Fig. 1. Defect detection and stratification based on the distortion score.

a Scheme of the defect detection. The training data set consists of the defect-free bulk structures. The structures for the training and test are represented in the same feature space ${\mathbb{R}}^D$ of atomic descriptors. To perform defect detection, each atomic environment from the test system is compared to the learned bulk structures and characterized by distortion score. Atoms with scores above the critical threshold are classified as structural outliers that form the defect. b Detection and stratification of four self-interstitials cluster with C15 morphology³⁸—$I_4^{\mathrm{C15}}$—based on the distortion score provided by robust MCD. Each point on the plot represents an atom in the simulation box. The colour of points corresponds to the colour of atoms in the inset structures. The threshold between the bulk (atoms-inliers) and defect (atoms-outliers) is indicated with a dashed red line, labelled as A. The grey dashed lines B and C indicate the possibilities for defect stratification. The defect structures A, B, C are obtained using the corresponding thresholds.