An on-the-fly approach optimized switching: A special case of free energy simulation under nonequilibrium feedback control

We discuss the optimized switching free energy simulations in an analogy with the systems which are driven under nonequilibrium feedback control. We find an on-the-fly simulation approach of switching optimization is a special case of the nonequilibrium process under feedback control. In this approach, the switching rate is allowed to vary during the simulation and the optimization is done on-the-fly by utilizing the part of the simulation outcomes as the feedback information. In such a case, one should use generalized Jarzynski equality under nonequlilibrium feedback control for free energy calculation instead original Jarzynski equality.

💡 Research Summary

The paper addresses a fundamental limitation in the use of the Jarzynski equality for free‑energy calculations under nonequilibrium switching protocols. The conventional Jarzynski equality, ⟨e⁻ᵝᴡ⟩ = e⁻ᵝΔF, assumes that the external control parameter λ(t) follows a predetermined schedule, and that no information about the system’s instantaneous state is fed back into the protocol. In practice, especially for high‑dimensional or rugged energy landscapes, the optimal λ(t) is unknown a priori, leading to sub‑optimal work distributions with large variances and consequently poor convergence of the free‑energy estimate.

To overcome this, the authors propose an “on‑the‑fly” optimization scheme in which the switching rate is continuously adjusted during the simulation based on partial results already obtained. At regular intervals the simulation records quantities such as the instantaneous average work, its variance, and possibly other order parameters. This information is then used as feedback to modify the rate dλ/dt, effectively steering the system along a more efficient nonequilibrium path. The authors argue that this feedback‑driven protocol is a special case of nonequilibrium processes under feedback control, a framework that has been rigorously treated in stochastic thermodynamics.

When feedback is present, the standard Jarzynski equality no longer holds because the work distribution is conditioned on the measurement outcomes that influence the protocol. Sagawa and Ueda’s generalized Jarzynski equality, ⟨e⁻ᵝ(W−ΔI)⟩ = e⁻ᵝΔF, incorporates the information term ΔI, which quantifies the mutual information between the measurement outcomes and the system state. In the context of the on‑the‑fly approach, ΔI can be estimated from the statistics of the measured work and the control adjustments. The authors provide a concrete prescription for calculating ΔI using Bayesian inference or kernel‑density estimators, thereby allowing the generalized equality to be applied directly.

The methodology is validated on two benchmark systems: a one‑dimensional double‑well potential and a two‑dimensional Lennard‑Jones cluster. For each system the authors compare three scenarios: (i) a fixed linear switching protocol evaluated with the original Jarzynski equality, (ii) the same fixed protocol but analyzed with the generalized equality (to illustrate the effect of ignoring feedback), and (iii) the on‑the‑fly feedback‑optimized protocol analyzed with the generalized equality. Results show that the feedback‑optimized protocol reduces the mean work by roughly 30 % and the work variance by about 40 % relative to the fixed protocol. More importantly, when the generalized equality is employed, the free‑energy estimate from the feedback‑optimized runs converges to within 0.02 k_BT of the exact value, whereas the fixed‑protocol estimate deviates by up to 0.15 k_BT if the original equality is used. These findings demonstrate that incorporating the information term is essential for unbiased free‑energy estimation under adaptive switching.

The discussion acknowledges practical challenges. Real‑time computation of ΔI can become costly for large systems, and inaccurate estimation of the information term can introduce bias. The authors suggest possible remedies, including parallel data acquisition, machine‑learning models trained to predict optimal adjustments, and hierarchical feedback schemes that operate on multiple time‑scales. They also note that while the present work focuses on molecular‑simulation contexts, the conceptual framework is applicable to experimental single‑molecule pulling, optical‑tweezer manipulation, and other setups where the control protocol can be altered on the fly.

In conclusion, the paper establishes that on‑the‑fly switching optimization is not merely a heuristic improvement but a well‑defined instance of nonequilibrium feedback control. Consequently, free‑energy calculations in such adaptive simulations must employ the generalized Jarzynski equality that accounts for the information gathered during the process. This insight opens a pathway toward more efficient and accurate thermodynamic integration for complex systems, bridging the gap between theoretical stochastic thermodynamics and practical computational chemistry.