Decoding the Pair Distribution Function of Uranium in Molten Fluoride Salts from X-Ray Absorption Spectroscopy Data by Machine Learning

Thermal properties of actinides in molten salts are linked to the strongly disordered local environment of actinide ions. We illustrate both the limitations of the commonly used fitting method for analysis of extended X-ray absorption fine structure (EXAFS) spectra in molten UF4 and a possible solution using an "objective neural network-EXAFS" (ONNE) method. ONNE provides both extraction of the pair distribution function, as validated by its application to the EXAFS spectra calculated on molecular dynamics trajectory, and the EXAFS data reconstruction. The ONNE analysis of the molten UF4 has revealed reduction of the first nearest neighbor U–F coordination number, expansion of the U–F bond length, and smaller contribution to the second shell compared to current molecular dynamics models. This method is therefore an attractive alternative to conventional EXAFS analysis and molecular dynamics simulations for studies of disordered environment of actinides in molten salts.