Was the reference simulation cell big enough for the energy model's reach?

EXPERIMENTS.md · lines 1519–1567

E31 — The reference cell was too small for its own clusters

Date 2026-09-12 · Question icet warns that the calculator self-interacts. Does it change the answer? · Provenance forager/thermo.py, …/{selfinteract,recalibrate}.py

The geometry. A 3x3x3 cell at a = 3.2935 A is 9.88 A across, so under the minimum image convention no cluster may reach beyond 4.94 A. The expansion has a 6.0 A pair cutoff. Clusters were wrapping onto their own periodic images and counting an interaction twice - in the Monte Carlo that serves as the reference, which everything else is calibrated and validated against. A 4x4x4 cell is 13.17 A across, limit 6.59 A, which clears it.

The bias, same compositions, same settings, 54 sites against 128:

This was larger than the error being reported against it. The fast path's agreement with the reference had been quoted as 0.96 meV/atom MAE; the ruler itself was bent by more than that.

Recalibrated at 128 sites, 40 compositions, and revalidated on the 15 compositions drawn after the model was chosen:

Calibration residual 2.14 meV/atom on the energy and 0.0059 on the entropy ratio. The uncertainty remains unbiased and conservative: z mean -0.13, sd 0.61, 15 of 15 inside two sigma. 1.33 meV/atom is the figure that stands; 0.96 is withdrawn. Nothing downstream breaks, because both are far below the expansion's own 5.67 (E16b) and the potential's 6.61 (E18) - but the difference between them is the difference between a measured number and a lucky one.

Fixed at the cause. CEThermo now refuses a cell too small for the expansion's cutoffs, with the required repeat in the error message. A warning in a log nobody reads is not a safeguard; this one had been printing throughout E21, E23 and E24.

Superseded. The accuracy figures in E23 and E24 were measured against the 54-site reference and are superseded by the table above. The methods in those entries stand - the variance basis of E24 was selected on correlations that the bias does not reverse - but their headline accuracy numbers should be read from here.