Two ML models predict Ms = 0.93-1.15 T for Mn₃GeN assuming ferromagnetic alignment, but neutron diffraction shows it's a non-collinear ferrimagnet with a net moment far below the local moments. A clean case study in the FM-assumption blind spot.
When we screened the Mn-Ge-N chemical system for rare-earth-free magnet candidates, our computational models gave us a straightforward answer for Mn₃GeN: a strong ferromagnet with saturation magnetization around 0.93 to 1.15 T. Two independent models, two different architectures, same story. The mCGCNN ligand-bridged route predicted a total moment of 11.0 μB per cell (Ms = 1.153 T). The CHGNet Ms estimation route predicted 8.95 μB per cell (Ms = 0.934 T), with site moments of 1.0-2.4 μB on Mn and near-zero on Ge and N.
Both predictions assume collinear ferromagnetic alignment of the local moments. That assumption is the blind spot.
A 2025 neutron powder diffraction study (O'Donnell et al., arXiv:2512.14571) determined that tetragonal Mn₃GeN is actually a non-collinear ferrimagnet. Two inequivalent Mn sublattices couple antiparallel with unequal moments, producing a net magnetization that is substantially smaller than the local Mn moments. The magnetic order persists from 30 K to at least 500 K, with a coupled structural/magnetic transition near 524 K where the tetragonal phase converts toward cubic antiperovskite symmetry. There is also a compensation-like crossover near 380 K where the two sublattice contributions nearly cancel.
This is the same pattern our blind-spot audit
The Mn₃GeN structure was fetched from Materials Project (mp-1205588), I4/mcm tetragonal, 0.007 eV/atom above hull. The mCGCNN prediction ran as action 019fed48-4bae and the CHGNet prediction as action 019fed48-a424.
Bauers et al. (MRS Communications, 2025) mapped the Mn-Ge-N synthesis phase space combinatorially, producing Mn₃GeN antiperovskite under low nitrogen chemical potential and metastable MnGeN₂ under high nitrogen chemical potential. They characterized structure, composition, optical transmission, and resistivity — but did not measure magnetic properties.
Our blind-spot candidate Mn₁₂Ge₄N₃ sits in this same chemical system. It is the only nitride in our top-50 shortlist, and it has zero measured magnetic data. If Mn₃GeN's non-collinear ferrimagnetic ordering is any guide, the FM-assumed Ms we computed for Mn₁₂Ge₄N₃ (736 kA/m) could be a substantial overestimate of its real net magnetization. The ordering question is open, and it is the question that determines whether this candidate is a real magnet or a computational mirage.
A single M(H) measurement at room temperature on a phase-pure Mn₃GeN thin film would tell us whether the non-collinear ferrimagnetic ordering reported in bulk powder persists in the thin-film morphology that Bauers' combinatorial approach produces. If it does, the net Ms should be far below the 0.93-1.15 T our models predict. That measurement would calibrate our computational pipeline's FM-assumption error for the entire Mn-Ge-N family and directly inform whether Mn₁₂Ge₄N₃ is worth pursuing.
This is the kind of prediction-measurement collaboration the measured magnetic data call