Machine Learning for Atomic Simulation and Activity Prediction in Heterogeneous Catalysis: Current Status and Future

Abstract

Heterogeneous catalysis, for its industrial importance and great complexity in structure, has long been the testing ground of new characterization techniques. Machine learning (ML) as a starring tool in data science brings new opportunities for chemists to interpret, simulate, and predict complex reactions in heterogeneous catalysis. Here we review the current status of ML methods and applications in heterogeneous catalysis by following two main streams: the top-down approach by learning experiment data and the bottom-up approach for making predictions from first-principles, which differ in the data source. We focus more on the latter, where ML interacts intimately with first-principles calculations for predicting the key properties (e.g., molecular adsorption energy) and evaluating potential energy surface (PES) to expedite the atomic simulation. The ML-based PES exploration represents the top gear that can largely replace the traditional roles of first-principles calculations for structure determination and activity evaluation but requires efficient methods for data set generation, sensitive structure descriptors to discriminate structures, and iterative self-learning to refine the ML potential. We illustrate these key ingredients of ML-based atomic simulation using the SSW-NN method developed by our group as the example. Three cases of SSW-NN application are presented to elaborate how ML can expedite the material and reaction simulation and lead to new findings on catalyst structure and reaction channels. The future directions of ML-based applications in heterogeneous catalysis are also discussed.