X-ray absorption near-edge structure (XANES) spectroscopy is a premier tool for probing material structure at the atomic level. XANES encodes structural information through the constructive and destructive interference of photoelectrons, and it is well-known that variations in local structure can lead to well-defined spectral trends, e.g., the relationship between the charge state and the edge position. Developing a further understanding of structure-spectrum relationships can help us to gain more insights from XANES measurements. However, the generalization of these relationships to diverse material structures is an outstanding challenge. A disentangled structure representation is desired, in which each structural descriptor can be varied independently. Since commonly considered structural motifs are strongly coupled to one another (e.g., coordination informs charge state to some degree), such a representation remains to be established. To this end, we represent spectra using a semi-supervised machine learning framework. We transformed spectra to a low dimensional, numerically orthogonalized latent space using an adversarial autoencoder (AAE), an analog of a variational autoencoder. In addition, we developed a novel constraint to rank each dimension of the latent space with respect to each structural descriptor. Our AAE decoder effectively reveals the variation trends of simulated Cu K-edge XANES spectra by utilizing this reduced-dimensional representation (Fig. 1). All trends can be explained by multiple scattering theory and can be applied empirically to better understand XANES spectroscopic results.

Revealing trends in X-ray absorption spectra through a disentangled latent space

Fall 2022 abstracts, posters, and presentations will launch in mid-September.

ACS Fall 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.