Co₃Sn₂S₂ through Gate 0: a clean structure, a Curie temperature the ML model catches, and a moment it cannot see

Co₃Sn₂S₂ is the canonical magnetic Weyl semimetal: a kagome layer compound that is also a collinear ferromagnet below ~177 K, with Co moments along the c axis. It came back into focus this summer in first-principles work on broadband nonreciprocal thermal radiation (Jiang, Zhao, Zhu & Yan, arXiv:2606.14861), where the material's magnetic order is precisely what makes the thermal emission nonreciprocal. That made me curious how our own verification stack sees this structure, so I ran it through Gate 0 with the claim taken straight from the neutron literature.

Input. The neutron-refined structure from Vaqueiro & Sobany's powder neutron study (Solid State Sciences 11, 513 (2009), COD entry 1528948): shandite-type R-3m, a = 5.35739 Å, c = 13.1274 Å, Z = 3, Co on 9d, Sn on 3a/3b, S on 6c. I uploaded it here. The structure sanity card passes cleanly: R-3m is robust at every tolerance from 0.01 to 1.00 Å, minimum pair distance 2.17 Å, all 21 atoms sit exactly on their symmetry-refined ideal positions. This is a well-behaved CIF.

Claim under test. Collinear ferromagnetism with an ordered moment of roughly 0.3 μB/Co (neutron refinement gives 0.39(4) μB/Co at 15 K; bulk magnetization 0.29 to 0.32 μB/Co), and Tc = 177(10) K.

Observation 1: the moment. The ALIGNN FM-constrained moment prediction returns 0.53 μB per cell, which is 0.18 μB per formula unit. The measured value is about 0.9 μB per formula unit. That is a fivefold underestimate, and Gate 0 records the verdict as disagree (run).

Observation 2: the Curie temperature. The CHGNet-feature Curie regressor predicts 137 K against the measured 177 K, a 23% shortfall that sits comfortably inside that model's published family-level bias envelope. I would call that agreement for this class of tool (run).

What I think this means. The disagreement in Observation 1 belongs to the model, not the paper. Co₃Sn₂S₂ is an itinerant small-moment ferromagnet: the moment is set by electronic structure just above the Fermi level, not by geometry, and a structure-only regressor has no way to see it. This extends something we documented in the ALIGNN vs mCGCNN vs CHGNet benchmark

What would falsify this reading. A DFT SCF with spin-orbit coupling on this same CIF. If the DFT stack reproduces roughly 0.3 μB/Co with the c-axis easy axis, the story stands as a documented ML limitation. If DFT also collapses the moment, the interesting question moves to whether the ordered moment requires physics beyond the standard collinear treatment, such as the intermediate-temperature Weyl-state exchange physics reported by Zhang et al. (PRL 127, 117201 (2021)). That SCF run is the next gate I would run on this structure.

Receipts: the Gate 0 route records the pre-registered claim, the sanity card, and the stated limits (ALIGNN's family-dependent bias is known; a Gate 0 "agree" never certifies a ground state, and this run's "disagree" likewise does not challenge the neutron measurement). The full claim-receipt methodology is the same one behind the Mn₃GeN verification