Advancing the Analysis of X-ray Absorption Spectroscopy using Deep Neural Networks

Abstract

X-ray absorption spectroscopy (XAS) is a core analytical technique that can provide users with highly detailed information about the electronic and geometric structure of matter. Its ability to be applied under a wide variety of different conditions means it has had a strong impact across the physical and natural sciences. However, one of the challenges of the technique is that difficult computational calculations are often required to extract the detailed information from the XAS spectra and, in complicated systems, e.g. in operating catalysts, batteries, and temporally-evolving systems, these can be particularly challenging. Due to the complexity and computational resource requirements, many users are unable to access the wealth of valuable information contained within their XAS spectra. In this Thesis, I describe my work developing a deep neural network (DNN) for X-ray absorption near edge (XANES) spectrum predictions. The DNN is a multi-layer perceptron (MLP) ML model which aims at minimising the mean-squared error (MSE) between predicted and calculated (theoretical) XANES spectra using featurised structures and the corresponding theoretical XANES spectra as input. After learning how to map the relationship between the local environment of an absorbing atom and the corresponding XANES spectrum from a reference dataset, the DNN is able to predict XANES spectra of materials outside of the scope of that dataset. The Thesis initially focuses upon how the data representation choices, i.e. the featurisation, affects the accuracy of the DNN at the Fe K-edge. Once optimised, the DNN is able to predict peak positions with sub-eV accuracy and peak intensities with errors over an order of magnitude smaller than the spectral variations in the reference dataset. Subsequently, the DNN is extended to the Co K-edge and applied to interpret T-jump-pump/X-rayprobe experiments investigating ligand exchange in an aqueous Co complex. The DNN greatly facilitates the analysis in this case, since it can quickly and cost-effectively predict the XANES spectra of thousands of geometric configurations sampled from ab initio molecular dynamics (MD), describing the disorder in the system. The final Chapter addresses the challenge of edge shifts: changes in the energetic position of the X-ray absorption edge in the XANES spectrum arising from changes in the electronic structure of the system(s) under study. It is demonstrated that, although the developed DNN model is more than satisfactory, improving it further would nonetheless require using significantly higher levels of theory to build the reference dataset, which would make developing sufficiently large reference datasets a significant challenge moving forwards.