Compositional optimization of rare earth permanent magnetic alloy by machine learning

Abstract

Machine learning (ML) is becoming increasingly crucial in the process of discovering and designing new materials. In this paper, we use the quaternary permanent magnetic alloy R(Co_1-x-yFe_xM_y)₅ (0 ≤ x ≤ 0.08, 0.08 ≤ y ≤ 0.16) as an example to show how ML can be used in rare earth materials research. Our density functional theory (DFT) high-throughput screening, guided by the Korringa-Kohn-Rostoker coherent potential approximation (KKR-CPA) method, ranks the synthesis difficulty of M−site substituted components as follows, from least to most challenging: Hf, Zr, Cu, Zn/Ni, Ti, Si, and Cr. The magnetic-property data is then used to train and test active ML models, supplemented by Markov chain Monte Carlo (MCMC) iterations. Our model forecasts that substituting rare earth sites with Gd and Eu, and the Co site with Cr or Ni, can result in magnetic moments on par with or exceeding SmCo₅. We further employ our model to optimize compositions and predict cost-effective, supply-reliable alternatives to SmCo₅, particularly for those with lower Sm and Co content. ML is thus beneficial for compositional optimization, especially when the underlying structure–property relationships are not fully understood.