Automatic Calculation of Kinetic Coverage Dependence using Pynta: Combining Active Learning and Automated Quantum Chemistry

Heterogeneous catalysts play a role in most industrial reaction processes. However, adsorbates on catalytic surfaces can be configured in combinatorially many ways and their inter-adsorbate interactions can have an enormous impact on the energetics of reactions. This makes it incredibly challenging to accurately compute coverage dependent kinetic parameters that account for these interactions without expensive enumerative calculations. In this work we present an active learning framework fusing Subgraph Isomorphic Decision Tree (SIDT) based machine learning with automated quantum chemistry calculations to efficiently iteratively learn the coverage dependence of adsorbates and transition states (TS) that we have implemented within our software Pynta. This framework exploits the internal readability of the SIDT using the node uncertainties and occurrence rates in the configurational space to select sets of calculations to run that are expected to best minimize the uncertainties in the SIDT each iteration. We validate our approach against enumerative DFT and cluster expansion approaches for the HO* dissociation reactions on Ru(0001) with O* as the co-adsorbate showing we are able to reach a mean absolute error of ~0.05 eV relative to DFT in only 43 DFT calculations, only about ~3x the cost of the running isolated calculations for the same configurations. Further, we demonstrate our approach on ammonia decomposition on Pt(111) with N* as a co-adsorbate and for the initial reactions for Fischer-Tropsch synthesis on Ru(0001) with OC* as the co-adsorbate. We show the importance for including these coverage dependent corrections in the ammonia decomposition case and examine the SIDT-predicted minimum energy configurations for O* on Ru(0001) in detail. Lastly, we discuss how to get the most accuracy out of the fewest calculations within this framework.