Compact and interpretable strength–ductility prediction of heat-treatable aluminum alloys using physics-retained explainable multi-target machine learning

Abstract

Predicting yield strength (YS), ultimate tensile strength (UTS), and elongation (El) in heat-treatable aluminum alloys is challenging because composition and heat-treatment parameters interact nonlinearly. This study presents a physics-retained, explainable, and uncertainty-aware multi-target machine-learning framework for simultaneous strength–ductility prediction. A hybrid strategy combining statistical relevance, embedded importance, and metallurgical retention reduced the inputs from 17–12 while preserving solution treatment temperature, aging temperature, and aging time. A matched 17-feature comparator used the same outer-validation splits, inner tuning, model families, search space, and budget. Across 25 outer evaluations, the 12-feature pipeline achieved macro-averaged R2, NRMSE, and NMAE values of 0.807±0.043, 0.424±0.047, and 0.293±0.033, respectively, compared with 0.800±0.047, 0.431±0.052, and 0.294±0.040 for the comparator. An exploratory post-selection comparison produced a mean macro- R2 difference of 0.0068, a corrected 95% interval of [−0.0198,0.0335], and p=0.602. Because k=12 was selected from the same outer-validation summaries, these results are descriptive rather than confirmatory. The selected pipeline therefore used 29.4% fewer descriptors with similar observed mean performance, although superiority or statistical equivalence was not established. On the row-wise held-out final-test partition, the locked 12-feature ExtraTrees model achieved R2 values of 0.857, 0.842, and 0.826 for YS, UTS, and El, respectively. SHAP assigned the largest model attributions to Zn, Cu, and selected heat-treatment conditions, without implying causal metallurgical effects. At a nominal target-wise marginal coverage of 90%, ensemble-conformal coverage was 89.01%, 80.22%, and 91.21% for YS, UTS, and El, respectively; UTS therefore showed under-coverage, whereas YS and El were closer to nominal coverage. The intervals supported uncertainty-aware retrospective prioritization of row-wise held-out conditions. Because external and experimental validation were not performed and performance decreased under composition-group-disjoint validation, the framework should be regarded as an internally evaluated workflow rather than a validated alloy-design tool.