Journal of Fluids and Structures 117 (2023) 103835

Contents lists available at ScienceDirect

Journal of Fluids and Structures

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

A numerical investigation on direct and data-driven flutter

prediction methods
∗
Nikolaos Simiriotis, Rafael Palacios
Imperial College, London, SW7 2AZ, United Kingdom

article info a b s t r a c t

Article history: Three alternative algorithms for dynamic aeroelastic stability analysis are investigated.
Received 24 September 2022 They are based either on direct search of oscillatory conditions of the coupled system or
Received in revised form 4 December 2022 on eigenvalue analysis from frequency-domain sampling of the unsteady aerodynamics.
Accepted 11 January 2023
The aerodynamic forcing is obtained with a harmonic-balanced solver developed for a
Available online 19 January 2023
general-purpose finite-volume fluid dynamics solver. This new implementation is first
Keywords: verified against experimental flutter results from the literature. A wing with relatively
Flutter low bending stiffness is then used to explore the relative performance of each approach,
Limit-cycle oscillations both in terms of the numerical robustness and of their usability to support aeroelastic
Harmonic balance design.
Unsteady aerodynamics © 2023 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC
Numerical methods BY license (http://creativecommons.org/licenses/by/4.0/).
Loewner framework

1. Introduction

Accurate prediction of flutter boundaries for air vehicles is crucial to ensure the safety of flight operations. Conservative
design approaches may lead to overly stiff and, hence, heavy designs, while preliminary designs resulting in overly flexible
wings may display aeroelastic problems during the aircraft certification stage. It is also not unusual for flutter issues to be
identified only at the final design and even during flight testing stages, at which point design changes can be extremely
costly (Jonsson et al., 2019).
Flutter instabilities are driven by the resonant characteristics of the wing or aircraft. The fluid acts as a feedback on
the small-amplitude vibratory response of the wing, introducing both modal coupling and damping, and instabilities
occur when the resulting coupled system is able to extract net power from the fluid (Livne, 2018; Kholodar, 2022).
For flutter simulation, this means that the structure properties can be well described using a few (low-frequency)
natural modes, even around a non-zero (geometrically-nonlinear) aeroelastic equilibrium point, which is very convenient
to simplify the analysis. Once the structure has been expressed in modal coordinates, there are multiple approaches
that can be followed for transonic flutter prediction using state-of-the-art computational fluids dynamics (CFD) solvers.
The most straightforward is time-domain simulation of the coupled aeroelastic system. This needs the fluid solver to
solve simultaneously with the equations to deform the solid walls at each time step. It is however very computational
expensive (Henshaw et al., 2007) and indeed the modal projection of the structural dynamics can be skipped at minimal
cost, and as a result it is rarely used beyond simple geometries. A much more computationally-efficient solution, that
leverages upon the relatively small number of vibration modes necessary for convergence, is to construct a (linear)
reduced-order model from time-domain flow solutions under different modal excitations (Silva and Bartels, 2004). This
identified aerodynamic model is then coupled to the structural equations to assess stability.

∗ Corresponding author.
E-mail address: [email protected] (R. Palacios).

https://doi.org/10.1016/j.jfluidstructs.2023.103835
0889-9746/© 2023 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/
licenses/by/4.0/).
N. Simiriotis and R. Palacios Journal of Fluids and Structures 117 (2023) 103835

Nomenclature
α0 angle of attack
κ reduced frequency
ν aeroelastic mass ratio
ρ∞ freestream air density
τF pseudo-time step in flow solver
τS pseudo-time step in structural solver
ωF flutter frequency
ωj natural frequency of structural mode j
Aq generalized aerodynamic forces
b characteristic length (semichord)
D̄ pseudo-spectral operator matrix
fq generalized forces
L Loewner matrix
NAERO Number of aeroelastic (CFD/CSD) iterations
NCFD Number of iterations within the fluid solver.
NT Number of time instances in harmonic balance solution
q generalized displacements
U∞ freestream air speed
VF flutter speed
w flow conservative variables
•∗ time instances of a variable
•ˆ Fourier coefficients for a variable
ℜ(•) Real part of a complex number
ℑ(•) Imaginary part of a complex number
Ma Mach number
Re Reynolds number

Alternatively, as flutter results in harmonic oscillations, it can be more natural to consider frequency-domain solutions
to both the fluid and coupled problem. This also links to traditional aeroelastic methods based on potential-flow theory,
which are mostly based on frequency domain models. With frequency-based methods, the time dependency of the state
variables in the fluid equations is approximated by a Fourier series and the time-domain problem is transformed into
a series of coupled steady-state problems. This results in the family of harmonic-balanced (HB) solution methods (Hall
et al., 2002; Dimitriadis, 2017). Importantly, when then number of harmonics is very small (in fact only one is necessary
for asymptotic stability analysis), the cost per simulation is comparable to that of a steady-state problem. The cost of the
fluid solution can still further reduced by using linear frequency-domain (LFD) (Thormann and Widhalm, 2013) solvers.
The drawback is the additional expense of bespoke implementations with increased algorithmic complexity, and the LFD
solver needs to be maintained in parallel with the solvers used for the aerodynamic design (and possibly differentiated,
for applications to design optimization (Jonsson et al., 2019)).
Once the frequency-domain fluid solver is available there are essentially two main approaches for flutter prediction,
which we can broadly characterize as sampling-based and direct-search strategies. In the sampling-based strategies prior
(offline) values of the aerodynamic forces are computed under different model inputs. They are then coupled (in the
online phase) with the structural dynamic equations to investigate the asymptotic stability conditions of the coupled
system. The most common strategy to achieve this is to use tabulated aerodynamic data in the solution of a modified
(nonlinear) eigenvalue problem (Hassig, 1971) from which the flutter conditions can be derived. This approach is typically
very robust and it is widely used in aeroelastic design (Henshaw et al., 2007). Alternatively, the sampled frequency-
domain forcing can be used to identify first a linear-time-invariant state–space model of the unsteady aerodynamics.
Flutter conditions are then established from the eigenvalues of the coupled (linear) system obtained from the feedback
interconnection of the identified aerodynamics and the structural admittance. Determining a linear dynamical system
for the aerodynamic forces is a data-driven inverse problem known as the realization problem, which for aeroelastic
applications is typically solved using Padé approximants with coefficients identified using least-square methods (Roger,
1977). The main challenge in that approach is that, for a robust numerical performance, it needs either many samples (CFD
runs) or additional constraints (typically, on the poles of the resulting model) for convergence. This can be avoided with
Loewner’s rational matrix approximation (Antoulas, 2005, Ch. 4), which has recently been demonstrated for applications
in unsteady aerodynamics (Quero et al., 2021). The Loewner matrix approach is the frequency-domain equivalent to
2
N. Simiriotis and R. Palacios Journal of Fluids and Structures 117 (2023) 103835

the time-domain eigenvalue realization algorithm (Silva and Bartels, 2004), built on a singular value decomposition of a
Hankel-type matrix. However, instead of the Markov’s parameters obtained from the step response of the fluid equations,
it uses the frequency samples obtained from the harmonic-balance solver.
Direct (or Hopf-bifurcation) methods do not rely on prior samples and instead seek the conditions at which the
coupled system, including the HB fluid solver, displays periodic oscillations. This was first proposed to accelerate flutter
prediction methods by Cardani and Mantegazza (1978). The integration of the approach with a HB fluid solver was
proposed by Thomas et al. (2002) to determine flutter boundaries and, more generally, periodic trajectories of finite
amplitude, or limit-cycle oscillations (LCO), when nonlinearities are included. The stability boundary of a two-degree-
of-freedom transonic aeroelastic system was determined with a prescribed pitching amplitude and a Newton–Raphson
search. Later modifications (Thomas and Dowell, 2018) have allowed for the aeroelastic governing equations to be solved
simultaneously with the aerodynamic solver. The computational cost of this approach is independent of the number of
structural degrees of freedom.
Variations of the classical balancing method include the nonlinear frequency domain (NLFD) form (McMullen and
Jameson, 2006) and the time spectral (TS) form (Gopinath and Jameson, 2005). He et al. (2019) applied the TS method
to create a high-fidelity flutter onset prediction methodology that could be used to capture the subcritical/supercritical
LCO response. The authors proposed a preconditioned, Jacobian-free, coupled Newton–Krylov solution strategy for the
TS flutter equations, which resolved both the computational structural dynamics (CSD) and CFD equations in a robust
and efficient manner. In particular, they assembled the dependent flow and structure variables into one global unknown
vector which was then solved using a Newton–Krylov method. The authors compared the proposed method against results
acquired by time-marching (TM) simulations to demonstrate the accuracy and the efficiency of the approach.
Based also on the HB approach, Li and Ekici (2017) have developed an efficient ‘‘one-shot’’ method for determining
self-excited LCO for 2-DOF airfoils, in inviscid and viscous transonic flow regimes. In their work, the fluid and the structure
fields were calculated by integrating the respective governing equations in pseudo-time, driving the residual of both fields
to convergence simultaneously, while an optimization search was employed for the frequency update, all in ‘‘one shot’’.
Later, the authors employed their one-shot method (Li and Ekici, 2018a) to capture the unstable LCO branch below the
flutter point, and to predict the onset of flutter (Li and Ekici, 2018b). In this improved approach, the amplitude of the
deformation was prescribed as an input and the reduced velocity was treated, along with the response frequency, as
unknown.
Direct and sampling-based methods are however not mutually exclusive and the key novelty of this paper is an
investigation on the complementarity of both approaches for flutter analyses. Several flutter prediction strategies are
explored, all stemming from a single harmonic-balance CFD solver, and their relative merit is assessed on a representative
wing configuration. In particular, the open-source SU2 CFD package (Economon et al., 2016), a second-order finite-volume
solver, is used in this work. The paper presents, next, a summary of the harmonic-balanced methods implemented in
the SU2 solver, which includes rather uniquely both a native finite-volume RANS and a finite-element solid mechanics
solver (both are solved using HB). Section 3 then discusses in some detail direct and sampling-based strategies to
predict flutter conditions using the HB solver. The sampling-based methods include the conventional solution using
tabulated aerodynamic forces and data-driven solutions in which a dynamical system model is constructed from the
frequency samples. All those strategies are then exercised in Section 4 on two tests cases. The first one is the well-
known AGARD445.6 aeroelastic wing (Yates, 1987), which is used to verify the HB implementation, while the second
one considers a more flexible wing with a more complex flutter mechanism. The computational trade-offs between the
different approaches to deal with those problems are finally discussed.

2. Methodology

2.1. Harmonic balancing of the fluid equations

The Arbitrary Lagrangian–Eulerian (ALE) formulation of the Reynolds-Averaged Navier–Stokes equations, with grid
displacement imposed at the solid boundary, is employed. For time-dependent problems, the mass and momentum
equations for compressible flow can be written as
∂w v v is
ALE − F ) = −R(w),
= −∇ · (Fin (1)
∂t
v v is
where w = {ρ; ρ u; ρ e} is the column of the conservative variables, and Fin ALE , F are the inviscid and viscous fluxes,
respectively (Economon et al., 2016). Source terms can also be added to consider, for example, volume changes on the
grid cells. The Euler equations can be derived by setting the viscous fluxes to zero.
Assuming now given periodic displacements at the solid boundaries, with a fundamental angular frequency ω, the
resulting nonlinear flow dynamics can be approximated as a truncated Fourier series with NH harmonics. Let ŵ =
{ŵ0 ; ŵA,1 ; ŵB,1 ; . . . ; ŵA,NH ; ŵB,NH } be the Fourier coefficients in the approximation of the conservative variable. A closed-
form transformation w∗ = E −1 ŵ, can also be determined, where w∗ = {w(t0 ); w(t1 ); . . . ; w(tNT −1 )} are the variable values
3
N. Simiriotis and R. Palacios Journal of Fluids and Structures 117 (2023) 103835

at time-instances with NT = 2NH + 1, tn = nθ/ω, θ = 2π/NT , and E the constant matrix

⎡1 1 1 ... 1 ⎤
2 2 cos θ 2 cos 2θ · 2 cos (NT −1)θ
2 2 sin θ 2 sin 2θ · 2 sin (NT −1)θ
1 ⎢
⎢. .. .. .. ⎥.
⎥
E= (2)
NT ⎣ .. . . . ⎦
2 2 cos NH θ 2 cos 2NH θ · 2 cos (NT −1)NH θ
2 2 sin NH θ 2 sin 2NH θ · 2 sin (NT −1)NH θ

For the case of one harmonic, which will be considered later, it is then NT = 3 and E a square matrix of dimension 3.
The definitions above allow for the balancing of the flow equations under the periodicity assumption. As a result, Eq. (1)
can be written as
ωD̄w∗ = −R∗ , (3)
where R are the NT time instances of the residual, and the pseudo-spectral (or balancing) operator ωD̄ has been
∗

introduced with D̄ = E −1 AE and

[ ]
0 1
A = diag(0, G1 , . . . , GNH ), with Gm = m . (4)
−1 0
The HB approach results in NT steady systems that need to be solved simultaneously as they are coupled through the
pseudo-spectral operator ωD̄. To improve numerical convergence, Eq. (3) is augmented with a pseudo time-marching
term, which is inherited from the steady RANS solver (Economon et al., 2016). Moreover, in order to consider arbitrary
surface deformations, the volume variation of the fluid cells in each time-instance needs to be taken into account as
additional source terms. For each instance n = 0, . . . , NT − 1 in Eq. (3), we finally solve the following set of equations,
δ wn wn
= −Rn − ωD̄w∗ − ω (D̄V ∗ ), (5)
δτF Vn
where wn = w(tn ) is the flow field vector at each instance, Vn = V (tn ) is the volume at each instance, and V ∗ =
{V (t0 ), V (t1 ), . . . , V (tNT −1 )}T . The present implementation of the HB solver treats the ωD̄w∗ part of the equation as a source
term and solves Eq. (5) iteratively (Rubino et al., 2018). The mesh deformation, with respect to the undeformed grid,
at each time-instance is calculated from the boundary displacement of the structure by solving a pseudo-linear elastic
problem of the form (Venkatesan-Crome et al., 2019)

KG xG − f G = 0, (6)
where KG is the fictitious stiffness matrix of the grid and f G is the vector of fictitious forces that enforces matching
boundary displacements at the fluid/solid interface. The grid velocity invoked in the ALE formulation, is then approximated
at each instance as:
ẋ∗G ≈ ωD̄x∗G . (7)

2.2. Structural equations

The linear governing equations of motion of the structure, on a suitable spatial discretization (typically through the
finite-element method), read

M η̈ + C η̇ + K η = f , (8)
where M is the mass matrix, C is the damping matrix, K is the stiffness matrix, η is the displacement vector, and f the
external loading, which includes the aerodynamic forces. Projection of the damped equations (8) onto the linear normal
modes of the (undamped) structure, results in

Mφ q̈ + Mφ Tφ Ω q̇ + Mφ Ω 2 q = f q , (9)

where Φ = φ1 , φ2 , . . . , φr is a set of r low-frequency mode shapes, q are the generalized displacements that
[ ]
approximate η = Φ q, and f q = Φ T f are the generalized forces. The matrix Mφ = Φ T M Φ contains the modal masses,
while Tφ contains the modal damping factors, and can be assumed Tφ = diag(2ξ1 , 2ξ2 , . . . , 2ξr ) if certain conditions apply
for C (Géradin and Rixen, 1997, Ch. 3). Finally, Ω = diag(ω1 , ω2 , . . . , ωr ) is the matrix containing the r natural frequencies.
The pseudo-spectral operator ωD̄ is also employed to acquire the HB set of equations, as

ω2 D̄2 q∗ + ωTφ Ω D̄q∗ + Ω 2 q∗ = Mφ−1 f ∗q , (10)

where q∗ and f ∗q contain the generalized displacements and forces, respectively, at discrete time-instances. Once q∗
is calculated, the values of the displacement at the boundary nodes of the fluid/solid interface is obtained from the
corresponding terms in the modal matrix Φ and a radial-basis function (RBF) interpolation to accommodate non-matching
discretizations between both domains (Beckert and Wendland, 2001). Its transposed transformation is used to obtain a
consistent mapping of the aerodynamic forces on the structural grid.
4
N. Simiriotis and R. Palacios Journal of Fluids and Structures 117 (2023) 103835

3. Flutter solution methods

Three different strategies to predict flutter onset are examined. All strategies employ the HB fluid solver described in
Section 2.1. First, we will seek direct solutions, in which the neutrally stable conditions are enforced on the coupled
problem. Second, we will seek sampling-based strategies, in which HB fluid solutions are first evaluated and then
interpolated to solve a nonlinear eigenvalue problem. Finally, the sampled frequency-domain fluid data will be used to
construct an identified linear time-invariant dynamical system (a data-driven model) of the fluid. The closed-loop stability
of the resulting coupled fluid–structure system determines flutter. Each method is described in some detailed next.

3.1. Direct prediction

Direct methods seek periodic trajectories, known as limit-cycle oscillations (LCOs), in the dynamics of the coupled
problem. LCOs in aeroelastic problems typically occur across a range of airspeeds and, for a given system without external
forcing, the amplitude of the oscillations is uniquely dependent on the airspeed. Flutter conditions can be then seen as a
particular case of an LCO, corresponding to the airspeed at which the amplitude of the oscillations can be made arbitrarily
small, and with the actual value of those amplitudes determined by the amount of energy injected onto the system at
t = 0. This insight enables a direct flutter prediction method, which is derived here from the one proposed by Li and
Ekici (2018a,b) with only some algorithmic changes to improve robustness, as described below.
For convenience, a nondimensional form of Eq. (10) is first derived by normalizing the aerodynamic forces with the
freestream dynamic pressure of the fluid, 21 ρ∞ U∞ 2
, and a reference mass m0 , surface S and volume, V of the solid. By
choosing next a characteristic length, b, and frequency, ω1 , of the structure, and mass normalized modes, the dimensionless
form of Eq. (10) reads
Sb ∗
ω̃F2 D̄2 q̃∗ + ω̃F Tφ Ω̃ D̄q̃∗ + Ω̃ 2 q̃∗ = ṼF2 f˜ q (Ma, Re), (11)
2V
∗ √
where ω̃F = ω/ω1 , q̃ = q∗ /b, and ṼF = U∞ /( µbω1 ) is known as the flutter speed index, with µ = m0 /(ρ∞ V) the mass
ratio. In order to improve convergence, Eq. (11) can also be augmented with a (non-dimensional) pseudo time-marching
term, as (Li and Ekici, 2017)
δ q̃∗ Sb ∗
+ ω̃F2 D̄2 q̃∗ + ω̃F Tφ Ω̃ D̄q̃∗ + Ω̃ 2 q̃∗ − ṼF2 f˜ q (Ma, Re) = 0. (12)
δτS 2V
In general, the fluid and structural pseudo time-steps in Eqs (5) and (12) do not have to coincide. When searching for
a flutter condition, both the flutter frequency ωF and speed VF are unknown a priori and are treated as independent
variables. For the aeroelastic prediction to converge to a unique flutter point (ωF , VF ) the amplitude and the phase of a
structural mode must be constrained, in particular (a) the phase will provide an initial time reference for the periodic
aeroelastic solution, and (b) the chosen amplitude (Ao , of sufficiently small value) will uniquely fix the solution along the
bifurcation diagram of the system.
Without loss of generality, the phase and amplitude of the first mode can be fixed. Moreover, as linear structural
dynamics are considered, only one harmonic is sufficient when expanding each structural degree of freedom, as

q1 (t) = q̂1,0 + q̂1,A1 cos(ωt) + q̂1,B1 sin(ωt). (13)

In this expression, q̂1,0 corresponds to the mean value of the mode, which is retained in general to compute flutter
around a deformed geometry (the assumption ( is then that)the aerodynamic mesh is given for the undeformed geometry).
1/2
The modal amplitude is obtained as A1 = q̂21,A1 + q̂21,B1 while the predicted phase of the mode is given by θ1 =
arctan 2(q̂1,A1 /q̂1,B1 ) − π/2. Only three time instances are needed for a problem with one harmonic, as

q∗1 = [q1 (t0 ), q1 (t1 ), q1 (t2 )]T = E −1 [q̂1,0 , q̂1,A1 , q̂1,B1 ]T , (14)

with tn = 2nπ /3ω. The phase of the mode can be fixed by setting q̂1,A1 = 0, while the amplitude is constrained by setting
q̂1,B1 = Ao . The rest of the structural modes should be accordingly shifted to retain the relative phase difference,

qk (t) = q̂k,0 + q̂k,A1 cos(ωt + θ1 ) + q̂k,B1 sin(ωt + θ1 ), for 1 < k ≤ r . (15)

w w
A relaxation factor can be included at this stage by scaling the modal amplitude with = q̂neand
1,A1 λq̂old
= λq̂old
1,A1 q̂ne
1,B1 ,1,B1
where λ = A1 /Ao . In this case, the chosen mode q1 should then be shifted along with the rest of the modes using Eq. (15).
The unknown (ωF , VF ) values are updated at every aeroelastic iteration NAERO in Algorithm 1 for the complete system
to converge to a unique solution. This is achieved by minimizing a figure of merit, based here on the structural residual
of Li and Ekici (2018a), namely, L(ωF , VF ) = 21 R Ts R s , with

∗ ∗ ∗ Sb ∗
R s = ω̃F2 D̄2 q̃ + ω̃F Tφ Ω̃ D̄q̃ + Ω̃ 2 q̃ − ṼF2 f˜ q = 0. (16)
2V
5
N. Simiriotis and R. Palacios Journal of Fluids and Structures 117 (2023) 103835

Algorithm 1 Direct prediction algorithm.

1: procedure Search for ωF , VF
2: Initialize ωF , VF and q∗0
3: for j ≤ NCFD do
4: Iterate HB/CFD (Eq. (5)) ▷ Initialize HB flow fields
5: end for
6: for i ≤ NAERO do
7: March forward HB/CSD (Eq. (12)) ▷ One pseudo time-step
8: Impose amplitude/phase constraints
9: Update VF (Eq. (17))
10: Deform HB/CFD grids ▷ Compute grid velocities
11: Update inlet BC
12: for j ≤ NCFD do
13: Iterate HB/CFD (Eq. (5)) ▷ Update aerodynamic forces
14: end for
15: Update ωF (Eq. (17))
16: if convergence then Break
17: end for
18: end procedure

The flutter frequency and flutter speed are finally updated by solving
∂L ∂L
= = 0. (17)
∂ωF ∂ VF
Based on the discussion in this section, a flutter/LCO prediction strategy has been developed and the resulting procedure
is described in Algorithm 1. First, the solver is initialized and the HB flow equations are iteratively solved. Since flutter
is determined from asymptotic stability analysis, only one harmonic is sufficient in the Fourier expansion of the fluid.
At every aeroelastic iteration, ωF and VF are sequentially updated following (17). The update on the airspeed is carried
out first, after the structural problem is solved and the amplitude/phase constraints are imposed. Based on the updated
airspeed, the inlet boundary conditions for the flow are updated. This may include updates on both Re or Ma number in
the fluid simulations, as appropriate. Then, the CFD grids are computed for each time-instance with respect to the initial
undeformed geometry. The fluid equations are then solved. After the update of the aerodynamic forces, the new frequency
is calculated. Finally, the updated aerodynamic forces and the new frequency are used to march forward the CSD problem
in the next aeroelastic iteration. For the case of structures undergoing large (geometrically-nonlinear) deformations, the
equilibrium shape associated to the mean aerodynamic forces, and the corresponding mode shapes, can be updated in
each aeroelastic iteration. The whole procedure is repeated until ωF and VF converge, and both the structural and the fluid
residuals sufficiently decrease. This simple sequential approach was found to be more robust for large problems than the
simultaneous update followed in Li and Ekici (2018b).

3.2. Nonlinear eigenvalue problem

Look-up tables of frequency-domain aerodynamic forces have been historically the basis for flutter prediction methods
and harmonic-balance methods provide a natural way to obtain those offline generalized forces from CFD under prescribed
kinematics at the fluid/solid interface (Badcock et al., 2011). As we seek to assess asymptotic stability, only the forces under
small amplitude harmonic oscillations need to be evaluated.
The offline sampling of the aerodynamic forces proceeds then as follows. For each mode shape, an imposed amplitude
and frequency are defined on the fluid boundaries, following Eq. (13) but without the constant terms, that is, q̂0 = 0. For
convenience, the problem is normally solved in the complex domain, with q = q̄ejωt , where q̄ = q̂A,1 − jq̂B,1 . Sampling
through frequencies, the non-stationary aerodynamic forces can be written as f¯ q (κi ) = 21 ρ∞ U∞ 2
Aq (κi )q̄ for a range of
values of the reduced frequency κ = bω/U∞ . We have expressed the forces in terms of the reduced frequency, as it can
be shown through dimensional analysis that it results in unique values of the aerodynamic coefficients for given Ma and
Re. The resulting matrix, Aq (κ, Ma, Re), is known as the matrix of generalized aerodynamic forces (GAFs).
As a result, the equations of motion in modal coordinates, Eq. (9) can be now written as
−ω2 Mφ + jωMφ Tφ Ω + Mφ Ω 2 + 21 ρ∞ U∞
2
Aq (κ, Ma, Re) q̄ = 0.
[ ]
(18)
where Aq (κ, Ma, Re) is only known at discrete values of the frequencies, Mach and Reynolds. The flutter onset conditions
are determined by the combination of airspeed, air density (or altitude), and frequency of oscillations that result in non-
trivial solutions to Eq. (18). As an alternative to the direct solution described in the previous section, it is more common to
seek solutions that, first, construct a look-up table of GAF matrices offline, and then, solve a modified but easier problem
6
N. Simiriotis and R. Palacios Journal of Fluids and Structures 117 (2023) 103835

for which Eq. (18) is a particular solution. This results in an nonlinear eigenvalue problem, which is solved iteratively
with GAFs interpolated from the given data at each iteration. A popular algorithm to solve this problem, which will be
exercised here, is the p–k method (Hassig, 1971). It seeks solutions to a generalized nonlinear eigenvalue problem in the
(unknown) flutter frequency. Assuming zero structural damping, it is
V2
[ ]
− ∞ p2 Mφ + Mφ Ω 2 + 12 ρ∞ V∞
2
Aq (κ, Ma, Re) q̄ = 0, (19)
b
where p = γ κ + jκ is a complex number that includes a non-dimensional rate of decay, γ . When γ = 0 the flutter
equation is recovered. For given air density and Mach number, pairs (κk , γk ) that satisfy (19) are found for increasing
values of the airspeed U∞ through an iterative procedure. First, an initial guess κk is set for the reduced frequency of
the kth aeroelastic mode. The corresponding GAF matrix Aq (κk ) is then obtained through interpolation from the sampled
frequencies in the offline calculations. Akima interpolation on real and imaginary parts for element of the matrix is used
here. Once the GAF matrix is fixed, Eq. (19) becomes a generalized eigenvalue problem in p, that can be readily solved at
little computational cost. Next κk is updated from the imaginary part of the closest eigenvalue, resulting in a fixed-point
iteration of the form κk = ℑ(pk ), in which the left hand side defines the reduced frequency on Aq , and the right-hand side
the solution to the resulting eigenvalue problem. This process is repeated for each mode and then the airspeed is updated.
The flutter speed is obtained when one of the roots satisfies ℜ(pk ) = 0. This results in unmatched flutter conditions, in
which the flutter velocity does not necessarily coincide with the Mach and (if viscous effects are considered) Reynolds
numbers used to obtained the sampled GAF matrices. An outer iteration loop may be at this point necessary to enforce
matched conditions between the flow parameters and the flutter point (Henshaw et al., 2007). For certification purposes,
however, the unmatched values are directly used to define velocity margins.

3.3. Rational interpolation

A brief summary of the Loewner’s rational matrix approximation, as implemented here for aeroelastic applications, is
included next.
Let Aq (κi ), with i = 1, . . . , n and κi+1 > κi , be the known (reduced) frequency samples of the generalized aerodynamic
matrix, obtained as in Section 3.2 above for a range of flow parameters (typically Ma and Re). For simplicity, we consider
here only non-zero positive frequencies (although they can be very small) and an even number of samples, that is,
mod (n, 2) = 0. We partition the data into two equal groups and for each frequency/GAF pair, we also introduce their
complex conjugates. A common split, which is followed here, is by odd and even order of the frequencies.
The first group, known as the left data, will be denoted as {µi , φi }i=1,n , where µ = {jκ1 , −jκ1 , jκ3 , −jκ3 , . . . , jκn−1 ,
]⊤
−jκn−1 } and similarly matrix φ = A⊤ q (κ1 ) Āq (κ1 ) Aq (κ3 ) Āq (κ3 ) . . . Aq (κn−1 ) Āq (κn−1 )
[ ⊤ ⊤ ⊤ ⊤ ⊤
is obtained from column
concatenation, with •¯ the complex conjugate. The second group, known as the right data, will be denoted as {νi , ψi }i=1,n ,
and it is defined in a similar manner using the even frequencies. The associated Loewner matrix, L, and the shifted Loewner
matrix, Lσ , are then computed from the divided differences between both sets, as
⎡φ φ1 −ψn
⎡µ
1 φ1 −λ1 ψ1 µ1 φ1 −λn ψn
⎤ ⎤
1 −ψ1
µ1 −λ1
··· µ1 −λn µ1 −λ1
··· µ1 −λn
⎢ .. .. .. ⎥ ⎢ .. .. .. ⎥.
⎥
.
⎥ and Lσ = ⎢
L=⎢
⎣ . . ⎦ ⎣ . . . ⎦ (20)
φn −ψ1 φn −ψn µn φn −λ1 ψ1 µn φn −λn ψn
µn −λ1
··· µn −λn µn −λ1
··· µn −λn

It can be proven (Karachalios et al., 2021) that, under certain rank conditions on both complex matrices, it is Aq (jκ ) ≈
ψ⊤ (Lσ − jκ L)−1 φ, that is, the Loewner framework defines a state–space model in descriptor form1 that interpolates
the given data. Moreover, as the interpolant is an analytic function, its domain of definition can be expanded to the full
complex plane by simply replacing jκ with the corresponding non-dimensional Laplace variable. Both matrices in (20) are
complex, with dimension the product of the number of frequency samples, n, times the number of structural modes, Nq .
As we seek a real-valued model, the following linear transformations are introduced
LR = J ⊤ LJ , Lσ R = J ⊤ Lσ J , φR = J ⊤ φ, and ψ R = J ⊤ ψ, (21)
with J = J1 ⊕ J2 ⊕ · · · ⊕ Jn being a block diagonal matrix built with n equal submatrices of the form
[ ]
1 Iq −jIq
Ji = √ , (22)
2 Iq jIq
where Iq is the unit matrix of dimension Nq . Finally, a SVD truncation is introduced to obtain a minimal interpolant of
the data. A two-sided projection is used with
[ ]
LR
Lσ R = U Σ̂ V̂ ⊤ = Û Σ V ⊤ .
[ ]
LR and (23)
Lσ R

1 Recall that for descriptor state–space with mass matrix E, state matrix A, input matrix B and output matrix C , the frequency response function
is G(ω) = C (jωE − A)−1 B.

7
N. Simiriotis and R. Palacios Journal of Fluids and Structures 117 (2023) 103835

Fig. 1. The first bending (left), and first torsion (right) mode shapes interpolated onto the aerodynamic surface mesh.

The descriptor state–space model, in physical time, is then obtained as

U∞
b
Ea ẋa = Aa xa + Ba q,
(24)
fq = Ca xa ,
with xa a set of aerodynamic states and

Ea = −U ⊤ LR V Aa = −U ⊤ Lσ R V , Ba = U ⊤ φR and Ca = ψ ⊤
R V. (25)

(For sflutter analysis,

)−1 we finally define the aerodynamic transfer function (modal amplitudes to modal forces) Ha (s) =
Ca U∞b
E a − A a Ba . Next, we obtain the structural transfer function (modal forces to modal amplitudes), Hs (s) from
Eq. (9). The dynamic stability of the aeroelastic system is finally determined by the poles of the closed-loop transfer
function (I − Hs Ha )−1 Hs . Stability analysis can finally be done in frequency or time domain using standard methods for
the analysis of linear-time invariant systems.

4. Results

The direct and sampling-based methodologies described above have been implemented within SU2 (Economon et al.,
2016). The implementation of the HB fluid solver has been verified in Simiriotis and Palacios (2022). A modal solver for
the native solid mechanics solver in SU2 has also been developed, which computes mode shapes within SU2 from solid
FE models. The solution of the eigenvalue problem employs the Spectra library (https://spectralib.org/), a C++ library for
large scale eigenvalue problems, which is built on top of Eigen (https://eigen.tuxfamily.org). Mode shapes are interpolated
onto the aerodynamic surfaces using radial basis functions. An Euclid’s hat basis function is used in this work.
Two test cases are investigated. First, the AGARD445.6 wing is considered to verify our implementation of the harmonic
balance solver on deforming meshes in SU2 and of the direct solution strategy. Second, a more flexible also transonic
wing is built, for which the structure is also modeled within the native solid mechanics implementation in SU2. That
wing displays a more complex stability diagram, and the performance of all three strategies discussed in Section 3 is
investigated.

4.1. AGARD445.6 test case

The weakened AGARD445.6 model 3 wing (Yates, 1987) is first considered to verify the direct prediction method of
Section 3.1. It employs a symmetric NACA 64A004 profile and has a sweep angle of 45 deg at the quarter chord line. The
span is 0.726 m, root chord is 2b = 0.5588 m, and the taper ratio is 0.66. Viscous effects are not included in the current
analysis. A rectangular unstructured computational grid is build around the wing. The dimensions of the computational
domain are Dx × Dy × Dz = 28b × 12b × 24b, where x is positive in the stream-wise direction, and y direction goes
along the span of the wing. The vertical displacement of the wing is along the z axis. The aerodynamic surface mesh of
the model consists of 65,000 boundary elements. The total grid size was about 320,000 volume elements. Convergence
studies can be found in Simiriotis and Palacios (2022).
The modal description of the structure is taken from Yates et al. (1963). The mode shapes are given on a 11 × 11 plane
grid, and were normalized to a unit modal mass in lbf-in-s2 , or 175.13 kg. The first four normal modes are used in this
study. The corresponding natural frequencies are 9.6, 38.17, 48.35, and 91.54 Hz. The first two modes on the aerodynamic
surface are shown in Fig. 1, with respect to the undeformed mean surface. As it can be seen, the interpolated displacement
accurately follows the imposed structural deformation.
To construct the flutter boundary, six Mach numbers are investigated. The specifications for every test case C # are
shown in Table 1. For each case, the freestream density ρ∞ (i.e. the flight altitude) is fixed and the mass ratio µ is
calculated accordingly to keep the mass of the wing constant. The freestream pressure at the flutter onset is searched for
through the updates of the flutter velocity. The first bending mode q1 was chosen to be constrained here. A sufficiently
small value |q1 |target = 0.001 (corresponding to a maximum displacement of 2.6% of the span) was given as an input
to the solver for all the calculations. The phase of q1 was always fixed at a zero phase. For each aeroelastic simulation,
8
N. Simiriotis and R. Palacios Journal of Fluids and Structures 117 (2023) 103835

Table 1
Specifications for the flutter boundary prediction (ρ∞ in kg/m3 ).
Case C1 C2 C3 C4 C5 C6
Ma 0.499 0.678 0.799 0.901 0.960 1.072
ρ∞ 0.4267 0.2077 0.1535 0.0992 0.0633 0.055
µ 33.456 68.753 93.024 143.92 225.82 259.59

Fig. 2. Flutter boundary of the AGARD 445.6 model 3, prediction of the (a) flutter velocity index, and (b) nondimensional flutter frequency, compared
with Yates et al. (1963), Lee-Rausch and Batina (1995), Li and Ekici (2019).

the displacement vector was initialized with the imposed |q1 |target , and zero generalized displacements for the rest of the
modes. No initial phase difference between the modes was employed.
An Euler solver has been employed for the fluid, with freestream boundary conditions set at the boundaries of the
computational domain, and slip (Euler) wall conditions on the wing. HB calculations for linear stability only need one
harmonic for both the fluid and the structural problem. The wing was clamped on its root on the xz plane which is
given a symmetry boundary condition. The JST central scheme with second and fourth order dissipation coefficients was
employed for this study. About NCFD = 100 iterations were carried out with the fluid solver in every aeroelastic iteration;
the pseudo time-step for the structure has been set at τS = 30. Four times more iterations were carried out for the
initialization of the HB flow fields. For all the calculations, the initial frequency was chosen at ωF ,0 /ω2 = 0.34, were ω2
is the natural frequency of the second mode, i.e. the first torsion mode. The initial reduced velocity was set at ṼF ,0 = 0.5
for most cases, while the sensitivity of the prediction on the initial VF ,0 and ωF ,0 values was evaluated.
The prediction of the flutter boundary is shown in Fig. 2 for varying Mach number. Results are plotted against Euler
computations from Lee-Rausch and Batina (1995), the ‘‘one-shot’’ approach of Li and Ekici (2019), and the experimental
data of Yates et al. (1963). The transonic dip at Ma = 0.96 is well captured with the current implementation and very good
agreement is found with the literature. In particular, the flutter velocity index follows exactly the prediction in Lee-Rausch
and Batina (1995) until the transonic dip and a slight underestimation in flutter velocity is seen at Ma = 0.96. Present
results slightly underpredict the flutter frequency compared to both (Lee-Rausch and Batina, 1995) and Li and Ekici (2019)
for every Ma < 1, but not compared to the experiments. Finally, the inviscid simulations give larger errors at supersonic
speeds, where shock-wave boundary layer interactions become significant.
The convergence of the direct solution procedure was also investigated. It was found that it strongly relates to the
aerodynamic force prediction and is affected by variations in the mean flow. The convergence of the flutter velocity index
and the flutter frequency for different Mach numbers is shown in Fig. 3. The solution procedure for 0.8 < Ma < 1.0
was found to be deeply converged with a frequency/velocity tolerance 10−6 at about NAERO = 800 aeroelastic iterations.
For lower Mach values, the flutter prediction converges to the same tight tolerance at about 200 aeroelastic iterations.
Outside the 0.8 < Ma < 1.0 range, shocks either do not form, or they form at the aft part of the wing profile sections. This
significantly accelerates the convergence of the flow solution and, thus, the convergence of the direct flutter prediction.
The convergence of VF and ωF prediction tracks the decrease of flow residuals, but can be further accelerated by reducing
the pseudo time-step of the structural solver. Moreover, the solution procedure was found to be numerically unstable
when the flow residuals were not decreased enough at every aeroelastic iteration, relative to the previous one. The stability
of the flow solution is governed by the number of NCFD iterations and the choice of pseudo time-step in the flow solver.
Finally, it was found that the procedure remained quite insensitive to changes of the initial VF (see Fig. 3(a)), but it was
numerically unstable when ωF ,0 ≤ ω1 and when ωF ,0 values approached ω2 . The next example shows that the direct
solution also performs well when the first torsional mode has a relatively higher natural frequency.
9
N. Simiriotis and R. Palacios Journal of Fluids and Structures 117 (2023) 103835

Fig. 3. Flutter boundary of the AGARD 445.6 model 3, convergence of the (a) flutter velocity index, and (b) the flutter frequency for different Mach
numbers.

Fig. 4. Flexible swept wing based on the NACA0012 airfoil, (a) aerodynamic surface mesh, and (b) pressure distribution on the surface of the wing
for M = 0.8 at a0 = 0.

4.2. Flexible swept wing

A longer and more flexible wing has also been created to investigate the three flutter solution algorithms. It employs
the symmetric NACA0012 profile and a sweep angle of 20 deg at the leading-edge line. The root chord is 2b = 0.7143 m,
while the wing semispan is 1.5 m (or 4.2b), with a taper ratio of 0.47. A rectangular unstructured computational grid is
build around the wing. The dimensions of the computational domain are Dx × Dy × Dz = 22.4b × 16.8b × 16.8b, where x is
positive in the stream-wise direction, while the y direction is placed along the span of the wing. The vertical displacement
of the wing is along the z axis. The aerodynamic surface mesh of the model consists of 54500 boundary elements and is
shown in Fig. 4(a). The total grid size was about 230000 tetrahedral volume elements. The specific mesh size has been
chosen after a mesh convergence study on steady-state Euler calculations. The surface pressure distribution along the
wing is shown in Fig. 4(b) for Ma = 0.8 and zero angle of attack.
A FE structural model is also constructed for this study. The solid wing of Fig. 5 is considered, with the structure
modeled with 3100 3D solid elements. Constant structural properties are considered, with density ρS = 347 kg/m3 ,
Young’s modulus E = 5 · 107 Pa, and Poisson’s ratio ν = 0.35. The mass and stiffness matrices for this FE model are built
using the native FE implementation in SU2 (Sanchez et al., 2017).
10
N. Simiriotis and R. Palacios Journal of Fluids and Structures 117 (2023) 103835

Fig. 5. Finite-element model of the flexible swept wing.

Fig. 6. Flexible swept wing structural modes interpolated onto the aerodynamic grid.

The modal description of the structural model is computed with these matrices by solving the associated eigenvalue
problem. The first four mass-normalized mode shapes are retained and shown in Fig. 6 after interpolation onto the
aerodynamic surface grid. The first and third mode are out-of-plane bending modes, the fourth mode is a torsion mode,
and the second mode displays in-plane bending. Their natural frequencies are 4.68, 18.52, 18.85, and 32.63 Hz.

4.2.1. Iterative eigenvalue search

Following the procedure described in Section 3.2, the GAF matrices are computed for this wing for Ma = 0.8 and
reduced frequencies in the range 0 < κ < 0.4. An HB-Euler calculation with one harmonic and imposed kinematics is
calculated for every actuation frequency. For every frequency, each mode is actuated individually with a given generalized
amplitude. The aerodynamic pressure forces produced with the simulations are then projected on the structural modes
to produce the generalized forces.
Freestream boundary conditions are set at the boundaries of the computational domain, and slip (Euler) wall conditions
on the wing. The wing is clamped on its root on the xz plane, which is given symmetry boundary conditions. The JST central
scheme with second and fourth order dissipation coefficients was employed for this study as well. About 1000 iterations
were necessary for the flow convergence with a tolerance of 10−15 .
To assess the sensitivity to the amplitude of the excitation, three sets of generalized forces have been produced for
three different amplitude levels on the mass-normalized modes, namely small (As ), medium (Am ), and large (Al ). They are
shown in Table 2, which includes both the generalized displacement given to each mode and the resulting maximum
nodal displacement in the finite-element model. The resulting generalized forces have been normalized by the dynamic
pressure 12 ρ∞ U∞ 2
and the amplitude of the mode that is actuated in each case, and are shown in Fig. 7.
By closely examining Fig. 7, some remarks can be made. As one would expect, the effect of the second mode,
corresponding to the in-plane motion, is much less significant. The respective force components are one to two orders of
magnitude smaller compared to the rest of the force components. For all components of the GAF matrix, the small and
medium levels of amplitude produce very similar results. This signifies a convergence for the choice of the amplitude
towards linear dynamics. For larger values, the real part is displaced in most cases by an offset, while the imaginary
11
N. Simiriotis and R. Palacios Journal of Fluids and Structures 117 (2023) 103835

Fig. 7. Generalized aerodynamic forces for the flexible swept wing, for varying modal amplitude (As with dashed lines, Am with solid lines, and Al
with dash-dotted lines).

Table 2
p–k flutter prediction for different amplitude levels (in parenthesis, the maximum nodal displacement normalized by
the semispan).
Modes 1 2 3 4 ωF (Hz) VF (m/s)
As /b 0.028 0.028 0.028 0.0056 57.6 368.5
(0.5%) (0.1%) (0.6%) (0.1%)
A /b
m
0.14 0.14 0.14 0.028 58.1 373.5
(2.5%) (1.6%) (3.0%) (0.5%)
Al /b 0.70 0.70 0.70 0.28 65.0 423.5
(12.3%) (10.6%) (15.1%) (2.7%)

part is mostly affected at the higher frequencies. As expected, the imaginary part of the force converges towards a zero
value at low frequencies, except for the in-plane mode when the amplitude increases. This can be justified since the
geometry differs significantly from the reference (undeformed) shape for larger amplitudes. The forces are computed
with the normal vectors of the deformed shape and then projected onto the relevant mode, but the changes in geometry
affect the projection of forces onto the in-plane mode.
Next, the CFD-produced data are interpolated to construct the functions that will be used in the p–k analysis. The
linear (1D) interpolation is carried out using the modified Akima cubic Hermite interpolation (Akima, 1974), which results
12
N. Simiriotis and R. Palacios Journal of Fluids and Structures 117 (2023) 103835

Fig. 8. Aeroelastic eigenvalues of the flexible swept wing for increasing values of the airspeed U∞ . Results from p–k method and GAFs with amplitude
As .

in a piece-wise function of third-order polynomials. Using the interpolated forces, the generalized nonlinear eigenvalue
problem (19) is solved as a function of the airspeed U∞ . Results are presented in Fig. 8, where the natural frequency of
the third mode is used to normalize the results.
The variation of the real part of λ is shown in Fig. 8(a). The aeroelastic modes are ranked by frequencies and identified
with roman numbers, with I being the mode stemming from the first bending mode at U∞ = 0, II from the in-plane
mode, III from the second out-of-plane mode, and IV the torsion mode. Mode I crosses the x axis at U∞ /c ω3 = 4.33 with
c = 2b, signaling flutter at 368.5 m/s. Soon after that the in-plane vibrations (which remain mostly unaffected by the
aerodynamics) become mode III. The slowly increasing damping of mode IV is reversed after the flutter onset.
The imaginary part of λ is shown in Fig. 8(b), and can be understood the modification of the natural modes of vibration
of the structure, under the influence of the aerodynamics. The frequency of mode IV decreases continuously with the
increase of the freestream airspeed, while that of mode I increases up to the flutter onset. A flutter frequency of 57.3 Hz
is obtained. When only the first bending and first torsion mode were considered in the p–k analysis, a flutter onset is
computed at a slightly modified airspeed, indicating that this is the dominant mechanism. Removing the in-plane bending
mode from the analysis, has no effect on flutter.
The p–k analysis was also repeated based on the forces obtained for medium Am and large Al levels of amplitudes, and
the resulting flutter onset predictions are included in Table 2. From Fig. 7, it can be seen that the out-of-plane bending
modes show little sensitivity to the amplitude of excitation up to values comparable to the semichord. This is not the case
for the torsional mode, since changes on the angle of attack do have a significant effect on shock wave dynamics, which
results in a significant reduction of the negative aerodynamic pitching moment about the elastic axis. This is manifested
in the flutter prediction, which is only mildly changed (≈ 1%) when moving from the small to medium amplitudes, for
which the amplitude of the torsional mode is about 3%. In the final case, when that amplitude has further increased by
another order of magnitude, the flutter prediction starts to diverge, although the discrepancy is only of 15% for the flutter
speed (compared to the results with As ), thus demonstrating rather robust behavior to the choice of modal amplitudes in
the generation of the GAF matrices.

4.2.2. Rational interpolation

Next, the small-amplitude CFD samplings are interpolated with the rational Loewner function. Fig. 9 shows the
generalized aerodynamic forces predicted by the Loewner rational interpolation model, which are compared to the original
sampled frequencies obtained from the harmonic balance calculation. No new sampling frequencies are added from those
used for the p–k to limit the computational effort. The implication is that there are small errors between the interpolated
values at the samples, which are mainly driven by the extrapolation that occurs at the higher frequencies. As we will see
next, this is sufficient to obtain a very good approximation for the flutter speed.
Following the procedure described in Section 3.3, a closed loop linear system is built at each velocity and its stability
is assessed. The resulting eigenvalues are shown in Fig. 10(b), which can be compared to the ones obtained by the p–k
method in Fig. 10(a). First, note that there are additional poles in the rational interpolation, which correspond to the
internal states in the identified LTI aerodynamic model. In the region of interest, where the system dynamics occur, the
aerodynamic poles are located along the x axis, although some high-frequency complex poles are also present and are not
shown. Importantly, Loewner’s rational interpolation on a limited frequency range is not stability preserving (Karachalios
et al., 2021) and some of those poles are indeed unstable in the current simulation. That does not hinder here the
identification of the physical solution in this case, but may need to be further investigated for a generic solver (Goizueta
et al., 2022a).
13
N. Simiriotis and R. Palacios Journal of Fluids and Structures 117 (2023) 103835

Fig. 9. Rational interpolation of the aerodynamic force matrices; data coming from the HB/CFD calculations are shown with circle (imaginary) and
square (real) markers, while solid and dashed lines are used for the Loewner matrices.

Fig. 10. Root locus plots of aeroelastic stability produced from the p–k method and closed-loop stability after rational interpolation for airspeed
varying between 150 m/s (squares) and 500 m/s (circles in increments of 50 m/s).

14
N. Simiriotis and R. Palacios Journal of Fluids and Structures 117 (2023) 103835

Fig. 11. Bifurcation diagrams with all three components of the modal displacement and vertical-only modal displacement.

The linear model derived from rational interpolation predicts flutter at VF = 371 m/s, which is slightly higher than
the p–k result of 368.5 m/s. This small difference is explained by the approximation introduced in identification of the LTI
aerodynamic system. The main difference between both methods is that the rational interpolation only includes physical
damping from the aerodynamic system, while a fictious damping is introduced by the p–k solver (its only physical solution
is, by construction, at the flutter speed(s)). The implication of that is different evolution of the eigenvalues away from the
flutter condition, which is apparent, in this case, in the crossing between the third and fourth modes.

4.3. Direct flutter solution

The flutter onset conditions can also be finally directly sought using the methods of Section 2.1. This approach does not
need pre-sampling of the CFD, thus accelerating the search (for a given accuracy), at the expense of a lack of insight of the
dynamic response of the wing below the flutter speed. However, it also characterizes post-flutter LCO thus characterizing
the dynamic features above the flutter speed.
To demonstrate this, the direct solution algorithm has been employed to construct the bifurcation diagram for the
flexible swept wing at Ma = 0.8. Solutions are based on one harmonic, as the focus is on asymptotic stability, but this
still gives relevant insight into the low-to-moderate amplitude LCOs at a minimum computational cost. In particular,
the first vibration mode q1 was chosen to be constrained in the search of the LCO frequency ω and reduced velocity
target
Ṽ = U∞ /ω3 c, where ω3 is the natural frequency of the second bending mode. Decreasing values of imposed q1 were
given as an input to the solver for the calculations.
The bifurcation diagram is shown in Fig. 11. For each point in the curves, the displacement vector was initialized with
target
the imposed generalized displacement q1 for the first bending, and zero generalized displacements for the rest of the
modes. No initial phase difference between the modes was employed. The initial frequency was chosen at ω0 /ω3 ≈ 0.5,
which is between the first and second natural frequency of the structure, and the initial reduced velocity was set at
Ṽ0 = 0.15. The sensitivity of the prediction on the values Ṽ0 and ω0 was also evaluated and the procedure was found to
be very robust with respect to the initial conditions.
As before, an Euler solver with the JST central scheme with second and fourth order dissipation coefficients was
employed for the flow. About NCFD = 40 − 50 iterations were carried out with the fluid solver in every aeroelastic iteration,
while four times more NCFD iterations were carried out for the initialization of the HB flow fields. The convergence of
the complete procedure to a very tight tolerance was achieved with about NAERO = 600 − 800 aeroelastic iterations.
To accelerate convergence, the pseudo time-step for the structure has been set at τS = 15 − 20. At every aeroelastic
computation, the inlet pressure is updated based on the ṼF prediction, while the freestream density ρ∞ = 0.09 kg/m3
and the mass ratio µ = 3856 are fixed to keep the mass of the wing constant. As no pre-sampling is needed, the simulation
cost does not increase with the number of modes (except for perhaps a slower convergence rate overall) and therefore
the in-plane bending mode was also retained in the analysis.
Fig. 11 shows the evolution of the normalized velocity and frequency of the LCOs as a function of the imposed modal
amplitude |q1 |. If only the vertical component (i.e., normal to the plane of the wing) of the displacement is used in the
natural modes of the structure, the predicted flutter speed is VF = 369.2 m/s. The figure also includes the p–k solution
from the previous section (VF =368.5 m/s), with the small difference being within the tolerance in the p–k computation.
Analysis of the modal components of the LCOs determines the flutter mechanics, which is again found to be due to the
first bending/torsion coupling.
15
N. Simiriotis and R. Palacios Journal of Fluids and Structures 117 (2023) 103835

Fig. 12. Bifurcation diagram produced with SU2 full modal displacements for a zero (C0) and a non-zero (C1) angle of attack, α0 .

Table 3
Flutter onset prediction comparison between direct and sampling-driven methods. 1D/3D: only vertical
displacements/all three components are transferred.
Method p–k Loewner Direct, 1D Direct, 3D Direct, 3D
α0 = 0 α0 = 0 α0 = 1.5 deg
ωF (Hz) 57.6 57.5 57.6 57.3 57.7
VF (m/s) 368.5 371 369.2 367.1 377.3

Possibly the main advantage of the direct solver is that it allows for sensitivity and parametric studies at a minimal
computational cost, as one can use the previously computed velocity/frequency pairs as initial guesses for the new
simulations. Two such studies are presented next as examples.
In the first example the three components of the displacement are considered for the mode shapes. As solid elements
have been used to built the structural model, this includes some additional resolution in the characterization of flutter
and the incipient LCO response. In particular, the in-plane bending mode is now fully represented, as well as Poisson
effects in the solid deformations. Results are also included in Fig. 11, and show a small but non-negligible impact on the
stability results. This is because using the full 3-D wing displacements slightly increases the aerodynamic forces and leads
to a lower flutter speed. Analysis of the modal contribution at the LCO solutions shows that the in-plane mode does not
participate in the instability. Importantly, these results are obtained with only a few additional CFD runs per point.
The wing considered here has a symmetric airfoil and did not produce steady lift in the results so far. In the second
example the angle of attack of the wing is set up at α0 = 1.5 deg. This produces non-symmetric forces and the solution
also includes now a mean displacement of the wing. Deformations are however still small and they are captured by a
linear structural model. The bifurcation diagram has been obtained around this new mean deformation and it is shown in
Fig. 12. Results indicate that around this new angle of attack the wing presents a subcritical bifurcation, thus highlighting
that relatively small changes in the equilibrium conditions can significantly affect flutter prediction (Goizueta et al.,
2022b). The flutter onset velocity is increased (see also Table 3); however, LCO can appear at airspeeds below the flutter
limit. Furthermore, the increase of the aerodynamic forces also leads to a significant increase of the oscillation amplitude
compared to the α0 = 0 case.
All flutter predictions in this section are summarized in Table 3 that attests to the accuracy of the solution procedure.

5. Conclusions

A harmonic-balanced algorithm for aeroelastic applications has been implemented in the SU2 open-source fluid solver,
and three strategies for predicting dynamic aeroelastic instabilities based on that aerodynamic solver have been explored
in some detail. Even though all the solution strategies yield virtually the same flutter predictions, there are some relative
advantages and disadvantages of each method.
Sampling-based approaches have been exemplified here by the p–k algorithm. It is seen to be a simple and robust
strategy, which justifies its widespread adoption. Employing the HB approach to construct the GAF matrices, even though
efficient compared to unsteady time-marching simulations, still entails significant computational cost. As one harmonic
is used, the cost of each HB calculation is about that of three steady CFD simulations. The sampling needs to be done
for each mode shape retained in the structural model and for the range of frequencies and Mach numbers of interest,
although it can be accelerated with strategies that simultaneously excite multiple modes (Silva, 2018). This approach is
16
N. Simiriotis and R. Palacios Journal of Fluids and Structures 117 (2023) 103835

ideally suited for exploration of a new configuration, as it produces a detailed map of the evolution of frequency and
damping of the aeroelastic system with the velocity.
We have also seen that flutter predictions using this approach are rather robust to the detailed choice of modal
amplitudes in the computation of the generalized aerodynamic forces. A drawback of the p–k and similar approaches
is however that the introduce non-physical damping into the problem. Even though this has no effect on the flutter
prediction (where that damping is zero), it may result in erroneous interpretations of the dynamic features under stable
conditions. This can be overcome by introducing a rational interpolation of the unsteady aerodynamic data set. In this
paper, this has been exercised here by means of the Loewner matrix, which produces a minimal data-driven dynamical
system model from the known frequency-response GAF samples. The aeroelastic model is then the feedback system
resulting from the LTI state–space models of the aerodynamic and structural subsystem, whose stability is determined
by its closed-loop eigenvalues. An additional advantage of the approach is that the resulting linear model is suitable for
the standard methods for control design. As with other SVD-based approaches, the main drawback of this strategy is
that it is not stability preserving, which may result in spurious flutter solutions by automatic search algorithms. Several
strategies for stabilization are readily available (Gosea and Antoulas, 2016; Yue and Zhao, 2020), although they have not
been necessary for this work.
A direct solution approach, which seeks oscillatory solutions of the coupled system, has also been investigated. As
opposed to the other two strategies, it does not need a training set and therefore can potentially result in much faster
computational times. The downside is that this strategy does not give information of the dynamic response of the system
as it approaches flutter (although as we have seen, it can be used to characterize the post-flutter response). In the current
implementation, the updates are based on a partial Jacobian for simplicity of evaluation, which results in relatively slow
convergence for poor initial guesses. Instead of being used for ‘‘blind’’ sampling, a strength of the direct approach is in
the evaluation of parametric dependencies of flutter with changes in the design, as it has been shown in the examples
above. In practice, a p–k type method would be used to identify the flutter characteristics in a baseline configuration, with
the direct solution then being used to perform parametric studies. Furthermore, the implementation of the direct search
algorithm is easily differentiable and in a follow-on project we are using the algorithmic differentiation features in SU2
to obtain flutter derivatives in a high-fidelity multi-disciplinary optimization context.
An additional advantage of the direct approach is that the number of CFD computations remains mostly independent of
the complexity of the design space. This includes the number of vibration modes needed to characterize the structure, but
also the search for matched solutions, in which the Mach (and possibly Reynolds) number in the flow solutions correspond
to the flutter conditions, as well as the geometry of the wing in aeroelastic equilibrium. All those variables can be updated
in each iteration of the direct algorithm. Finally, the direct method can easily explore postflutter behavior, even with a
single harmonic (although more can be then included, at a cost), as shown in the last set of results.
We finally note that in this work we have not linearized the Navier–Stokes equations and instead used a nonlinear
solver with small-amplitude excitations. While this clearly brings a computational penalty, it can be substantially
mitigated by picking suitably small inputs and solution restarts, with the big advantage that it enables the exploitation, in
aeroelastic stability analysis, of the full range of capabilities of fluid solver (here SU2, which includes multiple turbulent
models, engine models, or algorithmic differentiation, among others).

Declaration of competing interest

The authors declare the following financial interests/personal relationships which may be considered as potential
competing interests: Nikolaos Simiriotis reports financial support was provided by Commercial Aircraft Corporation of
China Ltd.

Data availability

Data will be made available on request.

Acknowledgments

This work has received funding from the Clean Sky 2 Joint Undertaking, under the European’s Union Horizon 2020
research and innovation Program under Grant Agreement number: 883670 - H2020 -CS2-CFP102019-01.

References

Akima, H., 1974. A method of bivariate interpolation and smooth surface fitting based on local procedures. Commun. ACM 17 (1), 18–20.
http://dx.doi.org/10.1145/360767.360779.
Antoulas, A.C., 2005. Approximation of Large-Scale Dynamical Systems. Society for Industrial and Applied Mathematics, Philadelphia, Pennsylvania,
USA, http://dx.doi.org/10.1137/1.9780898718713.
Badcock, K., Timme, S., Marques, S., Khodaparast, H., Prandina, M., Mottershead, J., Swift, A., Da Ronch, A., Woodgate, M., 2011. Transonic aeroelastic
simulation for instability searches and uncertainty analysis. Prog. Aerosp. Sci. 47 (5), 392–423. http://dx.doi.org/10.1016/j.paerosci.2011.05.002.
Beckert, A., Wendland, H., 2001. Multivariate interpolation for fluid-structure-interaction problems using radial basis functions. Aerosp. Sci. Technol.
5 (2), 125–134. http://dx.doi.org/10.1016/S1270-9638(00)01087-7.

17
N. Simiriotis and R. Palacios Journal of Fluids and Structures 117 (2023) 103835

Cardani, C., Mantegazza, P., 1978. Continuation and direct solution of the flutter equation. Comput. Struct. 8 (2), 185–192. http://dx.doi.org/10.1016/
0045-7949(78)90021-4.
Dimitriadis, G., 2017. Introduction to Nonlinear Aeroelasticity. John Wiley & Sons, Ltd, http://dx.doi.org/10.1002/9781118756478.
Economon, T.D., Palacios, F., Copeland, S.R., Lukaczyk, T.W., Alonso, J.J., 2016. SU2: An open-source suite for multiphysics simulation and design. AIAA
J. 54 (3), 828–846. http://dx.doi.org/10.2514/1.J053813.
Géradin, M., Rixen, D., 1997. Mechanical Vibrations – Theory and Applications to Structural Dynamics, second ed. John Wiley and Sons, Inc..
Goizueta, N., Wynn, A., Palacios, R., 2022a. Adaptive sampling for interpolation of reduced-order aeroelastic systems. AIAA J. http://dx.doi.org/10.
2514/1.J062050, In print.
Goizueta, N., Wynn, A., Palacios, R., Drachinsky, A., Raveh, D.E., 2022b. Flutter predictions for very flexible wing wind tunnel test. J. Aircr. 59 (4),
1082–1097. http://dx.doi.org/10.2514/1.C036710.
Gopinath, A., Jameson, A., 2005. Time spectral method for periodic unsteady computations over two- and three- dimensional bodies. In: 43rd AIAA
Aerospace Sciences Meeting. Reno, Nevada, USA, http://dx.doi.org/10.2514/6.2005-1220.
Gosea, I.V., Antoulas, A.C., 2016. Stability preserving post-processing methods applied in the Loewner framework. In: 20th Workshop on Signal and
Power Integrity. IEEE, Turin, Italy, http://dx.doi.org/10.1109/SaPIW.2016.7496283.
Hall, K.C., Thomas, J.P., Clark, W.S., 2002. Computation of unsteady nonlinear flows in cascades using a harmonic balance technique. AIAA J. 40 (5),
879–886. http://dx.doi.org/10.2514/2.1754.
Hassig, H.J., 1971. An approximate true damping solution of the flutter equation by determinant iteration. J. Aircr. 8 (11), 885–889. http:
//dx.doi.org/10.2514/3.44311.
He, S., Jonsson, E., Mader, C.A., Martins, J.R.R.A., 2019. A coupled Newton-Krylov time spectral solver for wing flutter and LCO prediction. In: AIAA
Aviation. Dallas, Texas, USA, http://dx.doi.org/10.2514/6.2019-3549.
Henshaw, M., Badcock, K., et al., 2007. Non-linear aeroelastic prediction for aircraft applications. Prog. Aerosp. Sci. 43 (4–6), 65–137. http:
//dx.doi.org/10.1016/j.paerosci.2007.05.002.
Jonsson, E., Riso, C., Lupp, C.A., Cesnik, C.E.S., Martins, J.R.R.A., Epureanu, B.I., 2019. Flutter and post-flutter constraints in aircraft design optimization.
Prog. Aerosp. Sci. 109 (100537), http://dx.doi.org/10.1016/j.paerosci.2019.04.001.
Karachalios, D.S., Gosea, I.V., Antoulas, A.C., 2021. Chapter 6, The Loewner framework for system identification and reduction. In: System- and
Data-Driven Methods and Algorithms. De Gruyter, pp. 181–228. http://dx.doi.org/10.1515/9783110498967-006.
Kholodar, D.B., 2022. Aircraft flutter and aerodynamic work. J. Aircr. http://dx.doi.org/10.2514/1.C036846, In print.
Lee-Rausch, E.M., Batina, J.T., 1995. Wing flutter boundary prediction using unsteady Euler aerodynamic method. J. Aircr. 32 (2), 416–422.
http://dx.doi.org/10.2514/3.46732.
Li, H., Ekici, K., 2017. Revisiting the one-shot method for modeling limit cycle oscillations: Extension to two-degree-of-freedom systems. Aerosp. Sci.
Technol. 69, 686–699. http://dx.doi.org/10.1016/j.ast.2017.07.037.
Li, H., Ekici, K., 2018a. Improved one-shot approach for modeling viscous transonic limit cycle oscillations. AIAA J. 56 (8), 3138–3152. http:
//dx.doi.org/10.2514/1.J056969.
Li, H., Ekici, K., 2018b. A novel approach for flutter prediction of pitch–plunge airfoils using an efficient one-shot method. J. Fluids Struct. 82, 651–671.
http://dx.doi.org/10.1016/j.jfluidstructs.2018.08.012.
Li, H., Ekici, K., 2019. Aeroelastic modeling of the AGARD 445.6 wing using the harmonic-balance-based one-shot method. AIAA J. 57 (11), 4885–4902.
http://dx.doi.org/10.2514/1.J058363.
Livne, E., 2018. Aircraft active flutter suppression: State of the art and technology maturation needs. J. Aircr. 55 (1), 410–452. http://dx.doi.org/10.
2514/1.C034442.
McMullen, M., Jameson, A., 2006. The computational efficiency of non-linear frequency domain methods. J. Comput. Phys. 212 (2), 637–661.
http://dx.doi.org/10.1016/j.jcp.2005.07.021.
Quero, D., Vuillemin, P., Poussot-Vassal, C., 2021. A generalized eigenvalue solution to the flutter stability problem with true damping: The p-L
method. J. Fluids Struct. 103, 103266. http://dx.doi.org/10.1016/j.jfluidstructs.2021.103266.
Roger, K.L., 1977. Airplane Math Modelling and Active Aeroelastic Control Design. Conference proceedings, AGARD-CP-228.
Rubino, A., Pini, M., Colonna, P., Albring, T., Nimmagadda, S., Economon, T., Alonso, J., 2018. Adjoint-based fluid dynamic design optimization in
quasi-periodic unsteady flow problems using a harmonic balance method. J. Comput. Phys. 372, 220–235. http://dx.doi.org/10.1016/j.jcp.2018.06.
023.
Sanchez, R., Albring, T., Palacios, R., Gauger, N., Economon, T., Alonso, J., 2017. Coupled adjoint-based sensitivities in large-displacement fluid-structure
interaction using algorithmic differentiation. Int. J. Numer. Methods Eng. 113 (7), http://dx.doi.org/10.1002/nme.5700.
Silva, W., 2018. AEROM: NASA’s unsteady aerodynamic and aeroelastic reduced-order modeling software. Aerospace 5 (2), http://dx.doi.org/10.3390/
aerospace5020041.
Silva, W.A., Bartels, R.E., 2004. Development of reduced-order models for aeroelastic analysis and flutter prediction using the CFL3Dv6.0 code. J.
Fluids Struct. 19 (6), 729–745. http://dx.doi.org/10.1016/j.jfluidstructs.2004.03.004.
Simiriotis, N., Palacios, R., 2022. A flutter prediction framework in the open-source SU2 suite. In: AIAA Science and Technology Forum and Exposition.
San Diego, California, USA, http://dx.doi.org/10.2514/6.2022-1327.
Thomas, J.P., Dowell, E.H., 2018. A fixed point iteration approach for harmonic balance based aeroelastic computations. In: AIAA Scitech. Kissimmee,
Florida, USA, http://dx.doi.org/10.2514/6.2018-1446.
Thomas, J.P., Dowell, E.H., Hall, K.C., 2002. Nonlinear inviscid aerodynamic effects on transonic divergence, flutter, and limit-cycle oscillations. AIAA
J. 40 (4), 638–646. http://dx.doi.org/10.2514/2.1720.
Thormann, R., Widhalm, M., 2013. Linear-frequency-domain predictions of dynamic-response data for viscous transonic flows. AIAA J. 51 (11),
2540–2557. http://dx.doi.org/10.2514/1.J051896.
Venkatesan-Crome, C., Palacios, R., Kattmann, T., Sanchez, R., Gauger, N.R., Economon, T.D., 2019. Discrete adjoint for unsteady incompressible flows
using a density-based formulation. In: AIAA Aviation. Dallas, Texas, USA, http://dx.doi.org/10.2514/6.2019-3669.
Yates, E.C., 1987. AGARD standard aeroelastic configurations for dynamic response. Candidate configuration I: Wing 445.6. Tech. rep., NASA, NASA
TM 100492.
Yates, E.C., Land, N.S., Foughner Jr., J.T., 1963. Measured and calculated subsonic and transonic flutter charactersitscs of a 45 deg sweptback wing
planform in air and in Freon-12 in the Langley Transonic Dynamics Tunnel. Tech. rep., NASA, NASA TN D-1616.
Yue, C., Zhao, Y., 2020. An improved aeroservoelastic modeling approach for state-space gust analysis. J. Fluids Struct. 99, 103148. http://dx.doi.org/
10.1016/j.jfluidstructs.2020.103148.

18