Prediction of hydrogen-related low-cycle fatigue life of stainless steel based on physics-guided and interpretable machine learning

With the advancement of hydrogen energy infrastructure and natural gas hydrogen blending technology, stainless steel, as the key lining material of high-pressure valves and hydrogen storage containers, is prone to low-cycle fatigue (LCF) failure under the coupling of alternating load and high-pressure hydrogen environment, which poses a serious threat to the pipeline network structure. In this study, a physics-guided interpretable machine learning framework is proposed. Two models, Gaussian process regression (GPR) and extreme gradient boosting (XGBoost), are used to replace discrete labels with continuous characterization parameters, and log-linearization feature reconstruction is performed based on the Manson-Coffin law. Logarithmic strain amplitude, hydrogen charging environment, nickel content, initial section shrinkage, and initial tensile strength are selected as input features to predict the LCF life of stainless steel under the coupling of high-pressure hydrogen environment and alternating load. The results show that GPR achieves better predictive performance and greater robustness than the XGBoost model, and effectively captures the internal mechanism of the fatigue life attenuation process, thereby reducing MAPE to about 12.5%. Further SHAP analysis shows that the logarithmic total strain amplitude (40.8%) and nickel content (32.1%) are the key factors driving fatigue life evolution, revealing the influence mechanisms of each feature on fatigue degradation at the physical level. This study partially integrates data-driven methods and physical mechanisms and provides a robust prediction method for the performance evaluation of hydrogen-related key structures.