Experiments · E41

Does more training data pin down the ordering temperature?

Withdrawn. It scattered 96–150 K however much data was used; that scatter was later traced to how the peak was read, not to the data.

In the log: The transition temperature does not converge with data, and E39 over-read one run

supersededDate 2026-09-12, as written in the logrung 1 · ordering0 predictions · 0 result paragraphsEXPERIMENTS.md lines 2074–2101
exp E41 diagram
What E41 did and how it came out, drawn from this record and the files it names (book/assets/diagrams/exp/E41.svg).

Results

No result paragraph for this entry was found in the log.

The full record

EXPERIMENTS.md · lines 2074–2101

E41 — The transition temperature does not converge with data, and E39 over-read one run

Date 2026-09-12 · Question How much data does a local expansion need before the short-range order and the transition temperature settle? · Provenance …/{local_ce,odt_converge,wl_test}.py

The literature said what to expect, and half of it applied. Xie, Zhou, Luan and Jiang (J. Chem. Theory Comput. 18, 3795, 2022) found that a training set large enough for the cross-validation score to converge still leaves the transition temperature badly uncertain, needing a few thousand configurations - which is why they generated theirs with a machine-learning force field rather than DFT. Our first pass reproduced their symptom exactly: fit error fell from 16.8 to 3.5 meV/atom while the transition wandered 450-900 K.

Their fix did not work here, which is the finding. Carried to 3,000 structures, with four independent Monte Carlo runs per fit so the transition comes with an error bar:

structures CV RMSE transition sd across runs strongest pair alpha
150 6.40 525 150 Mo-Ta -1.160
400 3.35 475 96 Mo-Ta -1.136
1000 3.14 425 96 Mo-Nb -0.769
2000 2.98 525 126 Mo-Nb -0.888
3000 3.02 500 129 Nb-V -0.815

The scatter is 96-150 K at every training size, while the fit converges by 400 structures. The uncertainty is Monte Carlo run-to-run, not the fit - so more training data cannot reduce it, and the bottleneck here is not the one the literature identified. Our transition is 500 +/- 130 K.

Related entries

Built with PRISMWebsite and visualizations made using Claude