Machine-Learning Discovery of Rare-Earth-Free Magnetic Materials in the Fe–Co–S/C Systems

We present a machine-learning–guided workflow for discovering rare-earth-free magnetic materials relevant to wind turbines, electric vehicles, hard drives, and defense. A crystal-graph convolutional neural network trained on formation energy rapidly screens >300,000 substitutional structures and down-selects 1,418 Fe–Co–S candidates for first-principles validation. Density-functional calculations identify three compounds with saturation magnetization J_s > 1.0 T and energies above hull ~0.1 eV/atom. Notably, Fe₁₇Co₄S₄ shows E_hull = 0.12 eV/atom, J_s = 1.9 T, and anisotropy K₁ = 0.64 MJ·m⁻³, indicating potential for permanent-magnet applications. We also find twelve low-energy metastable phases (E_hull< 0.01 eV/atom) and rediscover the electrochemically popular spinel FeCo₂S₄, providing an internal validation. Compared with prior Fe–Co–C studies, these results highlight how ML-accelerated screening coupled with targeted DFT efficiently surfaces high-magnetization, near-stable sulfides and focuses experimental efforts on the most promising compositions [1].

[1] W. Xia, M. Sakurai, T. Liao, et al. APL Mach. Learn. 2, 046103 (2024)