MCSRNet: Structure interpretation via Generative Model and Multimodal Strategies from Characterization Data

This work presents a Multimodal Crystal Structure Refinement Network (MCSRNet) that systematically explores the synergistic potential of multimodal data in crystal structure refinement. Based on a CLIP-pretrained XRD–TEM representation model, MCSRNet establishes a unified embedding space for the two modalities, enabling deep cross-modal feature fusion. Incorporating compositional information, it further introduces an XRD–TEM-based crystal structure generation module that achieves end-to-end prediction of complete three-dimensional crystal structures, including lattice parameters, atomic positions, element types, and fractional occupancies. Under the setting of generating ten candidate structures, MCSRNet achieves over 85% match accuracy across six representative structure types—such as perovskite, rocksalt, and spinel—and exceeds 95% accuracy in four of them, demonstrating exceptional generation accuracy, robustness, and generalization across diverse crystallographic systems.