Estimate the false-negative rate of relaxation-based validation criteria

Determine the fraction of erroneous Crystallography Open Database entries that the relaxation-based validation criteria fail to detect, using calibration against a sufficiently large ground-truth-labelled dataset.

Background

The study evaluates whether relaxation outcomes from universal machine-learning interatomic potentials can identify physically or chemically erroneous crystal structures in the Crystallography Open Database. Although the authors manually inspected 327 structures and found that convergence, volume change, energies, forces, and displacement descriptors provide varying information about structural validity, the labelled sample is too small to calibrate the criteria against definitive ground truth.

Consequently, the reliability of the screening procedure remains unresolved: in particular, the proportion of erroneous entries that pass the criteria undetected cannot be quantified from the present dataset. Establishing this false-negative rate would require substantially more labelled structures and an appropriate ground-truth validation protocol.