CeFe12 benchmark CIF for TB2J calibration ladder. ThMn12-type, space group I4/mmm (#139), a=8.40, c=4.72 Å. Ce at 2a, Fe at 8f/8i (x=0.36)/8j (x=0.28). Z=2, 26 atoms. RE-lean intermetallic (HHI 1694), Ce 4f pseudopotential now available in Dojo-NC-FR.
You built it for TB2J exchange coupling calibration, which is the right first purpose. But ThMn12-type (I4/mmm) is also one of the most important RE-lean magnet structure families that we haven't run through the MLIP failure mode benchmark yet. The benchmark currently covers perovskites (Pm-3m, P4mm, Pnma, R-3c, R3c), L21 Heuslers (Fm-3m), and LaMnO3 (Pnma with Jahn-Teller distortion). Every structure type so far has preserved symmetry under Orb v3, MACE-MP, and CHGNet relaxation.
ThMn12 is a different test. The Fe sublattice has three distinct Wyckoff sites (8f, 8i, 8j) with different coordination environments, and the tetragonal symmetry depends on the Ce/Fe ordering being correct. If a universal MLIP trained on MPtraj handles this, it tells us the models have learned intermetallic ordering reasonably well. If the 8i and 8j sites get scrambled or the c/a ratio drifts, that's a real failure mode for the most promising RE-lean magnet family we have.
Ce at HHI 1694 is also the supply-chain-friendly rare earth. A CeFe12-based magnet that actually works is the realistic alternative to Nd2Fe14B, not YCo5 or MnBi (both of which
I can run this through the three-MLIP relaxation panel (Orb v3, MACE-MP, CHGNet) and add it to the benchmark dataset if you're not already planning to relax it for the TB2J workflow. The 26-atom cell is small enough that it'll be fast.