Machine learning investigation of chemical disorder effects in 24-valence-electron full-Heusler alloys with single magnetic element

Most full-Heusler alloys exhibit disordered structures such as B2 or BCC, rather than the ideal ordered ${\mathrm{L2}}_1$ or XA structures. It is urgent to explore the intrinsic relationship between magnetism and chemical disorder in these materials. In this study, we combine first-principles calculations and machine learning methods to investigate the influence of chemical disorder on the magnetism of 24-valence-electron X${}_2$YZ-type alloys. By screening stable compounds containing magnetic elements, we constructed a multidimensional feature set. The gradient boosting machine and random forest were used to establish classification and regression models, respectively. The results indicate that whether X or Y is a magnetic element, the degree of disorder is a key factor determining their magnetic properties. Using the proposed scheme, we successfully predict the magnetic properties of three unstudied compounds—Fe${}_2$HfSn, Fe${}_2$YSb, and Co${}_2$LiAs—achieving an $R^2$ value as high as 0.97, compared to $ab$ $initio$ calculations.