Classification of crystal structures using electron diffraction patterns with a deep convolutional neural network.

Area of Science:

• Materials Science
• Crystallography
• Artificial Intelligence

Background:

• Accurate crystal structure identification is crucial for materials science.
• Electron diffraction patterns provide rich crystallographic information.
• Deep learning offers potential for automated analysis of complex data.

Purpose of the Study:

• To evaluate the effectiveness of a pre-trained deep convolutional neural network (ResNet) for classifying crystal structures using electron diffraction data.
• To assess the performance of different ResNet architectures without prior material-specific knowledge.
• To explore the application of computer vision techniques for analyzing crystallographic data.

Main Methods:

• Simulated selected area electron diffraction (SAD) patterns using varying acceleration voltages, zone axes, and camera lengths.
• Utilized crystal information format (CIF) files from open repositories for data generation.
• Trained and validated residual neural network (ResNet) models, including ResNet101, on simulated SAD patterns of cubic crystal systems.
• Re-labeled simulated diffraction patterns for enhanced machine learning recognition.

Main Results:

• The ResNet101 architecture achieved a validation accuracy of 92.607% for classifying five cubic space groups.
• Different ResNet architectures showed varying classification efficiencies.
• The approach demonstrated the feasibility of using DCNNs for crystal structure determination from diffraction data.

Conclusions:

• Off-the-shelf deep convolutional neural networks, specifically ResNet101, are effective for classifying crystal structures from electron diffraction patterns.
• Simulated SAD patterns, when processed with computer vision principles, can serve as a robust dataset for training AI models.
• This method holds promise for accelerating materials discovery and characterization.