GPSK-05 generated a 16-site Fe₁₆N₂ structure that dissolves into P1 symmetry upon relaxation — a complete structural collapse that produces something nowhere near the well-characterized α''-Fe₁₆N₂ phase.
The generated structure (from will's GPSK-05 run) starts at P-1 symmetry, but Orb v3 relaxation drives it to P1. The cell parameters tell the story before you even look at the coordinates: α=46.4°, β=37.5°, γ=42.8°. These are not tetragonal angles. The real α''-Fe₁₆N₂ is I4/mmm with a≈5.7 Å and c≈6.3 Å — a clean body-centered tetragonal cell that GPSK-05 should have been able to approximate.
What the model produced instead: a heavily distorted triclinic cell with fractional coordinates scattered across the unit cell in a way that bears no resemblance to the layered Fe-N ordering in the known structure. Fe atoms are randomly distributed rather than occupying the specific 8h and 8f Wyckoff sites that give α''-Fe₁₆N₂ its magnetic properties.
This is the same failure mode we've been tracking with other generative models on intermetallics — close-packed Laves phases, Heusler compounds — and now Fe₁₆N₂. The diffusion transformer can generate plausible-looking atomic configurations but cannot enforce the compositional and symmetry constraints that define real intermetallic structures.
The practical implication: GPSK-05 is not a reliable tool for targeted permanent magnet composition screening. If you're looking for Fe₁₆N₂, build from the known ICSD geometry and use ML for property prediction, not structure generation.
Cell + Ionic relaxation with Orb v3; 0.03 eV/Å threshold; final energy = -134.7232 eV; energy change = -14.6983 eV; symmetry: P-1 → P1
Generated structure (left) vs. the actual α''-Fe₁₆N₂ structure (ICSD reference). The distortion is structural, not computational.
GPSK-05 generates a P1-symmetry, triclinic-distorted Fe16N2 that bears no resemblance to the real tetragonal α''-Fe16N2 phase — another data point in the pattern of generative model failure on intermetallics.