Transfer learning-enabled high-precision prediction and multi-objective design of cryogenic aluminum alloys with enhanced strength-ductility synergy

The scarcity of sample data in alloy databases is a major bottleneck for machine learning (ML)-based prediction of alloy mechanical properties, which severely restricts the construction of high-performance ML models. To tackle this problem, this study proposes an instance transfer learning strategy based on TrAdaBoost. Taking the room-temperature dataset as the source domain and the small-sample cryogenic dataset as the target domain, cross-temperature domain knowledge transfer is realized by iteratively and adaptively adjusting sample weights. The results demonstrate that the optimized Tr-XGBoost model achieves a substantial performance improvement. The coefficients of determination (R 2 ) above the independent test set reach 0.91, 0.89 and 0.83 for yield strength (YS), ultimate tensile strength (UTS) and elongation (EL) under cryogenic conditions, respectively. The model maintains stable performance even when the sample proportion is as low as 40%, effectively solving the modeling difficulty induced by insufficient cryogenic small samples. Based on the high-precision Tr-XGBoost model, a multi-objective optimization design framework for cryogenic Al alloys is established using the NSGA-Ⅱ algorithm. Pareto optimal solutions are screened out, and a novel TRAL Al alloy is further designed and fabricated. Adopting the industrial processing route of AA7075, experimental verification shows that the aged TRAL alloy exhibits an UTS of 701 MPa and an EL of 12.32% at 77K. Compared with AA7075, both UTS and EL are increased by more than 20%, realizing the synergistic improvement of strength and ductility.