Data-driven analysis of hydrogen embrittlement in martensitic steels with interpretable machine learning

A data-driven analytical framework integrating machine learning (ML), SHAP (SHapley Additive exPlanations) analysis, and symbolic regression (SR) was developed and applied to quantitatively analyze and interpret hydrogen embrittlement behavior in martensitic steels. Slow strain rate tests (SSRT) were conducted on JIS-SCM440 steels with systematically varied diffusible hydrogen contents, phosphorus contents, and tempering conditions to construct an experimental database. The ML model accurately predicted the notch tensile strength, and SHAP analysis quantitatively evaluated the contributions and interactions of the key factors. The SHAP results were then simplified to extract dominant trends, which were interpreted from a metallurgical perspective to elucidate the characteristic dependencies of hydrogen embrittlement on hydrogen content, phosphorus content, and tempering conditions. Based on these insights, SR was applied to formulate explicit and interpretable equations representing these relationships. The resulting model not only reproduces the characteristic physical behavior associated with hydrogen embrittlement but also establishes a quantitative framework linking data-driven analysis and metallurgical understanding, providing a rational basis for the design and safety assessment of high-strength martensitic steels used in hydrogen infrastructure applications.