This study presents a data-driven framework that quantitatively predicts and physically interprets the corrosion behavior of stainless steels using X-ray fluorescence (XRF) compositional data. A deep neural network (DNN), coupled with SHAP (SHapley Additive exPlanations) analysis, provided physically interpretable insights into interactions between alloy composition and processing factors, surpassing basic statistical predictions. Input variables, including exposure time, XRF-measured composition, steel type, and manufacturer, were utilized to train a fully connected DNN model following systematic preprocessing. The framework attained high prediction accuracy (R 2 ≈ 0.97), enabling reliable estimation of corrosion behavior based on pre-exposure composition. SHAP analysis illustrated nonlinear contributions of Si, V, P, Nb, and Ti, consistent with metallurgical principles. Ni and Co displayed minor, yet composition-sensitive effects on the corrosion behavior, in accordance with known metallurgical trends. These results demonstrate that corrosion, previously deemed random and uncontrollable, can be systematically predicted and physically elucidated through data-driven learning, thereby establishing a foundation for AI-guided alloy design and process optimization.
Kim et al. (Sun,) studied this question.