You are on page 1of 29

Applied Mathematical Modelling 28 (2004) 211–239

www.elsevier.com/locate/apm

A finite volume unstructured mesh approach to dynamic


fluid–structure interaction: an assessment of the challenge
of predicting the onset of flutter
A.K. Slone *, K. Pericleous, C. Bailey, M. Cross, C. Bennett
Centre for Numerical Modelling and Process Analysis, University of Greenwich, The Old Royal Naval College,
Park Row, London SE10 9LS, UK
Received 2 April 2002; received in revised form 9 June 2003; accepted 31 July 2003

Abstract

Computational modelling of dynamic fluid–structure interaction (DFSI) is a considerable challenge. Our


approach to this class of problems involves the use of a single software framework for all the phenomena
involved, employing finite volume methods on unstructured meshes in three dimensions. This method
enables time and space accurate calculations in a consistent manner. One key application of DFSI simu-
lation is the analysis of the onset of flutter in aircraft wings, where the work of Yates et al. [Measured and
Calculated Subsonic and Transonic Flutter Characteristics of a 45 degree Sweptback Wing Planform in
Air and Freon-12 in the Langley Transonic Dynamic Tunnel. NASA Technical Note D-1616, 1963] on the
AGARD 445.6 wing planform still provides the most comprehensive benchmark data available. This paper
presents the results of a significant effort to model the onset of flutter for the AGARD 445.6 wing planform
geometry. A series of key issues needs to be addressed for this computational approach.

• The advantage of using a single mesh, in order to eliminate numerical problems when applying boundary
conditions at the fluid-structure interface, is counteracted by the challenge of generating a suitably high
quality mesh in both the fluid and structural domains.
• The computational effort for this DFSI procedure, in terms of run time and memory requirements, is
very significant. Practical simulations require even finer meshes and shorter time steps, requiring parallel
implementation for operation on large, high performance parallel systems.
• The consistency and completeness of the AGARD data in the public domain is inadequate for use in the
validation of DFSI codes when predicting the onset of flutter.

 2003 Elsevier Inc. All rights reserved.

*
Corresponding author. Tel.: +44-(0) 208-331-8554; fax: +44-(0) 208-331-8695.
E-mail address: a.slone@gre.ac.uk (A.K. Slone).

0307-904X/$ - see front matter  2003 Elsevier Inc. All rights reserved.
doi:10.1016/S0307-904X(03)00142-2
212 A.K. Slone et al. / Appl. Math. Modelling 28 (2004) 211–239

Keywords: Fluid–structure interaction; Finite volume; Transient structural dynamics; Geometric conservation law;
Newmark algorithm

1. Introduction

Fluid–structure interaction, as applied to flexible structures, has wide application in diverse


areas such as flutter in aircraft, wind response of buildings and structures, flows in elastic pipes
and blood vessels. A comprehensive computational model of these phenomena involving closely
coupled fluid flow and structural deformation, is a considerable challenge and, until recently,
work in this area typically focused on one phenomenon and represented the behaviour of the other
more simply. However, strategies are now emerging for solving the full coupling between the fluid
and solid mechanics behaviour. Solving dynamic fluid–structure interaction (DFSI) problems in a
generic time and space accurate fashion without making any simplifying assumptions in either the
fluid or structural sub-domains has acquired an increasing degree of attention in the last few years
[2–5]. Hughes et al. developed a family of FE based procedures with the SPECTRUM [2] code for
modelling DFSI problems in the context of a more general Ômulti-physicsÕ analysis framework
[6,7]. Wall and Ramm [3] have pursued the FE method, with some considerable success, for two-
dimensional dynamic fluid–structure problems, including Navier–Stokes flows, following the
overall arbitrary Lagrangian Eulerian (ALE) method of Farhat et al. [4], except that FarhatsÕ team
employed a mixed FE–FV formulation. The authors and their co-workers have pursued the de-
velopment of multi-physics modelling and software tools [8–10] based on finite volume unstruc-
tured mesh (FV-UM) techniques for two- and three-dimensional configurations. The low order
elements supported are constant strain triangular elements and bi-linear quadrilateral elements for
two-dimensional analysis plus linear tetrahedral, bi-linear pentahedral and tri-linear hexahedral
elements for three-dimensional analysis. The arising multi-physics simulation toolkit PHYSICA
[9], has been used as the environment for modelling a range of DFSI process [11–15].
The development of each of the above multi-physics software environments exhibits some
common features:

• Compatible methods of spatial discretisation, where either a FE or FV approach is used for


each of the physical phenomena.
• Similar solution procedures for all phenomena.
• A single mesh encompassing all sub-domains.
• A single database for the entire solution domain.

In the full spectrum of multi-physics modelling, the above features enable the implementation
of volume sources representing the impact of one phenomena on another (for example, electro-
magnetic fields coupled to fluid flow in magnetohydrodynamics (MHD) [16]) and in the imple-
mentation of internal boundary conditions between sub-domains, each with its own particular
phenomena (for example, in metal forming processes [17], which involves interaction between a
free surface fluid and a structural domain). The compatibility of the mesh and the discretisation
procedure enables the exchange of volume sources and internal boundary conditions to be de-
livered in a straightforward fashion in a space and time accurate manner. For the application of
A.K. Slone et al. / Appl. Math. Modelling 28 (2004) 211–239 213

multi-physics tools to DFSI simulation [2–5,11–15] there are a number of common features that
may be identified, i.e., a single family of numerical procedures is employed for both the fluid and
structural sub-domains, where the structural deformation is calculated from the solution of the
dynamic equilibrium equations. The calculated deformation subsequently may be used to com-
pute the strain and stress throughout the structural sub-domain. Also the flow field is solved as
either an Euler or a Navier–Stokes fluid for a dynamically changing mesh representing the dy-
namically changing geometrical domain. The procedure ensures that the mesh in the fluid sub-
domain obeys the geometric conservation law and the coupling between the flow and structural
sub-domains is explicit [11–13,15].
One application of these DFSI procedures is flow induced vibration of flexible elastic structures
and one classical example, in the context of aeroelasticity, is the consideration of the onset of
flutter. Aeroelastic simulation has been the subject of a large research investment over many years
[4,5,18–30]. The conventional approach to aeroelasticity simulation involves:

• The calculation of the natural frequencies and mass normalised mode shapes of the structure
geometry, (typically the wing planform) as a free undamped vibration problem [31].
• The re-formulation of the structural dynamic equilibrium equation in a normalised form, where
the newly defined principal co-ordinates de-couple the structural dynamic equilibrium equa-
tion.
• The load from the fluid on the structure, which for an Euler fluid is given by the fluid pressure,
is normalised within the uncoupled equation of motion.
• The degrees of freedom may be reduced to retain only the vibrational modes of interest.

The impact of the structural deformation is coupled to the flow field by one of two methods.
Either through a transpiration procedure [21,32], where the flow mesh is static and the effect of the
dynamic structural deformation is accounted for at the fluid-structure interface by a velocity
boundary condition or by directly reflecting the modification of the flow domain by deforming
the mesh [4,5]. Here, the mesh movement is accounted for using the static equilibrium equation with
a pseudo-stiffness matrix, where the mesh is modelled using a linear spring approximation [18,33].
Typically, the calculations are based upon the experimentally determined flutter conditions.
These calculations attempt to ensure that the model is consistent with the experimental data by
predicting that the oscillation will damp, be self-sustaining or grow, according to the flow con-
ditions. This approach is able to represent the whole dynamic envelope, but its predictive capa-
bility requires experimental data concerning the flutter conditions. Thus, its genuine predictive
capability is limited, with respect to the onset of flutter.
The authors have been involved in the modelling of DFSI problems for a number of years [11–
15] and the techniques have been embedded within an open object based multi-physics modelling
software toolkit, PHYSICA [9]. The broad approach to DFSI considered here is conventional in
that the fluid and structure are solved in an iterative staggered manner [13,14]. The pressure and
viscous stresses from the flow field provide load conditions for the structure, whilst, at the fluid–
structure interface, the deformed structure provides boundary conditions from the structure to the
fluid. The structure algorithm also provides the necessary mesh adaption for the flow field, the
effect of which is accounted for in the flow algorithm. The dynamic form of the solid equations is
solved using a predictor–corrector version of the Newmark algorithm [14,15]. The mesh is moved
214 A.K. Slone et al. / Appl. Math. Modelling 28 (2004) 211–239

within the solid algorithm using the static equilibrium equation with a pseudo-stiffness matrix,
which unlike the linear spring based methods [33,34], is achieved by defining separate values of
YoungÕs modulus for each of the deforming regions of the mesh.
This more generic multi-physics DFSI scheme does not require any experimental data con-
cerning the flutter envelope to predict the oscillatory behaviour of the structure, it only requires
the flow conditions together with the geometric and material properties of the structure. Thus, this
scheme may have the potential to provide a genuine prediction of the onset of flutter without a
priori knowledge of the experimentally measured flutter conditions. One set of experimental data
that has been extensively used in the verification of computational aeroelastic procedures is based
on the AGARD 445.6 wing planform carried out by Yates [1] at NASA in the early 1960s.
Although the AGARD 445.6 flutter problem has been studied by a number of researchers, this
has been with varying degrees of success. Typically, the method involves the solution of the fluid
flow using either the Euler or Navier–Stokes equations and a more simplistic representation of the
wing planform dynamics. This approach has been taken by Lee-Rausch and Batina [22,23], where
upwind-spatial differencing is used to solve the three-dimensional time dependent Euler equations
and thin layer Navier–Stokes equations for fluid flow. The structural analysis is performed using a
transient aeroelastic procedure, where the aeroelastic equations of motion are formed from a finite
modal series of free-vibration modes. Melville and Morton [35], employed a similar strategy to
Lee-Rausch and Batina, where the aeroelastic equations were solved on six overlapped sub-grids
for parallel implementation. Raveh [29] employed a Reduced Order Model of the aerodynamic
system, where the modal impulses from the first four elastic modes were introduced into the flow
analysis.
Farhat et al. [4,5] employed a three-field ALE finite volume/finite element approach to the
solution of large scale three-dimensional nonlinear aeroelastic problems on high performance
computational platforms. The Euler flow equations are solved using the monotonic upwinding
scheme for conservation laws (MUSCL) on unstructured grids. The structural dynamic equilib-
rium equations are solved using the finite element tearing and interconnecting (FETI) method.
By use of the mid-point rule, both the flow continuity equations and the grid conservation law
are enforced. Mesh movement is achieved using the Batina [33,34] method of a linear spring
system.
Rifai et al. [30] employed the commercial code SPECTRUM [2] for the AGARD 445.6 flutter
case, where the finite element treatment of both the fluid and the structure was based on the Hu-
Washizu variational principle. The computational domain was divided into three distinct regions,
i.e., the fluid region, the solid region and an interface between the fluid and solid, thus allowing
variable mesh densities and element topologies in the three regions. The interface was used to
enforce the coupling constraint between the fluid and the solid and the wing was perturbed with a
combination of the first two vibration modes applied as initial velocities.

2. Governing equations

The coupled DFSI problem may be considered as a three field problem, i.e., fluid flow,
structural deformation and the moving mesh [4,36], see Fig. 1. The governing equations for the
three-field problem are:
A.K. Slone et al. / Appl. Math. Modelling 28 (2004) 211–239 215

Fig. 1. Dynamic fluid–structure interaction.

• The Navier–Stokes equation for compressible flow given by:


o
ðqui Þ þ r  quui ¼ rp þ r  lrui þ F f ðtÞ;
ot
and the associated continuity equation given by:
oq
þ r  qu ¼ 0; ð1Þ
ot
where u is velocity, q is density, l is dynamic viscosity, p is pressure, ui is the Cartesian com-
ponent of velocity u corresponding to direction xi and F f ðtÞ is transient load vector defined for
the fluid. For dynamic meshes, the velocity is expressed relative to the movement of the mesh,
u ¼ ufluid  umesh .
• The equation governing the dynamic response of a structure or medium may be written in ma-
trix form [37] as:

M d€ þ C d_ þ Kd ¼ F s ðtÞ; ð2Þ
where d is the displacement vector, M is the mass matrix, C is damping matrix, K is the stiffness
matrix and F s ðtÞ is the transient load vector defined for the solid, where F s ¼ F ext þ F int . In the
context of fluid–structure interaction, the external load vector, F ext is composed of the traction
boundary condition imposed by the fluid and any other external applied loads on the structure.
F int is the internal force and is composed of the internal stresses within the structure.
Damping within the structure is approximated by Rayleigh damping, where the damping
matrix C in Eq. (2) is formed as a linear combination of the stiffness and mass matrices as:
C ¼ ar M þ br K; ð3Þ
where ar and br are the mass and stiffness proportional damping constants.
• Mesh movement may be modelled as a pseudo-structural problem with its own dynamics, with
a spring based mesh movement [34], governed by:
K m d m ¼ fm ðtÞ; ð4Þ
216 A.K. Slone et al. / Appl. Math. Modelling 28 (2004) 211–239

where K m is a pseudo-structural stiffness matrix which is defined for the whole domain and d m
is displacement of the mesh, which in the context of fluid–structure interaction, is the dis-
placement of the structure in the structural sub-domain.
• In the absence of any other boundary conditions, the method for integrating the flow equations
should preserve the solution of the Navier–Stokes and continuity equations. This condition is
satisfied only when the method for solving the flow equations and the algorithm for updating
the displacement and velocity of the mesh obey the geometric conservation law [24,36].

The solution algorithm has been implemented using FV-UM techniques within the PHYSICA
code [8,9] and is shown in Fig. 2, from which it may be seen that differential time stepping between
the fluid and the structure is allowed. This algorithm is described in detail elsewhere together with
its evaluation on a simple geometry cantilever [11–15]. Hence, we simply provide an outline of the
scheme here to inform the discussion below when considering its use in modelling the behaviour of
the AGARD 445.6 wing planform.

3. Navier–Stokes fluid flow solution procedure

The transient fluid flow is governed by the Navier–Stokes momentum and continuity equations.
The method of discretisation for the flow equations ensures that the scheme is fully conservative
for domains of arbitrary shape which move in any time-varying fashion [38]. The resulting
equations are integrated over each control volume such that the solution sets each of the integrals
to zero. By making assumptions concerning the required inter-nodal values, a linear equation is
obtained for each control volume in terms of the unknown scalar quantity at the nodes and for the
nodes on the adjacent control volumes. For fluid flow and heat transfer the standard boundary
conditions allowable in PHYSICA are those for an inlet, fixed pressure boundary, wall and
symmetry plane. Information in terms of external velocity components, pressure or temperature,
is introduced into the equations for those control volumes that have a face coincident with a
domain boundary. The set of linear equations for each control volume are assembled into a global
system matrix for each of the solved fluid variables, i.e temperature and velocity, using standard
computational finite volume, unstructured mesh techniques [39]. Based on the continuity equation
a pressure correction formulation is applied and the resulting discretised equations are solved
using an algorithm based on the semi-implicit method for pressure linked equations (SIMPLE) of
Patankar and Spalding [40].

3.1. Flow equations for a moving mesh

For a moving reference frame, the equations governing fluid flow are expressed in terms of the
fluid variables relative to the mesh movement i.e. u ¼ uf  um where uf is the velocity of the fluid
and um is the velocity of the mesh. Consider the momentum conservation equation for fluid flow
and associated continuity equation. For a moving reference frame the appropriate equations are:
o op
ðqui Þ þ r  ðqui ½uf  um Þ ¼ r  ðlrui Þ  þ S ui ; ð5Þ
ot oxi
A.K. Slone et al. / Appl. Math. Modelling 28 (2004) 211–239 217

Fig. 2. Solution algorithm.

with the associated continuity equation:


oq
þ r  ðq½uf  um Þ ¼ 0; ð6Þ
ot
where l is the dynamic viscosity, q is the density, p is the pressure, u is the resultant face velocity,
ui is the Cartesian component of velocity u corresponding to direction xi and Sui is the source term
218 A.K. Slone et al. / Appl. Math. Modelling 28 (2004) 211–239

for the ith component. At the fluid–solid interface the velocity of the fluid is that of the moving
mesh thus imposing a further boundary condition for the fluid:
ubnd ¼ um on Cs : ð7Þ
In addition, a typical DFSI simulation is also subject to the standard boundary conditions for
fluid flow, i.e., a mass inflow fixed velocity boundary, a fixed value pressure boundary, a no-slip
wall boundary and a symmetry plane, or in other words, an inlet, a free-boundary, symmetry
planes and a wall, see Fig. 4.
By comparison with the standard Navier–Stokes momentum and continuity equations, the
moving reference frame has introduced a further flux term of r  ðqui um Þ to the momentum
conservation equation and of r  ðqum Þ to the continuity equation.
It is essential that the calculation of the mesh velocity obeys the geometric conservation law
[4,41], which may be stated as: the change in volume (area) of each control volume between time tn
and tnþ1 must equal the volume (area) swept by the cell boundary during rt ¼ tnþ1  tn . Thus:
oV
þ r  um ¼ 0; ð8Þ
ot
where V is the volume of the control element. Integrating Eq. (8) over an arbitrary control volume
and applying GaussÕ divergence theorem gives the surface mesh flux as [41]:
I
DV
um  n dC ¼ : ð9Þ
C Dt
Eqs. (5) and (6) are integrated over an arbitrary control volume and are discretised using the
standard control volume unstructured mesh techniques [42] to give:
Z I I Z
oqui
dX þ qui ½uf  um  n dC ¼ lðrui  nÞ dC þ Sui dX; ð10Þ
cv ot Ccv Ccv Xcv

where n is the unit normal to the face of the control volume. Integration of the continuity equation
(6) gives:
Z I
oq
dX þ q½uf  um  n dC ¼ 0; ð11Þ
Xcv ot Ccv

where the face velocities are adjusted for mesh movement using Eq. (9).

4. Dynamic structural deformation solution procedure

The equilibrium equations of structural mechanics may derived from CauchyÕs equation of
motion, which is the governing equation concerning the conservation of momentum for either a
solid or a fluid in motion and are given by:

LT r þ b  qs d€ ¼ 0 in Xs ; ð12Þ
where b is the body force, d€ is the acceleration and qs is the density of the structure. The dis-
placement formulation is derived [43] by use of a finite set of weighting functions Wi to give:
A.K. Slone et al. / Appl. Math. Modelling 28 (2004) 211–239 219
Z Z Z Z
T
 ½LWi ½DLd  D dX þ0
WTi b dX  WTi qs d€ dX þ ½TT Wi T ½DLd  D0 dC
Xs Xs Xs Cd
Z
þ WTi tp dC; ð13Þ
Ct

for i ¼ 1; n where D is the constitutive matrix, L and T are respectively the linear differential and
the outward unit normal operators. r,  respectively, are the stress and strain vectors. Boundary
conditions on the surface Cs of the structural domain Xs are described in terms of prescribed
tractions tp on the boundary Ct and prescribed displacements dp on the boundary of the structure
Cd , where Cs ¼ Cd [ Ct [13]. The essential difference between the finite element method (FEM) and
the finite volume method (FVM) is that for FEM the weighting functions for a node are equal to
the shape functions for that node, i.e., Ni , whereas for cell-vertex FVM the weighting functions
are unity within the control volume and zero elsewhere, i.e. Wi ¼ I within the cv and Wi ; ¼ 0
elsewhere [13].
The standard boundary conditions in the structural algorithm are in terms of a specified dis-
placements, concentrated loads and nodal constraints. At the fluid–structure interface the
structure experiences a surface traction due to the fluid, hence the appropriate boundary condition
for the structure at the fluid–structure interface is expressed as:
 
oui oui 2
tp ¼ pdij þ l þ  lr  udij on Cs : ð14Þ
oxj oxi 3

4.1. Temporal discretisation

The structural algorithm employs a Newmark scheme for temporal discretisation [13,14,44].
The Newmark scheme is one of the most popular algorithms in structural dynamics. It is a
truncated Taylor series collocation algorithm with quadratic expansion, which is the minimum
requirement for second-order problems. The one-dimensional Newmark scheme is given below:
2
Dt
dnþ1 ¼ dn þ Dtd_ n þ ½ð1  2bs Þd€n þ 2bs d€nþ1 ; ð15Þ
2

d_ nþ1 ¼ d_ n þ Dt½ð1  cs Þd€n þ cs d€nþ1 ; ð16Þ


where bn and cn are specific to the Newmark scheme and are chosen to control stability and
accuracy. The scheme is implicit for bn > 0 and is unconditionally stable for 2bn P cn where
cn P 12. For bn ¼ 14 and cn ¼ 12, i.e. a constant-average-acceleration or trapezoidal scheme, it has
zero dissipation for all choices of time step, hence bn ¼ 14 and cn ¼ 12 have been chosen with
consideration of stability, accuracy and dissipation.
Applying linear interpolation to Eq. (2) to advance from time tn to time tnþ1 ¼ tn þ dts , where dts
is the structural time step, gives the dynamic solution at time tnþ1 as:
M d€nþ1 þ C d_ nþ1 þ K Dd ¼ F ext int
nþ1 þ F n ; ð17Þ
where F ext
nþ1 is the external force at tnþ1 which in the context of fluid–structure interaction is
composed of the traction boundary condition imposed by the fluid and any other external applied
loads, F int
n is the internal force at tn and is composed of the internal stresses within the structure.
220 A.K. Slone et al. / Appl. Math. Modelling 28 (2004) 211–239

Substituting the Newmark equations (15) and (16) in Eq. (17) gives:
 
1 c
Kþ M þ C Dd0 ¼ Fnþ1
0
; ð18Þ
b Dt2 b Dt
where

F 0nþ1 ¼ F ext int
nþ1 þ F n þ M þ C ; ð19Þ
   
1 _ 1 1
M ¼ M dn þ  b d€n ; ð20Þ
b Dt b 2
   
c_ c 1 €
C ¼C dn þ  b Dtdn : ð21Þ
b b 2
Solving Eq. (18) for the incremental displacement Dd0 , and rearranging Eq. (15), the nodal
acceleration, velocity and displacement are updated using standard Newmark methods [45],
yielding a first approximation to the acceleration, velocity and displacement for the current
time step, which is then updated until convergence is achieved using the equilibrium iteration
scheme:
 
1 c
Kþ Mþ C Ddi ¼ F inþ1 ; ð22Þ
b Dt2 b Dt
where
F inþ1 ¼ F ext int;i €i _i
nþ1 þ F nþ1 þ Midnþ1 þ C dnþ1 : ð23Þ
The convergence criterion used is based on the displacement Euclidean norm
T
kDdi k ðDdi Þ F inþ1
6 toldis and on the energy norm 6 tolenergy ;
dmax ðDd0 ÞT F 0nþ1

where kDdi k is the Euclidean norm of the incremental displacement for the ith equilibrium iter-
ation; dmax is the maximum total displacement up to and including the current iteration; toldis and
tolenergy are the tolerances set for each case, typically of the order 103 and 102 , respectively.
Divergence is defined in terms of an increasing residual energy norm kF 0nþ1 k 6 kF iþ1
nþ1 k and, if the
solution is divergent or is failing to converge within a specified number of equilibrium iterations,
the stiffness matrix is reformed using the updated geometry and the equilibrium iterations are
continued, thus forming a predictor–corrector variant of the standard Newmark scheme [13].

5. Dynamic mesh movement

Mesh movement is governed by the static displacement equation as given by Eq. (4) and may be
modelled as a pseudo-structural problem with its own dynamics. At each time step the following
static equilibrium equations are solved:
K m nðtÞ ¼ f ðdðtÞÞ; ð24Þ
A.K. Slone et al. / Appl. Math. Modelling 28 (2004) 211–239 221

Fig. 3. Mesh movement regions.

where K m is a pseudo-structural stiffness matrix which is defined for the entire fluid–structure
domain, nðtÞ is the mesh displacement at time t, the function f ðdðtÞÞ is the given displacement
function at time t, which for fluid–structure interaction is governed by the movement of the
structure.
The objective is to move the mesh close to the structure such that the integrity of the mesh is
preserved and to feed this mesh movement out smoothly to the remainder of the domain such that
far from the structure mesh movement is negligible. However, moving the entire mesh for a DFSI
problem would be a significant overhead in terms of run time and in practice it is only necessary to
move the mesh in the region close to the structure. This is achieved by defining four regions of the
pseudo-stiffness matrix K m , i.e. the structure, the boundary region, a flexible region, subject to
displacement but no additional stress, and a fixed region, see Fig. 3.
To ensure that the boundary region moves in the same manner as the structure it is given the
same YoungÕs modulus as the structure. The flexible region is given a YoungÕs modulus that is a
small proportion of the YoungÕs modulus of the structure so as to allow mesh movement such that
the mesh integrity is maintained and the deformation at the interface with the boundary region is
spread smoothly throughout. Typically, the flexible region is given a YoungÕs modulus 1% of the
structure, since this generally allows the mesh to deform without collapsing. The fixed region is
treated as a fixed boundary with zero displacement where d m ¼ d on Xs and d m ¼ 0 on Xfixed .

6. Test cases

The ultimate goal for this research is to investigate the oscillatory characteristics of complex
flexible structures subject to fluid loadings using a finite volume approach, where the onset of
aircraft flutter serves as a good example. A simple representation of an aircraft wing is a fixed-free
cantilever and a standard problem in structural mechanics is the fixed-free cantilever supporting
an applied load at the free end. The structural algorithm used here has been shown to be very
222 A.K. Slone et al. / Appl. Math. Modelling 28 (2004) 211–239

accurate in predicting the fundamental natural frequency of simple structures with respect to both
the amplitude and period of oscillation [14]. This simple problem was then extended to one of a
three-dimensional loaded fixed-free cantilever in fluid flow [13,15] and a sample of the results is
included below.

6.1. Simple geometry configurations

The three-dimensional loaded cantilever in fluid flow is shown in Fig. 4 and the boundary
conditions are a mass inflow fixed velocity boundary at the high y boundary (inlet), a fixed pressure
boundary at the low y boundary (free-boundary), a wall boundary at the low x boundary where
the structure has a zero displacement boundary condition and symmetry planes for the remaining
domain boundaries. The problem was meshed using 76,844 brick elements and the overall
problem dimensions were: depth 20.0 m, breadth 36.0 m and length 51.0 m. The dimensions of the
cantilever were: depth d ¼ 2:0 m, breadth b ¼ 2:0 m and length l ¼ 20:0 m and the cantilever was
subjected to an applied ramped load of 1.05 MN at the free end, which was subsequently re-
moved. The material properties of the cantilever were: YoungÕs modulus 21.0 GPa, PoissonÕs ratio
0.3 and density 2600 kg m3 .
The analytic solution to the structural problem in the absence of fluid is given by standard text
[46,47] as a period of oscillation of 0.4356 s and an amplitude of 0.1 m. To capture the sinusoidal
motion of the cantilever, the time step was taken as 0.022 s, which was expected to give ap-
proximately 20 time steps per cycle. Two Reynolds number cases of Re ¼ 200 and Re ¼ 2:0 105
were studied, where Re ¼ Uinm d and Uin ¼ 100:0 m s1 is the fluid inlet velocity, m is kinematic vis-
cosity. Four cases were studied for m ¼ 1.0 and 1 · 103 m2 s1 and fluid density q ¼ 1:0 and 10.0
kg m3 . The displacement at the tip of the cantilever in the neutral z plane is shown in Fig. 5,
where the effect of the fluid traction load on the cantilever may be seen for the differing Reynolds
number cases. There is little difference in the displacement of the tip for fluid density of 1.0 kg m3
but there is a significant difference in the displacement for the two Reynolds number cases at fluid

Fig. 4. Schematic three-dimensional cantilever in fluid flow.


A.K. Slone et al. / Appl. Math. Modelling 28 (2004) 211–239 223

Fig. 5. Comparison of cantilever tip displacement.

Fig. 6. (a) Re ¼ 200 and (b) Re ¼ 2:0 105 .

density 10.0 kg m3 . The shear xy stress for these two cases may be seen in Fig. 6. The greater
displacement of the tip for Re ¼ 200 is reflected in the increased shear stresses in comparison with
the Re ¼ 2:0 105 case. The flow field at Re ¼ 200 shows little difference for the two density cases.
The flow field for Re ¼ 2 105 in the yz plane at 0.75 cantilever lengths is shown in Figs. 7 and 8.
The above analysis of the simple cantilever illustrates that the FV-UM based DFSI procedure
can reliably represent the oscillatory behaviour of the structure, the flow field and their mutual
interaction.

6.2. Complex geometry case––the AGARD 445.6 wing planform

In real, three-dimensional, engineering applications the structure will have several degrees of
freedom, each of which will be subject to a traction load from the fluid, which will influence the
224 A.K. Slone et al. / Appl. Math. Modelling 28 (2004) 211–239

Fig. 7. Re ¼ 2 105 , velocity vectors, in yz plane at 0.75 cantilever lengths.

Fig. 8. Re ¼ 2 105 , pressure contours in yz plane at 0.75 cantilever lengths.

corresponding mode of vibration. Hence, the structure is unlikely to be in phase with the fluid for
all modes of vibration. It is phase difference, rather than amplitude, which governs the onset of
flutter [48] and, if the phase difference between the fluid force and the structure velocity exceeds
90, flutter will occur. Models based on a finite volume fluid analysis and a separate finite element
structural analysis, may introduce spurious phase mismatching between the fluid and the structure
[49], which may eliminated by using an integrated (finite volume) technique for both domains.
One of the most frequently used benchmarks in aeroelasticity is the AGARD 445.6 wing. This
test case has a panel aspect ratio of 4.0, a quarter-chord sweep of 45 and a taper ratio of 0.6, i.e.,
445.6, and a NACA 65A004 aerofoil section. The semi-span of this model was 2.5 feet and the
root chord was 1.833 feet. In the experimental test case, the wall mounted wing was constructed of
laminated mahogany. In order to obtain flutter data throughout the tunnel Mach number and
density ranges, the stiffness of the wing was reduced by drilling holes, normal to the chord plane,
A.K. Slone et al. / Appl. Math. Modelling 28 (2004) 211–239 225

Fig. 9. AGARD 445.6, weakened model 3.

see Fig. 9, to enable a total of six weakened models to be studied for various panel masses. The
flutter characteristics of this wing planform in air and freon-12 were investigated at the Langley
transonic dynamics tunnel, by Yates et al. in 1963 [1], for Mach numbers in the range 0.338–1.141.
The first four coupled vibrational modes had been previously evaluated by Jones and Unangst [50]
in 1956, where they describe a method for uncoupling these modes. It is unclear in the Yates et al.
paper of 1963 whether the uncoupled vibration frequencies quoted were measured by them or
were based on the results of Jones and Unangst. The distributions of bending and torsional
stiffness, EI and GI were measured and flutter characteristics were calculated using the method of
modified strip analysis.
Some 24 years later, based on his 1963 paper, Yates [51] calculated the flutter characteristics,
using finite element analysis, where the YoungÕs modulus of the aerofoil was calculated so as to
ensure that the calculated vibration modes agreed with those calculated/measured in his earlier
paper. Thus, the wing material properties were: density 4.15 · 102 kg m3 , YoungÕs modulus
3.2456 GPa and PoissonÕs ratio 0.3. In the subsequent flutter calculations, the material properties
for the weakened models were treated as homogeneous and the primary purpose of this work was
to provide aerofoil boundary co-ordinates for subsequent aeroelastic flow calculations.
Typically in aeroelastic simulations, the modal frequencies are taken as known boundary
conditions to the flow domain and Yates findings of 1963 and 1987 have been used for this
purpose to investigate the flutter characteristics of the AGARD 445.6 wing [22,23]. Thus this type
of simulation of the AGARD 445.6 wing flutter characteristics is guaranteed to observe the same
modal frequencies as Yates. But for present day researchers, who are seeking to perform a full
DFSI simulation of the AGARD 445.6 wing flutter characteristics and only have access to data
held in the public domain, the data presented by Yates is a source of uncertainty. The effective
value of YoungÕs modulus is dependent on YatesÕ calculations of 1987, which are themselves
dependent on data collected in 1955. There is uncertainty about the uncoupled modal frequencies,
since it is unclear whether they have been calculated or measured at some previous time, more-
over, the method of drilling holes in the mahogany must have implications for the homogeneity of
the aerofoil.
226 A.K. Slone et al. / Appl. Math. Modelling 28 (2004) 211–239

The work presented here sought only to capture the first mode of vibration and the time step
was calculated with this goal in mind. The first uncoupled bending frequency for the AGARD
445.6 wing, as given by Yates [1,51] for weakened model 3 was 9.6 Hz, hence the time step was set
at 0.005 s, yielding approximately 20 time steps per cycle and this time step was also applied to the
fluid flow calculations.
The test presented is for incompressible flow at Mach number, M1 ¼ 0:338. The fluid material
properties were: density 1.225 kg m3 and kinematic viscosity 1.461 · 103 m s2 and the wing
material properties were as given by Yates [51] for weakened model 3, i.e. YoungÕs modulus
3.2456 GPa and PoissonÕs ratio 0.3. The mesh for this problem was generated using the com-
mercial package GRIDPRO [52] with additional PHYSICA specific software [12,13] and may be
seen in Fig. 10. For the test case reported here, the wing is at a 5 angle of attack, with 25 nodes
from its root to its span-wise boundary and 146 nodes wrapped around the surface. The mesh for
the entire domain had 89,958 hexahedral elements with 95,064 nodes. As such it was near to the
maximum memory (l Gb) of a Dec Alpha 466 MHz processor of 100 K elements for single
processor implementation. The wing mesh had 5088 hexahedral elements and 7100 nodes. The
boundary layer region of the mesh, where the mesh movement is tightly coupled to the movement
of the wing, had 19,806 elements, the moving region of the mesh had 42,186 elements and the fixed
region, where no mesh movement is allowed, had 22,878 elements.
The boundary condition follow those for the schematic three-dimensional loaded cantilever in
fluid flow of Fig. 4, but in a different orientation. There is a mass inflow fixed velocity boundary at
the low x boundary (inlet), a fixed pressure boundary at the high x boundary (free-boundary), a
stationary wall boundary for the entire domain at the low z boundary, where the structure has
fixed zero displacement at its root and symmetry planes for the remaining domain boundaries. At
the start of the full DFSI simulation, the wing planform was subject only to the fixed displacement
condition at its root.
The full DFSI problem was run from the static flow only solution for 30 time steps, making a
total simulation time of 0.155 s and the average run time on a Dec Alpha processor was in excess
for 460 h per cycle of oscillation. The maximum displacements are at the trailing edge of the wing

Fig. 10. Fluid and wing mesh.


A.K. Slone et al. / Appl. Math. Modelling 28 (2004) 211–239 227

tip and are shown in Fig. 11, the shear xy stresses in the wing at 0.105 s are shown in Fig. 12. The
simulation was stopped at 0.55 s as the aerofoil displacements seemed to have reached an equi-
librium position of 0.4791 cm by 0.09 s and the solution was not significantly changing, as may be
seen in Fig. 11a. The pressure and velocity of the fluid at the wing tip at 0.105 s are shown in Figs.
13–15.
It became apparent in the flow simulation that the mesh was not sufficiently fine to capture the
physics of the flow field at the trailing edge of the wing and poor aspect ratios were observed in the
flow domain in the wake region, just downstream from the trailing edge of the wing tip [13]. These
poor aspect ratios give a CFL of approximately 3 · 104 based on the time step for the fundamental
frequency of the wing. A more detailed flow field may have implications for the oscillatory be-
haviour of the wing and when considering the portion of the mesh for the AGARD 445.6 wing in
isolation, it also became apparent that the quality of the mesh, in terms of aspect ratios and
orthogonality, was an issue.

Fig. 11. Mach 0.338, wing tip displacement at trailing edge.

Fig. 12. Shear XY stress, AGARD 445.6 wing.


228 A.K. Slone et al. / Appl. Math. Modelling 28 (2004) 211–239

Fig. 13. Mach 0.338, flow field at tip of AGARD 445.6 wing, 0.105 s.

Fig. 14. Close-up of Fig. 13 at leading edge, 0.105 s.

6.3. Torsionally loaded wing

For the case presented above, the fluid solution was not changing appreciably with time and
was providing an almost constant force to the aerofoil, so that the aerofoil eventually came to an
equilibrium position, as seen in Figs. 11a and b, and the aerofoil is not displaying flutter char-
acteristics.
For comparison purposes, it was desirable to use the data given by Yates [51] but one inter-
pretation of the displacements shown in Fig. 11 is that the time step is too large. The time step had
been set in accordance with the first coupled bending mode of 9.6 Hz, as calculated by Yates [51],
however, the fluid load on the aerofoil is bending and torsional. Thus, it was decided to maintain
the material properties as given by Yates and to investigate the behaviour of the aerofoil in
isolation from the flow, i.e. subjected to a physical load but no fluid flow. The aerofoil was treated
A.K. Slone et al. / Appl. Math. Modelling 28 (2004) 211–239 229

Fig. 15. Close-up of Fig. 13 at trailing edge, 0.105 s.

in isolation as a loaded cantilever subjected to an applied torsional load and the time step was
decreased.
The swept-back aerofoil was subjected to a variety of applied torsional loads at the centre of the
wing tip. After several static runs it was found that an applied load of Fs ¼ ð35i þ 35jÞN resulted
in a static displacement of 1.11 cm, which is close to the displacement of 0.5 in s or 1.27 cm
observed by Yates [1,51]. Several time steps were investigated and it was found that sinusoidal
motion was exhibited for a time step Dt ¼ 7:878 104 s. Lee-Rausch [22,23] employed numerical
damping to aid stability and it was found necessary to employ a stiffness damping parameter of
0.011% to ensure no amplitude growth. Using these parameters the problem was run for 80 time
steps, making a total simulation time of 0.063 s. The maximum amplitude of 1.11 cm is again close
to the displacement of 0.5 in s or 1.27 cm observed by Yates [1,51]. The displacements at the
trailing edge of the tip of the torsionally loaded wing are shown in Fig. 16. The displacements

Fig. 16. Torsional loaded aerofoil, wing tip displacement at trailing edge.
230 A.K. Slone et al. / Appl. Math. Modelling 28 (2004) 211–239

show no amplitude growth and after the first 0.0126 s the frequency of the aerofoil is unaffected by
the initial loading conditions and the period is virtually constant at 0.016 s.
A full DFSI simulation was then performed for the torsionally loaded AGARD 445.6, a ¼ 5:0,
in Mach 0.338 flow with a time step of 7.878 · 104 , for both the fluid and the aerofoil, but the
flow failed to converge. Hence, a further case with a 10:1 differential time stepping between the
fluid and that solid was performed, but again the flow failed to converge and relatively large mass
fluxes due to mesh movement were observed, indicating that some faces in the flow domain were
undergoing significant movement relative to their size. As the mesh moves, the volume of an
individual cell may change appreciably, even though mesh integrity may still be maintained. If the
dimensions in a cell are significantly decreased in the direction of the flow, the CFL condition may
be violated and the flow algorithm may fail to converge.
The default value of the pseudo-YoungÕs modulus for the movable region of the mesh was 1.0%
of the structure and this value was increased to 10% of the aerofoil. This had the effect of ensuring
that close to the aerofoil, where the mesh was fine, the mesh movement was almost a simple
translation, thus reducing the volume change in the movable region, which in turn limited the
mass flux due to mesh movement. To aid the flow solution the time step was decreased to
Dt ¼ 7:878 105 , which gave 200 time steps per cycle for the torsionally loaded wing and a CFL
for the flow of approximately 450. The problem was run as a full DFSI simulation for 0.015756 s,
i.e. 200 time steps, on a 466 MHz Dec Alpha processor where the total run time was 460.58 h. The
torsional load was linearly applied for the first 80 time steps and then released.
The maximum displacements are at the leading edge of the tip of the wing and are shown in Fig.
17, which also shows the maximum displacements for the torsionally loaded wing with no fluid
flow. The fluid is providing positive lift to the trailing edge of the wing tip and has increased the
maximum displacement from the equilibrium position from 1.11 cm at 0.007878 s to 1.27 cm at
0.009296 s. By the end of the simulation, at 0.015756 s, the difference in the displacement between
the two cases is 0.7776 cm. Fig. 17 shows the maximum amplitude at the tip of the wing occurring

Fig. 17. Comparison of displacement for torsional loaded aerofoil.


A.K. Slone et al. / Appl. Math. Modelling 28 (2004) 211–239 231

at 0.0093 s, where it may also be seen that the fluid load has increased the period of vibration to a
projected figure of 0.025 s for the first cycle of oscillation.
The maximum shear xy stresses occur close to the root of the wing near the trailing edge and the
minimum occurs near the leading edge close to the mid-point of the span. Fig. 18 shows the
maximum and minimum shear rxy stresses on the wing at 0.0093 s, when the tip is at its maximum
deflection. These figures also display the relative position of the root and mid-span of the wing.
Again, the fluid is providing positive lift to the wing so that mid-way along its span the wing has
been lifted, but the leading edge displacements are smaller than those at the trailing edge, thus
decreasing the angle of attack. Figs. 19 and 20 show the fluid pressures and velocity at the wing tip
for 0.0093 and 0.0158 s.
From examination of Fig. 17, it is possible that the fluid loaded wing will oscillate around a
non-zero equilibrium position with the fluid eventually damping the motion, but, in view of the
run time of over 460 h per cycle, it was not practical to perform the significantly longer runs
required to establish the full extent of the nature of the oscillatory behaviour.

Fig. 18. AGARD 445.6 wing, shear stress rxz at 0.0093 s.

Fig. 19. Flow field at tip of torsionally loaded wing 0.0093 s.


232 A.K. Slone et al. / Appl. Math. Modelling 28 (2004) 211–239

Fig. 20. Flow field at tip of torsionally loaded wing 0.0158 s.

7. Issues in modelling complex cases

DFSI problems generally require a finer mesh in the structure and surrounding boundary layer
than in the remainder of the mesh which for even moderately complex geometries may lead to a
rapid increase in the size of the problem and run times. There are few, if any, mesh generators
capable of providing a single unstructured mesh suitable for DFSI problems for anything other
than the simplest geometries.
For the dynamic structural mechanics problem, there are a number of analytic benchmark cases
for simple geometries, but to date no such standard benchmarks have been found for DFSI. How-
ever, for this work three main areas of concern in modelling DFSI problems have been identified:

• Mesh quality for complex cases.


• Excessive run times associated with such cases.
• Availability of experimental data.

These issues are addressed below.

7.1. Mesh quality for AGARD 445.6

The two main criteria for assessing the quality of the mesh are the aspect ratio and orthogo-
nality of the mesh. Poor aspect ratios decouple the solution in the direction of the smallest side for
both the fluid and the structure and may mask convergence in that the solution is converged only
in the direction of the largest side. Non-orthogonality in the mesh requires additional correction
terms in the flow equations, increasing the complexity of the calculation for fluid flow.
Skewness of the mesh, in terms of non-orthogonality, has implications for fluid flow. When
integrating the Navier–Stokes equations for fluid flow, the cell-centred finite volume discretisation
requires an estimate of the gradient of the face velocity, ou
on
, where n is the normal vector to the face.
For a structured mesh the line connecting the cell centroids either side of the face is parallel to the
face normal so that:
A.K. Slone et al. / Appl. Math. Modelling 28 (2004) 211–239 233

ou uA  uP
¼ ; ð25Þ
on d
where d is the distance form the cell centroid P to the neighbouring centroid A, sharing the face.
However, if the face is non-orthogonal, the line connecting the nodes P and A will not be parallel
to the face normal, see Fig. 21 so that:
ou uA  uP
¼ ; ð26Þ
ov d
where v is the vector from node P to node A. Hence, the non-orthogonality of the mesh requires an
additional source term, which is a function of the tangential gradient oto , where t is the tangent to the
face. At the start of any simulation the velocity gradients may be artificially high and inclusion of
the orthogonality correction terms would result in reducing the contribution to the system matrix
and introducing large source terms. Convergence may be slow and the solution may even diverge.
Despite much effort in generating the mesh, the maximum angle between the face normal and
the line connecting the adjacent cell centroids for the mesh in the fluid domain was found to be
71.5, which occurs four cells from the tip of the wing in the span-wise direction i.e., in the
movable region, however all the face angles in the same position along the span of the wing are in
excess of 69.0.
The minimum aspect ratio of the wing is 4.19 and is near the root. The maximum aspect ratio of
the wing is 2790 and is close to the tip on the trailing edge. Large aspect ratios will compromise
the accuracy of the solution for any structure and typically finite element analysis codes warn for
aspect ratios in excess of 20. The large aspect ratios in the AGARD 445.6 case were a direct
consequence of the need to avoid skewness in the mesh while preserving the geometry of the
aerofoil and to keep within single processor memory limits. Thus it was felt that the mesh for this
complex case was of as good quality as was possible within the limits of available mesh generation
technology and those imposed by single processor implementation.
Using the frequencies and material data claimed by the experimental data of Yates, the
AGARD 445.6 test case did not exhibit flutter, which may have been influenced by inadequacies
in the data and the resolution of the mesh. The block structured nature of GRIDPRO limited the

Fig. 21. Two-dimensional representation of orthogonality.


234 A.K. Slone et al. / Appl. Math. Modelling 28 (2004) 211–239

way in which the fine grid in the aerofoil could be expanded in the fluid domain, hence, the fine
grid was carried far into the fluid domain. There are two possible ways forward for addressing the
problem of mesh generation for complex geometries:

• Using two meshes i.e. one for the fluid and one for the structure, as used by Farhat et al. [4].
• Employment of high quality mesh generators.

The use of two, discontinuous meshes would eliminate the problem of feeding a detailed mesh
for the structure into the fluid domain and would allow an optimum mesh in each sub-domain.
However, using it would also require special effort in terms of an accurate interpolation between
the two sub-domains and the two-way exchange of information at the fluid-structure interface. To
ensure the accuracy of the particular DFSI boundary conditions, for those cases where the dis-
placement of the structure is small in comparison with the machine accuracy, it would be nec-
essary to execute the entire software in double precision. In addition, the transfer of data from one
sub-domain to the other would require maintaining a data map between the two sub-domains and
would increase the run time and the amount of storage required.
Obviously, the problems of large aspect ratios and non-orthogonality of the mesh may be re-
duced by increasing the size of the mesh, which is possible in the context of parallel implementa-
tion. But even in this case, the block structured mesh generators would carry the fine hexahedral
mesh of the structure into areas where such mesh density is not required. An alternative could have
been to use tetrahedral meshes which do not suffer in the same way and are capable of locally
refining the mesh, but it is frequently argued that tetrahedral meshes do not capture boundary layer
effects as well as hexahedral ones in the neighbourhood of the fluid-structure interface.
However, for wing structures, the flow in the boundary layer is predominately along the surface
of the wing from the leading to the trailing edge, thus the aspect ratios relative to the wing surface
from leading to trailing edge are not of primary importance. Ideally, the mesh normal to the
surface should be sufficiently aligned to capture the complex flow patterns and the faces of the cells
in the flow direction should be normal to the flow. This may be achieved by meshing the boundary
layer using hexahedral elements. Away from the wing, the mesh in the fluid domain may be
rapidly grown, which may be achieved using tetrahedral elements and the mesh in the wing may
be meshed with any combination of hexahedral, tetrahedral or wedge elements, providing the
aspect ratios are less than 20. However, to date there is no knowledge of mesh generators that can
easily transform a hexahedral mesh to a tetrahedral mesh.

7.2. Compute times for complex cases

The AGARD 445.6 wing case had approximately 95 K nodes and the average run time per time
step was approximately 2 h on a 466 MHz Dec Alpha processor. Typically, it is necessary to
simulate 10 complete cycles requiring 200 time steps to capture the cyclical motion of the struc-
ture, thus the total run time for a 100 K node problem would be of the order of 400 h. But as has
been shown above, the mesh for this complex case needed to be refined and if the mesh density
was only doubled in each direction, there would be an approximate 10-fold increase in the size of
the problem i.e. 1 million nodes, which would require about 10 Gb of processor memory. As mesh
density increases, the fluid time step would need to be reduced, in order to comply with the CFL
A.K. Slone et al. / Appl. Math. Modelling 28 (2004) 211–239 235

condition, i.e. the fluid time step would be decreased by a factor of 10, requiring 2000 fluid time
steps for a 10 cycle simulation.
Variable time stepping would have to be employed in order to avoid capturing multiple modes of
vibration, where 10 flow time steps are performed for one structural time step. Such a scheme is
currently allowed, where the traction load on the structure is calculated at each time step and the
average traction load is applied to the structure during the structural time step. Mesh movement is
calculated at the end of the structural time step and the mesh velocity used to modify the flow
equations is applied linearly over the fluid time steps. For such a case, the run time per cycle would
increase by a factor of 10, thus giving a total run time of 40,000 h. Even allowing for a five times
processor speed up, the total run time would still be of the order of 8000 h. The solution to the
memory and run time problems has to involve the implementation DFSI procedure in parallel.

7.3. Availability of experimental data

It is difficult to find reliable benchmarks for three-dimensional DFSI. Even in the case of
aeroelasticity, there are so few benchmarks that the AGARD 445.6 is still widely used for this
purpose. The original wind tunnel tests of the flutter dynamics of the AGARD 445.6 wing were
carried out in 1963 and by the standards of today, the information available in the public domain
is severely limited, e.g there is little information regarding boundary conditions and the effect on
the homogeneity of the wing in respect of its stiffness. This is not an issue for the aeroelastic
simulation of the flutter characteristics of the AGARD 445.6 wing, where the modal frequencies
are assumed and are used as boundary conditions to the flow, but it is of increasing importance as
more comprehensive, full DFSI simulations are carried out.
Material data is of particular concern, since it is evident that drilling holes in the wing and
replacing the mahogany with rigid foam plastic renders the wing a composite material and
compromises the assumption of homogeneity. The YoungÕs modulus and density of the wing are
essential material properties in Eq. (2) and failure to model these correctly will have serious
implications for the structural response. The fact that sinusoidal motion was observed for a re-
duced time step throws further doubt on the values of YoungÕs qffiffi modulus, E, and density, q, for
Yates et al. weakened model 3 data, since frequency f / Eq. The authors are not alone in being
unable to induce flutter for this case using the data reported by Yates, see, for example, Raveh [29]
and Rifai and Smith [53]. In view of the fact that a number of new techniques are emerging for
modelling three-dimensional DFSI, it seems appropriate to consider whether a new set of
benchmarks should be developed.

8. Closure

The simulation of DFSI is a considerable challenge to the computational modelling commu-


nity. The development of robust procedures to enable time and space accurate calculations is a
significant challenge, even for the simplest three-dimensional geometries and where such simu-
lations have been attempted either the physics or the geometries have usually been simplified.
In this work we have considered the application of the essentials of a full physics simulation, i.e.
Navier–Stokes flow, three-dimensional linear elastic structure, in the context of the AGARD
236 A.K. Slone et al. / Appl. Math. Modelling 28 (2004) 211–239

445.6 wing geometry. This research has highlighted three significant issues that need to be ad-
dressed if DFSI analysis is to be applied to ÔflutterÕ analysis of practical aerostructures.
Mesh structure and quality: the constraints of achieving a high quality mesh representation of
the structure, the surrounding boundary layer and the remainder of the flow domain seem to be
mutually incompatible. In this work, we have employed a single hexahedral mesh for the whole
domain, which considerably simplifies the mutual exchange of information at the fluid–structure
interface. Although the use of separate meshes should alleviate the meshing problem, it may not
eliminate the generation of poor quality cells in the flow domain at the edge of the boundary layer,
unless tetrahedral cells are employed outside the boundary layer. Of course if separate meshes are
used, the problem of interpolation between the meshes will have to be overcome.
Compute speeds: when full Navier–Stokes flow and three-dimensional structural calculations
are made, even for the simplest linear elastic solids and incompressible flows without a turbulence
model, the calculation times are essentially impractical for current, conventional, single processor
computers. Meshes with at least an order of magnitude more cells are required to resolve the flow
physics at the trailing edge of the wing. This leads to a time step for the fluid, which if applied to
the structure would result in capturing the higher order structural response modes and the sub-
sequent difficulties of resolving these and isolating any possible numerically spurious oscillations.
Adequate experimental data: the most comprehensive sets of experimental data to characterise
ÔflutterÕ are now well over 30 years old. This data was not collected to validate the kind of so-
phisticated computational DFSI models explored in this work and the information in the public
domain is not sufficiently detailed to enable the ÔflutterÕ characteristics of the flow to be clearly and
unambiguously identified. This work is not alone in highlighting these problems, see for example
the work of Bennett and Edwards [54].
Hence, successful modelling of the flutter characteristics of aerostructures requires:

• The development of a meshing strategy that enables a single mesh throughout the entire fluid–
structure domain, where the boundary layer is meshed using hexahedral element and the re-
mainder of the flow domain is meshed using tetrahedral elements. The structure sub-domain
could be any mix of tetrahedral and hexahedral elements.
• The extension of the model to compressible flows with two equation turbulence models, struc-
tures with non-linear material properties and their implementation in parallel with good scala-
bility up to 100+ processors.
• Sets of three-dimensional ÔflutterÕ data that have sufficient detail and are closed from the per-
spective of computational modelling requirements.

The future efforts of this programme will focus on the first two of the above issues to produce a
DFSI simulation tool that has the essential functionality to address a broad class of problems
involving the onset of flutter.

References

[1] E.C. Yates Jr., N.S. Land, J.T. Foughner Jr., Measured and Calculated Subsonic and Transonic Flutter
Characteristics of a 45 Sweptback Wing Planform in Air and Freon-12 in the Langley Transonic Dynamic Tunnel,
NASA Technical Note D-1616, 1963.
A.K. Slone et al. / Appl. Math. Modelling 28 (2004) 211–239 237

[2] SPECTRUM, Centric Technology, Inc., Sunnyvale, CA. Now owned by ANSYS.
[3] W. Wall, E. Ramm. A finite element based approach for nonlinear, unsteady fluid-structure interaction problems.
in: Proceedings of the MAFELAP Conference on the Mathematics of Finite Elements and Applications, 1999,
p. 130.
[4] C. Farhat, M. Lesoinne, N. Maman, Mixed explicit/implicit time integration of coupled aeroelastic problems:
three field formulation, geometric conservation and distributed solution, Int. J. Numer. Meth. Fluids 21 (1995)
807–835.
[5] C. Farhat, M. Lesoinne, P. Stern, S. Lanteri, High performance solution of three dimensional nonlinear aeroelastic
problems via parallel partitioned algorithms: methodology and preliminary results, Adv. Eng. Softw. 28 (1997) 43–
61.
[6] C.A. Taylor, T.J.R. Hughes, C.K. Zarins, Finite element modeling of blood flow in arteries, Comput. Meth. Appl.
Mech. Eng. 158 (1998) 155–196.
[7] C.A. Taylor, T.J.R. Hughes, C.K. Zarins, Finite element modeling of three-dimensional pulsatile flow in the
abdominal aorta: relevance to atherosclerosis, Ann. Biomed. Eng. 26 (1998) 975–987.
[8] C. Bailey, M. Cross, Discretisation procedures for multi-physics phenomena, J. Comput. Appl. Math. 103 (1999)
3–17.
[9] PHYSICA. Available from: http://www.multi-physics.com.
[10] M. Cross, Computational issues in the modelling of materials-based manufacturing processes, J. Comput.-aided
Mater. Des. 3 (1996) 100–116.
[11] A.K. Slone, K. Pericleous, C. Bailey, and M. Cross, Dynamic fluid-structure interaction using finite volume
unstructured mesh procedures, in: Proceedings of the International Forum on Aeroelasticity and Structural
Dynamics, Vol. 2, Rome, 1997, pp. 417–424.
[12] A.K. Slone, K. Pericleous, C. Bailey, M. Cross, C. Bennett, Dynamic fluid–structure interaction using finite volume
unstructured mesh procedures, AIAA Paper 2000-4788, 2000.
[13] A.K. Slone, A Finite Volume Unstructured Mesh Approach to Dynamic Fluid–Structure Interaction between
Fluids and Linear Elastic Solids. Ph.D. Thesis, The University of Greenwich, Greenwich, 2000.
[14] A.K. Slone, K. Pericleous, C. Bailey, M. Cross, Dynamic solid mechanics using finite volume methods, Appl.
Math. Model. 27 (2003) 67–89.
[15] A.K. Slone, K. Pericleous, C. Bailey, M. Cross, Dynamic fluid structure interaction using finite volume
unstructured mesh procedures, Comput. Struct. 80 (2001) 371–390.
[16] M. Hughes, L. Leboucher, V. Borjarevics, K. Pericleaous, M. Cross, Finite volume simulations in magnetohy-
drodynamics, in: R. Vilseier et al. (Eds.), Finite Volumes for Complex Applications II, Hermes Science, Paris, 1999,
pp. 453–458.
[17] A.J. Williams, T.N. Croft, M. Cross, Computational modelling of metals extrusion and forging processes, in:
Computational Modeling of Materials, Minerals and Metals Processing, The Minerals, Metals and Materials
Society, 2001, pp. 481–490.
[18] J.T. Batina, Unsteady Euler airfoil solutions using unstructured dynamic meshes, AIAA Paper 89-0115, 1989.
[19] J.T. Batina, E.M. Lee, W.L. Kleb, Unstructured-grid methods development for unsteady aerodynamic and
aeroelastic analysis, in: Specialist Meeting on Transonic Unsteady Aerodynamics and Aeroelasticity, number 73 in
Structures and Materials Panel, AGARD, 1991.
[20] O.O. Bendiksen, G. Wang, Nonlinear flutter calculations for transonic wings, Proceedings of the International
Forum on Aeroelasticity and Structural Dynamics, Vol. 2, 1997, pp. 105–114.
[21] K.K. Gupta, Development of a Euler equations aeroelastic analysis capability, J. Aircr. 33 (1996) 995–1002.
[22] E.M. Lee-Rausch, J.T. Batina, Calculation of AGARD wing 445.6 flutter using Navier–Stokes aerodynamics,
AIAA Paper 93-3476, 1993.
[23] E.M. Lee-Rausch, J.T. Batina, Wing flutter computations using an aerodynamic model based on the navier stokes
equations, J. Aircr. 33 (6) (1996) 1139–1147.
[24] M. Lesoinne, C. Farhat, Geometric conservation law for aeroelastic computations using unstructured dynamic
meshes, AIAA Paper 95-1709, 1995, pp. 721–735.
238 A.K. Slone et al. / Appl. Math. Modelling 28 (2004) 211–239

[25] M. Lesoinne, C. Farhat, Geometric conservation laws for flow problems with moving boundaries and deformable
meshes, and their impact on aeroelastic computations, Comput. Meth. Appl. Mech. Eng. 134 (1996) 71–90.
[26] R.B. Melville, S.A. Morton, D.P. Rizzetta, Implementation of a fully-implicit, aeroelastic Navier–Stokes solver,
AIAA Paper 97-2039, 1997.
[27] S.A. Morton, R.B. Melville, M.R. Visbal, Accuracy and coupling issues of the aeroelastic Navier–Stokes solutions
on deforming meshes, AIAA Paper 97-1085, 1997.
[28] R.D. Rausch, J.T. Batina, H.T.Y. Yang, Three dimensional time marching aeroelastic analyses using an
unstructured grid euler method, AIAA J. 31 (1993) 1626–1633.
[29] D.E. Raveh. Reduced order linear models for unsteady aerodynamics, AIAA Paper 2000-4785, 2000.
[30] S.M. Rifai, R.M. Ferencz, W-P. Wang, E.T. Spyropoulos, C. Lawrence, M.E. Melis. The role of multiphysics
simulation in multidisciplinary analysis, AIAA Paper 98-4863, 1998.
[31] J.G. Marshall, I. Imregun, A 3D time domain flutter prediction method for turbomachinery blades, in: Proceedings
of the International Forum on Aeroelasticity and Structural Dynamics, Royal Aeronautical Society, 1995.
[32] K.K. Gupta, L.S. Voelker, C. Bach, T. Doyle, E. Hahn, CFD-based aeroelastic analysis of the X-43 hypersonic
flight vehicle, AIAA Paper 2001-0712, 2001.
[33] B.A. Robinson, J.T. Batina, H.T.Y. Yang, Aeroelastic analysis of wings using euler equations with a deforming
mesh, J. Aircr. 28 (1991) 781–788.
[34] J.T. Batina, Unsteady airfoil solutions using unstructured dynamic meshes, AIAA Paper 0115, 1989.
[35] R. Melville, S. Morton, Fully implicit aeroelasticity on overset grid systems, AIAA Paper 98-0521, 1998.
[36] P.D. Thomas, C.K. Lombard, Geometric conservation law and its application to flow calculations on moving
grids, AIAA 17 (1979) 1030–1037.
[37] O.C. Zienkiewicz, R.L. Taylor, in: The Finite Element Method, Basic Formulation and Linear Problems, Vol. 1,
McGraw-Hill, New York, 1989.
[38] I. Demirdzic, M. Peric, Finite volume method for prediction of fluid flow in arbitrarily shaped domains with
moving boundaries, Int. J. Numer. Meth. Fluids 10 (1981) 771–790.
[39] H.K. Versteeg, W. Malalasekera, Computational Fluid Dynamics, Longman Scientific and Technical, Harlow,
1995.
[40] S.V. Patankar, Numerical Heat Transfer and Fluid Flow, Hemisphere, Dublin, 1980.
[41] I. Demirdzic, M. Peric, Space conservation law in finite volume calculations of fluid flow, Int. J. Numer. Meth.
Fluids 8 (1988) 1037–1050.
[42] N. Croft, K.A. Pericleous, and M. Cross, PHYSICA: a multiphysics environment for complex flow processes,
Numerical Methods in Laminar and Turbulent Flow, Vol. 9, Part 2, Pineridge Press, Swansea, 1995, pp. 1269–
1280.
[43] O.C. Zienkiewicz, R.L. Taylor, in: The Finite Element Method, Solid and Fluid Mechanics, Dynamics and Non-
linearity, Vols. 1 and 2, McGraw-Hill, New York, 1991.
[44] N.M. Newmark, A method of computation for structural dynamics, J. Eng. Mech. Div. 85 (1959) 67–94.
[45] R.D. Cook, D.D. Malkus, M.E. Plessha, Concepts and Applications of Finite Element Analysis, Wiley and Sons,
New York, 1989.
[46] R.W. Clough, J. Penzien, Dynamics of Structures, McGraw-Hill, New York, 1993.
[47] R.T. Fenner, Engineering Elasticity: Applications of Numerical and Analytical Techniques, Ellis Horwood,
Chichester, 1986.
[48] M.B. Giles, Stability, accuracy of numerical boundary conditions in aeroelastic analysis, Technical Report 13,
Oxford University Computing Laboratory, Oxford, 1995.
[49] G.A. Davis, O.O. Bendiksen, Unsteady transonic two dimensional euler solutions using finite elements, AIAA 31
(6) (1993) 1051–1057.
[50] G.W. Jones Jr., J.R. Unangst, Investigation to determine effects of centre of gravity location on transonic flutter
characteristics of a 45 sweptback wing, Technical Report NACA RM L55K30, 1956.
[51] E.G. Yates Jr., AGARD standard configuration for dynamic response, candidate configuration I.-Wing 445.6,
NASA TM 100492, 1987.
A.K. Slone et al. / Appl. Math. Modelling 28 (2004) 211–239 239

[52] GridProe, az3000 User Guide and Reference Manual, Program Development Corporation, White Plains, NY,
1993.
[53] S.M. Rifai, W.D. Smith, Multiphysics simulation of limit cycle aeroelastic oscillations, Technical Report NASA/
CP-1999, NASA, 1999.
[54] R.M. Bennett, J.W. Edwards, An overview of recent developments in computational aeroelasticity, AIAA Paper
98-2421, 1998.

You might also like