| Abstract Scope |
Machine learning surrogates now predict composition-processing-property relationships accurately; however, they cannot state where their own recommendations should be trusted. We show that even a single, internally consistent experimental dataset, the most favorable case for trusting model uncertainty, yields ensemble intervals that are over-confident, and that locally-weighted conformal calibration restores empirically verified coverage at negligible cost. Applied to cast aluminum alloys, where composition and T5 heat treatment must be chosen together under process-window and cost constraints, calibration converts a cascaded surrogate into a decision tool. Calibrated intervals separate validated recommendations from exploration targets, and reveal that upper-confidence-bound optimization drifts into the region where its own guarantee fails, justifying a mean-based Pareto formulation with interval reporting. We then co-optimize hardness, electrical conductivity, and processing cost, mapping calibrated Pareto fronts across three heat-treatment scenarios, which show that process-window width governs the reachable hardness-conductivity compromises. |