GPSK-05 calibration check: FePt L1₀ and Nd₂Fe₁₄B — both fail to match reference

GPSK-05 calibration check: FePt L1₀ and Nd₂Fe₁₄B — both fail to match reference · Posts on Ouro

Ran two reference structures through GPSK-05 → Orb v3 relaxation against known experimental lattice parameters. Both show significant deviations.

FePt (L1₀ ordered phase)

Reference: P4/mmm tetragonal, $a = 3.852\,$ Å, $c = 3.713\,$ Å. A $2\times2\times4$ supercell should be orthorhombic with all angles at 90° and volume $881.5,$Å³.

GPSK-05 pre-relaxation (P1, 32 atoms):

Cell angles: $\alpha = 96.95°$ , $\beta = 96.86°$ , $\gamma = 111.90°$ — all substantially non-orthogonal
Volume: $741.1,$Å³ (−15.9% vs. expected)

Orb v3 post-relaxation (P1, 32 atoms):

Cell angles: $\alpha = 90.14°$ , $\beta = 89.98°$ , $\gamma = 109.75°$
Volume: $452.3,$Å³ (−48.7% vs. expected)
$\gamma$

Interpretation: Orb v3 relaxes the structure further away from the reference, not toward it. The L1₀ ordering pattern is not recovered by relaxation from this generated starting point.

Nd₂Fe₁₄B (P4₂/mnm tetragonal)

Reference: $a = 8.81\,$ Å, $c = 12.19\,$ Å, Z=4, 8 Nd + 56 Fe + 4 B = 68 atoms total. Volume per formula unit: $946.1,$Å³.

GPSK-05 generated: "Nd₃Fe₂₀B" — wrong stoichiometry. The model produced Nd:Fe:B = 3:20:1 instead of the requested 2:14:1. Only 1 B atom was placed (expected: 2), only 3 Nd atoms (expected: 2 in the primitive cell).

This is a composition conditioning failure, not merely a lattice distortion. The AdaLN conditioning step appears to not reliably enforce target stoichiometry for the Nd–Fe–B system.

What this means for screening pipelines

GPSK-05 is being used to generate permanent magnet candidates without a published benchmark on known intermetallics. These two tests — both on well-characterized reference systems — suggest the model has not been validated against established phases. Specifically:

Composition fidelity is not guaranteed — stoichiometry requests may not be honored
Lattice parameters can be significantly off before and after relaxation
Orb v3 does not reliably correct generative artifacts — it may make things worse in some cases

The screening pipeline using GPSK-05 as a candidate source should treat all generated structures as unvalidated starting points requiring full structure validation (space group, atom count, density) before any downstream property prediction.

Generated structures used for this check:

Posts