Vol.26, Special Issue A, 2026, pp. S23–S29
https://doi.org/10.69644/ivk-2026-siA-0023 

Machine learning has become a powerful tool for prediction and discovery by learning patterns from observations, although its reliability depends on appropriate model selection, rigorous evaluation, and awareness of inherent limitations. This study investigates the prediction of S-N behaviour of Ck35 steel (DIN EN 10083) using three machine learning models: Support Vector Machine (SVM), Linear Regression (LR), and Gaussian Process Regression (GPR), implemented in MATLAB^®. A material identifier extracted from sample codes is used to partition the experimental dataset into distinct groups, with available fatigue data in each group alternately divided into training and test sets. To capture the expected S-N relationship, both stress amplitude and number of cycles to failure are transformed to a logarithmic space. Each model is trained separately to predict fatigue life as a function of applied stress, and performance is evaluated using Mean Absolute Percentage Error and Root Mean Square Error metrics. The SVM and GPR models demonstrate better predictive performance compared to classical LR, particularly for nonlinear data groups. Each predicted group and experimental S-N curve are compared to assess accuracy, illustrating the potential of machine learning approaches to reduce the scope of experimental fatigue testing through the application of intelligent algorithms.

Keywords: • S-N curve • machine learning • Support Vector Machine • Gaussian Process Regression