Harnessing Machine Learning to Enhance Transition State Search with Interatomic Potentials and Generative Models

Transition state (TS) search plays a crucial role in reaction pathway analysis, offering insights into reaction mechanisms and aiding in the optimization of chemical processes. Recently, machine learning interatomic potentials (MLIPs) and generative models have emerged as promising tools to accelerate TS search. In this study, we establish an end-to-end TS search workflow to benchmark seven MLIPs -- ANI-1x, CHGNet, DPA-2, LEFTNet, MACE, MatterSim, and Orb -- alongside React-OT as a state-of-the-art generative model. Our evaluation reveals that while current pre-trained foundation MLIPs show potential, they do not consistently excel in TS search tasks and require additional reactive data for effective fine-tuning. Furthermore, commonly used energy and force metrics for comparing MLIPs do not fully capture their performance in TS search. Notably, when LEFTNet is used for both React-OT and MLIP as the model architecture, React-OT often outperforms MLIP-based TS search, achieving a higher success rate in locating TSs. This work not only highlights the current capabilities of MLIPs and generative models but also provides valuable insights for future advancements in TS prediction and the exploration of new reaction mechanisms.