On data-driven parameterizations of multidimensional generalized Langevin dynamics in the presence of a quadratic potential

Abstract

We propose a numerical algorithm to construct a Markov model with an extended list of variables to parameterize the equation of motion of a multidimensional coarse-grained physical system in an external potential, when memory effects are relevant. Our method uses autocorrelation data of the stationary velocities, but it avoids the inverse problem of finding the corresponding memory kernel from these data in a first step. Rather, the data are used to construct a Prony series approximation of the autocorrelation function, and the parameters of this Prony series provide the corresponding Markov model. Numerical results for molecular dynamics data show a good match for parameterized models with five auxiliary variables for a one-dimensional, and twelve auxiliary variables for a two-dimensional system.