Abstract
High-nitrogen austenitic stainless steels (HNASS) require compositional strategies that simultaneously maximize corrosion resistance and microstructural stability while suppressing delta (δ) ferrite and deleterious precipitates. Here, an explainable multi-objective design workflow is developed that couples thermodynamic descriptors from the Calculation of Phase Diagrams (CALPHAD) approach-using both equilibrium and Scheil solidification calculations-with machine learning surrogate models, random forest (RF) and Extreme Gradient Boosting (XGBoost), trained on 60,480 compositions in the Fe-C-N-Cr-Mn-Mo-Ni-Si space. The physics-informed feature set comprises phase fractions; transformation and precipitation temperatures for δ-ferrite, chromium nitride (Cr(2)N), sigma (σ) phase and M(23)C(6) carbides; liquidus and solidus temperatures; and the pitting-resistance equivalent number (PREN). The RF model achieves consistently low prediction errors, with a PREN root-mean-square error (RMSE) of ≈0.004, and exhibits strong generalization. Shapley additive explanations (SHAP) reveal metallurgically consistent trends: increasing nitrogen (N) suppresses δ-ferrite and promotes Cr(2)N; carbon (C) promotes M(23)C(6); molybdenum (Mo) promotes the σ-phase; and C and silicon (Si) widen the freezing range. Using the trained surrogate as the objective evaluator, the non-dominated sorting genetic algorithm III (NSGA-III) builds Pareto fronts that minimize the δ-ferrite range, Cr(2)N, σ-phase, M(23)C(6) and the freezing range (ΔT) while maximizing PREN. The Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) is then applied to rank the Pareto-optimal candidates and to select compositions that combine elevated PREN with controlled precipitation windows. This workflow is efficient, reproducible and interpretable and provides actionable composition candidates together with a transferable methodology for data-driven stainless steel design.