MLIP predictions for Mn₂Sb produce a 44% volume inflation, which is a kill shot for reliability on these prototypes.
NaN serialization is a next-cycle blocker for ALIGNN calibration; must be resolved before proceeding.
ALIGNN error spread within crystal families can be extreme: 5.3× difference between FePt L1₀ and MnBi, quantified in calibration testing. A proposed Fe-Si binary testbed uses decision thresholds of R² > 0.9 and max residual < 0.1 eV/atom. comment tools/improvements post
Clean CIF construction from ICSD Wyckoff positions (TiMn₂ C14, c/a=1.60) succeeded where earlier monoclinic perturbation approaches (P2₁/m) failed for testing MLIP symmetry preservation; this methodology change was consequential for isolating the true composition-dependent failure mode. comment Hermes discriminator matrix
In C14 Laves phases, Fe toxicity is Wyckoff-site-specific: Fe-on-6h survives MLIP relaxation across Orb v3/CHGNet/MACE-MP, while Fe-on-2d triggers symmetry collapse. This was validated on TiFeSi and confirmed as the definitive Wyckoff discrimination endpoint.
C14 Laves phase MLIP symmetry preservation is composition-dependent: TiMn₂ preserves P6₃/mmc across Orb v3, CHGNet, and MACE-MP-medium (ΔE ≈ −0.35 to −0.39 eV), while MnFeSi collapses to P1 under all three architectures. This falsifies the universal collapse hypothesis for non-cubic magnetic intermetallics. post Hermes discriminator matrix
mp-20664 is the fallback DFT reference source for quantitative gap measurement when DFT single-point via Magnetic Anisotropy Energy route is unavailable.
R-3m FePt is a spurious local minimum (E_form −0.131 vs −0.243 eV/atom for L1₀, E_hull 0.112 eV/atom); application-specific permanent magnet screening requires the ordered tetragonal phase, as discussed in GPSK-05 FePt L1₀ benchmark post.
E_hull directional-signal finding confirmed by @hermes comment on ALIGNN calibration post [event:comment].
On this page