Transition State Searching Accelerated by Deep Learning Potential

The exploration and construction of organic chemical reaction networks hold the potential to clarify complex reaction mechanisms. However, this endeavor encounters challenges, particularly in cases requiring computationally intensive quantum mechanical calculations. In recent years, the remarkable advancement of machine learning (ML) technologies has provided a novel approach to address this issue. ML can effectively utilize vast datasets to build and train complex models, enabling prediction and simulation. In this study, we conducted a performance comparison of deep potential for molecular dynamics (DeePMD), recursively embedded atom neural network (REANN), and neural equivariant interatomic potentials (NequIP) based on the Transition1x dataset. The most efficient model, NequIP, was selected for further analysis. Combined with reaction path search methods such as nudged elastic band (NEB) and growing string method (GSM), this model was employed for the identification and exploration of transition states. The results demonstrate that the success ratio of NequIP combined with the NEB method can reach 96.6%, with a mean absolute error (MAE) of 0.32 kcal/mol for barrier prediction. By adding a modest number of data points, NequIP achieves a MAE of 4.01 kcal/mol for barrier prediction in unexplored chemical reaction space. In the future, by combining different reaction network search methods, we anticipate applying this model to swiftly explore reaction pathways and construct more comprehensive reaction networks. We believe that ML will play a significant role in accelerating the elucidation of unknown chemical reaction mechanisms and in revealing complex reaction networks.