Fusing Delta Learning and Machine Learned Interatomic Potentials for Efficient High Throughput Calculation of Surface Kinetic Parameters

Heterogeneous catalysis is critical in most industrial chemical processes. Microkinetic models can be used to greatly facilitate optimization of catalyst design and process conditions, but require thermochemical and kinetic parameters for all relevant species 1 and reactions. Our software Pynta enables fully automated calculation of thermochemical and kinetic parameters, however, the computational cost of density functional theory (DFT) calculations makes it difficult to calculate kinetic parameters at scale. In this work, we combine finetuning of the MACE multi-head v0 model with graph-based delta learning of stationary points using subgraph isomorphic decision trees (SIDT). We first generate a target chemical space on Pt111 by using the Reaction Mechanism Generator (RMG) software to generate all possible surface reactions between a set of 153 chemically adsorbed species containing C,H,N, and O. Applying Pynta to this reaction set we are able to generate geometries for many gas phase species, adsorbates and transition states in this chemical space allowing us to efficiently sample near stationary points for finetuning and stationary points for delta learning. In particular, we adapt Pynta's harmonically force saddle point search (HFSP) algorithm to enable efficient and reliable sampling of near transition state points. With this training data we show that SIDT-driven delta learning of the foundation MLIP alone can achieve similar accuracies on enthalpies to foundation model finetuning approaches. However, we show that by combining the two approaches into one framework we are able to significantly improve our accuracies over either approach allowing us to achieve 0.07 eV MAE relative to DFT on test barrier heights. Applying this framework across all reactions in our chemical space, we are able to obtain near DFT accuracy rate coefficients for 3111 reactions on Pt(111).