Well-Tempered Metadynamics: A Smoothly Converging and Tunable Free-Energy Method

This paper addresses a central problem in molecular simulation: how to compute the free-energy surface (FES) as a function of a small set of collective variables (CVs) when direct sampling is hindered by high barriers and slow transitions. Free-energy landscapes underpin quantitative predictions of thermodynamic stability, kinetics, and conformational populations in chemistry and biophysics. However, standard unbiased molecular dynamics (MD) or Monte Carlo (MC) often fail to explore relevant regions of configuration space within feasible simulation times.

Metadynamics is a prominent adaptive biasing method that accelerates sampling by adding a history-dependent potential built from a sequence of Gaussian “hills” deposited along the trajectory in CV space. After a transient, the accumulated bias compensates the underlying free energy along the CVs, enabling reconstruction of the FES. Despite its success, conventional metadynamics has two practical and theoretical shortcomings emphasized by the authors. First, convergence is problematic: in a single run, the reconstructed free energy does not settle to a fixed value but fluctuates around the correct result, with an average error that scales like the square root of the bias deposition rate. Reducing the deposition rate improves accuracy but increases the time needed to fill the landscape. Second, because the bias keeps growing, continuing a run can push the system into regions that may not be physically relevant, complicating decisions about when to stop.

The paper introduces “well-tempered metadynamics,” a modification that yields a bias potential whose growth rate decreases smoothly in time, leading to a reconstructed FES that converges in the long-time limit. The method also introduces a tunable parameter that controls which regions of the FES are explored, thereby focusing computational effort on physically relevant parts of CV space and making stopping criteria straightforward.

Methodologically, the system is defined by microscopic coordinates $q$ with potential energy $U(q)$, evolving under canonical sampling at temperature $T$. The goal is to estimate the free energy dependence on a selected CV $s(q)$. The authors express the FES (up to an additive constant) via the long-time limit of the unbiased histogram of $s$: $F(s)=-T{\lim }_{t\to \infty }\ln N(s,t)$, where $N(s,t)$ is the time-integrated delta-function histogram. To accelerate sampling, they add a history-dependent bias $V(s,t)$ defined as $$V(s,t)=\Delta T\ln \left(1+\frac{\omega N(s,t)}{\Delta T}\right),$$ where $\omega$ is an energy-rate parameter setting the initial bias deposition rate, and $\Delta T$ is a temperature-like parameter controlling the “tempering.” Differentiating with respect to time yields a key result for the bias update rate: $$\dot{V}(s,t)=\frac{\omega \Delta T{\delta }_{s,s(t)}}{\Delta T+\omega N(s,t)}=\omega \exp \left(-\frac{V(s,t)}{\Delta T}\right){\delta }_{s,s(t)}.$$ In practice, the delta-function deposition is replaced by finite-width Gaussians, so the algorithm can be implemented by rescaling the Gaussian heights according to the factor $\exp (-V/\Delta T)$. The authors interpret $\omega$ as the initial deposition rate and introduce the time interval ${\tau }_G$ between Gaussian depositions.

A central theoretical property is that because $N(s,t)$ grows linearly with time, the bias update rate decays asymptotically like $\propto 1/t$, ensuring that the bias eventually converges. Unlike simpler $1/t$ strategies, the decay is non-uniform in CV space: locations already visited more have larger $N$ and thus smaller $\dot{V}$, naturally reducing overfilling.

To connect the bias to free energy reconstruction, the authors assume that at large times the microscopic degrees of freedom equilibrate under the slowly varying bias. Under this quasi-equilibrium assumption, the distribution of $s$ in the biased system becomes $$P(s,t)ds\propto \exp \left(-\frac{F(s)+V(s,t)}{T}\right)ds.$$ Combining this with the bias evolution equation leads to the long-time limit $$V(s,t\to \infty )=-\frac{\Delta T}{\Delta T+T}F(s)(\text{up to a constant}).$$ Therefore, the bias does not fully cancel $F(s)$ as in standard metadynamics; instead, the sum satisfies $$F(s)+V(s)=\frac{T}{T+\Delta T}F(s).$$ Consequently, the asymptotic biased CV distribution corresponds to an effective temperature $T+\Delta T$: $$P(s,t\to \infty )\propto \exp \left(-\frac{F(s)}{T+\Delta T}\right).$$ The reconstructed free energy estimate used in practice is $$\tilde{F}(s,t)=-\frac{T+\Delta T}{\Delta T}V(s,t)=-(T+\Delta T)\ln \left(1+\frac{\omega N(s,t)}{\Delta T}\right).$$ The method recovers standard metadynamics in the limit $\Delta T\to \infty$ (where deposition rate becomes constant), while $\Delta T=0$ corresponds to no bias. Importantly, for finite $\Delta T$, the exploration is automatically limited to an energy range on the order of $T+\Delta T$, reducing the risk of spending time in irrelevant regions.

The paper tests the method on the alanine dipeptide free-energy landscape in vacuum as a function of backbone dihedral angles $(\Phi ,\Psi )$. This system is a standard benchmark because it has two minima, $C_{7eq}$ and $C_{7ax}$, separated by a barrier of approximately $9$ kcal/mol, which is difficult to cross with room-temperature dynamics. Using the CHARMM27 force field in the ORAC MD code, the authors perform canonical sampling at $300$ K. The Gaussian width is set to $20^{\circ }$, the deposition interval to $120$ fs, and the initial Gaussian height to $0.287$ kcal/mol, corresponding to an initial deposition rate $\omega =2.4$ cal/mol/fs.

They compute a reference $F(\Phi ,\Psi )$ using umbrella sampling, finding agreement with prior studies. They then run three well-tempered metadynamics trajectories starting from the same initial conditions at $C_{7eq}$ ($(-83,74)$ in the $(\Phi ,\Psi )$ plane), with $\Delta T$ set to $600$ K, $1800$ K, and $4200$ K. In all cases, the secondary metastable state $C_{7ax}=(70,-70)$ is frequently visited. To demonstrate convergence, they track the time evolution of the estimated free-energy difference between minima, $\Delta \tilde{F}(t)=\tilde{F}(C_{7ax},t)-\tilde{F}(C_{7eq},t)$. They report that $\Delta \tilde{F}(t)$ converges to the reference value $\Delta F\approx 2.2$ kcal/mol in all three trajectories. In contrast to standard metadynamics, the time derivative of the bias potential tends to zero and fluctuations around the correct value are progressively damped. They further state that all three simulations provide an accurate estimate of the free-energy difference within a few nanoseconds, even for the smallest $\Delta T$, where fewer barrier crossings lead to a “jumpier” evolution.

For quantitative error assessment, the authors define an RMS error measure over a relevant region $\Gamma$ of dihedral space: $$ϵ(t)={\left(\frac{1}{A}{\int }_{\Gamma }{\left(F(\Psi ,\Phi )-\tilde{F}(\Psi ,\Phi ,t)-C(t)\right)}^{2}d\Phi d\Psi \right)}^{1/2},$$ where $C(t)$ accounts for the arbitrary additive constant in free-energy reconstructions. They analyze the scaling of $⟨ϵ(t)⟩\sqrt{t}$ versus simulation time and report that, unlike standard metadynamics where the error does not converge to zero within a single simulation, well-tempered metadynamics shows behavior consistent with convergence (the error decreases as time increases). They also study how the asymptotic constant $k={\lim }_{t\to \infty }⟨ϵ(t)⟩\sqrt{t}$ depends on $\Delta T$, using this to optimize $\Delta T$. The optimal choice is reported to be close to $\Delta T=1200$ K, corresponding to a sampling temperature for the CVs of $T+\Delta T=1500$ K, which is on the order of the barrier height.

The authors discuss the role of the initial deposition rate $\omega$ through a time constant ${\tau }_B=\Delta T/\omega$, which sets the timescale for bias evolution. While ${\tau }_B$ is irrelevant in the long-time limit, it can affect the transient regime. They investigate this using an artificial model based on the alanine dipeptide FES, employing a high-friction Langevin dynamics on $(\Phi ,\Psi )$ with diffusion coefficients $D_{\Phi }$ and $D_{\Psi }$ taken from atomistic simulations. By tuning $D_{\Psi }$, they mimic fast versus slow relaxation of the orthogonal degree of freedom. In the fast case, a small ${\tau }_B$ is best because the orthogonal coordinate averages out quickly, yielding more Markovian effective dynamics. In the slow case, small ${\tau }_B$ increases transient time due to non-Markovian coupling, but the method remains robust: across a range spanning two orders of magnitude in ${\tau }_B$ and $D_{\Psi }$, they observe convergence to the same results on approximately the same timescale.

Limitations are primarily implicit in the methodology. The theoretical convergence argument relies on the assumption that at large times the biased system’s microscopic degrees of freedom equilibrate quickly relative to the bias evolution, enabling the quasi-equilibrium form of $P(s,t)$. While the authors do not provide a formal bound on separation of timescales, they emphasize that their result does not require an adiabatic separation between $s$ and other variables (unlike some earlier approaches). Practically, the method still depends on the choice of CVs: if the chosen CVs do not capture the slow modes, convergence may remain slow regardless of tempering. Additionally, while the paper provides guidance for tuning $\Delta T$, the optimal value can depend on CV relaxation times and the region of CV space one wants to explore.

The practical implications are substantial. Well-tempered metadynamics offers a unified framework that interpolates between standard metadynamics and unbiased sampling, while providing smoother convergence and a tunable exploration temperature. It enables users to focus computational effort on physically relevant regions of the FES, reduces the risk of overfilling, and makes stopping criteria simpler because the bias update rate naturally decreases and fluctuations damp. Researchers studying free-energy landscapes in complex molecular systems—protein conformational changes, ligand binding, phase transitions, and other rare-event processes—should care because the method improves reliability of FES reconstruction within a single run and provides a principled way to trade off exploration breadth against convergence speed.

Overall, the paper’s core contribution is the well-tempered biasing scheme that yields a long-time bias–free-energy relationship $V(s)\propto -\frac{\Delta T}{T+\Delta T}F(s)$, leading to an effective sampling distribution at temperature $T+\Delta T$ and enabling convergence behavior that is qualitatively different from standard metadynamics.