Artificial neural network approach for mechanical properties prediction of as-cast A380 aluminum alloy

Abstract

In this study, the dataset of as-cast A380 aluminum alloy involving chemical composition, processing parameter and mechanical properties was experimentally established. On this basis, the relationship between four input variables of Fe/Mn ratio, Sr content, cooling rate and porosity content, and two outputs of ultimate tensile strength (UTS) and elongation (El) was linked by two artificial neural network (ANN) models for comparison: (1) back propagation - artificial neural network (BP-ANN) model, and (2) back propagation optimized by particle swarm optimization - artificial neural network (PSO-BP-ANN) model. The results showed that the established PSO-BP-ANN model has better reliability and prediction accuracy than BP-ANN model, for lower root mean square errors (RMSE), 6.19 MPa for UTS and 0.08% for El, and higher determination coefficient (R2), 0.9775 for UTS and 0.9798 for El, respectively. Sensitivity level analysis result by the preferable PSO-BP-ANN prediction model showed that the order of relative importance for UTS is Sr content, cooling rate, Fe/Mn ratio, and porosity content while the importance ranking for El is Sr content, Fe/Mn ratio, cooling rate, and porosity content. The research results and dataset in this work lay a foundation for the control of casting performance in industrial practice and for secondary development of mechanical properties prediction module in commercial casting software.