AI-EXAFS: Physics-Informed Prediction of EXAFS from Atomic Structure

Understanding atomic structure across chemical systems relies heavily on extended X-ray absorption fine structure (EXAFS) because of its sensitivity to local atomic environments. However, quantitative interpretation remains inherently iterative, requiring structural parameters or models to be fitted against experimental spectra using explicit multiple-scattering calculations. Machine-learning approaches have accelerated EXAFS analysis, yet direct prediction of full spectra from atomistic structure across diverse…

By the Truvace deskOctober 20, 20261 min read1 sourcePeer-reviewedTRV-2026-1975· cite & verify

AI-EXAFS: Physics-Informed Prediction of EXAFS from Atomic Structure — Ritimukta Sarangi (8696740314) by ENERGY.GOV. Public domain

In brief

Researchers built AI-EXAFS, a physics-informed graph neural network that takes three-dimensional atomic coordinates and outputs extended X-ray absorption fine structure spectra. It uses an absorber-centered graph, distance-modulated message passing, and a training objective designed for EXAFS.

The model was trained and tested on 85,991 transition-metal complexes whose spectra were generated with FEFF calculations, and it reproduced held-out spectra with low R-factors. This matters because conventional EXAFS interpretation is iterative and fit-based, so a direct structure-to-spectrum predictor could speed screening of candidate structures. The reported results come from simulated spectra, so performance on experimental data is still to be established.

Main points

The model predicts spectra directly from 3D atomic coordinates without predefined structural descriptors or manual parameter selection.
It was trained and evaluated on 85,991 transition-metal complexes with FEFF-generated spectra.
Average held-out test-set R-factors were 0.043 in k-space and 0.019 in R-space.
The authors present it as a scalable foundation for increasingly complex and experimentally relevant environments.

The gain

A graph neural network can predict full EXAFS spectra directly from atomic coordinates, removing manual parameter selection and allowing inference across many candidate structures.

The rundown

EXAFS is valued for its sensitivity to local atomic environments, but quantitative interpretation normally requires fitting structural parameters or models to experimental spectra using explicit multiple-scattering calculations, which is iterative.

Earlier machine-learning methods have accelerated EXAFS analysis, yet predicting full spectra from atomistic structure across diverse chemical environments remained difficult. AI-EXAFS addresses this with an architecture combining an absorber-centered graph and distance-modulated message passing.

The authors describe the result as a scalable structure-to-spectrum framework that enables direct inference across large sets of candidate structures, and as a base for later work in more complex environments.