Application and feasibility analysis of knowledge-based machine learning in predicting fatigue performance of stainless steel

To better predict the fatigue-related S-N curves of different series of stainless steels, 570 sets of data including fatigue test results and other performance parameters for five common types of stainless steel materials were initially collected. Eight machine learning models were deployed and analyzed using the dataset, and their predictive performances were evaluated using assessment metrics. Based on the best-performing model, corresponding S-N curves were constructed. The SHapley Additive exPlanations (SHAP) method was then applied to the optimal model to comprehensively describe and analyze the influence mechanisms of various factors on the number of cycles than stainless steel could withstand. Finally, the results predicted by the optimal model were compared with multiple design standards to verify the feasibility and effectiveness of the model in predicting the S-N curves of stainless steel materials. The results show that a genetic algorithm–optimized artificial neural network (GA-ANN) model possesses higher prediction accuracy than other models, with a correlation coefficient $R^2$ of 0.98 and prediction data within a twofold error margin. The feature parameter constructed through feature engineering has the most significant impact on the number of cycles. The fatigue-related S-N curves predicted by the machine learning model can satisfy the requirements of design standards, demonstrating the feasibility of using this model to predict the fatigue property of stainless steel materials.