Machine Learning-Driven Deconvolution of Mixed X-ray Absorption Spectra

X-ray absorption spectroscopy (XAS) plays a crucial role in determining the oxidation states, local atomic structures, and electronic configurations of materials. However, the interpretation of complex spectra with multiple ion sites and varying local chemical environments can be labor-intensive and requires specialized expertise, presenting challenges for researchers. To address these issues, this study employs machine learning (ML) techniques to analyze mixed XAS data, with a particular focus on the cobalt (Co) element. We constructed a simulated data set of L-edge spectra, incorporating various combinations and energy shifts produced from multiplet calculations. Through a systematic evaluation of dimensionality reduction and ML methodologies, we identified a combination that achieved both high accuracy and efficiency in classification tasks. Our approach involved isolating individual components from the mixed spectra and implementing an automatic fitting algorithm to extract critical parameters, such as component proportions and Gaussian broadening. The fitting results closely aligned with the input spectra, thereby validating our methodology. This innovative process facilitates the determination of valence states, local crystal field parameters, and composition ratios, significantly enhancing the efficiency of spectral analysis.