AIAA JOURNAL

Vol. 62, No. 5, May 2024

Data-Driven Bayesian Inference for Stochastic Model Identification

of Nonlinear Aeroelastic Systems

Michael McGurk∗
University of Strathclyde, Glasgow, Scotland G1 1XJ, United Kingdom
Adolphus Lye†
National University of Singapore, Singapore S138602, Republic of Singapore
Ludovic Renson‡
Imperial College London, London, England SW7 2AZ, United Kingdom
and
Jie Yuan§
University of Southampton, Southampton, England SO17 1BF, United Kingdom
Downloaded by Northwestern Polytechnic University on April 29, 2024 | http://arc.aiaa.org | DOI: 10.2514/1.J063611

https://doi.org/10.2514/1.J063611
The objective of this work is to propose a data-driven Bayesian inference framework to efficiently identify
parameters and select models of nonlinear aeroelastic systems. The framework consists of the use of Bayesian
theory together with advanced kriging surrogate models to effectively represent the limit cycle oscillation response
of nonlinear aeroelastic systems. Three types of sampling methods, namely, Markov chain Monte Carlo, transitional
Markov chain Monte Carlo, and the sequential Monte Carlo sampler, are implemented into Bayesian model
updating. The framework has been demonstrated using a nonlinear wing flutter test rig. It is modeled by a two-
degree-of-freedom aeroelastic system and solved by the harmonic balance methods. The experimental data of the
flutter wing is obtained using control-based continuation techniques. The proposed methodology provided up to a
20% improvement in accuracy compared to conventional deterministic methods and significantly increased
computational efficiency in the updating and uncertainty quantification processes. Transitional Markov chain
Monte Carlo was identified as the optimal choice of sampling method for stochastic model identification. In
selecting alternative nonlinear models, multimodal solutions were identified that provided a closer representation
of the physical behavior of the complex aeroelastic system than a single solution.

Nomenclature x = system degrees of freedom

A = aerodynamic inertial matrix ρ = air density
a, b, c = aerodynamic constants
B = aerodynamic damping matrix
C = aerodynamic stiffness matrix I. Introduction
c = chord
ch
cα
=
=
heave damping coefficient
pitch damping coefficient A PRIMARY goal in the aerospace industry at present is to reduce
CO2 emissions. The use of lightweight materials and integrated
electric propulsion systems is considered crucial in enhancing the
D = structural damping matrix
Iα = wing moment of inertia about elastic axis efficiency and capabilities of aerospace systems. Modeling the
K = structural stiffness matrix dynamic response of lightweight aerospace structures is a challeng-
Kh = heave linear stiffness ing task because of their significant geometrical deformation, which
Kα = pitch linear stiffness introduces distributed nonlinearities into structures, resulting in com-
Kα2 = pitch second-order nonlinear stiffness plex nonlinear dynamic behavior [1,2]. Furthermore, with increasing
Kα3 = pitch third-order nonlinear stiffness integrated systems, new interfaces can introduce localized nonlinear-
Kα5 = pitch fifth-order nonlinear stiffness ities, as has been observed in tiltrotor systems dynamic responses
M = structural mass matrix [3,4]. The effects of these nonlinearities on dynamics and control can
mT = mass of wing and support be very significant, such as changing the flutter boundary in tiltrotor
mw = mass of wing systems [3,5] and shifting the aerodynamic center of certain wings,
t = time affecting control of the aircraft [6].
V = airspeed Due to nonlinearities in aeroelastic systems, subcritical postflutter
responses may exist [7,8], leading to a self-sustaining oscillation
occurring at velocities below the linear flutter velocity. These oscil-
Received 29 September 2023; revision received 19 December 2023; lations are referred to as limit cycle oscillations (LCOs) and cannot be
accepted for publication 14 January 2024; published online 15 March 2024.
estimated from purely linear analysis. These subcritical behaviors
Copyright © 2024 by the American Institute of Aeronautics and Astronautics,
Inc. All rights reserved. All requests for copying and permission to reprint have been observed experimentally in high-aspect-ratio flexible
should be submitted to CCC at www.copyright.com; employ the eISSN 1533- wings [9,10]. Recently, Drachinsky and Raveh [11,12] also discov-
385X to initiate your request. See also AIAA Rights and Permissions ered subcritical LCO formation in experimental testing of a highly
www.aiaa.org/randp. flexible wing in low-speed flutter experiments. However, the effects
*Ph.D. Candidate, Aerospace Centre of Excellence; michael.mcgurk@ of nonlinearities on aeroelasticity are often neglected in computa-
strath.ac.uk (Corresponding Author). tional analysis due to the added complexity and high computational
†
Research Fellow, Singapore Nuclear Research and Safety Initiative; cost during the design process, which greatly limits the design space
[email protected].
‡
Senior Lecturer, Dynamics Group, South Kensington Campus; l.renson@ of these aerospace systems. While LCO detected from nonlinear
imperial.ac.uk. analysis can be stable, in a real system, they could cause structural
§
Assistant Professor, Computational Engineering Design Group; j.yuan@ damage. It is therefore critical to validate mathematical models
soton.ac.uk (Co-corresponding Author). that accurately capture system’s nonlinear response. With reliable
1889
 1890 MCGURK ET AL.

mathematical models, the need for costly exterminates under differ- in the mathematical model that is selected or errors from experimental
ent flow conditions would not be required for certification. results. When using deterministic model updating methods, it is
This work proposes a probabilistic data-driven methodology to common to face issues with parameter nonidentifiability, as certain
minimize computational burden and maximize accuracy in identify- combinations of nonlinear parameters can lead to comparable results
ing nonlinear models to capture the LCO behavior of aeroelastic [33,34]. Recent deterministic approaches [35–37] have shown prom-
systems. Kriging models are developed with databases generated ise in determining aeroelastic LCO, but issues have been highlighted.
from nonlinear aeroelastic systems using harmonic balance method- In using universal approximates to replicate aeroelastic bifurcation
based bifurcation analysis. LCO experimental data are obtained from diagrams, Beregi et al. [19] found cases of overfitting data. This
a uniform wing test rig using control-based continuation (CBC) problem is often encountered when using deterministic approaches
testing methods [13]. Bayesian inference is then implemented with inherently noisy data.
between a kriging surrogate model and experimental data to both Considering the computational burden in LCO simulations, little
estimate probabilistic parameters of the nonlinear model and rank the work in this area related to model updating and UQ has been con-
evidence supporting model selection. In addition, the performance of ducted [38]. The formation of LCO occurs following a Hopf bifur-
three sampling techniques is effectively compared when stochasti- cation that can be detected with eigenvalue analysis of underlying
cally updating a nonlinear aeroelastic system for the first time. The linear systems [39,40]. It has been shown through both theory and
implementation of a Bayesian inference in conjunction with a surro- experimentation that Hopf bifurcations occur at the flutter points of
gate model to anticipate LCO behavior is a distinctive strategy in the aeroelastic systems [8,41]. Following a Hopf bifurcation, LCO char-
area of aeroelasticity. acteristics are tracked through numerical continuation to produce
Downloaded by Northwestern Polytechnic University on April 29, 2024 | http://arc.aiaa.org | DOI: 10.2514/1.J063611

Model updating has been widely used to calibrate the parameters iterative bifurcation diagrams. Recent work [42–44] has seen a focus
of mathematical or computation models based on experimental data on frequency domain techniques, namely, the harmonic balance
in the last few decades. A majority of existing techniques provide methods (HBMs). In approximating periodic motion with Fourier
deterministic estimates of parameters, such as least-squares minimi- series coefficients, HBMs provide computationally efficient estima-
zation, sensitivity-based model updating, and Levenberg–Marquardt tions in identifying the maximum response of LCO [45]. The use of
algorithm [14–17]. While direct deterministic methods are very HBM with continuation techniques is not essential for the proposed
effective for linear systems, they are subject to high computational framework but effectively reduces the computational burden of gath-
expense for complex and high-dimensional dynamic systems and ering training data compared to other numerical continuation alter-
have a very high sensitivity to the noise in the experimental data [18]. natives [45].
For nonlinear aeroelastic systems, Beregi employed a classical model Furthermore, the use of fast-running surrogate models can effec-
update approach to estimate nonlinear parameters in a subcritical tively reduce the high computational cost of potentially thousands of
airfoil test case using normal form theory. It was concluded that the nonlinear aeroelastic simulations required with BMU. Nonlinear
accuracy of the bifurcation diagram should be improved to account prediction can be achieved effectively through the use of data-driven
for different uncertainties in the experimental data [19,20]. Recently, methods such as polynomial chaos expansion (PCE) and neural
researchers have acknowledged the importance of quantifying uncer- networks [46–49]. PCE methods have been proposed specifically
tainties in the nonlinear behavior of systems, shifting the determin- for UQ in aeroelastic cases in [50]. With surrogate models based on
istic model updating approach to probabilistic ones [18,21]. orthogonal polynomials, however, the amount of training data
Bayesian model updating (BMU), first put forward in [22], is required has shown to increase factorially with the number of design
regarded as one of the most promising probabilistic nonlinear param- parameters [51]. Kernel-based surrogate models, namely, kriging
eter estimation techniques [23–25]. A significant advantage of BMU (Gaussian process [GP] regression), have been employed for cases
when compared to other probabilistic model updating techniques, with limited training data [52,53]. Kriging models have been found to
such as maximum likelihood estimation, is its capacity to seamlessly be remarkably effective in modeling complex nonlinear systems,
integrate prior knowledge [26]. With the Bayesian inference, one can with the added benefit of requiring only minimal training data [54].
assess the posterior probability density functions (PDFs) by utilizing A major benefit of kriging is the ability to seamlessly integrate new
the given measurement data. The posterior PDFs are instrumental in observations into the current model framework [55,56]. Kriging
providing numerical distributions for parameter estimations, along surrogate models have been employed to propagate uncertainty in
with their confidence bands [27]. The efficiency of the stochastic bifurcation diagrams of landing gear designs by Tartaruga et al. [57]
model updating is highly dependent on the quality of the sampling and ring damper designs by Sun et al. [54]. Lee et al. [58] developed a
methods. Advanced iterative sampling techniques are preferred due physics-informed machine-learnt model to improve the prediction of
to their ability to efficiently explore high-dimensional and correlated the bifurcation diagrams for an airfoil test case. It showed that
parameter spaces, handle nonstandard posterior shapes, facilitate deterministic estimation may lead to an overfitting problem when
iterative improvement, support model comparison, and provide reli- taking noisy data into consideration. The work also suggested that the
able uncertainty quantification (UQ). The most common methods are use of nonlinear solvers, specifically those suited for periodic sol-
Metropolis–Hastings Markov chain Monte Carlo (MCMC), transi- utions, could enhance the training algorithms.
tional chain Monte Carlo (TMCMC), and the sequential Monte Carlo When identifying a system based on experimental data, the two
sampler (SMC) [28]. Each method provides case-dependent benefits, key factors to consider are the quality of the data and the amount of it.
and experiments are often required to select an appropriate sampling In classical wind tunnel testing, only stable LCOs are detected,
method. Comparisons between sampling methods have been carried leaving a gap in unstable sections of bifurcation diagrams. CBC, first
out partially in the case of structural health monitoring [29–31]. A proposed by Sieber et al. [59], was introduced to perform bifurcation
direct comparison does not exist for an aeroelastic test case. An analysis directly in experiments. This is possible by combining New-
advantage of TMCMC is that the evidence function can be calculated ton iterations and feedback control embedded into the continuation,
as a byproduct of model selection. Song et al. utilized TMCMC in allowing for the detection of unstable LCO [60]. The capability of
successfully selecting a nonlinear model for a wing engine struc- CBC has been demonstrated on several mechanical systems, namely,
ture [32]. by Lee on a two-degree-of-freedom airfoil [61]. LCO amplitude and
Nonlinear system identification (NSI) is very challenging in aero- frequency for unstable LCO were gathered, presenting high-quality
elasticity. Firstly, the computational expense of nonlinear aeroelastic data for nonlinear aeroelastic system identification that will be used
simulations is highly demanding even for low-degree-of-freedom in this work.
systems, especially when using time-domain solvers. Secondly, This paper will first present the overall probabilistic model updat-
experimental testing of nonlinear autonomous systems is challenging ing methodology for a general nonlinear aeroelastic system. It will
due to the limitations of most linear experimental methods [13]. include the process of gathering training data from HBM continu-
Thirdly, the form of nonlinear functions can be uncertain and usually ation, the development of kriging-based surrogate models, and the
lacks explicit expressions, which can make it unclear if inaccuracies implementation of the Bayesian inference with MCMC, TMCMC,
between the experimental data and numerical data are down to errors and SMC sampling methods. The specific objectives behind the use
 MCGURK ET AL. 1891

of different MCMC-based sampling techniques are to provide a odology is divided into two parts: the development of a surrogate
comparative study on sampling performance and its robustness model describing the LCO behavior of Eq. (1) and the Bayesian
toward parameter identification and model updating for a highly model updating with forward UQ. The surrogate model developed in
nonlinear aeroelastic model under limited data. The three goals of Fig. 1a is created a priori and left unmodified during the model
the proposed methodology are 1) to provide distributions that updating process. With inputs gathered using Latin hypercube sam-
numerically estimate nonlinear parameters; 2) to produce bifurcation pling (LHS), HBM continuation runs are performed following a Hopf
plots with confidence bands; and 3) to rank different forms of non- bifurcation found in the underlying linear model. Data from HBM
linear models. In the following section, the mathematical model and continuation runs are used to construct surrogate models through
experimentation describing a simplified nonlinear flutter test rig will kriging. The second stage of the methodology in Fig. 1b describes the
be described. In the results section, nonlinear parameter estimations BMU process applied to the surrogate model to stochastically esti-
using the proposed methodology will be first presented, followed by a mate nonlinear parameters. Each step in both processes will be
discussion of the suitability of the nonlinear model and the appro- discussed as follows.
priate choice for the sampling method. Then, we will describe the
improvement of using alternative nonlinear models through Bayesian
A. Equation of Motion
model selection.
A general nonlinear aeroelastic system will be formulated by the
following:
II. Methodology
Downloaded by Northwestern Polytechnic University on April 29, 2024 | http://arc.aiaa.org | DOI: 10.2514/1.J063611

In this section, the methodology of the data-driven probabilistic

modeling identification framework for a general nonlinear aeroelas- Mx
Dx_
Kx
qnl fnl  Aρair x
Bρair Vx_
Cρair V 2 x
tic system will be presented. As shown in Fig. 1, the general meth- (1)

Fig. 1 The overall methodology of probabilistic identification framework for nonlinear aeroelastic systems.
 1892 MCGURK ET AL.

where x denotes the system’s degrees of freedom, and M, D, and K will be briefly introduced, and then the construction of kriging
are the structural mass, damping, and stiffness matrices, respectively. surrogate models will be described.
Matrices A, B, and C are aerodynamic matrices that are dependent on
air density and freestream velocity. All matrices are size of N × N, 1. HBM-Based Numerical Continuation
where N is the number of degrees of freedom of the system. The The computation of LCO behavior is carried out through HBM
nonlinear function fnl is used to represent different types of non- coupled with continuation techniques as formulated in [45]. It is first
linearities encountered in aeroelastic systems. The N × 1 vector qnl is required that the standard differential equation from Eq. (1) is
utilized to implement the nonlinear equations in the degrees of free- rearranged into first-order state Eq. (2) as
dom they impact.
The general nonlinear behavior of a subcritical aeroelastic system x_  Qx
qn fnl (2)
is laid out in Fig. 2. In the bifurcation diagram in Fig. 2b, a stable
solution exists until a Hopf bifurcation is detected at 17.8 m∕s. At where
velocities above this point, the response of the underlying linear
system to a perturbation is negatively damped. This can be deter- x_
x
mined without considering nonlinearities using standard eigenvalue x
analysis. Taking nonlinear effects into consideration, numerical
continuation is carried out iteratively, starting at the bifurcation M − ρA−1 ρVB − D M − ρA−1 ρV 2 C − K
Q
point. Unstable LCOs are identified at velocities below the bifurca- 0N×N I N×N
Downloaded by Northwestern Polytechnic University on April 29, 2024 | http://arc.aiaa.org | DOI: 10.2514/1.J063611

tion point and tracked until a turning point. Following the turning
point, stable LCOs are tracked. This describes a system where two −M − ρA−1 qnl
qn 
possible solutions exist between the turning point and the Hopf 0N×1
bifurcation point: a stable, damped solution and an LCO. Physi-
cally, unstable LCOs represent a nonconverging trajectory that where Q will be referred to as the linear matrix as it fully captures the
exhibits instability over time. This means that, following nonlinear linear behavior of the system. Flutter velocity can be found at the
aeroelastic analysis, the safe operating velocity of the system in Hopf bifurcation point considering only the linear part of the system.
Fig. 2 is shifted from 17.8 to 13.7 m∕s at the turning point of the Equation (2) is written as the eigenvalue problem x_ − Qx  0.
bifurcation diagram. Assuming an oscillatory response x  xo eψt , the eigenvalue problem
can be written as follows:
B. Development of Surrogate Models
As is observed in Fig. 2b, there is a unique velocity V solution for Q − Iψ i ϕ  0 (3)
each LCO amplitude point h. It is possible that the kriging surrogate
model fh; θ can be set up as V  fh; θ where θ represents the where ψ i are eigenvalues in the conjugate pair:
vector of nonlinear parameters. In this subsection, the development of
the kriging surrogate model will be detailed. The training data for the ψ i  −ζi ωi  iωi 1 − ζi (4)
surrogate model is first generated from deterministic simulations of
nonlinear aeroelastic systems using HBM with continuation tech- where i  1; : : : ; 2N, and ωi represents damped natural frequencies
niques. The methodology of HBM-based deterministic simulation and ζ i the damping ratios. Matrix ϕ contains the corresponding

a) Three-dimensional subcritical bifurcation diagram

b) Two-dimensional subcritical bifurcation diagram ( stable c) LCO time histories (14.5m/s ) (20m/s )
solution) ( unstable solution) ( hopf bifurcation) (
turning point)
Fig. 2 Subcritical bifurcation diagram with Hopf bifurcation at 17.8 m∕s ( , stable LCO; , unstable LCO).
 MCGURK ET AL. 1893

eigenvectors. Flutter is characterized by unstable negatively damped where h is a vector representing the separation in multiple dimensions
oscillations. From this definition it can be determined that if any of between x and X. Variance σ, smoothness parameter υ and range a are
the real parts of Eq. (4) is positive, the system is dynamically unstable determined through the optimization process described in [67]. A Bessel
[62]. If it is assumed that, following a Hopf bifurcation, the dynamic function is represented by K υ and Γυ is a generalization of the factorial
response of the system is an LCO, the time response of x and x_ can be function to noninteger values [68].
represented using the Fourier series. The unknowns can be expressed To begin the construction of the surrogate model, we first extract
by multiharmonic response and solved in the frequency domain: LHS samples from the uniform distributions for θ based on the initial
approximations of the nonlinear parameters. Every combination
l undergoes a HBM continuation run. The data obtained is then
xt  X0
X k;s sin kωt
X k;c cos kωt (5) segregated into individual data points that comprise velocity, ampli-
k1 tude, and nonlinear parameters. These points are preserved as train-
ing data and utilized to fuel the kriging algorithm. The accuracy of the
where l represents the number of Fourier modes, and X0 , Xk;s , and surrogate model is then validated using a set of separate data. This
Xk;c are Fourier coefficients. This assumed response is central to process is repeated until accuracy converges.
HBM, converting the system from the time domain to the frequency
domain. The number of unknowns will become 2 × l
1 × N. In C. Theory of Bayesian Inference and Sampling Techniques
addition to a set of Fourier coefficients, there is also the natural
The Bayesian inference is implemented to estimate posterior dis-
frequency of the system ω and the continuation parameter that can
tributions of nonlinear parameters from a prior distribution using
Downloaded by Northwestern Polytechnic University on April 29, 2024 | http://arc.aiaa.org | DOI: 10.2514/1.J063611

be any structural or aerodynamic parameter. In this case, velocity is

input data and a likelihood function including experimental data.
the continuation parameter. As is common in continuation, ortho-
Posterior distributions of the nonlinear parameters θ are estimated
gonality between a response and its differential is assumed to place an
with
additional value on the scheme [63].
Nonlinear force functions are usually modeled in the time domain. PVjθ; f ⋅ Pθjf
Since they are not linear functions of states or explicit functions of time, PθjV; f  (9)
it is not possible to transform directly to the frequency domain [64]. PVjf
The nonlinear force response is converted to the frequency domain
through the alternating-time-frequency procedure, which is able to where f is the surrogate model and V is the vector of experimental
calculate different types of nonlinear functions even with nonsmooth points. The prior distribution is represented by Pθjf, PVjθ; f is
the likelihood function, PVjf is the evidence, and PθjV; f is the
nonlinearities [65]. For each iteration, the predicted values of X 0 , Xk;s ,
posterior distribution. The evidence function acts as a normalizing
Xk;c , and ω are used in Eqs. (5) to obtain the time-domain response
constant to ensure the posterior integrates into one. Prior distributions
over a period. The time-domain nonlinear force response fnl t is then
reflect prior knowledge of the model parameters from observations.
found. A fast Fourier transform algorithm is used to estimate Fourier
In this work, distributions used to gather training data are selected.
coefficients based on the time-domain nonlinear force response. With
For a multidimensional problem such as the one discussed in this
the relationships described, the equation of motion shown in Eq. (2)
paper, distributions are combined as follows:
can be expressed into a set of algebraic residual equations, which are
solved numerically using a Newton–Raphson solver [1]. The full n
continuation scheme is laid out in [45]. Pθjf  Pθi jf (10)
i1
2. Kriging Algorithm
Based on the above HBM with continuation techniques, sets of The likelihood function reflects the degree of agreement between the
training input data Θ  θ1 ; θ2 ; : : : ; θn  and corresponding outputs obtained measurements V and the output of the surrogate model
V  V 1 ; V 2 ; : : : ; V n  can be gathered. In this study, LHS is used fθjf. Assuming the error between observations and the model
to gather the samples for the training data [66]. This choice of sampling follows a zero-mean normal distribution, the likelihood function
method is standard in surrogate development for exploring diverse can be written as
ranges of scenarios with minimal data. For each θ a set of h and
corresponding V values are generated with the HBM continuation
n
1 n
V i − fθi 2
PVjθ; f  p exp − (11)
scheme. Nonlinear parameters can be drawn from a distribution set by i1 σ i 2π i1
2σ 2i
the user. Assuming no knowledge of the system, wide, uniform distri-
butions can be typically set. This method ensures a diverse set of points where σ i is the variance of error ϵi, which controls the centralization
along each variable. The kriging predictor is formulated as follows: degree of the posterior distribution. Also referred to as the width
parameter, ϵi is predetermined and case dependent. The width param-
fθ  μθ
wT θK −1 V − μΘ (6) eter is selected based on the resulting acceptance rate of the samples.
The acceptance rate shows the tradeoff between accepting too many
where μx is the mean function that represents the expected value or small steps and rejecting too many large proposal steps. Typically, an
trend of the response variable. It is typically assumed to be constant or acceptance rate between 0.15 and 0.5 ensures that the algorithm’s
can be defined based on prior knowledge or domain expertise. The efficiency is above 80% so is aimed for [69]. The approximate
covariance matrix K is an n × n matrix, where K ij  Corrxi ; xj  Bayesian from Eq. (9) [where PVjf is assumed as a constant] is
represents the covariance or correlation between the ith and jth input evaluated by drawing samples from prior distributions until con-
points. The weight factor wx is defined as verged mean posteriors are reached. As the goal of the sampling is
to converge to an unknown stationary distribution, standard methods
wx  K−1 Corrx; X (7) (such as LHS used for surrogate development) are not suitable.
Advanced sampling techniques are, therefore, employed commonly
where Corrx; X is a vector representing the correlation or covariance for optimal efficiency [28], the three most common being MCMC,
between input x and training point X. In kriging, correlation is deter- TMCMC, and SMC. Each method provides unique benefits in terms
mined based on functions in the spatial domain. Here, the Matérn of computation time, number of samples required, and accuracy. The
covariance function is used: choice of sampling method is case dependent and should be consid-
ered carefully.
p υ p
21−υ 2υh 2υh The MCMC (Metropolis–Hastings method) sampler is composed
Corrx; X  σ 2 Kυ (8) of two main features: Monte Carlo simulation and Markov chains.
Γυ a a
Introduced by Markov, a chain is initiated from θ1 and a transition to
 1894 MCGURK ET AL.

θ2 that is carried out based on a transition probability distribution comparing bifurcation plots using BMU predictions of nonlinear
function [70]. The assumption is that by allowing the chain to extend parameters to experimental data for both configurations.
infinitely, the chain converges to a stationary distribution, which Forward UQ is performed by running the output distributions
corresponds to the posterior. The computation time per sample is through fθ. This produces a bifurcation plot with 95% confidence
the lowest with this method, but often a higher number of samples are bands. In theory, the mean solution should give a close approximation
required to reach a converged solution. It is common to discard a of the true bifurcation plot. It could also be argued, if all input
number of initial samples in what is known as the “burn-in” to experimental data fall within the confidence bands, a reasonable
improve the quality of the results. A tuning parameter is also required approximation of true uncertainty in the predicted parameters has
with MCMC; both tuning parameter and burn-in are predefined with been captured. If data points lie outside of the confidence bands, it
educated trial and error. may suggest that the mathematical model does not truly represent
The TMCMC sampler is an adaptive variation of MCMC [71] dynamic behavior or that there have been errors in obtaining the
and is also a specific variant of the SMC samplers (see [72]). A experimental data.
series of intermediate transitional distributions are used to gradu-
ally converge the samples toward the final posterior distribution. D. TMCMC-Based Class Selection
In addition, unlike the MCMC, it does not generate samples in One of the key advantages of the Bayesian interface is model
a serial manner. Instead, it generates N distinct Markov chains, selection. With the evidence term, the validity of models can be
each initiating from the N samples generated from the previous ranked. The evidence term is not required in the cases of MCMC
transition distribution, and updates each chain according to its but can be numerically estimated [75]. With regard to the TMCMC
Downloaded by Northwestern Polytechnic University on April 29, 2024 | http://arc.aiaa.org | DOI: 10.2514/1.J063611

associated statistical weight. Through such an approach, the sampling method, the evidence for specific model f, PVjf, is a
TMCMC sampler is applicable toward sampling from multimodal byproduct required for estimation of the statistical weight function w
posteriors and has also been utilized to sample from high- used to describe the importance of each sample [76]. The model
dimensional posteriors (i.e., 18 dimensions) [28]. While having selection with TMCMC sampling is formulated in [76]. The
a higher computational cost, TMCMC is automatically tuned and TMCMC algorithm draws samples from a series of intermediate
converges with fewer samples than MCMC commonly. Parallel distributions, generated iteratively, with the capability of converging
computing can also be taken advantage of to improve computa- to complex multimodal posterior distributions. These PDFs are
tional efficiency. defined as
The SMC sampler is a subset of the SMC methods that is based on
the particle filter methods typically employed for system identifi- Pj ∝ Pθjf ⋅ PVjf; θβj j  0; : : : ; m (14)
cation purposes [73,74]. The sampler utilizes the importance of the
sampling–resampling approach to generate samples sequentially where j is the stage number. Tempering parameter β evolves from
from a dynamic posterior. Like the TMCMC sampler, the sampling β0  0 to βm  1 to ensure a smooth transition between the prior and
procedure involves generating N distinct Markov chains, each posterior distribution so P0  Pθjf to Pm  Pθjf; V. The
initiating from the N samples obtained from the previous transi- weight of each sample is determined as
tional distribution. The difference from the TMCMC sampler is that
each Markov chain generates a new sample independently from one Pθjf ⋅ PVjf; θβj
1
wθj;k    PVjf; θβj
1 −βj k  1;: : : ; N j
another (i.e., each Markov chain generates one updated sample Pθjf ⋅ PVjf; θβj
from the starting seed sample), which allows the sampling pro-
cedure to be easily parallelized. As such, the SMC sampling (15)
approach is applicable for online Bayesian inference to infer either
time-invariant or time-varying parameters, while the MCMC and where N is the number of samples. Next, compute
the TMCMC samplers are only applicable for off-line Bayesian Nj
inference to infer time-invariant parameters. In addition, like the wθj;k 
TMCMC sampler, it is also able to generate samples from multi- Sj  k1
(16)
Nj
modal and high-dimensional posteriors. Computation time per
sample is on average less than TMCMC but more than MCMC, This process is repeated for each stage. At the end of the algorithm,
with convergence usually reached in fewer samples than MCMC. the evidence of the model can be computed as
Each method will be tested in this study to evaluate which sampling
method is suited to the methodology. Further details on the Baye- m−1
sian interface and each sampling method are laid out by Lye PVjf  Sj (17)
et al. [28]. j0
Three factors are used to determine the overall quality of the
results: coefficient of variance, bias, and quality [72]. Coefficient It is common to evaluate logPVjf for ease of computation,
of variance COV  std∕μ is used to determine the confidence mainly to avoid numerical problems (e.g., arithmetic underflow)
of a prediction, where std is the standard deviation and μ is the associated with the calculation of the full likelihood function [28].
mean of the prediction from BMU. Bias is calculated with The less negative logPVjf, the more evidence there is to suggest
that the specific model fits the provided data.
jμ − ej
bias  (12)
e
III. Model Description
where e is the corresponding true experimental point to μ. This In this section, the experimental flutter rig and the mathematical
gives a measure of accuracy only considering the mean prediction model that will be used to study the proposed probabilistic identi-
and not confidence bands. A completely accurate prediction fication framework are described.
would give a bias of zero. Taking confidence of prediction into
account, bias and COV are used to calculate quality. A. Experimental Configuration and Data Collection
Figure 3a shows the experimental flutter rig composed of a
quaility  bias2
COV 2 (13) NACA-0015 wing profile rigidly attached to a stainless steel shaft.
The airfoil is supported by rotational bearings on each end allowing
Combining both confidence and accuracy, quality is the main for rotation and a bearing system allowing vertical displacement.
factor used to determine overall accuracy. An ideal prediction The spring in the heavy degree of freedom behaves in a linear
would have a quality of zero. These factors will be calculated by manner. In the bearings, leaf springs are introduced to create a
 MCGURK ET AL. 1895

Fig. 3 Nonlinear flutter test rig.

nonlinear hardening effect in pitch motion. This mimics nonlinear B. Mathematical Model
effects at interfaces that are encountered in real aircraft. A detailed
Downloaded by Northwestern Polytechnic University on April 29, 2024 | http://arc.aiaa.org | DOI: 10.2514/1.J063611

The flutter rig is modeled as the two-degree-of-freedom system

description of the setup is found in [37]. An Omron ZX1-LD300 shown in Fig. 3b, coupled with an unsteady aerodynamic model
laser displacement sensor is used to measure heave displacement, described by Abdelkef et al. [77]. For the nonlinear flutter rig
and an RLS AksIM 18 bit absolute magnetic encoder captures pitch considered, x  h; α; w, where h is heave, pitch is α, and w is the
motion. The wind speed was recorded directly from the wind tunnel aerodynamic state. The structural matrices shown in Eq. (1) are
control system. Both stable and unstable LCOs are captured through configured as
control-based continuation. The raw data were fed into a fast
Fourier transform algorithm to minimize noise. This process was mT mw xα b 0 ch 0 0
carried out for two experimental configurations with different
spring constants. LCO amplitude and corresponding velocity for M mw xα b Iα 0 ; D 0 cα 0 ;
each configuration are displayed in Fig. 4, with design rig param- 0 0 1 −1∕b a − 1∕2 0
eters for both presented in Table 1. Data were collected using CBC,
providing access to both stable and unstable responses of the system Kh 0 0
(see [61] for more explanations about CBC and its exploitation in K 0 Kα 0 (18)
this context). Based on two sets of experimental data, K α2 and Kα3
were determined by implementing the deterministic approach to 0 0 0
normal form theory in [61]. The estimated values are shown in
Table 1 and will be used as the initial input points. The unsteady aerodynamic matrices are described as follows:

Fig. 4 Labeled experimental data ( , stable LCO; , unstable LCO).

Table 1 Measured parameters and estimated parameters, including initial estimates for nonlinear
parameters
Measured parameter Value Estimated parameter Initial estimated value
cm 0.15 Kh N∕m 3529.4 (configuration 1), 3318.3 (configuration 2)
a −0.5 Kα N∕rad 54.11 (configuration 1), 65.6 (configuration 2)
ρkg∕m3  1.204 Kα2 N∕rad2  751.6 (configuration 1), 774.7 (configuration 2)
mw kg 5.3 Kα3 N∕rad3  5006.7 (configuration 1), 3490.7 (configuration 2)
mT 16.9 cα kg ⋅ m2 ∕s2  0.5628 (configuration 1), 1.0338 (configuration 2)
xα 0.24
Iα 0.1724
ch kg∕s 14.5756
 1896 MCGURK ET AL.

−πb2 aπb3 0
A aπb3 −π1∕8
a2 b4
0 0 1
−πb −1
1∕2 − aπb2 ∕ρ −2πVbc1 c2
c3 c4 ∕ρ
B πa
1∕2b2 ∕V −1∕4 − a2 πb3 2πb2 Va
1∕2c1 c2
c3 c4 
0 0 −c2
c4 ∕ρb
0 −πb −2πVc2 c4 c1
c3 ∕ρ
C  0 πb2 1∕2
a 2πba
1∕2c2 c4 c1
c3  (19)
0 1∕ρVb c2 c4 ∕ρb2

where A; B; and C are the corresponding matrices in Eq. (1). To IV. Results: Comparison of Sampling Methods
model the nonlinear part of the system, it is initially assumed that
Downloaded by Northwestern Polytechnic University on April 29, 2024 | http://arc.aiaa.org | DOI: 10.2514/1.J063611

In this section, the described probabilistic identification method-

spring stiffness is approximated by quadratic and cubic terms in pitch ology proposed in the Sec. II is applied to the two configurations of
degree of freedom (which is typically used to represent geometrical the above nonlinear flutter test rig. It is assumed that original non-
nonlinearities [78]): linear function is in a form shown in Eq. (20) as in previous work
[19,37,58,61]. Considering the reference data in [58], we assume that
0 K α2 is in the range of 0–1500 N∕rad2 while Kα3 is in the range of
0–10;000 N∕rad3 . For each setup, three sampling methods, namely,
fnl  K α2 α2
Kα3 α3 ; qnl  1 (20) MCMC, TMCMC, and SMC sampling methods, are applied and
0 compared in terms of the quality factors of the model identification.
These results are also compared to the initial estimates obtained from
normal from theory [58] and listed in Table 1.
Nonlinear parameters K α2 and Kα3 are treated as unknowns. The Figure 5 shows the mean accuracy of kriging surrogate models
methodology described in this paper will work to estimate the non- with respect to the amount of training continuation runs carried out.
linear parameters. Other known parameter values are listed in Table 1. Here, mean accuracy is estimated based on the difference between
predictions from the kriging model and a separate set of comparison
data from continuation runs. Accuracy converges with 400 bifurca-
tion diagrams in training data at 99.8%. The same amount of training
data is again required for the second configuration. Figure 6 com-
pares the amplitude of LCO obtained from the HBM simulation and
the kriging surrogate model. The highlighted area reflects the degree
of uncertainty from the predictions using the kriging model within
95% confidence bands. Small discrepancies are observed at turning
points as the sharpness of the curve is increased, which was not
improved with increased sampling. The largest overshoot of the
turning point observed was 2% (shown in Fig. 6b), which was
considered within a reasonable margin of error for this study. Uncer-
tainty in model predictions close to estimated values for nonlinear
parameters proved to be less than 1% on average. It was therefore
assumed that the mean predictions of the surrogate model were
accurate in the subsequent Bayesian parameter estimation process,
Fig. 5 Accuracy of surrogate model with respect to sample size for and the uncertainty associated with the surrogate model was not
configuration 1. incorporated into the parameter estimation. With mean predictions,

Fig. 6 Surrogate comparison.

 MCGURK ET AL. 1897

it is observed that the kriging model is capable of accurately capturing Table 3 Setup 1 parameter estimation
the response of LCO.
MCMC TMCMC SMC

A. Configuration 1 Parameter Kα2 Kα3 Kα2 K α3 K α2 Kα3

The Bayesian inference framework is first applied to the first Mean 517.20 2555.8 524.50 2595.9 505.08 2420.6
COV [%] 5.52 9.44 9.72 16.20 6.29 3.47
configuration with the experimental data shown in Fig. 4a. Tuning
Convergence 2736 215 153
parameters presented in Table 2 are used to ensure an acceptance rate (samples)
between 0.15 and 0.5 for each sampling method. Prior distributions
are uniform between the stated bounds, as low prior knowledge of the
parameters is assumed. Figure 7 shows the convergence of nonlinear
parameters with three sampling methods. Convergence is defined as sampling methods is within 15%. It is worth noting that the estima-
when the mean and standard deviation of the posterior distribution tion of the mean is around 45% different from the initial estimates
become constant with further sampling. It shows that TMCMC and shown in Table 1. For both parameters, TMCMC provides the
SMC can converge within 300 samples, while MCMC required 2736 smallest confidence interval in parameter prediction. It is also found
samples with an additional 200 burn-in samples before the conver- that the two nonlinear parameters are linearly correlated from MCMC
gence. It is worth noting that SMC converges in the fewest samples, and TMCMC, which, however, is not observed with the SMC sam-
with only 153 mainly due to the benefit of using importance sam- pling method as no samples are rejected during the sampling process.
pling. With TMCMC and SMC, less samples are required for model Figure 9 shows the distributed bifurcation diagrams using the
Downloaded by Northwestern Polytechnic University on April 29, 2024 | http://arc.aiaa.org | DOI: 10.2514/1.J063611

updating than in the construction of the surrogate model. This seems estimated nonlinear parameters from the BMU. It is observed that
to suggest that there is no need to construct a surrogate model. both TMCMC and SMC results capture all experimental data points
However, with the goal of using the estimated parameter distributions within their 95% uncertainty bands. In the case of MCMC, the
to quantify the impact of uncertainty on bifurcation diagrams, the confidence interval is much narrower than that using TMCMC and
surrogate models are again implemented, significantly reducing SMC. There are two data points in the upper portion of the diagram
computation time. falling outside uncertainty bands. It can be also found that the lowest
Figure 8 shows the PDFs and scatter plots using three different band of the turning point in MCMC and TMCMC spans to 14 m∕s,
sampling methods. It can be observed that each sampling method while this point is at 12.20 m∕s for SMC. If we are to assume that the
produces a PDF that resembles a normal distribution. The mean and mathematical model is correct, it indicates that SMC appears to
COV of estimated nonlinear parameters are presented in Table 3. The overestimate the uncertainty of LCO.
estimated mean of the nonlinear parameter between these three Figure 10 shows a comparison of mean lines from three sampling
methods, the initial estimate, and experimental data. A clear improve-
Table 2 Tuning parameters ment can be observed with each of the sampling methods compared
to the one using the initial estimates. The estimation of the LCO
Prior distribution bounds Width parameter Burn-in Covariance matrix amplitude in each method closely follows the experimental data at the
Kα2 : 0–1500 MCMC: 0.6 200 [1000 0] low amplitude points but diverges after the turning point. The mean
Kα3 : 0–1e4 TMCMC: 1 0 [0 7000] lines from each method closely follow each other, only overshooting
SMC: 0.1 0 the turning point by 0.3 m∕s. Figure 10b shows the bias of these four
methods in relation to experimental data. Each sampling method is

Fig. 7 Configuration 1 trace plots ( , mean; , mean
std).

Fig. 8 Configuration 1 PDF and scatter plots.

 1898 MCGURK ET AL.

Fig. 9 Configuration 1 bifurcation plots (*, experimental points; , mean plot; , 95% confidence bounds).
Downloaded by Northwestern Polytechnic University on April 29, 2024 | http://arc.aiaa.org | DOI: 10.2514/1.J063611

Fig. 10 Configuration 1 comparison (*, experimental points; –, initial estimate; , MCMC; , TMCMC; , SMC).

more accurate than the initial estimate for all data points apart from observed with results from MCMC and TMCMC but not SMC. The
point 7. The accuracy for all sampling methods remains above 94% mean and COV from the estimations are shown in Table 4. It can be
without a sharp drop in accuracy. Figure 10c shows the quality of the seen that the mean results from the BMU are closer to the initial
prediction between these three sampling methods. It shows MCMC estimates shown in Table 1. The most confident predictions are again
and TMCMC provide the best overall estimation of physical behavior achieved, with SMC achieving the lowest average COV.
in terms of confidence and accuracy. The biases of all methods are Figure 13 shows the bifurcation plot with confidence interval using
within 3%. However, the relatively large COV in SMC predictions the estimated posterior distribution of estimations. Similar results are
lead to a higher quality than the other two sampling methods. observed in the first setup, with SMC capturing all data points with
wide confidence bands, while MCMC and TMCMC only capture
B. Configuration 2 LCO behavior at low amplitude before the turning point. Again, the
Using the same method for configuration 1, a surrogate model for MCMC does not capture two data points after the turning point, and
configuration 2 is constructed with 386 continuation runs required for the TMCMC misses only one point. The lower band of SMC signifi-
training data. The tuning parameters presented in Table 2 were cantly overshoots the turning point. As shown in Fig. 14a, the mean
implemented for model updating in setup 2. Figure 11 shows the line from each sampling method is compared to the response gen-
convergence of each sampling method, where a similar pattern is erated with the initial estimate. In this case, it is less obvious which
observed as the first setup. The SMC sampling method remains the prediction is more accurate. At lower amplitudes, the initial estimate
fastest-converging one with only 200 samples. Figure 12 shows the matches the behavior, but after the turning point, the accuracy is lost.
scatter plots and PDFs produced. In this case, each of the sampling The mean lines from BMU predictions offer a slight improvement
methods provides a shape similar to a normal distribution for both over the initial estimate. The initial estimate overshoots the turning
parameters. A strong correlation between the parameters is again point the most. Figure 14b shows a comparison of the bias of all the

Fig. 11 Setup 2 trace plots ( , mean; , mean
std).

 MCGURK ET AL. 1899

Fig. 12 Setup 2 PDF and scatter plots.

second configuration, the predicted results for both BMU and the
Table 4 Setup 2 parameter estimation
initial estimate appear to be in better agreement than in the first
Downloaded by Northwestern Polytechnic University on April 29, 2024 | http://arc.aiaa.org | DOI: 10.2514/1.J063611

MCMC TMCMC SMC configuration. This could mean the mathematical model holds
Parameter Kα2 Kα3 K α2 Kα3 Kα2 Kα3 better for the second setup or there were fewer errors in gathering
the experimental results. Another reason could be the increased
Mean 700.03 3122.4 678.32 2982.6 697.20 3040.8
COV [%] 4.00 7.66 5.26 9.41 3.76 3.32
number of experimental data points used in the estimation. Consid-
Convergence 2531 207 160 ering the overall mean accuracy, each sampling method outper-
(samples) formed the initial estimate, particularly at higher amplitudes. Each
sampling method can be argued for depending on whether the user
prioritizes the accuracy of the mean prediction or a combination of
accuracy and confidence. If confidence and mean prediction are
predictions compared to the experimental points. It is observed that valued, MCMC provides the best estimation. In contrast, if captur-
the initial estimate is most accurate until the turning point. For the ing all physical behavior is preferred, TMCMC and SMC may be
stable LCO, predictions from MCMC and TMCMC provide the most chosen despite being overshoots of the turning point, especially in
accurate predictions. Figure 14c confirms similar conclusions to the case of SMC. Parameter estimations have the lowest COV with
Fig. 10c, where MCMC and TMCMC show the best overall quality. SMC sampling but provide the widest confidence bands in their
bifurcation plots. This is likely due to SMC not rejecting samples
C. Conclusions of Sampling Method Comparison and not observing correlations between parameters. Both TMCMC
Considering both configurations, several conclusions are and SMC converge with fewer samples than training input runs are
reached. The experimental data are captured well using BMU, required; however, taking forward UQ into consideration, up to
leading to better estimations than deterministic results. In the 60% fewer HBM continuation runs are required.

Fig. 13 Setup 2 bifurcation plots (*, experimental points; , mean plot; , 95% confidence bounds).

Fig. 14 Setup 2 comparison (*, experimental points; –, initial estimate; , MCMC; , TMCMC; , SMC).
 1900 MCGURK ET AL.

Taking all factors into account, TMCMC provides the best overall whole design space. The turning points of all bifurcation plots
compromise in terms of convergence and accuracy for both configu- involving cubic hardening are below 12 m∕s for configuration 1,
rations. In configuration 1, the bifurcation plot produced captures all overshooting what would be expected from the experimental results
data points without significantly overshooting the turning point, by over 25%. The same behavior was observed in configuration 2.
compared to over 90% of the data in configuration 2. The estimations For this reason, cubic hardening design space is not considered. As
from TMCMC and SMC in configuration 1 capture all physical seen in [38], third- and fifth-order nonlinearity is also considered in
behavior within their confidence bands, which suggests that the subcritical airfoil systems. With the same criteria as in Fig. 15, this
parameter distributions accurately estimate the true uncertainty of nonlinearity’s only feasible design space is with a third-order soft-
the parameters and the mathematical model holds particularly in the ening and fifth-order hardening spring.
case of TMCMC. For these reasons, it may be argued that the
estimated distributions of nonlinear parameters can effectively fnl  K α3 α3  K α5 α5 (21)
represent physical behavior. Alternatively, there is less agreement
between all predicted values and experimental data after the turning
Table 5 shows the three possible nonlinear models to approximate the
point. There are two possible reasons: the model only holds for low-
experimental LCO data. The suitability of the models will be ranked
amplitude results or there was a measurement error in gathering
higher amplitude points. For the first case, the form of the model by both the average quality of the bifurcation results and the evi-
needs to be modified to capture higher amplitude behavior. Since the dence function logPDjM from BMU with TMCMC sampling.
errors become larger as amplitude increases, it is more likely the The kriging surrogate models for each nonlinear function are first
developed through the same methodology presented in Sec. II. The
Downloaded by Northwestern Polytechnic University on April 29, 2024 | http://arc.aiaa.org | DOI: 10.2514/1.J063611

aerodynamic model does not hold for higher-amplitude in this sce-

nario, but further investigation is required. TMCMC sampling method is selected for BMU since it has proved to
Based on the results from both setups, a clear pattern has emerged be the most effective method using the original nonlinearity in the
in that lower amplitude data points are well captured, but accuracy is previous section. It can also produce the evidence function as a by-
lost in the upper sections in both BMU and initial estimates. There product of the model selection. It was found that the evaluation of
may be two reasons why the experimental data may not match the logPDjM required additional samples to converge over mean and
results: 1) there was a significant amount of noise or error in gathering standard deviation values. Table 6 shows that logPDjM con-
the data; 2) the mathematical model, particularly the nonlinear func- verges at 10,000 samples. For each model with both configurations,
tion, may not fully capture the true physical behavior of the system. In
an effort to further improve the accuracy of mathematical models to
Table 6 Convergence of evidence
capture the experimental data, alternative nonlinear models are sug- function for configuration 1, model 1
gested and ranked in the next section. with TMCMC sampling

Samples logPDjM
V. Results: Nonlinear Model Selection 100 −7.83
Previously, it was assumed that both Kα2 and Kα3 were positive. 500 −7.94
1000 −7.72
Prior distributions and LHS for surrogate models for nonlinear
2500 −7.35
parameters could not draw samples from the negative design space. 5000 −7.44
In this section, prior distributions that include negative design space 7500 −7.46
are considered. The aim is to improve agreement with results from 10,000 −7.45
model updating and experimental data by altering the characteristics 125,000 −7.45
of the nonlinear function. Figure 15 shows an examination of the

Fig. 15 Full model selection design space. Each point on the scatter plot represents a full bifurcation diagram.

Table 5 Parameter estimations for model selection

No. Model Parameter, unit Mean COV, %

1 Kα2
Kα3 Kα2 N∕rad2 , K α3 N∕rad3  524.50,2595.9 (configuration 1) 9.72,16.20 (configuration 1)
678.32,2982.6 (configuration 2) 5.26,9.41 (configuration 2)
2 −Kα2
Kα3 Kα2 N∕rad2 , K α3 N∕rad3  −383.78,1442.1 (configuration 1) 12.20, 22.17 (configuration 1)
−530.99,1754.2 (configuration 2) 4.20,7.84 (configuration 2)
3 −Kα3
Kα5 Kα3 N∕rad3 , K α5 N∕rad5  −1822.5,1.136e5 (configuration 1) 22.01,31.14 (configuration 1)
−2343.5,1.061e5 (configuration 2) 19.27,21.55 (configuration 2)
 MCGURK ET AL. 1901

10,000 samples are taken to ensure evidence function convergence. evidence function indicates model 2 is the best choice, whereas
Tables 5 and 7 display the estimated parameters for each model and model 1 has the lowest quaility factor. Figure 16 shows that non-
the evidence function compared to the average quaility of pre- linear models 1 and 2 capture all data within confidence bands. The
dictions. bifurcation plots produced from all models show plots that closely
As shown in Table 6, model 3 provided the highest variation in resemble each other. The model 3 fails to capture one data point
predicted parameters for both configurations. For configuration 1, from the confidence bands after the turning point. Figure 16 shows a
model 1 has the highest confidence in estimations, and model 2 has closer examination of the PDFs of nonlinear parameters. It is very
the highest, with configuration 2 having the lowest average COV, clear that multimodal distributions are produced for models 2 and 3,
respectively. suggesting the existence of two possible solutions. The mode with
Table 7 shows the computed evidence of each model for the two the highest probability density is referred to as peak 1, while peak 2
configurations. It can be observed that, for configuration 1, the refers to the lower probability density peak. The model 3 has a
single-modal solution for K α3 but a multimodal solution for K α5.
Table 7 Evidence function for configurations 1 and 2 from 10,000 Figures 18a and 18b show the response of LCO using the nonlinear
TMCMC samples parameters estimated from each peak. For model 2, it can be
observed that peak 1 fails to capture the highest amplitude point
Configuration 1 Configuration 2 but has a closer overall fit to the experimental data. The response
quality, quality, using the nonlinear parameters from peak 2 reaches the highest
No. fnl logPDjM % logPDjM % amplitude point but stays away from the remaining data at ampli-
Downloaded by Northwestern Polytechnic University on April 29, 2024 | http://arc.aiaa.org | DOI: 10.2514/1.J063611

1 Kα2
Kα3 −7.45 5.24 −8.86 5.21 tudes above the turning point. In model 3, peak 1 captures the lower
2 −Kα2
Kα3 −5.30 5.61 −9.74 5.01 amplitude data and the highest amplitude point. Peak 2 captures
3 −Kα3
Kα5 −5.68 8.74 −8.77 4.78 data close to the turning point while failing to reach the highest
amplitude points.

Fig. 16 Bifurcation diagrams and PDF for configuration 1 [for multimodal PDF ( , peak 1; , peak 2)].

Fig. 17 Bifurcation diagrams and PDF for configuration 2 [for multimodal PDF ( , peak 1; , peak 2)].
 1902 MCGURK ET AL.

Fig. 18 Multimodal mean bifurcation diagrams ( , peak 1; , peak 2).

Considering just average quality, results suggest that model 1 is the convergence, it was determined that TMCMC produced the highest
optimal choice for configuration 1. This is in line with the bifurcation quality estimations, capturing all experimental points with much
Downloaded by Northwestern Polytechnic University on April 29, 2024 | http://arc.aiaa.org | DOI: 10.2514/1.J063611

diagram capturing all experimental points within confidence bands. higher convergence than MCMC. All predictions with estimations
Based on the TMCMC class selection, both models 2 and 3 are from BMU were up to 20% more accurate than those from the initial
recommended above model 1. However, the solutions to these mod- estimate. Accuracy in prediction of physical behavior following the
els are multimodal, as discussed. Since there are two suggested turning point dropped. From 5 to 22% of data points are omitted
solutions, it is not valid to just consider the mean and COV of the from confidence bands following the turning point. This suggests
posterior distribution estimations. This suggests considering a com- the mathematical model does not hold for higher amplitudes of
bination of the two different nonlinear functions at different LCO LCO, and a more sophisticated aerodynamic model might be
amplitudes might provide a better model than the single solution needed. Considering computational efficiency, for purely estimat-
provided by model 1. Quality measurement assumes there is a single ing mean nonlinear parameters, more HBM continuation runs are
estimated solution, whereas evidence function can account for multi- required with TMCMC and SMC with the data-driven approach. In
modal results. This is why quality and evidence functions recom- carrying out forward UQ, however, the proposed methodology can
mend alternative solutions in both configurations. cut the amount of costly continuation runs by up to 60%.
In configuration 2, both evidence function and quaility suggest that Two categories of nonlinear models were implemented and eval-
model 3 is the optimal choice. The examination of posterior distribu- uated through TMCMC class selection. For the first configuration, the
tions in Fig. 17 shows that models 1 and 2 converge to a unique solution, TMCMC process suggested selecting model 2, which includes quad-
while model 3 produces a multimodal PDF with two possible solutions. ratic softening and cubic hardening. As for the second configuration,
Figure 18c shows the amplitude of LCO with these two possible model 3 equipped with third-order softening and fifth-order hardening
solutions. It shows that peak 1 captures lower amplitude data, while was deemed the most suitable option. After analyzing the quality factor
peak 2 captures the higher amplitude points. As was the case with of the results obtained from the bifurcation diagrams along with the
configuration 1, TMCMC model selection favors a combination of confidence bands for both scenarios, it can be concluded that model 1 is
these two possible solutions, which outperforms a single solution from the most suitable option. An examination of posterior PDFs of nonlinear
the original model. Overall, this could mean that none of the models are parameters showed the presence of multimodal solutions in both con-
perfect, particularly with higher amplitude points, for which an updated figurations. Taking a single mean and standard deviation estimation
model is required. from either of the two estimated solutions did not capture the complete
set of data. Upon examining both solutions, it was noted that each
solution is capable of capturing a portion of the bifurcation diagram
VI. Conclusions either for lower or upper amplitude data. TMCMC class selection
In this paper, a probabilistic identification methodology for esti- suggests there is no unique solution to represent the nonlinear functions
mating parameters in a nonlinear aeroelastic system is presented. to capture all the physics across a wide range of amplitudes, particularly
An advanced kriging algorithm is used to develop accurate data- for higher amplitude points. A further improvement of the mathematical
driven surrogate models for LCO behavior. The surrogate models model for the nonlinear aeroelastic model is required. This could either
are employed with BMU to identify the nonlinear parameters and be achieved with an alternative nonlinear function or by implementing
functions. The performance of three advanced sampling methods nonlinearities in different degrees of freedom. Another area of improve-
is compared. Experimental data from the nonlinear flutter rig with ment could be in optimizing the amount of training data required for
two separate configurations was employed with the goal of estimat- constructing the surrogate models, potentially with a multilevel
ing a nonlinear model to fit the data. approach. This would integrate the kriging model with the model
Probabilistic bifurcation diagrams produced with BMU success- updating process.
fully captured the experimental data within the computed confi-
dence bands at low amplitude points. This proved a significant
improvement from the initial deterministic approach for each con- Acknowledgments
figuration. With higher amplitude points, portions of the data were
not captured within confidence bands but made an improvement M. McGurk acknowledges the support from the EPSRC Doctoral
over the initial estimate. Results from BMU for each sampling Training Partnership studentship for his Ph.D. study at the University
method had variations due to the different sampling philosophies of Strathclyde. J. Yuan acknowledges the funding support of the
associated with each respective approach. In the first configuration, Royal Academy of Engineering/Leverhulme Trust Research Fellow-
TMCMC and SMC could capture all data points within confidence ship (LTRF2223-19-150). The authors acknowledge the experimen-
bands, while SMC overshoots the estimation of the turning point by tal work by Kyoung Hyun Lee. Experimental data are available at
a large margin. The correlation between nonlinear parameters was https://github.com/Kyounghyunlee/.
also observed in the MCMC and TMCMC results. Only SMC could
capture all physical behavior while overshooting the turning point
in the second configuration. The estimations from the TMCMC References
are able to capture over 90% of the physical behavior, leading to [1] Dimitriadis, G., Introduction to Nonlinear Aeroelasticity, Wiley,
an average mean accuracy of 98%. Considering the accuracy and Hoboken, NJ, 2017, p. 4 (Sec. 1.1).
 MCGURK ET AL. 1903

[2] Ghaffari, R., and Sauer, R. A., “Modal Analysis of Graphene-Based Nonlinear Dynamical Systems in Aerospace and Mechanical Engi-
Structures for Large Deformations, Contact and Material Nonlinear- neering,” ASCE-ASME Journal of Risk and Uncertainty in Engineer-
ities,” Journal of Sound and Vibration, Vol. 423, June 2018, pp. 161– ing Systems, Part B: Mechanical Engineering, Vol. 9, No. 4, 2023,
179. Paper 040301.
https://doi.org/10.1016/j.jsv.2018.02.051 https://doi.org/10.1115/1.4063301
[3] Quintana, A. G., Saunders, B. E., Vasconcellos, R., and Abdelkefi, A., [22] Beck, J. L., and Katafygiotis, L. S., “Updating Models and Their
“Nonlinear Characterization of the Piecewise Structural Effects on Uncertainties. I: Bayesian Statistical Framework,” Journal of Engineer-
Whirl Flutter of a Rotor-Nacelle System,” AIAA Scitech 2022 Forum, ing Mechanics, Vol. 124, No. 4, 1998, pp. 455–461.
AIAA Paper 2022-2269, 2022. [23] Namdeo, V., and Manohar, C., “Nonlinear Structural Dynamical System
https://doi.org/10.2514/6.2022-2269 Identification Using Adaptive Particle Filters,” Journal of Sound and
[4] Mair, C., Rezgui, D., and Titurus, B., “Nonlinear Stability Analysis of Vibration, Vol. 306, Nos. 3–5, 2007, pp. 524–563.
Whirl Flutter in a Tiltrotor Rotor-Nacelle System,” Nonlinear Dynam- https://doi.org/10.1016/j.jsv.2007.05.040
ics, Vol. 94, No. 3, 2018, pp. 2013–2032. [24] Green, P. L., and Worden, K., “Bayesian and Markov Chain Monte
https://doi.org/10.1007/s11071-018-4472-y Carlo Methods for Identifying Nonlinear Systems in the Presence of
[5] Quintana, A., Vasconcellos, R., Throneberry, G., and Abdelkefi, A., Uncertainty,” Philosophical Transactions of the Royal Society A:
“Nonlinear Analysis and Bifurcation Characteristics of Whirl Flutter in Mathematical, Physical and Engineering Sciences, Vol. 373,
Unmanned Aerial Systems,” Drones, Vol. 5, No. 4, 2021, p. 122. No. 2051, 2015, Paper 2014-0405.
https://doi.org/10.3390/drones5040122 https://doi.org/10.1098/rsta.2014.0405
[6] Phillips, W., Alley, N., and Niewoehner, R., “Effects of Nonlinearities [25] Ebrahimian, H., Astroza, R., and Conte, J. P., “Extended Kalman Filter
on Subsonic Aerodynamic Center,” Journal of Aircraft, Vol. 45, No. 4, for Material Parameter Estimation in Nonlinear Structural Finite
Downloaded by Northwestern Polytechnic University on April 29, 2024 | http://arc.aiaa.org | DOI: 10.2514/1.J063611

2008, pp. 1244–1255. Element Models Using Direct Differentiation Method,” Earthquake
https://doi.org/10.2514/1.34241 Engineering & Structural Dynamics, Vol. 44, No. 10, 2015,
[7] Shukla, H., and Patil, M. J., “Nonlinear State Feedback Control Design pp. 1495–1522.
to Eliminate Subcritical Limit Cycle Oscillations in Aeroelastic Sys- https://doi.org/10.1002/eqe.2532
tems,” Nonlinear Dynamics, Vol. 88, May 2017, pp. 1599–1614. [26] Yin, M., Iannelli, A., and Smith, R. S., “Maximum Likelihood Estima-
https://doi.org/10.1007/s11071-017-3332-5 tion in Data-Driven Modeling and Control,” IEEE Transactions on
[8] Guckenheimer, J., and Holmes, P., Nonlinear Oscillations, Dynamical Automatic Control, Vol. 68, No. 1, 2023, pp. 317–328.
Systems, and Bifurcations of Vector Fields, Vol. 42, Springer Science & https://doi.org/10.1109/TAC.2021.3137788
Business Media, New York, 2013, p. 458. [27] Noël, J.-P., and Kerschen, G., “Nonlinear System Identification in
[9] Tang, D., and Dowell, E. H., “Experimental and Theoretical Study on Structural Dynamics: 10 More Years of Progress,” Mechanical Systems
Aeroelastic Response of High-Aspect-Ratio Wings,” AIAA Journal, and Signal Processing, Vol. 83, Jan. 2017, pp. 2–35.
Vol. 39, No. 8, 2001, pp. 1430–1441. https://doi.org/10.1016/j.ymssp.2016.07.020
https://doi.org/10.2514/2.1484 [28] Lye, A., Cicirello, A., and Patelli, E., “Sampling Methods for Solving
[10] Patil, M. J., Hodges, D. H., and Cesnik, C. E., “Limit-Cycle Oscillations Bayesian Model Updating Problems: A Tutorial,” Mechanical Systems
in High-Aspect-Ratio Wings,” Journal of Fluids and Structures, Vol. 15, and Signal Processing, Vol. 159, Oct. 2021, Paper 107760.
No. 1, 2001, pp. 107–132. https://doi.org/10.1016/j.ymssp.2021.107760
https://doi.org/10.1006/jfls.2000.0329 [29] Rocchetta, R., Broggi, M., Huchet, Q., and Patelli, E., “On-Line
[11] Avin, O., Raveh, D. E., Drachinsky, A., Ben-Shmuel, Y., and Tur, M., Bayesian Model Updating for Structural Health Monitoring,”
“An Experimental Benchmark of a Very Flexiblewing,” AIAA Scitech Mechanical Systems and Signal Processing, Vol. 103, March 2018,
2021 Forum, AIAA Paper 2021-1709, 2021. pp. 174–195.
https://doi.org/10.2514/6.2021-1709 https://doi.org/10.1016/j.ymssp.2017.10.015
[12] Drachinsky, A., Avin, O., Raveh, D. E., Ben-Shmuel, Y., and Tur, M., [30] Zhang, Y., and Yang, W., “A Comparative Study of the Stochastic
“Flutter Tests of the Pazy Wing,” AIAA Journal, Vol. 60, No. 9, 2022, Simulation Methods Applied in Structural Health Monitoring,” Engi-
pp. 5414–5421. neering Computations, Vol. 31, No. 7, 2014, pp. 1484–1513.
https://doi.org/10.2514/1.J061717 https://doi.org/10.1108/EC-07-2013-0185
[13] Barton, D. A., Mann, B. P., and Burrow, S. G., “Control-Based Con- [31] Fang, C., Liu, H.-J., Lam, H.-F., Adeagbo, M. O., and Peng, H.-Y.,
tinuation for Investigating Nonlinear Experiments,” Journal of Vibra- “Practical Model Updating of the Ting Kau Bridge Through the
tion and Control, Vol. 18, No. 4, 2012, pp. 509–520. MCMC-Based Bayesian Algorithm Utilizing Measured Modal
https://doi.org/10.1177/1077546310384004 Parameters,” Engineering Structures, Vol. 254, March 2022, Paper
[14] Kaba, A., Metin, E. Y., and Turan, O., “Thrust Modelling of a Target 113839.
Drone Engine with Nonlinear Least–Squares Estimation Based on https://doi.org/10.1016/j.engstruct.2022.113839
Series Expansions,” Aircraft Engineering and Aerospace Technology, [32] Song, M., Renson, L., Moaveni, B., and Kerschen, G., “Bayesian
Vol. 95, No. 1, 2022, pp. 38–52. Model Updating and Class Selection of a Wing-Engine Structure with
https://doi.org/10.1108/AEAT-08-2021-0236 Nonlinear Connections Using Nonlinear Normal Modes,” Mechani-
[15] Eiselt, H., and Sandblom, C.-L., Nonlinear Optimization, Springer, cal Systems and Signal Processing, Vol. 165, Feb. 2022, Paper
Berlin, 2019, p. 123. 108337.
[16] Friswell, M., and Mottershead, J. E., Finite Element Model Updating in https://doi.org/10.1016/j.ymssp.2021.108337
Structural Dynamics, Vol. 38, Springer Science & Business Media, [33] Walter, E., and Pronzato, L., “On the Identifiability and Distinguish-
Dordrecht, The Netherlands, 1995, pp. 98–100. ability of Nonlinear Parametric Models,” Mathematics and Computers
[17] Patelli, E., Govers, Y., Broggi, M., Gomes, H. M., Link, M., and in Simulation, Vol. 42, Nos. 2–3, 1996, pp. 125–134.
Mottershead, J. E., “Sensitivity or Bayesian Model Updating: A Com- https://doi.org/10.1016/0378-4754(95)00123-9
parison of Techniques Using the DLR AIRMOD Test Data,” Archive of [34] Quaranta, G., Lacarbonara, W., and Masri, S. F., “A Review on
Applied Mechanics, Vol. 87, May 2017, pp. 905–925. Computational Intelligence for Identification of Nonlinear Dynami-
https://doi.org/10.1007/s00419-017-1233-1 cal Systems,” Nonlinear Dynamics, Vol. 99, No. 2, 2020, pp. 1709–
[18] Bi, S., Beer, M., Cogan, S., and Mottershead, J., “Stochastic Model 1761.
Updating with Uncertainty Quantification: An Overview and Tutorial,” https://doi.org/10.1007/s11071-019-05430-7
Mechanical Systems and Signal Processing, Vol. 204, Dec. 2023, Paper [35] Gui, G., Liu, F., Sun, J., Yang, J., Zhou, Z., and Zhao, D., “Flight Delay
110784. Prediction Based on Aviation Big Data and Machine Learning,” IEEE
https://doi.org/10.1016/j.ymssp.2023.110784 Transactions on Vehicular Technology, Vol. 69, No. 1, 2019, pp. 140–
[19] Beregi, S., Barton, D. A., Rezgui, D., and Neild, S., “Using Scientific 150.
Machine Learning for Experimental Bifurcation Analysis of Dynamic https://doi.org/10.1109/TVT.2019.2954094
Systems,” Mechanical Systems and Signal Processing, Vol. 184, [36] Gábor, A., and Banga, J. R., “Robust and Efficient Parameter Estimation
Feb. 2023, Paper 109649. in Dynamic Models of Biological Systems,” BMC Systems Biology,
https://doi.org/10.1016/j.ymssp.2022.109649 Vol. 9, No. 1, 2015, pp. 1–25.
[20] Stolovitch, L., “Progress in Normal Form Theory,” Nonlinearity, Vol. 22, https://doi.org/10.1186/s12918-015-0219-2
No. 7, 2009, Paper R77. [37] Lee, K., Barton, D., and Renson, L., “Mathematical Model Identifica-
https://doi.org/10.1088/0951-7715/22/7/R01 tion of Self-Excited Systems Using Experimental Bifurcation Analysis
[21] Yuan, J., Denimal, E., Bi, S., Feng, J., Hu, Q., and Cicirello, A., Data,” Nonlinear Structures & Systems, Volume 1: Proceedings of the
“Special Section on Uncertainty Quantification and Management in 40th IMAC, A Conference and Exposition on Structural Dynamics
 1904 MCGURK ET AL.

2022, Springer, Berlin, 2022, pp. 61–63. [55] Lee, D., Kang, Y.-E., Kim, D.-H., and Yee, K., “Aeroelastic Design and
https://doi.org/10.1007/978-3-031-04086-3_9 Comprehensive Analysis of Composite Rotor Blades Through
[38] Riso, C., Cesnik, C. E., and Epureanu, B. I., “Bifurcation-Diagram-Free Cluster-Based Kriging,” AIAA Journal, Vol. 60, No. 10, 2022,
Postflutter Response Constraint for Design Optimization,” AIAA Jour- pp. 5984–6004.
nal, Vol. 61, No. 8, 2023, pp. 1–18. https://doi.org/10.2514/1.J061820
https://doi.org/10.2514/1.J062012 [56] Peng, X., Kou, J., and Zhang, W., “Multi-Fidelity Nonlinear Unsteady
[39] Seydel, R., Practical Bifurcation and Stability Analysis, Vol. 5, Springer Aerodynamic Modeling and Uncertainty Estimation Based on Hierar-
Science & Business Media, New York, 2009, p. 75 (Sec. 2.6). chical Kriging,” Applied Mathematical Modelling, Vol. 121, Oct. 2023,
[40] Lanza, V., Bonnin, M., and Gilli, M., “On the Application of the pp. 1–21.
Describing Function Technique to the Bifurcation Analysis of Nonlinear https://doi.org/10.1016/j.apm.2023.05.031
Systems,” IEEE Transactions on Circuits and Systems II: Express [57] Tartaruga, I., Cooper, J., Lowenberg, M., Sartor, P., and Lemmens, Y.,
Briefs, Vol. 54, No. 4, 2007, pp. 343–347. “Uncertainty and Sensitivity Analysis of Bifurcation Loci Characteriz-
https://doi.org/10.1109/TCSII.2006.890406 ing Nonlinear Landing-Gear Dynamics,” Journal of Aircraft, Vol. 55,
[41] Tartaruga, I., Barton, D., Rezgui, D., and Neild, S., “Experimental No. 1, 2018, pp. 162–172.
Bifurcation Analysis of a Wing Profile,” Proccedings of International https://doi.org/10.2514/1.C034252
Forum on Aeroelasticity and Structural Dynamics, Curran Assoc., Inc., [58] Lee, K., Barton, D., and Renson, L., “Modelling of Physical Systems
Red Hook, NY, 2019, pp. 9–13. with a HOPF Bifurcation Using Mechanistic Models and Machine
[42] Gudmundson, P., “On The Accuracy of the Harmonic-Balance Method Learning,” Mechanical Systems and Signal Processing, Vol. 191,
Concerning Vibrations of Beams with Nonlinear Supports,” Archive of May 2023, Paper 110173.
Applied Mechanics (1991), Vol. 59, No. 5, 1989, pp. 333–344. https://doi.org/10.1016/j.ymssp.2023.110173
[59] Sieber, J., and Krauskopf, B., “Control Based Bifurcation Analysis for
Downloaded by Northwestern Polytechnic University on April 29, 2024 | http://arc.aiaa.org | DOI: 10.2514/1.J063611

https://doi.org/10.1007/BF00534063
[43] Razzak, M. A., “A Simple Harmonic Balance Method for Solving Strongly Experiments,” Nonlinear Dynamics, Vol. 51, No. 3, 2008, pp. 365–
Nonlinear Oscillators,” Journal of the Association of Arab Universities for 377.
Basic and Applied Sciences, Vol. 21, Oct. 2016, pp. 68–76. https://doi.org/10.1007/s11071-007-9217-2
https://doi.org/10.1016/j.jaubas.2015.10.002 [60] Sieber, J., Krauskopf, B., Wagg, D., Neild, S., and Gonzalez-Buelga, A.,
[44] Detroux, T., Renson, L., Masset, L., and Kerschen, G., “The Harmonic “Control-Based Continuation of Unstable Periodic Orbits,” Journal of
Balance Method for Bifurcation Analysis of Large-Scale Nonlinear Computational and Nonlinear Dynamics, Vol. 6, No. 1, 2011, Paper
Mechanical Systems,” Computer Methods in Applied Mechanics and 011005.
Engineering, Vol. 296, Nov. 2015, pp. 18–38. https://doi.org/10.1115/1.4002101
https://doi.org/10.1016/j.cma.2015.07.017 [61] Lee, K., Barton, D., and Renson, L., “Model Identification of
[45] McGurk, M., and Yuan, J., “Computation of Limit Cycle Oscillations a Fluttering Aerofoil Using Control-Based Continuation and
and their Stabilities in Nonlinear Aeroelastic Systems Using Harmonic Normal Form Analysis,” Proceedings of the International Con-
Balance Methods,” Proceedings of the International Forum on Aero- ference on Noise and Vibration Engineering, ISMA, Curran Assoc.,
elasticity and Structural Dynamics, Curran Assoc., Inc., Red Hook, NY, Inc., Red Hook, NY, 2020, pp. 261–268.
2022, pp. 1169–1186. [62] Wright, J. R., and Cooper, J. E., Introduction to Aircraft Aeroelasticity
[46] Zhou, Y., and Lu, Z., “Active Polynomial Chaos Expansion for and Loads, Vol. 20, Wiley, Hoboken, NJ, 2008, Chap. 10.
Reliability-Based Design Optimization,” AIAA Journal, Vol. 57, [63] Govaerts, W., “Numerical Bifurcation Analysis for ODEs,” Journal of
No. 12, 2019, pp. 5431–5446. Computational and Applied Mathematics, Vol. 125, Nos. 1–2, 2000,
https://doi.org/10.2514/1.J058020 pp. 57–68.
[47] Agrawal, S., Gobiha, D., and Sinha, N., “Nonlinear Parameter Estima- [64] Cameron, T., and Griffin, J. H., “An Alternating Frequency/Time
tion of Airship Using Modular Neural Network,” Aeronautical Journal, Domain Method for Calculating the Steady-State Response of Non-
Vol. 124, No. 1273, 2020, pp. 409–428. linear Dynamic Systems,” Journal of Applied Mechanics, Vol. 56, No. 1,
https://doi.org/10.1017/aer.2019.125 1989, pp. 149–154.
[48] Verma, H., and Peyada, N., “Estimation of Aerodynamic Parameters https://doi.org/10.1115/1.3176036
Near Stall Using Maximum Likelihood and Extreme Learning Machine- [65] Yuan, J., Sun, Y., Schwingshackl, C., and Salles, L., “Computation of
Based Methods,” Aeronautical Journal, Vol. 125, No. 1285, 2021, Damped Nonlinear Normal Modes for Large Scale Nonlinear Systems
pp. 489–509. in a Self-Adaptive Modal Subspace,” Mechanical Systems and Signal
https://doi.org/10.1017/aer.2020.95 Processing, Vol. 162, Jan. 2022, Paper 108082.
[49] Yuan, J., Fantetti, A., Denimal, E., Bhatnagar, S., Pesaresi, L., Schwing- https://doi.org/10.1016/j.ymssp.2021.108082
shackl, C., and Salles, L., “Propagation of Friction Parameter Uncer- [66] Scarth, C., Cooper, J. E., Weaver, P. M., and Silva, G. H., “Uncertainty
tainties in the Nonlinear Dynamic Response of Turbine Blades with Quantification of Aeroelastic Stability of Composite Plate Wings Using
Underplatform Dampers,” Mechanical Systems and Signal Processing, Lamination Parameters,” Composite Structures, Vol. 116, Sept. 2014,
Vol. 156, July 2021, Paper 107673. pp. 84–93.
https://doi.org/10.1016/j.ymssp.2021.107673 https://doi.org/10.1016/j.compstruct.2014.05.007
[50] Hosder, S., Walters, R., and Balch, M., “Efficient Uncertainty Quanti- [67] “OpenTURNS,” Airbus-EDF-IMACS-ONERA-Phimeca, 2022, https://
fication Applied to the Aeroelastic Analysis of a Transonic Wing,” 46th openturns.github.io/www/.
AIAA Aerospace Sciences Meeting and Exhibit, AIAA Paper 2008- [68] Minasny, B., and McBratney, A. B., “Spatial Prediction of Soil Proper-
0729, 2008. ties Using EBLUP with the Matérn Covariance Function,” Geoderma,
https://doi.org/10.2514/6.2008-729 Vol. 140, No. 4, 2007, pp. 324–336.
[51] Weinmeister, J., Gao, X., and Roy, S., “Analysis of a Polynomial Chaos- https://doi.org/10.1016/j.geoderma.2007.04.028
Kriging Metamodel for Uncertainty Quantification in Aerodynamics,” [69] Roberts, G. O., and Rosenthal, J. S., “Optimal Scaling for Various
AIAA Journal, Vol. 57, No. 6, 2019, pp. 2280–2296. Metropolis-Hastings Algorithms,” Statistical Science, Vol. 16, No. 4,
https://doi.org/10.2514/1.J057527 2001, pp. 351–367.
[52] Palar, P. S., Zakaria, K., Zuhal, L. R., Shimoyama, K., and Liem, R. P., https://doi.org/10.1214/ss/1015346320
“Gaussian Processes and Support Vector Regression for Uncertainty [70] Serfozo, R., Basics of Applied Stochastic Processes, Springer Science &
Quantification in Aerodynamics,” AIAA Scitech 2021 Forum, AIAA Business Media, Berlin, Heidelberg, 2009, Chap. 1.
Paper 2021-0181, 2021. [71] Beck, J. L., and Au, S.-K., “Bayesian Updating of Structural Models
https://doi.org/10.2514/6.2021-0181 and Reliability Using Markov Chain Monte Carlo Simulation,”
[53] Yildiz, S., Cakmak, E., Kara, E., and Nikbay, M., “Uncertainty Quantifi- Journal of Engineering Mechanics, Vol. 128, No. 4, 2002,
cation of Aeroelastic Systems Using Active Learning Gaussian Process,” pp. 380–391.
Proceedings of the International Conference on Multidisciplinary Design https://doi.org/10.1061/(ASCE)0733-9399(2002)128:4(380)
Optimization of Aerospace Systems (AEROBEST 2021), IDMEC Instituto [72] Betz, W., Papaioannou, I., and Straub, D., “Transitional Markov Chain
Superior Técnico Universidade de Lisboa, Portugal, 2021, pp. 21–23. Monte Carlo: Observations and Improvements,” Journal of Engineering
[54] Sun, Y., Denimal, E., Yuan, J., and Salles, L., “Geometric Design of Mechanics, Vol. 142, No. 5, 2016, Paper 04016016.
Friction Ring Dampers in Blisks Using Nonlinear Modal Analysis and https://doi.org/10.1061/(ASCE)EM.1943-7889.0001066
Kriging Surrogate Model,” Structural and Multidisciplinary Optimiza- [73] Ristic, B., Arulampalam, S., and Gordon, N., Beyond the Kalman Filter:
tion, Vol. 65, No. 3, 2022, p. 98. Particle Filters for Tracking Applications, Artech House, Norwood,
https://doi.org/10.1007/s00158-021-03093-w MA, 2003, pp. 37–41.
 MCGURK ET AL. 1905

[74] Green, P. L., Devlin, L., Moore, R. E., Jackson, R. J., Li, J., and No. 7, 2007, pp. 816–832.
Maskell, S., “Increasing the Efficiency of Sequential Monte Carlo https://doi.org/10.1061/(ASCE)0733-9399(2007)133:7(816)
Samplers Through the Use of Approximately Optimal L-Kernels,” [77] Abdelkefi, A., Vasconcellos, R., Nayfeh, A. H., and Hajj, M. R., “An
Mechanical Systems and Signal Processing, Vol. 162, Jan. 2022, Paper Analytical and Experimental Investigation into Limit-Cycle Oscilla-
108028. tions of an Aeroelastic System,” Nonlinear Dynamics, Vol. 71, No. 1,
https://doi.org/10.1016/j.ymssp.2021.108028 2013, pp. 159–173.
[75] Chib, S., and Jeliazkov, I., “Marginal Likelihood from the Metropolis– https://doi.org/10.1007/s11071-012-0648-z
Hastings Output,” Journal of the American Statistical Association, [78] Lee, K. H., “Bifurcation-Based Ordinary Differential Equation Model
Vol. 96, No. 453, 2001, pp. 270–281. Identification,” Doctoral Dissertation, Univ. of Bristol, Bristol, England,
https://doi.org/10.1198/016214501750332848 U.K., 2023, pp. 36–39.
[76] Ching, J., and Chen, Y.-C., “Transitional Markov Chain Monte Carlo
Method for Bayesian Model Updating, Model Class Selection, and S. Shin
Model Averaging,” Journal of Engineering Mechanics, Vol. 133, Associate Editor
Downloaded by Northwestern Polytechnic University on April 29, 2024 | http://arc.aiaa.org | DOI: 10.2514/1.J063611