(777j) To Address Surface Reaction Network Complexity Using Machine Learning, Scaling Relations, and DFT Calculations

Surface reaction networks involving hydrocarbons exhibit enormous complexity with hundreds of species and thousands of reactions for all but the very simplest of chemistries. Among fields that regularly confront this challenge, such as catalysis, metabolic engineering, and combustion, catalysis is unique in that the most likely reaction pathway varies for each surface and active site and must be re-determined in every instance proving a fundamental bottleneck in new catalyst discovery. We present a framework to efficiently explore a reaction network from scratch using a hierarchy of well-established predictive methods such as group additivity and linear scaling relations as guides to traditional density functional theory (DFT) approaches. Unexplored regions of the reaction network are predicted using machine learning methods, and the mechanism is progressively refined using uncertainty-aware DFT calculations on the parts of the network that are likely rate-limiting. Applying these methods to the reaction of syngas on rhodium (111), we identify the most likely reaction mechanism along with bounds on the likelihood unexplored reaction pathways do not contribute, using DFT on just a fraction of the possible intermediates. Propagating uncertainty throughout this process allows for estimates that the final mechanism is complete given DFT-level uncertainty and measurements on only a subset of the entire network.