Machine-Learning Based Prediction of First-Principles XANES Spectra for Amorphous Materials

In recent years, x-ray absorption spectroscopy (XAS) has become increasingly important for structural characterization of materials. Especially, x-ray absorption near edge structure (XANES) can provide sensitive information on chemical bonds, valence states, and coordination. Conventional interpretation of XANES spectra is based on the so-called “fingerprint” approach in which an experimental spectrum of interest is compared with that of a reference crystalline material. When there are not enough reference spectra, reliable simulations of XANES spectra are necessary. We have already added the computational code of XANES [1] to Quantum Materials Simulator (QMAS) code [2] based on the density functional theory (DFT) within the projector augmented-wave (PAW) method, and applied it to the elucidations of amorphous structures [1] and chemical reactions at interfaces [3]. However, the computational cost of XANES simulation becomes an important issue when we deal with amorphous materials.<br/>In this study, we developed an efficient scheme based on machine learning to predict the theoretical XANES spectra obtained by DFT calculations using only the information of atomic configurations. Our scheme was based on the linear regression model of the structural descriptor.We present a comprehensive prediction of XANES spectra based on atom-centered symmetry functions (ACSF), smooth overlap of atomic positions (SOAP), local many-body tensor representation (LMBTR) and spectral neighbor analysis potential (SNAP). As a test case, we chose amorphous SiO material used as the negative electrode of rechargeable lithium-ion batteries (LIBs), in which the valence states of Si are distributed between 0 and +4. As a result, we could achieve enough prediction accuracy for Si K-edge XANES spectra. Furthermore, we proposed an efficient scheme for the compression of XANES spectral data and the efficient sampling of training data with the active learning approach.<br/><br/>References: [1] T. Tamura et al., Phys. Rev. B 85, 205210 (2012). [2] http://qmas.jp [3] T. Tamura et al., Phys. Rev. B 96, 035107 (2017).