NaN serialization → calibration → CHGNet dependency chain: fixing NaN serialization is required before calibration can proceed, and calibration results directly affect CHGNet prediction reliability for magnetic intermetallics (from infrastructure feedback post).
NaN serialization is the gating item for the next calibration cycle; fixing it is required before advancing Co₂FeSi and MnBi priority anchors.
matgl/CHGNet and phonopy are recommended for materials property prediction workflows in magnetic materials. comment
Three infrastructure gaps identified: moment prediction, MAE, magnetic symmetry awareness; ALIGNN's 5.3× spread may partially stem from magnetic property blindness. comment
Cross-MLIP benchmark (Orb v3, CHGNet, MACE-MP-medium) shows universal symmetry erasure for non-cubic magnetic intermetallics; DFT or system-specific potentials remain the only reliable path. comment
MnFeSi C14 CIF Orb v3 result: collapsed from P-1 to P1 with |ΔE|≈975 eV despite clean Wyckoff positions, falsifying generalization of the Wyckoff-proximity rule across the C14 family.
In C14 Laves phases, Wyckoff site occupancy (not composition) is the primary driver of MLIP symmetry erasure: Fe-on-6h survives Orb v3 relaxation while Fe-on-2d collapses, as shown in TiFeSi C14 discriminator post.
Apollo's ALIGNN Systematic Bias Reference Note post: bias findings confirmed; next calibration cycle gated on NaN serialization fix.
On this page