3.19. You can see that it covers only a part of the left
Im( ∆ t λ)
= stable
= unstable
2
−2 −1 1 2 Re( ∆ t λ)
−1
−2
half of the complex plane and this means that Forward Euler is only conditionally stable. In
the following two ways this is typical of all stability contours of explicit methods:
3.2.3 Matrix method 47
1. The contour with |σ| = 1 encloses a finite portion of the left-half of the complex λ∆t-plane.
2. The region of stability is inside this contour, and therefore the method is conditionally
stable.
This means that for a given spatial mesh you have to restrict the timestep to get all the values
of λ∆t in the stability region. When an eigenvalue λ lies on the imaginary axes (as for the
convection equation with central differences in space), the overall algorithm can never become
stable, because you have to take ∆t = 0 to get (λ∆t) in the stability region.
An example is shown in Figures 3.20 and 3.21. By choosing a different value for the time-step
exact exact
FE FE
1 1
0.8 0.8
0.6 0.6
u
u
0.4 0.4
0.2 0.2
0 0
−0.2 −0.2
0 0.5 1 1.5 2 0 0.5 1 1.5 2
x x
Figure 3.20: Stable solution with the Forward Figure 3.21: Unstable solution with the For-
Euler scheme for ∆t = 0.3. ward Euler scheme for ∆t = 0.4.
(1 − λ∆t)G − 1 = 0, (3.114)
1
σ= . (3.115)
1 − (∆tλ)
The stability region is shown in Figure 3.22 which covers the whole left half of the complex
plane and this means that backward Euler is unconditionally stable. However, just because a
method is implicit does not mean that it is unconditionally stable. Two examples of this are
the Adams-Moulton 3rd-order method and the 4th order Milne method (see reference [17]).
48 CHAPTER 3. TIME DEPENDENT PROBLEMS
Im( ∆ t λ)
= stable
= unstable
2
−2 −1 1 2 Re( ∆ t λ)
−1
−2
G2 − 2(λ∆t)G − 1 = 0. (3.118)
and the stability region for which for both roots yield |σ| ≤ 1 is shown in Figure 3.23 which
only covers the imaginary axis. You can see that the Leapfrog scheme can only be applied in
Im( ∆ t λ)
= stable
= unstable
2
−2 −1 1 2 Re( ∆ t λ)
−1
−2
special cases, because the method is only stable for imaginary values of λ. More characteristic
polynomials and stability regions for different schemes can for example be found on page 122
and 131–133 of [17].
Chapter 4
In FSI computations it is required that pressure loads are transmitted from the fluid side of the
fluid-structure interface to the structural nodes on that interface. Also, once the motion of the
structure has been determined, the motion of the fluid mesh points on the interface has to be
imposed. In FSI simulations it is usually not desirable to generate matching meshes at the fluid-
structure interface, because the flow generally requires a much finer mesh than the structure.
In additition, also different teams may take care of the different physical domains. This means
that the discrete interface between the domains may not only be non-conforming, but there
can also be gaps and/or overlaps between the meshes. The exchange of data over the discrete
interface becomes then far from trivial. In Figure 4.1 a 2D example of a non-matching discrete
interface between a flow and structure domain is shown. When the meshes are non-matching,
an interpolation/projection step has to be carried out to enable transfer of information between
the two domains. In literature different methods can be found to transfer data between non-
matching meshes, such as nearest neighbour interpolation [32], projection methods [8, 24, 26]
and methods based on interpolation by splines [3, 29, 30].
The general opinion is that energy should be conserved
over the interface leading to a conservative coupling approach Fluid
[14]. This approach is based on the global conservation of vir-
tual work over the interface, where one transformation ma- Interface
trix performs both the transfer of displacements and pressure
loads between the two discrete interfaces. However, for a gen-
Structure
eral coupling method this can lead to unphysical oscillations
in the pressure forces received by the structure as is briefly
mentioned by Ahrem et al [1]. Especially for flexible struc-
tures this can have a large negative influence on the accuracy
of the solution. Overlap Gap
Instead of using the same transformation matrix for both
transferring the displacement and pressure loads over the in- Figure 4.1: Non-matching
terface, two different transformation matrices can be defined. meshes in 2D.
This leads to a consistent coupling approach without unphys-
ical oscillations in the pressure forces. However, conservation of energy over the interface is not
guaranteed. When a partitioned coupling technique is used to advance in time this does not have
to be a problem. In unsteady partitioned computations energy is generally already not conserved
due to errors caused by the time lag between flow and structure. In general, when the error
introduced by the information transfer is smaller than the spatial and temporal discretization
error, this coupling error does not affect the stability and accuracy of the computation.
In this chapter we show the difference in accuracy and efficiency between the conservative
50 CHAPTER 4. COUPLING OF NON-MATCHING MESHES
and consistent approach for different coupling methods. First the consistent and conservative
approach are presented followed by a short discription of the different coupling methods. The
difference in the interpolation properties between the two approaches is investigated using an
analytical test problem. A simple quasi-1D FSI problem is used to investigate the performance
of the methods in FSI computations.
uf = us on Γ, (4.1a)
ps ns = pf nf on Γ, (4.1b)
with uf,s the displacement, pf,s the pressure or stress tensor and nf,s the outward normal of the
flow and structure interface, respectively. The continuous inteface between the flow and structure
is represented by Γ. The first of these two boundary equations expresses the compatibility
between the displacement fields of the structure and the fluid at the fluid-structure interface. The
second equation states that the tractions of the wet surface of the structure are in equilibrium
with those on the fluid side.
Whichever coupling method is chosen to define the discrete form of these conditions, its
outcome can be formulated as
Uf = Hsf Us (4.2a)
Ps = Hf s Pf , (4.2b)
with Hsf and Hf s transformation matrices between the flow and structure interface and U and
P are defined by the approximations
n u np
X X
i
u(x) = N (x)Ui , p(x)n(x) = Dj (x)Pj , (4.3)
i=1 j=1
where nu,p is the number of unknowns on the interface for the displacement and pressure, respec-
tively, N (x) a function depending on the spatial discretization method used for the displacement
(for example, a step function in the finite volume formulation or the basis function in the finite
element formulation) and D(x) a function depending on the discretization method used for the
pressure. When the row-sum of H is equal to one, the interpolation is consistent, which means
that constant values are interpolated exactly.
The general opinion is that energy should be conserved over the interface leading to a
conservative coupling approach [14]. We will start with the energy equation in continous form
which is given by
with E the energy and v the velocity. The total energy of the system is then defined by
Z T Z T I
E dµΩ = v · pn ds dt, (4.5)
Ω(t) 0 ∂Ω(t)
0
4.1. CONSISTENT AND CONSERVATIVE COUPLING APPROACH 51
where we assume that the velocity of the boundary ∂Ω(t) is equal to the velocity v at that
boundary. To conserve energy globally over the fluid-structure interface we need that
Z TZ Z TZ
vf · pf nf ds dt = vs · ps ns ds dt, (4.6)
0 Γf (t) 0 Γs (t)
with Γ the interface between flow and structure. By using the method of lines (separate solution
of time and space) this is satisfied when
Z tn Z Z tn Z
vf · pf nf ds dt = vs · ps ns ds dt. (4.7)
tn−1 Γf tn−1 Γs
So there is global conservation of energy over the interface if this equation is satisfied.
The overall conservation properties depend both on the time and the spatial coupling used,
which cannot be investigated separately if the system is solved in a partitioned way. However, in
this chapter we focus only on the spatial coupling. In the limit of very small time steps (virtual
displacements, or steady state solution) or when a monolitic solution procedure is used with the
same time integration method applied for both the flow and the structure, we can write
Z tn Z Z Z tn Z Z
v · pn ds dt = v dt · pn ds = ∆t vn − vn−1 · pn ds = xn − xn−1 · pn ds
tn−1 Γ Γ tn−1 Γ Γ
(4.8)
with u = xn − xn−1 the displacement of the interface. This allows us to anlyse the coupling in
space separately.
Writing out the left hand side of (4.9) using (4.3) gives
nu np
Z Z X f
X f
Substituting (4.11) and (4.12) in (4.9) shows that energy is globally conserved when
So choosing
−1 T
Hf s = Mf f Hsf Mss (4.15)
in (4.2b) for the transformation of pressure over the interface results in global conservation of
energy over the interface.
To obtain a consistent interpolation, a constant displacement and constant pressure should
be exactly interpolated over the interface (similar to the patch test criterion in Lagrange Mul-
tiplier methods). This means that in the conservative approach both the row-sum of Hsf and
−1 T should be equal to one. For a general transformation
the row-sum of Hf s = Mf f Hsf Mss
matrix Hsf this is not the case as we will see in the following section where different setups of
the transformation matrices are outlined. The main question is whether global conservation of
energy or a consistent interpolation is preferred in fluid-structure interaction computations.
This equality can be approximately satisfied by a weighted residual method. Both sides are
multiplied by a set of weighting functions φk and integrated over the fluid structure interface
resulting in
Z Z
φk (x)wB (x) dx = φk (x)wA (x) dx. (4.20)
Γ Γ
nB
X nA
X
wB (x) = NBi (x)WBi , wA (x) = NAj (x)WAj , (4.21)
i=1 j=1
with WA,B containing the values of wA,B in the points on the interface of mesh A and B,
respectively, NA,B the basis function of mesh A or B and nA,B the number of unknowns at the
interface of mesh A or B. Substituting this in Together this in (4.20) gives
Z nB
X Z nA
X
φk (x) NBi (x)WBi dx = φk (x) NAj (x)WAj dx. (4.22)
Γ i=1 Γ j=1
For the displacement this is equal to the compatibility equation obtained by a Lagrange Mul-
tiplier method. When a Galerkin method is used there are two possibilities for φ, the basis
function of the flow or the structure, so we can write
nB Z
X nA Z
X
Nαk NBi dx WBi = Nαk NAj dx WAj for k = 1, ..., nα , (4.23)
Γ j=1 | Γ
i=1 | {z } {z }
ki
CαB kj
CαA
−1
WB = CBB CBA WA . (4.25)
−1
As a consequence the transformation matrix is defined as HAB = CBB CBA .
54 CHAPTER 4. COUPLING OF NON-MATCHING MESHES
Consistent approach
To interpolate constant values exactly to obtain a consistent interpolation we need
with β A,B a vector of length nA or nB respectively with constant value β. Using (4.23) this
becomes
nB Z
X nA Z
X
k i k j
NB NB dx β = NB NA dx β for k = 1, ..., nB . (4.27)
i=1 Γ j=1 Γ
Conservative approach
We now investigate the consistency of the pressure when the conservative coupling approach
is used. We start again with the discretized kinematic boundary condition (4.2a), where the
weighted residual method gives us
Hsf = Cf−1
f Cf s . (4.31)
When the same discretisation is used for the pressure and displacement in the flow, so Nf = Df
and nuf = npf , then Mf f = Cf f and (4.32) becomes
−1 T
Ps = Cf s Mss Pf or Mss Ps = Csf Pf . (4.33)
4.2.3 Radial basis function interpolation (RBFI) 55
with β s,f a vector of length ns or nf respectively, with constant value β. We can derive in a
similar way as in (4.29) and (4.30) that (4.34) is equal to
Z Z
Nsk dx = Nsk dx. (4.35)
Γs Γf
So only when the meshes are matching, Γf = Γs , equation (4.34) is satisfied and the method is
both conservative for the global energy and consistent for the pressure values.
Gauss integration
For the integration in CBA over ΓB some kind of projection between the two meshes is needed,
because NA is only defined on ΓA . The integrals appearing in CBA can be computed using
Gauss integration (GI) [8, 26]. However, an overlay mesh has to be created to ensure that the
basis functions on both sides of the discrete interface are continuous within a cell, to assure an
exact evaluation of the integral. The overlay mesh is obtained by projecting the cells of mesh A
on mesh B and taking the intersection of both meshes. This results in the following evaluation
Z over n
nX X gp,i
kj
CBA = NBk (x)NAj (x) dx ≈ wg NBk (xg,i )NAj (ΠA (xg,i )), (4.36)
ΓB i=1 g=1
where nover is the number of overlay cells, ngp,i is the number of Gauss quadrature points xg
in overlay cell i; wg the weight of the gth quadrature point and ΠA (xg,i ) the projection of xg,i
from mesh B on mesh A. The number of Gauss points to be used should be chosen equal to
the underlying order of the discretisation. The Gauss points need to be orthogonally projected
onto mesh B. This projection has to be accurate and take into account the normals of the used
basis functions, otherwise the order of the total interpolation decreases.
where xAj are the centres in which the values are known, in this case the nodes at the interface
of mesh A, q a polynomial, and φ a given radial basis function with respect to the Euclidean
distance ||x||. Different suitable radial basis functions available in liturature are presented in the
next paragraph. The coefficients γj and the polynomial q are determined by the interpolation
conditions
with WA containing the discrete values of wA at the interface of mesh A, and the additional
requirements
nA
X
γj s(xAj ) = 0, (4.39)
j=1
for all polynomials s with a degree less than or equal to that of polynomial q. The minimal
degree of polynomial q depends on the choice of the basis function φ. A unique interpolant is
given if the basis function is a conditionally positive definite function. If the basis functions are
conditionally positive definite of order m ≤ 2, as is the case for the functions used in this paper,
a linear polynomial can be used [3]. A consequence of using a linear polynomial is that constant
values are exactly interpolated leading to a consistent interpolation.
For the known quantity at the interface of mesh A the interpolation conditions (4.38) and
(4.39) can be written in matrix form as follows
WA ΦAA QA γ
= , (4.40)
0 QTA 0 β
with γ containing the coefficients γj , β the coefficients of the linear polynomial q, ΦAA an
nA × nA matrix containing the evaluation of the basisfunction φAi Aj = φ(||xAi − xAj ||) and QA
a nA × 4 matrix with row j given by [ 1 xAj yAj zAj ].
To obtain the values for the unknown quantity at the interface of mesh B we have to evaluate
(4.37) in the nodes on the interface of mesh B which can be written in matrix form as
γ
WB = ΦBA QB . (4.41)
β
and we can define the transformation matrix HAB as the first nB rows and nA columns of matrix
He to obtain WB = HAB WA . To obtain HAB no orthogonal projection and search algorithm
is needed, but the computation involves the inversion of a relatively small matrix. The number
of rows and columns of this matrix is equal to the number of flow or structure points on the
fluid-structure interface, which is usually very small compared to the total number of structure
and flow points.
where the subscript + means that only positive values are taken into account and this function
is in the remainder of the paper abreviated by RBF. The radius r defines the support of the
4.3. OTHER TRANSFORMATION METHODS 57
radial basis function. A large support radius yields a good approximation order, but then a
full matrix system has to be solved. What is more, too large radia lead to singular matrices,
because then all the entries of ΦAA are approximately equal to one. A small support radius
leads to a stable system with a band matrix that can be easily solved, but the interpolation is
less accurate than with a large support radius. For an accurate computation the support radius
for a fluid-structure interaction problem should be chosen at least as large as the maximum
distance of all centres with their nearest neighbours in both meshes.
Several global radial basis functions have been tested and evaluated for analytical inter-
polation tests as well as real fluid-structure interaction computations by Smith et al [29, 30].
From this work the following two functions are shown to be the most robust, cost effective and
accurate of the methods tested:
• Multi-quadric Biharmonic splines (MQ)
p
φ(||x||) = kxk2 + a2 . (4.44)
• Thin-plate splines (TPS)be consistent we need
The MQ-method uses a parameter a that controls the shape of the basis functions. A large value
of a gives a flat sheetlike function, while a small value of a gives a narrow conelike function.
The value of a is typically chosen to be in the range 10−5 − 10−3 . In this paper we use the value
a = 10−3 . In contrast with the radial basis functions used by Beckert and Wendland, these two
functions are defined on the entire domain. As a result, always a full matrix system has to be
solved.
Due to the addition of the linear polynomial constant values are exactly recovered, and
therefore the interpolation is consistent. However, when the conservative coupling approach is
used, the interpolation is not consistent for the transformation of pressure values. The reason
for this is similar as shown for nearest neighbour interpolation.
2. Interpolate the normals in the structure points to the position of the fluid points.
The second option is more accurate and yields a smoother surface in the presence of sever
deformations. Both methods don’t have to be conservative.
−1 −0.0
−4
−2
−5 1.0 −3
error in displacement
error in pressure
−6 −4
−5
−7
−6
1.5 2.0
−8 2.5 −7
NN 2.0 NN
RBF 0.25 RBF 0.25
−9 RBF 5 −8 RBF 5
TPS 2.5 TPS
MQ −9 MQ
−10 GI GI
3.0
−10 3.0
1 1.5 2 2.5 1 1.5 2 2.5
number of structure cells number of structure cells
Figure 4.3: Error in displacement. Figure 4.4: Error in pressure (−: conserva-
tive, −−: consistent).
TPS method are second order accurate where the accuracy of the TPS method is higher. The
RBF method has an order of about 2.5, but the accuracy depends on the value of the radius: the
larger the non-dimensional radius r, the more accurate the method. With r = 5, RBF is more
accurate than TPS and with r = 0.25 it is comparable to MQ. The order of the GI method is
the same as the order of the discretization. For a discretization order higher than two it is the
most accurate method. In this paper we only show the results for a third order discretization.
4.4.2 Pressure received by structure 59
0.205 0.35
pressure
pressure
0.3
0.2
0.25
0.195 0.2
0.15
0.19
exact
conservative 0.1
consistent
0.185 0.05
−0.5 0 0.5 −0.5 0 0.5
x x
Figure 4.5: Pressure received by the structure Figure 4.6: Pressure received by the structure
obtained with the GI method for nf = 441 obtained with the RBF method with r = 5
and ns = 49. for nf = 441 and ns = 49.
The reason for the lower convergence for the conservative approach can be seen in Figures 4.5
and 4.6, where the exact pressure obtained by the structure and the ones obtained with the
conservative and consistent approach are shown for the GI method and RBF method with r = 5,
respectively. The difference between the exact solution and the one obtained with the consistent
approach is barely visible. However, the solution obtained with the conservative approach shows
large oscillations. Except for the GI method, the amplitude of these wiggles does not decrease
for finer meshes, leading to the zeroth order convergence.
−5
2.6
−10 3.0
NN 2.0
error in work
2.5
RBF 0.25
3.0
RBF 5
TPS
MQ
GI
−15
4.0
−20
1 1.5 2 2.5
number of structure cells
Figure 4.7: Difference in work of the different methods (−: conservative, −−: consistent).
In Figure 4.7 the difference in work excerted on the interface between the flow and structure
side is depicted. When energy is conserved over the interface, this difference should be zero, as
60 CHAPTER 4. COUPLING OF NON-MATCHING MESHES
is the case for the conservative approach. With the consistent approach the error decreases with
approximately one order higher than that of the error in displacement and pressure. So even
as the consistent approach is not strictly globally conservative for the energy transfer over the
interface, the error decreases consistently.
−3
−1
−4 −2
−3
−5
error in displacement
error in pressure
−4
−6
−5
−7 −6
−7
−8
NN NN
RBF 0.25 −8 RBF 0.25
−9 RBF 5 RBF 5
TPS −9 TPS
MQ MQ
−10 GI −10 GI
Figure 4.8: Efficiency of the displacement (−: Figure 4.9: Efficiency of the pressure (−: con-
conservative, −−: consistent). servative, −−: consistent).
4.4.3 Efficiency
The efficiency of the coupling method is not the most important issue, because the computation
time needed for the coupling is usually much smaller than, for example, the time needed for the
flow solve. However, we want to investigate if the difference in efficiency between the methods
is considerable. To obtain an estimation of the efficiency of the methods, the computation time
needed to obtain a certain accuracy using Matlab version 7.0.1 on a 3 GHz computer is shown
for the displacement and pressure in Figures 4.8 and 4.9, respectively. The closer the line is
to the lower left corner, the more efficient the method. For the displacement the conservative
approach is most efficient for all methods, where the highest efficiency is obtained with the
RBF method with r = 5. The GI and NN method are the least efficient because they need
a projection and search algorithm. For this test case a simple search algorithm is used which
performs a loop over the two closest elements in both directions. For the pressure the consistent
approach is most efficient, because the conservative approach converges with a lower order (if
it converges at all). This time NN is most efficient, closely followed by the radial basis function
methods. The main conclusion is that the GI method, although it is more accurate for higher
order discretizations, does need much more computation time than the methods based on radial
basis function interpolation. The computational costs of the GI method also increase when a
higher discretization order is used, because then the projection algorithm becomes more com-
plicated.
Also other analytic problems are investigated, with different configurations of the displace-
ment and pressure to be interpolated, different configurations of the interface and various ratios
between flow and structure cells. The conclusions that can be drawn from the results of these
test cases are similar to the ones described above. Overall it can be concluded that for this
simple analytic problem the consistent approach is preferred over the conservative approach.
NN is only first order accurate and therefore the least accurate method. When the discretization
order of the total system is higher than two, the GI method is the best choice. However, its im-
plementation is more difficult and the computation time is higher than for the RBFI methods.
4.5. QUASI-1D FSI PROBLEM 61
Therefore, when the discretization of the total system is of order two or lower, or less important,
the RBFI methods are preferred where the RBF with r = 5 is the best choice.
nel with a flexible curved wall. The main velocity of the q(x)
compressible flow is in the x-direction and the structure ∆z
qe = z0
is modelled as a membrane. The diameter of the channel pf(x)
z(x) = z0 + ∆ z
V0
may vary due to a pressure difference between the pressure Flow
pf = −ρ0 c0 V0 ∂x z, (4.46)
with ρ0 , c0 and V0 the density, speed of sound and velocity, respectively, assumed to be constant,
pf the pressure and z = z0 + ∆z the location of the panel which is equal to the initial location
of the panel, z0 , plus the displacement from this initial position, ∆z.
For convenience the variables are scaled as follows
x V0 pf z
x̄ = , V̄0 = , p¯f = , z̄ = , (4.47)
L c0 ρ0 c20 L
with L the length of the channel. This results in the following non-dimensional equation:
For notation purposes the bars are dropped in the remainder of the paper. To discretize the
equations, a third order finite element discretization is used.
κq − T ∂xx q = ps − pe , (4.49)
where q is the displacement from the ’dry’ equilibrium position, qe (x), when ps = pe ; ps is the
pressure acting on the wall, pe is the pressure behind the wall, assumed to be constant, κ the
62 CHAPTER 4. COUPLING OF NON-MATCHING MESHES
elasticity per unit length and T the longitudinal tension per unit length. Again the variables
are scaled using the non-dimensional variables of (4.47) and the additional variables
q ps pe κL T
q̄ = , p¯s = , p¯e = , κ̄ = , T̄ = . (4.50)
L ρ0 c20 ρ0 c20 ρ0 c20 Lρ0 c20
This results in an equation which has two non-dimensional physical parameters κ̄ and T̄ and
has the same form as (4.49). In the remainder of the paper the bars are dropped. Again a third
order finite element discretization is used to discretize the equations in space.
1. Calculate ps = Hf s pf .
2. Calculate the new displacement of the structure, q from (4.49).
3. Obtain ∆z = Hsf q and update the location of the wall z = z0 + ∆z.
4. Calculate the new pressure in the flow pf from (4.48).
These four steps are repeated until the change in q is smaller than a certain threshold. To obtain
a numerically ’exact’ solution the steady state problem is solved at once on very fine matching
meshes. This is equivalent to solving the following equations on a fine mesh
To obtaine the ’exact’ solution a fourth order finite element discretization is used to discretize
the equations in space.
4.5.4 Results
For the test cases the following configuration is used. The boundaries of the domain are xmin =
−0.5 and xmax = 0.5 and the initial shape of the tube wall is given by
2
z0 (x) = a0 − a1 e−a2 x , (4.52)
where the parameters have the values a0 = 0.5, a1 = 0.25 and a2 = 80. This means that we
have a smooth converging/diverging channel. The ’dry’ equilibrium position of the membrane,
qe , is equal to this initial shape. The values used for the non-dimensional structure parameters
are: κ = 50 and T = 0.04, which results in a rather flexible membrane. For the flow velocity
yields V0 = 3, corresponding to a supersonic flow of Mach 3. Initially the pressure in the flow,
pf , the pressure behind the wall, pe , and the displacement q are all equal to zero. We use
nf = 21 · 2k flow cells and ns = 6 · 2k structure cells, with k ∈ {0, 1, 2, 3, 4, 5}, leading to a ratio
of approximately 30%.
The L2 -error of the displacement in the flow points versus the number of structure points
is depicted in Figure 4.11. The solid line is obtained with the conservative and the dotted line
with the consistent approach. The gray lines are added to indicate the discretization error of the
total system and are generated with matching meshes: the solid gray line with ns = nf = 21 · 2k
4.5.4 Results 63
−1
−0.0 −0.0
0.5
−2
0 −0.0−0.0
−0.0 0.0
−3 −0.0
−0.5
−4 1.0 −1
−1.5
error z
s
−5 1.0
error p
2.0 2.0 −2
1.9 2.0
−6 −2.5
3.0 1.1
Direct −3 Direct
−7 0.9
NN 2.0 2.0 NN
RBF 0.25 2.2 −3.5 1.0
RBF 0.25 1.5
−8 RBF 5 3.0 RBF 5
3.0 −4 2.0 2.0
TPS TPS 1.5
MQ −4.5 MQ
−9
GI GI
3.0 2.0
1 1.5 2 2.5 3 1 1.5 2 2.5 3
number of structure cells number of structure cells
Figure 4.11: Error in displacement (−: con- Figure 4.12: Error in pressure (−: conserva-
servative, −−: consistent). tive, −−: consistent).
and the dotted line with nf = ns = 6 · 2k . Above these lines the coupling error of a method is
higher than the discretization error. NN is again the least accurate method. The methods based
on radial basis function interpolation are second order accurate with the consistent approach
giving the most accurate results. Only for higher values of ns the coupling error is higher than
the third order discretization error. When the discretization error of the total system is second
order or lower, the coupling error is always smaller than the discretization error. The coupling
error for the GI method is always smaller than the discretization error for both the conservative
and consistent approach.
The L2 -error of the pressure in the structure points versus the number of structure points is
depicted in Figure 4.4. Because the value for the pressure is obtained from the space derivative
of z, the order of convergence should be one lower than for the displacement. It can be seen
that the conservative approach leads again to a zeroth order error for all methods, except GI.
The coupling error for the GI method is always smaller than the discretization error for both
the conservative and consistent approach. The reduction in order for the conservative GI as can
be seen in the analytical test case in Figure 4.4 is not visible, because of the order reduction
caused by the space derivative.
10
exact 0.1
conservative exact
8
consistent 0.08 conservative
consistent
6
0.06
4
0.04
displacement
pressure
2
0.02
0
0
−2 −0.02
−4 −0.04
−6 −0.06
−8 −0.08
−0.5 0 0.5 −0.5 0 0.5
x x
Figure 4.13: Pressure received by the struc- Figure 4.14: Displacement obtained with the
ture obtained with the RBF method with RBF method with r = 5 for nf = 441 and
r = 5 for nf = 441 and ns = 36. ns = 36.
In Figures 4.13 and 4.14, the exact solution obtained by the structure and the ones obtained
with the conservative and consistent approach for the RBF method with r = 5 are shown for the
pressure obtained by the structure, ps , and the displacement q, respectively. It can be seen that
64 CHAPTER 4. COUPLING OF NON-MATCHING MESHES
the large oscillations felt by the structure also result in small deviations in the displacement.
The more flexible the structure, the larger these deviations become.
0.5
−2
0
−3 −0.5
−1
−4
error z
s
error p
−1.5
−5
−2
NN −2.5 NN
−6
RBF 0.25 RBF 0.25
RBF 5 −3 RBF 5
−7 TPS TPS
MQ −3.5 MQ
GI GI
−1 −0.5 0 0.5 1 1.5 2 2.5 3 3.5 −1 −0.5 0 0.5 1 1.5 2 2.5 3 3.5
cpu time cpu time
Figure 4.15: Efficiency for the displacement Figure 4.16: Efficiency for the pressure (−:
(−: conservative, −−: consistent). conservative, −−: consistent).
For this simple test problem, the computation time needed by the coupling algorithm is still
considerable compared to the overall computation time. From Figures 4.15 and 4.16 it can be
seen that the consistent RBF method with a large radius is the most efficient. The GI method
is again the least efficient, because of the projection algorithm. The results for the difference in
work are similar to the ones obtained by the analytical test case.
The main conclusion is that the NN method is not suitable for the coupling of non-matching
meshes, because it is only first order accurate. The conservative approach should be used with
the GI method, because it gives the highest accuracy and efficiency. The methods based on radial
basis function interpolation show large oscillations in the pressure obtained by the structure,
when the conservative approach is used. For these coupling methods the consistent approach
provides the best accuracy and efficiency.
4.6 Conclusions
In this paper we investigate the difference in accuracy and efficiency between the conservative
and consistent approach for different coupling methods. The performance is investigated for an
analytical test problem as well as a simple quasi-1D FSI problem. When the coupling method is
based on a weighted residual formulation of the coupling conditions, the highest accuracy and
efficiency are obtained with the conservative approach. For other coupling methods the conserv-
ative approach results in unphysical oscillations in the pressure received by the structure. When
the structure is flexible enough these oscillations can result in deviations in the displacement.
For these methods the consistent approach provides the best accuracy and efficiency.
Overall, when the discretization order of the total system higher than two, the conservative
GI method is the best choice. However, its implementation is more difficult and the computation
time is higher than for the methods based on radial basis function interpolation. This is because
the higher order of the GI method is only obtained when all the projection steps are accurately
performed. Therefore, when the discretization of the total system is of order two or lower, or
less important, the methods based on radial basis function interpolation are preferred where
the compact RBF with a high support radius is the best choice.
Up to now only a simple steady quasi-1D fluid-structure interaction problem has been con-
sidered. Future research will focus on the investigation of time-dependent and more realistic
4.6. CONCLUSIONS 65
Moving Meshes
Fluid-structure interaction computations typically involve moving boundaries for the flow due
to the deformation of the structure. To be able to perform the aeroelasticity computations
accurately and efficiently, a fast and reliable method is needed to adapt the computational
grid to the new domain. A natural choice is to completely regenerate the mesh at each time
step. However, this is in general an expensive process which requires user interaction. Because
the mesh must be updated frequently in an unsteady or design optimization computation,
an efficient and automatic method is needed to update the existing mesh by means of node
relocation, see Figure 5.1. It is desirable that the constantly updated mesh retains the essential
qualities of an initial mesh.
1111111111111111111111
0000000000000000000000
0000000000000000000000
1111111111111111111111
0000000000000000000000
1111111111111111111111
0000000000000000000000
1111111111111111111111
0000000000000000000000
1111111111111111111111
0000000000000000000000
1111111111111111111111
0000000000000000000000
1111111111111111111111
0000000000000000000000
1111111111111111111111
0000000000000000000000
1111111111111111111111
0000000000000000000000
1111111111111111111111
0000000000000000000000
1111111111111111111111
0000000000000000000000
1111111111111111111111
0000000000000000000000
1111111111111111111111
0000000000000000000000
1111111111111111111111
0000000000000000000000
1111111111111111111111
0000000000000000000000
1111111111111111111111
0000000000000000000000
1111111111111111111111
0000000000000000000000
1111111111111111111111
0000000000000000000000
1111111111111111111111
0000000000000000000000
1111111111111111111111
0000000000000000000000
1111111111111111111111
0000000000000000000000
1111111111111111111111
0000000000000000000000
1111111111111111111111
0000000000000000000000
1111111111111111111111
0000000000000000000000
1111111111111111111111
0000000000000000000000
1111111111111111111111
0000000000000000000000
1111111111111111111111
0000000000000000000000
1111111111111111111111
0000000000000000000000
1111111111111111111111
For structured meshes there are efficient methods available to deform the mesh, such as
transfinite interpolation. The displacements of nodes at the boundaries of the mesh are inter-
polated along grid lines to points in the interior of the mesh. However, these techniques are
unsuitable for unstructured meshes.
Unstructured grids have become more and more popular because of their flexibility to dis-
cretize arbitrary complex domains and to enhance the solution accuracy and efficiency by using
grid adaptation. Two different mesh movement strategies are known for unstructured grids. The
first exploits the connectivity of the internal grid points. The connection between the grid points
is represented for example by springs or as solid body elasticity. Special instances of this con-
tinuous approach include moving grids based on Laplacian and Biharmonic operators. All the
methods based on grid connectivity involve solving a system of equations involving all the flow
points and can therefore be quite expensive. Hanging nodes, often encountered in unstructured
68 CHAPTER 5. MOVING MESHES
X F = {xi,j | i = 1 · · · Ni , j = 1 · · · Nj }, (5.1)
where x = (x, y)T are the coordinates of the mesh vertices, Ni is the number of edges in the i
direction and Nj in the j direction. Normalized arclength variables s = (s, t)T can be defined
along the grid lines in respectively the i and j direction. The normalized arclength in the i
direction is defined as
Pi
k xm,j − xm−1,j k
si,j = Pm=1
Ni
. (5.2)
m=1 k xm,j − xm−1,j k
A similar formula as Equation (5.2) is used to define t as the normalized arclength in the j
direction. It follows from the preceding that
Each edge E has its own one-dimensional structured grid defined as:
X E = {xi | i = 1 · · · Ni }, (5.4)
where Ni is the number of vertices in the i direction. Furthermore, each vertex V has its own
position vector x.
The displacements d = (dx , dy )T of each internal vertex of a surface grid can be computed
with the TFI method in two steps. The first step computes the displacements d1 in the interior
by a straight line interpolation in the i direction:
where d1,j and dNi ,j are the known displacements on the edges of the face, where i = 1 and
i = Ni . These displacements have been determined by the additional mesh deformation method.
The second step computes the displacements d2 which add the mismatch of the displacements
along the second pair of opposite edges by straight line interpolation in the j direction:
where di,1 and di,Nj are the known displacements on the edges of the face, where j = 1 and
j = Nj . The displacement of the interior of the surface grid is finally defined as:
The TFI method for three-dimensional meshes is slightly more complex. A three-dimensional
grid consists of a set of blocks, faces, edges and vertices. Describing the method for three-
dimensional meshes is beyond the scope of this notes. A detailed description can be found in
[31].
of the displacement depends on the minimum distance between the master and slave node. The
displacement is scaled by a Gaussian distribution function which is defined as
−(x−a)2
e 2σ2
f (x) = √ , (5.8)
2πσ 2
where a is the mean of the function and σ the standard deviation. So if a slave node is very close
to a master node, it follows almost exactly its movement, however when the distance between
a slave node and master node is very large, the slave node will not move at all.
where Filinear is the force exerted on node i by its neighboring nodes j, Ni is the number of
neighboring nodes connected to node i, di and dj are the displacements of node i and its
neighboring node j, respectively. The spring stiffness kij for a given edge i to j is taken to be
inversely proportionally to the length of the edge and is defined as
1
kij = . (5.10)
||xi − xj ||
Due to mesh refinement or adaptation of cells in an unstructured mesh, hanging nodes can
arise (Figure 5.4). These hanging nodes can cause very poor quality cells after deformation,
because hanging nodes lack one or more neighboring nodes and this disturbs the equilibrium
balance. It is better to remove these hanging nodes before applying the spring analogy method.
Removing hanging nodes will increase the computational costs.
The spring analogy method can handle moderate deformations. Grid line crossovers can
occur for complex problems with relative large mesh motion amplitude. The next two paragraphs
present improvements of the spring analogy method by introdcucing (semi) torsional springs to
prevent grid line crossovers. The spring analogy method tends to be computationally expensive,
because a system of equations has to be solved as large as the number of internal grid points.
Spring analogy can be applied to both structured and triangular unstructured meshes, but for
structured meshes there are much more efficient methods available, like transfinite interpolation.
5.1.4 Spring analogy 71
Semi-torsional springs
The semi-torsional spring analogy model is similar to the lin-
eal formulation, with angle information incorporated into the k
spring stiffness [4, 41]. The stiffness of an edge is defined as
the sum of its lineal and semi-torsional stiffness. The semi- θ1
torsional stiffness depends on the angles facing the edge, so
lineal semi−torsional
kij = kij + kij , (5.12) i j
N Eij
X 1
semi−torsional
kij = 2 ij
(5.13)
m=1 sin θm θ2
where x and y are the coordinates of the vertices in the undeformed space, dx and dy are the
displacements in respectively x and y direction, Ωi is the cell area, n is a weighting function
and C is the set of grid cells. The weighting function n allows different weightings for cell
contribution. Each cell has the same weight for n = 0, but for n < 0, more importance is given
to small cells, which are normally located in critical regions of the domain.
Advantages are that this method runs fully automatically, requiring minimum user interaction
and that this method produces high quality grids, for example compared to the spring analogy
method (Section 5.1.4). A drawback is that this method is computationally expensive, because
a large system of equations has to be solved.
The minimum of the strain energy corresponds to the equilibrium position caused by the strain.
Advantages are that this method runs almost automatically and that this method produces high
quality grids. The grids produced with the least-squares approach (Section 5.1.5) seem to have
higher quality, but this method produces almost similar meshes [40]. Both methods produce
better meshes than the spring analogy method (Section 5.1.4). A drawback is that this method
is even more computationally expensive than the spring analogy method [39].
where X and Y are the mesh vertex coordinates after deformation and ∇2 is the Laplacian
∂2 ∂2
operator, defined in two dimensions as ∇2 = ∂x 2 + ∂y 2 . This system of equations consists of
second order partial differential equations. Therefore, only one condition along the boundary of
the mesh can be specified for each equation. Along the boundaries either the coordinates of the
boundary nodes (Dirichlet condition) can be fixed or the first derivative of these coordinates
in direction normal to the unperturbed boundary (Neumann condition) can be specified. It is
impossible to control the mesh cell sizes in direction normal to the boundary, if the coordinates
of the boundary nodes are fixed. And it is also impossible to control the mesh cell sizes on
the boundary, if the first derivative in normal direction is specified. Because of this limitation,
the Laplacian method can only be applied to very small boundary deformations and meshes
without boundary layers. The computational costs are high and approximately the same as for
the spring analogy method [19].
∇4 X = 0,
∇4 Y = 0. (5.17)
The biharmonic method consists of fourth order partial differential equations, instead of the
second order partial differential equations of the Laplacian method (Section 5.1.7). Therefore,
two boundary conditions along each boundary of the domain can be specified. This creates the
possibility to control the mesh cell sizes in direction normal to the boundary and to control the
mesh cell sizes on the boundary.
Numerical implementation of this method is performed by deriving the weak formulation of
Equations (5.17). The numerical implementation follows a Galerkin finite element procedure.
Further details are discussed in [19].
The ability to control the mesh cell sizes on the boundary and in direction normal to the
boundary is a significant advantage over the Laplacian method. The biharmonic method allows
larger deformations of the boundary and is therefore more robust. Test cases with boundary
layers can be handled with this method, because the mesh cell size near the boundaries is better
controllable. However, there are some serious drawbacks. One drawback is that the new coor-
dinates of the mesh nodes are not bounded by the values on the boundary of the domain. This
means that for very large deformations nodes can receive a position outside the boundaries,
which results in mesh cells with negative areas or volumes. Furthermore, this method is four
times more computationally expensive than the Laplacian method and gives only slightly bet-
ter results. Compared to the spring analogy method (Section 5.1.4), applying the biharmonic
operator gives almost the same results. However, the biharmonic method is roughly twice as
expensive in computational costs than the spring analogy method.
to the boundary of the aerodynamic mesh. But why not interpolate the displacement to all the
nodes of the flow mesh, instead of only to the boundary?
Radial basis function interpolation can be used to derive the displacement of the internal
fluid nodes given the displacement of the structural nodes on the interface. The interpolation
function, s, describing the displacement in the whole domain, can be approximated by a sum
of basis functions
nb
X
s(x) = αj φ(||x − xbj ||) + p(x), (5.18)
j=1
where xbj = [xbj , ybj , zbj ] are the centres in which the values are known, in this case the bound-
ary nodes, p a polynomial, nb the number of boundary nodes and φ a given basis function with
respect to the Euclidean distance ||x||. The coefficients αj and the polynomial p are determined
by the interpolation conditions
with db containing the discrete known values of the displacement at the boundary, and the
additional requirements
nb
X
αj q(xbj ) = 0, (5.20)
j=1
for all polynomials q with a degree less or equal than that of polynomial p. The minimal degree
of polynomial p depends on the choice of the basis function φ. A unique interpolant is given if the
basis function is a conditionally positive definite function. If the basis functions are conditionally
positive definite of order m ≤ 2, a linear polynomial can be used [3]. In this paper we only apply
basis functions that satisfy this criterion. A consequence of using a linear polynomial is that
rigid body translations are exactly recovered.
The values for the coefficients αj and the linear polynomial can be obtained by solving the
system
db Mb,b Pb α
= , (5.21)
0 PbT 0 β
with α containing the coefficients αj , β the coefficients of the linear polynomial p, Mb,b an
nb × nb matrix containing the evaluation of the basisfunction φbi bj = φ(||xbi − xbj ||) and Pb
an nb × 4 matrix with row j given by [ 1 xbj ybj zbj ]. Solving the system can be done by
fast iterative techniques [6, 16] or using partition of unity [38].
The values for the displacement in the interior of the flow mesh din , can then be derived by
evaluating the interpolation function (5.18) in the internal grid points:
The displacement can be interpolated separately for each spatial direction. The proces of mesh
deformation with RBF’s for one direction is visualized in Figures 5.7 and 5.8. Each individual
point is moved individually based on its position in space according to the interpolation function,
and this means that no mesh-connectivity information is needed at all. However, to be able to
see the effects on the mesh clearly, the points are connected in a very simple mesh as is shown
in Figure 5.7. The block in the middle is moved to the right and the resulting interpolation
function is shown in Figure 5.8. This function is equal to zero at the outer boundaries and equal
5.1.9 Radial basis function interpolation 75
Figure 5.7: Initial mesh. Figure 5.8: Interpolation function and re-
sulting deformed mesh.
to the displacement of the block at the location of the block boundaries. The displacement of
the nodes in the interior of the mesh can then be derived from this function and the resulting
deformed mesh is shown in Figure 5.8.
The size of the system that has to be solved in (5.21) is equal to (nb + 4) × (nb + 4) which
is usually very small compared to the systems that have to be solved in mesh-connectivity
schemes. The systems encounterd there are approximately as large as nin × nin , with nin the
total number of mesh points. The total number of mesh points is a dimension higher than the
number of points on the boundary of the mesh. The mesh movement technique based on RBF’s
is very easy to implement, even for 3D applications, because no mesh-connectivity information is
needed. Also the implementation for partitioned meshes, occuring in parallel flow computations,
is straightforward.
There are various radial basis function available in literature which are suitable for inter-
polating multivariate data. They can be divided in two groups: functions with compact and
functions with global support. Functions with compact support have the following property:
f (x) 0 ≤ x ≤ 1,
φ(x) = (5.23)
0 x > 1,
where f (x) ≥ 0. The function is generally scaled with a support radius r to control the compact
support, so φr = φ(x/r). When a radial basis function with compact support is used, mainly the
mesh nodes inside a circle (2D) or sphere (3D) with radius r around a centre xj are influenced
by the movement of this centre. Therefore, higher values for the support radius lead generally
to more accurate solutions. However, high values of the support radius r also result in dense
matrix systems, wheras low values of r result in sparse matrix systems which can be solved
more efficiently. Functions with global support are not equal to zero outside a certain radius,
but cover the whole interpolation space, which leads to dense matrix systems.
In [5] it is shown that the C 2 continuous basis function with compact support (CP C 2 )introduced
by Wendland [35] and the Thin Plate Spline (TPS) give meshes with the best quality. They are
given by:
More information about RBF’s and their error and convergence properties can be found in
[6, 37, 36].
Quadrilateral : Hexahedral :
xk+1 − xk xk+3 − xk xk+4 − xk
xk+1 − xk xk+3 − xk
Ak = , Ak = yk+1 − yk yk+3 − yk yk+4 − yk
yk+1 − yk yk+3 − yk
zk+1 − zk zk+3 − zk zk+4 − zk
(5.24)
where the indices are taken modulo n.
The determinant of Ak is proportional to the area of the element, so we define αk = det(Ak ).
Additional matrices (the metric tensors) are obtained by forming the combinations ATk Ak . The
5.2. MESH QUALITY METRICS 77
metric tensors are symmetric and have three (d = 2) or six (d = 3) distinct components λkij
with i, j = 1, .., d and are given by
d=2: d=3:
λk11 λk12 λk13
λk11 λk12 (5.25)
ATk Ak = , ATk Ak = λk12 λk22 λk23
λk12 λk22
λk13 λk23 λk33
The quality metric of a given element is denoted by f and is a scalar function of the Jacobian
matrix, which is a function of the positions of the element nodes. This quality metric f should
possess the following essential properties:
• f is dimensionless.
• f is referenced to an ideal element which has the desired configuration of the element.
• 0 ≤ f ≤ 1 with f = 1 if and only if the element is an ideal element, f = 0 if and only if
the element is degenerate.
• the value of f is independent of the way the nodes are numbered.
• the value of f is independent of translations.
• the value of f is independent of the orientation of the element.
• f is a function of node positions.
Skew metric
The skew metric measures the skewness and therefore the distortion of an element. If a node of
an element possesses a local negative area, this metric value is set to zero. The expressions for
the skew metric for triangular, quadrilateral, tetrahedral and quadrilateral elements are given
by:
√
3α
T riangular fskew = (5.27)
λ11 + λ22 − λ12
√ 2/3
3(α 2)
T etrahedral fskew = 3 (5.28)
2 (λ11 + λ22 + λ33 ) − (λ12 + λ23 + λ13 )
4
Quadrilateral fskew = P q (5.29)
3 k λk /α
k=0 λ 11 22 k
8
Hexahedral fskew = P q (5.30)
7
k=0 λk11 λk22 λk33 /αk
78 CHAPTER 5. MOVING MESHES
Size-skew metric
To measure both the change in element size and the distortion of an element, the size-skew
metric [21] is introduced which is defined as the weighted product of the relative size and skew
metrics:
p
fss = fsize fskew , (5.31)
. Since changes in element volume have a smaller influence on the mesh quality than element
distortion, the relative size metric has a smaller influence on the size-skew metric than the skew
metric. Essential properties of the quadrilateral size-skew metric are:
• fss = 1 ⇔ element has equal angles and same size as the initial element.
fss= 1 fss= 0.5 fss= 0.5 fss= 0 fss= 1 fss= 0.5 fss= 0.5 fss= 0
(a) Triangular. (b) Hexahedral.
The average value of the metric over all the elements indicates the average quality of the
mesh. The higher the average quality of the mesh, the more stable, accurate and efficient the
computation will be. The minimum value of the metric over all the elements indicates the quality
of the cell with the lowest quality. This value is required to be larger than zero, otherwise the
mesh contains degenerate cells. Degenerate elements have a very negative influence on the
stability and accuracy of numerical computations. In [5] it is shown that the minimum value of
the mesh quality metric is the best mesh quality indicator. Table 5.1 presents an easy way to
interpret the minimum value of the quality metric:
In the next section we will use minimal value of the size-skew metric to compare meshes
after mesh movement.
5.3 Examples
In this section we show some examples of differences between mesh deformation methods. First a
simple 2D example is given to show the difference in quality of the mesh obtained with the semi-
torsional spring analogy and radial basis function interpolation. Furthermore realistic results on
a distorted mesh are presented, where flow computations are performed around a NACA-0012
airfoil. There the difference is shown between completely regenerating a new mesh and moving
the mesh with radial basis function interpolation.
5.3.1 Example 1: Rotation and translation 79
The quality and robustness of the new method depend on the value of the support radius
when a radial basis function with compact support is used. When the support radius is chosen
large enough, the quality and robustness converge to an optimum. Therefore, a relatively high
value, r is 2.5 times the characteristic length of the computational domain, is used in the first
three test cases, where we only investigate the accuracy of the different RBF’s.
0.16
0.14
0.12
0.1
min(f )
ss
TPS
0.08
Semi−Tors
0.06
0.04
0.02
0
0 5 10 15
number of iterations
Figure 5.14: Quality of the worst cell of the Figure 5.15: Initial mesh.
mesh for the different RBF’s (test case 1).
The minimum value of fss after mesh movement with the semi-torsional spring analogy
and radial basis function interpolation is shown in Figure 5.14 for an increasing number of
intermediate steps. It can be seen that the minimum value of fss indeed increases when more
80 CHAPTER 5. MOVING MESHES
intermediate steps are taken. For the semi-torsional spring analogy method we impose one
layer of boundary stiffness by increasing the spring stiffness with a factor 3.5. More than 8
intermediate steps are needed to avoid degenerate cells and the minimum value of the mesh
quality metric is very low. The final mesh using 15 intermediate steps is shown in Figure 5.17.
It can be seen that the displacement does not spread very far into the domain and the cells
close to the moving block are heavily deformed. The mesh quality obtained with the Thin Plate
Spline is much higher, because the deformation is more evenly spread through the domain as
can be seen in Figure 5.16.
Figure 5.18: Final mesh using CTPS Cb2 af- Figure 5.19: Final mesh using CTPS C 1 af-
ter 40 intermediate steps. ter 40 intermediate steps.
the pressure field is shown in Figure 5.20. In the second computation the mesh is deformed in
Figure 5.20: Pressure field around wing on Figure 5.21: Pressure field around wing on
a new generated mesh. a deformed mesh.
one step with the thin plate spline from its original position. The result is shown in Figure 5.21.
The resulting pressure distributions over the wing are identical as can be seen in Figure 5.22.
The difference in lift computed on the two different meshes is only 0.8%.
0.8
re−mesh
TPS
0.78
0.76
pressure
0.74
0.72
0.7
0.68
0.66
0 0.2 0.4 0.6 0.8 1
cord
to avoid degenerate cells. In Fig. 5.23 the L2 -norms of the fluid density ρ, pressure p, velocity
in x-direction u and y-direction v for the different meshes and mesh deformation schemes are
shown.
Laplace smoothing RBF interpolation
0
-2
log ε
-4
10
|| ρ ||2
-6
|| u ||2
|| v ||2
|| p ||2
-8
-2 -1 0 -2 -1 0
10 10
log ∆t log ∆t
Figure 5.23: Convergence for the L2 -norm of the density, pressure and u- and v-velocity com-
ponents for the piston problem with Laplace smoothing and radial basis function interpolation.
We use the fourth order IMEX scheme [42] for the partitioned time integration. The results
obtained with the Laplace smoothing are not satisfactory. The fourth order of the scheme is
not observed and although the test problem is essentially one-dimensional, the v-velocity is
not zero due to the imperfect mesh. For the large time steps this perturbation is only small
compared to the other errors. However, since the convergence for the v-velocity is clearly not
fourth order, its influence becomes more apparent for the smaller time steps. RBF interpolation
for the imperfect mesh has the same nonzero v-velocity for the large time steps. This time,
however, the perturbation does converge with fourth order accuracy and therefore its influence
on the solution remains negligible. Therefore we can conclude that the RBF interpolation does
not aggravate the imperfections in the flow mesh.
In order to explain the bad convergence with the Laplace smoothing we study the mesh face
velocities for the cell faces which are displayed in Fig. 5.24. It shows that the Laplace smoothing
with optimization introduces irregularities (wiggles) in the mesh face velocities which are worse
for small ∆t. The mesh face velocity does not converge to a consistent solution so the design
order of the IMEX scheme can not be expected. Due to the high accuracy and regularity of the
displacement field obtained with RBF interpolation the RBF mesh deformation algorithm does
0,42 Tim 0
2
,
4
e[] Tim
e[]
5.4. IMPORTANCE OF SMOOTH MESH DEFORMATION FOR HIGHER ORDER TIME-INTEGRATION83
0,4 0,4
0,2 0,2
mesh velocity
mesh velocity
η=0 η=0
0 0
η=1 η=1
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
[1] R. Ahrem, A. Beckert & H. Wendland, A new multivariate interpolation method for large-
scale coupling problems in aeroelasticity, Conference proceedings IFADS, Munich, 2005.
[4] F. J. Blom, Considerations on the spring analogy, International Journal for Numerical
Methods in Fluids, vol. 32, pp. 647–668, 2000.
[5] A. de Boer, M. S. van der Schoot & H. Bijl, Mesh deformation based on radial basis function
interpolation, Computers and Structures, vol. 85, 2007.
[6] M. D. Buhmann, Radial basis functions, Acta Numerica, vol. 9, pp. 1–38, 2000.
[7] J. R. Cebral, Loose Coupling Algorithms for Fluid-Structure Interaction, Ph.D. thesis,
George Mason University, 1996.
[8] J. R. Cebral & R. Löhner, Conservative load projection and tracking for fluid-structure
problems, AIAA Journal, vol. 35, no. 4, pp. 687–692, 1997.
[9] C. Degand & C. Farhat, A three-dimensional torsional spring analogy method for unstruc-
tured dynamic meshes, Computers and Structures, vol. 80, pp. 305–316, 2002.
[10] C. Farhat, C. Degand, B. Koobus & M. Lesoinne, An improved method of spring analogy
for dynamic unstructured fluid meshes, in Structures, Structural Dynamics and Mater-
ial Conference and Exhibit, 39st AIAA/ASME/ASCE/AHS/ASC Structures, Structural
Dynamics, and Materials Conference and Exhibit, Long Beach, CA, April 20-23 1998.
[11] C. Farhat, C. Degrand, B. Koobus & M. Lesoinne, Torsional springs for two-dimensional
dynamic unstructured fluid meshes, Computer Methods in Applied Mechanics and Engi-
neering, vol. 163, pp. 231–245, 1998.
[12] C. Farhat, C. Harris & D. J. Rixen, Expanding a flutter envelope using accelerated flight
data - application to an f16 fighter configuration, in Structures, Structural Dynamics and
Material Conference and Exhibit, 41st AIAA/ASME/ASCE/AHS/ASC Structures, Struc-
tural Dynamics, and Materials Conference and Exhibit, Atlanta, GA, April 3-6 2000.
[13] C. Farhat, M. Lesoinne & P. LeTallec, A conservative algorithm for exchanging aerody-
namic and elastodynamic data in aeroelastic systems, in 36th AIAA Aerospace Sciences
Meeting and Exhibit, AIAA, Reno, Nevada, January 1998.
86 BIBLIOGRAPHY
[14] C. Farhat, M. Lesoinne & P. Tallec, Load and motion transfer algorithms for fluid/structure
interaction problems with non-matching discrete interfaces: Momentum and energy con-
servation, optimal discretization and application to aeroelasticity, Computer Methods in
Applied Mechanics and Engineering, vol. 157, pp. 95–114, 1998.
[15] C. Farhat, D. J. Rixen & L. D. Peterson, Simulation of the continuous parametric iden-
tification of an accelerating aeroelastic system, in 37th AIAA Aerospace Sciences Meeting
and Exhibit, AIAA, Reno, Nevada, January 1999, aIAA 99-0797.
[16] A. C. Faul & M. J. D. Powell, Proof of convergence of an iterative technique for thin plate
spline interpolation in two dimensions, Advances in Computational Mathematics, vol. 11,
pp. 183–192, 1999.
[18] P. M. Hartwich & S. Agrawal, Method for perturbing multiblock patched grids in aeroelastic
and design optimization applications, AIAA-97-2038, 1997.
[19] B. T. Helenbrook, Mesh deformation using the biharmonic operator, International Journal
for Numerical Methods in Engineering, vol. 56, pp. 1007–1021, 2003.
[21] P. M. Knupp, Algebraic mesh quality metrics for unstructured initial meshes, Finite Ele-
ments in Analysis and Design, vol. 39, pp. 217–241, 2003.
[22] M. Lesoinne & C. Farhat, Geometric conservation laws for aeroelastic computations using
unstructured dynamic meshes, in 12th AIAA Computational Fluid Dynamics Conference,
AIAA, San Diego, CA, June 1995.
[23] M. Lesoinne & C. Farhat, Higher-order subiteration-free staggered algorithm for nonlinear
transient aeroelastic problems, AIAA Journal, vol. 36, pp. 1754–1757, 1998.
[24] R. Löhner, C. Yang, J. Cebral, J. D. Baum, H. Luo, D. Pelessone & C. Charman, Fluid-
structure interaction using a loose coupling algorithm and adaptive unstructured grids, in
Computational Fluid Dynamics Review (edited by M. Hafez & K. Oshima), John Wiley,
1995.
[25] D. Lynch & K. ONeill, Elastic grid deformation for moving boundary problems in two
space dimensions, in Finite Elements in Water Resources (edited by S. Wang), 1980.
[26] N. Maman & C. Farhat, Matching fluid and structure meshes for aeroelastic computations:
A parallel approach, Computers and Structures, vol. 54, no. 4, pp. 779–785, 1995.
[27] S. Piperno & C. Farhat, Energy based design and analysis of staggered solvers for nonlinear
transient aeroelastic problems, in Structures, Structural Dynamics and Material Conference
and Exhibit, 41st AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and
Materials Conference and Exhibit, Atlanta, GA, April 3-6 2000.
[28] D. J. Rixen, Substructuring and Dual Methods in Structural Analysis, Ph.D. thesis, Uni-
versite de Liege, 1997.
BIBLIOGRAPHY 87
[29] M. J. Smith, C. E. S. Cesnik & D. H. Hodges, Evaluation of some data transfer algorithms
for noncontiguous meshes, Journal of Aerospace Engineering, vol. 13, no. 2, pp. 52–58,
2000.
[31] S. Spekreijse, B. Prananta & J. Kok, A simple, robust and fast algorithm to compute
deformations of multi-block structured grids, Tech. rep., 2002.
[32] P. Thévenza, T. Blu & M. Unser, Interpolation revisited, IEEE Transactions on Medical
Imaging, vol. 19, no. 7, pp. 739–758, 2000.
[33] H. M. Tsai & A. S. F. Wong, Unsteady flow calculations with a parallel multiblock moving
mesh algorithm, AIAA Journal, vol. 39, no. 6, pp. 1021–1029, 2001.
[34] Z. J. Wang & A. J. Przekwas, Unsteady flow computation using moving grid with mesh
enrichment, Tech. Rep. AIAA-94-0285, 1994.
[36] H. Wendland, Error estimates for interpolation by compactly supported radial basis func-
tions of minimal degree, Journal of Approximation Theory, vol. 93, pp. 258–272, 1998.
[37] H. Wendland, On the smoothness of positive definite and radial functions, Journal of
Computational and Applied Mathematics, vol. 101, pp. 177–188, 1999.
[38] H. Wendland, Fast evaluation of radial basis functions: Methods based on partition of unity,
in Approximation Theory X: Wavelets, Splines, and Applications (edited by C. K. Chui,
L. L. Schumaker & J. Stöckler), pp. 473–83, Vanderbilt University Press, 2002.
[39] A. Wick, Generation of dynamic grids using structural analogy, pp. 147–157, Leipzig, 2001.
[40] A. Wick, Grid deformation techniques for two-dimensional hybrid grids, Tech. rep., 2001.
[41] D. Zeng & C. R. Ethier, A semi-torsional spring analogy model for updating unstructured
meshes in 3d moving domains, Finite Elements in Analysis and Design, vol. 41, pp. 1118–
11139, 2005.
[42] A. H. van Zuijlen, Fluid-structure interaction simulations - Efficient higher order time
integration of partitioned systems, Ph.D. thesis, Delft University of Technology, 2006.