Is the new screening model accurate on independent quantum data for these alloys?

EXPERIMENTS.md · lines 768–821

E18 — MACE-MPA-0 validated on 300 independent DFT structures

Date 2026-09-12 · Question E13's verdict rested on ten points. Does it hold on a large independent DFT set? · Provenance …/rhea_validate.py, rhea_equil.py; RHEA database, Zenodo 10.5281/zenodo.18863415

Independence. MACE-MPA-0 trains on MPtrj + sAlex. MPtrj was counted exhaustively (E14) and contains zero materials with four or more of these elements, so RHEA's 3-to-9-component random solid solutions cannot be in it; sAlex is Alexandria prototype substitution, a different kind of object; and RHEA is dated March 2026, after the model. RHEA's own materials_project group (147 of 22,477 frames, 0.65%) is the only possible overlap and is excluded.

Method. RHEA bcc_alloys: 2,150 structures of 54 atoms, randomly ordered, randomly displaced, several volumes per composition. VASP-PBE. Absolute energies are not comparable across codes, so one constant per element is fitted by least squares - the standard element referencing - and the residuals judged. Restricted to the eight elements used here (1,454 of 2,150 structures).

Result — the error is almost entirely off-equilibrium.

meV/atom, strain measured against the Vegard volume for each structure's own composition. RHEA spans -42% to +24.8% volume strain by design, because it exists to train a potential; screening operates at the Vegard volume.

Interpretation. MACE-MPA-0 is accurate to 6.61 meV/atom where this project uses it, on 300 independent VASP structures at 54 atoms in the right chemistry. That is better than the optimistic ten-point estimate of 23 (E13) and about ten times better than the MACE-MP-0 the project started with. The RHEA-wide figure of 48 is a statement about compressed and expanded cells, not about screening.

6.61 meV/atom is below the DFT occupancy scatter of 4.03 measured in E17 to within a factor of two - that is, the potential is now approaching the intrinsic configurational noise of the quantity itself.

Consequences.

• Fine-tuning is not needed for near-equilibrium screening. It would chase an error comparable to the physics' own scatter.
• Fine-tuning on RHEA is the right move for relaxation, molecular dynamics or defects, which live in the 62 meV/atom regime. Freeze the foundation head and add a second head, per the documented recipe.
• A distortion proxy (spread of first-shell distances) did not separate the error cleanly, Spearman +0.09 with a non-monotonic pattern across quartiles. Volume strain did. The mechanism is recorded as volume, not as displacement disorder.

Caveat. The Spearman values reported by these scripts are near 1.000 in every regime because they correlate total energies dominated by composition; they carry no information here and MAE is the operative metric.