On July 3rd, 2020, at 3:30 pm
Velocity jumps: an alternative to multi-time step integrators.
Simulating motion of nuclei (following e.g. the Langevin dynamics) requires to compute forces at each time-step, which is numerically demanding. The goal of a multi-time step method is to evaluate different forces at different frequencies, reducing the overall number of computations. Nevertheless, they lead to an increased discretization error and may suffer from resonance problems. We present a continuous-time stochastic process, similar to the Langevin diffusion, that samples the correct equilibrium but can be simulated exactly (without discretization errors). It is particularly efficient for bounded forces, but not so well adapted to singular ones, which leads us to an hybrid process where short-range forces (singular but not very expensive) are discretized as usual but long-range forces (bounded but expensive) are tackled by velocity jumps. The performances are very promising.
More on the research by Pierre Monmarché: click here.