Multi-objective optimization design of Al-Si alloys based on machine learning

This study proposes a systematic inverse-design methodology for aluminum alloys, integrating machine learning (ML) with multi-objective optimization. Based on an industrial dataset comprising 3790 alloy records, a database was constructed, incorporating the mass fractions of 14 elements along with 160 weighted atomic descriptors. Forward feature selection identified optimal descriptor subsets-19 features for Brinell hardness(HB) and 14 for electrical conductivity(EC). A comparative assessment of several regression algorithms identified Extreme Gradient Boosting (XGBoost) as the most accurate predictor. The optimized XGBoost models were coupled with the expected-improvement (EI) criterion to construct a multi-objective expected-improvement (MOEI) function, which was subsequently maximized using particle swarm optimization (PSO). This iterative procedure converged on an as-cast alloy composition of Al-5.86Si-1.93Cu-0.56Mn-0.65Mg-0.28Cr-1.67Ni-1.36Zn-0.10Ti-0.92Fe-0.049Sr (wt%), striking an optimal balance between HB and EC. CALPHAD-based thermodynamic calculations and microstructural validation confirmed that the alloy achieves 96.1 HB and 24.4 % IACS. Experimental measurements deviated by less than 3.5 HB and 2.2 % IACS from the predictions, demonstrating that the inverse design workflow can reproduce target properties within experimental uncertainty.