Measuring the density of states in warm dense matter: a machine learning-based approach to the deconvolution of resonant inelastic x-ray scattering

Mapping the electronic structure of materials at extreme conditions is a challenge of high importance in modern physics and materials chemistry. Resonant inelastic x-ray scattering (RIXS) provides a uniquely powerful means of recovering the density of states (DOS) of warm dense matter (WDM). Recently, we demonstrated that by correlating the measured RIXS signal with the unique spectral profile of each incoming SASE XFEL pulse, it is possible to perform a deconvolution that improves the energy resolution of the density of states beyond the intrinsic SASE bandwidth — without sacrificing the SNR of the spectra. This was achieved using a neural network framework trained on measured XFEL and RIXS spectra. In this talk, we present our machine learning-based deconvolution framework, designed to robustly extract the DOS from K-edge and L-edge RIXS data. The method is demonstrated on experimental measurements of L-edge RIXS in x-ray heated gadolinium, used here as a test system due to its complex electronic structure and sensitivity to thermal perturbations. Additionally, we showcase promising results from dynamically compressed iron oxides, where DOS were extracted from K-edge RIXS spectra measured under laser-driven compression. These results demonstrate that our approach enables robust DOS reconstruction from noisy spectra, offering an attractive route towards meaningful measurements of electronic structure in WDM.