Machine Learning Prediction of Microhardness in TIG-Welded Al6061 Using Synthetic Data Augmentation and Bayesian Optimization

Predicting microhardness in TIG-welded Al6061 aluminum alloy is challenging due to limited experimental datasets and non-linear precipitation-hardening behavior across weld zones. A total of 144 Vickers microhardness measurements (HV0.5, 15 s dwell time) were collected from GTAW-welded Al6061-T6 plates (autogenous, 200 A AC) across 16 aging conditions (five temperatures, five durations, plus as-welded reference), sampled at nine positions spanning the base metal, heat-affected zone, and fusion zone. To address data scarcity, a polynomial synthetic augmentation strategy–calibrated to the experimental data and encoding aging temperature–time–welding-condition interactions–expanded the training set to 1,144 points. A Bayesian-optimized Gradient Boosting model trained on this augmented dataset achieved MAE = 5.23 HV and R 2 = 0.82, a 47% reduction in error relative to an experimental-only baseline (MAE = 9.95 HV, R 2 = 0.45). A Generative Adversarial Network (GAN) augmentation baseline was also evaluated but suffered mode collapse in this scarce-data regime (MAE = 22.49 HV, R 2 = − 0.49). Partial dependence analysis confirmed that model predictions align with known Al6061 precipitation kinetics, including the over-aging hardness drop above 185 ◦ C and a consistent +3.8 HV hardness recovery following post-weld heat treatment. The proposed polynomial augmentation framework is computationally inexpensive, physically interpretable, and transferable to other alloy systems where experimental data collection is costly or constrained.