Compact representation of the local atomic structure of matter for machine learning in XANES spectroscopy data processing

A method for representing data on the local structure of atoms in the form of histograms of paired radial distribution functions is proposed. This method is used to construct a structure descriptor needed to determine the structure of materials using machine learning and artificial intelligence techniques. A special feature of the method is the use of two sets of paired radial distribution functions simultaneously: for pairs of all types of atoms and for pairs with a selected absorbing atom. The developed approach was tested on the problem of determining the local atomic structure of the environment of the silver color center in sodium silicate glasses using data from X-ray absorption near-edge structure for the Ag K-edge. The information content of the proposed structure descriptor is demonstrated by the ability to reconstruct the three-dimensional structure of a silver color center model from the corresponding pairwise distance histograms. Using several machine learning methods, it was shown that the proposed descriptor allows to achieve high-quality reproduction (mean square error ~10^–3) of X-ray absorption near-edge structure spectra for silver color centers in glass, which makes it possible to reduce the time for calculating X-ray absorption near-edge structure spectra by 4 orders of magnitude. The resulting machine learning model allows us to establish a fundamental connection between the atomic structure of silver color centers in glasses and the Ag X-ray absorption near-edge structure spectrum, which is necessary for determining the structure of glasses.