Progress in Aerospace Sciences 109 (2019) 100537

Contents lists available at ScienceDirect

Progress in Aerospace Sciences

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

Flutter and post-flutter constraints in aircraft design optimization T

a,∗ a a a
Eirikur Jonsson , Cristina Riso , Christopher A. Lupp , Carlos E.S. Cesnik ,
Joaquim R.R.A. Martinsa, Bogdan I. Epureanub
a
Department of Aerospace Engineering, University of Michigan, Ann Arbor, MI, 48109, USA
b
Department of Mechanical Engineering, University of Michigan, Ann Arbor, MI, 48109, USA

ARTICLE INFO ABSTRACT

Keywords: Flutter is a dynamic aeroelastic instability driven by the interaction of inertial, elastic, and aerodynamic forces. It
Flutter is an undesirable phenomenon in aircraft because it causes divergent oscillations that may lead to structural
Post-flutter damage or failure, performance and ride comfort degradation, or loss of control. If flutter is discovered at the
Limit cycle oscillations aircraft certification stage, costly redesign is required. Performing flutter analysis early in the design process can
Aircraft design
mitigate this problem. Furthermore, including flutter analysis as a constraint in multidisciplinary design opti-
Multidisciplinary design optimization
mization reduces the risk of costly modifications late in the design cycle. We review the methods for flutter
analysis in the context of aircraft design optimization. We also include methods for predicting post-flutter limit
cycle oscillations due to the increasing impact of nonlinear effects on future aircraft. While there has been
extensive work in flutter and post-flutter analyses, developing design optimization constraints associated with
these analyses has additional requirements, such as acceptable computational cost, function smoothness, ro-
bustness, and derivative computation. We discuss these requirements and review efforts in the development,
implementation, and application of flutter and post-flutter constraints in aircraft design optimization. We con-
clude the paper by summarizing the current state of this field and the main open problems. Flutter constraints
have been included in structural optimizations, but optimizing both the structural sizing and the aerodynamic
shape remains a challenge due to the need to recompute the aerodynamic properties at each design iteration.
Additional difficulties arise in the presence of large structural deflections and transonic flow conditions due to
the dependency of the flutter point on the equilibrium state and the high cost of nonlinear computations. Post-
flutter constraints have rarely been included into design optimization, but they are crucial in the prevention of
undesirable limit cycle oscillations. Implementing such constraints requires the development of more efficient
and robust prediction methods that can handle realistic configurations. While this paper focuses on flutter and
post-flutter constraints for aircraft design optimization applications, the considerations and challenges are
broadly applicable to the optimization of engineering systems including stability and post-critical dynamic
constraints.

1. Introduction design in particular. Such a process would not only shorten the design
cycle, but also allow for advantageous design trade-offs between the
Flutter is a dynamic aeroelastic instability that causes divergent flutter requirements, the other constraints, and the aircraft perfor-
oscillatory vibrations [1]. It is an undesirable phenomenon in aircraft mance.
because it can cause structural damage or failure, performance and ride Performing multidisciplinary design optimization (MDO) that con-
comfort degradation, or loss of control. Flutter computations are typi- siders both aerodynamic shape and structural sizing simultaneously
cally performed only after an initial detailed design of the aircraft is while enforcing flutter constraints is a way to address this issue [3,4].
completed, because they require the vehicle stiffness, mass, and aero- Structural optimization alone, even if including aerostructural analyses
dynamic models to be available [2]. If the design does not satisfy the for enforcing flutter constraints, yields design solutions with suboptimal
flutter requirements at this stage, a redesign is necessary, which adds performance compared to the optimal designs resulting from MDO,
costs and delays to the aircraft development cycle. Thus, it is desirable where structural and aerodynamic sizing variables are optimized si-
to consider flutter concurrently with the aircraft design and the wing multaneously [4,5]. MDO can minimize structural weight, fuel

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

https://doi.org/10.1016/j.paerosci.2019.04.001
Received 8 October 2018; Received in revised form 1 April 2019; Accepted 4 April 2019
Available online 31 May 2019
0376-0421/ © 2019 Elsevier Ltd. All rights reserved.
E. Jonsson, et al. Progress in Aerospace Sciences 109 (2019) 100537

Nomenclature KS Kreisselmeier–Steinhauser
LCO Limit cycle oscillation
AD Automatic differentiation LFD Linear frequency domain
BFF Body-freedom flutter LTI Linear time invariant
BWB Blended wing-body MDO Multidisciplinary design optimization
CFD Computational fluid dynamics NC Numerical continuation
CNK Coupled Newton–Krylov NP Nonlinear perturbation
CSD Critical slowing down OML Outer mold line
DLM Doublet lattice method POD Proper orthogonal decomposition
DOF Degree of freedom RANS Reynolds-averaged Navier–Stokes
DV Design variable RFA Rational function approximation
EV Eigenvalue ROM Reduced-order modeling
FEM Finite element method TC Time cyclic
GAF Generalized aerodynamic force TD Time domain
GB Gradient-based TES Transonic equivalent strip
GF Gradient-free TM Time marching
HALE High-altitude long-endurance TS Time spectral
HARW High-aspect-ratio wing TSD Transonic small disturbance
HB Harmonic balance TBW Truss-braced wing
HOHB High order harmonic balance uCRM undeflected Common Research Model
IFT Implicit function theorem

consumption, or a combination of these two objectives with respect to practical aircraft configurations parameterized with a large number of
wing shape, internal structure arrangements, and sizing, while ac- design variables [11]. When using gradient-based algorithms, it is im-
counting for the interactions between aerodynamics, structures, and portant to consider the smoothness of the objective and constraint
other disciplines, and satisfying various constraints. MDO with flutter functions, as well as the accuracy and efficiency of the derivative
constraints results in designs with optimal aeroelastic tailoring. Omit- computations.
ting flutter constraints in the MDO process when minimizing fuel con- Optimizing HARW configurations subject to flutter constraints is
sumption tends to yield light-weight, high-aspect-ratio wing (HARW) even more challenging because it requires capturing couplings between
designs that despite being highly efficient may not be feasible [6,7]. aeroelasticity and flight dynamics along with geometric nonlinearities
After the aircraft has been designed and a prototype has been built, that arise in the presence of low natural vibration frequencies and
certification requires flight tests to demonstrate that the aircraft be free significant structural flexibility [12,13].
from flutter in the flight envelope with a 15% safety margin beyond the Nonlinearities in the structure (large deflections, free-play of control
dive speed. If flutter is discovered at the flight test certification stage, it surfaces, follower loading) or the aerodynamics (shock waves and flow
requires redesign to address it, incurring additional costs. The redesign separation) can cause self-sustained oscillations of limited amplitude
effort typically increases the structural weight, reducing the perfor- that remain constant in time, known as limit cycle oscillations (LCOs).
mance originally anticipated for the aircraft. For certain types of nonlinearities, LCOs may exist at flight conditions
The trend towards HARW aircraft is driven by better fuel efficiency, below the flutter point [14]. When nonlinear effects become important,
but their increased flexibility makes it all the more important to con- post-flutter analysis should be integrated into the design process in the
sider flutter accurately and early in the design process [8]. Another form of constraints to make sure that the optimal design be feasible. In
recent trend is the increasing use of control surfaces to suppress flutter. the context of aircraft design practice, LCO typically refers to oscilla-
Active flutter suppression systems can be incorporated late in the design tions due to control-surface free-play nonlinearities. From a certifica-
process when aeroelastic instabilities are encountered and a passive tion perspective, these are typically addressed by computing free-play
solution such as redesign is impractical and expensive [9,10]. Alter- tolerances. In this paper, we use the term LCO to refer more generally to
natively, MDO provides a way to obtain the best possible configuration a post-flutter response due to structural or aerodynamic nonlinearities.
by co-designing the wing shape and internal structure, which contribute While free-play LCOs are a particular case of post-flutter response, the
to passive flutter suppression that can then be augmented with an ac- paper focuses on LCOs involving vehicle components or the entire
tive flutter control system. airframe. Taking into account these global LCOs is challenging in design
While there has been extensive work in methods for flutter analysis, optimization due to the high cost of nonlinear post-flutter computations
integrating flutter constraints into design optimization requires addi- and the difficulties in formulating post-flutter constraints.
tional considerations. Models used for flutter prediction should capture There have been several review papers and textbooks on flutter and
the relevant physics with adequate accuracy to correctly drive the op- post-flutter analysis. Livne [15,16] reviewed the state-of-the-art and
timizer and inevitably there is a compromise between model fidelity future challenges in aeroelasticity of conventional and unconventional
and computational cost. To include flutter analysis in a numerical op- vehicles. A recent review by the same author focused on active flutter
timization cycle, speed of execution is particularly important to make suppression control systems [10]. Friedmann [17] reviewed the general
sure that the overall optimization process is tractable. challenges in nonlinear aeroelasticity, where the applications focused
Another important characteristic for integrating flutter analysis into on rotary wings. Dowell et al. [14] classified nonlinear aeroelastic be-
the optimization process is the robustness of the flutter prediction haviors and discussed theoretical, computational, and experimental
method. Since the optimization process automatically samples the de- analysis efforts. de C. Henshaw et al. [18] discussed traditional in-
sign space, it is likely to request for the analysis of designs that would dustrial linear flutter prediction and recent efforts for including non-
normally not be chosen by a human designer. Thus, it is important that linear effects, particularly due to transonic flows. More recently, Afonso
the flutter analysis converges for such designs so that the overall opti- et al. [8] reviewed nonlinear aeroelasticity of HARWs. Dimitriadis [19]
mization process is not interrupted. discussed nonlinear post-flutter behaviors in aeroelastic systems and
Gradient-based optimization algorithms are needed to optimize the related analysis methods. Beran et al. [20] reviewed methods for

2
E. Jonsson, et al. Progress in Aerospace Sciences 109 (2019) 100537

Fig. 1. Summary of flutter and post-flutter analysis methods.

3
E. Jonsson, et al. Progress in Aerospace Sciences 109 (2019) 100537

uncertainty quantification in aircraft aeroelasticity and their applica- computation is sometimes the bottleneck in the optimization cycle.
tion to formulate nondeterministic optimization problems. However, When it comes to methods for computing gradients, the finite-dif-
the field lacks a review on the integration of flutter and post-flutter ference method is a popular choice because it is easy to implement and
analysis as constraints in aircraft design optimization. can always be used, even with black-box codes. The major drawbacks of
In this paper, we address this shortcoming by reviewing methods for the finite-difference method is that it is inaccurate and its computa-
flutter and post-flutter prediction, and we discuss their advantages and tional cost scales poorly with the number of design variables. Unlike the
disadvantages in the context of aircraft design optimization. First, we finite-difference method, the complex-step method [28] is accurate, but
provide a brief background on multidisciplinary design optimization its cost still scales unfavorably with the number of design variables,
(Section 2) and on flutter and post-flutter modeling (Section 3). These making it prohibitive for wing design applications. Furthermore, the
sections emphasize the key aspects relevant to aircraft design and are complex-step method cannot be directly applied to programs that use
beneficial for readers not familiar with either of these topics. Next, we complex numbers. Such programs need to be made complex-step
present numerical methods for flutter and post-flutter analysis, which compatible by modifying them so that the real and imaginary part of
are summarized in Fig. 1. These methods and examples of their appli- the complex number are represented as two real numbers. Another
cation in aircraft design optimization problems are discussed in Section option for computing gradients is automatic differentiation (AD), which
4 and 5. In reviewing previous work we focus on deterministic opti- uses a software tool to parse the code of an analysis to produce a new
mizations, but we also present few examples where constraints are code that computes derivatives of that analysis [29,30]. Although AD
formulated using failure probabilities (reliability-based design optimi- can scale well with the number of variables, it does not handle iterative
zation) [20]. The paper concludes with remarks on the state of this field simulations efficiently in general. Finally, analytic methods are the
and the open challenges to be addressed for integrating flutter and post- most desirable because they are both accurate and efficient, especially
flutter considerations into the optimal design of aircraft configurations. for iterative simulations [31]. However, they require significant im-
While the paper focuses on aircraft design optimization, the methods plementation effort. There are two main approaches within the analytic
and challenges reviewed are more broadly applicable to the design methods: the direct approach and the adjoint approach. The adjoint
optimization of engineering systems subject to stability and post-critical approach is especially attractive because the computational cost is de-
dynamic constraints. pendent of the number of outputs of interest (objectives and con-
straints) but independent of the number of design variables [31,32]. A
2. Background on multidisciplinary design optimization coupled-adjoint approach can solve static aeroelastic problems [5,6]
and can be generalized to any multidisciplinary problem [33,34].
Multidisciplinary design optimization couples the relevant dis- In the context of aircraft design optimization subject to flutter or
ciplines of an engineering system and performing numerical optimiza- post-flutter constraints, most of the early efforts used gradient-based
tion to aid the design of that system [21]. MDO considers several dis- optimization with gradients computed with either finite differences or
ciplines simultaneously such that their interactions can be leveraged, the direct analytic method. However, recent efforts implemented the
resulting in a better optimum than if each discipline were optimized more efficient adjoint approach, and some also used AD techniques.
sequentially [22]. Thus, considering MDO early in the design process These applications are further discussed in Section 4.2 for flutter con-
allows engineers not only to improve the design, but also to minimize straints and Section 5.3 for post-flutter constraints.
development time and cost of the overall design.
Performing MDO of aircraft configurations by describing its outer 3. Background on flutter and post-flutter modeling
mold line (OML) and structural sizing with high fidelity requires a large
number of design variables. Detailed aerodynamic optimization of For aircraft designs to be useful and practical, the underlying
wings requires hundreds of shape variables [23] and structural sizing models used in the flutter and post-flutter analysis need to capture the
based on a detailed finite-element wingbox model that is best utilized correct physics involved. However, a simplification of the phenomena is
with an equally large number of sizing variables [24]. Gradient-based often necessary to make problems tractable to solve. Therefore, the
optimization methods are the feasible way to solve for high-dimen- choice of model should balance the fidelity needed to obtain accurate
sional problems within a reasonable computational time, especially predictions and the mathematical or computational tractability for
when using high-fidelity analyses [11,25]. Gradient-based methods design applications.
require derivatives of the objective and constraint functions with re- This section highlights the modeling aspects to be considered in
spect to the design variables to help the optimization algorithm find the flutter and post-flutter analysis, which are discussed in more detail in
most promising search directions and establish rigorous optimality Section 4 and 5 for obtaining meaningful results in a design optimiza-
conditions. tion. By flutter, we mean the onset of divergent oscillations as the flight
While gradient-free algorithms are typically more robust and some conditions of aircraft cross the critical stability boundary (flutter
of them explore the design space more widely, their cost is prohibitive boundary).
when the number of design variables is large. Although gradient-based Mathematically, flutter occurs at a Hopf bifurcation point [35] be-
methods only guarantee convergence to a local optimum, this can be yond which the system is in the post-flutter regime. Several post-flutter
mitigated by using a multi-start technique [26]. Furthermore, recent behaviors are possible, as discussed in detail by Dimitriadis [19] for a
studies failed to find multiple local minima (multimodality) in airfoil two-dimensional aeroelastic system with stiffness and damping non-
and wing shape design optimization [11,23], except the case of plan- linearities. Among these behaviors, we are particularly interested in
form optimizations where expected local minima were found related to self-sustained oscillations with limited amplitude that remains constant
choices such as upwards or downwards winglets [27]. in time, known as LCOs.
The efficacy of gradient-based algorithms relies on accurate and LCOs typically develop beyond the flutter boundary; however, for
efficient gradient computations. The gradient accuracy directly affects certain types of nonlinearities, they can also occur before reaching the
the ability to converge to the optimum with a specified tolerance and flutter boundary [12]. Integrating post-flutter analyses into the design
the order of convergence of the optimization. In the best case, in- process can prevent this undesirable situation. This places additional
accurate gradients increase the number of iterations required for con- challenges due to the difficulties in modeling nonlinear aeroelastic
vergence and in the worst case cause early stopping due to numerical systems accurately and the inherent high cost of nonlinear aeroelastic
issues. Efficiency gradient computation is also important because this computations.

4
E. Jonsson, et al. Progress in Aerospace Sciences 109 (2019) 100537

3.1. Flutter modeling Many possible sources of nonlinearities can be present simulta-
neously in aeroelastic systems [8,14,19]. In this paper, we focus on
Flutter is defined as a self-exciting dynamic instability that is asso- aerodynamic nonlinearities due to transonic flow regimes and geo-
ciated with the interaction of inertial, elastic, and aerodynamic forces metric structural nonlinearities due to large deflections, both of which
[1]. At the onset of flutter, this aeroelastic instability can be physically are critical in the design of next-generation transport aircraft.
described as an oscillation with a small amplitude that is constant in Aerodynamic nonlinearities due to shock waves and flow separation
time triggered by a small-amplitude disturbance, as shown in Fig. 2. significantly impact the flutter speed. This decreases dramatically in the
The flight condition in which the system damping vanishes, resulting in transonic regime, a phenomenon known as the transonic dip [36–40]
this self-sustained oscillation, represents the flutter point (or flutter illustrated in Fig. 5. Low-order, linear unsteady aerodynamic models
boundary). For linear systems, the flutter point is defined as the commonly used in flutter analysis are in general accurate enough for
minimum dynamic pressure at which at least one of the modes becomes subsonic and supersonic flows, but they severely overestimate the
unstable [1]. The dynamic pressure can be replaced by equivalent air- flutter speed for transonic conditions [41–43].
speed, and is a function of altitude and Mach number. As shown in Fig. 5 for a hypothetical wing, linear theory is non-
Past the flutter point, in the absence of restraining nonlinearities conservative when compared to nonlinear viscous models. Nonlinear
from the aerodynamics, the structure, or both, the amplitude of the inviscid models based on Euler equations can capture shock waves but
oscillations grow exponentially. Fluid-structure interactions may also they still fail to accurately predict the flutter boundary [44]. In many
result in a static instability called divergence [1], which is not asso- cases, the nonlinear inviscid theory predicts a highly conservative
ciated with oscillations. As for flutter, the structural response grows flutter speed at the dip, even though it is generally closer to viscous
unbounded past the onset point, eventually reaching a limited-ampli- theory predictions. Depending on the severity of shocks, models based
tude oscillation if restraining nonlinearities are present. on Navier–Stokes equations (which include viscous and turbulence ef-
In the following discussion, we focus mainly on flutter phenomena, fects like boundary layer thickening, flow separation, and interactions
because for many practical configurations flutter occurs before diver- between shocks and regions of separated flow) are necessary to obtain
gence. However, accounting for divergence and the associated post- accurate flutter points [42]. Studies on various geometries demon-
critical response in the design process shares many of the modeling and strated that taking into account viscous phenomena in the transonic
analysis aspects associated with flutter. Furthermore, some of the flow regime improves the numerical prediction of transonic dip
analysis methods and constraints discussed in Section 4 are applicable [45–48].
to divergence as well as flutter. Moreover, in this paper we focus on A common approach to improve the accuracy of transonic flutter
global wing or component flutter rather than localized effects such as computations while minimizing the increase in computational cost is to
panel flutter that typically occurs at supersonic flow conditions. use numerical or experimental corrections applied to potential-flow
The possible flutter characteristics are illustrated in Fig. 3, which linear models [49]. However, the correction data may not be available
shows the variation of the modal damping with flight speed at a fixed for optimization, either because it requires high-fidelity computations
altitude. This is known as V g diagram, which is a classical tool used that are too expensive or because it is obtained from wind-tunnel
in linear flutter analyses for determining the flutter point and inter- measurements. This problem can be addressed by analyzing flutter
preting the flutter characteristics. A similar representation can be ob- using time-accurate dynamic simulations and higher-fidelity aero-
tained by varying dynamic pressure at a fixed Mach number. dynamic models. On the other hand, flutter prediction based on time-
Damping changes with flight speed in different ways among dif- accurate computational fluid dynamics (CFD) is a challenge even for
ferent designs, leading to different flutter behaviors. Soft flutter occurs just analyzing the final configuration and is currently prohibitive for
when the damping decreases gradually with increasing flight speed, design space exploration.
while hard flutter occurs when this decrease is abrupt. Another possi- Methods exist that try to preserve the computational efficiency of
bility is that there is a gradual decrease in damping with increasing
flight speed, all the way to cross the zero value and beyond, followed by
a damping increase. This phenomenon is known as a hump mode. These
concepts are important when considering how to formulate a smooth
and continuous flutter constraint for gradient-based optimization and
are discussed in Section 4.
In flutter analysis, the physics described above is often represented
by less expensive linear models due to the number of conditions to be
considered for certification. However, nonlinear structural and aero-
dynamic effects or the interaction between elastic and rigid-body de-
grees of freedom (DOF), which become important in the presence of
low structural vibration frequencies, may significantly impact the
flutter point. Therefore, the flutter prediction accuracy depends on the
appropriate modeling of nonlinear effects and boundary conditions.
Furthermore, nonlinear effects impact not only the models used in
flutter analysis, but also the analysis methods themselves. For linear
systems, flutter characteristics do not depend on the deformation state.
Therefore, flutter is typically analyzed by considering the unloaded and
undeformed structure. For nonlinear systems, stability characteristics
vary with the deformation configuration. Therefore, flutter analysis
must be performed by computing the eigenvalues of the aeroelastic
system linearized around equilibrium points for each flight condition to Fig. 2. Aeroelastic system response before and past the flutter point. Prior to
identify at what point the damping vanishes [12] as shown in Fig. 4. reaching the flutter point, the aeroelastic response is damped. At the flutter
The eigenvalues can be computed by considering both the elastic and point, the system response is an oscillation with a small constant amplitude.
rigid-body DOFs (flutter in free flight) or by retaining only the elastic Past the flutter point, a linear system response diverges, while a system with
DOFs (traditional flutter) or the rigid-body DOFs (flight dynamic sta- structural or aerodynamic nonlinearities develops a stable response with finite
amplitude that remains constant in time, known as LCO.
bility).

5
E. Jonsson, et al. Progress in Aerospace Sciences 109 (2019) 100537

Fig. 3. V g diagrams for different types of flutter. Soft flutter is a gradual loss
of damping with increased speed while hard flutter occurs abruptly and vio-
lently. A hump mode manifests itself as a damping decrease followed by an Fig. 5. Characteristic transonic dip of a transport wing. Aerodynamic non-
increase, which may result in considerably lower flutter speed. linearities due to shock waves and flow separation have a significant impact on
the flutter speed, which may decrease dramatically. Linear theory (e.g., DLM) is
non-conservative when compared to nonlinear viscous theory (e.g., RANS).
Nonlinear inviscid theory (e.g., Euler) predicts highly conservative flutter speed
at the dip, but it is generally closer to viscous theory predictions.

[52]. Furthermore, several efforts have applied the time-linearization

directly to CFD solvers. This approach is appealing since they retain
nonlinear effects, provide accurate results in the transonic regime, and
intrinsically account for geometric properties of the body such as
thickness and camber.
Kreiselmaier and Laschka [53] developed an unsteady method
based on the small-disturbance Euler equations, which was later ex-
tended to small disturbance Navier–Stokes equations to include viscous
effects [54]. The proposed method produced good results in the
transonic flow regime [55,56]. Thormann and Winghalm [57] devel-
oped a linear frequency domain (LFD) solver taking advantage of pre-
conditioned Krylov GMRES [58] solution method. Later, Widhalm and
Thormann [59] improved the algorithm and provided the analytic de-
rivatives needed in the solution, improving the solver efficiency. The
method was shown to be accurate when compared to full unsteady
time-marching solutions.
These approaches consider unsteady aerodynamic models linearized
about nonlinear equilibrium states and thus can capture the impact of
static nonlinear effects on flutter. Moreover, they demonstrate compu-
tational savings well beyond an order of magnitude compared to fully
unsteady time-marching solutions [53,57]. However, computing deri-
vatives of such methods for optimization is challenging due to the need
for second-order derivative information.
Motivated by the interest in capturing key transonic flow physics
with low computational cost, recent efforts also developed low-order
unsteady transonic aerodynamic models suitable for integrating
transonic flutter analyses into aircraft design.
Fig. 4. Nonlinear flutter analysis process. The nonlinear equations of motion
Skujins and Cesnik [60] proposed a reduced-order unsteady aero-
are linearized around the equilibrium solutions for each flight condition and the
eigenvalues are computed by considering both the elastic and rigid-body DOFs
dynamic model for multiple Mach regimes based on linear convolution
(flutter in free flight), only the elastic DOFs (traditional flutter), or only the with a nonlinear static correction. The methodology included error
rigid-body DOFs (flight dynamic stability). estimation capabilities based on the newly developed method of seg-
ments, which represents a flexible wing as a collection of rigid spanwise
segments subject to local angle of attack and Mach number conditions.
lower-fidelity methods while retaining the nonlinear physics modeled The method of segments was also applied to transonic flutter analysis of
by the higher-fidelity CFD methods. One possibility is to use time-lin- a transport vehicle by Kitson and Cesnik [61].
earized transonic small disturbance (TSD) equations. Linear small-dis- Mallik et al. [62] developed a reduced-order model for HARW
turbance theory is inadequate for capturing strong transonic shocks, but configurations by combining time-linearized Leishman indicial func-
small-disturbance solutions about the steady nonlinear background tions [63] with lift-curve and moment-curve slopes obtained by solving
flow computed using high-fidelity CFD can provide acceptable perfor- the RANS equations around airfoils at various Mach number and angle
mance and accuracy [50]. Another possibility is to use the transonic of attack conditions and for various thickness ratios. They obtained a
equivalent strip (TES) theory [51] and a provided pressure distribution state-space formulation for the airfoil unsteady aerodynamics to be
from either experimental data or a high-fidelity CFD code to compute used for eigenvalue analysis, which was extended to three-dimensional
the small-disturbance transonic aerodynamic loads for flutter analysis HARW discretized in spanwise strips by accounting for sweep

6
E. Jonsson, et al. Progress in Aerospace Sciences 109 (2019) 100537

correction. Flutter results were compared with wind-tunnel experi- boundary conditions and state variables on flutter prediction.
ments for a truss-braced wing (TBW) configuration. The low-order Mazidi et al. [71] investigated the effect of engine placement and
model captured the transonic dip that was not predicted by potential- roll maneuver on flutter results. They observed that the roll maneuver
flow theory and presented good agreement with experiments at sig- has a destabilizing effect on the flutter boundary dependent on the bank
nificantly lower computational cost compared to unsteady RANS si- angle. Additionally, the location of the engine or external store greatly
mulations. These results showed the method suitability for conceptual affects the flutter boundary and the roll-induced effects. Nearly all ve-
HARW aircraft design including transonic flutter constraints. hicles perform roll maneuvers during turns, making the inclusion of
Opgenoord et al. [64] developed a physics-based two-dimensional these conditions relevant to the aircraft design process.
low-order model for transonic airfoils using the perturbations of the There has been further work on the effect of rigid-body DOFs on the
lowest-order volume-source and vorticity moments with respect to a flutter problem. Niblett [70] investigated the causes of BFF for a con-
known nonlinear background flow solution as the states. Evolution ventional wing-fuselage configuration using a linear analytical flutter
equations for these perturbations were derived and calibrated using method. Su and Cesnik [12] investigated the flutter behavior of a
data from high-fidelity Euler CFD simulations. A state-space unsteady blended wing-body (BWB) aircraft both for cantilevered and free-flying
aerodynamic model was obtained for airfoil flutter analysis which was conditions using the University of Michigan's Nonlinear Aeroelastic
later extended to three-dimensional HARW configurations [65] using Simulation Toolbox (UM/NAST). They observed a reduction in the
strip theory and sweep correction, as done by Mallik et al. [62]. The flutter speed when the rigid body DOFs were included compared to the
method was applied in conceptual design and optimization problems cantilever case. Moreover, the flutter mode changed to include pitch
including transonic flutter considerations [65,66]. and plunge motions, resulting in BFF. Similarly, Jones and Cesnik [72]
In addition to capturing transonic effects, a more recent flutter investigated the BFF characteristics of the X-56A experimental aircraft,
modeling need is to take into account geometric structural non- describing the entire modeling process used for the flutter prediction.
linearities. These are particularly important in the analysis of HARW Cesnik and Su [73] analyzed the University of Michigan's X-HALE very
configurations, which achieve higher energy efficiency at the cost of flexible aeroelastic testbed [74] and observed that significant wing
increased structural flexibility and thus experience large deflections deformations can drive lateral BFF due to the coupling of the Dutch roll
under normal operating load conditions. The changes in geometric and asymmetric wing bending modes. Su and Cesnik [69] investigated
shape and stiffness properties due to these deflections significantly af- the stability and dynamic response of a highly flexible flying wing for
fect the flutter boundary [13]. When deformations are large, traditional different payload configurations and gust disturbances. They found that
flutter analysis based on the vehicle undeformed shape does not capture wing deflections can lead to an unstable phugoid mode and an aper-
the actual behavior of the aircraft during flight. iodic short-period mode. Similar behaviors were observed by Patil and
Studies on isolated HARWs [67], high-altitude long-endurance Hodges [68] and Patil and Taylor [75]. Richards et al. [76] also ana-
(HALE) flying-wing configurations [12,68,69], and commercial trans- lyzed the coupled flight dynamics and aeroelasticity of flying wings.
port vehicles [61] pointed out the need to analyze very flexible aircraft They noted that BFF occurred due to a coupling of the short period
in statically deformed configuration at trim, which varies with the flight pitching mode and the first elastic bending mode. They compared dif-
condition. Including structural nonlinearities in flutter prediction is ferent inertial configurations of the aircraft and noted that BFF depends
challenging for both analysis and design due to the high computational largely on the inertia about the pitch axis. They found a boundary value
cost of nonlinear aeroelastic simulations and the flutter boundary de- for pitch inertia that uncoupled the pitch and bending modes, thus
pendency on the deformation state, which is not considered in linear replacing BFF with a more conventional flutter. After parameter stu-
approaches. dies, they concluded that BFF was caused by low fuselage inertia, which
Finally, classical wing flutter analyses typically assume the vehicle could be mitigated by redistributing the fuselage mass.
to be clamped at the wing root. While this may be an acceptable sim- BFF is not exclusive to flying-wing configurations; conventional
plification for some vehicles, it does not reflect the vehicle behavior in tube-and-wing aircraft can also encounter this type of instability.
free flight [15,16]. For some configurations, simply including rigid- Cavallaro et al. [77] investigated the flutter behavior of a Prandtl Plane
body DOFs influences the flutter solution substantially [12]. This occurs (boxed wing aircraft) including rigid-body DOFs using MSC Nastran.
due to the coupling between rigid-body motion and structural dynamics Similarly to Richards et al. [76], they observed a dependence on the
that arise in the presence of low natural vibration frequencies. These fuselage mass in causing BFF: while the baseline configuration en-
interactions usually result in lower flutter points than the cantilevered countered flutter without rigid-body contributions, increasing the fu-
configurations or different flutter mechanisms like body-freedom flutter selage mass resulted in BFF. Therefore, capturing these phenomena in
(BFF) [70]. Therefore, it is imperative to understand the effect of flutter analyses is advisable even for conventional configurations and

Fig. 6. Bifurcation diagrams for possible post-flutter scenarios. For supercritical bifurcation (a), LCOs occur only beyond the flutter point. For subcritical bifurcation
(b), LCOs can occur even before the flutter point if disturbances are sufficiently large. In presence of higher-order stabilizing nonlinearities, both a stable (solid line)
and an unstable (dashed line) solution branch exist for a range of flight conditions.

7
E. Jonsson, et al. Progress in Aerospace Sciences 109 (2019) 100537

imperative when investigating non-conventional ones [15,16]. lie within the flight envelope. This considerably reduces the effective-
ness of flutter constraints enforced during design.
3.2. Post-flutter modeling When higher-order stabilizing nonlinearities are present, as shown
in Fig. 6(b), the system suddenly responds with a large-amplitude LCO
When aircraft flutter, disturbances perturbing an equilibrium state when perturbations before the flutter point pass the amplitude limit
cause an oscillatory response with divergent amplitude. This may ul- given by the unstable solution branch. Similarly, the system suddenly
timately result in structural failure or, if the amplitude growth is gra- responds with a large amplitude LCO when it is perturbed beyond the
dual enough, the aircraft can be brought back at a stable operating flutter point. Therefore, subcritical bifurcations are characterized by
point by pilot or control action. Dynamic nonlinearities in the aero- discontinuous behavior with respect to both the initial conditions and
dynamics or the structure, however, usually play a stabilizing role and the flight parameters.
restrain divergent motions, leading to a self-sustained oscillatory re- In contrast, LCO amplitude grows gradually and with much smaller
sponse of limited amplitude that remains constant in time, known as amplitudes past the flutter point in supercritical systems (see Fig. 6(a)).
LCO. Unfortunately, nonlinear effects tend to be destabilizing and cause Furthermore, subcritical post-flutter response is characterized by hys-
LCOs even before reaching the flutter point [14]. Furthermore, LCOs teresis. Fig. 6(b) shows that once a stable LCO develops, the system does
may or may not develop for a given flight condition depending on the not recover the initial equilibrium state when the flight parameter is
amplitude of disturbances (initial conditions) and previous motion reduced below the critical (flutter) value.
history (hysteresis) [78]. Therefore, when nonlinear effects become To eliminate the LCO, it is necessary to reduce the flight parameter
significant, post-flutter phenomena need to be considered in addition to below the value at the folding point connecting the unstable and stable
the flutter boundary in the design process. solution branches. Depending on the type of nonlinearities present in
As mentioned in Section 1, in aircraft aeroelastic qualification, the the system, this may be significantly lower than the flutter value. The
term LCO is typically associated with control-surface free-play. In this above considerations make clear that when nonlinear effects become
paper, LCO has the more general meaning of post-flutter response significant, design optimization should include post-flutter constraints
leading to the development of self-sustained oscillations in vehicle for obtaining feasible optimal designs that do not experience subcritical
components or in the entire structure. When nonlinear behavior is an- LCOs.
ticipated in aircraft configurations, optimization should include con- The post-flutter behavior of wings with structural and aerodynamic
straints on the post-flutter response, in addition to flutter constraints, to dynamic nonlinearities was analyzed by various researchers
avoid the occurrence of undesirable LCOs at the operating conditions. [12,78,80–82]. This topic was also included in review papers [8,14,18].
Post-flutter behaviors are typically classified using bifurcation dia- However, only a few efforts integrated post-flutter analyses in design
grams. These diagrams represent the LCO amplitude variation with a optimization as a constraint, as discussed in Section 5.4.
chosen control parameter, called the bifurcation parameter. For aero-
elastic systems, this is usually the flight speed or the dynamic pressure 4. Flutter analysis in aircraft design optimization
[79]. Two qualitative post-flutter scenarios are possible, as illustrated in
Fig. 6. Because flutter is a safety-critical phenomenon, analyses and ex-
Fig. 6 (a) shows the bifurcation diagram for a system with super- perimental investigations are required for vehicle certification.
critical post-flutter behavior. When the system is perturbed before the Analyzing a configuration for flutter late in the development cycle is
flutter point, the response always converges towards the initial equili- likely to result in an inefficient design solution or pose challenges mi-
brium state regardless of the magnitude of the disturbance. When the tigating unexpected instabilities, resulting in performance decrease,
perturbation is applied beyond flutter, the response grows unbounded financial losses, or both. For this reason, flutter should be integrated
until dynamic stabilizing nonlinearities lead to an LCO with limited into the design process early in the form of a constraint.
amplitude that keeps constant in time. Each point on the bifurcation In this section, we review flutter prediction methods and previous
diagram gives the LCO amplitude as function of the flight parameter. In research that addressed optimization subject to flutter constraints. Past
the case of supercritical behavior, the LCO amplitude increases work has focused primarily on flutter analyses using linear structures
smoothly with the flight parameter and with a decreasing slope as and linear aerodynamic models. Further work considered transport
stabilizing nonlinearities become stronger [14]. aircraft operating in the transonic regime and thus used linear struc-
The bifurcation diagram of a system with subcritical post-flutter tures and nonlinear aerodynamics. The recent trend towards more
behavior is illustrated in Fig. 6(b). In this case, an unstable bifurcation flexible aircraft has led to research on constraining flutter for geome-
branch (dashed line) exists below the flutter point. As a result, when trically nonlinear structures.
disturbances are sufficiently large, the initial equilibrium configuration Due to the sheer number of design variables typically used in
is lost. If higher-order stabilizing nonlinearities are present, both an practical aircraft optimization, most of the previous work reviewed in
unstable and a stable (solid line) solution branches exist for a range of this paper used gradient-based methods. However, studies on geome-
flight conditions. In this case, the response grows until restraining trically nonlinear configurations were limited to simple structures
nonlinearities lead to an LCO with an amplitude given by the upper parametrized by few design variables. For this reason, efforts that
stable bifurcation branch. Supercritical bifurcation is characterized by considered gradient-free optimizations are also included.
discontinuous behavior with respect to the amplitude of initial dis-
turbance due to destabilizing nonlinear effects. 4.1. Prediction methods
Supercritical LCOs are undesirable in aircraft, since they cause
structural fatigue or failure, performance and ride comfort degradation, Flutter computations are typically performed in the frequency domain
or loss of control. However, if the flutter boundary is satisfied with the by solving an eigenvalue problem. Well-established eigenvalue-analysis
required safety margins, LCOs never develop since the aircraft always methods exist for flutter analysis of linear aeroelastic systems (e.g., k-, p-,
operates below the flutter speed by certification. Additionally, enfor- pk - and g-method) [83–87]. These methods are also applicable to non-
cing flutter constraints during design can guarantee that flutter does not linear systems by linearizing the equations of motion about the nonlinear
occur and therefore, supercritical LCOs are not encountered. equilibrium configuration at each flight condition for capturing static
However, subcritical post-flutter behavior is a more serious pro- nonlinearities due to the structure (large deflections) or the aerodynamics
blem, because instabilities may occur at flight conditions that are ex- (background transonic flow). Direct methods, based on the Hopf-bifurca-
pected to be stable according to a conventional flutter analysis. tion theory, can also be used to predict the flutter point of nonlinear
Depending on the type of nonlinearities, these flight conditions could aeroelastic systems directly in the frequency domain [18].

8
E. Jonsson, et al. Progress in Aerospace Sciences 109 (2019) 100537

It is also possible to predict flutter in the time domain, but this in- Fˆaero, r (s ) = q Qr (s ) qˆ (s ) (5)
curs a much higher computational cost. When analyzing the stability of
a system in the time domain, the flutter point is typically evaluated by where s is the Laplace variable, q = is the flight dynamic
U 2/2
perturbing equilibrium configurations at different flight conditions and pressure corresponding to the density ρ and flight speed U, Qr (s ) is the
by time-marching the equations of motion in order to verify the decay generalized aerodynamic force (GAF) matrix, and q̂(s ) is the Laplace
or growing of the response, or to extract damping values in a post- transform of q(t ) . Taking the Laplace transform of Eq. (3) yields
processing stage [88–95]. [s 2 Mr + Kr q Qr (s )] qˆ (s ) = 0, (6)
The computational cost of time-domain flutter analysis based on
transient simulations is currently prohibitive for optimization, parti- or
cularly in the presence of aerodynamic or structural nonlinearities. The U 2
high computational cost is due to the large number of computations p2 Mr + Kr q Qr (p) qˆ (p) = 0,
b (7)
required to evaluate the flutter speed by means of flight speed (or dy-
namic pressure) sweep or bisection. Additionally, ascertaining the sta- where p = sb/ U = g + ik is the nondimensional Laplace variable, b is
bility close to the flutter point requires long integration times due to the reference half chord, g is the nondimensional damping, and k is the
small damping values. reduced frequency [83].
For gradient-based optimization, another challenge is the efficient Eq. (7) requires computing the GAF in the Laplace domain. How-
computation of derivatives for time-marched systems. Adjoint methods ever, the GAF is typically given as a transcendental function, Qr (ik ) , of
are advantageous for optimizations with many design variables, but the reduced frequency. Two approaches are used for overcoming this
they are computationally expensive when applied to time marching problem [10]. One approach is to approximate Qr (p) Qr (ik ) and solve
solvers [96]. This is because the adjoint solution requires a reverse time the flutter equation, Eq. (6), or an equivalent form by computing the
integration following the forward integration for computing the system GAF matrix in the reduced frequency domain while enforcing the
response [97], which results in large computational times and storage condition I (p) = k . An alternate approach is to obtain a rational func-
memory requirements [98]. Due to the above limitations, we focus on tion approximation (RFA) of Qr (ik ) and use analytic continuation [105]
frequency-domain flutter prediction methods, but some fully coupled to extend its domain from the imaginary axis (reduced frequency) to the
unsteady aeroelastic adjoint implementations do exist in literature entire complex plane (nondimensional Laplace variable). The aero-
[99–104]. Eigenvalue-based methods are discussed in Section 4.1.1 and elastic system can be recast in state-space form by introducing addi-
direct methods in Section 4.1.2. tional aerodynamic states, such that flutter can be analyzed using a
standard eigenvalue analysis.
4.1.1. Eigenvalue analysis methods When the GAF is represented as a transcendental function of k,
The discrete equations of motion for a generic linear aeroelastic flutter analysis is performed using iterative or non-iterative methods
system can be written as that either compute the true damping only at the flutter point (k-
method) or at all flight speeds or dynamic pressure values (root locus,
Mu
¨ (t ) + Ku (t ) Faero (t ) = 0, (1) p-, pk -, and g-methods) [83–85,87,106]. As an example, using the
where Ns is the number of structural DOFs (nodal displacements and pk -method, assuming Qr (p) Qr (ik ) , Eq. (7) is rewritten as [83],
rotations), M, K Ns × Ns are the mass and stiffness matrices resulting
U 2
from the finite-element discretization, u Ns is the nodal displace- p2 Mr + Kr q Qr (ik ) qˆ (p) = 0.
b (8)
ment vector, and Faero Ns is the nodal aerodynamic load vector. With

no loss in generality, we omit the structural damping in Eq. (1) along This is a second-order nonlinear eigenvalue problem, where the
with other external loads acting on the structure. For the sake of con- nonlinearity stems from the dependency of Qr on the imaginary part of
ciseness, Eq. (1) omits the dependency of the structural matrices and p.
aerodynamic loads on the design variables d Nd
. Therefore, flutter Several decompositions of the GAF Qr in Eq. (8) exist in the litera-
and post-flutter analysis methods are presented for a fixed design and ture. Stanford [107] summarized these decompositions and showed
all the structural and aerodynamic quantities are updated at each op- that they predict the same flutter speed, but that the mode migration
timization step based on the current values of the design variables. and characteristics may be very different. This directly impacts the
Due to the sheer number of structural DOFs, it is common to reduce optimization, resulting in different optimal designs.
the computational effort by rewriting Eq. (1) in terms of a reduced set As an example, a possible approach is to split the GAF into its real
of Nr generalized coordinates (modal amplitudes) where Nr Ns . The and imaginary parts as Qr = QrR + iQrI . Assuming small damping, which
displacement field can be then approximated by gives p /(ik ) 1 [108], Eq. (8) can be rewritten in first-order form as the
generalized nonlinear eigenvalue problem [84,109,110],
u (x , y , z ; t ) r (x , y , z ) q (t ), (2)
Ir 0 qˆ 0 Ir qˆ
where q is the vector of retained generalized coordinates and
Nr
p = 0,
r is a matrix whose columns contain the corresponding ei-
Ns × Nr 0 () U 2
b
Mr pqˆ (K r q QrR)
q
k
QrI pqˆ
(9)
genvectors (mode shapes). Substituting Eq. (2) into Eq. (1) and pre-
multiplying by Tr yields where Ir Nr × Nr
is an identity matrix. For any given value of k, Eq. (8)
(or Eq. (9) with the above specific form of Qr ) can be solved using
Mr q¨ (t ) + Kr q (t ) Faero, r (t ) = 0, (3)
standard eigenvalue routines such as LAPACK [111] to compute the
where Faero, r = Nr is the vector of generalized aerodynamic
T
r Faero eigenvalues and eigenvectors. However, valid roots need to satisfy the
forces and, assuming that the eigenvectors are M-orthonormal, the equivalence I (p) = k . One approach to this problem is to use an
generalized mass and stiffness matrices take the form iterative solution algorithm as suggested by Hassig [83], similar to what
T Nr × Nr ,
is implemented in MSC Nastran [108]. As a disadvantage, iterative
Mr = rM = Ir
r
methods may converge to incorrect roots or not converge at all
T 2 Nr × Nr , (4) [112,113].
Kr = rK r = diag{ i }
In gradient-based optimization, robustness of analysis methods,
where i is the i-th natural angular frequency (i = 1, …, Nr ). continuity, and smoothness are of utmost importance. For this reason,
Assuming linear aerodynamics, the generalized aerodynamic forces noniterative methods [87,113] that mitigate the aforementioned issues
are written in the Laplace domain as were more recently proposed and applied in optimization settings

9
E. Jonsson, et al. Progress in Aerospace Sciences 109 (2019) 100537

[109,110,114]. theory for active flutter suppression [10]. Moreover, it enables state-
On the other hand, the GAF can be approximated by numerically space time-domain aeroservoelastic response analysis and facilitates
fitting data in the reduced frequency domain using a RFA. This includes synthesizing load alleviation control laws. RFAs have also been used in
polynomial terms up to the quadratic term for representing aero- aeroservoelastic stability and response sensitivity computation [124-
dynamic stiffness, damping, and mass effects along with a rational term 128].
representing unsteady aerodynamic lag effects. Once a RFA is com- The above treatments are typically applied to linear systems. For
puted, its validity can be extended from the reduced frequency domain these systems, stability is analyzed about the undeformed configuration
(imaginary axis of the complex plane) to the nondimensional Laplace by describing the structure using a detailed linear finite element model
domain (entire complex plane) using analytic continuation [105]. (FEM) and by computing the GAF using linear potential-flow lifting-
Several RFA approaches exist in the literature. Roger [115] pro- surface methods like the doublet-lattice method (DLM) [129] or panel
posed the following widely used form (given in nondimensional Laplace methods [8,18]. In the presence of nonlinear aerodynamic effects, GAF
domain below), matrices are computed by considering small perturbations with respect
Nl + 3 to the nonlinear background flow solution for each flight condition.
p Al One common approach to this problem is to use DLM [129] aug-
Q (p) = A 0 + p A1 + p2 A2 + ,
l=3
p+ l 2 (10) mented with higher-fidelity numerical data or experimental measure-
ments [49]. An alternate, more accurate approach is to use time-line-
where Nl is the number of lag terms (aerodynamic roots) l 2 , which are
arized CFD [57] to identify the GAF matrix while capturing static
specified by the user based on the reduced frequency range of interest.
nonlinear aerodynamic effects directly. If structural nonlinearities are
A common choice is to assume Nl = 3 [116]. The coefficient matrices
Nr × Nr are determined using a least-square fitting of the moderate, as in the case of modern transonic transport vehicles, this
A 0, …, An
statically nonlinear aerodynamic description can be still combined with
GAF as a function of k [115], whose accuracy depends on the number of
linear FEM structural models. However, when nonlinear effects in the
aerodynamic roots, as well as their values.
structure become significant, the equations of motion need to be line-
Karpel [116] reviewed early RFA methods and proposed a new
arized about the aeroelastic equilibrium configuration at each flight
technique based on the minimum number of aerodynamic states (Nm ),
condition to retain the effects of large static deflections or background
corresponding to the approximation
nonlinear steady flow. The eigenvalues of the linearized aeroelastic
Q (p) = A 0 + p A1 + p2 A2 + D [pI R ] 1 Ep , (11) system are then computed for each flight condition to determine the
flutter boundary, that is, the flight conditions where damping vanishes
which includes Eq. (10) as a special case [117]. The interpolating
[12].
matrices A 0, A1, A2 Nr × Nr , D Nr × Nm , R Nm × Nm , and
Due to the high computational cost of nonlinear aerostructural so-
E Nm × Nr
are again computed by fitting the GAF data at different
lutions and the number of flight conditions to be analyzed in the flutter
reduced frequencies. However, the least-square fitting requires to solve
search, structural nonlinearities are frequently captured with beam
a system of nonlinear algebraic equations due to unknown coefficients
models in nonlinear flutter analyses [7,12,61,68,69,130]. These models
in the denominator, which results in an iterative process. Morino et al.
incur low computational cost, while still being able to capture the be-
[118] later proposed a simpler and easier approach to RFA based on an
havior of very flexible HARW vehicles. Low-fidelity aerodynamic for-
interpolative structure similar to the one proposed by Karpel [116] but
mulations typically used in geometrically nonlinear flutter analyses are
requiring the solution of a linear system. Furthermore, they also pro-
potential-flow theories with tip and compressibility corrections
posed higher-order RFA models. More recently, Ripepi and Mantegazza
[7,12,68,69,130], vortex-lattice methods [131], or surrogate models
proposed an improved RFA interpolation structure for state-space
[61].
aeroelastic stability and gust response analysis based on three nonlinear
The previous section was a brief introduction to eigenvalue-analysis
least-squares identification techniques combined with a model-order
methods. We now discuss the critical aspects to be considered when
reduction [119].
using these methods for constraining flutter in design optimization. As
Eversman and Tewari [120] discussed methods to optimize the
previously mentioned, we mainly refer to gradient-based optimization
aerodynamic lag using gradient-free or gradient-based approaches for
methods, which require the evaluation of the constraint values and
improving the fitting accuracy. They also argued that the p in the nu-
their derivatives with respect to the design variables.
merator of the rational term in Eq. (10) can be omitted, resulting in the
The process of solving for the eigenvalues and computing deriva-
following alternate form
tives by differentiating the eigenvalue problem analytically is widely
Nl + 3
Al used [124,132–134] and it is included in several commercial analysis
Q (p) = A 0 + p A1 + p2 A2 + . and design software packages [135–137]. Depending on the form of the
p+ l 2 (12)
l=3
flutter equation, different levels of accuracy are considered when taking
They optimized the aerodynamic lag parameters and minimized the the derivative of the eigenvalue problem. If mode shapes are used to
number of these parameters for a given level of accuracy using a gra- reduce the model order, the fixed-mode approximation [134] is often
dient-free optimization method. Pasinetti and Mantegazza [121] also used, which neglects the derivatives of the mode shapes with respect to
used optimization to obtain a state-space aeroelastic model reduced to a the design variables. Such practice is commonly applied for the sake of
minimum number of states. More recently, Nissim [122,123] for- efficiency as these derivatives, depending on the number of structural
mulated a minimum-state approximation as a nonlinear optimization DOFs and mode shapes, can be very expensive to compute. Examples of
problem considering the aerodynamic lag terms and the two matrices this approximation are found in recent literature for structural opti-
that operate on them as design variables. mization (with no planform changes) [109,138,139] and topology op-
Using a RFA of the GAF and analytic continuation, the aeroelastic timization [114,140]. However, ignoring the derivatives of the mode
system Eq. (7) can be rewritten in the state-space form x = Ax . The shapes can result in inaccurate gradients [134].
state vector x = [qTr , qTr , aT]
T N
includes the generalized co- Another aspect that is critical to the efficacy of gradient-based op-
ordinates qr , their derivatives qr , and Na additional aerodynamic states timization is the smoothness of the constraint function, which should be
a (whose number varies depending on the type of RFA used), for a total at least C1 continuous. Several considerations are necessary to formulate
number of N = 2Nr + Na states. Representing the aeroelastic system in a smooth flutter constraint function. A naive approach would be to
terms of a linear time-invariant model with RFAs allows the aeroelastic specify the flutter point directly. However, this may introduce dis-
eigenvalues to be computed using standard linear eigenvalue analysis continuities in the constraint value between two consecutive design
techniques (root locus) and it facilitates the application of control iterations, di and di + 1, due to mode switching or to a hump mode

10
E. Jonsson, et al. Progress in Aerospace Sciences 109 (2019) 100537

becoming active at a significantly lower speed [3]. Therefore, addi- handled as well.
tional steps are required to ensure a continuous constraint. Although the above approach mitigates the discontinuity issues, it
Mode switching occurs when the mode that first becomes unstable requires one to apply a constraint for each speed increment and mode.
(that is, at the lower speed) changes between two consecutive design Thus, NU speed increments and Nr modes results in Nr NU constraints,
iterations di and di + 1. An example of mode switching is shown in dramatically increasing the total number of constraints. To mitigate the
Fig. 7(a), where the hypothetical damping of a system with two modes high cost of computing derivatives for so many constraints, the active
is plotted with respect to speed. Mode switching causes a C1 dis- set method can be used. This considers the full set of points, but reduces
continuity of the flutter point, which poses a challenge to gradient- them to a smaller set before evaluating derivatives based on the con-
based optimizers [141]. The imaginary part of the eigenvalue (fre- straint value. This approach was applied by Ringertz [141] and was also
quency) also switches, and in many cases the frequencies coalesce, applied in the time domain by Kang et al. [145]. An alternate approach
causing a mode to become unstable. A more serious problem is when a to reduce the number of constraints was proposed by Haftka [146], who
hump mode is present in design di and it becomes the critical mode in suggested replacing parametric constraints by minimum-value con-
the new design, di + 1, as shown in Fig. 7(b). The constraint value ex- straints. This reduces the number of constraints to the total number of
periences a C 0 discontinuity, which is even more challenging for gra- modes for the entire flight envelope. The approach can further benefit
dient-based optimizers. from the aggregation of the set of constraints for the individual modes
Techniques exist to mitigate these problems and they are summar- into a single constraint using for example the Kreisselmeier–Steinhauser
ized by Stanford et al. [138]. Frequency-separation constraints pro- (KS) function [147–149].
posed by Langthjem and Sugiyama [142] and also by Odaka and Furuya Other classical constraint aggregation techniques can be used, such
[143] can prevent mode switching by enforcing the critical mode to as the p-norm function or induced aggregates [150]. Kennedy and
remain the same. This approach is illustrated in Fig. 8 (left) for a hy- Hicken [150] compared the accuracy of classical aggregation functions
pothetical case. The disadvantage of this method is that Nm 2 con- with their proposed induced aggregates. They demonstrate that the
straints are needed for a case with Nm modes (with the expectation that induced aggregates provide better accuracy and properties compared to
two modes coalesce, and hence no constraint is needed for two modes). the classical aggregation functions (KS and p-norm). Lambe et al. [24]
Furthermore, specifying the unstable mode could over-constrain the studied the accuracy and efficiency of the above aggregation methods
optimization process. A more serious flaw with this approach is that a for structural mass minimization of a wingbox subject to failure con-
hump mode is still possible. straints. The induced exponential aggregation was shown to be more
Other techniques exist to handle both mode switching and hump accurate compared to the classical constraint aggregation functions
modes. One approach is to enforce the real part of each eigenvalue to [24].
remain below a preset bounding curve. Such a bounding curve is pro-
posed by Stanford et al. [144], which is a modified version of the one 4.1.2. Direct methods
proposed by Ringertz [141]. The boundary (G (U ) in Fig. 8) spans the Eigenvalue-analysis methods were originally introduced for linear
operating conditions of interest from wind-off to some maximum speed. flutter analysis and later applied to systems with structural or aero-
This approach mitigates the issues mentioned above and has further dynamic nonlinearities. Due to the growing interest in analyzing non-
benefits: 1) No constraint is placed on the flutter point itself, so there is linear aeroelastic systems, methods based on Hopf bifurcation theory
no need to compute it explicitly; 2) Modes that become unstable that compute the flutter point directly in the frequency domain while
abruptly (hard flutter) and have steeper slope than the boundary are enforcing the nonlinear aeroelastic equilibrium equations were more

Fig. 7. V g diagrams showing possible sources of discontinuities in the flutter constraint. Mode 1, which is the critical mode for design di , switches with mode 2 as
the critical mode in design di + 1 (adapted from Stanford et al. [138]).

11
E. Jonsson, et al. Progress in Aerospace Sciences 109 (2019) 100537

Fig. 8. V g diagrams showing two possible methods to prevent discontinuities in the flutter constraint: frequency-separation method (left) and damping boundary
(right, solid black). Structural frequencies for a panel are often not as well separated as for a wing, causing the frequency separation constraints to be often active.
(adapted from Stanford et al. [138]).

recently used. Approaches were also developed to compute derivatives solved, the solution vector at the next iteration is updated as
for direct methods to use them for the formulation of constraints in X (n+ 1) = X (n) + X (n) and the process is repeated until convergence is
gradient-based optimization (see Section 4.2). reached.
Here, we introduce the formulation of the Hopf-bifurcation ap- While the Newton–Raphson method usually exhibits fast con-
proach and highlight recent work that contributed to its development. vergence, it is not without some disadvantages. In particular, this
In particular, we focus on efforts that compute the Hopf point directly method requires a suitable initial guess X (0) that is sufficiently close to
[151]. This is in contrast to indirect methods that monitor eigenvalues, the Hopf point in order to converge successfully [152]. Furthermore,
look for sign changes, and then perform a local search (using a New- this method may converge to other Hopf points, such as divergence
ton–Raphson, secant, bisection, or interpolation method) [19], similar [153], or to a zero velocity.
to what was discussed in Section 4.1.1. A common strategy to overcome these issues is to solve the eigen-
Consider a generic nonlinear aeroelastic system with the dynamics value problem using large U steps and track when one eigenvalue
governing equations written as crosses the real axis. Once the crossing has been identified, the direct
Newton method is applied to identify the Hopf point, which typically
x = f (x; U ). (13)
converges in few iterations [154,155].
The state vector x = [qT , qT ]T includes the generalized co-
N
Computing the Jacobian J accurately is also a challenge. Various
ordinates assumed to describe the system and their derivatives, for a techniques can be employed to approximate the Jacobian when solving
total number of N states, f N
is a nonlinear vector field, and U is the Eq. (17): approximating it entirely using finite differences, approx-
flight speed (the bifurcation parameter). The system Jacobian matrix is imating or omitting individual entries at the expense of convergence, or
A: = f/ x N ×N
. A Hopf bifurcation point (flutter point) occurs when reducing the size of the Jacobian by taking advantage of symmetries in
the flight speed reaches the critical value UF , which corresponds to an the problem [152,153,155,156]. For practical aircraft configurations
equilibrium state xF such that f (xF ; UF ) = 0 . The Jacobian matrix modeled by CFD and FEM, the resulting Jacobian can be a large sparse
evaluated at the Hopf point A (xF ; UF ) has a pair of purely imaginary system (which requires a preconditioning strategy) that is also ill-con-
eigenvalues ± i F , where F is the flutter angular frequency. At this ditioned [153,157].
point, they cross the imaginary axis, causing the Hopf bifurcation. As Hui and Tobak [158] were among the first authors to apply the
explained in Section 3.2, this leads to an LCO. Hopf-bifurcation method to investigate the flutter behavior of nonlinear
Let the right complex eigenvectors associated with the bifurcating aeroelastic systems. They analyzed the stability of pitch motion about a
eigenvalues be u = ur + i ui , and u¯ = ur i ui , where ur , ui N . The
large mean angle of attack and showed that the aerodynamic damping
critical eigenvectors are computed by solving vanishes when the mean angle of attack exceeds a critical value.
Morton and Beran [152] studied a two-degree-of-freedom airfoil
(A iI F) u = 0, (14)
undergoing plunge and pitch in inviscid transonic flow. The aero-
where I is an identity matrix, and using the normalization
N ×N
dynamics was modeled with an Euler CFD solver, while the structural
condition cT u = i , where c N
is arbitrary and constant. The flutter model consisted of linear and torsional springs for the plunge and pitch
point is given by the solution of the nonlinear algebraic system, DOFs, respectively. They computed the flutter point directly using the
Hopf-bifurcation method and validated it with time-accurate analyses.
f (xF ; UF ) = 0
The Jacobian was computed using second-order finite differences and
Aur + F ui = 0
, the system, Eq. (16), was solved using a Newton–Raphson method.
Aui F ur = 0
They showed that the approach is computationally efficient compared
(15)
T
c ur = 0
to full time-accurate simulations, and that it produces accurate results
where the 3N + 2 unknowns are the components of xF , ur , and ui , the for inviscid transonic flow.
flutter speed UF , and the angular frequency F . This system can be re- Badcock et al. [159] extended their previous work [153] and ap-
written in the compact form as plied a modified Newton–Raphson method to solve the Hopf-bifurca-
T
tion problem. The method was demonstrated on the AGARD 445.6
F (X ) = 0 , where X = [ xTF uTr uTi UF F] , (16) aeroelastic wing [160] by computing the full flutter boundary in an
and solved iteratively via the Newton–Raphson method. Starting from a inviscid transonic flow. For this configuration, the direct approach was
first guess X (0) for the solution vector, at the nth iteration this is updated up to two orders of magnitude more efficient than time marching.
by solving Woodgate and Badcock [161] extended this work and analyzed
three geometries: the Goland wing [162], a transonic wing-body con-
J (n ) X (n ) = F (n ) (17) figuration [163], and the AGARD 445.6 wing. The direct method re-
where J = F/ X is the Jacobian of the extended system and is X (n ) duced the computational cost by one to three orders of magnitude
the Newton–Raphson update. Once the linear system, Section 17, is compared to time-accurate simulations.

12
E. Jonsson, et al. Progress in Aerospace Sciences 109 (2019) 100537

More recently, Badcock and Woodgate [157] further improved the number of design variables that were considered in the optimization
approach for large-scale CFD-based systems coupled with FEM struc- problem. Most of the efforts included only structural sizing variables
tures. They used a Schur-complement eigenvalue formulation to treat without considering changes in the aerodynamic shape. Fewer efforts
parts of the system Jacobian in a parallel computing environment. This optimized the airfoil or planform shapes. Including planform shape
procedure eliminated the process of solving large sparse systems, which variables is challenging because changes in the mode shapes and the
are almost singular, further increasing the method efficiency. This corresponding natural frequencies need to be considered when com-
method shows similarities to the eigenvalue methods applying mode puting derivatives, incurring additional computational cost.
tracking discussed in Section 4.1.1. The approach was applied to the In one of the early efforts to constrain flutter, Turner [167] for-
Goland wing, a supersonic transport arrow rudder buzz in transonic mulated a mass minimization problem by considering distribution of
flow [164], the wing-body configuration mentioned above [163], and a material rather than the structure topology to meet a specified flutter
generic fighter configuration based on publicly available F-16 data speed. Later, Bhatia and Rudisill [168] developed a numerical proce-
[157,165]. The computational cost of the proposed method varied from dure to minimize wing mass while satisfying a flutter constraint. They
5 to 60 times that of a steady-state solution. applied the procedure to a uniform cross-section box beam consisting of
In the context of gradient-based optimization, formulating a suc- three bays in order to minimize the mass while maintaining the flutter
cessful flutter constraint using the direct method needs careful con- speed. Rudisill and Bhatia [169] improved the rate of convergence of
sideration. Possible discontinuities due to competing flutter mechan- this procedure by computing second-order derivatives of the eigenva-
isms in the design process, such as mode swapping or hump modes, lues and of the flutter speed with respect to the design variables. Gwin
described in Section 4.1.2, may confuse the gradient-based optimizer. and Taylor [170] developed the method of feasible directions for the
However, techniques addressing these concerns in the context of direct mass minimization of a structure subject to a minimum flutter speed.
methods are deficient or nonexistent. They were able to handle up to 33 structural design variables in a su-
personic problem.
4.2. Applications to optimization Due to the limited computational capabilities of the time, these
early efforts used simple aerodynamic and structural models and em-
Despite early work by researchers such as Haftka [3,146] and Hajela ployed a similar strategy to enforce the flutter constraint. They all
[166] optimization subject to flutter constraints is still not a standard formulated the flutter problem in the frequency domain as an eigen-
design practice. More recently, several authors have integrated flutter value problem, which was then differentiated with respect to the design
constraints into design and investigated their effect on the optimal so- variables. The derivatives of the eigenvalues were obtained using the
lutions. These efforts are summarized in Table 1 and reviewed below. left and right eigenvectors and one of these efforts also considered
They differ in the use of eigenvalue analysis (EV), direct (Hopf), or second-order derivatives to better guide the optimization process and
time-domain (TD) prediction methods for flutter analysis, the fidelity of improve the rate of convergence [169].
structural and aerodynamic models, and the optimization problem Ringertz [141] applied the k-method to minimize the weight of a
formulation in terms of objective, type and number of design variables, cantilevered wing in incompressible flow subject to flutter and diver-
and use of gradient-based (GB) or gradient-free (GF) solution algo- gence constraints. The structure was modeled as a composite FEM
rithms. Efforts that used gradient-based algorithms also differ in the model, while the unsteady aerodynamic loads were computed using the
methods used for computing derivatives with respect to design vari- DLM. The eigenvalue problem was analytically differentiated in the
ables. Finally, previous work shown in Table 1 differs in the types and modal space. A continuous flutter constraint was formulated using a

Table 1
Summary of efforts on optimization or derivative computation considering flutter.
Design variables

Models Shape

a b c d
Effort Method Aerodynamics Structure Objective Structural Airfoil Planform N Algorithme

Turner [167] EV Strip theory Panel Min mass • 3 GB-Analytic

Bhatia and Rudisill [168] EV – Beam Min mass • 12 GB-Analytic
Rudisill and Bhatia [169] EV – Beam Min mass • 12 GB-Analytic
Gwin and Taylor [170] EV DLM Beam Min mass • 33 GB-Analytic
Ringertz [141] EV DLM Composite Min mass • 9 GB-Analytic
Mallik et al. [171] EV Strip theory Beam Min FB, Max TOGW • • • 19 GF-Genetic
Jonsson et al. [110] EV DLM Shell Max range • • 3 GB-Adjoint
Stanford et al. [144] EV TSD + Euler Shell Min mass • 92000 GB-Adjoint
Chen et al. [137] EV Euler Shell – Derivative computation only – GB-Complex-step
Bartels and Stanford [139] EV RANS Shell Min mass • 711 GB-Adjoint
Variyar et al. [7] EV Lifting line NL beam Min FB • • 12 GB-Finite difference
Xie et al. [131] EV DLM NL beam Min mass • 44 GF-Direct
Bhatia and Beran [172] EV Euler NL beam Min mass • 8 GB-Analytic
Lupp and Cesnik [130] EV Strip theory NL beam Min FB • • 5 GB-Adjoint
Kennedy et al. [156] Direct Panel Shell – Derivative computation only – GB-Adjoint
Beran et al. [173] Direct ONERA stall NL beam – Derivative computation only – GB-Adjoint
Mani and Mavriplis [99] TD Euler Mass-spring Max fl. speed • 32 GB-Adjoint
Zhang et al. [101] TD RANS Mass-spring Max fl. speed • 48 GB-Adjoint
Zhang et al. [102] TD RANS FEM Max fl. speed • • 120 GB-Adjoint

a
Methods: EV—Eigenvalue method, TD—Time-domain.
b
Aerodynamics: DLM—Doublet lattice method, TSD—Transonic small disturbance, RANS—Reynolds averaged Navier–Stokes.
c
Structures: NL—Nonlinear.
d
Objective: FB—Fuel burn, TOGW—Takeoff gross weight.
e
Algorithm: GB—Gradient based, GF—Gradient free.

13
E. Jonsson, et al. Progress in Aerospace Sciences 109 (2019) 100537

boundary to constrain the damping values, which resulted in a large structural design were metallic thickness variations, functionally
number of constraints. The method was demonstrated on a rectangular graded materials, balanced or unbalanced composite laminates, curvi-
wing and on a swept wing with taper, where the objective was to linear tow steering, and distributed trailing edge control surfaces. While
minimize weight with respect to element group thicknesses subject to there was a structural wing mass reduction for every case, many of the
flutter and divergence constraints. In both cases, a considerably lighter lighter designs had an active flutter constraint, while the buckling
design was achieved. constraint was active for the heavier cases.
Mallik et al. [171] investigated the impact of a previously developed Chen et al. [137] extended their previous work by [136] computing
flutter constraint [174] on the MDO of a TBW aircraft. The flutter speed derivatives of flutter constraints with respect to shape variables using
was computed using the k-method applied to a linear pre-stressed ZEUS coupled with a boundary layer code. The derivatives with respect
structural model. The unsteady aerodynamic loads were computed to shape were computed using the complex-step approach [28,31],
using Theodorsen theory [175] with a Prandtl–Glauert compressibility which is numerically exact. The flutter constraint was formulated using
correction. They implemented an iterative procedure to ensure con- the g-flutter method [87], which was differentiated analytically with
sistency of the flutter speed, Mach number, and altitude, and optimized respect to the design variables. The approach was verified for a canti-
a representative TBW configuration for minimum takeoff weight and levered planform, similar to the F-5 geometry [177,178], where a
fuel burn using a genetic algorithm. The flutter constraint was for- structure consisting of 10 spars, 10 ribs, and upper and lower skins was
mulated by enforcing a minimum flutter Mach number and was added modeled using MSC Nastran. While no optimization results were pre-
to several other mission constraints. Comparing the optimization results sented, the flutter derivatives were verified against exact results.
with those obtained by removing the flutter constraint, they showed Bartels and Stanford [139] proposed an approach to enforce a CFD-
that this is necessary to obtain a flutter-free optimal solution. based flutter constraint for gradient-based structural optimization of
Jonsson et al. [110] developed a flutter constraint suitable for high- transonic vehicles. The flutter analysis was performed as a standard
fidelity gradient-based optimization including wing planform variables eigenvalue analysis on the state-space representation of the aeroelastic
for the first time. They developed an efficient non-iterative root finding system obtained via RFA [115]. The GAF matrix of the baseline struc-
method to compute the flutter eigenvalues, based on the pk flutter ture was identified from time-linearized unsteady RANS simulations
equation that includes a scheme to track the mode migration between about nonlinear steady-state solutions [179]. The GAF matrix of the
subsequent reduced frequencies and designs. updated design was then computed by projecting the updated modes
The flutter-free flight envelope was implicitly defined using a con- onto the baseline modes. The methodology was used to minimize the
straint curve for the damping values to prevent discontinuities in the uCRM mass subject to structural and aeroelastic constraints. The flutter
flutter constraints and to allow the computation of derivatives. The was constrained by requiring that the real part of the system eigenva-
differences between damping values and the constraint curve were lues be below a bounding curve [166]. The optimization assumed a
aggregated into a single value using the KS function [147]. Accurate fixed-mode approximation. The authors compared the optimal solutions
and efficient derivatives of the aggregated flutter constraint value were obtained using unsteady RANS and DLM aerodynamics in the optimi-
computed with respect to both structural sizing and wing planform zation loop. They showed that the DLM-based solution was not con-
variables. The derivatives were computed using a combination of ana- servative and had a significantly different thickness distribution com-
lytic and automatic differentiation methods in reverse (adjoint) mode pared to the CFD-based optimal design.
and validated against the complex-step method. Opgenoord et al. [65] extended a low-order two-dimensional
The flutter constraint was used to maximizing the range of an transonic flutter prediction model [64] to wings and implemented the
idealized rectangular wing (flat plate) by varying structural thickness model into a conceptual aircraft design tool to investigate the impact of
and planform design variables. The optimal design had an increased geometric parameters and Mach number on the flutter boundary. Fur-
aspect ratio and an improved range while remaining flutter and di- thermore, they optimized the D8.0 configuration by minimizing the
vergence free. This approach provides a way to include airfoil shape maximum take-off weight and fuel burn with and without a transonic
variables in the future to perform full aerostructural design optimiza- flutter constraint. Enforcing the flutter constraints resulted in lower
tion [176] subject to flutter constraints. optimal aspect ratio and a weight penalty or lower fuel burn reduction
Several authors have focused on including aerodynamic non- compared to the case when the flutter constraint was omitted. Opgen-
linearities in flutter analysis to optimize transonic configurations. oord et al. [66] also used the developed flutter model to optimize the
Stanford et al. [144] evaluated six different novel tailoring schemes internal lattice structure of a wing by minimizing weight with and
employed in mass minimization optimization. They analyzed the flutter without enforcing a flutter constraint in addition to stress and buckling
characteristics using the pk method and the commercial software constraints. The optimal design including the flutter constraint showed
ZTRAN [50] to retain aerodynamic nonlinearities. The nonlinear only a slight mass increase thanks to the appropriate aeroelastic tai-
higher-fidelity Euler code, ZEUS, was used to compute steady back- loring of the lattice structure.
ground flow at multiple transonic Mach numbers for a fixed cruise While several authors have performed flutter-constrained optimi-
shape. Using these steady-state CFD solutions as an input, the linearized zations using nonlinear aerodynamic models, examples of flutter con-
unsteady loads were computed for a range of reduced frequencies using straints considering nonlinear structures are more rare due to the more
time-linearized transonic small disturbance (TSD) analyses about the recent interest in optimizing very flexible aircraft. Variyar et al. [7]
equilibrium solutions. developed a framework for MDO of geometrically nonlinear aircraft
The system damping values were forced to be under a stability subject to flutter constraints by coupling the SUAVE design tool [180]
boundary, similarly to the approach by Ringertz [141]. The transonic with the ASWING nonlinear aeroelastic solver [181]. They developed
aerodynamic loads were computed offline before the optimization and an interface to convert the arbitrary aircraft designs output by SUAVE
their derivatives were obtained by differentiating the eigenvalue pro- into equivalent nonlinear beam models for ASWING to drive structural
blem [134]. Flutter, stress, and buckling constraints computed in this and aeroelastic analyses and to post-process the results. The flutter
study were all aggregated using a KS function [147]. They considered speed was obtained iteratively by computing the eigenvalues of the
the fixed-mode derivatives to improve computational efficiency, an statically deformed aircraft at different flight conditions.
approach that does not allow for shape changes. This MDO framework was used to optimize the Sugar VOLT strut-
They studied the undeflected Common Research Model (uCRM) braced aircraft [182] for minimum fuel burn subject to mission, man-
wing [176] and obtained six different optimal wing structures corre- euver, gust, and flutter constraints. The flutter constraint was im-
sponding to different tailoring schemes, all for the same operating plemented by imposing a minimum flutter speed, and derivatives with
condition and setup. The six tailoring schemes considered for the respect to the design variables were obtained via finite differences. The

14
E. Jonsson, et al. Progress in Aerospace Sciences 109 (2019) 100537

authors performed three MDO cycles by adding the maneuver, gust, and requiring a search of the flutter point (which may be located well
flutter constraints to the mission constraints. Despite having better outside of the flight envelope) at each design iteration. The KS function
performance, the optimal solution achieved with only maneuver and [147] was used to smooth the effect of mode switching in the constraint
gust constraints experienced flutter within the flight envelope, high- value by aggregating the less damped modes, which yielded smooth
lighting the need for a flutter constraint in the design process. gradients. The authors presented preliminary results for a medium fi-
Lupp and Cesnik [130] studied the effect of a flutter constraint in- delity three-dimensional aerodynamic panel code coupled with the
cluding geometrical nonlinearities on the design of a BWB. They ex- structural finite-element code TACS [184]. Although no detailed opti-
tended the University of Michigan's Nonlinear Aeroelastic Simulation mization was performed, they performed a preliminary study on the
Toolbox (UM/NAST) to utilize AD in reverse mode to determine cou- uCRM benchmark [176].
pled aeroelastic derivatives including geometrical nonlinearities. They Beran et al. [173] developed a fast adjoint method to compute de-
proposed an algorithm to increase the computational efficiency of the rivatives of flutter points computed via the Hopf-bifurcation method for
gradient evaluation for a geometrically nonlinear aeroelastic analysis. gradient-based MDO. The approach was applied to the highly flexible
The sample optimization formulation included a geometrically non- cantilevered wing studied by Tang and Dowell [81,82]. Both geometric
linear beam-based flutter constraint using a KS aggregation [110] to nonlinearities due to the structure and aerodynamic effects described
obtain a scalar constraint for the entire flight envelope. The authors ran using the ONERA stall model [82] were considered when computing the
three fuel burn minimization cases: with a strength constraint, with a flutter point and its derivatives with respect to aerodynamic and
linear flutter constraint, and with a geometrically nonlinear flutter structural design variables. While no optimization study was presented,
constraint with the wing chord distribution, wing box size, and wing the authors outlined the future work required to apply the metho-
box thickness as design variables. While the linear flutter constraint dology: validate the derivatives, assess the computational cost com-
became active over the strength constraint, it was not conservative pared to alternate time- and frequency-domain flutter prediction
compared to the geometrically nonlinear constraint. They concluded methods, and develop the handling of mode switching to avoid dis-
that a geometrically nonlinear flutter constraint is needed for very continuous flutter points.
flexible aircraft. Some rare examples of optimizations with flutter constraints based
Xie et al. [131] used a flutter constraint to minimize the weight of a on time-accurate analyses are also found in the literature, but they have
very flexible wind tunnel model. The purpose of this constraint was to been restricted to simple problems. In an early work, Holden [185]
ensure flutter within the wind tunnel speed range. The optimization applied a colocation method to constrain the aeroelastic response en-
problem coupled a geometrically nonlinear beam solver with a vortex velope for a wing optimization problem. More recently, Mani and
lattice code for the static aeroelastic analysis, while a doublet lattice Mavriplis [99] were among the first to present a fully coupled unsteady
code was used to compute the unsteady loads. Since they use a gradient- adjoint for aeroelastic optimization. They demonstrated their method
free algorithm, no derivatives were computed. They compared opti- successfully in a shape optimization of two-dimensional airfoil section
mization results based on linear and geometric nonlinear beam models to suppress flutter, using a total of 32 design variables in the form of
subject to a flutter constraint. In addition to flutter, tip displacement Hicks–Henne bump functions [186]. Later, Mishra et al. [100] extended
and torsion angle constraints were also enforced. The linearly opti- previous work and presented the fully coupled unsteady adjoint for
mized configuration resulted in a wing lighter than the optimal solution three-dimensional aeroelastic problems, which was demonstrated in a
obtained with the nonlinear flutter constraint. Furthermore, the flutter shape optimization of a flexible rotorcraft configuration.
and displacements constraints were violated when the linear optimized Zhang et al. [101] developed a coupled adjoint method for unsteady
configuration was analyzed considering geometric nonlinear effects. aerostructural problems solved via time simulations. The method was
The difference in the results highlighted the need for a flutter constraint applied to an airfoil shape optimization problem with the goal of sup-
when optimizing very flexible vehicles and the importance of using pressing flutter. The aerodynamics was computed with an Euler CFD
nonlinear flutter prediction methods not only for analysis, but in design code coupled with a boundary layer code to account for viscous effects.
optimization as well. Both the continuous and discrete coupled adjoint were developed for
Bhatia and Beran [172] developed a framework to optimize ther- steady-state analyses. The discrete approach proved more promising
mally stressed nonlinear structures subject to transonic flutter con- and only this version was developed for unsteady cases. A damping
straints. They showed that including aerothermoelastic static non- objective function was proposed that used a Hilbert transform [92] of
linearities in the flutter analysis impacts the optimal design. These the nonlinear unsteady time-history. To achieve the required flutter
effects are particularly important for high-speed vehicles, which are boundary, the damping objective function was minimized to obtain a
subject to significant thermoelastic stresses when flying through the neutral response, indicating the flutter point. The authors applied the
transonic regime during reentry [183]. The authors optimized a skin methodology to the optimization of the two-dimensional (2D) Isogai
panel with respect to the thickness and density distributions to mini- airfoil [37,38] to suppress flutter. Only derivatives with respect to
mize the mass subject to a flutter constraint. The structure was modeled shape variables were computed and the airfoil shape was parametrized
using a Timoshenko beam finite element with nonlinear von Kármán using 48 Hicks–Henne bump functions [186]. A neutrally stable (zero-
strain, while the transonic flow was solved via a finite-element dis- damping) configuration was obtained for a given flight condition.
cretization of the Euler equations. The structure was linearized around Zhang et al. [102] extended their previous work [101], where the
the static thermoelastic response and the aerodynamics was linearized coupled adjoint was developed for time-marching simulations. Two
around the background steady transonic flow past the baseline geo- objective functions were used: one that maximized the flutter boundary
metry. Flutter was analyzed in the frequency domain as an iterative V-g and another that matched a given flutter boundary. To maximize the
solution [1] considered the system linearized around the nonlinear flutter boundary, they minimized the squared and time averaged his-
equilibrium configuration for each operating condition. The flutter tory of the lift coefficient. To match a given flutter boundary, the
speed was constrained directly. The optimal solution obtained by lin- damping value of a given time history was minimized to obtain a
earizing around the thermally stressed equilibrium configuration neutral response. B-spline curves were chosen as a parametrization
showed a mass lower than the result using an unstressed analysis. method due the large number of Hicks–Henne functions that were
The Hopf-bifurcation method has also been applied in design opti- previously needed. The authors presented steady-state optimization
mization. Kennedy et al. [156] proposed a variant of previous bi- results for a 2D airfoil optimized to match a given pressure distribution
furcation approaches [157,159] to optimize an aeroelastic system and for the 3D ONERA M6 case [187], where the objective was a
subject to a flutter constraint, which was formulated in terms of flight composite function of lift and drag. Configurations considered in the
speed rather than damping. The proposed method had the benefit of not unsteady optimization consisted of the 2D Isogai airfoil and the Goland

15
E. Jonsson, et al. Progress in Aerospace Sciences 109 (2019) 100537

wing as modeled by Kurdi et al. [188]. For the unsteady 2D airfoil, two flutter by considering the undeformed configuration at zero angle of
optimization cases were considered: flutter margin maximization and a attack for each flutter search point. Moreover, in the presence of non-
flutter boundary matching (i.e., a neutral response for the given flight linear effects multiple equilibrium points may also exist for each flight
condition). The design variables were the plunge and pitch stiffness condition, which further increases complexity and computational cost.
values. For the Goland wing, they optimized the aerodynamic shape to Stability must be analyzed about all equilibrium points, otherwise cri-
maximize the flutter speed with respect to 120 shape variables. A tical constraint values may be missed.
structural optimization of the Goland wing was also performed to For large high-fidelity models with both structural and aerodynamic
maximize the flutter speed with respect to the skin thickness. No op- nonlinearities, computing the aerostructural equilibrium points and the
timization was performed using simultaneously structural sizing and corresponding linearized systems may be computationally prohibitive.
aerodynamic shape variables. For moderately flexible configurations, structural nonlinearities can be
Kiviaho et al. [189] developed a flutter constraint using their pre- neglected, so eliminating the need to solve a nonlinear static aeroelastic
viously developed unsteady aeroelastic framework with adjoint sensi- problem at each flight condition. However, transonic aerodynamic
tives [103,104]. The flutter constraint is based on a matrix pencil nonlinearities still require identifying a linearized unsteady aero-
method [90] applied to a time-history, which estimates the damping dynamic model for each steady background flow solution. Computing
based on most critical aeroelastic modes. Two methods are proposed, a and retaining aerodynamic nonlinearities to accurately predict the
direct flutter point evaluation which finds the flutter point where the flutter point in the transonic regime remains a big challenge due to the
dynamic pressure lower bound is specified as the design flight condi- large computational cost associated with CFD. To address this problem,
tion, and a flutter margin or clearance approach where the dynamic some efforts tried to preserve the computational efficiency of lower-
pressure times some tolerance is fixed. The direct method was de- fidelity methods while retaining the nonlinear physics modeled by the
monstrated in a single design variable optimization of an airfoil were higher-fidelity CFD methods [50,52]. Other works proposed time-line-
the dynamic pressure is minimized subject to the flutter constraint in arized CFD solvers [53,57] or reduced- or low-order models calibrated
order to identify the flutter point. using CFD [62,64]. Despite these progresses, optimizing aircraft subject
to transonic flutter constraints is still an open problem, particularly
4.3. Open problems when using gradient-based methods and when seeking both structural
sizing and aerodynamic shape changes.
Now that we reviewed methods for flutter prediction and their ap- Due to these challenges, previous work mainly optimized linear
plications to optimization (see Table 1), we can summarize the open aircraft configurations or included only aerodynamic nonlinearities
problems in goal of integrating flutter constraints into aircraft design while limiting to structural sizing. Aerostructural optimizations con-
optimization. sidering both aerodynamic and geometric structural nonlinearities used
Gradient-based optimization is the preferred choice for optimiza- low-order models or optimized simple configurations to limit the
tions with respect to large numbers of design variables. When enforcing computational cost [7,130]. No previous work considered both struc-
a flutter constraint in a gradient-based optimization, a serious challenge tural and aerodynamic nonlinearities in a high-fidelity MDO setting.
is varying structural and aerodynamic design variables simultaneously Additionally, rigid-body DOFs were never included in the flutter con-
to optimize the aircraft external shape, planform, and internal sizing. straints, which may lead to unfeasible designs for configurations prone
Most of previous work optimized only structural sizing, and only a few to coupled rigid-elastic instabilities.
efforts included airfoil shape and wing planform design variables as Few studies used alternatives to frequency-domain eigenvalue
well. Furthermore, simplifying assumptions like the fixed-mode ap- analysis methods for nonlinear flutter prediction, like the Hopf-bi-
proximation were frequently used when computing derivatives to limit furcation method or time-domain simulations. Such studies are rare
computational cost [109,138,139]. These assumptions are adequate for because the high computational cost of time-accurate analyses makes
a structural optimization, but they can lead to inaccurate results when them prohibitive to optimize complex configurations. For this reason,
varying aerodynamic properties because this can cause significant these applications were limited to simple configurations and frequently
changes in the mode shapes at each design iteration. Therefore, gra- two-dimensional problems.
dient-based optimizations with respect to structural, planform, and
shape variables need approaches that take into account the derivatives 5. Post-flutter analysis in aircraft design optimization
of the mode shapes when computing the derivatives of the flutter
constraints. Few examples of these approaches applied to simplified As discussed in Section 3, flutter analysis is not sufficient to char-
configurations or using lower-fidelity models are available in the lit- acterize the dynamic response of nonlinear aeroelastic systems. For
erature [7,110,130]. However, they still have to be demonstrated on these systems, destabilizing nonlinear effects may cause LCOs even
practical configurations parametrized by large number of structural and before reaching the flutter boundary, if the disturbances are sufficiently
aerodynamic design variables. large (this is the subcritical post-flutter scenario shown in Fig. 6).
A second major challenge is developing efficient flutter analysis Preventing or mitigating this behavior through design optimization
models and methods applicable in the presence of aerodynamic or requires constraining not only flutter, but post-flutter characteristics as
structural nonlinearities. In previous aircraft optimizations including well.
linear flutter constraints, the natural choice was to analyze flutter in the Characterizing and constraining the LCO behavior of aircraft re-
frequency-domain due to the availability of well-established and com- quires the computation of bifurcation diagrams like the ones shown in
putationally efficient eigenvalue-analysis methods. The computational Fig. 6. Several methods exist for this purpose. Standard techniques to
cost of these methods is a big challenge when including aerodynamic or obtain bifurcation diagrams are dependent on the mathematical re-
structural nonlinearities because the flutter point depends on the presentation of the system dynamics in the form of Eq. (13). We refer to
equilibrium state. For each flight condition considered in the flutter these techniques as model-based prediction methods. However, a
search, the steady background flow solution (in the case of aero- mathematical representation of the system is not always available,
dynamic nonlinearities) or the coupled aerostructural equilibrium (in particularly for complex aircraft configurations that would require too
the case of both aerodynamic and structural nonlinearities) needs to be much theoretical development or computational effort. Therefore,
determined first. Next, the linearized model about each equilibrium model-free bifurcation prediction approaches that use only output data
point must be identified for computing the aeroelastic eigenvalues and have been recently proposed and applied to aeroelastic systems.
determine at which point modal damping vanished. This process must In this section, we summarize the theory of model-based and model-
be repeated for each flight condition, while linear methods analyze free methods to predict post-flutter behavior and discuss examples of

16
E. Jonsson, et al. Progress in Aerospace Sciences 109 (2019) 100537

their application to optimization problems and open challenges in this Stanford and Beran [195] developed a medium-fidelity nonlinear
field. aeroelastic framework for implicit time marching with direct compu-
tation of analytic derivatives and used it to optimize a wing structure
5.1. Model-based prediction methods for minimizing the LCO amplitude. Kennedy and Martins [196] devel-
oped a fully coupled adjoint-based derivative computation method for
Model-based post-flutter prediction methods can be broadly classi- time-dependent aeroelastic optimization of flexible aircraft, using a
fied under four categories: time-domain methods, frequency-domain panel code and a finite element code [184]. Recent work has focused on
methods, nonlinear perturbation methods, and continuation methods developing analytic adjoint derivatives for unsteady aeroelastic com-
[79]. Time- and frequency-domain methods compute limit cycles for a putations involving CFD solutions and demonstrating their feasibility
given flight speed (or other flight parameters). Each speed corresponds for optimization [99,104,189]. However, these approaches have not so
to a point in the bifurcation diagram, which must be constructed point- far been used for constraining LCOs of nonlinear aeroelastic systems.
by-point by means of a sweep on speed. Nonlinear perturbation In contrast to time-marching methods, time-periodic approaches
methods seek a reduced-order representation of Eq. (13) in a neigh- skip the initial transient phase and only compute fully developed per-
borhood of the Hopf bifurcation point, which corresponds to the flutter iodic solutions. Time-periodic methods consist of cyclic finite-difference
point, and use the reduced-order model to directly obtain a local ap- methods [79], cyclic spectral element methods [197–199], and
proximation of the bifurcation diagram. These methods are accurate shooting methods [19,79]. A related class of methods is based on finite
close to the flutter point, but they lose accuracy for large perturbations. elements in time and on combined methods for simultaneous spatial
Finally, continuation methods trace a branch of the bifurcation diagram and temporal model order reduction [200,201].
starting from a known point obtained from a time- or frequency-domain Cyclic finite differences apply a time-marching scheme to a period T
analysis method. of the LCO (between t = t 0 and t = t 0 + T ), and enforce the solution
periodicity at the time boundaries. The LCO period is discretized in M
5.1.1. Time-domain methods non-dimensional time intervals defined by the points 0, …, M = 0 + 1,
Time-domain methods can be either time marching or time periodic. where = t / T is the non-dimensional time variable. The governing
Both types discretize the time domain to compute the transient re- equations, Eq. (13), can be approximated at the mid-point of the ith
sponse. Time-marching methods compute the response to a perturba- time interval using a central-difference scheme for the time derivative
tion about an equilibrium point, while time-periodic methods compute and the trapezoidal rule for the right-hand side (single-step Adams–-
the fully developed LCO directly. Moulton method), yielding [79],
Time-marching methods are the most common approach to char-
acterize post-flutter behavior. These methods integrate Eq. (13) from xi xi = T [f (x i; U ) + f (xi 1; U )],
the initial time up to the development of a periodic solution. This can be
1
2 (18)
done using a variety of explicit or implicit time integration schemes,
where is the non-dimensional time step, and i = 1, …, M , and the
although implicit schemes are preferred due to the better stability
vector x lists the N system DOFs. Writing an equation like this for each
characteristics that allow larger time steps, and thus lower computa-
time interval and adding the periodicity condition x 0 = xM yields a
tional times.
system of NM nonlinear algebraic equations in the NM + 1 unknowns
The main advantage of time marching is the direct applicability to
x 0 , …, xM 1, and T. This system can be solved by applying a phase fixing
models of any fidelity, from analytic representations to high-fidelity
condition [19]. Stanford et al. [202] presented a formulation of the
models based on the coupling of FEM and CFD. Most early studies on
cyclic finite-difference method for the Crank–Nicholson time scheme
LCOs for aeroelastic systems were based on time-domain simulations
applied to a nonlinear beam in periodic motion.
[12,67,78,80,190–193]. There have also been studies that validated the
The need for a phase fixing condition stems from the periodic nature
numerical results against experiments [81,82,194]. The review papers
of the solution, and it is common to all the time-periodic and frequency-
cited previously also cover this topic [8,14,18].
domain LCO prediction methods. The periodicity condition is satisfied
Despite the insights obtained from these studies, they typically
by assuming any time t within the fully developed LCO as the reference
considered simple configurations such as 2D airfoils [190,192,193] or
initial time t 0 . Therefore, an additional condition is necessary for ob-
plates [191], cantilevered wings modeled using beam or shell structures
taining a unique solution. Several techniques are available for phase
[78,81,82,190], or low-order models of full-vehicle configurations
fixing [19]. One approach is to fix one of the unknown states and de-
[12,67,80]. This is because brute-force integration of Eq. (13) has
termine the remaining ones along with the LCO period by applying the
several drawbacks that makes it prohibitive for analyzing complex
Newton–Raphson method to the remaining system of NM equations. To
configurations and for optimizing even simple ones.
achieve convergence, the assumed value of the fixed state must lie
An obvious disadvantage of analyzing LCOs via transient simula-
within the amplitude range of the periodic solution. An appropriate first
tions is that periodic solutions may require long integration times to
guess for initializing the Newton–Raphson iterations is also necessary.
develop fully. Furthermore, tracing a portion of the bifurcation diagram
Another approach to phase fixing is to solve the undetermined system
for a fixed design requires performing several transient simulations. For
resulting from the application of a time-cyclic (or frequency-domain)
optimization, this procedure also needs to be repeated at each design
method using the Moore–Penrose pseudo-inverse [203]. Alternatively,
iteration. Therefore, time-marching LCO prediction methods are com-
one may enforce the integral orthogonality phase condition [204] or
putationally costly for design optimization of practical configurations.
the Poincaré phase fixing [205]. In any case, cyclic finite differences
Additionally, optimizing large high-fidelity models requires the use
requires the solution of a system of NM equations to determine the
gradient-based approaches, which need derivatives, typically computed
amplitudes of all the system states over the LCO period, which can
using adjoint methods. Despite their advantages for optimizations with
result in prohibitive computational costs for large-dimensional models.
respect to large numbers of design variables, adjoint methods are
The cyclic spectral element method discretizes the time domain into
computationally costly when computing gradients for dynamic re-
M elements that can have a variable time length. Transforming the time
sponses [96]. This is because a reverse time-integration is necessary to
interval t [ti 1, ti ] spanned by the ith element (which corresponds to a
obtain the adjoint vector of a time-marched system [97]. This cannot be
time step) into the non-dimensional interval [ 1,1], the solution
done simultaneously to the analysis computing the dynamic response
within the ith element can be approximated as a Kth-order Lagrange
and it requires large memory storage [98]. For low- and medium-fi-
polynomial
delity models, this issue can be mitigated by computing analytic deri-
vatives of unsteady response metrics directly.

17
E. Jonsson, et al. Progress in Aerospace Sciences 109 (2019) 100537

K
exists for a given value of flight speed and seek it in the form of a
xi ( ) = x ik k
i ( ), truncated Fourier series expansion [207],
k=0 (19)
M
where the unknown coefficients x ik are the solution vectors evaluated at x (t ) = x c 0 + [x ckcos(k t ) + x sksin(k t ) ],
the zeros of the Kth-order Lobatto polynomial, while ik ( ) is the kth- k=1 (21)
order Lagrange polynomial (k = 1, …, K ). Then, Eq. (19) is substituted
into the equations of motion written for each spectral element and the where k is the harmonic order, M is the number of harmonics, = 2 / T
system of nonlinear algebraic equations for the whole LCO period is is the fundamental angular frequency of the response, and x c0 , x ck , and
x sk N are the vectors of unknown coefficients. Assuming M = 1
assembled by enforcing the inter-element continuity and the periodicity
boundary condition. An example of this procedure applied to a non- yields the fundamental harmonic balance (HB) method, while higher-
linear beam in periodic motion is reported by Stanford et al. [202]. order harmonic balance (HOHB) methods are obtained for M > 1.
As for cyclic finite differences, spectral element methods result in an Variations of the classical balancing method include the nonlinear
algebraic system of dimension equal to the product of the number of frequency domain form [208] and the time spectral form [209], which
temporal and spatial DOFs (the additional unknown introduced by the are summarized by Hall [210].
LCO period is eliminated by phase fixing). However, since the solution To compute the limit cycle for a given condition, Eq. (21) is sub-
vector is evaluated at multiple inner points within each time interval stituted into Eq. (13) and the total coefficient multiplying each sine and
(that is, each spectral element), the number of temporal DOFs is gen- cosine function, along with the sum of the constant terms, is set to zero.
erally higher than for cyclic finite differences. This results in larger This yields a nonlinear algebraic system of N × (2M + 1) equations with
algebraic systems and thus higher computational costs. N × (2M + 1) + 1 unknowns, which can be written as
Finally, an alternative time-periodic approach for computing LCOs g (x c0, x ck , x sk , ; U ) = 0. (22)
is to couple a shooting strategy with time integration performed over a
single LCO period [206]. The shooting method starts with initial For a given value of U, this can be solved once an appropriate ad-
guesses for the initial condition, x (0)
0 : =x (t 0 ) , and the LCO period, T ,
(0) (0) ditional condition is enforced for phase fixing, as discussed for time-
and numerically integrates Eq. (13) from t = t 0 to t = t 0 + T (0) using any cyclic methods.
time-marching scheme (explicit or implicit, with fixed or variable time One advantage of HB and HOHB methods is their applicability to
step). Then, the residual systems of generic complexity, from two-dimensional airfoils
[211–219] to three-dimensional high-fidelity models [210,220]. Since
R (x 0, T )(0) : =x (t 0 + T )(0) x (t 0 )(0) (20)
they use time-periodic schemes, HB methods result in an algebraic
is evaluated and the procedure is repeated iteratively by updating the system of equations that may provide a starting point for numerical
guesses for the initial condition and the period using the New- continuation in order to obtain the bifurcation diagram for multiple
ton–Raphson method until the norm of R is below a desired threshold. values of the parameter U [19]. One weakness of HB methods is that the
Shooting methods result in N + 1 unknowns that can be solved after number of harmonics describing the limit cycle is assumed. HB solu-
adding one phase fixing condition. tions with few harmonics are inaccurate for complex bifurcation dia-
The main advantage of time-periodic methods (cyclic finite differ- grams having multiple folding points (for instance, Fig. 6(b) shows a
ences, cyclic spectral elements, and shooting) over time marching is case with one folding point), or when the response has a high harmonic
that they skip the initial transient phase and only compute the fully content [221]. HOHB methods become computationally expensive as
developed LCO. This is accomplished by solving an algebraic system, the number of unknowns in Eq. (22) increases. In general, truncating
which also allows to compute adjoint derivatives more efficiently than Eq. (21) up to an adequate number of terms requires previous knowl-
in the case of time-domain simulations [202]. On the other hand, sol- edge on the post-flutter behavior of the system, which limits the method
ving for a periodic response requires an appropriate phase fixing con- applicability for design optimization.
dition. Furthermore, adequate initial guesses are needed for con-
vergence when solving the algebraic system of equations, which 5.1.3. Nonlinear perturbation methods
requires previous knowledge of the post-flutter behavior. Another Nonlinear perturbation methods seek a reduced-order representa-
weakness of cyclic finite differences and cyclic spectral elements is that tion of the system dynamics close to the bifurcation point, which is used
the number of unknowns to be solved simultaneously may be very high, to approximate the bifurcation diagram locally [222]. These methods
because it equals the product of temporal and spatial DOFs. This pro- lose accuracy for large perturbations, but they may still provide suffi-
blem is mitigated when using shooting methods, for which the number cient information to determine whether the post-flutter behavior is
of unknowns is equal to the number of spatial DOFs. However, these subcritical or supercritical. Approximate solutions from nonlinear per-
approaches require time integrations over one LCO period in an itera- turbation analyses may be also used as first-guess for time-domain
tive process, which may be computationally expensive. Finally, time- methods, to choose the number of retained harmonics when using
periodic methods may experience convergence difficulties for highly frequency-domain methods, or to initialize numerical continuation
nonlinear problems and dynamic responses with high harmonic content schemes.
[202]. Nonlinear perturbation techniques for the analysis of nonlinear
Reduced-order modeling (ROM) techniques, such as proper ortho- systems are the method of multiple scales, the center-manifold reduc-
gonal decomposition (POD), can be used in combination with time- tion, and the method of normal forms [79,222,223]. The method of
domain methods to limit the cost of computing adjoint derivatives of multiple scales seeks an approximate solution as a series expansion
post-flutter behavior metrics [202]. However, these techniques are of function of different time scales, where the time scales are treated as
limited use in design optimization because a ROM is valid only for a independent variables. The center-manifold reduction projects the
single set of design variables, and it must be updated at each optimi- system dynamics onto the subspace defined by the bifurcating mode
zation step. ROMs can be trained with respect to the design variables as (the mode that causes flutter). The projected dynamic equations can be
well, but it is currently not feasible for large numbers of design vari- then manipulated for obtaining an analytical representation of the bi-
ables. furcation diagram by means of a coordinate transformation. This latter
procedure is known as the method of normal forms.
5.1.2. Frequency-domain methods Previous efforts have applied the method of multiple scales
Frequency-domain post-flutter prediction is based on the harmonic [155,224,225] and center-manifold reduction [161,226,227] to study
balance methods, which assume that a periodic solution of Eq. (13) nonlinear aeroelastic systems. Nayfeh [228] also compared the

18
E. Jonsson, et al. Progress in Aerospace Sciences 109 (2019) 100537

performance of the method of multiple scales and the center-manifold folding point of subcritical bifurcation diagrams, which may require the
reduction combined with the method of normal forms applied to non- computation of the post-flutter dynamic response for large disturbances
linear systems experiencing Hopf bifurcations. The two methods or for values of the bifurcation parameter much lower that the critical
yielded equivalent normal forms of the bifurcation diagram, but the one.
method of multiple scales required less manipulation of the mathema-
tical model and resulted in a compact solution for the type and am- 5.1.4. Continuation methods
plitude of LCO [79,228]. Continuation methods are a class of numerical methods that de-
Vio et al. [229] compared time-integration, frequency-domain, termine how solutions of a nonlinear system vary with certain para-
nonlinear perturbation, and continuation methods for predicting LCOs meters [230]. In the context of bifurcation theory, these methods trace
of a one-degree-of-freedom aeroelastic system. They concluded that branches in bifurcation diagrams starting from a known point along the
nonlinear perturbation methods were less accurate than other methods curve (for instance, the flutter point) and the unit vector tangent to the
for large amplitudes. Regardless, the method of multiple scales can be curve at that point. Continuation methods consists of predictor-cor-
used for obtaining a representation of bifurcation diagrams locally rector and piecewise-linear methods [79]. Predictor-corrector methods
around the flutter point suitable for enforcing post-flutter constraints in consist of a parametrization strategy for the branch of solutions to be
optimization [155] (see also Section 5.2). For this reason, we present continued, a predictor that gives an initial guess for the new point on
the key steps in the method below. the branch based on a previous known point, a corrector to adjust the
Consider the following expansion in a neighborhood of the Hopf prediction, and a step-control strategy for determining the increment in
bifurcation point (xF ; UF ) the continuation parameter as moving along the branch. On the other
x = xF + x̂, hand, piecewise-linear methods follow a piecewise-linear curve that
approximates a branch of solutions.
U= UF + ˆ,
2U (23) In this section, we discuss only predictor-corrector methods because
they are the most popular choice for tracing bifurcation diagrams using
where > 0 is a small non-dimensional quantity. Expanding Eq. (13) as
numerical continuation. Numerical continuation is frequently applied
a Taylor series about (xF ; UF ) and using the equilibrium condition yields
to trace Hopf bifurcation diagrams starting from a known point given
xˆ = Axˆ + ˆ Bxˆ
2U + C (ˆx , xˆ ) + 2 D (ˆ
x, xˆ , xˆ ) + …, (24) by a HB solution [18]. The nonlinear system (22) is used below as a
practical example. This system can be rewritten as the nonlinear alge-
where B: = f/ U , C is generated by a vector-valued symmetric bilinear
braic system
form, and D by a vector-valued symmetric trilinear form. These three
terms can be evaluated via finite differences [155] or any other method g (q; U ) = 0, (29)
for computing derivatives [31]. The solution of Eq. (24) close to the
where q is the vector of unknowns. Given a solution (q 0, U0) ,
M
Hopf bifurcation point is locally sought as
numerical continuation computes how q varies for small variations of U
xˆ = xˆ 1(T0, T2) + xˆ 2 (T0, T2) + 2x
ˆ 3 (T0, T2 ) + …, (25) such that Eq. (29) is satisfied. The existence of a unique solution
q1 = q 0 + q for U1 = U0 + U in a neighborhood of (q 0, U0) is ensured
where T0 = t and T2 = are different time scales treated as in-
2t
by the Implicit Function Theorem (IFT) as long as
dependent variables. To find the solution, we substitute Eq. (25) into
Eq. (24) and equate the coefficients of the like powers of ε. After ma- det J (q0, U0) 0, (30)
nipulation, the LCO peak-to-peak amplitude a as function of the speed
perturbation is given by where J = g/ q is the Jacobian of g .
Several predictor-corrector schemes exist with different para-
a= Û 1r / 2r , (26) metrizations, predictor and corrector methods, and parameter incre-
ment strategies. These schemes can be classified under two broad ca-
where
tegories: natural parameter and (pseudo) arc-length continuation
1: = 1r +i 1i = vT Bu , schemes [231–233]. Natural parameter schemes perform the con-
tinuation by varying the control parameter (sequential continuation),
2: = 2r +i 2i = vT [2C (u , z0) + C (u¯ , z2) + 3D (u , u , u¯ )/4], (27) which for the flutter problem would be U. Arc-length and pseudo arc-
length schemes perform the continuation with respect to the arc length
and the vectors u , v N are the right and left eigenvectors of A as-
s along the solution branch, while the control parameter is treated as an
sociated with the bifurcating eigenvalue. Finally, the vectors
N are computed from
additional unknown. Both methods assume that a starting point along
z 0 , z2 T
the branch (q 0, U0) and local unit tangent vector t 0 = [q 0T , U0] are
Az0 = C (u , u¯ )/2, known, where the prime denotes the derivative with respect to s.
Natural parameter continuation is the simplest continuation
(2iI F A) z2 = C (u, u)/2. (28) strategy. The range of U is discretized in M intervals defined by the
Equation (26) describes the type and amplitude of the LCO in a values U0, …, …UM , and the solution qi obtained for Ui is used as the
neighborhood of the bifurcation point. The bifurcation is subcritical if zero-th order predictor for the solution qi + 1 corresponding to the
parameter value Ui + 1. Then, the predictor is corrected using a
1r and 2r have opposite sign, since there exists a real solution for the
LCO amplitude for Û < 0 (below the flutter speed). The bifurcation is Newton–Raphson iteration scheme. Therefore, natural parameter con-
supercritical if 1r and 2r have the same sign, and this has smaller LCO tinuation requires that the system Jacobian matrix be nonsingular at
amplitude for larger 2r . any point along the solution path, which causes the method to fail at
One advantage of the method of the multiple scales is that it gives a folds in the bifurcation diagram.
compact solution for the LCO amplitude as a function of the speed In (pseudo) arc-length continuation schemes, the solution path is
perturbation Û and of the system parameters described by 1,2 . parametrized by the arc-length, while U is treated as an unknown
However, computing the matrices A , B , C , D is costly for large high- quantity. Therefore, Eq. (29) becomes
fidelity models and these matrices must be updated at each optimiza- g [q (s ); U (s )] = 0. (31)
tion step. Additionally, the method is not appropriate to constrain the
bifurcation diagram for large amplitudes (far from the bifurcation Starting from a known solution (q 0, U0) for s = s0 , a tangent (first-
point). For instance, the method may not be adequate to constrain the order) predictor is used to approximate the solution for s1 = s0 + s as

19
E. Jonsson, et al. Progress in Aerospace Sciences 109 (2019) 100537

Starting from this representation, the bifurcation diagram can be found

using different approaches. However, accurate models of complex
realistic systems are not always available; they may be inaccurate due
to simplifying assumptions and uncertainties, or they may require too
much theoretical development, computational effort, or both.
Model-free techniques that use only output data to forecast bi-
furcations have been proposed and applied to different types of systems
[234]. These methods exploit signals of approaching bifurcations in the
system response to predict the bifurcation point and diagram, or at least
a portion of it. Although these techniques are naturally oriented to-
wards forecasting bifurcations based on experimental measurements,
they can also use data from numerical models. In this case, the model is
treated as a black-box, where the model-free technique seeks to ap-
proximate the bifurcation diagram without operating on the mathe-
matical representation of the system itself, but only on the output data
it produces.
A model-free method for forecasting Hopf bifurcations in aeroelastic
systems was proposed based on the critical slowing down (CSD) phe-
Fig. 9. Bifurcation forecasting based on the CSD using the method (adapted nomenon [235]: when the system approaches the flutter point, tran-
from Ghadami et al. [237]). sient responses to perturbations last longer before the original equili-
brium is recovered. Measuring the recovery rate from perturbations in
the pre-bifurcation regime allows the forecasting of both the flutter
q1 = q 0 + q 0 s, speed and a portion of the bifurcation diagram. The method, originally
proposed by Lim and Epureanu [236] and further extended by Epur-
U1 = U0 + U 0 s. (32) eanu and Ghadami [234] to the application to aeroelastic systems,
presents some advantages compared to standard model-based predic-
Arc-length schemes use a Newton–Raphson iteration to correct the
tion techniques. The method is applicable to large disturbances, which
predictor in Eq. (32), which still causes convergence problems at
allows the prediction of large portions of the bifurcation diagram. The
folding points in the solution branch. To overcome these issues, pseudo
disturbances may be of any type, including realistic perturbations ex-
arc-length continuation schemes apply the Newton–Raphson correction
perienced by the structure during operation. The approach is applicable
the augmented system
to both low- and high-frequency oscillatory recoveries, and requires few
g (q1; U1) = 0 time histories of the system post-flutter response obtained only in the
, pre-bifurcation regime. The model is applicable to both experimental
h (q1; U1) = 0 (33)
and computational data. In this latter case, the model is treated as a
where h (q1; U1): =n 0 t 0 and black-box and its fidelity does not influence the application of the
prediction method. This model-free technique has been applied to
q1 q1 characterize the post-flutter behavior of a two-dimensional airfoil [234]
n0 =
U1 U1 (34) and of a idealized cantilevered wing [237]. The main steps in the for-
mulation of Ghadami and Epureanu [234] are summarized below and
is the vector from (q1 , U1 ) to the corrected solution (q1, U1) . The illustrated in Fig. 9.
system, Eq. (33), typically has nonsingular Jacobian at folding points, Considering a nonlinear system described by one degree-of-
such that the numerical continuation does not encounter convergence freedom, r, the change rate of response amplitude close to the bi-
issues. This procedure corresponds to seeking the corrected solution as furcation point can be expanded as
the intersection between n 0 , which is the direction normal to t 0 passing
through (q1 , U1 ) , and the solution branch. r = r [ p (r ) + 1 (r )(U UF ) + 2 (r )(U UF )2 + …], (35)
An advantage of numerical continuation schemes and of (pseudo) where p (r ) , 1 (r ) , and 2 (r ) are polynomial functions independent of U,
arc-length schemes in particular, is that they can compute large por- and UF is the flutter speed. This expansion is with respect to the flight
tions of bifurcation diagrams. This is particularly useful for character- speed and not with respect to the amplitude, which is not assumed to be
izing subcritical bifurcation branches for large-amplitude disturbances. small. From Eq. (35), the recovery rate λ is evaluated as
Shukla and Patil [219] presented an application of numerical con-
tinuation to nonlinear control optimization for suppressing subcritical d ln r
(r ; U ): = p (r ) + 1 (r )(U UF ) + 2 (r )(U UF )2 ,
LCOs (see Section 5.3). A disadvantage of continuation schemes is the dt (36)
need for a known starting point on the solution branch. A parameter The order of the polynomial defines the minimum number of re-
sweep on speed or arc length is necessary to trace the branch, followed coveries (time responses) from perturbations necessary for the bi-
by a Newton–Raphson solution to obtain each single point. The pro- furcation forecasting. Since Eq. (36) considers a second-order expansion
cedure must then be repeated at each optimization step, incurring a of the recovery rate with respect to the bifurcation parameter, the
high computational cost for large high-fidelity models. Finally, when forecasting requires at least three time responses. However, a higher
numerical continuation is directly applied to the algebraic systems re- number can be considered to improve the prediction accuracy. The
sulting from the application of an HB method, the predicted bifurcation perturbations can have different amplitudes and can also be of different
diagram is also a function of the number of retained harmonics (see types (e.g., gust loads, maneuver loads, control-surface inputs).
Section 5.1.2). The steps in the bifurcation forecasting are described below by as-
suming that M time responses to perturbations have been computed for
5.2. Model-free prediction methods M parameter values U = U1, …UM in the pre-flutter regime. For a fixed
amplitude r̃ , the recovery rates m = (˜; r Um ) (m = 1, …, M ) are com-
The post-flutter prediction methods described in Section 5.1 are puted using different techniques depending on the type and frequency
based on a mathematical model of the system in the form of Eq. (13). content of the recovery [234,237]. The recovery rates m = (˜; r Um )

20
E. Jonsson, et al. Progress in Aerospace Sciences 109 (2019) 100537

can be then fitted in the (U , ) plane according to Eq. (36). The inter- to one or the other of two categories—feasible or unfeasible—based on
section between the fitted curve and the axis = 0 yields the flight the classification of an initial set of training samples. For flutter, the
speed Ũ that corresponds to the point with amplitude r̃ on the bi- classification of the training samples is based on the system eigenvalues
furcation diagram (see Fig. 9). Repeating the procedure for different (frequency-domain approach) or on the time history of the mechanical
values of r̃ gives a portion of the bifurcation diagram and the flutter energy during the recovery from a perturbation (time-domain ap-
speed, where the flutter speeds corresponds to r̃ = 0 . The bifurcation proach). For subcritical post-flutter behavior, the classification was
diagram can be forecast up to the maximum response amplitude r̃max made by detecting discontinuities in the response amplitude using the
observed during the recoveries, which is not assumed to be small. K-means clustering technique [244]. The methodology was demon-
Handling large-dimensional systems requires additional considera- strated for a two-dimensional airfoil with cubic plunge and pitch stiff-
tions because the recoveries may involve contributions from many nesses. The subcritical post-flutter failure boundary was used to mini-
modes. To accurately evaluate the recovery rate, the contribution from mize the sum of the cubic stiffnesses under a probabilistic constraint,
the bifurcating mode—the most affected by the CSD [234,237]—must where the stiffness values were uncertain design variables.
be isolated beforehand in the recoveries for the DOFs of interest. Stanford et al. [202] compared the performance of different time-
Therefore, both projection onto the space spanned by the bifurcating domain methods for gradient-based optimization of nonlinear struc-
mode [234] and frequency-filtering [237] are used. However, the ap- tures undergoing limit cycles. They analyzed implicit time-marching,
propriate approach to isolate the contribution in the dynamic response cyclic finite differences, and cyclic spectral elements, each with and
that better shows the effect of the CSD needs further investigation in without a preliminary POD model order reduction. The test case was an
order to apply the method to general aeroelastic systems described by actuated nonlinear beam discretized in finite elements, which was op-
many DOFs. This is particularly true when recoveries are measured far timized with respect to the element cross-sectional properties to either
from the flutter point, because the system modes vary as the critical maximize or minimize the tip deflection, where no constraints were
mode may switch or be a hump mode when approaching the bifurca- enforced. The authors pointed out advantages and disadvantages of
tion. time-domain prediction methods (see Section 5.1.1) applied to opti-
mization. The main conclusion was that time-periodic schemes enable
5.3. Applications to optimization more efficient adjoint derivative computation than time marching, but
they require previous knowledge on the post-flutter response and suffer
Optimization studies with post-flutter constraints are rare in the convergence issues for highly nonlinear problems or high harmonic
literature. This is because modeling post-flutter efficiently is challen- content. In a later work, Stanford and Beran [154] performed an LCO-
ging and because contemporary aircraft configurations do not usually amplitude minimization of a cantilevered wing using a medium-fidelity
experience global LCOs within the flight envelope. Previous efforts are nonlinear time-simulation framework with analytic derivatives. They
summarized in Table 2 and reviewed below. These efforts used time pointed out that while computing LCOs via time-marching may be
marching (TM), time cyclic (TC),harmonic balance (HB), nonlinear feasible for low- and medium-fidelity gradient-based optimization, the
perturbation (NP), and numerical continuation (NC) methods for post- cost of this approach remains prohibitive for large high-fidelity models
flutter prediction and mainly considered simple systems parametrized parametrized by many design variables.
by few design variables. As mentioned previously in Table 1, optimi- Stanford and Beran [239] also used cyclic spectral elements com-
zations including flutter constraints considered combinations of struc- bined with POD to perform a gradient-based structural optimization of
tural, planform, and airfoil shape design variables. In contrast, opti- a geometrically nonlinear cantilevered wing. The wing was modeled as
mization including post-flutter constraints have considered either only a plate-type structure and discretized in finite elements, and was as-
structural sizing or only airfoil shape, but no planform changes. sumed to operate above its flutter speed in supersonic flow. The aero-
Missoum et al. [238] proposed a two-step methodology for the de- dynamics was modeled via piston theory. The optimization minimized
sign optimization of nonlinear aeroelastic systems under probabilistic the wing mass with respect to the thickness distribution for both a
flutter and subcritical post-flutter constraints. In the first step, explicit deterministic and a probabilistic constraint on the supercritical LCO
failure boundaries are constructed in the design space. In a second step, amplitude. The reliability-based optimization starting from the de-
these boundaries are used in a reliability-based optimization. The terministic optimal design increased the reliability index with a limited
failure boundaries are obtained from a support vector machine classi- mass increment with a similar thickness distribution pattern. The au-
fier [243]. This is a learning model that classifies samples as belonging thors discussed the scalability of the approach to high-fidelity

Table 2
Summary of sample efforts that included a post-flutter analysis in a design optimization problem.
Design variables

Models Shape

Effort Methoda Aerodynamicsb Structurec Objectived Structural Airfoil Planform N Algorithme

Missoum et al. [238] TM Theodorsen Mass-spring Min stiffness • 2 GB-Finite difference

Stanford et al. [202] TM, TC – NL beam Min/Max tip displ. • 50 GB-Adjoint
Stanford and Beran [154] TM VLM Shells Min LCO • 600 GB-Direct
Stanford and Beran [239] TC Piston theory Shells Min mass • 800 GB-Adjoint
Stanford and Beran [155] NP ONERA stall NL beam Min LCO • 40 GB-Finite difference
Thomas et al. [213,240] HB RANS Mass-spring Max fl. speed • 28 GB-Adjoint
Shukla and Patil [219] HB + NC Theodorsen Mass-spring Min control cost • 16 GB-Direct
He et al. [241,242] TS Euler Mass-spring Max fl. speed • 4 GB-Adjoint

a
Methods: TM—Time marching, TC—Time cyclic, HB—Harmonic balance.
b
Aerodynamics: VLM—Vortex lattice method, RANS—Reynolds average Navier–Stokes.
c
Structures: NL—Nonlinear.
d
Objective: LCO—Limit cycle oscillation, fl.—flutter, displ.—displacement.
e
Algorithm: GB—Gradient based, GF—Gradient free.

21
E. Jonsson, et al. Progress in Aerospace Sciences 109 (2019) 100537

multidisciplinary problems. Although the combination of a cyclic subcritical post-flutter behavior of the open-loop system, resulting in a
scheme and POD enabled efficient adjoint derivative computation, the closed-loop system with a supercritical bifurcation diagram.
methodology still relies on time marching. Computing the POD basis at He et al. [241] proposed a block-Jacobian preconditioned coupled
each optimization step and a first-guess for convergence of spectral Jacobian–free Newton–Krylov (CNK) method based on the time spectral
element solutions requires transient simulations. Therefore, the ap- (TS) method [209] to compute flutter and post-flutter efficiently. In
proach is not well scalable to large-dimensional optimization problems. contrast to the method proposed by Thomas et al. [212], O (NCSD ) CFD
Later, Stanford and Beran [155] performed a gradient-based struc- evaluations can be avoided. Here, NCSD is the structural DOFs multiplied
tural optimization of the cantilevered wing studied by Tang and Dowell by number of time instances. The state variables in the coupled system
[81,82], which showed subcritical post-flutter behavior. The wing was are thus the aerodynamic and structural states in addition to a given
modeled with beam finite elements with leading-order geometric non- motion magnitude and phase. More recently, He et al. [242] proposed a
linearities and was discretized with finite elements. The unsteady coupled adjoint method to compute flutter or post-flutter constraint
aerodynamics was modeled using the ONERA stall model [82]. The derivatives with respect to the design variables. Similar to the analysis
flutter speed was computed using a direct method [157] and the bi- method, O (NCSD ) CFD evaluations can be avoided by solving the cou-
furcation diagram in a neighborhood of the flutter point was approxi- pled adjoint, which directly computes the derivatives of the whole
mated using the method of multiple scales [79]. The optimization aeroelastic system. Preliminary results were provided for a 2D airfoil
problem minimized the supercritical post-flutter amplitude with respect section optimized to maximize the flutter speed index with respect to
the beam stiffness and inertia distribution subject to constant mass and four aerodynamic shape variables.
flutter speed constraints. The derivatives were computed by finite dif-
ferences. The authors showed the effectiveness of the methodology in 5.4. Open problems
turning the subcritical bifurcation into a supercritical one and pointed
out a contrast between the supercritical post-flutter response amplitude Post-flutter behavior has been the subject of research from a long
and the flutter speed. The optimal solution obtained using the method time, and several well-established theoretical and computational
of multiple scales was verified by comparing the bifurcation diagram methods have been developed to study the post-flutter nonlinear dy-
with the one obtained using the spectral element method, showing that namic response of aeroelastic systems. Time-domain simulations, peri-
the nonlinear perturbation analysis yielded a conservative result. odic time schemes, harmonic balance, nonlinear perturbation methods,
Thomas et al. [240] presented a discrete adjoint approach to com- and numerical continuation were all applied to analyze nonlinear
pute derivatives with respect to geometric design variables for an configurations. However, previous work mainly focused on simple
aeroelastic system analyzed using the HB method (see Section 5.1.2). systems like 2D airfoils or cantilevered wings. Furthermore, con-
The work extended a discrete adjoint previously developed for a CFD temporary aircraft typically show linear behavior, so integrating post-
code capable of solving the Reynolds-Averaged Navier–Stokes (RANS) flutter constraints into design has been more a research interest than a
equations for transonic flow conditions in the frequency domain [245]. practical industrial need.
Derivatives with respect to shape were obtained using AD and verified In order to include post-flutter constraint into aircraft design, one
using finite differences. The proposed method was used to redesign the major open problem is predicting the post-flutter behavior of realistic
NLR 7301 airfoil for improved flutter speed and LCO characteristics. full-scale aircraft with adequate fidelity and feasible computational
They showed that a small increase in thickness-to-chord ratio improved cost. This is challenging due to difficulties in developing accurate dy-
the flutter speed and LCO behavior, such that the reduced velocity was namically nonlinear models and the computational effort of performing
significantly increased while keeping the same baseline pitching am- repeated post-flutter computations, which is costly even for simple
plitudes. In contrast, small variations in camber reduced the flutter systems. For these reasons, only few efforts included post-flutter con-
speed and deteriorated the LCO characteristics, namely, the reduced straints in an optimization. These optimizations considered systems
velocity decreased for the same baseline pitching amplitudes. modeled analytically or simple beam-like or plate-like finite element
Thomas and Dowell [246] applied their previously developed models and used a small number of design variables. The post-flutter
aeroelastic adjoint implementation [240] to optimize NLR-7301 airfoil response was typically analyzed using time-domain methods, either
for a transonic flow condition. They used a class-shape function trans- direct time-marching or time-cycling schemes, which are hardly scal-
formation technique to parametrize the geometry [247], which con- able to large high-fidelity models. Furthermore, no optimization subject
sidered 14 parameters for each surface for a total of 28 design variables. to post-flutter constraints considered both aerodynamic shape and
Three optimization cases were considered: one maximizing the flutter structural sizing variables.
speed and two LCO cases maximizing the reduced velocity subject to a The performance of different post-flutter prediction methods ap-
pitch amplitude constraint. The geometry optimized for the flutter plied to optimization are still unclear, since limited comparisons and
speed demonstrated a subcritical LCO behavior, which may call into feasibility studies were performed. Therefore, identifying methods
question the usefulness of the results. However, the airfoil optimized for providing an adequate compromise of fidelity and computational cost is
the reduced velocity demonstrated improved supercritical LCO beha- a critical issue. The methods must be scalable to large-dimensional
vior when enforcing a large pitch amplitude (4 deg). The reduced ve- systems and suited for repeated computations within an optimization
locity was improved around the specified pitch amplitude (2–5 deg), cycle. Among the traditional post-flutter analysis methods, only the
but was made worse for lower amplitudes (0–2 deg). Flutter and LCO method of multiple scales and numerical continuation of harmonic-
characteristics were both improved when enforcing a medium pitch balance solutions seem scalable. However, these methods require
amplitude (2 deg). computing several matrices for each design point, which incurs a high
Shukla and Patil [219] recently performed a control optimization to computational cost for large-dimensional models. Both methods also
suppress the subcritical post-flutter behavior of a two-dimensional involve inner Newton–Raphson iterations and may require previous
airfoil with nonlinear pitch stiffness. The bifurcation diagram was knowledge on the system behavior for reliable convergence, which is
computed using the HB method and the Monroe–Penrose numerical not available during optimization. More recent model-free methods are
continuation [230]. The optimization sought to modify the shape of the an appealing alternative for formulating post-flutter designs con-
subcritical bifurcation diagram in order to minimize the cost of the straints. However, these methods need further development to de-
nonlinear feedback control while enforcing a subcritical post-flutter monstrate their suitability for practical wing or full-vehicle configura-
behavior. Derivatives of the objective with respect to the control vari- tions.
ables were computed analytically using the direct method. Using the Another critical problem is the choice of a suitable post-flutter
proposed method, the authors synthesized a control law to suppress the constraint metric that does not compromise performance. This is

22
E. Jonsson, et al. Progress in Aerospace Sciences 109 (2019) 100537

particularly important for avoiding subcritical oscillations before we also described the open problems to be addressed to make MDO
reaching the flutter boundary. A possible approach is to constrain the subject to dynamic aeroelastic constraints a standard practice.
folding point of the bifurcation diagram rather than the flutter point, Optimizations subject to flutter constraints have been addressed
but this may result in an overly conservative design solution or in- mainly using frequency-domain eigenvalue-analysis methods. Practical
troduce discontinuities in the constraint between subsequent optimi- MDO of complex aircraft configurations requires to use gradient-based
zation iterations. Alternatively, the type of bifurcation can be con- methods and compute derivatives using AD, adjoint methods, or a
strained instead such that subcritical behavior is changed into combination of the two to handle large numbers of design variables.
supercritical one [155,219]. However, this may have a detrimental Optimizing both the structural sizing and aerodynamic shape subject to
effect on the flutter point [155]. Additional investigations are required flutter constraints is challenging because the aerodynamic properties
to explore possible contrasts between flutter and subcritical post-flutter need to be recomputed at each design iteration. Moreover, in presence
constraints and how to effectively integrate them simultaneously in the of nonlinear effects the flutter boundary depends on the equilibrium
design process of configurations prone to nonlinear behavior. Opti- state. Therefore, the structural and aerodynamic properties must be
mizing realistic configurations using gradient-based methods also re- computed not only at each design iteration, but also for each flight
quires post-flutter constraints that are smooth with respect to the design condition for a given design. In the presence of aerodynamic or struc-
variables in order to compute their derivatives. tural nonlinearities, identifying the linearized characteristics about
Finally, uncertainties have a strong influence on post-flutter beha- nonlinear steady states at each optimization step is a challenge that has
vior: distinct, but nominally identical aircraft may experience different limited the scope of previous studies. Therefore, there is a need for new
response at the same flight condition [14], and small changes in the approaches to efficiently and robustly identify structural and aero-
natural frequencies can cause large variations in the response amplitude dynamic properties at each optimization step.
and in the bifurcation diagram [248]. Therefore, appropriate safety Optimizations subject to post-flutter (LCO) constraints have been
margins or probabilistic constraints are needed to ensure practical op- addressed only for very simple configurations and mostly using time-
timal solutions [238,239]. While several methods exist to compute domain methods, which are currently not scalable to large-dimensional
LCOs including uncertainties for a single configuration, design optimi- problems. Other techniques, like nonlinear perturbation methods and
zation requires effective approaches suited to repeated computations. numerical continuation, were also explored. However, these all require
significant computational effort for evaluating each design point and
6. Conclusions updating several quantities within the optimization cycle. Finally, the
frequency-domain methods, the harmonic balance, and the time-spec-
In this paper, we reviewed flutter and post-flutter prediction tral method have also been applied in optimization, but only for airfoil
methods and past work that addressed their integration into optimi- problems. Therefore, prediction methods for the efficient and robust
zation in the form of dynamic aeroelastic constraints. We pointed out LCO computation of practical configurations needs to be developed.
advantages and disadvantages of existing prediction approaches and Additionally, it is still unclear what metric of the nonlinear dynamic
outlined the open problems to be addressed for their application to response is the most appropriate to constrain the bifurcation diagram in
aircraft MDO, with a focus on next-generation transport vehicles. order to avoid subcritical behavior while not compromising the flutter
Optimizing these configurations for reducing fuel burn is likely to give boundary. Finally, this constraint needs to be formulated in a way that
light-weight, high-aspect-ratio structures that are prone to flutter. is suitable for gradient-based optimization.
Therefore, including flutter constraints in the optimization process is Future research should address the above challenges to effectively
critical to obtain feasible designs and avoid major modifications at apply design optimization to aircraft design subject to flutter con-
advanced design stages, which cause performance, time, and financial straints. Post-flutter and dynamic aeroelastic constraints will become
losses. Despite its potential to speed up and reduce the cost of aircraft increasingly important to address the current trend towards more
production cycle, including flutter constraints in design optimization is flexible vehicles and increasing aircraft wing spans.
challenging. Flutter analysis is computationally expensive due to the While this paper focuses on aircraft design optimization applica-
number of conditions that must be considered and the large design tions, the methodologies and challenges discussed here are broadly
space required to parametrize realistic configurations. Optimizing the applicable to the optimization of other engineering systems subject to
structural sizing and the aerodynamic shape simultaneously is chal- stability and post-critical constraints.
lenging because the aerodynamic properties need to be recomputed at
each design iteration. This is particularly critical for transonic flow Acknowledgments
conditions, due to the high computational cost of CFD analyses.
Next-generation high-aspect ratio aircraft are likely to experience The material is based upon work supported by Airbus in the frame
new phenomena, such as couplings between rigid-body and elastic of the Airbus–Michigan Center for Aero-Servo-Elasticity of Very
DOFs and geometrically nonlinear effects. These effects need to be Flexible Aircraft. Special thanks to Bret Stanford (NASA Langley
considered in the flutter constraints because they significantly impact Research Center) and Eli Livne (University of Washington) for their
the flutter point. This further increases the cost of flutter analyses helpful and comprehensive suggestions.
compared to traditional linear analyses based on the undeformed
structure and frequently not including rigid-body DOFs. Moreover, References
nonlinear effects can cause subcritical LCOs even before reaching the
flutter boundary. Therefore, additional constraints on the post-flutter [1] R.L. Bisplinghoff, H. Ashley, R.L. Halfman, Aeroelasticity, second ed., Dover
response are also needed to avoid undesirable subcritical oscillations. Publications, Mineola, NY, 1996.
[2] E. Garrigues, A review of industrial aeroelasticity practices at dassault aviation for
For the MDO to produce useful results, the above phenomena must military aircraft and business jets, J. Aerospace Lab 14 (2018) 1–34, https://doi.
be captured by the models used for dynamic aeroelastic prediction, org/10.12762/2018.AL14-09.
which requires better model fidelity and accuracy. For dynamic aero- [3] R.T. Haftka, Automated Procedure for Design of Wing Structures to Satisfy
Strength and Flutter Requirements, Tech. Rep. TN D-7264, NASA, July 1973.
elastic computations to be integrated in MDO, these must be performed [4] J. Sobieszczanski-Sobieski, R.T. Haftka, Multidisciplinary aerospace design opti-
with acceptable computational costs, which demands fast and robust mization: survey of recent developments, Struct. Optim. 14 (1) (1997) 1–23,
prediction methods and efficient optimization algorithms. https://doi.org/10.1007/BF011.
[5] G.K.W. Kenway, G.J. Kennedy, J.R.R.A. Martins, Scalable parallel approach for
Starting from the review of prediction methods and previous ap- high-fidelity steady-state aeroelastic analysis and derivative computations, AIAA
plications to optimization, we offered a detailed overview on the state J. 52 (5) (2014) 935–951, https://doi.org/10.2514/1.J052255.
of the art in the integration of flutter and LCO into design. Moreover, [6] G.K.W. Kenway, J.R.R.A. Martins, Multipoint high-fidelity aerostructural

23
E. Jonsson, et al. Progress in Aerospace Sciences 109 (2019) 100537

optimization of a transport aircraft configuration, J. Aircr. 51 (1) (2014) 144–160, [36] M. Farmer, P. Hanson, Comparison of supercritical and conventional wing flutter
https://doi.org/10.2514/1.C032150. characteristics, 17th AIAA/ASME/ASCE/AHS/ASC Structures, Structural
[7] A. Variyar, T. Economon, J. Alonso, Design and optimization of unconventional Dynamics, and Materials Conference, King of Prussia, PA, 1976, https://doi.org/
aircraft configurations with aeroelastic constraints, 55th AIAA Aerospace Sciences 10.2514/6.1976-1560.
Meeting, Grapevine, TX, 2017, https://doi.org/10.2514/6.2017-0463. [37] K. Isogai, On the transonic-dip mechanism of flutter of a sweptback wing, AIAA J.
[8] F. Afonso, J. Vale, É. Oliveira, F. Lau, A. Suleman, A review on non-linear aero- 17 (7) (1979) 793–795, https://doi.org/10.2514/3.61226.
elasticity of high aspect-ratio wings, Prog. Aero. Sci. 89 (Supplement C) (2017) [38] K. Isogai, Transonic dip mechanism of flutter of a sweptback wing. ii, AIAA J. 19
40–57, https://doi.org/10.1016/j.paerosci.2016.12.004. (9) (1981) 1240–1242, https://doi.org/10.2514/3.7853.
[9] E. Livne, Integrated aeroservoelastic optimization: status and direction, J. Aircr. [39] D.B. Kholodar, J.P. Thomas, E.H. Dowell, K.C. Hall, Parametric study of flutter for
36 (1999) 122–145, https://doi.org/10.2514/2.2419. an airfoil in inviscid transonic flow, J. Aircr. 40 (2) (2003) 303–313, https://doi.
[10] E. Livne, Aircraft active flutter suppression: state of the art and technology ma- org/10.2514/2.3094.
turation needs, J. Aircr. 55 (1) (2017) 410–452, https://doi.org/10.2514/1. [40] D.B. Kholodar, E.H. Dowell, J.P. Thomas, K.C. Hall, Improved understanding of
C034442. transonic flutter: a three-parameter flutter surface, J. Aircr. 41 (4) (2004)
[11] Y. Yu, Z. Lyu, Z. Xu, J.R.R.A. Martins, On the influence of optimization algorithm 911–917, https://doi.org/10.2514/1.467.
and starting design on wing aerodynamic shape optimization, Aero. Sci. Technol. [41] R.M. Bennett, J.T. Batina, H.J. Cunningham, Wing-flutter calculations with the
75 (2018) 183–199, https://doi.org/10.1016/j.ast.2018.01.016. CAP-TSD unsteady transonic small-disturbance program, J. Aircr. 26 (9) (1989)
[12] W. Su, C.E.S. Cesnik, Nonlinear aeroelasticity of a very flexible blended-wing-body 876–882, https://doi.org/10.2514/3.45854.
aircraft, J. Aircr. 47 (5) (2010) 1539–1553, https://doi.org/10.2514/1.47317. [42] D.M. Schuster, D.D. Liu, L.J. Huttsell, Computational aeroelasticity: success, pro-
[13] C.E.S. Cesnik, R. Palacios, E.Y. Reichenbach, Reexamined structural design pro- gress, challenge, J. Aircr. 40 (5) (2003) 843–856, https://doi.org/10.2514/2.
cedures for very flexible aircraft, J. Aircr. 51 (5) (2014) 1580–1591, https://doi. 6875.
org/10.2514/1.C032464. [43] N.V. Taylor, C.B. Allen, A.L. Gaitonde, D.P. Jones, G.A. Vio, J.E. Cooper,
[14] E. Dowell, J. Edwards, T. Strganac, Nonlinear aeroelasticity, J. Aircr. 40 (5) (2003) A.M. Rampurawala, K.J. Badcock, M.A. Woodgate, M. J. d. C. Henshaw, et al.,
857–874, https://doi.org/10.2514/2.6876. Aeroelastic analysis through linear and non-linear methods: a summary of flutter
[15] E. Livne, T.A. Weisshaar, Aeroelasticity of nonconventional airplane configura- prediction in the PUMA DARP, Aeronaut. J. 110 (1107) (2006) 333–343, https://
tions-past and future, J. Aircr. 40 (6) (2003) 1047–1065, https://doi.org/10. doi.org/10.1017/S0001924000013208.
2514/2.7217. [44] E.M. Lee-Rausch, J.T. Batina, Wing flutter boundary prediction using unsteady
[16] E. Livne, Future of airplane aeroelasticity, J. Aircr. 40 (6) (2003) 1066–1092, euler aerodynamic method, J. Aircr. 32 (2) (1995) 416–422, https://doi.org/10.
https://doi.org/10.2514/2.7218. 2514/3.46732.
[17] P.P. Friedmann, Renaissance of aeroelasticity and its future, J. Aircr. 36 (1) (1999) [45] B.A. Robinson, H.T. Yang, J.T. Batina, Aeroelastic analysis of wings using the euler
105–121, https://doi.org/10.2514/2.2418. equations with a deforming mesh, J. Aircr. 28 (11) (1991) 781–788, https://doi.
[18] M. J. de C. Henshaw, K.J. Badcock, G.A. Vio, C.B. Allen, J. Chamberlain, I. Kaynes, org/10.2514/3.46096.
G. Dimitriadis, J.E. Cooper, M.A. Woodgate, A.M. Rampurawala, D. Jones, [46] M.D. Gibbons, Aeroelastic calculations using CFD for a typical business jet model,
C. Fenwick, A.L. Gaitonde, N.V. Taylor, D.S. Amor, T.A. Eccles, C.J. Denley, Non- Tech. Rep., NASA Contractor Report 4753 (1996).
linear aeroelastic prediction for aircraft applications, Prog. Aero. Sci. 43 (4) [47] E. Lee-Rausch, J.T. Baitina, Wing flutter computations using an aerodynamic
(2007) 65–137, https://doi.org/10.1016/j.paerosci.2007.05.002. model based on the Navier–Stokes equations, J. Aircr. 33 (6) (1996) 1139–1147,
[19] G. Dimitriadis, Introduction to Nonlinear Aeroelasticity, John Wiley and Sons, https://doi.org/10.2514/3.47068.
2017, https://doi.org/10.1002/9781118756478. [48] R.E. Gordnier, R.B. Melville, Transonic flutter simulations using an implicit
[20] P. Beran, B. Stanford, C. Schrock, Uncertainty quantification in aeroelasticity, aeroelastic solver, J. Aircr. 37 (5) (2000) 872–879, https://doi.org/10.2514/2.
Annu. Rev. Fluid Mech. 49 (1) (2017) 361–386, https://doi.org/10.1146/ 2683.
annurev-fluid-122414-034441. [49] R. Palacios, H. Climent, A. Karlsson, B. Winzell, Assessment of strategies for cor-
[21] J.R.R.A. Martins, A.B. Lambe, Multidisciplinary design optimization: a survey of recting linear unsteady aerodynamics using CFD or experimental results,
architectures, AIAA J. 51 (9) (2013) 2049–2075, https://doi.org/10.2514/1. International Forum for Aeroelasticity and Structural Dynamics, 2001.
J051895. [50] P. C. Chen, X. W. Gao, L. Tang, Overset field-panel method for unsteady transonic
[22] I.R. Chittick, J.R.R.A. Martins, An asymmetric suboptimization approach to aerodynamic influence coefficient matrix generation, AIAA J. 42 (9). http://dx.
aerostructural optimization, Optim. Eng. 10 (1) (2009) 133–152, https://doi.org/ doi.org/10.2514/1.4390.
10.1007/s11081-008-9046-2. [51] D. Liu, Y.F. Kao, K.Y. Fung, An efficient method for computing unsteady transonic
[23] Z. Lyu, G.K.W. Kenway, J.R.R.A. Martins, Aerodynamic shape optimization in- aerodynamics of swept wings with control surfaces, J. Aircr. 25 (1) (1988) 25–31,
vestigations of the Common Research Model wing benchmark, AIAA J. 53 (4) https://doi.org/10.2514/3.45536.
(2015) 968–985, https://doi.org/10.2514/1.J053318. [52] P.C. Chen, D. Sarhaddi, D.D. Liu, Transonic-aerodynamic-influence-coefficient
[24] A.B. Lambe, J.R.R.A. Martins, G.J. Kennedy, An evaluation of constraint ag- approach for aeroelastic and MDO applications, J. Aircr. 37 (1) (2000) 85–94,
gregation strategies for wing box mass minimization, Struct. Multidiscip. Optim. https://doi.org/10.2514/2.2565.
55 (1) (2017) 257–277, https://doi.org/10.1007/s00158-016-1495-1. [53] E. Kreiselmaier, B. Laschka, Small disturbance euler equations: efficient and ac-
[25] Z. Lyu, Z. Xu, J.R.R.A. Martins, Benchmarking optimization algorithms for wing curate tool for unsteady load prediction, J. Aircr. 37 (5) (2000) 770–778, https://
aerodynamic design optimization, Proceedings of the 8th International Conference doi.org/10.2514/2.2699.
on Computational Fluid Dynamics, Chengdu, Sichuan, China, 2014 iCCFD8-2014- [54] A.N. Pechloff, B. Laschka, Small disturbance Navier–Stokes method: efficient tool
0203. for predicting unsteady air loads, J. Aircr. 43 (1) (2006) 17–29, https://doi.org/
[26] S.E. Cox, R.T. Haftka, C.A. Baker, B. Grossman, W.H. Mason, L.T. Watson, A 10.2514/1.14350.
comparison of global optimization methods for the design of a high-speed civil [55] D. Fleischer, C. Breitsamter, CFD Based Methods for the Computation of
transport, J. Glob. Optim. 21 (4) (2001) 415–432, https://doi.org/10.1023/ Generalized Aerodynamic Forces, Springer Berlin Heidelberg, Berlin, Heidelberg,
A:1012782825166. 2013, pp. 331–338, https://doi.org/10.1007/978-3-642-35680-3_40.
[27] N. Bons, X. He, C.A. Mader, J.R.R.A. Martins, Multimodality in aerodynamic wing [56] D. Fleischer, C. Breitsamter, Efficient computation of unsteady aerodynamic loads
design optimization, AIAA J. 57 (3) (2019) 1004–1018, https://doi.org/10.2514/ using computational-fluid-dynamics linearized methods, J. Aircr. 50 (2) (2013)
1.J057294. 425–440, https://doi.org/10.2514/1.C031851.
[28] J.R.R.A. Martins, P. Sturdza, J.J. Alonso, The complex-step derivative approx- [57] R. Thormann, M. Widhalm, Linear-frequency-domain predictions of dynamic-re-
imation, ACM Trans. Math Software 29 (3) (2003) 245–262, https://doi.org/10. sponse data for viscous transonic flows, AIAA J. 51 (11) (2013) 2540–2557,
1145/838250.838251. https://doi.org/10.2514/1.J051896.
[29] A. Griewank, Evaluating Derivatives: Principles and Techniques of Algorithmic [58] Y. Saad, Iterative Methods for Sparse Linear Systems, second ed., SIAM, 2003.
Differentiation, SIAM, Philadelphia, 2000, https://doi.org/10.1137/1. [59] M. Widhalm, R. Thormann, Efficient evaluation of dynamic response data with a
9780898717761. linearized frequency domain solver at transonic separated flow condition, 35th
[30] U. Naumann, The Art of Differentiating Computer Programs – an Introduction to AIAA Applied Aerodynamics Conference, Denver, CO, 2017, https://doi.org/10.
Algorithmic Differentiation, SIAM, Philadelphia, 2011, https://doi.org/10.1137/ 2514/6.2017-3905.
1.9781611972078. [60] T. Skujins, C.E.S. Cesnik, Reduced-order modeling of unsteady aerodynamics
[31] J.R.R.A. Martins, J.T. Hwang, Review and unification of methods for computing across multiple mach regimes, J. Aircr. 51 (6) (2014) 1681–1704, https://doi.org/
derivatives of multidisciplinary computational models, AIAA J. 51 (11) (2013) 10.2514/1.c032222.
2582–2599, https://doi.org/10.2514/1.J052184. [61] R. Kitson, C.A. Lupp, C.E.S. Cesnik, Modeling and simulation of flexible jet
[32] A. Jameson, Aerodynamic design via control theory, J. Sci. Comput. 3 (3) (1988) transport aircraft with high-aspect-ratio wings, 15th Dynamics Specialists
233–260, https://doi.org/10.1007/BF01061285. Conference, San Diego, CA, 2016, https://doi.org/10.2514/6.2016-2046.
[33] J.T. Hwang, J.R.R.A. Martins, A computational architecture for coupling hetero- [62] W. Mallik, J.A. Schetz, R.K. Kapania, Rapid transonic flutter analysis for aircraft
geneous numerical models and computing coupled derivatives, ACM Trans. Math conceptual design applications, AIAA J. 56 (6) (2018) 2389–2402, https://doi.
Software 44 (2018), https://doi.org/10.1145/3182393 37pp. org/10.2514/1.J056218.
[34] J.S. Gray, J.T. Hwang, J.R.R.A. Martins, K.T. Moore, B.A. Naylor, OpenMDAO: an [63] J. Leishman, Validation of approximate indicial aerodynamic functions for two-
open-source framework for multidisciplinary design, analysis, and optimization, dimensional subsonic flow, J. Aircr. 25 (10) (1988) 914–922, https://doi.org/10.
Struct. Multidiscip. Optim. 59 (2019) 1075–1104, https://doi.org/10.1007/ 2514/3.45680.
s00158-019-02211-z. [64] M.M.J. Opgenoord, M. Drela, K.E. Willcox, Physics-based low-order model for
[35] J.E. Marsden, M. McCracken, The Hopf Bifurcation and its Applicationss, Springer- transonic flutter prediction, AIAA J. 56 (4) (2018) 1519–1531, https://doi.org/10.
Verlag, 1976. 2514/1.J056710.

24
E. Jonsson, et al. Progress in Aerospace Sciences 109 (2019) 100537

[65] M.M. Opgenoord, K.E. Willcox, Aeroelastic tailoring using additively manu- Evaluation of Time-Domain Damping Identification Methods for Flutter-
factured lattice structures, 2018 Multidisciplinary Analysis and Optimization Constrained Optimization vol. 87, (2019), pp. 174–188, https://doi.org/10.1016/
Conference, Atlanta, GA, 2018, https://doi.org/10.2514/6.2018-4055. j.jfluidstructs.2019.03.011 0889–9746.
[66] M.M. Opgenoord, M. Drela, K.E. Willcox, Influence of transonic flutter on the [96] N. Kim, K. Choi, Design sensitivity analysis and optimization of nonlinear transient
conceptual design of next-generation transport aircraft, 2018 AIAA/ASCE/AHS/ dynamics, 8th Symposium on Multidisciplinary Analysis and Optimization, Long
ASC Structures, Structural Dynamics, and Materials Conference, Kissimmee, FL, Beach, CA, 2000, https://doi.org/10.2514/6.2000-4905.
2018, https://doi.org/10.2514/6.2018-0948. [97] A. Trier, S. Marthinsen, O. Sivertsen, Design sensitivities by the adjoint variable
[67] M.J. Patil, D.H. Hodges, C.E.S. Cesnik, Nonlinear aeroelasticity and flight dy- method in nonlinear structural dynamics, SIMS Simulation Conference,
namics of high-altitude long-endurance aircraft, J. Aircr. 38 (1) (2001) 88–94, Trondheim, Norway, 1996.
https://doi.org/10.2514/2.2738. [98] B. Beran, P. Stanford, M. Kurdi, Sensitivity analysis and optimization of dynamic
[68] M.J. Patil, D.H. Hodges, Flight dynamics of highly flexible flying wings, J. Aircr. systems with reduced order modeling, AIAA Aerospace Sciences Meeting, Orlando,
43 (6) (2006) 1790–1799, https://doi.org/10.2514/1.17640. FL, 2010, https://doi.org/10.2514/6.2010-1503.
[69] W. Su, C.E.S. Cesnik, Dynamic response of highly flexible flying wings, AIAA J. 49 [99] K. Mani, D.J. Mavriplis, Adjoint-based sensitivity formulation for fully coupled
(2) (2011) 324–339, https://doi.org/10.2514/1.53412. unsteady aeroelasticity problems, AIAA J. 47 (8) (2009) 1902–1915, https://doi.
[70] L.T. Niblett, The fundamentals of body-freedom flutter, Aeronaut. J. 90 (899) org/10.2514/1.40582.
(1986) 373–377, https://doi.org/10.1017/S0001924000015979. [100] A. Mishra, K. Mani, D. Mavriplis, J. Sitaraman, Time dependent adjoint-based
[71] A. Mazidi, S.A. Fazelzadeh, P. Marzocca, Flutter of aircraft wings carrying a optimization for coupled fluid–structure problems, J. Comput. Phys. 292 (2015)
powered engine under roll maneuver, J. Aircr. 48 (3) (2011) 874–883, https://doi. 253–271, https://doi.org/10.1016/j.jcp.2015.03.010.
org/10.2514/1.C031080. [101] Z. Zhang, P.C. Chen, S. Yang, Z. Wang, Q. Wang, Unsteady aerostructure coupled
[72] J.R. Jones, C.E.S. Cesnik, Nonlinear aeroelastic analysis of the x-56a multi-utility adjoint method for flutter suppression, AIAA J. 53 (8) (2015) 2121–2129, https://
aeroelastic demonstrator, 57th AIAA/ASCE/AHS/ASC Structures, Structural doi.org/10.2514/1.J053495.
Dynamics, and Materials Conference, San Diego, CA, 2016, https://doi.org/10. [102] Z. Zhang, P.-C. Chen, Q. Wang, Z. Zhou, S. Yang, Z. Wang, Adjoint based structure
2514/6.2016-1799. and shape optimization with flutter constraints, 57th AIAA/ASCE/AHS/ASC
[73] C. Cesnik, W. Su, Nonlinear aeroelastic simulation of x-hale: a very flexible uav, Structures, Structural Dynamics, and Materials Conference, San Diego, CA, 2016,
49th AIAA Aerospace Sciences Meeting, American Institute of Aeronautics and https://doi.org/10.2514/6.2016-1176.
Astronautics, Orlando, FL, 2011, , https://doi.org/10.2514/6.2011-1226. [103] J.F. Kiviaho, K. Jacobson, M.J. Smith, G. Kennedy, A robust and flexible coupling
[74] C.E.S. Cesnik, P.J. Senatore, E.M. Atkins, W. Su, C.M. Shearer, X-hale: a very framework for aeroelastic analysis and optimization, 18th AIAA/ISSMO
flexible unmanned aerial vehicle for nonlinear aeroelastic tests, AIAA J. 50 (12) Multidisciplinary Analysis and Optimization Conference, Denver, CO, 2017,
(2012) 2820–2833, https://doi.org/10.2514/1.j051392. https://doi.org/10.2514/6.2017-4144.
[75] M. Patil, Nonlinear gust response of highly flexible aircraft, 48th AIAA/ASME/ [104] K. Jacobson, J.F. Kiviaho, M.J. Smith, G. Kennedy, An aeroelastic coupling fra-
ASCE/AHS/ASC Structures, Structural Dynamics, and Materials and Co-located mework for time-accurate analysis and optimization, 2018 AIAA/ASCE/AHS/ASC
Conferences, Honolulu, HI, 2007, https://doi.org/10.2514/6.2007-2103. Structures, Structural Dynamics, and Materials Conference, Kissimmee, FL, 2018,
[76] P.W. Richards, Y. Yao, R.A. Herd, D.H. Hodges, P. Mardanpour, Effect of inertial https://doi.org/10.2514/6.2018-0100.
and constitutive properties on body-freedom flutter for flying wings, J. Aircr. 53 [105] M.J. Ablowitz, A.S. Fokas, Complex Variables: Introduction and Applications,
(3) (2016) 756–767, https://doi.org/10.2514/1.C033435. second ed., Cambridge University Press, Cambridge, UK, 2003, https://doi.org/10.
[77] R. Cavallaro, R. Bombardieri, L. Demasi, A. Iannelli, Prandtlplane joined wing: 1017/CBO9780511791246.
body freedom flutter, limit cycle oscillation and freeplay studies, J. Fluids Struct. [106] J.W. Edwards, C.D. Wieseman, Flutter and divergence analysis using the gen-
59 (2015) 57–84, https://doi.org/10.1016/j.jfluidstructs.2015.08.016 January. eralized aeroelastic analysis method, J. Aircr. 45 (3) (2008) 906–915, https://doi.
[78] M.J. Patil, D.H. Hodges, C.E.S. Cesnik, Limit-cycle oscillations in high-aspect-ratio org/10.2514/1.30078.
wings, J. Fluids Struct. 15 (1) (2001) 107–132, https://doi.org/10.1006/jfls.2000. [107] B.K. Stanford, Role of unsteady aerodynamics during aeroelastic optimization,
0329. AIAA J. 53 (12) (2015) 3826–3831, https://doi.org/10.2514/1.J054314.
[79] A. Nayfeh, B. Balachandran, Applied Nonlinear Dynamics, John Wiley and Sons, [108] W.P. Rodden, E.D. Bellinger, R.L. Harder, L.R. Center, Aeroelastic Addition to
New York, 1995, https://doi.org/10.1002/9783527617548. NASTRAN, National Aeronautics and Space Administration, Scientific and
[80] M.J. Patil, D.H. Hodges, C.E.S. Cesnik, Nonlinear aeroelastic analysis of complete Technical Information Branch, 1979.
aircraft in subsonic flow, J. Aircr. 37 (5) (2000) 753–760, https://doi.org/10. [109] P.D. Dunning, B.K. Stanford, H.A. Kim, C.V. Jutte, Aeroelastic tailoring of a plate
2514/2.2685. wing with functionally graded materials, J. Fluids Struct. 51 (Supplement C)
[81] D. Tang, E.H. Dowell, Experimental and theoretical study on aeroelastic response (2014) 292–312, https://doi.org/10.1016/j.jfluidstructs.2014.09.008.
of high-aspect-ratio wings, AIAA J. 39 (2001) 1430–1441, https://doi.org/10. [110] E. Jonsson, C.A. Mader, G.J. Kennedy, J.R.R.A. Martins, Computational modeling
2514/2.1484. of flutter constraint for high-fidelity aerostructural optimization, 2019 AIAA/
[82] D. Tang, E.H. Dowell, Limit-cycle hysteresis response for a high-aspect-ratio wing ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference, San
model, J. Aircr. 39 (2002) 885–888, https://doi.org/10.2514/2.3009. Diego, CA, 2019, https://doi.org/10.2514/6.2019-2354.
[83] H.J. Hassig, An approximate true damping solution of the flutter equation by [111] E. Anderson, Z. Bai, C. Bischof, S. Blackford, J. Demmel, J. Dongarra, J. Du Croz,
determinant iteration, J. Aircr. 8 (11) (1971) 885–889, https://doi.org/10.2514/ A. Greenbaum, S. Hammarling, A. McKenney, D. Sorensen, LAPACK Users' Guide,
3.44311. third ed., Society for Industrial and Applied Mathematics, Philadelphia, PA, 1999.
[84] J.R. Wright, E. Jonathan, Introduction to Aircraft Aeroelasticity and Loads, second [112] W.P. Rodden, E.D. Bellinger, Aerodynamic lag functions, divergence, and the
ed., Wiley, 2007, https://doi.org/10.1002/9781118700440. british flutter method, J. Aircr. 19 (7) (1982) 596–598, https://doi.org/10.2514/
[85] D.H. Hodges, G.A. Pierce, second ed., Introduction to Structural Dynamics and 3.44772.
Aeroelasticity vol. 15, Cambridge University Press, 2011, https://doi.org/10. [113] L.H. van Zyl, Aeroelastic divergence and aerodynamic lag roots, J. Aircr. 38 (3)
1017/CBO9780511997112. (2001) 586–588, https://doi.org/10.2514/2.2806.
[86] W.P. Rodden, Theoretical and Computational Aeroelasticity, Crest Pub., 2011. [114] B.K. Stanford, P.D. Dunning, Optimal topology of aircraft rib and spar structures
[87] P.C. Chen, Damping perturbation method for flutter solution: the g-method, AIAA under aeroelastic loads, J. Aircr. 52 (4) (2014) 1298–1311, https://doi.org/10.
J. 38 (9) (2000) 1519–1524, https://doi.org/10.2514/2.1171. 2514/1.C032913.
[88] R.M. Bennett, R.N. Desmarais, Curve Fitting of Aeroelastic Transient Response [115] K.L. Roger, Airplane math modeling methods for active control design, AGARD-
Data with Exponential Functions, Tech. Rep NASA Langley Research Center, 1976. CP- 288 (228) (1977) 4. 1–4. 11.
[89] W. Bousman, D. Winkler, Application on the moving-block analysis, Dynamics [116] M. Karpel, Design for active flutter suppression and gust alleviation using state-
Specialists Conference, Atlanta, GA, 1981, https://doi.org/10.2514/6.1981-653. space aeroelastic modeling, J. Aircr. 19 (3) (1982) 221–227, https://doi.org/10.
[90] Y. Hua, T.K. Sarkar, Matrix pencil method for estimating parameters of ex- 2514/3.57379.
ponentially damped/undamped sinusoids in noise, IEEE Trans. Acoust. Speech [117] M. Karpel, E. Strul, Minimum-state unsteady aerodynamic approximations with
Signal Process. 38 (5) (1990) 814–824, https://doi.org/10.1109/29.56027. flexible constraints, J. Aircr. 33 (6) (1996) 1190–1196, https://doi.org/10.2514/
[91] C.-G. Pak, P.P. Friedmann, New time-domain technique for flutter boundary 3.47074.
identification, 33rd AIAA/ASME/ASCE/AHS/ASC Structures, Structural [118] L. Morino, F. Mastroddi, R. De Troia, G.L. Ghiringhelli, P. Mantegazza, Matrix
Dynamics, and Materials Conference, Dallas, TX, 1992, https://doi.org/10.2514/ fraction approach for finite-state aerodynamic modeling, AIAA J. 33 (4) (1995)
6.1992-2102. 703–711, https://doi.org/10.2514/3.12381.
[92] N.E. Huang, Z. Shen, S.R. Long, M.C. Wu, H.H. Shih, Q. Zheng, N.-C. Yen, [119] M. Ripepi, P. Mantegazza, Improved matrix fraction approximation of aero-
C.C. Tung, H.H. Liu, The empirical mode decomposition and the hilbert spectrum dynamic transfer matrices, AIAA J. 51 (2013) 1156–1173, https://doi.org/10.
for nonlinear and non-stationary time series analysis, Proceedings of the Royal 2514/1.J052009.
Society of London A: Mathematical, Physical and Engineering Sciences, vol. 454, [120] W. Eversman, A. Tewari, Consistent rational-function approximation for unsteady
The Royal Society, 1998, pp. 903–995, , https://doi.org/10.1098/rspa.1998.0193. aerodynamics, J. Aircr. 28 (9) (1991) 545–552, https://doi.org/10.2514/3.46062.
[93] J.J. McNamara, P.P. Friedmann, Flutter boundary identification for time-domain [121] G. Pasinetti, P. Mantegazza, Single finite states modeling of aerodynamic forces
computational aeroelasticity, AIAA J. 45 (7) (2007) 1546–1555, https://doi.org/ related to structural motions and gusts, AIAA J. 37 (1999) 604–612, https://doi.
10.2514/1.26706. org/10.2514/2.760.
[94] B. Hallissy, C. Cesnik, High-fidelity aeroelastic analysis of very flexible aircraft, [122] E. Nissim, On the uniqueness of the minimum-state representation, J. Aircr. 42
52nd AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics and Materials (2005) 1339–1340, https://doi.org/10.2514/1.15105.
Conference, Structures, Structural Dynamics, and Materials and Co-located [123] E. Nissim, On the formulation of minimum-state approximation as a nonlinear
Conferences, Denver, CO, 2011, https://doi.org/10.2514/6.2011-1914. optimization problem, J. Aircr. 43 (2006) 1007–1013, https://doi.org/10.2514/1.
[95] Kevin E. Jacobson, Jan F. Kiviaho, Graeme J. Kennedy, Marilyn J. Smith, 17148.

25
E. Jonsson, et al. Progress in Aerospace Sciences 109 (2019) 100537

[124] M. Karpel, Sensitivity derivatives of flutter characteristics and stability margins for symmetric airfoil in transonic flow, AIAA J. 42 (5) (2004) 883–892, https://doi.
aeroservoelastic design, J. Aircr. 27 (1990) 368–375, https://doi.org/10.2514/3. org/10.2514/1.9584.
25281. [154] B. Stanford, P. Beran, Optimal structural topology of a platelike wing for subsonic
[125] A. Zole, M. Karpel, Continuous gust response and sensitivity derivatives using aeroelastic stability, J. Aircr. 48 (4) (2011) 1193–1203, https://doi.org/10.2514/
state-space models, J. Aircr. 31 (1994) 1212–1214, https://doi.org/10.2514/3. 1.C031185.
46632. [155] B. Stanford, P. Beran, Direct flutter and limit cycle computations of highly flexible
[126] E. Livne, W.-L. Li, Aeroservoelastic aspects of wing/control surface planform shape eings for efficient analysis and optimization, J. Fluids Struct. 36 (2013) 111–123,
optimization, AIAA J. 33 (1995) 302–311, https://doi.org/10.2514/3.12482. https://doi.org/10.1016/j.jfluidstructs.2012.08.008.
[127] M. Mor, E. Livne, Minimum-state unsteady aerodynamics for aeroservoelastic [156] G.J. Kennedy, G.K.W. Kenway, J.R.R.A. Martins, Towards gradient-based design
configuration shape optimization of flight vehicles, AIAA J. 43 (2005) 2299–2308, optimization of flexible transport aircraft with flutter constraints, 15th AIAA/
https://doi.org/10.2514/1.10005. ISSMO Multidisciplinary Analysis and Optimization Conference, Atlanta, GA,
[128] M. Mor, E. Livne, Sensitivities and approximations for aeroservoelastic shape 2014, https://doi.org/10.2514/6.2014-2726.
optimization with gust response constraints, J. Aircr. 43 (2006) 1516–1527, [157] K.J. Badcock, M.A. Woodgate, Bifurcation prediction of large-order aeroelastic
https://doi.org/10.2514/1.17467. models, AIAA J. 48 (6) (2010) 1037–1046, https://doi.org/10.2514/1.40961.
[129] E. Albano, W.P. Rodden, A doublet-lattice method for calculating lift distributions [158] W.H. Hui, M. Tobak, Bifurcation analysis of aircraft pitching motions about large
on oscillating surfaces in subsonic flows, AIAA J. 7 (2) (1969) 279–285, https:// mean angles of attack, J. Guid. Control Dyn. 7 (1) (1984) 113–122, https://doi.
doi.org/10.2514/3.5086. org/10.2514/3.8553.
[130] C.A. Lupp, C.E. Cesnik, A gradient-based flutter constraint including geometrically [159] K.J. Badcock, M.A. Woodgate, B.E. Richards, Direct aeroelastic bifurcation ana-
nonlinear deformations, 2019 AIAA SciTech Forum, American Institute of lysis of a symmetric wing based on the euler equations, J. Aircr. 42 (3) (2005)
Aeronautics and Astronautics, 2019, , https://doi.org/10.2514/6.2019-1212. 731–737, https://doi.org/10.2514/1.5323.
[131] C. Xie, Y. Meng, F. Wang, Z. Wan, Aeroelastic optimization design for high-aspect- [160] E.C. Yates Jr., Agard Standard Aeroelastic Configurations for Dynamic Response I-
ratio wings with large deformation, Shock Vib. (2017) 1–16, https://doi.org/10. Wing 445.6, Tech. Rep DTIC Document, 1988.
1155/2017/2564314. [161] M.A. Woodgate, K.J. Badcock, Fast prediction of transonic aeroelastic stability and
[132] C. Cardani, P. Mantegazza, Calculation of eigenvalue and eigenvector derivatives limit cycles, AIAA J. 45 (6) (2007) 1370–1381, https://doi.org/10.2514/1.25604.
for algebraic flutter and divergence eigenproblems, AIAA J. 17 (1979) 408–412, [162] P.S. Beran, N.S. Khot, F.E. Eastep, R.D. Snyder, J.V. Zweber, Numerical analysis of
https://doi.org/10.2514/3.61140. store-induced limit-cycle oscillation, J. Aircr. 41 (6) (2004) 1315–1326, https://
[133] P. Mantegazza, G. Bindolino, Aeroelastic derivatives as a sensitivity analysis of doi.org/10.2514/1.404.
nonlinear equations, AIAA J. 25 (1987) 1145–1146, https://doi.org/10.2514/3. [163] P. Girodroux-Lavigne, J.P. Grisval, S. Guillemot, M. Henshaw, A. Karlsson,
9758. V. Selmin, J. Smith, E. Teupootahiti, B. Winzell, Comparison of static and dynamic
[134] R.T. Haftka, Z. Gürdal, third ed., Elements of Structural Optimization vol. 11, fluid-structure interaction solutions in the case of a highly flexible modern
Springer Science & Business Media, Dordrecht, The Netherlands, 1992, https:// transport aircraft wing, Aero. Sci. Technol. 7 (2) (2003) 121–133, https://doi.org/
doi.org/10.1007/978-94-011-2550-5. 10.1016/S1270-9638(02)00007-X.
[135] D.J. Neill, E.H. Johnson, R. Canfield, Astros – a multidisciplinary automated [164] M. Tamayama, K. Saitoh, H. Matsushita, J. Nakamichi, NAL SST Arrow Wing with
structural design tool, J. Aircr. 27 (12) (1990) 1021–1027, https://doi.org/10. Oscillating Flap, Tech. Rep National Aerospace Laboratory of Japan, Tokyo,
2514/3.45976. Japan, 2000.
[136] P.C. Chen, D.D. Liu, E. Livne, Unsteady-aerodynamic shape sensitivities for air- [165] Open Source Fighter, http://www.cfd4aircraft.com/research_ecerta_osf.html, ac-
plane aeroservoelastic configuration optimization, J. Aircr. 43 (2006) 471–481, cessed: 2019-01-24.
https://doi.org/10.2514/1.10007. [166] P. Hajela, A root locus-based flutter synthesis procedure, J. Aircr. 20 (12) (1983)
[137] P.C. Chen, Z. Zhang, E. Livne, Design-oriented computational fluid dynamics- 1021–1027, https://doi.org/10.2514/3.48206.
based unsteady aerodynamics for flight-vehicle aeroelastic shape optimization, [167] M.J. Turner, Optimization of structures to satisfy flutter requirements, AIAA J. 7
AIAA J. 53 (12) (2015) 3603–3619, https://doi.org/10.2514/1.J054024. (5) (1969) 945–951, https://doi.org/10.2514/3.5248.
[138] B. Stanford, P. Beran, M. Bhatia, Aeroelastic topology optimization of blade-stif- [168] K.G. Bhatia, C.S. Rudisill, Optimization of complex structures to satisfy flutter
fened panels, J. Aircr. 51 (3) (2014) 938–944, https://doi.org/10.2514/1. requirements, AIAA J. 9 (8) (1971) 1487–1491, https://doi.org/10.2514/3.6389.
C032500. [169] C.S. Rudisill, K.G. Bhatia, Second derivatives of the flutter velocity and the opti-
[139] R.E. Bartels, B.K. Stanford, Aeroelastic optimization with an economical transonic mization of aircraft structures, AIAA J. 10 (12) (1972) 1569–1572, https://doi.
flutter constraint using Navier–Stokes aerodynamics, J. Aircr. 55 (4) (2018) org/10.2514/3.6690.
1522–1530, https://doi.org/10.2514/1.C034675. [170] L. Gwin, R. Taylor, A general method for flutter optimization, AIAA J. 11 (12)
[140] P.D. Dunning, B. Stanford, H.A. Kim, Level-set topology optimization with aero- (1973) 1613–1617, https://doi.org/10.2514/3.50657.
elastic constraints, 56th AIAA/ASCE/AHS/ASC Structures, Structural Dynamics, [171] W. Mallik, R.K. Kapania, J.A. Schetz, Effect of flutter on the multidisciplinary
and Materials Conference, Kissimmee, FL, 2015, https://doi.org/10.2514/6.2015- design optimization of truss-braced-wing aircraft, J. Aircr. 52 (6) (2015)
1128. 1858–1872, https://doi.org/10.2514/1.C033096.
[141] U.T. Ringertz, On structural optimization with aeroelasticity constraints, Struct. [172] M. Bhatia, P. Beran, Design of thermally stressed panels subject to transonic flutter
Optim. 8 (1) (1994) 16–23, https://doi.org/10.1007/BF01742928. constraints, J. Aircr. 54 (6) (2017) 2340–2349, https://doi.org/10.2514/1.
[142] M.A. Langthjem, Y. Sugiyama, Optimum shape design against flutter of a canti- C034301.
levered column with an end-mass of finite size subjected to a non-conservative [173] P. Beran, B.K. Stanford, K.G. Wang, Fast prediction of flutter and flutter sensi-
load, J. Sound Vib. 226 (1) (1999) 1–23, https://doi.org/10.1006/jsvi.1999.2211. tivities, 17th International Forum on Aeroelasticity and Structural Dynamics
[143] Y. Odaka, H. Furuya, Robust structural optimization of plate wing corresponding (IFASD), Como, Italy, 2017.
to bifurcation in higher mode flutter, Struct. Multidiscip. Optim. 30 (6) (2005) [174] M. Bhatia, R. Kapania, O. Gur, J. Schetz, W. Mason, R. Haftka, Progress towards
437–446, https://doi.org/10.1007/s00158-005-0538-9. multidisciplinary design optimization of truss braced wing aircraft with flutter
[144] B. Stanford, C.D. Wieseman, C. Jutte, Aeroelastic tailoring of transport wings in- constraints, 13th AIAA/ISSMO Multidisciplinary Analysis and Optimization
cluding transonic flutter constraints, 56th AIAA/ASCE/AHS/ASC Structures, Conference, Fort Worth, TX, 2010, https://doi.org/10.2514/6.2010-9077.
Structural Dynamics, and Materials Conference, Kissimmee, FL, 2015, https://doi. [175] T. Theodorsen, General theory of aerodynamic instability and the mechanism of
org/10.2514/6.2015-1127. flutter, (1935) Tech. rep., NACA Report 496.
[145] B.S. Kang, G.J. Park, J.S. Arora, A review of optimization of structures subjected to [176] T.R. Brooks, G.K.W. Kenway, J.R.R.A. Martins, Benchmark aerostructural models
transient loads, Struct. Multidiscip. Optim. 31 (2) (2006) 81–95, https://doi.org/ for the study of transonic aircraft wings, AIAA J. 56 (7) (2018) 2840–2855,
10.1007/s00158-005-0575-4. https://doi.org/10.2514/1.J056603.
[146] R.T. Haftka, Parametric constraints with application to optimization for flutter [177] J.W.G. Van Nunen, H. Tijdeman, K. A. N., Results of Transonic Wind Tunnel
using a continuous flutter constraint, AIAA J. 13 (4) (1975) 471–475, https://doi. Measurements on an Oscillating Wing with External Store (Data Report), Tech.
org/10.2514/3.49733. rep., DTIC Document, (1978) NLR TR 78030 U.
[147] G. Kreisselmeier, R. Steinhauser, Systematic control design by optimizing a vector [178] H. Tijdeman, J.W.G. Van Nunen, K. A. N, Transonic Wind Tunnel Tests of an
performance index, International Federation of Active Controls Symposium on Oscillating Wing with External Store, Parts I-Iv, Tech. rep., DTIC Document, NLR
Computer-Aided Design of Control Systems, Zurich, Switzerland, 1979. TR 78106 U, 1978.
[148] G.A. Wrenn, An Indirect Method for Numerical Optimization Using the [179] R. Bartels, Flexible launch vehicle stability analysis using steady and unsteady
Kreisselmeier–Steinhauser Function, Tech. Rep. CR-4220 NASA Langley Research computational fluid dynamics, J. Spacecr. Rocket. 49 (4) (2012) 644–650, https://
Center, Hampton, VA, 1989. doi.org/10.2514/1.A32082.
[149] N.M.K. Poon, J.R.R.A. Martins, An adaptive approach to constraint aggregation [180] T. Lukaczyk, A. Wendor, E. Boteroz, T. Macdonaldz, T. Momosez, A. Variyarz,
using adjoint sensitivity analysis, Struct. Multidiscip. Optim. 34 (1) (2007) 61–73, J. Veghz, M. Colonnox, T. Economon, J. Alonsok, T. Orra, C. Da Silvayy, Suave: an
https://doi.org/10.1007/s00158-006-0061-7. open-source environment for multi-fidelity conceptual vehicle design, 16th AIAA/
[150] G.J. Kennedy, J.E. Hicken, Improved constraint-aggregation methods, Comput. ISSMO Multidisciplinary Analysis and Optimization Conference, Dallas, TX, 2015,
Methods Appl. Mech. Eng. 289 (2015) 332–354, https://doi.org/10.1016/j.cma. https://doi.org/10.2514/6.2015-3087.
2015.02.017. [181] M. Drela, Integrated simulation model for preliminary aerodynamic, structural,
[151] A. Griewank, G. Reddien, The calculation of hopf points by a direct method, IMA J. and control-law design of aircraft, 40th AIAA/ASME/ASCE/AHS/ASC Structures,
Numer. Anal. 3 (3) (1983) 295–303, https://doi.org/10.1093/imanum/3.3.295. Structural Dynamics and Materials Conference, St. Louis, MO, 1999, https://doi.
[152] S.A. Morton, P.S. Beran, Hopf-bifurcation analysis of airfoil flutter at transonic org/10.2514/6.1999-1394.
speeds, J. Aircr. 36 (2) (1999) 421–429, https://doi.org/10.2514/2.2447. [182] M.K. Bradley, D. C. K., Subsonic ultra green aircraft research phase ii: N+4 ad-
[153] K.J. Badcock, M.A. Woodgate, B.E. Richards, Hopf bifurcation calculations for a vanced concept development, Tech. Rep., NASA Contract Report 2012-217556

26
E. Jonsson, et al. Progress in Aerospace Sciences 109 (2019) 100537

(2012). oscillation behavior using a harmonic balance approach, J. Aircr. 41 (6) (2004)
[183] R.J. Klock, C.E.S. Cesnik, Aerothermoelastic simulation of air-breathing hy- 1266–1274, https://doi.org/10.2514/1.9839.
personic vehicle, 55th AIAA/ASME/ASCE/AHS/SC Structures, Structural [214] L. Liu, E.H. Dowell, The secondary bifurcation of an aeroelastic airfoil motion:
Dynamics, and Materials Conference, National Harbor, MD, 2014, https://doi.org/ effect of high harmonics, Nonlinear Dynam. 37 (1) (2004) 31–49, https://doi.org/
10.2514/6.2014-0149. 10.1023/B:NODY.0000040033.85421.4d.
[184] G.J. Kennedy, J.R.R.A. Martins, A parallel aerostructural optimization framework [215] L. Liu, E.H. Dowell, Harmonic balance approach for an airfoil with a freeplay
for aircraft design studies, Struct. Multidiscip. Optim. 50 (6) (2014) 1079–1101, control surface, AIAA J. 43 (4) (2005) 802–815, https://doi.org/10.2514/1.
https://doi.org/10.1007/s00158-014-1108-9. 10973.
[185] M.E. Holden, Aeroelastic Optimization Using the Collocation Method, (1999). [216] B. Lee, L. Liu, K. Chung, Airfoil motion in subsonic flow with strong cubic non-
[186] R.M. Hicks, P.A. Henne, Wing design by numerical optimization, J. Aircr. 15 (7) linear restoring forces, J. Sound Vib. 281 (3) (2005) 699–717, https://doi.org/10.
(1978) 407–412, https://doi.org/10.2514/3.58379. 1016/j.jsv.2004.01.034.
[187] V. Schmitt, F. Charpin, Pressure Distributions on the ONERA-M6-Wing at [217] L. Liu, E.H. Dowell, J.P. Thomas, A high dimensional harmonic balance approach
Transonic Mach Numbers, Tech. Rep. Office National d'Etudes et Recherches for an aeroelastic airfoil with cubic restoring forces, J. Fluids Struct. 23 (3) (2007)
Aerospatiales, Chatillon, France, 1979 92320. 351–363, https://doi.org/10.1016/j.jfluidstructs.2006.09.005.
[188] M. Kurdi, N. Lindsley, P. Beran, Uncertainty quantification of the goland wings [218] M. Manetti, G. Quaranta, P. Mantegazza, Numerical evaluation of limit cycles of
flutter boundary, AIAA Atmospheric Flight Mechanics Conference and Exhibit, aeroelastic systems, J. Aircr. 46 (2009) 1759–1769, https://doi.org/10.2514/1.
Hilton Head, SC, 2007, https://doi.org/10.2514/6.2007-6309. 42928.
[189] J.F. Kiviaho, K. Jacobson, M.J. Smith, G. Kennedy, Application of a time-accurate [219] H. Shukla, M.J. Patil, Nonlinear state feedback control design to eliminate sub-
aeroelastic coupling framework to flutter-constrained design optimization, 19th critical limit cycle oscillations in aeroelastic systems, Nonlinear Dynam. 88 (3)
AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference, Atlanta, (2017) 1599–1614, https://doi.org/10.1007/s11071-017-3332-5.
GA, 2018, https://doi.org/10.2514/6.2018-2932. [220] E.H. Dowell, J.P. Thomas, K.C. Hall, C.M. Denegri, Theoretical predictions of F-16
[190] L. Tang, R. Bartels, P. Chen, D. Liu, Numerical investigation of transonic limit fighter limit cycle oscillations for flight flutter testing, J. Aircr. 46 (5) (2009)
cycle oscillations of a two-dimensional supercritical wing, J. Fluids Struct. 17 (1) 1667–1672, https://doi.org/10.2514/1.42352.
(2003) 29–41, https://doi.org/10.1016/S0889-9746(02)00114-7. [221] B. Liu, R.T. Haftka, L.T. Watson, Global-local structural optimization using re-
[191] E. Dowell, Nonlinear oscillations of a fluttering plate, AIAA J. 4 (7) (1966) sponse surfaces of local optimization margins, Struct. Multidiscip. Optim. 27 (5)
1267–1275, https://doi.org/10.2514/3.3658. (2004) 352–359, https://doi.org/10.1007/s00158-004-0393-0.
[192] S. Price, H. Alighanbari, B. Lee, The aeroelastic response of a two-dimensional [222] A.H. Nayfeh, Perturbation Methods, John Wiley and Sons, New York, 2000.
airfoil with bilinear and cubic structural nonlinearities, J. Fluids Struct. 9 (2) [223] A.H. Nayfeh, Method of Normal Forms, John Wiley and Sons, 1993.
(1995) 175–193, https://doi.org/10.1006/jfls.1995.1009. [224] H. Gilliatt, T. Strganac, A. Kurdila, An investigation of internal resonance in
[193] B. Lee, S. Price, Y. Wong, Nonlinear aeroelastic analysis of airfoils: bifurcation and aeroelastic systems, Nonlinear Dynam. 31 (1) (2003) 1–22, https://doi.org/10.
chaos, Prog. Aero. Sci. 35 (3) (1999) 205–334, https://doi.org/10.1016/S0376- 1023/A:1022174909705.
0421(98)00015-3. [225] A. Nayfeh, M. Ghommem, M.R. Hajj, Normal form representation of the aero-
[194] D. Tang, J.K. Henry, E.H. Dowell, Limit cycle oscillations of delta wing models in elastic response of the goland wing, Nonlinear Dynam. 67 (3) (2012) 1847–1861,
low subsonic flow, AIAA J. 37 (11) (1999) 1355–1362, https://doi.org/10.2514/ https://doi.org/10.1007/s11071-011-0111-6.
2.627. [226] L. Liu, Y. Wong, B. Lee, Application of the centre manifold theory in non-linear
[195] B.K. Stanford, B. P., Formulation of analytical design derivatives for nonlinear aeroelasticity, J. Sound Vib. 234 (4) (2000) 641–659, https://doi.org/10.1006/
unsteady aeroelasticity, AIAA J. 49 (3) (2011) 598–610, https://doi.org/10.2514/ jsvi.1999.2895.
1.J050713. [227] P. Shahrzad, M. Mahzoon, Limit cycle flutter of airfoils in steady and unsteady
[196] G.J. Kennedy, J.R.R.A. Martins, An adjoint-based derivative evaluation method for flows, J. Sound Vib. 256 (2) (2002) 213–225, https://doi.org/10.1006/jsvi.2001.
time-dependent aeroelastic optimization of flexible aircraft, 54th AIAA/ASME/ 4113.
ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference, [228] A.H. Nayfeh, Order reduction of retarded nonlinear systems – the method of
Boston, MA, 2013, https://doi.org/10.2514/6.2013-1530. multiple scales versus center-manifold reduction, Nonlinear Dynam. 51 (4) (2008)
[197] R. Kapania, S. Park, Nonlinear transient response and its sensitivity using finite 483–500, https://doi.org/10.1007/s11071-007-9237-y.
elements in time, Comput. Mech. 17 (5) (1996) 306–317, https://doi.org/10. [229] G. Vio, G. Dimitriadis, J. Cooper, Bifurcation analysis and limit cycle oscillation
1007/BF00368553. amplitude prediction methods applied to the aeroelastic galloping problem, J.
[198] P. Bar-Yoseph, F. D, G. O, Spectral element methods for nonlinear spatio-temporal Fluids Struct. 23 (7) (2007) 983–1011, https://doi.org/10.1016/j.jfluidstructs.
dynamics of an euler-Bernoulli beam, Comput. Mech. 19 (1) (1996) 136–151, 2007.03.006.
https://doi.org/10.1007/BF02824851. [230] E.L. Allgower, K. Georg, Numerical Continuation Methods: an Introduction,
[199] M. Kurdi, P. Beran, Spectral element method in time for rapidly actuated systems, Springer, Berlin, 2012.
J. Comput. Phys. 227 (3) (2008) 1809–1835, https://doi.org/10.1016/j.jcp.2007. [231] H.B. Keller, Numerical solution of bifurcation and nonlinear eigenvalue problems,
09.031. Applied Bifurcation Theory 1 (1977) 359–384.
[200] B.I. Epureanu, E.H. Dowell, Compact methodology for computing limit-cycle os- [232] H.B. Keller, Global homotopies and Newton methods, Recent Advances in
cillations in aeroelasticity, J. Aircr. 40 (5) (2003) 955–963, https://doi.org/10. Numerical Analysis, 1978, pp. 73–94, , https://doi.org/10.1016/B978-0-12-
2514/2.6880. 208360-0.50009-7.
[201] B.I. Epureanu, E.H. Dowell, Localized basis function method for computing limit [233] H.B. Keller, Lectures on Numerical Methods in Bifurcation Problems, Tata Institute
cycle oscillations, Nonlinear Dynam. 31 (2) (2003) 151–166, https://doi.org/10. Of Fundamental Research, 1986.
1023/A:1022081101766. [234] A. Ghadami, B.I. Epureanu, Bifurcation forecasting for large dimensional oscilla-
[202] B. Stanford, P. Beran, M. Kurdi, Adjoint sensitivities of time-periodic nonlinear tory systems: forecasting flutter using gust responses, J. Comput. Nonlinear Dyn.
structural dynamics via model reduction, Comput. Struct. 88 (19–20) (2010) 11 (2016), https://doi.org/10.1115/1.4033920 061009–061009–8.
1110–1123, https://doi.org/10.1016/j.compstruc.2010.06.012. [235] H. Strogatz, Nonlinear Dynamics and Chaos with Applications to Physics, Biology,
[203] P. Deuflhard, Computation of periodic solutions of nonlinear odes, BIT Numerical Chemistry, and Engineering, Westview Press, Boulder, 2001.
Mathematics 24 (4) (1984) 456–466, https://doi.org/10.1007/BF01934904. [236] J. Lim, B.I. Epureanu, Forecasting a class of bifurcations: theory and experiment,
[204] W. Govaerts, Numerical bifurcation analysis for odes, J. Comput. Appl. Math. 125 Phys. Rev. 83 (2011), https://doi.org/10.1103/PhysRevE.83.016203
(1) (2000) 57–68, https://doi.org/10.1016/S0377-0427(00)00458-1. 016203–016203–9.
[205] Y.A. Kuznetsov, Elements of Applied Bifurcation Theory, Springer, 1998. [237] A. Ghadami, C.E.S. Cesnik, B.I. Epureanu, Model-less forecasting of hopf bifurca-
[206] G. Sewell, Numerical Solution of Ordinary and Partial Differential Equations, John tions in fluid-structural systems, J. Fluids Struct. 76 (2018) 1–13, https://doi.org/
Wiley and Sons, 2005. 10.1016/j.jfluidstructs.2017.09.005.
[207] K.C. Hall, J.P. Thomas, W.S. Clark, Computation of unsteady nonlinear flows in [238] S. Missoum, C. Dribusch, P. Beran, Reliability-based design optimization of non-
cascades using a harmonic balance technique, AIAA J. 40 (5) (2002) 879–886, linear aeroelasticity problems, J. Aircr. 47 (3) (2010) 992–998, https://doi.org/
https://doi.org/10.2514/2.1754. 10.2514/1.46665.
[208] M. McMullen, A. Jameson, The computational efficiency of non-linear frequency [239] B. Stanford, P. Beran, Computational strategies for reliability-based structural
domain methods, J. Comput. Phys. 212 (2) (2006) 637–661, https://doi.org/10. optimization of aeroelastic limit cycle oscillations, Struct. Multidiscip. Optim. 45
1016/j.jcp.2005.07.021. (1) (2012) 83–99, https://doi.org/10.1007/s00158-011-0663-6.
[209] A. Gopinath, A. Jameson, Time spectral method for periodic unsteady computa- [240] J. Thomas, E. Dowell, K.C. Hall, Discrete adjoint method for nonlinear aeroelastic
tions over two- and three- dimensional bodies, 43rd AIAA Aerospace Sciences sensitivities for compressible and viscous flows, 54th AIAA/ASME/ASCE/AHS/
Meeting and Exhibit, Reno, NV, 2005, https://doi.org/10.2514/6.2005-1220. ASC Structures, Structural Dynamics, and Materials Conference, Boston, MA,
[210] K.C. Hall, K. Ekici, J.P. Thomas, E.H. Dowell, Harmonic balance methods applied 2013, https://doi.org/10.2514/6.2013-1860.
to computational fluid dynamics problems, Int. J. Comput. Fluid Dyn. 27 (2) [241] S. He, E. Jonsson, C.A. Mader, J.R.R.A. Martins, A coupled Newton–Krylov time
(2013) 52–67, https://doi.org/10.1080/10618562.2012.742512. spectral solver for flutter prediction, 2018 AIAA/ASCE/AHS/ASC Structures,
[211] D. Tang, E.H. Dowell, L.N. Virgin, Limit cycle behavior of an airfoil with a control Structural Dynamics, and Materials Conference, American Institute of Aeronautics
surface, J. Fluids Struct. 12 (7) (1998) 839–858, https://doi.org/10.1006/jfls. and Astronautics, Kissimmee, FL, 2018, , https://doi.org/10.2514/6.2018-2149.
1998.0174. [242] S. He, E. Jonsson, C.A. Mader, J.R.R.A. Martins, Aerodynamic shape optimization
[212] J.P. Thomas, E.H. Dowell, K.C. Hall, Nonlinear inviscid aerodynamic effects on with time spectral flutter adjoint, 2019 AIAA/ASCE/AHS/ASC Structures,
transonic divergence, flutter, and limit-cycle oscillations, AIAA J. 40 (4) (2002) Structural Dynamics, and Materials Conference, American Institute of Aeronautics
638–646, https://doi.org/10.2514/2.1720. and Astronautics, San Diego, CA, 2019, , https://doi.org/10.2514/6.2019-0697.
[213] J.P. Thomas, E.H. Dowell, K.C. Hall, Modeling viscous transonic limit cycle [243] E. Alpaydin, Introduction to Machine Learning, MIT Press, Cambridge, CA, 2004.

27
E. Jonsson, et al. Progress in Aerospace Sciences 109 (2019) 100537

[244] J. Hartigan, M. Wong, A k-means clustering algorithm, Appl. Stat. 28 (1979) Atlanta, GA, 2014, https://doi.org/10.2514/6.2014-2298.
100–108, https://doi.org/10.2307/2346830. [247] F. Zhu, N. Qin, Intuitive class/shape function parameterization for airfoils, AIAA J.
[245] J.P. Thomas, K.C. Hall, E.H. Dowell, Discrete adjoint approach for modeling un- 52 (1) (2013) 17–25, https://doi.org/10.2514/1.J052610.
steady aerodynamic design sensitivities, AIAA J. 43 (9) (2005) 1931–1936, [248] J.P. Thomas, E.H. Dowell, K.C. Hall, C.M. Denegri Jr., An investigation of the
https://doi.org/10.2514/1.731. sensitivity of F-16 fighter flutter onset and limit cycle oscillations to uncertainties,
[246] J. Thomas, E. Dowell, Discrete adjoint method for aeroelastic design optimization, 47th AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics and Materials
15th AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference, Conference, Newport, RI, 2006, https://doi.org/10.2514/6.2006-1847.

28