Learning EXAFS from atomic structure through physics-informed machine learning

Extended X-ray absorption fine structure (EXAFS) provides element-specific access to local atomic environments and is widely used to relate structure and reactivity across chemical systems. However, quantitative EXAFS interpretation still relies on manually constructed structural models and extensive parameter tuning, creating a growing bottleneck as experimental datasets increase in size and complexity. Addressing this bottleneck requires a direct and systematic mapping between atomic structure and EXAFS response. Here we introduce AI-EXAFS, a physics-informed graph neural network that predicts full EXAFS spectra directly from three-dimensional atomic coordinates. By formulating the learning problem around the physical principles governing EXAFS signal formation, the model learns transferable structure–spectrum relationships and eliminates the need for user-defined parameter selection at inference. Trained on 86,000 transition-metal complexes, AI-EXAFS reproduces reference theoretical spectra with accuracy consistent with established EXAFS analysis practice and generalizes to experimentally relevant systems, including platinum single-atom catalysts. AI-EXAFS provides an accurate and readily deployable forward model for EXAFS, enabling standardized first-pass structural screening and offering a scalable foundation for future extensions toward more realistic and data-rich EXAFS analysis.