Applied Mathematics and Computation xxx (2011) xxxxxx

Contents lists available at SciVerse ScienceDirect

Applied Mathematics and Computation

journal homepage: www.elsevier.com/locate/amc

Simulation of compressible viscous ow in time-dependent domains

Jan Cesenek a, Miloslav Feistauer a, Jaromr Horcek b, Vclav Kucera a,, Jaroslava Prokopov a
a
Charles University Prague, Faculty of Mathematics and Physics, Sokolovsk 83, 186 75 Praha 8, Czech Republic
b
Institute of Thermomechanics, Academy of Sciences of the Czech Republic, Dolejkova 5, 182 00 Praha 8, Czech Republic

a r t i c l e i n f o a b s t r a c t

Keywords: The paper is concerned with the simulation of viscous compressible ow in time depen-
Compressible NavierStokes equations dent domains. The dependence on time of the domain occupied by the uid is taken into
Time dependent domain account with the aid of the Arbitrary LagrangianEulerian (ALE) formulation of the com-
ALE method pressible NavierStokes equations. They are discretized by the discontinuous Galerkin
Discontinuous Galerkin method
nite element method using piecewise polynomial discontinuous approximations. The
Semi-implicit time discretization
Boundary conditions
time discretization is based on a semi-implicit linearized scheme, which leads to the solu-
Shock indicator tion of a linear algebraic system on each time level. A suitable treatment of boundary con-
Articial viscosity ditions and shock capturing are used, allowing the solution of ow with a wide range of
Mach numbers. The applicability of the developed method is demonstrated by computa-
tional results obtained for compressible viscous ow in a channel with moving walls and
ow induced airfoil vibrations.
2011 Elsevier Inc. All rights reserved.

1. Introduction

The interaction of uid ow with vibrating bodies plays a signicant role in many areas of science and technology. We
mention, for example, development of airplanes (vibrations of wings) or turbines (blade vibrations), some problems from
civil engineering (interaction of wind with constructions as bridges, TV towers or cooling towers of power stations), car
industry (vibration of various elements of a carosery), but also in medicine (hemodynamics or ow in the glottis with vibrat-
ing vocal folds). In a number of these examples the moving medium is a gas, i.e. compressible ow. For low Mach number
ows incompressible models are used (as e.g. in [1,14]), but in some cases compressibility plays an important role.
The solution of uidstructure interaction requires the coupling of the solution of equations describing the uid ow with
equations describing the structural behavior. Due to the deformation and/or vibrations of structures, the computational do-
main is time dependent. There exist several techniques of the solution of incompressible ow in time dependent domains.
See, e.g. [1,14] and references therein. The numerical simulation of compressible ow is much more difcult, particularly in
time dependent domains. It is necessary to overcome difculties caused by nonlinear convection dominating over diffusion,
which leads to boundary layers and wakes for large Reynolds numbers and shock waves and contact discontinuities for high
Mach numbers and instabilities caused by acoustic effects for low Mach numbers.
It appears that a suitable numerical method for the solution of compressible ow is the discontinuous Galerkin nite ele-
ment method (DGFEM). It employs piecewise polynomial approximations without any requirement on the continuity on
interfaces between neighboring elements. The DGFEM was used for the numerical simulation of the compressible Euler
equations, for example, by Bassi and Rebay in [2], where the space DG discretization was combined with explicit RungeKut-
ta time discretization. In [3] Baumann and Oden describe an hp version of the space DG discretization with explicit time

Corresponding author.
E-mail addresses: Jan.Cessa@seznam.cz (J. Cesenek), feist@karlin.mff.cuni.cz (M. Feistauer), jaromirh@it.cas.cz (J. Horcek), vaclav.kucera@email.cz (V.
Kucera), jarkaprokop@post.cz (J. Prokopov).

0096-3003/$ - see front matter 2011 Elsevier Inc. All rights reserved.
doi:10.1016/j.amc.2011.08.077

Please cite this article in press as: J. Cesenek et al., Simulation of compressible viscous ow in time-dependent domains, Appl. Math. Com-
put. (2011), doi:10.1016/j.amc.2011.08.077
2 J. Cesenek et al. / Applied Mathematics and Computation xxx (2011) xxxxxx

stepping to compressible ow. Van der Vegt and van der Ven apply spacetime discontinuous Galerkin method to the solu-
tion of the Euler equations in [15], where the discrete problem is solved with the aid of a multigrid accelerated pseudo-time-
integration. The papers [6,9] are concerned with a semi-implicit DGFEM technique for the solution of inviscid compressible
ow, which is unconditionally stable. In [11], this method was extended so that the resulting scheme is robust with respect
to the magnitude of the Mach number. The paper [5] is concerned with discontinuous Galerkin method for viscous com-
pressible ow.
The goal of our research is the numerical simulation of interaction of compressible ow with structures as, e.g. ow in-
duced airfoil vibrations or the ow past a vibrating elastic wall. We are concerned with the generalization of the method
from [11,9,5] to the solution of compressible inviscid and viscous ow in time dependent domains. The main ingredients
of the method is the discontinuous Galerkin space semidiscretization of the governing equations written in the ALE (arbi-
trary LagrangianEulerian, cf. [14]) form, semi-implicit time discretization, suitable treatment of boundary conditions and
the shock capturing avoiding Gibbs phenomenon at discontinuities. Numerical experiments prove the applicability of the
method.

2. Formulation of the problem

We shall be concerned with the numerical solution of compressible ow in a bounded domain Xt  R2 depending on time
t 2 [0, T]. Let the boundary of Xt consist of three different parts: oXt CI [ CO [ CW t , where CI is the inlet, CO is the outlet and
CW t denotes impermeable walls that may move in dependence on time.
The system describing compressible ow, consisting of the continuity equation, the NavierStokes equations and the en-
ergy equation, can be written in the form

ow X 2
of s w X 2
oRs w; rw
; 1
ot s1
oxs s1
oxs

where

w w1 ; . . . ; w4 T q; qv 1 ; qv 2 ; ET 2 R4 ;
w wx; t ; x 2 Xt ; t 2 0; T ;
T
f i w fi1 ; . . . ; fi4 qv i ; qv 1 v i d1i p; qv 2 v i d2i p; E pv i T ;
 T 2
Ri w; rw Ri1 ; . . . ; Ri4 T 0; sVi1 ; sVi2 ; sVi1 v 1 sVi2 v 2 koh=oxi ;
 
1 ov i ov j
sVij kdivv dij 2ldij v ; dij v :
2 oxj oxi
We use the following notation: q, density; p, pressure; E, total energy; v = (v1, v2), velocity; h, absolute temperature; c > 1,
Poisson adiabatic constant; cv > 0, specic heat at constant volume; l > 0, k = 2l/3, viscosity coefcients; k, heat conduc-
tion. The vector-valued function w is called state vector, the functions fi are the so-called inviscid uxes and Ri represent
viscous terms.
The above system is completed by the thermodynamical relations
   
E
1
p c  1 E  qjv j2 =2 ; h  jv j2 cv : 3
q 2
The resulting system is equipped with the initial condition

wx; 0 w0 x; x 2 X0 ; 4
and the following boundary conditions:

a qjCI qD ; b v jC I
v D v D1 ; v D2 T ; 5
X
2
oh
c sVij ni v j k 0 on CI ;
i;j1
on

oh
d v jC zD velocity of a moving wall; e 0 on CW t ;
Wt
on CW
t

X
2
oh
f sVij ni 0; j 1; 2; g 0 on CO :
i1
on

In order to treat the time dependence of the domain, we use the so-called arbitrary LagrangianEulerian ALE technique, see
e.g. [12]. We dene a reference domain X0 and introduce a regular one-to-one ALE mapping of X0 onto Xt:

At : X0 ! Xt ; i:e: X 2 X0 # x xX; t At X 2 Xt :

Please cite this article in press as: J. Cesenek et al., Simulation of compressible viscous ow in time-dependent domains, Appl. Math. Com-
put. (2011), doi:10.1016/j.amc.2011.08.077
 J. Cesenek et al. / Applied Mathematics and Computation xxx (2011) xxxxxx 3

Here we use the notation X for points in X0 and x for points in Xt .

Further, we dene the domain velocity:
o
~zX; t At X; t 2 0; T; X 2 X0 ;
ot
zx; t ~zA1 x; t; t 2 0; T; x 2 Xt
and the ALE derivative of a function f = f(x, t) dened for x 2 Xt and t 2 [0, T]:

DA o~f
f x; t X; t; 6
Dt ot
where
~f X; t f At X; t; X 2 X0 ; x At X:
As a direct consequence of the chain rule we get the relation

DA f of
divzf  f divz:
Dt ot
This leads to the ALE formulation of the NavierStokes equations

DA w X 2
og s w X2
oRs w; rw
w divz ; 7
Dt s1
ox s s1
oxs

where
g s w : f s w  zs w; s 1; 2;

are the ALE modied inviscid uxes.

We see that in the ALE formulation of the NavierStokes equations the time derivative ow/ot is replaced by the ALE deriv-
ative DAw/Dt, the inviscid uxes fs are replaced by the ALE modied inviscid uxes gs and a new additional reaction term
wdivz appears.

3. Discrete problem

3.1. Discontinuous Galerkin space semidiscretization

For the space semidiscretization we use the discontinuous Galerkin nite element method. We construct a polygonal
approximation Xht of the domain Xt. By T ht we denote a partition of the closure Xht of the domain Xht into a nite number
S
of closed triangles K with mutually disjoint interiors such that Xht K2T ht K.
By F ht we denote the system of all faces of all elements K 2 T ht . Further, we introduce the set of all interior faces
F Iht fC 2 F ht ; C  Xt g, the set of all boundary faces F Bht fC 2 F ht ; C  oXht g and the set of all Dirichlet boundary faces


F Dht C 2 F Bht ; a Dirichlet condition is prescribed on C . Each C 2 F ht is associated with a unit normal vector nC to C. For
C 2 F Bht the normal nC has the same orientation as the outer normal to oXht. We set d(C)= length of C 2 F ht .
L R R L
For each C 2 F Iht there exist two neighboring elements K C ; K C 2 T h such that C  oK C \ oK C . We use the convention
R L L R
that K C lies in the direction of nC and K C lies in the opposite direction to nC. The elements K C ; K C are called neighbors.
L
If C 2 F Bht , then the element adjacent to C will be denoted by K C .
The approximate solution will be sought in the space of piecewise polynomial functions


S ht Sht 4 ; with Sht v ; v jK 2 Pr K 8 K 2 T ht ; 8
where r P 0 is an integer and Pr(K) denotes the space of all polynomials on K of degree 6r. A function u 2 Sht is, in general,
L R
discontinuous on interfaces C 2 F Iht . By uC and uC we denote the values of u on C considered from the interior and the
L
exterior of K C , respectively, and set
 
L R L R
huiC uC uC =2; uC uC  uC : 9

The discrete problem is derived in the following way: We multiply system (7) by a test function uh 2 Sht, integrate over
K 2 T ht , apply Greens theorem, sum over all elements K 2 T ht , use the concept of the numerical ux and introduce suitable
terms mutually vanishing for a regular exact solution. In this way we get the following identity:
X Z DA w
 uh dx bh w; uh ah w; uh J h w; uh dh w; uh h w; uh :
K2T K Dt
ht

Please cite this article in press as: J. Cesenek et al., Simulation of compressible viscous ow in time-dependent domains, Appl. Math. Com-
put. (2011), doi:10.1016/j.amc.2011.08.077
4 J. Cesenek et al. / Applied Mathematics and Computation xxx (2011) xxxxxx

Here
X Z X
2
ouh X Z 
L R
 X Z 
L R

L
bh w; uh  g s w  dx Hg wC ; wC ; nC  uh C dS Hg wC ; wC ; nC  uhC dS
K2T ht K s1 oxs I C B C
C2F ht C2F ht

10
is the convection form, dened with the aid of a numerical ux Hg. We require that it is consistent with the uxes gs:
X
2
Hg w; w; n g s wns n n1 ; n2 ; jnj 1;
s1

conservative:
Hg u; w; n Hg w; u; n;
R
and locally Lipschitz-continuous. The determination of the boundary state wC in the case when C  oXht is described in
Section 3.4.
Further, we dene the viscous form

  X Z X 2
ouh X Z X 2
ah w; uh Rs w; rw  dx  hRs w; rwiC nC s  uh C dS
K2T K s1 ox s I C s1
ht C2F ht

X Z X
2
L
 Rs w; rwnC s  uhC dS; 11
C2F Dht C s1

(we use the incomplete discretization of viscous terms - the so-called IIPG version), the interior and boundary penalty terms
and the right-hand side form, respectively
X Z X Z
J h w; uh rwC  uh C dS rw  uL
hC dS; 12
C2F Iht C C2F Dht C

X Z X
2
h w; uh rwB  uhLC dS: 13
C2F Dht C s1

Here rjC = CWl/d(C) and CW > 0 is a sufciently large constant. The reaction form reads
X Z
dh w; uh w  uh divzdx: 14
K2T ht K

The boundary state wB is dened on the basis of the Dirichlet boundary conditions (5)(a), (b) and (d) and extrapolation:
 
1
wB qD ; qD v D1 ; qD v D2 ; cv qD hC qD jv D j2
L
on CI ; 15
2
L
wB wC on CO ; 16
 
1 L
wB qC ; qC zD1 ; qC zD2 ; cv qC hC qC jzD j2
L L L L L
on CW t : 17
2
The approximate solution is dened as wh(t) 2 Sht such that
X Z DA wh t          
 uh dx bh wh t; uh ah wh t; uh J h wh t; uh dh wh t; uh h wh t; uh
K2T K Dt
ht

holds for all uh 2 Sht, all t 2 (0, T) and wh 0 w0h is an approximation of the initial state w0.

3.2. Time discretization

Let us construct a partition 0 = t0 < t1 < t2. . . of the time interval [0, T] and dene the time step sk = tk+1  tk. We use the
^ kh wkh  Atk  A1
approximations wh t n  wnh 2 S htn ; zt n  zn ; n 0; 1; . . . and introduce the function w t k1 , which is dened
in the domain Xhtk1 . In order to approximate the ALE derivative at time tk+1, we start from its denition and then use the
backward difference:

DA wh ~h
ow ~ k1 X  w
w ~ kh X whk1 x  w
^ kh x
x; t k1 X; t k1  h ; x Atk1 X 2 Xhtk1 :
Dt ot sk sk

Please cite this article in press as: J. Cesenek et al., Simulation of compressible viscous ow in time-dependent domains, Appl. Math. Com-
put. (2011), doi:10.1016/j.amc.2011.08.077
 J. Cesenek et al. / Applied Mathematics and Computation xxx (2011) xxxxxx 5

By the symbol (, ) we shall denote the scalar product in L2 Xhtk1 . A possible full discretization reads: we seek wk1
h
2 S htk1
such that for all uh 2 Shtk1 ; k 0; 1; . . .,
!
wk1 ^ kh
w    k1   k1   k1   k1 
h
; uh bh wk1
h ; uh ah wh ; uh J h wh ; uh dh wh ; uh h wh ; uh : 18
sk
Scheme (18) is of the rst-order in time. In order to increase the accuracy in the time discretization, it is possible to use
the second-order backward difference formula (BDF) for the approximation of the ALE derivative:

DA wh 2sk sk1 sk sk1 ^ k sk

t k1  wk1  w ^ k1 :
w 19
Dt sk sk sk1 h sk sk1 h sk1 sk sk1 h
The problem for obtaining wk1h from (18) is equivalent to a strongly nonlinear algebraic system and its solution is rather
difcult.
Our goal is to develop a numerical scheme, which would be accurate and robust, with good stability properties and ef-
ciently solvable. Therefore, we proceed similarly as in [6] and use a partial linearization of the forms bh and ah. This approach
leads to a scheme that requires the solution of only one large sparse linear system on each time level. The efciency and
accuracy of this technique was analyzed in [6,9,11] in the case of inviscid ow, where advantages of this method are dem-
onstrated by numerical experiments. Here we show that this technique can be extended to the solution of viscous compress-
ible ow in time-dependent domains with applications to uidstructure interaction.
The linearization of the rst term of the form bh is based on the relations
    k1     k1  
g s wk1
h As w h  zk1
s I wk1
h
h
 As w  zk1
s I wk1
h ;

 hk1 we dene the state obtained by the extrapolation:

where As w is the Jacobi matrix of fs(w), cf. [10]. By w

 k1 ^ kh  hk1
sk sk1 ^ k sk ^ k1
w w and w wh  w ; 20
h
sk1 sk1 h
in the case of the rst-order time discretization and second-order time discretization, respectively.
The second term of bh is linearized with the aid of the Vijayasundaram numerical ux (cf. [16]) dened in the following
way. Taking into account the denition of gs, we have

Dg s w Df s w
 zs I As w  zs I; 21
Dw Dw

and can write:

X2
Dg s w X2
Pg w; n ns As wns  zs ns I: 22
s1
Dw s1

By [10], this matrix is diagonalizable. It means that there exists a nonsingular matrix T Tw; n such that

Pg T T1 ; diagk1 ; . . . ; k4 ; 23

where ki = ki(w, n) are eigenvalues of the matrix Pg . Now we dene the positive and negative parts of the matrix Pg by
 
Pg T 
T1 ; 
diag k1 ; . . . ; k4 ; 24

where k+ = max(k, 0), k = min(k, 0). Using the above concepts, we introduce the modied Vijayasundaram numerical ux (cf.
[16] or [10]) as:
w w  w w 
L R L R
H g wL ; wR ; n Pg ; n wL Pg ; n wR : 25
2 2
Using the above denition of the numerical ux, we introduce the approximations:
   k1  k1L  k1  k1R
k1L k1R
Hg whC ; whC ; nC  Pg wh
C
; nC whC Pg wh
C
; nC w h C :

for C 2 F Ihtk1 and

   k   k  k1R
k1L k1R
Hg whC ; whC ; nC  Pg w h ; nC whk1L
C Pg w  hC :
 h ; nC w
C C

Please cite this article in press as: J. Cesenek et al., Simulation of compressible viscous ow in time-dependent domains, Appl. Math. Com-
put. (2011), doi:10.1016/j.amc.2011.08.077
6 J. Cesenek et al. / Applied Mathematics and Computation xxx (2011) xxxxxx

for C 2 F Bhtk1 . In this way we get the form

X Z X
2
^ w k1 k1
   k1    ouh x
 x  zsk1 xI wk1 x 
bh  h ; wh ; u h  As w dx
K2T ht K s1 oxs
k1

X Z  
 k1   k1 
Pg w h ; nC whk1L Pg w h ; nC whk1R  uh dS
C2F Iht C
k1
Z 
X  k1   k1  k1R 
Pg w h ; nC whk1L Pg w h
 h ; nC w  uh dS: 26
C2F Bht C
k1

For the determination of the boundary state w  k1R

h , if C  oXhtk1 , we refer to Section 3.4.
The linearization of the form ah is based on the fact that Rs(wh, rwh) is linear in rw and nonlinear in w. We get the lin-
earized viscous form
X Z X
2 X Z X
2
 k1 k1   k1  ouh   k1   
 h ; wh ; uh
^h w
a  h ; rwk1
Rs w h  dx   h ; rwk1 nC s  uh dS
Rs w
K2T ht K s1 oxs I C s1
k1 C2F ht
k1

X Z X
2
 k1 
  h ; rwk1
Rs w h nC s  uh dS: 27
C2F Dht C s1
k1

3.3. Articial viscosity

In high-speed gas ow with large Mach numbers, discontinuities called shock waves or contact discontinuities
appear. In viscous high-speed ow these discontinuities may be smeared due to viscosity and heat conduction. Near
shock waves and contact discontinuities, the so-called Gibbs phenomenon, manifested by nonphysical spurious over-
shoots and undershoots, usually occurs in the numerical solution. In order to avoid this undesirable phenomenon, it
is necessary to apply a suitable limiting procedure. Here we use the approach proposed in [11] based on the disconti-
nuity indicator
Z  
 2
g k K q^ kh dS= hK jKj3=4 ; K 2 T htk1 ; 28
oK
 k
introduced in [7]. By q ^ kh on the boundary o K and jKj denotes the area of the element
^ h we denote the jump of the function q
K. Then we dene the discrete discontinuity indicator

Gk K 0 if g k K < 1; Gk K 1 if g k K P 1; K 2 T htk1 ; 29

and the articial viscosity forms

X Z
 k k1 
^h w
b ^ h ; wh ; uh m1 hK Gk K rwk1
h  ruh dx;
K2T ht K
k1

  X 1   Z     30
bJ h w
^ kh ; wk1 Gk K C Gk K C
L R
wk1
h ; uh m2 h  uh dS;
2 C
C2F Iht
k1

with parameters m1, m2 = O(1).

The resulting scheme has the following form: We seek whk1 2 S htk1 such that for all uh 2 S htk1 ; k 0; 1; . . .,
!
whk1  w
^ kh ^ w k k1
  k   k1   k1  
^ ^ k k1

; uh bh ^ h ; wh ; u h a ^ h ; wk1
^h w h ; uh J h wh ; uh dh wh ; uh bh wh ; wh ; uh
sk
 k k1   k1 
bJ h w
^ h ; wh ; uh w
 B ; uh ; 31
in
 the case of the rst-order time discretization. The second-order time discretization is obtained by replacing the expression
whk1  w
^ kh ; uh =sk by the approximation (19) and replacing w ^ kh in the forms a ^ by w
^h and b  hk1 .
h
This method successfully overcomes problems with the Gibbs phenomenon in the context of the semi-implicit scheme. It
is important that the indicator Gk(K) vanishes in regions, where the solution is regular and, therefore, the numerical solution
does not contain any nonphysical entropy production in these regions. If the described articial viscosity is not applied, then
in the case of high-speed ow with shock waves and contact discontinuities the computational process collapses, because
negative values of the approximation of the density and pressure appear.

Please cite this article in press as: J. Cesenek et al., Simulation of compressible viscous ow in time-dependent domains, Appl. Math. Com-
put. (2011), doi:10.1016/j.amc.2011.08.077
 J. Cesenek et al. / Applied Mathematics and Computation xxx (2011) xxxxxx 7

^
3.4. Treatment of boundary states in the form b h

k1R
If C 2 F Bhtk1 , it is necessary to specify the boundary state w
 hC appearing in the numerical ux Hg in the denition of the
inviscid form b ^ . For simplicity we shall use the notation w(R) for values of the function w
 k1R which should be determined at
h hC
(L)
individual integration points on the face C. Similarly, w will denote the values of w  hk1L
C at the corresponding points.
On the inlet, which is assumed xed, we proceed in the same way as in [11], Section 4. Using rotational invariance, we
transform the Euler equations

ow X 2
of s w
0;
ot s1
oxs

to the coordinates ~
x1 , parallel with the normal direction n = nC to oX, and ~
x2 , tangential to the boundary. In this way we ob-
tain a system for an unknown function q = Q(n)w, where
0 1
1; 0; 0; 0
B 0; n ; n ; 0 C
B 1 2 C
Q n B C 32
@ 0; n2 ; n1 ; 0 A
0; 0; 0; 1
is the rotational matrix. Now we neglect the derivative with respect to ~
x2 and linearize the system around the state
q(L) = Q(n)w(L). In this way we obtain the linear system

oq oq
A1 qL 0; 33
ot o~x1

for the vector-valued function q q~

x1 ; t, considered in the set (1, 0) (0, 1) and equipped with the initial and boundary
conditions:

q~x1 ; 0 qL ; ~x1 < 0; and q0; t qR ; t > 0: 34

(R)
The goal is to choose q in such a way that this initial-boundary value problem is well posed, i.e. has a unique solution. The
method of characteristics leads to the following process:
Let us put q = Q(n)w, where w is a given boundary state at the inlet. We calculate the eigenvectors rs corresponding to
the eigenvalues ks, s = 1, . . . , 4, of the matrix A1 qL , arrange them as columns in the matrix T and calculate T1 . Now we set:

a T1 qL ; b T1 q
35
(R)
and dene the state q by the relations
X
4 
as ; ks P 0;
qR : cs rs ; cs 36
s1
bs ; ks < 0:

Finally, the sought boundary state w(R) is dened as

wR Q 1 nqR : 37
On the impermeable moving wall we prescribe the normal component of the velocity
v  n zD  n; 38
where n is the unit outer normal to CW t and zD is the wall velocity. This implies that two eigenvalues of Pg w; n are equal to
zero, one eigenvalue is positive and one eigenvalue is negative. Then, in analogy to [10], Section 3.3.6, we should prescribe
one quantity, namely v  n, and extrapolate three quantities tangential velocity, density and pressure.
However, here we dene the numerical ux on CW t as the physical ux through the boundary with the assumption (38)
taken into account. Thus, on CW t we write
X
2
g s wns v  n  zD  nw p0; n1 ; n2 ; v  nT p0; n1 ; n2 ; zD  nT : H g : 39
s1

On the outlet (which does not depend on time) the pressure is prescribed and other variables are extrapolated. However,
numerical experiments show that this treatment of the outlet boundary conditions can lead to the reection of a strong
intensity vortex on an articial outlet in the numerical simulation. This problem, which does not appear in the examples
presented in Section 4, will require a special analysis in the future.

^ ; d ; J ; bJ and b
^h ; b
Remark 1. In practical computations, integrals appearing in the denitions of the forms a ^h are evaluated
h h h h
with the aid of quadrature formulas.

Please cite this article in press as: J. Cesenek et al., Simulation of compressible viscous ow in time-dependent domains, Appl. Math. Com-
put. (2011), doi:10.1016/j.amc.2011.08.077
8 J. Cesenek et al. / Applied Mathematics and Computation xxx (2011) xxxxxx

The developed numerical scheme can also be used for the numerical solution of inviscid ow, if we set l = k = k = 0. See
[11].
The linear algebraic system equivalent to (31) is solved either by a direct solver UMFPACK ([4]) or by the GMRES method
with block diagonal preconditioning.
If we set r = 0, we get a nite volume scheme.

4. Numerical experiments

In order to demonstrate the applicability of the developed method, we present here results of two numerical experiments.
In both cases piecewise quadratic nite elements (r = 2) in the space discretization are used. For the time discretization the
second-order BDF formula from Section 3.2 is used.

4.1. Flow through a channel with moving walls

In the rst example we present results of numerical experiments carried out for viscous compressible ow in a channel
with geometry from [13] inspired by the shape of the human glottis and a part of supraglottal spaces as shown in Fig. 1. The
walls are moving in order to mimic the vibrations of vocal folds during voice production. The lower channel wall between the
points A and B and the upper wall symmetric with respect to the axis of the channel are vibrating up and down periodically
with frequency 100 Hz. This movement is interpolated into the domain resulting in the ALE mapping At .
The width of the channel at the inlet (left part of the boundary) is H = 0.016 m and its length is L = 0.16 m. The width of
the narrowest part of the channel (at the point C) oscillates between 0.0004 and 0.0028 m with frequency 100 Hz. We con-
sider the following input parameters and boundary conditions: magnitude of the inlet velocity vin = 4 m/s, the viscosity
l = 15  106 kg m1 s1, the inlet density qin = 1.225 kg m3, the outlet pressure pout = 97611 Pa, the Reynolds number Re =
qinvinH/l = 5227, heat conduction coefcient k = 2.428  102 kg m s2 K1, the specic heat cv = 721.428 m2 s2 K1, the
Poisson adiabatic constant c = 1.4. The inlet Mach number is Min = 0.012.
In [13], the described channel ow was solved by the rst-order nite volume method under the assumption that the ow
is symmetric with respect to the axis of the channel. This means that the results presented in [13] do not reect the behavior
of real ow. Here, we use piecewise quadratic nite elements and we do not require the symmetry of the ow eld.
Fig. 2 shows computed streamlines of the solution at different time instants 2.016, 2.124, 2.448, 2.664 s during the fourth
period of the motion. In the solution we can observe large vortex formation convected through the domain. The ow eld is
neither periodic, nor axisymmetric, in spite of the fact that the computational domain is axisymmetric and the motion of the
channel walls is periodic and symmetric as well. We can observe the so-called Coanda effect, when the main ow is attached
to one of the walls. This effect is not present in results of the paper [13].

4.2. Flow induced airfoil vibrations

The second example is concerned with the simulation of vibrations of elastically supported airfoil NACA 0012 induced by
compressible viscous ow. The airfoil has two degrees of freedom: the vertical displacement H (positively oriented down-
wards) and the angle of rotation around an elastic axis a (positively oriented clockwise), cf. Fig. 3. The motion of the airfoil
is described by the system of nonlinear ordinary differential equations for unknowns H, a:
kHH H Sa a
mH cos a  Sa a_ 2 sin a dHH H_ Lt; cos a Ia a
Sa H kaa a daa a_ Mt: 40
The dot and two dots denote the rst-order and second-order time derivative, respectively. We use the following notation:
Lt, aerodynamic lift force (upwards positive); Mt, aerodynamic torsional moment (clockwise positive); m, mass of the
airfoil; Sa, static moment around the elastic axis EA; Ia, inertia moment around the elastic axis EA; kHH, bending stiffness;
kaa, torsional stiffness; dHH, structural damping in bending; daa, structural damping in torsion; c, length of the chord of
the airfoil; l, airfoil depth.
_
System (40) is equipped with the initial conditions prescribing the values H(0), a(0), H0; a_ 0. It is transformed to a rst-
order ODE system and solved numerically by the fourth-order RungeKutta method. For the derivation of Eqs. (40), see [14].
The aerodynamic lift force L acting in the vertical direction and the torsional moment M are dened by
Z X
2 Z X
2
L l s2j nj dS; M l sij nj rort
i dS; 41
CWt j1 CWt i;j1

Fig. 1. Computational domain for ow in human vocal folds.

Please cite this article in press as: J. Cesenek et al., Simulation of compressible viscous ow in time-dependent domains, Appl. Math. Com-
put. (2011), doi:10.1016/j.amc.2011.08.077
 J. Cesenek et al. / Applied Mathematics and Computation xxx (2011) xxxxxx 9

Fig. 2. Streamlines at time instants t = 2.016, 2.124, 2.448, 2.664 s.

Fig. 3. Computational domain for ow around a vibrating airfoil.

where
 
ov i ov j
sij p k divv dij l ;
oxj oxi 42
r ort
1 x2  xEA2 ; r ort
2 x1  xEA1 :

By sij we denote the components of the stress tensor, dij denotes the Kronecker symbol, n = (n1, n2) is the unit outer normal to
oXt on CWt (pointing into the airfoil) and xEA = (xEA1, xEA2) is the position of the elastic axis (lying in the interior of the airfoil).
Relations (41) and (42) dene the coupling of the uid dynamical model with the structural model.

4.2.1. Algorithm of the ow induced airfoil vibrations simulation

In the solution of the complete coupled uidstructure interaction problem we apply the following algorithm:

(1) Assume that the approximate solution of the discrete ow problem (31) at time levels tk1 and tk is known and the
force L and torsional moment M are computed from (41).
(2) Extrapolate L and M on the time interval [tk, tk+1].
(3) Compute the displacement H and angle a at time tk+1 as the solution of system (40).

Please cite this article in press as: J. Cesenek et al., Simulation of compressible viscous ow in time-dependent domains, Appl. Math. Com-
put. (2011), doi:10.1016/j.amc.2011.08.077
10 J. Cesenek et al. / Applied Mathematics and Computation xxx (2011) xxxxxx

(4) Determine the position of the airfoil at time tk+1, the domain Xtk1 , the ALE mapping and the domain velocity at time tk+1.
(5) Solve the discrete problem (31) at time level tk+1.
(6) Compute L and M at time tk+1 and interpolate L and M on [tk, tk+1].
(7) Is higher accuracy needed? YES: go to (3); NO: k : k + 1, go to (2).

If in step (7) one goes to (2), the so-called loose (weak) coupling is applied. In our numerical experiments the stronger
coupling was applied with 45 loops for obtaining the difference between two approximations of H and a less than 105.
The ALE mapping and the domain velocity are computed in the same way as in [8].

4.2.2. Results of numerical experiments

(I) The simulation of ow induced airfoil vibrations was carried out for the following data: m = 0.086622 kg,
Sa = 0.000779673 kg m, Ia = 0.000487291 kg m2, kHH = 105.109 Nm1, kaa = 3.696682 Nm1, l = 0.05 m, c = 0.3 m,
l = 1.8375  105 kg m1 s1, far-eld density q = 1.225 kg m3, H(0) = 0.02 m, a(0) = 6 degrees, H0 _ 0; a_ 0. We
neglect the structural damping. The elastic axis is placed on the airfoil chord at the 40% distance from the leading edge.
The computational process starts at time t = d < 0 by the solution of the ow, keeping the airfoil in a xed position
given by the prescribed initial translation H and the angle of attack a. Then, at time t = 0 the airfoil is released and we
continue by the solution of a complete uidstructure interaction problem.

20 8
15 6
10 4
5
2
H[mm]

[]

0
0
-5
-10 -2

-15 -4
-20 -6
0 1 2 3 4 0 1 2 3 4

15 7
6
10 5
5 4
3
0
H[mm]

2
[]

-5 1
0
-10 -1
-15 -2
-3
-20 -4
0 0.5 1 1.5 2 0 0.5 1 1.5 2

10 7
5 6
5
0
4
-5
H[mm]

3
[]

-10 2
1
-15
0
-20 -1
-25 -2
0 0.5 1 1.5 2 0 0.5 1 1.5 2

40 12
10
20
8
0 6
-20 4
H[mm]

[]

2
-40 0
-60 -2
-4
-80
-6
-100 -8
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

Fig. 4. Displacement H (left) and rotation angle a (right) of the airfoil in dependence on time for far-eld velocity 10, 20, 30 and 40 ms1.

Please cite this article in press as: J. Cesenek et al., Simulation of compressible viscous ow in time-dependent domains, Appl. Math. Com-
put. (2011), doi:10.1016/j.amc.2011.08.077
 J. Cesenek et al. / Applied Mathematics and Computation xxx (2011) xxxxxx 11

0 12

-2 10

8
-4

H[mm]

[]
6
-6
4
-8 2

-10 0
0 0.02 0.04 0.06 0.08 0.1 0 0.02 0.04 0.06 0.08 0.1
t[s] t[s]

Fig. 5. Displacement H (left) and rotation angle a (right) of the airfoil in dependence on time for far-eld velocity 290 m/s and far-eld Mach number 0.85.

Fig. 6. Flow past an airfoil: Mach number isolines for far-eld velocity 290 m/s and far-eld Mach number 0.85 at time instants t = 0.00261, 0.00661,
0.00831, 0.00961 s, ordered from left to right in rows.

Fig. 4 shows the displacement H and the rotation angle a in dependence on time for the far-eld velocity 10, 20, 30 and
40 m/s. The corresponding Reynolds number was in the range 2  105  8  105. We see that for the velocities 10, 20 and
30 m/s the vibrations are damped, but for the velocity 40 m/s we get the utter instability when the vibration ampli-
tudes are increasing in time. The monotonous increase and decrease of the average values of H and a, respectively,
shows that the utter is combined with a divergence instability in the presented example.
(II) In the above examples the ow was subsonic. The described method was also applied to transonic ow with far-eld
velocity 290 m/s, far-eld Mach number 0.85, Reynolds number 5000 and initial data H(0) = 0, a(0) = 4 degrees,
_
H0 a_ 0 0. In this case it was necessary to consider harder bending and torsional stiffnesses. We set
kHH = 105109 Nm1 and kaa = 36.956 Nm rad1. Fig. 5 shows the time dependence of H and a. In Fig. 6, Mach number
isolines at time instants t = 0.00261, 0.00661, 0.00831, 0.00961 s are shown. We see an interesting system of shock
waves, separated boundary layer, wake moving in time and vortices leaving the airfoil.

5. Conclusion

We have presented an efcient numerical scheme for the solution of the compressible NavierStokes equations in time
dependent domains and the simulation of ow induced airfoil vibrations. It is based on several important ingredients:

 the ALE method applied to the compressible NavierStokes equations,

 the application of the discontinuous Galerkin method for the space discretization,
 semi-implicit time discretization,

Please cite this article in press as: J. Cesenek et al., Simulation of compressible viscous ow in time-dependent domains, Appl. Math. Com-
put. (2011), doi:10.1016/j.amc.2011.08.077
12 J. Cesenek et al. / Applied Mathematics and Computation xxx (2011) xxxxxx

 suitable treatment of boundary conditions,

 articial viscosity applied in the vicinity of discontinuities.

The developed method behaves as unconditionally stable and appears to be robust with respect to the magnitude of the
Mach number. The presented examples demonstrate that the method can be applied to the numerical solution of compress-
ible ow with very low Mach numbers as well as high-speed ow with shock waves and contact discontinuities.
Future work will be concentrated on the following topics:

 further analysis of the robustness and accuracy of the method with respect to the Mach number and Reynolds number,
 investigation of various types of boundary conditions,
 the realization of a remeshing in case of closing the channel during the oscillation period of the channel walls,
 the coupling of the developed method with the solution of elasticity equations describing the deformation of vocal folds,
 the use of a suitable turbulence model.

Acknowledgements

This work was supported by the research project MSM 0021620839 (M. Feistauer, V. Kucera) and by the Necas Center for
Mathematical Modelling, project LC06052 (J. Cesenek), both nanced by the Ministry of Education of the Czech Republic. It
was also partly supported by the grants No. 201/08/0012 (M. Feistauer, V. Kucera) and P101/11/0207 (J. Horcek in the year
2011) of the Czech Science Foundation, and by the grant SVV-2010261316 nanced by the Charles University in Prague (J.
Prokopov).

References

[1] S. Badia, R. Codina, On some uid-structure iterative algorithms using pressure segregation methods. Application to aeroelasticity, Int. J. Numer.
Methods Eng. 72 (2007) 4671.
[2] F. Bassi, S. Rebay, High-order accurate discontinuous nite element solution of the 2D Euler equations, J. Comput. Phys. 138 (1997) 251285.
[3] C.E. Baumann, J.T. Oden, A discontinuous hp nite element method for the Euler and NavierStokes equations, Int. J. Numer. Methods Fluids 31 (1999)
7995.
[4] T.A. Davis, I.S. Duff, A combined unifrontal/multifrontal method for unsymmetric sparse matrices, ACM Trans. Math. Softw. 25 (1999) 119.
[5] V. Dolej, Semi-implicit interior penalty discontinuous Galerkin methods for viscous compressible ows, Commun. Comput. Phys. 4 (2008) 231274.
[6] V. Dolej, M. Feistauer, A semi-implicit discontinuous Galerkin nite element method for the numerical solution of inviscid compressible ow, J.
Comput. Phys. 198 (2004) 727746.
[7] V. Dolej, M. Feistauer, C. Schwab, On some aspects of the discontinuous Galerkin nite element method for conservation laws, Math. Comput. Simul.
61 (2003) 333346.
[8] L. Dubcov, M. Feistauer, J. Horcek, P. Svcek, Numerical simulation of interaction between turbulent ow and a vibrating airfoil, Comput. Visual. Sci.
12 (2009) 207225.
[9] M. Feistauer, V. Dolej, V. Kucera, On the discontinuous Galerkin method for the simulation of compressible ow with wide range of Mach numbers,
Comput. Visual. Sci. 10 (2007) 1727.
[10] M. Feistauer, J. Felcman, I. Strakraba, Mathematical and Computational Methods for Compressible Flow, Clarendon Press, Oxford, 2003.
[11] M. Feistauer, V. Kucera, On a robust discontinuous Galerkin technique for the solution of compressible ow, J. Comput. Phys. 224 (2007) 208221.
[12] T. Nomura, T.J.R. Hughes, An arbitrary LagrangianEulerian nite element method for interaction of uid and a rigid body, Comput. Methods Appl.
Mech. Eng. 95 (1992) 115138.
[13] P. Puncochrov, J. Furst, K. Kozel, J. Horcek, Numerical solution of compressible ow with low Mach number through oscillating glottis, in: I.
Zolotarev, J. Horcek (Eds.), Proceedings of the 9th International Conference on Flow-Induced Vibration (FIV2008), Institute of Thermomechanics
ASCR, Prague, 2008, pp. 135140.
[14] P. Svcek, M. Feistauer, J. Horcek, Numerical simulation of ow induced airfoil vibrations with large amplitudes, J. Fluids Struct. 23 (2007) 391411.
[15] J.J.W. van der Vegt, H. van der Ven, Spacetime discontinuous Galerkin nite element method with dynamic grid motion for inviscid compressible
ow, J. Comput. Phys. 182 (2002) 546585.
[16] G. Vijayasundaram, Transonic ow simulation using upstream centered scheme of Godunov type in nite elements, J. Comput. Phys. 63 (1986) 416
433.

Please cite this article in press as: J. Cesenek et al., Simulation of compressible viscous ow in time-dependent domains, Appl. Math. Com-
put. (2011), doi:10.1016/j.amc.2011.08.077