Jingli Ren1*

1Zhengzhou University, Henan Academy of Big Data, Zhengzhou 450052, China

EXTENDED ABSTRACT: How to identify multi-principal element alloys (MPEAs) with desired properties from a huge composition space with faster and lower cost remains a grand challenge. Here, we introduce two machine learning approaches to efficiently design alloys with multiple excellent properties. Method 1: Build a high-precision multi-objective machine learning model based on experimental data of strength and ductility in the literature. We find that lower valence electron concentration, higher melting points, and near-zero mixing entropy exert the strongest contributions to the strength­ductility trade-off, and use polynomial fitting to engineer the characteristic contribution trends of the alloys. Method 2: A loop ensemble is formed in the order of molecular dynamics simulation, machine learning, and multi-objective efficient global optimization algorithms. Some surrogate models with R2>99% are established, and the surrogate models are applied in the virtual space, combined with a multi-objective efficient global optimization algorithm to screen excellent performance components. The current work shows that properties fitting can be accomplished effectively by machine learning, which is expected to accelerate the discovery of novel MPEAs with multiple properties at low cost.
Keywords: multi-principal alloy, multi-objective performance, machine learning, alloys design.