Identifying structures and defects with machine learning using MultiSOM

Identification of defects in crystalline structures is of vital importance when describing the plastic behavior of metals. Despite the increasing number of tools available in the literature, additional techniques are still required to classify different types of defects. We present a novel tool based on Multi-layer Self Organizing Maps (MultiSOM) [1], which can be employed in the identification of vacancies, dislocations and free surfaces at the atomic scale. Three case-studies were analyzed and the performance of MultiSOM was compared to traditional tools, such as common neighbor analysis and polyhedral template matching, revealing that our algorithm is able to identify local structures while other methods are unable to. We demonstrate that excellent results can be obtained for identifying local structural patterns in solids through the MultiSOM framework when combining per-atom properties such as coordination numbers and centrosymmetry parameters; though alone, each property is insufficient. Overall, MultiSOM is an open-source software, that works under the Python programming language with user-selected features. The software is available at the following GitHub repository: https://github.com/SIMAF-MDZ/MultiSOM