A structure score reduces DFT checks under an error budget

When a pretrained molecular model screens many configurations, which predictions should be checked with costly density functional theory (DFT)?

The actual error is unavailable until DFT is run, and a score that ranks likely errors does not by itself establish how much error can remain among configurations kept with the model.

The method ranks molecular errors from structure and calibrates a cutoff so DFT checks can be selected under an expected capped-error budget.

To screen molecular configurations with a pretrained model, how can a workflow decide which predictions need costly density functional theory (DFT) checks when the true error is hidden until DFT is run? A ranking can prioritize likely errors, but it cannot alone set a cutoff with a stated error limit.

The method pairs a structure-based error score with calibration under the same risk-control procedure used for committee and random scores.

Across eight dataset–model combinations, its mean DFT-call share was lower than those alternatives at the tested target; this supports the conclusion that the score can reduce calls under that shared expected-error budget.

The guarantee is an expected bound on capped, identity-level retained error under exchangeability within the evaluated dataset–MLIP setting. It does not guarantee each prediction or pool, and it does not cover relative energies or conformer selection unless those objectives are calibrated directly.