Pattern Matching and Classification of Clusters in Collision Cascades

Abstract

The structure of defect clusters formed in a displacement cascade plays a significant role in the micro-structural evolution during irradiation. Molecular dynamics simulations have been widely used to study collision cascades and subsequent clustering of defects. We present a novel method to pattern match and classify defect clusters. A cluster is characterized by the geometrical and topological histograms of its angles and distances which can then be used as similarity metrics. The technique is demonstrated by matching similar clusters for different cluster shapes like ring, crowdions etc. in a database of cascade damage configurations in Fe and W at different energies. We further use graph based dimensionality reduction techniques and unsupervised machine learning on the features of all the clusters present in the database to find classes of clusters. The classification successfully separates out many already known categories of clusters such as crowdions, planar crowdion pairs, rings and perpendicular crowdions. The dimensionality and size of different classes provides a broad categorization of classes. The distribution of different classes of shapes among cascades of different elements and energies shows the exclusivity of shapes to elements and energies. We discuss the key points and computational efficiency of the algorithms along with the various prominent results of their application. We discuss the motivation for using machine learning and statistics for the problems and compare different techniques. The algorithms along with the supporting analysis and visualizations give an unsupervised approach for classification and study of defect clusters in cascades. The distribution of cluster shapes and structures along with the shape properties like diffusivity, stability, etc. can be used as input to higher scale models in a multi-scale radiation damage study.