Classification of crystal structure based on XRD patterns using deep learning

Zhatong Du MS Defense, Advised by Niaz Abdolrahim

Tuesday, August 16, 2022
1 p.m.

Deformation behaviors of materials under extreme conditions have attracted interest from material scientists. Researchers use X-ray diffraction methods to reveal microscopic material structure changes. The traditional ways of decoding an XRD pattern rely on software of programmed algorithms based on ab initio first principle indexing and Rietveld refinement, where correct peak indexing and expertise are vital. In the last decade, machining learning and artificial intelligence technology developed rapidly with the rise of computational capabilities. One of the most common methods is a deep learning based on artificial neural networks. Classification task allows deep learning models to learn and assign labels to feature inputs from the problem domain.
This work applies a deep learning model to identify material structures, where crystal symmetries and full XRD patterns are labels and input features. First, we simulate ideal XRD patterns from more than 170,000 known structures from the ICSD database. Then we train deep neural network models to learn from all calculated XRD patterns and their symmetry labels. The model results demonstrate that deep learning can decode XRD patterns to the structure symmetries without any material science insights or programmed algorithms. Moreover, we pushed the models to real-world experimental XRD patterns from the RRUFF database and achieved 77% and 63.5% on the 7-way crystal system and 230-way space group classifications.