Machine-Learning Screening of Heat-Treatment Effects on the Mechanical Properties of PBF-LB/M AlSi10Mg: Synthetic-Data Fidelity and Validation Limits

Qualification of PBF-LB/M components remains experimentally intensive, motivating data-efficient prediction of mechanical performance. This study evaluates nine machine-learning models, from Decision Tree to Transformer architectures, for predicting yield strength (YS), ultimate tensile strength (UTS), elongation (TE), and hardness (HV) of AlSi10Mg from a previously published, single-source 40-configuration dataset spanning five heat treatments, four aging durations, and two build orientations, with and without training-set-only SMOGN, CTGAN, and TVAE augmentation. On the primary stratified held-out split, Decision Tree achieved R2 = 0.97 (YS), 0.98 (UTS), and 0.99 (HV), and SVM reached 0.97 for TE. Fifty repeated stratified splits with nested hyperparameter tuning gave lower median R2 values of 0.95, 0.97, 0.93, and 0.95, respectively, and XGBoost, Gaussian Process Regression, and Decision Tree differed by less than the split-to-split variability. Augmentation did not significantly improve tree-based models (median ΔR2 within ±0.02), and a +0.50 elongation gain observed on one split was not reproducible; SMOGN improved SVM (median ΔR2 up to +0.09, p < 0.01), whereas CTGAN and TVAE often degraded accuracy and produced fewer than 2% of rows consistent with the heat-treatment routes. Group-aware validation retained R2 = 0.84–0.91 for unseen aging levels but gave R2 < 0 when an entire heat treatment was withheld, and jackknife+ prediction intervals under-covered such conditions. Model errors exceeded the specimen scatter 1.5–4.4-fold, physically constrained augmentation did not benefit tree-based models, and out-of-sample screening identified target-meeting aging times and orientations within, but not beyond, the tested routes. ML can therefore support data-efficient screening within the sampled processing space, but not qualification of untested processing routes. Independent experimental validation is required before process-parameter selection.