Machine learning in computational magnetic materials discovery

Abstract

Computational design of a new compound is a big challenge in condensed-matter physics overyears. High-throughput computational screening is recently attracting interest as a possibletool for solving this issue. The basic idea is the following. (1) We first collect structure data ofexisting compounds from database. (2) Hypothetical compounds are generated by elementsubstitution. (3) Physical properties of these compounds are evaluated by first-principlescalculation. If good properties are obtained, the compound is recommended as a candidatecompound. A bottleneck of this method is that the step (3) is computationally demanding. Inthis talk, I will discuss how machine learning can be utilized to accelerate the computationalscreening by taking transition-metal compounds as an example. We propose orbital field matrix(OFM) [1] as a descriptor to represent a compound (Fig.1). Application to thousands oftransition-metal compounds reveals that the kernel ridge regression using OFM reproducesboth the formation energy and magnetic moments in reasonable accuracy. Virtual screening ofNd-Fe-B systems using this technique will be presented. We also discuss how machinelearning can be used to extract controlling parameters of magnetic properties. We analyzeexperimental Curie temperature of 108 rare-earth transition-metal bimetals using kernel ridgeregression. Full-search analysis reveals that the rare-earth concentration is the most relevantparameter (good descriptor) [2].
[1] T.L. Pham et al., STAM 18, 756 (2017); JCP 148, 204106 (2018).
[2] H.C. Dam et al., arXiv:1705.00978.