Quasi-classical trajectories-based calculation of rate constants using ab-initio trained machine learning force field (aML-MD)

Machine learning (ML) provides a great opportunity for the construction of molecular dynamics (MD) potentials with almost as high accuracy as quantum mechanical methods and as high efficiency as classical molecular dynamics. In this work, two ab-initio trained ML based MD (aML-MD) potentials or models are developed for hydrogen combustion using two different sets of DFT data (system-wide data and reaction-specific data) within the Deep Potential MD (DPMD) framework. Both aML-MD models exhibit excellent accuracy in capturing the potential energy surface from training data. The accuracy of the aML-MD models in predicting the reaction dynamics is demonstrated by calculating rate constants for the singlet reaction H + HO2 ® OH + OH using quasi classical trajectories (QCT). We show that both models underpredict the rate constants compared to the existing state-of-the-art QCT predictions. It is shown that the system-wide data trained aML-MD model significantly underpredicts the rate constants whereas the reaction specific data trained model improves the rate constant prediction. We show that an accurate and comprehensive dataset with wideranging energy levels is critical for an aML-MD model to capture diverse reaction dynamics, whichcanencompass multiple energybarriers andintermediates. Future work willbe focused on using transfer learning to improve the accuracy, efficiency, and generalization of aML-MD models.