AI and Quantum Approaches for an Informed Search of Rare-Earth-Free Magnets: Fe₃CoB₂ Predicted and Realized

Rare-earth permanent magnets underpin many modern technologies, but their reliance on critical elements (neodymium, dysprosium) raises supply and cost concerns. We present a data-driven framework to accelerate discovery of rare-earth-free permanent magnets by combining machine learning and first-principles simulations. In our approach, a Crystal Graph Convolutional Neural Network (CGCNN) model rapidly predicts formation energies across broad chemical spaces for efficient candidate screening, while an adaptive genetic algorithm explores favorable crystal structures. The most promising candidates are subsequently evaluated via density functional theory (DFT) to determine their thermodynamic stability and intrinsic magnetic properties, including saturation magnetization, magnetocrystalline anisotropy, and Curie temperature. We demonstrate this strategy on ternary Fe–Co–X systems (X = B, C, Si, P, or S). Fe–Co alloys provide high magnetization but often crystallize in high-symmetry phases with insufficient anisotropy for hard magnet applications. Incorporating a main-group element X breaks this symmetry, a well-known approach to enhance anisotropy. Using our CGCNN-guided search and DFT validation, we identified several promising candidates. Among them, Fe₃CoB₂ emerged as a particularly promising compound with strong magnetocrystalline anisotropy approaching that of Nd₂Fe₁₄B. This prediction was later confirmed by experimental synthesis. Other compounds were identified as potential rare-earth-free magnets, though none are predicted to rival Fe₃CoB₂. Our results illustrate how an AI-guided exploration, grounded in known physical principles and validated by quantum simulations, can accelerate the development of sustainable permanent magnets.