Progress in Aerospace Sciences 125 (2021) 100725

Contents lists available at ScienceDirect

Progress in Aerospace Sciences

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

Data-driven modeling for unsteady aerodynamics and aeroelasticity

Jiaqing Kou a, b, Weiwei Zhang a, *
a
School of Aeronautics, Northwestern Polytechnical University, Xi’an, 710072, China
b
ETSIAE-UPM-School of Aeronautics, Universidad Politécnica de Madrid, Plaza Cardenal Cisneros 3, E-28040, Madrid, Spain

A R T I C L E I N F O A B S T R A C T

Keywords: Aerodynamic modeling plays an important role in multiphysics and design problems, in addition to experiment
Data-driven method and numerical simulation, due to its low-dimensional representation of unsteady aerodynamics. However, in the
Unsteady aerodynamics traditional study of aerodynamics, developing aerodynamic and flow models relies on classical theoretical
System identification
(potential flow) and empirical investigation, which limits the accuracy and extensibility. Recently, with signif­
Mode decomposition
Data fusion
icant progress in high-fidelity computational fluid dynamic simulation and advanced experimental techniques,
Machine learning very large and diverse fluid data becomes available. This rapid growth of data leads to the development of data-
driven aerodynamic and flow modeling. Through advanced mathematical methods from control theory, data
science and machine learning, a lot of data-driven aerodynamic models have been proposed. These models are
not only more accurate than theoretical models, but also require very low computational cost compared with
numerical simulation. At the same time, they help to gain physical insights on flow mechanism, and have shown
great potential in engineering applications like flow control, aeroelasticity and optimization. In this review
paper, we introduce three typical data-driven methods, including system identification, feature extraction and
data fusion. In particular, main approaches to improve the performance of data-driven models in accuracy,
stability and generalization capability are reported. The efficacy of data-driven methods in modeling unsteady
aerodynamics is described by several benchmark cases in fluid mechanics and aeroelasticity. Finally, future
development and potential applications in related areas are concluded.

1. Introduction steady aerodynamics, unsteady aerodynamics is more challenging to

model, since these models need to account for the time-delayed effects
In fluid mechanics, recent advances in high-performance computing on the predictions at current time. At first, aerodynamic models are
and experimental facilities make the amount of flow data increase derived based on the mathematical formulation of simplified flow
rapidly. These data include flow field quantities sampled in space and physics and some assumptions. This classical modeling approach is then
time, as well as the integrated aerodynamic forces for various test cases followed by the computational modeling thanks to the widespread use of
and conditions. They are not only useful for analyzing flow physics, but Computational Fluid Dynamics (CFD). Although CFD and experiment
also can be utilized to develop compact and useful models that describe often give accurate and reliable prediction of the aerodynamic loads, the
the dynamics of flow. In combination with the progress in data science, cost in computational resource and experimental facilities for a group of
machine learning and control theory, this provides attractive opportu­ runs still make it necessary to use simpler and less costly aerodynamic
nities for research. The central topic of this paper is to review data- models.
driven methods which produce simple and accurate aerodynamic Recently, there has been an increasing research interest in modeling
models that support the study of flow physics, control, design and unsteady aerodynamics in a data-driven manner. The purpose of data-
optimization. driven modeling is to obtain reduced-order models (ROMs) for un­
Unsteady aerodynamic modeling is a classical and important steady aerodynamics [1,2]. The term, ‘reduced-order modeling’, has
research field in fluid mechanics. An ideal aerodynamic model should been widely used before the prevalence of ‘data-driven modeling’ in
reflect flow physics properly with a low level of complexity, thus fluid dynamics. However, data-driven modeling refers to a wider range
allowing accurate and efficient aerodynamic prediction. Different from of topics, including data-assisted turbulent models [3] which will not be

* Corresponding author.
E-mail addresses: [email protected] (J. Kou), [email protected] (W. Zhang).

https://doi.org/10.1016/j.paerosci.2021.100725
Received 15 January 2021; Received in revised form 20 April 2021; Accepted 24 April 2021
Available online 20 June 2021
0376-0421/© 2021 Elsevier Ltd. All rights reserved.
J. Kou and W. Zhang Progress in Aerospace Sciences 125 (2021) 100725

covered, and data fusion which is one of the main topics in this paper. analysis methods [14] and their applications [15] have been overviewed
ROM refers to the low-dimensional mathematical models built from a by Taira and his colleagues. Rowley and Dawson [16] summarized
large flow dataset based on CFD or experiment, in order to model steady model reduction methods for flow analysis and control. Brunton et al.
or unsteady aerodynamics in the range of input parameters used to [17] discussed the applications and potential of artificial intelligence in
create ROMs. These models can predict not only the flow fields but also fluid mechanics. Pandey et al. [18] discussed a perspective on machine
the integrated forces and moments for different geometries. Generally, learning in turbulent flows. Distinct from these works, the main focus of
reduced-order modeling serves two purposes, including efficient aero­ the present paper is to review different data-driven methods for un­
dynamic simulation and the extraction of the underlying physics. It has steady aerodynamic modeling, including approaches based on system
been reported that ROMs can predict the linear and nonlinear aero­ identification, feature extraction and data fusion. In particular, some
dynamic phenomenon accurately within a fraction of time compared existing challenges and solutions are discussed, as well as how to better
with that of CFD-based analysis, where a speedup of one or two orders of adapt to these methods from a practical point of view.
magnitude can be reached [4]. This high accuracy and low computa­
tional cost is attractive for various engineering applications. By 2. Construction of aerodynamic models
analyzing the dynamics based on ROM, it is also possible to extract the
dominant dynamic features from a large flow dataset. In general, the construction of aerodynamic models has gone
Data-driven modeling has potential applications in several fields: 1) through three main stages, including theoretical, computational and
Aeroelasticity [5]. Aeroelasticity includes the flutter phenomenon data-driven modeling. This development is mainly attributed to the
which is a self-excited oscillation behavior due to the interaction be­ increasing computational capability and storage resources in the past
tween unsteady flows and flexible structures. Stability and response of decades [19]. However, all of these strategies still remain widely used,
aircraft wings, offshore structures and bridges under the disturbance of therefore are worth discussing and will be briefly introduced in this
turbulent flows are main examples of this area. However, a large section.
computational resource is needed to simultaneously simulate the
fluid-structure interaction (FSI) phenomena when an Euler or 2.1. Theoretical modeling
Navier-Stokes fluid model and a finite element model for the structure is
used. ROMs based on CFD for the flow and finite element model for the Early research in fluid mechanics mainly focused on deriving theo­
structure not only speed up the aerodynamic and aeroelastic simulation retical (potential flow) and empirical aerodynamic models. Although
but also can be used to analyze the mechanism of the FSI phenomenon. some assumptions and simplifications are made, these models are still
2) Flight dynamics [2]. Flow around a rapid maneuvering fighter has a often used for a specific application. Most of these models are based on
strong impact on its flight performance and quality. However, the simple algebraic expressions, ordinary differential or integral equations,
nonlinearity and hysteresis of the aerodynamic loads cannot be properly which give a reasonable approximation of the aerodynamic loads. In the
described by traditional models based on the concept of a dynamic de­ early 1930s, the Grossman [20,21], Theodorsen [22] and Wagner [23]
rivative. ROM provides a solution to this problem, by creating aero­ models were proposed, which remain useful even in recent years. The
dynamic models directly from flight testing or simulation data. 3) Grossman model is a quasi-steady aerodynamic model for thin plates,
Optimization design [6]. Optimization design is a many-query problem, which considers the effect of displacement and velocity of the wing
where multiple calls of a CFD solver are needed. Compared with CFD motion, but ignores the influence of the wake vortex. Thereafter, the
simulation, a fast and accurate ROM accelerates the optimization iter­ influence of the vortices in the wake was taken into account in the
ations by reducing the time of aerodynamic simulation. The data-driven Theodorsen model that describes the unsteady aerodynamic loads for
model reduction approach also helps to obtain a reduced number of the sinusoidal motion of thin plates in the frequency domain. A similar
design variables [7] or a parametric ROM [8], thus reducing the model in the time domain was developed by Wagner, which is an inte­
complexity and cost of the problem at hand. 4) Flow control [9,10]. gral model based on the step response of the airfoil. Since these models
Data-driven modeling builds low-dimensional flow models for designing are linear, they are not suitable to describe the flows over an airfoil at
control laws to reduce the drag forces and avoid instabilities. Although large angles of attack with stall phenomena. These limitations are
designing passive and open-loop control strategies can be achieved by addressed by aerodynamic models that describe dynamic stall, like the
experiment and numerical simulation, model-based flow control pro­ ONERA [24] and the Beddoes–Leishman [25] models. They are empir­
vides an efficient tool to design closed-loop control laws, allowing better ical models derived from the essential flow physics but contain several
manipulation of flow behaviors. 5) Mechanism and characteristic anal­ coefficients that need to be determined from experimental measure­
ysis [11]. Gaining physical insights from flow data based on feature ments. Therefore these models can be seen as early versions of a
extraction and data-driven modeling methods, has been an active data-driven model, where the model structure remain fixed but the co­
research topic. Examples include the application of ROMs in analyzing efficients are obtained from data. These models give very clear repre­
the lock-in phenomenon in FSI [67,69], as well as modal analysis to sentations for different aerodynamic effects. For example, the attached
capture coherent structures in turbulent flows [155,157]. and separated flows are described by the linear and nonlinear sub­
So far, data-driven modeling methods in fluid mechanics have been models, respectively. The first-order and second-order differential
discussed in several review papers from different perspectives. In early equations are used separately for inviscid and viscous effects. They have
2000, Dowell and Hall [5] and Lucia et al. [1] introduced the system been used extensively in aerodynamic analysis of airfoils and helicop­
identification and proper orthogonal decomposition (POD) methods ters. Different dynamic stall models were compared by Holierhoek et al.
used in FSI problems. Da Ronch et al. [12] summarized the use of sur­ [26], indicating that improvement of these models is still needed espe­
rogate models for steady aerodynamics in flight dynamics, with appli­ cially in the deep stalled regime. Another theoretical, semi-empirical
cations to different aerodynamic configurations. Aerodynamic and model worth mentioning is the wake oscillator model, which describes
aeroelastic modeling in hypersonic flows has been reviewed by McNa­ the aerodynamics of flow past elastic bodies [27]. It has been extensively
mara and Friedmann [13]. ROMs in flight dynamics and stability anal­ used for describing vortex-induced vibrations of a circular cylinder [28].
ysis of different aircraft have been discussed by Ghoreyshi et al. [2], Various wake oscillator models have been proposed and are discussed by
where the system identification methods from input-output aero­ Dowell [29]. Although all of these theoretical models are attractive due
dynamic data are mainly focused. It should be mentioned that some of to their conceptual and mathematical simplicity, they still have limita­
the methods discussed in this review, have also been detailed in these tions because of the assumptions in model derivation.
review articles. Closed-loop control of unsteady flows based on different
fluid models is introduced in Brunton and Noack [9]. Recently, modal

2
J. Kou and W. Zhang Progress in Aerospace Sciences 125 (2021) 100725

2.2. Computational modeling velocity, density and pressure, etc. Extraction of features from these flow
data helps to capture the coherent structures of complex flows, making
Computational modeling in fluid mechanics derives the governing the explanation and understanding of flow physics straightforward.
equations of flow physics from basic conservation laws, where the Since these data can be obtained from multiple sources, which vary in
resulting partial differential equations need to be solved numerically cost and fidelity, research on data fusion also became popular in recent
with proper discretization schemes. An early version of these models is years. The performance of data-driven modeling is basically evaluated
derived from the potential flow theory that provides Laplace’s equation based on accuracy and generalization capability. Model accuracy re­
for the velocity potential. The resulting vortex-lattice method and quires the resulting aerodynamic model to reproduce the dynamics in
doublet-lattice method have been a fundamental tool for low-speed the training data, which are used to generate the model. However,
aircraft aeroelasticity and dynamic-load analysis [30,31]. A similar generalization capability is a more strict requirement that requires the
model has been used to formulate the transonic small disturbance theory model to keep a good prediction capability for new inputs not encoun­
that characterizes the behavior of transonic flows [32]. With the tered in the training stage. These are the main factors that should be
development of advanced computational facilities, solving the Euler and taken into account in data-driven modeling.
Navier-Stokes equations becomes possible, where different CFD
methods have been extensively developed, like the Direct Numerical 3. Data-driven modeling methods
Simulation (DNS) [33], Large Eddy Simulation (LES) [34] and
Reynolds-Averaged Navier-Stokes (RANS) [35] methods. These methods 3.1. System identification for input-output data
allow high-fidelity numerical simulation to investigate the flow dy­
namics in detail, and also produce a very large flow dataset that can be System identification is the theory to construct mathematical models
used for data-driven modeling. These CFD-based approaches are usually from system measurements, which lays the foundation for system con­
called full-order aerodynamic models and are expensive for design and trol and analysis [45]. As mentioned before, in aerodynamic modeling,
optimization when a CFD solver with a larger grid number is used. the input data is usually the flow state variable (e.g., Mach number,
The other branch of computational modeling aims at efficiently Reynolds number, mean angle of attack, etc.) for steady flows or the
solving these governing equations. Two groups of methods, harmonic motion of the body for unsteady flows, while the output data is usually
balance method in the frequency domain and collocation method in the the integrated aerodynamic loads like lift and drag coefficients. The
time domain, have been proposed [36]. Due to the periodic nature of unknown aerodynamic system can be treated as a black-box without any
unsteady flows, the governing equations for unsteady flows can be physical insight, or a grey-box with partial prior knowledge. The general
transformed to steady equations in the frequency domain with Fourier workflow of system identification is to firstly obtain the training data
transformation. Popular harmonic methods in the frequency domain are from numerical simulation or experimental measurements. Secondly,
linear [37] and nonlinear [38,39] harmonic methods. Similarly, FSI the model unknowns are determined by different identification
problems can be directly solved based on the aerodynamic describing methods. This model is finally validated and tested by different input
functions [40]. For the collocation method in the time domain, the signals and compared with the reference data to evaluate the accuracy.
transient responses of unsteady flow are firstly projected onto the pre­ The performance of system identification depends on three main
defined orthogonal polynomials. After that, the governing equations can factors: data, model structure and identification method. Data used to
be transformed into steady problems, characterized by the coupling train the model can be predesigned by the user to include sufficient
between several time instants of flow. Examples of these methods dynamical information of interest. The model structure needs to be
include harmonic balance in the time domain [41] and time spectral determined based on the flow physics, which leads to different available
[42–44] methods. Following the pioneering works reviewed in Ref. [5], model structures, as discussed in this section. The model identification
it is now widely understood that for periodic unsteady flows either linear method can be further grouped into parametric and nonparametric
or nonlinear, a harmonic balance method is much more computationally methods [433]. In the parametric method, the entire system has a
efficient than a time marching solution. known mathematical structure, where only the unknown coefficients for
each term are identified. The identification of empirical aerodynamic
2.3. Data-driven modeling models, like dynamic stall models, can be classified into this group. In
the nonparametric method, the system is not assumed to be a parame­
Motivated by the emergence of flow data, data-driven modeling has terized model, therefore the user has to choose the existing model
become a novel paradigm to construct a simple and concise represen­ structure that is known to have good flexibility, like the neural network.
tation of unsteady flow dynamics. Due to lower complexity and cost, System identification not only provides a model structure but also the
traditional aerodynamic modeling can be complemented, even replaced mathematical foundation to obtain the model. Different models will be
by data-driven modeling approaches. The fundamental consideration for detailed in the following sections.
data-driven modeling lies in the distinction between a linear dynamic The unsteady aerodynamic system is indeed a nonlinear system, as
(but nonlinear static) aerodynamic model and a fully nonlinear model reflected by its governing equations. However, it can be linearized about
[5], which provides guidance for the selection of appropriate an equilibrium state under a small dynamic perturbation [5]. From a
data-driven models. At any level of aerodynamic modeling, from theo­ system identification perspective, unsteady aerodynamics has two main
retical modeling to computational modeling, a linear dynamic model features: 1) the nonlinearity of a dynamic system depends on the
(linearized about a steady state flow) can be constructed by a variety of magnitude of disturbance; 2) the unsteadiness of flow is reflected by its
methods as discussed by Dowell and Hall [5] who did the early research time-delayed memory effects. Therefore, under different flow conditions
on this concept. Based on this concept, nonlinear models come after and motions, unsteady aerodynamics can be described by statically
linear ones to describe the strong aerodynamic nonlinearity when the linear, dynamically linear (statically nonlinear) or dynamically
linear model fails. So far, data-driven models can be classified into two nonlinear models [46]. For example, in aeroelasticity, the flutter phe­
groups according to two types of aerodynamic data: 1. nomenon is characterized by the dynamically linear model where the
system-identification-based model for input-output aerodynamic data; flow variables vary in a linear fashion with the body motion [5]. How­
2. feature-extraction-based model for high-dimensional flow data. ever, the limit-cycle oscillations are nonlinear phenomena induced by
Input-output data include the generalized displacement, the flow state either structural or dynamical nonlinearity of the unsteady aero­
parameters and the integrated aerodynamic coefficients, which reflect dynamics. Based on this prior knowledge of the dynamics, the model
the dynamic and response behavior of the system. High-dimensional structure can be selected accordingly. The nonlinearity of the model is
flow data comes from measurement of flow field quantities, like reflected by different system identification models, while the

3
J. Kou and W. Zhang Progress in Aerospace Sciences 125 (2021) 100725

unsteadiness effect is reflected by the time-delayed input and output

∑
n m∑
− 1
terms. Existing dynamically linear models mainly include Eigensystem y(k) = Ai y(k − i) + Bi u(k − i) (2)
Realization Algorithm (ERA), AutoRegressive with eXogenous input i=1 i=0
(ARX) model and indicial functions. In addition, numerous dynamically
nonlinear models have been proposed, including Volterra series, Krig­ where A and B are the coefficient matrices to be identified, whose sizes
ing, neural network and block-oriented models. Generally, nonlinear are determined by input and output delay orders m and n. Considering
system identification is to represent unsteady nonlinear aerodynamic the difference in the error term, there are other similar models, like the
systems by the Nonlinear ARX (NARX) structure. Note that a similar AutoRegressive Moving Average (ARMA) model and the AutoRegressive
structure is the nonlinear autoregressive moving average model with Integrated Moving Average (ARIMA) model. In practice, all of these
exogenous input (NARMAX), which considers time-delayed evolution of models can be identified by the least-squares method and its variants
prediction error [45]. Without loss of generality, NARX formulation for [56]. The only parameters that determine the model structure are delay
a multi-input multi-output dynamical system is as follow: orders of input and output, which account for the memory effects of the

y(k) = Φ (u(k), u(k − 1), u(k − 2), …, u(k − m), y(k − 1), y(k − 2), …, y(k − n) ) (1)

unsteady aerodynamics. Like ERA, FSI analysis can be achieved by

where u(k) and y(k) denote the input and output vector of the NARX coupling the aerodynamic ROM with the structural motion equations in
model, respectively. k is the time index, where the discrete time is the state-space form. Cowan et al. [57] showed one of the early works to
defined as t = kΔt sampled with a constant sampling time. m and n use such models for unsteady aerodynamic modeling. They used ARMA
refer to the delay orders of input and output. The selection of these or­ model to predict the flutter boundary of AGARD 445.6 wing and the
ders will be discussed in Sec. 4.1.1. The system state vector x is char­ aeroelastic response of hypersonic aircraft. Thereafter, extensive studies
acterized by all the input terms, where x(k) = [u(k), u(k on analyzing and explaining the mechanism of different aeroelastic and
− 1), u(k − 2), …, u(k − m), y(k − 1), y(k − 2), …, y(k − n) ]T . flow-induced vibration (FIV) problems have been conducted by Zhang
The goal of modeling nonlinear unsteady aerodynamics is to obtain et al. and his coworkers [11]. They firstly applied the ARX-based aero­
proper formulation of Φ to represent unknown dynamical systems. elastic model to describe stable flows, including aeroelastic analysis of a
wing at large angles of attack [58], aeroservoelastic analysis of an airfoil
3.1.1. Eigensystem realization algorithm with controlling flap [59–61], aeroelastic modeling [62] and optimiza­
The ERA model was proposed by Juang and Pappa in 1985 [47]. It tion [63] considering arbitrary structural modes, aeroelastic analysis of
obtains the state-space model of the dynamically linear system from the turbomachinery [64] and transonic buzz [65], as well as modeling the
impulse response data: thrust force of flapping wings [66]. For FSI problems based on unstable
flows or critical flow states, the more complicated interaction between
x(k + 1) = Ax(k) + Bu(k) fluid mode and structure mode is clarified by the ARX-based FSI model.
(2)
y(k) = Cx(k) + Du(k)
This includes mechanisms of vortex-induced vibration [67] and
galloping [68] for flow past cylinders at low Reynolds numbers, as well
where the system matrices are A, B, C, D. After collecting the impulse
as frequency lock-in Ref. [69] and active flow control [70] in the buffet
response data, ERA forms two Hankel matrices from the system output,
flow. From the work of He et al. [71,72], this linear aerodynamic model
and performs singular value decomposition (SVD) to compute the sys­
can be coupled with a nonlinear structure to investigate the bifurcation
tem matrices. The order of system should be defined by the user. More
behaviors of an airfoil with free-play in transonic flow. The general
details are given in Refs. [16,47]. Unsteady aerodynamic modeling
framework of CFD-based and data-driven ROM based aeroelastic anal­
based on ERA can be traced back to the early 2000s. Silva and Raveh
ysis framework is shown in Fig. 1. It should be noted that, both ERA
[48] constructed the 32nd-order state-space model from ERA, where the
model and ARX model can be used to model unstable flow dynamics,
unsteady aerodynamics for the AGARD 445.6 wing at different fre­
while meaningful analysis relies on the data that describes the linear
quencies is accurately predicted. Gaitonde and Jones [49] modeled the
dynamics. Therefore, the aerodynamic data should be obtained based on
linearized unsteady aerodynamics of a NACA64A010 airfoil based on
a proper base flow, like the linear equilibrium state with a small
ERA. They pointed out that if the order of the state-space model is too
disturbance, in order to include accurate linear dynamic behaviors.
low, the resulting aerodynamic model can be less accurate and unstable.
Silva and Bartels [50] not only identified the unsteady aerodynamic
3.1.3. Indicial function
model of the AGARD 445.6 wing, but also coupled the aerodynamic
Indicial response refers to the response of the aerodynamic loads to a
model with the structural motion equations to predict the flutter
step change in the motion, such as angle of attack or pitch rate [2]. The
boundary in the transonic flow. It is also worth mentioning that Kim
indicial model is an integral formulation which shares the mathematical
[51] combined ERA with the single-composite-input algorithm, which
similarity with the classical Wagner model [73]. The only difference is
extends the original ERA approach to allow arbitrary inputs. Applica­
that the indicial function is identified in a data-driven manner from the
tions of ERA include aerodynamic modeling for a NACA0006 airfoil at
step response data, rather than from analytical functions. In indicial
low Reynolds numbers based on noisy experimental measurement [52],
theory, the temporal response of any input signal can be represented as
robust control for flow past a D-shaped body [53], and mechanism of
the convolution of the input with the indicial function of the aero­
vortex-induced vibration of flow past bluff bodies [54,55].
dynamic force. This method is detailed in Tobak [74] and Reisenthel
[75], where the resulting models are used for modeling flight dynamics
3.1.2. Autoregressive with exogenous input model
of aircraft. The linear indicial function of the aerodynamic coefficients is
The ARX model provides a general description for linear dynamic
shown as follow:
systems in the discrete time domain, which is described by linear dif­
ferential equations. The model formulation is as follow:

4
J. Kou and W. Zhang Progress in Aerospace Sciences 125 (2021) 100725

Fig. 1. CFD-based and ROM-based aeroelastic analysis framework [11]. Structural equations (S) can be coupled with either CFD solver (F1) for CFD-based aeroelastic
simulation, or ROM (F2) for ROM-based aeroelastic analysis.

⎡ t ⎤ ⎡ t ⎤
∫ ∫ which is capable of describing casual, time-invariant, finite-memory
d⎣ d⎣
Cj (t) = Cj0 + ⎦
Cjα (t − τ)α(τ) dτ + Cjq (t − τ)q(τ) dτ⎦ systems in both time domain and frequency domain. Take a single-input
dt dt
0 0 single-output system as an example, the Volterra series defined in
(3) continuous and discrete time domains are as follows:
∫t
where Cj is the jth aerodynamic coefficient, and Cj0 is the zero-angle-of-
y(t) = H0 + H1 (t − τ)u(τ) dτ+
attack coefficient. α and q refer to angle of attack and normalized pitch
rate, respectively. Cjα (t) and Cjq (t) are the response functions identified ∫ t ∫t
0

from the step response. This formulation equals to the first-order Vol­ H2 (t − τ1 , t − τ2 )u(τ1 )u(τ2 )dτ1 dτ2 +
terra series based on the step response. A nonlinear model can be further
0 0
considered when the variation with respect to angle of attack and Mach
number is taken into account. This model is achieved by building indi­ ∫t ∫t ∏
s (5)
cial functions changing with state variables, while the function for in­ … Hs (t − τ1 , …, t − τs ) u(τi )dτi
termediate flow state can be interpolated from functions of neighboring
i=1
0 0

states. The nonlinear indicial function is shown as follow:

⎡ t ⎤ ∑
k
∫ y(k) = H0 + H1 (k − n)u(n)+
d⎣
Cj (t) = Cj0 + Cjα (t − τ, α, Ma)α(τ) dτ⎦ n=0
dt ∑
k ∑
k
0
⎡ t ⎤ H2 (k − n1 , k − n2 )u(n1 )u(n2 )+
∫ n1 =0 n2 = 0
d⎣
+ Cjq (t − τ, Ma)q(τ) dτ⎦ (4) ∑
k ∑
k ∏
s
(6)
dt … Hs (k − n1 , …, k − ns ) u(ni )
0
n1 =0 ns = 0 i=1
Note that for such nonlinear models, we can assume that the
response functions about the pitch rate only change with freestream
Mach number, at small angles of attack [2]. A toolkit based on nonlinear where Hs is the sth-order Volterra operator, and the first term H0 is the
initial functions is developed by Reisenthel et al. [76], where the un­ steady-state term. Each Volterra kernel is denoted as an s-fold convo­
steady aerodynamics of a 65-degree sweep delta wing at large angles of lution between the input and the sth-order Volterra operator Hs. It
attack is modeled. Details for constructing an aerodynamic model with should be noted that the discrete-time formulation is more suitable for
indicial functions are discussed in Ghoreyshi [77–79], where nonlinear numerical and experimental applications, since the continuous-time
indicial functions are used to predict the aerodynamic loads of the response is defined in an ideal condition where the input signal has an
SACCON aircraft, T-38 aircraft and Lambda wing. In addition, the amplitude approaching infinity while its width is zero [83]. The Volterra
indicial function can also be calculated from combining theoretical kernel can be identified based on impulse response, step response or
formulation and parameter identification. An example is shown in Da random motions with different identification methods [84]. When this
Ronch et al. [80], where the analytical indicial functions are extended model is used for dynamically linear systems, only the first-order Vol­
with aerodynamic data to predict the aerodynamic loads of delta wings terra kernel is needed. However, the first-order Volterra model may
in subsonic flows. encounter difficulty in modeling unsteady flows approaching instability
[84], due to the fact that the output is only a function of time-delayed
3.1.4. Volterra series inputs without autoregressive, time-delayed output terms. Therefore,
Volterra models represent nonlinear systems by a functional infinite the Volterra model has also been augmented with autoregressive terms
series, which was first proposed by mathematician Vito Volterra [81]. It [85,86] for better accuracy. The nonlinearity of dynamical systems is
has been extensively applied to modeling many nonlinear systems [82], reflected by high-order Volterra kernels. However, for nonlinear system

5
J. Kou and W. Zhang Progress in Aerospace Sciences 125 (2021) 100725

identification, time cost of obtaining a high-order Volterra kernel in­ for flow control based on synthetic jets is detailed in Li et al. [105] and
creases exponentially with order. To address this issue, pruned Patterson and Friedmann [106].
high-order Volterra series [87] and sparse Volterra series [88] have been
proposed to reduce the computational complexity. 3.1.6. Neural network
An early application of the Volterra model was proposed by Silva Neural network, also known as artificial neural network, is the
[89], where the aeroelastic analysis of an aircraft wing in subsonic and computing system which mimics the function of biological neural net­
transonic flows is shown. Subsequently, a more stable identification works. It has universal approximate capability that is able to fit any
method based on the step response was proposed by Raveh [90]. The smooth function exactly, under certain conditions [107]. Generally, the
identification of the Volterra series for an aeroelastic system in the fre­ standard neural network has a three-layer structure, including input,
quency domain was investigated by Marzocca et al. [91]. Applications of hidden and output layers, which can be represented as follows:
Volterra models in aeroelastic analysis were reviewed by Silva [83]. The
∑
l
Volterra model has also been extended to model parameter-varying y(k) = w0 + wj ⋅hj (x(k) ) (8)
aerodynamics across multiple Mach numbers, as shown in Skujins and j=1
Cesnik [92]. With the need of modeling strong aerodynamic nonline­
arity, higher-order Volterra series have been used. Balajewicz and where w0 and wj are the bias value and the weight in the output layer. h
Dowell [88] used sparse Volterra series to model nonlinear unsteady is the nonlinear activation function in the hidden layer, which varies for
aerodynamics and predict the limit-cycle oscillations of the nonlinear different neural networks, and l is the number of hidden neurons. The
aeroelastic system. In this model, only diagonal terms of higher-order output is approximated as a weighted sum of a set of activation functions
Volterra series are kept, in combination with pseudo-inverse identifi­ in the hidden layer. An early version of neural network is the Multilayer
cation method to train the model from random training data. Recently, Perceptron (MLP) [107]. Here we take radial basis function neural
de Paula et al. [93] proposed time-delayed neural networks to identify network (RBFNN) as an example, which adopts RBF as the activation
the Volterra kernels. With the neural network structure, high-order function and has been shown to have good generalization capability
Volterra series can be computed much more easily to reproduce the [108,109]. This model is firstly proposed by Broomhead and Lowe
unsteady aerodynamic loads under a large motion. [110]. The Gaussian basis function is usually used for RBFNN:
( ⃦ ⃦
⃦)
⃦
3.1.5. Kriging ⃦x(k) − v2j ⃦
hj (x(k) ) = exp − (9)
Kriging is a multivariable interpolation method to approximate the 2σ 2j
unknown functions of interest from data, which stems from geophysical
research [94]. Kriging constructs a response surface for the underlying where σ and v are two hyperparameters, named as the width parameter
function between input x and output y, as a realization of a regression and the center vector. It should be noted that the performance of RBFNN
model g(x) and a stochastic process z(x) with zero mean and variance relies largely on the types of activation functions [111], as well as the
σ 2var : hyperparameters of the activation function [112,113]. RBFNN and
Kriging are equivalent under certain conditions [114]. Applications of
y(x) = g(x) + z(x) (7)
neural network in modeling the aerodynamic loads for flight dynamics
where the regression model provides a globally valid trend function, can be traced back to the 1990s [115]. Linse and Stengel [115] used
while the stochastic process term reflects the deviation from the neural network to estimate the aerodynamic coefficients obtained from
regression model to ensure that sample points are interpolated exactly. flight testing for nonlinear control of aircraft. Marques and Anderson
The regression model is a polynomial assumed to be either constant, [116] constructed the unsteady aerodynamic model with a finite
linear or quadratic, whose coefficients are calculated in a least-squares memory neural network, which reflects the memory effect by the finite
impulse response function in the hidden layer. Suresh et al. [117]
sense. It is defined as g(x) = f(x)T β, where f(x) contains the poly­
modeled the lift coefficient of an OA012 airfoil from experimental data,
nomial basis functions and β contains the unknown coefficients. The
based on the Recurrent Neural Network (RNN). This work also included
stochastic process is described by a correlation function characterized by
the generalization error for model training, in order to guarantee good
the covariance matrix to measure the correlation between sample points
generalization capability. Motivated by the NARX framework for
[95]. The correlation function between sample xi and xj is defined as
( ) modeling nonlinear dynamical systems, Zhang et al. [4] proposed a
R θ, xi , xj , where θ is the hyperparameter of the correlation function
Recursive RBFNN (RRBFNN) to identify the nonlinear unsteady aero­
that needs to be optimized by maximizing a likelihood function from dynamics of a NACA0012 airfoil at large amplitudes in transonic flow.
training data. The stochastic process is usually modeled as a Gaussian The LCO response induced by aerodynamic nonlinearity was then pre­
process model [96]. dicted by the ROM-based FSI simulation. Ghoreyshi et al. [118] con­
An advantage of the Kriging model is that it not only gives the mean structed the unsteady aerodynamic ROM for the airfoil motion in the
value, but also the confidence interval of output, allowed by the prob­ subsonic flow based on RBFNN. Mannarino and Mantegazza [119]
ability distribution of the stochastic process. This makes Kriging a reli­ adapted RNN for the prediction of nonlinear aeroelastic behaviors of two
able tool in uncertainty quantification [97,98] and aerodynamic typical airfoil models. Based on a fuzzy neural network, Winter and
optimization [99]. Using the Kriging model to make aerodynamic tables Breitsamer [120] predicted the flutter boundary at a range of Mach
for flight dynamics and stability control has been proposed by Da Ronch numbers for AGARD 445.6 wing. Kou and Zhang [121] proposed an
et al. [12]. The first work that used Kriging to model unsteady nonlinear approach to improve the generalization capability of nonlinear aero­
dynamics was conducted by Glaz et al. [100], where the Kriging-based dynamic ROMs, through introducing validation data and particle swarm
NARX framework was named as Surrogate-Based Recurrence Frame­ optimization (PSO) algorithm to determine hyperparameters. An effi­
work (SBRF). Here the input state vector x is defined based on the NARX cient algorithm for RBF center selection was also proposed [122] for
framework describe in Sec. 3.1. This model was then extended to predict accurate LCO prediction. For aerodynamic prediction across multiple
dynamic stall with swept flow effects [101]. Liu et al. [102] used SBRF to Mach numbers, Kou and Zhang [112] proposed a multi-kernel neural
model the unsteady aerodynamics and predict flutter boundaries across network model, which introduces more flexibility in the model to ac­
multiple Mach numbers. Li et al. [103] performed unsteady aero­ count for more complicated nonlinear behaviors. As shown in Wang
dynamic prediction with a light stall phenomenon based on SBRF. et al. [123], more flexibility for RRBFNN can be added through the
Brouwer and McNamara [104] used Kriging for aeroelastic loads pre­ scalar-valued RBF to achieve a better scaling property of input. The
diction in the presence of shock-induced separation. The Kriging model

6
J. Kou and W. Zhang Progress in Aerospace Sciences 125 (2021) 100725

Fig. 2. RRBFNN aerodynamic modeling framework across multiple Mach numbers [113]. The model is trained from dataset containing dynamic information across
different Mach numbers, based on efficient methods like POD, PSO, and cross-validation. With Mach number as an additional input, the aerodynamic and aeroelastic
prediction at various Mach numbers can be achieved by this ROM.

nonlinear ROM framework for multiple Mach numbers based on aerodynamics and nonlinear aeroelasticity in several applications. Bal­
RRBFNN is illustrated in Fig. 2 [113]. In addition, neural network for delli et al. [129] predicted the flutter and LCO behavior of a 2D wing
aerodynamic optimization has also become an active research area in section by using a Hammerstein nonlinear model to include structural
recent years [124]. nonlinearity in the stiffness of the pitch motion. Lum and Lai [130] also
utilized a Hammerstein model to construct nonlinear ROM. The flutter
3.1.7. Block-oriented models boundary of the AGARD445.6 wing was predicted. Using block-oriented
Modeling nonlinear dynamic systems is a difficult task in system aeroelastic models with structural nonlinear feedback, Dai et al. [131]
identification because the constructed dynamic model could very well performed uncertainty analysis for flutter and LCO. Huang et al. [132]
exhibit instability in simulation [125–127], and the purely nonlinear proposed a Wiener-type cascade model and used it for open-loop and
model is not suitable for the identification of dynamic systems with a closed-loop aeroservoelastic analysis [133], where multiple Wiener
dominant linear behavior [128]. These problems motivate the devel­ blocks are constructed in serial and connected in parallel. The perfor­
opment of block-oriented nonlinear models. Block-oriented models mance of the model was demonstrated by modeling transonic
provide nonlinear model frameworks that are composed of a serial compressible flow over a two-degrees-of-freedom wing section with the
connection of dynamically linear and statically nonlinear subsystems. NACA 64A010 airfoil. Due to the fact that the standard Wiener model
Each subsystem can be identified by any proper system identification only considers a statically nonlinear subsystem, Kou et al. [134] pro­
method. They can effectively represent a wide range of nonlinear be­ posed improved Wiener models that are better suited for describing
haviors as well as several nonlinear systems with a dominant linear nonlinear and unsteady aerodynamics. In particular, the statically
behavior. There are two typical block-oriented models, i.e., Wiener and nonlinear block was modified as a quasi-dynamic block, in order to
Hammerstein models. Wiener models have a linear dynamic block fol­ capture the strong nonlinear effects of a moving airfoil at large angles of
lowed by a static nonlinear function, while Hammerstein models have a attack with improved accuracy.
static nonlinear block followed by a dynamically linear block. Defining
the dynamically linear transfer function G(s) and statically nonlinear 3.1.8. Long short-term memory
function f N , the Hammerstein and the Wiener models are represented as Recently, deep learning has attracted widespread attention, which
follows, respectively: outperforms other machine learning methods in numerous important
{ applications [135]. As a typical deep neural network, Long Short-Term
x(k) = f N (u(k) )
(10) Memory (LSTM) has shown its advantages in time series forecasting
y(k) = G(s)x(k)
problems, where the memory effect of dynamics needs to be modeled.
{ LSTM was proposed by Schmidhuber and his coworkers in 1997 [136,
x(k) = G(s)u(k)
(11) 137], which is a variant of RNN but overcomes stability bottlenecks
y(k) = f N (x(k) )
encountered in traditional RNNs like the vanishing gradient [138].
Block-oriented models have been used for modeling unsteady Different from traditional unsteady aerodynamic modeling based on

7
J. Kou and W. Zhang Progress in Aerospace Sciences 125 (2021) 100725

RNN [4,119], the user does not need to assign delay orders for the LSTM, POD, in order to describe the dynamics of flow past a cylinder and a
since the time-delayed effect is considered by the network itself. This wind-driven ocean gyre. Besides, Mohan and Gaitonde [146] and Rah­
dramatically reduces the difficulty in model training for unsteady man et al. [147] used a similar method to model turbulent flows. A
aerodynamic modeling. hybrid analysis and modeling approach is proposed by Pawar et al.
The structure of the LSTM network for three discrete time steps is [148]. ROMs based on POD and Galerkin projection are firstly obtained,
shown in Fig. 3. It is clearly seen that this model is a deep multi-layer and LSTM is then introduced to further correct the prediction. Hou et al.
neural network in time, in which all the layers share the same weights [149] modeled the influence of a disturbance to flow states from surface
[140]. The memory effect is considered by introducing input and output pressure measurement based on LSTM. In modeling unsteady aero­
at the previous time step to the present LSTM cell. Backpropagation dynamics, Li et al. [150] developed LSTM for nonlinear unsteady
algorithm can be used to train LSTM, which is an algorithm informing aerodynamics of bridge motion, thus constructing the FSI model for
the network to change its internal parameters [140]. Moreover, the flutter prediction. Recently, a significant contribution in data-driven
memory block of LSTM learns the underlying relationship of aerodynamic modeling based on LSTM was made by Li et al. [139].
input–output data at each time step. The LSTM cell contains three gates: Nonlinear unsteady aerodynamic ROM with varying Mach numbers was
the input gate, output gate and the forget gate. The following equations successfully constructed for nonlinear aeroelasticity. This model shows
are used to define the network at discrete time index k: an improved accuracy over classical neural network models in predict­
⎧ ing the aerodynamic loads and aeroelastic responses. From this work,
⎪ ( )
⎪ T
⎪ ik = sig Wi ⋅[yk − 1 , uk ] + bi
⎪ the capability of deep learning in handling large numbers of dataset is
⎪ ( )
⎪
⎪
⎪
T
̃ k = tanh WC ⋅ [yk − 1 , uk ] + bC
C shown, indicating that LSTM is more suitable to model complex dy­
⎪
⎨ ( )
T
fk = sig Wf ⋅ [yk − 1 , uk ] + bf namics with varying operation conditions. To the authors’ knowledge,
(12)
⎪ Ck = fk *( Ck − 1 + ik *C
⎪ ̃k this is the first application of a deep neural network to aeroelastic
⎪ )
⎪
⎪
⎪ T
⎪ ok = sig Wo ⋅ [yk − 1 , uk ] + bo problems of transonic flow. Li et al. [151] applied LSTM to model
⎪
⎩ yk = ok *tanh(Ck )
⎪ nonlinear aeroelastic behaviors of bluff bodies based on experimental
data and reproduced experiment results with a high level of accuracy.
where * denotes pointwise multiplication of two vectors. i, f and o are
the output of input, forget and output gates, respectively. C denotes the 3.2. Feature extraction for flow field data
cell state which changes slowly in time to represent long term memory,
as seen by its linear formulation. The cell input and output are given by u Unsteady flows are high-dimensional, nonlinear and multiscale in
and y, respectively, where y also means the short-term memory state nature. Understanding the complex flow physics relies on the extraction
which changes dramatically. W denotes the weight matrices, and sig and of main features known as coherent structures or modes. Coherent
tanh are sigmoid and hyperbolic tangent activation functions. The structures are characteristic and recurrent flow patterns that not only
purpose of unsteady aerodynamic modeling through LSTM neural reveal the essential flow physics, but also provide subspaces to construct
network is to identify a nonlinear mathematical relationship between low-dimensional flow models. Feature extraction methods mainly refer
system input and output: yk = f(uk ,yk− 1 ,Ck− 1 ). LSTM can be trained and to the mode decomposition methods that identify coherent structures
simulated by the Keras libraries in python [141]. Activation functions from a large flow dataset. Generally, flow modes can be extracted based
are defined as follows: on either energy or frequency, resulting in two main mode decomposi­
tion methods: POD [152] and Dynamic Mode Decomposition (DMD)
sig(x) =
1
(13) [153,154]. POD obtains spatially orthogonal flow modes according to
the energy content, while DMD obtains dynamic modes with a single
1 + e− x

frequency and growth rate to represent important dynamic features.

ex − e− x
tanh(x) = (14) To perform modal analysis, we need to first collect the flow snapshot
ex + e− x
data including the flow quantities of interest, like velocity, pressure and
LSTM has been used in modeling different dynamical systems. Vla­ temperature, etc. Each snapshot is a column vector with a large
chas et al. [142] utilized LSTM to forecast the evolution of dimension M (e.g., number of grids or sensors), xk ∈ RM , taken from one
high-dimensional chaotic systems, where LSTM was shown to outper­ time instant of unsteady simulation, or one flow state of steady simu­
form the Gaussian process model. Wan et al. [143,144] proposed lation, which is arranged into the snapshot matrix: X = [ x1 , x2 , .., xN ],
data-assisted reduced-order modeling that combines LSTM and analyt­ where N is the number of snapshot. Feature extraction methods
ical dynamical model, in order to predict the dynamics of extreme events decompose the snapshot matrix, sometimes in combination with the
in complex dynamical systems and the motion of spherical particles in governing equations, to obtain low-dimensional structures and features
fluid flows. Wang et al. [145] developed a ROM based on LSTM and that characterize the underlying flow dynamics.

Fig. 3. LSTM cell with various gates describing short and long term memory effects. Output of the present time step depends on current input, the output of the
previous time step (short-term memory), as well as previous state (long-term memory) [139].

8
J. Kou and W. Zhang Progress in Aerospace Sciences 125 (2021) 100725

3.2.1. Proper orthogonal decomposition r ( )

d ( ) ∑ ∑ r
POD is a procedure to extract a basis for a modal decomposition from ai = Ci + Cis + Lij + Lsij aj + Qijk aj ak (20)
a set of flow measurements [155], which was firstly used in the analysis dt j=1 j,k = 1

of turbulent flows by Lumley [152]. It is also known as Karhunen-Loève

transform or principal component analysis (PCA). POD aims at finding a where C, L and Q are coefficients for constant, linear, and quadric terms,
proper low-dimensional subspace that best approximates the data in the respectively. Coefficients with superscript s is the stabilization term to
l2-norm manner. In other words, for a given number of modes, POD can calibrate the long-term dynamics. Note that the quadric nonlinearity in
fit the data with the minimum l2 error, characterizing its energy opti­ the Galerkin model is induced by the convection term, while the diffu­
mality. For POD, each snapshot xk ∈ RM is represented by a linear sion term only works as a linear damping. This low-dimensional model
combination of the spatially orthogonal modes: respects flow physics since the governing equations are considered,
therefore theories on convergence and error estimate can be derived. A
∑
r ∑
r
pioneering work was done by Aubry et al. [162], where the
xk = u0 + aj,k uj = aj,k uj (15)
j=1 j=0 low-dimensional model for transition in the turbulent boundary layer is
modeled. Deane et al. [163] extended the Galerkin model in reproducing
where k is the snapshot index. u0 ∈ RM is the mean value of snapshot the flow in a periodically grooved channel and the wake of an isolated
matrix, where a0,k = 1. The jth POD mode vector and coefficients are aj,k circular cylinder. Ma and Karniadakis [164] generalized Galerkin
and uj ∈ RM , respectively. r is the number of modes to be retained and models for a three-dimensional flow past a circular cylinder. In order to
the dimension of low-dimensional subspace. For unsteady flows, a is a model transient dynamics, rather than the long-term periodic dynamics,
function of time, while u is a basis in space fixed for all snapshots. POD a landmark work for Galerkin models was done by Noack et al. [165]. A
mode is obtained through minimizing the projection residual between shift mode is introduced that significantly improves the resolution of the
the snapshot and the POD basis, through maximizing the following transient dynamics from the onset of vortex shedding to the periodic von
objective function [155,156]: Kármán vortex street. Finally, a least-order Galerkin model that de­
scribes the laminar periodic flow around a circular cylinder with an
∑ unstable steady solution and stable periodic vortex shedding with only
N ∑
r ⃒
( )⃒ ( )
max ⃒ xk , uj ⃒2 , subject to ui , uj = δi,j (16)
u1 ,…,ur
k=1 j=1 three POD modes is proposed. The energy transfer process between
modes was analyzed by Couplet et al. [166]. The influence of the pres­
where (⋅, ⋅) denotes the inner product between two snapshots. δi,j is the sure term in Galerkin models for incompressible flows was discussed by
Kronecker symbol satisfying δi,j = 1 for i = j and δi,j = 0 otherwise, Noack et al. [167]. Recently, the least-order Galerkin model with only
which defines the orthogonality of POD mode. Based on the snapshot five modes was proposed by Deng et al. [168], in order to model an
method [157], this optimization problem can be solved via SVD of the incompressible flow undergoing two successive supercritical bi­
data matrix X ∈ RM×N , or via an eigenvalue decomposition of the furcations of Hopf and pitchfork type. A two-dimensional wake flow
covariance matrix C = XT X [14]. SVD gives the POD modes directly around a cluster of three equidistantly spaced cylinders, named as the
from its left singular vectors U ∈ RM×r , while the mode coefficients are fluidic pinball, was investigated. For compressible flows, a stable inner
given by the product of singular value matrix Σ ∈ Rr×r and right singular product was proposed by Rowley et al. [169], which successfully models
vectors V ∈ RN×r : and controls [170] the cavity flow oscillations. Lucia and Beran [171]
extended the POD-Galerkin method to linearized Euler equations. Ad­
X = UΣVT = Ua (17) vantages of POD-Galerkin models were shown by comparing with the
direct projection method based on Taylor series expansion. Bourguet
It should be noted that since only energy optimality is considered,
et al. [161] used this intrusive ROM to model the transonic buffet
mixed frequency information is included in the POD mode coefficient.
problem. Huang et al. [172] and Xu et al. [173] extended
Recently, there is a growing interest in performing POD in the frequency
projection-based ROMs for modeling reacting flows and combustion
domain, resulting in modes that are both orthogonal in time and space,
instability problems. A Petrov-Galerkin projection method that better
like the Spectral POD (SPOD) [158,159]. POD modes and mode co­
handles nonlinear terms of projection-based ROMs was proposed by
efficients can give very useful information about spatio-temporal details
Carlberg et al. [174]. In addtion to the construction of Galerkin models,
of coherent structures and the interaction between dominant flow
model-based flow control has also been studied, like controling flow past
modes. In addition, the POD modes are also orthogonal subspaces that
an airfoil [175] or the stabilization of jet flows [176]. To guarantee the
can be used to perform model reduction of high-dimensional dynamical
observability and controllability for projection-based ROMs, Willcox
systems and construct ROMs to describe unsteady flows. Different ways
and Peraire [177] and Rowley [178] introduced balanced truncation
to model time coefficients lead to either intrusive or nonintrusive ROMs.
and proposed balanced POD method to construct better models for
Intrusive ROMs are utilized when the underlying governing equations
system control. Uncertainty quantification of flow dynamics based on
are known a priori, which are also named as projection-based ROMs
Galerkin models was studied in Ref. [179].
[160]. They are built from projecting the governing equations onto POD
Engineering applications of intrusive ROMs for complex geometries
subspaces through the Galerkin projection. The full-order dynamics can
have also been investigated in different literatures. Early versions of
be written as:
such models are based on the eigenmode approach, where the large and
d sparse discrete operator is directly reduced by its eigenvectors without
x(t) = f(x(t) ) (18)
dt any data [180,181]. Romanowski [182] constructed efficient and ac­
The reduced-order dynamics is obtained by using the mode basis x = curate ROMs for the discretized Euler equations based on POD, in order
Ua, resulting in a set of ordinary differential equations for the mode to achieve fast time accurate unsteady aerodynamic and aeroelastic
coefficient: simulation. POD-based ROMs for linear Euler equations that enable
aeroelastic simulation in the frequency domain for an airfoil and
d
a(t) = UT f(Ua(t) ) (19) turbomachinery have been developed by Kim [183] and Hall et al.
dt [184]. Thomas et al. [185] extended such models for the aeroelastic
This formulates a quadratic polynomial ordinary differential equa­ analysis of three-dimensional geometries in the frequency domain.
tion system for the temporal dynamics of mode coefficient [161]: LeGresley and Alonso [186] explored the applications of POD in aero­
dynamic optimization. Coupling both intrusive ROMs for compressible
Euler equations and reduced-order structural models for aeroelastic

9
J. Kou and W. Zhang Progress in Aerospace Sciences 125 (2021) 100725

prediction has been reported in Ref. [187]. Da Ronch et al. [188] pro­ based on either eigenmodes or POD modes for linear dynamic flows are
posed a direct reduction method for the linearized Jacobian matrix to very effective [5], but for nonlinear dynamic flows this topic is still an
predict the aeroelastic responses accurately. Chen et al. [189] used POD area of active research [222,223].
and balanced truncation to model unsteady transonic flows and coupled
with state-space structural models for flutter suppression. A novel 3.2.2. Dynamic mode decomposition
concept that transforms the POD-based ROM into a residual- DMD is a data-driven and equation-free algorithm that extracts
minimization formulation was developed in Refs. [190,191]. space-time dynamic information from experimental measurement and
Although extensive studies have been reported, a drawback for numerical simulation, which was proposed by Schmid et al. [154,224].
intrusive ROM approaches is that it usually involves the modification of Derived from the linear tangent approximation or Koopman theory
the source code, which may be impossible in commercial software. [153], DMD describes the evolution of flow observables in a linear
Moreover, it is also not applicable when the exact governing equations manner, therefore the resulting dynamic modes have a single frequency
of the dynamic system is unknown a priori. To handle these problems, and growth rate characterized by the eigenvalue (Ritz value) of system
purely data-driven methods that only require flow snapshots, known as matrix. This makes DMD very attractive in analyzing dynamically linear
nonintrusive ROMs, have also been proposed [192]. For modeling or periodic flow dynamics. In addition, based on the linear formulation
steady flows, the mapping between flow states, e.g., Mach number and of DMD, low-dimensional flow models can be obtained directly, without
angles of attack, and the mode coefficient can be constructed from an additional modeling process for time coefficients like POD. In other
existing interpolation methods. Fossati and Habashi [193] combined words, DMD is a purely data-driven method so that both the spatial
POD and Kriging interpolation to predict the aero-icing problems. The modes and temporal dynamics can be directly obtained from flow data.
performance of POD in modeling transonic flows with shock waves was These advantages lead to the extensive usage of DMD in analyzing
evaluated in Ref. [194], where strong oscillations near the shock wave different flow phenomena and modeling flow dynamics. So far, DMD has
were observed. In aerothermoelastic analysis, POD-Kriging models can been widely used in many fluid dynamic problems, including boundary
be used to model the distribution of pressure and heat flux [195–197]. layer flows [225], high-subsonic cavity flow [226], transitional jets
Huang et al. [198] developed an aerothermoelastic analysis framework [227], and wall-boundary turbulence [228], etc. Applications of DMD
based on POD-Kriging models, whose flexibility is enhanced using a are reviewed in Refs. [16,229]. A book of DMD refers to Ref. [230].
novel correction and scaling technique. Combining POD with Kriging An important theory that allows DMD to analyze nonlinear flow
interpolation or neural networks for aerodynamic optimization has been dynamics is the Koopman theory [231,232]. Rowley et al. [153] intro­
reported in Refs. [199,200]. Nonintrusive ROMs for unsteady flows can duced the Koopman operator and connected this theory with DMD.
be constructed based on the POD and the system identification method. Koopman theory is the nonlinear generalization of eigendecomposition,
Lucia et al. [201] used POD and Volterra series to model aeroelastic indicating that the dynamics of observable in finite-dimensional
systems and predict the LCO. Attar et al. [202] proposed unsteady nonlinear systems can be represented by the Koopman operator that
aerodynamic models based on POD and multivariable polynomials. describes the dynamics of infinite-dimensional linear systems. DMD is
Walton et al. [203] modeled the motion of the ONERA M6 wing in un­ an efficient algorithm to achieve the numerical approximation of the
steady flows based on POD and RBF interpolation. Lindhorst et al. [204] Koopman operator. An example of Koopman analysis for flow past a
constructed nonlinear unsteady aerodynamic models based on POD and cylinder is shown in Ref. [233]. For a nonlinear dynamical system xi+1 =
RBFNN, where the POD bases for surface pressure distribution are used. f(xi ) defined in the discrete time domain, an infinite-dimensional linear
Xiao et al. [205] proposed POD-RBF models for fluid-structure interac­ operator Ut can be used to describe the evolution of vector-valued ob­
tion simulation. Winter and Breitsamter [206] used POD and a fuzzy servables g(xi ) defined as a nonlinear transformation of the state vector
neural network to generate the unsteady aerodynamic models for load x:
prediction. Hesthaven and Ubbiali [207] proposed a non-intrusive ROM
Ut g(xi ) = g( f(xi ) ) = g(xi+1 ) (21)
based on POD and the neural network to predict the cavity flow with
different geometries. San et al. [208] also used POD with artificial neural From Koopman theory, once the observables are defined properly,
network to model transient flows. Wang et al. [209] proposed a greedy the linear operator can be accurately approximated. The selection of
nonintrusive ROM for parameterized time-dependent fluid dynamic nonlinear observables g is the key to obtain accurate Koopman eigen­
problems, where POD is used to extract temporal and spatial reduced values, eigenvectors, and modes, which still remains a challenge. A
bases, and RBF interpolation is used for predicting dynamics of unde­ comprehensive review of Koopman theory in fluid dynamics can be
termined coefficients. found in Mezić [234]. DMD approximates the evolution of state or
In addition, numerous studies have been conducted to extend the observable vectors from the current to the next time step with a
usage of POD. Gappy POD [210–212] is a typical method to recover the high-dimensional linear approximation:
entire flow information from incomplete flow measurements. It also
xi+1 = Axi (22)
motivates the development of the hyperreduction approach, where only
a fraction of flow data is used to approximate the nonlinear term of DMD aims at estimating the eigenvalues and eigenvectors of system
projection-based ROMs, thus reducing the computational cost [174, matrix A efficiently, which can be achieved through two types of algo­
213–215] of ROM generation. Projection-based reduced-order modeling rithms based on either a companion matrix or a similar matrix [235]. To
can also be transformed into an optimization problem where optimal perform DMD, an additional snapshot matrix Y = [ x2 , x3 , .., xN+1 ] is
subspaces can be derived for different objective functions based on collected by shifting the dataset X = [ x1 , x2 , .., xN ] by one time step.
either the integrated force [216] or the temporal sparsity of modes Then following DMD we can reach Y = AX. Note that here the snap­
[217]. For rigid body motions, Placzek et al. [218] proposed a ROM shot pair (xi , xi+1 ) is considered to form X and Y, which can be written in
treatment in the moving reference frame. Freno et al. [219] proposed a more generalized form (xi , f(xi )) [236].
dynamic basis functions to handle body motions. Yu et al. [220] pro­ For the first algorithm based on the companion matrix, when the
posed a body-fitted coordinate based ROM approach that enables snapshot number N is beyond a critical value, the subsequent flow field
reduced-order modeling for different boundary conditions and geome­ can be expressed as a linear combination of the previous snapshots,
tries. For model adaptation of dynamic systems with parameter varia­ indicating that the snapshots become linearly dependent:
tion, Peherstorfer and Willcox [221] proposed a dynamic reduced-order
xN+1 = c1 x1 + c2 x2 + c3 x3 + … + cN xN = Xc (23)
modeling approach to update the projection-based model in the online
operation condition. These works largely extended the applicability of After obtaining c from snapshots, a companion matrix S is formed:
projection-based ROMs. In summary, low-dimensional fluid models

10
J. Kou and W. Zhang Progress in Aerospace Sciences 125 (2021) 100725

⎡ ⎤
0 0 … 0 c1 suboptimal modes with less contribution to the flow energy, or numer­
⎢1
⎢ 0 … 0 c2 ⎥
⎥ ical modes with high damping rate that decay fast, can be retained with
S = ⎢
⎢0 1 … 0 c3 ⎥
⎥ (24) priority. This can lead to the failure of DMD especially for complex flows
⎣⋮ ⋮ ⋱ ⋮ ⋮ ⎦ involving strong nonlinearity or transient dynamics. Therefore, different
0 0 … 1 cN mode selection criteria have been proposed, thereby improving the
Then we have AX = Y = XS. The eigendecomposition of A is then predictive capability of standard DMD in complex flows [235,257,258].
transformed to the eigendecomposition of companion matrix S [153, An efficient mode selection that considers the global con­
154]. tribution/energy of mode in the sampling region was proposed by Kou
For the second method based on the similar matrix, SVD is firstly and Zhang [235]. This criterion is defined as the temporal integration of
performed for snapshot matrix X to provide the subspace U: X = the absolute value of time coefficient. To explain the criterion, DMD is
UΣVT . A minimization of the error between Y and AX is then formu­ firstly expressed as the superposition of normalized mode and
lated as follows: coefficient:
⃦ ⃦2 ∑
r
⃦
min⃦Y − UAΣṼ T⃦ ⃦ (25) xk = b j (k) φNorm, j (27)
̃ F j=1
A

where à is the similar matrix of A that is able to approximate the ei­ where φNorm, j is the jth dynamic mode normalized by its Frobenius norm
genvalues of A, while the eigenvectors of A can be obtained from U and with ‘‘Norm’’ representing the term ‘‘normalized’’, and bj (k) is the time-
eigenvectors of A.
̃ ‖⋅‖ is the Frobenius norm. More details for both al­ dependent mode coefficient of the jth normalized mode at time index k.
F ⃒ ⃒
gorithms are given in Ref. [235]. Finally, the flow state vector at time The criterion Ij is computed by integrating the coefficient ⃒bj (k)⃒ in time,
index k is represented as follows: resulting in continuous and discrete form as follows:

∑
r ∫t ∑
N ⃒
⃒ ⃒ ⃒
xk = αj φj μk−j 1
(26) Ij = ⃒b j (t)⃒dt = ⃒b j (k)⃒Δt (28)
j=1 k =1
0

where αj , φj and μj are the amplitude, basis function, and eigenvalue of

where Δt is the discrete time step. The above criterion can be understood
the jth DMD mode. A low-dimensional model with a very small
as an energy index that describes the global energy/contribution of each
dimension r is usually retained. The frequency and growth rate of each
DMD mode. It should be highlighted that the absolute value, rather than
mode are obtained from the eigenvalue. After DMD is proposed, many
the value itself, should be selected. For example, integrating the time
succeeding studies have been conducted to improve the preprocessing or
coefficient itself for purely periodic flows can lead to zero value, making
the postprocessing capability of the original DMD algorithm. Typical
the integration meaningless. As shown in Ref. [235], with this criterion,
methods can be briefly categorized into two groups, including the
improved performance compared with SPDMD can be reached, espe­
improvement of robustness and the selection of dominant modes. The
cially for transient flow dynamics. An explanation to this is that this
robustness of the standard DMD algorithm can be improved in many
criterion is a better measurement to evaluate the importance of mode, as
ways, resulting in different variants of DMD: (1) Residual minimization.
opposed to the traditional amplitude criterion used in the first step of
It transforms the DMD approximation into an optimization problem and
SPDMD (i.e., determining sparse structures). Following this line, an
corrects the eigenvalue, mode, or the mode amplitude, including Opti­
adaptive SPDMD approach is developed by Lai et al. [259], which takes
mized DMD [237], Optimal Mode Decomposition [238] or
advantages of both SPDMD in getting the sparse structure with ampli­
Sparsity-Promoting DMD (SPDMD) [239]. (2) Sampling conditions. It
tude correction, and the mode selection criterion in selecting important
improves DMD to handle complicated dataset, like the low-rank update
dynamic modes. Combination of HODMD with the criterion (HODMDc)
under large and streaming dataset [240], the spatial agglomeration of
has been investigated in Ref. [260], where the data for transient flows
high-dimensional data [241], the composite snapshot data [242], online
are used to extrapolate the limit-cycle states in transonic buffet flows.
estimation with parameter-varying dynamics [243], the debiasing of the
Other applications based on this criterion include shock buffet motions
noisy data [244,245], nonuniform sampling [246], and sub-Nyquist-rate
for the supercritical airfoil [261,262], acoustic field in radial compres­
sampling conditions [247]. (3) Mode calculation. Advantages of energy
sors [263], as well as automotive aerodynamics [264,265]. In the
optimality and mode orthogonality of POD, as well as the pure frequency
Koopman framework, this criterion may fail [266,267], since the most
characteristic of DMD are both taken into account, and a balance be­
energetic modes may not be the most critical modes, which should be
tween these two methods is maintained, like recursive DMD [248] or
the Koopman mode with invariant eigenfunction in time [267].
spectral POD [249]. (4) Mode coefficient correction. The mode coeffi­
Another important application is to extend the DMD-based low-
cient can be corrected by temporal integration [250] or modeled in a
dimensional models to flow control or multiphysics problems, which
nonintrusive way [251]. (5) Improved frequency resolution. Fre­
requires the construction of data-driven dynamic models with external
quencies of dynamic modes can be captured more accurately by multi­
inputs. To this end, some variants of DMD have been proposed in recent
resolution DMD [252]. Time-delayed coordinates can be used to
years [268,269], as well as the Koopman operator for model predictive
generalize DMD to model higher-order time derivatives, like
control [270]. Among them, there are two typical methods, named DMD
Higher-Order DMD (HODMD) [253]. (6) Extension to nonlinear sys­
with control (DMDc) and Input-Output DMD (IODMD) [269]. These
tems. To achieve this, one needs to select appropriate nonlinear ob­
methods extend the standard DMD framework to a state-space formu­
servables for Koopman analysis, from a priori physical knowledge or a
lation and identify the system and control matrices directly with
library of candidate functions, like extended DMD [254] or kernel-based
DMD-like algorithms, which have been validated by modeling the forces
DMD [255]. (7) Spatio-temporal mode decomposition scheme that
and pressures of a rapidly pitching airfoil [271], aerodynamic identifi­
captures both waves in space and frequencies in time, like
cation for aeroelastic system [272], flow over wind farms [269] and
Spatio-Temporal Koopman Decomposition (STKD) [256].
closed-loop separation control [273]. Since the system and control
Moreover, the selection of modes is also important to the perfor­
matrices are identified simultaneously, the intrinsic dynamics without
mance of DMD since the analysis and reduced-order modeling of flow
input may not be accurately identified. A modification based on these
dynamics depend on dominant dynamic modes. Unlike POD whose
methods was proposed by Kou and Zhang [274], named as DMD with
modes can be easily ranked from the energy, the selection of dominant
exogenous input (DMDX), where a two-step identification method to
DMD modes is not straightforward. If the modes are not ranked properly,

11
J. Kou and W. Zhang Progress in Aerospace Sciences 125 (2021) 100725

obtain the intrinsic flow dynamics and the controlled flow dynamics predictive performance can be improved compared with any model built
separately is developed. This method improves the accuracy of DMD in from only one data source, while the total cost for data generation still
modeling controlled flow dynamics that are close to the unstable state. remains low. These multi-fidelity models can be applied to similar
In addition to the applications and improvement of DMD, some ex­ problems that traditional data-driven modeling can handle with better
tensions are also worth mentioning. The link between DMD and global performance.
stability analysis is discussed by Ferrer et al. [275], where linear modes Multi-fidelity methods have been used in many applications, such as
for stability and sensitivity can be obtained without relying on the un­ mechanism analysis, optimization design, statistic inference, and un­
stable base flow. Performing DMD in space enables the prediction of certainty quantification [283,289]. Most of them are many-query
transition for flow past airfoils [276]. For steady flows, Koopman problems that require multiple model evaluations. Co-Kriging [290,
operator can be applied in a static way, leading to nonintrusive 291] is a typical multi-fidelity model which has been widely used in
modeling [277] of flow over airfoils in transonic flows. A recent work recent years. Multi-fidelity methods have also been used to solve prob­
shows that DMD is able to perform modal analysis for structural dy­ lems in fluid dynamics. Zheng et al. [292] and Badrya et al. [293]
namics [278], as well as to obtain fluid and structure modes simulta­ analyzed the flapping wing dynamics by coupling both low-fidelity and
neously for FSI systems [279]. DMD-based convergence acceleration high-fidelity flow solvers. Keane [294], Leffison and Koziel [295], Han
strategies, including unsteady flow simulation [280], steady flow et al. [296] and March and Willcox [297] all developed multi-fidelity
simulation [281] and adjoint-based optimization [282], have been surrogate models for aerodynamic optimization. Multi-fidelity aero­
developed for efficient CFD-based simulation. However, with all of these dynamic modeling has been studied by Ghoreyshi et al. [298], Mifsud
efforts and achievement, the extension of DMD to nonlinear et al. [288] and Murphy et al. [299] for flight simulation problems.
reduced-order modeling is still challenging, due to the fact that this Rokita and Friedmann [300] proposed a multi-fidelity model based on
concept is based on a linear approximation. To address this issue, POD and co-Kriging for hypersonic aerodynamic load prediction, where
Koopman theory [234] will still play an important role in the near POD bases are constructed by both low-fidelity and high-fidelity snap­
future. Furthermore, advantages of DMD over POD or eigenmodes in shots. Multi-fidelity methods are also considered for developing
constructing linear dynamic ROMs can be case-dependent, since in data-driven turbulence models, by using high-fidelity experimental or
practice there has to be a compromise between pure frequency content simulation data to correct the RANS model [301,302]. Uncertainty
and the energy optimality when modeling complex flows. quantification of fluid dynamic systems is also achieved using the
multi-fidelity modeling framework [303–306]. For example, Feldstein
3.3. Data fusion et al. [307] used the Gaussian process model to perform
blended-wing-body multidisciplinary analysis under uncertainty. It
Unsteady aerodynamic forces and flow field data can be obtained should be noted that most of these studies focus on the evaluation of
from different sources, such as numerical simulation, experiment, and steady aerodynamics for fast simulation and optimization, while the
flight testing. Most of the existing data-driven modelling approaches study of unsteady multi-fidelity aerodynamic modeling is lacking.
assume that the dataset only comes from one of these sources. However, Although very important, multi-fidelity modeling for unsteady
in practice, aerodynamic data always comes from different sources with aerodynamic is a relative new research topic. Since the generation of
different cost and accuracy. To maintain a good balance between model unsteady aerodynamic data is much more expensive than the steady
accuracy and cost of data generation, data-driven modeling motivated aerodynamic data, it is more important to explore the availability of
by data fusion is gaining attention. Although its definition varies in multi-fidelity methods for unsteady aerodynamics. To this end, Ghor­
different scientific communities, data fusion generally refers to the eyshi et al. [118] firstly showed that combining Euler and RANS data
process of combining data and information from multiple sources, in will lead to improved prediction that outperform models built upon a
order to refine, estimate, and achieve a greater understanding of the single data source. The model uses the NARX framework to construct the
data [244–247]. high-fidelity, low-fidelity and multi-fidelity ROMs separately. The
Data-driven modeling based on data fusion mainly refers to the multi-fidelity ROM is constructed from augmenting low-fidelity time-­
multi-fidelity modeling methods [283,284] from aerodynamic data with delayed predictions in the input term of the high-fidelity NARX model:
different fidelities. The other popular research topic in data fusion is the
y(k) = Ψ (u(k), u(k − 1), u(k − 2), …, u(k − m), yHF (k
data assimilation [285–287] that incorporates experimental observation
into numerical simulation. Here this concept will not be discussed since − 1), yHF (k − 2), …, yHF (k − n), yLF (k − 1), yLF (k
this method is more related to numerical methods rather than − 2), …, yLF (k − n) ) (29)
data-driven modeling. Generally, high-fidelity data are provided by
flight test, wind tunnel experiment or direct numerical simulation, while where the subscript HF and LF refer to high-fidelity data and low-fidelity
low-fidelity data can be obtained from numerical simulation with some data, respectively. This modeling framework is a straightforward
types of simplification, like coarser discretization, relaxed convergence extension of traditional unsteady ROMs to allow multi-fidelity data. One
tolerance, lower order of accuracy and omitted physics [288]. Among possible issue for this framework is the use of dynamic structure, which
different sources of data, there are two features: 1) as the fidelity of data may lead to stability problem in simulation. To address, an alternative
increases, the cost of data acquisition becomes higher; 2) due to the formulation is proposed recently by Kou and Zhang [308]. In this
limitation of the cost, available amount of data decreases with formulation, only time-delayed outputs from low-fidelity model are
increasing fidelity. Moreover, the information contained in high-fidelity included as additional input terms, thus the stability issue can be avoi­
data is sometimes limited, making aerodynamic modeling expensive and ded. This formulation is motivated by the co-Kriging model:
unsatisfactory based on these data. For example, most experimental data
yHF (x(k) ) = ρ ⋅ yLF (x(k) ) + z(x(k) ) (30)
are ensemble-averaged results from harmonic motions rather than
random motions, which have better quality but limited dynamic infor­
where ρ and z(x(k) ) are the constant multiplicative term and the addi­
mation. Motivated by these facts, multi-fidelity methods have been
tive/correction term at time index k. Following NARX formulation, the
developed and widely used, providing general frameworks to fuse data
additive term is a function of system state composed of time-delayed
with different fidelities in the modeling process. The basic idea is to use
input and output of the low-fidelity model. The state vector x(k) can
low-fidelity data or models (either physics-driven or data-driven) to
be represented as a whole, or separated into the quasi-steady state x1 (k)
provide the general trend of the underlying dynamics, while
and the unsteady state x2 (k):
high-fidelity data are used to correct the model to better reproduce
high-fidelity results. Through multi-fidelity modeling, the overall

12
J. Kou and W. Zhang Progress in Aerospace Sciences 125 (2021) 100725

⎧
⎨ x(k) = [x1 (k), x2 (k) ]T using the framework that combines linear and nonlinear models. These
x (k) = [u(k), u(k − 1), …, u(k − m) ]T (31) challenges and solutions will be detailed in this section.
⎩ 1
x2 (k) = [yLF (k − 1), yLF (k − 2), …, yLF (k − n) ]T
4.1.1. Model structure and stability
This results in two formulations of the correction term, where the
Due to the recurrent structure which includes time-delayed output
first is represented as a single function of the state, while the second is
feedback in the dynamic model, a stability issue can exist in model
represented as a combination of the quasi-steady term z1 and the un­
simulation [125–127]. It is even more pronounced in multiphysics
steady term z2 :
simulation, when the aerodynamic ROM is coupled with other physical
yHF (k) = ρ ⋅ yLF (k) + z(x(k) ) (32) models. Accurately modeling the unsteady effect and stability property
mainly relies on the design of the model structure. The goal of the model
yHF (k) = ρ ⋅ yLF (k) + z1 (x1 (k) ) + z2 (x2 (k) ) (33) structure determination is to find a parsimonious structure with an
optimal balance between model accuracy and model complexity. The
In the work of Kou and Zhang [308], the correction term is modeled
complexity of the model structure relies on delay orders and the
by a multi-kernel neural network. The multi-fidelity models are vali­
hyperparameters (i.e., unknown parameters that need to be optimized)
dated by modeling the aerodynamic loads of transonic flow past a
in the model equation. The delay orders can be selected by some crite­
NACA0012 airfoil, where Euler and Unsteady RANS (URANS) solvers
rions, e.g., the Akaike [316] or the Bayes [317] information criterion.
are used to provide low-fidelity and high-fidelity data, respectively.
Moreover, the user can also adopt a trial and error approach by gradu­
Accurate prediction of high-fidelity results is reported when only several
ally increasing the delay orders and evaluating the model performance
high-fidelity training cases from three sets of harmonic motion are
[318] to determine the optimal values. For example, for the linear ARX
available. It also outperforms the ROM only based on high-fidelity data
model, the model structure totally depends on delay orders m and n. If
when the same dataset is available, and maintains reasonable overall
the flow is stable, relatively small delay orders can be used. In transonic
computation cost.
aeroelastic problems where the flow is stable enough (i.e., far below the
Data fusion has also been introduced to handle high-dimensional
buffet boundary), these orders can be smaller than 10. However, when
flow field data. For example, the fusion of sparse sensor observations
the flow damping is close to 0 (e.g., near the buffet onset), or the fluid
and CFD solution is studied by Jiang et al. [309]. They determined a set
system is unstable, the delay order should be much larger to accurately
of dominant POD basis from a limited number of snapshots, and
model the stability property. For example, in the study of transonic
recovered the flow field with a new boundary condition from only
single-degree-of-freedom flutter near the buffet onset angle, these delay
several sensor measurements. Ruscher [310] fused the data from
orders are increased to 80 [65]. For the supercritical flow past a circular
experiment and CFD, in order to recover the missing experimental data
cylinder, i.e., the Reynolds number larger than 47 with periodic vortex
at the same spatial location in all snapshots. The proposed approach is
shedding, these orders can be 100 [67].
named as fused POD. Williams et al. [311] applied data fusion to
However, ensuring the stability of nonlinear systems is not
Koopman analysis and fused measurement data from different sources.
straightforward. Instability of nonlinear models can be induced by large
Zimmermann et al. [312] proposed a data fusion framework for steady
delay orders or overly complex model structure. For example, for
flow prediction based on POD and constrained least-squares method.
nonlinear RBFNN models, the number of hidden neurons and delay or­
Using high-fidelity integrated force data as the constraint, better pre­
ders will determine the model structure and complexity, which will
diction of surface pressure distribution is achieved. Following this work,
affect the model stability. Although large delay orders will increase the
Mifsud et al. [313] utilized gappy POD with constrained least-squares
approximation accuracy, these quantities should be limited in order to
method to fuse CFD simulation and sparse measurement of wind tun­
avoid the over-fitting and instability problems. Typically, delay orders
nel data. With this approach, the surface load distribution is recovered
between 2 and 6 will provide accurate approximation and good gener­
with improved accuracy. Recent works also focus on the reconstruction
alization capability for stable flow dynamics [100,121,122]. The num­
of the flow field based on the multi-fidelity flow field data. From Yu and
ber of hidden neurons is usually chosen between 50 and 200, which
Hesthaven [314] and Wang et al. [315], a mapping between
depends on the complexity of the problem. If a deep learning model is
high-fidelity and low-fidelity POD mode coefficient is constructed by
used, the number of both hidden layers and hidden neurons should be
either neural network or Kriging. The flow can be reconstructed based
taken into consideration [139]. The user may need to adjust these
on precomputed high-fidelity POD basis and the predicted high-fidelity
quantities to adapt to different problems. A proper design of the training
mode coefficients. This takes advantages of partially physical informa­
signal will also alleviate the stability problem [319]. Another effective
tion from low-fidelity data, resulting in improved accuracy compared
approach to obtain stable aerodynamic models is to switch off the
with traditional methods only based on high-fidelity data.
autoregressive term, making the system responses depend only on the
inputs, like the nonlinear Volterra series. This model can be regarded as
4. Improvement and challenges
a quasi-steady model, which improves the stability, but has to use large
input delay orders to achieve comparable (usually lower) accuracy [84].
4.1. System identification
An even more challenging task is to model nonlinear aerodynamic sys­
tems that are unstable or close to instability, like modeling
System identification is essentially a data-driven method. The ac­
large-amplitude motions in transonic buffet flows. Although significant
curacy, generalization capability and stability of system identification
progresses have been made through predicting the nodal-shaped oscil­
rely hevaily on the training dataset and the model structure. Accuracy
lations induced by the interaction between flutter and buffet [320,321],
and stability are basic requirements for the performance of data-driven
more in-depth investigations are still needed.
nonlinear unsteady aerodynamic models, while generalization capa­
bility is the key to engineering applications. Generalization capability
4.1.2. Generalization capability on new data
refers to the capability of data-driven models to adapt to new samples, i.
Data-driven aerodynamic models should have good generalization
e., to produce reasonable outputs for inputs not encountered during
capability in order to adapt to different types of input encountered in
training [109]. Moreover, nonlinear models are usually not able to
real applications. The generalization capability is reflected by the ac­
capture linear dynamics accurately, due to the intrinsic nonlinear
curacy in predicting the dynamics under new inputs with different
structures. For example, the nonlinear neural network models may not
amplitudes and frequencies, and the coupled response for multiphysics
be able to provide accurate prediction for the bifurcation point, which is
simulation. For parameter-varying models, generalization capability
characterized by linear dynamics [4]. This problem can be overcome by
also refers to the prediction performance under different operation

13
J. Kou and W. Zhang Progress in Aerospace Sciences 125 (2021) 100725

conditions, like the Mach number, mean angle of attack or geometrical 4.1.3. Consideration of linear and nonlinear dynamics
parameter. The training data, model structure and identification algo­ Since unsteady aerodynamics exhibits different characteristics under
rithm will all influence the generalization capability. The most impor­ different input amplitudes, neither a purely linear nor a purely nonlinear
tant factors among them is the design of training samples and the model is sufficient to capture all aerodynamic features. It has been re­
objective function used for training nonlinear aerodynamic ROMs. ported that nonlinear ROMs may produce an unsatisfactory prediction
A training case is needed for any data-driven model. A training case for purely linear dynamics [4,119]. A straightforward solution is to
contains all the available input-output data for ROM construction, which include samples with linear dynamics when training the nonlinear
should cover the frequency and amplitude ranges of interest. The fre­ ROMs [122]. However, the underlying ROM dynamics is still nonlinear.
quency range can be roughly determined based on the problem of in­ In addition, there are better alternative solutions that adapt to linear and
terest. For example, for aeroelastic applications, instability occurs due to nonlinear dynamics by constructing a hybrid ROM framework composed
frequency coalescence, whose frequency usually lies between the fre­ of both linear and nonlinear models. As emphasized in Ref. [319], the
quency of the dominant structural modes. Therefore, the training signal key to modeling nonlinear dynamics is to first construct a linear dynamic
should cover the frequency of the first several structural modes. As for model. Then, with the linear dynamic model fixed, a nonlinear model
amplitude, if the user is interested in linear stability problems like the (like Volterra series or neural network) is used to represent the differ­
flutter phenomenon, the amplitude of the training signal should be small ence between the nonlinear dynamics of the flow and the linear dy­
enough to ensure linear dynamics. A useful approach to determine the namics of the flow. An advantage of such a combination lies in the
linearity of the signal is to compare the output of the same signal with possibility of automatically switching to linear models under a small
different amplitudes [121], as will be shown in Sec. 5.1. If the user is input, rather than using nonlinear models to describe linear dynamics.
interested in LCO behaviors induced by aerodynamic nonlinearity, the Different hybrid modeling frameworks have been developed to
amplitude range should be enlarged to include the nonlinear aero­ consider both linear and nonlinear aerodynamics, which include both
dynamic characteristics. Thereafter, the number of training samples and linear and nonlinear subsystems with either a serial or a parallel
the shape of the training signal are focused. An excessive number of data connection. The serial structure leads to the block-oriented models
will lead to the degradation of performance when the model structure is discussed in Sec. 3.1.7, which have been widely used in nonlinear sys­
not flexible enough to learn all the information. Therefore, for different tem identification. Since block-oriented models are combinations of
applications, the number of samples varying from 500 to 5000 is usually dynamically linear and statically nonlinear subsystems, the time-
acceptable. However, with the development of deep learning, deep delayed output feedback does not exist, therefore the stability issue
neural networks are usually capable of learning from a very large can be avoided. Applications of Hammerstein models are given in Refs.
dataset. For example in the LSTM-based ROM [139] for unsteady dy­ [129,130]. Recent studies have focused on Wiener models [132,134,
namics across a range of flow conditions, 3500 samples are selected for 325], where in Ref. [134] improved Wiener models with quasi-dynamic
each Mach number, resulting in 17500 training samples in total. How­ nonlinear blocks are introduced, as discussed in Sec. 3.1.7.
ever for such deep learning models, care must be taken to avoid the The basic idea of parallel-structured modeling is as follows [56].
over-fitting problem. Firstly, a linear dynamic model is constructed, whose deviation with the
The shape of the training signal should be designed carefully to cover original system will contain all unmodeled nonlinear effects. Secondly, a
enough dynamic information with a relatively small number of samples. nonlinear model is used to approximate this deviation to account for
Different training signals have been proposed. For linear ROMs, step- unmodeled nonlinearity. This idea is attractive, since the linear model
type [92], multistep [59] and 3211 [57] signals have been proven to obtained in the first step has good robustness, while the nonlinear model
capture the aerodynamic characteristics accurately under small pertur­ further improves the overall performance. As noted in time-series fore­
bations. However, the design of input signals for nonlinear ROMs is casting studies [326], using a parallel hybrid modeling technique will
much harder, because the nonlinear aerodynamic features change with use each model’s unique feature to capture different patterns in the data,
amplitude and frequency in a complicated manner. At present, ampli­ thereby improving the forecasting performance. This framework has
tude sweep with frequency modulation signals [204], filtered white been investigated in structural dynamic identification [327], geomag­
Gaussian noise (FWGN) signals [132], randomlike signals [319], the netic field identification [328], time-series prediction [329], etc. The
amplitude-modulated pseudorandom binary signals [120], frequency basic idea of parallel modeling of unsteady aerodynamic has been
sweep signals [320] and self-excited vibration responses [4] have been validated in Ref. [330]. Mannarino and Dowell [319] proposed a par­
used for training different nonlinear ROMs. These signals have shown allel framework trained by a two-level training procedure, which is then
broad frequency and amplitude ranges with multiple types of motion applied to flutter suppression [331]. This model is applied to identify the
and a small number of sample. LCO behaviors of an airfoil with four degrees of freedom (DOFs). Huang
Choosing an objective function is to find a criterion to assess et al. [332], Yang et al. [321] and Liu and Gao [333] developed ROMs
meaningful model construction. ROMs can easily overfit data when only based on nonlinear state-space models, in order to predict transonic
the approximation of training data is set as the objective function. It has aeroservoelastic and complicated aeroelastic responses. Another
been shown that when the model is evaluated and calibrated by pre­ contribution to parallel modeling includes the layered ROM and the
dicting a validation case, whose samples are different from the training hybrid ROM proposed by Kou and Zhang [318,320]. Structures of these
signal, a good generalization capability is maintained [121]. This is models are shown in Fig. 4, where ARX and neural network are used to
exactly the cross-validation strategy [322] used in the machine learning describe linear and nonlinear subsystems, respectively. Each subsystem
community. is identified by a particular training signal exhibiting either dynamically
With the need of data-driven modeling for parameter-varying sys­ linear or nonlinear features. The structure of the layered ROM is shown
tems, additional considerations to guarantee generalization capability in Fig. 4a, which adopts isolated linear and nonlinear subsystems and
are required. Different approaches have been applied to model assumes that the nonlinear deviation makes no contribution to the linear
parameter-varying systems, including direct interpolation of model pa­ model. The hybrid ROM shown in Fig. 4b is a fully coupled
rameters [323], augmenting input with the parameter [100,113,139], parallel-structured framework where time-delayed output feedback is
and correcting the model output [92,324]. All of these works require an introduced to both subsystems, in order to better reflect unsteady flow
additional sampling procedure in the parameter space, in order to cap­ physics. This variation in model structure, however, makes a big change
ture the dynamics under multiple system parameters. This has moti­ in model response, especially the nonlinear residual [320]. Results
vated new ways to design training signals and objective functions, which showed that the hybrid ROM gave better accuracy in predicting the
remains an open question. flutter boundary and LCO responses [320]. In addition, the hybrid ROM
also reproduced the complex aeroelastic responses with reasonable

14
J. Kou and W. Zhang Progress in Aerospace Sciences 125 (2021) 100725

Fig. 4. The structure of (a) the layered ROM and (b) the hybrid ROM. The layered ROM adopts isolated linear and nonlinear subsystems, while the hybrid ROM
utilizes the autoregressive structure for the whole system [320].

accuracy, like the beat phenomena at large reduced velocities and the introduced to represent the unsolved modes. Rempfer and Fasel [334]
nodal-shaped oscillation induced from the interaction between buffet further extended this model by proposing the modal eddy viscosities,
and flutter in transonic flow. arguing that the eddy viscosity should be scale-dependent thus needs to
be calibrated with the scale of mode. Noack et al. [167] introduced the
4.2. Feature extraction modal power balance to Galerkin models and used the pressure term for
closure purpose. These closure methods can be used to correct the linear
Feature extraction of flow dynamics usually involves dimensionality part of the Galerkin model, but the energy transfer is in nature a
reduction of the high-dimensional dynamics based on mode decompo­ nonlinear mechanism [335]. Therefore, state-dependent nonlinear eddy
sition, which has unresolved dynamics that need additional closure viscosity has been developed, which is derived from a finite-time ther­
modeling. In addition, some research began to focus on constructing modynamics formalism [336]. This nonlinear eddy viscosity model has
parameter-varying projection-based data-driven models, with effective been shown to produce accurate and robust Galerkin models for LES of
algorithms to ensure the generalization capability under different pa­ the Ahmed body [223]. Wang et al. [337] compared different closure
rameters. To handle more complicated nonlinear flow dynamics, models for turbulent flows past a three-dimensional circular cylinder,
nonlinear feature extraction methods need to be developed in the future. including dynamic subgrid-scale, variational multiscale, mixing length
These three topics will be discussed in detail. and Smagorinsky closure models. With different criteria, the dynamic
subgrid-scale and the variational multiscale models are shown to be the
4.2.1. Closure modeling most accurate. The closure modeling can also be transformed into an
Data-driven models based on feature extraction, like projection- optimization problem to obtain optimal accuracy and stability
based ROMs, are generated with a certain level of approximation. This [338–340]. Balajewicz et al. [222] proposed a constrained POD
approximation comes from the mode decomposition itself, where only formulation for unsteady lid-driven cavity flows, where optimal modes
dominant low-order modes are retained in the low-dimensional model. with power balance and stability can be obtained by rotating the main
Although high-order modes are important to reflect the dissipation ef­ POD modes. In addition, closure and stability can also be achieved
fect in dynamics and stability, retaining too many of these modes will numerically, where different methods have been proposed, including
increase complexity and computational cost of the model. Therefore, stable inner product [169,341], artificial viscosity [171], pole assign­
these fine-scale structures have to be omitted, which are referred to as ment [342], shooting method [343], Petrov-Galerkin projection [174],
unsolved dynamics. This can make POD-Galerkin models become un­ constrained projection with prior information about the long-time
stable when they are integrated for long-term dynamics, since these attractor [344] and Jacobian-type stabilization [345], etc. Stability of
models ignore the energy transfer between resolved and unsolved dy­ POD-Galerkin models for shock-vortex interaction and combustion
namics. A proper closure modeling is needed for the unsolved dynamics, problems are discussed by Rezaian and Wei [346] and Huang et al.
which not only stabilizes the numerical simulation, but also adds addi­ [347], respectively.
tional physics to data-driven models. Data-driven methods have also motivated the development of data-
Extensive works have been proposed to model the closure term of driven closure models. A straightforward data-driven closure approach
projection-based ROMs. Physics-based closure models are constructed is to directly apply regression methods to compute and correct the co­
from basic physical insights or closure strategies for Navier-Stokes efficients in Galerkin models. These coefficients can be obtained by
equations. Since the instability of long-term dynamics is mainly either the Tikhonov regularization [348] or sparse regression [349]
caused by the positive eigenvalues of linear terms, it is straightforward method. Moreover, the unsolved dynamics can also be modeled by
to introduce numerical dissipation that will move the unstable eigen­ data-driven modeling methods as well. This takes advantages of both
values to the stable region. To this end, an eddy viscosity term can be physics-driven and data-driven modeling, where the main dynamics is
introduced to correct the eigenvalues. Closure modeling has been dis­ resolved by known physics and modes, while the omitted physics is
cussed in the pioneering work of Aubry et al. [162] in modeling tur­ described by the closure model. San and Maulik [350] used the neural
bulent boundary layer flows. By mimicking the Smagorinsky-type model network as the closure model for POD-Galerkin models, in order to
for the Reynolds stress in LES, a single constant eddy viscosity is capture the effect of truncated modes. Parish and Duraisamy [351]

15
J. Kou and W. Zhang Progress in Aerospace Sciences 125 (2021) 100725

utilized the Mori-Zwanzig (M-Z) formalism to develop closure models of parametrized partial differential equations. Low-rank libraries
LES, which represents nonlocal behavior of coarse-grained scale via a composed of POD modes at different bifurcation regimes are con­
finite memory approximation. Pan and Duraisamy [352] proposed a structed, while compressive sensing is used to select the most important
data-driven discovery method of closure models based on operator modes from sparse flow measurements, and construct low-dimensional
inference to extract the governing dynamics of closure. Wang et al. Galerkin models. In addition, local basis functions can be constructed
[353] developed a LSTM-based closure for parametric POD-Galerkin in time to model time-dependent flow dynamics, such as dynamic basis
models based on the Mori-Zwanzig formalism. Another LSTM-based update [221], time-dependent basis [369] and online system update
closure model for POD-Galerkin models is proposed by Ahmed et al. [243].
[354]. Hijazi et al. [355] merged projection-based and data-driven Variable separation is a common approach to solve partial differen­
strategies to create POD-Galerkin models for turbulent flows. In sum­ tial equations. It also lays the foundation of reduced-order modeling,
mary, closure modeling is a critical step needed for data-driven models which separates the space-time dynamics into spatial modes and tem­
based on feature extraction, since the unsolved dynamics must be poral coefficients. Similarly, for parameterized dynamical systems, the
handled properly. variation with respect to parameter can be incorporated into a
parameter-dependent term in the variable separation formulation. This
4.2.2. Generalization capability for parameterized systems motivates the reduced basis method [370,371] and proper generalized
As discussed in Sec. 4.1, generalization capability is required for decomposition (PGD) [372]. The solution is separated into the product
data-driven modeling, especially for modeling parameter-varying of space, time and parameter bases. The variation with respect to
models. These parameter-varying data-driven models should repro­ parameter is included in the parameter basis, which is decoupled from
duce the dynamics at different operation conditions with a good level of spatial and temporal dimensions. As an example, a greedy nonintrusive
accuracy. Projection-based ROMs for parametrized dynamical systems ROM based on POD and RBF interpolation was proposed for parame­
are reviewed in Ref. [160]. The basic idea is to obtain data-driven dy­ terized time-dependent unsteady flow dynamics [209].
namic models for unseen states from neighboring system information
built from the training data and system. Therefore the design of exper­ 4.2.3. Nonlinear feature extraction
iments in the parameter space is crucial. Latin hypercube and greedy Nonlinear feature extraction, including novel machine learning
algorithms are popular methods in sampling in the parameter space. methods and nonlinear extension of mode decomposition, is being
So far, there are three approaches to achieve parameter-varying focused to handle more complicated flow physics with nonlinear inter­
data-driven modeling: interpolation-based method, local basis function action in highly turbulent flows. It should be mentioned that, even
and separation of variable. The interpolation-based method performs though POD is a linear representation of high-dimensional flows, the
interpolation to mode basis or system coefficient from those of neigh­ resulting POD-Galerkin model is a low-dimensional nonlinear model
boring parameters. Amsallem and Farhat [356] introduced an interpo­ because of the quadratic term coming from the convection term in the
lation method of POD bases in the tangent space of the Grassmann Navier-Stokes equations. Nonlinear surrogates can also be used to model
manifold. This approach has been used to predict the unsteady aero­ the nonlinearity of the POD coefficient in a nonintrusive manner. As for
dynamic loads for an F-16 aircraft model with varying Mach numbers. DMD, the nonlinear extension has been given by the Koopman theory.
The method is then extended to many applications, including adaptive As introduced in Sec. 3.2.2, the Koopman operator describes the evo­
model update [357], time-dependent dynamics of flow past a cylinder lution of observables in nonlinear dynamical systems [231]. It points out
[358], unsteady aerodynamics for hypersonic vehicles [359], and that the evolution of the observables of a finite-dimensional nonlinear
non-intrusive model interpolation for parametric flow dynamics [360]. system over time can be represented by an infinite-dimensional linear
The other approach is to directly interpolate the local reduced-order Koopman operator. Observable refers to the state variable defined in the
state-space matrices. Bui-Thanh et al. [361] used an optimization state space, which is the nonlinear transformation of the primitive var­
formulation to explicitly derive nonlinear parameter dependence for iables. Mezić et al. [232,234] discussed the use of Koopman theory in
parametrized reduced-order modeling, which is used for unsteady reduced-order modeling of stochastic or deterministic dynamical pro­
aerodynamics with geometric variability. The matrix interpolation is cesses, and found that the spectral decomposition of the Koopman
also discussed in Degroote et al. [362], where a parametric ROM based operator is related to the spatial mode of the dynamical system. This
on spline interpolation in the Reimann manifold was proposed. The mode has a single frequency, which provides a basis for estimating the
parametrized ROMs are compared with Kriging models for a thermal fin Koopman operator using DMD. Rowley et al. [153] therefore discussed
design problem and for prediction of contaminant transport. Son et al. the relationship between DMD and Koopman mode decomposition, and
[363] used interpolation on Grassmann manifolds to construct system showed that the DMD works as the numerical approximation of the
matrices of parametric ROMs. De Sturler et al. [364] proposed Koopman operator. It is evident from extended DMD [254] and
system-theoretic model order reduction methods to construct inter­ kernel-based DMD [255] that once suitable observables are chosen,
polatory parametric ROMs based on rational interpolation. DMD can be used to identify nonlinear models. The Koopman theory
Local basis function methods combine both classification and local essentially states that for a non-linear system of finite dimensionality, a
subspaces. In the offline phase where the model is being built, classifi­ set of coordinate systems can be found to make the system linear [373].
cation in the parameter space is first performed, followed by con­ This theory provides an interesting perspective for feature extraction of
structing local bases and dynamic models for each class. In the online non-linear systems. However, the main challenge in the Koopman theory
phase where the model-based simulation is performed, the input is first is the selection of observables from data and the underlying physics.
classified, then the ROM in the corresponding class is called [365]. In Machine learning methods, like neural network [374], cluster anal­
Ref. [366], a local reduced-order model is proposed, where the solutions ysis [434], manifold learning [435], gene expression programming
are clustered by an unsupervised algorithm, and are represented by local [375], etc., can reveal the intrinsic connections for a large flow dataset
reduced-order bases in each class. To evaluate the nonlinear terms in the from new perspectives. Typical machine-learning-based methods that
Galerkin model, Peherstorfer [367] proposed a localized discrete have been applied in fluid dynamics include cluster-based ROM [376],
empirical interpolation method to construct online ROMs for parame­ cluster-based network model [377], Kernel POD (Kernel POD) [378],
terized systems. The training data is classified by the k-means cluster Isomap [379] and local linear embedding [380]. Cluster analysis is a
method, and a set of local reduced basis is constructed for each class. method to group samples into several central states, which can be used
Machine learning is used to generate a classifier to identify the local to analyze the similarity of different dynamical processes, and design
subspaces adaptively as the computation proceeds. Brunton et al. [368] control laws [381]. Manifold learning theory indicates that the
combined compressive sensing and POD for reduced-order modeling of high-dimensional flow comes from the mapping between the

16
J. Kou and W. Zhang Progress in Aerospace Sciences 125 (2021) 100725

low-dimensional manifold and the high-dimensional physical space. time-delayed history of either high-fidelity or low-fidelity data, as pro­
Therefore, high-dimensional data can be represented by the posed by Ghoreyshi et al. [118] and Kou and Zhang [308]. However,
low-dimensional space defined in the manifold. The essence of manifold when time-delayed high-fidelity output is considered [118], feedback
learning is to relate the flow states in the manifold based on a certain terms of the whole multi-fidelity model output are introduced, which
metric of similarity. Using the similarity relationship, the new flow state may exhibit instability in simulation. In contrast, when only low-fidelity
can be represented by known flow information. The most polular time-delayed effects are introduced [308], the aerodynamic prediction
manifold learning method is the Isomap algorithm [436]. Compressive is stable, but unsteady effects are partially reflected which may lead to
sensing and sparse coding methods that extract dominant and sparse lower accuracy. Both problems indicate that future works are still
information are another novel tools for feature extraction. Using needed to maintain both stability and accuracy of the multi-fidelity
compressive sensing to reconstruct surface pressure distribution from unsteady aerodynamic model. In addition, it is worth mentioning that
limited flow measurements has been reported by Zhao et al. [382]. the data-assisted reduced-order modeling [143,144] framework is able
Sparse regression has been shown to identify the governing equations of to inform the development of the multi-fidelity modeling framework,
physical laws from data [373]. Deshmukh et al. [217] combined sparse where the leading-order approximation is given by a low-fidelity phys­
coding to build a low-dimensional Galerkin model, resulting in flow ical model, while the high-order effect is then modeled in a data-driven
modes that are sparse in time. Loiseau et al. [383] proposed a sparse manner. If more than two sources of data are present, one has to consider
ROM framework, where the dynamical model is directly identified from the framework that can be adapted to more data sources. Straightfor­
sensor measurements or flow snapshots. Network and graph theory ward extension of existing models [118,308] and a recursive framework,
[384,385] have been used to analyze the vortex interactions in complex e.g., recursive co-Kriging [303], can be considered. The quantification of
fluid flows. Furthermore, a nonlinear feature can be learnt directly from uncertainty and reliability for multi-fidelity unsteady aerodynamic
data by a neural network, i.e. by constructing an autoencoder to learn a models is another issue to be solved, where the Kriging model which
self-mapping of the flow snapshot, where the model reduction is ach­ provides the confidence interval of prediction can be a good option.
ieved in the hidden layer with a reduced number of neurons. This idea For multi-fidelity modeling based on feature extraction, one way is to
was first used by Milano and Koumoutsakos in the modelling of compute dominant modes from multi-fidelity snapshots composed of
near-wall turbulence [386]. In recent years, with the development of both low-fidelity and high-fidelity flow data, and use surrogate models
deep learning, such a feature extraction method has retained attention in to predict the high-fidelity mode coefficient [288,300]. The other option
fluid dynamic problems [387]. Autoencoders based on deep learning is to construct mappings between mode coefficients of low-fidelity and
have been used as a dimensionality reduction method to obtain ROMs high-fidelity bases, i.e., aHF = f(aLF ) [314,315]. Multi-fidelity modeling
for nonlinear flow patterns [388,389]. Autoencoder can also be used to from a few experimental measurement and full-field simulation data can
find the observables of Koopman operator [390]. This modelling be used to predict the load distribution based on methods like gappy
approach has now been validated for fundamental flow problems by POD [313]. However, a prerequisite for these methods is that the POD
Brunton et al. [391], Otto and Rowley [392], as well as Pan and bases should be available, thus the accuracy is limited by the POD
Duraisamy [393]. The autoencoder-based dimensionality reduction is method itself. Therefore, nonlinear extension of other feature extraction
also suitable for geometric dimensionality reduction in aerodynamic methods needs to be studied in the future. When both input-output data
optimization [394]. In addition, the use of deep learning to directly and flow field data with different fidelities are available, multi-fidelity
extract information from geometry and body motion for flow field modeling can benefit from data assimilation and nonintrusive
reconstruction is becoming attractive. Sekar et al. [395] proposed a modeling. For example, Zimmermann et al. [312] fused the integrated
data-driven approach for the prediction of the incompressible laminar aerodynamic coefficient data with flow modes for steady flow recon­
steady flow field over airfoils based on deep neural networks. The pre­ struction. Yao and Marques [402] proposed a nonintrusive ROM based
diction of unsteady flow based on deep neural network is discussed by on a combination of discrete empirical interpolation method, POD and
Han et al. [396] and Bukka et al. [397]. The wake structure with peri­ RBF neural network, which obtains both the aerodynamic loads and the
odic vortex shedding of flow past a cylinder and an airfoil is predicted flow details from sparse measurements on the wing surface.
precisely [396]. Recently, neural network has been used to examine the In practice, a small amount of high-fidelity samples with less di­
causal interactions of energy-eddies in wall-bounded turbulence [398]. versity is usually available. The small amount is due to the high cost of
Almost all of these works make use of machine learning to perform data generation, which may take lots of experimental or computational
regression, while some works have also been conducted to utilize the resources. The lack in diversity means that high-fidelity data may only
classification technique of machine learning, like classifying vortex include part of the dynamic information of interest. This leads to a
wakes [399] and detecting turbulence in the boundary layer [400] or typical few-shot learning problem in machine learning [403], where the
flow past a cylinder [401]. high-fidelity training data are not enough. Online adaptation of the
multi-fidelity model, where the model is updated when a new
4.3. Data fusion models high-fidelity sample becomes available [404], is also worth investi­
gating. Recently, physics-informed data-driven modeling has become a
The essence of fusing data from multiple sources is to model the new trend. Leveraging the underlying physics of low-fidelity and
correlation between multi-fidelity data, thus improving the overall high-fidelity models will be very helpful to inform the construction of
predictive capability. A mapping or calibration term that transfers low- multi-fidelity models. By introducing conservation laws or physical in­
fidelity data to high-fidelity data should be constructed, which requires formation to the modeling process, the resulting models can be more
a proper model framework. Once this relation is modeled, high-fidelity consistent with flow physics and thus give reasonable predictions. This
data can be recovered from low-fidelity results computed at similar approach is rarely considered in multi-fidelity modeling, especially for
operation conditions. unsteady aerodynamics. However, it has been studied in other related
For input-output data, the steady aerodynamics has been successfully area, like the introduction of Galilean invariants in machine learning of
modeled based on co-Kriging models [284]. However, the unsteady turbulence [405] or the calibration of coefficients in turbulence models
multi-fidelity modeling still remains to be a new research topic, and a lot [406]. In addition, the physics-informed neural network [407–410] has
of investigations need to be done. As discussed for the system identifi­ been developed to learn the differential equation, as well as inferring
cation method, the time-delayed effects of unsteady aerodynamics solutions using noisy multi-fidelity data [411] or sparse and noisy data
should be accounted in the multi-fidelity unsteady aerodynamic model. [412]. Recently, this physics-informed deep-learning framework is
Currently, the multi-fidelity unsteady aerodynamic models are based on extended with data assimilation [413] to simultaneously exploit the
the NARX framework, where the unsteady effect is considered by the information from flow visualization and the Navier-Stokes equations. All

17
J. Kou and W. Zhang Progress in Aerospace Sciences 125 (2021) 100725

of these provide lots of opportunities to multi-fidelity modeling of un­ Therefore, the deviation between the normalized lift coefficient and that
steady aerodynamics in the future. at A = 0.001 is also compared in Fig. 6b. From Fig. 6, when the relative
amplitude is at A = 0.020, 0.005 and 0.001, the normalized lift co­
5. Applications efficients are nearly coincident. This indicates that when A is lower than
0.020, the lift forces vary in a linear fashion, which is appropriate for
This section overviews several benchmark cases to show the appli­ modeling dynamically linear behaviors. When the relative amplitude
cations of data-driven models in efficient aerodynamic simulation and increases to A = 0.050 (pitching motion at approximately 2.5◦ ), the
flow physics analysis. Two typical flow phenomena, including incom­ normalized aerodynamic coefficients cannot keep the coincidence,
pressible laminar flow past a cylinder and compressible transonic flow where aerodynamic nonlinearity starts to arise. As amplitude increases,
past airfoils and wings, are investigated. The flow past a circular cyl­ the deviation also increases, indicating that the nonlinearity is becoming
inder is a classical and representative problem in fluid dynamics, where stronger. Therefore, when the amplitude reaches these values, dynam­
typical instabilities are present, like Hopf bifurcation, von Kármán ically nonlinear aerodynamic models should be considered. In addition,
vortex shedding, shear layer separation, etc. In addition, due to the aerodynamic nonlinearity also depends on the frequency of motion,
shock-boundary layer interaction, transonic flow exhibits massive sep­ where at lower frequency, nonlinear effect is stronger [414].
aration, strong nonlinearity and unsteady flow physics. Therefore, ap­ As an example, a nonlinear RRBFNN model is constructed based on
plications of data-driven modeling in compressible flows can be well the random training dataset at relative amplitude A = 0.150. This
represented by transonic flow phenomena, like buffet and various training case ensures that the resulting model will be able to capture
aeroelastic responses. nonlinear and unsteady effects at a large amplitude. To improve the
generalization capability, a validation case that is different from the
training case is introduced to evaluate the objective function, while PSO
5.1. Unsteady aerodynamic prediction is used to find the optimal width parameter for each activation function.
The aerodynamic ROM is used to predict the moment coefficient of the
The first case considered is the nonlinear unsteady aerodynamic same random signal at different amplitudes, as shown in Fig. 7. PSO-
modeling of airfoils in transonic flows based on system identification, Training and PSO-Validation refer to two ROMs trained without and
which models the input-output mapping between motion and the with the validation case. Results show that by introducing the validation
aerodynamic coefficient. As discussed in Sec. 4.1.2, the FWGN signal case for model training, the aerodynamic ROM based on RBFNN can
shown in Fig. 5 is used as the training signal, which covers sufficient capture both linear and nonlinear aerodynamics at different amplitudes
information about frequency and amplitude. This signal is used to with an improved accuracy.
generate the training data of a pitching NACA0012 airfoil at Mach For aerodynamic prediction across multiple Mach numbers, reduced-
number 0.8 and Reynolds number 3 × 106. The maximum pitching angle order modeling becomes more challenging, where the ROM should be
in the figure is about 74.5◦ , which will be scaled by the relative ampli­ more flexible to adapt to the nonlinearity changing with parameter.
tude A. Reduced frequency kf is used to quantify the frequency of the Based on the RRBFNN aerodynamic model, this increase in flexibility
signal, defined as: can be achieved by introducing more nonlinear characteristics to the
ωb activation function. For example, Kou and Zhang [112] proposed
kf = (34) multi-kernel neural networks, which capture the complex nonlinear
V∞
characteristics under multiple flow conditions by linearly combining the
where ω is the oscillatory frequency, b denotes the half chord length, V∞ Gaussian and wavelet basis functions as the hidden basis functions. The
is the freestream velocity. In Fig. 5a, T = t/(2b/a) is the non-dimensional unsteady aerodynamic prediction of the moment coefficient from a
time, where t denotes the physical time, and a is the sound speed. Power typical harmonic pitching motion of a NACA64A010 airfoil is shown in
spectrum analysis of the training signal is shown in Fig. 5b. From Fig. 5b, Fig. 8. The model is constructed by five datasets of random motion at
high energy is concentrated between reduced frequency ranging from Mach numbers ranging from 0.74 to 0.82 with interval 0.02. In these
0 to 0.5, which is wide enough to ensure good generalization capability. figures, ROM 1 refers to the standard RBFNN model, while ROM 2 and
To investigate the aerodynamic nonlinearity, a simple approach in ROM 3 refer to the multi-kernel neural network composed of symmetric
Ref. [121] is used by comparing the aerodynamic responses for the same and asymmetric basis functions, respectively. From the results, it is
motion at different amplitudes. Based on the signal in Fig. 5, six training clearly seen that the prediction is improving as more complexity is
cases at relative amplitudes of A = 0.150, 0.100, 0.050, 0.020, 0.005, added to the model. In addition, such nonlinear and parameter-varying
and 0.001 are simulated. The outputs are normalized by the relative dynamics can also be learnt from deep learning, resulting in a more
amplitude, which are compared in Fig. 6. Results at A = 0.001 for small accurate aerodynamic ROM [139].
disturbances are used as the benchmark case for purely linear dynamics.

Fig. 5. FWGN signal α(t) and power spectral density plot [121]. (a) Pitching motion history of α(t) (A = 1). (b) Power spectrum of training case.

18
J. Kou and W. Zhang Progress in Aerospace Sciences 125 (2021) 100725

Fig. 6. Comparison of nonlinear characteristics of lift coefficient for a pitching NACA0012 airfoil [121]. (a) Comparisons of normalized Cl. (b) Deviation of
normalized Cl with normalized Cl of A = 0.001 (linear aerodynamics).

Fig. 7. Moment coefficient predictions for a pitching NACA0012 airfoil at Mach number 0.8. PSO-Training and PSO-Validation refer to two ROMs trained without
and with the validation case [121]. (a) A = 0.100 (b) A = 0.050. (c) A = 0.020. (d) A = 0.005.

5.2. Transonic aeroelastic simulation Fig. 11, LCO responses at V * = 0.67 from the hybrid ROM are shown,
while the layered ROM fails to predict this phenomenon. The hybrid
The second case is the application of system-identification-based ROM captures the initial development of the beat phenomenon very
nonlinear unsteady aerodynamic ROMs to aeroelastic simulation in precisely, while the error in amplitude and phase gradually grows after
transonic flow. Aeroelastic analysis is an important research topic for physical time 130. These predictions demonstrate the potential of the
data-driven unsteady aerodynamic modeling. Coupling the aero­ hybrid ROM to simulate complex aeroelastic phenomena.
dynamic ROM with structural equations of motion, fast linear FSI In addition, the capability of the hybrid ROM in predicting more
analysis and nonlinear aeroelastic simulation can be achieved. A typical complicated aeroelastic phenomena is also validated by reproducing the
aeroelastic model with two degrees of freedom is shown in Fig. 9. For nodal-shaped oscillation (divergent-damping-divergent-damping oscil­
ROM-based FSI analysis in transonic flow, the reader is referred to a lation). This phenomenon is induced by the interaction between buffet
review of Gao and Zhang [11]. Here nonlinear aeroelastic simulation is and flutter [419], which is observed for the BACT model at Mach
performed for the BACT wing (the NACA0012 airfoil) in transonic flow number 0.71 and reduced velocity 0.65 where flutter happens. An
[415], based on the layered and hybrid ROMs illustrated in Fig. 4. explanation of this phenomenon is that when the pitching angle of the
The LCO response at Mach number 0.8 and reduced velocity V* = airfoil experiencing flutter exceeds the buffet onset angle of attack (4.8◦
√̅̅̅
0.61 is shown in Fig. 10, where V * = V∞ /(ωα ⋅b ⋅ μ), ωα and μ are the in this case), the high-frequency aerodynamic loads from the buffeting
uncoupled natural frequency of pitching motion and mass ratio. As flow will destroy the flutter mode, making the amplitude decrease.
mentioned in Ref. [320], the hybrid ROM captures the flow physics However, as the angle of attack decays to values less than the buffet
more accurately, therefore it reproduces the LCO responses with a onset angle, flutter will happen again and the structural motion con­
higher level of accuracy. Since the physical frequency predicted by the tinues to diverge. As time evolves, these behaviors will repeat, leading to
layered ROM is larger than that of the CFD-based aeroelastic analysis, a the nodal-shaped oscillation.
larger phase difference from this model is seen. The predicted amplitude To model this phenomenon, the reduced-order modeling approach
of structure oscillations for the layered ROM is also lower than that of faces some new challenges. Due to the buffet condition, the flow
the hybrid ROM which is more accurate. damping changing with angle of attack should be accurately captured.
As reduced velocity V * increases to larger values like 0.65, structural The ROM should also capture both the linear flutter behavior under
responses show an evident beat phenomenon. This means that there are small disturbances and nonlinear aerodynamics at large angles of attack.
two or more dominant frequencies competing and the maximal ampli­ The hybrid ROM is designed to overcome these challenges by incorpo­
tudes of the resulting responses change periodically [416–418]. In rating both linear and nonlinear subsystems. The predicted nodal-

19
J. Kou and W. Zhang Progress in Aerospace Sciences 125 (2021) 100725

Fig. 8. Moment coefficients of a harmonic pitching motion for a NACA64A010 airfoil at three different Mach numbers 0.7619, 0.7805, 0.8 (reduced frequency k =
0.1636, amplitude αmax = 0.05 rad). (a) ROM 1 (Single-kernel RRBF network). (b) ROM 2 (Symmetric multi-kernel neural network). (c) ROM 3 (Asymmetric multi-
kernel neural network) [112].

frequency aerodynamics at large structural amplitudes, as shown in

Fig. 12c and d. To further investigate the model performance in
capturing the transonic buffet phenomenon, temporal evolutions of
pitching motion and lift coefficient in a typical period of the nodal-
shaped oscillation are compared in Fig. 13. The formation of the high-
frequency aerodynamic loads is marked in blue circles. In Fig. 13a,
when the angle of attack exceeds the buffet onset angle 4.8◦ , the high-
frequency aerodynamic loads are imposed to the airfoil, which agrees
well with the results from Zhang et al. [419]. However, for the
Fig. 9. A typical aeroelastic model with pitching (α) and plunging (h) degrees ROM-based aeroelastic simulation, buffet loads are observed when the
of freedom [320]. K refers to the spring stiffness. structure amplitude is about 4.4◦ , which means the buffet onset angle
predicted from the hybrid ROM is slightly lower. As the pitching angle
shaped oscillations, obtained by coupling the structural motion equa­ increases, both models capture the buffet loads with high frequency,
tions with the hybrid ROM or the CFD solver, are shown in Fig. 12. The even though some error is noticed at the peak values. Although there is
overall accuracy of the hybrid ROM is acceptable. However, due to the still work to be done to improve the prediction capability of ROMs,
complexity of flow physics in this type of oscillation, there are still some current results still show the potential of data-driven models in pre­
errors in the prediction of amplitudes and phases, which can be further dicting complicated aeroelastic responses.
improved in the future. Although larger errors are noticed in the aero­ Recently, LSTM has become a promising model for nonlinear aero­
dynamic coefficients, the hybrid ROM can still imitate the high- elastic simulation across multiple Mach numbers. In Li et al. [139],

Fig. 10. LCO temporal responses for BACT model at Mach number 0.8 and V * = 0.61 [320]. (a) Plunging motion. (b) Pitching motion.

20
J. Kou and W. Zhang Progress in Aerospace Sciences 125 (2021) 100725

Fig. 11. LCO temporal responses of beat phenomenon for BACT model at Mach number 0.8 and V * = 0.67 [320]. (a) Plunging motion. (b) Pitching motion.

Fig. 12. Nodal-shaped oscillation of the structural motion and the aerodynamic coefficient for BACT model [320]. (a) Plunging motion. (b) Pitching motion. (c) Lift
coefficient. (d) Moment coefficient.

Fig. 13. Temporal evolution of pitching motion and lift coefficient in a typical period of the nodal-shaped oscillation for BACT model [320]. (a) CFD-based
aeroelastic simulation. (b) ROM-based aeroelastic simulation.

LSTM is used to model the nonlinear unsteady aerodynamics of a about 47, accompanied by the periodic von Kármán vortex shedding
NACA64A010 airfoil at different Mach numbers varying from 0.74 to phenomenon [420]. However, when the cylinder is elastically supported
0.82. The resulting model is coupled with the structural motion equa­ [421,422] or undergoing forced vibrations [423,424], the periodic
tions for LCO prediction. LCO trends computed by CFD-based and vortex shedding occurs at a subcritical Reynolds number much lower
ROM-based simulations at Mach numbers 0.77, 0.78, 0.79, 0.80 and than Recr. An example of this unstable phenomenon is the
0.81 are compared in Fig. 14. A good match between the two methods is vortex-induced vibration (VIV) in subcritical flow regime [67,421,422,
seen, while errors exist near the flutter boundaries. Compared with the 425,426], where it is shown that VIV can occur at a Reynolds number of
shallow neural network [113], the accuracy of this ROM has been about 20. This mechanism can be easily elaborated through linear sta­
greatly improved. As Mach number decreases, the LCO amplitude be­ bility analysis based on ROM. Zhang et al. [67] performed a ROM-based
comes larger, indicating a weaker aerodynamic nonlinearity. Such trend FSI analysis by the ARX model, where it is demonstrated that in
is also captured accurately by the LSTM-based ROM. These results subcritical flow regimes, the instability of the VIV is due to the coupled
indicate that deep learning is well suited in modeling nonlinear un­ system being linearly unstable to the structure mode [427]. After con­
steady aerodynamics, and has great potential in developing aero­ structing several FSI models at different Reynolds numbers, instability
dynamic models with improved generalization capability and boundaries for two typical mass ratios in the reduced velocity versus Re
robustness to different operating conditions. plane are shown in Fig. 15. It can be seen that the lower instability
boundary is about Re = 18 for both cases. However, the reason why the
FSI instability does not occur at Re < 18 remains unknown. This will be
5.3. Modal analysis for flow past a cylinder explained from the perspective of flow mode, based on the flow mode
testing approach in Kou et al. [428].
The third case is to apply the feature extraction method, in particular The flow mode testing approach is based on an impulse excitation
DMD, to analyze the physics of two-dimensional flow past a cylinder at and DMD. At subcritical Reynolds numbers Re < Recr , when a small-
low Reynolds numbers. It is well known that the flow around a sta­ amplitude impulse in the transverse direction is imposed to the cylin­
tionary cylinder becomes unstable at critical Reynolds number Recr

21
J. Kou and W. Zhang Progress in Aerospace Sciences 125 (2021) 100725

Fig. 14. LCO trends for a NACA64A010 airfoil at different Mach numbers. (a) Plunging amplitude versus LCO velocity. (b) Pitching amplitude versus LCO velocity.
(c) Plunging amplitude versus LCO frequency. (d) Pitching amplitude versus LCO frequency [139].

Fig. 15. Instability boundaries of VIV for two specific mass ratios analyzed by the ROM-based FSI model and the results of Mittal and Singh [422] calculated by
direct numerical simulations. (a) Mass ratio m* = 4.73 (b) Mass ratio m* = 50 [428].

der, the flow will recover to the steady state with its intrinsic frequency 18, this mode contour is nearly constant near the wake. In Fig. 16c, as Re
and decaying rate [424]. Snapshots describing the flow decaying are increases to 18, the anti-symmetrical vortex shedding mode appears and
collected for DMD analysis, and the dominant mode is selected as the the pressure difference is observed from the top and bottom surfaces of
mode corresponding to the intrinsic flow damping and frequency. the cylinder. This pair of vortex, which is near to the upper and lower
Strouhal number St = fs D / V∞ (fs is the vortex shedding frequency and cylinder surfaces, develops as Re increases, as illustrated in Fig. 16b–f.
D is the cylinder diameter) is used to quantify the frequency. Flow When Re = 45, this vortex pair is distinct, characterizing the closeness to
damping and frequency are obtained by fitting the temporal responses of the onset of the Hopf bifurcation point at Re = Recr . The vortex pair
the lift or the moment coefficient. For the consistency of frequency and continues to move back from the cylinder as the flow evolves to the
damping coefficient, the characteristic frequency for a supercritical supercritical unstable equilibrium state. The complete separation of this
Reynolds number, i.e., Re > Recr , is the frequency of unstable equilib­ vortex pair can be noticed in Fig. 16g for Re = 55, which describes the
rium state, rather than the saturated limit-cycle oscillation (LCO) state. natural flow instability. The fully developed unstable flow mode (LCO
In the LCO regime, the frequency is larger than that of unstable equi­ state) at this Reynolds number is shown in Fig. 16h, whose near-wall
librium [237]. Note that for supercritical Reynolds numbers, the impulse vortex pair shows a larger magnitude. This DMD mode is unstable and
excitation is not necessary, since the flow will develop linearly starting it characterizes the periodic von Kármán vortex street shedding.
from the unstable steady state. Although unsteady vortex shedding does not happen for purely fluid
Pressure modes captured by DMD at Reynolds numbers 12, 16, 18, systems at a subcritical Re, i.e., Re < Recr , the von Kármán mode
30, 35, 45 and 55 are shown in Fig. 16, where all the modes are scaled to pattern can still be observed. This indicates that the von Kármán mode
the same magnitude. Since the flow is unstable for Re = 55, we provided still exists in supercritical regimes. However, from the growth rate ob­
the DMD modes at both unstable equilibrium and LCO states in Fig. 16g tained by DMD, this mode is stable, therefore it will not induce vortex
and h, respectively. As shown in Fig. 16a and b, when Re is smaller than shedding for flow past a static cylinder. For FSI systems, this von Kármán

22
J. Kou and W. Zhang Progress in Aerospace Sciences 125 (2021) 100725

Fig. 16. Von Kármán modes obtained from DMD at different flow conditions for flow past a cylinder. (a) Re = 12. (b) Re = 16. (c) Re = 18. (d) Re = 30. (e) Re = 35.
(f) Re = 45. (g) Re = 55 (unstable equilibrium state) (h) Re = 55 (LCO state) [428].

mode corresponds to eigenvalues almost identical to the eigenvalues of These results and analysis explain why the VIV in subcritical flow re­
the purely fluid system with a stationary structure [421]. The coupling gimes happens and why the lowest Reynolds number for VIV is 18.
between this stable mode and the structure mode, however, will induce So far it has been shown that DMD can be used to extract the
FSI instabilities, making the structural branch of eigenvalues unstable dominant flow mode at subcritical Reynolds numbers. Now an example
within a particular frequency range [67,427]. Moreover, this von of DMD for supercritical flows is briefly shown. The flow past a cylinder
Kármán mode disappears exactly at Re = 18, which is also the lowest Re at Reynolds number 60 is studied [235]. The evolution of the absolute
of the FSI instability. Since such von Kármán mode shows the difference value of lift coefficient and the Strouhal number, starting from the un­
of pressure in the transverse direction, it will contribute to the transverse stable steady based flow, is shown Fig. 17. Three typical states can be
aerodynamic forces. When this mode pattern disappears at Re ≤ 18, the identified, including the unstable equilibrium state, the transient state
generation of the transverse aerodynamic forces is very difficult, thus and the periodic limit cycle state. We sample the linear regime (the
the coupled FSI instability will not happen. unstable equilibrium state) and compute the dominant flow modes
From the above-mentioned analysis, it is revealed that the stable von based on DMD. This is the linear exponential growth process and four
Kármán mode arises from Reynolds number 18 and it becomes unstable dominant modes are captured from the criterion in Ref. [235], as shown
at Recr . We can then draw two conclusions: 1) Even though the von in Fig. 18. Note that the dynamic mode with non-zero frequency con­
Kármán mode is stable for subcritical flows at 18 < Re ≤ Recr , VIV still tains two conjugate modes with a phase shift in the imaginary part.
can be triggered due to the coupling between this stable von Kármán Mode 1 is the mean flow mode which represents the mean value of the
mode and the elastic structure. 2) The disappearance of this von Kármán dynamic process. Mode 2 is the main vortex shedding mode which has
mode pattern, i.e., Re ≤ 18, makes the fluid system less possible to the same frequency as the oscillation of the lift coefficient. It is an un­
interact with the elastic structure, therefore VIV will not be induced. stable mode which leads to the exponential growth of the equilibrium

23
J. Kou and W. Zhang Progress in Aerospace Sciences 125 (2021) 100725

Fig. 17. Amplitude and frequency characteristics of lift coefficient of flow past a cylinder at Re = 60 [235]. (a) Absolute value of lift coefficient changes with time.
(b) Instantaneous Strouhal number.

Fig. 18. Dominant DMD modes from DMD with criterion technique of a cylinder at the equilibrium state of Re = 60, where the DMD modes are the same as standard
method and only the selected dominant DMD modes are different [235]. (a) Mode 1. (b) Mode 2. (c) Mode 3. (d) Mode 5.

state. Mode 3 is the static shift mode, which indicates the nonlinearity in the first one; thus, in total, seven modes are selected) in the present
the flow that leads to the change in mean flow [165]. Mode 5 is the drag study. All Ritz eigenvalues and the seven selected dominant eigenvalues
mode, which doubles the frequency of the first oscillatory mode. This are shown in Fig. 19. The dashed line represents the unit circle, which
mode has a symmetric pattern which influences the drag force in the indicates the mode stability. We notice that the eigenvalues are located
streamwise flow direction. Note that this mode has a small mode in the vicinity of this circle, indicating these modes have nearly zero
amplitude, therefore needs to be selected by the improved mode selec­ growth rate thus are nearly periodic. This is consistent with the fact that
tion criterion [235]. snapshots are taken from the fully developed limit cycle state. The first

5.4. Modal analysis for transonic buffet

The fourth case focuses on applying DMD to analyze the transonic

buffet flows. Transonic buffet is a typical flow instability phenomenon,
with periodic oscillation of shock waves over the wing surface [429,
430]. It is similar to a cylinder flow at low Reynolds numbers since they
are both flow instabilities that develop from an unstable steady base
flow to a periodic limit cycle oscillation state. From an engineering point
of view, periodic loads due to buffet may cause structural fatigue or even
flight accidents. Therefore, it is of great importance to study the
mechanism of buffet and understand the nonlinear interactions between
the shock wave and the boundary layer. Because transonic buffet ex­
hibits periodic features, using DMD will help to capture the dominant
frequency information and stability characteristics, thus allowing
further investigations in characteristic analysis or buffet suppression.
The first example is the DMD analysis of a NACA0012 airfoil in
transonic buffet flow [70]. The flow state is set to Mach number 0.70,
angle of attack 5.5◦ , and Reynolds number 3 × 106 . From the power
spectrum density for the lift coefficient, the buffet reduced frequency is
0.2. 300 pressure snapshots in the limit cycle state are recorded as the
sampling dataset for the DMD analysis. From previous studies [154,
235], it is sufficiently accurate to illustrate the dynamics of an unstable
periodic flow using no more than 10 dominant modes. Therefore, we Fig. 19. The Ritz eigenvalues for a transonic buffet flow in the limit cycle state.
select the first four dominant modes (all are conjugate modes except for Four dominant global modes among all modes by DMD are shown [70].

24
J. Kou and W. Zhang Progress in Aerospace Sciences 125 (2021) 100725

four dynamic mode contours are shown in Fig. 20, and the corre­ Table 1
sponding reduced frequencies and growth rates are shown in Table 1. Growth rates and reduced frequencies of the dominant DMD modes [70].
We notice that both the growth rate and the frequency are zero for the Mode Growth Rate Reduced Frequency
first mode. It is a static mode, close to the mean flow field. All of the
1 0 0
other modes reflect the periodic oscillating features resulting from shock 2 3.75 × 10− 6 0.196
waves, therefore they have zero growth rates. Mode 2 is the base dy­ 3 5 0.393
3.86 × 10−
namic mode which has the buffet frequency 0.196, characterizing the
4 1.20 × 10 − 6 0.588
main flow pattern of transonic buffet. Other high-order modes are
high-order harmonic modes where the frequency is double and triple of
the buffet frequency. Note that these modes can also be well predicted
by extrapolating the dynamic modes from snapshots in the transient
flow states [260].
Through DMD analysis for three-dimensional configurations, more
complex dynamic behaviors can be analyzed. Gao and Zhang [431]
studied three-dimensional buffet phenomena based on a rectangular
OAT15A wing model. The two-dimensional airfoil is extended in the
spanwise direction, where different spanwise lengths are considered,
defined by the aspect ratio between the spanwise length and the chord
length. The flow state is set to Mach number 0.73, angle of attack 4◦ and
Reynolds number 3 × 106 . In Fig. 21, the root mean square of the surface
pressure coefficient is compared for two aspect ratios 0.25 (Model 1) and
3.0 (Model 2) respectively. Time evolution of the pressure coefficients
and power spectral analysis at three typical chordwise and spanwise
positions are shown in Fig. 22. The chordwise positions are highlighted
in Fig. 22, and η refers to the spanwise position, defined as the ratio
between the position of wing section and the spanwise length of the
wing. For Model 1, the shock wave oscillation is mainly determined by
the flow dynamics in the chordwise direction, as evidenced by very
similar pressure evolutions across different spanwise positions. In
addition, the frequency of the pressure oscillation is the same as the
frequency of the shock wave oscillation in the chordwise direction.
However, dynamic behavior in the spanwise direction is evident for
Model 2, which has a longer spanwise length. The chordwise charac­
teristic is dominant near the shock wave, while the spanwise charac­
teristic can be clearly seen near the separation region behind the shock
wave position. The spanwise characteristic is shown by the evolution of
pressure coefficient, where a dominant phase lag exists for chordwise
sections x/c = 40% and x/c = 90%. Fig. 21. Root mean square distribution of surface pressure coefficient of the
In addition, the first four DMD pressure modes for two wing models OAT15A wing model [431]. (a) Aspect ratio 0.25. (b) Aspect ratio 3.0.
are captured and shown in Fig. 23 and Fig. 24. When the aspect ratio is

Fig. 20. The first four DMD modes for a transonic buffet flow in the limit cycle state [70]: (a) mode 1. (b) mode 2. (c) mode 3. (d) mode 4.

25
J. Kou and W. Zhang Progress in Aerospace Sciences 125 (2021) 100725

(a) (b)
Fig. 22. Temporal evolution and power spectrum density of pressure coefficient at typical sample positions of the OAT15A wing model [431]. (a) Aspect ratio 0.25.
(b) Aspect ratio 3.0.

0.25, the pressure distribution is nearly parallel in the spanwise direc­ 5.5. Multi-fidelity modeling of a NACA0012 airfoil
tion, indicating that the buffet is dominated by the instability in the
chordwise direction. The chordwise instability is characterized by the The last case is the multi-fidelity modeling for nonlinear unsteady
second to the forth mode. The second mode is the dominant one oscil­ aerodynamics by fusing low-fidelity and high-fidelity input-output data,
lating with the buffet frequency. The third and fourth modes are both which has been reported by Kou and Zhang [308]. We consider the
double-frequency modes which reflect the higher-order instability fea­ NACA0012 airfoil pitching in transonic flow. The selected Mach number
tures. For Model 2 with aspect ratio 3.0, a spanwise instability is seen is 0.8, and the Reynolds number for the viscous flow is 3 × 106 . The
from the mode patterns. Although Mode 2 is still dominated by the pitching axis is set to 0.25c, where c is the chord length. The low-fidelity
chordwise instability, the third and fourth modes exhibit oscillatory flow aerodynamic loads are calculated from the Euler solver for inviscid
patterns in the spanwise direction, which describe the spanwise buffet flows, while the high-fidelity aerodynamics are obtained from the
phenomenon. These results indicate that for Model 2, the buffet insta­ URANS solver considering the viscous effects.
bility is more complicated, since it is induced by the interaction between To show that the unsteady multi-fidelity modeling method can be
the chordwise and spanwise modes. This also explains the phase lag of extended to experimental data, a more realistic training dataset is
the pressure in the spanwise direction, as illustrated in Fig. 22b. considered. These data are obtained from harmonic motions, rather than
Combining the studies for different aspect ratios, it is found that the the random training case which is difficult to obtain in experimental
spanwise length is the main factor that determines the buffet in the settings. Here we use limited high-fidelity data from only three types of
spanwise direction. As the spanwise length increases, the buffet flow harmonic motion, where the mean angle of attack is zero, and the
characteristic becomes more evident, which interacts with the chord­ amplitude of harmonic motion is defined as αmax . The sampling space
wise buffet and produces a more complicated buffet response. This contains two variables of interest, i.e., amplitude and reduced frequency
conclusion is consistent with the numerical simulation of Iovnovich and of harmonic motion, as shown in Fig. 25. The parameter range is large
Raveh [432]. enough to cover a range of dynamic behaviors in transonic flow. Three
training cases are chosen from the Latin hypercube method, which
covers different dynamic behaviors. Note that both low-fidelity and

Fig. 23. First four DMD modes for the OAT15A wing model with aspect ratio 0.25 [431].

26
J. Kou and W. Zhang Progress in Aerospace Sciences 125 (2021) 100725

Fig. 24. First four DMD modes for the OAT15A wing model with aspect ratio 3 [431].

periods are simulated, therefore each point in Fig. 25 corresponds to

hundreds of data pairs, depending on the reduced frequency and the
number of cycles. Finally, three test cases have generated 3680 data
pairs in total, which are all used for model training.
We have generated two models according to Eq. (32) and Eq. (33),
and the multi-kernel neural network [112] is used to identify the
correction term. The input and output delay orders m and n are set to m
= n = 6. The number of hidden neurons in the multi-kernel neural
network is set to 74. Predicted results of Training Case 3 and Test Case 3
are compared in Fig. 26 and Fig. 27, respectively. Results indicate that
both multi-fidelity models improve the low-fidelity results to a large
extent. It should be noted that when unsteady aerodynamics shows a
dynamically linear behavior, the shape of the hysteresis loop is nearly
elliptical. Therefore, it is observed that the aerodynamic nonlinearity is
strong for Training Case 3, where a big difference is observed between
low-fidelity and high-fidelity data. This is in accordance with previous
studies that nonlinear aerodynamics is dominant at high amplitude and
low reduced frequency [121]. If a low-fidelity model is used, such
nonlinear features cannot be captured accurately. Not surprisingly, the
proposed multi-fidelity models are accurate for all training cases, lead­
ing to at least 70% reduction of root mean squared error compared with
Fig. 25. Sampling space and the selected training and test cases for multi-
fidelity modeling of a NACA0012 airfoil [308]. a single low-fidelity model [308]. For Test Case 3, since this case is not
encountered for model training, its accuracy will be a bit worse than the
prediction for training cases, which is still reasonable. Compared with
high-fidelity data for the same type of motion are needed for model
the low-fidelity results, the prediction can still be improved a lot by the
training. Parameters (αmax , kf) for three training cases are (2.4930,
multi-fidelity models. Note that due to the stronger nonlinear effect of
0.2607), (1.2629, 0.1459) and (2.8648, 0.0654), while those for test
the aerodynamic moment, the prediction of the moment coefficient is
cases are (2.8648, 0.1636), (1.4324, 0.2045), (1.1459, 0.0818), and
not as good as that of the lift coefficient. The root mean squared errors
(2.2918, 0.0545). For each training case, harmonic motions over fifteen

Fig. 26. Prediction of aerodynamic responses of Training Case 3 (αmax = 2.8648◦ , k = 0.0654) for a NACA0012 airfoil [308]. (a) Lift coefficient. (b)
Moment coefficient.

27
J. Kou and W. Zhang Progress in Aerospace Sciences 125 (2021) 100725

Fig. 27. Fig. 26 Prediction of aerodynamic responses of Test Case 3 (αmax = 1.1459◦ , k = 0.0818) for a NACA0012 airfoil [308]. (a) Lift coefficient. (b)
Moment coefficient.

for lift and moment coefficients, based on low-fidelity method, Model 1 the proposed multi-fidelity models only have about 20% of the error
and Model 2, are (2.5343e-2, 6.2344e-3), (4.7450e-3, 1.4922e-3), and from low-fidelity models, which are only trained from three harmonic
(5.9490e-3, 1.0080e-3), respectively. Therefore, for this particular case, motions.

Fig. 28. Prediction of aerodynamic responses of two test cases (random motion) for a NACA0012 airfoil [308]. Pitching motion of (a) Test Case 1 and (b) Test Case 2.
Power spectrum of (c) Test Case 1 and (d) Test Case 2. Lift coefficient versus time of (e) Test Case 1 and (f) Test Case 2. Moment coefficient versus time of (g) Test
Case 1 and (h) Test Case 2.

28
J. Kou and W. Zhang Progress in Aerospace Sciences 125 (2021) 100725

Since previous test cases only contain some typical harmonic mo­ Declaration of competing interest
tions, which are not enough to validate the model performance, we then
provide another two test cases of the airfoil experiencing random mo­ The authors declare that they have no known competing financial
tions. These motions are generated from FWGN signals. Both small- interests or personal relationships that could have appeared to influence
amplitude and large-amplitude motions are considered. The predicted the work reported in this paper.
aerodynamic responses are given in Fig. 28. The power spectral analysis
based on Welch method is used to identify the frequency ranges of these Acknowledgements
signals. As shown in the power spectral density diagram in Fig. 28c and
d, the reduced frequency that contains the main dynamic information This work was supported by the National Natural Science Foundation
ranges from 0 to 0.25, which lies within the parameter space. As seen of China (Grant No. 12072282 and No. 91852115), National Numerical
from the figure, the proposed models are still able to predict the high- Wind-Tunnel (Grant No. NNW2018-ZT1B01 and No. NNW2019ZT2-
fidelity data based on low-fidelity results, even though the considered A05). The authors also wish to thank anonymous reviewers of the pre­
motion has multiple amplitudes and frequencies. Besides, the proposed sent paper, who have shared essential suggestions and comments to
models give similar prediction of both linear and nonlinear aero­ improve its contents. The first author would like to dedicate this work to
dynamics, which may overcome the limitation that traditional nonlinear Mr. Bingqian Kou, for his invaluable support in the preparation and
ROMs have difficulty in capturing linear aerodynamics accurately writing of the manuscript.
[121]. However, compared with Model 2, Model 1 predicts the moment
coefficient with larger error. This is because Model 2 separates the un­ References
steady and quasi-steady terms, which may result in better accuracy. But
from all the test cases, both models exhibit reasonable performance. In [1] D.J. Lucia, P.S. Beran, W.A. Silva, Reduced-order modeling: new approaches for
computational physics, Prog. Aero. Sci. 40 (1–2) (2004) 51–117.
summary, these test cases show the efficacy of multi-fidelity modeling [2] M. Ghoreyshi, A. Jirásek, R.M. Cummings, Reduced order unsteady aerodynamic
methods for unsteady aerodynamics. For detailed comparison with modeling for stability and control analysis using computational fluid dynamics,
ROMs built only from high-fidelity data, please refer to Ref. [308]. These Prog. Aero. Sci. 71 (2014) 167–217.
[3] K. Duraisamy, G. Iaccarino, H. Xiao, Turbulence modeling in the age of data,
results indicate that it is possible to use data from multiple sources to Annu. Rev. Fluid Mech. 51 (2019) 357–377.
improve the prediction of high-fidelity unsteady aerodynamics. [4] W. Zhang, B. Wang, Z. Ye, J. Quan, Efficient method for limit cycle flutter
analysis based on nonlinear aerodynamic reduced-order models, AIAA J. 50 (5)
(2012) 1019–1028.
6. Conclusions [5] E.H. Dowell, K.C. Hall, Modeling of fluid-structure interaction, Annu. Rev. Fluid
Mech. 33 (1) (2001) 445–490.
This paper details popular data-driven methods that have been [6] J.R.R.A. Martins, A.B. Lambe, Multidisciplinary design optimization: a survey of
architectures, AIAA J. 51 (9) (2013) 2049–2075.
developed to model unsteady aerodynamics and aeroelasticity. Signifi­
[7] X. Wu, W. Zhang, X. Peng, Z. Wang, Benchmark aerodynamic shape optimization
cant progress has been made by data-driven methods, in order to better with the pod-based cst airfoil parametric method, Aero. Sci. Technol. 84 (2019)
understand the flow physics, and derive low-dimensional models that 632–640.
are useful for multiphysics simulation, optimization and control. [8] M.J. Zahr, Adaptive Model Reduction to Accelerate Optimization Problems
Governed by Partial Differential Equations, Stanford University, 2016.
Recently, with the development of machine learning and data science, [9] S.L. Brunton, B.R. Noack, Closed-loop turbulence control: progress and
data-driven modeling will continue to play an important role in related challenges, Appl. Mech. Rev. 67 (5) (2015), 050801.
research areas. Main conclusions are drawn as follows: [10] F. Ren, H. Hu, H. Tang, Active flow control using machine learning: a brief
review, J. Hydrodyn. 32 (2) (2020) 247–253.
[11] C. Gao, W. Zhang, Transonic aeroelasticity: a new perspective from the fluid
1) Research in system identification and feature extraction has been mode, Prog. Aero. Sci. 113 (2020) 100596.
conducted for years, with successful applications in modeling linear [12] A. Da Ronch, M. Ghoreyshi, K.J. Badcock, On the generation of flight dynamics
aerodynamic tables by computational fluid dynamics, Prog. Aero. Sci. 47 (8)
dynamic (but nonlinear static) behaviors of the unsteady flow [5]. (2011) 597–620.
However, future works are still needed to develop nonlinear dynamic [13] J.J. McNamara, P.P. Friedmann, Aeroelastic and aerothermoelastic analysis in
reduced-order models robust to large-amplitude motion, massive hypersonic flow: past, present, and future, AIAA J. 49 (6) (2011) 1089–1122.
[14] K. Taira, S.L. Brunton, S.T.M. Dawson, C.W. Rowley, T. Colonius, B.J. McKeon, O.
flow separation, and different flow conditions. In addition, data T. Schmidt, S. Gordeyev, V. Theofilis, L.S. Ukeiley, Modal analysis of fluid flows:
fusion remains a new and active research topic in modeling unsteady an overview, AIAA J. 55 (12) (2017) 4013–4041.
aerodynamics and will gain more attention in the future. This is [15] K. Taira, M.S. Hemati, S.L. Brunton, Y. Sun, K. Duraisamy, S. Bagheri, S.T.
M. Dawson, C.-A. Yeh, Modal analysis of fluid flows: applications and outlook,
because combining experimental measurement and numerical
AIAA J. 58 (3) (2020) 998–1022.
simulation for accurate flow prediction will be a general trend in the [16] C.W. Rowley, S.T.M. Dawson, Model reduction for flow analysis and control,
near future, which balances the overall cost and resources and pro­ Annu. Rev. Fluid Mech. 49 (2017) 387–417.
vides improved accuracy. [17] S.L. Brunton, B.R. Noack, P. Koumoutsakos, Machine learning for fluid
mechanics, Annu. Rev. Fluid Mech. 52 (2020) 477–508.
2) Incorporating flow physics in data-driven modeling remains an op­ [18] S. Pandey, J. Schumacher, K.R. Sreenivasan, A perspective on machine learning in
portunity and an open question. Firstly, augmenting physics will add turbulent flows, J. Turbul. 21 (9–10) (2020) 567–584.
more interpretability with accessible physical insights to data-driven [19] J. Jiménez, Computers and turbulence, Eur. J. Mech. B Fluid 79 (2020) 1–11.
[20] G. Chen, C. Zou, C. Yang, Aeroelastic Design Basis, Beijing University of
models. Secondly, more physics will reduce the dependence on data, Aeronautics and Astronautics Press, Beijing, China, 2004.
making it possible to achieve few-shot learning that constructs [21] Z. Ye, W. Zhang, A. Shi, Fundamentals of Fluid-Structire Coupling and its
models only from a fraction of the flow data. Application, Harbin Institute of Technology Press, Harbin, China, 2010.
[22] T. Theodorsen, General Theory of Aerodynamic Instability and the Mechanism of
3) Quantifying the reliability and uncertainty for data-driven models Flutter, Technical Report, 1935.
still needs more effort. Although the performance of data-driven [23] H. Wagner, Über die entstehung des dynamischen auftriebes von tragflügeln,
models relies largely on the training data, it is still important to Z. Angew. Math. Mech. 5 (1) (1925) 17–35.
[24] C.T. Tran, D. Petot, Semi-empirical model for the dynamic stall of airfoils in view
have theories or methods that can evaluate the applicability of these of the application to the calculation of responses of a helicopter in forward flight,
models, especially for nonlinear unsteady aerodynamics. Vertica 5 (1) (1981) 35–53.
4) Applications of data-driven aerodynamic modeling with increasing [25] J.G. Leishman, T.S. Beddoes, A semi-empirical model for dynamic stall, J. Am.
Helicopter Soc. 34 (1989) 3–17.
complexity, including complex geometries and more realistic prob­
[26] J.G. Holierhoek, J.B. de Vaal, A.H. van Zuijlen, H. Bijl, Comparing different
lems, need to be explored. Data-driven models that handle complex dynamic stall models, Wind Energy 16 (1) (2013) 139–158.
and three-dimensional configurations, as well as highly nonlinear, [27] R.T. Hartlen, I. Currie, Lift-oscillator model of vortex-induced vibration, J. Eng.
stochastic and multiscale turbulent flows, can be considered. Mech. Div. 96 (5) (1970) 577–591.

29
J. Kou and W. Zhang Progress in Aerospace Sciences 125 (2021) 100725

[28] M.L. Facchinetti, E. de Langre, F. Biolley, Coupling of structure and wake [63] Z. Wang, W. Zhang, X. Wu, K. Chen, A novel unsteady aerodynamic reduced-
oscillators in vortex-induced vibrations, J. Fluid Struct. 19 (2) (2004) 123–140. order modeling method for transonic aeroelastic optimization, J. Fluid Struct. 82
[29] E.H. Dowell, A Modern Course in Aeroelasticity, Springer, 2015. (2018) 308–328.
[30] J. Murua, R. Palacios, J.M.R. Graham, Applications of the unsteady vortex-lattice [64] D. Su, W. Zhang, Z. Ye, A reduced order model for uncoupled and coupled
method in aircraft aeroelasticity and flight dynamics, Prog. Aero. Sci. 55 (2012) cascade flutter analysis, J. Fluid Struct. 61 (2016) 410–430.
46–72. [65] C. Gao, W. Zhang, Z. Ye, A new viewpoint on the mechanism of transonic single-
[31] C. Xie, L. Wang, C. Yang, Y. Liu, Static aeroelastic analysis of very flexible wings degree-of-freedom flutter, Aero. Sci. Technol. 52 (2016) 144–156.
based on non-planar vortex lattice method, Chin. J. Aeronaut. 26 (3) (2013) [66] X. Li, Y. Liu, J. Kou, W. Zhang, Reduced-order thrust modeling for an efficiently
514–521. flapping airfoil using system identification method, J. Fluid Struct. 69 (2017)
[32] E.M. Murman, J.D. Cole, Calculation of plane steady transonic flows, AIAA J. 9 137–153.
(1) (1971) 114–121. [67] W. Zhang, X. Li, Z. Ye, Y. Jiang, Mechanism of frequency lock-in in vortex-
[33] P. Moin, K. Mahesh, Direct numerical simulation: a tool in turbulence research, induced vibrations at low Reynolds numbers, J. Fluid Mech. 783 (2015) 72–102.
Annu. Rev. Fluid Mech. 30 (1998) 539–578. [68] X. Li, Z. Lyu, J. Kou, W. Zhang, Mode competition in galloping of a square
[34] M. Leiseur, O. Métais, New trends in large-eddy simulations of turbulence, Annu. cylinder at low Reynolds number, J. Fluid Mech. 867 (2019) 516–555.
Rev. Fluid Mech. 28 (1996) 45–82. [69] C. Gao, W. Zhang, X. Li, Y. Liu, J. Quan, Z. Ye, Y. Jiang, Mechanism of frequency
[35] C.G. Speziale, Analytical methods for the development of Reynolds-stress closures lock-in in transonic buffeting flow, J. Fluid Mech. 818 (2017) 528–561.
in turbulence, Annu. Rev. Fluid Mech. 23 (1991) 107–157. [70] C. Gao, W. Zhang, J. Kou, Y. Liu, Z. Ye, Active control of transonic buffet flow,
[36] W. Zhang, Y. Gong, Y. Liu, Time discretization methods in the computation of J. Fluid Mech. 824 (2017) 312–351.
unsteady flow, Adv. Mech. 49 (1) (2019) 201907. [71] S. He, Z. Yang, Y. Gu, Nonlinear dynamics of an aeroelastic airfoil with free-play
[37] K.C. Hall, E.F. Crawley, Calculation of unsteady flows in turbomachinery using in transonic flow, Nonlinear Dynam. 87 (4) (2017) 2099–2125.
the linearized euler equations, AIAA J. 27 (6) (1989) 777–787. [72] S. He, S. Guo, W. Li, D. Yang, Y. Gu, Z. Yang, Nonlinear aeroelastic behavior of an
[38] L. He, W. Ning, Efficient approach for analysis of unsteady viscous flows in airfoil with free-play in transonic flow, Mech. Syst. Signal Process. 138 (2020).
turbomachines, AIAA J. 36 (11) (1998) 2005–2012. [73] J.G. Leshiman, Validation of approximate indicial aerodynamic functions for two-
[39] J.P. Thomas, E.H. Dowell, K.C. Hall, Nonlinear inviscid aerodynamic effects on dimensional subsonic flow, J. Aircraft 25 (10) (1988) 914–922.
transonic divergence, flutter, and limit-cycle oscillations, AIAA J. 40 (4) (2002) [74] M. Tobak, Mathematical Modeling of the Aerodynamic Characteristics in Flight
638–646. Dynamics, NASA Technical Memorandum, 1984.
[40] S. He, Z. Yang, Y. Gu, Transonic limit cycle oscillation analysis using aerodynamic [75] P. Reisenthel, Development of a Nonlinear Indicial Model Using Response
describing functions and superposition principle, AIAA J. 52 (7) (2014) Functions Generated by a Neural Network, 1997.
1393–1403. [76] P. Reisenthel, M. Bettencourt, J. Mayatt, D. Grismer, A Nonlinear Indicial
[41] K.C. Hall, J.P. Thomas, W.S. Clark, Computation of unsteady nonlinear flows in Prediction Tool for Unsteady Aerodynamic Modeling, 1998.
cascades using a harmonic balance technique, AIAA J. 40 (5) (2002) 879–886. [77] M. Ghoreyshi, A. Jirasek, R. Cummings, M. Post, K. Bergeron, R. Decker,
[42] A. Gopinath, A. Jameson, Time spectral method for periodic unsteady A Computional Investigation into the Use of Response Functions for Aerodynamic
computations over two- and three- dimensional bodies, In: 43rd AIAA Aerospace Loads Modeling, 2011.
Sciences Meeting and Exhibit. AIAA 6.2005-1220, Reno, Nevada. [78] M. Ghoreyshi, R.M. Cummings, Unsteady aerodynamic modeling of aircraft
[43] Y. Gong, W. Zhang, Efficient aeroelastic solution based on time-spectral control surfaces by indicial response methods, AIAA J. 52 (12) (2014)
fluid–structure interaction method, AIAA J. 57 (7) (2019) 3014–3025. 2683–2700.
[44] S. He, E. Jonsson, C. A. Mader, J. R. R. A. Martins, Aerodynamic shape [79] M. Ghoreyshi, M.E. Young, A.J. Lofthouse, A. Jirásek, R.M. Cummings, Numerical
optimization with time spectral flutter adjoint, In: AIAA Scitech 2019 Forum. simulation and reduced-order aerodynamic modeling of a lambda wing
AIAA 2019-0697, San Diego, California. configuration, J. Aircraft 55 (2) (2018) 549–570.
[45] L. Ljung, System identification, in: A. Procházka, J. Uhlíř, P.J.W. Rayner, N. [80] A. Da Ronch, A. Ventura, M. Righi, M. Franciolini, M. Berci, D. Kharlamov,
G. Kingsbury (Eds.), Signal Analysis and Prediction, Birhiiuser Boston, 1998, Extension of analytical indicial aerodynamics to generic trapezoidal wings in
pp. 163–173. subsonic flow, Chin. J. Aeronaut. 31 (4) (2018) 617–631.
[46] E. Dowell, J. Edwards, T. Strganac, Nonlinear aeroelasticity, J. Aircraft 40 (5) [81] V. Volterra, Theory of Functionals and of Integral and Integro-Differential
(2003) 857–874. Equations, Dover, New York, 1959.
[47] J.N. Juang, R.S. Pappa, An eigensystem realization algorithm for modal [82] C.M. Cheng, Z.K. Peng, W.M. Zhang, G. Meng, Volterra-series-based nonlinear
parameter identification and model reduction, J. Guid. Contr. Dynam. 8 (5) system modeling and its engineering applications: a state-of-the-art review, Mech.
(1985) 620–627. Syst. Signal Process. 87 (2017) 340–364.
[48] W. Silva, D. E. Raveh, Development of unsteady aerodynamic state-space models [83] W. Silva, Identification of nonlinear aeroelastic systems based on the volterra
from cfd-based pluse responses, In: 19th AIAA Applied Aerodynamics Conference. theory: progress and opportunities, Nonlinear Dynam. 39 (1–2) (2005) 25–62.
AIAA 2001-1213, Anaheim,CA,U.S.A. [84] J. Kou, W. Zhang, Comparisons of Two Reduced Order Models for Linearized
[49] A.L. Gaitonde, D. Jones, Reduced order state-space models from the pulse Unsteady Aerodynamic Identification, arXiv, 2019. arXiv preprint: 1905.03250.
responses of a linearized cfd scheme, Int. J. Numer. Methods Fluid. 42 (6) (2003) [85] B.R. Maner, F.J. Doyle III, Polymerization reactor control using autoregressive-
581–606. plus volterra-based mpc, AIChE J. 43 (7) (1997) 1763–1784.
[50] W.A. Silva, R.E. Bartels, Development of reduced-order models for aeroelastic [86] R. Díaz-Mendoza, H. Budman, Structured singular valued based robust nonlinear
analysis and flutter prediction using the cfl3dv6.0 code, J. Fluid Struct. 19 (6) model predictive controller using volterra series models, J. Process Contr. 20 (5)
(2004) 729–745. (2010) 653–663.
[51] T. Kim, Efficient reduced-order system identification for linear systems with [87] A. Abdelhafiz, A. Kwan, O. Hammi, F.M. Ghannouchi, Digital predistortion of lte-
multiple inputs, AIAA J. 43 (7) (2005) 1455–1464. a power amplifiers using compressed-sampling-based unstructured pruning of
[52] S.L. Brunton, C.W. Rowley, D.R. Williams, Reduced-order unsteady aerodynamic volterra series, IEEE Trans. Microw. Theor. Tech. 62 (11) (2014) 2583–2593.
models at low Reynolds numbers, J. Fluid Mech. 724 (2013) 203–233. [88] M. Balajewicz, E. Dowell, Reduced-order modeling of flutter and limit-cycle
[53] T.L.B. Flinois, A.S. Morgans, Feedback control of unstable flows: a direct oscillations using the sparse volterra series, J. Aircraft 49 (6) (2012) 1803–1812.
modelling approach using the eigensystem realisation algorithm, J. Fluid Mech. [89] W.A. Silva, Application of nonlinear systems theory to transonic unsteady
793 (2016) 41–78. aerodynamic responses, J. Aircraft 30 (5) (1993) 660–668.
[54] W.G. Yao, R.K. Jaiman, Model reduction and mechanism for the vortex-induced [90] D.E. Raveh, Reduced-order models for nonlinear unsteady aerodynamics, AIAA J.
vibrations of bluff bodies, J. Fluid Mech. 827 (2017) 357–393. 39 (8) (2001) 1417–1429.
[55] W.G. Yao, R.K. Jaiman, Feedback control of unstable flow and vortex-induced [91] P. Marzocca, L. Librescu, W. Silva, Aeroelastic response of nonlinear wing
vibration using the eigensystem realization algorithm, J. Fluid Mech. 827 (2017) sections using a functional series technique, AIAA J. 40 (5) (2002) 813–824.
394–414. [92] T. Skujins, C.E.S. Cesnik, Reduced-order modeling of unsteady aerodynamics
[56] J. Sjőberg, Q.H. Zhang, L. Ljung, A. Benveniste, B. Delyon, P.Y. Glorennec, across multiple mach regimes, J. Aircraft 51 (6) (2014) 1681–1704.
H. Hjalmarsson, A. Juditsky, Nonlinear black-box modeling in system [93] N.C.G. de Paula, F.D. Marques, W.A. Silva, Volterra kernels assessment via time-
identification: a unified overview, Automatica 31 (12) (1995) 1691–1724. delay neural networks for nonlinear unsteady aerodynamic loading identification,
[57] T.J. Cowan, A.S. Arena, K.K. Gupta, Accelerating computational fluid dynamics AIAA J. 57 (4) (2019) 1725–1735.
based aeroelastic predictions using system identification, J. Aircraft 38 (1) (2001) [94] R.L. Hardy, Multiquadric equations of topography and other irregular surfaces,
81–87. J. Geophys. Res. 76 (8) (1971) 1905–1915.
[58] W. Zhang, Z. Ye, Reduced-order-model-based flutter analysis at high angle of [95] S.N. Lophaven, H.B. Nielsen, J. Søndergaard, Dace: A Matlab Kriging Toolbox
attack, J. Aircraft 44 (6) (2007) 2086–2089. (Version 2.0), 2002.
[59] W. Zhang, Z. Ye, Control law design for transonic aeroservoelasticity, Aero. Sci. [96] M. Seeger, Gaussian processes for machine learning, Int. J. Neural Syst. 14 (2)
Technol. 11 (2–3) (2007) 136–145. (2004) 69–106.
[60] W. Zhang, Z. Ye, Effect of control surface on airfoil flutter in transonic flow, Acta [97] R. Dwight, Z. Han, Efficient uncertainty quantification using gradient-enhanced
Astronaut. 66 (7–8) (2010) 999–1007. kriging, In: 50th AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics,
[61] W. Zhang, Z. Ye, C.a. Zhang, Aeroservoelastic analysis for transonic missile based and Materials Conference. AIAA 2009-2276.
on computational fluid dynamics, J. Aircraft 46 (6) (2009) 2178–2183. [98] S. Bhattrai, J.H.S. de Baar, A.J. Neely, Efficient uncertainty quantification for a
[62] W. Zhang, K. Chen, Z. Ye, Unsteady aerodynamic reduced-order modeling of an hypersonic trailing-edge flap, using gradient-enhanced kriging, Aero. Sci.
aeroelastic wing using arbitrary mode shapes, J. Fluid Struct. 58 (2015) 254–270. Technol. 80 (2018) 261–268.
[99] S. Jeong, M. Murayama, K. Yamamoto, Efficient optimization design method
using kriging model, J. Aircraft 42 (2) (2005) 413–420.

30
J. Kou and W. Zhang Progress in Aerospace Sciences 125 (2021) 100725

[100] B. Glaz, L. Liu, P.P. Friedmann, Reduced-order nonlinear unsteady aerodynamic [134] J. Kou, W. Zhang, M. Yin, Novel wiener models with a time-delayed nonlinear
modeling using a surrogate-based recurrence framework, AIAA J. 48 (10) (2010) block and their identification, Nonlinear Dynam. 85 (4) (2016) 2389–2404.
2418–2429. [135] J. Schmidhuber, Deep learning in neural networks: an overview, Neural Network.
[101] B. Glaz, P.P. Friedmann, L. Liu, J.G. Cajigas, J. Bain, L.N. Sankar, Reduced-order 61 (2015) 85–117.
dynamic stall modeling with swept flow effects using a surrogate-based [136] S. Hochreiter, J. Schmidhuber, Long short-term memory, Neural Comput. 9 (8)
recurrence framework, AIAA J. 51 (4) (2013) 910–921. (1997) 1735–1780.
[102] H. Liu, H. Hu, Y. Zhao, R. Huang, Efficient reduced-order modeling of unsteady [137] K. Greff, R.K. Srivastava, J. Koutník, B.R. Steunebrink, J. Schmidhuber, Lstm: a
aerodynamics robust to flight parameter variations, J. Fluid Struct. 49 (2014) search space odyssey, IEEE T. Neur. Net. Lear. 28 (10) (2016) 2222–2232.
728–741. [138] R. Pascanu, T. Mikolov, Y. Bengio, On the difficulty of training recurrent neural
[103] X. Li, Z. Guo, G. Dana, Z. Hou, Efficient reduced-order modeling of unsteady networks, in: 30th International Conference on International Conference on
aerodynamics under light dynamic stall conditions, Proc. Inst. Mech. Eng. Part G- Machine Learning, 2013, pp. 1310–1318.
J. Aerosp. Eng. 233 (6) (2019) 2141–2151. [139] K. Li, J. Kou, W. Zhang, Deep neural network for unsteady aerodynamic and
[104] K.R. Brouwer, J.J. McNamara, Surrogate-based aeroelastic loads prediction in the aeroelastic modeling across multiple mach numbers, Nonlinear Dynam. 96 (3)
presence of shock-induced separation, J. Fluid Struct. 93 (2020) 102838. (2019) 2157–2177.
[105] W. Li, D. Jin, Y. Zhao, Efficient nonlinear reduced-order modeling for synthetic- [140] Y. LeCun, Y. Bengio, G. Hinton, Deep learning, Nature 521 (7553) (2015)
jet-based control at high angle of attack, Aero. Sci. Technol. 62 (2017) 98–107. 436–444.
[106] R. P. Patterson, P. P. Friedmann, Application of a cfd-based surrogate approach [141] F. Chollet, Keras, Available from: https://github.com/fchollet/keras, 2015.
for active flow control modeling, In: AIAA SciTech 2019 Forum. AIAA 2019-1706, [142] P.R. Vlachas, W. Byeon, Z.Y. Wan, T.P. Sapsis, P. Koumoutsakos, Data-driven
7-11 January 2019, San Diego, California. forecasting of high-dimensional chaotic systems with long, short-term memory
[107] S. Haykin, Neural Networks: A Comprehensive Foundation, Prentice-Hall, Upper networks 474 (2213) (2018) 20170844.
Saddle River, NJ, 1999. [143] Z.Y. Wan, T.P. Sapsis, Machine learning the kinematics of spherical particles in
[108] H. Yu, T.T. Xie, S. Paszczynski, B.M. Wilamowski, Advantages of radial basis fluid flows, J. Fluid Mech. 857 (2018) R2.
function networks for dynamic system design, IEEE Trans. Ind. Electron. 58 (12) [144] Z.Y. Wan, P.R. Vlachas, P. Koumoutsakos, T. Sapsis, Data-assisted reduced-order
(2011) 5438–5450. modeling of extreme events in complex dynamical systems, PloS One 13 (5)
[109] S. Haykin, Neural Networks and Learning Machines, Prentice Hall, Upper Saddle (2018), e0197704.
River, 2009. [145] Z. Wang, D. Xiao, F. Fang, R. Govindan, C.C. Pain, Y. Guo, Model identification of
[110] D.S. Broomhead, D. Lowe, Multivariable functional interpolation and adaptive reduced order fluid dynamics systems using deep learning, Int. J. Numer. Methods
networks, Complex Syst. 2 (3) (1988) 321–355. Fluid. 86 (4) (2018) 255–268.
[111] C. Harpham, C.W. Dawson, The effect of different basis functions on a radial basis [146] A.T. Mohan, D.V. Gaitonde, A Deep Learning Based Approach to Reduced Order
function network for time series prediction: a comparative study, Modeling for Turbulent Flow Control Using Lstm Neural Networks, arXiv
Neurocomputing 69 (16–18) (2006) 2161–2170. preprint, 2018. arXiv:1804.09269.
[112] J. Kou, W. Zhang, Multi-kernel neural networks for nonlinear unsteady [147] S.M. Rahman, S. Pawar, O. San, A. Rasheed, T. Iliescu, Nonintrusive reduced
aerodynamic reduced-order modeling Aerosp, Sci. Technol. 67 (2017) 309–326. order modeling framework for quasigeostrophic turbulence, Phys. Rev. 100 (5)
[113] J. Kou, W. Zhang, Reduced-order modeling for nonlinear aeroelasticity with (2019), 053306.
varying mach numbers, J. Aerospace Eng. 31 (6) (2018), 04018105. [148] S. Pawar, S.E. Ahmed, O. San, A. Rasheed, Data-driven recovery of hidden physics
[114] D.E. Myers, Kriging, cokriging, radial basis functions and the role of positive in reduced order modeling of fluid flows, Phys. Fluids 32 (2020), 036602.
definiteness, Comput. Math. Appl. 24 (12) (1992) 139–148. [149] W. Hou, D. Darakananda, J.D. Eldredge, Machine-learning-based detection of
[115] D.J. Linse, R.F. Stengel, Identification of aerodynamic coefficients using aerodynamic disturbances using surface pressure measurements, AIAA J. 57 (12)
computational neural networks, J. Guid. Contr. Dynam. 16 (6) (1993) (2019) 5079–5093.
1018–1025. [150] T. Li, T. Wu, Z. Liu, Nonlinear unsteady bridge aerodynamics: reduced-order
[116] F.D. Marques, J. Anderson, Identification and prediction of unsteady transonic modeling based on deep lstm networks, J. Wind Eng. Ind. Aerod. 198 (2020)
aerodynamic loads by multi-layer functionals, J. Fluid Struct. 15 (1) (2001) 104116.
83–106. [151] W. Li, S. Laima, X. Jin, W. Yuan, H. Li, A novel long short-term memory neural-
[117] S. Suresh, S.N. Omkar, V. Mani, T.N. Guru Prakash, Lift coefficient prediction at network-based self-excited force model of limit cycle oscillations of nonlinear
high angle of attack using recurrent neural network, Aero. Sci. Technol. 7 (8) flutter for various aerodynamic configurations, Nonlinear Dynam. 100 (3) (2020)
(2003) 595–602. 2071–2087.
[118] M. Ghoreyshi, A. Jirásek, R.M. Cummings, Computational approximation of [152] J.L. Lumley, The structure of inhomogeneous turbulence, in: Proceedings of the
nonlinear unsteady aerodynamics using an aerodynamic model hierarchy, Aero. International Colloquium on the Fine Scale Structure of the Atmosphere and its
Sci. Technol. 28 (1) (2013) 133–144. Influence on Radio Wave Propagation, 1967.
[119] A. Mannarino, P. Mantegazza, Nonlinear aeroelastic reduced order modeling by [153] C.W. Rowley, I. Mezić, S. Bagheri, P. Schlatter, D.S. Henningson, Spectral analysis
recurrent neural networks, J. Fluid Struct. 48 (2014) 103–121. of nonlinear flows, J. Fluid Mech. 641 (2009) 115–127.
[120] M. Winter, C. Breitsamter, Neurofuzzy-model-based unsteady aerodynamic [154] P.J. Schmid, Dynamic mode decomposition of numerical and experimental data,
computations across varying freestream conditions, AIAA J. 54 (9) (2016) J. Fluid Mech. 656 (2010) 5–28.
2705–2720. [155] G. Berkooz, P. Holmes, J.L. Lumley, The proper orthoghnal decomposition in the
[121] J. Kou, W. Zhang, An approach to enhance the generalization capability of analysis of turbulent flows, Annu. Rev. Fluid Mech. 25 (1) (1993) 539–575.
nonlinear aerodynamic reduced-order models, Aero. Sci. Technol. 49 (2016) [156] R. Dupuis, J.-C. Jouhaud, P. Sagaut, Surrogate modeling of aerodynamic
197–208. simulations for multiple operating conditions using machine learning, AIAA J. 56
[122] W. Zhang, J. Kou, Z. Wang, Nonlinear aerodynamic reduced-order model for (9) (2018) 3622–3635.
limit-cycle oscillation and flutter, AIAA J. 54 (10) (2016) 3304–3312. [157] L. Sirovich, Turbulence and the dynamics of coherent structures part i: coherent
[123] X. Wang, J. Kou, W. Zhang, Unsteady aerodynamic modeling based on fuzzy structures, Q. Appl. Math. 45 (3) (1987) 561–571.
scalar radial basis function neural networks, Proc. Inst. Mech. Eng. Part G-J. [158] O.T. Schmidt, T. Colonius, Guide to spectral proper orthogonal decomposition,
Aerosp. Eng. 233 (14) (2019) 5107–5121. AIAA J. 58 (3) (2020) 1023–1033.
[124] G. Sun, S. Wang, A review of the artificial neural network surrogate modeling in [159] A. Towne, O.T. Schmidt, T. Colonius, Spectral proper orthogonal decomposition
aerodynamic design, Proc. Inst. Mech. Eng. Part G-J. Aerosp. Eng. 233 (16) and its relationship to dynamic mode decomposition and resolvent analysis,
(2019) 5863–5872. J. Fluid Mech. 847 (2018) 821–867.
[125] L. Ljung, Identification of Nonlinear Systems, 2007. [160] P. Benner, S. Gugercin, K. Willcox, A survey of projection-based model reduction
[126] V. Verdult, L. Ljung, M. Verhaegen, Identification of composite local linear state- methods for parametric dynamical systems, SIAM Rev. 57 (4) (2015) 483–531.
space models using a projected gradient search, Int. J. Contr. 75 (16–17) (2002) [161] R.m. Bourguet, M. Braza, A. Dervieux, Reduced-order modeling for unsteady
1385–1398. transonic flows around an airfoil, Phys. Fluids 19 (11) (2007) 111701.
[127] L. Ljung, Perspectives on system identification, Annu. Rev. Contr. 34 (1) (2010) [162] N. Aubry, P. Holmes, J.L. Lumley, E. Stone, The dynamics of coherent structures
1–12. in the wall region of a turbulent boundary layer, J. Fluid Mech. 192 (1988)
[128] J. Schoukens, J.G. Nemeth, P. Crama, Y. Rolain, R. Pintelon, Fast approximate 115–173.
identification of nonlinear systems, Automatica 39 (7) (2003) 1267–1274. [163] A.E. Deane, I.G. Kevrekidis, G.E. Karniadakis, S.A. Orszag, Low-dimensional
[129] D.H. Baldelli, R. Lind, M. Brenner, Control-oriented flutter/limit-cycle-oscillation models for complex geometry flows: application to grooved channels and circular
prediction framework, J. Guid. Contr. Dynam. 31 (6) (2008) 1634–1643. cylinders, Phys. Fluids 3 (10) (1991) 2337–2354.
[130] K.-Y. Lum, K. L. Lai, Identification of a hammerstein model for wing flutter [164] X. Ma, G.E. Karniadakis, A low-dimensional model for simulating three-
analysis using cfd data and correlation method, In: 2010 American Control dimensional cylinder flow, J. Fluid Mech. 458 (2002) 181–190.
Conference. vols. 1314–1319, Marriott Waterfront, Baltimore, MD, USA. [165] B.R. Noack, K. Afanasiev, M. Morzyński, G. Tadmor, F. Thiele, A hierarchy of low-
[131] Y. Dai, Z. Wu, C. Yang, Identification and robust limit-cycle-oscillation analysis of dimensional models for the transient and post-transient cylinder wake, J. Fluid
uncertain aeroelastic system, Sci. China Technol. Sci. 54 (7) (2011) 1841–1848. Mech. 497 (2003) 335–363.
[132] R. Huang, H. Hu, Y. Zhao, Nonlinear reduced-order modeling for multiple-input/ [166] S.W. Choi, D.W. Lee, J.H. Park, I.B. Lee, Nonlinear regression using rbfn with
multiple-output aerodynamic systems, AIAA J. 52 (6) (2014) 1219–1231. linear submodels, Chemomet. Intell. Lab. 65 (2) (2003) 191–208.
[133] R. Huang, H. Li, H. Hu, Y. Zhao, Open/closed-loop aeroservoelastic predictions [167] B.R. Noack, P. Papas, P.A. Monkewitz, The need for a pressure-term
via nonlinear, reduced-order aerodynamic models, AIAA J. 53 (7) (2015) representation in empirical galerkin models of incompressible shear flows,
1812–1824. J. Fluid Mech. 523 (2005) 339–365.

31
J. Kou and W. Zhang Progress in Aerospace Sciences 125 (2021) 100725

[168] N. Deng, B.R. Noack, M. Morzyński, L.R. Pastur, Low-order model for successive [200] K.H. Park, S.O. Jun, S.M. Baek, M.H. Cho, K.J. Yee, D.H. Lee, Reduced-order
bifurcations of the fluidic pinball, J. Fluid Mech. 884 (2020) A37. model with an artificial neural network for aerostructural design optimization,
[169] C.W. Rowley, T. Colonius, R.M. Murray, Model reduction for compressible flows J. Aircraft 50 (4) (2013) 1106–1116.
using pod and galerkin projection, Phys. Nonlinear Phenom. 189 (1–2) (2004) [201] D.J. Lucia, P.S. Beran, W.A. Silva, Aeroelastic system development using proper
115–129. orthogonal decomposition and volterra theory, J. Aircraft 42 (2) (2005) 509–518.
[170] C.W. Rowley, D.R. Williams, T. Colonius, R.M. Murray, D.G. Macmynowski, [202] P.J. Attar, E.H. Dowell, J.R. White, J.P. Thomas, Reduced order nonlinear system
Linear models for control of cavity flow oscillations, J. Fluid Mech. 547 (2006) identification methodology, AIAA J. 44 (8) (2006) 1895–1904.
317–330. [203] S. Walton, O. Hassan, K. Morgan, Reduced order modelling for unsteady fluid
[171] D.J. Lucia, P.S. Beran, Projection methods for reduced order models of flow using proper orthogonal decomposition and radial basis functions, Appl.
compressible flows, J. Comput. Phys. 188 (1) (2003) 252–280. Math. Model. 37 (20–21) (2013) 8930–8945.
[172] C. Huang, K. Duraisamy, C. Merkle, Challenges in reduced order modeling of [204] K. Lindhorst, M.C. Haupt, P. Horst, Efficient surrogate modelling of nonlinear
reacting flows, In: 2018 Joint Propulsion Conference. AIAA 2018-4675. aerodynamics in aerostructural coupling schemes, AIAA J. 52 (9) (2014)
[173] J. Xu, C. Huang, K. Duraisamy, Reduced-order modeling framework for 1952–1966.
combustor instabilities using truncated domain training, AIAA J. 58 (2) (2020) [205] D. Xiao, P. Yang, F. Fang, J. Xiang, C.C. Pain, I.M. Navon, Non-intrusive reduced
618–632. order modelling of fluid–structure interactions, Comput. Methods Appl. Mech.
[174] K. Carlberg, C. Bou-Mosleh, C. Farhat, Efficient non-linear model reduction via a Eng. 303 (2016) 35–54.
least-squares petrov-galerkin projection and compressive tensor approximations, [206] M. Winter, C. Breitsamter, Efficient unsteady aerodynamic loads prediction based
Int. J. Numer. Methods Eng. 86 (2) (2011) 155–181. on nonlinear system identification and proper orthogonal decomposition, J. Fluid
[175] D.M. Luchtenburg, B. Günther, B.R. Noack, R. King, G. Tadmor, A generalized Struct. 67 (2016) 1–21.
mean-field model of the natural and high-frequency actuated flow around a high- [207] J.S. Hesthaven, S. Ubbiali, Non-intrusive reduced order modeling of nonlinear
lift configuration, J. Fluid Mech. 623 (2009) 283–316. problems using neural networks, J. Comput. Phys. 363 (2018) 55–78.
[176] L. Alvergue, H. Babaee, G. Gu, S. Acharya, Feedback stabilization of a reduced- [208] O. San, R. Maulik, M. Ahmed, An artificial neural network framework for reduced
order model of a jet in crossflow, AIAA J. 53 (9) (2015) 2472–2481. order modeling of transient flows, Commun. Nonlinear Sci. Numer. Simulat. 77
[177] K. Willcox, J. Peraire, Balanced model reduction via the proper orthogonal (2019) 271–287.
decomposition, AIAA J. 40 (11) (2002) 2323–2330. [209] C. Wang, J.S. Hesthaven, J. Bai, Y. Qiu, Y. Zhang, T. Yang, Greedy nonintrusive
[178] C.W. Rowley, Model reduction for fluids, using balanced proper orthogonal reduced order model for fluid dynamics, AIAA J. 56 (12) (2018) 4927–4943.
decomposition, Int. J. Bifurcat. Chaos 15 (2005) 997–1013, 03. [210] D. Ventur, G.E. Karniadakis, Gappy data and reconstruction procedures for flow
[179] T.P. Sapsis, P.F.J. Lermusiaux, Dynamically orthogonal field equations for past a cylinder, J. Fluid Mech. 519 (2004) 315–336.
continuous stochastic dynamical systems, Phys. Nonlinear Phenom. 238 (23–24) [211] K. Willcox, Unsteady flow sensing and estimation via the gappy proper orthogonal
(2009) 2347–2360. decomposition, Comput. Fluids 35 (2) (2006) 208–226.
[180] K.C. Hall, Eigenanalysis of unsteady flows about airfoils, cascades, and wings, [212] R. Everson, L. Sirovich, Karhunen–loeve procedure for gappy data, J. Opt. Soc.
AIAA J. 32 (12) (1994) 2426–2432. Am. A 12 (8) (1995) 1657–1664.
[181] E.H. Dowell, Eigenmode analysis in unsteady aerodynamics: reduced-order [213] S. Chaturantabut, D. Sorensen, Nonlinear model reduction via discrete empirical
models, AIAA J. 34 (8) (1996) 1578–1583. interpolation, SIAM J. Sci. Comput. 32 (5) (2009) 2737–2764.
[182] M. C. Romanowski, Reduced order unsteady aerodynamic and aeroelastic models [214] D. Amsallem, M.J. Zahr, Y. Choi, C. Farhat, Design optimization using hyper-
using karhunen-loeve eigenmodes, In: 6th Symposium on Multidisciplinary reduced-order models, Struct. Multidiscip. Optim. 51 (2015) 919–940.
Analysis and Optimization. AIAA 1996-3981. [215] H. Gao, J.-X. Wang, M.J. Zahr, Non-intrusive model reduction of large-scale,
[183] T. Kim, Frequency-domain karhunen-loeve method and its application to linear nonlinear dynamical systems using deep learning, Phys. Nonlinear Phenom. 412
dynamic systems, AIAA J. 36 (11) (1998) 2117–2123. (2020).
[184] K.C. Hall, J.P. Thomas, E.H. Dowell, Proper orthogonal decomposition technique [216] T. Bui-Thanh, K. Willcox, O. Ghattas, B. van Bloemen Waanders, Goal-oriented,
for transonic unsteady aerodynamic flows, AIAA J. 38 (10) (2000) 1853–1862. model-constrained optimization for reduction of large-scale systems, J. Comput.
[185] J.P. Thomas, E.H. Dowell, K.C. Hall, Three-dimensional transonic aeroelasticity Phys. 224 (2) (2007) 880–896.
using proper orthogonal decomposition-based reduced-order models, J. Aircraft [217] R. Deshmukh, J.J. McNamara, Z. Liang, J.Z. Kolter, A. Gogulapati, Model order
40 (3) (2003) 541–551. reduction using sparse coding exemplified for the lid-driven cavity, J. Fluid Mech.
[186] P. LeGresley, J. J. Alonso, Airfoil design optimization using reduced order models 808 (10) (2016) 189–223.
based on proper orthogonal decomposition, In: Fluids 2020 Conference and [218] A. Placzek, D.M. Tran, R. Ohayon, A nonlinear pod-galerkin reduced-order model
Exhibit. AIAA 2000-2545. for compressible flows taking into account rigid body motions, Comput. Methods
[187] M.R. Brake, M.F. Barone, D.J. Segalman, Nonlinear model reduction of von Appl. Mech. Eng. 200 (49) (2011) 3497–3514.
kármán plates under linearized compressible fluid flow, AIAA J. 50 (5) (2012) [219] B.A. Freno, N.R. Matula, R.L. Fontenot, P.G.A. Cizmas, The use of dynamic basis
1047–1059. functions in proper orthogonal decomposition, J. Fluid Struct. 54 (2015)
[188] A. Da Ronch, K.J. Badcock, Y. Wang, A. Wynn, R. Palacios, Nonlinear model 332–360.
reduction for flexible aircraft control design, in: AIAA Atmospheric Flight [220] B. Yu, G. Yu, Z. Cao, D. Han, Q. Shao, Fast calculation of the soil temperature field
Mechanics Conference, August 13-16, AIAA, Minneapolis, Minnesota, 2012, around a buried oil pipeline using a body-fitted coordinates-based pod-galerkin
pp. 2012–4404. reduced-order model, Numer. Heat Tran., Part A: Applications 63 (10) (2013)
[189] G. Chen, J. Sun, Y. Li, Active flutter suppression control law design method based 776–794.
on balanced proper orthogonal decomposition reduced order model, Nonlinear [221] B. Peherstorfer, K. Willcox, Dynamic data-driven reduced-order models, Comput.
Dynam. 70 (1) (2012) 1–12. Methods Appl. Mech. Eng. 291 (1) (2015) 21–41.
[190] R. Zimmermann, S. Görtz, Improved extrapolation of steady turbulent [222] M. Balajewicz, E.H. Dowell, B.R. Noack, Low-dimensional modelling of high-
aerodynamics using a non-linear pod-based reduced order model, Aeronaut. J. Reynolds-number shear flows incorporating constraints from the Navier-Stokes
116 (1184) (2012) 1079–1100. equation, J. Fluid Mech. 729 (2013) 285–308.
[191] M. Ripepi, M.J. Verveld, N.W. Karcher, T. Franz, M. Abu-Zurayk, S. Görtz, T. [223] J. Östh, B.R. Noack, S. Krajnović, D. Barros, J. Borée, On the need for a nonlinear
M. Kier, Reduced-order models for aerodynamic applications, loads and mdo, subscale turbulence term in pod models as exemplified for a high-Reynolds-
CEAS Aeronautical Journal 9 (2018) 171–193. number flow over an ahmed body, J. Fluid Mech. 747 (2014) 518–544.
[192] J. Yu, C. Yan, M. Guo, Non-intrusive reduced-order modeling for fluid problems: a [224] P. J. Schmid, J. Sesterhenn, Dynamic mode decomposition of numerical and
brief review, Proc. Inst. Mech. Eng. Part G-J. Aerosp. Eng. 233 (16) (2019) experimental data, In: Sixty-First Annual Meeting of the APS Division of Fluid
5896–5912. Dynamics. 208, San Antonio, Texas, USA.
[193] M. Fossati, W.G. Habashi, Multiparameter analysis of aero-icing problems using [225] G. He, J. Wang, C. Pan, Initial growth of a disturbance in a boundary layer
proper orthogonal decomposition and multidimensional interpolation, AIAA J. 51 influenced by a circular cylinder wake, J. Fluid Mech. 718 (2013) 116–130.
(4) (2013) 946–960. [226] Z. Chen, N.A. Adams, Mode interactions of a high-subsonic deep cavity, Phys.
[194] J. Li, W. Zhang, The performance of proper orthogonal decomposition in Fluids 29 (5) (2017), 056102.
discontinuous flows, Theoretical & Applied Mechanics Letters 6 (5) (2016) [227] Z.-H. Wan, L. Zhou, B.-F. Wang, D.-J. Sun, Dynamic mode decomposition of
236–243. forced spatially developed transitional jets, Eur. J. Mech. B Fluid 51 (2015)
[195] N.J. Falkiewicz, C.E.S. Cesnik, A.R. Crowell, J.J. McNamara, Reduced-order 16–26.
aerothermoelastic framework for hypersonic vehicle control simulation, AIAA J. [228] Z. Tang, N. Jiang, Dynamic mode decomposition of hairpin vortices generated by
49 (8) (2011) 1625–1646. a hemisphere protuberance, Sci. China Phys. Mech. 55 (1) (2012) 118–124.
[196] A.R. Crowell, J.J. McNamara, Model reduction of computational [229] J. Kou, W. Zhang, Dynamic mode decomposition and its applications in fluid
aerothermodynamics for hypersonic aerothermoelasticity, AIAA J. 50 (1) (2012) dynamics, Acta Aerodyn. Sin. 36 (2) (2018) 163–179.
74–84. [230] J.N. Kutz, S.L. Brunton, B.W. Brunton, J.L. Proctor, Dynamic Mode
[197] X. Chen, L. Liu, T. Long, Z. Yue, A reduced order aerothermodynamic modeling Decomposition: Data-Driven Modeling of Complex Systems, Society for Industrial
framework for hypersonic vehicles based on surrogate and pod, Chin. J. Aeronaut. and Applied Mathematics (SIAM), 2016.
28 (5) (2015) 1328–1342. [231] B.O. Koopman, Hamiltonian systems and transformation in hilbert space, Proc.
[198] D. Huang, P.P. Friedmann, An aerothermoelastic analysis framework with Natl. Acad. Sci. Unit. States Am. 17 (5) (1931) 315–318.
reduced-order modeling applied to composite panels in hypersonic flows, J. Fluid [232] I. Mezić, Spectral properties of dynamical systems, model reduction and
Struct. 94 (2020) 102927. decompositions, Nonlinear Dynam. 41 (1) (2005) 309–325.
[199] J. Li, M.A. Bouhlel, J.R.R.A. Martins, Data-based approach for fast airfoil analysis [233] S. Bagheri, Koopman-mode decomposition of the cylinder wake, J. Fluid Mech.
and optimization, AIAA J. 57 (2) (2019) 581–596. 726 (2013) 596–623.

32
J. Kou and W. Zhang Progress in Aerospace Sciences 125 (2021) 100725

[234] I. Mezić, Analysis of fluid flows via spectral properties of the koopman operator, [268] J.L. Proctor, S.L. Brunton, J.N. Kutz, Dynamic mode decomposition with control,
Annu. Rev. Fluid Mech. 45 (1) (2013) 357–378. SIAM J. Appl. Dyn. Syst. 15 (1) (2016) 142–161.
[235] J. Kou, W. Zhang, An improved criterion to select dominant modes from dynamic [269] J. Annoni, P. Seiler, A method to construct reduced-order parameter-varying
mode decomposition, Eur. J. Mech. B Fluid 62 (2017) 109–129. models, Int. J. Robust Nonlin. 27 (4) (2017) 582–597.
[236] J.H. Tu, C.W. Rowley, D.M. Luchtenburg, S.L. Brunton, J.N. Kutz, On dynamic [270] M. Korda, I. Mezić, Linear predictors for nonlinear dynamical systems: koopman
mode decomposition: theory and applications, J. Comput. Dynam. 1 (2) (2014) operator meets model predictive control, Automatica 93 (2018) 149–160.
391–421. [271] S. T. M. Dawson, N. K. Schiavone, C. W. Rowley, D. R. Williams, A data-driven
[237] K.K. Chen, J.H. Tu, C.W. Rowley, Variants of dynamic mode decomposition: modeling framework for predicting forces and pressures on a rapidly pitching
boundary condition, koopman, and fourier analyses, J. Nonlinear Sci. 22 (6) airfoil, In: 45th AIAA Fluid Dynamics Conference, AIAA AVIATION Forum.
(2012) 887–915. Dallas, TX, USA.
[238] A. Wynn, D.S. Pearson, B. Ganapathisubramani, P.J. Goulart, Optimal mode [272] M.M. Opgenoord, M. Drela, K. Willcox, Physics-based low-order model for
decomposition for unsteady flows, J. Fluid Mech. 733 (2013) 473–503. transonic flutter prediction, AIAA J. 56 (4) (2018) 1519–1531.
[239] M.R. Jovanović, P.J. Schmid, J.W. Nichols, Sparsity-promoting dynamic mode [273] E.A. Deem, L.N. Cattafesta, M.S. Hemati, H. Zhang, C. Rowley, R. Mittal, Adaptive
decomposition, Phys. Fluids 26 (2) (2014), 024103. separation control of a laminar boundary layer using online dynamic mode
[240] M.S. Hemati, M.O. Williams, C.W. Rowley, Dynamic mode decomposition for decomposition, J. Fluid Mech. 903 (2020).
large and streaming datasets, Phys. Fluids 26 (11) (2014) 111701. [274] J. Kou, W. Zhang, Dynamic mode decomposition with exogenous input for data-
[241] B. Li, J. Garicano-Mena, Y. Zheng, E. Valero, Dynamic mode decomposition driven modeling of unsteady flows, Phys. Fluids 31 (5) (2019), 057106.
analysis of spatially agglomerated flow databases, Energies 13 (9) (2020). [275] E. Ferrer, J. De Vicente, E. Valero, Low cost 3d global instability analysis and flow
[242] J. Garicano-Mena, B. Li, E. Ferrer, E. Valero, A composite dynamic mode sensitivity based on dynamic mode decomposition and high-order numerical
decomposition analysis of turbulent channel flows, Phys. Fluids 31 (11) (2019). tools, Int. J. Numer. Methods Fluid. 76 (3) (2014) 169–184.
[243] H. Zhang, C.W. Rowley, E.A. Deem, L.N. Cattafesta, Online dynamic mode [276] M. Wu, Z. Han, H. Nie, W. Song, S.L. Clainche, E. Ferrer, A transition prediction
decomposition for time-varying systems, SIAM J. Appl. Dyn. Syst. 18 (3) (2019) method for flow over airfoils based on high-order dynamic mode decomposition,
1586–1609. Chin. J. Aeronaut. 32 (11) (2019) 2408–2421.
[244] M.S. Hemati, C.W. Rowley, E.A. Deem, L.N. Cattafesta, De-biasing the dynamic [277] S.A. Renganathan, Y. Liu, D.N. Mavris, Koopman-based approach to nonintrusive
mode decomposition for applied koopman spectral analysis of noisy datasets, projection-based reduced-order modeling with black-box high-fidelity models,
Theor. Comput. Fluid Dynam. 31 (4) (2017) 349–368. AIAA J. 56 (10) (2018) 4087–4111.
[245] S.T.M. Dawson, M.S. Hemati, M.O. Williams, C.W. Rowley, Characterizing and [278] A. Saito, T. Kuno, Data-driven experimental modal analysis by dynamic mode
correcting for the effect of sensor noise in the dynamic mode decomposition, Exp. decomposition, J. Sound Vib. 481 (2020).
Fluid 57 (3) (2016) 42. [279] A. Goza, T. Colonius, Modal decomposition of fluid–structure interaction with
[246] F. Guéniat, L. Mathelin, L.R. Pastur, A dynamic mode decomposition approach for application to flag flapping, J. Fluid Struct. 81 (2018) 728–737.
large and arbitrarily sampled systems, Phys. Fluids 27 (2) (2015), 025113. [280] Y. Liu, G. Wang, Z. Ye, Dynamic mode extrapolation to improve the efficiency of
[247] J.H. Tu, C.W. Rowley, J.N. Kutz, J.K. Shang, Spectral analysis of fluid flows using dual time stepping method, J. Comput. Phys. 352 (2018) 190–212.
sub-nyquist-rate piv data, Exp. Fluid 55 (9) (2014) 1805. [281] Y. Liu, W. Zhang, J. Kou, Mode multigrid - a novel convergence acceleration
[248] B.R. Noack, W. Stankiewicz, M. Morzyński, P.J. Schmid, Recursive dynamic mode method, Aero. Sci. Technol. 92 (2019) 605–619.
decomposition of transient and post-transient wake fows, J. Fluid Mech. 809 [282] W. Chen, W. Zhang, Y. Liu, J. Kou, Accelerating the convergence of steady adjoint
(2016) 843–872. equations by dynamic mode decomposition, Struct. Multidiscip. Optim. 62 (2)
[249] M. Sieber, C.O. Paschereit, K. Oberleithner, Spectral proper orthogonal (2020) 747–756.
decomposition, J. Fluid Mech. 792 (2016) 798–828. [283] B. Peherstorfer, K. Willcox, M. Gunzburger, Survey of multifidelity methods in
[250] T. Sayadi, P.J. Schmid, F. Richecoeur, D. Durox, Parametrized data-driven uncertainty propagation, inference, and optimization, SIAM Rev. 60 (3) (2018)
decomposition for bifurcation analysis, with application to thermo-acoustically 550–591.
unstable systems, Phys. Fluids 27 (3) (2015), 037102. [284] R. Yondo, E. Andrés, E. Valero, A review on design of experiments and surrogate
[251] D.A. Bistrian, I.M. Navon, Randomized dynamic mode decomposition for non- models in aircraft real-time and many-query aerodynamic analyses, Prog. Aero.
intrusive reduced order modelling, Int. J. Numer. Methods Eng. 112 (1) (2017) Sci. 96 (2018) 23–61.
3–25. [285] I.M. Navon, Data assimilation for numerical weather prediction: a review, in: L.X.
[252] J.N. Kutz, X. Fu, S.L. Brunton, Multiresolution dynamic mode decomposition, S.K. Park (Ed.), Data Assimilation for Atmospheric, Oceanic and Hydrologic
SIAM J. Appl. Dyn. Syst. 15 (2) (2016) 713–735. Applications, Springer, Berlin, Heidelberg, 2009, pp. 21–65.
[253] S. Le Clainche, J.M. Vega, Higher order dynamic mode decomposition, SIAM J. [286] V. Mons, J.C. Chassaing, T. Gomez, P. Sagaut, Reconstruction of unsteady viscous
Appl. Dyn. Syst. 16 (2) (2017) 882–925. flows using data assimilation schemes, J. Comput. Phys. 316 (2016) 255–280.
[254] M.O. Williams, I.G. Kevrekidis, C.W. Rowley, A data-driven approximation of the [287] Z. Belligoli, R.P. Dwight, G. Eitelberg, Reconstruction of turbulent flows at high
koopman operator: extending dynamic mode decomposition, J. Nonlinear Sci. 25 Reynolds numbers using data assimilation techniques, AIAA J. 59 (3) (2021)
(6) (2015) 1307–1346. 855–867.
[255] M.O. Williams, C.W. Rowley, I.G. Kevrekidis, A kernel-based method for data- [288] M.J. Mifsud, D.G. MacManus, S.T. Shaw, A variable-fidelity aerodynamic model
driven koopman spectral analysis, J. Comput. Dynam. 2 (2) (2017) 247–265. using proper orthogonal decomposition, Int. J. Numer. Methods Fluid. 82 (10)
[256] S. Le Clainche, J.M. Vega, Spatio-temporal koopman decomposition, J. Nonlinear (2016) 646–663.
Sci. 28 (2018) 1793–1842. [289] M.G. Fernández-Godino, C. Park, N.H. Kim, R.T. Haftka, Review of Multi-Fidelity
[257] G. Tissot, L. Cordier, N. Benard, B.R. Noack, Model reduction using dynamic Models, 2016 eprint arXiv: 1609.07196.
mode decomposition, CR. Mecanique 342 (6–7) (2014) 410–416. [290] D.E. Myers, Matrix formulation of co-kriging, J. Int. Assoc. Math. Geol. 14 (3)
[258] D.A. Bistrian, I.M. Navon, An improved algorithm for the shallow water equations (1982) 249–257.
model reduction: dynamic mode decomposition vs pod, Int. J. Numer. Methods [291] M.C. Kennedy, A. O’Hagan, Predicting the output from a complex computer code
Fluid. 78 (9) (2015) 552–580. when fast approximations are available, Biometrika 87 (1) (2000) 1–13.
[259] X. Lai, X. Wang, Y. Nie, Q. Li, Characterizing complex flows using adaptive sparse [292] L. Zheng, T.L. Hedrick, R. Mittal, A multi-fidelity modelling approach for
dynamic mode decomposition with error approximation, Int. J. Numer. Methods evaluation and optimization of wing stroke aerodynamics in flapping flight,
Fluid. 92 (6) (2019) 587–602. J. Fluid Mech. 721 (2013) 118–154.
[260] J. Kou, S. Le Clainche, W. Zhang, A reduced-order model for compressible flows [293] C. Badrya, A. Sridharan, J.D. Baeder, C.M. Kroninger, Multi-fidelity coupled trim
with buffeting condition using higher order dynamic mode decomposition with a analysis of a flapping-wing micro air vehicle flight, J. Aircraft 54 (5) (2017)
mode selection criterion, Phys. Fluids 30 (1) (2018). 1614–1630.
[261] L. Poplingher, D.E. Raveh, E.H. Dowell, Modal analysis of transonic shock buffet [294] A.J. Keane, Wing optimization using design of experiment, response surface, and
on 2d airfoil, AIAA J. 57 (7) (2019) 2851–2866. data fusion methods, J. Aircraft 40 (4) (2003) 741–750.
[262] N.F. Giannelis, O. Levinski, G.A. Vio, Origins of atypical shock buffet motions on a [295] L. Leifsson, S. Koziel, Multi-fidelity design optimization of transonic airfoils using
supercritical aerofoil, Aero. Sci. Technol. 107 (2020). physics-based surrogate modeling and shape-preserving response prediction,
[263] A. Broatch, J. García-Tíscar, F. Roig, S. Sharma, Dynamic mode decomposition of J. Comput. Sci. 1 (2) (2010) 98–106.
the acoustic field in radial compressors, Aero. Sci. Technol. 90 (2019) 388–400. [296] Z. Han, R. Zimmermann, S. Görtz, Alternative cokriging model for variable-
[264] M. Kiewat, L. Haag, T. Indinger, V. Zander, Low-memory reduced-order fidelity surrogate modeling, AIAA J. 50 (5) (2012) 1205–1210.
modelling with dynamic mode decomposition applied on unsteady wheel [297] A. March, K. Willcox, Provably convergent multifidelity optimization algorithm
aerodynamics, In: ASME 2017 Fluids Engineering Division Summer Meeting. Vol. not requiring high-fidelity derivatives, AIAA J. 50 (5) (2012) 1079–1089.
vol. 58066, p. V58001CT58023A58006, Waikoloa, Hawaii, USA. [298] M. Ghoreyshi, K.J. Badcock, M.A. Woodgate, Accelerating the numerical
[265] M. Kiewat, D. Matsumoto, L. Haag, V. Zander, T. Indinger, Online dynamic mode generation of aerodynamic models for flight simulation, J. Aircraft 43 (3) (2009)
decomposition methods for the investigation of unsteady aerodynamics of the 972–980.
drivaer model (second report), International Journal of Automotive Engineering 9 [299] P.C. Murphy, V. Klein, N.T. Frink, Nonlinear unsteady aerodynamic modeling
(2) (2018) 72–78. using wind-tunnel and computational data, J. Aircraft 54 (2) (2017) 659–683.
[266] H. Zhang, S.T.M. Dawson, C.W. Rowley, E.A. Deem, L.N. Cattafesta, Evaluating [300] T. Rokita, P.P. Friedmann, Multifidelity cokriging for high-dimensional output
the Accuracy of the Dynamic Mode Decomposition, Arxiv preprint, 2017. arxiv functions with application to hypersonic airloads computation, AIAA J. 56 (8)
1710.00745. (2018) 3060–3070.
[267] S. Pan, A.-M. Nicholas, K. Duraisamy, Sparsity-promoting Algorithms for the [301] E.J. Parish, K. Duraisamy, A paradigm for data-driven predictive modeling using
Discovery of Informative Koopman Invariant Subspaces, Arxiv preprint, 2020. field inversion and machine learning, J. Comput. Phys. 305 (2016) 758–774.
arXiv:2002.10637.

33
J. Kou and W. Zhang Progress in Aerospace Sciences 125 (2021) 100725

[302] J.-X. Wang, J.-L. Wu, H. Xiao, Physics-informed machine learning for predictive [336] B.R. Noack, M. Schlegel, B. Ahlborn, G. Mutschke, M. Morzynski, P. Comte,
turbulence modeling: using data to improve rans modeled Reynolds stresses, G. Tadmor, A finite-time thermodynamics of unsteady fluid flows, J. Non-
Phys. Rev. Fluids 2 (3) (2016) 1–22. Equilibrium Thermodyn. 33 (2) (2008) 103–148.
[303] P. Perdikaris, D. Venturi, J.O. Royset, G.E. Karniadakis, Multi-fidelity modelling [337] Z. Wang, I. Akhtar, J. Borggaard, T. Iliescu, Proper orthogonal decomposition
via recursive co-kriging and Gaussian-markov random fields, P. Roy. Soc. A-Math. closure models for turbulent flows: a numerical comparison, Comput. Methods
Phy. 471 (2179) (2015) 20150018. Appl. Mech. Eng. 237–240 (2012) 10–26.
[304] P.S. Palar, T. Tsuchiya, G.T. Parks, Multi-fidelity non-intrusive polynomial chaos [338] M. Couplet, C. Basdevant, P. Sagaut, Calibrated reduced-order pod-galerkin
based on regression, Comput. Methods Appl. Mech. Eng. 305 (2016) 579–606. system for fluid flow modelling, J. Comput. Phys. 207 (1) (2005) 192–220.
[305] H. Xiao, J.-L. Wu, J.-X. Wang, R. Sun, C.J. Roy, Quantifying and reducing model- [339] L. Chen, W.N. Edeling, S.J. Hulshoff, Pod enriched boundary models and their
form uncertainties in Reynolds-averaged Navier-Stokes simulations: a data- optimal stabilisation, Int. J. Numer. Methods Fluid. 77 (2) (2015) 92–107.
driven, physics-informed bayesian approach, J. Comput. Phys. 324 (2016) [340] D. Amsallem, C. Farhat, Stabilization of projection-based reduced-order models,
115–136. Int. J. Numer. Methods Eng. 91 (4) (2012) 358–377.
[306] J. De Baar, S. Roberts, R. Dwight, B. Mallol, Uncertainty quantification for a [341] M.F. Barone, I. Kalashnikova, D.J. Segalman, H.K. Thornquist, Stable galerkin
sailing yacht hull, using multi-fidelity kriging, Comput, Fluid 123 (2015) reduced order models for linearized compressible flow, J. Comput. Phys. 228 (6)
185–201. (2009) 1932–1946.
[307] A. Feldstein, D. Lazzara, N. Princen, K. Willcox, Multifidelity data fusion: [342] I. Kalashnikova, B. van Bloemen Waanders, S. Arunajatesan, M. Barone,
application to blended-wing-body multidisciplinary analysis under uncertainty, Stabilization of projection-based reduced order models for linear time-invariant
AIAA J. 58 (2) (2020) 889–906. systems via optimization-based eigenvalue reassignment, Comput. Methods Appl.
[308] J. Kou, W. Zhang, Multi-fidelity modeling framework for nonlinear unsteady Mech. Eng. 272 (2014) 251–270.
aerodynamics of airfoils, Appl. Math. Model. 76 (2019) 832–855. [343] I. Akhtar, A.H. Nayfeh, C.J. Ribbens, On the stability and extension of reduced-
[309] C. Jiang, Y.C. Soh, H. Li, Sensor and cfd data fusion for airflow field estimation, order galerkin models in incompressible flows, Theor. Comput. Fluid Dynam. 23
Appl. Therm. Eng. 92 (2016) 149–161. (3) (2009) 213–237.
[310] C.J. Ruscher, Application of Data Fusion to Fluid Dynamic Data, Syracuse [344] L. Fick, Y. Maday, A.T. Patera, T. Taddei, A stabilized pod model for turbulent
University, 2014. flows over a range of Reynolds numbers: optimal parameter sampling and
[311] M.O. Williams, C.W. Rowley, I. Mezić, I.G. Kevrekidis, Data fusion via intrinsic constrained projection, J. Comput. Phys. 371 (2018) 214–243.
dynamic variables: an application of data-driven koopman spectral analysis, [345] H. Imtiaz, I. Akhtar, Nonlinear closure modeling in reduced order models for
Europhys. Lett. 109 (4) (2015) 40007. turbulent flows: a dynamical system approach, Nonlinear Dynam. 99 (1) (2019)
[312] R. Zimmermann, A. Vendl, S. Görtz, Reduced-order modeling of steady flows 479–494.
subject to aerodynamic constraints, AIAA J. 52 (2) (2014) 255–266. [346] E. Rezaian, M. Wei, A hybrid stabilization approach for reduced-order models of
[313] M. Mifsud, A. Vendl, L.-U. Hansen, S. Görtz, Fusing wind-tunnel measurements compressible flows with shock-vortex interaction, Int. J. Numer. Methods Eng.
and cfd data using constrained gappy proper orthogonal decomposition, Aero. Sci. 121 (8) (2020) 1629–1646.
Technol. 86 (2019) 312–326. [347] C. Huang, W. E. Anderson, C. L. Merkle, V. Sankaran, Investigation of the stability
[314] J. Yu, J.S. Hesthaven, Flowfield reconstruction method using artificial neural of pod galerkin techniques for reduced order model development, In: AIAA
network, AIAA J. 57 (2) (2019) 482–498. SciTech Forum. AIAA Paper 2016-1937.
[315] X. Wang, J. Kou, W. Zhang, Multi-fidelity surrogate reduced-order modeling of [348] L. Cordier, B.A. El Majd, J. Favier, Calibration of pod reduced-order models using
steady flow estimation, Int. J. Numer. Methods Fluid. 92 (12) (2020) 1826–1844. tikhonov regularization, Int. J. Numer. Methods Fluid. 63 (2) (2010) 269–296.
[316] H. Akaike, A new look at the statistical model identification, IEEE Trans. [349] J.C. Loiseau, S.L. Brunton, Constrained sparse galerkin regression, J. Fluid Mech.
Automat. Contr. 19 (6) (1974) 716–723. 838 (2018) 42–67.
[317] G. Schwarz, Estimating the dimension of a model, Ann. Stat. 6 (2) (1978) [350] O. San, R. Maulik, Neural network closures for nonlinear model order reduction,
461–464. Adv. Comput. Math. 44 (2018) 1717–1750.
[318] J. Kou, W. Zhang, Layered reduced-order models for nonlinear aerodynamics and [351] E.J. Parish, K. Duraisamy, Non-markovian closure models for large eddy
aeroelasticity, J. Fluid Struct. 68 (2017) 174–193. simulations using the mori-zwanzig formalism, Phys. Rev. Fluids 2 (1) (2017).
[319] A. Mannarino, E.H. Dowell, Reduced-order models for computational-fluid- [352] S. Pan, K. Duraisamy, Data-driven discovery of closure models, SIAM J. Appl.
dynamics-based nonlinear aeroelastic problems, AIAA J. 53 (9) (2015) Dyn. Syst. 17 (4) (2018) 2381–2413.
2671–2685. [353] Q. Wang, N. Ripamonti, J.S. Hesthaven, Recurrent neural network closure of
[320] J. Kou, W. Zhang, A hybrid reduced-order framework for complex aeroelastic parametric pod-galerkin reduced-order models based on the mori-zwanzig
simulations, Aero. Sci. Technol. 84 (2019) 880–894. formalism, J. Comput. Phys. 410 (2020) 109402.
[321] Z. Yang, R. Huang, H. Liu, Y. Zhao, H. Hu, An improved nonlinear reduced-order [354] S.E. Ahmed, O. San, A. Rasheed, T. Iliescu, A long short-term memory embedding
modeling for transonic aeroelastic systems, J. Fluid Struct. 94 (2020) 102926. for hybrid uplifted reduced order models, Phys. Nonlinear Phenom. 409 (2020).
[322] S. Arlot, A. Celisse, A survey of cross-validation procedures for model selection, [355] S. Hijazi, G. Stabile, A. Mola, G. Rozza, Data-driven pod-galerkin reduced order
Stat. Surv. 4 (2010) 40–79, 0. model for turbulent flows, J. Comput. Phys. 416 (2020) 109513.
[323] A. Omran, B. Newman, Full envelope nonlinear parameter-varying model [356] D. Amsallem, C. Farhat, Interpolation method for adapting reduced-order models
approach for atmospheric flight dynamics, J. Guid. Contr. Dynam. 35 (1) (2012) and application to aeroelasticity, AIAA J. 46 (7) (2008) 1803–1813.
270–283. [357] R. Zimmermann, B. Peherstorfer, K. Willcox, Geometric subspace updates with
[324] H. Liu, R. Huang, Y. Zhao, H. Hu, Reduced-order modeling of unsteady applications to online adaptive nonlinear model reduction, SIAM J. Matrix Anal.
aerodynamics for an elastic wing with control surfaces, J. Aerospace Eng. 30 (3) Appl. 39 (1) (2018) 234–261.
(2017), 04016083. [358] J.C. Loiseau, S.L. Brunton, B.R. Noack, From the pod-galerkin method to sparse
[325] M. Winter, C. Breitsamter, Nonlinear identification via connected neural networks manifold models, Handbook of Model-Order Reduction (2) (2019) 1–47.
for unsteady aerodynamic analysis, Aero. Sci. Technol. 77 (2018) 802–818. [359] Z. Chen, Y. Zhao, R. Huang, Parametric reduced-order modeling of unsteady
[326] J. Wu, C.K. Chan, Prediction of hourly solar radiation using a novel hybrid model aerodynamics for hypersonic vehicles, Aero. Sci. Technol. 87 (2019) 1–14.
of arma and tdnn, Sol. Energy 85 (5) (2011) 808–817. [360] M. Oulghelou, C. Allery, Non intrusive method for parametric model order
[327] S.F. Masri, A hybrid parametric/nonparametric approach for the identification of reduction using a bi-calibrated interpolation on the grassmann manifold,
nonlinear systems, Probabilist. Eng. Mech. 9 (1–2) (1994) 47–57. J. Comput. Phys. 426 (2020), 109924.
[328] S.A. Billings, H.L. Wei, The wavelet-narmax representation: a hybrid model [361] T. Bui-Thanh, K. Willcox, O. Ghattas, Parametric reduced-order models for
structure combining polynomial models with multiresolution wavelet probabilistic analysis of unsteady aerodynamic applications, AIAA J. 46 (10)
decompositions, Int. J. Syst. Sci. 36 (3) (2005) 137–152. (2008) 2520–2529.
[329] G.P. Zhang, Time series forecasting using a hybrid arima and neural network [362] J. Degroote, J. Vierendeels, K.E. Willcox, Interpolation among reduced-order
model, Neurocomputing 50 (1) (2003) 159–175. matrices to obtain parameterized models for design, optimization and
[330] J. Kou, W. Zhang, Z. Ye, Dynamic nonlinear aerodynamics modeling method probabilistic analysis, Int. J. Numer. Methods Fluid. 63 (2010) 207–230.
based on layered model, Acta Aeronautica Astronautica Sinica 36 (12) (2015) [363] N.T. Son, A real time procedure for affinely dependent parametric model order
3785–3797. reduction using interpolation on grassmann manifolds, Int. J. Numer. Methods
[331] A. Mannarino, E.H. Dowell, P. Mantegazza, An adaptive controller for nonlinear Eng. 93 (8) (2013) 818–833.
flutter suppression and free-play compensation, J. Vib. Contr. 23 (14) (2017) [364] E. de Sturler, S. Gugercin, M.E. Kilmer, S. Chaturantabut, C. Beattie,
2269–2290. M. O’Connell, Nonlinear parametric inversion using interpolatory model
[332] R. Huang, H. Liu, Z. Yang, Y. Zhao, H. Hu, Nonlinear reduced-order models for reduction, SIAM J. Sci. Comput. 37 (3) (2015) B495–B517.
transonic aeroelastic and aeroservoelastic problems, AIAA J. 56 (9) (2018) [365] D. Amsallem, M.J. Zahr, C. Farhat, Nonlinear model order reduction based on
3718–3731. local reduced-order bases, Int. J. Numer. Methods Eng. 92 (10) (2012) 891–916.
[333] H. Liu, X. Gao, Identification of nonlinear aerodynamic systems with application [366] Z. Zhan, W.G. Habashi, M. Fossati, Local reduced-order modeling and iterative
to transonic aeroelasticity of aircraft structures, Nonlinear Dynam. 100 (2) (2020) sampling for parametric analyses of aero-icing problems, AIAA J. 53 (8) (2015)
1037–1056. 2174–2185.
[334] D. Rempfer, H.F. Fasel, Evolution of three-dimensional coherent structures in a [367] B. Peherstorfer, D. Butnaru, K. Willcox, H.-J. Bungartz, Localized discrete
flat-plate boundary layer, J. Fluid Mech. 260 (1994) 351–375. empirical interpolation method, SIAM J. Sci. Comput. 36 (1) (2014) A168–A192.
[335] B.R. Noack, M. Morzynski, G. Tadmor, Reduced-order Modelling for Flow [368] S.L. Brunton, J.H. Tu, I. Bright, J.N. Kutz, Compressive sensing and low-rank
Control, Springer, 2011. libraries for classification of bifurcation regimes in nonlinear dynamical systems,
SIAM J. Appl. Dyn. Syst. 13 (4) (2014) 1716–1732.

34
J. Kou and W. Zhang Progress in Aerospace Sciences 125 (2021) 100725

[369] H. Babaee, T. Sapsis, A minimization principle for the description of modes [401] B. Li, Z. Yang, X. Zhang, G. He, B.-Q. Deng, L. Shen, Using machine learning to
associated with finite-time instabilities, P. Roy. Soc. A-Math. Phy. 472 (2016) detect the turbulent region in flow past a circular cylinder, J. Fluid Mech. 905
20150779. (2020).
[370] A. Quarteroni, A. Manzoni, F. Negri, Reduced Basis Methods for Partial [402] W. Yao, S. Marques, Nonlinear aerodynamic and aeroelastic model reduction
Differential Equations: an Introduction, Springer, 2015. using a discrete empirical interpolation method, AIAA J. 55 (2) (2017) 1–14.
[371] J.S. Hesthaven, G. Gozza, B. Stamm, Certified Reduced Basis Methods for [403] F. Sung, Y. Yang, L. Zhang, T. Xiang, P. H. S. Torr, T. M. Hospedales, Learning to
Parametrized Partial Differential Equations, Springer, 2016. compare: relation network for few-shot learning, In: Proceedings of the IEEE
[372] F. Chinesta, P. Ladevèze, E. Cueto, A short review on model order reduction based Conference on Computer Vision and Pattern Recognition (CVPR). 1199-1208.
on proper generalized decomposition, Arch. Comput. Methods Eng. 18 (4) (2011) [404] B. Peherstorfer, K. Willcox, Dynamic data-driven model reduction: adapting
395. reduced models from incomplete data, Advanced Modeling and Simulation in
[373] S.L. Brunton, J.L. Proctor, J.N. Kutz, Discovering governing equations from data Engineering Sciences 3 (1) (2016).
by sparse identification of nonlinear dynamical systems, Proc. Natl. Acad. Sci. [405] J. Ling, A. Kurzawski, J. Templeton, Reynolds averaged turbulence modelling
Unit. States Am. 113 (5) (2016) 3932–3937. using deep neural networks with embedded invariance, J. Fluid Mech. 807 (2016)
[374] C. Xie, J. Wang, H. Li, M. Wan, S. Chen, Artificial neural network mixed model for 155–166.
large eddy simulation of compressible isotropic turbulence, Phys. Fluids 31 (8) [406] A.P. Singh, S. Medida, K. Duraisamy, Machine-learning-augmented predictive
(2019). modeling of turbulent separated flows over airfoils, AIAA J. 55 (7) (2017)
[375] Y. Zhao, H.D. Akolekar, J. Weatheritt, V. Michelassi, R.D. Sandberg, Rans 2215–2227.
turbulence model development using cfd-driven machine learning, J. Comput. [407] M. Raissi, P. Perdikaris, G.E. Karniadakis, Physics-informed neural networks: a
Phys. 411 (2020). deep learning framework for solving forward and inverse problems involving
[376] E. Kaiser, B.R. Noack, L. Cordier, A. Spohn, M. Segond, M. Abel, G. Daviller, nonlinear partial differential equations, J. Comput. Phys. 378 (2019) 686–707.
J. Östh, S. Krajnović, R.K. Niven, Cluster-based reduced-order modeling of a [408] X. Jin, S. Cai, H. Li, G.E. Karniadakis, Nsfnets (Navier-Stokes flow nets): physics-
mixing layer, J. Fluid Mech. 754 (2014) 365–414. informed neural networks for the incompressible Navier-Stokes equations,
[377] H. Li, D. Fernex, R. Semaan, J. Tan, M. Morzyński, B.R. Noack, Cluster-based J. Comput. Phys. 426 (2021).
network model, J. Fluid Mech. 906 (2021). [409] M. Yin, X. Zheng, J.D. Humphrey, G. Em Karniadakis, Non-invasive inference of
[378] M. Mifsud, Reduced-order Modelling for High-Speed Aerial Weapon thrombus material properties with physics-informed neural networks, Comput.
Aerodynamics, Cranfield University, 2008. Methods Appl. Mech. Eng. 375 (2021).
[379] T. Franz, R. Zimmermann, S. Görtz, N. Karcher, Interpolation-based reduced- [410] C. Rao, H. Sun, Y. Liu, Physics-informed deep learning for incompressible laminar
order modelling for steady transonic flows via manifold learning, Int. J. Comput. flows, Theoretical and Applied Mechanics Letters 10 (3) (2020) 207–212.
Fluid Dynam. 28 (3–4) (2014) 106–121. [411] M. Raissi, P. Perdikaris, G.E. Karniadakis, Inferring solutions of differential
[380] A. Ehlert, C.N. Nayeri, M. Morzynski, B.R. Noack, Locally Linear Embedding for equations using noisy multi-fidelity data, J. Comput. Phys. 335 (2017) 736–746.
Transient Cylinder Wakes, Arxiv preprint, 2019. arXiv:1906.07822. [412] L. Sun, J.-X. Wang, Physics-constrained bayesian neural network for fluid flow
[381] A.G. Nair, C.-A. Yeh, E. Kaiser, B.R. Noack, S.L. Brunton, K. Taira, Cluster-based reconstruction with sparse and noisy data, Theoretical and Applied Mechanics
feedback control of turbulent post-stall separated flows, J. Fluid Mech. 875 Letters 10 (3) (2020) 161–169.
(2019) 345–375. [413] M. Raissi, A. Yazdani, G.E. Karniadakis, Hidden fluid mechanics: learning velocity
[382] X. Zhao, L. Du, X. Peng, Z. Deng, W. Zhang, Research on refined reconstruction and pressure fields from flow visualizations, Science 367 (6481) (2020)
method of airfoil pressure based on compressed sensing, Theoretical and Applied 1026–1030.
Mechanics Letters (2021), 100223. [414] E.H. Dowell, M. Ilgamov, Studies in Nonlinear Aeroelasicity, Springer, New York,
[383] J.C. Loiseau, B.R. Noack, S.L. Brunton, Sparse reduced-order modelling: sensor- 1988.
based dynamics to full-state estimation, J. Fluid Mech. 844 (2018) 459–490. [415] R.H. Landon, Naca0012 Oscillatory and Transient Pitching, 1982.
[384] K. Taira, A.G. Nair, S.L. Brunton, Network structure of two-dimensional decaying [416] K. Washizu, A. Ohya, Y. Otsuki, K. Fujii, Aeroelastic instability of rectangular
isotropic turbulence, J. Fluid Mech. 795 (2016). cylinders in a heaving mode, J. Sound Vib. 59 (2) (1978) 195–210.
[385] A.G. Nair, S.L. Brunton, K. Taira, Networked-oscillator-based modeling and [417] D. Tang, E.H. Dowell, Experimental and theoretical study on aeroelastic response
control of unsteady wake flows, Phys. Rev. E 97 (6–1) (2018), 063107. of high-aspect-ratio wings, AIAA J. 39 (8) (2001) 1430–1441.
[386] M. Milano, P. Koumoutsakos, Neural network modeling for near wall turbulent [418] W.M. Zhang, Y.J. Ge, Flutter mode transition of a double-main-span suspension
flow, J. Comput. Phys. 182 (1) (2002) 1–26. bridge in full aeroelastic model testing, J. Bridge Eng. 19 (7) (2014), 06014004.
[387] T. Murata, K. Fukami, K. Fukagata, Nonlinear mode decomposition with [419] W. Zhang, C. Gao, Y. Liu, Z. Ye, Y. Jiang, The interaction between flutter and
convolutional neural networks for fluid dynamics, J. Fluid Mech. 882 (2020) A13. buffet in transonic flow, Nonlinear Dynam. 82 (4) (2015) 1851–1865.
[388] K. Lee, K.T. Carlberg, Model reduction of dynamical systems on nonlinear [420] C.H.K. Williamson, Vortex dynamics in the cylinder wake, Annu. Rev. Fluid Mech.
manifolds using deep convolutional autoencoders, J. Comput. Phys. 404 (2020) 28 (1) (1996) 477–539.
108973. [421] C. Cossu, L. Morino, On the instability of a spring-mounted circular cylinder in a
[389] Y. Zhu, N. Zabaras, Bayesian deep convolutional encoder–decoder networks for viscous flow at low Reynolds numbers, J. Fluid Struct. 14 (2) (2000) 183–196.
surrogate modeling and uncertainty quantification, J. Comput. Phys. 366 (2018) [422] S. Mittal, S. Singh, Vortex-induced vibrations at subcritical re, J. Fluid Mech. 534
415–447. (2005) 185–194.
[390] N. Takeishi, Y. Kawahara, T. Yairi, Learning koopman invariant subspaces for [423] E. Buffoni, Vortex shedding in subcritical conditions, Phys. Fluids 15 (3) (2003)
dynamic mode decomposition, in: Advances in Neural Information Processing 814.
Systems, vol. 30, NIPS, 2017, pp. 1130–1140. [424] S.-S. Chen, R.-H. Yen, A.-B. Wang, Investigation of the resonant phenomenon of
[391] B. Lusch, J.N. Kutz, S.L. Brunton, Deep learning for universal linear embeddings flow around a vibrating cylinder in a subcritical regime, Phys. Fluids 23 (1)
of nonlinear dynamics, Nat. Commun. 9 (2018) 4950. (2011), 014105.
[392] S.R. Otto, C.W. Rowley, Linearly recurrent autoencoder networks for learning [425] C.H.K. Williamson, R. Govardhan, Vortex-induced vibrations, Annu. Rev. Fluid
dynamics, SIAM J. Appl. Dyn. Syst. 18 (1) (2019) 558–593. Mech. 36 (1) (2004) 413–455.
[393] S. Pan, K. Duraisamy, Physics-informed probabilistic learning of linear [426] S. Mittal Navrose, Lock-in in vortex-induced vibration, J. Fluid Mech. 794 (2016)
embeddings of nonlinear dynamics with guaranteed stability, SIAM J. Appl. Dyn. 565–594.
Syst. 19 (1) (2020) 480–509. [427] P. Meliga, J.M. Chomaz, An asymptotic expansion for the vortex-induced
[394] D. D’Agostino, A. Serani, E.F. Campana, M. Diez, Deep autoencoder for off-line vibrations of a circular cylinder, J. Fluid Mech. 671 (2011) 137–167.
design-space dimensionality reduction in shape optimization, in: AIAA/ASCE/ [428] J. Kou, W. Zhang, Y. Liu, X. Li, The lowest Reynolds number of vortex-induced
AHS/ASC Structures, Structural Dynamics, and Materials Conference, AIAA, vibrations, Phys. Fluids 29 (4) (2017).
Kissimmee, Florida, 2018, 2018-1648. [429] N.F. Giannelis, G.A. Vio, O. Levinski, A review of recent developments in the
[395] V. Sekar, Q. Jiang, C. Shu, B.C. Khoo, Fast flow field prediction over airfoils using understanding of transonic shock buffet, Prog. Aero. Sci. 92 (2017) 39–84.
deep learning approach, Phys. Fluids 31 (5) (2019). [430] B.H.K. Lee, Self-sustained shock oscillations on airfoils at transonic speeds, Prog.
[396] R. Han, Y. Wang, Y. Zhang, G. Chen, A novel spatial-temporal prediction method Aero. Sci. 37 (2) (2001) 147–196.
for unsteady wake flows based on hybrid deep neural network, Phys. Fluids 31 [431] C. Gao, W. Zhang, Numerical simulation and modal analysis of transonic buffet
(12) (2019). flow over wings, Acta Aeronautica Astronautica Sinica 40 (7) (2019), 122597-
[397] S.R. Bukka, R. Gupta, A.R. Magee, R.K. Jaiman, Assessment of unsteady flow 122597.
predictions using hybrid deep learning based reduced-order models, Phys. Fluids [432] M. Iovnovich, D.E. Raveh, Numerical study of shock buffet on three-dimensional
33 (1) (2021). wings, AIAA J. 53 (2) (2015) 449–463.
[398] A. Lozano-Durán, H.J. Bae, M.P. Encinar, Causality of energy-containing eddies in [433] Fouad Giri, Er-Wei Bai. Block-oriented Nonlinear System Identification, Springer,
wall turbulence, J. Fluid Mech. 882 (2019). London, 2010.
[399] B. Colvert, M. Alsalman, E. Kanso, Classifying vortex wakes using neural [434] Benjamin S. Duran, Patrick L. Odell. Cluster Analysis: a Survey, Springer Science
networks, Bioinspiration Biomimetics 13 (2) (2018), 025003. & Business Media, 2013.
[400] Z. Wu, J. Lee, C. Meneveau, T. Zaki, Application of a self-organizing map to [435] Yunqian Ma, Yun Fu. Manifold Learning Theory and Applications, Boca Raton:
identify the turbulent-boundary-layer interface in a transitional flow, Phys. Rev. CRC press, 2012.
Fluids 4 (2) (2019). [436] Mukund Balasubramanian, Eric L. Schwartz, Joshua B. Tenenbaum, Vin de Silva,
John C. Langford, The isomap algorithm and topological stability, Science 295
(5552) (2002) 7.

35