46th AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics & Materials Conference AIAA 2005-2295

18 - 21 April 2005, Austin, Texas

POD-based Aeroelastic Analysis of a Complete F-16

Configuration: ROM Adaptation and Demonstration

Thuan Lieu∗
Stanford University, Stanford, CA, 94305-3035, USA

Charbel Farhat†
Stanford University, Stanford, CA, 94305-3035, USA

Michel Lesoinne‡
University of Colorado, Boulder, CO, 80303-0429, USA

he proper orthogonal decomposition method (POD) is applied to the computational

T fluid dynamics (CFD)-based reduced-order aeroelastic modeling of a complete F-16
fighter configuration, in order to assess its potential for the solution of realistic aeroelastic
problems. The limitation of such a computational approach to a fixed free-stream Mach
number is addressed by a Mach-adaptation strategy that interpolates the angle between
two POD subspaces rather than the POD basis vectors directly. The predicted aeroelastic
frequencies and damping ratio coefficients are compared with counterparts obtained from
full-order nonlinear simulations and from flight test data. The results of these comparisons,
including in the transonic regime, reveal a good potential of POD-based reduced-order
modeling for the near real-time prediction of aircraft flutter using CFD technology.

I. Introduction
he ability to accurately predict flutter is essential for developing high-performance, safe, aircraft designs.
T There are a number of computational methods for achieving this objective. The advantages and disad-
vantages of each of them depend primarily on the flight regime of interest. Many computational methods are
based on the linear aeroelastic theory which assumes that the aerodynamic forces can be reliably predicted by
a linear operator. In the subsonic regime, this operator is often computed using the doublet-lattice method,1
while methods derived from the piston theory2 are more suitable in the supersonic regime. In both cases,
such linear methods are attractive because they appear to offer an accurate and yet fast computational mean
for identifying flutter speeds.
However, most modern aircraft, especially high-performance fighters, operate in the transonic regime
where complex nonlinear flow patterns preclude the exclusive use of linear aerodynamic theories for predicting
the unsteady aerodynamic forces. Consequently, scaled wind tunnel testing is often performed to obtain
∗ Post-DoctoralScholar, Institute for Computational and Mathematical Engineering, Building 500, 488 Escondido Mall,
Stanford University, Stanford, CA 94305-3035; AIAA Member
† Professor, Department of Mechanical Engineering, Institute for Computational and Mathematical Engineering, and De-

partment of Aeronautics and Astronautics (by courtesy), Building 500, 488 Escondido Mall, Stanford University, Stanford, CA
94305-3035; AIAA Fellow
‡ Assistant Professor, Department of Aerospace Engineering Sciences and Center for Aerospace Structures, University of

Colorado, Boulder, CO 80309-0429; AIAA Member

1 of 16

American Institute of Aeronautics and Astronautics

Copyright © 2005 by the American Institute of Aeronautics and Astronautics, Inc. All rights reserved.
corrections to the flutter speeds predicted by linear theories. However, the design of scaled wind tunnel
models and the subsequent data analysis typically require more than one year of time.3 State-of-the-art,
computational fluid dynamics (CFD)-based, nonlinear aeroelastic simulation capabilities have shown that for
low to moderate angles of attack, they can be a reliable alternative to scaled wind tunnel testing provided that
adequate computing resources are made available. For example, using a 128-processor computing system, the
AERO code was reported to accurately predict the aeroelastic parameters of a complete F-16 configuration
at five different Mach numbers in the transonic regime, and in less than one day.4, 5 More practically, an
engineer with access to a six-processor computing platform can meet the same objective using AERO in
less than three days. Hence, nonlinear simulation technologies such as AERO and others6, 7 seem to offer a
partial viable alternative to scaled wind tunnel testing for flutter prediction in the transonic regime.
The major computational cost incurred by CFD-based nonlinear aeroelastic simulations is attributable
to the need for high-fidelity fluid models in order to resolve the complex flow patterns present in the tran-
sonic regime. Because of this computational cost, the potential of CFD-based nonlinear aeroelastic codes is
currently limited to the analysis of a few, carefully chosen configurations rather than routine analyses. In
some instances such as those discussed in this paper, it is possible however to address this limitation with
the use of reduced-order models (ROMs). For example, it was recently shown8, 9, 10, 11, 12, 13, 14, 15, 16, 17 that
ROMs constructed by a variety of methods, including the popular proper orthogonal decomposition method
(POD),18, 19 can produce numerical results that compare well with those generated by full-order nonlinear
models. However, since its first application to aeroelastic problems,8 the POD method was primarily applied
to simple airfoils,9, 10, 11, 12 panels,20 wings,13, 14, 15 turbine blades,16, 17 but not to complete aircraft config-
urations. Furthermore, a ROM constructed by POD or any other similar technique is usually not robust
with respect to change in a model parameter.10, 21 Some progress in this area has been recently reported for
the case of structural parameters;14 however, little has been reported for changes in the free-stream Mach
number.
Hence, the objectives of this paper are two fold: (a) to demonstrate the potential of a POD-based ROM
methodology for the CFD-based aeroelastic analysis of a complete fighter configuration, including in the
transonic regime, and (b) to present an algorithm for rapidly adapting aeroelastic ROMs to different free-
stream Mach numbers. To this effect, the remainder of this paper is organized as follows. In Section II, the
computational framework adopted for constructing a POD-based ROM is presented. In Section III, a novel
algorithm for adapting two given ROMs to a varying free-stream Mach number is described. In Section IV,
both computational technologies are applied to the aeroelastic analysis of a complete F-16 configuration at
a low angle of attack, and the aeroelastic parameters obtained using the ROMs are compared with those
obtained using full-order nonlinear simulations and flight test data. Finally, conclusions pertaining to the
performance characteristics, merits, limitations, and potential of the described computational methodologies
are formulated in Section V.

II. Computational Framework

order to construct a POD subspace, an ensemble of sample data representative of some physical system,
Itheninfrequency
this case a flexible aircraft, is required. This sample data is generated here by numerical simulation in
domain of the linearized flows associated with a set of excitations of the displacement field of
the structural model of the aircraft. To this effect, the transient aeroelastic problem is first formulated as a
coupled, nonlinear, fluid-structure interaction problem. However, in order to facilitate the application of the
POD method, this nonlinear formulation is then linearized and simplified. Although the main objective of
aeroelastic reduced-order modeling is to reduce the size of the discrete aeroelastic system, model reduction
is performed here separately for the individual fluid and structural subsystems. This decoupled approach
has several advantages. First, it allows using the most suitable ROM approach for each separate subsystem.
For example, a modal-based ROM is as effective for a structural subsystem as a POD-based one and simpler
to construct. Furthermore, this approach allows formulating the fluid subsystem in adimensionalized form
so that the resulting POD basis is independent of the free-stream pressure and density, thus making the

2 of 16

American Institute of Aeronautics and Astronautics

resulting ROM applicable over a wide range of flight conditions. This robustness with respect to changes in
the free-stream density and/or pressure would be more difficult to achieve if reduced-order modeling were to
be applied to the coupled aeroelastic system in a monolithic fashion, because the state of the coupled system
depends on these free-stream parameters.
Throughout the remainder of this paper, the flow is assumed to be inviscid.

A. Governing Equations
A nonlinear aeroelastic system can be represented by the three-field arbitrary Lagrangian-Eulerian (ALE)
formulation.22 After semi-discretization by a finite element (FE) or finite volume method, this formulation
gives rise to three coupled ordinary differential equations. This coupled system of equations can be written
as

A(x)w ,t
+ F(w, x, ẋ) = 0 (1)
Mü + f int (u, u̇) = f ext (w, u) (2)
K̃x = K̃c u (3)
Here, equation (1) represents a finite volume discretization of the ALE conservation form of the fluid
equations: A denotes the diagonal matrix of cell volumes, F the nonlinear numerical flux function, and w
the conservative state vector of the fluid subsystem. Equation (2) is a FE discretization of the structural
equations of dynamic equilibrium: M denotes the FE mass matrix, f int the vector of internal forces, f ext the
vector of external forces applied on the structure, and u the structural displacement vector. Both of the , t
and dot notations represent a derivative with respect to time. The various Dirichlet and Neumann boundary
conditions associated with the fluid and structural subproblems are embedded in the above system and, for
simplicity, are not explicitly stated. In many aeroelastic applications, the boundary of the fluid domain is
required to deform according to the motion of the wet surface of the structure. This is represented here by
equation (3) which models the fluid mesh as a pseudo-structure with a piece-wise static behavior: x denotes
the fluid mesh motion, K̃ a fictitious stiffness matrix, and K̃c a transfer matrix describing the effect of
structural motions on the fluid mesh at the fluid-structure interface.23
For the purpose of constructing a ROM, the above formulation of an aeroelastic problem is simplified as
follows.
First, both the fluid and structural equations are linearized around an equilibrium point designated by
the subscript o and satisfying ẇo = ẋo = 0. As outlined at the beginning of this section, the fluid equation is
also adimensionalized in order to eliminate from the sought-after ROM any dependence on the dimensional
free-stream pressure and density, so that the computation for a given angle of attack of any “base” nonlinear
steady state flow solution depends only on the free-stream Mach number, M∞ .
The linearization of the fluid subsystem adopted here follows the development described first in refer-
ence,24, 25 then enhanced for adimensionalization in reference.21 This linearization transforms equation (1)
into
Āw̄,τ + H̄w̄ + (Ē + C̄)x̄˙ + Ḡx̄ = 0 (4)
where
∂ F̄
H̄ = |o
∂ w̄
∂ Ā
Ē = |o w̄o
∂ x̄
∂ F̄
C̄ = |o
∂ x̄˙
∂ F̄
Ḡ = |o
∂ x̄

3 of 16

American Institute of Aeronautics and Astronautics

To keep the notation as compact as possible, the same variables w and x are used to denote in the above
equation the perturbations of the fluid state and mesh motion vectors, respectively, around the chosen
equilibrium point. The matrices H̄, Ḡ, Ē, and C̄ are the first-order terms of a Taylor expansion of the
adimensionalized numerical flux function around the adimensionalized operating point (w̄o , x̄o , x̄˙ o ). The
bar notation indicates that a quantity is an adimensionalized one. The matrix H̄ is the gradient of the
adimensionalized numerical flux function with respect to the adimensionalized fluid state vector and thus has
in general a rank equal to the number of fluid degrees of freedom (dof). The coupling matrices Ḡ and C̄ are
gradients of the adimensionalized flux function with respect to the adimensionalized fluid mesh motion. The
coupling matrix Ē is the gradient of the adimensionalized cell volumes with respect to the adimensionalized
fluid mesh motion. Further details regarding the adimensionalization of the above linearization are given in
appendix A.
The linearization of the structural subsystem represented by equation (2) around an equilibrium state is
straightforward and leads to
∂f ext
Mü + Ko u = Pw̄ + |o u (5)
∂u
where
∂f ext
P= |o (6)
∂ w̄
Again, for the sake of notational simplicity, the same notation u is used in the above equation to denote the
perturbation of the structural displacement around its equilibrium point uo . The matrix Ko denotes the
structural stiffness matrix at the operating point uo . The external aerodynamic force applied to the structure
is dependent on the fluid state as well as the attitude of the structure. The motion at the fluid-structure
interface will determine the direction of the surface normals which will alter the direction of the applied
force. Thus, P arises from the linearization of the external loading with respect to the adimensionalized
ext
fluid state vector and for simplicity, the additional linearization term ∂f∂u |o is ignored in the remainder of
this paper. Provided that this term is negligible, it permits the expression of the structural system in terms
of its dry natural modes. The fluid mesh position and velocity variables x̄ and x̄˙ are eliminated from the
coupled system of linearized equations by introducing

K̄ = K̃−1 K̃c (7)

so that

x̄ = K̄ū (8)
x̄˙ = K̄ū˙ (9)

where ū is the adimensionalized displacement of the structure. The above algebraic manipulations allow
re-writing equation (4) as
Āw̄,τ + H̄w̄ + (Ē + C̄)K̄ū˙ + ḠK̄ū = 0 (10)
Next, equation (5) is projected on a basis of dry, natural, structural modes and therefore is transformed
into
Iü + Ω2 u = Pw̄ (11)
where Ω is the diagonal matrix of the squared natural pulsations of the structure, and I is the identity
matrix. Again, for the sake of notational compactness, the same notation u and P is used in the above
equation to denote the generalized displacement (modal) coordinates of the structure and the “modalized”
external force matrix, respectively. The use of a modal basis to represent the structure accomplishes two
important goals. First, it reduces the number of degrees of freedom for the structure, thus contributing to
a more compact aeroelastic ROM. Second, it reduces the size of the coupling matrices in the linearized fluid
equation (10).

4 of 16

American Institute of Aeronautics and Astronautics

 Equation (10) and equation (11) define the linearized aeroelastic formulation adopted in this paper.
Note that the linearized fluid subsystem (10) provides a good approximation of the nonlinear formulation
only if the aeroelastic system is not perturbed far from the linearization point. The computational expense
associated with the solution of the linearized coupled system of equations (10,11) can be expected to be
less than that associated with the solution of its nonlinear counterpart; however, this expense can still be
considerable because of the size of the discrete fluid subsystem. The aim of a POD-based ROM is therefore
to address the latter issue.

B. Construction of a POD-based ROM

POD is a method that provides a basis for representing a given data set from which a lower-dimensional
subspace can be identified. When the given data set is, in some way, representative of a physical system,
the resulting reduced basis can be deemed a low-order model of the original full-order model representing
that system. The theory and application of POD is covered in many publications.26, 27, 28, 14 For the sake
of brevity, readers are referred to these references. To keep however this paper as self-contained as possible,
the POD procedure used within the aeroelastic computational framework described in the previous section
is summarized below. This procedure is similar in its spirit to the approaches exposed in earlier works.11, 12

1. Generate complex-valued snapshot solutions of equation (10) in the frequency domain for a varying
reduced frequency k
w̄j (k) = (ik Ā + H̄)−1 ik(Ē + C̄) + Ḡ K̄ūj


(12)
Equation (12) above is obtained by assuming a periodic solution of the form w̄ = w̄j eikτ and a periodic
√
excitation of the form ū = ūj eikτ , where ūj is a prescribed structural displacement field and i = −1
is the imaginary number. For each specified value of ūj , a sweep is performed on the reduced frequency
k and several snapshots w̄j (k) are generated. Typically, ūj is chosen as a dry natural mode of the
structure and therefore the total number of generated snapshots is equal to the product of the number
of excitation modes and the number of considered reduced frequencies.
2. Form the real-valued correlation matrix
R = SST (13)
where the superscript T designates the transpose,
h i
S = Re(W̄) Im(W̄) (14)

and each column of W̄ contains a complex-valued snapshot of the form given in (12).
3. Compute the eigenvalues and eigenvectors of the correlation matrix R. However, since all non-zero
eigenvalues of the matrix ST S are also eigenvalues of the correlation matrix R, and the size of ST S
is significantly smaller than that of R, it is more attractive to replace this step by the solution of the
alternative generalized eigenvalue problem ST SΨ = ΨΛ. Then, form the following POD basis
1
Φ = SΨΛ− 2 (15)

which satisfies
ΦT Φ = I (16)

4. Form a truncated POD-basis, Φr , by reducing the size of the matrix Φ to a few r columns. Usually,
the magnitude of the eigenvalue value associated with a column of Ψ (and therefore Φ) is used as a
criterion for deciding which POD vector to retain and which to discard.

5 of 16

American Institute of Aeronautics and Astronautics

 5. Project the snapshots on the truncated POD basis

w̄ ≈ Φr w̄r ← w̄r = ΦTr w̄ (17)

6. Project the governing fluid equation (10) onto the POD basis. This step leads to

(w̄r ),τ = −ΦTr Ā−1 H̄Φr w̄r − ΦTr Ā−1 (Ē + C̄)K̄ū˙ + ḠK̄ū


(18)

which is re-written here as

(w̄r ),τ = H̄r w̄ − B̄r ȳr (19)
where

= −ΦTr Ā−1 H̄ Φr


H̄r
h i
B̄r = ΦTr Ā−1 (Ē + C̄)K̄ ḠK̄
" #
ū˙
ȳr =
ū

Depending on the size of the truncated POD basis defined in step 4, the adimensionalized reduced-order
fluid state vector, w̄r , the fluid subsystem matrix H̄r , and the coupling matrix B̄r can be significantly
smaller than their full-order counterparts.

In summary, the POD process outlined above leads to a reduced basis that can be used for constructing
a fluid ROM for a specified free-stream Mach number. The corresponding aeroelastic ROM is obtained by
coupling equation (19) with equation (11). The fluid ROM, and therefore the aeroelastic ROM, may be
used for computing flows at the specified free-stream Mach number, but for variable free-stream pressure
and density and therefore variable altitude.

III. Adaptation of the Aeroelastic ROM

he main drawback of a POD-based ROM, and for that matter any similarly constructed ROM, is the
T lack of robustness over an entire parameter space. Essentially, this is because the data samples are
collected only within a small region of the state space. While this focused data sampling leads to a very
accurate ROM, it does not lead to a reduced-order basis that can accurately capture the solution space for
a range of the parameter space. In particular, since the steady-state, adimensionalized fluid solution is a
function of the free-stream Mach number, the POD basis is sensitive to changes in the free-stream Mach
number. Thus, a ROM generated by the procedure outlined above cannot be expected to approximate well
the fluid subsystem for Mach numbers away from the operating (equilibrium) point.
Previous approaches for adapting a POD basis to address a parameter variation include the global
POD method29, 30, 31 and the direct interpolation of the basis vectors.15 The former approach has only
been demonstrated at low free-stream Mach numbers outside the transonic regime. The latter approach was
shown to deliver poor results.15, 21 The adaptation strategy proposed next is named here the “subspace angle
interpolation” method.15 This method relies on the concepts of principal angles between two subspaces and
principal vectors for a pair of subspaces.32
Definition 1. Let M denote a subspace of dimension p equipped with the usual Hermitian norm, and N
another subspace of dimension q with q ≤ p. The smallest angle, θ1 ∈ [0, π/2], between M and N is defined
by
cos θ1 = max max u∗ v (20)
u∈M v∈N
kuk2 =1 kvk2 =1

6 of 16

American Institute of Aeronautics and Astronautics

where u is not to be confused with its previous meaning, and the * operation denotes the usual Hermitian
inner product. Denote by the subscript 1 the pair of vectors u and v defining θ1 . Each subsequent angle,
θk ∈ [0, π/2], k = 2, · · · , q, is defined recursively as the smallest angle between the orthogonal complement
in M to {u1 , · · · , uk−1 }, denoted here by M⊥ k−1 , and the orthogonal complement in N to {v1 , · · · , vk−1 },
⊥
denoted here by Nk−1 . This can be written as

cos θk = max max u∗ v = u∗k vk (21)

u∈M⊥ ⊥
k−1 v∈Nk−1
kuk2 =1 kvk2 =1

Definition 2. The principal vectors of the pair of subspaces (M, N ) are the two vector sets U =
{u1 , · · · , uq } and V = {v1 , · · · , vq }, where uk and vk are the vectors introduced in Definition 1. U forms a
unitary basis of M, and V may be complemented by (p − q) unitary vectors to obtain a unitary basis of N .
The principal angles, θk , can be interpreted as the set of angles providing a series of rotations that
transform one of the two considered subspaces into the other one. These angles can be computed by the
following algorithm.32 Let Φ1 and Φ2 denote two matrices storing the POD basis vectors constructed for two
different values of a parameter such as the free-stream Mach number, and let Y, Σ, and Z be the matrices
obtained by the following Singular Value Decomposition (SVD)

ΦT1 Φ2 = YΣZ (22)

The cosines of the principal angles, θk , are given by the singular values stored in the diagonal matrix Σ, and
the principal vectors are given by

U = YΦ1 V = ZΦ2 (23)

If for the sake of simplicity Φ1 and Φ2 have the same size, r, the SVD described above yields r principal
angles and r pairs of principal vectors. As stated above, each principal angle represents the rotation through
which a basis vector of one subspace can be transformed into a basis vector of the second subspace.
Suppose that the two bases Φ1 and Φ2 correspond to two sets of POD vectors constructed for two
different values of the free-stream Mach number, M∞1 and M∞2 . To construct a fluid POD basis — and
then an aeroelastic ROM — for an arbitrary free-stream Mach number M∞1 < M∞ < M∞2 , it is proposed
to first linearly interpolate the principal angles, θk , between Φ1 (M∞1 ) and Φ(M∞ )

M∞ − M∞1

θk (M∞1 , M∞ ) = θk (M∞1 , M∞2 ) M∞ ∈ [M∞1 , M∞2 ] (24)
M∞2 − M∞1

Then, each principal vector, uk ∈ U(M∞1 ), is rotated towards

its corresponding principal vector, vk ∈ V(M∞2 ), through the in-
terpolated rotation angle, θk (M∞1 , M∞ ), to obtain the interpolated
basis vector
qk (M∞ ) = uk cos θk (M∞1 , M∞ )
vk − (uTk vk )uk (25)
+ sin θk (M∞1 , M∞ )
kvk − (uTk vk )uk k2

The interpolated vectors {qk } form the new interpolated POD

basis at the intermediate free-stream Mach number. The method
outlined above for adapting a pair of POD bases to a variable free-
stream Mach number is graphically depicted in Fig. 1.

Figure 1. Rotation of principal vector,

u, through principal angle, θ, towards
principal vector, v

7 of 16

American Institute of Aeronautics and Astronautics

 IV. Aeroelastic Parametric Identification of a Complete F-16 Configuration
n this section, the two ROM methodologies described in the previous sections are applied to the aeroelastic
Iauthors,
parametric identification of a complete F-16 configuration with clean wings. To the best knowledge of the
this is the first time the potential of an aeroelastic ROM methodology is assessed for such a realistic
aeronautical engineering problem. The aeroelastic full-order computational model chosen here consists of a
168,799-dof FE structural model built with bar, beam, solid, plate, shell, metallic, and composite elements,
and a CFD unstructured (Euler) grid with 403,919 vertices and more than 2 million dofs (Fig. 2). This
full-order model is used, together with the AERO numerical simulator,4, 5 to generate the various snapshots
needed for constructing the aeroelastic ROMs, and to compute full-order linearized as well as nonlinear
aeroelastic reference solutions. In all cases, the angle of attack is set to α = 0◦ , nine low-frequency dry
structural modes are used to construct the aeroelastic ROMs, H̄ is evaluated by a semi-analytical method,25
and Ē, C̄, and Ḡ are computed by a finite difference scheme.
All computations are performed on a Linux cluster with 15 Pentium 4, 4.0 GHz processors.

(a) Detailed FE structural model (b) Fluid surface grid

Figure 2. F-16 aeroelastic model

A. Numerical Results and Flight Test Data

First, the free-stream Mach number is set to M∞ = 0.9 and several snapshots are generated for 0 ≤ k ≤ 0.02
(see section B). An aeroelastic ROM is then constructed using 60 POD vectors. The performance of this
ROM is assessed by applying it to the prediction of the transient aeroelastic response of the F-16 to an initial
displacement excitation proportional to the first dry mode of this aircraft. Fig. 3 reports for various values of
the free-stream dynamic pressure the lift time-histories predicted using this ROM, and an aeroelastic model
in which the full-order linearized fluid is coupled to the modalized structure. The reader can observe that in
all cases, the ROM performs well. Fig. 4 focuses on the free-stream dynamic pressure q∞ = 825 lb/f t2 . Fig.
4.a compares the lift time-histories predicted by the ROM with those predicted by the full-order nonlinear
aeroelastic model. These time-histories exhibit the same frequency but variations in the amplitude. Since
the results reported in Fig. 3 suggest that the ROM reproduces faithfully the response predicted by the
coupled full-order linearized fluid and modalized structure system, the amplitude errors reported in Fig. 4.a
can be attributed to either the linearization of the fluid or the representation of the structure by the nine
selected modes. However, Fig. 4.b shows that the ROM and computational model built by coupling the

8 of 16

American Institute of Aeronautics and Astronautics

full-order nonlinear fluid model and the modalized structure model predict the same lift time-histories. It is
therefore concluded that the errors observable in Fig. 4.a are attributable to a poor modal representation of
the structure. Further studies have revealed that these errors are induced by inaccuracies in the computation
of the first two dry bending modes of the F-16.

(a) q∞ = 825 lb/f t2 , h = 10, 000 (b) q∞ = 858 lb/f t2 , h = 9, 000 f t (c) q∞ = 891 lb/f t2 , h = 8, 000 f t
ft

Figure 3. F-16 lift time-histories generated for M∞ = 0.9 and various free-stream dynamic pressures, q∞ , using
the aeroelastic ROM and the full-order linearized fluid model coupled to the modalized structural model

(a) Comparison of aeroelastic ROM and the full-order (b) Comparison of aeroelastic ROM and the computa-
nonlinear aeroelastic model tional model coupling the full-order nonlinear fluid and
the modalized structure

Figure 4. F16 lift time-histories generated for M∞ = 0.900 and q∞ = 825 lb/f t2

9 of 16

American Institute of Aeronautics and Astronautics

 Fig. 5 reports the numerically pre-
dicted variations with the free-stream
Mach number of the aeroelastic damp-
ing ratio coefficient associated with the
first wet torsional mode, and contrasts
them with results obtained from flight
test data. It shows that the numeri-
cal results generated by the aeroelastic
ROM are in reasonable agreement with
those predicted by the full-order nonlin-
ear model and those obtained from flight
test data.
Next, two POD bases are pre-
computed for M∞ = 0.611 and M∞ =
0.799, and a third one is constructed
for M∞ = 0.710 by adapting the pre- Figure 5. Variation with the free-stream Mach number of the
computed ROMs using the subspace an- damping ratio coefficient associated with the first torsional mode:
gle interpolation method. The perfor- comparison of the results generated by the ROM, full-order non-
linear simulations, with the counterparts obtained from flight test
mance of this adaptation is confirmed data
by Fig. 6 which compares the lift time-
histories produced by the Mach-adapted
ROM with those generated by the coupled full-order linearized fluid (at M∞ = 0.710) and modalized struc-
ture system.
Finally, the same two pre-computed ROMs mentioned above are continuously adapted for 0.611 ≤ M∞ ≤
0.799 to predict the variation with the free-stream Mach number of the damping ratio coefficient of the first
aeroelastic torsional mode. Because flight test data is not available in this window of the free-stream Mach
number, these predictions are compared in Fig. 7 with their counterparts obtained from full-order nonlinear
aeroelastic simulations. Again, the reader can observe that the ROM Mach-adaptation performs well in this
case too.

B. Computational Cost
A summary of the computational cost as-
sociated with constructing a fluid ROM POD (Frequency Domain)
at M∞ = 0.9 for the F-16 aircraft consid-
Snapshot interval ∆k = 0.004
ered in this paper is given in Table 1. As
it can be expected, the CPU time asso- Total number of snapshots 99
ciated with generating the snapshots ac- CPU time: snapshot generation 3.651 hrs
counts for the vast majority of this com- CPU time: POD basis construction 631 secs
putational cost. CPU time: total 3.845 hrs
From the performance results recently
reported for the CFD-based nonlinear Table 1. Construction of a POD basis at M∞ = 0.9 for a complete
aeroelastic analysis of the same F-16 con- F-16 configuration: performance results on 15 3.0 GHz Pentium 4
figuration,5 one can estimate that the processors
parametric identification of this aircraft
using the full-order nonlinear aeroelastic computational model described at the beginning of this section
requires about five CPU hours per free-stream Mach number on the same 15-processor system. From this
observation, the already emphasized limitation of a fluid ROM to one free-stream Mach number, and the
performance results reported in Table 1, it follows that as far as CFD-based aeroelastic analysis is concerned,
constructing a ROM is a worthwhile effort only if a sweep on altitude is anticipated. Such a sweep — or a

10 of 16

American Institute of Aeronautics and Astronautics

Figure 6. Comparison of F-16 lift time-histories generated by adapting ROMs for M∞ = 0.611 and M∞ = 0.799
to M∞ = 0.710, with those generated by coupling a full-order linearized fluid at M∞ = 0.710 to a modalized
structure

Figure 7. Comparison of F-16 first-torsional aeroelastic damping ratio coefficients generated by ROMs con-
tinuously adapted from M∞ = 0.611 and M∞ = 0.799 to 0.611 ≤ M∞ ≤ 0.799, with counterparts generated from
full-order nonlinear simulations

11 of 16

American Institute of Aeronautics and Astronautics

sweep on the free-stream density or pressure — is usually part of most numerical procedures for extracting
flutter speeds. However, the subspace angle interpolation method highlighted in this paper boosts the poten-
tial of the ROM technology for speeding up further the CFD-based flutter as well as more general aeroelastic
analysis of an aircraft. Indeed, once two POD bases associated with two different free-stream Mach numbers
M∞1 and M∞2 have been pre-computed, this interpolation method constructs a POD basis at any specified
Mach number M∞1 < M∞ < M∞2 without generating any additional snapshots — that is, very rapidly.

V. Conclusions
In recent years, significant progress has been made in advancing the state-of-the-art of computational fluid
dynamics (CFD)-based aeroelastic reduced-order modeling (ROM), to the point where the introduction of
this computational technology in a design environment is often discussed in the literature. To this effect, this
paper contributes the first demonstration and discussion of the application of a CFD-based aeroelastic ROM
methodology to a complete aircraft configuration — namely, an F-16 fighter with clean wings at zero angle of
attack — its suitability for the transonic regime, and its potential for speeding up CFD-based flutter analysis.
It also addresses the limitation of a fluid ROM to a specified free-stream Mach number by contributing a fast
Mach-adaptation algorithm. Given two ROMs generated for two free-stream Mach numbers M∞1 and M∞2 ,
this algorithm efficiently constructs a ROM for any M∞1 ≤ M∞ ≤ M∞2 , without computing any additional
snapshots. The aeroelastic identification of the F-16 predicted by the considered ROM methodology is
found to be in excellent agreement with that predicted by full-order nonlinear aeroelastic simulations and in
reasonable agreement with flight test data, including in the transonic regime. It is also found that the basic
ROM technology described in this paper has the potential for speeding up linearized flutter analysis by a
factor five. The Mach-adaptation algorithm is also found to perform well. It produces accurate results and
offers hope for near real-time CFD-based flutter analysis of complete aircraft configurations.

Acknowledgements
The authors acknowledge the support by the Air Force Office of Scientific Research under Grant F49620-
01-1-0129. They also thank ICEM CFD Engineering Inc. for providing its mesh generation software.

APPENDIX

A. Linearization and Adimensionalization of the ALE form of the Euler

Equations
The vector of conservative fluid variables w is given by
 
ρ
 ρu 
 1
w =  ρu2  (26)
 
 
 ρu3 
ρeT

where ρ is the density, u is the fluid velocity vector whose components in a Cartesian coordinate system
are denoted by the appropriate subscripts, and eT is the total internal energy of the fluid. If ρ∞ and U∞
denote the free-stream density and velocity, respectively, and Lo denotes a characteristic length scale, w can

12 of 16

American Institute of Aeronautics and Astronautics

be adimensionalized as follows
   ρ 
ρ̄ ρ∞
 ρ̄ū   ρu1 
 1  ρ∞ U∞ 
  ρu2  −1
w̄ =  ρ̄ū2  =  ρ∞ U∞  = D w (27)
 
 
 ρ̄ū3   ρu3 
  
ρ∞ U∞
ρ̄ēT ρeT
ρ U2
∞ ∞

where  
ρ∞ 0 0 0 0
 0 ρ∞ U∞ 0 0 0 
 
D≡ 0 0 ρ∞ U∞ 0 0 
 
 
 0 0 0 ρ∞ U∞ 0 
2
0 0 0 0 ρ∞ U∞
and a bar indicates an adimensionalized quantity. From equation (27), it follows that
∂F ∂F −1
= D (28)
∂w ∂ w̄
The three dimensional flux function, F, can be written as
 
ρuj
ρu u + pδ 
 j 1 j1 
Fj = ρuj u2 + pδj2  , j = 1, 2, 3 (29)
 
 
ρuj u3 + pδj3 
ρ(eT + p)uj
where p is the pressure and δij is the Kronecker delta. It can be adimensionalized as follows
1
F̄ = D−1 F (30)
U∞ L2o
Starting from equation (1) and linearizing all terms around (wo , xo , ẋo ) (with ẇo = ẋo = 0) leads to
 
∂A ∂F ∂F ∂F
Aẇ + |o wo ẋ + |o w + |o x + |o ẋ = 0 (31)
∂x ∂w ∂x ∂ ẋ
where the same variables w, x, and ẋ are re-used here to denote the perturbations around wo , xo , and ẋo ,
respectively. Substituting equation (27), equation (28), and equation (30) into equation (31) yields
   
∂A ∂ F̄ ∂ F̄ ∂ F̄
ADw̄ ˙ + |o Dw̄o ẋ + U∞ L2o D |o DD−1 w̄ + |o x + |o ẋ = 0 (32)
∂x ∂ w̄ ∂x ∂ ẋ
D−1
Premultiplying the above equation by leads to
U∞ L2o
 
1 −1 1 −1 ∂A ∂ F̄ ∂ F̄ ∂ F̄
˙
D ADw̄ + D |o Dw̄o ẋ + |o w̄ + |o x + |o ẋ = 0 (33)
U∞ L2o U∞ L2o ∂x ∂ w̄ ∂x ∂ ẋ
The volume matrix, A, has the dimension of a cubic length and therefore can be adimensionalized as
1 −1
Ā = D AD (34)
L3o

13 of 16

American Institute of Aeronautics and Astronautics

Hence, equation (33) can be re-written as

L3o L3o
 
∂ Ā ∂ F̄ ∂ F̄ ∂ F̄
˙
Āw̄ + |o w̄o ẋ + |o w̄ + |o x + |o ẋ = 0 (35)
U∞ Lo 2 2
U∞ Lo ∂x ∂ w̄ ∂x ∂ ẋ

The time, t, can be adimensionalized as

U∞
τ= t (36)
Lo
and therefore partial derivative with respect to time can be expressed as
∂ U∞ ∂
= (37)
∂t Lo ∂τ
The fluid mesh motion vector, x, can be adimensionalized as
x
x̄ = (38)
co
where co has the dimension of a length. Hence, the gradient with respect to x can be expressed as
∂ 1 ∂
= (39)
∂x co ∂ x̄

and an adimensionalized expression of x̄˙ is given by

∂ x̄ Lo
x̄˙ = = ẋ (40)
∂τ co U∞
so that
∂ Lo ∂
= (41)
∂ ẋ co U∞ ∂ x̄˙
Substituting the above adimensionalizations in equation (35) yields

L3o U∞ L3o
 
∂ Ā co U∞ ∂ F̄ ∂ F̄ ∂ F̄
Ā w̄,τ + |o w̄o ( ˙ +
x̄) |o w̄ + |o x̄ + |o x̄˙ = 0 (42)
2
U∞ Lo Lo 2
U∞ Lo co ∂ x̄ Lo ∂ w̄ ∂ x̄ ∂ x̄˙
which can be re-arranged as
Āw̄,τ + H̄w̄ + (Ē + C̄)x̄˙ + Ḡx̄ = 0 (43)
where
∂ F̄
H̄ = |o
∂ w̄
∂ Ā
Ē = |o w̄o
∂ x̄
∂ F̄
C̄ = |o
∂ x̄˙
∂ F̄
Ḡ = |o
∂ x̄

14 of 16

American Institute of Aeronautics and Astronautics

 References
1 Albano, E. and Rodden, W., “A doublet-lattic method for calculating lift distribution on oscillating surfaces in subsonic

flow,” AIAA Journal, Vol. 7, 1969, pp. 279–285.

2 Ashley, H. and Zartarian, G., “Piston theory - A new aerodynamic tool for the aeroelastician,” Journal of Aerospace

Sciences, Vol. 23, 1956, pp. 1109–1118.

3 Yurkovich, R. N., Liu, D. D., and Chen, P. C., “The state-of-the-art of unsteady aerodynamics for high performance

aircraft,” AIAA Paper 2001-0428 , 2001.

4 Geuzaine, P., Brown, G., Harris, C., and Farhat, C., “Aeroelastic dynamic anlysis of a full F-16 configuration for various

flight conditions,” AIAA Journal, Vol. 41, 2003, pp. 363–371.

5 Farhat, C., Geuzaine, P., and Brown, G., “Application of a three-field nonlinear fluid-structure formulation to the

prediction of the aeroelastic parameters of an F-16 fighter,” Computers and Fluids, Vol. 32, No. 1, 2003, pp. 3–29.
6 Sadeghi, M. and Liu, F., “Investigation of non-linear flutter by a coupled aerodynamics and structural dynamics method,”

AIAA Paper 2001-0573 , 2001.

7 Liu, F., Cai, J., and Zhu, Y., “Calculation of wing flutter by a coupled fluid-structure method,” Journal of Aircraft,

Vol. 38, No. 2, 2001, pp. 334–342.

8 Romanowski, M. C., “Reduced order unsteady aerodynamic and aeroelastic models using Karhunen-Loeve eigenmodes,”

6th AIAA/USAF/NASA/ISSMO Symposium on Multidisciplinary Analysis and Optimization, September 1996, pp. 7–13.
9 Willcox, K. and Peraire, J., “Balanced model reduction via the proper orthogonal decomposition,” AIAA Paper 2001-

2611 , 2001.
10 Epureanu, B. I., “A parametric analysis of reduced order models of viscous flows in turbomachinery,” Journal of Fluids

and Structures, Vol. 17, 2003, pp. 971–982.

11 Hall, K. C., Thomas, J. P., and Dowell, E. H., “Reduced-order modelling of unsteady small-disturbance flows using a

frequency domain proper orthogonal decomposition technique,” AIAA Paper 99-16520 , 1999.
12 Hall, K. C., Thomas, J. P., and Dowell, E. H., “Proper orthogonal decomposition technique for transonic unsteady

aerodynamic flows,” AIAA Journal, Vol. 38, No. 2, October 2000, pp. 1853–1862.
13 Kim, T. and Bussoletti, J. E., “An optimal reduced-order aeroelastic modeling based on a response-based modal analysis

of unsteady CFD models,” AIAA Paper 2001-1525 , 2001.

14 Thomas, J. P., Dowell, E. H., and Hall, K. C., “Three-dimensional transonic aeroelasticity using proper orthogonal

decomposition-based reduced order models,” Journal of Aircraft, Vol. 40, No. 3, 2003, pp. 544–551.
15 Lieu, T. and Lesoinne, M., “Parameter adaptation of reduced order models for three-Dimensional flutter analysis,” AIAA

Paper 2004-0888 , 2004.

16 Cizmas, P. G. A. and Palacios, A., “Proper orthogonal decomposition of turbine rotor-stator interaction,” Journal of

Propulsion and Power , Vol. 19, No. 2, March 2003, pp. 268–281.
17 Spitler, J. E., “Initial studies of low-order turbulence modeling of the wind turbine in-flow environment,” AIAA Paper

2004-1004 , 2004.
18 Karhunen, K., “Zur Spektraltheorie stochastischer prozesse,” Ann. Acad. Sci. Fennicae, Vol. 34, 1946.
19 Loeve, M., “Functions aléatoire de second ordere,” Comptes Rendus Acad. Sci. Paris, Vol. 220, 1945.
20 Beran, P. S. and Pettit, C. L., “Prediction of nonlinear panel response using proper orthogonal decomposition,” AIAA

Paper 2001-1292 , 2001.

21 Lieu, T., Adaptation of reduced order models for applications in aeroelasticity, Ph.D. thesis, University of Colorado at

Boulder, 2004.
22 Farhat, C., Lesoinne, M., and Maman, N., “Mixed explicit/implicit time integration of coupled aeroelastic problems:

three-field formulation, geometric conservation and distributed solution,” International Journal For Numerical Methods in
Fluids, Vol. 21, 1995, pp. 807–835.
23 Farhat, C., Lesoinne, M., and LeTallec, P., “Load and motion transfer algorithms for fluid/structure interaction prob-

lems with non-matching discrete interfaces: momentum and energy conservation, optimal discretization and application to
aeroelasticity,” Computer Methods in Applied Mechanics and Engineering, Vol. 157, 1998, pp. 95–114.
24 Lesoinne, M. and Farhat, C., “CFD-based aeroelastic eigensolver for the subsonic, transonic, and supersonic regimes,”

Journal of Aircraft, Vol. 38, No. 4, 2001, pp. 628–635.

25 Lesoinne, M., Sarkis, M., Hetmaniuk, U., and Farhat, C., “A linearized method for the frequency analysis of three-

dimensional fluid/structure interaction problems in all flow regimes,” Computer Methods in Applied Mechanics and Engineering,
Vol. 190, 2001, pp. 3121–3146.
26 Lumley, J. L., “The structure of inhomogeneous turbulence,” Atmospheric Turbulence and Wave Propagations, edited

by A. Yaglom and V. Tatarski, Nauka, Moscow, 1967, pp. 166–178.

27 Holmes, P., Lumley, J. L., and Berkooz, G., Turbulence, coherent structures, dynamical systems and symmetry, Cam-

bridge University Press, 1996.

28 Beran, P., “Reduced-order modeling: new approaches for computational physics,” AIAA Paper 2001-0853 , 2001.
29 Taylor, J., Dynamics of large scale structures in turbulent shear layers, Ph.D. thesis, Clarkson University, 2001.

15 of 16

American Institute of Aeronautics and Astronautics

 30 Taylor, J. and Glauser, M., “Towards practical flow sensing and control via POD and LSE based low-dimensional tools,”

2002 ASME Fluids Engineering Division Summer Meeting, 2002, pp. 1–9.
31 Schmidt, R. and Glauser, M., “Improvements in low dimensional tools for flow-structure interaction problems: using

global POD,” AIAA Paper 2004-0889 , 2004.

32 Björck, Å. and Golub, G. H., “Numerical methods for computing angles between linear subspaces,” Mathematics of

Computation, Vol. 27, No. 123, 1973, pp. 579–594.

16 of 16

American Institute of Aeronautics and Astronautics