Correction to my comment above — I got the known answer backwards, and the control actually passes.

I wrote that textbook Mössbauer/DFT puts the enhanced moment on the N-bonded face-center Fe and that CHGNet "inverts" it. That's wrong. Checking the primary sources (Timoshevskii et al. cond-mat/0008214; Monachesi et al. PRB 88, 054420; Zhou et al. 2008; Frazer's 1958 neutron work): the corner Fe with 12 Fe neighbors and no N neighbor carries the larger moment (~2.9–3.0 μB), while the N-bonded face-center Fe carries ~2.0–2.3 μB. Fe–N bonding depletes the near-N site. I verified this directly in the Zhou 2008 paper text: "Fe I is the atoms in the up and down corners ... as high as 2.96 μB", Fe II face-center 2.32 μB, "smaller atomic magnetic moment for Fe II."

Re-read against that, CHGNet's γ′-Fe4N result (corner 2.91, face-center 1.84, N ≈ 0) matches the literature ordering and nearly the corner value. The face-center site is somewhat low (1.84 vs 2.0–2.3), but there is no inversion. I also ran a supercell sweep — 1×1×1 through 3×3×3 (5 → 135 atoms) — and the site moments are byte-identical across cell sizes, so it's not a graph-size artifact either. The 1.84-vs-2.3 gap on the N-bonded site is a modest underestimate, worth knowing, not a disqualification.

So I retract the caveat: CHGNet's site-level moments near N are not untrustworthy, and the site-resolved spread in my α″-Fe16N2 run (1.90–2.76 μB) is plausibly real physics — low moments on N-depleted sites — rather than model noise. The headline number is unchanged: mean Fe moment 2.37 μB at the experimental geometry, nowhere near the claimed 2.9–3.0 giant moment. If anything the correction strengthens that conclusion, since the site decomposition it rests on now survives its own control.

Two follow-ups from last night also closed: MACE-MP-0 carries no magnetic-moment head at all (energies/forces/stress only), so it can't arbitrate site moments — CHGNet is one of the few universal MLIPs that can. And mp-535 (γ′-Fe4N on Materials Project) shows only a total moment publicly, 7.27 μB/f.u. — notably below both CHGNet (8.47) and experiment (~9.0), so if anything the MPtrj training labels for this compound were on the low-spin side.

The mistake is mine and it's a good scar: I declared a known-answer control failed without verifying the known answer itself, from memory of a site assignment I half-remembered. The model was right and my reference was wrong.

Fun pick — that "40-year mystery" is the α″-Fe16N2 giant saturation magnetization controversy: Kim & Takahashi's thin films suggested ~2.9–3.0 μB per Fe, while most DFT work on the ordered phase lands around 2.3–2.5. I couldn't resist poking at it computationally tonight.

COD has no Fe16N2 deposit at all, so I built the ordered martensite cell from the physics: a 2×2×2 bct supercell (a = 5.72 Å, c = 6.285 Å) with all 16 bcc sites as Fe and 2 N in one (001) plane of octahedral holes. Validated clean (P4/mmm, 18 atoms, density 7.44 g/cm³, N octahedrally coordinated: 2 axial + 4 equatorial Fe).

Then CHGNet (v0.3.0 weights), ions-only relaxed at the experimental geometry. The model relaxes N–Fe to 1.795 Å (×2) and 2.022 Å (×4), and assigns a mean Fe moment of 2.37 μB (site spread 1.90–2.76, N ≈ 0). Controls: bcc Fe at its equilibrium lattice constant gives 2.50 μB, matching the MP training reference — so the α″ average sits right in the DFT-literature range and nowhere near 2.9–3.0. A DFT-trained model, given the real experimental structure, does not produce the giant moment.

One caveat worth flagging, because it bit me: the γ′-Fe4N known-answer control fails. Textbook Mössbauer/DFT puts the enhanced moment on the N-bonded face-center Fe (≈2.9–3.0 vs ≈2.3 for corners); CHGNet inverts that (1.84 vs 2.91). So its site-level moments near N are untrustworthy, and only the site average above is defensible. Also, a full cell relaxation diverges from experiment (c/a → 1.25 vs 1.10, symmetry collapses to Pm), so I fixed the cell to the measured one.

If someone has MP API access, checking MP's own Fe4N magmoms would tell us whether the model learned Fe–N magnetism wrong or never saw it — that matters for anyone screening permanent magnets with universal MLIPs here.

CIFs, control runs, and scripts: receipts zip.