Abstract
Microkinetic models (MKMs) are widely used within the computational heterogeneous catalysis community to investigate complex reaction mechanisms, to rationalize experimental trends, and to accelerate the rational design of novel catalysts. However, constructing these models requires computationally expensive and manually tedious density functional theory (DFT) calculations for identifying transition states for each elementary reaction within the MKM. To address these challenges, we demonstrate a novel protocol that uses the open-source kinetics workflow tool Pynta to automate the iterative training of a reactive machine learning potential (rMLP). Specifically, using the silver-catalyzed partial oxidation of methanol as a prototypical example, we first demonstrate our workflow by training an rMLP to accelerate the parallel calculation of DFT-quality transition states for all 53 reactions, achieving a 7× speedup compared to a DFT-only strategy. Detailed analysis of our training curriculum reveals the shortcomings of using an adaptive sampling scheme with a single rMLP model to describe all reactions within the MKM simultaneously. We show that these limitations can be overcome using a balanced "reaction class" approach that uses multiple rMLP models, each describing a single class of similar transition states. Finally, we demonstrate that our Pynta-based workflow is also compatible with large pretrained foundational models. For example, by fine-tuning a top-performing graph neural network potential trained on the OC20 dataset, we observe an impressive 20× speedup with an 89% success rate in identifying transition states. This work highlights the synergistic potential of integrating automated tools with machine learning to advance catalysis research.