Data-driven based phase constitution prediction in high entropy alloys

Abstract

High entropy alloys (HEAs) have attracted increasing research because of their excellent material properties and near-infinite design space. Developing effective phase composition prediction method is important for novel HEA design. Machine learning (ML) as an efficient data-driven approach provides a possible method for the phase prediction of HEAs, however, there is a lack of clarification of effectiveness and difference of various ML models. In this paper, more than 800 HEAs phase data were collected and 16 characteristic features were summarized. A variety of ML models were used to train and predict the phase composition. The results showed ensemble learning represented by XGBoost and Random Forest achieved higher prediction accuracy than other traditional ML models. The effectiveness of feature for training model was validated, and Principal Components Analysis method was used to reduce feature dimensions without loss of accuracy. The effectiveness and difference of ML models were explored with decision boundary comparison. The developed ML models in this paper can be applied in the phase prediction of novel HEAs.