ChronoCast: A time-series paradigm for molecular dynamics simulation using equivariant graph neural networks.

We propose a time-series paradigm named ChronoCast for molecular dynamics simulation using an advanced autoregressive Equivariant Graph Neural Network, which reformulates the conventionally time-consuming integration process into a forecasting task. This reformulation in turn offers a promising solution to overcome the challenge of tremendous computational cost for studying complex physical problems based on traditional algorithms. By incorporating velocity as a node feature and enforcing momentum conservation, the proposed model achieves exceptionally high accuracy in reproducing diverse physical properties. Using the simple Si crystal and complex van der Waals NbSe_{3} nanowires as two examples, we demonstrate that the radial distribution function, mean-squared displacement, and vibrational density of states in both systems can be accurately reproduced by long-term autoregressive forecasting. More importantly, ChronoCast significantly reduces the trajectory generation time by orders of magnitude compared to the ab initio method and by more than half compared to state-of-the-art machine learning potential, demonstrating superior efficiency over these prevailing molecular dynamics simulations. This work offers an accurate and efficient time-series approach for studying the statistical physics of complex dynamical processes.