Mapping Pair Distribution Functions of Zirconium in NaF-ZrF4 Molten Salt from X-ray Absorption Spectroscopy Data via Machine Learning

Molten salts are central to advanced clean energy technologies yet elucidating their highly disordered local atomic environments is especially challenging. In this work, we apply a model-agnostic machine learning method, to invert experimental extended X-ray Absorption Fine Structure (EXAFS) of NaF–ZrF₄ (53–47 mol%) directly into pair distribution functions (PDFs). Thereby circumventing the difficulty of conventional analysis that requires structural assumptions or molecular dynamics (MD) simulations, which are limited by the accuracy of force fields and computational expense. Our workflow generates training data from diverse statistical PDFs, enabling robust learning across varied configurations. Our results resolve key structural features for local atomic coordination of Zr directly from experiments, that align with previous MD and X-ray scattering studies providing a rapid and scalable tool for characterizing disordered systems. This workflow offers a pathway to in-situ and operando monitoring of atomic structure evolution for various material classes.