Chunguang Shen¹, Chenchong Wang¹, Xiaolu Wei¹, Wei Xu¹; ¹Northeastern University

With the progression of the Materials Genome Initiative, material design by physical metallurgy (PM) and machine learning (ML) models has received much attention. However, neither PM nor ML models can perfectly deal with a dataset with small samples. Therefore, the combination of two models is a promising solution to overcome this limitation. In this study, datasets of maraging steels with small samples were established based on literature. Then, hardness of maraging steels was predicted and designed by different PM and ML models. For property prediction, the differences of prediction accuracy, microstructure sensitivity and data dependence between two models were systematically compared, and it is believed that combination of PM and ML models is helpful to alleviate small sample and make ML model have physical significance. For alloy design, a PM-guided ML model constructed by combining physical principles and ML is a promising method for producing novel, rational alloy designs.