Séminaire Théorie

Mardi 18 Octobre 2016 à 11h00.

Long­time atomistic simulations with Parallel Trajectory Splicing

''Molecular Dynamics (MD) is a workhorse of computational materials science. Indeed, MD can in principle be used to obtain any thermodynamic or kinetic quantity, without introducing approximation or assumptions beyond the adequacy of the interaction potential. This enviable quality however comes at a steep computational price, hence limiting the system sizes and simulation times that can be achieved in practice. While the size limitation can be efficiently addressed with massively parallel implementations of MD based on spatial decomposition strategies, the same approach usually cannot extend the timescales much beyond microseconds. In this talk, we discuss a novel approach --- Parallel Trajectory Splicing (ParSplice) --- that aims at addressing the timescale limitation of MD for systems that evolve through rare state-to­state transitions by parallelizing in the time domain. A key innovation of ParSplice lies in the use of speculation on the future evolution of the dynamical trajectory. This exposes additional parallelism by allowing work to simultaneously proceed in many states. Predictions are carried out through a concurrent stochastic simulation parameterized on the fly. We review the formal underpinnings of the method and demonstrate that it can provide arbitrarily accurate results while providing significant improvement in computational efficiency compared to its predecessors. We then illustrate the usefulness of ParSplice by presenting different examples of materials simulations where access to long timescales was essential to access the physical regime of interest.''