PSI - Issue 72

Oleh Yasniy et al. / Procedia Structural Integrity 72 (2025) 181–187

182

1. Introduction The prediction of fatigue crack growth rate (FCG) is a crucial problem in fracture mechanics that holds substantial importance in maintaining the reliability and safety of structures subjected to cyclic loading. These structures, such as aircraft, bridges, and mechanical components, often experience repeated stresses that can lead to the propagation of cracks over time, resulting in potential failures. Accurate prediction models for FCG are necessary to ensure timely maintenance, optimize structural design, and prevent catastrophic failures. Traditional models have been widely used to predict crack growth based on parameters like stress intensity factor range, and crack length. However, experimental data often exhibits a certain degree of scatter due to the complex nature of material behavior and external factors. Consequently, traditional models may not always capture the full scope of the crack growth process, leading to deviations between predicted and actual results. In recent years, machine learning techniques, particularly neural networks (NNs), have gained attention for their ability to model complex relationships in large datasets. Unlike traditional models, neural networks can incorporate various variables and automatically learn patterns from data, making them a powerful tool for predicting fatigue crack growth. Neural networks can consider the influence of multiple factors that affect crack propagation, such as stress ratio, loading cycles, material properties, and environmental conditions. Neural network is used in many areas of science and technology, including fracture mechanics by Pidaparti et al. (1995), Mohanty et al. (2009), and Yasnii et al. (2018, 2020). In addition, the paper by Zarrabi et al. (2008) proposes a three-layer artificial neural network (ANN) to predict fatigue crack length under constant amplitude mode I cyclic loading. The model showed higher accuracy than traditional methods, with a less than 0.05% prediction error. In particular, in the paper of Liu et al. (2005) there was introduced a novel ANN-based algorithm to evaluate fatigue crack growth processes. The study establishes the ANN model and validates its effectiveness in predicting crack growth under various loading conditions. In the work of Murakami et al. (1996) there was modeled the fatigue crack growth rate of nickel-base superalloys using a neural network within a Bayesian framework. The study highlights the model's ability to predict FCG rate with consideration of uncertainties. The paper by Zhang et al. (2021) presents an increment learning scheme based on a fully connected neural network to predict fatigue – crack growth in middle tension specimens. The proposed model adapts to varying loading conditions, enhancing prediction accuracy. In the study of Pratoori (2023), random forest and neural network frameworks were compared to predict fatigue crack growth rates in nickel superalloys. Both models exhibit strong predictive capabilities, with the neural network showing a slightly higher coefficient of determination (r² = 0.9831). In the paper of Moussouni et al. (2023), a fatigue crack growth model is developed based on the artificial neural networks (ANN) for the V-notch Charpy specimen. In addition, it is known that neural networks are used in materials science, namely, in the study of Monticeli et al. (2022) there were predicted the bending properties of 3D-printed carbon/epoxy composites with different processing parameters using an artificial neural network and statistical methods. In addition, machine learning methods as applied to modeling the thermal conductivity of epoxy-based composites with different fillers for aircraft by Yasniy et al. (2024). In the study of Konovalenko et al. (2020) it was classified steel surface defects using a deep residual neural network. In the paper of Tymoshchuk et al. (2024) it was explored machine learning methods, particularly MLP neural networks, to predict loading frequency in nickel-titanium shape memory alloys based on cyclic tensile test data. In particular, in the work of Yasniy et al. (2023) it was demonstrated that machine learning methods, particularly neural networks, effectively predict the jump-like creep behavior of AMg6 aluminum alloy.

Nomenclature a

the crack length the stress ratio

R N

the number of loading cycles n the test data set size ( ) the true value of material strain in the test data set . ( ) the predicted value of material strain in the test data set