Noise reduction and automated molecule detection in atomistic diffusion calculations of correlated systems

When using molecular dynamics to compute diffusion coefficients for correlated systems, the fully correlated analysis of molecular trajectories leads to very noisy estimates because it is based on the center of mass motion rather than motion of individual atoms. We propose a new method that systematically reduces the noise in the diffusion constant calculation. By assuming that the underlying correlation structure of a system is the same for the entire trajectory, we leverage the information learned from the well-converged short-time position-position correlation matrix to reduce the noise at longer times. The proposed method allows to significantly decrease the uncertainty of diffusivity estimates and works for both non-bonded and bonded correlation scenarios. In the latter case we are able to identify automatically the diffusing degrees of freedom of the system, i.e. identify the molecular species. Furthermore, we demonstrate that using our method the estimation of the diffusion constant is never worse than the widely-adopted center-of-mass approach.