Inverse Analysis of Near-edge Spectra: Toward the Prediction of Chemical Bonding and Atomic Structure.

Near-edge structure of core-loss spectra, including electron energy-loss near-edge structure (ELNES) measured by transmission electron microscopy and X-ray absorption near-edge structure (XANES) measured by synchrotron radiation, provides powerful probes of chemical bonding and atomic coordination in materials. Conventional analysis of ELNES/XANES relies on a forward approach, where candidate atomic structures are assumed and theoretical spectra are compared with experimental data. While powerful, this strategy becomes increasingly inefficient for complex systems with large structural degrees of freedom, or for materials where the structural space is vast and poorly constrained. Here we review recent advances in machine learning-based inverse analysis that directly extract ground-state electronic structure, including partial density of states (PDOS) as a descriptor of chemical bonding, and atomic configurations from near-edge structure of core-loss spectra, bypassing the need for trial-and-error structural modeling. We demonstrate two major capabilities using both simulated and experimental spectra: 1) prediction of PDOS including both occupied and unoccupied states from a single ELNES/XANES spectrum, enabling direct extraction of valence band structure and chemical bonding information; and 2) conditional generation of three-dimensional atomic coordinates from a single ELNES/XANES spectrum. We further examine the robustness of these approaches against spectral noise and energy range variations, and confirm their applicability to experimental spectra. Through systematic validation, we show that deep neural networks can learn the spectroscopy-structure relationships applicable to both ELNES and XANES measurements. This unified inverse analysis framework offers a direct and scalable pathway for high-throughput materials characterization across multiple spectroscopic platforms, enabling the extraction of quantitative chemical bonding and coordination information from near-edge structure of core-loss spectra.