atmosphere

Article
Intelligent Identification and Verification of Flutter Derivatives
and Critical Velocity of Closed-Box Girders Using Gradient
Boosting Decision Tree
Neyu Chen 1 , Yaojun Ge 1, * and Claudio Borri 2

1 State Key Lab of Disaster Reduction in Civil Engineering, Tongji University, Shanghai 200092, China;
chenneyu@tongji.edu.cn
2 Department of Civil and Environmental Engineering, University of Florence, 50139 Florence, Italy;
claudio.borri@unifi.it
* Correspondence: yaojunge@tongji.edu.cn

Abstract: Flutter derivatives (FDs) of the bridge deck are basic aerodynamic parameters by which
flutter analysis determines critical flutter velocity (CFV), and they are traditionally identified by
sectional model wind tunnel tests or computational fluid dynamics (CFD) numerical simulation.
Based on some wind tunnel testing results and numerical simulation data, the machine learning
models for identifying FDs of closed-box girders are trained and developed via a gradient boosting
decision tree in this study. The models can explore the underlying input–output transfer relationship
of datasets and realize rapid intelligent identification of FDs without wind tunnel tests or numerical
simulation. This method also provides a convenient and feasible option for expanding datasets of
FDs, and the distribution of FDs can be analyzed through the post-interpretation of trained models.
Combined with FD sensitivity analysis, the models can be verified by the calculation error of CFV. In
addition, the proposed method can help determine the appropriate shape of the box girder cross-
section in the preliminary design stage of long-span bridges and provide the necessary reference for
aerodynamic shape optimization by modifying the local geometric features of the cross-section.
Citation: Chen, N.; Ge, Y.; Borri, C.
Intelligent Identification and
Keywords: flutter derivative; intelligent identification; closed-box girder; machine learning; gradient
Verification of Flutter Derivatives and
boosting decision tree; sensitivity analysis
Critical Velocity of Closed-Box
Girders Using Gradient Boosting
Decision Tree. Atmosphere 2023, 14,
1165. https://doi.org/10.3390/
atmos14071165
1. Introduction
With rapid advancements in the building materials and construction techniques over
Academic Editors: Bowen Yan,
recent years, there has been an upward trend in the number of long-span bridges being
Jinhui Yan, Chao Li, Chaorong Zheng,
proposed to cross wide canyons, rivers, and straits [1]. Long-span bridges are characterized
Xiao Li and Qiusheng Li
by lighter weight, higher flexibility, and lower damping; hence, they are vulnerable to
Received: 1 June 2023 wind effects and usually dominated by the wind-resistant performance in the structural
Revised: 16 July 2023 design process. To describe the wind-resistant performance of the bridge structure, the
Accepted: 17 July 2023 aerodynamic parameters related to various wind-induced vibrations of the bridge girder
Published: 18 July 2023 need to be identified. They play a vital role in the analysis of the wind-induced static
stability, flutter, vortex-induced vibration, and buffeting of the bridge. Among them, flutter
instability, which would activate violent oscillations and could result in the collapse of
bridge structures, should be prevented at all times.
Copyright: © 2023 by the authors.
In order to predict the flutter performance of bridge deck sections, the identification of
Licensee MDPI, Basel, Switzerland.
flutter derivatives (FDs) is the most critical step. Currently, the FDs can be well identified
This article is an open access article
distributed under the terms and
using wind tunnel tests or numerical simulation based on computational fluid dynamics
conditions of the Creative Commons
(CFD). Based on the vibration state of the sectional model, there are three methods of
Attribution (CC BY) license (https:// identifying FDs in the wind tunnel tests: forced vibration [2]; free vibration [3]; and random
creativecommons.org/licenses/by/ vibration [4]. Each of these methods can be divided into two types: time domain and
4.0/). frequency domain. The coupled free vibration method has commonly been the preference to

Atmosphere 2023, 14, 1165. https://doi.org/10.3390/atmos14071165 https://www.mdpi.com/journal/atmosphere

Atmosphere 2023, 14, 1165 2 of 27

extract the bridge FDs in wind tunnel tests due to its instrumental simplicity and operational
convenience [5–7]. A series of methods were developed over the past several decades based
on the free decay time histories of the bridge deck at various wind speeds [8,9]. CFD
methods can achieve numerical calculations for various fluid problems by solving fluid
control equations. Numerical tests require an adequate load model with a turbulent
stochastic structure and a suitable structural model for each bridge in order to describe
the wind action [10]. Borri et al. established multiple wind load models for time-domain
simulation of the wind-induced vibration of bridges [11,12]. Fang et al. modeled and
simulated extreme winds and their effects in different ways [13,14]. Similar to the wind
tunnel test, all the identification methods for FDs can be also achieved by CFD [15–17].
At present, the numerical identification of FDs is mainly based on the sub-state forced
vibration method: the structure is forced to vibrate at a certain frequency in the single DoF.
Aerodynamic forces can be obtained, and the least square method is used for the exaction
of FDs.
Although the wind tunnel test is considered to be the most effective and reliable means
for the FD identification of bridges, and the advancement of computer capacity enables
more rapid simulation of CFD, many aspects of these traditional research methods still
face challenges. Traditional wind tunnel tests have the issue of high cost and can only be
used to analyze specific bridge decks. The experimental results are not universal since
these aerodynamic parameters are bridge-deck-dependent. For numerical simulation, the
process still requires a large amount of computational resources to reach a high precision.
Although CFD analysis has its potential, certain characteristics of bluff body flutter are
not yet completely understood (for example, turbulence), and there still remain great
challenges to modeling the phenomenon solely by numerical simulation. In recent years,
as an advanced alternative technique, machine learning has provided a novel solution
for the identification of aerodynamic parameters and the wind-resistant performance
analysis of bluff body sections. Machine learning is a technique of data science that helps
computers learn from existing data to forecast unknown behaviors, outcomes, and trends.
There are already some mature open-source wind engineering databases, which have
mainly been established by wind engineering groups from Canada, China, Japan, the
USA, etc. Zhou and Kareem first established an online aerodynamic database for tall
buildings in 2003 (http://aerodata.ce.nd.edu/) (accessed on 1 September 2022) [18]. Wang
and Cheng conducted high-frequency force balance wind tunnel tests on more high-rise
building models and built a wind resistance database for them based on the expert system
(http://windexpert.ce.tku.edu.tw/) (accessed on 1 September 2022) [19]. In terms of an
aerodynamic database for low-rise buildings, Ho and Kopp built the NIST aerodynamic
database from a large number of pressure measuring tests of low-rise building models
conducted at the University of Western Ontario [20]. Quan and Tamura developed a
publicly accessible database of surface wind pressure for low-rise buildings (http://www.
wind.arch.t-kougei.ac.jp/info_center/weic.html) (accessed on 1 September 2022) [21]. The
existence of these data and databases is a prerequisite for researchers to carry out wind
resistance analysis using machine learning methods.
In the past 20 years, machine learning methods have been successfully applied to
the identification of aerostatic coefficients and flutter derivatives (FDs), as well as to the
analysis of wind-resistant stability. Jung et al. took the lead in realizing the estimation
of six FDs of a rectangular section using an artificial neural network (ANN) based on
17 sets of experimental data [22]. The ANN was utilized by Chen et al. to predict eight
FDs of bluff body sections [23]. The support vector machine (SVM) was introduced by
Lute et al. to identify FDs of the main girder before estimating the critical flutter velocity
(CFV) of cable-stayed bridges [24]. Chung et al. used CFD simulations and the forced
vibration test in a wind tunnel to estimate eight FDs of a rectangular section by the back
propagation neural network (BP neural network) [25]. The ANN model based on the
Levenberg–Marquardt algorithm was derived and trained using various geometrical and
mechanical parameters of the bridge deck cross-section, and CFV of suspension bridge with
 parameters of the bridge deck cross-section, and CFV of suspension bridge with clos
box deck sections can be predicted directly [26]. Abbas et al. used the normalized lift fo
and torsional moment coefficients at current time step as the output of ANN to pred
the aeroelastic response of bridge decks [27]. Mei proposed a machine learning strate
Atmosphere 2023, 14, 1165 for flutter prediction based on four widely used machine learning algorithms 3 of[28].
27 Li et
applied ANNs to establish the relationship between aerostatic coefficients and flut
performance for fast prediction of CFV [29]. All these studies confirm the feasibility a
closed box deck sections can be predicted directly [26]. Abbas et al. used the normalized lift
effectiveness of machine learning methods in the FDs identification and flut
force and torsional moment coefficients at current time step as the output of ANN to predict
performance analysis
the aeroelastic response [30]. However,
of bridge decks [27].these
Mei attempts
proposed a are not learning
machine really strategy
independent
specially designed
for flutter wind
prediction tunnel
based tests.
on four Current
widely research
used machine lacks algorithms
learning the integration
[28]. Liand da
mining of the existing results of wind tunnel tests or numerical simulations.flutter
et al. applied ANNs to establish the relationship between aerostatic coefficients and Furthermo
performance for fast prediction of CFV [29]. All these studies confirm the feasibility and
most of the existing research describes a direct application of common machine learni
effectiveness of machine learning methods in the FDs identification and flutter performance
methods
analysisto [30].
a specific
However, scenario in parameter
these attempts identification
are not really independentor wind resistance
of specially designed analy
without
windillustrating
tunnel tests. the applicability
Current research lacksof the algorithms
the integration andand improving
data-mining them
of the according
existing
The existing machine
results of wind tunnellearning applications
tests or numerical are mostly
simulations. employed
Furthermore, most oftothe
build a black b
existing
lacking the interpretation and extension of the identification model.
research describes a direct application of common machine learning methods to a specific
scenario in parameter identification or wind resistance analysis without illustrating the
The objective
applicability foralgorithms
of the this paper andisimproving
to utilizethempreviously
accordingly.measured experimental
The existing machine resu
that contain invaluableare
learning applications information
mostly employed about flutter
to build mechanics.
a black box, lackingApproximations
the interpretation based
the cumulated
and extension existing data are expected
of the identification model. to produce acceptable results, and imprecisi
tolerance of the gradient boosting decision
The objective for this paper is to utilize tree (GBDT)
previously is a beneficial
measured characteristic
experimental results wh
that contain invaluable information about flutter mechanics. Approximations based on
solving the problems of parameter identification without big data. In this study, the F
the cumulated existing data are expected to produce acceptable results, and imprecision
of closed-box
tolerance ofgirders are boosting
the gradient trained decision
and predicted
tree (GBDT)by isGBDT based
a beneficial on a specialized
characteristic when brid
windsolving
resistance performance
the problems database,
of parameter and without
identification the prediction
big data. Inresults arethe
this study, applied
FDs to t
related analysis of flutter performance. Figure 1 shows the technology roadmap of t
of closed-box girders are trained and predicted by GBDT based on a specialized bridge
study.wind resistance performance database, and the prediction results are applied to the related
analysis of flutter performance. Figure 1 shows the technology roadmap of this study.

FigureFigure
1. Technology roadmap.
1. Technology roadmap.

2. Identification Methods of Flutter Derivatives

2. Identification Methods of Flutter Derivatives
This section introduces the traditional identification methods for FDs. It begins with
This section
physical introduces
identification, the traditional
including identification
the theoretical definition of methods
FDs and theforwind
FDs.tunnel
It begins w
physical identification, including the theoretical definition of FDs and the wind tunnel t
test measurement methods. This is followed by numerical identification, i.e., numerical
simulation by CFD methods to extract the FDs.
measurement methods. This is followed by numerical identification, i.e., numeri
simulation by CFD
2.1. Physical methods to extract the FDs.
Identification
For a bridge deck that is immersed in the incoming air flow, the oscillation was mainly
2.1. Physical Identification
driven by the self-excited force. The mechanical system of flutter is described by a two
degrees of freedom (DoFs) linear oscillator, free to vibrate in heaving h(t) and pitching
For a bridge
θ(t) modes, deckinthat
as shown is 2immersed
Figure in the incoming
(where U represents the incomingair flow,
wind theL and
speed; oscillation
M w
mainly drivenself-excited
represent by the self-excited
lift force andforce.
torque,The mechanical system of flutter is described b
respectively).
two degrees of freedom (DoFs) linear oscillator, free to vibrate in heaving h(t) and pitchi
θ(t) modes, as shown in Figure 2 (where U represents the incoming wind speed; L and
represent self-excited lift force and torque, respectively).
Atmosphere 2023, 14, x FOR PEER REVIEW 4 of 2
Atmosphere 2023, 14, 1165 4 of 27

Figure2.2.Reference
Figure Reference system
system for for displacements
displacements and self-excited
and self-excited forces. forces.

Based on the study performed by Scanlan and Tomko, the wind-induced linear self-
Based on the study performed by Scanlan and Tomko, the wind-induced linear self
excited force on a bridge deck is modeled as the function of the vibration state, i.e., dis-
excited force
placements and on a bridge
velocities deck
in two is using
DoFs modeled
eight as
FDsthe
[31]:function of the vibration state, i.e.
displacements and velocitieshin two
.
DoFs using
.
eight FDs [31]:i
Lse = 12 ρU 2 B KH1∗ Uh + KH ∗ Bθ + K 2 H ∗ θ + K 2 H ∗ h ,
1 h2  . * h 2 U . * Bθ 3 2 * 4 B i2 * h  (1)
Lse1=ρU 2ρBU2 KA
B  ∗KH
h 1 +∗KHBθ 2 2 +∗ K H 32θ +∗ Kh H4 
Mse = 2 2  1 U +UKA 2 U +K U A3 θ + K A4 B B,
(1

where U is the wind speed; ρ1is the2 air2 density; 
h B is theBθwidth2of*the bridge 
hdeck; K is the
M se = ρU B  KA1* + KA2* + K A3θ + K 2 A4* 
ωB/U which is the reduced 2
frequency of
 U
oscillation; ω U
is circular frequency B
of the coupled
mode; and Hi∗ and Ai∗ (i = 1,2,3,4) are the FDs.
where U is study,
In this the windthe speed;
unifying ρ is thesquare
least air density;
method B is the widthbyofGu
developed theetbridge and K is the
al. [32]deck;
ωB/U which
improved is theetreduced
by Ding al. [6], Lifrequency
et al. [33] and Bartoli et al.ω[34]
of oscillation; is circular frequency
was utilized of the
to extract the coupled
FDs in wind ∗
tunnel ∗ which was referred to as the modified unifying least square
tests,
mode; and 𝐻𝐻𝑖𝑖 and 𝐴𝐴𝑖𝑖 (i = 1,2,3,4) are the FDs.
(MULS) approach.
In this study, the unifying least square method developed by Gu et al. [32] and
improved byIdentification
2.2. Numerical Ding et al. [6], Li et al. [33] and Bartoli et al. [34] was utilized to extract the
FDs The
in wind tunnel tests,
FDs calculated which
by CFD was simulation
numerical referred to
in as
thisthe modified
paper are basedunifying least square
on the forced
(MULS) approach.
vibration method. It avoids solving the motion equation, which saves time and is easy
to be realized by software (ANSYS Fluent 2021 R1). The two degrees of freedom of cross-
2.2. Numerical
section Identification
motion are decoupled, and the model is designated to perform vertical and torsional
motions respectively.
The FDs calculatedBasedby on CFD
Scanlan’s theoretical
numerical framework
simulation in of thepaper
this superposition
are basedof on the
linear self-excited forces, it is assumed that the model performs multi-frequency forced
forced vibration method. It avoids solving the motion equation, which saves time and i
vibration of a single degree of freedom:
easy to be realized by software (ANSYS Fluent 2021 R1). The two degrees of freedom o
cross-section motion are decoupled,
h(t) = ∑and
n the model is designated to perform vertical and

i =1 h0 sin(2π f i t ) , (2)
torsional motions respectively. α(t) =Based
∑in=1 α0on Scanlan’s
sin(2π f i t) theoretical framework of the
superposition of linear self-excited forces, it is assumed that the model performs multi
where h0 and α0 are the vertical bending and torsional amplitudes. The FDs are obtained
frequency forced vibration of a single degree of freedom:
by identifying the amplitude of aerodynamic force and the phase difference between force

h(t ) = ∑ i =1 h0 sin(2π fi t ) ,
 domainn method.
and displacement through the frequency
 Derivatives (2
α (t ) = ∑ i =1α 0 sin(2π f i t )
3. Intelligent Identification of Flutter n

The machine learning algorithm applied to predict the FDs is introduced in this section.
Typical
where machine
h0 and αlearning methods mainly include decision trees, random forests, artificial
0 are the vertical bending and torsional amplitudes. The FDs are obtained
neural networks, and Bayesian learning. In the case of big data, the accuracy that can be
by identifying the amplitude of aerodynamic force and the phase difference between force
achieved by existing machine learning methods shows insignificant difference. However,
and
in thedisplacement
face of a small through
amount ofthe frequency
data, domain
it is necessary method.various machine learning
to compare
algorithms before selecting the most favorable one. The comparison of various algorithms
3. Intelligent
has Identification
been performed of Flutter
in our previous study Derivatives
[35], showing that the error back propagation
(EBP) The machine
neural network learning algorithm
presents serious danger applied
of fallingtointo
predict the FDs since
local minimum is introduced
it is based in thi
section. Typical machine learning methods mainly include decision trees, random
on a gradient descent method. Support vector regression (SVR) requires very rigorous data forests
pre-processing and hyperparameters tuning, so the model is difficult to check and adjust.
artificial neural networks, and Bayesian learning. In the case of big data, the accuracy tha
Gradient boosting decision tree (GBDT) has the advantages of good training effect and less
can be achieved by existing machine learning methods shows insignificant difference
overfitting on low-dimensional data, and its framework allows post-interpreters to be run
However,
for in the face ofand
model representation a small amount of data, it is necessary to compare various machine
improvement.
learning algorithms before selecting the most favorable one. The comparison of variou
algorithms has been performed in our previous study [35], showing that the error back
propagation (EBP) neural network presents serious danger of falling into local minimum
since it is based on a gradient descent method. Support vector regression (SVR) require
very rigorous data pre-processing and hyperparameters tuning, so the model is difficul
 to check and adjust. Gradient boosting decision tree (GBDT) has the advantages of good
training effect and less overfitting on low-dimensional data, and its framework allows
post-interpreters
Atmosphere 2023, 14, 1165
to be run for model representation and improvement. 5 of 27
A hybrid model combining GBDT and the linear regression method is therefore used
for the training of FDs in this study. Its main idea is to use weak classifiers (decision trees)
to iteratively train input data before obtaining the optimal model. The architecture of the
A hybrid model combining GBDT and the linear regression method is therefore used
model is shownfor in the
Figure 3. Input features are transformed by means of GBDT which
training of FDs in this study. Its main idea is to use weak classifiers (decision trees)
consists of m decision trees.
to iteratively trainThe input
number ofbefore
data trees obtaining
and the number of branches
the optimal model. Theand leaf
architecture of the
nodes per tree aremodel is shown in Figure 3. Input features are transformed by means ofofGBDT which
determined based on the amount of input data. In the construction
the decision tree, if it cannot
consists further improve
of m decision trees. The thenumber
generalization ability
of trees and theof the model,
number the
of branches and leaf
creation of leaf nodes
nodes will be stopped.
per tree To avoid
are determined overfitting,
based a threshold
on the amount can also
of input data.beInset
thetoconstruction
limit the creationof theof decision
leaf nodes,tree, which is called
if it cannot furthertoimprove
pre-pruning. After successfully
the generalization ability of the model,
constructing a complete
the creationdecision tree,
of leaf the will
nodes sub-trees can be checked
be stopped. To avoidby replacingathem
overfitting, with can also be
threshold
leaf nodes; then,settheto sub-trees will be of
limit the creation replaced if the
leaf nodes, model
which can betoimproved
is called withAfter
pre-pruning. leaf successfully
nodes, which is constructing
called post-pruning.
a completeDuring
decisionthe whole
tree, training can
the sub-trees process of GBDT,
be checked a new them with
by replacing
tree is added toleafthenodes;
existing trees
then, theto reduce will
sub-trees the beresiduals
replaced ofifthe
the current
model can model in the with leaf
be improved
gradient direction, and the negative gradient of the loss function is taken as GBDT,
nodes, which is called post-pruning. During the whole training process of the a new tree
is added to the existing trees to reduce the residuals of the
approximate value of the residual to replace the residual. Finally, the transformed features current model in the gradient
direction, and the negative gradient of the loss function is taken as the approximate value
wi (i = 0, 1, 2,…, n) are trained in a linear regression model to obtain the final prediction
of the residual to replace the residual. Finally, the transformed features wi (i = 0, 1, 2, . . . , n)
results.
are trained in a linear regression model to obtain the final prediction results.

Figure 3. Architecture of3.GBDT.

Figure Architecture of GBDT.

3.1. Negative Gradient Fitting Gradient Fitting

3.1. Negative
GBDT is one of GBDT is oneused
the widely of themodels
widelyunder
used models under boosting
the gradient the gradient boosting framework.
framework. It It
accomplishes the task of classification by defining a logarithmic loss function for logistic
accomplishes the task of classification by defining a logarithmic loss function for logistic
regression. To solve the problem of measuring the approximation of the loss function,
regression. To solve the problem of measuring the approximation of the loss function, the
the gradient boosting algorithm uses the negative gradient of the loss function to fit the
gradient boosting algorithm uses the negative gradient of the loss function to fit the
approximation of the loss and fits a classification and regression tree (CART) so that the
approximation of the
loss inloss
eachand
roundfitsbecomes
a classification
as smalland regression tree (CART) so that the
as possible.
loss in each round becomes as small
The minimum as possible.
value of the loss function is:
The minimum value of the loss function is:
Cmj = argmin∑ L(yi , f m−1 ( xi ) + c) ( x ∈ Rmj ), (3)
= ∑
Cmj arg min L( yi , f m−1 ( xi ) + c) (x ∈ Rmj ) , (3)
where the function L is a logarithmic loss function; xi is the eigenvalue of the input; yi is
the output; and c is a constant. Rmj is the zone of the decision tree j domain; and i, m, j are
counting variables.
Atmosphere 2023, 14, 1165 6 of 27

The decision tree fitting function is:

J
hm ( x ) = ∑ Cmj I (x) ( x ∈ Rmj ), (4)
j =1

where the function I(x) is an indicator function that returns 0 when the equation in paren-
theses is false; otherwise, it returns to 1. J is the number of iterations.
The updated boosting decision tree is the sum of the previously fitting decision trees
and the latest fitting function:

f ( x ) = f m −1 ( x ) + h m ( x ). (5)

3.2. GBDT Regression

GBDT can be used for classification and regression, and the principles of both are
similar. A weak learner is generated in each round, and a strong predictive model is
achieved through continuous accumulation. Based on the above definitions of negative
gradient loss function and update process, the entire GBDT regression process is expressed
as follows.
1. Initialize loss function:

N
1 1+y
f 0 ( x ) = argmin ∑ L(yi , c) = log , (6)
i =1
2 1−y

where f 0 (x) is the initial value of the algorithm; N denotes the length of the training
set; y represents the mean of outputs; and the others are defined as above.
2. For each number of iteration rounds, m = 1, 2, . . . , M.
For the training set (i = 1, 2, . . . , N,) compute the negative gradient which is also the
residual rmi :  
∂L(yi , f ( xi ))
rmi = − . (7)
∂ f ( xi ) f ( x )= f m−1 ( x )

3. Fit a decision tree with rmi , obtain the leaf node region Rmj of the m-th tree.
For j = 1, 2, . . . , J, use linear search to obtain the minimum value of loss function:

Cmj = argmin∑ L(yi , f m−1 ( xi ) + c) ( x ∈ Rmj ). (8)

4. Update f (x):
J
f ( x ) = f m −1 ( x ) + ∑ Cmj I (x) (x ∈ Rmj ). (9)
j =1

3.3. Algorithm Optimization

The GBDT model has the advantages of robustness, applicability to low-dimensional
problems, and fast tuning. In the case of small samples, GBDT has poor generalization and
usually faces the problem of overfitting. To solve this issue, a regular term is added to the
original loss function to improve the generalization of the trained model. The loss function
after adding the regular term is shown as follows:

N
L0 = ∑ L ( y i , f m −1 ( x i ) + c ) + ∑ Ω ( g m ), (10)
n =1 m

where Ω( g) = δT + 12 λkwk2 ; δ and λ are hyperparameters; T is the number of leaf nodes;

and w is a vector consisting of all leaf node values.
Atmosphere 2023, 14, 1165 7 of 27

4. Datasets and Data Processing for Intelligent Identification

This section begins with the introduction of a specialized wind resistance performance
database for long-span bridges. Next, two types of datasets for intelligent identification of
FDs are presented, including hybrid datasets and pure numerical simulation datasets. The
selection of wind tunnel test data from the database, the collected numerical simulation data
from literature, and the supplementary data in the hybrid datasets are introduced. Then,
the pure numerical simulation datasets obtained through CFD recalculation are described.

4.1. Bridge Wind Resistance Performance Database

Although there are some mature open-source wind engineering databases in the world,
these databases are all specific to building structures and are self-contained. Even if they
can be used for bridge wind resistance, the data types may not be matched completely.
Therefore, a specialized bridge wind resistance performance database is developed in this
study before performing the machine learning of parameters of interest. This database
collected the test data of nearly 100 long-span bridges from the wind tunnel laboratory of
Tongji University since a large quantity of wind tunnel test data for long-span bridges have
been accumulated at Tongji University up to now, and various experimental data have
been unified to the same criterion.
This database uses Microsoft Access as the underlying database system. To facilitate
the users’ visual operation, the Java programming language is adopted to establish the
foreground application interface. The connection between the underlying database and
the foreground application is driven by a local protocol to achieve platform independence
and high execution efficiency [36]. After classification, all the data are summarized into
three modules: basic information; dynamic characteristics; and aerodynamic parameters.
The basic information includes the type of bridge, span, the material of girder, the type
and size of cross-section, mass, stiffness, basic wind speed, critical velocity, amplitude, etc.
The dynamic characteristics include the natural frequency and mode of vibration of each
bridge. Aerodynamic parameters include the aerostatic coefficients and flutter derivatives
(FDs), but aerodynamic admittance is not involved.

4.2. Closed-Box Girder Dataset

This subsection introduces the hybrid datasets of closed-box girders, which include
20 sets of wind tunnel test data from the database, 20 sets of collected numerical simulation
data from the open-source literature, and 14 sets of supplementary data.

4.2.1. Physical Identification Data through Wind Tunnel Tests

Specifically, FDs and the flutter performance of bridges using closed-box girder are
extracted and studied. For machine learning modeling, it is necessary to establish a
specialized sample set of closed-box girders, including the dimensions of the cross-section,
the wind attack angle, the reduced wind velocity, FDs, etc. The wind tunnel test data of
20 long-span bridges with streamlined closed-box girders are selected from the database as
mentioned above. The specific involved parameters include the dimensions of cross-section,
Atmosphere 2023, 14, x FOR PEER REVIEW 8 of 29
the FDs under different reduced wind velocities, and the airflow field conditions of wind
tunnel tests. The dimensions of the cross–section are shown in Figure 4, including the
width of deck—B, box height of beam—H, wind fairing extension length—b, wind fairing
length—b,and
angle—α, wind fairingweb
inclined angle—α, and inclined
slope—β. web
These five slope—β. can
parameters These five parameters
determine can
the unique
determine the unique closed-box girder section, and they are independent of each
closed-box girder section, and they are independent of each other. The specific dimensions other.
The
of specific
these dimensions
20 sets of these are
of cross-sections 20 sets of cross-sections
illustrated in Table 1,are illustrated
and the FDs in
of Table
these 1,
20and
setsthe
of
FDs of these 20
cross-sections sets
are of cross-sections
given in Figure 5. are given in Figure 5.

Figure 4.
Figure 4. Architecture
Architecture of
of aa closed-box
closed-box girder.
girder.

3.2 1.5

2.4 1.0

0.5
1.6
 Atmosphere 2023, 14, 1165 8 of 27

Table 1. Geometric size of 20 sets of cross-sections collected for wind tunnel test.

Section No. B (m) H (m) b (m) α (◦ ) β (◦ )

1 35.2 3.535 1.79 40 15
2 28.5 3.0 0.97 60 8
3 30.0 3.5 2.0 39 14
4 46.7 5.83 1.66 63 18
5 40.8 3.5 0.6 72 13
, 14, x FOR PEER REVIEW 6 33.5 2.875 2.9 50 8 of 26 20
7 36.2 3.8 0.9 71 20
8 40.0 3.74 2.25 48 21
9 45.2 3.0 1.15 64.8 18.6
10 43.0 2.0 1.0 55.8 10.8
fairing angle—α, and inclined
11 web 30.0
slope—β. These 5.0 five parameters
5.5 can determine
49 the 24
unique closed-box girder 12 section, and
36.9they are independent
3.6 of1.0
each other. The66 specific 14
13 of cross-sections
dimensions of these 20 sets 43.0 3.0
are illustrated 1.0 1, and the FDs
in Table 62 of these 16
14 30.4 3.5 0.6 73 23
20 sets of cross-sections 15
are given in 18.0
Figure 5. 2.5 1.8 48 22
16 50.0 2.5 0.8 60 10
17 44.0 4.5 1.0 79 26
18 33.5 3.0 1.7 48.1 19.9
19 38.0 4.0 1.3 59 15
20 28.6 3.5 0.7 72.6 17.6
Range 18.0–50.0 2.0–5.8 0.6–5.5 39–79 8–26
Figure 4. Architecture of a closed-box girder.

3.2 1.5

2.4 1.0

0.5
1.6
No.1 No.11 0.0
0.8 No.2 No.12
A*4
A*1

No.3 No.13 -0.5

0.0 No.4 No.14
No.5 No.15 -1.0
-0.8 No.6 No.16 No.1 No.6 No.11 No.16
No.7 No.17 -1.5 No.2 No.7 No.12 No.17
No.8 No.18 No.3 No.8 No.13 No.18
-1.6 -2.0
No.9 No.19 No.4 No.9 No.14 No.19
No.10 No.20 No.5 No.10 No.15 No.20
-2.4 -2.5
0 5 10 15 20 25 30 35 0 5 10 15 20 25 30
U/fB U/fB
3 6

0 3

-3 0

-6 -3
H*4
H*1

-9 -6

-12 No.1 No.6 No.11 No.16 -9 No.1 No.6 No.11 No.16

No.2 No.7 No.12 No.17 No.2 No.7 No.12 No.17
-15 No.3 No.8 No.13 No.18 -12 No.3 No.8 No.13 No.18
No.4 No.9 No.14 No.19 No.4 No.9 No.14 No.19
No.5 No.10 No.15 No.20 No.5 No.10 No.15 No.20
-18 -15
0 5 10 15 20 25 30 0 5 10 15 20 25 30
U/fB U/fB
0.3 7
Figure 5. Cont. No.1 No.11
No.2 No.12
0.0 6
No.3 No.13
No.4 No.14
5
-0.3 No.5 No.15
No.6 No.16
No.1 No.11 4 No.7 No.17
-0.6 No.2 No.12
No.8 No.18
A*2

A*3

No.3 No.13 3 No.9 No.19

No.4 No.14
-0.9 No.5 No.15
No.10 No.20
2
No.6 No.16
-1.2 No.7 No.17
1
No.8 No.18
No.9 No.19
-1.5 0
 H

H
-9 -6

-12 No.1 No.6 No.11 No.16 -9 No.1 No.6 No.11 No.16

No.2 No.7 No.12 No.17 No.2 No.7 No.12 No.17
-15 No.3 No.8 No.13 No.18 -12 No.3 No.8 No.13 No.18
No.4 No.9 No.14 No.19 No.4 No.9 No.14 No.19
Atmosphere 2023, 14, 1165 No.5 No.10 No.15 No.20 No.5 No.10 No.15 No.20 9 of 27
-18 -15
0 5 10 15 20 25 30 0 5 10 15 20 25 30
U/fB U/fB
0.3 7
No.1 No.11
No.2 No.12
0.0 6
No.3 No.13
No.4 No.14
5
-0.3 No.5 No.15
No.6 No.16
No.1 No.11 4 No.7 No.17
-0.6 No.2 No.12
No.8 No.18
A*2

A*3
No.3 No.13 3 No.9 No.19
No.4 No.14
-0.9 No.5 No.15
No.10 No.20
2
No.6 No.16
-1.2 No.7 No.17
1
No.8 No.18
No.9 No.19
-1.5 0
No.10 No.20
0 5 10 15 20 25 0 5 10 15 20 25
U/fB U/fB
4.8 5
No.1 No.6 No.11 No.16
4.0 No.2 No.7 No.12 No.17
0
No.3 No.8 No.13 No.18
3.2 No.4 No.9 No.14 No.19
No.5 No.10 No.15 No.20 -5
2.4 No.1 No.11
-10
No.2 No.12
1.6
H*2

H*3

No.3 No.13
-15 No.4 No.14
0.8 No.5 No.15
-20 No.6 No.16
0.0 No.7 No.17
-25 No.8 No.18
-0.8 No.9 No.19
No.10 No.20
-1.6 -30
0 5 10 15 20 0 5 10 15 20 25
U/fB U/fB

Figure 5. 8 FDs under different

Figure reduced
5. 8 FDs wind speeds
under different of 20
reduced setsspeeds
wind of cross-sections from database.
of 20 sets of cross-sections from database.

4.2.2. Numerical Identification Data Based on CFD Simulation

Table 1. Geometric size of 20 sets of cross-sections collected for wind tunnel test.
To improve the machine learning effect, another 20 sets of numerical simulation data
Section No. forBclosed-box
(m) H (m)
girders b (m)literature were
from open-source α (°)added as mixed
β (°)datasets to jointly
drive the training process of machine learning. The involved parameters are consistent with
the wind tunnel test data. The specific dimensions of these 20 sets of data are illustrated in
Table 2, and the FDs of these 20 sets of cross-sections are given in Figure 6.

Table 2. Geometric size of the collected 20 sets of cross-sections for numerical simulation.

Section No. B (m) H (m) b (m) α (◦ ) β (◦ )

21 [37] 38.5 4.39 1.25 65.3 12.26
22 [38] 33.9 3.0 1.8 47.7 21.14
23 [39] 12.0 2.0 0.9 50.8 22.2
24 [40] 29.7 3.9 0.15 101.4 17.84
25 [41] 27.0 4.3 2.0 53.6 27.1
26 [42] 45.2 3.1 2.44 36.4 9.8
27 [43] 35.5 3.5 3.1 34.7 18.5
28 [44] 29.9 3.0 0.8 79.1 19.95
29 [45] 26.0 2.8 1.7 40.9 15.7
30 [46] 35.5 3.5 2.18 44 21
31 [47] 35.0 3.5 1.8 48.6 20.77
32 [48] 41.0 3.5 0.5 76.7 13.12
33 [49] 23.2 3.36 1.92 48.1 19.28
34 [50] 35.3 3.54 4.2 30 15.6
35 [51] 45.2 3.0 1.152 65 19
36 [52] 31.1 2.8 1.8 43.1 16.5
37 [53] 39.0 2.0 3.0 37 18
38 [54] 38.0 3.5 1.3 59.1 17.2
39 [55] 35.5 3.54 3.12 38.2 18.15
40 [56] 35.9 3.5 1.3 59.5 11.42
Range 12.0–45.2 2.0–4.39 0.15–4.2 30–101.4 9.8–27.1
Atmosphere
Atmosphere2023,
2023,14,
14,x 1165
FOR PEER REVIEW 10 10
of of2627

2.5 1.0

2.0
0.5
1.5
0.0
1.0

A*4
A*1
0.5 -0.5

0.0
-1.0
No.21 No.26 No.31 No.36 No.21 No.26 No.31 No.36
-0.5 No.22 No.27 No.32 No.37 No.22 No.27 No.32 No.37
No.23 No.28 No.33 No.38 -1.5 No.23 No.28 No.33 No.38
-1.0 No.24 No.29 No.34 No.39 No.24 No.29 No.34 No.39
No.25 No.30 No.35 No.40 No.25 No.30 No.35 No.40
-1.5 -2.0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
U/fB U/fB
1 2
No.21 No.31
0 No.22 No.32
No.23 No.33
-1 0
No.24 No.34
-2 No.25 No.35
No.26 No.36
-3 No.27 No.37 -2 No.21 No.31
No.28 No.38 No.22 No.32
H*1

-4

H*4
No.29 No.39 No.23 No.33
-5 -4 No.24 No.34
No.30 No.40
No.25 No.35
-6 No.26 No.36
-6 No.27 No.37
-7
No.28 No.38
-8 No.29 No.39
No.30 No.40
-9 -8
0 5 10 15 20 25 30 35 0 5 10 15 20 25 30 35

U/fB U/fB
0.5 10
No.21 No.31
0.0 No.22 No.32
8 No.23 No.33
-0.5 No.24 No.34
-1.0 No.25 No.35
No.21 No.31 6 No.26 No.36
-1.5 No.27 No.37
No.22 No.32
A*2

No.28 No.38
A*3

No.23 No.33 4
-2.0 No.29 No.39
No.24 No.34
No.30 No.40
-2.5 No.25 No.35
No.26 No.36 2
-3.0 No.27 No.37
No.28 No.38
-3.5 No.29 No.39 0
No.30 No.40
-4.0 0 5 10 15 20 25
0 5 10 15 20 25
U/fB U/fB
14 5
No.21 No.31
12 No.22 No.32 0
No.23 No.33
10 No.24 No.34 -5
No.25 No.35
8 No.26 No.36 -10 No.21 No.31
No.27 No.37
No.22 No.32
No.28 No.38
H*2

H*3

6 -15 No.23 No.33

No.29 No.39
No.24 No.34
4 No.30 No.40 -20 No.25 No.35
No.26 No.36
2 -25 No.27 No.37
No.28 No.38
0 -30 No.29 No.39
No.30 No.40
-2 -35
0 5 10 15 20 25 0 5 10 15 20 25
U/fB U/fB

Figure6.6.8 8FDs
Figure FDsunder
underdifferent
differentreduced
reducedwind
windspeeds
speedsofof20
20sets
setsofofcross-sections
cross-sectionsfrom
fromliterature.
literature.

4.2.3. Supplementary Data

Table 2. Geometric size of the collected 20 sets of cross-sections for numerical simulation.
This study focuses on the machine learning prediction of flutter derivatives and flutter
Section No. of closed-box
performance B (m) H (m)
girders. The value brange
(m) of shapeαparameters
(°) ofβthe
(°) cross-
21 [37]
sections 38.5set will have4.39
in the sample 1.25on the prediction
a decisive impact 65.3 results. The
12.26existing
40 22
sets of data are33.9
[38] not uniform in3.0 terms of dimensions
1.8 such 47.7
as the width and
21.14height,
etc.,
23of the cross-section.
[39] 12.0 Moreover,
2.0 considering0.9the actual requirements
50.8 of 22.2
the bridge
Atmosphere 2023, 14, 1165 11 of 27

structure, each aerodynamic shape parameter has a concentrated distribution range, so the
representativeness of the existing data is also not good enough. For example, most of the
wind fairing angles used in projects are between 45◦ and 60◦ . Most of the inclined web
slopes are between 15◦ and 25◦ , and 11◦ and 32◦ are also often used in engineering [57].
Based on the above considerations, an additional 14 sets of cross-sections were added in this
study, whose size distributions remain within these concentrated distribution ranges and fill
in the gaps in some size ranges. The specific dimensions of these 14 sets of supplementary
data are illustrated in Table 3, and the FDs of these 14 sets of cross-sections are given in
Figure 7 by CFD numerical simulation. Therefore, a total of 54 sets of closed-box girder
cross-sections form a complete sample set for machine learning training. The distribution
of the whole sample set tends to be closer to a Gaussian distribution than a uniform
distribution, as shown in Figure 8, with the aim of making subsequent machine learning
applications more engineering-practical.

Table 3. Geometric size of 14 sets of supplementary cross-sections.

Section No. B (m) H (m) b (m) α (◦ ) β (◦ )

41 30.5 3.0 1.7 57 28
42 29.5 3.0 1.5 57 17
43 34.3 3.2 1.5 50 13
44 26.2 2.8 0.7 70 20
45 25.5 3.0 1.5 51 19
46 51.4 5.0 1.5 57 14
47 54.0 5.8 2.0 65 15
48 32.8 3.2 0.6 60 12
49 39.1 4.6 2.1 53 23.5
50 42.8 4.5 2.0 53 17.5
51 41.8 4.4 2.0 53 16.5
52 58.1 4.3 1.9 43 16.5
53 56.7 4.2 1.9 43 11.5
Atmosphere 2023, 14, x FOR PEER REVIEW 54 59.5 4.1 1.8 37 13 of 28
16.5
Range 25.5–59.5 2.8–5.8 0.6–2.1 37–70 12–28

1.2 0.25
No.41 No.46 No.51 No.41 No.46 No.51
No.42 No.47 No.52 No.42 No.47 No.52
1.0 No.43 No.48 No.53 0.20 No.43 No.48 No.53
No.44 No.49 No.54 No.44 No.49 No.54
No.45 No.50 No.45 No.50
0.8 0.15

0.10
A*4

0.6
A*1

0.4 0.05

0.2 0.00

0.0 -0.05
2 4 6 8 10 12 2 4 6 8 10 12
U/fB U/fB
0.0 0.8
-0.5 0.6
-1.0 0.4
-1.5
0.2
-2.0
0.0
H*1

H*4

-2.5
-0.2
-3.0
No.41 No.46 No.51 No.41 No.46 No.51
-3.5 No.42 No.47 No.52 -0.4 No.42 No.47 No.52
No.43 No.48 No.53 No.43 No.48 No.53
-4.0 No.44 No.49 No.54 -0.6 No.44 No.49 No.54
No.45 No.50 No.45 No.50
-4.5 -0.8
2 4 6 8 10 12 2 4 6 8 10 12
U/fB U/fB
0.1 2.4
Figure 7. Cont. No.41 No.46 No.51
2.1 No.42 No.47 No.52
0.0 No.43 No.48 No.53
1.8 No.44 No.49 No.54
-0.1 No.45 No.50
1.5
-0.2
1.2
A*3
A*2
 -1.5
-2.0
0.2

H*1 H1

H*4 H*4
0.0

*
-2.0
-2.5
0.0
-2.5 -0.2
-3.0
No.41 No.46 No.51 -0.2 No.41 No.46 No.51
-3.0
-3.5 No.42 No.47 No.52 -0.4 No.42 No.47 No.52
No.41 No.46 No.51 No.41 No.46 No.51
No.43 No.48 No.53 -0.4 No.43 No.48 No.53
-3.5 No.42 No.47 No.52 -0.6 No.42 No.47 No.52
-4.0 No.44 No.49 No.54 No.44 No.49 No.54
No.43 No.48 No.53 No.43 No.48 No.53
Atmosphere 2023, 14, 1165 -4.0 No.45
No.44
No.50
No.49 No.54 -0.6 No.45
No.44
No.50
No.49 No.54 12 of 27
-4.5 -0.8
2 No.454 6
No.50 8 10 12 2 No.45 4 No.506 8 10 12
-4.5 -0.8
2 4 6 U/fB 8 10 12 2 4 6 U/fB8 10 12
U/fB 2.4 U/fB
0.1
2.4 No.41 No.46 No.51
0.1 2.1 No.42 No.47 No.52
No.41 No.46 No.51
0.0 No.43 No.48 No.53
2.1 No.42 No.47 No.52
0.0 1.8 No.44
No.43 No.49
No.48 No.54
No.53
-0.1 1.8 No.45
No.44 No.50
No.49 No.54
-0.1 1.5 No.45 No.50
-0.2 1.5
1.2

A*3A*3
A*2A2
*
-0.2
1.2
-0.3 0.9
-0.3 0.9
-0.4 No.41 No.46 No.51 0.6
-0.4 No.42
No.41 No.47
No.46 No.52
No.51 0.6
-0.5 No.43
No.42 No.48
No.47 No.53
No.52 0.3
No.44
No.43 No.49
No.48 No.54
No.53 0.3
-0.5
No.45
No.44 No.50
No.49 No.54 0.0
-0.6 No.45 No.50 0.0
-0.6 2 4 6 8 10 12 2 4 6 8 10 12
2 4 6
U/fB 8 10 12 2 4 6
U/fB
8 10 12
U/fB U/fB
0.9 2
0.9 No.41 No.46 No.51 2
No.41
No.42 No.46
No.47 No.51
No.52
0.6 No.42 No.47 No.52 0
0.6 No.43 No.48 No.53
No.43
No.44 No.48
No.49 No.53
No.54 0
0.3 No.44
No.45 No.49
No.50 No.54
0.3 No.45 No.50 -2
-2
0.0
0.0
* 2

3 3
-4
*

* *
HH

HH
2

-0.3 -4
-0.3
-6
-0.6 -6 No.41 No.46 No.51
-0.6 No.41
No.42 No.46
No.47 No.51
No.52
-8 No.42
No.43 No.47
No.48 No.52
No.53
-0.9 -8 No.43 No.48 No.53
-0.9 No.44 No.49 No.54
No.44 No.49 No.54
No.45 No.50
-1.2 -10 No.45 No.50
-1.2 2 4 6 8 10 12 -10 2 4 6 8 10 12
2 4 6 8 10 12 2 4 6 8 10 12
U/fB
U/fB U/fB
U/fB

7.7.888FDs
Figure7.
Figure
Figure FDsunder
FDs underdifferent
under differentreduced
different reducedwind
reduced windspeeds
wind speedsof
speeds ofof14
14sets
14 setsof
sets ofofsupplementary
supplementary
supplementary cross-sections.
cross-sections.
cross-sections.

16
16 12
12
Experimental
Experimental data
data Experimentaldata
Experimental data
14
14 Numerical
Numerical simulation data
simulation data Numericalsimulation
Numerical simulationdata
data
Supplementary 10
10 Supplementarydata
data
Supplementary data
data Supplementary
12
12
10 88
10
Count
Count

Count
Count

88 66
66
44
44
Atmosphere 2023, 14, x FOR PEER REVIEW 22 22 13 of 26
00 00
66 88 10
10 12
12 14
14 16 18 20 22 0.0 0.2
0.0 0.2 0.4
0.4 0.6
0.6 0.8
0.8 1.0
1.0 1.2
1.2 1.4
1.4 1.6
1.6
B/H
B/H b/H
b/H
9 14
Experimental data Experimental data
8 Numerical simulation data Numerical simulation data
12
7 Supplementary data Supplementary data
10
6
8
Count

5
Count

4 6
3
4
2
2
1
0 0
30 40 50 60 70 80 90 100 6 8 10 12 14 16 18 20 22 24 26 28 30
α (°) β (°)

Cross-sectiondistribution
Figure8.8.Cross-section
Figure distributionhistogram
histogramofofsample
sampleset.
set.

4.3. Two-Dimensional CFD Numerical Simulation

Table 3. Geometric size of 14 sets of supplementary cross-sections.
In this paper, it was found through preliminary trial calculations that good prediction
Section No. not be
results could B (m) H (m)
obtained when driven by 20b pure
(m) sets of wind
α (°)tunnel testβdata
(°) or by
20 sets
41 of numerical simulation data
30.5 3.0 due to too 1.7
little data. Under
57 the mixed driving
28 of
two data
42 modes, there 29.5 may be significantly
3.0 different
1.5 underlying 57mapping relationships
17
43 34.3 3.2 1.5 50 13
44 26.2 2.8 0.7 70 20
45 25.5 3.0 1.5 51 19
46 51.4 5.0 1.5 57 14
 fore, the FDs of the existing 40 sets of cross-sections, as well as the 14 sets of supplemen-
tary
Atmosphere cross-sections,
2023, 14, 1165 are recalculated by CFD numerical simulation uniformly to obtain a13 of 27
new sample set more suitable for machine learning.
The numerical calculation domain is a two-dimensional flow field. Pointwise is used
between similar sections due to differences in recognition methods, making it difficult for
for geometric rendering and mesh generation, and ANSYS Fluent is used for numerical
machine learning models to distinguish them and obtain good prediction results. Therefore,
simulation. The Reynolds
the FDs number of each
of the existing sectional
40 sets rigid model
of cross-sections, as wellis the same
as the 14 setsas of
that of the
supplementary
corresponding wind tunnel test.areInrecalculated
cross-sections, this paper, by the
CFDFDs at 0°simulation
numerical wind attack angletoofobtain
uniformly eacha new
cross-section are calculated, and all the numerical simulation results are compared with
sample set more suitable for machine learning.
The numerical calculation domain is a two-dimensional flow field. Pointwise is used
the wind tunnel test results or calculation results by other researchers in the literature to
for geometric rendering and mesh generation, and ANSYS Fluent is used for numerical
check the validity ofsimulation.
the data.The Reynolds number of each sectional rigid model is the same as that of the
Figure 9 showscorresponding
the results of wind8 FDs
tunnelfortest.
twoInofthisthepaper,
54 sets
the ofFDscross-sections
at 0◦ wind attack changing
angle of each
with the reduced wind speed under
cross-section 0° wind
are calculated, attack
and all theangle. Figure
numerical 9a shows
simulation resultsthearecomparison
compared with the
between the numerical
windsimulation results
tunnel test results and the wind
or calculation resultstunnel
by other test results in
researchers ofthe
cross-section
literature to check
the validity of the data.
1 from the database (asFigure
shown in Table
9 shows 1). Figure
the results 9b for
of 8 FDs shows
two of the
thecomparison between changing
54 sets of cross-sections the
numerical simulationwithresults in this
the reduced windstudy
speed (CFD1)
under 0◦ and
wind other researcher’s
attack angle. Figure 9acalculation re-
shows the comparison
sults (CFD2) of cross-section 21 from simulation
between the numerical the literatureresults(as
andshown
the windintunnel
Table test2)results
[37]. ofItcross-section
can be 1
seen from the figures that the numerical simulation results of these two closed-box girders
from the database (as shown in Table 1). Figure 9b shows the comparison between the
numerical simulation results in this study (CFD1) and other researcher’s calculation results
are in good agreement with the wind tunnel test results and those of other researcher’s
(CFD2) of cross-section 21 from the literature (as shown in Table 2) [37]. It can be seen from
calculations. This verification
the figures thatprocess is applicable
the numerical simulation toresults
all other collected
of these cross-sections.
two closed-box girders are In
in good
the following, theseagreement
numerical simulation
with data will
the wind tunnel be used
test results andfor machine
those of other learning
researcher’s training
calculations.
and prediction so asThistoverification
form an process
intelligent identification
is applicable method
to all other collectedfor FDs of closed-box
cross-sections. In the following,
girders. these numerical simulation data will be used for machine learning training and prediction
so as to form an intelligent identification method for FDs of closed-box girders.
2.5 2
A*1−CFD A*1−Experiment
2.0 A*4−CFD A*4−Experiment
0

1.5
-2
A*1, A*4

H*1, H*4

1.0

-4
0.5

-6
0.0 H*1−CFD H*1−Experiment
*
H −CFD
4
H*4−Experiment
-0.5 -8
2 4 6 8 10 12 14 16 18 2 4 6 8 10 12 14 16 18

U/fB U/fB
3.0 1
A2* −CFD A*2−Experiment
0
2.5
A*3−CFD A*3−Experiment
-1
2.0
-2
1.5
-3
H*2, H*3
A*2, A*3

1.0 -4

-5
0.5
-6
0.0
-7
H*2−CFD H*2−Experiment
-0.5
-8 *
H −CFD
3
H*3−Experiment
-1.0 -9
0 2 4 6 8 10 12 0 2 4 6 8 10 12

U/fB U/fB
(a)

Figure 9. Cont.
 -5
0.5
-6
0.0
-7
H*2−CFD H*2−Experiment
-0.5
-8
H*3−CFD H*3−Experiment
-1.0 -9
0 2 4 6 8 10 12 0 2 4 6 8 10 12
Atmosphere 2023, 14, 1165 14 of 27
U/fB U/fB
(a)
2.0 2
A*1−CFD1 A*1−CFD2
1.6 A*4−CFD1 A*4−CFD2
0

1.2
-2
A*1, A*4

H*1, H*4
0.8
-4

0.4
-6
H*1−CFD1 H*1−CFD2
0.0 *
H −CFD1
4
H*4−CFD2
-8
0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20

U/fB U/fB
8
0
6

4 -5

2
H*2, H*3
A *2, A*3

-10
0

-2 -15

-4
-20
A*2−CFD1 A*2−CFD2 H*2−CFD1 H*2−CFD2
-6
A*3−CFD1 A*3−CFD2 H*3−CFD1 H*3−CFD2
-8 -25
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20

U/fB U/fB
(b)

Figure 9. Trend of FDs changing with reduced wind speed: (a) the comparison between the numerical
simulation results and the wind tunnel test results of cross-section 1; (b) the comparison between the
numerical simulation results in this study (CFD1) and other researcher’s calculation results (CFD2) of
cross-section 21.

5. Intelligent Identification for FDs of Closed-Box Girders

This section focuses on the intelligent identification of FDs by GBDT. Firstly, the input
and output of the machine learning model are introduced, and the data division pattern is
also explained. Next, the training and prediction results of the model are presented. Finally,
a post-interpretation is provided to further validate the effectiveness of the model.

5.1. Input and Output

When the machine learning method is used to study the aerodynamic characteristics
of the bridge, the section geometry is a key factor, and the geometric dimensions of the
cross-section must be used as the input for the neural network. For a closed-box girder, as
shown in Figure 3, five parameters determine the unique cross-section, including B, H, b,
α, and β. Since the amount of data is not big in this study, without being able to increase
the amount of data significantly, the dimensionality of the input parameters should be
reduced as much as possible. Considering that FDs are dimensionless parameters that are
only related to the shape of the cross-section and are independent of the absolute size of
the cross-section, four shape feature parameters (B/H, b/H, α, β), as well as reduced wind
velocity, are finally used as input parameters for the model. The outputs of the model are
FDs, but only one FD at a single reduced wind velocity is output each time. In addition, all
inputs and outputs need to be pre-processed for normalization to reflect the essence of the
data and avoid numerical problems before conducting machine learning training.
A regression analysis was performed before the machine learning training by creating
an equation to determine how the dependent variable varies with multiple independent
Atmosphere 2023, 14, 1165 15 of 27

variables and testing the equation and parameters for statistical significance. The T-test
is used here for significance testing of the parameters, which can also be regarded as a
parameter sensitivity analysis. When judging the regression analysis results, the effect of
the corresponding variable is significant when |t| > 2, and the probability of error will not
exceed 0.05. If |t| is much greater than 2, the probability of error is even smaller. Table 4
gives the t-statistics of each input feature for the output parameter (FD). It can be seen that
most of the values of |t| are greater than 2, which proves that when the machine learning
method is used to study the aerodynamic characteristics of bridges, the dimensions of
cross-section and wind velocity are very important factors, and they must be used as inputs
of the model.

Table 4. T-test of regression analysis.

t-Statistic A*1 A*4 H*1 H*4 A*2 A*3 H*2 H*3

B/H 2.539 1.237 −2.950 −3.798 −3.066 2.635 3.267 −2.977
b/H −2.351 1.601 1.573 2.648 2.468 −1.829 −1.812 1.989
α −4.517 1.407 1.632 2.681 2.802 −1.210 −1.079 1.277
β 4.147 −2.478 −1.033 −1.609 −2.339 0.847 0.380 −0.815
U/fB 115.834 47.381 −82.835 −27.241 −37.922 38.996 5.875 −37.282

5.2. Model Construction and Trial Calculation

The GBDT method used to identify FDs in this study is based on Python and its
toolkit. The programming process is conducted on Jupyter Notebook which is a web-based
interactive computing platform. After establishing the preliminary model, 20 sets of pure
wind tunnel test data and 54 sets of hybrid data were used for trial calculations to evaluate
the basic performance of the GBDT model. One of the evaluation indices of the model is
fitting degree (R2 statistic), which is defined as follows:
2
∑in=1 (yi − ŷi )
R2 (y, ŷi ) = 1 − 2
, (11)
∑in=1 (yi − y)

where yi is the true value of the sample; ŷi is the predicted value; and n is the number of
samples. R2 reflects the proportion of the variation in the dependent variables that can be
explained by the independent variables through the regression relationship, and R2 takes a
value between [0,1]. The larger the R2 , the better the model.
To judge the prediction effect of trained machine learning models, the extrapolation
ability of the test cross-section should also be evaluated, which is also called the generaliza-
tion ability evaluation. The prediction accuracy can be reflected by the mean relative error
(MRE), which is defined as follows:

1 n yi − ŷi
n i∑
MRE(y, ŷ) = , (12)
=1
yi

where yi is the true value of the sample; ŷi is the predicted value; and n is the number of
samples. The smaller the value of MRE, the more accurate the prediction.
Taking cross-section 1 as the test section, Figures 10 and 11 show the fitting degree
and generalization ability of the model under two different datasets, and Table 5 gives
the specific prediction errors. It can be seen that the GBDT model struggles to capture the
potential distribution pattern of FDs under only 20 sets of wind tunnel test data. It is still
unable to achieve very satisfactory prediction results under 54 sets of hybrid data after
repeated hyperparameters tuning. To further test and improve the performance of the
GBDT model, pure numerical simulation data is considered for subsequent model training
and optimization.
 and generalization ability of the model under two different datasets, and Table 5 gives the
specific prediction errors. It can be seen that the GBDT model struggles to capture the
potential distribution pattern of FDs under only 20 sets of wind tunnel test data. It is still
unable to achieve very satisfactory prediction results under 54 sets of hybrid data after
repeated hyperparameters tuning. To further test and improve the performance of the
Atmosphere 2023, 14, 1165 16 of 27
GBDT model, pure numerical simulation data is considered for subsequent model train-
ing and optimization.

1.00 0.973 1.00 0.963 0.955 0.952 0.959

0.927
0.891 0.923 0.9210.938
0.855 0.834
0.8 0.824
0.748 0.716
0.75 0.75

2
R

R
0.50 0.50

0.25 0.25

0.00 * *
0.00 * *
* * * *
A*1 A*4 H1 H*4 A2 A3 H*2 H3 A*1 A*4 H1 H*4 A2 A3 H*2 H3

(a) (b)
Atmosphere 2023, 14, x FOR PEER REVIEW 17 of 27
Figure 10. Fitting
Figure degreedegree
10. Fitting of training set: (a) based
of training onbased
set: (a) 20 setson
of wind tunnel
20 sets test data;
of wind tunnel(b)test
based on 54
data; (b) based on
sets of hybrid data.
54 sets of hybrid data.
1 1
1.8
1.8
A* −Predicted value 0
A*1−Predicted
1
value 0
1.5 *
A −Predicted value
1.5 A*4−Predicted
4
value -1
*
A1−True value -1
1.2 *
A −True value
A* −True value -2
1.2 1

H*1 , H*4
A*1 , A*4

4
0.9
*
A −True value -2
-3

H , H*4
A*1 , A*4

4
0.9 -3
0.6 -4 H*1−Predicted value

1
*
0.6 -5 -4 H*4−Predicted
H*1−Predicted
value value
0.3
H*1−True value
H*4−Predicted value
-6 -5
0.3 0.0
H*4−True
* value
-7 H1−True value
2 4 6 8 10 12 14 16 -6
2 4 6 8 10 12 14 16
0.0 H*4−True value
U/fB -7 U/fB
2 4 6 8 10 12 14 16
2 4 6 8 10 12 14 16
2.0 1
U/fB
A*2−Predicted value U/fB
1.6 A*3−Predicted value 0

A*2−True value -1
1.2
A*3−True value
3

H*2 , H*3

-2
*
A ,A

0.8
-3
2
*

H*2−Predicted value
0.4
-4 H*3−Predicted value
0.0 -5 H*2−Ture value
H*3−Ture value
-0.4 -6
2 4 6 8 10 12 14 16 2 4 6 8 10 12 14 16

U/fB U/fB
(a)
1
1.8
*
A −Predicted value
1 0
1.5 *
A −Predicted value
4
* -1
A −True value
1.2 1
*
A −True value -2
H*1 , H*4

4
4
*
A ,A

0.9 -3
1
*

0.6 -4 H*1−Predicted value

-5 H*4−Predicted value
0.3
H*1−True value
-6
0.0 H*4−True value
-7
2 4 6 8 10 12 14 16
2 4 6 8 10 12 14 16
U/fB U/fB
2.0 1
A*2−Predicted value
1.6
*
A −Predicted value 0
3
*
A −True value
2 -1
1.2 *
A −True value
3
3

H*2 , H*3

-2
*
A ,A

0.8
-3
2
*

H*2−Predicted value
0.4
-4 H*3−Predicted value
0.0 -5 H*2−Ture value
H*3−Ture value
-0.4 -6
2 4 6 8 10 12 14 16 2 4 6 8 10 12 14 16

U/fB U/fB
(b)
Figure 11. Prediction results of cross-section 1: (a) based on 20 sets of wind tunnel test data (model
Figure 11. Prediction results of cross-section 1: (a) based on 20 sets of wind tunnel test data (model
1); (b) based on 54 sets of hybrid data (model 2).
1); (b) based on 54 sets of hybrid data (model 2).
Table 5. Prediction errors of cross-section 1.

MRE Mean
*
A1 A4* H1* H4* A2* A3* H2* H3*
Model 1 0.1656 0.5174 0.1491 0.4644 0.2105 0.2171 0.6673 0.2825 0.3342
Model 2 0.0595 0.2033 0.1150 0.2750 0.1480 0.1624 0.5310 0.1702 0.2081
 Atmosphere 2023, 14, 1165 17 of 27

Table 5. Prediction errors of cross-section 1.

MRE A*1 A*4 H*1 H*4 A*2 A*3 H*2 H*3 Mean
Model 1 0.1656 0.5174 0.1491 0.4644 0.2105 0.2171 0.6673 0.2825 0.3342
Model 2 0.0595 0.2033 0.1150 0.2750 0.1480 0.1624 0.5310 0.1702 0.2081

5.3. Model Optimization and Validation

Atmosphere 2023, 14, x FOR PEER REVIEW
The ultimate goal of machine learning is to deploy trained models into real environ-
ments with the hope that the models can achieve good prediction effect in real scenarios.
In this subsection, the training and prediction of GBDT models are based on pure numer-
and stability of
ical simulation theand
data, model
furtherthrough different
analysis of dataperformance
GBDT model division patterns. Each sample
is given through
used, in turn, as the training set, validation set, and test set. The cross validation m
complete cross validation.
can
5.3.1.help
Data to monitor the changes of training effect under different patterns and
Division
overfitting problems.
The sample set of machine learning is usually divided into three parts: training set;
There
validation are
set; anda total ofIn54this
test set. sets of pure
paper, crossnumerical
validation issimulation
used to evaluatedatathe
inaccuracy
this study. Sin
samples were
and stability farmodel
of the fromthrough
the mean of the
different datadatasets
division(samples in the
patterns. Each dashed
sample box in Fi
will be
i.e., cross-section 10, 16, 24, and 37), they were removed from the sample set accor
used, in turn, as the training set, validation set, and test set. The cross validation method
can help to monitor the changes of training effect under different patterns and avoid
the statistical
overfitting results. A total of 50 sets of data were available in the sample set. T
problems.
sets of data
There arewere
a totalused as the
of 54 sets testnumerical
of pure set by turns, resulting
simulation data inin 50study.
this workingSince conditions.
four
each working condition, we can take any one of the remaining 49 sets of8,samples
samples were far from the mean of the datasets (samples in the dashed box in Figure
validation set, and
i.e., cross-section then
10, 16, 24, all
andthe37),remaining sets as from
they were removed the training
the sample set,
setresulting
according in 49 w
to the statistical results. A total of 50 sets of data were available in the sample set. These
conditions. This will ultimately result in a total of 50 × 49 = 2450 working conditio
50 sets of data were used as the test set by turns, resulting in 50 working conditions. Under
GBDT models are trained
each working condition, we canunder
take anyallone
working conditions,
of the remaining from
49 sets which as
of samples the
thebest and
working
validation conditions
set, and thenare displayed
all the remainingbelowsets as to
theverify
trainingthe
set,applicability
resulting in 49of the existing
working
set for machine learning modeling under all working conditions.
conditions. This will ultimately result in a total of 50 × 49 = 2450 working conditions. The
GBDT models are trained under all working conditions, from which the best and worst
working conditions are displayed below to verify the applicability of the existing sample
5.3.2. Fitting Accuracy of Training Set
set for machine learning modeling under all working conditions.
According to the data division pattern mentioned above, machine learning
5.3.2. Fitting Accuracy of Training Set
were trained under 2450 working conditions. It can be found that the GBDT model
According to the data division pattern mentioned above, machine learning models
fits
werethe training
trained underset with
2450 a high
working accuracy,
conditions. regardless
It can of the
be found that data division
the GBDT pattern us
model always
average training
fits the training effect
set with of FDs
a high are shown
accuracy, regardlessinofFigure
the data12. It canpattern
division be seen from
used. Thethe figu
all the models
average can obtain
training effect of FDs aretheshown
potential input–output
in Figure transfer
12. It can be seen relationship
from the figure that allof the t
set. All values
the models can
2
of R are larger than 0.9 except
obtain2 the potential input–output
∗
for
transfer
∗ H 2 and H 4 .
relationship
* of
* the training set.
All values of R are larger than 0.9 except for H2 and H4 .

0.985 0.991 0.966 0.960 0.960

1.00
0.900 0.894
0.828
0.75
R2

0.50

0.25

0.00
A*1 A*4 H*1 H*4 A*2 A*3 H*2 H*3

Figure 12.Average
Figure 12. Average fitting
fitting degree.
degree.

5.3.3. Generalizability of Test Set

To verify the generalizability of models for the test cross-sections, the predicte
is compared with the true value, and Figure 13 gives the best and worst pre
Atmosphere 2023, 14, 1165 18 of 27

5.3.3. Generalizability of Test Set

To verify the generalizability of models for the test cross-sections, the predicted value
is compared with the true value, and Figure 13 gives the best and worst prediction outcomes
for all test sets (predicted value means the predicted FD by the trained model; true value
means the recalculated CFD result). Table 6 gives the specific errors of the best and worst
Atmosphere 2023, 14, x FOR PEER REVIEW 19 of 26
predictions. It shows that the machine learning models are able to predict the distribution
of FDs to a large extent under the current data condition.
1
1.2
A*1−Predicted value
1.0 0
A*4−Predicted value
0.8 A*1−True value
-1
A*4−True value

H*1 , H*4
A*1 , A*4

0.6
-2
0.4
H*1−Predicted value
-3
0.2
H*4−Predicted value
0.0 -4 H*1−True value
H*4−True value
-0.2 -5
2 4 6 8 10 12
2 4 6 8 10 12
U/fB U/fB
2.4 1
A*2−Predicted value
2.0 0
*
A −Predicted value
3
1.6 *
-1
A −True value
2
* -2
1.2 A −True value
3
H*2 , H*3
A *2 , A *3

-3
0.8
-4
0.4 H*2−Predicted value
-5
0.0 H*3−Predicted value
-6
H*2−Ture value
-0.4 -7
H*3−Ture value
-0.8 -8
2 4 6 8 10 12 2 4 6 8 10 12

U/fB U/fB
(a)
1.0
0.6
A*1−Predicted value
0.5 0.5
A*4−Predicted value
*
0.4 A −True value
1
0.0
*
A −True value
H*1 , H*4

4
-0.5
A*1 , A*4

0.3

0.2 -1.0
H*1−Predicted value
0.1 -1.5 H*4−Predicted value
0.0 -2.0 H*1−True value
H*4−True value
-0.1 -2.5
0 1 2 3 4 5 6 7
0 1 2 3 4 5 6 7
U/fB U/fB
0.6 0.3
A*2−Predicted value
0.5 * 0.0
A −Predicted value
3
0.4 *
A −True value
2 -0.3
*
0.3 A −True value
3
H*2 , H*3
A *2 , A *3

-0.6
0.2
-0.9
0.1 H*2−Predicted value
-1.2 H*3−Predicted value
0.0
-1.5 H*2−Ture value
-0.1
H*3−Ture value
-0.2 -1.8
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7

U/fB U/fB
(b)
Figure
Figure13.
13.Generalization
Generalization ability of of
ability testtest
set:set:
(a) the bestbest
(a) the prediction results
prediction (cross-section
results 29); (b)
(cross-section the(b) the
29);
worst prediction results (cross-section 26).
worst prediction results (cross-section 26).

5.4. Model Interpretation

The process of machine learning is often considered a “black box”. The output is only
obtained from the input, and the middle training process is difficult to explain. To express
the machine learning model in some explicit way, the Shapley additive explanation
Atmosphere 2023, 14, 1165 19 of 27

Table 6. Prediction errors of test set.

MRE A*1 A*4 H*1 H*4 A*2 A*3 H*2 H*3 Mean
Best 0.0584 0.1745 0.0662 0.1924 0.0985 0.1089 0.2042 0.0833 0.1233
Worst 0.1755 0.4128 0.0910 0.1851 0.1411 0.1573 0.2400 0.1730 0.1970

5.4. Model Interpretation

The process of machine learning is often considered a “black box”. The output is only
obtained from the input, and the middle training process is difficult to explain. To express
the machine learning model in some explicit way, the Shapley additive explanation (SHAP)
is applied here. SHAP is an additive model interpreter constructed by Lundberg and Lee
and inspired by cooperative game theory, which focuses on calculating
Atmosphere 2023, 14, x FOR PEER REVIEW 20 of 26 the SHAP values
of each feature as a reflection of how much the feature contributes to the prediction of the
model [58]. SHAP interprets the predictive value of the model as the sum of SHAP values
and Lee and
of each inspired
input by cooperative game theory, which focuses on calculating the SHAP
feature:
values of each feature as a reflection of how much the feature contributes
M to the prediction

values of each input feature:

fi , ∑
of the model [58]. SHAP interprets the predictive value of the model as the sum of SHAP
ŷ = f 0 + (13)
i =1
M
where ŷ is the predictive value; 0 
y = f + f isf the
i i
i =1
, SHAP value of each feature; (13) and f0 is the mean
predictive value of all training samples. SHAP reflects the impact of each feature on the
where y is the predictive
final prediction andvalue;
can show
fi is the theSHAP positivity
value of each and negativity
feature; and f0 is of thethe
mean impact, increasing the
interpretability
predictive value of allof the model.
training samples. SHAP It also provides
reflects the impactpowerful data on
of each feature visualization
the functions to
final
show prediction
the model and can and show
thetheprediction
positivity and negativity
results, whichof theare
impact,
widelyincreasing
usedthe to explain models with
interpretability of the model. It also provides powerful data visualization functions to
complex algorithms.
show the model and the prediction results, which are widely used to explain models with
complex Figure
algorithms.14 shows the SHAP values of shape parameters of the cross-section mentioned
aboveFigure (H,14 B, b, α,theand
shows SHAP for the
β)values prediction
of shape parameters results of the model.
of the cross-section They are ranked from top
mentioned
above (H, B, b,according
to bottom α, and β) for the prediction
to the effectresults
magnitude of the model.on eachThey are
FD.ranked
Overall, from except
top H2∗ , the influence
to bottom according to the effect magnitude on each FD. Overall, except H 2* , the influence
of these shape parameters on the other FDs is obvious. H, α, and β play a major role in the
of these shape parameters on the other FDs is obvious. H, α, and β play a major role in the
FDs,
FDs, butbut the influence
the influence directiondirection and the degree
and the contribution contribution
of different degree of different shape parameters
shape parameters
∗ ∼ A∗ , H has the greatest effect on them. They have the
totothetheFDs FDs are different.
are different. For A1* ~ForA 4* , A
H 1has the 4greatest effect on them. They have the
same
same change
change direction
direction as H. his
as H. his means A1* ~ means
A 4* increase A1∗with∼ theA4∗height
increase
of beam,withbut theA1* height of beam, but
∗
, AA12 , and ∗
A2 , Aand ∗
areAmore ∗
likelyAto
3 arelikely
more likely
to be positive,to be andpositive,
A 4 is moreand 4 is
be more
negative. likely
For to be negative. For
* * *
3
* ∗ * , the∗impact of β is large. Except that the situation of * ,∗
HH ~
1 1 H ∼4
H 4 , the impact of β is large. Except that Hthe situation
* is unclear,
2 H *,
1
of H H
3 2
is unclear, H1∗ , H3∗ , and
and
HH
∗ ∗
H1 ,HH ∗ ∗
4 4 all change in the opposite direction to i.e., 4 3 , and H4 all decrease as the inclined
* all change in the opposite direction to β, i.e., , H * , and * all decrease as the
H 1*β, 3

inclined
web slope web slope increases,
increases, andandititisismore
more likely
likelythat that they are all
they arenegative. The specific
all negative. The specific contribution
contribution degree of these shape parameters to the FDs is given in a semi-quantitative
degree of these shape parameters to the FDs is given in a semi-quantitative form (SHAP
form (SHAP values). SHAP analysis enables the interpretability of black box models, and
itvalues). SHAPthe
can even provide analysis
necessaryenables
reference the interpretability
for further of black
explicit representation of box models, and it can even
the train-
provide
ing process the necessary
of machine reference for further explicit representation of the training process of
learning.
machine learning.

(a) A1* (b) A 4*

(c) H 1* (d) H 4*

Figure 14. Cont.

Atmosphere 14, 1165
2023,2023,
Atmosphere 14, x FOR PEER REVIEW 21 of 26 20 of 27

(e) A 2* (f) A3*

(e) A 2* (f) A3*

(g) H 2* (h) H 3*
Figure 14. SHAP model explanation.

2
(g) H* (h) H 3*
6. Prediction to Flutter Performance Analysis
Figure 14. SHAP model explanation.
Figure 14. SHAP model explanation.
The existing data-driven model can perform the identification of FDs out of physica
6. Prediction to Flutter Performance Analysis
and numerical wind tunnel
6. Prediction to Flutter Performancetests to some extent. It can also provide a convenient and fea
Analysis
The existing data-driven model can perform the identification of FDs out of physical
sible
and option
The
numerical windfor
existing expanding
data-driven
tunnel datasets
tests to somemodel
extent. Itcan
canofperform
aerodynamic
also the
provide parameters.
identification
a convenient The
and fea-of FDs outFDs are dimension
of physical
less parameters related to the shape of the cross-section.
and numerical wind tunnel tests to some extent. It can also provide a convenient and feasi-
sible option for expanding datasets of aerodynamic parameters. The FDs are When
dimension- the shape of the cross
less parameters related to the shape of the cross-section. When the shape of the cross-
section
ble optionchanges, the FDs
for expanding will inevitably
datasets of aerodynamic be affected.
parameters.In the
section changes, the FDs will inevitably be affected. In the process of calculating the CFV
Theprocess
FDs areofdimensionless
calculating the CFV
parameters
ofof a long-span
a long-span related to
bridge,
bridge, by the
modifying shape
bythe of
modifyingthe cross-section.
local size of thethe local size
cross-section, When
theof the shape of
the cross-section,
influence of the the cross-section
the influence of the
changes,
section
sectionshape the
shape FDs
on the onwill
CFV thebeinevitably
can evaluated.
CFV can At
bebe
the affected. InAt
same time, the
evaluated. the
machine
the process
learning
same ofmodel
calculating
time, the CFV
the machine of a mode
learning
can be reasonably
long-span verifiedby
bridge, andmodifying
optimized. the local size of the cross-section, the influence of the
can be reasonably verified and optimized.
section shape on the CFV can be evaluated. At the same time, the machine learning model
6.1. Intelligent Identification of Flutter Derivatives
canThis
be paper
reasonably verified and optimized.
takes the Runyang Yangtze River Bridge (south branch) as an example,
6.1. Intelligent Identification of Flutter Derivatives
which is a suspension bridge with the main span of 1490 m. Figure 15 shows the cross-
6.1. Intelligent
section diagram of Identification
This paper maintakes
girder, the
andof itRunyang
Flutter Derivatives
is a steel Yangtzeclosed-box
streamlined River Bridge (south
branch) as an example
girder with a
height of 3 m and a width of 36.9 m.
whichThisis paper takes thebridge
a suspension RunyangwithYangtze Riverspan
the main Bridge
of (south branch)
1490 m. as 15
Figure an shows
example,
the cross
which is a suspension bridge with the main span of 1490 m. Figure 15 shows
section diagram of main girder, and it is a steel streamlined closed-box girder with athe cross-
section
height diagram of main
of 3 m and girder,
a width ofand
36.9it m.
is a steel streamlined closed-box girder with a height
of 3 m and a width of 36.9 m.
Figure 15. Cross-section diagram of Runyang Bridge.

The trained GBDT models under the best working condition are used to obtain the
FDs of this main girder cross-section. The comparison of predictions with CFD calcula-
tions and experimental results are shown in Figure 16, and the specific prediction error
(MRE) is shown in Table 7.
Figure 15.Cross-section
Figure15. Cross-sectiondiagram of Runyang
diagram Bridge.
of Runyang Bridge.
Table 7. Prediction error of Runyang Bridge.
The trained GBDT models under* the best working condition are used to obtain the
A 1*
MREThe trained
*
A 4GBDT H 1* models
H 4* A2
under A 3* best
the H 2* working
H 3* condition
Mean are used to obtain the
FDs
CFD
of this main girder cross-section. The comparison
0.0082 0.1401 0.1383 0.3958 0.0746 0.0435 0.2094 0.0109 0.1520
of predictions with CFD calculations
FDs of this
and experimental main girder
results cross-section.
are shown The comparison of predictions with CFD is calcula
Experiment 0.1606 0.2113 0.1950 0.2174 in Figure
0.1812 16, and
0.1828 0.2070the specific
0.1445 prediction error
0.1875 (MRE)
tions and
shown experimental
in Table 7. results are shown in Figure 16, and the specific prediction error
(MRE) is shown in Table 7.
Table 7. Prediction error of Runyang Bridge.
Table 7. Prediction error* of Runyang* Bridge.
MRE A*1 A*4 * H1 H4 A2 A*3 H*2 H*3 Mean
* * * * * * * *
CFD 0.0082 0.1401 MRE0.1383 A 0.3958 A 1 H H A4 A 0.0109H
0.07461 0.0435
4 20.2094 3 2 H3
0.1520 Mean
Experiment 0.1606 0.2113 CFD 0.1950 0.0082 0.1401 0.1383 0.3958 0.0746 0.04350.1445
0.2174 0.1812 0.1828 0.2070 0.2094 0.1875
0.0109 0.1520
Experiment 0.1606 0.2113 0.1950 0.2174 0.1812 0.1828 0.2070 0.1445 0.1875
6.2. Sensitivity Analysis of Critical Flutter Velocity
This subsection analyzes the influence of FD prediction error on FCV calculation error
and further analyzes the main source of error via sensitivity analysis of eight FDs.
3, 14, x FOR PEER REVIEW 22 of 26
Atmosphere 2023, 14, 1165 21 of 27

A*1−Experiment A*1−CFD A*1−Prediction H*1−Experiment H*1−CFD H*1−Prediction

* *
A*4−Experiment A*4−CFD A*4−Prediction H −Experiment
4
H −CFD
4
H*4−Prediction
1.2 1
1.0 0
0.8
-1
A1 , A4

0.6

H1 , H4
*

*
-2
0.4
*

*
-3
0.2
0.0 -4

-0.2 -5
2 4 6 8 10 12 2 4 6 8 10 12
U/fB U/fB
A*2−Experiment A*2−CFD A*2−Prediction H*2−Experiment H*2−CFD H*2−Prediction
* *
*
A −Experiment
3
*
A −CFD
3
*
A −Prediction
3
H −Experiment
3
H −CFD
3
H*3−Prediction
0.8 0.5
0.0
0.6
-0.5
0.4
A2 , A3

-1.0
H2 , H3
*

-1.5
*

0.2
-2.0
0.0
-2.5
-0.2 -3.0
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8
U/fB U/fB
Figure
Figure 16. Prediction Prediction
16. of
results FDs for results
Runyangof FDs for Runyang Bridge.
Bridge.

6.2.1. Prediction
6.2. Sensitivity Analysis of Critical Error of Velocity
Flutter Flutter Derivatives
As there are some errors in the FDs predicted by machine learning; we need to deter-
This subsection analyzes the influence of FD prediction error on FCV calculation er-
mine the extent to which these errors affect the CFV calculations before the aerodynamic
ror and further analyzes the mainanalysis.
shape sensitivity source of error
Using via
the sensitivity
predicted FDs analysis
of Runyangof eight
BridgeFDs.
for the 2-D 3-DoF
numerical analysis [59], the CFV of this structure at 0◦ wind attack angle can be obtained.
6.2.1. Prediction Error of Flutter
Actually, several Derivatives
machine learning models based on different types of datasets (pure wind
tunnelerrors
As there are some test datasets,
in the pure
FDs numerical
predictedsimulation
by machine datasets, hybrid
learning; wedatasets,
need to etc.) were trained
deter-
in this study. The FDs prediction results with different errors
mine the extent to which these errors affect the CFV calculations before the aerodynamic can be obtained by different
trained models. These errors are ultimately reflected in the CFV calculation, as shown in
shape sensitivity analysis. Using the predicted FDs of Runyang Bridge for the 2-D 3-DoF
Figure 17. It can be seen that even with an error of up to 25% in FDs prediction, the error
numerical analysis [59], the CFV of this structure at 0° wind attack angle can be obtained.
in CFV calculation is only 10%. When the prediction error of FDs can be reduced to less
Actually, several machine
than 15%,learning models
the calculation based
result on different
of CFV types
is very close to of
thedatasets (pure
true value wind by wind
(measured
tunnel test datasets, pure numerical simulation datasets, hybrid datasets, etc.)
tunnel test), and the calculation error is less than 2%. This is due to the fact were
that different
trained in this study.
FDs The
haveFDs prediction
a major results
and minor effectwith different
on the CFV. Iferrors can be obtained
the prediction accuracy of byimportant
different trained models. These
FDs is high, the errors are high
CFV with ultimately reflected
calculation in the
accuracy canCFV calculation,
be obtained. as there is
It means
no need to predict every FD accurately in the case of limited
shown in Figure 17. It can be seen that even with an error of up to 25% in FDs prediction, computing resources. This
analysis further demonstrates that it is entirely feasible to identify
the error in CFV calculation is only 10%. When the prediction error of FDs can be reduced FDs and perform flutter
performance
to less than 15%, the calculationanalysis
resultby
ofmachine learning
CFV is very methods.
close to the true value (measured by
wind tunnel test), and the calculation error is less than 2%. This is due to the fact that
different FDs have a major and minor effect on the CFV. If the prediction accuracy of im-
portant FDs is high, the CFV with high calculation accuracy can be obtained. It means
there is no need to predict every FD accurately in the case of limited computing resources.
This analysis further demonstrates that it is entirely feasible to identify FDs and perform
flutter performance analysis by machine learning methods.
 Atmosphere 2023, 14, x FOR PEER REVIEW 23 of 26

Atmosphere 2023,
Atmosphere 2023,14,
14,x1165
FOR PEER REVIEW 22 of 27 23 of 2
12

of CFV (%)
10
12
8

CFV (%)
10

error of error
6
8
4

Calculation
6
2

04
Calculation

2 0 5 10 15 20 25
Prediction error of FDs (%)
0
Figure 17. Error analysis. The x value of the blue dot represents the average error of the predicted
0
FDs of each 5
model, 10
while the15y value
20 represents
25 the calculation error of the corresponding CFV. The
red dashed linePrediction error
indicates theofoverall
FDs (%)trend of the CFV calculation error change with the FDs predic-
tion error.
Figure
Figure 17.Error
17. Error analysis.
analysis. x value
TheThe of the
x value ofblue dot represents
the blue the average
dot represents error of error
the average the predicted
of the predicted
FDs
FDs of
of each
eachmodel,
model,while they yvalue
whilethe valuerepresents
representsthethe
calculation error
calculation of the
error corresponding
of the CFV.CFV. The
corresponding
6.2.2. Sensitivity Analysis of
indicates Flutter
the Derivatives
red dashed line indicates the overall trend of the CFV calculation error change with theFDs
The red dashed line overall trend of the CFV calculation error change with the FDs predic
For the
tion error.
prediction cross-section shown in Figure 15, we can further analyze which FDs have a
error.
major impact on CFV by changing the magnitude of the FD sequentially. As shown in
6.2.2.
Figure Sensitivity Analysis axis
18, the horizontal of Flutter Derivatives
6.2.2. Sensitivity Analysis ofrepresents the change ratio of FDs. R > 1 indicates the FD
Flutter Derivatives
For the
increases, andcross-section
R < 1 indicates shown
the FD in Figure
decreases. 15, weThecan further analyze
longitudinal which FDs
axis indicates have a
the change
major For
of CFVimpact
the cross-section
causedon byCFV
shown
by changing
parameter change.the
in Figure 15, wethe
RV magnitude
> 1 indicatesofthat
can FDfurther
the
analyze
sequentially.
CFV increases afterAswhich
shown
parame-
FDs
in have a
major
Figure
ter change, impact
18, the on CFV
andhorizontal by changing
axis represents
RV < 1 indicates that the CFV the magnitude
the decreases.
change ratio of the
of FDs.
It can be seenFD sequentially.
R >from
1 indicates
the figure As
thethat
FD shown in
Figure
the main18,
increases, andtheR horizontal
factors < 1 indicates
affecting theaxis
the
CFVFD represents
A1* , A2* ,the
aredecreases. A3* ,change
The H 3*ratio
longitudinal
and ofconclusion
. Thisaxis FDs. R >is1the
indicates indicates
change the FD
consistent
of CFV caused by parameter
increases, and R < 1 indicates the FD change. R > 1 indicates that the
V decreases. The longitudinal axis CFV increases after parameter
indicates * the change
with theand
change, testRand analysis
V <by
1 indicates results of Hong
thatchange.
the CFVR from Tongji
decreases. University
It can be seen [60]. theA1* figure
and Hthat 3 rep-
of CFV caused parameter V > 1 indicates that thefromCFV increases theparame
after
resent
main the coupled
factors
ter change, Rvertical
affecting
and the CFVmotion
V < 1 indicates that
A1∗ , Athe
are speed ∗ , A∗ , and
2excited
CFV
3 by Hthe∗ aerodynamic
3 . This
decreases. conclusion
It can∗ be liftseen
isgenerated
consistent
from the bywith
the
figure tha
torsional motion displacement and the aerodynamic ∗
the test and analysis results of Hong from Tongji
* * damping
University * formed
[60]. *A1 andby the
H3 aerodynamic
represent the
the mainvertical
moment
coupled factorsmotion
caused byaffecting theexcited
the feedback.
speed CFVTheyare make
by A1 ,aerodynamic
the theA2torsional liftH
, A3 , andvibration 3 . This
generated modal conclusion
by damping isde-
the torsional consisten
motion
with the
crease displacement
from testpositive and
and analysis
to theresults
aerodynamic
negative, of Hong
indicating damping
thatfrom
flutter formed
Tongji by the aerodynamic
University
has reached the critical 1
*
[60]. Astate. andA2*H 3* rep
moment
caused
represents by the feedback. They make theformed torsional byvibration modal damping momentdecrease from
resent the the aerodynamic
coupled verticaldamping
motion speed excited the aerodynamic
by the aerodynamic directly
lift gen- by the
generated
positive to negative, indicating that flutter has reached the* critical state. A2∗ represents
erated
torsional by the torsional
motion motion speed,
displacement and and the increase in A2 with the increase byinthevelocity
the aerodynamic damping formed bythetheaerodynamic
aerodynamic damping moment directly formed generated aerodynamic
by
*
moment
will
the torsional caused
be beneficial toby
motion thethe
speed, feedback.
stability
andofthe They
theincrease
system. makeThe
in Athe∗change
with torsional A3 also
of increase
the vibration
affects
in modal
the CFV,
velocity damping
will but
be de
2
crease
beneficial from
the influence to thepositive
stability
is smaller to
than *
negative, * indicating
A1 , system.
of the A2 , and The * that
H 3 .change
The CFV flutter∗ has
A3 also affects
of decreases reached
with the the
theincrease critical
CFV, but in the *state.
A1 A2*
influence is smaller than A1∗ , A2∗damping
, and H3∗ . The CFV by in A1∗directly
, A3∗ ,
A3* , and
,represents
∗
*
Hthe
3
aerodynamic
and increases with the increase formed
∗ in A2*decreases
. The theother with the increase
aerodynamic
FDs have almost moment no effect gen
and H3 and increases with the increase in A2 . The other FDs have * almost no effect on
erated
on
the the
CFV.
by the
CFV.
TheThe
torsional
conclusion
conclusion
motion
above above speed,
is not not and
is universal. the increase
universal.
It is It onlyis only in for
for the
A2given
with
the giventhecross-section
increase in velocity
cross-section and
and cross-sections
will be beneficial
cross-sections with a
with atosimilar similar
the stability shape
shape and of theand dimension.
dimension. *
system. The change of A3 also affects the CFV, bu
the2.4influence is smaller
A H
than A1* , A2* , and H 3* . The CFV decreases with the increase in A1*
* *
* H3 1 1
2.1*
, A3 , and H and A increases
H with the increase in A2* . The other FDs have almost no effec
*
3
*
2
*
2
* *
*
A1 A H
on1.8the CFV. The conclusion
A H
3 3
above is not universal. It is only for the given cross-section
* *
4 4
1.5 cross-sections
and with a similar shape and dimension.
A*3
RV

1.2
2.4
0.9 A*1 H*1
H3*
2.1 *
A H*2
0.6
A*2
2
*
1.8 A1
* A
3
H*3
0.3
0.00 0.25 0.50 0.75 1.00 * *
A 1.25 H1.50 1.75 2.00
4 4
1.5 R
A* 3
RV

1.2 18.
Figure 18. Influence of FDs on CFV.
CFV.

6.3.0.9Sensitivity Analysis of Aerodynamic Shape

0.6The main
factors affecting the flutter stability of bridges can be divided into two cate-
A* 2
gories:
0.3
(1) the aerodynamic shape of the cross-section, such as the width-to-height ratio,
wind0.00 0.25 angle,
fairing 0.50 0.75
and1.00 1.25 1.50
inclined web1.75 2.00 (2) the dynamic characteristics of the structure,
slope;
R

Figure 18. Influence of FDs on CFV.

 6.3. Sensitivity Analysis of Aerodynamic Shape
The main factors affecting the flutter stability of bridges can be divided into two cat-
egories: (1) the aerodynamic shape of the cross-section, such as the width-to-height ratio,
Atmosphere 2023, 14, 1165 wind fairing angle, and inclined web slope; (2) the dynamic characteristics of the structure, 23 of 27
such as mass and frequency. The impact of these factors on the CFV is uncertain, and any
parameter change can cause a change in the CFV.
such This
as massstudy does
and not consider
frequency. the impact
The impact of dynamic
of these factors on characteristics on flutter stability
the CFV is uncertain, and any
currently, but only focuses on the
parameter change can cause a change in the CFV.impact of aerodynamic shape on flutter stability. Since
CFV This
calculation can be
study does notcombined
consider the with the machine
impact of dynamiclearning predictionon
characteristics offlutter
FDs and high
stability
calculation accuracy can be obtained, the relationship between the
currently, but only focuses on the impact of aerodynamic shape on flutter stability. Since shape of the cross-sec-
tion and
CFV the CFVcan
calculation canbebecombined
analyzed withby the thetrained
machine machine
learning learning models.
prediction of FDsTheand
analysis
high
results can accuracy
calculation provide acan reference for the
be obtained, theaerodynamic
relationship shape
between optimization
the shape ofofthe the closed-box
cross-section
girder.
and the CFV can be analyzed by the trained machine learning models. The analysis results
For closed-box
can provide girders,
a reference for thethere are not many
aerodynamic shape factors affectingofthe
optimization theaerodynamic shape
closed-box girder.
without considering the influence of the ancillary facilities on the
For closed-box girders, there are not many factors affecting the aerodynamic shape CFV in the construction
stage. It is
without time consuming,
considering and it of
the influence may thenot lead tofacilities
ancillary better calculation
on the CFV results
in theifconstruction
every detail
of the It
stage. cross-section is takenand
is time consuming, into account.
it may Therefore,
not lead to betterthis study only
calculation discusses
results if everythree
detailim-of
portant
the parameters:
cross-section width-to-height
is taken into account.ratio; wind fairing
Therefore, this studyangle;
onlyand inclined
discusses webimportant
three slope.
parameters: width-to-height ratio; wind fairing angle; and inclined web slope.
6.3.1. Width-to-Height Ratio
6.3.1. Width-to-Height Ratio
The width-to-height ratio of a closed-box girder varies greatly. The width of cross-
Thevaries
section width-to-height
from 30 m ratio to 50ofma and closed-box
the height girder varies
ranges greatly.
from 2 m Theto 5 width of cross-
m. Taking the
section varies from 30 m to 50 m and the height ranges from 2 m to
Runyang Bridge with the width-to-height ratio of 12.3 as an example, to reduce the work- 5 m. Taking the Runyang
Bridge
load of with the width-to-height
numerical calculation, it is ratio of 12.3that
assumed as an theexample,
CFVs aretoproportional
reduce the workload
to differentof
numerical calculation, it is assumed that the CFVs are proportional
wind fairing angles and inclined web slopes, so the wind fairing angle and inclined web to different wind fairing
angles andkept
slope are inclined web slopes,
unchanged. The so the wind fairing
width-to-height angle
ratio is and inclined
adjusted onlywebby slope are kept
changing the
unchanged.
width of the The width-to-height
cross-section. The FDsratiowithis adjusted
differentonly by changingratios
width-to-height the width of the cross-
are identified by
section. The FDs with different width-to-height ratios are identified
the trained machine learning model, and the CFV under different width-to-height ratios by the trained machine
learning model,
is calculated and the CFV under
by two-dimensional different
flutter numericalwidth-to-height
analysis. Theratios
resultisshows
calculated by CFV
that the two-
dimensional flutter numerical analysis. The result shows that the
decreases with the increase in the width-to-depth ratio, which is consistent with the con- CFV decreases with the
increase in the width-to-depth ratio, which is consistent with
clusion obtained by Zhang of Chang’an University [61]. The variation law of CFV with the conclusion obtained by
Zhang of Chang’an University [61]. The variation law of CFV with the width-to-depth
the width-to-depth ratio and the comparison with Zhang’s analysis are shown in Figure
ratio and the comparison with Zhang’s analysis are shown in Figure 19a. In this study, it
19a. In this study, it can also be found that the CFV decreases sharply when the width-to-
can also be found that the CFV decreases sharply when the width-to-height ratio is lower
height ratio is lower than 11, and the decrease trend of CFV gradually slows down with
than 11, and the decrease trend of CFV gradually slows down with the increase in the
the increase in the width-to-depth ratio when the ratio is greater than 11.
width-to-depth ratio when the ratio is greater than 11.
3.0 This study 1.6
This study 1.4 This study
Zhang (2018) Xian and Liao (2008) 1.5 Li (2018)
2.5 1.2
1.4
1.0
2.0 1.3
0.8
RV

1.2
RV

RV

1.5 0.6
1.1
0.4
1.0
1.0
0.2 0.9
0.5 0.0 0.8
7 8 9 10 11 12 13 14 15 16 17 30 40 50 60 70 80 90 8 10 12 14 16 18 20 22 24 26

Ratio of width to depth Wind fairing angle (°) Inclined web slope (°)
(a) (b) (c)
Figure 19. Influence of shape parameters on CFV: (a) width-to-height ratio [61]; (b) wind fairing
Figure 19. Influence of shape parameters on CFV: (a) width-to-height ratio [61]; (b) wind fairing
angle [62]; (c) inclined web slope [63].
angle [62]; (c) inclined web slope [63].

6.3.2. Wind Fairing

6.3.2. Fairing Angle
Angle
Figure 19b shows the CFV changing with the wind fairing angle. It can be seen that
the CFV first increases and then decreases with the increase in the wind fairing angle and
reaches the peak at around 60◦ . Xian and Liao designed several wind fairings with different
angles for a closed-box girder and analyzed the impact of the wind fairing angle on the
CFV based on wind tunnel test [62]. The result showed that the CFV first increased and
then decreased with the increase in wind fairing angle (α) (i.e., from 30◦ to 60◦ ). This trend
is consistent with the results obtained in this study (40◦ < α < 90◦ ). The experimental result
of a wind fairing angle larger than 60◦ is not given by Xian and Liao. Since the shape
Atmosphere 2023, 14, 1165 24 of 27

dimensions of the closed-box girder in their study are somewhat different from those in this
paper, the optimal wind fairing angles are also different. Although the conclusion is limited,
which is only applicable to the closed-box girder section with a specific width-to-depth
ratio and inclined web slope, the influence regularity of the wind fairing angle on CFV can
still be used as a reference for the optimization of aerodynamic shape.

6.3.3. Inclined Web Slope

Li et al. studied the influence of inclined web slope on CFV by wind tunnel tests [63].
The conclusion is that for a closed-box girder with the width-to-depth ratio of 11, the
larger the inclined web slope is, the smaller the CFV is. For the closed-box girder with
the width-to-depth ratio of 7, the influence of inclined web slope on CFV is very small.
The width-to-depth ratio of the prototype structure in this paper (the main girder cross-
section of Runyang Bridge shown in Figure 15) is 12.3. When the inclined web slope is
changed from 10◦ to 25◦ , the influence regularity of inclined web slope on CFV is shown in
Figure 19c. It can be seen that the CFV decreases with the increase in inclined web slope,
which is consistent with Li’s research, and this trend is almost linear. However, during
the actual design and construction, the specific inclined web slope should be determined
according to the actual engineering because it will significantly increase the manufacturing
difficulty if the inclined web slope is too small.

7. Conclusions
In this paper, based on a database of flutter derivatives obtained by wind tunnel tests
and numerical simulation, the flutter derivatives of closed-box girders are identified by
machine learning methods. Modeling is implemented with a gradient boosting decision
tree algorithm. The model is trained and analyzed under different data division patterns,
and model interpretation after training is realized. The sensitivity of flutter derivatives
and the influence of the geometric shape of the cross-section on critical flutter velocity are
also analyzed, which provides a reference for aerodynamic shape optimization and further
verifies the validity of the models. Some conclusions are summarized as follows:
• The model trained by gradient boosting decision tree is able to predict the flutter
derivatives under different data division patterns with the mean relative error of less
than 0.2. The machine learning prediction error of flutter derivatives will be weakened
in the numerical calculation of critical flutter velocity. If the prediction error of flutter
derivatives can be reduced to less than 15%, the calculation result of critical flutter
velocity will be very close to the true value (i.e., the calculation error is less than 2%).
• For closed-box girders, A1∗ , A2∗ , A3∗ , and H3∗ have great influence on critical flutter
velocity. The critical flutter velocity decreases with the increase in A1∗ , A3∗ , and H3∗ and
increases with the increase in A2∗ . The other flutter derivatives almost do not affect the
critical flutter velocity. This conclusion is not universal; it is only applicable for the
prototype structure in this study or for cross-sections with similar shape dimensions.
• The critical flutter velocity decreases with the increase in the width-to-depth ratio of
the closed-box girder. In the range of 40◦ –90◦ of wind fairing angles, the critical flutter
velocity first increases and then decreases with the increase in angle and reaches the
peak at around 60◦ . The critical flutter velocity decreases with an increase in inclined
web slope and is almost linear for a particular width-to-height ratio.
The machine learning method proposed in this paper can provide a new means of
identifying flutter derivatives and rapidly evaluating flutter stability in the preliminary
wind-resistant design stage. However, this study has some limitations:
• The proposed methods are only applicable to closed-box girder cross-sections, and the
sensitivity analysis of the critical flutter wind speed with the shape of cross-sections
only focuses on three feature dimensions: width-to-height ratio; wind fairing angle;
and inclined web slope. In the future, we expect to implement more refined research
Atmosphere 2023, 14, 1165 25 of 27

on the change of shape and even on the ancillary facilities of the main girder and
expand the research to other types of main girder cross-sections.
• The identification method of flutter derivatives proposed in this paper can achieve
rapid evaluation of flutter stability without physical and numerical wind tunnel tests,
but a good prediction effect can only be achieved based on pure numerical simulation
data at present. It is necessary to further improve the accuracy and robustness of
intelligent identification and achieve machine learning predictions based on complete
wind tunnel test data, which depends on larger datasets and better data conditions.
• The existing data and proposed methods only achieve the identification of eight flutter
derivatives and two-dimensional flutter stability analysis. In the future, more works
on three-dimensional flutter stability analysis based on the predicted eighteen flutter
derivatives via machine learning methods will also be needed.

Author Contributions: Conceptualization, Y.G. and C.B.; Methodology, Y.G.; Investigation, N.C.;
Data curation, N.C.; Writing—original draft, N.C.; Writing—review and editing, Y.G. and C.B.;
Supervision, Y.G. and C.B.; Funding acquisition, Y.G. All authors have read and agreed to the
published version of the manuscript.
Funding: This research was funded by the National Natural Science Foundation of China (51978527,
52278520) and China Scholarship Council (201906260198).
Data Availability Statement: The data presented in this study are available on request from the
corresponding author.
Acknowledgments: The authors gratefully acknowledge the technical support of the University of
Florence. We also appreciate the writing suggestions from Genshen Fang and the computational
assistance provided by Yizhe Lan, Zihang Liu, and Miaomiao Wei.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Xiang, H.F.; Ge, Y.J. Aerodynamic Challenges in Span Length of Suspension Bridges. Front. Archit. Civ. Eng. China 2007, 1, 153–162.
[CrossRef]
2. Liu, S.Y.; Zhao, L.; Fang, G.S.; Hu, C.X.; Ge, Y.J. Numerical Aerodynamic Characteristics and Modeling of a Quasi-Flat Plate at
Torsional Vibration: Effects of Angle of Attack and Vibration Amplitude. Nonlinear Dyn. 2022, 107, 2027–2051. [CrossRef]
3. Fang, G.S.; Cao, J.X.; Yang, Y.X.; Zhao, L.; Cao, S.Y.; Ge, Y.J. Experimental Uncertainty Quantification of Flutter Derivatives for a
PK Section Girder and Its Application on Probabilistic Flutter Analysis. J. Bridge Eng. 2020, 25, 04020034. [CrossRef]
4. Qin, X.R.; Gu, M. Stochastic System Method for Identification of Aerodynamic Derivatives of Bridge Decks in Turbulent Flow.
China Civ. Eng. J. 2005, 38, 73–77. (In Chinese) [CrossRef]
5. Sarkar, P.P.; Jones, N.P.; Scanlan, R.H. Identification of Aeroelastic Parameters of Flexible Bridges. J. Eng. Mech. 1994, 120, 1718–1742.
[CrossRef]
6. Ding, Q.S.; Zhou, Z.Y.; Zhu, L.D.; Xiang, H.F. Identification of Flutter Derivatives of Bridge Decks with Free Vibration Technique.
J. Wind. Eng. Ind. Aerodyn. 2010, 98, 911–918. [CrossRef]
7. Xu, F.Y.; Zhu, L.D.; Ge, X.M.; Zhang, Z. Some New Insights into the Identification of Bridge Deck Flutter Derivatives. Eng. Struct.
2014, 75, 418–428. [CrossRef]
8. Yamada, H.; Miyata, T.; Ichikawa, H. Measurement of Aerodynamic Coefficients by System Identification Methods. J. Wind. Eng.
Ind. Aerodyn. 1992, 41, 1255–1263. [CrossRef]
9. Chowdhury, A.G.; Sarkar, P.P. A New Technique for Identification of Eighteen Flutter Derivatives Using a Three-degree-of-freedom
Section Model. Eng. Struct. 2003, 25, 1763–1772. [CrossRef]
10. Borri, C.; Costa, C. Quasi-steady Analysis of a Two-dimensional Bridge Deck Element. Comput. Struct. 2004, 82, 993–1006.
[CrossRef]
11. Borri, C.; Zahlten, W. Fully Simulated Nonlinear Analysis of Large Structures Subjected to Turbulent Artificial Wind. Mech. Struct.
Mach. 1991, 19, 213–250. [CrossRef]
12. Borri, C.; Höffer, R. Aeroelastic Wind Forces on Flexible Bridge Girders. Meccanica 2000, 35, 1–15. [CrossRef]
13. Fang, G.S.; Pang, W.; Zhao, L.; Rawal, P.; Cao, S.Y.; Ge, Y.J. Toward a Refined Estimation of Typhoon Wind Hazards: Parametric
Modeling and Upstream Terrain Effects. J. Wind. Eng. Ind. Aerodyn. 2021, 209, 104460. [CrossRef]
14. Hu, X.N.; Fang, G.S.; Yang, J.Y.; Zhao, L.; Ge, Y.J. Simplified Models for Uncertainty Quantification of Extreme Events Using
Monte Carlo Technique. Reliab. Eng. Syst. Saf. 2023, 230, 108935. [CrossRef]
Atmosphere 2023, 14, 1165 26 of 27

15. Zhu, Z.W.; Gu, M. Wind Tunnel and CFD Study on Identification of Flutter Derivatives of a Long-span Self-anchored Suspension
Bridge. Comput. Aided Civ. Infrastruct. 2007, 22, 541–554. [CrossRef]
16. Wang, X.J. Identification of Aerodynamic Parameters of Bridge Decks by Stochastic Approaches. Ph.D. Thesis, Tongji University,
Shanghai, China, 2008. (In Chinese) [CrossRef]
17. Xu, F.Y.; Zhang, Z.B. Free Vibration Numerical Simulation Technique for Extracting Flutter Derivatives of Bridge Decks. J. Wind.
Eng. Ind. Aerodyn. 2017, 170, 226–237. [CrossRef]
18. Zhou, Y.; Kijewski, T.; Kareem, A. Aerodynamic Loads on Tall Buildings: Interactive Database. J. Struct. Eng. 2003, 129, 394–404.
[CrossRef]
19. Wang, J.; Cheng, J.; Teng, E. Design Wind Loads on Tall Buildings: A Wind Tunnel Data Based Expert System Approach. In
Proceedings of the 11th International Conference on Wind Engineering, Lubbock, TX, USA, 2–5 June 2003.
20. Ho, T.; Surry, D.; Morrish, D.; Kopp, G. The UWO Contribution to the NIST Aerodynamic Database for Wind Loads on Low
Buildings: Part 1. Archiving Format and Basic Aerodynamic Data. J. Wind. Eng. Ind. Aerodyn. 2005, 93, 1–30. [CrossRef]
21. Quan, Y.; Tamura, Y.; Matsui, M.; Cao, S.Y. Introduction to Aerodynamic Database of Low Buildings. In Proceedings of the 7th
National Conference on Wind Engineering and Industrial Aerodynamics, Chengdu, China, 1–3 August 2006. (In Chinese).
22. Jung, S.; Ghaboussi, J.; Kwon, S.D. Estimation of Aeroelastic Parameters of Bridge Decks Using Neural Networks. J. Eng. Mech.
2004, 130, 1356–1364. [CrossRef]
23. Chen, C.H.; Wu, J.C.; Chen, J.H. Prediction of Flutter Derivatives by Artificial Neural Networks. J. Wind. Eng. Ind. Aerodyn.
2008, 96, 1925–1937. [CrossRef]
24. Lute, V.; Upadhyay, A.; Singh, K.K. Support Vector Machine Based Aerodynamic Analysis of Cable Stayed Bridges. Adv. Eng.
Softw. 2009, 40, 830–835. [CrossRef]
25. Chung, J.; Lee, S.W.; Chang, S.; Kim, Y.S. Estimation of Flutter Derivatives of Various Sections Using Numerical Simulation and
Neural Network. In Proceedings of the 2012 World Congress on Advances in Civil, Environmental, and Materials Research,
Seoul, Republic of Korea, 26–30 August 2012.
26. Rizzo, F.; Caracoglia, L. Artificial Neural Network Model to Predict the Flutter Velocity of Suspension Bridges. Comput. Struct.
2020, 233, 106236. [CrossRef]
27. Abbas, T.; Kavrakov, I.; Morgenthal, G.; Lahmer, T. Prediction of Aeroelastic Response of Bridge Decks Using Artificial Neural
Networks. Comput. Struct. 2020, 231, 106198. [CrossRef]
28. Mei, H.Y. Deep Neural Network-Based Method for Calculating Nonlinear Aerodynamic Forces and Responses of Bridge Section.
Ph.D. Thesis, Southwest Jiaotong University, Chengdu, China, 2021. (In Chinese) [CrossRef]
29. Li, Y.; Li, C.; Liang, Y.D.; Li, J.W. Fast Prediction of the Flutter Critical Wind Speed of Streamlined Box Girders by Using Aerostatic
Force Coefficients and Artificial Neural Networks. J. Wind. Eng. Ind. Aerodyn. 2022, 222, 104939. [CrossRef]
30. Li, H.; Spencer, B.F.; Bao, Y. Machine Learning Paradigm for Structural Health Monitoring. Struct. Health Monit. 2021, 20, 1353–1372.
[CrossRef]
31. Scanlan, R.H.; Tomko, J.J. Airfoil and Bridge Deck Flutter Derivatives. J. Eng. Mech. 1971, 6, 1717–1737. [CrossRef]
32. Gu, M.; Zhang, R.; Xiang, H. Identification of Flutter Derivatives of Bridge Decks. J. Wind. Eng. Ind. Aerodyn. 2000, 84, 151–162.
[CrossRef]
33. Li, Y.; Liao, H.; Qiang, S. Weighting Ensemble Least-square Method for Flutter Derivatives of Bridge Decks. J. Wind. Eng. Ind.
Aerodyn. 2003, 91, 713–721. [CrossRef]
34. Bartoli, G.; Contri, S.; Mannini, C.; Righi, M. Toward an Improvement in the Identification of Bridge Deck Flutter Derivatives. J.
Eng. Mech. 2009, 135, 771–785. [CrossRef]
35. Chen, N.Y.; Ge, Y.J. Aerodynamic Parameter Identification of Typical Bridge Sections Based on Artificial Neural Network. China
Civ. Eng. J. 2019, 52, 91–97. (In Chinese) [CrossRef]
36. Chen, N.Y. Database Integration and Intelligent Evaluation of Wind-resistant Performances of Long-Span Bridges. Master’s
Thesis, Tongji University, Shanghai, China, 2017. (In Chinese).
37. Wang, X. Numerical Simulation Study on Wind-induced Vibration of Long-span Bridges. Master’s Thesis, Guizhou University,
Guiyang, China, 2018. (In Chinese).
38. Zhang, Q. Research on Flutter Derivatives of Bridge Girder Sections by Numerical Simulation. Master’s Thesis, Southwest
Jiaotong University, Chengdu, China, 2008. (In Chinese).
39. Zhu, A.D. Wind-resistant Performance Analysis and Experimental Investigation on Long-span Cable-stayed Bridges. Master’s
Thesis, Dalian University of Technology, Dalian, China, 2014. (In Chinese).
40. Zhang, Z.B. Numerical Study on Nonlinear Aerodynamic Forces on Typical Bridge Decks. Master’s Thesis, Dalian University of
Technology, Dalian, China, 2016. (In Chinese).
41. Qi, Y.H. Numerical Analysis on 3D Flutter Instability of Long-span Suspension Bridges and Effects from Active Controlled Wind
Barrier. Master’s Thesis, Central South University, Changsha, China, 2013. (In Chinese) [CrossRef]
42. Bai, H.; Xia, Y.; Liu, J.X.; Li, J.W. Numerical Simulation for Flutter Stability of Streamlined Bridge Decks. J. Chang. Univ.
2011, 31, 45–50. (In Chinese) [CrossRef]
43. Hong, G. Identifying Long-Span Bridge Flutter Derivatives via the Free Vibration Method Based on the Flutter Software. Master’s
Thesis, Chang’an University, Xi’an, China, 2012. (In Chinese) [CrossRef]
Atmosphere 2023, 14, 1165 27 of 27

44. Zhu, Y.Q. Numerical Simulation and Experimental Study on Flutter Characteristics of Long Span Suspension Bridge. Master’s
Thesis, Southwest Jiaotong University, Chengdu, China, 2017. (In Chinese).
45. Bai, Y.L.; Ou, J.P. Numerical simulation of flutter derivatives of CFRP streamlined box girders and determination of critical wind
speed. In Proceedings of the 14th National Conference on Structural Wind Engineering, Beijing, China, 27–28 August 2009.
(In Chinese).
46. Hao, D. The Study of the Effect of Tuyere on Bridge Flutter Stability. Master’s Thesis, Chang’an University, Xi’an, China, 2011. (In
Chinese) [CrossRef]
47. Jiao, L.N. The Effects of Characteristics of Mean Wind and Fluctuate Wind on Long-span Bridge. Master’s Thesis, Harbin Institute
of Technology, Harbin, China, 2009. (In Chinese).
48. Fu, J. Wind Tunnel Experimental Investigation and Numerical Simulation on the Flutter Characteristic of a Long-span Suspension
Bridge. Master’s Thesis, Chongqing University, Chongqing, China, 2016. Available online: https://kns.cnki.net/KCMS/detail/
detail.aspx?dbname=CMFD201701&filename=1016908037.nh (accessed on 1 September 2022). (In Chinese).
49. Peng, Y. The Numerical Analysis for Flutter Stability on Streamlined Box Girder and Pneumatic Control Measures. Master’s
Thesis, Central South University, Changsha, China, 2012. (In Chinese).
50. Gao, W. Flutter Derivatives Research of Large Span Bridges under Skew Wind. Master’s Thesis, Southwest Jiaotong University,
Chengdu, China, 2013. (In Chinese) [CrossRef]
51. Pang, W. The Identification of Flutter Derivatives for Long-Span Bridges Based on Fluent. Master’s Thesis, Southwest Jiaotong
University, Chengdu, China, 2007. (In Chinese) [CrossRef]
52. Hong, C.J. Numerical Calculation Method for Flutter Critical Wind Speed and Flutter Derivatives of Long-Span Bridge Streamlined
Steel Box Girder. Master’s Thesis, Southwest Jiaotong University, Chengdu, China, 2019. (In Chinese) [CrossRef]
53. Lv, L.S.; Meng, Q.B.; Li, T. Study on Flutter Derivative Identification Method Based on MIMO System. Bridge Tunn. Eng.
2011, 2, 42–47. (In Chinese)
54. Wang, D.L. Numerical Simulation of Aeroelastic Problems on Bridges Based on Discrete Vortex Method. Master’s Thesis, Tongji
University, Shanghai, China, 2003. (In Chinese).
55. Xia, Y. Numerical Calculation of the Flutter Stability for Long-span Suspension Bridge. Master’s Thesis, Chang’an University,
Xi’an, China, 2011. (In Chinese) [CrossRef]
56. Chao, C.R. Research on Flutter Bridges with Three Towers and Its Whole Process Simulation. Master’s Thesis, Southeast
University, Nanjing, China, 2015. (In Chinese).
57. Chen, Y.Q. Aerodynamic Shape Optimization of Central Slotted Box Girders Based on Performance of Wind-induced Self-excited
Vibration. Master’s Thesis, Tongji University, Shanghai, China, 2016. (In Chinese).
58. Lundberg, S.M.; Lee, S.I. A Unified Approach to Interpreting Model Predictions. In Proceedings of the 31st International
Conference on Neural Information Processing Systems, Long Beach, CA, USA, 25 November 2017. [CrossRef]
59. Zhu, Z.W.; Chen, Z.Q. Numerical Simulations for Aerodynamic Derivatives and Critical Flutter Velocity of Bridge Deck. China J.
Highw. Transp. 2004, 17, 41–45. (In Chinese) [CrossRef]
60. Hong, L.Z. Influence of Wind Barrier on Flutter Performance of Bridge. Master’s Thesis, Tongji University, Shanghai, China, 2020.
(In Chinese).
61. Zhang, H. Estimation for Critical Flutter Velocity of the Bridge with Flat Box Girder. Master’s Thesis, Chang’an University,
Xi’an, 2018. Available online: https://kns.cnki.net/KCMS/detail/detail.aspx?dbname=CMFD201901&filename=1018791251.nh
(accessed on 1 September 2022). (In Chinese).
62. Xian, R.; Liao, H.L. Wind Tunnel Test Study of Aerodynamic Optimization Measures for Flutter Stability of Critical Flat Steel Box
Girder. World Bridges 2008, 138, 44–47. (In Chinese)
63. Li, Z.G.; Wang, Q.; Liao, H.L.; Wei, Y.F. Effects of Inclined Web Slope on Flutter Performance of Flat Box Girders and Their
Quantification. J. Vib. Shock. 2018, 37, 17–24. (In Chinese) [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.