Engineering Fracture Mechanics 325 (2025) 111319

Contents lists available at ScienceDirect

Engineering Fracture Mechanics

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

Data-assisted physics-informed neural network for predicting

fatigue life under various strain ratios and pre-strain effect
Qixuan Zhang , Wei Zhang * , Wen Chu , Dengdeng Rong , Qiaofa Yang , Tianhao Ma ,
Changyu Zhou
School of Mechanical and Power Engineering, Nanjing Tech University, Nanjing 211816, China

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

Keywords: Hydrogen storage cylinders generally experience repetitive pressurization and depressurization
DA-PINN cycles, leading to cyclic stresses that can cause progressive fatigue damage. Accurate fatigue life
Fatigue life prediction prediction is essential for ensuring the structural integrity and operational safety of these cylin­
Strain ratio
ders. Given the challenges posed by limited experimental data and poor accuracy with small
Pre-strain
datasets, this study introduces a data-assisted physics-informed neural network (DA-PINN) for
low cycle fatigue life prediction under various strain ratios and pre-strain effect. Bootstrapping
and spline interpolation techniques are used to enrich both the training and testing data set of the
neural network, ensuring a more robust learning process. Simultaneously, physical knowledge of
strain-controlled fatigue behavior is embedded into the model through constrained loss functions,
enabling physics-consistent learning. Results demonstrate that the proposed DA-PINN achieves
prediction accuracy predominantly within a ±2.0-fold error band, outperforming conventional
empirical models. Moreover, comparative analysis reveals that the bootstrapping augmentation
method yields more consistent predictions than spline interpolation. The developed DA-PINN
exhibits strong material adaptability, suggesting promising applications for other metallic ma­
terials under various strain ratios and pre-strain effect.

1. Introduction

In the burgeoning field of hydrogen energy development, the storage and transportation of hydrogen emerge as critical components
within the hydrogen energy industry chain. Central to this process are hydrogen storage cylinders, which play a pivotal role in enabling
high-density storage, portability, and mobility of hydrogen. These cylinders not only ensure the safety and durability of stored
hydrogen but also significantly enhance its utilization efficiency [1,2]. As the hydrogen economy advances and technological in­
novations continue to emerge, hydrogen storage cylinders are expected to retain their central role. Their ongoing development and
optimization are essential for promoting the widespread adoption and sustainable advancement of hydrogen energy, thereby
contributing to a cleaner and more resilient energy future. As a medium-carbon high-strength steel, 4130X steel exhibits exceptional
metallurgical characteristics, including superior tensile strength, high yield resistance, and remarkable ductility. These characteristics
make it a promising candidate for high-pressure applications [3]. However, under operational conditions, hydrogen storage systems
are subjected to cyclic pressurization and depressurization, leading to varying strain ratios and pre-strain effects. These cyclic loading

* Corresponding author.
E-mail address: [email protected] (W. Zhang).

https://doi.org/10.1016/j.engfracmech.2025.111319
Received 24 March 2025; Received in revised form 16 May 2025; Accepted 31 May 2025
Available online 2 June 2025
0013-7944/© 2025 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
Q. Zhang et al. Engineering Fracture Mechanics 325 (2025) 111319

Nomenclature

ML Machine learning
SWT Smith–Watson–Topper
ANN Artificial neural network
DNN Deep neural network
PINN Physics-informed neural network
DA-PINN Data-assisted PINN
P Pre-strain
R Strain ratio
σ max Maximum stress
σm Mean stress
εa Strain amplitude
σʹf Fatigue strength coefficient
εʹf Fatigue ductility coefficient
b Fatigue strength exponent
c Fatigue ductility exponent
Nf Fatigue life
MSE Mean squared error
MAE Mean absolute error
RL2 L2 regularization term
R2 Coefficient of determination
ECDFs Empirical cumulative distribution functions
K-S Kolmogorov-Smirnov
KSD Kolmogorov-Smirnov statistic
KSC Kolmogorov-Smirnov critical value
H0 Null hypothesis
SELU Scaled exponential linear unit
Dmax mt Maximum distributional divergenceFirst-order moment estimates of the gradient
vt Second-order moment estimates of the gradient
m
̂t Corrected values for mt
v̂t Corrected values for vt
( ( ) )
L f x(i) ; ωt , y(i) Training set samples
gt Gradient
a Learning rate

conditions induce fatigue damage that pose significant challenges to long-term structural integrity and reliability. Consequently,
robust fatigue life prediction models are required to ensure safe and durable performance [4,5].
Over the past decades, the fatigue behavior of high-pressure cylinder materials has been extensively studied under various
influencing factors, including stress amplitude [6–11], stress ratio [12–14], strain amplitude [12–16], strain ratio [17–19], pre-strain
[20,21] and environment conditions [22,23]. It has been observed that increases in strain and stress ratios often intensify plastic
deformation, accelerate crack initiation, and reduce fatigue life. Conversely, negative stress and strain ratios can impede crack
propagation and prolong fatigue life. Additionally, higher stress/strain amplitudes and pre-strain levels consistently shorten fatigue
life. Environmental factors such as elevated temperature, corrosive atmospheres, and humidity further exacerbate material degra­
dation, reducing fatigue resistance. In response, a range of empirical fatigue life prediction models have been proposed, including the
Smith–Watson–Topper model (SWT) [24,25], the Morrow model [26], the Manson-Coffin-Basquin model [27].
More recently, fatigue life prediction has increasingly been enhanced by the application of machine learning (ML) techniques,
which offer superior predictive accuracy and computational efficiency compared to traditional empirical models. ML models excel at
capturing nonlinear interactions among loading parameters, degradation mechanisms, and environmental variables—capabilities that
are often limited in physics-based approaches. For instance, Abdalla et al. [28] employed an artificial neural network (ANN) to predict
energy dissipation in steel under varying strain amplitudes and ratios, achieving high consistency with experimental results. Zhang
et al. [29] developed a hybrid model combining integrated learning with domain knowledge to predict fracture toughness in pipeline
steels under hydrogen exposure, demonstrating the superiority of the CatBoost algorithm over traditional methods. Similarly, Campari
et al. [30] introduced an ML-enhanced risk-based inspection framework to assess fatigue crack growth in hydrogen-exposed pipelines.
Jiang et al. [31] incorporated the fatigue crack propagation law of 316 stainless steel into a predictive model accounting for tem­
perature and strain rate variations, thereby improving prediction consistency. He et al. [32] combined ML techniques with Bayesian
inverse analysis to estimate fatigue life and fatigue limits of materials such as AISI 316, AISI 4140, and CA6NM, achieving strong
agreement with experimental results. Shi et al. [33] expanded the dataset of selective laser-melted AlSi10Mg alloys using

2
Q. Zhang et al. Engineering Fracture Mechanics 325 (2025) 111319

interpolation, while Duan et al. [34] utilized a deep neural network (DNN) to predict low cycle fatigue life in 316 stainless steel, using
strain amplitude, temperature, and dwell time as inputs. Qian et al. [35] applied Monte Carlo simulations to enrich the very high cycle
fatigue dataset of Ti60 alloy and used dynamic recurrent neural networks, random forest, and XGBoost models for life prediction under
different stress ratios, maximum stresses, and temperatures. Bhardwaj et al. [36] evaluated feature importance in austenitic stainless
steels using a classification and regression tree algorithm, followed by ANN-based prediction, which outperformed conventional
models in accuracy and robustness.
Addressing both forward and inverse problems governed by nonlinear partial differential equations, Raissi et al. [37] introduced
physics-informed neural network (PINN), demonstrating their effectiveness across various domains [38–41]. Karniadakis et al. [42]
outlined three principal strategies for constructing PINNs: (1) data integration or augmentation aligned with physical laws, (2) physics-
based network architecture design, and (3) loss function modification with physical constraints. A notable extension, the data-assisted
physics-informed neural network (DA-PINN), enhances model performance by incorporating empirical or simulation-generated data
and embedding governing equations into the loss function. This method has shown promise in fatigue life prediction across a range of
materials and conditions [43,44]. Feng et al. [45] improved model interpretability and accuracy by embedding physically meaningful
loss functions related to critical defect features and fracture mechanics. Salvati et al. [46] proposed a PINN for fatigue life prediction in
defective materials, offering insights into morphologically unexplained defect behaviors. Jiang et al. [47] embedded physical laws
governing 316 stainless steel at various temperatures and strain rates into the loss function, aligning predictions with physical ex­
pectations. Zhang et al. [48] compared predicted creep-fatigue life of 316 stainless steel using three conventional ML models and a
PINN. The PINN, enhanced by dual constraints in the loss function, demonstrated superior performance. Despite these advancements,
most existing ML-based fatigue life prediction models for hydrogen storage cylinder materials primarily focus on stress amplitude,
stress ratio, and mean stress as input features. Models incorporating pre-strain and strain ratio remain relatively rare. Moreover, few
studies report on integrating physical laws within deep neural network optimization frameworks under such conditions.

Fig. 1. Flowchart of the proposed DA-PINN.

3
Q. Zhang et al. Engineering Fracture Mechanics 325 (2025) 111319

In this work, a DA-PINN is proposed to predict fatigue life under strain-controlled loading conditions, accounting for the effects of
strain ratio and pre-strain. The methodological framework is illustrated in Fig. 1. Initially, fatigue test datasets of 4130X steel are
collected. Given the limited amount of data available for training, validation, and testing, the dataset is augmented through boot­
strapping and spline interpolation techniques. This augmentation serves to mitigate data sparsity while enhancing the role of data-
driven loss constraints within the hybrid loss framework of the DA-PINN. To demonstrate the performance of the proposed model,
predictions are compared against those from two conventional empirical models that consider pre-strain and strain ratio effects. The
superiority of the bootstrapping-based data augmentation method is also evaluated against spline augmentation. Furthermore, the
applicability of the DA-PINN is validated using fatigue test data from AZ61A alloy under strain-controlled loading. Comparative
analyses against standard PINN further highlight the advantages of the proposed approach.

2. Methods

2.1. Materials and experiment dataset

The experimental investigation employed 4130X steel as the test material. Fatigue tests were conducted using a hydraulic servo-
controlled MTS-809 testing machine (Fig. 2) with standard cylindrical specimens, the geometry of specimen is detailed in Fig. 2(c). The
experiments were conducted under strain-controlled mode with four distinct amplitude levels (0.45 %, 0.4 %, 0.35 %, and 0.3 %) and
four strain ratio (R) conditions (− 1, − 0.5, 0, 0.5). Triangular waveform loading was applied under maintained constant strain rate
conditions and different pre-strain (P) conditions (P = − 0, 2, 4). The statistical distribution of fatigue life results is listed in Table 1 and
presented in Fig. 3.
Fig. 3(a) and (b) present the experimental fatigue life under different pre-strains and strain ratios, respectively. An increase in pre-
strain typically implies that the material has undergone greater plastic deformation, resulting in the accumulation of microstructural
defects and a reduced resistance to fatigue failure. In the logarithmic coordinate system, a clear negative correlation is observed. As
pre-strain increases, fatigue life significantly decreases, as shown in Fig. 3(a). This result indicates that fatigue life exhibits a negative
relationship with respect to pre-strain. Moreover, previous studies have consistently shown that larger strain amplitudes (εa ) lead to
larger plastic deformation during fatigue cycling, promoting faster fatigue crack growth [12–16]. Consequently, fatigue life is
markedly reduced with increasing strain amplitude. The strain ratio, which characterizes the asymmetry of the loading cycle, also
plays a critical role. An increase in strain ratio raises the mean strain, accelerating damage accumulation during the tensile phase and
further shortening fatigue life [21], as shown in Fig. 3(b).

2.2. Traditional empirical formulas for fatigue life prediction

Traditional fatigue life prediction models, such as the Manson-Coffin model, had previously neglected the influence of mean stress
resulting from different strain ratios and pre-strain on fatigue life prediction. This limitation hindered their ability to accurately predict
fatigue life under asymmetric loading conditions. Consequently, researchers developed various models that integrated the impact of
mean stress into predictions by expanding upon the foundation laid by the Manson-Coffin model.

Fig. 2. (a, b) MTS-809 testing machine and (c) dimensions of the fatigue test samples.

4
Q. Zhang et al. Engineering Fracture Mechanics 325 (2025) 111319

Table 1
Fatigue test data for 4130X steel under strain-controlled conditions.
Specimen No. εa /2 P R Nf

1 0.3 0 − 1 24,966
2 0.3 0 − 0.5 23,467
3 0.3 0 0 21,737
4 0.3 0 0.5 16,914
5 0.35 0 − 1 14,229
6 0.35 0 − 0.5 13,547
7 0.35 0 0 12,989
8 0.35 0 0.5 10,589
9 0.4 0 − 1 7728
10 0.4 0 − 0.5 7765
11 0.4 0 0 7538
12 0.4 0 0.5 7527
13 0.45 0 − 1 4871
14 0.45 0 − 0.5 4474
15 0.45 0 0 4294
16 0.45 0 0.5 4624
17 0.3 2 − 1 18,703
18 0.3 4 − 1 17,220
19 0.35 2 − 1 10,916
20 0.35 4 − 1 8546
21 0.4 2 − 1 6862
22 0.4 4 − 1 6554
23 0.45 2 − 1 4183
24 0.45 4 − 1 4514

Fig. 3. Experimental life of 4130X stainless steel with different (a) pre-strains and (b) strain ratios.

2.2.1. SWT model

Smith et al. [49] incorporated a σ max εa stress–strain relationship to consider the effect of mean stress in their analysis. They
included the maximum stress term on the left-hand side of the Manson-Coffin equation and fitted other parameters based on results
from the aforementioned Manson-Coffin formula.
( )2
σ,f ( )2b ( )b+c
σmax εa = 2Nf + σ ,f εʹf 2Nf (1)
E

where σ,f is the fatigue strength coefficient, εʹf is the fatigue ductility coefficient, b is the fatigue strength exponent, c is the fatigue
ductility exponent, σ max is the maximum stress.

2.2.2. Morrow model

Morrow and Sinclair [50] proposed that due to stress relaxation effects, the mean stress primarily affected only the elastic portion of
a material’s behavior, with negligible impact on its plastic region. Consequently, they modified the Manson-Coffin formula by
incorporating an additional term representing the mean stress to address corrections in the elastic part while retaining other pa­
rameters based on results from the original Manson-Coffin equation. This modification allowed for more accurate predictions of fatigue

5
Q. Zhang et al. Engineering Fracture Mechanics 325 (2025) 111319

life under effect of mean stress.

( ) b
σ,f σm ( ) ( )c
εa = 1− , 2Nf + εʹf 2Nf (2)
E σf

where σm is the mean stress.

2.3. ANN

ANN is a computational model that emulates the structural and functional properties of biological neural networks [51]. It is a
highly complex nonlinear dynamic model constructed by processing intricate information, comprising three key layers: input layer,
hidden layer, and output layer. These layers receive and process data as follows: the input layer receives initial feature vectors, which
are transmitted to the hidden layer where weighted summation calculations occur during computations. The network is trained by
adjusting its weights and biases to achieve precise predictions for new, unseen data [52]. The architecture of ANN is illustrated in
Fig. 4, and the operational principle of a specific neuron follows:
(∑ )
y=f ωx + t (3)

where x is the input variables, ω is the weight parameters, t is the bias, f( • ) is the activation function, and y is the neuron’s response.

2.4. PINN

Neural networks are progressively recognized as universal models for function approximation, owing to their computational power
and inherent flexibility. However, the process of training relies heavily on a large number of experimental datasets, which can be
particularly challenging in the field of materials fatigue. In this field, available fatigue data often falls short of meeting the re­
quirements for training high-performance machine learning models due to the significant time and economic costs involved. This
makes purely data-driven approaches unsuitable for modeling fatigue problems effectively. To address these limitations, researchers
have developed frameworks that integrate a priori physics knowledge with data-driven methods into predictive models, namely PINN
[53–56]. PINN typically consists of deep neural networks, enabling them to learn from limited data while incorporating physical
principles through the training process. This allows PINN not only to leverage less data but also to predict based on fundamental
physical laws, thereby discovering empirical relationships that might otherwise be overlooked. These models make predictions more
interpretable and physically meaningful, enhancing their utility in practical applications. In this study, PINN is employed to develop a
neural network that incorporates constraints derived from macroscopic experimental phenomena, ensuring the physical consistency of
the model. The structural framework of the PINN is illustrated in Fig. 5.
Traditional empirical models for fatigue life prediction typically consider only the independent effects of either strain ratio or pre-
strain, while neglecting the combined influence of multiple parameters. This limitation often leads to unreliable predictions under

Fig. 4. Schematic of the structure of ANN.

6
Q. Zhang et al. Engineering Fracture Mechanics 325 (2025) 111319

complex loading conditions. By embedding physical derivative constraints into the neural network, the model can account for the
synergistic effects of multiple parameters and capture the sensitivities associated with both their independent and combined actions.
This enables a more accurate representation of the damage accumulation process under complex loading. It has been concluded above
that the pre-strain, strain amplitude, and strain ratio have significant effects on fatigue life. Therefore, incorporating these physical
relationships into the loss function of the neural network can enhance the model’s physical consistency and generalization capability.
These physical effects are quantitatively reflected in Eqs. (4)–(6). It is important to note that fatigue life values were log-
transformed during model training. Consequently, the fatigue life expressed in Eqs. (4)–(6) is in logarithmic form.
∂Nf
≤0 (4)
∂P

∂Nf
≤0 (5)
∂εa

∂Nf
≤0 (6)
∂R
The basic framework of the model is illustrated in Fig. 5. Compared to ANN models, this approach introduces physical losses
alongside data losses within the loss function to ensure physically consistent training results. The loss function is designed to quantify
the discrepancy between the predicted and actual values, optimally, achieving a loss value of zero would indicate perfect agreement.
To enable a smooth and differentiable integration of the partial derivatives into the neural network, the sigmoid function [57] is
employed as shown in Eqs. (7) and (8). The introduction of the parameter β modulates the curvature of the function, thereby regulating
the severity of the physical constraints, while G(x) functions as a physical penalty mechanism. Consequently, when the partial de­
rivatives significantly deviate from the physical constraints, the associated loss term approaches zero, thus preventing the model from
overfitting these regions that seriously violate the physical laws. Transforming these into a suitable form for the model, which is
defined as:
1
F(x) = (7)
1 + e− βx

G(x) = F( − |x| ) • x (8)

where, as β increases, the value of Eq. (7) approaches zero asymptotically as long as the independent variable remains below zero.
Consequently, ensuring that the input of F( • ) in Eq. (8) is negative enables adjustment of partial derivatives toward approaching zero,
β = 100 is adopted in this work.
Through Eqs. (7) and (8), the loss function is composed of two components: one part constitutes data loss calculated based on the
experimental fatigue life and predicted fatigue life, which is given by the following equation.
( )
Lossdata = H Nfexp , Nfpred (9)

Fig. 5. Schematic of the structure and propagation of PINN.

7
Q. Zhang et al. Engineering Fracture Mechanics 325 (2025) 111319

where Nfexp is experimental fatigue life, Nfpred is predicted fatigue life.

The other part constitutes physical loss, which is derived from the physical bias transformation inequality and zero, defined as
LossPDE , given by the following equation:
[ ( )] [ ( )] [ ( )]
∂Nf ∂ Nf ∂Nf
LossPDE = H 0, G + H 0, G + H 0, G (10)
∂P ∂εa ∂R
where PDE means partial differential equation, H( • ) represents the loss function, which can be formulated as follows:
⎧ (
⎪1 y − y
)2 ⃒ ⃒
⃒ ⃒
( ) ⎪⎨ true pred ⃒ytrue − ypred ⃒ ≤ δ
2
H ytrue , ypred = (⃒ ⃒ ) (11)
⎪
⎩ δ ⃒⃒ytrue − ypred ⃒⃒ − δ otherwise
⎪
2

where ytrue is the true value and ypred is the predicted value. The advantages of selecting the Huber loss function in this context are as
follows [58]. This loss function effectively combines the strengths of mean squared error (MSE) and mean absolute error (MAE). Within
the δ-boundary region, it behaves like MSE, offering sensitivity to smaller errors and preventing large outliers from dominating the
optimization process. Outside the δ-boundary, it switches to MAE, providing robustness against outliers while maintaining smoothness
at the transition point between these two behaviors. Compared to MAE, the Huber loss function offers a balance by being differentiable
everywhere, which ensures more stable and faster convergence during optimization, making it an ideal choice for this work, and δ = 1
is adopted in this work.
In order to constrain the complexity of the model and improve the generalization ability, a regular term is added to this loss
function. L2 regularization term is adopted in this work, which can be expressed as follows:

RL2 = λ‖w‖22 (12)

where λ is the regularization parameter, which is used to control the strength of the regularization, λ = 0.01 is adopted in this work.
The final loss function is as follows:
Losscombined = Lossdata + LossPDE + RL2 (13)

2.5. DA-PINN

In contrast to standard PINN, DA-PINN is characterized by incorporating real or augmented data into the training process. This
effectively mitigates the limitation of model performance caused by scarce experimental data, while maintaining physical consistency.
It also enhances the contribution of the data-driven term in the hybrid loss function. The incorporation of limited yet reliable data has
been demonstrated to enhance data constraints, improve model accuracy, and promote convergence and stability, particularly in
complex physical modeling conditions [59,60]. Furthermore, the augmented data introduces diversity, thereby enhancing the model’s
capacity to generalize to unseen cases. The present study methodically details the construction of the DA-PINN through the following
steps and is shown in Fig. 1: The initial step in the procedure is to obtain raw data from the fatigue experiment. The second step
involves the implementation of data augmentation techniques, which serve to expand the dataset and enhance data diversity. The third
step in the process entails the training and prediction of models based on an expanded data set. This is achieved by employing the
PINN, which was previously introduced. In comparison with the standard PINN, the proposed DA-PINN method demonstrates
accelerated convergence, enhanced prediction accuracy, and augmented generalization ability.

2.6. Data augmentation

2.6.1. Bootstrapping
Bootstrapping is a statistical method that involves random sampling with replacement, which belongs to the broader category of
resampling methods [61,62]. It provides a method to assess the accuracy of sample estimates by facilitating the estimation of the
sampling distribution for nearly any statistic through random sampling techniques. This method estimates statistical properties by
evaluating them within an approximating distribution, which is often chosen as the empirical distribution function of the observed
data. When the observed data can be assumed to originate from an independent and identically distributed (i.i.d.) population,

Fig. 6. Workflow of bootstrapping.

8
Q. Zhang et al. Engineering Fracture Mechanics 325 (2025) 111319

bootstrapping is implemented by generating multiple resamples with replacement from the original dataset. A simple bootstrap
process is shown in Fig. 6. A sample is drawn from a population. From this sample, resample are generated by drawing with
replacement. From the resample, the statistic x is calculated, allowing for the construction of a histogram to estimate the distribution of
x. The bootstrapping method works by treating inference of the true probability distribution J, given the original data, as analogous to
inference of the empirical distribution ̂
J, given the resampled data. In this paper, bootstrapping is performed using the’resample’
function based on ‘sklearn’ framework. The size of each sample is the same as the original data, the procedure was repeated n times for
each of the two materials to obtain 400 sets of extended fatigue test data.

2.6.2. Spline interpolation

Cubic spline interpolation divides known data points into intervals and constructs cubic polynomials for each interval [63]. These
polynomials have second-order continuous derivatives at every interior node, it should be satisfied at the interior nodes:

Sj (xi − 0) = Sj (xi + 0)(j = 0, 1, 2) (14)

The change in the function value of a cubic spline interpolation node exerts a limited influence on the sub-interval segments imme­
diately adjacent to that point, while its effect diminishes significantly for more distant sub-interval segments.

2.7. Evaluation of prediction accuracy

( )
The prediction accuracy of the machine learning model is assessed by selecting the MSE and the coefficient of determination R2 .
The formulas of MSE and R2 are expressed in Eqs. (15) and (16), respectively:
n (
1∑ )2
MSE = Nfexp
i
− Nfpre
i
(15)
n i=1

∑n ( i )2
i=1 Nfexp − Nfpre
i

R = 1− ∑ (
2
)2 (16)
n
i=1 Nfexp − Nfexp
i

n
1∑
Nfexp = Nfexp (17)
n i=1

i
where n is the number of values, Nfexp is the experimental fatigue life of sample i, Nifpre is the predicted life. The closer the MSE value is to
0 and the closer the difference between the predicted and observed values, the better the predictive performance. The R2 approaching
unity signifies a higher proportion of variance explained by the model, reflecting greater accuracy in fatigue life prediction.

2.8. Two-sample Kolmogorov-Smirnov test

The two-sample Kolmogorov-Smirnov (K-S) test is employed to assess whether the distributions of samples remain consistent before
and after enhancement. This statistical method operates on the principle that the difference between two distributions can be
quantified by calculating the maximum disparity observed in their empirical cumulative distribution functions (ECDFs) for these two
sample groups [64]. It is assumed that the null hypothesis between ECDF of the original data set EXdata,i and augmented data set EXaug,i is
{ }
tested to determine if they come from the same distribution, where i denotes the i − th variable, i ∈ P, εa , R, Nf . The two-sample
Kolmogorov-Smirnov statistic (KSD ) was employed to ascertain the statistical significance of any potential disparities in two uni­
variate ECDF. KSD is defined as:
⃒ ⃒
KSDndata, naug , i = supx ⃒EXdata, ndata ,i (x) − EXaug ,naug ,i (x) ⃒ (18)

where supx ( • ) is a maximum value, ndata is the number of samples in the original data set, and naug is the number of samples in the
augmented set.
The null hypothesis (H0 ) is as follows:
H0 : EXdata,i ( • ) = EXaug,i ( • ) (19)

The H0 is rejected at significance level α if:

KSDndata, naug , i > KSCndata, naug (20)
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
(α) 1 + naug
ndata
KSCndata, naug = − ln (21)
2 2naug

9
Q. Zhang et al. Engineering Fracture Mechanics 325 (2025) 111319

where KSCndata, naug is called a Kolmogorov-Smirnov critical value. If KSD,i > KSc , then the original hypothesis will be rejected and cannot
be represented well. If KSD,i < KSc , then the original hypothesis will be accepted and can be well represented.

3. Network training

3.1. Data augmentation

The original dataset comprising 24 fatigue test specimens of 4130X steel proves insufficient for developing robust data-driven
prognostic models. To address this limitation, bootstrapping technique (Fig. 6) was used for experimental data augmentation, sys­
tematically expanding the dataset to 400 virtual specimens through computational replication. The efficacy of this enhancement
protocol is rigorously validated via two-sample K-S hypothesis testing, which statistically compares the empirical distribution func­
tions of original and augmented datasets. Quantitative validation was performed through computation of the KSD using the formu­
lation in Eqs. (18)–(21). Fig. 7 presents the critical analysis of the maximum distributional divergence (Dmax ) at 95 % confidence level
(α = 0.05) across all material response variables. The statistical results demonstrate that the H0 is not rejected for any compared
parameter (p > 0.05), confirming that the bootstrap-augmented dataset maintains distributional fidelity to the original experimental
population. This empirical validation establishes the enhanced dataset as statistically representative for subsequent fatigue life
modeling applications.

3.2. Data preprocessing and segmentation

The augmented dataset generated through data augmentation techniques was subsequently employed for life prediction model
( )
development. The predictive model utilizes three strain-related parameters as input features: εa , P, and R, with fatigue life Nf serving
as the target output variable. To address dimensional heterogeneity among input features, accelerate model convergence, and enhance
computational stability, a critical preprocessing step was implemented through feature normalization following Eq. (22):
(X − Xmin )
Ẋ = (22)
Xmax − Xmin

X represents the feature, Ẋ is the normalized feature. To reduce the data span, the fatigue life was logarithmized by Eq. (23) and then
normalized.

Ṅf = log10 Nf (23)

where Ṅf represents the logarithmized fatigue life, Nf is the fatigue life. The dataset was initially partitioned into two subsets, with 70
% allocated for training and the remaining 30 % reserved for testing. Subsequently, 20 % of the training set was further separated to
form a validation set. Details of the data partitioning are provided in Table 2.

3.2. Network training and hyperparameter configuration

This section details the implementation of DA-PINN architecture. Its network architecture is illustrated in Fig. 5. Hyperparameter
tuning using the enhanced dataset of 4130X steel. Several configurations of hidden layers (1–3), number of neurons per layer (16–64),
and learning rates (1e-4 to 1e-2) were tested using a validation set. The final configuration, as shown in Table 3, was chosen based on the
best validation performance. To address persistent gradient instability issues common in deep architectures, the scaled exponential

Fig. 7. Value of two-sample Kolmogorov-Smirnov test for 4130X.

10
Q. Zhang et al. Engineering Fracture Mechanics 325 (2025) 111319

Table 2
Data partitioning of the 4130X steel dataset under the DA-PINN.
Material Training set Validation set Test set

4130X 224 56 120

linear unit (SELU) activation function-a self-normalizing transformation was implemented with closed-form solution:
{
xx > 0
SELU(x) = λ (24)
α(ex − 1)x ≤ 0

where λ and α are two constants, usually set to: λ = 1.0507, α = 1.67326. This activation mechanism [65] intrinsically maintains
mean-zero activation outputs with unit variance through its self-normalizing properties, effectively stabilizing gradient propagation
during backpropagation while preserving higher-order feature representations.
The training process employs the Adam optimizer [66], which synergistically integrates the advantages of both Momentum and
RMSprop algorithms to achieve accelerated convergence with minimal fluctuations. This optimization algorithm dynamically regu­
lates the learning rate by estimating and applying bias correction to the first and second moments of the gradient. The mathematical
formulation of this process is presented in Eq. (25).
⎧ mt = β1 mt− 1 + (1 − β1 )gt
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪ vt = β2 vt− 1 + (1 − β2 )gt 2
⎪
⎪
⎪
⎪ mt
⎪
⎪
⎪ m
̂t =
⎪
⎪
⎪ 1 − β1 t
⎪
⎨
vt (25)
v̂t =
⎪
⎪
⎪ 1 − β2 t
⎪
⎪
⎪
⎪
⎪ a
⎪
⎪
⎪ ωt = ωt− 1 + √̅̅̅̅̅̅̅̅̅̅̅̅̅ m
̂t
⎪
⎪
⎪ vt + ε
̂
⎪
⎪
⎪
⎩ g = 1 ∇ ∑ L( f ( x(i) ; ω ), y(i) )
⎪
⎪
m t
m t i

where mt , vt are first-order moment estimates and second-order moment estimates of the gradient, m ̂t , v̂t are the corrected values,
( ( ) )
L f x(i) ; ωt , y(i) is the training set samples, gt is the gradient, a is the learning rate, which is set to 0.001, β1 and β2 are the exponential
decay rates of 0.9 and 0.99, respectively and ε is 10-8. During training, early stopping was employed with a patience of 20 epochs,
monitoring the validation loss. Training was halted when no improvement was observed, and the model weights were restored to those
corresponding to the lowest validation loss. Furthermore, Table 3 also provides the detailed hyperparameters used for early stopping.

4. Prediction results and discussion

4.1. Fatigue life prediction based on empirical formulas

Prior to the implementation of the DA-PINN, a comprehensive analysis is conducted to evaluate the predictive performance of two

Table 3
Parameters used in DA-PINN for 4130X steel.
Component DA-PINN

Inputs 3(strain amplitude, strain ratio, pre-strain)

Hidden layers 2
Neurons per layer 32, 16
Activation function SELU
Output 1(log10 Nf )
Learning rate 1e-3
Optimizer Adam
Normalization Min-Max-Scaler
Epochs(max) 1000
Early stopping Enabled
Patience 20
Monitor Validation loss
Mode ’min’
Restore best weights True
Loss function Losscombined
Data augmentation Bootstrapping or spline interpolation

11
Q. Zhang et al. Engineering Fracture Mechanics 325 (2025) 111319

distinct Manson-Coffin formulation models that incorporate mean stress effects. The Manson-Coffin formula, having been originally
formulated for symmetric cyclic loading conditions, necessitates independent fitting of the elastic and plastic deformation components
in the cyclic loading behavior of 4130X steel. This analytical approach facilitates the determination of fatigue parameters of σʹf , εʹf , b,
and c. The experimental curves of elastic strain amplitude, plastic strain amplitude, and total strain amplitude under low cycle fatigue
conditions for 4130X steel, obtained based on the Manson-Coffin model, are shown in Fig. 8. By fitting the elastic and plastic regions
during symmetric cyclic loading of 4130X steel, σʹf , εʹf , b, and c were determined.

(1) SWT model

By substituting the respective parameters into Eq. (26), the SWT model applicable to 4130X steel can be obtained, with the specific
expression as follows:
( )− 0.32 ( )− 0.532
σmax εa = 30.18 2Nf + 130.79 2Nf (26)

(2) Morrow model

By substituting the respective parameters into Eq. (27), the Morrow model applicable to 4130X steel can be obtained, with the
specific expression as follows:
( σm )( )− 0.16 ( )− 0.372
εa = 0.012 1 − 2Nf + 0.052 2Nf (27)
2515.2
Table 4 lists the fitted empirical formula parameters and the corresponding analytical results are comprehensively illustrated in
Fig. 9. The results reveal distinct performance characteristics of the SWT model across different strain conditions. Notably, the model
demonstrates superior predictive accuracy under high strain amplitudes, where the prediction error is significantly minimized.
However, the model’s performance is compromised at low strain ratios, where it systematically overestimates the mean stress in­
fluence, leading to larger deviations from experimental data with consistently conservative predictions. The observed behavior can be
attributed to the inherent characteristics of mean stress relaxation under varying strain conditions. At high strain amplitudes, the
accelerated rate of mean stress relaxation contributes to the SWT model’s accuracy, despite the Morrow model’s inherent neglect of
mean stress relaxation effects. Conversely, at low strain amplitudes, the reduced relaxation rate amplifies the significance of mean
stress relaxation in fatigue damage accumulation, resulting in the Morrow model’s tendency to produce overestimated predictions
compared to experimental values. A comprehensive assessment of model performance indicates that the Morrow model maintains
superior overall prediction accuracy compared to the SWT model, as evidenced by the comparative data presented in Fig. 9.

4.2. Predictive performance of DA-PINN

Fig. 10(a) delineates the convergence behavior of the DA-PINN through the loss function evolution for 4130X steel. The stabili­
zation of Total Loss beyond 45 training epochs provides explicit evidence of model convergence. Notably, the validation set achieves a
minimized Total Loss of 0.0256, indicative of high-fidelity alignment between predicted and experimental fatigue life. The fatigue life
prediction results for the 4130X steel under the DA-PINN are shown in Fig. 10(b), all prediction results fall within an ±1.5-fold error
band. The model shows MSE of 0.0041 and R2 of 0.9354 for 4130X steel. This indicates that compared to traditional empirical formula
models, the DA-PINN approach achieves superior performance in terms of predictive accuracy and model precision. The increase in R2

Fig. 8. Elasticity, plastic and total strain amplitude versus fatigue life.

12
Q. Zhang et al. Engineering Fracture Mechanics 325 (2025) 111319

Table 4
Empirical formula parameters.
σ,f /E E εʹf b c

0.012 209,600 0.052 − 0.16 − 0.372

Fig. 9. Comparison of experimetal fatigue life and predicted fatigue life using SWT and Morrow model.

Fig. 10. (a) Loss function evolution curves and (b) fatigue life prediction results of DA-PINN for 4130X steel.

while a significant reduction in MSE reflects enhanced prediction capability and better fitting effect, showcasing the advantages of the
DA-PINN method over conventional models.

4.3. Comparison of results with different materials

To validate the applicability of the model, AZ61A alloy was introduced to test its performance. A total of 25 datasets were retrieved
from the literature of AZ61A steel [67], as illustrated in Fig. 11(a) and Table 5. To ensure statistical robustness, a bootstrapping al­
gorithm was used to perform data augmentation, expanding the dataset to 400 experimental samples. Subsequently, a two-sample K-S
test was conducted to statistically compare the enhanced data distribution against the original dataset, with comprehensive test results
detailed in Fig. 11(b). Following consistent validation methodology, the dataset partitioning and network configuration of the DA-
PINN are strictly aligned with those used for 4130X steel, enabling direct comparative analysis between different materials.

13
Q. Zhang et al. Engineering Fracture Mechanics 325 (2025) 111319

Fig. 12(a) presents the convergence behavior of the DA-PINN through the loss function evolution for AZ61A alloy. The stabilization
of total loss beyond 61 training epochs provides explicit evidence of model convergence. Notably, the validation set achieves a
minimized Total Loss of 0.034, indicative of high-fidelity alignment between simulated and experimental fatigue cycle data. The
fatigue life prediction results for AZ61A using the DA-PINN are shown in Fig. 12(b). The error bands for the distribution of prediction
results reveal that most data points lie within a ±3-fold error band. The model achieves a MSE of 0.1066 and a R2 of 0.8521, indicating
strong predictive capability that surpasses benchmark values in metallic fatigue prediction literature. These results suggest the DA-
PINN has the potential to be generalized for predicting the fatigue life of other materials under similar cyclic loading conditions.

4.4. Comparison of results with PINN

To verify the superiority of the propose DA-PINN, comparative analysis were conducted on two materials using PINN. As shown in
Table 6, the network configuration and hyperparameters are consistent with those used in the DA-PINN. Furthermore, the dataset was
partitioned following the same strategy as that applied to 4130X steel under the DA-PINN. As shown in Fig. 13(a) and (b), the
convergence behaviors of the DA-PINN for 4130X steel and AZ61A alloy are presented, where the total loss stabilizes after 274 and 297
iterations, respectively. The corresponding validation losses for 4130X steel and AZ61A alloy are 0.0086 and 0.2735. Compared with
the DA-PINN, the PINN requires more iterations to converge and yet exhibited higher MSE losses, indicating that the DA-PINN achieves
faster convergence and better generalization capability. The fatigue life prediction results for 4130X steel and AZ61A alloy using the
DA-PINN and PINN are illustrated in Fig. 13(c) and (d). In contrast, the DA-PINN demonstrates superior predictive performance, with
all predicted points for 4130X steel falling within the ±1.5-fold error band. For AZ61A alloy, the predictions of the DA-PINN are also
closer to the experimental data compared to those of the PINN, which deviates significantly. Furthermore, the R2 values for both
materials are lower for the PINN than for the DA-PINN, and the MSE values for the PINN are higher in both cases. These results
collectively demonstrate that, compared with the PINN, the DA-PINN provides predictions that are more consistent with experimental
results and exhibits superior performance in fatigue life prediction.

4.5. Comparison of results under different data augmentation

Fig. 14(a) and (b) present a comparative analysis of data augmentation efficacy between bootstrapping and spline interpolation for
4130X steel and AZ61A alloy, respectively. For the spline interpolation, the network configuration and hyperparameters are the same
as those of the DA-PINN, as presented in Table 3. The dataset was also partitioned in the same manner as for 4130X steel in the DA-
PINN. The predictive performance of these augmentation strategies is summarized in Table 7. Notably, the bootstrapping approach
demonstrates prediction errors predominantly within a ±1.5-fold error band for 4130X steel, while spline interpolation-based
augmentation shows relatively larger deviations with most predicted values exhibiting up to ±2-fold error band (Fig. 14(a)). For
AZ61A alloy, the bootstrapping-enhanced data augmentation demonstrates prediction accuracy mostly within a ±1.5–3-fold error
band, whereas a lot of data of spline interpolation exceed ±3-fold error band (Fig. 14(b)). This comparative analysis reveals the su­
perior efficacy of bootstrapping in prediction error mitigation. Furthermore, DA-PINN trained with bootstrapping-augmented datasets
exhibits statistically significant performance enhancement over spline interpolation counterparts for both 4130X steel and AZ61A
alloy (Table 7). This evidence suggests that the bootstrapping mechanism better preserves the inherent stochastic characteristics and
original data distribution patterns critical for maintaining physical consistency in material deformation modeling.
Comparative analysis of Fig. 15 indicates that the predictive performance of the SWT and Morrow models degrades following data
augmentation compared to the DA-PINN. Specifically, the Morrow model’s predictions after data augmentation exceed the ±2-fold
error band, while the SWT model, under bootstrapping data augmentation, still exhibits inferior predictive accuracy relative to the DA-
PINN. Quantitatively, the differences in R2 and MSE under bootstrapping data augmentation are 14 % and 0.008, respectively. These
findings provide further empirical validation that the DA-PINN demonstrates superior predictive accuracy and generalization capa­
bility compared to conventional empirical models, highlighting its potential for more reliable fatigue life estimation across different
datasets.

Fig. 11. (a) Fatigue data of AZ61A alloy with different strain ratio and (b) the value of Two-sample Kolmogorov-Smirnov test for AZ61A.

14
Q. Zhang et al. Engineering Fracture Mechanics 325 (2025) 111319

Table 5
Fatigue test data for AZ61A under strain-controlled conditions [67].
Specimen No. εa /2 P R Nf

1 1 0 − 1 400
2 0.8 0 − 1 490
3 0.7 0 − 1 1020
4 0.6 0 − 1 1380
5 0.52 0 − 1 2620
6 0.5 0 − 1 4500
7 0.45 0 − 1 5860
8 0.4 0 − 1 8740
9 0.35 0 − 1 16,000
10 0.3 0 − 1 16,750
11 0.25 0 − 1 39,600
12 0.2 0 − 1 113,000
13 0.19 0 − 1 128,100
14 0.8 0 0 390
15 0.7 0 0 540
16 0.6 0 0 605
17 0.55 0 0 895
18 0.5 0 0 2000
19 0.4 0 0 3296
20 0.35 0 0 4037
21 0.3 0 0 11,952
22 0.25 0 0 15,107
23 0.2 0 0 50,070
24 0.15 0 0 144,300
25 1 0 0 340

Fig. 12. (a) Evolution curves of the loss function and (b) predicted fatigue life by DA-PINN for the AZ61A alloy.

5. Conclusion

In this study, a DA-PINN was developed for predicting fatigue life under varying strain ratios and pre-strain conditions, with
physical constraints incorporated into the network to enhance predictive accuracy. Due to the limited size of the original dataset, data
augmentation techniques were employed to improve the model’s generalization capability. The findings of the study are summarized
as follows:

(1) By integrating physical partial derivatives into the loss function of a deep neural network as part of the model’s physics-
informed framework, the predicted outcomes adhere to established physical principles, thereby accurately capturing the fac­
tors influencing fatigue behavior.
(2) Validation across different materials and comparison with the standard PINN model confirm that the proposed DA-PINN
achieves high predictive accuracy under various pre-strain and strain ratio conditions. Most predicted values fall within the
±2-fold error band, with high coefficients of determination, underscoring the model’s strong generalization capability.

15
Q. Zhang et al. Engineering Fracture Mechanics 325 (2025) 111319

Table 6
Parameters used in PINN for two materials.
Component PINN

Inputs 3(strain amplitude, strain ratio, pre-strain)

Hidden layers 2
Neurons per layer 32, 16
Activation function SELU
Output 1(log10 Nf )
Learning rate 1e-3
Optimizer Adam
Normalization Min-Max-Scaler
Epochs(max) 1000
Early stopping Enabled
Patience 20
Monitor Validation loss
Mode ’min’
Restore best weights True
Loss function Losscombined
Data augmentation none

Fig. 13. Evolution curves of the loss function and prediction results of fatigue life: (a) evolution curves of the loss function for 4130X steel (b)
evolution curves of the loss function for AZ61A alloy (c) predicted fatigue life by DA-PINN and PINN for the 4130X (d) predicted fatigue life by DA-
PINN and PINN for the AZ61A.

(3) Comparative analysis reveals that the proposed DA-PINN significantly outperforms conventional empirical formulas, demon­
strating superior predictive precision, enhanced material generality, and robust extrapolation capabilities across variable
loading scenarios.
(4) Bootstrapping and spline interpolation methods were utilized to expand the dataset and generate sufficient training samples.
Comparative analysis shows that bootstrapping yields more accurate and consistent fatigue life predictions for both 4130X steel
and AZ61A alloy than spline interpolation.

16
Q. Zhang et al. Engineering Fracture Mechanics 325 (2025) 111319

Fig. 14. Predicted fatigue life with different data-augmentation of 4130X steel (a) and AZ61A alloy (b).

Table 7
The prediction performance of the PINN with different data-augmentation.
Materials Data-augmentation R2 MSE

4130X Bootstrapping 0.9354 0.0041

4130X Spline interpolation 0.8650 0.0070
AZ61A Bootstrapping 0.8521 0.1066
AZ61A Spline interpolation 0.8366 0.1332

Fig. 15. Fatigue life prediction of (a) SWT and (b) Morrow models with data augmentation of 4130X steel.

CRediT authorship contribution statement

Qixuan Zhang: Writing – original draft, Methodology, Investigation. Wei Zhang: Writing – review & editing, Supervision, Funding
acquisition, Conceptualization. Wen Chu: Methodology, Investigation. Dengdeng Rong: Methodology, Investigation. Qiaofa Yang:
Methodology, Investigation. Tianhao Ma: Methodology, Investigation. Changyu Zhou: Writing – review & editing, Supervision.

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.

17
Q. Zhang et al. Engineering Fracture Mechanics 325 (2025) 111319

Acknowledgements

The authors gratefully acknowledge the financial support of the National Natural Science Foundation of China (No. 52005250) and
Postgraduate Research &Practice Innovation Program of Jiangsu Province (College Project) (No. JXSS-037).

Data availability

Data will be made available on request.

References

[1] Usman MR. Hydrogen storage methods: review and current status. Renew Sustain Energy Rev 2022;167:112743.
[2] Niaz S, Manzoor T, Pandith AH. Hydrogen storage: materials, methods and perspectives. Renew Sustain Energy Rev 2015;50:457–69.
[3] Bakhshi S, Mirak A. The effect of low temperature transformation time on microstructural & textural evolution, mechanical properties and fracture behavior of a
low alloy, medium carbon, super strength AISI 4340 steel. Mater Sci Engng A 2022;831:142247.
[4] Wu E, Zhao Y, Zhao B, Xu W. Fatigue life prediction and verification of high-pressure hydrogen storage vessel. Int J Hydrogen Energy 2021;46:30412–22.
[5] Zhang M, Lv H, Kang H, Zhou W, Zhang C. A literature review of failure prediction and analysis methods for composite high-pressure hydrogen storage tanks. Int
J Hydrogen Energy 2019;44:25777–99.
[6] Ahlström J, Karlsson B. Fatigue behaviour of rail steel—a comparison between strain and stress controlled loading. Wear 2005;258:1187–93.
[7] He L, Zheng J, Li T, Zhou H, Xia L, Jiang B. The effect of precipitates on low cycle fatigue behavior of WE54 magnesium alloys under stress-controlled mode. Int
J Fatigue 2024;179:108061.
[8] Pahlevanpour AH, Behravesh SB, Adibnazari S, Jahed H. Characterization of anisotropic behaviour of ZK60 extrusion under stress-control condition and notes
on fatigue modeling. Int J Fatigue 2019;127:101–9.
[9] Zhang H, Wang Q, Gong X, Wang T, Pei Y, Zhang W, et al. Comparisons of low cycle fatigue response, damage mechanism, and life prediction of MarBN steel
under stress and strain-controlled modes. Int J Fatigue 2021;149:106291.
[10] Koh SK. Fatigue damage evaluation of a high pressure tube steel using cyclic strain energy density. Int J Press Vessel Pip 2002;79(12):791–8.
[11] Li Y, Wang W, Pan M, et al. Fatigue life analysis of high-pressure seamless steel cylinder for hydrogen using autofrettage design. Int J Press Vessel Pip 2023;206:
105065.
[12] Golański G, Mroziński S. Low cycle fatigue and cyclic softening behaviour of martensitic cast steel. Engng Fail Anal 2013;35:692–702.
[13] Gong X, Wang T, Li Q, Liu Y, Zhang H, Zhang W, et al. Cyclic responses and microstructure sensitivity of Cr-based turbine steel under different strain ratios in
low cycle fatigue regime. Mater Des 2021;201:109529.
[14] Maharaja H, Das B, Singh A, Mishra S. Comparative assessment of strain-controlled fatigue performance of SS 316L at room and low temperatures. Int J Fatigue
2023;166:107251.
[15] Shao CW, Zhang P, Liu R, Zhang ZJ, Pang JC, Zhang ZF. Low cycle and extremely-low cycle fatigue behaviors of high-Mn austenitic TRIP/TWIP alloys: property
evaluation, damage mechanisms and life prediction. Acta Mater 2016;103:781–95.
[16] Wang Q, Wang Q, Gong X, Wang T, Zhang W, Li L, et al. A comparative study of low cycle fatigue behavior and microstructure of Cr-based steel at room and
high temperatures. Mater Des 2020;195:109000.
[17] Chen G, Lu L-t, Cui Y, Xing R-s, Gao H, Chen X. Ratcheting and low cycle fatigue characterizations of extruded AZ31B Mg alloy with and without corrosive
environment. Int J Fatigue 2015;80:364–71.
[18] Lin YC, Liu Z-H, Chen X-M, Chen J. Uniaxial ratcheting and fatigue failure behaviors of hot-rolled AZ31B magnesium alloy under asymmetrical cyclic stress-
controlled loadings. Mater Sci Engng A 2013;573:234–44.
[19] Yang X. Low cycle fatigue and cyclic stress ratcheting failure behavior of carbon steel 45 under uniaxial cyclic loading. Int J Fatigue 2005;27:1124–32.
[20] Branco R, Costa JD, Borrego LP, Wu SC, Long XY, Antunes FV. Effect of tensile pre-strain on low cycle fatigue behaviour of 7050-T6 aluminium alloy. Engng Fail
Anal 2020;114:104592.
[21] Rong D, Zhang W, Chen W, Li X, Zhao G, He X, et al. Effect of strain ratio and pre-strain on the low cycle fatigue behavior of 4130X steel at different strain
amplitudes. Mater Today Commun 2024;41:110340.
[22] Akid R. The influence of environment upon the accumulation of damage under corrosion fatigue conditions. Fatigue Fract Engng Mater Struct 1996;19(2–3):
277–85.
[23] Prowse RL, Wayman ML. Effect of environment on the fatigue behavior of a medium Carbon steel. Corrosion 2013;30:280–4.
[24] Zhou J, Huang H-Z, Peng Z. Fatigue life prediction of turbine blades based on a modified equivalent strain model. J Mech Sci Technol 2017;31:4203–13.
[25] Yu A, Huang H-Z, Li Y-F, Li H, Zeng Y. Fatigue life prediction of rolling bearings based on modified SWT mean stress correction. Chinese Journal of Mechanical
Engineering 2021;34:110.
[26] Ince A, Glinka G. A modification of Morrow and Smith–Watson–Topper mean stress correction models. Fatigue Fract Engng Mater Struct 2011;34(11):854–67.
[27] Usabiaga H, Muniz-Calvente M, Ramalle M, Urresti I, Fernández CA. Improving with probabilistic and scale features the basquin linear and bi-linear fatigue
models. Engng Fail Anal 2020;116:104728.
[28] Abdalla JA, Hawileh RA. Assessment of effect of strain amplitude and strain ratio on energy dissipation using machine Learning. Cham: Springer International
Publishing; 2021. p. 98–108.
[29] Yingjie Z, Yibo A, Weidong Z. Physics informed ensemble learning used for interval prediction of fracture toughness of pipeline steels in hydrogen environments.
Theor Appl Fract Mech 2024;130:104302.
[30] Campari A, Vianello C, Ustolin F, Alvaro A, Paltrinieri N. Machine learning-aided risk-based inspection strategy for hydrogen technologies. Process Saf Environ
Prot 2024;191:1239–53.
[31] Jiang L, Hu Y, Liu Y, Zhang X, Kang G, Kan Q. Physics-informed machine learning for low cycle fatigue life prediction of 316 stainless steels. Int J Fatigue 2024;
182:108187.
[32] Gan L, Wu H, Zhong Z. Fatigue life prediction considering mean stress effect based on random forests and kernel extreme learning machine. Int J Fatigue 2022;
158:106761.
[33] Shi T, Sun J, Li J, Qian G, Hong Y. Machine learning based very-high-cycle fatigue life prediction of AlSi10Mg alloy fabricated by selective laser melting. Int J
Fatigue 2023;171:107585.
[34] Duan H, Yue S, Liu Y, He H, Zhang Z, Zhao Y. A deep learning-based method for predicting the low cycle fatigue life of austenitic stainless steel. Mater Res
Express 2023;10:086506.
[35] Qian H, Huang Z, Xu Y, Zhou Q, Wang J, Shen J, et al. Very high cycle fatigue life prediction of Ti60 alloy based on machine learning with data enhancement.
Engng Fract Mech 2023;289:109431.
[36] Kumar Bhardwaj H, Shukla M. Low cycle fatigue life prediction of austenitic stainless steel alloys: a data-driven approach with identification of key features. Int
J Fatigue 2024;187:108454.
[37] Raissi M, Perdikaris P, Karniadakis GE. Physics-informed neural networks: a deep learning framework for solving forward and inverse problems involving
nonlinear partial differential equations. J Comput Phys 2019;378:686–707.

18
Q. Zhang et al. Engineering Fracture Mechanics 325 (2025) 111319

[38] Jagtap AD, Kawaguchi K, Karniadakis GE. Adaptive activation functions accelerate convergence in deep and physics-informed neural networks. J Comput Phys
2020;404:109136.
[39] Mao Z, Jagtap AD, Karniadakis GE. Physics-informed neural networks for high-speed flows. Comput Methods Appl Mech Engng 2020;360:112789.
[40] Rao C, Sun H, Liu Y. Physics-informed deep Learning for computational elastodynamics without labeled data. J Engng Mech 2021;147:04021043.
[41] Shukla K, Di Leoni PC, Blackshire J, Sparkman D, Karniadakis GE. Physics-informed neural network for ultrasound nondestructive quantification of Surface
breaking Cracks. J Nondestr Eval 2020;39:61.
[42] Karniadakis GE, Kevrekidis IG, Lu L, Perdikaris P, Wang S, Yang L. Physics-informed machine learning. Nat Rev Phys 2021;3:422–40.
[43] Ma S, Dang Y, Sun Y, Yang Z. A data-assisted physics-informed neural network for predicting fatigue life of electronic components under complex shock loads.
Int J Fatigue 2025;197:108933.
[44] Wang C, Xiao Q, Zhou Z, Yang Y, Kosec G, Wang L, et al. A data-assisted physics-informed neural network (DA-PINN) for fretting fatigue lifetime prediction. Int
J Mech Syst Dynam 2024;4:361–73.
[45] Feng F, Zhu T, Yang B, Zhou S, Xiao S. A physics-informed neural network approach for predicting fatigue life of SLM 316L stainless steel based on defect
features. Int J Fatigue 2024;188:108486.
[46] Jiang L, Hu Y, Liu Y, Zhang X, Kang G, Kan Q. Physics-informed machine learning for low-cycle fatigue life prediction of 316 stainless steels. Int J Fatigue 2024;
182:108187.
[47] Salvati E, Tognan A, Laurenti L, Pelegatti M, De Bona F. A defect-based physics-informed machine learning framework for fatigue finite life prediction in
additive manufacturing. Mater Des 2022;222:111089.
[48] Zhang X-C, Gong J-G, Xuan F-Z. A physics-informed neural network for creep-fatigue life prediction of components at elevated temperatures. Engng Fract Mech
2021;258:108130.
[49] Smith KN, Topper TH, Watson P. A stress–strain function for the fatigue of metals (stress-strain function for metal fatigue including mean stress effect). J Mater
1970;5:767–78.
[50] Morrow J, Sinclair GM. Cycle-dependent stress relaxation symposium on basic mechanisms of fatigue. ASTM International; 1959.
[51] Krogh A. What are artificial neural networks? Nat Biotechnol 2008;26:195–7.
[52] Fathalla E, Tanaka Y, Maekawa K. Remaining fatigue life assessment of in-service road bridge decks based upon artificial neural networks. Engng Struct 2018;
171:602–16.
[53] Wang Q, Ma Y, Zhao K, Tian Y. A comprehensive survey of loss functions in machine Learning. Ann Data Sci 2022;9:187–212.
[54] Wang L, Zhu S-P, Luo C, Liao D, Wang Q. Physics-guided machine learning frameworks for fatigue life prediction of AM materials. Int J Fatigue 2023;172:
107658.
[55] Lian Z, Li M, Lu W. Fatigue life prediction of aluminum alloy via knowledge-based machine learning. Int J Fatigue 2022;157:106716.
[56] Hao WQ, Tan L, Yang XG, Shi DQ, Wang ML, Miao GL, et al. A physics-informed machine learning approach for notch fatigue evaluation of alloys used in
aerospace. Int J Fatigue 2023;170:107536.
[57] Haghighat E, Abouali S, Vaziri R. Constitutive model characterization and discovery using physics-informed deep learning. Engng Appl Artif Intel 2023;120:
105828.
[58] Giannella V, Bardozzo F, Postiglione A, Tagliaferri R, Sepe R, Armentani E. Neural networks for fatigue crack propagation predictions in real-time under
uncertainty. Comput Struct 2023;288:107157.
[59] Hervé de Beaulieu M, Jha MS, Garnier H, Cerbah F. Remaining useful life prediction based on physics-informed data augmentation. Reliab Engng Syst Saf 2024;
252:110451.
[60] Horňas J, Běhal J, Homola P, Doubrava R, Holzleitner M, Senck S. A machine learning based approach with an augmented dataset for fatigue life prediction of
additively manufactured Ti-6Al-4V samples. Engng Fract Mech 2023;293:109709.
[61] Affonso TB, Conceição SV, Muniz LR, de Freitas Almeida JF, de Lima JC. A new hybrid forecasting method for spare part inventory management using heuristics
and bootstrapping. Dec Anal J 2024;10:100415.
[62] Zhang M, Liu X, Wang Y, Wang X. Parameter distribution characteristics of material fatigue life using improved bootstrap method. Int J Damage Mech 2019;28
(5):772–93.
[63] Farid M. Data-driven method for real-time prediction and uncertainty quantification of fatigue failure under stochastic loading using artificial neural networks
and Gaussian process regression. Int J Fatigue 2022;155:106415.
[64] Mora-López L, Mora J. An adaptive algorithm for clustering cumulative probability distribution functions using the Kolmogorov–Smirnov two-sample test.
Expert Syst Appl 2015;42:4016–21.
[65] Klambauer G, Unterthiner T, Mayr A, Hochreiter S. Self-normalizing neural networks. Adv Neural Inf Proces Syst 2017;30.
[66] Long X, Lu C, Su Y, Dai Y. Machine learning framework for predicting the low cycle fatigue life of lead-free solders. Engng Fail Anal 2023;148:107228.
[67] Yu Q, Zhang J, Jiang Y, Li Q. Effect of strain ratio on cyclic deformation and fatigue of extruded AZ61A magnesium alloy. Int J Fatigue 2012;44:225–33.

19