Overcoming Time Step Limitations in Molecular Dynamics Simulation with Machine Learning

Molecular dynamics simulations are usually limited to small time steps of about 1 fs, a major limitation on the capability of atomistic simulations to reach large time scales. We present a machine-learning method to accelerate molecular dynamics simulations by taking very large time steps. The data-driven approach, built with equivariant graph neural networks, is trained from MD trajectories to faithfully reproduce the same dynamics. The method was demonstrated on a few representative case studies, including single-particle in a 1D double-well potential, and solvated butane dihedral angle dynamics. We show that quantitative agreement on these barrier-crossing events can be achieved with our approach.