(290b) Unsupervised Machine Learning to Extract the Electronic and Chemical Properties of Alloy and Metal Oxide Surfaces

Catalyst design greatly benefits from the identification of the atomic, chemical, and electronic descriptors that map to catalytic figures of merit (e.g., stability, activity, selectivity). One important and widely used class of descriptors in catalysis is electronic-structure descriptors, which have been extensively used as site-specific chemisorption descriptors for a variety of different materials. Nonetheless, there exists no unified framework to systematically identify electronic-structure descriptors. We demonstrate a simple, yet informative unsupervised machine learning strategy using principal component analysis (PCA) that rapidly identifies accurate and interpretable electronic-structure descriptors of catalysts. We first demonstrate the approach for transition metal alloys and show that the principal component descriptors yield accurate ML models of chemisorption of adsorbates (e.g., O*, C*, H*, N*) that outperform ML models built using traditional electronic structure descriptors. We use an ML model interpretation method called partial dependence plots to visualize the relationships between the identified electronic-structure descriptors and the chemisorption trends. Importantly, we are able to interpret the meaning of these descriptors and connect them to the alloys’ geometric structure and composition using signal reconstruction methods from the cognitive neuroscience and machine intelligence literature. Interpretation of the PCs suggests that they capture physical trends that are consistent with prior efforts to develop electronic-structure chemisorption descriptors for alloys. We show that the approach is general by finding descriptors of surface oxygen reactivity for metals and metal oxides. Ultimately, this unsupervised learning approach provides a pathway for finding electronic-structure descriptors for catalytic materials that readily connect to geometric structure and composition.