Abstract
Neutron irradiation produces, within a few picoseconds, displacement cascadesthat are sequences of atomic collisions generating point and extended defectswhich subsequently affects the long-term evolution of materials. The diversityof these defects, characterized morphologically and statistically, defines whatis called the "primary damage". In this work, we present a fully unsupervisedmachine learning (ML) workflow that detects and classifies these defectsdirectly from molecular dynamics data. Local environments are encoded by theSmooth Overlap of Atomic Positions (SOAP) vector, anomalous atoms are isolatedwith autoencoder neural networks (AE), embedded with Uniform ManifoldApproximation and Projection (UMAP) and clustered using HierarchicalDensity-Based Spatial Clustering of Applications with Noise (HDBSCAN). Appliedto 80 keV displacement cascades in Ni, Fe$_7$0Ni$_{10}$Cr$_{20}$, and Zr, theAE successfully identify the small fraction of outlier atoms that participatein defect formation. HDBSCAN then partitions the UMAP latent space ofAE-flagged SOAP descriptors into well defined groups representing vacancy- andinterstitial-dominated regions and, within each, separates small from largeaggregates, assigning 99.7 % of outliers to compact physical motifs. A signedcluster-identification score confirms this separation, and cluster size scaleswith net defect counts (R2 > 0.89). Statistical cross analyses between the MLoutlier map and several conventional detectors (centrosymmetry, dislocationextraction, etc.) reveal strong overlap and complementary coverage, allachieved without template or threshold tuning. This ML workflow thus providesan efficient tool for the quantitative mapping of structural anomalies inmaterials, particularly those arising from irradiation damage in displacementcascades.