An interpretable and reliable framework for alloy discovery in thermomechanical processing

In thermomechanical controlled processing, both the alloy composition and the processing strategy shape the mechanical properties of metals. In this study, we present a data-driven approach to discover alloys with optimized strength-ductility trade-off in the thin slab direct rolling process. We evaluate seven different supervised machine learning algorithms to predict two mechanical properties, namely Ultimate Tensile Strength and % Elongation. SHapely Additive exPlanations (SHAP) augments interpretability to the best performing models. NSGA-II, an evolutionary genetic algorithm, is employed with ML models as objective functions to obtain the optimal Pareto Front solutions. Further, we incorporate manifold learning and unsupervised clustering to screen the Pareto Front and to select a few unique solutions which can facilitate added analysis for implementation. Furthermore, we introduce the application of conformal predictions for uncertainty quantification, ensuring reliability of the framework. Overall, the proposed approach enables interpretable and reliable property prediction, thus accelerating alloy design in thermomechanical processing. • Conformal Predictions for Reliability of ML Framework. • Alloy Discovery using Multi-objective Optimization. • Data-driven Post-Pareto Analysis to Choose Unique Alloys.