Target-adaptive, physics-informed machine learning for the simultaneous prediction of strength, hardness, and ductility in precipitation-hardening stainless steels

The mechanical properties of precipitation-hardening stainless steels (PHSS) are governed by the coupled effects of alloy chemistry, aging, matrix-phase stability, and atomic-scale thermodynamics, and therefore exhibit nonlinear behavior and pronounced trade-offs among strength, hardness, and ductility. Drawing on published literature and unpublished records from the authors' research group, the datasets contain 310, 282, and 305 records for tensile strength, hardness, and elongation, respectively. Here, we develop a target-adaptive, interpretable machine-learning framework that couples composition and aging variables with descriptors of precipitation potential, phase stability, composition–aging interactions, and atomic, electronic, and thermodynamic characteristics. Independent feature-selection pipelines define property-specific descriptor spaces for tensile strength (Rm), Rockwell hardness (HRC), and elongation (A). Among six regressors, gradient-boosting regression trees (GBRT) achieve the most balanced performance across the three targets using an 80:20 record-random training/test split, with test-set R² values of 0.954, 0.948, and 0.937 and training-set five-fold out-of-fold cross-validation R² values of 0.880, 0.875, and 0.915 for Rm, HRC, and A, respectively. Shapley additive explanations (SHAP) reveal property-dependent, nonlinear attributions consistent with precipitation strengthening, aging-induced recovery, matrix stability, and solid-solution and diffusion-related effects, although these attributions do not by themselves constitute mechanistic evidence. The literature comparison shows predictive agreement for the sampled records within the represented composition–aging domain. The framework thus enables physically interpretable, multi-property prediction from heterogeneous, small-sample PHSS datasets. Given the limited AM coverage, this composition–aging framework may serve as a reference baseline for future AM-specific models that incorporate processing and microstructural variables.