Parallel replica method for dynamics of infrequent events

Abstract

Although molecular-dynamics simulations can be parallelized effectively to treat large systems (${10}^{6}$--10${}^{8}$ atoms), to date the power of parallel computers has not been harnessed to make analogous gains in time scale. I present a simple approach for infrequent-event systems that extends the time scale with high parallel efficiency. Integrating a replica of the system independently on each processor until the first transition occurs gives the correct transition-time distribution, and hence the correct dynamics. I obtain $>$90% efficiency simulating Cu(100) surface vacancy diffusion on 15 processors.