3D Structure Analysis from XAS using Diffusion Models

The local three-dimensional (3D) atomic structure of materials and molecules can be interrogated by X-ray absorption near-edge structure (XANES), a powerful analytical technique. Here, we introduce a diffusionmodel-based approach that rapidly infers candidate 3D structures from X-ray absorption spectroscopy (XAS) data. The method is validated on multi-element hard X-ray XAS datasets and generalizes across a broad range of molecular and material systems without imposing restrictions on the number of constituent element types. Compared with multi-round iterative optimization methods for 3D structure determination that use a structure-to-spectrum model as the core spectral calculation engine, our approach substantially improves data analysis efficiency, enabling 3D structure determination within a few reverse diffusion steps. On low-end consumer GPUs, 3D structure analysis takes 0.5–5s. We anticipate that this method will significantly facilitate online data analysis at beamlines and associated real-time experimental planning and optimization, enable batch screening of atomic-level 3D structures, and improve the efficiency of conventional 3D structure fitting.