(684f) Theory-Infused Neural Network for Interpretable Reactivity Prediction | AIChE

(684f) Theory-Infused Neural Network for Interpretable Reactivity Prediction


Wang, S. H. - Presenter, Virginia Tech
Pillai, H., Virginia Tech
Wang, S., Virginia Polytechnic Institute and State University
Achenie, L. E. K., National Science Foundation
Xin, H., Virginia Tech
The adsorption of reactants or their fragments at solid catalyst surfaces is the fundamental step and its adsorption energy often serves as reactivity descriptors. Quantum chemical calculation is a great tool for discovering site structures with kinetics-favorable descriptor values, but it remains a daunting task for exploring the enormous size of the accessible design space. In recent years, machine learning (ML) has become an alternative approach to predicting the chemical reactivity of catalytic sites. However, most of the modern machine learning models, particularly deep learning, are black boxes. The nature of their mathematical construction leads to limited extrapolation and poor transferability to unseen systems, in spite of learning some correlations that look good on both training and test datasets. In this study, we developed a theory-infused neural network (TinNet), which integrates deep learning algorithms with the well-established d-band theory of chemisorption for reactivity prediction of transition-metal surfaces [1]. The well-established d-band theory makes TinNet inherently interpretable. A simple adsorbate, *OH, at the active site is used as an example to demonstrate that the TinNet is on par with purely data-driven ML methods in prediction performance. The key feature is that TinNet not only opens the black box of ML and reveals the nature of chemical bonding, but also opens up a new way for ML to discover new motifs with desired catalytic properties.

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