International Journal of Mechanical Sciences 284 (2024) 109766

Contents lists available at ScienceDirect

International Journal of Mechanical Sciences

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

Physics-informed deep learning for structural dynamics under moving load

Ruihua Liang a,b, Weifeng Liu a, Yuguang Fu b,* , Meng Ma a
a
Key Laboratory of Urban Underground Engineering of Ministry of Education, Beijing Jiaotong University, Beijing 100044, China
b
School of Civil and Environmental Engineering, Nanyang Technological University, Singapore 639798, Singapore

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

Keywords: Physics-informed deep learning has emerged as a promising approach that incorporates physical constraints into
Structural dynamics the model, reduces the amount of data required, and demonstrates robustness and potential in dealing with
Deep learning limited datasets for a variety of studies. However, several key challenges still exist, with one being the spectral
Parameter identification
bias problem of deep learning in the simulation of functions with multi-frequency features. To overcome the
Physics-informed neural network (PINN)
Moving load
challenge, this study proposes a novel physics-informed deep learning method, which integrates physics-
Frequency domain informed neural network with Fourier transform so as to solve partial differential equations in the frequency
domain, thus alleviating the problem of spectral bias of neural networks in the simulation of multi-frequency
functions. In addition, the proposed method is used to focus on the forward simulation and parameter inverse
identification issues in structural dynamics under moving loads. To illustrate the superiority of the method, the
issues of dynamic response of simply supported beams under moving loads are presented as case studies, and the
performance of the method in multiple cases is analysed and discussed. The research results demonstrate the
feasibility and effectiveness of the method for structural dynamics simulation and parameter inverse identifi­
cations using limited datasets.

1. Introduction simulations of structural mechanical behavior [15]. This data-driven

approach has shown excellent performance in mechanical behavior
The forward simulation of structural mechanical behavior and the simulation [16], structural health monitoring [17,18], load identifica­
inverse identification of their parameters are significant research areas tion [19,20], source localization [21], uncertainty analysis [22], and
with applications in mechanical fields, including civil infrastructures other research tasks [23,24]. However, in many real-world engineering
[1], soils [2], and mechanical systems [3]. Over the past decades, scenarios, data acquisition for physical systems is often highly complex
numerous scientific numerical computational methods and their de­ and costly [25]. This is a significant drawback of deep learning methods,
rivatives have been proposed. These methods solve partial differential which rely heavily on data, and it poses challenges to their feasibility
equations (PDEs) that describe structural mechanical behavior [4], and reliability in practical applications [26].
enabling both forward simulation of structures and inverse identifica­ To mitigate the data-dependency issue in deep learning, embedded
tion of structural parameters [5]. The most widely used methods include physical information has been introduced, incorporating physical con­
the finite element method [6], the finite difference method [7], the straints into the optimization of model parameters [27]. Models devel­
discrete element method [8], the spectral element method [9], the finite oped based on these concepts are known as physics-informed neural
volume method [10], the meshless method [11], and the discontinuous networks (PINNs) [28]. Due to these physical information-based con­
deformation analysis method [12]. straints, PINNs can address forward and inverse problems involving
In recent years, deep learning research has flourished, demonstrating nonlinear physical systems with small datasets or even without labeled
strong potential in the simulation of structural mechanical behavior data. This method has demonstrated excellent performance and prom­
[13]. Specifically, deep learning, grounded in mathematical optimiza­ ising potential in various research areas [29], such as simulation of heat
tion and statistics, enables the construction of models with complex transfer problems [30,31], fluid flows [32,33], seismic wavefield [34,
functions and numerous parameters [14]. These models are trained and 35], solid mechanics [36,37], etc. However, existing studies have shown
optimized using vast amounts of data, thereby achieving accurate that neural networks struggle to simulate functions with

* Corresponding author.
E-mail address: [email protected] (Y. Fu).

https://doi.org/10.1016/j.ijmecsci.2024.109766
Received 23 June 2024; Received in revised form 29 September 2024; Accepted 29 September 2024
Available online 2 October 2024
0020-7403/© 2024 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
R. Liang et al. International Journal of Mechanical Sciences 284 (2024) 109766

multi-frequency features [38], common in structural dynamics, a phe­ 2. Methodology

nomenon known as spectral bias [39]. Multi-frequency features are
prevalent in structural mechanical behavior, particularly in structural In this section, the concept of structural dynamics under moving
dynamics, where the structural response under dynamic loading typi­ loads is described in Section 2.1, which provides the preliminary con­
cally exhibits both the modal properties of the structure and the spectral ditions for the subsequent work. Section 2.2 presents the concept of
properties of the loading [40]. Moreover, structural dynamics has PINN in addressing forward and inverse problems, including a descrip­
numerous real-world engineering applications [41], including identi­ tion of the model structure, embedded physical constraints, and the
fying the dynamic response and structural health of bridges [42], tun­ training procedure. Additionally, Section 2.3 details the proposed FD-
nels [43], and other infrastructure [44]. However, traditional PINNs PINN, covering its solution concept, specific governing equations,
often encounter spectral bias when solving these structural dynamics implementation procedure, and more.
problems, hindering the use of interpretable AI techniques that incor­
porate physical information in mechanics and engineering [45].
2.1. Theory of structural dynamics under moving load
Therefore, research on mitigating the spectral bias in neural networks is
necessary to support further engineering applications of deep learning
The motion of a simply supported Euler-Bernoulli beam under a
methods in tasks like structural mechanical behavior simulation and
moving load p with velocity v is studied, and the PDE of motion is given
parameter identification [46].
by [57]:
In recent years, several studies have been conducted to mitigate the
impact of spectral bias in neural network-based solutions. For example, ∂4 u(x, t) ∂u(x, t) ∂2 u(x, t)
EI + c + m = pδ(x − vt), (1)
Chai et al. [47] proposed strategies such as multiscale Fourier feature ∂x4 ∂t ∂t2
mapping and revised neuron splitting to solve the spectral bias of the
neural network in high-frequency wavefield predictions. Zhao et al. [48] where E is Young’s modulus; I is the moment of inertia of the cross-
proposed an adaptive method to control spectral bias by regulating section; u(x, t) is the transverse deflection of the beam; c is the damp­
spectral norms across convolutional layers through guided gradient ing coefficient; m is the mass per unit length; δ is the Dirac delta func­
flow, thereby reducing overfitting and improving model performance on tion; x is the distance from one support in the direction of motion; t is
denoising and inpainting tasks. Ding et al. [49] utilized the modal in­ time; L is the length of the beam.
formation of the target system obtained from the numerical analysis, For a beam with simple supports, the boundary conditions and initial
thus realized an accurate neural network solution for the dynamical conditions are as follows,
system based on the modal superposition method. Waheed et al. [50]
∂2 u ∂2 u
introduced Kronecker neural networks with adaptive activation func­ u(0, t) = u(L, t) = (0, t) = 2 (L, t) = 0, (2)
∂x2 ∂x
tions to overcome the spectral bias in neural networks for wavefield
modeling. Existing studies on addressing spectral bias are typically ∂2 u
tailored to specific equations or problems, and most of these approaches u(x, 0) = (x, 0) = 0. (3)
∂t 2
require the incorporation of labeled data into the construction of the loss
function during the solution process. Furthermore, in some researches, The use of the separation of variables method to solve the above PDE
the integration of additional techniques makes it challenging for PINN to is a classical analytical method [58], which assumes the beam
simultaneously handle both forward and inverse computations [51]. displacement solution in the following form [59]:
In the solution of differential equations, integral transform methods ∑
∞

[52], such as the Fourier transform [53], provide an efficient approach u(x, t) = ϕn (x)qn (t), (4)
n=1
by converting the original equation into a domain that is more amenable
to solution [54]. As a general differential equation processing technique, where ϕn(x) is the nth vibration mode of the beam that satisfies the
the Fourier transform can be used to effectively circumvent the boundary conditions, and qn(t) is the generalized coordinate corre­
multi-frequency characteristics present in solutions to structural dy­ sponding to the nth mode, it describes the intensity or participation level
namics problems [55], enabling accurate solutions without relying on of the system in the nth mode.
labeled data [56]. The aim of this paper is to propose a novel PINN
model to mitigate the spectral bias of neural networks and to apply it in
solving forward and inverse problems in structural dynamics, thus 2.2. Physics-informed neural network
providing support for their broader application in the future. Specif­
ically, frequency domain physics-informed neural networks (FD-PINN) PINN is a deep learning framework that integrates physical infor­
method is proposed to address the spectral bias of neural networks. This mation and data, which allows approximating the solution of partial
method uses the Fourier transform to convert the original equations into differential equations using various neural network models. The PDE to
the frequency domain, thereby reducing the difficulty of neural network be solved in PINN can be expressed in the following form [45],
simulations and improving accuracy. Several case studies involving the N [u](x) = f(x), x ∈ Ω, (5)
dynamic response analysis of a simply supported beam structure under
moving loads is conducted to thoroughly analyze and discuss the for­ B [u](x) = g(x), x ∈ ∂Ω, (6)
ward and inverse simulation performances of the FD-PINN models.
The rest of the paper is organized as follows: Section 2 provides a where N [⋅] is a differential operator; B [⋅] are the boundary conditions; u
detailed description of the mathematical representation of the structural is the unknown quantity in the equations, which is the deflection of
dynamics problem and the establishment and calculation of the PINN in beams in this study; x is the independent variable in the equation, which
both the time and frequency domains. Section 3 presents the computa­ is the spatial and time variables in this study; Ω is the calculation
tional results of the PINN model for forward simulation of structural domain; ∂Ω is the boundary of the computational domain; f(x) and g(x)
dynamics in both domains, with an analysis and discussion of the results. are specific functions regarding the governing equations and boundary
Section 4 illustrates the performance of the PINN models in the inverse conditions, respectively.
identification of structural dynamics parameters, including a discussion The effect of physical information on the PINN is mainly manifested
of the results. Finally, Section 5 concludes the research and presents an in the loss function, and by optimizing the trainable parameters in the
outlook for future work. neural network to minimize the loss function, the neural network can
approximate the solution of the PDEs. The loss function of PINN in

2
R. Liang et al. International Journal of Mechanical Sciences 284 (2024) 109766

solving the forward problem can usually be expressed as follows [60, constraints when the network is updated, and Self-adaptive phys­
61]: ics-informed neural network (SA-PINN) [64], which facilitates adaptive
updates of coefficients in the loss function during model training. These
L (θ) = wDE L DE (θ) + wBC L BC (θ), (7)
networks empower the solution of PDE equations without the need for
Nr ⃒ labeled data.
1 ∑ ( ) ( )⃒
⃒N [u] xi − f xi ⃒2 ,
L DE (θ) = (8)
Nr i=1 r r

2.3. Frequency domain physics-informed neural network

Nb
1 ∑ ⃒ ( ) ( )⃒
⃒B [u] xi − g xi ⃒2 ,
L BC (θ) = (9) Structural dynamics problems often involve multi-frequency features
Nb i=1 b b
[65], traditional neural networks might struggle to accurately simulate
them. To enhance the capability of PINNs in addressing structural dy­
where θ is the trainable parameter vector; L DE (θ) and L BC (θ) are the
namics issues, this paper introduces the FD-PINN method. Specifically,
loss functions for the governing equations and boundary conditions
the Fourier transform is employed to convert the original time-domain
respectively; xr and xb are the training data randomly sampled in the
PDE into the frequency domain, where the frequency components of
respective computational domains; Nr and Nb are the number of training
the solution become simpler. Subsequently, the response spectrum ob­
data. The parameters {wDE, wBC} are the coefficients for each sub-term in
tained by solving the PINN in the frequency domain is inverse Fourier
the loss function, which need to be specified in advance.
transformed back to the time domain, thereby obtaining the solution to
The network structure for the forward solving of the beam response
the original PDE. The network structures of FD-PINN in solving the
using PINN is depicted in Fig. 1 (a), and the mechanical information in
forward and inverse problem are shown in Fig. 1 (c) and (d).
Eq. (1) is used to form the loss function, where the differentiation cal­
In the case of the dynamic response of beams under moving loads
culations can be implemented by the automatic differentiation algo­
studied in this paper, the formulation of Eq. (1) in the frequency domain
rithm in PINN [62].
is obtained using the Fourier transform as shown below [66]:
For inverse problems such as inverse identification of PDE parame­
∫ L/v
ters using measured data, the parameters to be identified can be ∂4 ̂
u (x, ω) p − ixω
EI + icω ̂
u (x, ω) − mω2 ̂
u (x, ω) = pδ(x − vt)e− iωt dt = e v ,
considered as trainable parameters in PINN and the difference between ∂x 4
0 v
the model output and the measured data can be used to form a loss (12)
function. The network structure of PINN in solving the inverse problem
is shown in Fig. 1 (b), and the loss function can be represented as: ∫ L/v
u (x, ω) =
̂ u(x, t)e− iωt
dt, (13)
L (θ) = wDE L DE (θ) + wBC L BC (θ) + wME L ME (θ), (10) 0

Nm ⃒ where ̂ u (x, ω) represents the spectrum of beam displacements, which is

1 ∑ ⃒
⃒uacutal − uoutput ⃒2 ,
L ME (θ) = (11) the result of u(x, t) in the frequency domain obtained using the Fourier
Nm i=1
transform; ω is the angular frequency. In addition, when solving the
above equation using PINN, an immediate problem is that the neural
where L ME (θ)is the loss functions for the measurement data; uacutal is
network is unable to deal with complex numbers. Therefore, the solution
the actual measurement result; uoutput is the model output result; Nm is
of the PDE in the frequency domain is represented in the following form:
the number of measurement data, and the parameter wME is the corre­
sponding coefficient in the loss function. u (x, ω) = ̂
̂ u r (x, ω) + i ̂
u i (x, ω), (14)
In addition, recent years have seen the development of various PINN
variants, such as Gradient-enhanced physics-informed neural network substituting this Eq. (14) into Eq. (12) gives the real and complex parts
(gPINN) [63], which incorporates the derivatives of the PDE into the of the equation, respectively, which can be expressed in the following
form:

Fig. 1. The framework of PINN models for (a) solving forward problems and (b) solving inverse problems, and the framework of proposed FD-PINN models for (c)
solving forward problems and (d) solving inverse problems. In solving the forward problem, the spatial and temporal or frequency variables x and t or ω serve as
inputs to the neural network, while the corresponding displacements are its outputs. The PDE governing the problem is integrated into the model and used for
training. In solving the inverse problem, the parameters to be identified are treated as training parameters, and actual data is incorporated into the training process.
This enables the model to identify unknown parameters while simulating the structural dynamic response.

3
R. Liang et al. International Journal of Mechanical Sciences 284 (2024) 109766

∂4 ̂
u r (x, ω) 1 (xω) Appendix A.
EI − cω ̂ u r (x, ω) = p cos
u i (x, ω) − mω2 ̂ , (15) After calculating the response spectrum of the beam in the frequency
∂x 4 v v
domain using FD-PINN, the inverse Fourier transform can be used to
∂4 ̂
u i (x, ω) 1 (xω) obtain the solution of the original equations in the time domain, which is
EI u i (x, ω) = − p sin
u r (x, ω) − mω2 ̂
+ cω ̂ . (16)
∂x4 v v shown below:
For the above governing equations in the frequency domain, a PINN ∫
1 +∞
can be developed for forward solving the response of simply supported u(x, t) = u (x, ω)eiωt dω.
̂ (18)
2π − ∞
beams under moving load, as shown in Fig. 1 (c). Also, the boundary
conditions in Eq. (2) need to be transformed to the frequency domain Similarly, for solving inverse problems such as parameter identifi­
and considered in the loss function of the neural network, which is as cation using FD-PINN, both the PDE and the measurement data need to
follows: be transformed into the frequency domain using the Fourier transform,
including their real and imaginary parts. Specifically, the discrete
∂2 ̂
u ∂2 ̂
u Fourier transform (DFT) [68] can be used to transform the measurement
u (0, ω) = ̂
̂ u (L, ω) = (0, ω) = 2 (L, ω) = 0. (17)
∂x 2 ∂x data into the frequency domain as follows:
It should be noted that it is usually difficult to consider the special N− 1
∑ 2π
i kn
initial conditions in the frequency domain, and the solution results in the u (k) =
̂ u(n)⋅e− N , (19)
frequency domain are consistent with those for stationary initial con­ n=0

ditions. Considering that the response of the beams under the moving
where ̂ u (k) is the k-th frequency component of the frequency domain
load usually takes the stationary state as the initial condition, which
signal; u(n) is the n-th sample of the time domain signal; N is the total
response can be considered as a periodic function on the time scale. 2π
Therefore, it is feasible to use the FD-PINN for solving the structural number of samples in the signal; e− i N kn is the Fourier basis function,
dynamic issue in this study. The schematic of the computational domain used to project the time domain signal into the frequency domain.
and sampling points used in model training for PINN and FD-PINN is The procedure for solving simply supported beams under moving
shown in Fig. 2, it can be noticed that different random sampling stra­ loads using the FD-PINN method is summarized in Fig. 3.
tegies are used in FD-PINN since the results of the equations at different
frequencies are decoupled. 3. Forward simulation of structural dynamics
In addition, since the network uses frequencies as inputs, which
means that the frequency range of the computational domain needs to be In this section, the structural dynamics under various moving load
set appropriately before the computation. It should be noted that the conditions are presented, and traditional PINN, gPINN, SA-PINN, and
range of frequencies used in the actual calculation cannot be infinite. the proposed FD-PINN are implemented to address these problems.
Therefore, the original problem can be approximated by selecting the Specifically, Section 3.1 describes the details of the structural dynamics
range of frequencies based on its major frequency components. The cases, and Section 3.2 introduces the implementation details of the
principles for selecting calculation frequency ranges in spectral nu­ models. Section 3.3 presents the calculation results of the models.
merical methods [67] can be used as references for FD-PINN. Consid­
ering that the dynamic response of simply supported beams under 3.1. Calculation setup
moving loads is mainly composed of the first-order mode [57], the fre­
quency input range of the neural network in this paper is expected to To evaluate the performance of the FD-PINN method in forward
contain the first-order natural frequency of the structure. For other is­ solving the aforementioned problems, three cases are presented in this
sues, the computational frequency range should also be set by refer­ section. Additionally, according to the basic theory of dynamic response
encing the frequency components of the load. In addition, in some of simply supported beams under moving loads presented in Section 2.1,
engineering applications, such as simulating train induced vibrations, the calculation results of the analytical model based on the modal su­
the frequency range can be determined based on specific requirements perposition method are utilized as a reference, and the forward simu­
in engineering. lation performance of each model on its dynamic response is
In addition, given that the Dirac functions in the control equations investigated. The parameters used in various cases are shown in Table 1.
are not feasible in numerical computations, the Gaussian function is In addition, the models were also built on another set of parameters, the
employed in this study as an approximation, the details are described in details of which and the results of the model calculations are shown in

Fig. 2. Schematic of the computational domain and sampling points used in model training for (a) PINN and (b) FD-PINN. The training points are randomly sampled
from this time-space domain, and in order for the solution of the model to satisfy the initial and boundary conditions in the PDE, additional training points need to be
sampled at the corresponding times and boundaries.

4
R. Liang et al. International Journal of Mechanical Sciences 284 (2024) 109766

Fig. 3. Solution procedure for simply supported beams under moving loads using the FD-PINN method. Specifically, the equations and data need to be represented in
the frequency domain before being fed into the neural network, and then the PINN is used to solve for the spectrum of the dynamic response, after which the inverse
Fourier transform is utilized to obtain the solution of the original PDE.

Neural networks with the same structure were developed respec­

Table 1
tively. Specifically, the built neural network contains three hidden
Mechanical parameters used in the cases.
layers, each containing 60 hidden cells, and the hyperbolic tangent
v (m/ L EI m (kg/ ξ ω1 (rad/ p function is used as the activation function in the network, which is
s) (m) (N⋅m2) m) s) (kN)
considered to be a suitable activation function for PINN computation
Case 50 20 2.4 × 3000 0.025 22.069 12 [69].
1 109
For model training, the Glorot uniform method [70] is used to
Case 20 20 2.4 × 3000 0.025 22.069 12
2 109 initialize networks, 20,000 collection points were randomly sampled
Case 15 20 2.4 × 3000 0.025 22.069 12 over the computational domain to calculate the model losses and for
3 109 updating the model weights, the Limited-memory Broyden-­
Fletcher-Goldfarb-Shanno (L-BFGS) optimiser [71] was utilised to train
the model, the number of iterations for model training was chosen to be
Appendix B.
400,000. Additionally, the training point resampling technique has been
adopted [72], allowing the model to randomly resample the training
3.2. Model implementation points on the computational domain after every 100 training iterations,
which helps the model to avoid overfitting to a specific set of training
Following the methodology presented in Section 2, neural network points, thereby promoting a more generalized learning of the solution
models including FD-PINN, PINN, gPINN, SA-PINN are developed across the entire domain [73]. The coefficients of each sub-term of the
respectively to solve the dynamic response of simply supported beams loss function were set to the reciprocal of their initial values, ensuring
mentioned above. Since the forced vibration response of the beam is the that the initial values of each sub-term within the loss function are
main concern in the above problem, the time domain used for the cal­ approximately equal. Furthermore, the output transform technique was
culations is [0, L/v], which is the period of the beam subjected to the utilized to scale the network output to more sensitive values of the
moving load. Whereas, in the neural network developed using the FD- activation function, thereby improving the performance of the models
PINN method, the calculations are carried out in the spatial-frequency [74].
domain, and the frequency domain used for the calculations is consid­ In recent years, numerous platforms for PINN computation have
ered as [0, 5] Hz, which envelopes main frequency components of the been proposed, greatly facilitating the implementation of PINN, such as
dynamic response of the structure [57]. The parameter σ in the Gaussian Neurodiffeq [75], NVIDIA SimNet [76], FastVPINNs [77]. DeepXDE
function used to approximate the Dirac function in moving loads is set to [78] was utilized to implement the PINN in this paper. All numerical
0.1.

5
R. Liang et al. International Journal of Mechanical Sciences 284 (2024) 109766

calculations were performed on a standard workstation equipped with transformation of the original PDE in the spatial-time domain to the
an NVIDIA Tesla V100 GPU (32 G). After each model was computed, the spatial-frequency domain in the FD-PINN method. Consequently, the
displacements of 101 × 101 = 10,201 points in the spatial- temporal neural network is relieved from simulating the original complex func­
domain were uniformly extracted for result comparison. tion containing multi-frequency features in the solution, thereby
significantly enhancing the accuracy of the simulation.
3.3. Result
4. Parameter identification of structural dynamics
Fig. 4 shows the variations of the loss function of the developed
In this section, the parameter identification problem in structural
models with the number of iterations during the training process, and it
dynamics is addressed, employing traditional PINN, gPINN, SA-PINN,
can be found that the training loss of the models gradually decreases and
and the proposed FD-PINN. Section 4.1 elaborates on the details of the
almost no longer decreases as it stabilizes around 400,000 iterations,
parameter identification problem, while Section 4.2 introduces the
which suggests that the selected number of training iterations is
implementation details of models. Subsequently, Section 4.3 presents
reasonable.
the calculation results of the models.
The results of the displacement field calculated by the fully trained
model on various cases are shown in Fig. 5 and Fig. 6, and the response
results calculated by the PINN model and the FD-PINN model are shown 4.1. Calculation setup
in Fig. 7. The response of the beam under a moving load comprises two
components: a low-frequency component, associated with the moving The inverse identification of structural parameters using the
velocity, and a high-frequency component, related to the natural fre­ measured structural response under known moving loads is a classical
quency of the beam. inverse problem, and the performance of the FD-PINN models on this
In simulating the dynamic response of the beam using the PINN problem is investigated in this section. The calculation example for the
model, it often accurately captures the low-frequency components of the inverse identification of structural parameters is presented. Specifically,
response but struggles to simulate the high-frequency components the parameter of bending stiffness EI of the beam is unknown and needs
associated with the beam’s natural frequency. Fig. 7 illustrates that to be identified. A sensor is positioned in the middle of the beam to
while the overall trend of the dynamic response calculated by PINN access the beam’s response under a moving load at this location. The
aligns with the actual result, there remains a significant error in the parameters of the beams and loads in the cases are consistent with those
simulation. This phenomenon is also observed in the simulation results in Section 3, as detailed in Table 1.
of gPINN and SA-PINN. The underlying reason for this limitation lies in a
common limitation of neural network simulation, where model weights 4.2. Model implementation
are optimized to local optimal solutions during training but cannot
achieve the global optimal solution [38,45]. As a result, when a neural In the inverse problem solution of PINN, the difference between the
network simulates a multi-frequency function, it might only be able to measurement response and the model simulation result is considered in
simulate the low-frequency components of the function while ignoring the loss function of the model, and the parameter to be identified is
its high-frequency components, this phenomenon is also referred to as considered as a weight of the model, which is optimized and updated
spectral bias in some research. Although gPINN adds additional con­ during the model training. The inverse identification results of the pa­
straints related to the derivatives of the PDE in the model training loss rameters and the solution of the model can be obtained after the training
and SA-PINN uses adaptive loss coefficients, both still cannot avoid is completed. In addition, the parameters to be identified need to be
spectral bias in the simulation. initialized before the training of the model, and the initial value of EI is
In contrast, the FD-PINN model proposed in this paper exhibits chosen as 5 × 109 N⋅m2 in this study. The remaining details of the model
relatively low error compared to the analytical solution in nearly all implementation, such as the network size, model training method,
cases, suggesting its feasibility in simulating the dynamic response of the chosen optimizer, computational platform, and other relevant infor­
beam under moving loads. This improvement arises from the mation, are consistent with those outlined in Section 3.2 for the forward

Fig. 4. Model training loss on various computational cases: (a) FD-PINN model, (b) PINN model, (c) gPINN model, (d) SA-PINN model. The loss functions of the
models built based on all methods converged after being trained for 400,000 iterations.

6
R. Liang et al. International Journal of Mechanical Sciences 284 (2024) 109766

Fig. 5. The displacement field calculated by models across various cases (units: mm). In each displacement field, the horizontal coordinates represent time, the
vertical coordinate represents positions on the beam, and the colors represent the amplitude of the structural dynamic response at different locations and moments.

Fig. 6. The calculation errors of displacement fields by various models across different cases (units: mm). In each displacement field, the horizontal coordinates
represent time, the vertical coordinate represents positions on the beam, and the colors depict the amplitude of the simulation error of structural dynamic response at
different locations and moments.

simulation. In addition, the magnitude of the displacement field on the compu­

To compare the performance of the PINN model in forward and in­ tational domain in different cases is varied, the min-max normalization
verse solutions, the mean square error (MSE) between the model method was used before calculating the MSE to compare the model er­
calculation results and the analytical solution was used to evaluate the rors in the different cases, which is calculated as follows.
accuracy of the model simulation, which is calculated as follows.
y i − ymin
̂
yi = (21)
∗
̂ ,
1 ∑N ymax − ymin
MSE = y i − yi )2 ,
(̂ (20)
N i=1
where ymin is the minimum value of the exact solution for all sampling
where yi is the exact result of the ith sampling point in the calculation points, ymax is the maximum value of the exact solution for all sampling
points, ̂
y i is the output of the ith sampling point, ̂
y i is the normalized
∗
domain, ̂y i is the value of the ith sampling point calculated by the
models, and N is the total number of sampling points. result of the model output for the ith sampling point.

7
R. Liang et al. International Journal of Mechanical Sciences 284 (2024) 109766

Fig. 7. Comparison of the dynamic response at the locations x=5 m of the beams with the (a) analytical solution computed by the (b) PINN model and (c) FD-PINN
model in Cases. The blue line represents the simulation results obtained from the PINN and FD-PINN calculations, and the orange dashed line represents the actual
analytical results.

Moreover, in order to quantitatively compare the accuracy of each where EIidentified is the parameter result identified by various models,
model on parameter identification, the indicator of accuracy is defined and EItruth is the true parameter value.
as follows.
⃒ ⃒
⃒EIidentified − EItruth ⃒ 4.3. Result
Accuracy = 1 − , (22)
EIidentified
The results of parameter inverse identification by the models are
shown in Fig. 8. It can be found that the two fully trained models can

Fig. 8. Inverse identification results of parameters on various cases by (a) FD-PINN, (b) PINN, (c) gPINN, and (d)SA-PINN. The true parameter values are represented
by black dashed lines, the variation of the parameter identification results with the number of iterations for the neural networks built based on the two methods is
represented by solid lines.

8
R. Liang et al. International Journal of Mechanical Sciences 284 (2024) 109766

almost identify the true parameter values in various cases, but there is a leading to the highest proportion of the high-frequency response
difference in the convergence speed and final accuracy of the identifi­ component of the beam, whereas it is the smallest in Case 3. Given that
cation results. Specifically, Table 2 demonstrates the identification re­ the high-frequency response component is challenging for PINN to
sults and accuracy of the four models on different iterations, and it can simulate, this results in a notable disparity in the simulation error of
be seen that the PINN model could achieve relatively acceptable iden­ PINN for each case.
tification results on about 1 × 104 iterations, and the final identification In contrast, the error of the FD-PINN model in the inverse simulation
accuracy of the model after training 4 × 105 iterations is about 95 % in is lower than that of the PINN model, and is also marginally lower than
various cases. The performance of gPINN and SA-PINN is slightly better that of the FD-PINN model in the forward simulation. This highlights the
than that of the traditional PINN, especially after 4 × 105 iterations of greater potential of the PINN method in inverse analysis compared to
training, the parameter identification results of both of them are almost forward analysis.
100 %, but the convergence speed of their identification results are
relatively poor, and neither of them achieves effective parameter iden­ 5. Conclusions
tification when the model has been trained for 1 × 104 iterations. The
FD-PINN model proposed in this paper achieved significantly faster In this paper, a PINN framework for both forward simulation and
convergence in parameter identification, and it converged to a relatively inverse parameter identification of structural dynamics is developed.
great identification result at about 1 × 104 iterations, and the identifi­ Additionally, the FD-PINN method is proposed to address the challenge
cation accuracy of the model is more than 99 % on both 1 × 105 itera­ of spectral bias in neural networks when simulating multi-frequency
tions and 4 × 105 iterations of training. functions. Specifically, this method involves transforming the original
Overall, the above calculation results demonstrate the feasibility of PDE from the time domain to the frequency domain using the Fourier
the PINN methods for structural dynamic parameter identification. transform, thereby reducing the simulation complexity for the neural
Moreover, the FD-PINN method proposed in this paper achieves higher network and enhancing its accuracy. Furthermore, typical structural
accuracy and faster convergence than the traditional PINN method in dynamics cases are presented, focusing on the dynamic responses of
structural dynamic parameter identification. simply supported beams under moving loads. The calculations con­
The displacement field calculated by the models in parameter inverse ducted with the proposed FD-PINN method are compared with those
identification is shown in Fig. 9, and the MSE of the model for forward using the traditional PINN method across these cases.
simulation and inverse simulation are shown in Table 3. It can be seen The calculation results indicate that the traditional PINN method was
that the performance of almost all models in the inverse simulation is only able to simulate the low-frequency components of the structural
superior to their performance in the forward simulation. Specifically, on dynamic response, failing to capture the high-frequency components
Cases 1 and 2, all the PINN, gPINN, and SA-PINN were able to achieve associated with the natural frequencies of the structure. Consequently,
simulation of the natural frequencies of the structure at higher fre­ this resulted in lower simulation accuracy. In contrast, the proposed FD-
quencies in the inverse simulation, which is significantly better than the PINN method effectively overcomes this challenge, achieving an accu­
original results on the forward simulation. This could be attributed to rate simulation of the structural dynamic response. Specifically, the
the fact that in the inverse simulation, the model is provided with the proposed FD-PINN can reduce the mean-square error of the simulation
dynamic response of the beam at certain positions, thereby reducing the to below about 7 × 10–5 in various cases, while that of the traditional
complexity of the simulation and facilitating the attainment of the global PINN is about 1.5 × 10–3 or more. In addition, for the inverse identifi­
optimal solution rather than settling for a local optimal solution. These cation of structural parameters, both the PINN and FD-PINN methods
results suggest that the addition of measured data to the training of PINN can achieve better identification results, while the proposed FD-PINN
is able to improve the ability of the model to deal with the spectral bias can achieve higher convergence speed and identification accuracy. For
problem. However, in Case 3, the three models still struggle to simulate example, after the same training of 1 × 104 iterations, the proposed FD-
the high-frequency component of the dynamic response, which is PINN can achieve a parameter identification accuracy of more than 94 %
dependent on the natural frequency, despite minor errors in parameter across all cases, while the accuracy of the PINN is below 68 %. In
identification. This may be due to the fact that the solution of the conclusion, the results demonstrate the feasibility and the superiority of
equations in that case is more complicated, or the measurement data the proposed FD-PINN method for forward simulation and inverse
provided is not the most beneficial information for the model parameter identification of structural dynamics.
simulation. Furthermore, it’s important to highlight that the FD-PINN proposed
Additionally, it is noted that among the cases, the PINN simulation in this paper alleviates the complexity of solving the original PDE
error is highest for Case 1 and lowest for Case 3. This discrepancy arises through integral transformation, thereby addressing the spectral bias in
from the fact that the load is moving at its maximum speed in Case 1, neural network simulation. However, the mechanism underlying

Table 2
Inverse identification results of parameters by models on various iterations.
Model Training Iteration Case 1 Case 2 Case 3
( × 104)
Identified result Accuracy Identified result Accuracy Identified result Accuracy
(109 N⋅m2) (109 N⋅m2) (109 N⋅m2)

FD-PINN 1 2.409 99.6 % 2.500 95.8 % 2.534 94.4 %

10 2.400 100.0 % 2.405 100.2 % 2.405 99.8 %
40 2.400 100.0 % 2.400 100.0 % 2.400 100.0 %
PINN 1 1.632 68.0 % 0.557 23.2 % 0.328 13.7 %
10 2.323 96.8 % 2.386 99.4 % 2.213 92.2 %
40 2.299 95.8 % 2.414 99.4 % 2.362 98.4 %
gPINN 1 1.301 54.20 % 0.300 12.52 % 0.027 1.12 %
10 2.280 95.00 % 2.410 99.60 % 1.306 54.40 %
40 2.386 99.40 % 2.405 99.80 % 2.410 99.60 %
SA-PINN 1 0.285 11.88 % 0.401 16.72 % 0.002 0.09 %
10 2.290 95.40 % 2.390 99.60 % 2.381 99.20 %
40 2.405 99.80 % 2.400 100.00 % 2.410 99.60 %

9
R. Liang et al. International Journal of Mechanical Sciences 284 (2024) 109766

Fig. 9. Displacement fields calculated by the model in parameter inverse identification in various cases (units: mm). In each displacement field, the horizontal
coordinates represent time, the vertical coordinate represents positions on the beam, and the colors depict the amplitude of the simulation error of structural dynamic
response at different locations and moments.

Methodology, Investigation, Data curation. Weifeng Liu: Writing – re­

Table 3
view & editing, Supervision, Resources. Yuguang Fu: Writing – review
The MSE results of the model in forward simulation and inverse simulation on
& editing, Supervision, Resources. Meng Ma: Writing – review & edit­
various cases.
ing, Resources.
Model Case 1 Case 2 Case 3

Forward Inverse Forward Inverse Forward Inverse

Declaration of competing interest
FD- 4.72E- 4.69E- 7.15E-05 6.47E- 7.01E- 7.01E-
PINN 05 05 05 05 05
The authors declare that they have no known competing financial
PINN 1.39E- 1.10E- 2.43E-03 1.00E- 1.56E- 1.56E-
02 04 04 03 03 interests or personal relationships that could have appeared to influence
gPINN 1.83E- 1.10E- 3.99 E- 1.02E- 3.50E- 1.70E- the work reported in this paper.
02 04 03 04 03 03
SA- 1.41E- 1.09 E- 2.53E-03 1.02 E- 1.61E- 1.61E- Data availability
PINN 02 04 04 03 03

The source codes to reproduce the results in this study are available
spectral bias in neural networks, as well as more comprehensive solution on GitHub at https://github.com/railleon/FD-PINN_dynamics upon
methods, warrant further investigation in future research endeavors. final publication.
Moreover, while incorporating physical constraints into the model via
PINN enhances the physical system simulations, it inevitably introduces
additional complexity. Future research is needed to address this chal­ Acknowledgements
lenge, focusing on strategies to balance the integration of physical
constraints with the optimization of simulation efficiency, ultimately This research/project was supported by Beijing Natural Science
leading to more accurate and computationally efficient simulations. Foundation (Grant No. L221021), the National Natural Science Foun­
dation of China (Grant No. 52178404), the National Research Founda­
CRediT authorship contribution statement tion, Singapore under its AI Singapore Programme (AISG Award No:
AISG2-TC-2021-001), and The Ministry of Education Tier 1 Grants,
Ruihua Liang: Writing – review & editing, Writing – original draft, Singapore (No. RG121/21 and No. RS04/21).

Appendix A. Numerical approximation of the Dirac function

The Gaussian function is employed in this study as an approximation of Dirac functions, the relationship between these two functions can be
expressed as follows [79,80],
1 x2
δ(x) = lim √̅̅̅̅̅̅̅̅̅̅e− 2σ 2 . (23)
a→∞ 2πσ 2

x2
1 ̅ −
Therefore, √̅̅̅̅̅̅̅
2πσ 2
e 2σ 2 is used in this research to approximate the Dirac function, where the parameter σ is assigned a small value.

10
R. Liang et al. International Journal of Mechanical Sciences 284 (2024) 109766

Appendix B. Supplementary calculations

The parameters of the structure and loads used in the supplementary calculations were selected with reference to the research [81], which is shown
in Table B1. The proposed FD-PINN and the traditional PINN were established to compute the dynamic response of the structure. The implementation
details and parameters of the models can be found in Section 3.2, the calculated displacement field of the models is shown in Fig. B1. It can be seen that
the proposed FD-PINN still achieves significantly more accurate simulations than the PINN, especially on the high-frequency component of the
response, which is similar to the results shown in Section 3, which further proves the superiority of the proposed FD-PINN.

Table B1
Mechanical parameters used in the cases.

v (m/s) L (m) EI (N⋅m2) m (kg/m) ξ p (kN)

Case B1 30 20 29.4 × 109 34,088 0.025 215.6

Case B2 20 20 29.4 × 109 34,088 0.025 215.6
Case B3 10 20 29.4 × 109 34,088 0.025 215.6

Fig. B1. The displacement field calculated by models across various cases and the corresponding calculation error (units: mm). In each displacement field, the
horizontal coordinates represent time, the vertical coordinate represents positions on the beam, and the colors represent the amplitude of the structural dynamic
response at different locations and moments.

References [12] Fan H, Li L, Chen G, Liu H, Ji X, Jiang X, Zhou S. An improved 3D DDA method
considering the unloading effect of tunnel excavation and its application. Comput
Geotech 2023;154:105178. https://doi.org/10.1016/j.compgeo.2022.105178.
[1] An Y, Chatzi E, Sim S-H, Laflamme S, Blachowski B, Ou J. Recent progress and
[13] Peng H, Yan J, Yu Y, Luo Y. Time series estimation based on deep learning for
future trends on damage identification methods for bridge structures. Struct
structural dynamic nonlinear prediction, in: structures. Elsevier; 2021. p. 1016–31.
Control Health Monit 2019;26:e2416.
https://www.sciencedirect.com/science/article/pii/S2352012420306901.
[2] Liu W, Tong L, Li H, Wang Z, Sun Y, Gu W. Multi-parameter intelligent inverse
accessed March 31, 2024.
analysis of a deep excavation considering path-dependent behavior of soils.
[14] Liang R, Liu W, Ma M, Liu W. An efficient model for predicting the train-induced
Comput Geotech 2024;174:106597.
ground-borne vibration and uncertainty quantification based on Bayesian neural
[3] Dessena G, Civera M, Zanotti Fragonara L, Ignatyev DI, Whidborne JF. A loewner-
network. J Sound Vib 2021;495:115908.
based system identification and structural health monitoring approach for
[15] Tsialiamanis G, Dervilis N, Wagg DJ, Worden K. Towards a population-informed
mechanical systems, struct. Control Health Monit 2023;2023:1–22. https://doi.
approach to the definition of data-driven models for structural dynamics. Mech
org/10.1155/2023/1891062.
Syst Signal Process 2023;200:110581.
[4] Jarfi H, Eskandari M, KA KA. Development of governing partial differential
[16] Liang R, Liu W, Li C, Li W, Wu Z. A novel efficient probabilistic prediction
equations of reinforcing thin films. Int J Mech Sci 2024;276:109407.
approach for train-induced ground vibrations based on transfer learning. J Vib
[5] Xu YF, Huang GL. Modal sensitivity analysis of acoustic metamaterials for
Control 2023:10775463221148792.
structural damage detection. Int J Mech Sci 2023;259:108571.
[17] Bao Y, Tang Z, Li H, Zhang Y. Computer vision and deep learning–based data
[6] Ding W, Semperlotti F. A multimesh finite element method for integral nonlocal
anomaly detection method for structural health monitoring. Struct Health Monit
elasticity using mesh-decoupling technique. Int J Mech Sci 2024;275:109260.
2019;18:401–21.
[7] Yang X-D, Zhang W, Chen L-Q, Yao M-H. Dynamical analysis of axially moving
[18] Ye XW, Jin T, Yun CB. A review on deep learning-based structural health
plate by finite difference method. Nonlinear Dyn 2012;67:997–1006. https://doi.
monitoring of civil infrastructures. Smart Struct Syst 2019;24:567–85.
org/10.1007/s11071-011-0042-2.
[19] Li C, Liu W, Liang R. Identification of vertical wheel-rail contact force based on an
[8] Jin D, Guo Y, Li X, Yuan D, Shu J, Chen J, Li F. Modeling of reinforced-concrete
analytical model and measurement and its application in predicting ground-borne
cutting with shield rippers using FEM-DEM-coupling method. Int J Mech Sci 2024;
vibration. Measurement 2021;186:110182.
282:109619.
[20] Yang H, Jiang J, Chen G, Zhao J. Dynamic load identification based on deep
[9] Ji W, Ma H, Liu F, Sun W, Wang D. Dynamic analysis of cracked pipe elbows:
convolution neural network. Mech Syst Signal Process 2023;185:109757.
Numerical and experimental studies. Int J Mech Sci 2024;281:109580.
[21] Liang R, Liu W, Li W, Wu Z. A traffic noise source identification method for
[10] Pfeil S, Gravenkamp H, Duvigneau F, Woschke E. Semi-analytical solution of the
buildings adjacent to multiple transport infrastructures based on deep learning.
Reynolds equation considering cavitation. Int J Mech Sci 2023;247:108164.
Build Environ 2022;211:108764.
[11] Zhao Y, Zhou Z, Bi J, Wang C. Simulation of brittle fractures using energy-bond-
based smoothed particle hydrodynamics. Int J Mech Sci 2023;248:108236.

11
R. Liang et al. International Journal of Mechanical Sciences 284 (2024) 109766

[22] Liu W, Liang R, Zhang H, Wu Z, Jiang B. Deep learning based identification and [50] Waheed UB. Kronecker Neural Networks Overcome Spectral Bias for PINN-Based
uncertainty analysis of metro train induced ground-borne vibration. Mech Syst Wavefield Computation. IEEE Geosci Remote Sens Lett 2022;19:1–5. https://doi.
Signal Process. 2023;189:110062. org/10.1109/LGRS.2022.3209901.
[23] Park S, Marimuthu KP, Han G, Lee H. Deep learning based nanoindentation method [51] Zhang R, Liu Y, Sun H. Physics-guided convolutional neural network (PhyCNN) for
for evaluating mechanical properties of polymers. Int J Mech Sci 2023;246: data-driven seismic response modeling. Eng Struct 2020;215:110704. https://doi.
108162. org/10.1016/j.engstruct.2020.110704.
[24] Liang R, Liu W, Kaewunruen S, Zhang H, Wu Z. Classification of external vibration [52] Nuruddeen RI, Muhammad L, Nass AM, Sulaiman TA. A review of the integral
sources through data-driven models using hybrid CNNs and LSTMs, struct. Control transforms-based decomposition methods and their applications in solving
Health Monit 2023;2023:1–18. https://doi.org/10.1155/2023/1900447. nonlinear PDEs. Palest J Math 2018;7:262–80.
[25] Zhang Z, Li H, Hu Y, Liu Y, Li Y, Li B. Few-sample information-enhanced inverse [53] Duhamel P, Vetterli M. Fast Fourier transforms: a tutorial review and a state of the
design framework for customizing transmission-modulated elastic metasurfaces. art. Signal Process 1990;19:259–99.
Int J Mech Sci 2024;279:109507. [54] Helms H. Fast Fourier transform method of computing difference equations and
[26] Xiong Z, Yang P, Zhao P. Physics-driven neural networks for nonlinear simulating filters. IEEE Trans Audio Electroacoustics 1967;15:85–90.
micromechanics. Int J Mech Sci 2024;273:109214. [55] Sallam O, Fürth M. On the use of Fourier Features-Physics Informed Neural
[27] Shang L, Zhao Y, Zheng S, Wang J, Zhang T, Wang J. Quantification of gradient Networks (FF-PINN) for forward and inverse fluid mechanics problems. Proc Inst
energy coefficients using physics-informed neural networks. Int J Mech Sci 2024; Mech Eng Part M J Eng Marit Environ 2023;237:846–66. https://doi.org/10.1177/
273:109210. https://doi.org/10.1016/j.ijmecsci.2024.109210. 14750902231166424.
[28] Batuwatta-Gamage CP, Rathnayaka CM, Karunasena HCP, Jeong H, Karim MA, [56] Zhang T, Yan R, Zhang S, Yang D, Chen A. Application of Fourier feature physics-
Gu YT. A physics-informed neural network framework to investigate nonlinear and information neural network in model of pipeline conveying fluid. Thin-Walled
heterogenous shrinkage of drying plant cells. Int J Mech Sci 2024;275:109267. Struct 2024;198:111693.
https://doi.org/10.1016/j.ijmecsci.2024.109267. [57] Ouyang H. Moving-load dynamic problems: A tutorial (with a brief overview).
[29] Liu C-X, Wang X, Liu W, Yang Y-F, Yu G-L, Liu Z. A physics-informed neural Mech Syst Signal Process. 2011;25:2039–60. https://doi.org/10.1016/j.
network for Kresling origami structures. Int J Mech Sci 2024;269:109080. https:// ymssp.2010.12.010.
doi.org/10.1016/j.ijmecsci.2024.109080. [58] Wanxie Z, Xiangxiang Z. Method of separation of variables and Hamiltonian
[30] Zhang B, Wu G, Gu Y, Wang X, Wang F. Multi-domain physics-informed neural system. Numer Methods Partial Differ Equ 1993;9:63–75. https://doi.org/
network for solving forward and inverse problems of steady-state heat conduction 10.1002/num.1690090107.
in multilayer media. Phys Fluids 2022;34. [59] Xia H, Li HL, Guo WW, De Roeck G. Vibration Resonance and Cancellation of
[31] Cai S, Wang Z, Wang S, Perdikaris P, Karniadakis GE. Physics-informed neural Simply Supported Bridges under Moving Train Loads. J Eng Mech 2014;140:
networks for heat transfer problems. J Heat Transf 2021;143:060801. 04014015. https://doi.org/10.1061/(ASCE)EM.1943-7889.0000714.
[32] Eivazi H, Tahani M, Schlatter P, Vinuesa R. Physics-informed neural networks for [60] Raissi M. Deep hidden physics models: deep learning of nonlinear partial
solving Reynolds-averaged Navier-Stokes equations. Phys Fluids 2022;34. https:// differential equations. J Mach Learn Res 2018;19:932–55.
doi.org/10.1063/5.0095270. [61] Raissi M, Perdikaris P, Karniadakis GE. Physics-informed neural networks: a deep
[33] Ranade R, Hill C, Pathak J. DiscretizationNet: A machine-learning based solver for learning framework for solving forward and inverse problems involving nonlinear
Navier–Stokes equations using finite volume discretization. Comput Methods Appl partial differential equations. J Comput Phys 2019;378:686–707.
Mech Eng 2021;378:113722. https://doi.org/10.1016/j.cma.2021.113722. [62] Karniadakis GE, Kevrekidis IG, Lu L, Perdikaris P, Wang S, Yang L. Physics-
[34] Liang R, Liu W, Xu L, Qu X, Kaewunruen S. Solving elastodynamics via physics- informed machine learning. Nat Rev Phys 2021;3:422–40.
informed neural network frequency domain method. Int J Mech Sci 2023:108575. [63] Yu J, Lu L, Meng X, Karniadakis GE. Gradient-enhanced physics-informed neural
[35] Rao C, Sun H, Liu Y. Physics-informed deep learning for computational networks for forward and inverse PDE problems. Comput Methods Appl Mech Eng
elastodynamics without labeled data. J Eng Mech 2021;147:04021043. 2022;393:114823.
[36] Bai J, Rabczuk T, Gupta A, Alzubaidi L, Gu Y. A physics-informed neural network [64] McClenny LD, Braga-Neto UM. Self-adaptive physics-informed neural networks.
technique based on a modified loss function for computational 2D and 3D solid J Comput Phys 2023;474:111722. https://doi.org/10.1016/j.jcp.2022.111722.
mechanics. Comput Mech 2023;71:543–62. https://doi.org/10.1007/s00466-022- [65] Ghayesh MH, Farokhi H, Zhang Y, Gholipour A. Nonlinear coupled moving-load
02252-0. excited dynamics of beam-mass structures. Arch Civ Mech Eng 2020;20:45.
[37] Zheng B, Li T, Qi H, Gao L, Liu X, Yuan L. Physics-informed machine learning https://doi.org/10.1007/s43452-020-00040-2.
model for computational fracture of quasi-brittle materials without labelled data. [66] Li J, Zhang H, Zhu D, Li C. A moving load amplitude spectrum for analyzing the
Int J Mech Sci 2022;223:107282. resonance and vibration cancellation of simply supported bridges under moving
[38] Tancik M, Srinivasan P, Mildenhall B, Fridovich-Keil S, Raghavan N, Singhal U, loads. Eur J Mech-ASolids 2022;92:104428.
Ramamoorthi R, Barron J, Ng R. Fourier features let networks learn high frequency [67] Burns KJ, Vasil GM, Oishi JS, Lecoanet D, Brown BP. Dedalus: A flexible framework
functions in low dimensional domains. Adv Neural Inf Process Syst 2020;33: for numerical simulations with spectral methods. Phys Rev Res 2020;2:023068.
7537–47. https://doi.org/10.1103/PhysRevResearch.2.023068.
[39] Fridovich-Keil S, Gontijo Lopes R, Roelofs R. Spectral bias in practice: the role of [68] Wang Z. Fast algorithms for the discrete W transform and for the discrete Fourier
function frequency in generalization. Adv Neural Inf Process Syst 2022;35: transform. IEEE Trans Acoust Speech Signal Process 1984;32:803–16.
7368–82. [69] Wang S, Wang H, Perdikaris P. Learning the solution operator of parametric partial
[40] Yang YB, Chen L, Wang ZL, Liu ZY, Liu D-H, Yao H, Zheng Y. Cancellation of differential equations with physics-informed deeponets. Sci Adv 2021;7:eabi8605.
resonance for elastically supported beams subjected to successive moving loads: [70] Hanin B, Rolnick D. How to start training: The effect of initialization and
Optimal design condition for bridges. Eng Struct 2024;307:117950. https://doi. architecture. Adv Neural Inf Process Syst 2018:31.
org/10.1016/j.engstruct.2024.117950. [71] Liu DC, Nocedal J. On the limited memory BFGS method for large scale
[41] Xu L, Ma M, Cao R, Tan X, Liang R. Effect of longitudinally varying characteristics optimization. Math Program 1989;45:503–28.
of soil on metro train-induced ground vibrations based on wave propagation [72] Lu L, Pestourie R, Yao W, Wang Z, Verdugo F, Johnson SG. Physics-informed neural
analysis. Soil Dyn Earthq Eng 2022;152:107020. networks with hard constraints for inverse design. SIAM J Sci Comput 2021;43:
[42] Hester D, Brownjohn J, Huseynov F, Obrien E, Gonzalez A, Casero M. Identifying B1105–32. https://doi.org/10.1137/21M1397908.
damage in a bridge by analysing rotation response to a moving load. Struct [73] Wu C, Zhu M, Tan Q, Kartha Y, Lu L. A comprehensive study of non-adaptive and
Infrastruct Eng 2020;16:1050–65. https://doi.org/10.1080/ residual-based adaptive sampling for physics-informed neural networks. Comput
15732479.2019.1680710. Methods Appl Mech Eng 2023;403:115671. https://doi.org/10.1016/j.
[43] Ma M, Xu L, Liu W, Tan X. Semi-analytical solution of a coupled tunnel-soil cma.2022.115671.
periodic model with a track slab under a moving train load. Appl Math Model [74] Wu C, Zhu M, Tan Q, Kartha Y, Lu L. A comprehensive study of non-adaptive and
2024;128:588–608. residual-based adaptive sampling for physics-informed neural networks. Comput
[44] Shao D, Li B, Cao Y, Tao Y, Sun N, Shi Y. Dynamic responses of composite plate- Methods Appl Mech Eng 2023;403:115671.
shell coupled structures under moving excitations. Int J Mech Sci 2024;265: [75] Chen F, Sondak D, Protopapas P, Mattheakis M, Liu S, Agarwal D, Di Giovanni M.
108884. Neurodiffeq: a python package for solving differential equations with neural
[45] Wang S, Wang H, Perdikaris P. On the eigenvector bias of Fourier feature networks: networks. J Open Source Softw. 2020;5:1931.
From regression to solving multi-scale PDEs with physics-informed neural [76] Hennigh O, Narasimhan S, Nabian MA, Subramaniam A, Tangsali K, Fang Z,
networks. Comput Methods Appl Mech Eng 2021;384:113938. Rietmann M, Byeon W, Choudhry S. NVIDIA SimNetTM: An AI-accelerated multi-
[46] Wang S, Yu X, Perdikaris P. When and why PINNs fail to train: a neural tangent physics simulation framework. In: Int. Conf. Comput. Sci. Springer; 2021.
kernel perspective. J Comput Phys 2022;449:110768. p. 447–61.
[47] Chai X, Cao W, Li J, Long H, Sun X. Overcoming the spectral bias problem of [77] Anandh T, Ghose D, Ganesan S. FastVPINNs: An efficient tensor-based Python
physics-informed neural networks in solving the frequency-domain acoustic wave libraryfor solving partial differential equations using hp-Variational
equation. IEEE Trans Geosci Remote Sens 2024. https://doi.org/10.1109/ PhysicsInformed Neural Networks. J Open Source Softw 2024;9:6764. https://doi.
TGRS.2024.3440471. 1–1. org/10.21105/joss.06764.
[48] Zhao Z, Shi D. Adaptive control of spectral bias in untrained neural network priors [78] Lu L, Meng X, Mao Z, Karniadakis GE. DeepXDE: a deep learning library for solving
for inverse problems. Expert Syst Appl 2024;255:124516. https://doi.org/ differential equations. SIAM Rev 2021;63:208–28.
10.1016/j.eswa.2024.124516. [79] Rivera MJ, Trujillo M, Romero-García V, López Molina JA, Berjano E. Numerical
[49] Ding W, He Q, Tong H, Wang Q, Wang P. Solving coupled differential equation resolution of the hyperbolic heat equation using smoothed mathematical functions
groups using PINO-CDE. Mech Syst Signal Process. 2024;208:111014.

12
R. Liang et al. International Journal of Mechanical Sciences 284 (2024) 109766

instead of Heaviside and Dirac delta distributions. Int Commun Heat Mass Transf [81] Kumar CPS. Vibration of simply supported beams under a single moving load_ a
2013;46:7–12. https://doi.org/10.1016/j.icheatmasstransfer.2013.05.017. detailed study of cancellation phenomenon. Int J Mech Sci 2015.
[80] Kärnä T, Deleersnijder E, De Brauwere A. Simple test cases for validating a finite
element unstructured grid fecal bacteria transport model. Appl Math Model 2010;
34:3055–70. https://doi.org/10.1016/j.apm.2010.01.012.

13