Quantification of sampling uncertainty for molecular dynamics simulation: Time-dependent diffusion coefficient in simple fluids

Abstract

We analyze two standard methods to compute the diffusion coefficient of a tracer particle in a medium from molecular dynamics (MD) simulation, the velocity autocorrelation function (VACF) method, and the mean-squared displacement (MSD) method. We show that they are equivalent in the sense that they provide the same mean values with the same level of statistical errors. We obtain analytic expressions for the level of the statistical errors present in the time-dependent diffusion coefficient as well as the VACF and the MSD. Under the assumption that the velocity of the tracer particle is a Gaussian process, all results are expressed in terms of the VACF. Hence, the standard errors of all relevant quantities are computable once the VACF is obtained from MD simulation. By using analytic models described by the Langevin equations driven by Gaussian white noise and Poissonian white shot noise, we verify our theoretical error estimates and discuss the non-Gaussianity effect in the error estimates when the Gaussian process approximation does not hold exactly. For validation, we perform MD simulations for the self-diffusion of a Lennard-Jones fluid and the diffusion of a large and massive colloid particle suspended in the fluid. Our theoretical framework is also applicable to mesoscopic simulations, e.g., Langevin dynamics and dissipative particle dynamics.