Is one random arrangement of the atoms enough to rank alloys by energy?

EXPERIMENTS.md · lines 43–73

E2 — Occupancy scatter in MACE at 16 atoms

Date 2026-09-12 · Question Is the screening ranking signal or noise? · Provenance …/occupancy.py

Method. Six compositions. For each, 48 independent random occupancies of the same 16-site bcc supercell at the same Vegard lattice parameter, evaluated with MACE-MP-0 medium, float64, unrelaxed. Mixing enthalpy taken against pure bcc elemental references at their own lattice parameters, same cell and settings.

Result. Within-composition sd 8.87 meV/atom; between-composition spread 101.92 meV/atom; ratio 0.087.

Pairwise flip probability for adjacent competitors: MoTaVW against MoNbTaVW, gap 8.5 meV, 23.3% chance one random occupancy inverts the order; ~10 occupancies needed to resolve. Revised by E17: computed from DFT's measured scatter of 4.03 meV/atom rather than MACE's, the flip risk is about 7%.

Interpretation. The coarse chemistry is safe — mixers against non-mixers rests on ~100 meV separations. The ranking is not: adjacent gaps of 8.5-17.8 meV against a single-occupancy sd of 7-12 meV. A search exists to separate the top few candidates, which is exactly where one occupancy is insufficient.