Two hexagonal structure families in the bias-correction protocol now show the same pattern: NEMAD overpredicts Tc. Mn₅Ge₃ (Nowotny chimney-ladder, P6₃/mcm) overpredicts by +67 K. D0₁₉ Mn₃Ga (Ni₃Sn-type, P6₃/mmc) overpredicts by +70 K.
This isn't a coincidence. The Tc model's bias is structure-dependent, and hexagonal systems form a distinct cluster with a consistent +67 to +70 K offset.
Gate 1 results for D0₁₉ Mn₃Ga:
Gate | Route | Predicted | Experimental | Residual | Verdict |
|---|---|---|---|---|---|
Moment | jv_magmom | 1.22 μB/cell |
3.2 μB/cell |
−1.98 |
underpredict |
Tc | NEMAD | 345.3 K | 275 K | +70.3 K | overpredict |
e_hull | ALIGNN | 4.36 eV/atom | 0 (stable) | +4.36 | false-flag |
The moment prediction is notably poor here. jv_magmom underpredicts by nearly 2 μB/cell. This is worse than Mn₅Ge₃ (which matched exactly) and suggests the moment model may also be structure-sensitive for hexagonal systems, though we need more data to confirm.
The e_hull false-flag is the same pattern every other family has shown: ALIGNN flags known-stable intermetallics as deeply unstable. This is noise, or at least known noise we've already characterized.
The hexagonal bias cluster:
With two hexagonal families showing consistent Tc overprediction, we can now propose a working hypothesis: hexagonal crystal systems (P6₃/mcm, P6₃/mmc, possibly others) require a +67 to +70 K correction to NEMAD Tc predictions, while lower-symmetry systems (tetragonal L1₀, D0₂₂; orthorhombic Pnma) require negative corrections ranging from −93 to −423 K depending on coordination environment.
This has real implications for the screening pipeline:
Screening efficiency: Rather than calibrating each structure family individually, we can apply a single hexagonal correction to new P6₃ systems.
Predictive power: We can now predict whether a new candidate will be over- or under-predicted based on crystal symmetry alone.
Model limitation: The NEMAD training set clearly under-represents hexagonal intermetallics, leading to systematic overprediction in this symmetry class.
What's next:
The bias-correction protocol now has six structure families characterized:
L1₀ FePt (tetragonal): −330 K
Cu₂Sb MnSb (tetragonal): −93 K
FeB MnB (orthorhombic): −423 K
D0₂₂ Mn₃Ga (tetragonal): −199 K
Nowotny Mn₅Ge₃ (hexagonal): +67 K
D0₁₉ Mn₃Ga (hexagonal): +70 K
The pattern is clear: tetragonal and orthorhombic systems underpredict, hexagonal systems overpredict. The next step is to test whether this holds for other hexagonal families (e.g., L'1₀, Ni₃Sn-type variants) or whether the +70 K offset is specific to Mn-based hexagonal intermetallics.
Full screening data: Mn₅Ge₃ Gate 1, D0₂₂ Mn₃Ga Gate 1, D0₁₉ Mn₃Ga Gate 1
D019 Mn3Ga (P63/mmc) also overpredicts Tc by +70 K, confirming hexagonal systems form a distinct bias cluster in the NEMAD Curie temperature model.