Energy extrapolation schemes for adaptive multi-scale molecular dynamics simulations.

This paper evaluates simple schemes to extrapolate potential energy values using the set of energies and forces extracted from a molecular dynamics trajectory. In general, such a scheme affords the maximum amount of information about a molecular system at minimal computational cost. More specifically, schemes like this are very important in the field of adaptive multi-scale molecular dynamics simulations. In this field, often the computation of potential energy values at certain trajectory points is not required for the simulation itself, but solely for the a posteriori analysis of the simulation data. Extrapolating the values at these points from the available data can save considerable computational time. A set of extrapolation schemes are employed based on Taylor series and central finite difference approximations. The schemes are first tested on the trajectories of molecular systems of varying sizes, obtained at MM and QM level using velocity-Verlet integration with standard simulation time steps. Remarkably good accuracy was obtained with some of the approximations, while the failure of others can be explained in terms of the distinct features of a molecular dynamics trajectory. We have found that, for a Taylor expansion of the potential energy, both a first and a second order truncation exhibit errors that grow with system size. In contrast, the second order central finite difference approximation displays an accuracy that is independent of the size of the system, while giving a very good estimate of the energy, and costing as little as a first order truncation of the Taylor series. A fourth order central finite difference approximation requires more input data, which is not always available in adaptive multi-scale simulations. Furthermore, this approximation gives errors of similar magnitude or larger than its second order counterpart, at standard simulation time steps. This leads to the conclusion that a second order central finite difference approximation is the optimal choice for energy extrapolation from molecular dynamics trajectories. This finding is confirmed in a final application to the analysis of an adaptive multi-scale simulation.