Accelerated design of Al−Zn−Mg−Cu alloys via machine learning

Abstract

A machine learning-based alloy rapid design system (ARDS) was proposed to customize the preparation strategies for the desired properties or predict the alloy properties following the preparation strategies. For achieving this, three regression algorithms: linear regression (LR), support vector regression (SVR), and back propagation neural network (BPNN), were employed separately to train the multi-property prediction model, in which the machine learning (ML) model built using SVR was proved to be the best. Then, inspired by the generative adversarial network (GAN) algorithm, the ARDS was constructed. The predictive reliability of ARDS was examined, and for the accurate prediction of the preparation strategies, the upper limits of ultimate tensile strength (UTS), yield strength (YS), and elongation (EL) are about 790 MPa, 730 MPa, and 28%, respectively. Moreover, an ARDS-designed aluminum alloy with superior mechanical properties (764 MPa for UTS, 732 MPa for YS, and 10.1% for EL) was experimentally fabricated, further verifying the reliability of ARDS.