Long Time-Scale Accuracy of Neural Network Potentials in Molecular Dynamics Simulations

Molecular dynamics (MD) has been extensively employed in understanding a wide range of physical and chemical problems. To overcome the limitations of the high computational cost in ab initio MD (AIMD) and relative lower predictive accuracy in classical MD (CMD), machine learning (ML) force fields with accuracy comparable to AIMD and computational efficiency comparable to CMD have been of interest in recent years. However, most ML force fields still suffer from limitations such as poor interpretability and extrapolation as well as less accuracy in long-range interactions for large simulations. In this work, we compared three ML potentials (DPMD, TorchANI, SchNet) with different architectures, and found a systematically accumulated error in the MD simulations for all these ML potentials. To further reduce the error between the reference trajectories and predicted trajectories across the simulated period, we proposed a graph neural network time stepper that can be used for direct prediction of molecular trajectories after being trained by AIMD trajectories, and a penalty for the discrepancy between the reference and predicted trajectory over simulation time is considered during GNN surrogate training. This could further improve the predictive power of our model compared to traditional single-step ML potentials. We will also illustrate our recent progress in designing a physically informed ML potential surrogate to maintain predictive accuracy and improved transferability by informing the physical nature of interatomic interactions. We anticipate that our development will be easily generalizable to many systems and it will effectively accelerate research of atomistic scale dynamics simulations.