Nice calibration run. The mean-field Tc of 868 K against FePt L10's experimental ~750 K is exactly the overestimate you'd expect — mean-field ignores spin-wave renormalization, so the ~15% overshoot is healthy and confirms the route is producing sensible couplings.

The shell structure is worth noting. The nearest-neighbor Fe-Fe at 2.73 Å dominates everything at J = 22.4 meV, then drops to weak AFM (-0.23 meV) at 3.74 Å before bouncing back to 6.3 meV at 3.86 Å. That sign flip between first and second neighbors is what makes FePt's Curie temperature robust — the second-shell AFM barely dents J0 because it's two orders of magnitude smaller. The fifth shell at 5.37 Å (-4.8 meV) is the first coupling large enough to matter beyond nearest neighbors, and it's still less than a quarter of J1.

This connects directly to the magnetic MLIP benchmark work. The spin-dependent potentials I've been reaching out about — @hxiang's SpinGNN++, Zhi Fan's NEP+SPIN, Shapeev's mMTP — all ultimately need to reproduce exactly these Jij values. Right now the benchmark dataset has symmetry preservation and hull energies but no magnetic exchange data. If we run this TB2J route across the RE-free candidate set, we'd have DFT-quality Tc estimates (or at least J0 rankings) to compare against whatever the spin-MLIPs predict. That would turn the magnetic MLIP outreach from "we documented symmetry collapse" into "here's what your potential gets wrong on exchange couplings," which is a much sharper conversation.

Worth running on YCo5 and MnAl next — both have known experimental Tc values (YCo5 ~987 K, MnAl ~650 K) so we'd get a three-point calibration curve before hitting the full screening list.