A comparative study of query strategies in Pareto active learning for heat treatment optimization of medium-Mn steel

This study employs a Pareto Active Learning (PAL) framework to optimize multi-step heat treatment conditions for medium-Mn steels, aiming to maximize both ultimate tensile strength (UTS) and total elongation (TE) while minimizing the number of experimental iterations. A systematic comparison of three query strategies—Expected Improvement (EI), Upper Confidence Bound (UCB), and Greedy Search (GS)—revealed that the UCB-based approach produced a superior Pareto front with greater breadth and diversity, as quantified by hypervolume and coverage relation metrics. SHapley Additive exPlanations (SHAP) analysis confirmed the UCB-based model's ability to identify key relationships between processing parameters and mechanical properties. Experimental validation demonstrated high predictive accuracy (93.8% for UTS and 88.5% for TE). Microstructural analysis of optimized specimens revealed structure–property relationships underlying their mechanical performance: high-strength specimens exhibited predominantly martensitic structures, those with enhanced ductility featured ferritic structures, and specimens with balanced strength and ductility displayed optimized multiphase microstructures composed of tempered martensite, ferrite, and retained austenite. The UCB-based PAL framework demonstrated high efficiency in identifying optimal processing conditions with minimal experimental iterations, while achieving high predictive accuracy for both strength and ductility. • PAL provided an efficient framework for multi-objective heat treatment optimization. • UCB-based PAL successfully identified optimal conditions with minimal experiments. • The optimal UCB model showed 93.81% and 88.49% accuracy for UTS and TE, respectively.