Coarse graining from variationally enhanced sampling applied to the Ginzburg–Landau model

Contributed by Michele Parrinello, February 14, 2017 (sent for review November 7, 2016; reviewed by Daan Frenkel and John D. Weeks)
March 14, 2017
114 (13) 3370-3374


A multiscale description is believed to be the only feasible route toward the simulation of complex systems. Yet moving from one scale to another is not a trivial task, and many different approaches can be taken. We propose here a systematic approach to go from a fine scale to a coarser one, based on the recently introduced variationally enhanced sampling method. Contrary to other coarse-graining methods, it combines in the same algorithm the sampling enhancement and the optimization of the coarse-graining parameters. We illustrate the method in the case of the vapor–liquid critical transition of a Lennard–Jones fluid. This is done in the context of the Ginzburg–Landau field theoretical description of second-order phase transitions.


A powerful way to deal with a complex system is to build a coarse-grained model capable of catching its main physical features, while being computationally affordable. Inevitably, such coarse-grained models introduce a set of phenomenological parameters, which are often not easily deducible from the underlying atomistic system. We present a unique approach to the calculation of these parameters, based on the recently introduced variationally enhanced sampling method. It allows us to obtain the parameters from atomistic simulations, providing thus a direct connection between the microscopic and the mesoscopic scale. The coarse-grained model we consider is that of Ginzburg–Landau, valid around a second-order critical point. In particular, we use it to describe a Lennard–Jones fluid in the region close to the liquid–vapor critical point. The procedure is general and can be adapted to other coarse-grained models.
Computer simulations of condensed systems based on an atomistic description of matter are playing an ever-increasing role in many fields of science. However, as the complexity of the systems studied increases, so does the awareness that a less detailed, but nevertheless accurate, description of the system is necessary.
This has been long since recognized, and branches of physics like elasticity or hydrodynamics can be classified in modern terms as coarse-grained (CG) models of matter. In more recent times, a field theoretical model suitable to describe second-order phase transitions has been introduced by Landau (1) and later perfected by Ginzburg (2). In recent decades a number of coarse-grained models that aim at describing polymers or biomolecules have also been proposed (3, 4). In all these approaches some degrees of freedom, deemed not essential to study the phenomenon at hand, are integrated out and the resulting reduced description contains a number of parameters that are not easily determined.
Here we use the recently developed variationally enhanced sampling (VES) method (5) to suggest a procedure that allows the determination of such parameters, starting from the microscopic Hamiltonian. This illuminates a somewhat unexpected application of VES, which has been introduced as an enhanced sampling method. We show that it also provides a powerful framework for the optimization of CG models, due to the combination of its enhanced sampling capabilities and its variational flexibility. Moreover, VES takes advantage of its deep connection with relative entropy, a quantity that has been shown to play a key role in multiscale problems (6, 7).
As a first test case for our procedure we consider the Ginzburg–Landau (GL) model for continuous phase transitions. An advantage of using this model is that its strengths and limits are well known and that other researchers have already attempted to perform such a calculation (814). A system that undergoes a second-order phase transition is described in the GL model by the following free energy, valid in a rotationally invariant one-component real-order parameter scenario,
where the field ψ(r) describes the order parameter fluctuations, and g, a, and b are phenomenological quantities that we call Landau’s parameters. To be defined, we apply our method close to the liquid–vapor critical point of a Lennard–Jones fluid.

Variationally Enhanced Sampling

Before discussing our method, we briefly recall some of the ideas at the basis of VES. The VES method shares with metadynamics (15, 16) the idea of enhancing the sampling by adding an external bias potential, which is a function of a few coarse-grained order parameters, so-called collective variables (CVs). Contrary to metadynamics and other similar methods, VES does not generate the bias potential by periodically adding on the fly small contributions, but as the result of the minimization of a convex functional. Various extensions have been proposed, where VES has been used in different ways (1722).
Let us suppose that the physical behavior of interest is well described by a set of CVs 𝐬=𝐬(𝐑), which are function of the coordinates 𝐑 of the N particles that compose the system. Then we can write the associated free-energy function as F(𝐬)=(1/β)logd𝐑δ[𝐬𝐬(𝐑)]eβU(𝐑), where β=1/(kBT) is the inverse temperature and U(𝐑) is the interparticle potential energy.
In VES a functional that depends on an externally applied bias V(𝐬) is introduced,
where p(𝐬) is a chosen probability distribution that we call target distribution. The functional is convex and for the value of V(𝐬) that minimizes Ω[V], we have
is the distribution of 𝐬 in the ensemble biased by V(𝐬). Neglecting immaterial constants, Eq. 3 can also be written as
Thus, finding the minimizing bias potential V(𝐬) is equivalent to finding the free energy F(𝐬).
We note here that the functional Ω[V] has close connections to the relative entropy or Kullback–Leiber divergence DKL (2325), in particular βΩ[V]=DKL(pPV)DKL(pP0). Furthermore, at the minimum we have Ω[V]0, which is a rewrite of the Gibbs–Bogoliubov inequality (26).
To minimize Ω[V] we express V(𝐬) as a linear function of a finite set of variational parameters 𝜶={αi} and thus V(𝐬;𝜶). By doing so, our functional becomes a function of 𝜶 that can be minimized using the gradient and the Hessian
where the averages in the right-hand side of the equations are calculated either in the biased ensemble V(𝜶) or in the p(𝐬) ensemble p, and second derivatives in 𝜶 have been omitted due to the linearity assumption.
Because the forces in Eqs. 6 and 7 are calculated as statistical averages, a stochastic optimization method (27) is needed. We describe in detail the minimization procedure in Minimization Procedure.

Minimization Procedure

The minimization of the VES functional Ω[V], Eq. 2, was carried out following the iterative procedure of Bach and Moulines (27), as described also in refs. 5 and 17. In this scheme we consider two distinct sequences of parameters: the instantaneous 𝜶(n) and the averaged 𝜶¯(n). The instantaneous sequence is updated using the gradient and the full Hessian (Eqs. 6 and 7), evaluated at the last available average parameter values 𝜶¯(n),
where μ is a fixed step size that we keep on the order of 0.1 in all our simulations. An important feature of this algorithm is that it does not require a long sampling time for each iteration; in fact, we typically estimate the gradient and the Hessian with a stride S=2,000 simulation time steps. We also implemented a multiple-walkers framework (37) to further improve the sampling, using two independent walkers that share the same bias.
To accelerate convergence, we found it useful to calculate 𝜶¯ as an exponentially decaying average; namely,
which is the discretized equivalent of
The weight W is linked to the decaying time τ by the relation 1/W=1eSΔt/τSΔt/τ, where Δt is the simulation time step and S is the stride between the parameter updates. In the runs reported here we aimed at a τ of about 1/10th of the total simulation length, which amounts to taking W=500.
Within reasonable bounds, the precise choice of the parameters μ, S, and τ does not affect much the rate of convergence of the optimization algorithm. In fact, we found that the choice of the target distribution parameters gt and at plays a more important role in this.

Free-Energy Model

We consider a Lennard–Jones fluid of N particles confined in a periodically repeated cubic box of volume V=L3. The order parameter of such a system is linked to the density ρ(r),
where ρ0=N/V and ρc is the critical density. The order parameter can be expanded in a discrete Fourier series as
where k=2πLn, with n3. The Ginzburg–Landau model implicitly relies on a characteristic CG length Λ. This is defined by including in the series expansion, Eq. 9, only those wave vectors k whose modulus is less than kM=2π/Λ. We refer to this as “wave vector cutoff of order nM,” where nM is an integer such that |k|2kM2=nM(2π/L)2 for all of the included k. Close to Tc the system is dominated by long-wavelength fluctuations and the presence of a cutoff should become physically irrelevant.
We consider as collective variables these Fourier components of the order parameter, 𝐬={ψk}|k|kM. In terms of 𝐬 the GL free energy can be rewritten as
where the integrals in Eq. 1 are rewritten in Fourier space as
and in the last convolution it is implied that if |qk|>kM, then ψqk=0. We note that because the order parameter is a real quantity, its Fourier transform has the property ψk=ψk. This symmetry property is used to reduce the number of independent collective variables. With a cutoff of order nM=3 we have 26 CVs, but in our simulations we have dealt with up to 364 CVs, in the nM=19 case (see Table S1).
Table S1.
Number of independent CVs for a given wave vector cutoff order nM

Bias Potential and Target Distribution

To proceed, we have to choose a variational form for V(𝐬) and introduce a suitable p(𝐬). As a general principle we use for V(𝐬) and p(𝐬) expressions that, when combined as in Eq. 5, lead to an F(𝐬) with exactly the functional form suggested by the considered CG model. In the present case we have as reference the GL free energy, Eq. 10, and thus we take
These two expressions can be used in Ω[V], which then can be minimized relative to the variational parameters gV, aV, and bV. At the minimum the estimated F(𝐬) has the desired GL structure, with parameters g=gtgV, a=ataV, and b=bV.
Because V(𝐬) in Eq. 14 has only a limited variational flexibility, our estimate of the free energy F(𝐬) is only approximated. However, given that our V(𝐬) is based on sound physical considerations, we still expect the F(𝐬) obtained to be a good approximation.
Using the p(𝐬) in Eq. 15 has some advantages. It is a product of univariate Gaussian distributions, so the averages over p(𝐬) that are needed to minimize Ω[V] (Eqs. 6 and 7) can easily be evaluated. And, more importantly, it can be used to guide sampling and accelerate the optimization convergence.
To understand this point we refer to Fig. 1, which gives an artist’s impression of a Landau free energy as a function of a one-dimensional order parameter as the system crosses the critical temperature Tc. For T>Tc the order parameter fluctuations are predominantly Gaussian and the effect of the non-Gaussian quartic term is hidden in the tail of the distribution, as pointed out also by other authors (10, 13). By using a p(𝐬) that is broader than the natural Gaussian fluctuations, we can enhance the probability with which the tails are sampled, thus improving the computation of the parameter b proportional to the quartic term I4(𝐬).
Fig. 1.
A simplified one-dimensional representation of the chosen quadratic target free energy Ft=(1/β)logp(𝐬) (green dashed line) and the expected Landau free energy (orange solid line).
For T<Tc we enter into the coexistence region, where different problems arise. The symmetry is broken, and the system spontaneously separates into two slab-shaped regions of finite but different value of the order parameter. These two phases, vapor and liquid, correspond to the two minima in Fig. 1, Right. By using a p(𝐬) that focuses on the transition region we can keep the system homogeneous and avoid to a certain extent some finite-size drawbacks (28). However, as the temperature is lowered, the interface between the two phases sharpens up with respect to the CG length Λ. This situation cannot be described quantitatively by a GL model that takes into account only long-wavelength fluctuations of the order parameter. In our simulations there is some clear evidence that the underlying model is no longer suited for describing the system. In fact, the convergence in the optimization of Ω[V] slows dramatically, and the obtained parameters no longer follow the expected linear behavior.
The use of a higher wave vector cutoff, and thus a smaller CG length Λ, can mitigate this issue, but only marginally. In fact, there is another important feature that is not taken into account by this simple model, that is, the so-called field mixing due to the lack of particle–hole symmetry (29), which induces an asymmetry between the two phases at low T. Nevertheless the GL model we adopted can still give us some relevant information, as shown at the end of Results.

Computational Setup

We simulate an Argon fluid, described by the classical two-body Lennard–Jones potential ϕLJ(r)=4ϵ[(σ/r)12(σ/r)6] truncated and shifted at Rc=2.5σ:
This simple system has been intensively studied (2831) and exhibits in its phase diagram the needed second-order phase transition, at the liquid–vapor critical point. All of the physical quantities in this paper are expressed in terms of Lennard–Jones reduced units.
We perform canonical molecular dynamics simulations at a fixed average density equal to the critical one, ρc=0.317 (30). We are aware that in a finite-size system the critical density is not exactly ρc, but for our purposes we just need to be close to that value. The temperature T is kept constant by the velocity-rescaling stochastic thermostat (32). The time step used is 0.002, and typically we used 107 steps for each run.
All of the simulations are performed with the molecular dynamics package Gromacs 5.0.4 (33), patched with a development version of the Plumed 2 (34) enhance sampling plugin, in which we implemented the VES method for our specific GL free-energy model.
The choice of gt and at, which define p(𝐬)in Eq. 15, was based on the amplitude of the fluctuations observed in the unbiased runs and on some trial-and-error simulations performed on small systems (N=128,256). The values thus obtained were then used as reference for the larger systems. The adopted values were between 0.05 and 0.2, and generally gt>at.


The Ginzburg–Landau model expects the parameter a to have a linear temperature dependence close to the critical point, and thus
whereas the other two parameters should be constant, g(T)=gc and b(T)=bc. We run multiple simulations varying both the size of the system and the cutoff order and consequently also the CG length Λ. As noted at the end of Bias Potential and Target Distribution, depending on the system size and Λ, convergence can become problematic at low temperatures. Nevertheless we can, somewhat conservatively, refer only to the region T>0.95Tc, where convergence is safely reached for all of the systems studied here. In this region our estimates are always in qualitative agreement with Landau’s ansatz, as can be seen in Fig. 2. The temperature dependence of the three parameters can be well fitted linearly, and all of the results obtained are provided in Table S2. Here we note only that the slope of g(T) and b(T) is roughly 10 times smaller than the one of a(T). This nonconstant behavior has been reported also in previous literature, for other Landau free-energy models (10).
Fig. 2.
(A–D) Landau’s parameters optimized for systems with (A) N=128, nM=3 and thus Λ=4.267; (B) N=2,048, nM=3 and thus Λ=10.75; (C) N=2,048, nM=19 and thus Λ=4.272; and (D) N=8,192, nM=3 and thus Λ=17.07. Data at temperatures where convergence was not clearly reachable are omitted. The linear fits are obtained using only data at T>0.95Tc, and error bars are smaller than the data points.
Table S2.
Landau parameters for different systems sizes N and wave vector cutoff orders nM
We recall that the coarse-grained length is Λ=N1/3/(ρ01/3nM) and that in the thermodynamic limit the critical temperature is Tc =1:08(5).
It is possible to define an effective critical temperature Tc as the temperature at which the parameter a(T) becomes zero. This is different from the critical temperature in the thermodynamic limit, Tc=1.085±0.005 (30), due to finite-size effects. Contrary to the effective critical temperature that could be defined by looking at the specific heat, Tc is always greater than Tc, as noted already in ref. 8. In Fig. 3 we show our estimate of Tc/Tc as a function of both the system size N and the wave vector cutoff order nM. As expected, increasing the size at fixed cutoff, we have that Tc approaches Tc. Maybe less intuitively, when we increase the number of wave vectors, Tc moves away from Tc. What we observe is that Tc is mainly linked to the coarse-graining length Λ; in fact, we find that roughly (TcTc)1/Λ. Despite the great variations in the value of Tc, all our data on a(T) can be collapsed into a single line when plotted as a function of (TTc)/Tc (Fig. 4). These are all empirical observations and do not necessarily imply rigorous scaling laws.
Fig. 3.
The effective critical temperature Tc/Tc for different system sizes N and wave vector cutoff orders nM. The dimension of the points is proportional to (Tc/Tc1). The background colors are related to the coarse-graining length ΛN1/3/nM, according to the scale on the right.
Fig. 4.
Parameter a, the one proportional to the quadratic term in Eq. 10, as a function of the rescaled temperature (TTc)/Tc. The effective critical temperature Tc is different for each system, as shown in Fig. 3. Only points with T>0.95Tc are plotted.

Simulations from the Coarse-Grained Model

Having obtained a coarse-grained description, we can ask to what an extent the GL model can represent the microscopic properties of the system. To try to answer this question, we run Monte Carlo simulations, using as variables the Fourier components ψk and as Hamiltonian the GL free energy. We define the model at all temperatures by assuming that the parameters depend linearly on T. Despite the limitations of the GL model (end of Bias Potential and Target Distribution), we find that some of the properties of our system are well described for a wide range of temperatures.
The most remarkable one is maybe the description of the Binder cumulant (35, 36), a quantity often used to obtain a good estimate of the critical temperature Tc, from finite-size simulations. Using our notation, we can handily express the Binder cumulant as
where is the ensemble average at temperature T, and I2 and I4 are defined in Eqs. 12 and 13, respectively. The important property of the Binder cumulant is that U(Tc) does not depend on the system size, and thus it is possible to evaluate Tc by looking at the crossing point of the curves obtained with different N.
The behavior of U(T) for different system sizes is reported in Fig. 5, obtained with unbiased molecular dynamics and Monte Carlo coarse-grained simulations. Remarkably the CG simulations well reproduce the U(T) curve, even at T<Tc, and give an estimate of TC in good agreement with the literature (30). To calculate I2 and I4 for Fig. 5 we used a wave vector cutoff of order nM=3, but it is important to note that the inclusion of higher-order wave vectors does not change the crossing point, even though the shape of the curves U(T) may change. This reflects the fact that the critical behavior is dominated by long-wavelength fluctuations.
Fig. 5.
The Binder cumulant U(T) defined in Eq. 18 obtained with full molecular dynamics simulations and with Monte Carlo simulations using the optimized GL free energy as a Hamiltonian. In both cases a wave vector cutoff of order nM=3 was used, and thus for the CG simulations only 26 independent variables were used for each system size. The interception of the curves gives an estimate of the critical temperature. We take as reference Tc=1.085 (30). Error bars are smaller than the data points, and lines are drawn only to guide the eyes.


In this paper we have shown that VES not only is a method for enhanced sampling, but also provides a formidable tool to approach coarse graining in a systematic and rigorous way. It combines the strength of relative entropy optimization methods with the intriguing ability to orient the sampling toward a desired statistical distribution.
In the present application we treated the case of a second-order phase transition, for which the well-established GL theory provided a good CG model, with a clear functional form for the free energy. In other fields of application careful consideration will have to be paid, and the structure of the CG model and the corresponding free energy might not be easily determinable. However, we can envisage several cases in which this might be possible. For instance, we could use as a CG model the classical density functional theory or models such as those introduced in refs. 3 and 4. Then VES provides a way to link rigorously and effectively the parameters of the coarse-graining model to the microscopic Hamiltonian.

Collective Variables

We report here explicitly the CVs used for our Lennard–Jones system. We can write the density of the system as a sum of delta functions centered in the atom’s positions Rj:
This is a safe approximation, because the wave vector cutoff will then actually limit our study to distances greater than Λ. Being the order parameter ψ(r)=(ρ(r)ρc)/ρ0, its Fourier components ψk are
and it is now easy to note that ψk=ψk, which halves the number of independent Fourier components.
The number of independent CVs is directly related to the cutoff order nM. For each independent Fourier component ψk, we have two independent CVs, its real and imaginary parts. The number of CVs as a function of nM is reported in Table S1.

Landau Parameters

We report here the Landau parameters obtained with the variational enhanced sampling method, as described in the main text. The parameters in Table S2 are obtained with a linear fit of the optimized parameters at temperatures
The lowest temperature for the fit is chosen in such a way that a good convergence is ensured at any T even for the bigger systems. Given Ginzburg–Landau free energy, Eq. S10, the fitting parameters are defined as follows:
The errors in our estimate of the GL parameters arise from two sources: the statistical error coming from the stochastic optimization and the error provided by the linear fit. The statistical error is significantly smaller than the fit error.


The authors thank Haiyang Niu for carefully reading the manuscript. The computational time for this work was provided by the Swiss National Supercomputing Center. This research was supported by the National Center for Computational Design and Discovery of Novel Materials MARVEL, funded by the Swiss National Science Foundation, and European Union Grant ERC-2014-AdG-670227/VARMET.

Supporting Information

Supporting Information (PDF)


LD Landau, On the theory of phase transitions. Collected papers of L. D. Laudau (Gordon and Breach, New York), pp. 193–216 (1965).
LD Landau, EM Lifshitz Statistical Physics (Elsevier, Oxford) Vol 5 (2013).
W Noid, et al., The multiscale coarse-graining method. I. A rigorous bridge between atomistic and coarse-grained models. J Chem Phys 128, 244114 (2008).
MG Saunders, GA Voth, Coarse-graining methods for computational biology. Annu Rev Biophys 42, 73–93 (2013).
O Valsson, M Parrinello, Variational approach to enhanced sampling and free energy calculations. Phys Rev Lett 113, 090601 (2014).
MS Shell, The relative entropy is fundamental to multiscale and inverse thermodynamic problems. J Chem Phys 129, 144108 (2008).
A Chaimovich, MS Shell, Coarse-graining errors and numerical optimization using a relative entropy framework. J Chem Phys 134, 094112 (2011).
K Kaski, K Binder, J Gunton, A study of a coarse-grained free energy functional for the three-dimensional Ising model. J Phys A Math Gen 16, L623–L627 (1983).
AP Giddy, MT Dove, V Heine, What do Landau free energies really look like for structural phase transitions? J Phys Condens Matter 1, 8327–8335 (1989).
S Radescu, I Etxebarria, J Perez-Mato, The Landau free energy of the three-dimensional Phi 4 model in wide temperature intervals. J Phys Condens Matter 7, 585–595 (1995).
M Gracheva, J Rickman, JD Gunton, Coarse-grained Ginzburg-Landau free energy for Lennard-Jones systems. J Chem Phys 113, 3525–3529 (2000).
J Íñiguez, S Ivantchev, J Perez-Mato, A García, Devonshire-Landau free energy of BaTiO3 from first principles. Phys Rev B 63, 144103 (2001).
A Tröster, C Dellago, W Schranz, Free energies of the ϕ4 model from Wang-Landau simulations. Phys Rev B 72, 094103 (2005).
A Tröster, C Dellago, Coarse graining the ϕ4 model: Landau-Ginzburg potentials from computer simulations. Ferroelectrics 354, 225–237 (2007).
A Laio, M Parrinello, Escaping free-energy minima. Proc Natl Acad Sci USA 99, 12562–12566 (2002).
O Valsson, P Tiwary, M Parrinello, Enhancing important fluctuations: Rare events and metadynamics from a conceptual viewpoint. Annu Rev Phys Chem 67, 159–184 (2016).
O Valsson, M Parrinello, Well-tempered variational approach to enhanced sampling. J Chem Theory Comput 11, 1996–2002 (2015).
J McCarty, O Valsson, P Tiwary, M Parrinello, Variationally optimized free-energy flooding for rate calculation. Phys Rev Lett 115, 070601 (2015).
J McCarty, O Valsson, M Parrinello, Bespoke bias for obtaining free energy differences within variationally enhanced sampling. J Chem Theory Comput 12, 2162–2169 (2016).
P Shaffer, O Valsson, M Parrinello, Enhanced, targeted sampling of high-dimensional free-energy landscapes using variationally enhanced sampling, with an application to chignolin. Proc Natl Acad Sci USA 113, 1150–1155 (2016).
P Shaffer, O Valsson, M Parrinello, Hierarchical protein free energy landscapes from variationally enhanced sampling. J Chem Theory Comput 12, 5751–5757 (2016).
PM Piaggi, O Valsson, M Parrinello, A variational approach to nucleation simulation. Faraday Discuss 195, 557–568 (2016).
R Rubinstein, The cross-entropy method for combinatorial and continuous optimization. Methodol Comput Appl Probab 1, 127–190 (1999).
I Bilionis, PS Koutsourelakis, Free energy computations by minimization of Kullback–Leibler divergence: An efficient adaptive biasing potential method for sparse representations. J Comput Phys 231, 3849–3870 (2012).
W Zhang, H Wang, C Hartmann, M Weber, C Schütte, Applications of the cross-entropy method to importance sampling and optimal control of diffusions. SIAM J Sci Comput 36, A2654–A2672 (2014).
A Isihara, The Gibbs-Bogoliubov inequality dagger. J Phys A Gen Phys 1, 539–548 (1968).
F Bach, E Moulines, Non-strongly-convex smooth stochastic approximation with convergence rate O (1/n). Advances in Neural Information Processing Systems 26 (NIPS 2013) (Curran Associates, Inc., Red Hook, NY), pp. 773–781 (2013).
K Binder, BJ Block, P Virnau, A Tröster, Beyond the van der Waals loop: What can be learned from simulating Lennard-Jones fluids inside the region of phase coexistence. Am J Phys 80, 1099–1109 (2012).
NB Wilding, Critical-point and coexistence-curve properties of the Lennard-Jones fluid: A finite-size scaling study. Phys Rev E 52, 602–611 (1995).
B Smit, Phase diagrams of Lennard-Jones fluids. J Chem Phys 96, 8639–8640 (1992).
J Pérez-Pellitero, P Ungerer, G Orkoulas, AD Mackie, Critical point estimation of the Lennard-Jones pure fluid and binary mixtures. J Chem Phys 125, 054515 (2006).
G Bussi, D Donadio, M Parrinello, Canonical sampling through velocity rescaling. J Chem Phys 126, 014101 (2007).
MJ Abraham, et al., GROMACS: High performance molecular simulations through multi-level parallelism from laptops to supercomputers. SoftwareX 1–2, 19–25 (2015).
GA Tribello, M Bonomi, D Branduardi, C Camilloni, G Bussi, PLUMED 2: New feathers for an old bird. Comput Phys Commun 185, 604–613 (2014).
K Binder, Finite size scaling analysis of Ising model block distribution functions. Z Med Phys B Condens Matter 43, 119–140 (1981).
DP Landau, K Binder A Guide to Monte Carlo Simulations in Statistical Physics (Cambridge Univ Press, Cambridge, UK, 2014).
P Raiteri, A Laio, FL Gervasio, C Micheletti, M Parrinello, Efficient reconstruction of complex free energy landscapes by multiple walkers metadynamics. J Phys Chem B 110, 3533–3539 (2006).

Information & Authors


Published in

Go to Proceedings of the National Academy of Sciences
Proceedings of the National Academy of Sciences
Vol. 114 | No. 13
March 28, 2017
PubMed: 28292890


Submission history

Published online: March 14, 2017
Published in issue: March 28, 2017


  1. coarse graining
  2. enhanced sampling
  3. Ginzburg–Landau free energy
  4. second-order phase transitions
  5. Lennard–Jones


The authors thank Haiyang Niu for carefully reading the manuscript. The computational time for this work was provided by the Swiss National Supercomputing Center. This research was supported by the National Center for Computational Design and Discovery of Novel Materials MARVEL, funded by the Swiss National Science Foundation, and European Union Grant ERC-2014-AdG-670227/VARMET.



Michele Invernizzi
Department of Physics, Eidgenössische Technische Hochschule (ETH) Zurich c/o Università della Svizzera Italiana (USI) Campus, 6900 Lugano, Switzerland;
Facoltà di Informatica, Instituto di Scienze Computationali, and National Center for Computational Design and Discovery of Novel Materials MARVEL, Università della Svizzera Italiana, 6900 Lugano, Switzerland;
Omar Valsson
Facoltà di Informatica, Instituto di Scienze Computationali, and National Center for Computational Design and Discovery of Novel Materials MARVEL, Università della Svizzera Italiana, 6900 Lugano, Switzerland;
Department of Chemistry and Applied Biosciences, ETH Zurich c/o USI Campus, 6900 Lugano, Switzerland
Michele Parrinello1 [email protected]
Facoltà di Informatica, Instituto di Scienze Computationali, and National Center for Computational Design and Discovery of Novel Materials MARVEL, Università della Svizzera Italiana, 6900 Lugano, Switzerland;
Department of Chemistry and Applied Biosciences, ETH Zurich c/o USI Campus, 6900 Lugano, Switzerland


To whom correspondence should be addressed. Email: [email protected].
Author contributions: M.I., O.V., and M.P. designed research, performed research, analyzed data, and wrote the paper.
Reviewers: D.F., University of Cambridge; and J.D.W., Institute for Physical Science and Technology, University of Maryland.

Competing Interests

The authors declare no conflict of interest.

Metrics & Citations


Note: The article usage is presented with a three- to four-day delay and will update daily once available. Due to ths delay, usage data will not appear immediately following publication. Citation information is sourced from Crossref Cited-by service.

Citation statements



If you have the appropriate software installed, you can download article citation data to the citation manager of your choice. Simply select your manager software from the list below and click Download.

Cited by


    View Options

    View options

    PDF format

    Download this article as a PDF file


    Get Access

    Login options

    Check if you have access through your login credentials or your institution to get full access on this article.

    Personal login Institutional Login

    Recommend to a librarian

    Recommend PNAS to a Librarian

    Purchase options

    Purchase this article to access the full text.

    Single Article Purchase

    Coarse graining from variationally enhanced sampling applied to the Ginzburg–Landau model
    Proceedings of the National Academy of Sciences
    • Vol. 114
    • No. 13
    • pp. 3265-E2799







    Share article link

    Share on social media