Simulating Molecular Dynamics with Large Timesteps using Recurrent Neural Networks

Molecular dynamics simulations rely on numerical integrators such as Verlet to solve the Newton's equations of motion. Using a sufficiently small timestep to avoid discretization errors, Verlet integrators generate a trajectory of particle positions as solutions to the equations of motions. We introduce an integrator based on recurrent neural networks that is trained on trajectories generated using Verlet integrator and learns to propagate the dynamics of particles with timestep up to 4000 times larger compared to the Verlet timestep. We demonstrate significant net speedup of up to 32000 for few-particle (1 - 16) 3D systems and over a variety of force fields.