Determining Local Structure from X-ray Absorption Near-Edge Spectra with Diffusion Models.

26^th September 2025, 15:00
Tanaporn Na Narong
Columbia University, US

Abstract

X-ray near-edge absorption spectra (XANES) contain rich information about the local atomic environment, but they are rarely used to directly extract full 3D structures beyond simple descriptors like oxidation state or coordination number. Inspired by advances in generative AI, we train a latent diffusion model conditioned on XANES to reconstruct local geometries around transition metal atoms in oxides. Parameterized by equivariant graph neural networks, the model outputs cartesian coordinates of the metal’s neighbors up to the second coordination shell (~6 Å). But how well does it actually work? To answer this, we evaluate the reconstructions with Pair Distribution Functions and developed a tool to transform, align, and visualize overlapping local structures. In this talk, I will present the model, our evaluation framework, as well as further challenges for deployment.