Accelerating electron diffraction analysis using graph neural networks and attention mechanisms

Abstract Electron diffraction(ED) often used to solve for unknown structures or refine existing ones. Existing methods for automated ED analysis often struggle with challenges such as computational expense and experimental noise. This study introduces a deep learning framework to accelerate and improve crystal structure determination from diffraction patterns. The methodology treats each diffraction pattern as a relational graph of Bragg spots. Spot features are encoded using a 1D convolutional network, from which a relational attention aggregator constructs an orientation-agnostic graph. This graph is processed by a Graphormer encoder enhanced with Mixture-of-Experts layers, allowing the model to learn complex crystallographic relationships efficiently. Trained and tested on a large dataset of simulated diffraction patterns, the model achieved a crystal system classification accuracy of 89.2% and a space group accuracy of 70.2% from single patterns, significantly outperforming a state-of-the-art random forest baseline (74.2% and 57.8%, respectively). By aggregating predictions across multiple zone axes, these accuracies improved to 96.5% and 79.5%. The model also demonstrated robust performance on experimental data of gold nanoparticles, producing plausible classifications consistent with known orientation degeneracies. By unifying relational graph reasoning with specialized expert networks, this work presents a robust and automated framework for high-throughput materials characterization.