Local structural information about nanoparticles during catalysis can be measured quantitatively by X-ray near edge absorption spectra (XANES) studies using Neural Network methods. However, the extent of information encoded in XANES spectra is not known a priori. We developed an approach to solving the information content problem: Latent Space Analysis of Spectra (LSAS).

 

LSAS approach compresses the information in the input (spectrum) to a much smaller dimensionality of the latent space, which is then used to analyze and extract structural descriptors. Our information centric approach identifies and extracts the key descriptors in the hydrogen disassociation and Pd hydride formation reaction. Such an unsupervised approach was demonstrated to be crucial for learning factors behind the real-time changes in catalytic activity.

For more details, read our manuscript: Latent Representation Learning for Structural Characterization of Catalysts, Phys. Chem. Lett. , 12, 2021, DOI: 10.1021/acs.jpclett.0c03792