First shell EXAFS data analysis of nanocatalysts via neural networks

Understanding the mechanisms of work of nanoparticle catalysts requires the knowledge of their structural and electronic descriptors, often measured in operando X-ray absorption fine structure (XAFS) spectroscopy experiments. We introduce a neural-network-based framework for rapidly mapping the extended XAFS (EXAFS) spectra onto structural parameters as an alternative to the commonly used non-linear least-squares fitting approaches. Our method leverages a multilayer perceptron trained on theoretical EXAFS and validated against theoretical test data and experimental spectra of frequently used nanoparticle types. The network helps lower the correlation between parameters, achieves high accuracy in the presence of noise and glitches, and can provide real-time parameter predictions with minimal user intervention. Parameter uncertainties are estimated as well. This method can be readily integrated into beamline pipelines or laboratory data analysis workflow and has the potential to accelerate high-throughput catalyst characterization and testing.