Transfer learning-assisted multi-objective optimization of mechanical properties for particle reinforced aluminum matrix composites

Abstract Particle reinforced aluminum matrix composites (PAMCs) exhibit high specific strength and processability, demonstrating promising potential for lightweight high-strength applications in advanced structural components. However, achieving multi-objective optimization of mechanical properties in PAMCs remains challenging due to the complexities of compositions and processing parameters. Given the relatively small size of the curated PAMCs dataset (192 entries) sourced from peer-reviewed literature, we proposed a hybrid machine learning workflow named the mechanical property prediction of PAMCs (PAMCs-MP) to predict mechanical properties of PAMCs by integrating transfer learning with transformer-based neural networks. This approach leveraged an Al alloy dataset comprising 1089 entries to overcome data limitations, effectively pre-train feature extractors for predicting matrix-dependent mechanical properties in PAMCs. Comparative evaluation against conventional machine learning models revealed the superior predictive accuracy of PAMCs-MP, achieving coefficients of determination of 92.4 ± 3.7% for ultimate tensile strength and 90.8 ± 4.4% for elongation. Perturbation analysis indicates electronic interactions among Si, Mg and modification elements (Ce, B), as well as particle-driven dislocation strengthening are key determinants of PAMCs’ mechanical properties. The established hybrid workflow provides an effective strategy for performance optimization of complex material systems with limited datasets, offering valuable insights for transfer learning application in material design.