CatWalk: Reaction Pathway Generation and Transition State Sampling via Score-based Diffusion Models

Abstract

Computing transition states for heterogeneous catalyst surface reactions is computationally demanding due to the relatively high cost in precisely locating a saddle point using ab initio electronic structure calculations. In this work, we present a new generative score-based diffusion model—CatWalk—that predicts the initial structures for reaction pathways of heterogeneous catalysts with surface adsorbates. Given only the desired reaction type and reactant structure, CatWalk produces nine pathway structures by iteratively perturbing the coordinates of the previous output. CatWalk was trained on 731 nudged elastic band (NEB) calculations from the CatTSunami dataset, which contains diverse substrate elements with adsorbates undergoing three types of chemical reactions: desorption, dissociation, and atom transfer. The model can be directed to target one of these three reaction types during generation, and due to its stochastic nature can generate various products via multiple pathways from the same initial state. Optimizing CatWalkgenerated pathways using a pretrained foundation machine-learned interatomic potential (MLIP) with the nudged elastic band (NEB) method leads to several lower-energy pathways compared to what is obtained by the traditional linear interpolation initialization method for NEBs. Additionally, repeated intermediate structure generations using CatWalk enabled sampling of different pathways and final states which can lead to multiple saddle points with variation in energy by up to 0.2 eV. Overall, CatWalk provides a general and scalable framework for accelerating reaction-pathway discovery, enabling broader exploration of catalytic mechanisms at significantly reduced computational cost.