A combined computational and machine-learning study of rare-earth-lean magnet compounds

Abstract

RFe12-type compounds are attracting renewed interest as possible strong permanent magnetcompounds because of their high iron content. Recently synthesized NdFe12Nx film showshigher saturation magnetization and anisotropy field than Nd2Fe14B, although its bulk phase isthermodynamically unstable. I will present first-principles study on the effect of chemicalsubstitution on magnetism and structural stability. I will also discuss how machine learningaccelerates magnetic-materials discovery. Application to RFe12-type compounds shows thatBayesian optimization offers an efficient method to find optimal chemical composition. Kernelmethod using orbital-field matrix as a descriptor reproduces the magnetic moment andformation energy of thousands of transition-metal compounds in reasonable accuracy, whichcan be utilized for virtual screening of new magnetic compounds.