The composition optimization of Al-Si alloy for the optimal strength-ductility combination is traditionally made by the trail-and-error experiments, which is costly and unpredictable. In this work, we proposed a multimodal adaptive learning framework for clearer guidance to design high-performance non-heat treatable die casting (NHT-HPDC) Al-Si alloys. Two feature approaches, namely the alloy element and alloy factor approaches, were adopted to train the machine learning (ML) model in this framework. Firstly, based on the ML model trained by alloy element approach, the distributed SHapley Additive exPlanations analysis was newly proposed to understand the role of alloying elements, offering valuable guidance for the selection of feature range. Secondly, the alloy factor approach with three-steps feature engineering was beneficial for training ML model with the higher prediction accuracy. By integrating above two procedures, a multi-objective optimization process was conducted to search for the composition with the superior comprehensive properties. In our study, the optimized NHT-HPDC Al-Si alloy exhibited an exceptional combination of yield strength (YS=140.3±4.2 MPa) and elongation (EL=9.6±0.4%). The underlying linkage between the alloy microstructures and critical features in the ML model is constructed. Generally, our proposed framework is significantly effective in the composition optimization for superior mechanical properties, and can be applied to a wide range of Al alloys.
Zhang et al. (Tue,) studied this question.