AI-EXAFS: Physics-Informed Prediction of EXAFS from Atomic Structure.
5 affiliations
- 1Institute of Chemical Research of Catalonia (ICIQ-CERCA), The Barcelona Institute of Science and Technology (BIST), Tarragona, Spain.
- 2Departament de Química Inorgànica i Orgànica (Secció de Química Inorgànica), Institut de Química Teòrica i Computacional (IQTC), Universitat de Barcelona, Barcelona, Spain.
- 3Department of Chemistry and Applied Biosciences, Institute for Chemical and Bioengineering, ETH Zurich, Zurich, Switzerland.
- 4NCCR Catalysis, Zürich, Switzerland.
- 5Paul Scherrer Institute, Villigen, Switzerland.
Abstract
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 chemical environments remains challenging. Here, we introduce AI-EXAFS, a physics-informed graph neural network that predicts spectra directly from three-dimensional atomic coordinates. Combining an absorber-centered graph representation with distance-modulated message passing and an EXAFS-specific training objective enables direct learning of structure-spectrum relationships without predefined structural descriptors or manual parameter selection. Trained and evaluated across a dataset of 85,991 transition-metal complexes with FEFF-generated spectra, AI-EXAFS achieves average held-out test-set R-factors of 0.043 in k -space and 0.019 in R -space, demonstrating accurate structure-to-spectrum prediction across a broad range of local chemical environments. AI-EXAFS establishes a scalable structure-to-spectrum framework for EXAFS, enabling direct inference across large sets of candidate structures and providing a foundation for data-driven prediction across increasingly complex and experimentally relevant environments.
Related
- From Local Atomic Motifs to Thermodynamic State: An Interpretable Physics-Informed Framework for Cu-Zr Metallic Glasses
- Physics-Informed Latent Neural Operator for Three-Dimensional Compressor Cascade Flow Prediction
- Small-data physics-informed ML model for magnetic exchange based on chemical composition and structure
- Hereditary Effective Atom Theory for Large-Scale Atomic Structure Prediction
- Toward Reliable Railway-Bogie Response Prediction Using Multifidelity TDNN and Physics-Informed Residual Learning
BibTeX
@article{herasdomingo2026aiexafs,
title = {{AI-EXAFS: Physics-Informed Prediction of EXAFS from Atomic Structure.}},
author = {Javier Heras-Domingo and Andrea Ruiz-Ferrando and Adam H Clark and Stephan Pollitt and Sharon Mitchell and Javier Pérez-Ramírez and Núria López},
journal = {Advanced science (Weinheim, Baden-Wurttemberg, Germany)},
year = {2026},
month = oct,
doi = {10.1002/advs.78226},
url = {https://europepmc.org/article/MED/42893450},
note = {CC BY},
}
RIS
TY - JOUR
TI - AI-EXAFS: Physics-Informed Prediction of EXAFS from Atomic Structure.
AU - Javier Heras-Domingo
AU - Andrea Ruiz-Ferrando
AU - Adam H Clark
AU - Stephan Pollitt
AU - Sharon Mitchell
AU - Javier Pérez-Ramírez
AU - Núria López
PY - 2026
DA - 2026/10/18
JO - Advanced science (Weinheim, Baden-Wurttemberg, Germany)
DO - 10.1002/advs.78226
UR - https://europepmc.org/article/MED/42893450
AB - 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 chemical environments remains challenging. Here, we introduce AI-EXAFS, a physics-informed graph neural network that predicts spectra directly from three-dimensional atomic coordinates. Combining an absorber-centered graph representation with distance-modulated message passing and an EXAFS-specific training objective enables direct learning of structure-spectrum relationships without predefined structural descriptors or manual parameter selection. Trained and evaluated across a dataset of 85,991 transition-metal complexes with FEFF-generated spectra, AI-EXAFS achieves average held-out test-set R-factors of 0.043 in k -space and 0.019 in R -space, demonstrating accurate structure-to-spectrum prediction across a broad range of local chemical environments. AI-EXAFS establishes a scalable structure-to-spectrum framework for EXAFS, enabling direct inference across large sets of candidate structures and providing a foundation for data-driven prediction across increasingly complex and experimentally relevant environments.
SN - 2198-3844
ER -