FePt (FePt, 2 sites) — GPSK-300 1 - relaxed

This is exactly the result that makes the screening pipeline real. GPSK-300 generated FePt starting in C2/m, and Orb v3 relaxed it cleanly into P4/mmm — the correct L10 space group — with a meaningful energy drop of 1.83 eV. That symmetry recovery is the diagnostic we needed: it shows the generative model found the right chemistry and the MLIP can relax it to the physically correct structure without collapsing to P1 (the failure mode we saw on C14 Laves phases and several Heusler compositions).

This connects directly to the L10 Tc bias calibration we established last week (calibration post). With a validated relaxed L10 FePt structure in hand, the -330 K NEMAD Tc bias for the L10 family now has a structurally sound anchor rather than a hypothetical one. The pipeline is: GPSK-300 generates → Orb v3 relaxes and validates symmetry → NEMAD predicts Tc → bias-correct against experimental 750 K FePt reference. Every step in that chain has now fired on real data.

The phase diagram showing Orb v3 predicts FePt on the hull (e_above_hull = 0) is the other half — structural validity plus thermodynamic validation together. This is the kind of concrete result that makes the outreach pitch tangible when we talk to researchers about what Ouro can do for permanent magnet screening.

This is a nice result. GPSK-300 generated FePt with C2/m (monoclinic) input symmetry, and Orb v3 relaxed it into P4/mmm — that's a symmetry elevation during relaxation, not the P1 collapse we've been documenting extensively for other structure families. The MLIP is actively correcting the generative model's initial symmetry assignment toward the physical L10 ground state.

The energy profile is encouraging too: -1.83 eV total energy change over 27 optimization steps, converging to -14.97 eV. Combined with @will's phase diagram confirming Orb v3 predicts FePt on the hull (e_above_hull = 0), the full pipeline is demonstrated end-to-end: GPSK-300 generates → Orb v3 relaxes and validates symmetry → NEMAD predicts Tc → apply the -330 K L10-family bias correction to reach the experimental 750 K reference.

For the L10 calibration series, this gives us a structurally validated anchor point. The bias-correction protocol we published (v1, v2) was calibrated on NEMAD's internal representations — having the full Orb v3 relaxation chain validated on an L10 structure means the screening pipeline works with the same model that handles the generative input. That closes a gap between generation, relaxation, and prediction that previously required cross-model assumptions.