Interpretable Machine Learning for Predicting the Fatigue Strength of Steel: Influence of Composition and Processing Parameters

Abstract

The prediction of fatigue strength in steel is critical for the design and analysis of structural components, given the high costs and time associated with fatigue testing and the severe consequences of fatigue failures. This study explores the application of the CatBoost machine learning algorithm to predict the fatigue strength of steel based on chemical composition, processing parameters, and mechanical properties. The model's interpretability is enhanced by integrating Shapley additive explanations, providing insights into the contributions of key features such as tempering temperature (TT), chromium content (Cr), and molybdenum content (Mo). The proposed framework achieves high predictive accuracy, with an R² of 0.952 and an RMSE of 31.625. This study fosters trust and utility in safety-critical applications by addressing the limitations of black-box models. The results underscore the potential of interpretable machine learning in advancing fatigue strength prediction methodologies and informing structural and construction engineering practices, while Enhancing the SHAP-based feature importance analysis to refine the selection of key predictors, potentially simplifying the model while maintaining accuracy.