GRADE-AWARE MULTI-TARGET MACHINE LEARNING OF STEEL MECHANICAL PROPERTIES: ENSEMBLE, ANN, AND SELF-ATTENTION TRANSFORMER MODELS UNDER UNSEEN-GRADE SHIFT

Predicting steel mechanical properties from heterogeneous chemistry and heat-treatment metadata is a domain-transfer problem because repeated observations are hierarchically clustered by grade and property reporting is incomplete. This study develops a leakage-resistant multi-target learning framework using 1,360 Steel Bench records representing 562 unique grades, 12 chemistry and processing descriptors, and five responses: tensile strength, yield strength, elongation, reduction of area, and impact energy. Missing targets are not imputed. Numeric imputation, standardization, and categorical encoding g are fitted within each training fold, while neural models use a masked multi-response objective. Twelve conventional regressors are benchmarked against a 256-128-64 artificial neural network and a two-block, four-head self-attention Transformer. Model selection uses three repeats of five-fold grade-grouped cross-validation, followed by a completely grade-disjoint locked test. Extra Trees provides strong cross-target development performance, with grouped-cross-validation R² values of approximately 0.572, 0.462, 0.558, 0.366, and 0.354 for tensile strength, yield strength, elongation, reduction of area, and impact energy, respectively, while Gradient Boosting leads the development ranking for elongation and impact energy. On the locked unseen-grade test, elongation remains transferable (R² = 0.646) and impact energy retains moderate skill (R² = 0.370), whereas tensile strength, yield strength, and reduction of area exhibit negative R². Distribution-shift and descriptor-ablation analyses show that chemistry redistribution, missingness shift, and incomplete process-state descriptors constrain transfer more strongly than nominal model complexity.