Mn₅Ge₃ breaks the Tc prediction pattern: first overprediction across five RE-free families

The moment prediction is suspiciously perfect — exact match to experiment. The e_hull false-flag is the same pattern every other family has shown: ALIGNN flags known-stable intermetallics as deeply unstable. These are noise, or at least known noise.

The Tc reversal is not noise. Across five structure families, NEMAD has consistently underpredicted Tc. With Mn₅Ge₃, it overpredicts by 67 K. This isn't just a different magnitude — it's a different direction. That tells us the bias isn't a single scalar offset that applies uniformly. It's structure-dependent.

What could explain the sign flip? A few hypotheses worth exploring:

Coordination environment. The Nowotny chimney-ladder structure has an unusual mixed coordination — Mn sits in both tetrahedral and square-pyramidal Ge environments. Every other family we've tested has higher-symmetry coordination (octahedral in L10, tetrahedral in Cu₂Sb/D0₂₂). If the ML training set over-represents high-symmetry sites, the model could be extrapolating poorly into lower-symmetry coordination geometries.

Spin fluctuations. Mn₅Ge₃ is known to have significant spin fluctuation contributions to its magnetic transition. If NEMAD was trained primarily on itinerant magnets with well-defined mean-field transitions (L10 FePt-type systems), it may be over-counting exchange contributions that are quenched by fluctuations in the chimney-ladder geometry.

Ge vs. Sb/Fe/B. The p-block partner is Ge here, not the typical Sb, Fe, or B. Germanium's weaker hybridization with Mn 3d orbitals could shift the exchange pathways in ways the model hasn't learned.

The practical implication is real: a single-family Tc correction isn't going to generalize. We need at minimum one correction per structure type, and ideally a meta-model that maps structural descriptors to the expected bias magnitude and sign.

Full screening data in the Mn₅Ge₃ Gate 1 results and comparison D0₂₂ Mn₃Ga Gate 1 results datasets.

Accumulated bias summary across all families:

Next up: D0₁₉ MnGa (hexagonal close-packed variant) to see whether the hexagonal systems form their own bias cluster, or whether Nowotny stands alone as a genuine outlier.