Crystal phase identification of“odd shaped” particles with increasing roundness using machine learning.

Abstract

Crystalline structure identification is a widely studied topic in condensed matter physics. Understanding how the structure of materials are determined by the shape of the particles it is build from is important for the understanding of the properties of these materials and can potentially help us tweak these properties to develop useful applications. In this research, we investigate how we can use machine learning tools for the identification of crystal structures by looking at systems of rounded polyhera. The bond order parameters of these systems are obtained from Floppy Box Monte Carlo simulations that produce large, high dimensional data sets. This data is analyzed with use of unsupervised machine learning, specifically dimensionality reduction techniques PCA and diffusion maps. We find that these techniques can identify different crystal structures for simple systems. For more complicated ones, a clear distinction of all crystal structures is hard to make with the naked eye. Additional techniques, like clustering algorithms, are needed to make a complete identification in more complex systems. Although the differences in results are minimal, diffusion maps reduces dimensionality slighlty better than PCA.