(532ed) Theory-Infused Neural Network for Interpretable D-Band Moments Prediction

For heterogeneous catalytic reactions, higher reactivity and selectivity are always advantageous. The adsorption of reactants or their fragments at solid catalyst surfaces is the fundamental step of heterogeneous catalytic reactions. Electronic structures of the substrate determine the strength of adsorption, and thus the activity and selectivity of the reaction. The well-established d-band theory explains trends between electronic structures and adsorption properties. Ab-initio calculations have helped people discover new catalysts with kinetics-favorable electronic structures. However, exploring the enormous size of the accessible design space remains a daunting task. To simplify the problem, d-band moments are often used as descriptors to the electronic structure and explain trends in activity and selectivity. In recent years, ML has emerged as an alternative method for predicting d-band moments of catalytic sites. However, there are pros and cons. Complex network structures make ML algorithms a black box, which limits its interpretability and transferability to unseen systems. In this study, we developed a theory-infused neural network (TinNet) [1], which integrates deep learning algorithms with the tight-binding theory to predict d-band moments, e.g. d-band center (1^st order moment) and d-band width (2^nd order moment). The machine-learned coordination number not only bridges the relationship between the substrate structure and its d-band moments and makes TinNet inherently interpretable, but also helps us to interpret the trend of d-band moments and to design new materials with desired properties.

[1] Wang, S. H., Pillai, H. S., Wang, S., Achenie, L. E., and Xin, H. Nature communications, 12(1), 1-9 (2021).