Free energy reconstruction from steered dynamics without post-processing (Q2638241)

The authors propose, basing on the Bayes theorem, a more direct method to construct steered dynamics, which is implemented in two steering schedules. One of the important application of molecular simulation is the problem of estimation Landau \(F\) free energy of a given multi-particle system by the equation \[ F(\xi)=-k_{B}T\ln P(\xi) \] where \(k_B\) stands for Boltzmann's constant, \(T\) denotes temperature and \(P(\xi)\) is the probability to observe the system with value \(\xi\). To restore numerical ergodicity many simulation techniques have been proposed, based on umbrella sampling. Usually in this technique we have restoring to a judicious steering or restraining potential that enhances exploration of regions of phase space that would be poorly sampled otherwise: a series of umbrella sampling simulations are performed to cover the area of interest, then the collected averages are combined using reweighing procedures related to Bennett's acceptance ratio method and based on likelihood maximization. Hummer and Szabo proposed a method, in which they reconstruct free energy profiles by applying the histogram reweighing procedure to nonequilibrium simulations (instead of equilibrium ones) and by introduction of an additional variable \(\xi^{\text{add}}\). They reconstruct the equilibrium properties by two-step procedure. The authors of considered paper proposed an estimator that enables to retrieve the equilibrium information from nonequilibrium trajectories basing on reverse-to-forward probability ratios and retrieving the information in all time-slices. This estimator has two important properties: (i) it does not involve any post-processing, (ii) it can be used with more general steering schedules comparing to the Hummer and Szabo approach. This can be achieved thanks to the considerations of Langevin dynamics in which steering arises from additional restraining variables that evolve stochastically and autonomously out of equilibrium into unexplored regions of phase space. The paper consist of 7 sections and two appendixes, in which the reader can find the following aspects. Section 2 shows the theoretical framework for the steered dynamics of a particle system and its additional variables in the context of extended Hamiltonian (Section 2.1), presents the steered Langevin dynamics and reverse-to-forward probability ratios as a generalized balance equation involving the forward and backward Kolmogorov operator associated to Langevin dynamics. This is used to discuss two steering schedules for the dynamics in Section 2.4 and to construct the two-state estimators that are taken into account for calculations of free energy differences with non-autonomous steering in Section 3. In the next section we have the proposal of an extended sampler and estimator that enable the reconstruction of free energy -- as it turn-out this proposal works for both autonomous and non-autonomous steering schedules. In Section 5 we have the numerical illustration of the performance of proposed estimator in the case of 1D reconstruction problem -- the achieved result confirmed the exactness of authors' method. The next section shows 2D case for the Lennard-Jones cluster with 38 atoms. This system presents a rugged energy landscape with two energy funnels separated by a high free energy barrier -- once again the proposed method succeeds, however the comparison of numerical efficiency to the other methods is quite difficult because estimators and steering schedules of distinct typed were used. Paper is concluded in Section 7. The appendixes shows: A -- the discussion about residence algorithms and optimized path-sampling; B -- the discussion about algorithms efficiency in the context of used time step \(\Delta t\).