Unsupervised learning of XAS spectra to enhance machine learning techniques

Core-level spectroscopy, including X-ray absorption spectra (XAS) and electron energy loss spectra (EELS), can be simulated using DFT-based approaches. However, a given measurement at finite temperature, even for a well-defined, unambiguous system, may not be well reproduced using one atomic structure as input, but rather needs an appropriate sampling of the structural space. Predicting spectra relying uniquely on ab-initio techniques is therefore time-consuming and resource intensive. In this context, there is increasing interest in the development of machine learning models to understand the relationship between atomic structures and XAS spectra. Due to the complex and non-linear nature of the problem, this challenge is far from being solved. In this work, we present the application of unsupervised learning algorithms and clustering techniques to assist the task of fitting XAS spectra of isolated molecules. The goal is dual: we want to learn what are the characteristic features of particular transition states, and we aim to employ machine learning to fit the XAS spectra using only easily obtainable structural features. The end product is a pipeline that can predict XAS spectra given only the atomic positions as an input.