The structure of defect clusters formed in a displacement cascade plays a significant role in the micro-structural evolution during irradiation. We present a novel method to pattern match and classify defect clusters from Molecular Dynamics simulations of collision cascades. The methods are applied on a database of collision cascades in Fe and W at energies ranging from 10 keV to 200 keV. The results show families and classes of cluster shapes providing new insights and parameters that can be used in simulations at higher scales. We discuss each step, starting from efficient identification of defects from simulation output to reduction of physics problems to machine learning stages viz. feature engineering, dimensionality reduction and unsupervised classification. The open-source software implementation of the exploratory data analysis and visualizations of the cluster patterns provides new approach to the study of defect clusters in cascades.
We find the geometrical histograms of angles and distances between neighboring defects in a cluster to characterize different qualitative cluster shapes. We show that the histograms can be effectively used to find clusters of specific shapes in a database of cascades irrespective of differences in superfluous details. We further use these histogram representations to classify the clusters with density based unsupervised classification. We find many already known categories of clusters such as crowdions, planar crowdion pairs, rings or C15 cluster, etc. as different classes. The results also show new cluster shape classes. We analyse quantitative properties of the classes such as their sizes, dimensionality and preferences to elements and energies. The cluster shape preferences for Fe and W agree with previous studies. The distribution of cluster shapes along with their properties like diffusivity, stability, etc. can be used as input to higher scale models in a multi-scale radiation damage study.