Copyright

c 2006 Tech Science Press CMES, vol.12, no.2, pp.121-136, 2006

A Meshless Spatial Coupling Scheme for Large-scale Fluid-structure-interaction

Problems

R. Ahrem1 , A. Beckert2 , and H. Wendland3

Abstract: We present a new efficient scheme for loose Farhat (2003); Fornasier, Rieger, Tremel, and der Weide
coupling in fluid-structure-interaction problems as they (2002)). This is particularly crucial for highly critical
typically appear in the context of aircraft design. This flight conditions representing specific parts of the flight
coupling scheme is based upon a multivariate scattered envelope. The behavior of the elastic aircraft in flight
data interpolation approach, based on radial basis func- is influenced by the interaction between deformations of
tions and partition of unity methods. It allows us to cou- the elastic structure caused by fluid flow and the impact
ple arbitrary meshes on fluid and structure side. It con- of the aerodynamic forces on the structure. Hence, the in-
serves virtual work and forces. It is designed for large teraction between the fluid flow and the elastic structure
scale problems and allows the coupling of entire aircraft has to be studied in depth (Done (1996); Edward (1993);
meshes. Försching (1994, 1995)).
In principle, there are two different approaches to the
1 Introduction numerical solution of FSI problems. In a monolithic
approach, one tries to model the fluid, the structure,
In fluid-structure interaction (FSI) the reciprocal action and the interaction in one single model. While this
of a flexible structure with a flowing fluid, in which it is has from the mathematical point of view certain ad-
submersed or by which it is surrounded, is studied. Natu- vantages, its drawbacks in practical applications come
rally, FSI has applications in many fields of engineering, in particular from the fact that the new mathematical
such as the stability and response of aircrafts, the flow of model forces a completely new programming of the
blood through arteries, the vibration of turbine and com- solver. In particular, for complex aeroelastic problems,
pressor blades, and the response of bridges and tall build- the fluid and the structural domain show different mathe-
ings to winds. matical and numerical properties, requiring highly tuned
In this paper we study the mutual interaction between and adapted solvers. Hence, the simultaneous solution
aerodynamical and elastic forces for an aerospace vehi- by a monolithic scheme is in general computationally
cle. An airborne aircraft structure is subjected to sur- challenging, mathematically and economically subop-
face pressures induced by its surrounding flow. For an timal, and software-wise almost unmanageable (Farhat
introduction into the field of aeroelasticity we refer the and Lesoinne (1998)).
reader for example to Bisplinghoff and Ashley (1962); A more practical point of view is taken in the loose cou-
Fung (1955); Försching (1974); Dowell, Crawley, Jr., Pe- pling approach. Here, for each time step, the fluid and the
ter, Scanlan, and Sisto (1995). structural problem are solved independently, and the in-
In the development process of future air vehicles, mul- fluence of the other problem is restricted to the exchange
tidisciplinary simulation has become a key technology. of boundary conditions. To be more precise, a loose cou-
For the design and clearance of modern aircrafts, such pling scheme consists mainly of the iteration of the fol-
multi-disciplinary simulations are used to predict and lowing four steps:
to analyze the behavior of the elastic aircraft in flight
and during maneuver (Geuzaine, Brown, Harris, and 1. Compute the fluid solution and the resulting forces
on fluid side.
1 Institut für Numerische und Angewandte Mathematik, Universität
Göttingen 2. Transfer the forces to the structure.
2 European Aeronautic Defence and Space Company (EADS)
3 Institut für Numerische und Angewandte Mathematik, Universität 3. Compute the solution of the structure problem re-
Göttingen sulting in displacements on the structural side.
122 Copyright

c 2006 Tech Science Press CMES, vol.12, no.2, pp.121-136, 2006

4. Transfer the deformations to the fluid model. ∂(ρu)

+ ∇ · (ρu)u + ∇p = 0 (2)
∂t
This has the important advantage that problem-specific
models and existing solvers for both the fluid and the ∂(ρE)
+ ∇ · (ρu(E + p)) = 0 (3)
structure problem can still be used in FSI. However, since ∂t
now we cannot expect the underlying discretizations of
coming from the conservation of mass (1), of momentum
the single problems to match, special methods have to be
(2), and energy (3).
developed to exchange these boundary conditions.
Here, u denotes the velocity field, p the pressure, ρ the
In this paper we will concentrate on this interface prob-
density, and E = ρe + ρx22 /2 the energy (where e is
lem and describe a procedure which allows us to couple
the internal energy and x ∈ R3 the position vector). If, in
arbitrary meshes. In fact, the discretizations do not even
addition, T denotes the temperature then these unknowns
have to be based upon a mesh, since our coupling proce-
are related by an equation of state which, for atmospheric
dure is using only point-based information. Our method
flight conditions, is the perfect gas law
will be based upon radial basis function interpolation.
Recent reviews on meshless scattered data methods can, p
= RT [1 + α], (4)
for example, be found in Atluri (2004, 2005); Atluri and ρ
Shen (2005); Buhmann (2001); Wendland (2005).
where α = α(p, T ) is an effective mass fraction of di-
This paper is organized as follows. In the next section,
atomic molecules in the dissociated condition and R is
we describe the three field formulation for a typical FSI
the “undissociated” gas constant. If dissociation is ne-
problem. The third section is devoted to our coupling
glected, then (4) implies that the specific internal energy,
scheme, where special emphasis is put on efficiency and
e, and enthal py
the conservation of virtual work and forces. The final
section deals with a real-world example. Our coupling p
h = e+ (5)
scheme is applied to the spatial coupling of an entire air- ρ
craft model. We also compare our method to other ones
previously used. are functions of T only, which can be expressed by the
caloric equation of state for a calorically perfect gas:

2 The Three Field Formulation e = cv T. (6)

Fluid-structure-interaction problems are generally de- For more details see for example Bisplinghoff and Ash-
scribed in a so called three field formulation, referring to ley (1962).
the fluid domain, the structural domain and the interface In our application, theses equations have to be solved in
between both of them. the outer field around an aircraft, i.e. ΩF (t) = R3 \ ΩS (t).
To be more precise, let Ω(t) be a time dependent domain The solution of the Euler equations produces a pressure
of R3 . We will assume that Ω(t) = ΩF (t) ∪ ΩS (t) for all distribution on the structure, from which forces result,
/ Here, ΩF (t) denotes the do-
time and ΩF (t) ∩ ΩS (t) = 0. which act on the structure (in the case of an aircraft they
main occupied by the fluid at time t, while ΩS (t) denotes are responsible for the lift). However, for elastic struc-
the structural domain. Finally, let Γ(t) = ΩF (t) ∩ ΩS (t) tures these forces lead also to a deformation of (parts) of
be the fluid-structure interface. the structure, which, in practical computations, cannot be
On ΩF (t) the governing equations that describe the flow neglected.
are given by the Navier-Stokes or Euler equations. These These deformations are typically described by the Lamé
equations are based on the conservation of mass, momen- or Navier-Cauchy equations
tum, and energy. In our application, we will restrict our-
selves to the Euler equations for compressible flow, ρs̈ = µ∆s + (µ + λ)∇(∇ · s) + b, (7)

∂ρ where λ and µ are the Lamé constants and represent the

+ ∇ · (ρu) = 0 (1) elastic properties of the structure, s is the displacement
∂t
Coupling in fluid-structure-interaction 123

field, b are the body and external forces, and ρ is the 3 The Coupling Procedure
density of the body.
We will now describe our coupling procedure. This
In particular in aeroelastic computations, an eigenmode means that we have to describe two steps in the coupling
formulation of (7) is used. Suppose the displacements s process. The first step is concerned with transferring dis-
are expressed as placement vectors from the structural grid to the aerody-
∞
namical mesh, while the second step deals with transfer-
s(x,t) = ∑ qn (t)ψ ψn (x), ring forces from the aerodynamical mesh to the structural
n=1 one. Our coupling scheme will be based only upon point
information giving the most possible flexibility.
where the functions ψ n (x) are the natural eigenmodes of To make this more precise we will from now on use the
vibration of the structure and qn (t) are the Lagrange vari- following notation:
ables. Then, using only a finite number of Lagrange vari-
ables, (7) can be expressed as • Structural side:

N N – X = {x1 , . . ., xN } is the set of nodes,

∑ Mnpq̈ p (t) + ∑ Knp q p (t) = en (8) – f = (f(x1 ), . . ., f(xN )) are the forces that have
p=1 p=1
to be computed during the coupling process,
for 1 ≤ n ≤ N, where M = (Mnp ) is the mass and K = – g = (g(x1 ), . . ., g(xN )) are the displacements
(Knp ) the stiffness matrix, and en results from the body that are given by the structural solver.
and external forces. • Aerodynamical side:
Both problems have to be completed by appropriate
boundary and initial conditions. However, in the context – Y = {y1 , . . ., yM } is the set of nodes,
of FSI problems (some of) these boundary conditions are – F = (F(y1 ), . . ., F(yM )) are the forces that are
given by the other problem. To be more precise, on Γ(t) computed by the fluid solver,
we have to transfer – G = (G(y1 ), . . ., G(yM )) are the displace-
ments that have to be computed during the
• forces from the fluid problem, which are integrated coupling process.
pressures, to the structural domain,
Note that the involved forces and displacements
• displacements, resulting from the structural solver, f(x j ), g(x j ), F(y j), and G(y j ) are 3D vectors. Hence,
to the fluid domain. from now on fi = ( fi (x1 ), . . ., fi (xN ))T ∈ RN denotes the
ith component of the Matrix f, and so on.
This has to be done in a conservative way, such that vir-
3.1 Exchanging Forces
tual work and forces are globally preserved. We will de-
scribe this exchange of boundary conditions only for al- The entire coupling process can be described by employ-
ready discretized problems. ing a coupling matrix H ∈ RM×N , which relates the dis-
It is important to see that discretized models of the fluid placement vectors via
and structural problem differ in almost all practical appli-
Gi = Hgi , 1 ≤ i ≤ 3. (9)
cations. First of all, already the continuous model differ,
since, in general, the Navier-Stokes or Euler equations We will discuss how to derive such a coupling matrix in
are described in Eulerian coordinates while the struc- the next subsection. But for the time being, let us assume
tural problem is designed using Lagrangian coordinates. that we have such a coupling matrix H ∈ RM×N at hand.
Moreover, often, finite volume schemes are used to dis- Then, we can use H also to exchange the forces simply
cretize the fluid problem while finite element methods by setting
dominate discretization schemes for the structural prob-
lem. fi = H T Fi , 1 ≤ i ≤ 3. (10)
124 Copyright

c 2006 Tech Science Press CMES, vol.12, no.2, pp.121-136, 2006

This simple approach already guarantees conservation on the choice of the coupling matrix H. For example,
of virtual work and has been used often before (see for a global coupling matrix H built with radial basis func-
example Beckert (2000); Harder and Desmarais (1972); tions as proposed in Hounjet and Meijer (1994); Beckert
Hounjet and Meijer (1994)). However, to conserve also and Wendland (2001) leads often to an undesired and un-
forces, the following terminology will prove to be useful. physical distribution of forces on the structural side. Our
new local method will also remedy this kind of problem.
Definition 3.1. A coupling matrix H ∈ RM×N will be
called conservative if 3.2 Exchanging Displacements

HeN = eM In this section we describe how the coupling matrix H

can generally be built. This is done in the context of
where the k-dimensional vector ek is given by ek = scattered data interpolation, where only node informa-
(1, . . ., 1)T ∈ Rk . tion is required such that the coupling can be carried out
between arbitrary meshes.
For any coupling process of the form (9) and (10) it is
To be more precise, to transfer displacements from the
true that the virtual work is conserved. This, and the con-
structural model to the aerodynamical model we first as-
servation of forces is subject of
sume that each coordinate of the displacement field can
Theorem 3.2. Suppose the coupling process is based be modeled independently. This assumption is problem-
upon the coupling matrix H ∈ RM×N via (9) and (10). atic if additional information like rotations are given, but
Then, it conserves virtual work: it is appropriate as long as the deformations are small.
However, our approach can also be generalized to incor-
N M
porating rotations. But we will not pursue this here.
∑ fi(x j )gi(x j) = ∑ Fi(y j )Gi(y j ), 1 ≤ i ≤ 3. A radial basis function interpolant is a function of the
j=1 j=1
form
Moreover, if the coupling matrix H is conservative in the
N
sense of Definition 3.1, then the coupling process also
s(x) = ∑ α j Φ(x − x j ) + p(x).
conserves forces: j=1

N M
Here, Φ : R3 → R is a (conditionally) positive definite
∑ fi(x j ) = ∑ Fi(y j ), 1 ≤ i ≤ 3.
function of a certain order m and p ∈ πm−1 (R3 ) is a 3-
j=1 j=1
variate polynomial of degree at most m − 1. The coef-
Proof. For the first part, one simply has to realize that ficient vector α ∈ RN as well as the polynomial p are
determined by the interpolation conditions
fTi gi = (H T Fi )T gi = FTi Hgi = FTi Gi
s(x j ) = v j , 1 ≤ j ≤ N,
for any 1 ≤ i ≤ 3. The conservation of forces follows in
the same way. Since our coupling matrix H is supposed where v j = gi (x j ) are the given deformation for the ith
to be conservative, we have coordinate (i between 1 and 3 is from now on considered
N fixed) together with the additional conditions
∑ fi(x j) = fTi eN = (H T Fi)T eN N
j=1
M ∑ α j q(x j) = 0, for all q ∈ πm−1 (R3 ).
∑ Fi (y j)
j=1
= FTi HeN = FTi eM =
j=1
Typical examples of radial basis functions together with
for 1 ≤ i ≤ 3. their order m are given in Table 1. The names are ab-
breviations for Volume Spline, Thin-Plate Spline, Multi-
While a conservative coupling matrix has the just stated Quadric, Inverse Multi-Quadric, Gaussian, Wendland
conservation properties which are crucial to any coupling C0, C2, C4, and Euclid’s Hat, respectively. For W1–W3
process, the global approach using (10) depends heavily and EH we used the notation (r)+ = max{0, r}. Hence,
Coupling in fluid-structure-interaction 125

Table 1 : Conditionally positive definite functions tionally positive definite of order m and provided that the
set X is πm−1 (Rd ) unisolvent. The latter means that the
Φ(x) = φ(r) with r = x2 Name CPD only polynomial from πm−1 (Rd ) vanishing on X is the
r VS m≥1 zero polynomial.
r√2 logr TPS m≥2
2 +1 If we additionally define, for given evaluation points Y =
r√ MQ m≥1
{y1 . . ., yM } ⊆ Ω, the matrices AΦ,X,Y = (Φ(yi − x j )) ∈
1/ r2 + 1 IMQ m≥0
RM×N and PY = p j (yi ) ∈ RM×Q , then we can read off the
exp(−r2 ) G m≥0
coupling matrix H : RN → RM from
(1 − r)2+ W1 m≥0
(1 − r)4+ (1 + 4r) W2 m≥0
−1

  AΦ,X PX v
(1 − r)6+ (35r2 + 18r + 3) W3 m≥0 s|Y = AΦ,X,Y PY
π PXT 0 0
12 (r + 2)(1 − r)+ m ≥ 0.
2 EH

  v
=: 
H H (12)
0

Hence, for small data sets the coupling matrix H can be

W1–W3 and EH have a compact support. The support
precomputed and, in each step of the iterative coupling
radius can be simply changed by scaling φ(·/δ). Note
algorithm, the deformations and also the forces can be
that these compactly supported functions are only posi-
exchanged by a simple matrix by vector multiplication.
tive definite on Rd with d ≤ 3, which is sufficient for our
purposes, while all the others are conditionally positive The resulting coupling matrix is indeed conservative as
definite on every Rd . More information on the theoreti- long as constant polynomials are added to the interpolant.
cal background can be found in Wendland (2005). Here, Thus, this must (and indeed can) be done also for (uncon-
only two things matter. First, a smoother basis function ditionally) positive definite functions.
leads to a smother displacement field and hence to a more Lemma 3.3. Suppose the interpolation process by ra-
realistic deformation. But a smoother basis function of- dial basis function employs at least a constant polyno-
ten also suffers from numerical instability. Hence, the mial (i.e. m ≥ 1). Suppose further that the data sites
function should be chosen as smooth as possible, depend-
are πm−1 (R3 ) unisolvent. Then, the resulting coupling
ing on the given data set. matrix is conservative and the coupling (9) and (10) pre-
Second, a conditionally positive definite function of or- serves virtual work and forces.
der m is also conditionally positive definite of order
>
m. In particular, in the case of an unconditionally pos- Proof. In the light of Theorem 3.2 it suffices to show that
itive definite function, where no polynomial part for the the resulting coupling matrix H is conservative. How-
interpolant is necessary, it does not harm to add a polyno- ever, since the interpolant is unique, polynomials from
mial part. On the contrary, in the light of Definition 3.1 πm−1 (R3 ) are exactly recovered. This means that if the
all computations should be done using at least a constant data values are generated by such a polynomial the in-
polynomial. terpolant is the polynomial. Hence, if we have v j = 1
The resulting coupling matrix can be set up in the follow- for 1 ≤ j ≤ N, then the interpolant s is identical one, i.e.
ing way. Suppose p1 , . . ., pQ form a basis for the space s = 1 meaning in particular s(y j ) = 1 for all 1 ≤ j ≤ M.
of polynomials πm−1 (R3 ). Define AΦ,X = (Φ(x j − xk )) ∈ But this means that H is conservative.
RN×N and PX = (p j (xi )) ∈ RN×Q . Then, the coefficient
vector α can, together with the coefficient vector β ∈ RQ Note that the proof actually leads to a stronger con-
of p = ∑k βk pk , be determined from the linear system servation property. Employing polynomials of degree



 πm−1 (Rd ) leads to the moment condition
AΦ,X PX α v
= (11) N M
PXT β
0 0
∑ fi(x j )p(x j) = ∑ Fi(y j )p(y j)
j=1 j=1
It is well known (cf. Wendland (2005)) that the linear
system (11) has a unique solution provided Φ is condi- for 1 ≤ i ≤ 3, p ∈ πm−1 (R3 ).
126 Copyright

c 2006 Tech Science Press CMES, vol.12, no.2, pp.121-136, 2006

So far we have mainly described the method from Beck- interpolant is now formed by weighting the local inter-
ert and Wendland (2001); Harder and Desmarais (1972), polants. To this end, a partition of unity is chosen. This is
which has successfully been used for smaller models. a family of smooth, nonnegative functions {w j }Kj=1 such
However, problems arise if the size of the models be- that each w j is supported only on its associated patch Ω j ,
comes larger. This is mainly due to two facts. On the i.e. w j vanishes outside Ω j . Moreover, the weight func-
one hand, to compute the coupling matrix H we have to tions are supposed to sum up to one:
invert the (N + Q) × (N + Q) matrix in (11). Since Q is K
negligible when compared to N, the number of nodes on ∑ w j (x) = 1, x ∈ Ω.
an O (N 3 ) complexity. Multi-
the structural side,this has  j=1

plying the matrix A  P  with this inverse needs an- Then, the global interpolant is formed by
other O (NM) operations. For N and M too large both K
operations, even if they have to be done only once, need s(x) = ∑ w j (x)s j(x).
j=1
too much time. Moreover, for large M and N it is also
not reasonable, sometimes not even possible to keep the It inherits many of the properties of the local interpolants.
M × (N + Q) matrix H in the main memory of a modern For example, it is really an interpolant and it preserves
computer. For example, if we have N = 10, 000 structural the approximation orders of the local interpolants. More-
nodes and M = 1, 000, 000 dynamical surface nodes, stor- over, if the constant number of patches that contain a sin-
ing the matrix in double format using 8 bytes per num- gle evaluation point can be located in log(N) time, which
ber would need more than 74 gigabytes of memory. is the case for the above mentioned data structures, then
Hence, for larger models it is necessary to design an en- each evaluation also takes only log(N) time.
hanced coupling method for exchanging these data be- The evaluation of the interpolant can, as in the global
tween the two models. We will achieve this here, by com- case (12) be described by a matrix vector multiplica-
bining the radial basis function interpolation approach tion. Using the transposed of this new coupling matrix
with a partition of unity, see Wendland (2002). The idea for transferring forces from the aerodynamic to the struc-
behind this method can be described as follows. First, a tural model leads again to a conservative scheme.
region of interest Ω is defined. This region must contain Lemma 3.4. Suppose that on each patch Ωk the inter-
the node sets X and Y . Next, Ω is decomposed into a polation process employs at least a constant polynomial
number of overlapping subregions: Ω ⊆ ∪Kj=1 Ω j . This (i.e. m ≥ 1) and that the local data sets Xk = X ∩ Ωk are
decomposition has to be done in such a way that each πm−1 (Rd ) unisolvent. Suppose finally that H is the result-
patch Ω j contains only a small number N j of points, ing coupling matrix from our partition of unity approach.
which, when compared to N, can be considered to be Then, H is conservative and the coupling scheme (9) and
constant. Consequently, the number of patches K has to (10) preserves virtual work and forces.
be proportional to the number N. Moreover, it is impor-
tant that each y ∈ Y is contained in only a small number Proof. Given input data eN = (1, . . ., 1)T ∈ RN lead on
of patches and that these patches can be found fast. These every patch Ωk to local input data eNk , and hence to a lo-
requirements make it necessary to build “intelligent” data cal interpolant sk which is identical one. Thus, the global
structures for both the points in X and the patches Ω j . interpolant s satisfies
For each patch Ω j we collect the points from X in X j , i.e. s(x) = w (x)s (x) = w (x) = 1
X j = X ∩ Ω. Then, we solve the interpolation problem on
∑ k k ∑ k
k k
each patch in the previously described way, resulting in for all x ∈ Ω, meaning that the coupling matrix is conser-
a local interpolant s j . Since the number of points in each vative.
X j can be considered to be constant, this can be done in
constant time for each patch and, since we have O (N) Employing such a local strategy has the additional ad-
patches the overall time is also O (N) plus the time nec- vantage that forces and work are also locally preserved
essary to build the associated data structures. Such data resulting automatically into a more local distribution of
structures are in general based upon binary trees and can the forces, which will be demonstrated in the next sec-
be constructed in O (N logN) time. Finally, the global tion.
Coupling in fluid-structure-interaction 127

Figure 1 : The structural model

Figure 2 : The aerodynamical model

4 Numerical Results

In this section, we present computational results for the

just described aeroelastic problem in the static aeroelastic
equilibrium.
We applied our new coupling technique as well as the
global one based only on radial basis functions to Ale- Figure 3 : The difference between the structural and
nia’s SMJ model, which has been developed by Alenia aerodynamical model
Aeronautica and provided as a test case for the TAURUS
project. It describes a reference for a 70 seats passenger
aircraft. The difference between both models can best be seen
from Figure 3. Obviously, there are large regions, in par-
4.1 Alenia’s SMJ Model ticular on the wings, where no structural information is
available, which makes the coupling process difficult and
The structural model consists of a fuselage, fixed in its causes standard projection methods to fail. As a matter of
plane of symmetry, a wing structure, and a horizontal and fact, any interpolation method turns into an extrapolation
a vertical tail (see Figure 1). For the numerical compu- method.
tation of the static aeroelastic equilibrium of the aircraft,
symmetric boundary conditions have been imposed. 4.2 Results of the Aeroelastic Analysis
The discretized model consists of 5630 nodes and 12 388
In this section, we present the results of the aeroelastic
finite elements of different types. The structural finite el-
analysis compared to a rigid nonelastic computation.
ement equations are solved by an analysis code using the
formulation of the structural model in a generalized form. For simplicity, all computations were done with the CG
The related generalized coordinates, stiffness and mass (center of gravity) fixed. After convergence of the itera-
matrices result from a normal-modes solution, which has tion was reached, the deflected shape showed a maximum
been performed using the finite element analysis code vertical translation of 0.41 mt at the wing tip (upward)
MSC.Nastran (Msc (1987)). and a 0.15 mt downward deflection at the horizontal tail-
plane (see Figure 4).
For the aerodynamic model of the transport aircraft, the
fluid flow is described by the nonlinear three-dimensional The deformation of the structure leads to a decrease of
Euler equations. which are solved by a specific upwind- the total lift as shown in Figure 5. Beside this decreased
scheme based on finite volumes. In our situation, spatial lift, a different behavior of the spanwise Cp distribution
discretization is given by an unstructured finite volume at the wing tip is given.
mesh consisting of four million tetrahedrons. The surface This phenomenon can be explained by a vortex at the
mesh of the complete aircraft (see Figure 2) is based on trailing edge of the wing tip. In this region, the fluid par-
about 300 000 triangles. ticles undergo a rapid acceleration when passing the tip
128 Copyright

c 2006 Tech Science Press CMES, vol.12, no.2, pp.121-136, 2006

Figure 4 : Maximum deflections

and produce a local vortex that increases the Mach num-

ber to a maximum value of 1.4. Table 2 summarizes the
differences between the elastic and rigid computations in Figure 5 : Pressure Distribution
terms of lift and drag. Clearly, the influence of the elas-
ticity cannot be neglected.
• Load distribution
Table 2 : Influence of elasticity
• Computational time
elastic rigid ∆% • Boundary effects
Lift Cl 0.431 0.508 -15
Drag Cd 0.0299 0.0368 -18 4.3.1 Vertical Translation
We start with the vertical translation along the wing span.
The results for different basis functions are shown in
4.3 Comparison of the PoU and the Global Coupling Figure 6 for the global method and in Figure 7 for the
Method PoU method. The maximum difference between the ba-
sis functions is 7.5% in the global and 0.5% in the local
We will now compare the global coupling method based setting. Hence, the local method produces more reliable
on RBF interpolation with the local method using also results being almost independent of the chosen radial ba-
a partition of unity. Moreover, we analyze the influence sis function.
of the employed radial basis function. Significant differ-
ences in the interpolated displacement fields result into 4.3.2 Pressure Distribution
undesired different aerodynamical models and eventually
Next we take a look at the pressure distribution. Figures
into different aeroelastic equilibriums. Hence, to avoid
8 to 12 show the results for the global method and dif-
ambiguity a method which is indifferent with respect to
ferent basis functions. The results for the local method
the chosen basis function is preferable.
are presented in Figures 13 to 17. Again, in the case of
In detail, our comparison is based upon the following the global method significant differences appear. For ex-
quantities: ample, in the area of the wing near its root (where most
of the lift is generated), the results for W2 and TPS are
• Vertical translation along wing span
quite different, in particular where the flow expands be-
• Pressure distribution hind the leading edge. Moreover, in the case of TPS,
there is a thin region of low pressure, which is parallel to
• Variation of sweep angle (i.e. local angle of attack) the wing axis and extends from the tip to the mid span.
Coupling in fluid-structure-interaction 129

0.4
0.4

0.35
0.35 W2 C4
EH C4 EH C4
VS C4 W2 C4
0.3 0.3 TPS C4
W3 C4
TPS C4 VS C4
W1 C4 W1 C4
0.25 0.25
T3

T3
0.2 0.2

0.15 0.15

0.1
0.1

0.05
0.05

5 10
Y 5 10
Y

Figure 6 : Influence of the chosen RBF on the global

Figure 7 : Influence of the chosen RBF on local method
method

Figure 8 : Pressure distribution, global, TPS Figure 9 : Pressure distribution, global, VS

Figure 10 : Pressure distribution, global, W1 Figure 11 : Pressure distribution, global, W2

130 Copyright

c 2006 Tech Science Press CMES, vol.12, no.2, pp.121-136, 2006

Figure 12 : Pressure distribution, global, EH Figure 13 : Pressure distribution, PoU, TPS

Figure 14 : Pressure distribution, PoU, VS Figure 15 : Pressure distribution, PoU, W1

Figure 16 : Pressure distribution, PoU, W2 Figure 17 : Pressure distribution, PoU, EH

This region does not appear in the case of the other basis distribution is almost independent of the employed basis
functions. But it is visible for all basis functions in the function.
case of the local method. Moreover, again, the pressure
Coupling in fluid-structure-interaction 131

Global methods Local methods

0 0
-0.1 W2 C4 -0.1 EH C4
-0.2 EH C4 -0.2 W2 C4
VS C4 TPS C4
-0.3 W3 C4 -0.3 VS C4
-0.4 TPS C4 -0.4 W1 C4
W1 C4
-0.5 -0.5

-0.6 -0.6
alpha

alpha
-0.7
-0.7
-0.8
-0.8
-0.9
-0.9
-1
-1
-1.1
-1.1
-1.2
-1.2
-1.3
-1.3 -1.4
-1.4 -1.5
5000 10000 5000 10000
Y Y

Figure 18 : Angle of Attack: Global method Figure 19 : Angle of Attack: PoU method

X Y

X
Y

Figure 20 : Transformed forces on the FE model, global Figure 21 : Transformed forces on the FE model, global
method method

4.3.3 Variation of the Local Angle of Attack wing span).

Again, the local method behaves more uniformly over all
In order to analyze the torsional deformation of the wing basis functions, even in the case of smooth basis func-
in terms of a local angle of attack, we used a linear inter- tions.
polation between the vertical displacements of points at
the LE (leading edge) and the corresponding points at the 4.3.4 Load Distribution
TE (trailing edge). The results are presented in Figures
18 and 19 and confirm once again that the global method Even though the conservation of forces and work is guar-
is significantly dependent on the underlying basis func- anteed for each method, an advantage of the local algo-
tion. In particular for smoother basis functions (W2 and rithm comes from the local way in which it couples the
W3), the angle of attack represents a curve with a higher force fields.
concavity and, when approaching the wing tip, the slope In most cases, global methods result in nonsparse cou-
of the curve changes even its sign (at around 10 mt of pling matrices so that using the transposed matrix for ex-
132 Copyright

c 2006 Tech Science Press CMES, vol.12, no.2, pp.121-136, 2006

X Y

X
Y

Figure 22 : Transformed forces on the FE model, local Figure 23 : Transformed forces on the FE model, local
method method

changing forces will eventually destroy any local charac-

ter of the force distribution. This leads, in certain cases, 1
to a non-physical distribution of loads. Local methods
instead, result in a distribution which is similar to the
0.8
”physical” one. Global
Local
As shown in Figures 20, 21 (global) and 22, 23 (PoU),
0.6
residual ε

the global coupling algorithm returns a certain number

of concentrated high modulus forces, where most of the
lift is produced. On the other hand, with the local method 0.4

high peaks of forces are more uniformly distributed to the

leading and trailing edges of the lifting surfaces which is 0.2
physically more reasonable.

1 2 3 4 5 6 7 8
4.3.5 Computational Time iteration #

The new local method leads to convergence after 5 iter- Figure 24 : Convergence of static aeroelastic solution
ations with a residual of ε = 3.44e−4 . Here, the residual
is defined to be the relative difference in displacements
Table 3 : Computational Time
of two consecutive solution steps. For comparison, the
classical, global spatial coupling approach needed 6 iter- Times/Method Global PoU
ations to stop with an even worse residual of ε = 1.16e−3 .
Spatial Coupling [sec.] 5000 20
Figure 24 visualizes the residual over the iteration steps
CFD Solution [sec.] 7500 6700
and shows once again the superiority of the local method.
Total [sec.] 12700 7000
Table 3 shows the different computational times for both
methods. The aeroelastic computations have been per-
formed on a Linux workstation cluster, using 8 Intel
Xeon 2.66 GHz CPUs. The CFD solution was parti-
tioned into 6 domains and distributed on 6 CPUs, while each computed on one of the remaining CPUs.
the structural solution and the coupling algorithm were It is remarkable that even the time for the CFD solver im-
Coupling in fluid-structure-interaction 133

proves with the new local method. This is apparently due

to the fact that a physically more relevant distribution of
forces is obtained by the local method on the structural
mesh resulting in more reasonable deformations. The
latter produce a better deformation of the aerodynami-
cal mesh resulting also into an improved behavior of the
CFD solver.

4.3.6 Boundary Effects

In this subsection, we analyze the effects of the cou-
pling algorithm on the deformation of the wing pro-
file more thoroughly by looking at vertical displacement
along chordwise slices. These slices have been sampled
at different locations, namely 18%, 50%, 75% and 95%,
of the wing span.
Fig. 25 shows a typical behavior of the chordwise dis-
placement encountered in several computations using the Figure 26 : Real scale deformation.
global method. Such a displacement pattern is responsi-
ble for the deformation of the wing profile. Note that
the maximum displacement occurs in a location different
from the trailing edge (TE). Since points behind the lo-
cation of the maximum are subject to a smaller displace-
ment this results in a deformation of the profile. Even if
this is feasible it stands in contrast to the design criteria
of the specific model, which require the profile shape to
be conserved by means of the rib’s stiffness. To obtain a
rigid rotation, the maximum vertical displacement must
be attained at the TE.

Figure 27 : Real scale deformation, 10 x amplified.

Though evident on the 2D plot, this deformation can

be spotted from the 3D representation (Figure 26) only
through an amplification by a factor 10 as shown in Fig-
ure 27. This demonstrates that the differences in com-
puted displacements are quite small; their order is of frac-
tions of millimeters. However, the increase in the camber
of the profile produces an increase in the lift at the rear
section of the wing resulting in an increase of the pitch-
ing moment.
Figure 25 : Chordwise vertical translation pattern This phenomenon was encountered only when coupling
the meshes by the global method and mainly for the slice
134 Copyright

c 2006 Tech Science Press CMES, vol.12, no.2, pp.121-136, 2006

Frame 001 ⏐ 19 May 2004 ⏐ DISPLACEMENT at y/s = 0.18 Frame 001 ⏐ 19 May 2004 ⏐ DISPLACEMENT at y/s = 0.5

0.031

0.03 0.145
0.029

0.028 0.14
DISPLACEMENTZ

DISPLACEMENTZ
0.027
0.135
0.026

0.025
0.13
0.024
0.023 0.125
0.022
0.021 0.12

0.02
0.115
0.019
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1
x/c x/c

Figure 28 : Chordwise T3, TPS global method, 18% of Figure 29 : Chordwise T3, TPS global method, 50% of
wing span wing span
Frame 001 ⏐ 19 May 2004 ⏐ DISPLACEMENT at y/s = 0.7 Frame 001 ⏐ 19 May 2004 ⏐ DISPLACEMENT at y/s = 0.95

0.26
0.43

0.255
0.425
0.25
DISPLACEMENTZ

DISPLACEMENTZ

0.42
0.245

0.24 0.415

0.235
0.41

0.23
0.405
0.225

0.4
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1
x/c x/c

Figure 30 : Chordwise T3, TPS global method, 70% of Figure 31 : Chordwise T3, TPS global method, 95% of
wing span wing span

at 18% wing span. (see Figures 28 to 31). It seems that using our coupling libraries.
this is caused by the global character of the interpolant
and the resulting influence of interpolation points from
the fuselage. References
To eliminate this effect, the influence of different parts Atluri, S. N. (2004): The Meshless Method(MLPG)
of the aircraft has to be minimized. Figures 32 to 35 for Domain & BIE Discretizations. Tech Science Press,
illustrate that this is automatically achieved by the local Forsyth, USA. 680 pages.
method.

Acknowledgement: The results in the fourth section Atluri, S. N. (2005): Methods of Computer Modeling
were produced by Filippo Mattioni in Mattioni (2004) in Engineering & the Sciences, Volume I. Tech Science
Press, Forsyth, USA.
Coupling in fluid-structure-interaction 135

Frame 001 ⏐ 19 May 2004 ⏐ DISPLACEMENT at y/s = 0.18 Frame 001 ⏐ 19 May 2004 ⏐ DISPLACEMENT at y/s = 0.5

0.14

0.135
0.03

0.13
DISPLACEMENTZ

DISPLACEMENTZ
0.125

0.025
0.12

0.115

0.02
0.11

0.105
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1
x/c x/c

Figure 32 : Chordwise T3, W1 local method, 18% of Figure 33 : Chordwise T3, W1 local method, 50% of
wing span wing span
Frame 001 ⏐ 19 May 2004 ⏐ DISPLACEMENT at y/s = 0.7 Frame 001 ⏐ 19 May 2004 ⏐ DISPLACEMENT at y/s = 0.95

0.245
0.4
0.24
0.395
0.235
DISPLACEMENTZ

DISPLACEMENTZ

0.39
0.23

0.225 0.385

0.22 0.38

0.215
0.375

0.21
0.37
0.205
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1
x/c x/c

Figure 34 : Chordwise T3, W1 local method, 70% of Figure 35 : Chordwise T3, W1 local method, 95% of
wing span wing span

Atluri, S. N.; Shen, S. P. (2005): The basis of meshless radial basis functions. Aerosp. Sci. Technol., vol. 5, pp.
domain discretization: the meshless local petrov-galerkin 125–134.
(mlpg) method. Adv. Comput. Math., vol. 23, pp. 73–93.
Bisplinghoff, R. L.; Ashley, H. (1962): Principles of
Beckert, A. (2000): Coupling fluid (CFD) and structural
Aeroelasticity. John Wiley & Sons, Inc., New York.
(FE) models using finite interpolation elements. Aerosp.
Sci. Technol., vol. 1, pp. 13–22.
Buhmann, M. D. (2001): A new class of radial basis
Beckert, A.; Wendland, H. (2001): Multivariate in- functions with compact support. Math. Comput., vol. 70,
terpolation for fluid-structure-interaction problems using pp. 307–318.
136 Copyright

c 2006 Tech Science Press CMES, vol.12, no.2, pp.121-136, 2006

Done, G. T. S. (1996): Past and future progress in fixed MacNeal-Schwendler Corp.. MSC.Nastran - Hand-
and rotary wing aeroelasticity. Aeronautic Journal, vol. book for Aeroelastic Analysis I/II, 1987. ISSN 0741-
100, pp. 260–279. 8043.

Dowell, E. H.; Crawley, E. F.; Jr., H. C.; Peter, D. A.; Mattioni, F. (2004): Studies on the effects of global and
Scanlan, R. H.; Sisto, F. (1995): A Modern Course in local spatial coupling methods in static aeroelstic anal-
Aeroelasticity. Kluwer Acad. Pub., Dordrecht/Boston. ysis. Diploma thesis, Universita Romà ”La Sapienza”,
2004.
Edward, J. W. (1993): Computational aeroelasticity.
In Flight-Vehicles, Materials, Structures and Dynam- Wendland, H. (2002): Fast evaluation of radial basis
ics – Assessment and Future Directions, volume 5. The functions: Methods based on partition of unity. In Chui,
American Society of Mechanical Engineers, New York. C. K.; Schumaker, L. L.; Stöckler, J.(Eds): Approxima-
tion Theory X: Wavelets, Splines, and Applications, pp.
Farhat, C.; Lesoinne, M. (1998): Higher-order stag-
473–483, Nashville. Vanderbilt University Press.
gered and subiteration free algorithms for coupled dy-
namic aeroelasticity problems. In 36th Aerospace Sci- Wendland, H. (2005): Scattered Data Approximation.
ences Meeting and Exhibit,AIAA 98-0516, Reno/NV. Cambridge Monographs on Applied and Computational
Mathematics. Cambridge University Press, Cambridge,
Fornasier, L.; Rieger, H.; Tremel, U.; der Weide, E. V.
UK.
(2002): Time-dependent aeroelastic simulation of rapid
manoeuvring aircraft. In In 40th AIAA Aerospace Sci-
ences Meeting & Exhibit, 14–17 January 2002, Reno
Nevada. AIAA.

Försching, H. (1974): Grundlagen der Aeroelastik.

Springer, Heidelberg-Berlin-New York.

Försching, H. (1994): New ultra high capacity air-

craft (uhca) - challenges and problems from an aeroelas-
tic point of view. ZFW, vol. 18, pp. 219–231.

Försching, H. (1995): Challenges and perspectives in

computational aeroelasticity. In Proc. of the Interna-
tional Forum on Aeroelasticity and Structural Dynamics,
volume 1, pp. 1.1 – 1.9, Manchester (UK).

Fung, Y. C. (1955): An introduction to the theory of

aeroelasticity. John Wiley, New York.

Geuzaine, P.; Brown, G.; Harris, C.; Farhat, C.

(2003): Aeroelastic dynamic analysis of a full f-16 con-
figuration for various flight conditions. AIAA Journal,
vol. 41, pp. 363–371.

Harder, R. L.; Desmarais, R. N. (1972): Interpolation

using surface splines. Journal of Aircraft, vol. 9, pp.
189–197.

Hounjet, M. H. L.; Meijer, J. J. (1994): Evaluation of

elastomechanical and aerodynamic data transfer methods
for non-planar configurations in computational aeroelas-
tic analysis. Technical report, ICAS-Publication, 1994.