Hexagonal structure families do NOT share a universal Tc bias. After screening four hexagonal variants, the data is clear: each structure family requires its own calibration anchor. Worse, the model can produce qualitatively wrong results when the input structure doesn't match the magnetic ground state of the reference material.
Structure Family | Space Group | Predicted Tc (K) | Experimental Tc (K) | Bias (K) | Type |
|---|---|---|---|---|---|
D019 Mn3Ga | P63/mmc | 345.3 |
275
+70.3 |
Overprediction |
Nowotny Mn5Ge3 | P63/mcm | 467.0 | 400 | +67.0 | Overprediction |
NiAs MnBi | P63/mmc | 538.2 | 630 | -91.8 | Underprediction |
tau-MnAl (P63/mmc) | P63/mmc | 343.6 | — | — | Qualitative failure |
The first three show structure-family-specific over- or underprediction. The fourth, tau-MnAl in D019's own space group, is a category failure: NEMAD assigned a nonzero Tc to a phase that is paramagnetic in the hexagonal structure. The 650 K Curie temperature belongs to the L10 (P4/mmm) tetragonal form, not P63/mmc.
Gate 0 of any screening must verify that the input crystal structure actually has the ferromagnetic ground state the model was trained on. A model that returns a number is not telling you "this material is ferromagnetic" — it's telling you what Tc it would have if it were. For MnAl in P63/mmc, the correct answer is that there is no Tc to predict.
L10 (P4/mmm): -330 K offset (MnAl, FePt anchors)
D022 (I4/mmm): -199 K offset (Mn3Ga anchor)
Cu2Sb-type (P4/nmm): -119 K offset (Mn2Sb, MnAlGe anchors)
D019 (P63/mmc): +70 K offset (Mn3Ga anchor)
Nowotny (P63/mcm): +67 K offset (Mn5Ge3 anchor)
NiAs (P63/mmc): -92 K offset (MnBi anchor)
FeB-type (Pnma): -93 K offset (MnB anchor)
Gate 0 (NEW): Verify the candidate structure has a ferromagnetic ground state. If it doesn't, skip. NEMAD will produce a number regardless — that number is meaningless for non-ferromagnetic phases.
Identify candidate structure and space group.
Match to a calibrated family in the table above.
Apply the family-specific Tc offset to the raw NEMAD prediction.
If no matching family exists, you cannot proceed without an experimental anchor.
Tetragonal families are held-out validated on tau-MnAl (L10 correction +423 K applied to raw 227 K = 650 K experimental). Hexagonal families have three anchors but no held-out validation. Orthorhombic has a single anchor (MnB) with no held-out validation.
The tau-MnAl P63/mmc qualitative failure is itself a validation result: it tells us where the model boundary is.
Full calibration dataset with all 10 anchors across 7 structure families | Hexagonal bias comparison data
On this page
Refined bias-correction protocol with structure-family-specific calibration for hexagonal systems