Heterogeneous catalysis plays a major role in many different industrial chemical processes. It is impractical to run expensive and time-consuming experiments at all conditions of interest so typically it is most practical to construct a microkinetic model resolving the elementary reaction steps to help design, screen, and optimize these processes. Microkinetic models require estimates of many thermodynamic and kinetic parameters. The presence of nearby co-adsorbates on catalytic surfaces, however, can dramatically affect reaction energetics making the kinetic parameters dependent on adsorbate coverages. However, the space of possible co-adsorbed configurations on a surface is combinatorically large causing enumerative quantum chemistry approaches to range from expensive to intractable. We present a workflow tool for efficient automated calculation of coverage dependent thermodynamic and kinetic parameters within our software Pynta. This workflow fuses two synergistic active learning (AL) frameworks. The outer AL framework uses MACE machine learning potentials to compute configurations from the inner AL framework, automatically correcting by training on DFT calculations at select configurations. The inner AL framework uses PySIDT interpretable machine learning to learn the co-adsorbed association energy and its uncertainty as a function of the surface graph (sites, adsorbate atoms, and bonds) using uncertainty- and association-energy-based weighting to select new configurations. These configurations are automatically calculated using the outer AL framework with MACE and used to retrain the SIDT. We validate this dual AL workflow against enumerative methods and our original single AL workflow (only using the inner loop). We then leverage the dual AL workflow to develop an SIDT model that can accurately predict coverage dependence for a wide range of Pt111 adsorbates and reactions.
