The Role of Deep Learning and High-Throughput Dynamical Simulations in Space Group Determination from Kikuchi Patterns

Abstract The design of novel materials hinges on the understanding of structure–property relationships. However, in recent times, our capability to synthesize a large number of materials has outpaced our speed of characterizing them, and thorough crystal structure analysis of newly synthesized samples remains a bottleneck in high-throughput nanomaterials discovery. Thus, rapid methods for crystal symmetry determination that can analyze many material samples within a short time frame are especially needed. Kikuchi diffraction in the scanning electron microscope (SEM) is a promising technique for this due to its sensitivity to dynamical scattering, which may provide space group information. In this study, we train neural networks to classify the space group from cubic Kikuchi patterns to enable its use for high-throughput symmetry classification of novel material samples. Due to the low number of material phases in the experimental training data, an unsupervised deep learning-based domain adaptation method was utilized to train neural networks on both simulated and experimental data. This helped to remedy the low diversity of the experimental training data. Additionally, instead of predicting the true space group, we relabel the data based on space group information that is realistically obtainable given physical limitations of Kikuchi diffraction, which enables our models to achieve accuracy scores higher than 95% for simulated and experimental data. This suggests that adapting to limitations imposed by diffraction physics is crucial for developing working models for symmetry classification.