Designing magnetism in Fe-based Heusler alloys: a machine learning approach

Abstract

Combining material informatics and high-throughput electronic structure calculations offers the possibility of a rapid characterization of complex magnetic materials. Here we demonstrate that datasets of electronic properties calculated at the ab initio level can be effectively used to identify and understand physical trends in magnetic materials, thus opening new avenues for accelerated materials discovery. Following a data-centric approach, we utilize a database of Heusler alloys calculated at the density functional theory level to identify the ideal ions neighbouring Fe in the $X_2$Fe$Z$ Heusler prototype. The hybridization of Fe with the nearest neighbour $X$ ion is found to cause redistribution of the on-site Fe charge and a net increase of its magnetic moment proportional to the valence of $X$. Thus, late transition metals are ideal Fe neighbours for producing high-moment Fe-based Heusler magnets. At the same time a thermodynamic stability analysis is found to restrict $Z$ to main group elements. Machine learning regressors, trained to predict magnetic moment and volume of Heusler alloys, are used to determine the magnetization for all materials belonging to the proposed prototype. We find that Co$_2$Fe$Z$ alloys, and in particular Co$_2$FeSi, maximize the magnetization, which reaches values up to 1.2T. This is in good agreement with both ab initio and experimental data. Furthermore, we identify the Cu$_2$Fe$Z$ family to be a cost-effective materials class, offering a magnetization of approximately 0.65T.