A General Data-Driven Framework for Materials Discovery: Application to Thermodynamically Favored Magnetic Intermetallics

Relaxing large pools of hypothetical crystal structures with density functional theory (DFT) is the chief bottleneck in high-throughput materials discovery. We present a simple, modular pretrain-screen-DFT-refine framework that couples a CEAL-backbone ChemGNN pretrained on the Materials Project to iterative, targeted DFT feedback. Applied to a 1.25 M hypothetical Ce–Co–Cu pool generated by element permutation and lattice scaling, the loop enriches low-formation-energy candidates across generations: precision at −0.2 eV/atom improves from 31.2% → 74.3% → 86.2%, while DFT-verified medians shift 0.016 → −0.331 → −0.386 eV/atom, cutting the relaxations needed to recover 1,000 stable compounds by ~58–64%. We quantify stability relative to the reference Ce–Co–Cu convex hull via ΔE_ref, mapping model-identified compositions onto the ternary triangle and showing candidates lie on/near tie-planes in the Co-rich, Ce-lean region, highlighting rare-earth-lean opportunities likely accessible with modest finite-T stabilization. Spin-polarized DFT+SOC on selected candidates reveals substantial magnetization at ≥67% Co, including Ce₂Co₈Cu₂ variants with Js ≈ 0.65–0.74 T; 300 K ab initio molecular dynamics exhibit rapid RMSD plateaus, supporting ambient robustness. Because the framework is model- and target-agnostic, it readily extends to multi-objective discovery (e.g., adding magnetic proxies, MAE, or T_C) and to other chemistries under constrained DFT budgets.