CatBench framework for benchmarking machine learning interatomic potentials in adsorption energy predictions for heterogeneous catalysis

The rapid advancement of machine learning interatomic potentials (MLIPs) has opened new avenues for accelerating computational catalysis research, particularly in adsorption energy predictions—a key descriptor that efficiently correlates with catalytic activity and selectivity. While density functional theory remains the gold standard for first-principles calculations, its computational cost limits large-scale applications. MLIPs accelerate adsorption energy prediction but require rigorous validation to ensure accuracy. Here we report CatBench, a benchmarking framework designed to systematically evaluate the adsorption energy prediction performance of MLIPs. By applying CatBench to extensive adsorption reaction datasets including both small- and large-molecule adsorption, we analyze the predictive capabilities of widely used universal MLIPs (uMLIPs). The benchmarking results provide a comprehensive comparison of uMLIPs, offering critical insights into the practical use of these models for catalysis research. This work highlights the importance of systematic benchmarking in the adoption of machine learning-driven approaches to catalytic system modeling.