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

Abstract

Transition state (TS) search is crucial for illuminating chemical reaction mechanisms but remains the major bottleneck in automated discovery because of the high computational cost. Recently, machine learning interatomic potentials (MLIPs) and generative models have shown promise in accelerating TS search, but their comparative strengths and limitations remain unclear. In this study, the first systematic and rigorous benchmarking framework is established to evaluate the effectiveness of ML methods in TS search, enabling a standardized and application-relevant assessment of their performance. Using an end-to-end TS search workflow, seven representative MLIPs are benchmarked alongside React-OT, a state-of-the-art generative model. These results demonstrate that pre-trained foundation MLIPs frequently fall short in reliably localizing TSs without task-specific fine-tuning. Furthermore, traditional energy and force metrics alone do not reliably predict TS search success, underscoring the need for more tailored evaluation criteria. Notably, with the same graph neural network architecture, React-OT frequently outperforms its MLIP counterpart, highlighting the potential of generative approaches for TS discovery. This benchmark serves as a critical foundation for the development and evaluation of future ML methods in chemical reactions, offering guidance for improving their generalizability and reliability in reactive chemistry.