DataXAS

Motivation

Investigation of catalytic active sites is challenging, because they are often made up of only a few atoms amongst a majority of atoms of the same elements that are spectator species and often requires time-resolved synchrotron-based X-ray tools to identify them. Advanced synchrotron spectroscopy techniques give direct access to geometric structural and electronic characteristics of materials under operation. Within the broad ensemble of synchrotron techniques, X-ray absorption spectroscopy (XAS) is particularly well suited to the study of catalytic active sites elucidating, uniquely with element specificity, the local atomic configuration and electronic state of catalyst actives sites in operando conditions.

‍

‍

The application of state-of-the-art data science techniques to catalysis is still in its infancy and is an important short-, medium- and long-term goal in catalysis. This project aims at providing a fundamental step towards this goal by establishing robust and self-consistent workflows for data collection, processing, and analysis, applied to high-throughput spectroscopic data. The general objective of this project is to develop a data-driven methodology for data handling and processing of XAS data.

Proposed Approach / Solution

For the first task of denoising [2], we propose an efficient denoising approach of noisy XAS spectra based on careful Gaussian Processes (GP) regression with adaptive length scales. We propose using a warping procedure to preprocess the data and make the signals more stationary, which already improves all the classical denoising methods, see Figure 2. For the data with higher levels of noise, we propose GP regression for denoising of the preliminary warped signals.

‍

In the second task of blind signal separation [3], we encounter the challenge of rotational ambiguity inherent in the classical Nonnegative Matrix Factorization (NMF) formulation. In particular, like shown on Figure 3, multiple pure component sets can lead to the same quality of the NMF representation. We propose to address this issue by regularizing the NMF formulation using a database of real pure elements.

‍

‍

Thirdly, using the labeled dataset of real spectra measurements, we use a classification approach [4], [5], [6], [7], for X-ray absorption near-edge structure (XANES) spectra, possibly NN based, to determine the key geometrical and structural features of the investigated chemical compounds directly from the spectrum.

Impact

Impact. The combined denoising, classification, and decomposition techniques will help domain scientists determine the particular components of the time-resolved XAS spectra with better certainty and verify them by checking fewer potential candidates for the components hence facilitating to streamline the process of compound characterization.