PSI - Issue 81

Oleh Yasniy et al. / Procedia Structural Integrity 81 (2026) 116–122

117

One of the most common steel materials used in automotive manufacturing is QSTE340TM steel. Its chemical composition and thermomechanical treatment ensure high strength and good fatigue life characteristics, making it the optimal choice for components subject to cyclic loading. However, even such materials can fail due to the accumulation of damage from prolonged loading. Classic approaches to crack growth assessment are usually based on fracture mechanics models, but these approaches may not be accurate enough for complex loading conditions, such as overload, which lead to the phenomenon of crack growth retardation. Machine learning methods open up new opportunities for modelling crack growth based on empirical data. Machine learning algorithms reveal dependencies between experimental parameters and crack behaviour. Ensemble methods, which combine the predictions of many models into a single powerful prediction, are particularly promising. A study by Zhang et al. (2023) showed that the use of ensemble algorithms, in particular random forest, XGBoost, and LightGBM, allows for highly accurate modelling of the growth rate of short fatigue cracks in K477 nickel alloy under variable temperature and stress conditions. The highest coefficient of determination was achieved using the random forest method, and the results of the machine prediction were in good agreement with the experimental data. In a study by Konda et al. (2022), a series of machine learning models were constructed to predict the growth rate of fatigue cracks in Ti6Al4V alloy manufactured by additive manufacturing. The authors showed that the use of algorithms, in particular random forest and extreme gradient boosting, allows the patterns of fatigue crack growth rate to be reproduced with high accuracy. The work by Ben Seghier et al. (2023) showed that ensemble machine learning models, in particular random forest, AdaBoost, and gradient boosting regression tree, effectively predict the growth rate of cracks in steels and alloys based on a large database of experimental data from the literature. The resulting models showed high accuracy and consistency with real crack growth curves, confirming the feasibility of using ML approaches to analyse fatigue problems in complex engineering structures. A study by Badora et al. (2021) demonstrated that even with a limited sample size of 31 observations, ensemble methods such as random forest, AdaBoost, and extreme gradient boosting can provide reliable predictions of the maximum length of fatigue cracks in gas turbine components. The authors proposed special loss functions adapted to the task of detecting the longest cracks and emphasised the importance of feature selection and reducing uncertainty in data when working with small samples. The work by Gbagba et al. (2024) provided a thorough review of modern machine learning methods used to predict the fatigue life of welded structures. The authors noted that among the various approaches, ensemble methods, in particular random forest algorithms and hybrid models, are most often used to process parameters such as stress intensity, weld thickness, load history, and number of cycles to failure. It should be noted that machine learning methods are successfully applied not only in fracture mechanics and fatigue of materials, but also in a wider range of engineering and applied fields. For example, in a study by Yasniy et al. (2025a), a neural network was proposed for predicting the service life of structural elements made of titanium alloys. The model provided high accuracy in reproducing the dependence of crack length on the number of cycles and the load asymmetry coefficient, and the prediction error on the test data did not exceed 0.4%, confirming the effectiveness of machine learning in long-term material life tasks. Meanwhile, in the work by Yasniy et al. (2025b), artificial intelligence was used to classify epoxy composites for aviation applications, which made it possible to increase the accuracy and speed of material identification. In studies by Tymoshchuk et al. (2024) and Yasniy et al. (2025c), ML methods were effectively applied to analyse the behaviour of shape memory alloys under cyclic loading, in particular for modelling hysteresis curves and classifying load frequencies. Of particular note is the study by Acar et al. (2015), which experimentally investigated the superelastic response of shape memory alloys as a function of temperature and loading rate, highlighting the complexity of the behaviour of such materials and the need for accurate modelling. In addition, the work by Tymoshchuk et al. (2025) demonstrates the capabilities of machine learning in composite structure classification tasks, confirming the versatility and flexibility of these methods for a wide range of tasks in materials science and mechanics. Furthermore, a series of studies by Stukhliak et al. (2024a; 2024b; 2025) demonstrated the effectiveness of neural networks for modelling the mechanical, tribotechnical, and anti-friction characteristics of epoxy composites. A significant number of studies also demonstrate the effectiveness of ML algorithms in the field of environmental monitoring. For example, the works by Stanko et al. (2025) and Pawul et al. (2016) consider the use of artificial intelligence to predict air pollution levels, in particular CO concentrations, based on the UV index and general atmospheric parameters. In the field of energy, Ayaz Atalan et al. (2025) applied ML to predict wind energy depending on changes in the environment, and Forootan et al. (2022) conducted a thorough review of the use of machine and deep learning in complex energy systems. Research on intelligent transport deserves special attention, where, as shown in the work by Li et al. (2021), machine learning allows for effective real-time prediction of traffic flows. In the field of cybersecurity, Park et al. (2025) demonstrated the effectiveness of unsupervised methods for detecting anomalies in network traffic. In financial analytics, Ahmed et al. (2022) summarised current approaches to the application of AI and ML in finance in a bibliometric review. In biomedicine, Urbina Fredes et al. (2025) applied a hybrid deep neural network to analyse sleep phases, while Hassan et al. (2025) considered the prospects for using artificial intelligence in the diagnosis and treatment of cancer in paediatrics. All these examples highlight the growing role of machine learning in science and technology, opening up prospects for interdisciplinary analysis and data-driven decision-making. The aim of this work is to build machine learning models based on ensemble methods, in particular, random forest and boosted trees for predicting the growth of fatigue cracks in QSTE340TM steel under cyclic loading and single overload. The proposed models estimate the crack length depending on the number of cycles and load parameters. The work compares the accuracy of the models, analyses their generalisation ability, and interprets the results obtained.