Physics-guided deep learning strategy for 2D structure reconstruction from diffraction patterns

Two-dimensional (2D) materials have garnered significant attention due to their tunable electronic and optical properties and exceptional mechanical performance. Reconstructing 2D structures from diffraction patterns without prior assumptions or comprehensive knowledge is challenging, especially for heterogeneous stacking and quantum 2D materials. Here, we introduce DD2D (diffraction pattern deep-reconstruction 2D structures), a physics-guided deep learning method that predicts 2D structures directly from diffraction patterns. DD2D employs a twin-tower framework, integrating a crystallographic geometric encoder and a site texture encoder, and uses a self-attention mechanism to identify intrinsic correlations in physical information and corresponding areas in the diffraction pattern. The results demonstrate high anti-interference, robust recognition capabilities, reliable interpretability, and prediction accuracy of up to 99.0%, highlighting its potential for future 2D materials discoveries. The authors introduce DD2D, a physics-guided deep learning method that predicts 2D structures directly from diffraction patterns using a twin-tower framework. The method demonstrates high anti-interference, robust recognition, and up to 99.0% prediction accuracy, showing promise for future 2D materials discoveries.