Path-Accelerated Molecular Dynamics: Parallelizing Molecular Dynamics in Time using Path Integrals

Ample availability of parallel computing resources has instigated efforts to accelerate all-atom molecular dynamics (MD) simulations by exploiting their inherent spatial parallelizability. However, as spatial decomposition schemes approach their parallel scaling limits, the need arises for new algorithms that can harness parallel computing architectures in orthogonal and complimentary ways to enable further computational speed-ups and the generation of longer MD trajectories. Here, we use the fact that the transition kernel for the time evolution of a system admits a path-integral representation to introduce an algorithm that employs path-sampling techniques to generate long MD trajectories in short wall-clock times. By employing massively parallelizable and rapidly convergent path-sampling techniques to identify paths for the system to evolve along, we effectively parallelize propagation of the system's equilibrium dynamics with respect to the discretization timestep. We apply the algorithm to simulate various diffusive systems including a Lennard-Jones liquid, and attain hundredfold speedups over a standard Euler integrator in terms of the length of equilibrium trajectory generated per wall-clock time unit.