Case Studies in Construction Materials 18 (2023) e01962

Contents lists available at ScienceDirect

Case Studies in Construction Materials

journal homepage: [Link]/locate/cscm

Case study

Machine learning-assisted fatigue performance optimization for

cutout geometry of orthotropic steel bridge decks
Xudong Wang a, b, Changqing Miao a, b, *, DiDi Hao a, b
a
Key Laboratory of Concrete and Prestressed Concrete Structures of Ministry of Education, Southeast University, Nanjing 211189, China
b
School of Civil Engineering, Southeast University, Nanjing 211189, China

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

Keywords: Diaphragm cutout is a typical fatigue detail of orthotropic steel decks (OSDs). The cutout ge­
Diaphragm cutout ometry could alter the structural responses and fatigue performance, significantly challenging the
Fatigue design process. This study proposes an efficient computational framework addressing fatigue
Orthotropic steel deck
performance optimization for the cutout geometry design. Sensitivity analysis is first carried out
Artificial neural network
to identify the significant parameters, and the datasets are established according to the random
Optimization
Finite element analysis sampling technique and the finite element (FE) model. Then, a comparison of prediction per­
formance between the back-propagation neural networks (BPNNs) and radial basis function
neural networks (RBFNNs) is employed to present the applicability of two types of artificial
neural networks (ANNs) in the prediction of structural responses. Finally, the cutout geometry
optimization is performed by integrating the prediction model and the multi-objective particle
swarm optimization (MOPSO) algorithm. The validity and applicability of the framework are
demonstrated with a real-world application. The optimization results show that the fatigue life is
increased from 119.8 to 150.2 years for cutout detail 1 and from 37.4 to 70.2 years for cutout
detail 2. The proposed framework can significantly reduce the computational burden and deliver
an optimized scheme for the cutout design.

1. Introduction

Orthotropic steel deck (OSD) has been widely used in steel bridges due to its advantages, such as light self-weight, structural
redundancy, and convenient construction [1]. However, due to the complex geometry and welded connections in OSDs, the fatigue
cracks are prone to occur at the positions of stress concentration under the joint action of repeated wheel load and environment, which
seriously affects the fatigue life of OSDs [2,3]. To reduce mutual constraint between the longitudinal ribs and transverse floor beam
(FB) or diaphragm, and minimize the out-of-plane distortion of the diaphragm plate under the passage of vehicles, a diaphragm cutout
beneath the rib is usually adopted in the OSDs. However, this detail could weaken the diaphragm plate and interfere with the in-plane
stress flow, resulting in stress concentration and fatigue cracking at the edge of the cutout [4]. At present, fatigue cracking of the cutout
detail has been reported worldwide, including at the Williamsburg Bridge in the United States [5], the West Gate Bridge in Australia
[6], and the Pingsheng Bridge in China [4].
Small-scale and full-scale experiments are the most direct methods for fatigue analysis [7,8]. However, the experiments usually

* Corresponding author at: Key Laboratory of Concrete and Prestressed Concrete Structures of Ministry of Education, Southeast University,
Nanjing 211189, China.
E-mail address: chqmiao@[Link] (C. Miao).

[Link]
Received 1 November 2022; Received in revised form 16 February 2023; Accepted 25 February 2023
Available online 26 February 2023
2214-5095/© 2023 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license
([Link]
X. Wang et al. Case Studies in Construction Materials 18 (2023) e01962

lead to high financial, time costs, and potential safety problems [9]. The numerical simulation based on finite element analysis (FEA)
increasingly provides reasonable and rapid performance assessment for the initial design phase. Connor [10] conducted a parameters
analysis on the fatigue performance of cutouts based on a three-dimensional (3D) FE model. It was found that the stiffness of cutout
types specified in the AASHTO LRFD specification can be excessive. Based on FEA, De Corte [11] conducted a parametric study for
eight cutout types and recommended several cutout shapes for highway and railway bridges. Wang et al. [12] analyzed the stress
responses from four cutout types. The research demonstrated that the arc radius greatly influenced the stress distribution of cutout, and
the fatigue performance can be improved with the increase of web thickness. Zhu et al. [13] compared the stress behavior of two types
of cutout, which were used in the bridge maintenance of the Pingsheng Bridge in China. The results showed that enlarging the cutout
using a large arc radius could decrease the stiffness near the cutout and may not improve the fatigue resistance compared to the
original cutout. These studies indicate that the shape and size of the cutout have significant impacts on fatigue performance. Increasing
the arc radius of cutout could weaken the stress raiser, but it may also increase the in-plane stress and decrease the fatigue strength of
cutout. Therefore, it is necessary to determine the appropriate cutout geometry according to the practical engineering application.
Structural optimization is a critical step in designing advanced and efficient structures [14,15]. In recent years, coupling analysis
methods based on machine learning approaches and heuristic algorithms have gradually gained attention in solving the optimization
problems of complex structures, which involve nonexplicitly known performance functions and high computational costs. Cheng [16]
studied the reliability of bridges and presented an analysis framework integrating artificial neural networks (ANNs) and genetic al­
gorithms (GA). The results indicated the superiority and reliability of the proposed method in response prediction of complex
structures. Ke et al. [17] performed an optimization analysis on the cutout type composed of round arcs and straight lines. It was found
that the cutout with a larger arc radius showed a better fatigue performance than other typical shapes of the cutouts when the cutout
areas were similar. Morfidis and Kostinakis [18] made a comparison of seismic damage prediction accuracy between the multi-layer
feedforward perceptron neural networks (MFPNNs) and radial basis function neural networks (RBFNNs). The results showed both
networks had high efficiencies in damage prediction. Baandrup et al. [19] performed parametric optimizations on the OSDs using the
girder weight and manufacturing costs as the objective function and the stress and stiffness as the constraint function. Although the
results showed the design parameters could meet the final design requirements, the cutout geometry was not changed, and the stress
concentration still existed. Based on the support vector machine model and multi-objective grey wolf optimizer algorithm, Liu et al.
[20] performed the design optimization of carbon fibre reinforced plastic square tubes. Few systematic studies, however, have been
reported to perform the optimization analysis on fatigue performance of cutout detail.
At present, many reasonable shapes of cutout have been proposed [10,12,17]. However, with the rapid development of infra­
structure, there could be many changes in the bridge’s structural forms, traffic load, and lane layouts, leading to the existing shapes of
the cutout may no longer meet the requirement of structural design. In order to provide rapid decision-making reference and find a
balance between design efficiency and design quality, this study presents a novel machine learning-assisted optimization framework
by integrating ANNs and the multi-objective particle swarm optimization (MOPSO) algorithm. The sensitivity analysis is first con­
ducted to improve the optimization efficiency before establishing the datasets. Then, a comparison between the BPNNs and RBFNNs is
employed to determine the proper prediction model for structural responses. Finally, the BP-MOPSO-based optimization method is
carried out on the multi-objective optimization of cutout geometry. The proposed method is used in the cutout geometry optimization
of a real-world bridge. The difference between the optimized scheme and the initial scheme is compared to illustrate the validity of the
proposed method.

2. Machine learning-assisted fatigue performance optimization method

2.1. Optimization problem statement

Based on the FE analysis and field measurement [13], the arc radius of cutout is usually in a compression state under wheel loads. In
addition, the cracking of the cutout is not only at the place with the arc radius but also at the weld toe of the butt weld between the
diaphragm and U-rib where the tensile stress usually occurs [21]. Therefore, the proposed optimization problem is to find the

Fig. 1. Geometrical parameters of the diaphragm cutout.

2
X. Wang et al. Case Studies in Construction Materials 18 (2023) e01962

geometrical parameters that minimize the tensile stress (fs1) at the butt weld between the diaphragm and U-rib (i.e. detail 1) and the
compressive stress (fs3) at the arc radius of cutout (i.e. detail 2). In order to find the proper cutout geometry, five independent design
variables: cutout width (d), cutout height (h), included angle (θ), the radius at the highest portion of cutout (r1 ), and the radius at the
lowest portion of cutout (r2 ), are considered, as shown in Fig. 1. The cutout geometry can be determined according to the combination
of the above five variables. In addition, two reference points, i.e., point A and point B, are defined to determine the constraints function
conveniently. To meet the requirement of geometric elements of the cutout, constraint functions are defined, such as restricting the
range of the parameters and the Euclidean distance of reference points. Based on the relevant optimization studies [20,22], the
optimization problem of the cutout geometry is defined as:
⎧
⎪
⎪ find D = [d, h, θ, r1 , r2 ]T
⎪
⎨ min [fs1 (D), f√
s3 (D)]
̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ (1)
⎪
⎪ s.t. L(D) = (y1 − y2 )2 + (z1 − z2 )2 ≥ r1
⎪
⎩
DL ≤ D ≤ DU

where D is the vector of design variables, L(D) is the distance between point A and point B, DU and DL are the vectors of the upper and
the lower bounds of the design variables, respectively.

2.2. Back-propagation neural network (BPNN)

BPNN is a multi-layer forward network with a back-propagation learning algorithm, which consists of an input layer, an output
layer, and one or more hidden layers [23]. It is a powerful mathematical tool for complex nonlinear relations, where explicit for­
mulates are difficult to obtain [24]. Furthermore, it has been proved that a BPNN with three layers can approximate any nonlinear
continuous function [25].
In this study, a three-layer BPNN is adopted to predict structural responses, as shown in Fig. 2. The samples X = (x1 , x2 , ., xi ) are
taken as the input nodes, and the expected values (also called the actual values) Z = (z1 , z2 , ., zn ) are taken as the output nodes. The
output (O) of the networks and the error function (Er) can be defined as Eqs. (2) and (3).
( ) ( ( ) )
∑ q ∑q ∑p
(2)
′ ′
On = ψ wmn yi − θk = ψ wmn φ wim xi − θm − θk
m=1 m=1 i=1

1 ∑ s
Er = (Zn − On )2 (3)
2S n=1

where ψ (x) and φ(x) are the transfer function of hidden and output layer, respectively; w is the weight; θ and θ′ are the threshold of
hidden and output layer, respectively; x and y are the input and output of hidden layer, respectively; S is the sample size.

Fig. 2. Configuration of BPNN.

3
X. Wang et al. Case Studies in Construction Materials 18 (2023) e01962

2.3. Radial basis function neural network (RBFNN)

RBFNN was first proposed by Broomhead and Lowe [26], which has the advantage of good generalization ability, simple structure,
and high efficiency. Different from BPNN, RBFNN is a three-layer feedforward neural network with only one hidden layer. The
structure of RBFNN is shown in Fig. 3. There are three main stages in the training process of RBFNNs [27]：(1) Centre selection of
neurons in the hidden layer; (2) Calculation of width of neurons in the hidden layer; (3) Update of the network weight matrix.
The Gaussian function is used as the RBF function in the hidden layer, and the self-organizing learning method is adopted to
determine the center point. The output value of the hidden layer is accumulated according to the weight. Hence, the jth output neuron
of the networks (yj) can be expressed as:
( )
∑l ∑l
1
yj = vhj Gh = vhj exp − ‖xi − ch ‖ , j = 1, 2, ., k (4)
h=1 h=1
2σ h 2

where,vhj is the weight, Gh is the Gaussian function at the hth neuron in the hidden layer, ch and σ h are the centre and width of Gh,
respectively.

2.4. MOPSO algorithm

Due to the mutual constraints between different objectives, multi-objective optimization problems usually have many optimum
solutions, called the Pareto set. The objective functions corresponding to these solutions are Pareto fronts [28]. In order to find the
Pareto set, the MOPSO algorithm [29], which is inspired by the behavior of bird populations, is adopted in this study. The individual in
the population is regarded as a particle, and each particle search for the optimum solution by learning the optimum historical position.
Compared with the classical optimization algorithms, the MOPSO algorithm has great advantages in solving multi-objective opti­
mization problems, such as high efficiency and good robustness [30,31]. Fig. 4 illustrates the flow chart of the MOPSO algorithm, and
detailed solution procedures are as follows.

(1) Initialization: a population containing m particles is generated in the N-dimensional search space. Then, the speed, position, and
archive set of each particle are initialized.
(2) The fitness value of each particle is calculated, and the Pareto solution is saved to the archive set according to the dominance
relation.
(3) The personal best performance (pbest) of each particle and global best performance (gbest) are generated.
(4) The updated velocity and position of particles are calculated according to Eqs. (5) and (6), respectively. The new pbest is
generated according to the evaluation of updated particles. Then, the archive set is updated, and the gbest is developed ac­
cording to the updated fitness value.
vt (i + 1) = a · vt (i) + c1 · b1 [pbestt (i) − pt (i)] + c2 · b2 [gbest(i) − pt (i)] (5)

pi+1 = pi + vi+1 (6)

where vt (i +1) is the velocity vector of tth particle in the ith iteration; pt(i) is the position of tth particle in the ith iteration; a is the
weighting coefficient; c1 and c2 are acceleration coefficients; b1 and b2 are random variables from 0 to 1;
(5) If the stopping criterion is satisfied, the search is terminated, and the Pareto solution set is obtained. Else, return to step (3).

2.5. The framework of the machine learning-assisted optimization method

In this section, a modular analysis framework integrating ANNs and MOPSO algorithm is proposed to optimize the cutout fatigue

Fig. 3. Configuration of BRFNN.

4
X. Wang et al. Case Studies in Construction Materials 18 (2023) e01962

Fig. 4. MOPSO flow chart.

performance, as shown in Fig. 5. The proposed machine learning-assisted optimization method includes three essential modules: the
data generation module, the prediction module, and the optimization module, mainly consisting of the following procedures.

(1) Generate datasets: according to the geometric parameters and optimum objective, the sensitivity analysis is first conducted to
identify the sensitive parameters. Then, the datasets are established, and the input and output samples are obtained by the
random sampling technique and FEA, respectively, as shown in Eq. (7). The samples are randomly assigned to the training and
test sets after the normalization.
{
In = [d, h, θ, r1 , r2 ]T
(7)
Out = [f s1 , f s3 ]T

Fig. 5. The framework of the machine learning-assisted optimization method.

5
X. Wang et al. Case Studies in Construction Materials 18 (2023) e01962

(2) Establish the prediction models: first, the proper neural networks should be selected to improve the prediction accuracy. As the
popular neural networks in the past decades, the BPNNs and RBFNNs are employed in this study for comparison. The training
parameters and structure of BPNNs and RBFNNs are determined based on the performance analysis. Then, the training set and
test set are applied in the establishment of BPNNs and RBFNNs. The prediction model is established until the convergence
condition is satisfied. Finally, two types of neural networks are evaluated according to statistical criteria.
(3) Solve the optimization problem: closed-form objective functions are extracted from the well-trained prediction models to
remain computationally feasible between the prediction and optimization analysis. The MOPSO algorithm is employed to
obtain the Pareto set, and the detailed procedures are introduced in Section 2.4.
(4) Evaluate the feasible solution: the feasible solutions selected from the Pareto set are evaluated based on the performance in­
dexes, such as von Mises stress, tensile stress, compressive stress, and displacement. The final optimum solution is obtained
according to the requirements of the design code [32] and experience.

3. Engineering application

3.1. Bridge description

The Weier Road bridge (as shown in Fig. 6), located in Hebei Province, China mainland, is a symmetrical curved-pylon cable-stayed
bridge. The overall length is 162 m. The OSD steel box girder with a length of 115 m is adopted in the main span, and the prestressed
concrete box girder with a length of 47 m is adopted in the side span. The steel box girder at the main span is 51 m wide and 1.9 m high,
containing eight traffic lanes (four in east-bound and four in west-bound), as shown in Fig. 6(b).
The design scheme of the OSD system is shown in Fig. 7. The U-shaped ribs (hereinafter called U-ribs) are employed with 320 mm
height and 300 mm width. The thickness of the deck plate, the rib web, the horizontal stiffener, and the diaphragm are 16, 8, 10, and
12 mm, respectively. Trapezoidal diaphragm cutouts, which are used as the design scheme, are designed on the diaphragm around the
lower portion of the U-ribs. It should be noted that the design scheme of cutout has been adopted in many bridges in China, and some
studies have proven that the cutout with this geometry is prone to fatigue cracking [13,33]. Fig. 8 shows the typical cracks observed on
the same type of cutout of an OSD bridge [13].

3.2. Finite element model

To investigate the structural response of the OSD system and generate the dataset for the prediction model, a 3D FE model, which
consists of about 450,000 Shell63 elements and 420,000 nodes, is established using ANSYS software, as shown in Fig. 9. Five di­
aphragms are modeled in the FE model. The stay cables are anchored at the diaphragms of D1, D3, and D5. The steel material has
Young’s modulus and Poisson ratio equal to 206 GPa and 0.3, respectively. The validation of the FE model is based on a comparison
between the measured values in the static load tests and the calculated values in the FE analysis. More detailed information on the
validation can be found in Ref. [34].
In this study, about 500 times transient analyses are performed on a workstation with Intel(R) Xeon(TM) Gold 5218 CPU
@ 2.30 GHz and 32 GB of RAM, and the total CPU time of FEA is about 3.2 h, which can be very time-consuming. For boundary
conditions, the nodes in the diaphragms of D1, D3, and D5 are all fixed. Besides, the nodes at the southern end are both restricted from
translations and rotations in the X and Y directions. The nodes at the northern end are restricted from translations in the Y direction and

Fig. 6. Bridge layouts. (a) overview; (b) cross-section of steel box girder (Unit: m).

6
X. Wang et al. Case Studies in Construction Materials 18 (2023) e01962

Fig. 7. Cutout geometry in design scheme (Unit: mm).

Fig. 8. Typical fatigue cracks at cutout, (a) short crack; (b) long crack [13].

Fig. 9. FE model of the OSD using shell elements.

the rotations in the X and Y directions.

Due to the structural responses of the cutout details being dependent on the loading position of the vehicle [35], several load steps
(i.e., 480 load steps) are defined for determining the most unfavorable loading conditions. A wheel load of 100 kN is applied to the
selected area of 200 mm (in the longitudinal direction) by 600 mm (in the transverse direction) at each step. As illustrated in Fig. 10
(a), the transverse loading begins on the northern side of the deck with load case LC1, then moves to the southern side until reaching
load case LC9. As illustrated in Fig. 10(b), the longitudinal loading begins at D1 and moves northward with 200 mm intervals until it
reaches D5.
The nominal stress approach, which is convenient for comparing the fatigue behavior of different design schemes, is used to study

7
X. Wang et al. Case Studies in Construction Materials 18 (2023) e01962

Fig. 10. Loading arrangement: (a) transverse direction; (b) longitudinal direction.

the fatigue performance of cutout details [36]. By referring to the research conducted by Zhu et al. [13], the nominal stress extraction
position is defined as 6 mm away from the weld toe or free edge of the cutout. According to the FEA results of the initial design scheme,
the stress concentration phenomenon at the free edge of the cutout is obvious, as illustrated in Fig. 11, which means the cutout ge­
ometry optimization should be performed. Meanwhile, this stress concentration position is consistent with the fatigue cracking that
occurs in the cutout of a real-world bridge [13] (see Fig. 8).

3.3. Sensitivity analysis

The sensitivity analysis is performed to quantitatively identify the significance of the design variables X = [d, h, θ, r1 , r2 ]T on the
objectives, fs1 and fs3, which could avoid interference of irrelevant variables on the prediction accuracy of ANNs. In this study, the
orthogonal test method based on the combinatorial theory is used for sensitivity analysis. The orthogonal test is deemed an effective
and economical design method for multifactorial evaluation [37]. The sensitivity of factors could be easily measured by reasonably
arranging the sample factors on the orthogonal test table.
Based on the design code [32] and experience, the characteristic of design variables is listed in Table1. Three different levels which
are selected for each variable are determined, including, μ − σ, μ, and μ + σ. Hence, a five-factor and three-level orthogonal test is
conducted based on the arrangement of orthogonal test table L18 (35), and the results are listed in Table 2. Moreover, the range (Ra) is
selected as the evaluation index, and Ra can be formulated as:
Ra = max(K1 , K2 , …, Ki ) − min(K1 , K2 , …, Ki ) (8)

1∑ n
Ki = yk (9)
n k=1

where Ki and n are the average and test number of the variables at level i, respectively; yk is the kth value of the corresponding objective.
It can be seen from Fig. 12(a) that the variable θ is the most significant influence variable for objective fs1, followed by h, d, and r1 ,
and the variable r2 is insignificant for objective fs1. Similarly, it can be shown in Fig. 12(b) that the variable h has the most significant
effect on the sensitivity of fs3, followed by θ, d, and r1 , and the variable r2 has a minor effect on objective fs3. The above results show
that the fs1 and fs3 are more sensitive to the variables θ, h, d, and r1 than the variable r2 in the design domain. Therefore, four significant
variables (i.e., θ, h, d, and r1 ) are considered in the optimized procedure.

3.4. Optimization of diaphragm cutout geometry

3.4.1. Comparison of prediction models

In order to improve the prediction accuracy and avoid high computational costs from Monte Carlo simulations in conventional
optimization procedures [38], the prediction models of objective fs1 and fs3 (hereafter called models S1 and S3) are established using

Fig. 11. 3rd principal stress contour plot of the cutout.

8
X. Wang et al. Case Studies in Construction Materials 18 (2023) e01962

Table 1
Characteristics of the design variables.
Variable DT MV (μ) SD (σ)

d (mm) Uniform 280 17.32

h (mm) Uniform 95 20.21
θ (rad) Uniform 1.29 0.26
r1 (mm) Uniform 140 69.28
r2 (mm) Uniform 70 17.32

Note. DT = distribution type; MV = mean value; SD = standard deviation

Table 2
Orthogonal test results.
NO. d (mm) h (mm) θ (rad) r1 (mm) r2 (mm) fs1 (MPa) fs3 (MPa)

1 1 (262.68) 1 (74.79) 1 (1.028) 1 (70.72) 1 (52.68) 20.83 62.13

2 1 2 (95.00) 2 (1.290) 2 (140.00) 2 (70.00) 20.47 83.49
3 1 3 (115.21) 3 (1.552) 3 (209.28) 3 (87.32) 20.37 92.22
4 2 (280.00) 1 1 2 2 20.59 82.75
5 2 2 2 3 3 20.60 86.83
6 2 3 3 1 1 20.77 87.74
7 3 (297.32) 1 2 1 3 20.69 58.65
8 3 2 3 2 1 20.22 75.35
9 3 3 1 3 2 22.61 107.65
10 1 1 3 3 2 19.36 52.25
11 1 2 1 1 3 21.73 86.29
12 1 3 2 2 1 21.39 81.82
13 2 1 2 3 1 19.96 71.88
14 2 2 3 1 2 20.20 64.08
15 2 3 1 2 3 22.60 100.19
16 3 1 3 2 3 19.79 62.73
17 3 2 1 3 1 21.52 85.13
18 3 3 2 1 2 22.24 104.81

Fig. 12. Sensitivity analysis results of (a) fs1, and (b) fs3.

Table 3
Configuration parameters of neural networks.
Parameter BPNN RBFNN

Number of input neurons 4 4

Number of hidden neurons 2, 4, …, 16 500 (Maximum number)
Number of output neurons 1 1
Learning rate 0.06 N/A
Training accuracy 1e-5 1e-5
Spread of radial basis functions N/A 5,10, …,40
Number of neurons to add between displays N/A 10

9
X. Wang et al. Case Studies in Construction Materials 18 (2023) e01962

BPNNs and RBFNNs, respectively. The levenberg-marquardt algorithm [39] and orthogonal least square algorithm [40] are adopted as
training functions of BPNNs and RBFNNs, respectively. Table 3 lists the configuration parameters of the BPNNs and RBFNNs. A dataset
with 1000 groups of samples is established, which consists of input parameters Xinput = [d, h, θ, r1 ]T and the output parameters fs1 (for
model S1) and fs3 (for model S3). 750 groups of samples (75% of the dataset) are randomly selected as the training set, and the other
250 groups of samples (25% of the dataset) form the test set.
The goodness of fit (R2) in Eq. (10) is selected to evaluate the prediction accuracy of BPNNs and RBFNNs in models S1 and S3. It
should be noted that the fitting accuracy between the predicted and expected values is higher when the R2 is closer to 1.0.
∑
S
(Zi − Oi )2
2
R = i=1
S (
(10)
∑ )2
Oi − Oavg
i=1

where Oavg is the average of the predicted values.

For defining the number of neurons in the hidden layer of BPNNs, a statistical analysis of model performance based on index R2 is
conducted, as shown in Fig. 13. It can be seen that the precision of both models shows a rapid increase with the increase of hidden
neurons. The best model performance is obtained when hidden neurons are 8 for model S1 and 10 for model S3. Hence, the BPNNs
structure of 4–8–1 and 4–10–1 for models S1 and S3 are adopted, respectively. Fig. 14 shows the statistical analysis results of RBFNNs
performance. It can be found that the index R2 of the training set and the test set slowly decrease and increase with the increase of
spread value, respectively. The corresponding RBFNN obtains the best performance when the spread is 40 for model S1 and model S3.
The prediction models are established by MATLAB R2020a, and the average CPU time for BPNNs training and RBFNNs training is 180 s
and 260 s, respectively, which dramatically improves the computational efficiency compared with FEA.
In this study, closed-form expressions of fs1 and fs3 with respect to the design variables (i.e., θ, h, d, and r1 ) are extracted from the
well-trained prediction models. The performance assessment of the different prediction models is shown in Figs. 15 and 16. It can be
seen from Fig. 15 that the BPNN-based prediction models have a good correlation between the predicted value and the actual value. In
contrast, the RBFNN-based prediction models have relatively weak performance in prediction accuracy, as shown in Fig. 16.
Furthermore, the evaluation index of BPNNs and RBFNNs in different models is listed in Table 4. The R2 values of BPNN-based
models are higher than 0.9, which indicates that the accuracy of the BPNN-based prediction models can guarantee the requirement
of engineering applications [41,42]. Hence, the BPNNs are adopted in this research as a useful mathematical tool in dealing with the
prediction problems of structural response.

3.4.2. Optimization results analysis

Since the connection length between the diaphragm and longitudinal rib could affect the stiffness of the longitudinal rib [33], this
study keeps the coordinates of point A unchanged and only change the coordinates of point B (see Fig. 1). Therefore, the geometry
optimization problem is simplified as follows:
⎧
⎪
⎪ T
⎪
⎪ find D = [ d, h, θ, r1 ]
⎪
⎪
⎪
⎪
⎨ min [ fs1 (D), fs3 (D) ]
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ (11)
⎪
⎪
⎪ s.t.
⎪ [122 − (d/2 + h/ tan θ)]2 + (130 − h)2 ≥ r1
⎪
⎪
⎪
⎪
⎩ D ≤D≤D L U

Fig. 13. Prediction performance of BPNN with different hidden neurons in (a) model S1 and (b) model S3.

10
X. Wang et al. Case Studies in Construction Materials 18 (2023) e01962

Fig. 14. Prediction performance of RBFNN-based model with different spread values.

Fig. 15. Performance analysis of BPNN-based prediction model in (a) model S1, and (b) model S3.

Fig. 16. Performance analysis of RBFNN-based prediction model in (a) model S1, and (b) model S3.

11
X. Wang et al. Case Studies in Construction Materials 18 (2023) e01962

Table 4
The performance comparison in different networks for the structure response prediction.
Networks R2 in Model S1 R2 in model S3

Training set Test set Training set Test set

BPNN 0.983 0.963 0.979 0.951

RBFNN 0.826 0.669 0.895 0.763

where the DU and DL can be expressed as follows:

DL = [ 270, 60, 0.837, 50 ]T and

(12)
DU = [ 310, 130, 1.744, 300 ]T
According to the proposed method, the optimization results are calculated after 100 iterations of the MOPSO algorithm, and the
parameters employed for MOPSO are given in Table 5. Fig. 17(a) shows the Pareto front obtained from the multi-objective optimi­
zation. It can be found that the optimum solution can not be directly determined, and a comparison of the feasible solutions needs to be
conducted.
Considering there is no best optimal solution for the Pareto set, three schemes, named schemes 1–3, are selected from the Pareto set
based on optimization results and manufacturing experience. The corresponding geometry is plotted in Fig. 17(b). Three performance
indexes are employed to compare schemes, including tensile stress (i.e., fs1), compressive stress (i.e., fs3), and displacement. Table 6
shows the values of each index of the selected schemes. It can be seen that the index values of the three selected schemes are
significantly reduced compared with the initial scheme, and the difference among the three schemes is small.
For further comparison, the difference percentage between the different schemes is illustrated in Fig. 18. It is found that scheme 1
has the best performance in stress, and the maximum values of fs1 and fs3 are reduced by 32.3% and 47.9%, respectively. The maximum
displacement of scheme 1, scheme 2, and scheme 3 is reduced by 7.7%, 9.6%, and 7.7%, respectively. The deflection-to-span ratio of
the deck for scheme 1, scheme 2, and scheme 3 are 0.00074, 0.00083, and 0.00074, respectively, which are less than the specified
value (i.e., 0.0014) in the design code [32]. The stiffness of the three optimized schemes is good enough for engineering applications.
Thus, scheme 1 which has the best performance index is determined as the final optimized scheme.

3.5. Advantages of the optimized geometry scheme

Due to the complexity of the cutout geometry, the optimum scheme is difficult to determine in the usual design process. The
machine learning-assisted optimization method proposed in this study provides an efficient way to optimize cutout geometry. Table 7
shows the 1st principal stress and 3rd principal stress of the cutout in different schemes. There is an obvious stress concentration in the
initial scheme, which can be detrimental to fatigue performance. On the contrary, the distribution of principal stress at the cutout edge
is more uniform in the optimized scheme. The maximum tensile stress decreases from 24.27 to 16.41 MPa, and the maximum
compressive stress decreases from 58.65 to 30.54 MPa in the optimized scheme, which is a remarkable improvement for fatigue
performance.
The schemes proposed by Zhu et al. [13] and Fang et al. [43] are selected for comparison in this research. It can be seen from
Table 7 that the optimized scheme has a better structural response performance than reference schemes I and II, which validate the
applicability of the proposed method.
In order to evaluate the fatigue life of the cutout, a fatigue vehicle model specified in the design code [32] is used, and the
equivalent constant-amplitude stress ranges (Seq) are calculated using the rain-flow method [44]. According to the Palmgren-Miner
linear damage rule [45], Eurcode 3 [46], and Ref. [47], the fatigue life of cutout can be determined as follows:
( )5
∑ Y
(1 + η)i nyi Seqi
D= ≥1 (13)
i=1
KD

where Y is the fatigue life (year); nyi is the number of effective stress cycles; Seqi is the equivalent stress ranges; η is the average traffic
growth rate; KD is the category constant. In this study, details 1 and 2 both belong to category 71 [46], hence KD = 1.9 × 1015.

Table 5
Parameter setting of MOPSO algorithm.
MOPSO Parameter Value

Population size 200

Repository size 200
Maximum number of generations 100
Inertia weight 0.4
Individual confidence factor 2
Swarm confidence factor 2
Number of grids in each dimension 20

12
X. Wang et al. Case Studies in Construction Materials 18 (2023) e01962

Fig. 17. Cutout schemes. (a) Pareto fronts obtained by BP-MOPSO-based optimization method; (b) cutout geometry of different schemes
(Unit: mm).

Table 6
Comparison of four schemes with different indexes.
Scheme Performance index

Max. fs1 (MPa) Max. fs3 (MPa) Max. displacement (mm)

Initial scheme 24.27 58.65 0.052

Scheme 1 16.41 30.54 0.048
Scheme 2 17.34 33.28 0.047
Scheme 3 19.54 37.40 0.048

Fig. 18. Performance assessment for three optimized schemes.

Meanwhile, η = 3% is considered based on the real traffic situation.

Fig. 19 shows the fatigue life of different schemes for two details. It can be seen that there is a nonlinear relation between index D
and fatigue life, and the fatigue life is significantly increased in the optimized scheme. For detail 1, the fatigue life of the initial scheme,
scheme 2, scheme 3, and optimized scheme (i.e., scheme 1) is 119.8, 135.5, 140.6, and 150.2 years, respectively. For detail 2, the
fatigue life of the same schemes is 37.4, 55.2, 62.8, and 70.2 years, respectively. The results demonstrate that the proposed method is
beneficial in improving the fatigue performance of cutouts.

13
X. Wang et al. Case Studies in Construction Materials 18 (2023) e01962

Table 7
Comparison of different schemes.

4. Conclusion

In this study, a machine learning-assisted optimization design framework is proposed to improve the fatigue performance of the
diaphragm cutout. Based on modularization design, the framework realizes the convenient and fast invocation of multiple data sources
and algorithms. Sensitivity analysis is first conducted to identify the significant parameters, and a comprehensive dataset is built
considering the uncertainty of parameters. Subsequently, prediction models of the fatigue performance indexes (fs1, fs3) in terms of the
geometry parameters (d, h, θ, r1.) are established and evaluated. Finally, multi-objective optimization is carried out to obtain the
optimum parameters using the BP-MOPSO-based optimization method. Principal conclusions drawn from this study are given below:

(1) The sensitivity analysis results show that the tensile stress fs1 is most sensitive to the angle θ, followed by the included height h,
width d, radius r1 , and radius r2 . Instead, the included height h is the most critical parameter influencing the compressive stress
fs3, followed by the angle θ, width d, radius r1 , and radius r2 .
(2) The prediction performance of the RBFNN and BPNN models are compared. The results indicate that the BPNN model could
provide a nonlinear mathematical function with better accuracy and efficiency than the RBFNN model. The BPNN model with
optimal structures can perform a good prediction for the structural response.
(3) The proposed framework has been applied in the cutout optimization of a real-world bridge. The results show that maximum
tensile stress and compressive stress are reduced by 32.3% and 47.9%, respectively. The fatigue life is increased from 119.8 to

14
X. Wang et al. Case Studies in Construction Materials 18 (2023) e01962

Fig. 19. Comparison of fatigue life.

150.2 years for cutout detail 1 and from 37.4 to 70.2 years for cutout detail 2. The results show the proposed method can be
effectively employed in optimizing the fatigue performance of cutout detail.
(4) Considering the complexity of vehicle load, service environment, and constructional quality of OSDs, there could be a large
scatter between the calculated fatigue life and the actual fatigue life, hence some field tests are needed for further research.
Furthermore, future studies can be conducted to verify the applicability of the proposed method in different types of cutouts.

Disclosure statement

No potential conflict of interest was reported by the authors.

Funding

This research has been supported by the National Natural Science Foundation of China [Grants No. 52178119, 51778135], the
National Key R&D Program Foundation of China [Grant No. 2017YFC0806001], and the Postgraduate Research & Practice Innovation
Program of Jiangsu Province [Grant No. KYCX22_0216].

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to
influence the work reported in this paper.

Data availability

Data will be made available on request.

References

[1] B.M. Kozy, R.J. Connor, D. Paterson, D.R. Mertz, Proposed revisions to AASHTO-LRFD bridge design specifications for orthotropic steel deck bridges, J. Bridge
Eng. 16 (2011) 759–767, [Link]
[2] C. Cui, Y.-L. Xu, Q.-H. Zhang, Multiscale fatigue damage evolution in orthotropic steel deck of cable-stayed bridges, Eng. Struct. 237 (2021), 112144, https://
[Link]/10.1016/[Link].2021.112144.
[3] Q. Zhang, Y. Bu, Q. Li, Review on fatigue problems of orthotropic steel bridge deck, China J. Highw. Transp. 30 (2017) 14–30.
[4] Z. Zhu, Z. Xiang, Fatigue cracking investigation on diaphragm cutout in a self-anchored suspension bridge with orthotropic steel deck, Struct. Infrastruct. Eng.
15 (2019) 1279–1291, [Link]
[5] P.A. Tsakopoulos, J.W. Fisher, Full-scale fatigue tests of steel orthotropic decks for the Williamsburg Bridge, J. Bridge Eng. 8 (2003) 323–333, [Link]
10.1061/(ASCE)1084-0702(2003)8:5(323).
[6] Silla Sanchez J., Noonan J., Percy R. West Gate Bridge: management of fatigue cracking. Proceedings of Austroads Bridge Conference, 9th, 2014, Sydney, New
South Wales, Australia 2014.6.
[7] B. Cheng, X. Ye, X. Cao, D.D. Mbako, Y.-s Cao, Experimental study on fatigue failure of rib-to-deck welded connections in orthotropic steel bridge decks, Int. J.
Fatigue 103 (2017) 157–167, [Link]
[8] H. Nakamura, W. Jiang, H. Suzuki, K.-i Maeda, T. Irube, Experimental study on repair of fatigue cracks at welded web gusset joint using CFRP strips, Thin-
Walled Struct. 47 (2009) 1059–1068, [Link]
[9] R. Ma, C. Cui, M. Ma, A. Chen, Performance-based design of bridge structures under vehicle-induced fire accidents: basic framework and a case study, Eng.
Struct. 197 (2019), 109390, [Link]

15
X. Wang et al. Case Studies in Construction Materials 18 (2023) e01962

[10] R.J. Connor, Influence of cutout geometry on stresses at welded rib-to-diaphragm connections in steel orthotropic bridge decks, Transp. Res. Rec. 1892 (2004)
78–87, [Link]
[11] W. De Corte, Parametric study of floorbeam cutouts for orthotropic bridge decks to determine shape factors, Bridge Struct. 5 (2009) 75–85, [Link]
10.1080/15732480903142518.
[12] C. Wang, Q. Zhang, T. Zhang, Y. Feng, Floor-beam web cutout shape analysis to improve the fatigue resistance in orthotropic steel bridge decks, Key Eng. Mater.
452 (2011) 161–164, [Link]
[13] Z. Zhu, Z. Xiang, J. Li, Y. Huang, S. Ruan, Fatigue behavior of orthotropic bridge decks with two types of cutout geometry based on field monitoring and FEM
analysis, Eng. Struct. 209 (2020), 109926, [Link]
[14] H. Liu, D. Yang, X. Wang, Y. Wang, C. Liu, Z. Wang, Smooth size design for the natural frequencies of curved Timoshenko beams using isogeometric analysis,
Struct. Multidiscip. Optim. 59 (2019) 1143–1162, [Link]
[15] Y. Wang, Z. Wang, Z. Xia, L.H. Poh, Structural design optimization using isogeometric analysis: a comprehensive review, Comput. Model. Eng. Sci. 117 (2018)
455–507, [Link]
[16] J. Cheng, An artificial neural network based genetic algorithm for estimating the reliability of long span suspension bridges, Finite Elem. Anal. Des. 46 (2010)
658–667, [Link]
[17] L. Ke, J.Q. Lin, C.X. Li, Y. Liu, Z.Y. Chen, Fatigue performance and structural detail optimization of arc-shape cutouts in diaphragm of steel box girder, Bridge
Constr. 47 (2017) 18–23.
[18] K. Morfidis, K. Kostinakis, Comparative evaluation of MFP and RBF neural networks’ ability for instant estimation of r/c buildings’ seismic damage level, Eng.
Struct. 197 (2019), 109436, [Link]
[19] M. Baandrup, P. Noe Poulsen, J. Forbes Olesen, H. Polk, Parametric optimization of orthotropic girders in a cable-supported bridge, J. Bridge Eng. 24 (2019),
04019118, [Link]
[20] Q. Liu, K. Liufu, Z. Cui, J. Li, J. Fang, Q. Li, Multiobjective optimization of perforated square CFRP tubes for crashworthiness, Thin-Walled Struct. 149 (2020),
106628, [Link]
[21] F. Jiang, Y. Ding, Y. Song, F. Geng, Z. Wang, CFRP strengthening of fatigue cracks at U-rib to diaphragm welds in orthotropic steel bridge decks: Experimental
study, optimization, and decision-making. Structures, Elsevier,, 2022, pp. 1216–1229.
[22] X. Wang, W. Wu, Y. Liu, Z. Ran, Surrogate-assisted two-phase tensioning strategy optimization for the system transformation process of a cable-stayed bridge,
Eng. Optim. (2019), [Link]
[23] M.T. Hagan, H.B. Demuth, O.D. Jesús, An introduction to the use of neural networks in control systems, Int. J. Robust. Nonlinear Control.: IFAC-Affiliated J. 12
(2002) 959–985, [Link]
[24] Y. Xu, T. You, C. Du, An integrated micromechanical model and BP neural network for predicting elastic modulus of 3-D multi-phase and multi-layer braided
composite, Compos. Struct. 122 (2015) 308–315, [Link]
[25] K. Cui, X. Jing, Research on prediction model of geotechnical parameters based on BP neural network, Neural Comput. Appl. 31 (2019) 8205–8215, [Link]
org/10.1007/s00521-018-3902-6.
[26] D. Broomhead, D. Lowe, Multivariable functional interpolation and adaptive networks, Complex Syst. 11 (1988) 321–355.
[27] Z.M. Yaseen, A. El-Shafie, H.A. Afan, M. Hameed, W.H.M.W. Mohtar, A. Hussain, RBFNN versus FFNN for daily river flow forecasting at Johor River, Malaysia,
Neural Comput. Appl. 27 (2016) 1533–1542, [Link]
[28] R.S. Burachik, C.Y. Kaya, M.M. Rizvi, Algorithms for generating pareto fronts of multi-objective integer and mixed-integer programming problems, Eng. Optim.
(2021) 1–13, [Link]
[29] C.A.C. Coello, G.T. Pulido, M.S. Lechuga, Handling multiple objectives with particle swarm optimization, IEEE Trans. Evolut. Comput. 8 (2004) 256–279,
[Link]
[30] E. Fallah-Mehdipour, O.B. Haddad, M.M.R. Tabari, M.A. Mariño, Extraction of decision alternatives in construction management projects: Application and
adaptation of NSGA-II and MOPSO, Expert Syst. Appl. 39 (2012) 2794–2803, [Link]
[31] G. Sun, G. Li, Z. Gong, G. He, Q. Li, Radial basis functional model for multi-objective sheet metal forming optimization, Eng. Optim. 43 (2011) 1351–1366,
[Link]
[32] MOT, JTG D64-2015 Specifications for Design of Highway Steel Bridge, China Communications Press,, Beijing, 2015.
[33] Z. Chen, C. Li, J. He, H. Xin, Retrofit fatigue cracked diaphragm cutouts using improved geometry in orthotropic steel decks, Appl. Sci. 10 (2020) 3983, https://
[Link]/10.3390/app10113983.
[34] X. Wang, C. Miao, Y. Mao, Y. Ding, Fatigue Performance of Orthotropic Steel Decks in Super-Wide Steel Box Girder Considering Transverse Distribution of
Vehicle Load, Struct. Durab. Health Monit. 15 (2021) 299, [Link]
[35] J. Di, X. Ruan, X. Zhou, J. Wang, X. Peng, Fatigue assessment of orthotropic steel bridge decks based on strain monitoring data, Eng. Struct. 228 (2021), 111437,
[Link]
[36] M. Liu, D.M. Frangopol, K. Kwon, Fatigue reliability assessment of retrofitted steel bridges integrating monitored data, Struct. Saf. 32 (2010) 77–89, [Link]
org/10.1016/[Link].2009.08.003.
[37] X. Wang, C. Miao, X. Wang, Prediction analysis of deflection in the construction of composite box-girder bridge with corrugated steel webs based on MEC-BP
neural networks, Structures 32 (2021) 691–700, [Link]
[38] G. Renner, A. Ekárt, Genetic algorithms in computer aided design, Comput. -Aided Des. 35 (2003) 709–726, [Link]
[39] I. Mukherjee, S. Routroy, Comparing the performance of neural networks developed by using Levenberg–Marquardt and Quasi-Newton with the gradient
descent algorithm for modelling a multiple response grinding process, Expert Syst. Appl. 39 (2012) 2397–2407, [Link]
[40] S. Chen, C. Cowan, P. Grant, Orthogonal least squares learning algorithm for radial basis function networks, IEEE Trans. Neural Netw. 2 (1991) 302–309,
[Link]
[41] B. Do, A. Lenwari, Optimization of Fiber-Reinforced Polymer Patches for Repairing Fatigue Cracks in Steel Plates Using a Genetic Algorithm, J. Compos. Constr.
24 (2020), 04020006, [Link]
[42] R. Jin, W. Chen, T.W. Simpson, Comparative studies of metamodelling techniques under multiple modelling criteria, Struct. Multidiscip. Optim. 23 (2001) 1–13,
[Link]
[43] H. Fang, N. Iqbal, G. Van Staen, H. De Backer, Structural optimization of rib-to-crossbeam joint in orthotropic steel decks, Eng. Struct. 248 (2021), 113208,
[Link]
[44] I. Rychlik, A new definition of the rainflow cycle counting method, Int. J. Fatigue 9 (1987) 119–121, [Link]
[45] W. Wu, H. Liou, H. Tse, Estimation of fatigue damage and fatigue life of components under random loading, Int. J. Press. Vessels Pip. 72 (1997) 243–249,
[Link]
[46] ECS. Eurocode 3: Design of steel structures—Part 1–9: Fatigue. EN1993–1-9. Brussels, Belgium, 2005.
[47] N. Lu, M. Noori, Y. Liu, Fatigue reliability assessment of welded steel bridge decks under stochastic truck loads via machine learning, J. Bridge Eng. 22 (2017),
04016105, [Link]

16