Simulation Techniques for Calculating Free Energies

Size: px
Start display at page:

Download "Simulation Techniques for Calculating Free Energies"

Transcription

1 Simulation Techniques for Calculating Free Energies M. Müller 1 and J.J. de Pablo 2 1 Institut für Theoretische Physik, Georg-August- Universität, Göttingen, Germany mmueller@theorie.physik.uni-goettingen.de 2 Department of Chemical and Biological Engineering, University of Wisconsin-Madison, Madison Wisconsin , USA depablo@engr.wisc.edu Marcus Müller and Juan J. de Pablo M. Müller and J.J. de Pablo: Simulation Techniques for Calculating Free Energies, Lect. Notes Phys. 703, (2006) DOI / c Springer-Verlag Berlin Heidelberg 2006

2 68 M. Müller and J.J. de Pablo 1 Introduction Methods Weighted Histogram Analysis Multicanonical Simulations Wang-Landau Sampling and DOS Simulations Successive Umbrella Sampling Configurations Inside the Miscibility Gap and Shape Transitions Applications Liquid-Vapor Coexistence Demixing of Binary Blends Compressible Blends Crystallization InterfaceFreeEnergies Protein Simulations Conclusions and Outlook References...119

3 Simulation Techniques for Calculating Free Energies 69 1 Introduction The study of phase transitions has played a central role in the study of condensed matter. Since the first applications of molecular simulations, which provided some of the first evidence in support of a freezing transition in hardsphere systems, to contemporary research on complex systems, including polymers, proteins, or liquid crystals, to name a few, molecular simulations are increasingly providing a standard against which to measure the validity of theoretical predictions or phenomenological explanations of experimentally observed phenomena. This is partly due to significant methodological advances that, over the past decade, have permitted study of systems and problems of considerable complexity. The aim of this chapter is to describe some of these advances in the context of examples taken from our own research. The application of Monte Carlo simulations for the study of phase behavior in fluids attracted considerable interest in the early 90 s, largely as a result of Monte Carlo methods, such as the Gibbs ensemble technique [1], which permitted direct calculation of coexistence properties (e.g., the orthobaric densities of a liquid and vapor phases at a given temperature) in a single simulation, without a need of costly thermodynamic integrations for the calculation of chemical potentials. Perhaps somewhat ironically, some of the latest and most powerful methods for the study of phase behavior have reverted back to the use of thermodynamic integration, albeit in a different form from that employed before the advent of the Gibbs ensemble technique. One of the central concepts that paved the way for the widespread use of free-energy based methods in simulations of phase behavior was the conceptually simple, but highly consequential realization that histograms of data generated at one particular set of thermodynamic conditions could be used to make predictions about the behavior of the system over a wide range of conditions [2, 3]. The so-called weighted histogram analysis or histogram-reweighing technique constitutes an essential component of the methods presented in this chapter, and we therefore begin with a brief description of its implementation. We then discuss multicanonical simulation techniques, Wang-Landau sampling and extensions, and successive umbrella sampling. We close this section on Methods with a discussion of the configurations that one encounters when one uses the order parameter of a first order phase transition as a reaction coordinate between the two coexisting phases. The methodological section is followed by several applications that illustrate the computational methods in the context of examples drawn from our own research. The chapter closes with a brief look ahead.

4 70 M. Müller and J.J. de Pablo 2 Methods 2.1 Weighted Histogram Analysis In the simplest implementation of a molecular simulation, data corresponding to a specific set of thermodynamic conditions (e.g., number of particles, n, volume, V, and temperature, T, for a canonical ensemble) are used to estimate the ensemble or time average of a property of interest. For a canonical ensemble, for example, the outcome of such a simulation would be a pressure p corresponding to n, V,andT. Histogram reweighting techniques permit calculation of ensemble averages over a range of thermodynamic conditions without having to perform additional simulations. The underlying ideas, introduced by Ferrenberg and Swendsen in the late 80 s [2, 3], provide a simple means for extrapolating data generated at one set of conditions to nearby points in thermodynamic space. For a canonical ensemble, the probability of observing a configuration having energy E at an inverse temperature β =1/k B T (where k B is Boltzmann s constant) takes the form P β (E) = Ω(E)exp( βe) Z(β), (1) where Ω(E) is the density of states of the system and Z(β) is the canonical partition function, given by Z(β) = E Ω(E)exp( βe). (2) The probability distribution at a nearby inverse temperature, denoted by β, can be expressed in terms of the distribution at β according to P β (E) = P β(e)exp[(β β )E] E P β(e)exp[(β β )E]. (3) In other words, by using (3) one can extrapolate data generated at temperature T to nearby temperature T. The same idea can be used to interpolate data generated from multiple simulations [3]. Consider a series of canonical ensemble Monte-Carlo simulations conducted at r different temperatures. The n th simulation is performed at β n, and the resulting data are stored and sorted in N n (E) histograms, where the total number of entries is n n. The probability distribution corresponding to an arbitrary temperature β is given by P β (E) = r n=1 N n(e)exp( βe) r n=1 n n exp( β n E f n ) (4) where exp[f n ]= E P βn (E), (5)

5 Simulation Techniques for Calculating Free Energies 71 and where f n is a measure of the free energy of the system at temperature β n. The value of f n can be found self-consistently by iterating (4) and (5). Multiple histogram reweighing can therefore yield probability distributions over a broad range of temperatures, along with the corresponding relative free energies corresponding to that range. Equation (4) was derived by minimizing the error that arises when all histograms are recombined to provide thermodynamic information over the entire range of conditions spanned by distinct simulations. Note that histograms corresponding to neighboring conditions must necessarily overlap in order for (4) to be applicable. 2.2 Multicanonical Simulations The range of applicability of histogram extrapolation is large if the fluctuations that arise during the course of the simulation sample states that are also representative of neighboring thermodynamic conditions. This is often the case for small systems, or in the vicinity of a critical point. For large systems, or remote from a critical point, fluctuations are smaller and it is advantageous to sample configurations according to a non-boltzmann statistical weight, thereby coercing the system to sample configurations that are representative of a wide interval of thermodynamic conditions. Thermal averages, such as the internal energy, E, can then be obtained via a time average over the weighted sequence of visited states. Note, however, that in such a sampling scheme the entropy, S, or free energy, F, cannot be calculated in a direct manner because those quantities cannot be expressed as a function of the particle coordinates; special simulation techniques are required to estimate S and F. We now discuss several techniques that generate configurations according to weights that are generally constructed in such a way as to provide a more uniform sampling of phase space than the Boltzmann weight. It is instructive to note that a number of simulation methods e.g., multicanonical [4, 6, 21] or Wang-Landau techniques [7,8], have been originally formulated in terms of the pair of thermodynamically conjugated variables consisting of the temperature and the energy. Such methods, however, can be carried over to arbitrary pairs of order parameter or reaction coordinate and conjugated field, e.g., (magnetization and magnetic field), (composition and exchange potential), or (number of particles and chemical potential), by simply replacing the energy by the order parameter and the temperature by the thermodynamically conjugated field. The free energy, F, can be obtained by thermodynamic integration of the specific heat, c V, from a reference state interval along a thermodynamically reversible path: βf = βe S k B = βe ( ) S T 0 + dt c V (T ) k B T 0 k B T. (6) The specific heat can in turn be obtained in canonical-ensemble simulations from the fluctuations of the internal energy,

6 72 M. Müller and J.J. de Pablo c V de dt = E2 E 2 k B T 2. (7) In order to estimate the free energy many canonical simulations at different temperatures are necessary; furthermore, it is often difficult to define a suitable reference state with a known entropy S 0. Two alternatives can be followed to overcome these difficulties: (i) expanded ensemble methods and (ii) multicanonical methods. In the expanded ensemble method [9], one considers the conjugated field as a Monte Carlo variable and one introduces moves that allow for altering the conjugated field. Each thermodynamic integration can be cast into this form. The conjugated field over which one integrates (e.g., temperature) can adopt discrete values T 1 T i T 2 in the interval of interest, including the boundary values. The values have to be chosen such that the order parameter distributions in the canonical ensemble overlap for neighboring values of the conjugated fields, T i and T i+1. States of the expanded ensemble are characterized by the particle coordinates, {r i }, and the conjugated field, T i.the partition function of the expanded ensemble takes the form Z ex = e w(ti) Z(n, V, T i )= e w(ti) D[{r i }]exp( β i E[{r i }]) T 1 T i T 2 T 1 T i T 2 (8) where Z(n, V, T i )=e βif denotes the canonical-ensemble partition function. Here we have introduced weight factors, w(t i ), that depend only on the value of the conjugated field but not on the microscopic particle configuration, {r i }. They have to be chosen in such a way as to generate an approximately uniform sampling of the different conjugated fields, T i, throughout the course of the simulation. The probability of finding the system in state T i is given by P (T i )= e w(ti) Z(n, V, T i ) Z ex. (9) To achieve uniform sampling of the different canonical ensembles within an expanded ensemble simulation, the weights should ideally obey P (T i ) const w(t i ) ln Z(n, V, T i ) + const. (10) The free energy is then given by: βf = w(t ) ln P (T ) ln Z ex. (11) Formally, these equations are valid for an arbitrary choice of the weight factors; a bad choice, however, does not allow for sufficient sampling of all values of the conjugated field and dramatically reduces the efficiency of the method. The problem of calculating the free energy is therefore shifted to that of obtaining appropriate weights. An optimal choice of the weights, however, requires a working estimate of the free energy difference between the states.

7 Simulation Techniques for Calculating Free Energies 73 The second scheme multicanonical simulations [4, 6] generates configurations according to a non-boltzmann distribution designed to sample all values of the order parameter (energy) that are pertinent to the interval of the thermodynamic field (temperature) of interest. Multicanonical methods are advantageous for the study of first order phase transitions, where the order parameter connects the two coexisting phases via a reversible path (see Sect. 2.5), or in circumstances where the conjugated field does not have a simple physical interpretation (e.g., crystalline order in Sect. 3.4). Typically, a canonical simulation can only be reweighted by histogram extrapolation techniques [2] in the vicinity of the temperature at which the canonical distribution was sampled. In contrast, a single multicanonical simulation permits calculation of canonical averages over a range of temperatures which would require many canonical simulations. This feature led to the name multicanonical [10]. A multicanonical simulation generates configurations according to the partition function Z muca = D[{r i }]e w(e[{ri}]) (12) where w(e) is a weight function that only depends on the internal energy but not explicitly on the particle coordinates, {r i }.IfΩ(E) denotes the density of states that corresponds to a given value of the order parameter (i.e., the microcanonical partition function, if we utilize the energy), then the probability of sampling a configuration with order parameter E is given by: P muca (E) Ω(E)e w(e) (13) and the choice w(e) = lnω(e) + C, where C is a constant, leads to uniform sampling of the energy range of interest. Boltzmann averages of an observable O at temperature T can be obtained by monitoring the average of this observable, O(E), at a given energy, E, using O nv T deω(e)e βe O(E) deω(e)e βe = depmuca (E)e w(e) βe O(E) depmuca (E)e w(e) βe (14) Note that for the specific choice, w(e) = E/k B T 0, a multicanonical simulation corresponds to a canonical simulation at temperature T 0, and the equation above corresponds to the Ferrenberg-Swendsen weighted histogram analysis [2, 3] (cf. (4)). The goal of multicanonical sampling is to explore a much wider range of energies in the course of the simulation, thereby enlarging the extrapolation range. The free energy can be obtained from βf ln de Ω(E)e βe = ln de P muca (E)e w(e) βe +C. (15) Formally, these equations are valid for an arbitrary choice of the weight function, w(e), but a failure to sample all configurations that significantly contribute to canonical averages for a temperature in the interval T 1 <T <T 2

8 74 M. Müller and J.J. de Pablo introduces substantial sampling errors. A necessary condition is that the simulation samples all energies in the interval [ E nv T1 : E nv T2 ] with roughly equal probability. Again, the problem of efficient multicanonical sampling is shifted to that of obtaining the appropriate weights. The optimal choice of the weights, however, requires a working estimate of the density of states with a given order parameter. Several methods to obtain these weights are discussed in the following section. 2.3 Wang-Landau Sampling and DOS Simulations As described above, traditional multicanonical algorithms [5] provide an estimate of the density of states through weights, w(e), constructed to facilitate or inhibit visits to particular state points according to the frequency with which they are visited. The calculation of the weights is necessarily iterative and, depending on the nature of the problem (e.g., the roughness of the underlying free energy profile), can require considerable oversight. In recent years, a different class of algorithms has emerged for direct calculation of the density of states from Monte Carlo simulations [7,8,11,12]. We refer to these algorithms as density-of-states (DOS) based techniques. In a recent, powerful implementation of such algorithms by Wang and Landau, a random walk in energy space has been used to visit distinct energy levels [7]; the density of states corresponding to a particular energy level is modified by an arbitrary factor when that level is visited. By controlling and gradually reducing that factor in a systematic manner, an estimate of the density of states is generated in a self-consistent and self-monitoring manner. The algorithm relies on a histogram of energies to dictate the rate of convergence of a simulation [7]. A random walk is generated by proposing random trial moves, which are accepted with probability p = min [1,Ω(E 1 )/Ω(E 2 )], where E 1 and E 2 denote the potential energy of the system before and after the move, respectively. Every time that an energy state E is visited, a running estimate of the density of states is updated according to Ω(E) =fω(e), where f is an arbitrary convergence factor. The energy histogram is also updated; once it becomes sufficiently flat, a simulation stage is assumed to be complete, the energy histogram is reset to zero, and the convergence factor f is decreased in some prescribed manner (e.g., f = f). The entire process is repeated until f is very close to 1 (the original literature recommends that ln(f) attain a value of 10 9 ). Wang and Landau s original algorithm focused on an Ising system in the canonical ensemble [7]. It was later extended to systems in a continuum and to other ensembles [8, 13 16] and found to be highly effective. Because the running estimate of the density of states changes at every step of the simulation, detailed balance is never satisfied. In practice, however, the convergence factor decreases exponentially, and its final value can become so small as to render the violation of detailed balance essentially non-existent. Given that the convergence factor decreases as the simulation proceeds, configurations generated at different stages of the simulation do not contribute

9 Simulation Techniques for Calculating Free Energies 75 equally to the estimated density of states. In the final stages of the simulation the convergence factor is so small that the corresponding configurations make a negligible contribution to the density of states. In other words, many of the configurations generated by the simulation are not used efficiently. As described below, one can alleviate this problem by integrating the temperature [17, 18]. The internal energy of a system is related to entropy S and volume V through de = T ds p dv, (16) where p is the pressure. The temperature of the system is related to the density of states Ω(n, V, E) by Boltzmann s equation [19] according to ( ) ( ) 1 S ln Ω(n, V, E) T = = k B. (17) E E V Equation (17) can be integrated to determine the density of states from knowledge of the temperature: ln Ω(n, V, E) = E E 0 1 k B T V de. (18) Equation (18) requires that the temperature be known as a function of energy. Evans et al. [20] have shown that an intrinsic temperature can be assigned to an arbitrary configuration of a system. This so-called configurational temperature is based entirely on configurational information and is given by i F i 1 i =, (19) k B T config F i 2 where subscript i denotes a particle, and F i represents the force acting on that particle. This configurational temperature can be particularly useful in Monte Carlo simulations, where kinetic energy is not explicitly involved. In the past it has been proposed as a useful tool to diagnose programming errors [21]. The estimator for the configurational entropy can be exploited in the context of DOS simulations in the following way: Four histograms are collected during a simulation; one for the density of states, one for the potential energy, one for the numerator of (19), and one for its denominator [17]. Note that two independent sets of density of states are available at the end of each stage: one from the original density-of-states histogram, and one from the configurational temperature, which can be integrated to provide Ω. In principle, either set can be used as a starting point for the next stage of a simulation. In practice, however, using a combination of these is advantageous. i

10 76 M. Müller and J.J. de Pablo In the early stages of the simulation, detailed balance is grossly violated as a result of the large value of f. The resulting estimates of configurational temperature are therefore incorrect. In order to avoid contamination of late stages by the transfer of incorrect information, the temperature accumulators can be reset at the end of the early stages, once the density of states is calculated from the temperature. For small enough convergence factors (e.g., ln f<10 5 ), the slight violations of detailed balance incurred by the method become negligible, and the temperature accumulators need no longer be reset at the end of each stage. Clearly, in the proposed configurational temperature algorithm, the dynamically modified density of states only serves to guide a walker through configuration space. The true, thermodynamic density of states is calculated from the configurational temperatures accumulated in the simulation. Because all configurations generated in the simulation contribute equally to the estimated temperature, configurations generated at various stages of the simulation contribute equally to Ω. For lattice systems, or for systems interacting through discontinuous potential energy functions (e.g., hard-spheres), (19) cannot be used. In such cases, one can introduce a temperature by resorting to a microcanonical ensemble formulation [22, 23] of the density-of-states algorithm [18]. Figure 1 shows results for a truncated-and-shifted Lennard-Jones fluid (the potential energy is truncated at r c =2.5σ); σ and ɛ denote the length and energy scale of the potential. Results are shown for a system of n = 400 particles at a density of ρσ 3 = , well within the liquid regime [17]. For the random walk algorithm with configurational temperature, the energy window is set to 1930 E/ɛ < 1580; for the multi-microcanonical ensemble simulation, the energy window is 1500 E/ɛ < 500. In both cases, the energy window corresponds roughly to temperatures in the range 0.85 <T k B T/ɛ < 1.5 (i.e. above and below the critical point). In random walk simulations with configurational temperature, the calculations are started with a convergence factor f = exp(0.1). When f > exp(10 5 ), the density of states calculated from the temperature is used as the initial density of states for the next stage, the convergence factor is reduced by f k+1 = f 1/10 k, and the temperature accumulators are reset to zero at the end of each stage. For later stages, e.g., f<exp(10 5 ), the density of states generated by the random walker is carried over to the next stage and the convergence factor is reduced according to f k+1 = f k. The simulation is terminated when the convergence factor satisfies f<exp(10 10 ). To estimate the statistical errors in the estimated density of states, 7 independent simulations are conducted, with exactly the same code, but with different random-number generator seeds. The calculated density of states always contains an arbitrary multiplier; the estimated densities of states from these 7 runs are matched by shifting each ln Ω(E) in such a way as to minimize the total variance.

11 Simulation Techniques for Calculating Free Energies Statistical Errors in lnω(e) Convergence Factor (ln f) Fig. 1. Statistical errors in the density of states as a function of convergence factor [17] The diamonds in Fig. 1 show results from the Wang and Landau original algorithm. These errors exhibit two distinct regimes. For large f (e.g., f>10 4 ), the statistical error is proportional to f. In the small f regime (f < 10 6 ), the error curve levels off, and asymptotically converges to a limiting value of approximately 0.1. The main reason for this is that in the Wang-Landau algorithm, configurations generated at different stages of the simulation do not contribute equally to the histograms. The Wang-Landau algorithm leaves the density of states essentially unchanged once f is reduced to less than 10 6 ; additional simulations with smaller f only polish the results locally, but do little to decrease the overall quality of the data. If phase space has not been ergodically sampled by the time f reaches about 10 6,the final result of the simulation is likely to be inaccurate. Using a more stringent criterion for flatness only alleviates the problem partially, because the computational demands required to complete a stage increase dramatically. The squares in Fig. 1 show the statistical errors in the density of states calculated from configurational temperature. In contrast to the curve generated through a conventional random walk, the errors steadily decrease as the simulation proceeds. The figure also shows that the statistical errors from the configurational temperature method proposed here are always considerably smaller than those from the Wang-Landau technique. At the end of a simulation, the statistical error from a Wang-Landau calculation is approximately 5 times larger than that from configurational temperature. In other words, thermodynamic-property calculations of comparable accuracy would be 25 times faster in the proposed configurational algorithm than in existing random-walk algorithms.

12 78 M. Müller and J.J. de Pablo 0.10 Systematic Errors in lnω(e) Potential Energy (E) Fig. 2. Systematic errors in configurational temperatures in the first 4 stages of the simulation. With decreasing magnitude of the errors, the corresponding modification factors are ln f = 0.1, 0.01, 0.001, and , respectively Figure 2 shows the difference between the configurational temperature corresponding to the first 4 stages and the final configurational temperature. The figure shows that in the first few stages, the calculated configurational temperature differs from the true value in a systematic manner. The reason for such deviations is that in the first few stages the detailed balance condition is grossly violated. For a large convergence factor, the system is expelled into neighboring energy levels, even if the trial configuration does not conform to that energy level. The figure also shows that at the fourth stage (ln f =10 4 ), such systematic deviations have already become negligible and are smaller than the random errors; the configurational temperature accumulators need not be reset at the end of each stage once ln f Figure 3 compares the statistical errors in the heat capacity calculated from a Wang-Landau algorithm, from multi-microcanonical ensemble simulations, and from configurational temperature, as a function of actual CPU time. The error in the original Wang-Landau algorithm becomes almost constant once the simulation is longer than about 8 hours [17]. The error from multi-microcanonical ensemble and configurational-temperature simulations are smaller, and continue decreasing as the simulation proceeds. Recently, it has been proposed that a more efficient version of the original Wang-Landau algorithm can be devised by removing the flatness constraint from the energy histograms [24]. The underlying premise behind non-flat histograms is that the algorithm should maximize the number of round trips of a random walk within an energy range. While that idea has been shown to be effective in the context of spin systems, it is of limited use in the context of complex fluids. The method requires a good initial guess of the density of

13 Simulation Techniques for Calculating Free Energies Statistical Errors in Cv* CPU Time (Hour) Fig. 3. Statistical errors in heat capacity as a function of CPU time. The squares are the results from the Wang-Landau algorithm; the circles are the results from configurational temperature calculations, and the diamonds are the results from multi-microcanonical ensemble simulations [17] states, which can only be obtained after considerable effort (or multiple iterations of the Wang-Landau algorithm). Our experience suggests that it is more effective to resort to a configurational temperature approach, with multiple overlapping windows [17, 18] whose width and number is adjusted to sample difficult regions of phase space more or less exhaustively. The implementation of configurational temperature density-of-states simulations is illustrated here in the context of a protein. The particular protein discussed in this case is the C-terminal fragment of protein G, which has often been used as a benchmark for novel protein-folding algorithms [14]. This 16-residue peptide is believed to assume a β-hairpin configuration in solution. In this example it is modelled using an atomistic representation and the CHARMM19 force field with an implicit solvent [14, 25, 26]. The total energy range of interest is broken up into smaller but slightly overlapping energy windows [E k : E+ k ] (see Fig. 4). In the spirit of parallel tempering [27], neighboring windows are allowed to exchange configurations. The width of individual windows can be adjusted in such a way as to have more configurations in regions where sampling is more demanding. As shown in Fig. 5, high-energy windows include unfolded configurations, whereas low-energy windows are populated by folded configurations. This density of states can subsequently be used to determine any thermodynamic property of interest over the energy range encompassed by the simulation. The heat capacity, for example, which can be determined from fluctuations of the energy according to (7), can be determined to high accuracy and used to extract a precise temperature for the folding transition of this polypeptide. Furthermore, these high accuracy

14 80 M. Müller and J.J. de Pablo (a) Potential Energy (U) kj/mol MC Steps x 10 6 (b) Histogram H (U) Potential Energy (U) kj/mol Fig. 4. (a) Evolution of energy in different energy windows during a configurationaltemperature DOS simulation of the C-terminal fragment of Protein G. (b) Representative histogram of visited energy states in overlapping windows [18] estimates of c v over a relatively wide energy range can be generated on the order of 100 hours of CPU time on a Pentium architecture. Our description thus far of DOS-based methods has centered on calculation of the density of states. A particularly fruitful extension of such methods involves the calculation of a potential of mean force (Φ), or PMF, associated with a specified generalized reaction coordinate, ξ(r) [28,29]. A PMF measures the free energy change as a function of ξ(r) (where r represents a set of Cartesian coordinates). This potential is related to the probability density of finding the system at a specific value ξ of the reaction coordinate ξ(r): P (ξ(r) =ξ) Ce βφ(ξ), (20) where C is a normalization constant. As we shall later see, the methods of umbrella sampling rely on (20) for estimating free energy changes by altering the potential function with a biasing function designed to sample phase space more efficiently. This bias is later removed and the simulation data are

15 Simulation Techniques for Calculating Free Energies 81 Fig. 5. Representative configurations of the C-terminal fragment of protein G corresponding to different energy ranges reweighted to arrive at the correct probability distribution. The potential of mean force is then calculated from: Φ(ξ) = k B T ln P (ξ)+c. (21) Other methods arrive at the potential of mean force by calculating the derivative of the free energy with respect to the constrained generalized coordinate ξ in a series of simulation runs. A mean force, F ξ = (Φ(ξ)) (ξ), can be integrated numerically to yield the corresponding PMF. These simulations involve constraints, and an appropriate correction term must therefore be accounted for in the calculation of the PMF. For a system with no constraints and consisting of N particles with Cartesian coordinates r =(r 1, r 2,...,r N ), the mean force can be written as: [ r E ] ξ r + k ln J BT δ(ξ(r) ξ) r F ξ =, (22) δ(ˆξ(r) ξ)

16 82 M. Müller and J.J. de Pablo where E is the potential energy function and J is the Jacobian associated with the transformation from Cartesian to generalized coordinates. To compute the mean force acting on the end-to-end distance of a molecule, for example, a suitable reaction coordinate is provided by ξ = r ij = r i r j, where r ij represents the distance between the two terminal sites i and j. The Jacobian, J, is a function of the separation r ij and can be taken out of the ensemble average to yield [29]: F = E + 2k BT. (23) ξ ξ ξ ξ The mean force therefore includes a contribution from the average mechanical force and another contribution arising from the variations of the volume element associated with the reaction coordinate ξ. The free energy change between two states ξ 1 and ξ 2 can be obtained by integrating (23) according to ξ2 ( ) E ξ2 Φ(ξ 2 ) Φ(ξ 1 )= dξ 2k B T ln. (24) ξ ξ 1 Constrained simulations rely on the calculation of the first term of (23) from a series of simulations conducted at different values of ξ. This average force is then corrected by adding the second term of (23), and then numerically integrated to give a potential of mean force for the desired range of ξ. As noted above, the density-of-states method described earlier can be extended to yield accurate estimates of the potential of mean force. The weight factors that dictate the walk in the ξ space can be computed on the fly during a simulation in a self-consistent manner. The simulation can be performed without any constraints, which means that the resulting weights can be used directly as in (21) to give the potential of mean force. One can also accumulate the forces acting on the particles that define the reaction coordinate and then use (24) to get the PMF. The computed PMF is therefore available as a continuous function of ξ. In recent work [28,29] we have explored the use of DOS methods in the context of expanded ensembles, where intermediate states are introduced to facilitate the transition between configurations separated by large energy barriers. We refer to the resulting methods through the acronym EXEDOS (for Expanded Ensemble Density of States). The expanded states are usually defined by some reaction coordinate, ξ, and the sampling in ξ space is governed by unknown weights. This so-called expanded-ensemble density of states method has been employed for studies of suspensions of colloidal particles in liquid crystals [28], colloidal suspensions in polymer solutions [30], and proteins on a surface [31]. In one of the examples at the end of this chapter we discuss it in the context of reversible, mechanical stretching of proteins [32]. In that example, the reaction coordinate, ξ, is chosen to be the end-to-end distance between the N and C terminus of the molecule being stretched. In a different ξ 1

17 Simulation Techniques for Calculating Free Energies 83 example we discuss it in the context of crystallization [33], where a more elaborate order parameter is necessary. The goal of the method is to perform a random walk in ξ space. Consider a system consisting of N particles interconnected to form a molecule, and having volume V and temperature T. The end-to-end distance of the molecule (ξ) can be discretized into distinct states; each state is characterized by its end-to-end distance, ξ, in some specified range of interest [ξ,ξ + ]; ξ and ξ + represent a lower and an upper bound, respectively. The partition function Ω of this expanded ensemble is given by Ω = Q(N,V,T,ξ)g(ξ)dξ = Q ξ g(ξ)dξ, (25) where Q ξ is the canonical partition function for that particular state ξ, and g(ξ) is the corresponding weight factor. The probability of visiting a state having extension ξ can be written as P (ξ) = Q ξg(ξ) Ω. (26) The free energy difference between any two states can therefore be calculated from the weight factors, and the population density can be determined from Φ(ξ 2 ) Φ(ξ 1 ) k B T ln Q ξ 2 Q ξ1 = k B T [ ln g(ξ 1) g(ξ 2 ) +lnp (ξ ] 2) P (ξ 1 ). (27) If each state is visited with equal probability, the second term on the right of (27) disappears and the PMF can be computed from Φ(ξ) = k B T ln g(ξ)+c. (28) In the method discussed here, a running estimate of the weight factors can be computed and refined in a self-consistent and self-monitoring manner. At the beginning of a simulation, g(ξ) is assumed to be unity for all states. Trial Monte Carlo moves are accepted with probability [ P acc (ξ 1 ξ 2 ) = min 1, g(ξ ] 1) g(ξ 2 ) exp( β E), (29) where ξ 1 and ξ 2 are the end-to-end distances of the system before and after a trial move. After each trial move, the corresponding weight factor is updated by multiplying the current, existing value by a convergence factor, f, that is greater than unity (f > 1), i.e., g(ξ) g(ξ) f. Every time that g(ξ) is modified, a histogram H(ξ) is also updated. As before, this g(ξ) refinement process continues until H(ξ) becomes sufficiently flat. Once this condition is satisfied, the convergence factor is reduced by an arbitrary amount. We use again f new = f old. The histogram is then reset to zero (H(ξ) = 0), and a

18 84 M. Müller and J.J. de Pablo new simulation stage is started. The process is repeated until f is sufficiently small. In addition to computing these weight factors from the histograms of visited states, one can obtain a second estimate from the integration of the mean force, as given by (24). The first term on the right hand side of (23) can be estimated in a density of states simulation. The component of the total force acting on the two sites that define the reaction coordinate along the end-toend vector, ˆξ, is accumulated as a function of ξ. At the end of the simulation this mean force is corrected by adding the corresponding second term of (23), and then integrated to yield the PMF. As mentioned above, in the earlier stages of the simulation (ln f>10 5 ), when the convergence factor is large, detailed balance is severely violated. As a result, thermodynamic quantities computed during this time (including average forces) are incorrect. To avoid carrying this error into later stages, the accumulators for average forces are reset at the end of early stages. As the convergence factor decreases (e.g., ln f<10 5 ), the violation of detailed balance has a smaller effect, and the accumulators need not be reset anymore. 2.4 Successive Umbrella Sampling A complementary and alternative technique to reweighting methods are umbrella sampling strategies. The guiding idea of umbrella sampling [34] is to divide the pertinent range of the order parameter, E, into smaller windows and to investigate one small window after the other. If the windows are sufficiently narrow, the free energy profile will not substantially vary within a window, and a crude estimate of the weights is often sufficient to ensure uniform sampling [35]. In the limiting case, the windows just consist of two neighboring values of the order parameter. A histogram H k (E) monitors how often each state is visited in the k th window [E k : E+ k ]. Care must be exercised at the boundaries of a window to fulfill detailed balance [36]: If a move attempts to leave the window, it is rejected and H(window edge) is incremented by unity. Another question which may arise from the discussion of the boundary is the optimum amount of overlap to minimize the uncertainty of the overall ratio. Here we choose the minimal overlap of one state at the interval boundaries, i.e., E + k = E k 1. This is simple to implement and sufficient to match the probability distributions at their boundaries. A larger overlap may reduce the uncertainty but requires a higher computational effort. Let H k H k (E k )andh k+ H k (E + k ) denote the values of the kth histogram at its left and right boundary, and R k H k+ /H k characterize their ratio. After a predetermined number of Monte Carlo steps per window, the (unnormalized) probability distribution can recursively be estimated according to:

19 Simulation Techniques for Calculating Free Energies 85 P (E) P (E0 ) = H 0+ H1+ Hk(E) = Π k 1 i=1 H 0 H 1 H R i Hk(E) with R i H i+, k H k H i (30) when E [E k : E+ k ]. The ratios in (30) correspond to the Boltzmann factor associated with the free energy difference across the order parameter interval i. We now consider how the overall error depends on the choice of the window size [35]: For clarity, we assume that no sampling difficulties are encountered in our system, i.e., the order parameter in which we reweight (e.g., the energy, E, or the particle number, n) is suitable to flatten and overcome all barriers in the (multidimensional) free energy landscape and the restriction of fluctuations of the average order parameter by the window size does not impart sampling errors onto the simulation [37]. Under these circumstances, it has been suggested [38, 39] that small windows reduce computational effort by a factor of N k, where N k denotes the total number of windows into which the sampling range is subdivided: The time τ to obtain a predetermined standard deviation δ of the ratio R k in a single window is proportional to the square of the window size, k = E + k E k. With τ 2 k, we get a total computation time (for all windows) of t cpu N k 2 k, implying that the overall error of the simulation is also δ. This contrasts the behavior of a single large window; N m =1and m = N k k yield t cpu =(N k k ) 2 = N k t cpu,which suggests that a window size as small as possible should be chosen [38, 39]. Being interested in localizing phase coexistence, however, the pertinent error is related to the free energy difference of the two end points of the order parameter interval that corresponds to the distinct phases. In this case, we have to account for error propagation in (30) and we obtain for the error,, ofthe ratio P (E + N k )/P (E1 ) ( ) P (E + N δ k ) P (E1 ) = N k δrk 2 O(δ N k ). (31) Due to error propagation across the windows, the error of each individual subinterval has to be smaller than the total error,, by a factor of N k.thus, the time that must be spent in each window to achieve an error of δ = / N k is N k τ and the total simulational effort is N 2 k τ (N k k ) 2 N k t cpu,whichis identical to the time required for a single large window. This argument implies that the statistical error for a given computational effort is independent from the window size, i.e., the number of intervals the range of order parameter is divided into. A more detailed analysis including possible systematic errors due to (i) estimating probability ratios and (ii) correlations between successive windows can be found in [35]. The computational advantage from subdividing the range of order parameter into windows does not stem from an increased statistical accuracy due to the small correlation times within a window, but from the fact that successive umbrella sampling does not require the independent and computationally costly generation of high-quality weights. Making the window size small we k=1

20 86 M. Müller and J.J. de Pablo reduce the variation of the free energy across a single interval and achieve sufficiently uniform sampling without or with a very crude approximation of the weights. In this sense, successive umbrella sampling is as efficient as a multicanonical simulation with very good weights, except that now the weights need not be known beforehand. Additionally, the scheme is easy to implement on parallel computers and the range of order parameter can be easily enlarged by simply adding more windows. This computational scheme has successfully been applied to study phase equilibria and interface properties in polymer-solvent mixtures [40,41], polymer-colloid mixtures [42 45] and liquid crystals [46, 47]. As we are sampling one window after the other, efficiency can be increased by combining the scheme with the multicanonical concept. In a multicanonical simulation we replace H(E) in (30) by H(E)exp[w(E)]. In principle, one could use Wang-Landau sampling to estimate the weights on the fly. Given that the free energy differences are small a rather crude method often suffices: After P (E) is determined according to (30), w(e) = ln[p (E)] is extrapolated into the next window. The first window is usually unweighted. If, in this case, states are not accessible, w(e) can be altered by a constant amount of k B T in each iteration step. In the limit that each window only contains two states we use linear extrapolation for the second and quadratic extrapolations for all subsequent windows. Then, extrapolated and true values for P (E) only differ by a few percent. The basic idea behind this kind of extrapolation is to flatten the free energy landscape even in the small windows. Depending on the steepness of the considered landscape, this may lead to a significant reduction of the error vis-à-vis unweighted umbrella sampling. Additionally, one should keep in mind that multicanonical simulations in a single, one-dimensional order parameter are not always sufficient when free energy landscapes become more complex, and that the restriction of fluctuations of the average order parameter in a window can impart sampling problems. The problem of simulating too small windows with umbrella sampling is well known, but difficult to quantify [34]. Hence, in practice, the optimum choice of the window size is a compromise between a small value, which allows for a efficient sampling even in the absence of accurate weights, and a value large enough to avoid getting trapped kinetically. This compromise depends on the specific system of interest and its size. A particular aspect of these general issues is discussed further in the next section. 2.5 Configurations Inside the Miscibility Gap and Shape Transitions In the previous section we have outlined computational methods to accurately obtain free energy profiles as a function of an order parameter, E. On the one hand, these one-dimensional free energy profiles contain a wealth of information. On the other hand, the configuration space is of very high dimension and a one-dimensional order parameter might not be sufficient to construct

21 Simulation Techniques for Calculating Free Energies 87 a thermodynamically reversible path, i.e., even if the histogram of the order parameter is flat the system might still be forced to overcome barriers in order to sample the entire interval of order parameters. This is readily observable in the time sequence of the order parameter: For a good choice of the order parameter, E, the system performs a random walk in E. For a bad choice of the order parameter the histogram of E is flat, but the range of E can be divided into subintervals. Within a subinterval the different values of the order parameter are sampled in a random walk-like fashion but there is a barrier associated with crossing from one subinterval to another. The seldom crossing of the barriers between subintervals can slow down the simulation considerably even if the weights are optimal. The success of the multicanonical method depends on an appropriate identification of the order parameter (or reaction coordinate) that resolves all pertinent free energy barriers. We illustrate the behavior for a first order transition between a vapor and a dense liquid in the framework of a simple Lennard-Jones model. The condensation of a vapor into a dense liquid upon cooling is a prototype of a phase transition that is characterized by a single scalar order parameter the density, ρ. The thermodynamically conjugated field is the chemical potential, µ. The qualitative features, however, are general and carry over to other types of phase coexistence, e.g., Sect At a given temperature, T, and chemical potential, µ coex (T ), a liquid and a vapor coexist if they have the same pressure, p coex : p liq (T,µ coex )=p vap (T,µ coex )=p coex (32) The two coexisting phases are characterized by the two values of their order parameter, ρ liq and ρ vap. If one prepares a macroscopic system of volume V, at a fixed density, ρ vap <ρ<ρ liq, inside the miscibility gap, it will phase separate into two macroscopic domains in which the densities attain their coexistence values. The volumes of the domains, V liq and V vap, are dictated by the lever rule: V liq ρ liq V + ρ V vap vap = ρ. (33) V The pressure of the system inside the miscibility gap is p coex independent from density. This macroscopic behavior is in marked contrast to the vander-waals loop of the pressure, p, which is predicted by analytic theories that consider the behavior of a hypothetical, spatially homogeneous system inside the miscibility gap. Using multicanonical simulation techniques we can sample all configurations in the pertinent interval of density (order parameter) and determine their free energy. It is important to note that the simulation samples all states at a fixed order parameter with the Boltzmann weight of the canonical ensemble. What are the typical configurations that a finite system of volume V adopts inside the miscibility gap in the canonical ensemble [48 55]? Let us consider the condensation of the vapor phase as we increase the density at coexistence. If the excess number of particles, n =(ρ ρ vap )V,is

22 88 M. Müller and J.J. de Pablo small, they will homogeneously distribute throughout the simulation cell and form a supersaturated vapor. In this case the excess free energy defined by F (ρ) F (ρ) F vap F liq F vap ρ liq ρ vap (ρ ρ vap ) (34) takes the form: F sv (ρ) = V 2κ ρ2. (35) where κ denotes the compressibility of the vapor and ρ = ρ ρvap ρ liq ρ vap is the normalized distance across the miscibility gap. Increasing the excess number of particles (or ρ) we increase the excess free energy quadratically and the curvature is proportional to the compressibility. In a macroscopic system a supersaturated vapor, µ µ µ coex > 0, is metastable and the excess number of particles will condense into a drop of radius R that consists of the thermodynamically stable liquid. In the framework of classical nucleation theory, the excess free energy of such a spatially inhomogeneous system can be decomposed into a surface and a volume contribution: F drop (R) F vap +4πγR 2 4π 3 R3 (ρ liq ρ vap ) µ. (36) The first term describes the increase of the free energy due to the formation of a liquid-vapor interface and γ denotes the interface free energy per unit area (interface tension). The second term describes the free energy reduction by the formation of the thermodynamically stable phase. In the simplest approximation we assume that (i) all excess particles condense into a single drop and (ii) the density of the drop s interior corresponds to the liquid, ρ liq,and its interface tension is given by the macroscopic value γ. These assumptions are reasonable for large drops. Then the size of the drop is given by the lever rule: 4π 3 R3 (ρ liq ρ vap )=(ρ ρ vap )V. (37) Since F liq F vap = µv (ρ liq ρ vap ) the excess free energy of a droplet is given by the surface contribution F drop (R) 4πγR 2. (38) Note that the conjugated field, µ, does not have a direct interpretation in a multicanonical simulation. The two equations, (37) and (38), allow us to calculate the excess free energy as a function of the excess density. To first order we obtain, R ρ 1/3 and F drop = g(v ρ) 2/3 with g =(4π/9) 1/3 γ. A more accurate expression can be obtained by not condensing all excess particles into the drop but allowing the density, ρ, of the surrounding vapor to increase [56]. This increases the free energy of the vapor but it decreases the drop s radius and the associated

23 Simulation Techniques for Calculating Free Energies 89 interface free energy. Then one obtains the excess free energy by minimizing with respect to the drop s radius, R and the vapor density, ρ F drop ( ρ) = min [4πγR 2 + V 4π ] 3 R3 (ρ ρ vap ) 2 (39) R,ρ 2κ under the constraint ρ V + 4π 3 R3 (ρ liq ρ )=Vρ. (40) This refinement qualitatively captures that the chemical potential for a drop configuration is not µ = µ µ coex = 0 but rather µ = F/ n. The shift of the chemical potential increases with the interface tension and decreases as the size becomes larger (Kelvin s equation). If ρ increases further, the drop grows until its size becomes comparable to the linear dimension V 1/3 of the simulation cell. At that point it becomes favorable to form a liquid slab which is separated from the vapor by two interfaces of area V 2/3. In this case the excess free energy F slab =2γV 2/3 (41) is independent from the excess density ρ. As both, the drop and the slab excess free energies scale like γv 2/3, the transition from a drop to a slab occurs at a fixed ρ which depends on the aspect ratio of the simulation cell, but which is independent from its volume or the interface tension. Hence, the largest drops that are observable have a radius R max V 1/3. Increasing the excess density further one observes the reverse set of configurations: From a slab-like configurations one goes to a bubble of vapor surrounded by liquid, and finally to an undersaturated but spatially homogeneous liquid. The free energy profile obtained from Monte Carlo simulations of a small Lennard-Jones monomer system and the above approximations utilizing values of the interface tension and compressibility extracted from independent simulations are shown in Fig. 6. Of course, the simple expressions overestimate the value of the excess free energy, but the qualitative shape of the free energy profile is predicted well. Snapshots of the simulations are presented in Fig. 7. These corroborate the correct identification of the dominant system configurations. From the dependence of the free energy profile on the excess density it is apparent that the free energies of the supersaturated vapor and the drop will exhibit an intersection point. For small excess densities the homogeneous supersaturated vapor, F sv ρ 2, has the lower free energy while for larger excess densities the drop configuration, F drop ρ 2/3, will be more stable. The transition between the supersaturated vapor and the configuration containing a drop is called droplet evaporation/condensation [50,54,55]. From the two simple expressions we can readily read off that droplet condensation occurs at

24 90 M. Müller and J.J. de Pablo 80 F=2L 2 γ 60 4πR 2 γ F k 1 ρ 2 k 2 ρ ρ Fig. 6. Grand-canonical simulation of a Lennard-Jones monomer system. Particles interact [ via a shifted and truncated Lennard-Jones potential of the form: E LJ = ( 4ɛ σ ) 12 ( ) ] σ for distances, r 2 6 2σ, ande = 0 for larger particle r r separations. The free energy profile F (in units of ɛ) has been obtained from the probability distribution of the order parameter, ρ, in grand-canonical simulations for a cubic simulation cell of linear dimension L = 11.3 σ and k BT/ɛ = The phenomenological expressions (35), (38), and (41) for the free energy are also indicated using γ =0.291 ɛ. Circles mark densities at which typical configurations σ 2 are visualized in Fig. 7 Fig. 7. Typical system configurations for the same parameters as in Fig. 6. The density is ρ =0, 0.184, 0.286, Only a thin slice of the simulation box is shown in the right-most image. Each monomer is represented by a sphere of diameter 1.12 σ, which corresponds to the minimum of the Lennard-Jones potential ρ dc =(2κg) 3/4 V 1/4 (42) where g is defined on p. 88. Of course, for any finite simulation cell going from a supersaturated vapor to a drop is not a sharp transition in a thermodynamic sense because the free energy difference between the two states remains finite. Thus the transition will be rounded over a range of densities δρ = ρ ρ dc where the free energy difference between the two states supersaturated vapor and drop is comparable to the thermal energy scale, δf = F drop

25 Simulation Techniques for Calculating Free Energies 91 F sv O(k B T ). Using the above expressions one can expand the free energy difference between the drop and the supersaturated vapor as a function of the distance from the droplet evaporation/condensation: δf 3 ( ) 3/4 ( [2κ] 1/3 gv ρ (2κg) 3/4 V 1/4), (43) 4 This estimate yields δρ ( κ 1/3 gv ) 3/4 for the width of the transition region. As we increase the system size, the excess density at which the supersaturated homogeneous vapor is stable decreases like V 1/4 but the smallest drops that are observable at that density are of radius R min V 1/4, i.e., they increase with system size. In the following we illustrate in somewhat more detail the droplet evaporation/condensation in a finite-sized system of Lennard-Jones monomers. In order to accurately locate the droplet condensation, we regard the derivative of the free energy µ F/ n. The results for such a system are shown in Fig. 8. Inside the miscibility gap and for finite V, µ does not remain constant inside the miscibility gap as suggested by macroscopic arguments. First the chemical potential, µ, linearly increases with the excess density, ρ. This behavior characterizes the homogeneous, supersaturated vapor. Further inside the miscibility gap, µ exhibits an s-shaped variation as a function of ρ. At the densities marked by the symbol we store configurations for further analysis. After the simulation, the distribution, P (N c ), of cluster sizes was determined. Any ensemble of particles whose distance is smaller than 1.5σ is assumed to belong to the same cluster (Stillinger criterion) [57]. For β µ / n β µ n n Fig. 8. Plot of µ/k BT vs. the number of particles n, for a Lennard-Jones fluid in a cubic simulation box of size L =22.5σ at k BT/ɛ =0.68. µ = µ µ coex is the distance from bulk coexistence. denote states at which configurations have been stored for further analysis (Figs. 9 and 10). The inset shows the derivative of µ w.r.t. the number of particles. From MacDowell et al. [55]

26 92 M. Müller and J.J. de Pablo N=380 N=375 ln P(N c ) N=370 N=365 P(µ) n=355 n=365 N=360 n=370 N= N c n= βµ Fig. 9. Distribution P (N c) of the cluster size N c for several choices of particle number, n, (left) and the corresponding distribution of the chemical potential of the supersaturated gas P ( µ) obtained by Widom s particle insertion attempts (right). System parameters are the same as in Fig. 8. From MacDowell et al. [55] n 350, i.e., on the ascending branch of the µ vs. n curve in Fig. 8, P (N c )is monotonically decreasing with cluster size, N c (not shown). This distribution characterizes the homogeneous supersaturated vapor. For larger supersaturation, n 355, a peak around N c 120 appears. This maximum becomes more pronounced and moves to larger sizes, N c,asn increases. It corresponds to a single large liquid drop. Such a liquid drop, however, cannot be found in all sampled configurations; but rather some configurations contain a drop while others correspond to a supersaturated vapor. This is most clearly observed when we investigate the probability distribution of the chemical potential, µ. This quantity has been calculated by performing Widom s particle insertion attempts into the stored configurations. For small particle numbers, n, the distribution of chemical potentials consists of a single peak that characterizes the supersaturated vapor. Upon increasing the particle number a second peak occurs at more negative values of µ. These lower values correspond to configurations with a drop: Some of the excess particles have condensed into the drop and the concomitant reduction of the density of the surrounding vapor gives rise to a lower chemical potential. Therefore the decrease of the chemical potential upon increasing of the particle number in Fig. 8 indicates the droplet evaporation/condensation and gives rise to the s-shaped behavior of the µ vs. ρ curve. From the broad distributions and the fact that supersaturated vapor and drop configurations can be observed over an extended interval of particle numbers (or temperature) it is apparent that in

27 Simulation Techniques for Calculating Free Energies 93 Fig. 10. Two snapshots of configurations at the transition point (n = 365 particles): homogeneous, supersaturated vapor (left) and drop (right). System parameters are the same as in Fig. 8. From MacDowell et al. [55] a finite-sized system the transition is not a sharp one but rather a gradual crossover from configurations dominated by a homogeneous distribution of particles to configurations that contain a drop occurs. Two typical snapshots of these configurations at the same number of particles are presented in Fig. 10. In the course of the simulations the system switches from one to the other conformation forth and back. The dependence on the system size is explored in Fig. 11: In the left panel we plot the chemical potential vs. density for system sizes ranging from L =11.3σ to 22.5σ. The turning point of the curves shifts closer to the coexistence density of the vapor, ρ 0, as we increase the system size. Also the maximum slope increases with increasing L, indicating that for L a sharp transition occurs. The right panel of Fig. 11 presents the probability distribution of the energy, U, at the droplet evaporation/condensation for different system sizes. In qualitative agreement with the expectations the droplet evaporation/condensation becomes sharper and both states (the supersaturated vapor and drop) become more separated as we increase the system size. From the turning points of the µ vs. ρ curves we estimate the location of the droplet evaporation/condensation. The inset of Fig. 11 (right) shows the dependence of ρ dc on the system size. The data are compatible with an effective power law ρ dc V 0.35, while the phenomenological consideration (see (42)) yields an exponent 1/4. The deviations can be traced back to the small system sizes: (i) the drop evaporation/condensation still occurs far inside the miscibility gap such that the simple compressibility approximation in (35) breaks down and (ii) the drop is not large enough to neglect additional contributions to the excess free energy, e.g., due to curvature effects. Similarly, only the gross qualitative features of F ( ρ) in Fig. 6 are described by equations

28 94 M. Müller and J.J. de Pablo ρ-ρ c ρ-ρ c = 0.604*L β µ L=22.5σ L=20.3σ L=18.3σ ρ-ρ c [1/σ 3 ] L=11.3σ L=13.5σ L=15.8σ P(-βU) L=22.5σ L=20.3σ L=18σ L=15.8σ βu [ε] Fig. 11. Left panel: Chemical potential vs density loops for T/k BT =0.68 and several system sizes as indicated in the key. Right panel: Distribution of the energy per particle for different system sizes at the droplet condensation. Inset: ρ ρ vap as a function L = V 1/3.(ρ ρ vap(l =15.8 σ) = ,ρ ρ σ vap(l =18σ) = ,ρ ρ σ vap(l =20.3 σ) = ,ρ ρ 3 σ vap(l =22.5 σ) = ). 3 σ 3 From MacDowell et al. [55] L (35,38,41), but there are quantitative differences. If one insists on identifying the drop s radius according to (38) (and thereby lumps all approximations into the estimate of the interface tension), these deviations correspond to a reduction of the effective interface tension of small drops of the order 20%. This simulation study illustrates what kind of qualitatively different conformations are sampled inside the miscibility gap and what macroscopic quantities, γ and κ, can be obtained from the free energy profile. It is important to realize that the interval of excess densities in which supersaturated vapor or drops can be observed depends on the system size. In the limit of large system size, the droplet evaporation/condensation shifts towards the coexistence curve, ρ dc 0. This corresponds to the macroscopic phase coexistence where the excess free energy F vanishes across the miscibility gap. For a finite system, however, the interface contribution will be important the system chooses a balance between an inhomogeneous density distribution with the corresponding free energy costs associated with the interfaces and the free energy cost of increasing the bulk density homogeneously. The free energy will vary with density and its scale can be estimated by the excess free energy at the droplet evaporation/condensation, F V 1 V. There are two important consequences for simulations: In a finite size simulation box of volume, V, drops of linear dimension R can only be observed in a very limited range of excess densities. The size is limited by V 1/4 R min <R<R max V 1/3. Consequentially, to study the dependence of the free energy on the drop s radius one also has to vary the system size. Note that the equilibrium drops observed in the canonical ensemble correspond to critical drops in the nucleation

29 Simulation Techniques for Calculating Free Energies 95 theory at a supersaturation µ = F/ n. At fixed volume, the density of the surrounding vapor (mother phase) does depend on the drop s size. Hence, one cannot study the growth of a drop at fixed supersaturation the situation assumed in nucleation theory without a systematic study of finite size effects. The transitions from the supersaturated vapor to the drop and from the drop to the slab configurations represent barriers in the configuration space which are not removed by the multicanonical reweighting scheme. Hence, these shape transitions limit the applicability of reweighting methods in the study of phase equilibria. The last issue can be clearly observed in the time sequence of the order parameter in the simulation. With a very good reweighting function the probability distribution sampled in the simulation is flat but the system does not perform a random walk in the order parameter. Rather one finds a banded structure of the time evolution where the system explores the configurations within a typical shape homogenous vapor, drop, slap, bubble, or homogeneous liquid in a random walk-like manner but only occasionally jumps between different shapes [58]. Typically the free energy barrier associated with transitions from one shape to another is a small fraction of the free energy barrier associated with the formation of the interfaces in the slab-like configurations. Neuhaus and Hager [58] pointed out the slowing down of simulations due to these shape transitions. For a two-dimensional simulation cell of square geometry, L L, they investigated the barrier for the transition between a drop (spot) and a slab (stripe) and compared simulation results of the Ising model with a transition state that had earlier been suggested by Leung and Zia [59]. It consists of a lens-shaped domain whose two arc ends touch each other via the periodic boundaries. The opening angle of the arc is chosen such that the enclosed area of the lens equals the area, L 2 /π, atthe drop to slab transition. Simple geometric considerations yield the arc length 1.135L which is larger than the length of the liquid-vapor interface (line), L. Thus the free energy barrier associated with the shape transition is 13.5% of the free energy barrier due to the interfaces. Given that this free energy barrier should not exceed the thermal free energy scale k B T by an order of magnitude total free energy barriers, i.e., 2γL 2 on the order 10 2 k B T can be efficiently sampled. 3 Applications 3.1 Liquid-Vapor Coexistence In the first application will will exemplify the accurate localization of the condensation transition for a coarse-grained bead-spring polymer model. As discussed above, the order parameter of the liquid-vapor transition is the monomer number density, ρ. The equation of state of the polymer solution is

30 96 M. Müller and J.J. de Pablo pσ 3 /ε k B T/ε=5 k B T/ε=4 k B T/ε=3 k B T/ε=2.5 k B T/ε=1.68 TPT ρσ 3 k B T/ε θ TPT1 MC p= ρσ 3 glassy behavior Fig. 12. Equation of state (a) and phase diagram (b) of a bead-spring polymer model. Monomers interact via a truncated and shifted Lennard-Jones potential as in Fig. 6 and neighboring monomers along a molecule are bonded together via a finitely ( extensible ) non-linear elastic potential of the form: E FENE(r) = 15ɛ(R 0/σ) 2 ln 1 r2 with R R0 2 0 = 1.5σ. Each chain is comprised of N = 10 monomeric units. The pressure of the bulk system in panel (a) has been extracted from canonical Monte Carlo simulations using the virial expression (symbols). Lines show the result of Wertheim s perturbation theory (TPT1). The phase diagram (b) has been obtained from grand-canonical Monte Carlo simulations. Binodal compositions are shown as full lines, while the corresponding results from Wertheim s perturbation theory are presented by dashed lines. Filled and open circles mark the location of the critical point in the simulations and the TPT1 calculations, respectively. The Θ-temperature is indicated by an arrow on the left hand side. The diamonds present the results of constant pressure simulations indicating the densities that correspond to vanishing pressure, which are a good approximation for the coexistence curve at low temperatures. Adapted from Müller and Mac Dowell [60] presented in Fig. 12 (a) and the Monte Carlo results (symbols) are compared to Wertheim s thermodynamic perturbation theory (TPT1) represented by lines without adjustable parameter [60]. As we reduce the temperature, the pressure decreases and the analytic theory predicts a van-der-waals loop with negative pressures for a hypothetical homogeneous system. In the simulation, however, the system spatially separates into two coexisting phases a lowdensity vapor and a dense liquid. The monomer number density, ρ = nn/v (n being the number of chains and N the number of monomers per chain), distinguishes the two phases. The sequence of typical configurations inside the miscibility gap has been discussed in Sect Upon increasing the system size the equation of state does not exhibit a loop but develops a plateau, p coex, in the miscibility gap. The grand-canonical ensemble is particularly well suited for studies of liquid-vapor phase coexistence: (i) Fluctuations of the order parameter, i.e., the density, are efficiently relaxed. Since the density is not conserved, spatial fluctuations do not decay via slow diffusion of polymers but relax much faster through insertion/deletion moves. In the grand-canonical ensemble one controls the temperature, T, the volume, V, and the chemical potential, µ,

31 Simulation Techniques for Calculating Free Energies 97 in the simulations and the number of particles, n, fluctuates. To realize this ensemble in Monte Carlo simulations, insertion/deletion moves are utilized to supplement canonical moves that displace particles and alter the conformations of the polymers. Note that insertion of extended particles into a dense liquid is difficult, and several methods have been devised to overcome the concomitant sampling difficulties. In our application we utilize configurational bias Monte Carlo [61] to grow the polymer chains into the system. (ii) The key quantity to monitor in the simulation is the probability distribution of the number of particles in the simulation cell. It contains much information about the coexisting phases and the interface that separates them (c.f. Sect. 2.5). Additionally, there exist sophisticated methods to control finite-size effects and to accurately locate critical points. In the vicinity of the phase coexistence, µ = µ coex (T ), the distribution P (ρ) exhibits two pronounced peaks that are separated by a deep valley. The corresponding free energy profile, F (ρ) = k B T ln P (ρ), for polymers that are comprised of N = 10 effective segments and temperature k B T/ɛ =1.68, is presented in Fig. 13. The location of phase coexistence can be accurately estimated by the equal-weight rule [63]: Two phases coexist if the weights of the two corresponding peaks in the distribution function of the order parameter are equal. It is important for the system to tunnel often between the two phases in F/2L 2 k B T vapor interface liquid γ LV /k B T ρ V ρ L ρσ probability Fig. 13. Illustration of the grand-canonical simulation technique for temperature k BT/ɛ =1.68 and µ = µ coex. A cuboidal system geometry 13.8σ 13.8σ 27.6σ is used with periodic boundary conditions in all three directions. The solid line corresponds to the negative logarithm of the probability distribution, P (ρ), in the grand canonical ensemble. The two minima correspond to the coexisting phases and the arrows on the ρ axis mark their densities. The height of the plateau yields an accurate estimate for the interfacial tension, γ LV.Thedashed line is a parabolic fit in the vicinity of the liquid phase employed to determine the compressibility. Representative system configurations are sketched schematically. From [62]

32 98 M. Müller and J.J. de Pablo the course of the simulation in order to equilibrate the weight between the two phases. Without multicanonical techniques, the large free energy barrier between the two phases would render the simultaneous sampling of both phases in a single simulation infeasible. If we assume that in a multicanonical simulation the system performs a random walk along the order parameter then the statistical uncertainty of the peak weights will scale as 1/ #, where # is the number of transitions (or tunneling events) from one phase to the other. The concomitant statistical uncertainty in the free energy difference of the two coexisting phases is on the order O(k B T/ #). The rule for phase coexistence can be rationalized as follows: Using the relation between the the probability distribution, P (n), and the grandcanonical partition function, Z gc, we obtain for the ratio of the partition functions of the two phases: Zgc liq Zgc vap = exp(βµn) n liq n! exp(βµn) n! n vap Dn [{r}]exp( βe[{r}]) Dn [{r}]exp( βe[{r}]) (44) where D n [{r}] sums over all configurations of the system with n polymers, and E[{r}] denotes the energy of this microscopic configuration, {r}. The condition n liq, characterizes all particle numbers that correspond to the liquid phase; if n vap, the configuration belongs to the vapor phase. The detailed way in which the interval of particle numbers is divided into vapor and liquid phases affects the estimate of the ratio only by an very small amount. Furthermore, that amount decreases exponentially with the system size because the configurations between the peaks are strongly suppressed due to the free energy cost of the interface between the coexisting phases (cf. Sect. 3.5). The pressure, p, is related to the grand-canonical partition function via pv = k B T ln Z gc. Therefore one obtains: Zgc liq Zgc vap =1 p liq = p vap (equal weight rule) (45) Thus the equal weight-rule rephrases the macroscopic definition of phase coexistence: Two phases coexists if they have equal pressure at equal chemical potential. Using histogram extrapolation [2] we determine the coexistence value of the chemical potential. The properties of the individual phases can be extracted by taking averages over the corresponding regions of the order parameter. The densities of the two coexisting phases the binodal densities are presented in Fig. 12(b). From the curvature of the free energy profile at the minima we can also extract the compressibility of the liquid and the vapor. Additional information about interfaces between the coexisting phases can also be obtained as discussed in Sect. 3.5.

33 3.2 Demixing of Binary Blends Simulation Techniques for Calculating Free Energies 99 Another example of phase coexistence that is described by a single scalar onecomponent order parameter is provided by incompressible binary mixtures. One considers a dense liquid of two components, A and B; the composition of the mixture, φ = number of particles, n A + n B, is assumed to be independent of composition. The more general case of compressible mixtures will be discussed in the following section. Sariban and Binder [64] employed simulations in the semi-grand canonical ensemble for investigating the phase behavior of an incompressible binary polymer blend at constant volume. In this ensemble, the total monomer density is fixed, the composition of the blend fluctuates, and the chemical potential difference, µ, between the species is controlled. The Monte Carlo scheme comprises two types of moves. Canonical updates relax the conformation of the macromolecules, whereas semi-grand canonical moves convert A polymers into B polymers and vice-versa. This ensemble is particularly well suited for the study of strictly symmetric mixtures where the two components have the same chain architecture and intramolecular potentials, but different species repel each other. In this limit, semi-grand canonical moves consist of a mere identity exchange (a switch of labels). Note that the algorithm implies that both phases have identical densities but it does not automatically ensure that they have identical pressure. The algorithm can be extended to some degree of structural asymmetry (e.g., different chain lengths between the species [65]). Overall speaking, it is reasonably efficient for a modest degree of structural asymmetry between the different constituents, but the extension to pronounced structural asymmetries is a challenging task. Improvement might be achieved via gradually mutating one species into another [66]. Figure 14 exemplifies two computational methods to determine the probability distribution of composition for binary polymer blends described by the bond fluctuation model [67]. Phase coexistence can be extracted from these data via the equal-weight rule. For the specific example of a symmetric blend, the coexistence value of the exchange chemical potential, µ, isdic- tated by the symmetry. One can simply simulate at µ coex 0 and monitor the composition. Nevertheless, the probability distribution contains additional information, as discussed in Sect Panel (a) illustrates the application of the Wang-Landau algorithm [7]. The different probability distributions obtained at successive steps of the modification parameter, f are shown. As f decreases the distribution gradually converges towards the Boltzmann distribution. Note that the starting value of the modification factor, f =1.02, is chosen much smaller than in the original application to a spin system in order to match the relaxation time of the system (chain conformations) with the time scale for exploring the entire composition range. na n A+n B, distinguishes the two coexisting phases. The total

34 100 M. Müller and J.J. de Pablo f=1.02 f= f= f= final 10 0 P(φ) P(φ) φ φ Fig. 14. (a) Probability distribution of the composition, φ, of a symmetric binary polymer blend within the framework of the bond fluctuation model. In this coarse-grained lattice model monomeric units are represented by unit cubes on a three-dimensional cubic lattice. Each monomer blocks the 8 sites of a cube from double occupancy. This mimics excluded volume interactions. Monomers along a polymer chain are connected via one of 108 bonding vectors that can adopt lengths, b =2, 5, 6, 3, and 10 in units of the lattice spacing. This represents the connectivity along the polymer backbone. Monomers of the same type attract each other via a square-well potential of depth ɛ that extends over the nearest 54 lattice sites. Unlike monomers repel each other via a potential of opposite sign. The simulation data are obtained at the different stages of the Wang-Landau sampling for chain length N =32andɛ/k BT =0.02 in a simulation cell of geometry in units of the lattice spacing (R e = 17). The convergence factors, f, are indicated in the key. A flatness of 50% was required for the histogram of visited to particle numbers to reduce the convergence factor from f f = f.(b) Piecewise probability distributions obtained by successive umbrella sampling. By patching the distributions together in regions of mutual overlap it is possible to construct the probability distribution over the complete composition range Panel (b) illustrates the application of successive umbrella sampling [35], where each piece of the curve corresponds to several intervals used in the successive sampling. The intervals overlap just by one particle number, which is sufficient if one takes due account of detailed balance at the interval boundaries [36]. Matching the pieces at their boundaries one obtains the distribution across the entire composition range. Of course, both methods yield identical results for the final probability distribution and, provided that one starts with an appropriate modification factor, f, they require comparable amounts of CPU time. One advantage of these Monte Carlo simulations is that both the thermodynamic and structural properties of the binary polymer mixture are simultaneously accessible, and can quantitatively be compared to analytic approaches. The mean field theory of polymers makes detailed predictions for the bulk [68 70] and interface properties [71 77] as a function of the incompatibility, χn, the spatial extension of the molecules R e as measured by the root mean squared end-to-end distance of the Gaussian coils, and the

35 Simulation Techniques for Calculating Free Energies 101 invariant degree of polymerization N ( ρr 3 /N ) 2. The last parameter describes the strength of fluctuation effects that are neglected by mean field theory; the analytic approach is thought to describe the limit N. One question that these simulations can address is the relation between different estimates of the incompatibility between polymer species, i.e.: How can one identify the mutual repulsion between polymer species parameterized by the product of the Flory-Huggins parameter and chain length, χn, for a specific microscopic polymer model [65, 78]? In simulations, as well as in experiments, it is common practice to measure the Flory-Huggins parameter by comparing the results of simulations to the predictions of the mean field theory. Since the predictions are affected by fluctuation effects to different extents, not all quantities yield mutually compatible estimates of the Flory-Huggins parameter, χ. Within mean field theory, for a symmetric blend the excess free energy of mixing per volume is given by the Flory-Huggins expression: F mix k B T (V/R 3 e) = N [φ ln φ +(1 φ)ln(1 φ)+ χnφ(1 φ)] (46) From this expressions one can calculate the binodal curves (composition of the two coexisting phases) ln φ 1 φ + χn(1 2φ) = µ coex =0, (47) k B T the location of the critical point, χ c N =2andφ c =1/2, which marks the onset of phase separation, and the composition fluctuations in the one phase region N S(k 0) = 1 φ χn for χn < 2 (48) 1 φ as measured by the structure factor, S(k), of composition fluctuations in the limit where the wavevector k vanishes. The latter quantity is often used in experiments to determine the Flory-Huggins parameter for miscible blends. A quantitative comparison between the mean field prediction and the Monte Carlo results is presented in Fig. 15. The main panel plots the inverse scattering intensity vs. χn. At small incompatibility, the simulation data are compatible with a linear prediction (cf. (48)). From the slope, it is possible to estimate the relation between the Flory-Huggins parameter, χ, and the depth of the square well potential, ɛ, in the simulations of the bond fluctuation model. As one approaches the critical point of the mixture, deviations between the predictions of the mean field theory and the simulations become apparent; the theory cannot capture the strong universal (3D Ising-like) composition fluctuations at the critical point [64, 79, 80] and it underestimates the incompatibility necessary to bring about phase separation. If we fitted the behavior of composition fluctuations at criticality to the mean field prediction, we would obtain a quite different estimate for the Flory-Huggins parameter.

36 ~ 102 M. Müller and J.J. de Pablo N/S(k=0) χn MC (L=96) FH 3D Ising (FSS fit) φ A ~ 8 χn FH ~ χn=2nz c ε/kt with z c =2.44 ln[φ A /(1-φ A )]/(2φ A -1) 8 4 MC Fig. 15. Inverse maximum of the collective structure factor of composition fluctuations, N/S(k 0), as a function of the incompatibility, χn. Symbols correspond to Monte Carlo simulations of the bond fluctuation model, the dashed curve presents the results of a finite-size scaling analysis of simulation data in the vicinity of the critical point, and the straight, solid line indicates the prediction of the Flory-Huggins theory. The critical incompatibility, χ cn = 2 predicted by the Flory-Huggins theory and that obtained from Monte Carlo simulations of the bond fluctuation model ( N 240, N = 64, ρ =1/16 and R e =25.12) are indicated by arrows. The left inset compares the phase diagram obtained from simulations with the prediction of the Flory-Huggins theory (c.f. (47)). The right inset depicts the compositions at coexistence such that the mean field theory predicts them to fall onto a straight line. From Müller [78] This estimate, however, would not characterize the incompatibility between the polymer species too well, but rather quantify the inability of the mean field theory to cope with Ising-like order parameter fluctuations. In the insets of Fig. 15 we show binodal curves for the symmetric blend. Again, we find deviations in the immediate vicinity of the critical point but for larger incompatibilities, χn 2, the mean field predictions provide an adequate description of the phase boundary utilizing the Flory-Huggins parameter extracted from the composition fluctuations in the one-phase region, χn < Compressible Blends The phase behavior of binary blends becomes much more complex if one relaxes the assumption of incompressibility. Then both liquid-liquid demixing and liquid-vapor phase separation are possible and the system is described by two scalar order parameters density and composition (ρ, φ) or the two densities of the species (ρ CO2,ρ HD ). The interplay between the two types of phase

37 Simulation Techniques for Calculating Free Energies 103 Table 1. Critical points of the pure components from MC simulations and experiments. By comparing the critical temperatures from simulation and experiment, we identify ɛ CO2 = J, ɛ HD = Jandɛ CO2 /ɛ HD = From the critical densities we derive σ CO2 = m, σ HD = mand σ CO2 /σ HD = ρ refers to monomer number densities T crit ρ crit p crit MC CO ɛ k σ 3 EXP CO K g cm 3 MC HD ɛ k σ 3 EXP HD 723 K g cm ɛ σ bar ɛ σ bar separation gives rise to 6 qualitatively different types of phase behavior in the classification scheme of Konynenburg and Scott [81]. An example of a phase diagram of type I and type III (in this classification) is presented in Fig. 16. The simulations study the phase behavior of a short polymer (hexadecane, HD) in a supercritical solvent (carbon dioxide, CO 2 ) within a coarse-grain model. The polymer is represented by a Lennard-Jones bead spring model such as that discussed in Sect We utilize a chain length of N =5,which roughly mimics the geometrical shape of the monomer. Comparing the two parameters of the Lennard-Jones potential, ɛ and σ, with experimental results for the critical temperature and pressure, we identify energy and length scales [40, 41]. Carbon dioxide (CO 2 ) is represented by a single Lennard-Jones bead; the interactions between CO 2 -beads are adjusted to reproduce the experimental critical point. A modified Lorentz-Berthelot mixing rule is used to describe the interactions between hexadecane and CO 2 beads [82, 83]: σ HD CO2 = σ HD + σ CO2 2 and ɛ HD CO2 = ξ ɛ HD ɛ CO2 (49) where ξ describes the deviations from the standard mixing rule. Figure 16 shows projections of the phase diagram in the plane of temperature, T,and pressure, p, for two choices of ξ. The value ξ = 1 corresponds to the standard Berthelot rule which describes the mixing of van-der-waals interactions, while the value ξ = matches the experimental behavior [84, 85] most closely. Solid lines represent the liquid-vapor phase equilibria of the two pure components that end in critical points marked by arrows. When a small amount of solvent is added to the pure polymer, the liquid-vapor coexistence shifts and so does the critical point. The loci of critical points for the binary system form a critical line that is shown by the dashed line with squares for ξ =1 and triangles for ξ = In the former case phase behavior of type I the critical line connects the critical points of the two pure components and the two coexisting phases gradually change from vapor and solvent-rich liquid

38 104 M. Müller and J.J. de Pablo 300 sim: lv c ξ=1.00 ξ= pressure [bar] p 100 critical point CO 2 critical point C 16 H temperature T [K] Fig. 16. Projection of the global phase diagram for a compressible mixture of hexadecane, C 16H 34, and carbon dioxide, CO 2, into the temperature-pressure plane for two values of the mixing parameter, ξ = 1(square) andξ = (triangle). Simulation results for the liquid-vapor coexistence of the pure components are shown by solid lines and end in critical points that are indicated by arrows. The line of critical points that emerges from the critical point of the less volatile polymer component is indicated by symbols. From Virnau et al. [40] to vapor and polymer-rich liquid. This is not, however, what is observed in experiments where a phase diagram of type III is encountered. In a type III diagram, the critical line that emerges from the liquid-vapor critical point of the polymer is not connected to the critical point of the pure solvent but gradually changes its character from a liquid-vapor coexistence to a liquid-liquid coexistence between a dense solvent-liquid and a polymer-liquid. Just a slight change of the interaction between the different components from ξ = 1 to ξ = brings about a qualitative change of the phase diagram. The reduction of the attractions between CO 2 and hexadecane can be traced back to quadrupolar interactions between CO 2 molecules [86]. The well depth ɛ CO2 of the pure solvent parameterized the attraction between carbon dioxide molecules that is caused by the joined effect of both, van-der-waals and quadrupole interactions. The Berthelot mixing rule, however, implicitly assumes that all attractions result from van-der-waals interactions and utilizes the full ɛ CO2, which only partially stems from dipolar interactions. The techniques (e.g., weighted histogram analysis, equal weight rule, etc.) utilized for one-component systems or incompressible binary mixtures can be readily carried over to compressible systems. Since the system is described by two order parameters one monitors the joint probability distribution of

39 Simulation Techniques for Calculating Free Energies 105 Fig. 17. Joint probability distribution of the solvent and the polymer density along the critical line in a finite-size simulation cell for ξ = The distributions correspond to temperatures T = 314 K, 356 K, 398 K, 486 K, 650 K and 713 K from up/left to down/bottom. For temperatures T = 314 K and 356 K the box size is L = 6.74 σhd, while L = 9 σhd was employed for other temperatures. The arrow in the middle right panel indicates our finite-size estimate for the location of a critical point as an average over both peaks. From Virnau et al. [40]

40 106 M. Müller and J.J. de Pablo ρ HD and ρ CO2. These two-dimensional probability distributions are shown in Fig. 17 for various temperatures. For a finite-size system, the probability distribution at the critical point is bimodal. The two peaks indicate the densities of the two coexisting phases slightly below the critical temperature. Unlike a simulation at (a first-order) phase coexistence, however, the positions of the peaks depend on the system size, and they collapse onto a single density in the thermodynamic limit. The average over the composition of the bimodal distribution of the finite-size system yields an estimate for the critical density. The two-dimensional order parameter distribution [40] at the estimate of the critical point clearly reveals the gradual crossover from a phase coexistence between a polymer-liquid and a vapor characterized by a horizontal ridge in the distribution to a liquid-liquid coexistence marked by a diagonal ridge. If the phase diagram were of type I, however, the distribution would rotate counterclockwise and end up vertically at low temperatures. In order to locate phase coexistence we utilize the equal-weight rule, i.e., we divide the plane of order parameters into regions which correspond to the coexisting phases and tune both chemical potentials as to fulfill the equal weight condition. Looking for phase coexistence in this larger parameter space is hardly feasible without the help of weighted histogram analysis. In practice, one often just chooses a suitable linear combination of the two order parameters that clearly distinguishes between the two coexisting phases, i.e., one divides the order parameter plane by a line. At low pressure the total density will be a suitable order parameter; at higher pressure one should use the composition. The equal weight condition will provide us with µ HDcoex (T,µ CO2 ), i.e., at constant temperature there is a set of phase coexistences characterized by the chemical potential µ CO2 or the pressure, p. If we sampled the entire interval of both order parameters we could not only locate the critical point but we could construct an entire slice of the phase diagram as a function of pressure and composition at constant temperature from a single simulation. 3.4 Crystallization The simulation techniques presented above can be applied to all first order phase transitions provided that an appropriate order parameter is identified. For vapor-liquid equilibria, where the two coexisting phases of the fluid have the a similar structure, the density (a thermodynamic property) was an appropriate order parameter. More generally, the order parameter must clearly distinguish any coexisting phases from each other. Examples of suitable order parameters include the scalar order parameter for study of nematic-isotropic transitions in liquid crystals [87], a density-based order parameter ψ for block copolymer systems [88], or a bond order parameter for study of crystallization [89]. Having specified a suitable order parameter, Φ, we now show how the EXEDOS technique introduced earlier can be used to obtain P β,p (Φ(q n )) in a particularly effective manner for simulations of crystallization [33]. The Landau free energy of the system Λ(Φ) can then be related to P β,p (Φ(q n ))

41 by (50) [90]: Simulation Techniques for Calculating Free Energies 107 Λ(Φ(q n )) = constant k B T ln{p (Φ(q n ))}. (50) Depending on whether P (Φ) is obtained from EXEDOS in a constant (n, V, T ) or constant (n, p, T ) ensemble, Λ will correspond to the Helmholtz free energy or the Gibbs free energy of the system, respectively. Away from a phase transition, Λ is expected to exhibit a single minimum as a function of Φ. Close to coexistence, however, Λ develops two minima; the barrier between the minima corresponds to the free energy associated with the formation of an interface [91] (cf. Sect. 3.5). The method described above is capable of providing a free energy profile (including free energy barriers) at any given temperature and pressure. In general, however, the coexistence temperature (i.e., melting temperature) or pressure are not known. In order to obtain the precise conditions under which two phases coexist it is therefore necessary to conduct a series of exploratory simulations, until a set of conditions is found for which the free energy minima corresponding to the two equilibrium states became identical. This process can be facilitated by resorting to the following weighted histogram analysis [2, 33]. Energy and volume data from EXEDOS simulations must be saved and sorted in histograms according to the order parameter. The probability of a system to be in bin m, denoted P β2,p(φ m,φ m+1 ), at temperature β 2 can be obtained from a simulation at a different temperature β 1 via weighted histogram analysis (cf. (4)) P β1,p(φ N m,φ m+1) m i=1 exp( (β1 β2)(ei,m + pvi,m)) P β2,p(φ m,φ m+1) = Nb m=1 P β 1,p(φ N m,φ m+1) m i=1 exp( (β1 β2)(ei,m + pvi,m)) (51) Where E m,i and V m,i are the energy and volume entries in the mth order parameter histogram bin, N b is the total number of histogram bins, and N m is the number of entries in bin m. Equation (51) can be used to determine the precise values of coexistence temperature T and pressure p for which the free energy (50) exhibits two minima of equal weight [63]. This weighted histogram analysis works particularly well in conjunction with multicanonical methods because we do not only sample configurations which have high Boltzmann weights but configurations which might have a low Boltzmann weight but are important for free energy estimations. Having determined the free energy Λ as a function of Φ(q n ), configurations corresponding to the top of the free energy barrier can be used to explore the transition pathways from one phase to another, as discussed later in this work in the context of nucleation mechanisms [92]. Here we consider a system of particles interacting through a repulsive Lennard-Jones (LJ) potential E(r) = ɛ( σ r )12 truncated at a distance of r c = 2σ. We measure temperature and pressure in Lennard-Jones units, T k B T/ɛ and p = pσ 3 /ɛ, respectively. The long-range contribution to the potential was calculated under the assumption of constant density. For this system, the stable nuclei at melting are believed to be face-centered-cubic

42 108 M. Müller and J.J. de Pablo (fcc) [93 95]. Previous simulations by ten Wolde et al. [92] used umbrella sampling to generate free energy curves and to examine the structure of the critical nuclei that form in a deeply supercooled liquid. The order parameter used is the bond orientational order parameter, Q 6, originally introduced by Steinhardt, Nelson and Ronchetti [89] and later used by van Duijneveldt and Frenkel [96] to simulate crystallization. It is sensitive to the overall degree of crystallinity in the system, irrespective of the crystal structure, i.e., it distinguishes the liquid from the crystal. The values of Q 6 for pure fcc, body-centered-cubic (bcc), and for the liquid are , , and 0.0, respectively. Previous work [96] has shown that a defective fcc crystal, which crystallizes from the liquid in simulations, has an order parameter below 0.5 and that finite-size effects lead to small positive values for the order parameter in the liquid phase. Our simulations were performed on systems of various sizes, ranging from 108 and 1364 particles. The order parameter range explored here goes from 0.05 to 0.5 for the small system and from 0.02 to 0.45 for the larger systems. In all simulations the order parameter was calculated at every step and was used in the Monte Carlo acceptance criteria. Following our previous work, the range of order parameter was split into multiple overlapping windows (intervals of order parameter) [32, 97]. 10 overlapping windows with a 50% of overlap between adjacent windows were chosen in all our simulations. Configuration swaps were implemented to facilitate convergence. During the simulation we observed the tunneling of individual replicas between the liquid and the crystalline state. The total number of tunneling events for each system size exceeds Thus the statistical ( ) uncertainty of the free energies we calculate is on the order Λ O kbt # 0.01k B T. The corresponding errors T and p are on the order of and 0.003, respectively. The accuracy of the coexistence temperature and pressure in EXEDOS is comparable to that achieved in techniques such as phase switch Monte-Carlo [98], self-referential Monte-Carlo [99], and techniques using thermodynamic integration and Gibbs-Duhem integration [100]. The added advantage with EXE- DOS is that, in addition of obtaining accurate coexistence conditions, it also provides an accurate free energy curve as a function of the chosen reaction coordinate. An alternative technique, transition path sampling, can also provide both coexistence conditions and a free energy curve. In a recent study [101] on crystallization, a free energy barrier and a transition path were also obtained for the Lennard-Jones system; transition path sampling, however, requires that a large number of trajectories be run, and is therefore highly computationally demanding. For a system of 864 particles, the starting guess for coexistence temperature and pressure was T = 1.14 and p = The order parameter range was set between 0.02 and EXEDOS simulations were performed in multiple windows. The free energy curves obtained at this temperature and pressure (see Fig. 18) show that, as a result of finite-size effects, T = 1.14

43 Simulation Techniques for Calculating Free Energies Direct simulation at T * = 1.14 Predicted at T * =1.03 Direct simulation at T * =1.03 G (LJ units) Order Parameter (Q 6 ) Fig. 18. Free energy curve for crystallization of an 864-particle repulsive LJ system. EXEDOS simulations at T =1.14solid line, predicted free energy profile by reweighing T = 1.14 data to T = 1.03:dashed line, and EXEDOS simulations at T =1.03dot-dot-dashed line. From Chopra et al. [33] is not a good estimate of the phase transition temperature for the larger system. To obtain the coexistence point the free energy curves were reweighted to arrive at a melting temperature of Tσ = The histogram-predicted free energy profile was confirmed by a subsequent EXEDOS simulation at the same temperature. The above result is indicative of strong finite-size effects. To verify the finite-size scaling and to predict the coexistence temperature for an infinite system we also performed calculations on systems of 256 and 500 particles. Figure 19 shows the variation of Tσ with the volumetric length scale; a clear finite-size scaling can be seen in the behavior of Tσ with volume. The extrapolation to infinite size predicts a coexistence temperature of Tσ, =0.99(1). Note that in the EXEDOS algorithm the nucleation barrier necessary to expand a solid nucleus is removed but the barrier associated with transforming an ordered but homogeneous liquid into an inhomogeneous liquid with an initial solid nucleus at the same value of the order parameter, Q 6, remains [55, 58]. This formation of solid bodies of high Q 6 allows the larger system to move more easily between high and low order-parameter, thereby accelerating the convergence of the simulation. A subsequent structural examination of the trajectories can be performed using Voronoi analysis and using the analysis outlined by ten Wolde et al. [92]. In this procedure, a combined distribution of local invariants q 4, q 6, w 4, and w 6 was generated for a given configuration. A detailed explanation of local invariants q 4, q 6, w 4,andw 6 can be found in the literature [102]. The distribution of local invariants was then decomposed into contributions for thermally equilibrated bcc, fcc and liquid configurations according to (52): 2 =[ṽ cl (f liq ṽ liq + f bcc ṽ bcc + f fcc ṽ fcc )] 2, (52)

44 110 M. Müller and J.J. de Pablo Data T coex = /L T coex /L 3 Fig. 19. Melting temperature T as a function of system size. The arrow points to the limiting (infinite-size) value. From Chopra et al. [33] where ṽ cl,ṽ liq,ṽ fcc, and ṽ bcc are combined distributions for the cluster to be analyzed, thermally equilibrated liquid, fcc, and bcc configurations respectively. In (52) f liq, f bcc,andf fcc are weight coefficients for liquid, bcc and fcc configurations. The analysis confirmed that the first minimum in the free energy profile corresponds to a liquid phase, and the second minimum corresponds to a thermally equilibrated fcc phase. The analysis also provides an explanation for the asymmetric shape and the kinks that appear in the free energy curve of Figure 18. These features were not observed in previous simulations of the same system, which used a different simulation approach, and it is therefore of interest to discuss their origin. This can be achieved by close inspection of the configurations that the system adopts for specific values of the order parameter [48,49,51,52,55,58]. For the order parameter range of , we observe the occurrence of spherical nuclei in the configurations. In the range from 0.15 to 0.25 we observe the formation of solid-fluid interfaces. An interesting point to note here is that the interface has a predominately bcc character, hinting at the formation of thin interfacial layers which are structurally different from the bulk solid. The fact that we observe a bcc interface coincides with the fact that the liquidbcc tension is smaller than the liquid-fcc tension [103]. Moving further along the free energy profile to the range we observe the growth of an interface that spans the entire simulation box, with a defective bcc-like solid. Upon further increasing the order parameter we observe that, beyond 0.35, this defective bcc-like solid rearranges into a defective fcc-like solid. Figure 20 shows snapshots of various representative configurations extracted from our simulations. The asymmetric shape of the free energy profile can be attributed to this two-step transition from an interface between a liquid and defective bcc-like structures, and then from defective bcc-like structures to defective fcc-like structures.

45 Simulation Techniques for Calculating Free Energies 111 Fig. 20. Snapshots of various representative configurations observed during a typical simulation trajectory. Deffective bcc configuration (left) and fcc configuration (right) 3.5 Interface Free Energies The probability distribution of the order parameter does not only characterize the coexisting phases in the bulk (i.e., their density, compressibility, and coexistence pressure) but additional information about spatially inhomogeneous system configurations can be extracted. Let us first discuss the case of liquidvapor coexistence (c.f. Sect. 3.1) where the total number of polymers is used as order parameter along which a free energy profile is obtained. Recall that at each fixed value of the number of polymers the simulation samples configurations according to the Boltzmann weight independent of the reweighting function. As discussed in Sect. 2.5 the typical configurations consist of (1) supersaturated homogeneous vapor, (2) liquid drop, (3) slab of liquid separated by two planar interfaces from the coexisting vapor, (4) bubble of vapor in liquid and (5) undersaturated spatially homogeneous liquid as one increases the density. The excess free energy of the configuration (3) that contains a slab of liquid is dominated by the interface free energy [91], F =2L 2 γ provided the system is large enough for the two interfaces not to interact. Here, L 2 is the area of the interface, γ is the interface tension and the factor 2 arises because in a system with periodic boundary conditions two liquid-vapor interfaces are present. The independence of the two interfaces can be verified by the probability distribution itself: If the interfaces do not interact, one can change the distance between them by varying the amount of liquid at no free energy costs. Therefore one will observe a plateau in the free energy profile [104]. No such plateau is observed in small, cubic systems (cf. Fig. 6) but a clear plateau is observed in Figs. 12 or 14. Often it is useful to utilize a cuboidal simulation box where one dimension, L z is larger than the two lateral extensions, L [105,106]. This

46 112 M. Müller and J.J. de Pablo geometry displaces the two interfaces farther apart and thereby reduces their interaction without increasing their free energy costs and the concomitant free energy barrier. If one observes a plateau in the free energy profile and verifies that the typical configurations consist of a slab of the two coexisting bulk phases one obtains the interface tension via [91]: γl 2 k B T = 1 2 ln P bulk P slab (53) There are several corrections to the interface tension which scale like γ 1 L 2 or ln L L which stem inter alia from the translation entropy of the interfaces, 2 capillary waves, bulk compressibility, etc. A careful consideration of finite size effects is necessary to obtain accurate estimates. The application of this technique to symmetric binary blends is discussed in Fig. 21, where we plot the results of simulations of the bond fluctuation model. The interface tension has been obtained for different chain lengths and the results are compared to mean field theory [104, 107, 108]. This comparison utilizes the identification of the incompatibility χn obtained from the simulation of the spatially homogeneous system [65] in Sect The mean field theory suggests a simple picture of the structure of the interface between two immiscible polymers which is illustrated in Fig. 22. The cost of each loop into the hostile phase is comparable to the thermal energy γ I /γ SSL 0.6 N=16 N= N=64 N=32 (capillary waves) SCF 0.2 ~ 1-4ln2/χN 1-4ln2/χN /(χN) ~ ~ 3/ ~ 1/χN Fig. 21. Ratio between the interface tension γ and the simple expression for the strong segregation limit γ SSL in (54) as a function of inverse incompatibility. Symbols correspond to Monte Carlo results for the bond fluctuation model, the solid line shows the result of the SCF theory, and the dashed line presents first corrections to (54) calculated by Semenov. Also an estimate of the interface tension from the spectrum of capillary waves is shown to agree well with the results of the reweighting method. Adapted from Schmid and Müller [107]

47 Simulation Techniques for Calculating Free Energies 113 composition fluctuation loop Fig. 22. (a) Snapshot of an interface between two coexisting phases in a binary polymer blend in the bond fluctuation model (invariant polymerization index N = 91, incompatibility χn 17, linear box dimension L 7.5Re, or number of effective segments N = 32, interaction /kb T = 0.1, monomer number density ρ0 = 1/16.0). (b) Cartoon of the configuration illustrating loops of a chain into the domain of opposite type, fluctuations of the local interface position (capillary waves) and composition fluctuations in the bulk and the shrinking of the chains in the minority phase. From Mu ller [109] scale, kb T. Each monomer along the loop contributes to this cost an amount χ kb T. Thus the typical number of monomers of a loop is 1/χ. If one assumes that loops contain many monomers and that the spatial statistics of loops corresponds to the Gaussian behavior of the entire coil the spatial extent of a loop is given by w/re (1/χ )/N (Re being the end-to-end distance). This characterizes the (intrinsic) width of the interface. Each monomer within this interfacial zone contributes to the free energy cost of the interface an amount χ. The free energy per unit area can be calculated to γ ρwχ. The mean field theory corroborates this simple picture of the interface at strong segregation (1 χ N N ) and yields [71] # N # (54) γssl = ρre χ /6N = 2 χ N/6 Re This estimate has been used to normalize the interface tension in Fig. 21. The collapse of the data for different chain lengths onto a common curve shows that the interface tension indeed only depends on the combination χ N and the data are well described by numerical mean field calculations [107] and analytic predictions by Semenov [110]. Monitoring a joint histogram of the order parameter and other quantities of interest (e.g., energy or surfactant concentration) one can calculate excess properties of the interface. An example is shown in Fig. 23 where we consider a

Lecture V: Multicanonical Simulations.

Lecture V: Multicanonical Simulations. Lecture V: Multicanonical Simulations. 1. Multicanonical Ensemble 2. How to get the Weights? 3. Example Runs (2d Ising and Potts models) 4. Re-Weighting to the Canonical Ensemble 5. Energy and Specific

More information

Multicanonical parallel tempering

Multicanonical parallel tempering JOURNAL OF CHEMICAL PHYSICS VOLUME 116, NUMBER 13 1 APRIL 2002 Multicanonical parallel tempering Roland Faller, Qiliang Yan, and Juan J. de Pablo Department of Chemical Engineering, University of Wisconsin,

More information

Generalized Ensembles: Multicanonical Simulations

Generalized Ensembles: Multicanonical Simulations Generalized Ensembles: Multicanonical Simulations 1. Multicanonical Ensemble 2. How to get the Weights? 3. Example Runs and Re-Weighting to the Canonical Ensemble 4. Energy and Specific Heat Calculation

More information

Advanced sampling. fluids of strongly orientation-dependent interactions (e.g., dipoles, hydrogen bonds)

Advanced sampling. fluids of strongly orientation-dependent interactions (e.g., dipoles, hydrogen bonds) Advanced sampling ChE210D Today's lecture: methods for facilitating equilibration and sampling in complex, frustrated, or slow-evolving systems Difficult-to-simulate systems Practically speaking, one is

More information

Melting line of the Lennard-Jones system, infinite size, and full potential

Melting line of the Lennard-Jones system, infinite size, and full potential THE JOURNAL OF CHEMICAL PHYSICS 127, 104504 2007 Melting line of the Lennard-Jones system, infinite size, and full potential Ethan A. Mastny a and Juan J. de Pablo b Chemical and Biological Engineering

More information

Multicanonical methods

Multicanonical methods Multicanonical methods Normal Monte Carlo algorithms sample configurations with the Boltzmann weight p exp( βe). Sometimes this is not desirable. Example: if a system has a first order phase transitions

More information

Lecture 8: Computer Simulations of Generalized Ensembles

Lecture 8: Computer Simulations of Generalized Ensembles Lecture 8: Computer Simulations of Generalized Ensembles Bernd A. Berg Florida State University November 6, 2008 Bernd A. Berg (FSU) Generalized Ensembles November 6, 2008 1 / 33 Overview 1. Reweighting

More information

PHASE TRANSITIONS IN SOFT MATTER SYSTEMS

PHASE TRANSITIONS IN SOFT MATTER SYSTEMS OUTLINE: Topic D. PHASE TRANSITIONS IN SOFT MATTER SYSTEMS Definition of a phase Classification of phase transitions Thermodynamics of mixing (gases, polymers, etc.) Mean-field approaches in the spirit

More information

Order-parameter-based Monte Carlo simulation of crystallization

Order-parameter-based Monte Carlo simulation of crystallization THE JOURNAL OF CHEMICAL PHYSICS 124, 134102 2006 Order-parameter-based Monte Carlo simulation of crystallization Manan Chopra Department of Chemical Engineering, University of Wisconsin, Madison, Wisconsin

More information

Phase Equilibria and Molecular Solutions Jan G. Korvink and Evgenii Rudnyi IMTEK Albert Ludwig University Freiburg, Germany

Phase Equilibria and Molecular Solutions Jan G. Korvink and Evgenii Rudnyi IMTEK Albert Ludwig University Freiburg, Germany Phase Equilibria and Molecular Solutions Jan G. Korvink and Evgenii Rudnyi IMTEK Albert Ludwig University Freiburg, Germany Preliminaries Learning Goals Phase Equilibria Phase diagrams and classical thermodynamics

More information

Understanding Molecular Simulation 2009 Monte Carlo and Molecular Dynamics in different ensembles. Srikanth Sastry

Understanding Molecular Simulation 2009 Monte Carlo and Molecular Dynamics in different ensembles. Srikanth Sastry JNCASR August 20, 21 2009 Understanding Molecular Simulation 2009 Monte Carlo and Molecular Dynamics in different ensembles Srikanth Sastry Jawaharlal Nehru Centre for Advanced Scientific Research, Bangalore

More information

Available online at ScienceDirect. Physics Procedia 57 (2014 ) 82 86

Available online at  ScienceDirect. Physics Procedia 57 (2014 ) 82 86 Available online at www.sciencedirect.com ScienceDirect Physics Procedia 57 (2014 ) 82 86 27th Annual CSP Workshops on Recent Developments in Computer Simulation Studies in Condensed Matter Physics, CSP

More information

Monte Carlo simulation of proteins through a random walk in energy space

Monte Carlo simulation of proteins through a random walk in energy space JOURNAL OF CHEMICAL PHYSICS VOLUME 116, NUMBER 16 22 APRIL 2002 Monte Carlo simulation of proteins through a random walk in energy space Nitin Rathore and Juan J. de Pablo a) Department of Chemical Engineering,

More information

CE 530 Molecular Simulation

CE 530 Molecular Simulation CE 530 Molecular Simulation Lecture 20 Phase Equilibria David A. Kofke Department of Chemical Engineering SUNY Buffalo kofke@eng.buffalo.edu 2 Thermodynamic Phase Equilibria Certain thermodynamic states

More information

ChE 210B: Advanced Topics in Equilibrium Statistical Mechanics

ChE 210B: Advanced Topics in Equilibrium Statistical Mechanics ChE 210B: Advanced Topics in Equilibrium Statistical Mechanics Glenn Fredrickson Lecture 1 Reading: 3.1-3.5 Chandler, Chapters 1 and 2 McQuarrie This course builds on the elementary concepts of statistical

More information

Wang-Landau sampling for Quantum Monte Carlo. Stefan Wessel Institut für Theoretische Physik III Universität Stuttgart

Wang-Landau sampling for Quantum Monte Carlo. Stefan Wessel Institut für Theoretische Physik III Universität Stuttgart Wang-Landau sampling for Quantum Monte Carlo Stefan Wessel Institut für Theoretische Physik III Universität Stuttgart Overview Classical Monte Carlo First order phase transitions Classical Wang-Landau

More information

Phase Transition in a Bond. Fluctuating Lattice Polymer

Phase Transition in a Bond. Fluctuating Lattice Polymer Phase Transition in a Bond arxiv:1204.2691v2 [cond-mat.soft] 13 Apr 2012 Fluctuating Lattice Polymer Hima Bindu Kolli and K.P.N.Murthy School of Physics, University of Hyderabad Hyderabad 500046 November

More information

Hanoi 7/11/2018. Ngo Van Thanh, Institute of Physics, Hanoi, Vietnam.

Hanoi 7/11/2018. Ngo Van Thanh, Institute of Physics, Hanoi, Vietnam. Hanoi 7/11/2018 Ngo Van Thanh, Institute of Physics, Hanoi, Vietnam. Finite size effects and Reweighting methods 1. Finite size effects 2. Single histogram method 3. Multiple histogram method 4. Wang-Landau

More information

Gas-liquid phase separation in oppositely charged colloids: stability and interfacial tension

Gas-liquid phase separation in oppositely charged colloids: stability and interfacial tension 7 Gas-liquid phase separation in oppositely charged colloids: stability and interfacial tension We study the phase behaviour and the interfacial tension of the screened Coulomb (Yukawa) restricted primitive

More information

A New Method to Determine First-Order Transition Points from Finite-Size Data

A New Method to Determine First-Order Transition Points from Finite-Size Data A New Method to Determine First-Order Transition Points from Finite-Size Data Christian Borgs and Wolfhard Janke Institut für Theoretische Physik Freie Universität Berlin Arnimallee 14, 1000 Berlin 33,

More information

Copyright 2001 University of Cambridge. Not to be quoted or copied without permission.

Copyright 2001 University of Cambridge. Not to be quoted or copied without permission. Course MP3 Lecture 4 13/11/2006 Monte Carlo method I An introduction to the use of the Monte Carlo method in materials modelling Dr James Elliott 4.1 Why Monte Carlo? The name derives from the association

More information

2m + U( q i), (IV.26) i=1

2m + U( q i), (IV.26) i=1 I.D The Ideal Gas As discussed in chapter II, micro-states of a gas of N particles correspond to points { p i, q i }, in the 6N-dimensional phase space. Ignoring the potential energy of interactions, the

More information

IV. Classical Statistical Mechanics

IV. Classical Statistical Mechanics IV. Classical Statistical Mechanics IV.A General Definitions Statistical Mechanics is a probabilistic approach to equilibrium macroscopic properties of large numbers of degrees of freedom. As discussed

More information

Phenomenological Theories of Nucleation

Phenomenological Theories of Nucleation Chapter 1 Phenomenological Theories of Nucleation c 2012 by William Klein, Harvey Gould, and Jan Tobochnik 16 September 2012 1.1 Introduction These chapters discuss the problems of nucleation, spinodal

More information

Markov Chain Monte Carlo Method

Markov Chain Monte Carlo Method Markov Chain Monte Carlo Method Macoto Kikuchi Cybermedia Center, Osaka University 6th July 2017 Thermal Simulations 1 Why temperature 2 Statistical mechanics in a nutshell 3 Temperature in computers 4

More information

Evaporation/Condensation of Ising Droplets

Evaporation/Condensation of Ising Droplets , Elmar Bittner and Wolfhard Janke Institut für Theoretische Physik, Universität Leipzig, Augustusplatz 10/11, D-04109 Leipzig, Germany E-mail: andreas.nussbaumer@itp.uni-leipzig.de Recently Biskup et

More information

Phase transitions for particles in R 3

Phase transitions for particles in R 3 Phase transitions for particles in R 3 Sabine Jansen LMU Munich Konstanz, 29 May 208 Overview. Introduction to statistical mechanics: Partition functions and statistical ensembles 2. Phase transitions:

More information

How the maximum step size in Monte Carlo simulations should be adjusted

How the maximum step size in Monte Carlo simulations should be adjusted Physics Procedia Physics Procedia 00 (2013) 1 6 How the maximum step size in Monte Carlo simulations should be adjusted Robert H. Swendsen Physics Department, Carnegie Mellon University, Pittsburgh, PA

More information

Crystal nucleation for a model of globular proteins

Crystal nucleation for a model of globular proteins JOURNAL OF CHEMICAL PHYSICS VOLUME 120, NUMBER 17 1 MAY 2004 Crystal nucleation for a model of globular proteins Andrey Shiryayev and James D. Gunton Department of Physics, Lehigh University, Bethlehem,

More information

AGuideto Monte Carlo Simulations in Statistical Physics

AGuideto Monte Carlo Simulations in Statistical Physics AGuideto Monte Carlo Simulations in Statistical Physics Second Edition David P. Landau Center for Simulational Physics, The University of Georgia Kurt Binder Institut für Physik, Johannes-Gutenberg-Universität

More information

Monte Carlo. Lecture 15 4/9/18. Harvard SEAS AP 275 Atomistic Modeling of Materials Boris Kozinsky

Monte Carlo. Lecture 15 4/9/18. Harvard SEAS AP 275 Atomistic Modeling of Materials Boris Kozinsky Monte Carlo Lecture 15 4/9/18 1 Sampling with dynamics In Molecular Dynamics we simulate evolution of a system over time according to Newton s equations, conserving energy Averages (thermodynamic properties)

More information

Phase Transitions. µ a (P c (T ), T ) µ b (P c (T ), T ), (3) µ a (P, T c (P )) µ b (P, T c (P )). (4)

Phase Transitions. µ a (P c (T ), T ) µ b (P c (T ), T ), (3) µ a (P, T c (P )) µ b (P, T c (P )). (4) Phase Transitions A homogeneous equilibrium state of matter is the most natural one, given the fact that the interparticle interactions are translationally invariant. Nevertheless there is no contradiction

More information

Contents. 1 Introduction and guide for this text 1. 2 Equilibrium and entropy 6. 3 Energy and how the microscopic world works 21

Contents. 1 Introduction and guide for this text 1. 2 Equilibrium and entropy 6. 3 Energy and how the microscopic world works 21 Preface Reference tables Table A Counting and combinatorics formulae Table B Useful integrals, expansions, and approximations Table C Extensive thermodynamic potentials Table D Intensive per-particle thermodynamic

More information

Monte Carlo (MC) Simulation Methods. Elisa Fadda

Monte Carlo (MC) Simulation Methods. Elisa Fadda Monte Carlo (MC) Simulation Methods Elisa Fadda 1011-CH328, Molecular Modelling & Drug Design 2011 Experimental Observables A system observable is a property of the system state. The system state i is

More information

An Introduction to Two Phase Molecular Dynamics Simulation

An Introduction to Two Phase Molecular Dynamics Simulation An Introduction to Two Phase Molecular Dynamics Simulation David Keffer Department of Materials Science & Engineering University of Tennessee, Knoxville date begun: April 19, 2016 date last updated: April

More information

Polymer Solution Thermodynamics:

Polymer Solution Thermodynamics: Polymer Solution Thermodynamics: 3. Dilute Solutions with Volume Interactions Brownian particle Polymer coil Self-Avoiding Walk Models While the Gaussian coil model is useful for describing polymer solutions

More information

Understanding temperature and chemical potential using computer simulations

Understanding temperature and chemical potential using computer simulations University of Massachusetts Amherst ScholarWorks@UMass Amherst Physics Department Faculty Publication Series Physics 2005 Understanding temperature and chemical potential using computer simulations J Tobochnik

More information

Sparse Sampling of Water Density Fluctuations. near Liquid-Vapor Coexistence

Sparse Sampling of Water Density Fluctuations. near Liquid-Vapor Coexistence Sparse Sampling of Water Density Fluctuations arxiv:183.5279v1 [cond-mat.stat-mech] 14 Mar 218 near Liquid-Vapor Coexistence Erte Xi, Sean M. Marks, Suruchi Fialoke, and Amish J. Patel Department of Chemical

More information

Chapter 2 Ensemble Theory in Statistical Physics: Free Energy Potential

Chapter 2 Ensemble Theory in Statistical Physics: Free Energy Potential Chapter Ensemble Theory in Statistical Physics: Free Energy Potential Abstract In this chapter, we discuss the basic formalism of statistical physics Also, we consider in detail the concept of the free

More information

4.1 Constant (T, V, n) Experiments: The Helmholtz Free Energy

4.1 Constant (T, V, n) Experiments: The Helmholtz Free Energy Chapter 4 Free Energies The second law allows us to determine the spontaneous direction of of a process with constant (E, V, n). Of course, there are many processes for which we cannot control (E, V, n)

More information

Optimized statistical ensembles for slowly equilibrating classical and quantum systems

Optimized statistical ensembles for slowly equilibrating classical and quantum systems Optimized statistical ensembles for slowly equilibrating classical and quantum systems IPAM, January 2009 Simon Trebst Microsoft Station Q University of California, Santa Barbara Collaborators: David Huse,

More information

Introduction Statistical Thermodynamics. Monday, January 6, 14

Introduction Statistical Thermodynamics. Monday, January 6, 14 Introduction Statistical Thermodynamics 1 Molecular Simulations Molecular dynamics: solve equations of motion Monte Carlo: importance sampling r 1 r 2 r n MD MC r 1 r 2 2 r n 2 3 3 4 4 Questions How can

More information

Computer Simulation of Peptide Adsorption

Computer Simulation of Peptide Adsorption Computer Simulation of Peptide Adsorption M P Allen Department of Physics University of Warwick Leipzig, 28 November 2013 1 Peptides Leipzig, 28 November 2013 Outline Lattice Peptide Monte Carlo 1 Lattice

More information

Chapter 4: Going from microcanonical to canonical ensemble, from energy to temperature.

Chapter 4: Going from microcanonical to canonical ensemble, from energy to temperature. Chapter 4: Going from microcanonical to canonical ensemble, from energy to temperature. All calculations in statistical mechanics can be done in the microcanonical ensemble, where all copies of the system

More information

Rare event sampling with stochastic growth algorithms

Rare event sampling with stochastic growth algorithms Rare event sampling with stochastic growth algorithms School of Mathematical Sciences Queen Mary, University of London CAIMS 2012 June 24-28, 2012 Topic Outline 1 Blind 2 Pruned and Enriched Flat Histogram

More information

Modeling the Free Energy Landscape for Janus Particle Self-Assembly in the Gas Phase. Andy Long Kridsanaphong Limtragool

Modeling the Free Energy Landscape for Janus Particle Self-Assembly in the Gas Phase. Andy Long Kridsanaphong Limtragool Modeling the Free Energy Landscape for Janus Particle Self-Assembly in the Gas Phase Andy Long Kridsanaphong Limtragool Motivation We want to study the spontaneous formation of micelles and vesicles Applications

More information

V.E Mean Field Theory of Condensation

V.E Mean Field Theory of Condensation V.E Mean Field heory of Condensation In principle, all properties of the interacting system, including phase separation, are contained within the thermodynamic potentials that can be obtained by evaluating

More information

Computational Physics (6810): Session 13

Computational Physics (6810): Session 13 Computational Physics (6810): Session 13 Dick Furnstahl Nuclear Theory Group OSU Physics Department April 14, 2017 6810 Endgame Various recaps and followups Random stuff (like RNGs :) Session 13 stuff

More information

Grand Canonical Formalism

Grand Canonical Formalism Grand Canonical Formalism Grand Canonical Ensebmle For the gases of ideal Bosons and Fermions each single-particle mode behaves almost like an independent subsystem, with the only reservation that the

More information

MONTE CARLO METHODS IN SEQUENTIAL AND PARALLEL COMPUTING OF 2D AND 3D ISING MODEL

MONTE CARLO METHODS IN SEQUENTIAL AND PARALLEL COMPUTING OF 2D AND 3D ISING MODEL Journal of Optoelectronics and Advanced Materials Vol. 5, No. 4, December 003, p. 971-976 MONTE CARLO METHODS IN SEQUENTIAL AND PARALLEL COMPUTING OF D AND 3D ISING MODEL M. Diaconu *, R. Puscasu, A. Stancu

More information

There are self-avoiding walks of steps on Z 3

There are self-avoiding walks of steps on Z 3 There are 7 10 26 018 276 self-avoiding walks of 38 797 311 steps on Z 3 Nathan Clisby MASCOS, The University of Melbourne Institut für Theoretische Physik Universität Leipzig November 9, 2012 1 / 37 Outline

More information

9.1 System in contact with a heat reservoir

9.1 System in contact with a heat reservoir Chapter 9 Canonical ensemble 9. System in contact with a heat reservoir We consider a small system A characterized by E, V and N in thermal interaction with a heat reservoir A 2 characterized by E 2, V

More information

VSOP19, Quy Nhon 3-18/08/2013. Ngo Van Thanh, Institute of Physics, Hanoi, Vietnam.

VSOP19, Quy Nhon 3-18/08/2013. Ngo Van Thanh, Institute of Physics, Hanoi, Vietnam. VSOP19, Quy Nhon 3-18/08/2013 Ngo Van Thanh, Institute of Physics, Hanoi, Vietnam. Part III. Finite size effects and Reweighting methods III.1. Finite size effects III.2. Single histogram method III.3.

More information

Statistical Thermodynamics and Monte-Carlo Evgenii B. Rudnyi and Jan G. Korvink IMTEK Albert Ludwig University Freiburg, Germany

Statistical Thermodynamics and Monte-Carlo Evgenii B. Rudnyi and Jan G. Korvink IMTEK Albert Ludwig University Freiburg, Germany Statistical Thermodynamics and Monte-Carlo Evgenii B. Rudnyi and Jan G. Korvink IMTEK Albert Ludwig University Freiburg, Germany Preliminaries Learning Goals From Micro to Macro Statistical Mechanics (Statistical

More information

Chemical Potential. Combining the First and Second Laws for a closed system, Considering (extensive properties)

Chemical Potential. Combining the First and Second Laws for a closed system, Considering (extensive properties) Chemical Potential Combining the First and Second Laws for a closed system, Considering (extensive properties) du = TdS pdv Hence For an open system, that is, one that can gain or lose mass, U will also

More information

Computer simulation methods (1) Dr. Vania Calandrini

Computer simulation methods (1) Dr. Vania Calandrini Computer simulation methods (1) Dr. Vania Calandrini Why computational methods To understand and predict the properties of complex systems (many degrees of freedom): liquids, solids, adsorption of molecules

More information

An improved Monte Carlo method for direct calculation of the density of states

An improved Monte Carlo method for direct calculation of the density of states JOURNAL OF CHEMICAL PHYSICS VOLUME 119, NUMBER 18 8 NOVEMBER 2003 An improved Monte Carlo method for direct calculation of the density of states M. Scott Shell, a) Pablo G. Debenedetti, b) and Athanassios

More information

Efficient Combination of Wang-Landau and Transition Matrix Monte Carlo Methods for Protein Simulations

Efficient Combination of Wang-Landau and Transition Matrix Monte Carlo Methods for Protein Simulations Journal of Computational Chemistry, in press (2006) Efficient Combination of Wang-Landau and Transition Matrix Monte Carlo Methods for Protein Simulations Ruben G. Ghulghazaryan 1,2, Shura Hayryan 1, Chin-Kun

More information

Thermodynamics of nuclei in thermal contact

Thermodynamics of nuclei in thermal contact Thermodynamics of nuclei in thermal contact Karl-Heinz Schmidt, Beatriz Jurado CENBG, CNRS/IN2P3, Chemin du Solarium B.P. 120, 33175 Gradignan, France Abstract: The behaviour of a di-nuclear system in

More information

Nanoscale simulation lectures Statistical Mechanics

Nanoscale simulation lectures Statistical Mechanics Nanoscale simulation lectures 2008 Lectures: Thursdays 4 to 6 PM Course contents: - Thermodynamics and statistical mechanics - Structure and scattering - Mean-field approaches - Inhomogeneous systems -

More information

to satisfy the large number approximations, W W sys can be small.

to satisfy the large number approximations, W W sys can be small. Chapter 12. The canonical ensemble To discuss systems at constant T, we need to embed them with a diathermal wall in a heat bath. Note that only the system and bath need to be large for W tot and W bath

More information

8.333: Statistical Mechanics I Problem Set # 5 Due: 11/22/13 Interacting particles & Quantum ensembles

8.333: Statistical Mechanics I Problem Set # 5 Due: 11/22/13 Interacting particles & Quantum ensembles 8.333: Statistical Mechanics I Problem Set # 5 Due: 11/22/13 Interacting particles & Quantum ensembles 1. Surfactant condensation: N surfactant molecules are added to the surface of water over an area

More information

Wang-Landau Sampling of an Asymmetric Ising Model: A Study of the Critical Endpoint Behavior

Wang-Landau Sampling of an Asymmetric Ising Model: A Study of the Critical Endpoint Behavior Brazilian Journal of Physics, vol. 36, no. 3A, September, 26 635 Wang-andau Sampling of an Asymmetric Ising Model: A Study of the Critical Endpoint Behavior Shan-o sai a,b, Fugao Wang a,, and D.P. andau

More information

arxiv:cond-mat/ v1 [cond-mat.stat-mech] 5 Jun 2003

arxiv:cond-mat/ v1 [cond-mat.stat-mech] 5 Jun 2003 Performance imitations of Flat Histogram Methods and Optimality of Wang-andau Sampling arxiv:cond-mat/030608v [cond-mat.stat-mech] 5 Jun 2003 P. Dayal, S. Trebst,2, S. Wessel, D. Würtz, M. Troyer,2, S.

More information

CH 240 Chemical Engineering Thermodynamics Spring 2007

CH 240 Chemical Engineering Thermodynamics Spring 2007 CH 240 Chemical Engineering Thermodynamics Spring 2007 Instructor: Nitash P. Balsara, nbalsara@berkeley.edu Graduate Assistant: Paul Albertus, albertus@berkeley.edu Course Description Covers classical

More information

Many proteins spontaneously refold into native form in vitro with high fidelity and high speed.

Many proteins spontaneously refold into native form in vitro with high fidelity and high speed. Macromolecular Processes 20. Protein Folding Composed of 50 500 amino acids linked in 1D sequence by the polypeptide backbone The amino acid physical and chemical properties of the 20 amino acids dictate

More information

Multiple time step Monte Carlo

Multiple time step Monte Carlo JOURNAL OF CHEMICAL PHYSICS VOLUME 117, NUMBER 18 8 NOVEMBER 2002 Multiple time step Monte Carlo Balázs Hetényi a) Department of Chemistry, Princeton University, Princeton, NJ 08544 and Department of Chemistry

More information

Phase transitions and finite-size scaling

Phase transitions and finite-size scaling Phase transitions and finite-size scaling Critical slowing down and cluster methods. Theory of phase transitions/ RNG Finite-size scaling Detailed treatment: Lectures on Phase Transitions and the Renormalization

More information

Decoherence and the Classical Limit

Decoherence and the Classical Limit Chapter 26 Decoherence and the Classical Limit 26.1 Introduction Classical mechanics deals with objects which have a precise location and move in a deterministic way as a function of time. By contrast,

More information

Statistical Mechanics

Statistical Mechanics Franz Schwabl Statistical Mechanics Translated by William Brewer Second Edition With 202 Figures, 26 Tables, and 195 Problems 4u Springer Table of Contents 1. Basic Principles 1 1.1 Introduction 1 1.2

More information

(Crystal) Nucleation: The language

(Crystal) Nucleation: The language Why crystallization requires supercooling (Crystal) Nucleation: The language 2r 1. Transferring N particles from liquid to crystal yields energy. Crystal nucleus Δµ: thermodynamic driving force N is proportional

More information

Phase transitions of quadrupolar fluids

Phase transitions of quadrupolar fluids Phase transitions of quadrupolar fluids Seamus F. O Shea Department of Chemistry, University of Lethbridge, Lethbridge, Alberta, Canada, T1K 3M4 Girija S. Dubey Brookhaven National Laboratory, Upton, New

More information

A Brief Introduction to Statistical Mechanics

A Brief Introduction to Statistical Mechanics A Brief Introduction to Statistical Mechanics E. J. Maginn, J. K. Shah Department of Chemical and Biomolecular Engineering University of Notre Dame Notre Dame, IN 46556 USA Monte Carlo Workshop Universidade

More information

Non equilibrium thermodynamics: foundations, scope, and extension to the meso scale. Miguel Rubi

Non equilibrium thermodynamics: foundations, scope, and extension to the meso scale. Miguel Rubi Non equilibrium thermodynamics: foundations, scope, and extension to the meso scale Miguel Rubi References S.R. de Groot and P. Mazur, Non equilibrium Thermodynamics, Dover, New York, 1984 J.M. Vilar and

More information

Monte Carlo study of the Baxter-Wu model

Monte Carlo study of the Baxter-Wu model Monte Carlo study of the Baxter-Wu model Nir Schreiber and Dr. Joan Adler Monte Carlo study of the Baxter-Wu model p.1/40 Outline Theory of phase transitions, Monte Carlo simulations and finite size scaling

More information

Collective behavior, from particles to fields

Collective behavior, from particles to fields 978-0-51-87341-3 - Statistical Physics of Fields 1 Collective behavior, from particles to fields 1.1 Introduction One of the most successful aspects of physics in the twentieth century was revealing the

More information

Physics 115/242 Monte Carlo simulations in Statistical Physics

Physics 115/242 Monte Carlo simulations in Statistical Physics Physics 115/242 Monte Carlo simulations in Statistical Physics Peter Young (Dated: May 12, 2007) For additional information on the statistical Physics part of this handout, the first two sections, I strongly

More information

Principles of Equilibrium Statistical Mechanics

Principles of Equilibrium Statistical Mechanics Debashish Chowdhury, Dietrich Stauffer Principles of Equilibrium Statistical Mechanics WILEY-VCH Weinheim New York Chichester Brisbane Singapore Toronto Table of Contents Part I: THERMOSTATICS 1 1 BASIC

More information

LECTURE 11: Monte Carlo Methods III

LECTURE 11: Monte Carlo Methods III 1 LECTURE 11: Monte Carlo Methods III December 3, 2012 In this last chapter, we discuss non-equilibrium Monte Carlo methods. We concentrate on lattice systems and discuss ways of simulating phenomena such

More information

THE DETAILED BALANCE ENERGY-SCALED DISPLACEMENT MONTE CARLO ALGORITHM

THE DETAILED BALANCE ENERGY-SCALED DISPLACEMENT MONTE CARLO ALGORITHM Molecular Simulation, 1987, Vol. 1, pp. 87-93 c Gordon and Breach Science Publishers S.A. THE DETAILED BALANCE ENERGY-SCALED DISPLACEMENT MONTE CARLO ALGORITHM M. MEZEI Department of Chemistry, Hunter

More information

Potts And XY, Together At Last

Potts And XY, Together At Last Potts And XY, Together At Last Daniel Kolodrubetz Massachusetts Institute of Technology, Center for Theoretical Physics (Dated: May 16, 212) We investigate the behavior of an XY model coupled multiplicatively

More information

Khinchin s approach to statistical mechanics

Khinchin s approach to statistical mechanics Chapter 7 Khinchin s approach to statistical mechanics 7.1 Introduction In his Mathematical Foundations of Statistical Mechanics Khinchin presents an ergodic theorem which is valid also for systems that

More information

Lecture 14: Advanced Conformational Sampling

Lecture 14: Advanced Conformational Sampling Lecture 14: Advanced Conformational Sampling Dr. Ronald M. Levy ronlevy@temple.edu Multidimensional Rough Energy Landscapes MD ~ ns, conformational motion in macromolecules ~µs to sec Interconversions

More information

CE 530 Molecular Simulation

CE 530 Molecular Simulation CE 530 Molecular Simulation Lecture 0 Simple Biasing Methods David A. Kofke Department of Chemical Engineering SUNY Buffalo kofke@eng.buffalo.edu 2 Review Monte Carlo simulation Markov chain to generate

More information

Pressure Dependent Study of the Solid-Solid Phase Change in 38-Atom Lennard-Jones Cluster

Pressure Dependent Study of the Solid-Solid Phase Change in 38-Atom Lennard-Jones Cluster University of Rhode Island DigitalCommons@URI Chemistry Faculty Publications Chemistry 2005 Pressure Dependent Study of the Solid-Solid Phase Change in 38-Atom Lennard-Jones Cluster Dubravko Sabo University

More information

Intro. Each particle has energy that we assume to be an integer E i. Any single-particle energy is equally probable for exchange, except zero, assume

Intro. Each particle has energy that we assume to be an integer E i. Any single-particle energy is equally probable for exchange, except zero, assume Intro Take N particles 5 5 5 5 5 5 Each particle has energy that we assume to be an integer E i (above, all have 5) Particle pairs can exchange energy E i! E i +1andE j! E j 1 5 4 5 6 5 5 Total number

More information

4/18/2011. Titus Beu University Babes-Bolyai Department of Theoretical and Computational Physics Cluj-Napoca, Romania

4/18/2011. Titus Beu University Babes-Bolyai Department of Theoretical and Computational Physics Cluj-Napoca, Romania 1. Introduction Titus Beu University Babes-Bolyai Department of Theoretical and Computational Physics Cluj-Napoca, Romania Bibliography Computer experiments Ensemble averages and time averages Molecular

More information

arxiv: v1 [cond-mat.soft] 22 Oct 2007

arxiv: v1 [cond-mat.soft] 22 Oct 2007 Conformational Transitions of Heteropolymers arxiv:0710.4095v1 [cond-mat.soft] 22 Oct 2007 Michael Bachmann and Wolfhard Janke Institut für Theoretische Physik, Universität Leipzig, Augustusplatz 10/11,

More information

Metropolis Monte Carlo simulation of the Ising Model

Metropolis Monte Carlo simulation of the Ising Model Metropolis Monte Carlo simulation of the Ising Model Krishna Shrinivas (CH10B026) Swaroop Ramaswamy (CH10B068) May 10, 2013 Modelling and Simulation of Particulate Processes (CH5012) Introduction The Ising

More information

THE TANGO ALGORITHM: SECONDARY STRUCTURE PROPENSITIES, STATISTICAL MECHANICS APPROXIMATION

THE TANGO ALGORITHM: SECONDARY STRUCTURE PROPENSITIES, STATISTICAL MECHANICS APPROXIMATION THE TANGO ALGORITHM: SECONDARY STRUCTURE PROPENSITIES, STATISTICAL MECHANICS APPROXIMATION AND CALIBRATION Calculation of turn and beta intrinsic propensities. A statistical analysis of a protein structure

More information

PHYSICAL REVIEW LETTERS

PHYSICAL REVIEW LETTERS PHYSICAL REVIEW LETTERS VOLUME 86 28 MAY 21 NUMBER 22 Mathematical Analysis of Coupled Parallel Simulations Michael R. Shirts and Vijay S. Pande Department of Chemistry, Stanford University, Stanford,

More information

arxiv:cond-mat/ v1 10 Aug 2002

arxiv:cond-mat/ v1 10 Aug 2002 Model-free derivations of the Tsallis factor: constant heat capacity derivation arxiv:cond-mat/0208205 v1 10 Aug 2002 Abstract Wada Tatsuaki Department of Electrical and Electronic Engineering, Ibaraki

More information

Distance Constraint Model; Donald J. Jacobs, University of North Carolina at Charlotte Page 1 of 11

Distance Constraint Model; Donald J. Jacobs, University of North Carolina at Charlotte Page 1 of 11 Distance Constraint Model; Donald J. Jacobs, University of North Carolina at Charlotte Page 1 of 11 Taking the advice of Lord Kelvin, the Father of Thermodynamics, I describe the protein molecule and other

More information

5.1 2D example 59 Figure 5.1: Parabolic velocity field in a straight two-dimensional pipe. Figure 5.2: Concentration on the input boundary of the pipe. The vertical axis corresponds to r 2 -coordinate,

More information

1. Thermodynamics 1.1. A macroscopic view of matter

1. Thermodynamics 1.1. A macroscopic view of matter 1. Thermodynamics 1.1. A macroscopic view of matter Intensive: independent of the amount of substance, e.g. temperature,pressure. Extensive: depends on the amount of substance, e.g. internal energy, enthalpy.

More information

Equilibrium sampling of self-associating polymer solutions: A parallel selective tempering approach

Equilibrium sampling of self-associating polymer solutions: A parallel selective tempering approach THE JOURNAL OF CHEMICAL PHYSICS 123, 124912 2005 Equilibrium sampling of self-associating polymer solutions: A parallel selective tempering approach Chakravarthy Ayyagari, a Dmitry Bedrov, and Grant D.

More information

summary of statistical physics

summary of statistical physics summary of statistical physics Matthias Pospiech University of Hannover, Germany Contents 1 Probability moments definitions 3 2 bases of thermodynamics 4 2.1 I. law of thermodynamics..........................

More information

MACROSCOPIC VARIABLES, THERMAL EQUILIBRIUM. Contents AND BOLTZMANN ENTROPY. 1 Macroscopic Variables 3. 2 Local quantities and Hydrodynamics fields 4

MACROSCOPIC VARIABLES, THERMAL EQUILIBRIUM. Contents AND BOLTZMANN ENTROPY. 1 Macroscopic Variables 3. 2 Local quantities and Hydrodynamics fields 4 MACROSCOPIC VARIABLES, THERMAL EQUILIBRIUM AND BOLTZMANN ENTROPY Contents 1 Macroscopic Variables 3 2 Local quantities and Hydrodynamics fields 4 3 Coarse-graining 6 4 Thermal equilibrium 9 5 Two systems

More information

Free energy simulations

Free energy simulations Free energy simulations Marcus Elstner and Tomáš Kubař January 14, 2013 Motivation a physical quantity that is of most interest in chemistry? free energies Helmholtz F or Gibbs G holy grail of computational

More information

Advanced Monte Carlo Methods Problems

Advanced Monte Carlo Methods Problems Advanced Monte Carlo Methods Problems September-November, 2012 Contents 1 Integration with the Monte Carlo method 2 1.1 Non-uniform random numbers.......................... 2 1.2 Gaussian RNG..................................

More information