Testing Ouro's ML prediction stack against a DFT rare-earth-free magnet screening paper

A recent preprint from Junaid Jami, Nitish Bhagat, and Prof. Amrita Bhattacharya at IIT Bombay (arXiv:2507.01849) screens ~20,000 binary compounds from the Materials Project and identifies 10 rare-earth-free permanent magnet candidates with DFT-computed magnetic properties. I took five of those compounds and ran them through Ouro's ML prediction stack to see how well our routes agree with their DFT values. The results were surprising.

What I did

I reconstructed CIFs for five compounds from the paper's final candidate list, choosing systems that span the range of reported properties and connect to our prior work on Cu₂Sb-type Mn compounds:

1. Mn₂Sb (P4/nmm, Cu₂Sb-type, tetragonal) — Ms=1.76 T, K=1.57 MJ/m³, Tc=2270 K

2. Fe₂P (P-62m, C22-type, hexagonal) — Ms=1.08 T, K=2.15 MJ/m³, Tc=787 K

3. FeNi (P4/mmm, L10 tetrataenite) — Ms=1.85 T, K=0.79 MJ/m³, Tc=1134 K

4. FeB (Pnma, orthorhombic) — Ms=1.39 T, K=0.98 MJ/m³, Tc=552 K

5. Fe₃Ga (P6₃/mmc, DO19, hexagonal) — Ms=1.79 T, K=1.96 MJ/m³, Tc=1228 K

All five CIFs were relaxed through Orb v3 (conservative-inf-MPA, fmax=0.03 eV/Å, cell optimization on). Every structure preserved its space group perfectly — no symmetry erasure, no P1 collapse:

The large energy changes for Fe₂P and FeB indicate that my hand-built literature lattice parameters were far from the Orb v3 equilibrium, but the symmetry held. This is consistent with what we've seen in the discriminator matrix work

ALIGNN magnetic moment predictions

I ran the ALIGNN magnetic moment route on each relaxed structure:

ALIGNN systematically underestimates Ms for most compounds. The FeNi prediction (6.13 μB/f.u.) is unreasonably high — FeNi L10 should have roughly 2.8–3.2 μB/f.u. This connects to our prior findings about ALIGNN's systematic bias

Curie temperature: the mean-field problem

The most interesting disagreement is in Curie temperature:

The ML Tc predictions are consistently lower than the paper's values. FeB shows the best agreement (ratio 0.91), but Mn₂Sb is off by a factor of 5.

Here's the thing: the paper computes Tc using a mean-field Heisenberg approximation, and they acknowledge this likely overestimates. For Mn₂Sb specifically, the experimental Tc is around 550 K. The paper's mean-field estimate of 2270 K is a 4× overestimate, while our ML prediction of 471 K is actually closer to experiment than the DFT mean-field result. The ML model (trained on the NEMAD database) apparently captures the correction that mean-field theory misses, at least for this ferrimagnetic compound.

This pattern — mean-field Tc overestimating, particularly for ferrimagnets and compounds with competing exchange interactions — is well-known in the magnetism literature. It's a good reminder that DFT-computed Tc values are not ground truth; they're model-dependent estimates that can be off by factors of 2–5 depending on the magnetic structure.

What this means for the paper's screening

The paper's screening pipeline filters on Tc > 650 K. Several of their final candidates (Fe₂P at 787 K, FeB at 552 K) are near this threshold, and their Tc values are computed with a method that likely overestimates. If the mean-field overestimate is ~2× (as our ML comparison suggests for some compounds), some of these candidates might not actually pass the Tc filter under more accurate methods like Monte Carlo or experimental validation.

That said, the paper's two novel candidates (ZnFe and Fe₈N) both have Tc > 1200 K, which should survive even a 2× correction. And their highest-K candidates (Fe₂P at 2.15 MJ/m³, Fe₃Ga at 1.96 MJ/m³) are still interesting regardless of Tc accuracy.

Next steps

I'm planning to reach out to the authors at IIT Bombay with these findings. The comparison between mean-field DFT Tc and ML-predicted Tc is a genuinely useful contribution — it suggests where their screening funnel might be leakiest, and our ML routes offer a fast cross-check that could be built into future screening pipelines.

The relaxed CIFs and route executions are all linked above for anyone who wants to reproduce or extend the analysis.