Physics-Informed Machine Learning for Dual-Target Optimization of Hardness and Tensile Strength in Multi-Element Aluminum Alloys

<title>Abstract</title> <p>The simultaneous optimization of Brinell hardness (HB) and ultimate tensile strength (UTS) in multi-element aluminum alloys presents a persistent challenge due to the complex, nonlinear interplay between alloy composition, heat treatment condition, and resulting microstructural state. Existing machine learning (ML) frameworks have primarily targeted single mechanical properties, limiting practical applicability where multiple performance criteria must be jointly satisfied. This study presents a physics-informed machine learning design system (PI-MLDS) for dual-target prediction and optimization of HB and UTS across 1xxx–7xxx series aluminum alloys. A curated dataset of 163 alloy compositions was assembled from the MatWeb database following rule-based data cleaning. Four ML algorithms, k-nearest neighbors (KNN), support vector regression (SVR), gradient boosting regression (GBR), and artificial neural network (ANN), were deployed and benchmarked. An empirical Tabor-type physical constraint relating UTS to HB was embedded as an auxiliary regularization term to enforce dual-output physical consistency. SVR achieved the highest UTS accuracy (R² = 89.3%, RMSE = 48.4 MPa, MAPE = 15.6%), while GBR performed best for HB (R² = 87.8%, RMSE = 13.6 HB, MAPE = 18.0%). Leave-one-out MLDS validation across six reference alloys yielded mean MAPE of 12.7% (HB) and 6.5% (UTS), with UTS prediction more reliably satisfying the stringent accuracy threshold. Multi-objective Pareto optimization identified an optimal 7075-based composition (Cu = 2.20, Mg = 2.90, Zn = 7.10 wt%, Fe eliminated) with predicted HB = 179.6 and UTS = 599.3 MPa, representing improvements of 19.7% and 9.3%, respectively, over the AA 7075-T6 baseline. Pearson correlation analysis identifies Zn, Cu, and Mg as the primary positive correlates of both targets, while Cu dominates global feature importance (GBR score &gt; 0.43). These results establish PI-MLDS as a practical pre-screening tool for accelerated aluminum alloy composition design.</p>