GNoME found it, Kitagawa measured it, our models undershoot it: MnFeCo4Si2

GNoME found it, Kitagawa measured it, and our structure-only models undershoot it. That is the short version of what happened when I ran MnFeCo4Si2 through our verification pipeline.

The paper is Experimental investigation of magnetic properties of MnFeCo4Si2 discovered by GNoME by Shupei Naganuma and Jiro Kitagawa at the Fukuoka Institute of Technology. They picked a layered rhombohedral compound out of the GNoME catalog (Materials Project mp-3203253), synthesized it, and confirmed the predicted structure: R-3m, refined to a = 3.9978 Å and c = 19.583 Å against the GNoME values of 3.95 and 19.36 Å. The magnetism: a soft ferromagnet with a Curie temperature of 1039 K and a saturation moment of 11.63 μB per formula unit (161.35 emu/g at 50 K), with a coercive field of just 3.7 Oe.

I rebuilt the CIF from their Rietveld refinement (here), validated it, and ran it through our routes. What follows is observation first, interpretation after.

What the routes produced

The structure sanity card passes cleanly: R-3m holds at every tolerance from 0.01 to 1.00 Å, all 24 sites ordered, minimum pair distance 2.42 Å, density 7.40 g/cm3. The paper CIF is a solid artifact.

CHGNet, aligned ferromagnetically, gives per-site moments of 2.47 μB on Mn, 2.85 μB on Fe, 0.58 and 0.75 μB on the two Co sites, and essentially nothing on Si. Summed, that is 8.06 μB per formula unit, or 111.7 emu/g, or 1.04 T. The Curie-temperature regressor (CHGNet structural features into CatBoost) returns 577 K.

Observation versus interpretation

Observed: both fast models get the character right and the numbers wrong. The moment distribution matches the paper's own KKR picture, parallel spins with Mn and Fe carrying most of the moment and Co contributing weakly. But CHGNet undershoots the measured Ms by about 31 percent, and the Curie model undershoots 1039 K by roughly 460 K, far outside a generous plus-or-minus 200 K band. The formal Gate 0 receipt records a "disagree" verdict on the Tc claim (run).

My interpretation: for this layered, Co-rich intermetallic, structure-only models act as directional filters, not property predictors. They correctly said "ferromagnet with Mn and Fe as the moment carriers" and they badly missed both magnitudes. Notably, the measured 11.63 μB/f.u. exceeds even the fully FM-aligned sum of CHGNet's local moments, which means the real Co sites carry more moment than CHGNet assigns them. And the Curie miss runs the same direction as our documented family-level biases, but much larger than any correction envelope we have published. This is the first measured Tc above 1000 K in our calibration set for this family of layered intermetallics, which makes it genuinely valuable: it is exactly the kind of anchor that tells us where the regressor breaks.

What would falsify the "Co carries more moment" reading: a site-resolved measurement (neutron or XMCD) showing Co moments near 0.6-0.8 μB would push the excess moment onto Mn and Fe instead, and CHGNet's site assignments would then be right at the element level while still missing the total.

Why this matters for the community

We verified a GNoME-discovered magnet before anyone cited its properties as settled, and the honest answer is that our fast screens would not have predicted how good this compound is. Experimental ground truth like Kitagawa's measurement is the scarcest resource in AI-driven materials discovery. One measured point like this is worth more than a thousand model predictions of the same number. If you have measured magnetic data on computed-first candidates, the measured-data call

The CIF, the moment run, and the Gate 0 receipt are all public. Corrections welcome, especially from anyone who has site-resolved moments on this system.