From database to prediction: Machine learning for 5-f elements coordination using actinide x-ray experimental spectra (AXES) collection

Abstract

The Actinide X-ray Experimental Spectra (AXES) database has been presented as a comprehensive database of x-ray absorption spectroscopy (XAS) spectra. It is the largest database of experimental spectra of actinides with a diverse range of measurement techniques (standard resolution x-ray absorption near edge structure and high energy resolution fluorescence detected XAS), absorption edges (L3, M4, M5, and less common edges), and absorber types (Th, U, Np, Pu, and Am). The spectra have been aligned and normalized to facilitate further analysis, while the original unprocessed data have been retained for reference. Coordination information derived from the spectra and their corresponding structures has been integrated into a convolutional neural network to construct a structural property model capable of predicting the presence or absence of uranium atoms in various coordination environments. The model’s predictive accuracy and reliability can be further enhanced by expanding the AXES database or employing transfer learning techniques. In this study, the model has been pre-trained using Fe K-edge XAS spectra. In addition, key spectral regions critical for coordination number prediction have been identified using the Shapley Additive Explanations (SHAP) approach. The SHAP-value distribution indicates that spectral features associated with six-coordination uranium primarily appear in the edge and post-edge regions, while those linked to eight-coordination uranium predominantly influence only the edge shape. This analysis underscores the model's potential for advancing actinide coordination studies.