(306d) Resolving Coverage Dependence By Combining Automated Quantum Chemistry Workflows with Interpretable Machine Learning

Heterogeneous catalysis is incredibly important in many different chemical processes. Running experiments on these systems at many different conditions can be incredibly expensive and time consuming so it is usually most practical to build microkinetic models to describe the associated chemistry at an elementary level. However, these microkinetic models require estimates of many thermodynamic and kinetic parameters. In these systems the presence of nearby co-adsorbates can dramatically affect energetics causing these parameters to be dependent on species coverage fractions. In this work we combine Pynta, our toolkit for automating quantum chemistry calculations, with PySIDT, our software for interpretable machine learning using subgraph isomorphic decision trees (SIDT).

In our first thrust, we considered a diverse set of 9 reactions on Cu111 with 4 possible co-adsorbates. We used Pynta to generate and optimize 2400 randomly selected surface configurations involving the associated adsorbates, co-adsorbates and transition states that pass a pair-wise filter. We then trained an SIDT classifier and a regular SIDT to predict the stability and (if stable) the association energy of arbitrary configurations. This SIDT architecture allows us to estimate energetic coverage corrections for arbitrary combinations of adsorbates and transition states and thus barriers on Cu111. Additionally, we develop techniques to extend our model on Cu111 to estimate coverage dependence parameters on other simple surfaces.

In our second thrust we utilized transfer learning and active learning schemes to accurately calculate specific coverage dependent parameters on scales as small as individual reactions (and associated reactants/products) with a minimal number of calculations. We validate our active learning technique against a comprehensive enumeration and calculation of configurations for OH* dissociation on Ru0001 for both H* and O* co-adsorbates and demonstrate our technique on a small set of reactions.