Machine learning reveals the importance of the formation enthalpy and atom-size difference in forming phases of high entropy alloys

Abstract

Despite outstanding and unique properties, the structure-property relationship of high entropy alloys (HEAs) is not well established. The machine learning (ML) is used to scrutinize the effect of nine physical quantities on four phases. The nine parameters include formation enthalpies determined by the extended Miedema theory, and mixing entropy. They are highly related to the phase formation, common ML methods cannot distinguish accurately. In this paper, feature selection and feature variable transformation based on Kernel Principal Component Analysis (KPCA) are proposed, the feature variables are optimized, the distinction of phases is carried out by Support vector machine (SVM) model. The results indicate that elastic energy and atom-size difference contribute significantly in the formation of different phases. The accuracy of testing set predicted by SVM based on four feature variables and KPCA (4V-KPCA) is 0.9743. The F1-scores predicted detailedly by SVM based on 4V-KPCA for the considered alloy phases are 0.9787, 0.9463, 0.9863 and 0.8103, corresponding to solid solution, amorphous, the mixture of solid solution and intermetallic, and intermetallic respectively. The extended Miedema theory provides accurate thermodynamic properties for the design of HEAs, and ML methods (especially SVM combined KPCA) are powerful in the prediction of alloy phases.