日本ゼオライト学会　刊行物 Publication of Japan Zeolite Association

Characterizing the local atomic structure of amorphous materials is a significant challenge due to the absence of long-range order. While X-ray Absorption Near Edge Structure (XANES) spectroscopy is sensitive to the local environment, interpreting its spectra is often complex and computationally intensive. This study presents an advanced analytical framework that integrates simulation and machine learning to overcome these limitations, using amorphous silicon suboxide (SiO_x) as a model system. We first generated a diverse range of amorphous SiO_x structures with varying compositions using molecular dynamics simulations. Subsequently, a large-scale dataset of over 5,000 Si K-edge XANES spectra was created from these structures via first-principles calculations. This comprehensive dataset was then used to train an artificial neural network (ANN) to solve the inverse problem: predicting local structural parameters directly from a XANES spectrum. The resulting model demonstrated high accuracy in determining the atomic valence and radial distribution function (RDF) for individual Si atoms. More importantly, the ANN showed excellent generalization performance, successfully predicting the average valence and RDF from averaged spectra, which correspond to experimentally measured data. This measurement informatics approach enables rapid and quantitative analysis of the local structure in amorphous SiO_x, establishing a robust inverse model applicable to material design. The proposed framework is not limited to SiO_x and holds significant potential for accelerating the structural analysis and rational design of a wide variety of amorphous materials crucial to industrial applications such as battery electrodes, barrier coatings, and optical components.