We present a machine-learning-guided first-principles study of rare-earth-free magnetic materials in the Fe-Co-S ternary system. A crystal-graph convolutional neural network (CGCNN) fine-tuned through three successive generations of iterative retraining, combined with adaptive genetic algorithm (AGA) structure searches, rapidly screens more than 300 000 substitutional structures and down-selects approximately 1400 candidates for density-functional-theory (DFT) validation. Spin-polarized DFT screening followed by phonon calculations identifies two dynamically stable tetragonal candidates, ${\mathrm{Fe}}_{17}{\mathrm{Co}}_4{\mathrm{S}}_4$ ($I4/m$) and ${\mathrm{Fe}}_2{\mathrm{Co}}_9{\mathrm{S}}_2$ ($I\overline{4}m2$), with saturation polarization $J_s\ge 1\mathrm{T}$ and energies approximately 0.10 eV/atom above the convex hull. These compounds emerge as the leading synthesis targets. Notably, ${\mathrm{Fe}}_{17}{\mathrm{Co}}_4{\mathrm{S}}_4$ exhibits $E_{\mathrm{hull}}=0.107$ eV/atom, $J_s=1.88\mathrm{T}$, and magnetocrystalline anisotropy $K_1=0.71 {\mathrm{MJ}/\mathrm{m}}^3$. We also identify eight near-hull phases ($E_{\mathrm{hull}}<0.01$ eV/atom) and independently rediscover the electrochemically prominent spinel ${\mathrm{FeCo}}_2{\mathrm{S}}_4$, providing an internal validation of the search strategy. Compared with our prior Fe-Co-C/Si/P studies, Fe-Co-S exhibits distinct stability and magnetization regimes consistent with the larger sulfur anion. The combination of machine learning with quantum simulations narrows synthesis targets for Fe-rich sulfides with high $J_s$ and moderate $K_1$, which are competitive with hard ferrites.

Close

PrevNext

View figure in article

• Functional materials
• Permanent magnets
• Genetic algorithms
• Machine learning