An explainable “family bucket” model for simultaneous prediction of K-edge XANES for multiple light transition metals

Abstract

X-ray near-edge structure (XANES) is a crucial bridge between the local structures and chemical properties of materials. Although there have been a number of studies devoted to the development of predictive K-edge XANES spectral models, existing methods are usually still limited to separate modeling for a specific absorbing element. Currently, there is a lack of a K-edge XANES spectra prediction model that can be broadly applied to a wide range of elements, which would enable data dispersed in terms of absorbing elements to be integrated and well utilized. In this work, we develop an innovative “family bucket” model based on a multi-head graph attention convolutional neural network by combining a multi-element mixed dataset with a crystal topology approach for the localized environment. The model is able to predict the K-edge XANES spectra for a wide range of light transition metals (periods 3 and 4) simultaneously. Moreover, it is demonstrated that the training scheme not only improves the accuracy of the model but also the efficiency of its training. In terms of interpretability, several fascinating insights were gained, uncovering the underlying mechanisms of the model for spectral prediction. We investigate the collective behavior of neurons by employing a range of responses to different samples as descriptive features. Notably, the analysis revealed that neurons in the neural network exhibit functional differentiation characteristics analogous to Brodmann areas in the cerebral cortex. The homology of data analysis indicates that the mutual learning of samples from different absorbing elements is occurring between close elements of the same period. Additionally, the attention scores of the samples are determined by both the absorbing element and its surrounding atomic environment. In conclusion, this research advances the understanding of the relationship between XANES spectra and material structures while providing valuable insights into neural networks, enhancing the comprehension of neuronal behavior.