Improving fast-switching free energy estimates by dynamical freezing

An important limitation of nonequilibrium pulling experiments/simulations in recovering free energy differences is the poor convergence of path-ensemble averages. Therefore, a large number of fast-switching trajectories needs to achieve free energy estimates with acceptable accuracy. We propose a method to improve free energy estimates by drastically lowering the computational cost of steered molecular dynamics simulations employed to realize such trajectories. This is accomplished by generating trajectories where the particles not directly involved in the driven process are dynamically frozen. Such a freezing is dynamical rather than thermal because it is reached by a synchronous scaling of atomic masses and velocities keeping the kinetic energy of each particle unchanged. The forces between dynamically frozen particles can then be calculated rarely. Thus, it is possible to generate realizations of a process whose computational cost is not correlated with the size of the whole system, but only with that of the reaction site. The method is illustrated on a simple model system and its general applicability is discussed.