Advancing Catalysis Theory with Theory-infused Deep Learning

Abstract

Despite recent advances of data acquisition and algorithms development, machine learning (ML) faces tremendous challenges to being adopted in practical catalyst design, largely due to its limited generalizability and poor explainability. We developed a theory-infused neural network (TinNet) approach that integrates deep learning algorithms with catalysis theory. Incorporation of scientific knowledge of physical interactions into learning from data opens up new avenues for interpretable discovery of novel motifs with desired catalytic properties. The TinNet framework offers a robust platform for transforming ab initio data into physicochemical insights, enabling the design of novel catalytic materials. Its architecture, deeply rooted in the physics of electronic interactions, transcends algorithmic boundaries. TinNet not only sheds light on the fundamental characteristics of active sites but also enhances prediction accuracy in harmony with the physical principles governing catalytic surfaces. In this dissertation, We will highlight the role of TinNet in the rapid development of new catalytic materials, emphasizing its crucial contribution to sustainable chemical processes. The framework is particularly flexible at predicting surface reactivity, electronic structures, and cohesive energies, thus guiding the design and synthesis of an array of metallic systems, from single-atom alloys to complex high-entropy alloys. TinNet's unique fusion of theory and data-driven algorithms signifies a transformative step in the field of heterogeneous catalysis, one that could redefine the future of material design and sustainability.