Numerical simulation of fluid–structure interaction using

the finite element method
P.R.F. Teixeira
a
, A.M. Awruch
b,
*
a
Department of Materials and Construction, Federal University of Rio Grande-FURG,
96201-900 Rio Grande, RS, Brazil
b
Graduate Program in Civil Engineering, Applied and Computational Mechanics Center,
Federal University of Rio Grande do Sul, 90035-190 Porto Alegre, RS, Brazil
Received 20 March 2002; received in revised form 8 December 2003; accepted 31 March 2004
Available online 20 July 2004
Abstract
An algorithm to simulate 3D fluid–structure interaction problems using the finite element technique is
presented in this work. A two-step Taylor–Galerkin scheme and linear tetrahedra elements are employed to
analyze the fluid flow, which may be high or slightly compressible. An arbitrary Lagrangean–Eulerian
(ALE) formulation is adopted, which must be compatible with the motion of the fluid–structure interface.
A fractional method with velocity correction is used for incompressible fluids. The structure is analyzed
using triangular elements with three nodes and six degrees of freedom in each node (three displacement
components and three rotation components). Geometrically non-linear effects are included. The Newmark
method is employed to integrate in time the dynamic equilibrium equations using an updated Lagrangean
description. The algebraic system of equations is solved using the conjugated gradient method and an
incremental-iterative scheme is used to solve the non-linear system resulting from finite displacements and
rotations. The code is optimized to take advantages of vector processors. Some cases studies have been
considered for validation of the computational algorithm. A two-dimensional supersonic flow over a
clamped flat plate is analyzed in order to study the aeroelastic behavior of this plate. Vibrations due to wind
action of an inflated membrane as well as vortex inducing vibrations in a panel immersed in a slightly
compressible fluid are also studied.
Ó 2004 Elsevier Ltd. All rights reserved.
*
Corresponding author. Tel.: +55-51-3316-3587; fax: +55-51-3316-3999.
E-mail address: amawruch@vortex.ufrgs.br (A.M. Awruch).
0045-7930/$ - see front matter Ó 2004 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compfluid.2004.03.006
Computers & Fluids 34 (2005) 249–273
www.elsevier.com/locate/compfluid
1. Introduction
Important progress has been obtained in the solution of complex fluid–structure interaction
problems in recent years, allowing the study of multidisciplinary applications in different engi-
neering areas. This progress is due mainly to the increase of the speed of modern computers, to the
evolution of multidisciplinary solution algorithms and of pre- and post-processing tools. The
methods of simulation of fluid–structure interaction problems are divided basically in two groups
called partitioned and monolithic schemes [1,2]. In the partitioned schemes the governing equa-
tions of the fluid and the structure are integrated in time alternately in an isolated way. In the
monolithic schemes the two fields are considered as a single entity, allowing to integrate in time
the two subdomains simultaneously. In partitioned schemes such as proposed by Soria and Ca-
sadei [3], Rifai et al. [4], Farhat et al. [5], Cebral and L€ ohner [6] and many other authors, the
kinematic and dynamic boundary conditions in the interface are the unique information changed
between the subdomains. In these schemes each sub-domain can be solved by discretization
techniques and by efficient solution algorithms in an individual way. New methods and models
can be introduced in a modular way in order to get more flexibility. Besides, the meshes of each
subdomain can be built without the coincidence of fluid and structural nodes at the interfaces
[6,7]. These modular characteristics and flexibility may be extremely convenient.
An algorithm to simulate fluid–structure interaction problems using a partitioned scheme is
presented in this work. A two-step explicit Taylor–Galerkin scheme [8,9], with linear tetrahedral
finite elements is employed. An arbritary Lagrangean–Eulerian (ALE) description is adopted for
the fluid domain, while for the structural domain an updated Lagrangean formulation is con-
sidered. The ALE description was first presented by Hirt et al. [10] and after this method was used
by several authors such as Hughes et al. [11], Donea et al. [12], Benson [13], Liu et al. [14] among
others. The structure is analyzed using generalized conforming triangular plates and shell ele-
ments with drilling degrees of freedom [15]. Geometrically non-linear effects are included. The
Newmark method [16] is employed to integrate in time the dynamic equilibrium equation. The
non-linear and the algebraic systems are solved using an incremental-iterative scheme and
the conjugate gradient method, respectively. The code was vectorized to take advantages of
vectorial processors. Several problems are used to validate the methods developed in this work
and illustrate the difference between linear and non-linear aeroelastic computations.
2. The fluid dynamic solver
2.1. The explicit two-step Taylor–Galerkin scheme for high compressible flows
In the ALE description, the computational frame is a reference independent of the particle
movement and may be moving with an arbitrary velocity in the laboratory system (this motion is
called the ‘‘mesh’’ motion in the finite element formulation); the continuum view from this ref-
erence is denoted as X
v
and the coordinates of any point are denoted as v [14]. The equations
expressing mass, momentum and energy conservation in ALE formulation may be written in a
compact form as
250 P.R.F. Teixeira, A.M. Awruch / Computers & Fluids 34 (2005) 249–273
oV
ot
¸
¸
¸
¸
v
þ
oFi
oxi
¼ wi
oV
oxi
ði ¼ 1; 2; 3Þ ð1Þ
with F
i
¼ F
iA
þ F
iD
and
V ¼
q
qv
1
qv
2
qv
3
qe
_
¸
¸
¸
¸
_
¸
¸
¸
¸
_
_
¸
¸
¸
¸
_
¸
¸
¸
¸
_
; F
iA
¼
qv
i
qv
1
v
i
þ pd
i1
qv
2
v
i
þ pd
i2
qv
3
v
i
þ pd
i3
v
i
ðqe þ pÞ
_
¸
¸
¸
¸
_
¸
¸
¸
¸
_
_
¸
¸
¸
¸
_
¸
¸
¸
¸
_
; F
iD
¼
0
Às
1i
Às
2i
Às
3i
Às
ij
v
j
À k
ij
oT=ox
j
_
¸
¸
¸
¸
_
¸
¸
¸
¸
_
_
¸
¸
¸
¸
_
¸
¸
¸
¸
_
ði; j ¼ 1; 2; 3Þ ð2Þ
where v
i
and w
i
are the fluid and the mesh velocity components in the direction of the spatial
coordinate x
i
respectively, q is the density, p is the thermodynamic pressure, s
ij
are the compo-
nents of the deviatoric stress tensor, T is the temperature, e is the total specific energy and k
ij
are
the components of the conductivity tensor; vector V contains the conservation or the field vari-
ables and F
i
are the components of the flux variables (F
iA
contains the advective terms and F
iD
the
viscous terms). Finally, d
ij
is the Kronecker delta. Eq. (1) is complemented by the equation of state
for an ideal gas and by the constitutive equations. Initial and boundary conditions must be added
to these equations in order to define uniquely the problem.
In the Taylor–Galerkin scheme, conservation equations are expanded in time by Taylor series,
and after, space discretization is accomplished by the classical Bubnov–Galerkin scheme [16–18].
A two-step method is used, as indicated by Zienkiewicz et al. [8]; this approach can be interpreted
as the finite element version of the Lax–Wendroff scheme used in finite differences [19]. In the first
step, corresponding to the time interval ½t
n
; t
nþ1=2
Š, the unknown vector V at t ¼ t
nþ1=2
is expanded
in Taylor series. Using a linear shape function N associated with each node to interpolate V
n
, a
constant shape function P
E
¼ 1 associated with element E to interpolate V
nþ1=2
and applying the
classical Galerkin weighted residual method to the expression resulting from Taylor series
expansion, the following equation is obtained:
X
nþ1=2
E
V
nþ1=2
E
¼
_
X
n
E
NdX
_ _
V
n
À
Dt
2
_
X
n
E
oN
ox
i
dX
_ _
F
n
i
_
À
_
X
n
E
Nw
n
i
oN
ox
i
dX
_ _
V
n
_
ði ¼ 1; 2; 3Þ ð3Þ
where X
nþ1=2
E
is the element volume, V
nþ1=2
E
is a constant value at element level and the upper bar
indicates nodal variables.
In the second step, the unknown vector at t ¼ t
nþ1
is expanded in Taylor series. Using again the
same shape functions, and applying the Bubnov–Galerkin method, the following equation is
obtained for the second step:
_
X
nþ1
E
N
T
NdX
_ _
V
nþ1
¼
_
X
n
E
N
T
NdX
_ _
V
n
þDt
_
X
nþ1=2
E
oN
ox
i
dX
_ _
F
nþ1=2
i
_
À w
nþ1=2
i
V
nþ1=2
_
À Dt
_
C
nþ1=2
E
Nl
i
dC
_ _
F
nþ1=2
Bi
_
À w
nþ1=2
Bi
V
nþ1=2
B
_
ði ¼ 1; 2; 3Þ ð4Þ
P.R.F. Teixeira, A.M. Awruch / Computers & Fluids 34 (2005) 249–273 251
where C
nþ1=2
E
is the boundary of the element domain X
nþ1=2
E
and l
i
is the cosine of the angle formed
by the outward normal axis to C
nþ1=2
E
with the positive direction of the reference axis x
i
. Index B is
referred to values at the boundary of the element domain. In Eq. (4), the consistent mass matrix is
substituted by the lumped mass matrix, and then this equation is solved iteratively. The proposed
scheme is conditionally stable, and the local stability condition is applied.
In order to stabilize numerically the solution, specially in the presence of strong shocks, it is
necessary to add numerical damping to the flow solver. In this work the viscosity model, as given
by Peraire et al. [20], is adopted. An artificial viscosity is added explicitly to the non-smoothed
solution, as follows
V
nþ1
s
¼ V
nþ1
þ ðM
nþ1
L
Þ
À1
D ð5Þ
where V
nþ1
s
and V
nþ1
are the smoothed and non-smoothed solution at t ¼ t
nþ1
respectively. M
nþ1
L
is
the assembled lumped mass matrix at t ¼ t
nþ1
The vector D is given by
D ¼

E
CFL
E
CCS
E
½M
n
E
C
À M
n
E
L
ŠV
n
E
ð6Þ
where E is an index referred to a specific element, CFL
E
is the local Courant number, CC is a
global constant specified by the user, S
E
is a coefficient of pressure distribution (see Refs. [8,9]),
M
n
E
C
and M
n
E
L
are the consistent and lumped mass matrices, respectively. The constant CC must be
specified with care in order to avoid interferences of the artificial and physical viscosities. CC
varies between 2.0 and 10.0.
2.2. The two-step Taylor–Galerkin scheme for slightly compressible flows
Mass conservation for slightly compressible fluids, assuming constant entropy, may be ex-
pressed by the following equation:
oq
ot
¼
1
c
2
op
ot
¼ À
oU
i
ox
i
ði ¼ 1; 2; 3Þ ð7Þ
where c is the sound speed and U
i
¼ qv
i
ði ¼ 1; 2; 3Þ.
Expanding the momentum conservation equations in Taylor series, the following expression is
obtained for the first step:
U
nþ1=2
i
¼ U
n
i
þ
Dt
2
oU
n
i
ot
¼ U
n
i
À
Dt
2
of
n
ij
ox
j
_
À
os
n
ij
ox
j
þ
op
n
ox
i
þ
1
2
oDp
ox
i
À w
n
j
oU
n
i
ox
i
_
ði; j ¼ 1; 2; 3Þ
ð8Þ
where f
ij
¼ v
j
U
i
ði; j ¼ 1; 2; 3Þ and p
nþ1=2
¼ p
n
þ 1=2Dp, with Dp ¼ p
nþ1
À p
n
. Using
¯
U
nþ1=2
i
¼ U
n
i
À
Dt
2
of
n
ij
ox
j
_
À
os
n
ij
ox
j
þ
op
n
ox
i
À w
n
j
oU
n
i
ox
i
_
ði; j ¼ 1; 2; 3Þ ð9Þ
Eq. (8) is given by the following expression:
U
nþ1=2
i
¼
¯
U
nþ1=2
i
À
Dt
4
oDp
ox
i
ði ¼ 1; 2; 3Þ ð10Þ
252 P.R.F. Teixeira, A.M. Awruch / Computers & Fluids 34 (2005) 249–273
Discretizing Eq. (7) in time and using Eq. (10), it is obtained:
Dq ¼
1
c
2
Dp ¼ ÀDt
oU
nþ1=2
i
ox
i
¼ ÀDt
o
¯
U
nþ1=2
i
ox
i
_
À
Dt
4
o
ox
i
oDp
ox
i
_
ði ¼ 1; 2; 3Þ ð11Þ
The second time step is given by the following expression:
U
nþ1
i
¼ U
n
i
þDt
oU
nþ1=2
i
ot
¼ U
n
i
ÀDt
of
nþ1=2
ij
ox
j
_
À
os
nþ1=2
ij
ox
j
þ
op
nþ1=2
ox
i
À w
nþ1=2
j
oU
nþ1=2
i
ox
i
_
ði; j ¼ 1; 2; 3Þ ð12Þ
Then the flow is analyzed, after space discretization, by the following algorithm: (1) determine
¯
U
nþ1=2
i
with Eq. (9); (2) determine Dp with Eq. (11) and calculate p
nþ1
¼ p
n
þ Dp; (3) determine
U
nþ1=2
i
with Eq. (10); (4) determine U
nþ1
i
with Eq. (12).
Considering the same shape functions used in compressible flows and applying the classical
Galerkin method for space discretization, the following matrix expressions are obtained for Eqs.
(9), (11), (10) and (12), respectively:
X
nþ1=2
E
¯
U
nþ1=2
iE
¼
_
X
n
NdX
_ _
U
n
i
À
Dt
2
_
X
n
oN
ox
j
dX
_ _

f
n
ij
_
À
_
X
n
oN
ox
j
dX
_ _
s
n
ij
þ
_
X
n
oN
ox
i
dX
_ _
p
n
À
_
X
n
Nw
n
j
oN
ox
j
dX
_ _
U
n
i
_
ði; j ¼ 1; 2; 3Þ ð13Þ
_
X
nþ1=2
N
T
1
c
2
NdX
_
þ
Dt
2
4
_
X
nþ1=2
oN
T
ox
i
oN
ox
i
dX À
_
C
nþ1=2
N
T
oN
ox
i
n
i
dX
_
Dp
¼ Dt
_
X
nþ1=2
oN
T
ox
i
dX
_ _
¯
U
nþ1=2
iE
_
À
_
C
nþ1=2
N
T
n
i
dC
_ _
¯
U
nþ1=2
iE
_
ði ¼ 1; 2; 3Þ ð14Þ
X
nþ1=2
E
U
nþ1=2
iE
¼ X
nþ1=2
E
¯
U
nþ1=2
iE
À
Dt
4
_
X
nþ1=2
oN
ox
i
dX
_ _
Dp ði ¼ 1; 2; 3Þ ð15Þ
_
X
nþ1
N
T
NdX
_ _
U
nþ1
i
¼
_
X
n
N
T
NdX
_ _
U
n
i
þDt
_
X
nþ1=2
oN
T
ox
j
dX
_ _
f
nþ1=2
ij
_
À w
nþ1=2
j
U
nþ1=2
i
_
E
À Dt
_
X
nþ1=2
oN
T
ox
j
NdX
_ _
s
n
ij
þDt
_
X
nþ1=2
oN
T
ox
i
NdX
_ _
p
n
_
þ Dp=2
_
À Dt
_
C
nþ1=2
N
T
n
j
dC
_ _
ðf
nþ1=2
ij
À w
nþ1=2
j
U
nþ1=2
i
Þ
E
þ Dt
_
C
nþ1=2
N
T
Nn
j
dC
_ _
s
n
ij
ÀDt
_
C
nþ1=2
N
T
Nn
i
dC
_ _
_
p
n
þ Dp=2
_
ði; j ¼ 1; 2; 3Þ ð16Þ
P.R.F. Teixeira, A.M. Awruch / Computers & Fluids 34 (2005) 249–273 253
In Eqs. (13)–(16) the index E indicates that the corresponding variables are taken with a constant
value over the element domain. Eq. (14) is solved using the conjugate gradient method with
diagonal pre-conditioning. In this equation the boundary integral involving the pressure incre-
ment in the left hand side could be omitted. One of the reason for this simplification is apparent if
Eq. (10) is examined. It is observed that on boundaries where the velocities are prescribed
oDp=ox
i
¼ 0. The other reason is that on traction boundaries, no change of traction occurs and
then Dp ¼ 0 [18].
2.3. The algorithm for mesh movement
The mesh velocity field w is computed looking for small element distortions, conserving pre-
scribed velocities in moving and stationary boundary surfaces. The mesh movement algorithm
adopted in this work uses a smoothing procedure for the velocities based in these boundary
surfaces. The updating of the mesh velocity at a point i of the finite element domain is based on
the mesh velocity of the points j belonging to the boundary surfaces in the following way
(see Fig. 1):
w
i
¼

ns
j¼1
a
ij
w
j

ns
j¼1
a
ij
ð17Þ
where ns is the total number of points belonging to the boundary surfaces and a
ij
are the influence
coefficients between the point i inside the domain and the point j of the boundary surface given by
the following expression:
a
ij
¼
1
d
4
ij
ð18Þ
with d
ij
being the distance between the points i and j. In other words, a
ij
represents the weight that
each point j of the boundary surface has on the value of the mesh velocity at points i inside the
domain. When d
ij
is small, a
ij
has an high value, favouring the influence of points i, located closer
to the boundary surface containing point j.

Fig. 1. Distances from the boundary surfaces to a point ‘‘i’’ in the fluid domain.
254 P.R.F. Teixeira, A.M. Awruch / Computers & Fluids 34 (2005) 249–273
In the present work a variation to simplify this method and safe computer time was adopted for
the case in which the displacement of the body surface is not significantly high. It consists of
choosing previously for each point inside the fluid domain at least two surface points having
important influence in the calculation of the mesh velocity according to Eq. (17). One of these
points belongs to the moving body surface and the other one to the external fixed surface. Each of
these two points have the largest values of coefficient a
ij
among all the other points in its respective
surface because each one is the closest point of the respective surface with respect to the node
where the mesh velocity will be determined. Therefore, this procedure uses the same Eq. (17), but
considering only one point in each boundary surface, corresponding to the node which have the
smallest distance with respect of the point inside the fluid domain where the mesh velocity will be
computed. The method can be generalized when there are two or more body surfaces. In this case,
it must be used a point of each body surface to calculate the mesh velocities field inside the fluid
domain.
3. The structural dynamic solver with a triangular thin plate element
A generalized conforming triangular thin plate element with six degrees of freedom in each
node (three displacements and three rotations components), in which the compatibility conditions
at each node and along each side are applied, is used in the present work (see Fig. 2) [15].
The total stiffness matrix of the element is obtained by the overlap of the membrane stiffness
matrix with the bending stiffness matrix. A typical membrane triangular element is adopted where
each point has two degrees of freedom of translation u
xi
and u
yi
ði ¼ 1; 2; 3Þ and one of rotation h
zi
ði ¼ 1; 2; 3Þ in the plane of the element middle surface. The membrane displacements are expressed
in the following way:
u
m
¼ ½ u
x
u
y
Š
T
¼ N
m
u
e
m
ð19Þ
where u
e
m
is the nodal membrane generalized displacements vector given by

Fig. 2. The thin plate triangular element.
P.R.F. Teixeira, A.M. Awruch / Computers & Fluids 34 (2005) 249–273 255
u
e
mi
¼ u
xi
u
yi
h
zi
½ Š
T
ði ¼ 1; 2; 3Þ ð20Þ
and N
m
is the membrane interpolation function defined as follows:
N
mi
¼
L
i
0 N
uhi
0 L
i
N
vhi
_ _
ði ¼ 1; 2; 3Þ ð21Þ
being L
i
the area coordinates and
N
uhi
¼
1
2
L
i
ðb
m
L
j
À b
j
L
m
Þ; N
vhi
¼
1
2
L
i
ðc
m
L
j
À c
j
L
m
Þ;
b
i
¼ y
j
À y
m
; c
i
¼ x
m
À x
j
ði; j; m ¼ 1; 2; 3Þ ð22Þ
The transverse displacement field is discretized by
u
z
¼ N
b
u
e
b
ð23Þ
where u
e
b
is the nodal bending generalized displacements vector, which is written in the following
way:
u
e
bi
¼ ½ u
zi
h
xi
h
yi
Š
T
ði ¼ 1; 2; 3Þ ð24Þ
being u
zi
the nodal transverse displacements, and h
xi
and h
yi
the nodal rotations around the axes x
and y, respectively. N
b
is the bending interpolation function given by
N
bi
¼ ½ N
i
N
xi
N
yi
Š
T
ði ¼ 1; 2; 3Þ ð25Þ
with
N
i
¼ L
i
À 2F
i
þ ð1 À r
j
ÞF
j
þ ð1 þ r
m
ÞF
m
N
xi
¼ À
1
2
½b
m
L
i
L
j
À b
j
L
m
L
i
þ ðb
j
À b
m
ÞF
i
þ ðr
j
b
j
þ b
m
ÞF
j
þ ðr
m
b
m
À b
j
ÞF
m
Š
N
yi
¼ À
1
2
½c
m
L
i
L
j
À c
j
L
m
L
i
þ ðc
j
À c
m
ÞF
i
þ ðr
j
c
j
þ c
m
ÞF
j
þ ðr
m
c
m
À c
j
ÞF
m
Š
ð26Þ
F
i
¼ L
i
ðL
i
À 0:5ÞðL
i
À 1Þ; r
i
¼
1
l
2
jm
l
2
im
_
À l
2
ij
_
; l
ij
¼
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
x
2
ij
þ y
2
ij
_
;
x
ij
¼ x
i
À x
j
; y
ij
¼ y
i
À y
j
ði; j; m ¼ 1; 2; 3Þ
The linear stiffness matrix at element level due to membrane and bending effects are given,
respectively, by the following expressions:
K
Lmi
¼
_
X
ðeÞ
B
T
mi
D
m
B
mi
dX; K
Lbi
¼
_
X
ðeÞ
B
T
bi
D
b
B
bi
dX ði ¼ 1; 2; 3Þ ð27Þ
where X
ðeÞ
is the element domain. The constitutive matrices D
m
and D
b
are given, respectively, by
D
m
¼
Eh
1 À m
2
D; D
b
¼
Eh
3
12ð1 À m
2
Þ
D; D ¼
1 m 0
m 1 0
0 0 ð1 À mÞ=2
_
_
_
_
ð28Þ
where h is the elements thickness, m the Poisson’s ratio and E the Young’s modulus. Finally, the
strain–displacement relations are
256 P.R.F. Teixeira, A.M. Awruch / Computers & Fluids 34 (2005) 249–273
B
mi
¼
1
4X
ðeÞ
2b
i
0 b
i
ðb
m
L
j
À b
j
L
m
Þ
0 2c
i
c
i
ðc
m
L
j
À c
j
L
m
Þ
2c
i
2b
i
ðc
i
b
m
þ b
i
c
m
ÞL
j
À ðc
i
b
j
þ b
i
c
j
ÞL
m
_
_
_
_
ði; j; m ¼ 1; 2; 3Þ ð29Þ
and
B
bi
¼ À
N
i;xx
N
xi;xx
N
yi;xx
N
i;yy
N
xi;yy
N
yi;yy
2N
i;xy
2N
xi;xy
2N
yi;xy
_
_
_
_
ði ¼ 1; 2; 3Þ ð30Þ
The non-linear geometric stiffness matrix is given by
K
NLi
¼
_
X
ðeÞ
G
T
Gi
T
i
G
Gi
dX ði ¼ 1; 2; 3Þ ð31Þ
where T
i
contains the membrane internal forces and is given by
T
i
¼
T
xxi
T
xyi
T
yxi
T
yyi
_ _
ði ¼ 1; 2; 3Þ ð32Þ
with
T
xxi
¼

3
k¼1
D
m
ð1; kÞ Á B
mi
ðk; 1Þ
_ _
Á u
mi
ð1Þ
T
yyi
¼

3
k¼1
D
m
ð2; kÞ Á B
mi
ðk; 2Þ
_ _
Á u
mi
ð2Þ ði ¼ 1; 2; 3Þ
T
xyi
¼

3
k¼1
D
m
ð3; kÞ Á B
mi
ðk; 3Þ
_ _
Á u
mi
ð1Þ ¼ T
yxi
ð33Þ
while G
Gi
is given by
G
Gi
¼
N
i;x
N
xi;x
N
yi;x
N
i;y
N
xi;y
N
yi;y
_ _
ði ¼ 1; 2; 3Þ ð34Þ
The dynamic equilibrium equations of the structural system are obtained by the space discreti-
zation of the virtual work equations, which must be solved for each time step using a numerical
integration scheme. In the present work the well-known Newmark method is used [16], with
parameters d ¼ 1=2 and a ¼ 1=4. In geometrically non-linear problems the solution of these
balance equations is accomplished for each time step with an incremental iterative procedure
using an updated Lagrangean formulation. The resulting system of algebraic equations are solved
using the gradient conjugate method with incomplete Cholesky factorization [21,22].
4. The fluid–structure interaction algorithm
Commonly, fluid and structure fields have different scales of time. The global time step is
usually commanded by the fluid. Although the use of the same time step for the fluid and the
P.R.F. Teixeira, A.M. Awruch / Computers & Fluids 34 (2005) 249–273 257
structure may provide some implementation advantages, the procedure with subcycles of factor
n
SF
¼ Dt
S
=Dt
F
(where Dt
S
and Dt
F
are the time intervals adopted for the structure and the fluid,
respectively) may offer substantial computational advantages, including an economy of CPU time
due to the smaller number of time steps in the structural analysis and an economy in the infor-
mation transfer. The partitioned algorithm with subcycles adopted in the present work consists in
the following steps (see Fig. 3):
(a) Set the initial conditions for the structure and the fluid.
(b) Update the structure displacements, velocities, accelerations and stresses.
(c) Update the fluid velocities, specific mass, pressure and total energy using subcycles.
(c.1) Compute the new mesh taken into account the structural motion.
(c.2) Update the fluid flow variables with the new boundary conditions.
(d) Update the structural variables with the loads transferred by the fluid.
(e) Repeat steps (c) and (d) until the objectives of the simulation are reached.
Kinematics boundary conditions, taken into account the structural motion, must be applied to
the fluid domain. Therefore, at the fluid–structure interface the following condition must be
satisfied:
x
nþ1
i
¸
¸
C
F
¼
tþDt
x
i
¸
¸
C
E
ði ¼ 1; 2; 3Þ ð35Þ
where x
nþ1
i
j
C
F
and
tþDt
x
i
j
C
E
are the updated coordinates for the instant t þ Dt at the fluid and solid
interfaces C
F
and C
E
, respectively. For viscous fluids the same velocity components are prescribed
for the mesh, the fluid and the structure at the interface. Therefore,
w
nþ1
i
¸
¸
C
F
¼ v
nþ1
i
¸
¸
C
F
¼
tþDt
_ u
i
¸
¸
¸
C
E
ði ¼ 1; 2; 3Þ ð36Þ
For non-viscous fluids the corresponding prescribed boundary conditions are:
w
nþ1
i
¸
¸
C
F
Á n
i
¼ v
nþ1
i
¸
¸
C
F
Á n
i
¼
tþDt
_ u
i
¸
¸
¸
C
E
Á n
i
ði ¼ 1; 2; 3Þ ð37Þ
where n
i
is the normal vector with respect to the interface surface at time t þ Dt. Eq. (37) is applied
to each node located in the interface surface.

Fig. 3. Algorithm for fluid–structure interaction with subcycles.
258 P.R.F. Teixeira, A.M. Awruch / Computers & Fluids 34 (2005) 249–273
5. Numerical applications
5.1. Supersonic inviscid flow over a flat plate
In this example a supersonic inviscid flow over a flat plate, which is clamped on both ends, as
shown in Fig. 4, is analyzed. The plate length, width and thickness are 0.5 m, 0.2 m and 1.35 mm,
respectively. The density of the fluid is q
F
¼ 0:339 kg/m
3
, while the plate has the following
properties: specific mass q
E
¼ 2710 kg/m
3
, Young’s modulus E ¼ 77:28 GPa and Poisson’s ratio
m ¼ 0:33. In this problem, a value equal to 5.0 was adopted for CC, defined in Eq. (6).
Initially the plate is subject to the same pressure on both sides. It is assumed that the pressure
on the inferior face of the plate falls 0.1% instantaneously and remains in this condition for 4 ls,
originating displacements in the plate and disturbances in the flow. After this period, this pressure
x
y
z
v
2
=0.0
v
2
=0.0
v.n=0
v
3
=0.0
ρ

, v
i∞
, ε

Fig. 4. Geometry and boundary conditions in the fluid domain for the example of a supersonic inviscid flow over a flat
plate.
Fig. 5. General view of the finite element mesh for the fluid domain.
P.R.F. Teixeira, A.M. Awruch / Computers & Fluids 34 (2005) 249–273 259
0 0.02 0.04 0.06 0.08 0.1
-0.002
-0.0015
-0.001
-0.0005
0
0.0005
0.001
0.0015
0.002
Mach = 1.9
Mach = 2.0
Mach = 2.02
Mach = 2.05
displacement (m)
in x=0.35m
time (s)
Fig. 6. Displacements of the plate in x ¼ 0:35 m along the time, for Mach ¼1.9, 2.0, 2.02 and 2.05.
0 0.02 0.04 0.06 0.08 0.1
-0.002
-0.0015
-0.001
-0.0005
0
0.0005
0.001
0.0015
0.002
Mach = 2.0
Rifai et al., 1999
displacement (m)
in x=0.35m
time (s)
Fig. 7. Comparison of the displacements in x ¼ 0:35 m obtained by the present work for Mach¼2.0 and by Rifai et al.
[4] for Mach¼1.98.
260 P.R.F. Teixeira, A.M. Awruch / Computers & Fluids 34 (2005) 249–273
returns to its initial value of 28 kPa. The solution of the linear aeroelastic problem indicates a limit
of stability with a critical Mach number equal to 2.0 using the piston theory [23]; above this value,
continuous growth of oscillations amplitudes are observed and instability of the plate occurs.
The domain discretization is accomplished such that the fluid tetrahedral faces adjacent to the
plate are coincident with the triangles belonging to the structure, although this fact would not be
necessary for the solution scheme. 13303 nodes and 38915 tetrahedral elements for the mesh of the
fluid domain and 506 nodes and 508 triangular elements for the structural mesh are used. This
inviscid flow has a two-dimensional behavior, only one layer of elements is adopted to discretize
the domain in the direction of the axis z, transverse to the flow. A general view of the mesh is
shown in Fig. 5. The boundary conditions imposed to the fluid are shown in Fig. 4, while in the
plate are imposed the following conditions: u ¼ v ¼ w ¼ hx ¼ hy ¼ hz ¼ 0:0 in the clamped ends
and w ¼ hx ¼ hy ¼ 0:0 in the plane of symmetry xy.
Fig. 8. Pressure distribution for Mach ¼2.0 at t ¼ 0:07 s.
Fig. 9. Configurations of the clamped plate for Mach ¼2.0 at t ¼ 0:03, 0.07 and 0.1 s.
P.R.F. Teixeira, A.M. Awruch / Computers & Fluids 34 (2005) 249–273 261
In a first analysis, simulations are accomplished with Mach numbers around the critical value
predicted by the analytical solution using the piston theory [22]. Results of the displacement of
the plate in x ¼ 0:35 m along the time are shown in Fig. 6 for Mach numbers ¼1.9, 2.0, 2.02 and
2.05. These results were obtained using to analyze the plate time steps equal to 1.8 ·10
À7
s for
0 0.02 0.04 0.06 0.08 0.1
-0.02
-0.015
-0.01
-0.005
0
0.005
0.01
0.015
0.02
Linear
Non linear
displacement (m)
in x=0.35m
time (s)
Fig. 10. Comparison of the displacements in x ¼ 0:35 m between the geometrically linear and non-linear solutions with
Mach¼2.3.
Fig. 11. Description of the problem to study the interaction between air flow and a flexible membrane.
262 P.R.F. Teixeira, A.M. Awruch / Computers & Fluids 34 (2005) 249–273
Mach¼1.9 and 1.75 · 10
À7
s for Mach¼2.0, 2.02 and 2.05. It is observed that the critical Mach
number obtained here is practically equal to the value predicted by the solution using the piston
theory (Mach¼2.0). In Fig. 7 these results are compared with those obtained by Rifai et al. [4] (in
this work a critical Mach number equal to 1.98 was found). It is noticed that, although a little
discrepancy stays among these results, the amplitudes are very close. The pressure distribution in
the domain for a Mach number ¼2.0 at the instant t ¼ 0:07 s is shown in Fig. 8. Fig. 9 presents
the plate configurations and values of vertical displacements at t ¼ 0:03, 0.07 and 0.1 s.
A comparison between the linear and the non-linear behaviour of the plate is performed for
Mach number ¼2.3 and a time step Dt ¼ 1:6 Â 10
À7
s, whose displacements in x ¼ 0:35 m are
shown in Fig. 10. The geometrically linear solution, as well as the solution based in the piston
theory [22], presents an exponential growth of the oscillations, while the geometrically non-linear
solution presents limited displacement amplitudes. The difference between the two solutions is
attributed to the coupling between membrane and bending effects in the panel when non-linear
elasticity due to finite displacements is considered. This phenomenon increases the effective
stiffness of the plate, modifying the behaviour and the response of the system.
Fig. 12. Finite element mesh. (a) General view. (b) Detail of the region close to the membrane.
Table 1
Air and membrane properties
Membrane properties Value Air properties Value
Young’s modulus, E 3.333 ·10
8
N/m
2
Sound speed, C
1
345 m/s
Poisson’s ratio, m 0.0 Kinematic viscosity, l 17.9 · 10
À6
Pa s
Specific mass, q
E
1000 kg/m
3
Specific mass, q
F
1.21 kg/m
3
P.R.F. Teixeira, A.M. Awruch / Computers & Fluids 34 (2005) 249–273 263
As the structure of this example presents relatively small displacements, the analysis may be
carried out using the simplified mesh movement algorithm described previously in Section 2.3. It
is obtained the same results for the displacements along the time, in simulations with Mach¼2.20,
using the complete and simplified mesh movement algorithms. The speed-up obtained with the
simplified algorithm in relation to the complete algorithm is 1.83, even with the fall of Mflops
(from 888 to 670 in a CRAY T-94 computer). The reason of this result is the difference of
computational cost that the mesh movement algorithm represents with respect to the total CPU
time. While the complete algorithm represents 45.82% of the total CPU time, the simplified
algorithm represents only 0.74%.
Fig. 13. Pressure distribution at the instants: (a) 1.5 s, (b) 3.0 s, (c) 6.0 s, (d) 9.0 s, (e) 10.5 s and (f) 12.0 s.
264 P.R.F. Teixeira, A.M. Awruch / Computers & Fluids 34 (2005) 249–273
5.2. Air flow past an inflated membrane
The interaction between air flow (considered as an incompressible flow) and a flexible mem-
brane is analyzed. The problem is described in Fig. 11. The membrane has a semi-cylindrical
format with diameter d and thickness h equal to 20 m and 3.0 mm, respectively, and is clamped at
both ends.
The external boundary of the fluid domain is a semi-cylinder with radius R and width L in the
perpendicular direction to the plane where the flow take place equal to 200 and 1 m, respectively.
At the external boundary a velocity profile, which is a function of the distance with respect to the
ground level, is prescribed. This velocity is referred to the velocity at 10 m above the ground level,
Fig. 14. Velocity vectors at the instants: (a) 1.5 s, (b) 3.0 s, (c) 6.0 s, (d) 9.0 s, (e) 10.5 s and (f) 12.0 s.
P.R.F. Teixeira, A.M. Awruch / Computers & Fluids 34 (2005) 249–273 265
which varies with time as shown in Fig. 11. A reference pressure p ¼ 0 is prescribed in the point
located in the external boundary condition, 200 m above the ground level. The air and membrane
properties are given in Table 1. The internal pressure p
0
which keeps the membrane inflated, is
taken equal to 60% of the air stagnation pressure at a velocity of 28 m/s (p
0
¼ 285:0 Pa).
Due to the existence of air inside the space covered by the membrane, a damping matrix is
included to analyze the structure. A Rayleigh viscous damping is adopted, which means that
C ¼ aM þ bK ð38Þ
Fig. 15. Mesh configurations in the fluid domain at: (a) 1.5 s, (b) 3.0 s, (c) 6.0 s, (d) 9.0 s, (e) 10.5 s and (f) 12.0 s.
266 P.R.F. Teixeira, A.M. Awruch / Computers & Fluids 34 (2005) 249–273
where C, M and K are the damping, mass and stiffness matrices, respectively. Considering a linear
behaviour, adopting damping ratios equal to n
1
¼ 1:0% and n
2
¼ 2:0% for the first and second
modes, respectively, and calculating the corresponding circular frequencies x
1
and x
2
the coef-
ficient a and b can be determined by the procedure described by Bathe [16]. For this case,
a ¼ 0:003529 and b ¼ 0:443162 are obtained.
Although this case is essentially a two-dimensional problem and beam elements for the
structure may be used, a three-dimensional code is employed, where the membrane model was
built with triangular flat elements and the fluid domain was discretized with tetrahedral elements.
In the perpendicular direction to the flow, only one layer of elements is used.
The non-structured finite element mesh in the fluid domain has 20 550 tetrahedral elements and
7097 nodes, as indicated in Fig. 12. The structural mesh has 124 nodes and 124 triangular ele-
ments (these elements are coincident with the faces of the tetrahedral elements at the fluid–
structure interface). The prescribed boundary conditions on the lateral ends are w ¼ h
x
¼ h
y
¼ 0
for the plate and v
3
¼ 0 for the fluid domain.
As Reynolds number in this problem is relatively high (Re ¼ 1:9 Â 10
7
), if it is determined with
a reference velocity equal to 28 m/s and a reference length equal to the radius of the structure
(10 m), a simple algebraic turbulence model is used. The Reynolds stresses are modeled using a
generalization of the Prandtl mixing length hypothesis [24]. This model was also employed by
Mittal and Tezduyar [25], and consists in the addition of a eddy viscosity l
T
to the molecular
kinematic viscosity l. The eddy viscosity is given by the following expression:
l
T
¼ ðKlÞ
2
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1
2
ov
i
ox
j
þ
ov
j
ox
i
_ _
:
ov
i
ox
j
þ
ov
j
ox
i
_ _
¸
ð39Þ
where K ¼ 0:41 is the Von Karman constant and l is the shortest distance between the point
where l
T
will be determined and the closest wall to this point.
A time interval Dt
F
¼ 1:25 Â 10
À3
s is adopted for the fluid domain, while a time step
Dt
S
¼ 1:25 Â 10
À4
s is used to analyze the membrane. Pressure distributions, velocity vectors and
Fig. 16. Geometry and boundary conditions for the panel vibration induced by vortex shedding originated by a rigid
fixed rectangular prism.
P.R.F. Teixeira, A.M. Awruch / Computers & Fluids 34 (2005) 249–273 267
mesh configurations in the fluid domain are shown in Figs. 13, 14 and 15, respectively for t ¼ 1:5,
3.0, 6.0, 9.0, 10.5 and 12.0 s. An overpressure on the left-hand side of the field and an under-
pressure over the middle and the right-hand side of the field are shown in Fig. 13. The re-
bouncing of the membrane, while a vortex is being generated downstream of the structure, is
Table 2
Fluid and panel properties
Properties Value
Young’s modulus, E 2.5 MN/cm
2
Poisson’s ratio, m 0.35
Specific mass, q
E
0.1 g/cm
3
Sound speed, C
1
34 500 cm/s
Specific mass, q
F
1.18 ·10
À3
g/cm
3
Kinematic viscosity, l 1.82 ·10
À4
g/cms
Fig. 17. Finite element mesh for the fluid domain. (a) General view. (b) Detail in the region close to the panel.
268 P.R.F. Teixeira, A.M. Awruch / Computers & Fluids 34 (2005) 249–273
shown in Fig. 14 at first 3 s. Within the subsequent 3 s, at a constant free-stream velocity, nearly
steady conditions are attained. When the free stream velocity increases from v
10
¼ 14 to 28 m/s
(t ¼ 6–9 s), the membrane moves again to the right pushed by the air flow and the vortex for-
mation downstream, behind the membrane, becomes pronounced. Following this initial phase the
free-stream air velocity remains constant and the periodic development and detachment of vor-
tices in correspondence with the membrane structure movement is indicated in Fig. 14. It is ob-
served in Fig. 15 the distortion of the finite element mesh at different times and a good
performance of the mesh movement algorithm was obtained.
tempo (s)
d
i
s
p
l
a
c
e
m
e
n
t
(
c
m
)
d
i
s
p
l
a
c
e
m
e
n
t
(
c
m
)
0 1 2 3
-0.5
-0.4
-0.3
-0.2
-0.1
0
0.1
0.2
0.3
0.4
0.5
Present work
Wall and Ramm, 1998
tempo (s)
0 1 2 3
-1.5
-1
-0.5
0
0.5
1
1.5
Present work
Wall and Ramm, 1998
time (s) time (s)
(a)
(b)
Fig. 18. Comparison of results obtained in this work and those obtained by Wall and Ramm [17]. (a) Displacements of
the panel at the end. (b) Displacements of the panel at the mid span.
P.R.F. Teixeira, A.M. Awruch / Computers & Fluids 34 (2005) 249–273 269
Membrane configurations obtained in the present work are very similar to those presented by
Argyris et al. [26]. However, pressure distributions and velocity vectors here show vortices gen-
erated downstream of the structure that was not well captured by the above mentioned reference,
probably because the finite element mesh was poorly refined (taking into account the charac-
teristics of this problem).
5.3. Panel vibration induced by vortex shedding originated by a fixed rectangular prism
A flexible panel is clamped in a fixed rigid rectangular prism, which generates vortex shedding
inducing vibrations on the deformable structure. This problem is described in Fig. 16 using
Fig. 19. Mesh configurations at (a) t ¼ 3:765 s and (b) t ¼ 3:85899 s.
270 P.R.F. Teixeira, A.M. Awruch / Computers & Fluids 34 (2005) 249–273
dimensionless values. The fluid is considered as slightly compressible and Re ¼ 332. Although this
case is essentially a two-dimensional problem, a three-dimensional code is used, where tetrahedral
elements are employed for the fluid domain and flat plate elements for the flexible panel. The
boundary conditions on the lateral ends are w ¼ h
x
¼ h
y
¼ 0 for the plate and v
3
¼ 0 for the fluid
domain. Fluid and panel properties are given in Table 2. Bearman [27] performed a detailed
experimental study and clarified the various regimes of the flow encountered by varying the
splitter plate length for range of Reynolds number.
A non-structured mesh with 19198 tetrahedral elements and 6544 nodes, with only one layer of
elements in the perpendicular direction to the flow, is used for the fluid domain and is shown in
Fig. 17, while the structural mesh has 84 triangular plate elements and 84 nodes. The fluid–
structure interaction takes place at both sides of the panel (it means that there are two fluid–
structure interfaces).
Fig. 20. Pressure distributions at (a) t ¼ 3:765 s and (b) t ¼ 3:85899 s.
P.R.F. Teixeira, A.M. Awruch / Computers & Fluids 34 (2005) 249–273 271
Simulation of the dynamic behaviour of the panel was carried out using a time interval
Dt ¼ 3:0 Â 10
À5
s. Displacement at the free and at the mid span are shown in Fig. 18, and are
compared with those obtained by Wall and Ramm [28]. These authors use a two-dimensional
code.
Some discrepancies can be observed between the results of the present work and those obtained
by Wall and Ramm [28]. Reasons of these differences are very difficult to establish because both
works use different schemes, different meshes and the structure was modeled with beam elements,
whereas in the present work thin plate elements were used. Mesh configurations and pressure
distributions at t ¼ 3:765 and 3.8589 s are shown in Fig. 19 and Fig. 20, respectively.
6. Conclusions
A partitioned algorithm for the solution of fluid–structure interaction problems using a two-
step Taylor–Galerkin method for the fluid in the ALE formulation and the Newmark implicit
scheme for the solution of the structural dynamic equations is presented in this work. The po-
tential of the method was shown through several examples of compressible and slightly com-
pressible flows acting on flexible structures (plates and shells) and good results were obtained. The
differences between the dynamic responses of a linear and a non-linear system were studied in the
example of a supersonic inviscid flow acting on a deformable flat plate. It was also verified that
using procedures with subcycles, the precision of the results are reduced and the number of
iterations to obtain the convergence of the solution of the equilibrium equations of the structure
increases. But, it was also observed that the decrease in computer processing time may become
expressive when a non-structured mesh with very different size of elements are used. The method
of mesh movement adopted here was shown to be efficient and the complete and simplified
algorithms were tested. The complete scheme has much more accuracy, however it requires a
considerable CPU time, disadvantage that can be avoided using the simplified scheme, which has
reasonable precision for problems with small displacements. Finally, the good performance of the
vectorized algorithms was verified. Speeds going from 670 to 888 Mflops in a Cray T94 computer
and with a CPU time of the order of 1.2–2.3 · 10
À5
s/Dt node have been obtained.
Acknowledgements
The authors wish to thank CNPq and CAPES for the financial support.
References
[1] Felker FF. Direct solution of two-dimensional Navier–Stokes equations for static aeroelasticity problems. AIAA J
1993;31(1):148–53.
[2] Blom FJ. A monolithical fluid–structure interaction algorithm applied to the piston problem. Comput Methods
Appl Mech Eng 1998;167:369–91.
[3] Soria A, Casadei F. Arbitrary Lagrangian–Eulerian multicomponent compressible flow with fluid–structure
interaction. Int J Numer Methods Fluids 1997;25:1263–84.
272 P.R.F. Teixeira, A.M. Awruch / Computers & Fluids 34 (2005) 249–273
[4] Rifai SM et al. Multiphysics simulation of flow-induced vibrations and aeroelasticity on parallel computing
platforms. Comput Methods Appl Mech Eng 1999;174:393–417.
[5] Farhat C, Lesoinne M, Maman N. Mixed explicit/implicit time integration of coupled aeroelastic problems: three-
field formulation, geometric conservation and distributed solution. Int J Numer Methods Fluids 1995;21:807–35.
[6] Cebral JR, L€ ohner R. Conservative load projection and tracking for fluid–structure problems. AIAA J
1997;35:687–92.
[7] Farhat C, Lesoinne M, LeTallec P. Load and motion transfer algorithms for fluid/structure interaction problems
with non-matching discrete interfaces: momentum and energy conservation, optimal discretization and application
to aeroelasticity. Comput Methods Appl Mech Eng 1998;157:95–114.
[8] Zienkiewicz OC, L€ onher R, Morgan KO, Peraire J. High speed compressible flow and other advection-dominated
problems of fluid dynamics. In: Gallagher R, et al., editors. Finite elements in fluids, vol. 6. 1988. p. 41–8.
[9] Teixeira PRF, Awruch AM. Three-dimensional simulation of high compressible flows using a multi-time-step
integration technique with subcycles. Appl Math Modell 2001;25/7:613–27.
[10] Hirt CW, Amdsen AA, Cook JL. An arbitrary Lagrangian–Eulerian computing method for all flow speeds.
J Comput Phys 1974;14:227–53.
[11] Hughes TJR, Liu WK, Zimmerman TK. Lagrangian Eulerian finite element formulation for incompressible
viscous flows. Comput Methods Appl Mech Eng 1981;29:329–49.
[12] Donea J, Giuliani S, Halleux JP. An arbitrary Lagrangian–Eulerian finite element method for transient dynamic
fluid–structure interactions. Comput Methods Appl Mech Eng 1982;33:689–723.
[13] Benson DJ. An efficient accurate simple ALE method for non-linear finite element programs. Comput Methods
Appl Mech Eng 1989;72:305–50.
[14] Liu WK, Chang H, Chen JS, Belytschko T. Arbitrary Lagrangean–Eulerian Petrov–Galerkin finite elements for
non-linear continua. Comput Methods Appl Mech Eng 1988;68:259–310.
[15] Zhang Q, Um L, Wenqi K. Geometric non-linear analysis of space shell structures using generalized conforming
flat shell elements––for space shell structures. Commun Numer Methods Eng 1998;14:941–57.
[16] Bathe K. Finite element procedures. Englewood Cliffs, NJ, USA: Prentice Hall; 1996.
[17] Chung TJ. Finite element analysis fluid dynamic. New York: McGraw-Hill; 1978.
[18] Zienkiewicz OC, Taylor RL. The finite element method solid and fluid mechanics dynamics and nonlinearity, vol.
2. 4th ed. Singapore: McGraw-Hill; 1991.
[19] Richtmeyer RD, Morton KW. Difference methods for initial value problems. New York: Interscience; 1967.
[20] Peraire J et al. Finite element Euler computations in three dimensions. Int J Numer Methods Eng 1988;26:2135–59.
[21] Monteuffel TA. An incomplete factorization technique for positive definite linear systems. Math Comput
1980;34:473–97.
[22] Meijerink JA, Van der Vorst HA. Guidelines of the usage of incomplete decomposition in solving sets of linear
equations as they occur in practical problems. J Comp Phys 1981;44:134–55.
[23] Dowell EH. Aeroelasticity of plates and shells. The Netherlands: Noordhof Int. Publ., Leyden; 1975.
[24] Schlichting H. Boundary-layer theory. New York: McGraw-Hill; 1968.
[25] Mittal S, Tezduyar TE. Parallel finite element simulation of 3D incompressible flows: Fluid–structure interactions.
Int J Numer Methods Fluids 1995;21:933–53.
[26] Argyris J et al. Ta pamsa qei. Comput Meth Appl Mech Eng 1985;21:289–362.
[27] Bearman PW. Investigation of the flow behind a two dimensional model with a blunt trailing edge and fitted with
splitter plates. J Fluid Mech 1965;21:241–55.
[28] Wall WA, Ramm E. Fluid–structure interaction based upon a stabilized (ALE) finite element method. In: Idelshon
S, editor. Comput Mech. Barcelona: CIMNE; 1998.
P.R.F. Teixeira, A.M. Awruch / Computers & Fluids 34 (2005) 249–273 273

250

P.R.F. Teixeira, A.M. Awruch / Computers & Fluids 34 (2005) 249–273

1. Introduction Important progress has been obtained in the solution of complex fluid–structure interaction problems in recent years, allowing the study of multidisciplinary applications in different engineering areas. This progress is due mainly to the increase of the speed of modern computers, to the evolution of multidisciplinary solution algorithms and of pre- and post-processing tools. The methods of simulation of fluid–structure interaction problems are divided basically in two groups called partitioned and monolithic schemes [1,2]. In the partitioned schemes the governing equations of the fluid and the structure are integrated in time alternately in an isolated way. In the monolithic schemes the two fields are considered as a single entity, allowing to integrate in time the two subdomains simultaneously. In partitioned schemes such as proposed by Soria and Casadei [3], Rifai et al. [4], Farhat et al. [5], Cebral and L€hner [6] and many other authors, the o kinematic and dynamic boundary conditions in the interface are the unique information changed between the subdomains. In these schemes each sub-domain can be solved by discretization techniques and by efficient solution algorithms in an individual way. New methods and models can be introduced in a modular way in order to get more flexibility. Besides, the meshes of each subdomain can be built without the coincidence of fluid and structural nodes at the interfaces [6,7]. These modular characteristics and flexibility may be extremely convenient. An algorithm to simulate fluid–structure interaction problems using a partitioned scheme is presented in this work. A two-step explicit Taylor–Galerkin scheme [8,9], with linear tetrahedral finite elements is employed. An arbritary Lagrangean–Eulerian (ALE) description is adopted for the fluid domain, while for the structural domain an updated Lagrangean formulation is considered. The ALE description was first presented by Hirt et al. [10] and after this method was used by several authors such as Hughes et al. [11], Donea et al. [12], Benson [13], Liu et al. [14] among others. The structure is analyzed using generalized conforming triangular plates and shell elements with drilling degrees of freedom [15]. Geometrically non-linear effects are included. The Newmark method [16] is employed to integrate in time the dynamic equilibrium equation. The non-linear and the algebraic systems are solved using an incremental-iterative scheme and the conjugate gradient method, respectively. The code was vectorized to take advantages of vectorial processors. Several problems are used to validate the methods developed in this work and illustrate the difference between linear and non-linear aeroelastic computations.

2. The fluid dynamic solver 2.1. The explicit two-step Taylor–Galerkin scheme for high compressible flows In the ALE description, the computational frame is a reference independent of the particle movement and may be moving with an arbitrary velocity in the laboratory system (this motion is called the ‘‘mesh’’ motion in the finite element formulation); the continuum view from this reference is denoted as Xv and the coordinates of any point are denoted as v [14]. The equations expressing mass, momentum and energy conservation in ALE formulation may be written in a compact form as

Awruch / Computers & Fluids 34 (2005) 249–273 251 .M.P. A. Teixeira.F.R.

oV .

.

þ oFi ¼ wi oV ot .

corresponding to the time interval ½tn . the unknown vector V at t ¼ tnþ1=2 is expanded in Taylor series. In the Taylor–Galerkin scheme. 3Þ ð1Þ with Fi ¼ FiA þ FiD and 8 8 9 9 qvi > q > > > > > > > > > > > > qv1 vi þ pdi1 > > > < qv1 = < = V ¼ qv2 . qe vi ðqe þ pÞ 9 0 > > > > Às1i = Às2i ði. 2. Àsij vj À kij oT =oxj 8 > > > > < ð2Þ where vi and wi are the fluid and the mesh velocity components in the direction of the spatial coordinate xi respectively. 2. this approach can be interpreted as the finite element version of the Lax–Wendroff scheme used in finite differences [19]. Using a linear shape function N associated with each node to interpolate Vn . j ¼ 1. conservation equations are expanded in time by Taylor series. the following equation is obtained: " Z ! ! ! # Z Z Dt oN n n n nþ1=2 nþ1=2 n oN ¼ N dX V À dX Fi À Nwi dX V XE VE n n oxi n 2 oxi XE XE XE ði ¼ 1. tnþ1=2 Š. e is the total specific energy and kij are the components of the conductivity tensor. and after. Initial and boundary conditions must be added to these equations in order to define uniquely the problem. q is the density. T is the temperature. VE is a constant value at element level and the upper bar indicates nodal variables.v oxi oxi ði ¼ 1. Eq. p is the thermodynamic pressure. Finally. [8]. > > > qv3 > > qv3 vi þ pdi3 > > > > > > > > > > > : : . . Using again the same shape functions. 2. 3Þ nþ1=2 nþ1=2 ð3Þ where XE is the element volume. (1) is complemented by the equation of state for an ideal gas and by the constitutive equations. sij are the components of the deviatoric stress tensor. as indicated by Zienkiewicz et al. A two-step method is used. 2. FiA ¼ qv2 vi þ pdi2 . the unknown vector at t ¼ tnþ1 is expanded in Taylor series. vector V contains the conservation or the field variables and Fi are the components of the flux variables (FiA contains the advective terms and FiD the viscous terms). In the first step. 3Þ ð4Þ CE nþ1=2 . In the second step. and applying the Bubnov–Galerkin method. a constant shape function PE ¼ 1 associated with element E to interpolate Vnþ1=2 and applying the classical Galerkin weighted residual method to the expression resulting from Taylor series expansion. 3Þ FiD ¼ > > > > Às3i > > > > : . dij is the Kronecker delta. the following equation is obtained for the second step: ! ! ! Z Z Z  nþ1=2  oN nþ1 n nþ1=2 nþ1=2 T T À wi N N dX V ¼ N N dX V þ Dt dX Fi V nþ1=2 ox i Xnþ1 Xn XE E E ! Z  nþ1=2  nþ1=2 nþ1=2 Nli dC FBi À wBi VB À Dt ði ¼ 1. space discretization is accomplished by the classical Bubnov–Galerkin scheme [16–18].

3Þ ð10Þ . The proposed scheme is conditionally stable. the consistent mass matrix is substituted by the lumped mass matrix. may be expressed by the following equation: oq 1 op oUi ¼ 2 ¼À ot c ot oxi ði ¼ 1.2. j ¼ 1. it is necessary to add numerical damping to the flow solver. (8) is given by the following expression: Ui nþ1=2 ð9Þ e nþ1=2 À Dt o Dp ¼ Ui 4 oxi ði ¼ 1.252 P. In Eq. 2. Expanding the momentum conservation equations in Taylor series. 3Þ þ À wj i 2 oxj oxj oxi oxi Eq. Awruch / Computers & Fluids 34 (2005) 249–273 where Cnþ1=2 is the boundary of the element domain Xnþ1=2 and li is the cosine of the angle formed E E nþ1=2 with the positive direction of the reference axis xi . and the local stability condition is applied.M.0. 3Þ ð7Þ where c is the sound speed and Ui ¼ qvi ði ¼ 1. assuming constant entropy. 3Þ Ui 2 ot 2 oxj oxj oxi 2 oxi oxi ð8Þ where fij ¼ vj Ui ði. [20]. Index B is by the outward normal axis to CE referred to values at the boundary of the element domain. the following expression is obtained for the first step:  n  n Dt oUin Dt ofij osn opn 1 o Dp ij nþ1=2 n n n oUi ¼ Ui À ¼ Ui þ þ þ À wj À ði. specially in the presence of strong shocks. CC is a global constant specified by the user. An artificial viscosity is added explicitly to the non-smoothed solution.R. and then this equation is solved iteratively. The constant CC must be E E specified with care in order to avoid interferences of the artificial and physical viscosities.0 and 10. A. The two-step Taylor–Galerkin scheme for slightly compressible flows Mass conservation for slightly compressible fluids. (4). respectively. 3Þ. Mn C and Mn L are the consistent and lumped mass matrices. Teixeira. 2. with Dp ¼ pnþ1 À pn . Using  n  n ofij osn opn ij n oUi e inþ1=2 ¼ U n À Dt U À ði. j ¼ 1. is adopted. 2. [8.9]). Mnþ1 L is where and V are the smoothed and non-smoothed solution at t ¼ t the assembled lumped mass matrix at t ¼ tnþ1 The vector D is given by X n D¼ CFLE CCS E ½Mn C À Mn L ŠVE E E E nþ1 Vs ð6Þ where E is an index referred to a specific element. as follows Vs nþ1 ¼V nþ1 þ ðMnþ1 ÞÀ1 D L nþ1 nþ1 ð5Þ respectively. 2. CFLE is the local Courant number. 2.F. 2. In order to stabilize numerically the solution. In this work the viscosity model. as given by Peraire et al. 3Þ and pnþ1=2 ¼ pn þ 1=2 Dp. j ¼ 1. S E is a coefficient of pressure distribution (see Refs. CC varies between 2. 2.

j ¼ 1. Teixeira. (3) determine Ui Uinþ1=2 with Eq. (7) in time and using Eq. 2. (11) and calculate pnþ1 ¼ pn þ Dp. A. (12). 2. 3Þ ð16Þ . 2. 2. 3Þ p XE UiE 4 Xnþ1=2 oxi T Z ð14Þ ð15Þ Z Xnþ1  Z  Z nþ1 n T N N dX Ui ¼ N N dX Ui þ Dt T   oNT dX f nþ1=2 À wnþ1=2 Unþ1=2 ij j i E Xn Xnþ1=2 oxj Z  Z  T T  oN oN À Dt N dX n þ Dt sij N dX n þ D=2 p p Xnþ1=2 oxj Xnþ1=2 oxi Z  nþ1=2 nþ1=2 nþ1=2 NT nj dC ðf ij À wj Ui ÞE À Dt nþ1=2 C Z  Z   T T n p þ Dt N Nnj dC ij À Dt s N Nni dC n þ D=2 p Cnþ1=2 Cnþ1=2 ði. (9). Awruch / Computers & Fluids 34 (2005) 249–273 253 Discretizing Eq. after space discretization. (10). (9).F.M. 2. (11). 3Þ dX U iE À N ni dC U iE Xnþ1=2 oxi Cnþ1=2 Z  Dt oN nþ1=2 nþ1=2 nþ1=2 e nþ1=2 ¼ XE U iE À dX D ði ¼ 1. it is obtained: " # e 1 oUinþ1=2 o U inþ1=2 Dt o o Dp Dq ¼ 2 Dp ¼ ÀDt À ði ¼ 1. respectively: Z   Z   Z Dt oN oN n nþ1=2 e nþ1=2 n À sij ¼ N dX Ui À dX f ij dX n XE U iE 2 Xn Xn oxj Xn oxj Z  Z   oN n n n oN þ dX  À p Nwj dX Ui ði. (4) determine Uinþ1 with Eq. (10). 3Þ ð13Þ oxj Xn oxi Xn  Z Z 1 Dt2 oNT oN T oN N 2 N dX þ dX À N ni dX D p c oxi 4 Xnþ1=2 oxi oxi Xnþ1=2 Cnþ1=2  Z   Z  oNT nþ1=2 nþ1=2 T e e ¼ Dt ði ¼ 1. j ¼ 1. (2) determine Dp with Eq. 2.P. by the following algorithm: (1) determine e nþ1=2 with Eq. 3Þ ¼ ÀDt c 4 oxi oxi oxi oxi The second time step is given by the following expression: Uinþ1 ¼ Uin þ Dt oUi ot nþ1=2 ð11Þ ofij ¼ Uin À Dt oxj nþ1=2 osij À oxj nþ1=2 opnþ1=2 oU þ À wnþ1=2 i j oxi oxi nþ1=2 ! ði. (10) and (12). 3Þ ð12Þ Then the flow is analyzed. Considering the same shape functions used in compressible flows and applying the classical Galerkin method for space discretization. the following matrix expressions are obtained for Eqs.R. j ¼ 1.

254 P. favouring the influence of points i. conserving prescribed velocities in moving and stationary boundary surfaces. .3. 1): ns P aij wj j¼1 ð17Þ wi ¼ P ns aij j¼1 where ns is the total number of points belonging to the boundary surfaces and aij are the influence coefficients between the point i inside the domain and the point j of the boundary surface given by the following expression: aij ¼ 1 4 dij ð18Þ with dij being the distance between the points i and j. Fig. In this equation the boundary integral involving the pressure increment in the left hand side could be omitted. The mesh movement algorithm adopted in this work uses a smoothing procedure for the velocities based in these boundary surfaces. 2. located closer to the boundary surface containing point j. Distances from the boundary surfaces to a point ‘‘i’’ in the fluid domain. (13)–(16) the index E indicates that the corresponding variables are taken with a constant value over the element domain.F. aij has an high value. (10) is examined. Eq. The other reason is that on traction boundaries. A. Awruch / Computers & Fluids 34 (2005) 249–273 In Eqs. aij represents the weight that each point j of the boundary surface has on the value of the mesh velocity at points i inside the domain. The algorithm for mesh movement The mesh velocity field w is computed looking for small element distortions. It is observed that on boundaries where the velocities are prescribed o Dp=oxi ¼ 0. 1. One of the reason for this simplification is apparent if Eq. In other words. no change of traction occurs and then Dp ¼ 0 [18].R. When dij is small. Teixeira. (14) is solved using the conjugate gradient method with diagonal pre-conditioning.M. The updating of the mesh velocity at a point i of the finite element domain is based on the mesh velocity of the points j belonging to the boundary surfaces in the following way (see Fig.

M. The thin plate triangular element. (17). It consists of choosing previously for each point inside the fluid domain at least two surface points having important influence in the calculation of the mesh velocity according to Eq. . Each of these two points have the largest values of coefficient aij among all the other points in its respective surface because each one is the closest point of the respective surface with respect to the node where the mesh velocity will be determined. (17). in which the compatibility conditions at each node and along each side are applied. A typical membrane triangular element is adopted where each point has two degrees of freedom of translation uxi and uyi ði ¼ 1. The total stiffness matrix of the element is obtained by the overlap of the membrane stiffness matrix with the bending stiffness matrix. 2) [15]. 3Þ in the plane of the element middle surface. 3. The structural dynamic solver with a triangular thin plate element A generalized conforming triangular thin plate element with six degrees of freedom in each node (three displacements and three rotations components). it must be used a point of each body surface to calculate the mesh velocities field inside the fluid domain. Therefore. Teixeira. 2. but considering only one point in each boundary surface.F. this procedure uses the same Eq.P. A. The method can be generalized when there are two or more body surfaces. One of these points belongs to the moving body surface and the other one to the external fixed surface. In this case.R. Awruch / Computers & Fluids 34 (2005) 249–273 255 In the present work a variation to simplify this method and safe computer time was adopted for the case in which the displacement of the body surface is not significantly high. 2. 3Þ and one of rotation hzi ði ¼ 1. 2. is used in the present work (see Fig. The membrane displacements are expressed in the following way: um ¼ ½ ux uy ŠT ¼ Nm ue m ð19Þ where ue is the nodal membrane generalized displacements vector given by m Fig. corresponding to the node which have the smallest distance with respect of the point inside the fluid domain where the mesh velocity will be computed.

The constitutive matrices Dm and Db are given. 3Þ ð25Þ ði. KLbi ¼ BT Db Bbi dX ði ¼ 1. 2 2 bi ¼ yj À ym . 2. m ¼ 1. 2. m ¼ 1.R. Finally. ci ¼ xm À xj ði. Teixeira. Nb is the bending interpolation function given by Nbi ¼ ½ Ni with Ni ¼ Li À 2Fi þ ð1 À rj ÞFj þ ð1 þ rm ÞFm Nxi ¼ À1½bm Li Lj À bj Lm Li þ ðbj À bm ÞFi þ ðrj bj þ bm ÞFj þ ðrm bm À bj ÞFm Š 2 Nyi ¼ À1½cm Li Lj À cj Lm Li þ ðcj À cm ÞFi þ ðrj cj þ cm ÞFj þ ðrm cm À cj ÞFm Š 2 Fi ¼ Li ðLi À 0:5ÞðLi À 1Þ. respectively. 2. Awruch / Computers & Fluids 34 (2005) 249–273 ue ¼ ½ uxi mi uyi hzi ŠT ði ¼ 1. respectively. the strain–displacement relations are .M. by the following expressions: Z Z T KLmi ¼ Bmi Dm Bmi dX. 2. respectively. D ¼ 4 m 1 0 ð28Þ 1 À m2 12ð1 À m2 Þ 0 0 ð1 À mÞ=2 where h is the elements thickness. 2. and hxi and hyi the nodal rotations around the axes x and y. 3Þ The transverse displacement field is discretized by u z ¼ Nb u e b ð21Þ ð22Þ ð23Þ where ue is the nodal bending generalized displacements vector. 2. A. 3Þ Nmi ¼ 0 Li Nvhi being Li the area coordinates and 1 1 Nuhi ¼ Li ðbm Lj À bj Lm Þ. Nvhi ¼ Li ðcm Lj À cj Lm Þ. j. which is written in the following b way: ue ¼ ½ uzi bi hxi hyi ŠT ði ¼ 1. ij l2 jm lij ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2 x2 þ yij .F. by 2 3 1 m 0 Eh Eh3 5 Dm ¼ D. j. xij ¼ xi À xj . 3Þ ð27Þ bi XðeÞ XðeÞ where XðeÞ is the element domain. 3Þ The linear stiffness matrix at element level due to membrane and bending effects are given.256 P. ij ð26Þ Nxi Nyi ŠT ði ¼ 1. m the Poisson’s ratio and E the Young’s modulus. 3Þ ð24Þ being uzi the nodal transverse displacements. yij ¼ yi À yj ri ¼  1 2 lim À l2 . Db ¼ D. 3Þ ð20Þ and Nm is the membrane interpolation function defined as follows:   Li 0 Nuhi ði ¼ 1. 2.

F.22]. 3Þ Tyxi Tyyi with Txxi ¼ " Tyyi ¼ " Txyi ¼ " # Dm ð1. 3Þ # Dm ð3. 2. The global time step is usually commanded by the fluid. In geometrically non-linear problems the solution of these balance equations is accomplished for each time step with an incremental iterative procedure using an updated Lagrangean formulation.x Nyi. 3Þ ð34Þ The dynamic equilibrium equations of the structural system are obtained by the space discretization of the virtual work equations.P.xy ð30Þ The non-linear geometric stiffness matrix is given by Z KNLi ¼ GT Ti GGi dX ði ¼ 1. 4.xx Bbi ¼ À4 Ni.M. 2Þ Á umi ð2Þ ði ¼ 1. 2. fluid and structure fields have different scales of time. Teixeira. 2. Although the use of the same time step for the fluid and the .yy 2Nxi. Awruch / Computers & Fluids 34 (2005) 249–273 257 2b 1 4 i Bmi ¼ ðeÞ 0 4X 2c i 2 0 2ci 2bi bi ðbm Lj À bj Lm Þ 5 ði.xy 2 Nxi. 2.xx Nyi. which must be solved for each time step using a numerical integration scheme.x GGi ¼ Ni.yy 2Ni. j. The resulting system of algebraic equations are solved using the gradient conjugate method with incomplete Cholesky factorization [21.R. The fluid–structure interaction algorithm Commonly. 2. 3Þ Gi XðeÞ ð31Þ where Ti contains the membrane internal forces and is given by   Txxi Txyi Ti ¼ ði ¼ 1.x Nxi. 3Þ Á umi ð1Þ ¼ Tyxi ð32Þ 3 X k¼1 3 X k¼1 3 X k¼1 ð33Þ while GGi is given by  Ni. kÞ Á Bmi ðk. 2. 3Þ ci ðcm Lj À cj Lm Þ ðci bm þ bi cm ÞLj À ðci bj þ bi cj ÞLm 3 Nyi. 3Þ 2Nyi. m ¼ 1. with parameters d ¼ 1=2 and a ¼ 1=4.y Nxi.xy 3 ð29Þ and Ni.yy 5 ði ¼ 1.xx Nxi.y Nyi. A. In the present work the well-known Newmark method is used [16].y  ði ¼ 1. 1Þ Á umi ð1Þ # Dm ð2. kÞ Á Bmi ðk. kÞ Á Bmi ðk.

structure may provide some implementation advantages.1) Compute the new mesh taken into account the structural motion. Algorithm for fluid–structure interaction with subcycles. pressure and total energy using subcycles. taken into account the structural motion. velocities.2) Update the fluid flow variables with the new boundary conditions. Teixeira. A. must be applied to the fluid domain. specific mass. accelerations and stresses. (e) Repeat steps (c) and (d) until the objectives of the simulation are reached. Therefore. (c. (d) Update the structural variables with the loads transferred by the fluid.F. 3): (a) Set the initial conditions for the structure and the fluid.258 P. respectively) may offer substantial computational advantages. (c) Update the fluid velocities.M. (c. (b) Update the structure displacements. 3. at the fluid–structure interface the following condition must be satisfied: . Awruch / Computers & Fluids 34 (2005) 249–273 Fig. including an economy of CPU time due to the smaller number of time steps in the structural analysis and an economy in the information transfer. The partitioned algorithm with subcycles adopted in the present work consists in the following steps (see Fig. Kinematics boundary conditions. the procedure with subcycles of factor nSF ¼ DtS =DtF (where DtS and DtF are the time intervals adopted for the structure and the fluid.R.

.

ð35Þ xnþ1 .

CF ¼ tþDt xi .

CE ði ¼ 1. For viscous fluids the same velocity components are prescribed for the mesh. . Therefore. 3Þ i where xnþ1 jCF and tþDt xi jCE are the updated coordinates for the instant t þ Dt at the fluid and solid i interfaces CF and CE . the fluid and the structure at the interface. 2. respectively.

.

.

.

3Þ ð36Þ wnþ1 . _ ði ¼ 1. 2.

CF ¼ vnþ1 .

CF ¼ tþDt ui .

i i CE For non-viscous fluids the corresponding prescribed boundary conditions are: .

.

.

.

_ wnþ1 .

CF Á ni ¼ vnþ1 .

CF Á ni ¼ tþDt ui .

3Þ i i CE ð37Þ where ni is the normal vector with respect to the interface surface at time t þ Dt. (37) is applied to each node located in the interface surface. 2. . Eq. Á ni ði ¼ 1.

Young’s modulus E ¼ 77:28 GPa and Poisson’s ratio m ¼ 0:33. is analyzed.n=0 x z v2=0. Supersonic inviscid flow over a flat plate In this example a supersonic inviscid flow over a flat plate. The density of the fluid is qF ¼ 0:339 kg/m3 . width and thickness are 0. v i∞ . 4.P. Awruch / Computers & Fluids 34 (2005) 249–273 259 5. respectively. originating displacements in the plate and disturbances in the flow. After this period. General view of the finite element mesh for the fluid domain.0 Fig. A. Teixeira. 4.2 m and 1. It is assumed that the pressure on the inferior face of the plate falls 0.M. . a value equal to 5.35 mm. 0.1. this pressure ρ ∞ . as shown in Fig. (6). The plate length.R. Numerical applications 5. Geometry and boundary conditions in the fluid domain for the example of a supersonic inviscid flow over a flat plate. which is clamped on both ends. while the plate has the following properties: specific mass qE ¼ 2710 kg/m3 . Fig. 5. In this problem. defined in Eq. Initially the plate is subject to the same pressure on both sides.0 y v. ε ∞ v2=0.0 v3=0.5 m.F.1% instantaneously and remains in this condition for 4 ls.0 was adopted for CC.

Awruch / Computers & Fluids 34 (2005) 249–273 displacement (m) in x=0.02 and 2.0 and by Rifai et al.0.1 Fig.0015 -0.02 0.0005 -0.08 0.0015 -0. A. . Displacements of the plate in x ¼ 0:35 m along the time. [4] for Mach ¼ 1.001 -0.002 Mach = 1.0005 0 -0.F.260 P.9. 2.0005 -0.08 0. Comparison of the displacements in x ¼ 0:35 m obtained by the present work for Mach ¼ 2.002 0. displacement (m) in x=0.0015 0. 7. 1999 0 0.0005 0 -0. 2.06 0.04 time (s) 0. Teixeira.1 Fig.0015 0.M.05 0 0.02 0.05.0 Mach = 2. 6.02 Mach = 2.001 0.0 Rifai et al.002 0.R.06 0.98.002 Mach = 2.35m 0. for Mach ¼ 1.001 0.35m 0.001 -0..04 time (s) 0.9 Mach = 2.

A general view of the mesh is shown in Fig. while in the plate are imposed the following conditions: u ¼ v ¼ w ¼ hx ¼ hy ¼ hz ¼ 0:0 in the clamped ends and w ¼ hx ¼ hy ¼ 0:0 in the plane of symmetry xy. 0.0 at t ¼ 0:07 s. 8. The domain discretization is accomplished such that the fluid tetrahedral faces adjacent to the plate are coincident with the triangles belonging to the structure.P. Pressure distribution for Mach ¼ 2.F.R. The boundary conditions imposed to the fluid are shown in Fig. 13303 nodes and 38915 tetrahedral elements for the mesh of the fluid domain and 506 nodes and 508 triangular elements for the structural mesh are used. Fig. The solution of the linear aeroelastic problem indicates a limit of stability with a critical Mach number equal to 2. This inviscid flow has a two-dimensional behavior. 4. only one layer of elements is adopted to discretize the domain in the direction of the axis z. above this value.1 s. 9.0 at t ¼ 0:03.M. Teixeira. transverse to the flow. 5. A. continuous growth of oscillations amplitudes are observed and instability of the plate occurs. although this fact would not be necessary for the solution scheme.0 using the piston theory [23].07 and 0. . Fig. Awruch / Computers & Fluids 34 (2005) 249–273 261 returns to its initial value of 28 kPa. Configurations of the clamped plate for Mach ¼ 2.

02 0.262 P. Awruch / Computers & Fluids 34 (2005) 249–273 In a first analysis.005 -0. 2. Comparison of the displacements in x ¼ 0:35 m between the geometrically linear and non-linear solutions with Mach ¼ 2. simulations are accomplished with Mach numbers around the critical value predicted by the analytical solution using the piston theory [22].08 0. Fig. 2.005 0 -0.05.02 0.R.1 Fig. Description of the problem to study the interaction between air flow and a flexible membrane. A. Results of the displacement of the plate in x ¼ 0:35 m along the time are shown in Fig. 11.015 -0.3.04 time (s) 0. 10.01 -0.35m 0. These results were obtained using to analyze the plate time steps equal to 1.015 0.01 0. 6 for Mach numbers ¼ 1.M. Teixeira.0.02 Linear Non linear 0 0.F.02 and 2.8 · 10À7 s for displacement (m) in x=0. .06 0.9.

although a little discrepancy stays among these results.0. E Poisson’s ratio.9 and 1. m Specific mass. This phenomenon increases the effective stiffness of the plate. It is observed that the critical Mach number obtained here is practically equal to the value predicted by the solution using the piston theory (Mach ¼ 2.3 and a time step Dt ¼ 1:6 Â 10À7 s. modifying the behaviour and the response of the system. 9 presents the plate configurations and values of vertical displacements at t ¼ 0:03.0 at the instant t ¼ 0:07 s is shown in Fig. 10. Fig. C1 Kinematic viscosity. the amplitudes are very close.21 kg/m3 Fig. 0.0 1000 kg/m3 8 2 Air properties Sound speed. [4] (in this work a critical Mach number equal to 1. In Fig. A comparison between the linear and the non-linear behaviour of the plate is performed for Mach number ¼ 2. Awruch / Computers & Fluids 34 (2005) 249–273 263 Mach ¼ 1.M. while the geometrically non-linear solution presents limited displacement amplitudes.P. (b) Detail of the region close to the membrane. 7 these results are compared with those obtained by Rifai et al.9 · 10À6 Pa s 1.02 and 2.07 and 0.R. .333 · 10 N/m 0. whose displacements in x ¼ 0:35 m are shown in Fig.0). 12. The difference between the two solutions is attributed to the coupling between membrane and bending effects in the panel when non-linear elasticity due to finite displacements is considered. 2. It is noticed that. l Specific mass. (a) General view. qF Value 345 m/s 17. qE Value 3. Finite element mesh.F. as well as the solution based in the piston theory [22]. Table 1 Air and membrane properties Membrane properties Young’s modulus.05. 8. Teixeira.98 was found). A. The pressure distribution in the domain for a Mach number ¼ 2.75 · 10À7 s for Mach ¼ 2. presents an exponential growth of the oscillations. The geometrically linear solution.1 s.

20. (e) 10.R. A. . 13.5 s and (f) 12. (b) 3.0 s.M. While the complete algorithm represents 45. Awruch / Computers & Fluids 34 (2005) 249–273 As the structure of this example presents relatively small displacements.74%.5 s. The speed-up obtained with the simplified algorithm in relation to the complete algorithm is 1.0 s. using the complete and simplified mesh movement algorithms. It is obtained the same results for the displacements along the time.83.0 s. Fig. Teixeira. the simplified algorithm represents only 0. (d) 9.82% of the total CPU time. The reason of this result is the difference of computational cost that the mesh movement algorithm represents with respect to the total CPU time. in simulations with Mach ¼ 2.3.0 s.264 P.F. the analysis may be carried out using the simplified mesh movement algorithm described previously in Section 2. (c) 6. even with the fall of Mflops (from 888 to 670 in a CRAY T-94 computer). Pressure distribution at the instants: (a) 1.

Awruch / Computers & Fluids 34 (2005) 249–273 265 5. Teixeira.5 s and (f) 12. Fig.F.0 s.0 s.0 s. (d) 9. and is clamped at both ends. which is a function of the distance with respect to the ground level. respectively. respectively. A. At the external boundary a velocity profile. Air flow past an inflated membrane The interaction between air flow (considered as an incompressible flow) and a flexible membrane is analyzed. (c) 6. (e) 10.M. The external boundary of the fluid domain is a semi-cylinder with radius R and width L in the perpendicular direction to the plane where the flow take place equal to 200 and 1 m. 11.2. is prescribed.0 mm. 14. This velocity is referred to the velocity at 10 m above the ground level. The membrane has a semi-cylindrical format with diameter d and thickness h equal to 20 m and 3. (b) 3.P. The problem is described in Fig.5 s. Velocity vectors at the instants: (a) 1.R.0 s. .

5 s. A Rayleigh viscous damping is adopted. 200 m above the ground level.5 s and (f) 12. a damping matrix is included to analyze the structure.0 s.M.266 P.F. . (c) 6. A. The internal pressure p0 which keeps the membrane inflated.0 s. Mesh configurations in the fluid domain at: (a) 1. A reference pressure p ¼ 0 is prescribed in the point located in the external boundary condition.R. Due to the existence of air inside the space covered by the membrane. Teixeira. The air and membrane properties are given in Table 1. 11. (e) 10.0 s.0 s. is taken equal to 60% of the air stagnation pressure at a velocity of 28 m/s (p0 ¼ 285:0 Pa). Awruch / Computers & Fluids 34 (2005) 249–273 which varies with time as shown in Fig. which means that C ¼ aM þ bK ð38Þ Fig. 15. (b) 3. (d) 9.

mass and stiffness matrices. The eddy viscosity is given by the following expression: sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi     1 ovi ovj ovi ovj 2 : ð39Þ þ þ lT ¼ ðKlÞ 2 oxj oxi oxj oxi where K ¼ 0:41 is the Von Krmn constant and l is the shortest distance between the point a a where lT will be determined and the closest wall to this point. For this case. As Reynolds number in this problem is relatively high (Re ¼ 1:9 Â 107 ). respectively. Although this case is essentially a two-dimensional problem and beam elements for the structure may be used. This model was also employed by Mittal and Tezduyar [25]. Considering a linear behaviour. a simple algebraic turbulence model is used. Pressure distributions.R. as indicated in Fig. a three-dimensional code is employed. The Reynolds stresses are modeled using a generalization of the Prandtl mixing length hypothesis [24].P. velocity vectors and Fig. Awruch / Computers & Fluids 34 (2005) 249–273 267 where C.M. The non-structured finite element mesh in the fluid domain has 20 550 tetrahedral elements and 7097 nodes. A time interval DtF ¼ 1:25 Â 10À3 s is adopted for the fluid domain. while a time step DtS ¼ 1:25 Â 10À4 s is used to analyze the membrane. respectively. Teixeira. The structural mesh has 124 nodes and 124 triangular elements (these elements are coincident with the faces of the tetrahedral elements at the fluid– structure interface).F. 12. The prescribed boundary conditions on the lateral ends are w ¼ hx ¼ hy ¼ 0 for the plate and v3 ¼ 0 for the fluid domain. adopting damping ratios equal to n1 ¼ 1:0% and n2 ¼ 2:0% for the first and second modes. and consists in the addition of a eddy viscosity lT to the molecular kinematic viscosity l. . if it is determined with a reference velocity equal to 28 m/s and a reference length equal to the radius of the structure (10 m). and calculating the corresponding circular frequencies x1 and x2 the coefficient a and b can be determined by the procedure described by Bathe [16]. M and K are the damping. a ¼ 0:003529 and b ¼ 0:443162 are obtained. 16. In the perpendicular direction to the flow. Geometry and boundary conditions for the panel vibration induced by vortex shedding originated by a rigid fixed rectangular prism. A. only one layer of elements is used. where the membrane model was built with triangular flat elements and the fluid domain was discretized with tetrahedral elements.

qF Kinematic viscosity. 6.35 0. 14 and 15. Awruch / Computers & Fluids 34 (2005) 249–273 mesh configurations in the fluid domain are shown in Figs. 13. m Specific mass.F. 13. The rebouncing of the membrane. is Table 2 Fluid and panel properties Properties Young’s modulus.0. .268 P. l Value 2. An overpressure on the left-hand side of the field and an underpressure over the middle and the right-hand side of the field are shown in Fig. 9.82 · 10À4 g/cm s Fig. C1 Specific mass. A.0 s. respectively for t ¼ 1:5. qE Sound speed. (b) Detail in the region close to the panel.0. 17.5 MN/cm2 0.1 g/cm3 34 500 cm/s 1. 3. Teixeira.5 and 12. Finite element mesh for the fluid domain.0.M. 10.R. while a vortex is being generated downstream of the structure.18 · 10À3 g/cm3 1. E Poisson’s ratio. (a) General view.

3 -0. 15 the distortion of the finite element mesh at different times and a good performance of the mesh movement algorithm was obtained. (a) Displacements of the panel at the end. Comparison of results obtained in this work and those obtained by Wall and Ramm [17].R. 14. Awruch / Computers & Fluids 34 (2005) 249–273 269 shown in Fig.5 0 -0. 14 at first 3 s.P.3 0. the membrane moves again to the right pushed by the air flow and the vortex formation downstream.5 -1 -1.1 -0.2 displacement (cm) 0.5 1 displacement (cm) 0.F. 1998 0.5 0. (b) Displacements of the panel at the mid span. A. nearly steady conditions are attained.2 -0. at a constant free-stream velocity. When the free stream velocity increases from v10 ¼ 14 to 28 m/s (t ¼ 6–9 s). Present work Wall and Ramm. Within the subsequent 3 s.1 0 -0.5 0 1 tempo (s) 2 3 (a) time (s) time (s) Present work Wall and Ramm.M. behind the membrane. 1998 1. It is observed in Fig.5 0 1 tempo (s) 2 3 (b) Fig.4 -0.4 0. Following this initial phase the free-stream air velocity remains constant and the periodic development and detachment of vortices in correspondence with the membrane structure movement is indicated in Fig. . 18. Teixeira. becomes pronounced.

Panel vibration induced by vortex shedding originated by a fixed rectangular prism A flexible panel is clamped in a fixed rigid rectangular prism.F. A. 19. pressure distributions and velocity vectors here show vortices generated downstream of the structure that was not well captured by the above mentioned reference. Awruch / Computers & Fluids 34 (2005) 249–273 Membrane configurations obtained in the present work are very similar to those presented by Argyris et al. However. 5. 16 using Fig.M. Mesh configurations at (a) t ¼ 3:765 s and (b) t ¼ 3:85899 s.R.3. . This problem is described in Fig. probably because the finite element mesh was poorly refined (taking into account the characteristics of this problem). [26]. Teixeira. which generates vortex shedding inducing vibrations on the deformable structure.270 P.

A non-structured mesh with 19198 tetrahedral elements and 6544 nodes. . The boundary conditions on the lateral ends are w ¼ hx ¼ hy ¼ 0 for the plate and v3 ¼ 0 for the fluid domain. Pressure distributions at (a) t ¼ 3:765 s and (b) t ¼ 3:85899 s.M. is used for the fluid domain and is shown in Fig. 20. a three-dimensional code is used. where tetrahedral elements are employed for the fluid domain and flat plate elements for the flexible panel. The fluid is considered as slightly compressible and Re ¼ 332.P. Awruch / Computers & Fluids 34 (2005) 249–273 271 dimensionless values. Although this case is essentially a two-dimensional problem. The fluid– structure interaction takes place at both sides of the panel (it means that there are two fluid– structure interfaces). A. with only one layer of elements in the perpendicular direction to the flow. Teixeira.F. 17.R. Bearman [27] performed a detailed experimental study and clarified the various regimes of the flow encountered by varying the splitter plate length for range of Reynolds number. while the structural mesh has 84 triangular plate elements and 84 nodes. Fluid and panel properties are given in Table 2. Fig.

31(1):148–53.R. The differences between the dynamic responses of a linear and a non-linear system were studied in the example of a supersonic inviscid flow acting on a deformable flat plate. AIAA J 1993. A monolithical fluid–structure interaction algorithm applied to the piston problem. The complete scheme has much more accuracy. Reasons of these differences are very difficult to establish because both works use different schemes. Casadei F. These authors use a two-dimensional code.F. Displacement at the free and at the mid span are shown in Fig. Arbitrary Lagrangian–Eulerian multicomponent compressible flow with fluid–structure interaction. The method of mesh movement adopted here was shown to be efficient and the complete and simplified algorithms were tested. disadvantage that can be avoided using the simplified scheme. Some discrepancies can be observed between the results of the present work and those obtained by Wall and Ramm [28]. however it requires a considerable CPU time. [2] Blom FJ.M. different meshes and the structure was modeled with beam elements.8589 s are shown in Fig. and are compared with those obtained by Wall and Ramm [28]. But. the precision of the results are reduced and the number of iterations to obtain the convergence of the solution of the equilibrium equations of the structure increases. It was also verified that using procedures with subcycles.3 · 10À5 s/Dt node have been obtained. Int J Numer Methods Fluids 1997. Finally. A. 19 and Fig. Speeds going from 670 to 888 Mflops in a Cray T94 computer and with a CPU time of the order of 1. it was also observed that the decrease in computer processing time may become expressive when a non-structured mesh with very different size of elements are used. Acknowledgements The authors wish to thank CNPq and CAPES for the financial support. [3] Soria A. respectively.25:1263–84. Teixeira. Awruch / Computers & Fluids 34 (2005) 249–273 Simulation of the dynamic behaviour of the panel was carried out using a time interval Dt ¼ 3:0 Â 10À5 s.167:369–91. the good performance of the vectorized algorithms was verified.2–2. which has reasonable precision for problems with small displacements. Direct solution of two-dimensional Navier–Stokes equations for static aeroelasticity problems. whereas in the present work thin plate elements were used. . Comput Methods Appl Mech Eng 1998. 6. The potential of the method was shown through several examples of compressible and slightly compressible flows acting on flexible structures (plates and shells) and good results were obtained. 20. References [1] Felker FF. 18. Conclusions A partitioned algorithm for the solution of fluid–structure interaction problems using a twostep Taylor–Galerkin method for the fluid in the ALE formulation and the Newmark implicit scheme for the solution of the structural dynamic equations is presented in this work. Mesh configurations and pressure distributions at t ¼ 3:765 and 3.272 P.

An arbitrary Lagrangian–Eulerian computing method for all flow speeds. Geometric non-linear analysis of space shell structures using generalized conforming flat shell elements––for space shell structures. Comput Methods Appl Mech Eng 1998. p. Lagrangian Eulerian finite element formulation for incompressible viscous flows. Um L. Awruch AM. [6] Cebral JR. Comput Methods Appl Mech Eng 1989. Halleux JP.26:2135–59. [28] Wall WA. Int J Numer Methods Fluids 1995. 6.29:329–49. J Fluid Mech 1965.174:393–417.. Math Comput 1980. Morton KW. Investigation of the flow behind a two dimensional model with a blunt trailing edge and fitted with splitter plates. 1998. [27] Bearman PW.68:259–310.P. [8] Zienkiewicz OC.14:227–53. 1991. [25] Mittal S. [26] Argyris J et al. New York: McGraw-Hill. [15] Zhang Q. Finite element analysis fluid dynamic. T a pamsa qei. editor. [19] Richtmeyer RD. Three-dimensional simulation of high compressible flows using a multi-time-step integration technique with subcycles.35:687–92. Multiphysics simulation of flow-induced vibrations and aeroelasticity on parallel computing platforms.72:305–50.34:473–97. Conservative load projection and tracking for fluid–structure problems. Van der Vorst HA.21:289–362. [16] Bathe K. Barcelona: CIMNE. Comput Meth Appl Mech Eng 1985.21:807–35. [14] Liu WK. [23] Dowell EH. [22] Meijerink JA.14:941–57. J Comp Phys 1981. [20] Peraire J et al. Appl Math Modell 2001. Chang H. Belytschko T.157:95–114. Liu WK. [17] Chung TJ. Commun Numer Methods Eng 1998. Comput Methods Appl Mech Eng 1999. vol. LeTallec P. Cook JL. L€nher R. 2. An incomplete factorization technique for positive definite linear systems. Teixeira. Maman N. Difference methods for initial value problems. Comput Methods Appl Mech Eng 1981. Singapore: McGraw-Hill.44:134–55. 1967. Aeroelasticity of plates and shells. J Comput Phys 1974. Lesoinne M. In: Idelshon S. [24] Schlichting H. [7] Farhat C. [18] Zienkiewicz OC.33:689–723. [13] Benson DJ. [10] Hirt CW.F. Morgan KO. Awruch / Computers & Fluids 34 (2005) 249–273 273 [4] Rifai SM et al. Guidelines of the usage of incomplete decomposition in solving sets of linear equations as they occur in practical problems. 1968. Chen JS. The Netherlands: Noordhof Int. In: Gallagher R. optimal discretization and application to aeroelasticity. Mixed explicit/implicit time integration of coupled aeroelastic problems: threefield formulation. High speed compressible flow and other advection-dominated o problems of fluid dynamics. Tezduyar TE. Lesoinne M. Int J Numer Methods Fluids 1995. 1975. [9] Teixeira PRF. New York: Interscience. A.R. [12] Donea J. [21] Monteuffel TA. Zimmerman TK. Int J Numer Methods Eng 1988. Wenqi K. Peraire J. 1988.M. Comput Mech. Arbitrary Lagrangean–Eulerian Petrov–Galerkin finite elements for non-linear continua. Parallel finite element simulation of 3D incompressible flows: Fluid–structure interactions. Boundary-layer theory. USA: Prentice Hall. Comput Methods Appl Mech Eng 1988.25/7:613–27. editors. AIAA J o 1997. Ramm E. An arbitrary Lagrangian–Eulerian finite element method for transient dynamic fluid–structure interactions.. Finite elements in fluids. et al. vol. Fluid–structure interaction based upon a stabilized (ALE) finite element method. . Load and motion transfer algorithms for fluid/structure interaction problems with non-matching discrete interfaces: momentum and energy conservation. [11] Hughes TJR. NJ. 1978. Leyden. The finite element method solid and fluid mechanics dynamics and nonlinearity. geometric conservation and distributed solution. Taylor RL. 41–8. L€hner R. Englewood Cliffs. An efficient accurate simple ALE method for non-linear finite element programs. Comput Methods Appl Mech Eng 1982. Finite element Euler computations in three dimensions. 4th ed. Giuliani S.21:933–53. Publ. Finite element procedures. Amdsen AA. 1996. New York: McGraw-Hill.21:241–55. [5] Farhat C.

Sign up to vote on this title
UsefulNot useful