Fig. 6: Reliability of outlier detection for the different strategies.
From: Outlier-detection for reactive machine learned potential energy surfaces
Given the 1000 structures with the largest variance/uncertainty, it is evaluated whether they correspond to the structures that also have the largest errors from comparison with reference data for Ndata = [25, 50, 100, 200, 400, 800, 1000]. i.e., it is evaluated whether the Ndata structures with the actual highest errors are contained in the 1000 that are predicted to have high errors.
