Teaching Core-Hole Spectroscopy to a Deep Neural Network

Deep neural networks (DNNs) – multilayer machine-learning models that are able to extract and learn patterns represented in data without hand-coded heuristics – are transforming what we can do, and the way we do it, across the physical sciences.

XANESNET is a DNN for instantaneous simulations of X-ray absorption spectra (XAS); the XANESNET Project is about addressing the challenge of delivering detailed, high- level theoretical simulations that can capture the complex underlying physics of these experiments but that are – at the same time – fast, affordable, and accessible enough to appeal to beamline users. Using DNNs like XANESNET, we can reduce the time taken to simulate XAS from hours/days to a fraction of a second, democratise data analysis, and enable beamline users to better plan beamtime allocations by facilitating ‘on-the-fly’, ‘real-time’ analysis. We’ve already deployed XANESNET in the practical

arena to take on open questions in physical and materials chemistry with success.

This talk will showcase how XANESNET delivers ‘black-box’, qualitative predictions of XAS at the transition metal K-edges using nothing more than the local geometries of arbitrary absorption sites, bypassing time- and resource-intensive quantum-chemical calculations and reproducing peak positions to sub-eV accuracy wrt. reference XAS data across the transition metal K-edges. Lifting the lid on XANESNET reveals that the predictions it produces are informed by the expected physics, giving us an insight into how, what, and why the model learns.