All-around local structure classification with supervised learning: The example of crystal phases and dislocations in complex oxides

To accurately identify local structures in atomic-scale simulations of complex materials is crucial for the study of numerous physical phenomena including dynamic plasticity, crystal nucleation and glass formation. In this work, we propose a data-driven method to characterize local atomic environments, and assign them to crystal phases or lattice defects. After constructing a reference database, our approach uses descriptors based on Steinhardt's parameters and a Gaussian mixture model to identify the most probable environment. This approach is validated against several test cases: polymorph identification in alumina, and dislocation and grain boundary analysis in the olivine structure.