View metadata, citation and similar papers at core.ac.

uk brought to you by CORE

provided by Open Archive Toulouse Archive Ouverte

Book Title:
Puigjaner, Heyen (Eds.): Computer Aided Process & Product Engineering - CAPE

Title of the chapter

Molecular modelling for physical property prediction

Corresponding author:

Vincent GERBAUD
Laboratoire de Génie Chimique
BP 1301, 5 rue Paulin Talabot
31106 Toulouse Cedex 1
France
[email protected]

Coauthor:

Xavier JOULIA
ENSIACET – Laboratoire de Génie Chimique
117 Route de Narbonne
31077 Toulouse Cedex 4
France
[email protected]

1
 3.2. Molecular modeling for physical property prediction

3.2.1. INTRODUCTION
Multiscale modeling is becoming the standard approach for process study in a broader framework that
promotes computer aided integrated product and process design. In addition to usual purity
requirements, end products must meet new constraints in terms of environmental impact, safety of
goods and people, specific properties. Engineering achievements can be startling from the user side
like aqueous solvent paint still washable after drying!
This can only be done by improving process knowledge and performance at all scales, down to the
atomic one. Current experimental and modeling approaches assess with difficultly such sub-micronic
scales. In experiments, how to conceive experimental devices small enough and introduce them in
molecular systems without affecting irreversibly the phenomena that they look at? In modeling and
simulation, which hypotheses are still relevant? how to handle boundary effects? Numerical difficulties
may arise along with the necessity of defining new parameters… that will be adjustable ones as no
experiments can obtain them. This latter statement is particularly true for energetic interaction
parameters like binary interaction parameters in current liquide – vapor equilibrium macroscopic
thermodynamic models based on activity coefficient approach or on equation of state approach. In all
processes, study of phenomena attributed to energetic interactions has always been left over for a
time … that has come:
Indeed, molecular modeling is a field of study that in interested in the behavior of atomic and
molecular systems subject to energetic interactions. It is then a natural complement of experimental
and modeling approaches to expand multiscale approaches towards smaller scales. Besides, process
flows concern primarily molecules from raw materials to end products. Therefore, at any process
development step, the challenge of knowing the physical properties and thermodynamical state of
molecules is critical. But, the future of this challenge is dim when one thinks about the millions of
chemical compounds referenced in the chemical abstract series. Neither experimental approaches nor
current themodynamic models can handle the combination of properties needed. In some cases
experiments are not even practical because of materials decomposition or safety issues. Universal
group contribution methods are a pipe dream and existing ones are efficient but are restricted in use to
specific areas like petrochemical and small molecular systems.
As providers of accurate physico-chemical data, molecular modeling methods offer an alternative to
an intensive and expensive experimental campaign, once molecular models are available, which is
becoming increasingly the case [CAS 04]. But this first goal is nothing compared to the main interest
of molecular methods that is, probing the matter at the molecular scale [CHE 02 ; DEP 02, SAN 03].
Indeed, molecular modeling can be seen as a “third way to explore real matter” [ALL 87]. Like a
theoretical approach, it is based on a model system of the real one. But unlike theory, no hypothesis,
no transcription of key phenomena into equations or correlations is performed. Rather, molecular
modeling performs a numerical experiments to simulate directly the behavior of the model system.
The concept of numerical experiment is strong. Fisrt, the model system is made of a boundered
molecular system and of an interaction model analogous to an experimental sensor that enables to
compute the internal energy of the model system. Second, think of the pseudo constant thermometer
temperature and of the brownian motion of atoms in a liquid that generates a fluctuating temperature.
More generally, any macroscopic property value measured by an experimental probe is a time
average over many instantaneous fluctuating values. Statistical thermodynamic postulates that this
time average equals an ensemble average over a statistically significant numbers of configurations of
the model system. Molecular modeling generates them numerically using methods like molecular

1
dynamics or Monte Carlo Methods. Any property of interest is then derived using thermodynamical
laws from instantaneous property value averages and correlation factors. Thirdly, numerical standard
deviation associated to the ensemble average is the equivalent of experimental accuracy.
This chapter presents molecular modeling concepts so as to demystify them and stress their interests
for chemical engineers. Multiscale approach including molecular modeling are not illustrated due to
restricted space. Rather, routine examples on the use of several molecular techniques suitable to get
accurate vapor – liquid equilibrium data when no data is available are provided.

3.2.2. WHAT IS MOLECULAR MODELING?

Molecular modeling includes computer theoretical chemistry and molecular simulation.
Computer theoretical chemistry calculations are carried out at 0K and solve Schrödinger equation to
obtain nuclear and electronic properties such as conformation, orbital, density of load, electrostatic
potential surface in a fundamental or excited states. Computation time is huge, being proportional at
best to N electrons 2.5 which restricts its use to small systems. The precision of the results is significant
because the only assumptions are linked to approximations carried out to solve the equation of
Schrödinger. In particular, there are no adjustable parameters. Besides, it provides crucial information
on the electronic distribution that enables to evaluate electrostatic interactions in molecular simulation.
Molecular simulation is a numerical technique to get the physicochemical properties of macroscopic
systems from the description on an atomic scale of the elementary interactions and from the
application of statistical thermodynamics principles. It concerns the calculation of a model system
internal energy at a positive temperature. Computation time is proportional to N molecules , which makes it
a technique adapted to the study of real systems: phase properties, interfaces, reaction, transport
phenomena.... Molecular simulation carries out a dynamic modeling of the system subjected to
realistic temperature and pressure conditions thanks to an adequate sampling of the system
configurations. A configuration is a set of particles coordinates and connections. Inaccuracy may arise
from the energetic models that contain fitted but physically meaningful parameters or from system
configurations sampling techniques that must comply with statistical thermodynamic principles.
Molecular simulation offers the most potentialities for process engineering. Wherever energetic
interaction related phenomena have a prevalent place, molecular simulation deserves to be
considered to study and look further into the knowledge of the phenomena in the heart of the
processes. In particular, it is suitable for the study of phase equilibrium, interfacial properties (specific
adsorption on catalyst…), transport coefficients, chemical reactivity, activity coefficients ….

3.2.2.1 Scientific challenges of molecular modeling in Process

Engineering.

The use of molecular simulation in Process Engineering lies mainly in the difficulty of establishing the
link between the macroscopic properties and their energetic description or that of significant
parameters at mesoscopic or molecular scale. The micro – macro relation can be simple: in distillation
the knowledge of phase equilibrium data enables to run an extensive study and design of the process.
In tablet processing, the relation is more complex: the tablet properties (compactness, friability,
dissolution) are related to the pellets cohesion and to the substrates solubility. Obviously energetic
interaction is a key phenomena and is taken into account through solubility parameters which can be
broken down into primarily energy contributions (Van der Waals repulsion - attraction, Coulombic
interaction…), precisely the applicability of molecular simulation. But particle size and solvent effects
on the aggregates size and homogeneity are equally important notwithstanding operational process
parameters and are still difficult to address at a molecular scale. So, identifying the limiting
phenomena is a priority before any molecular simulation.

2
Size of the model systems is not an unsolvable problem as periodic boundary conditions can be
applied to replicate the original system box and mimic a homogeneous macroscopic phase. Rather,
the scientific challenges concern issues often encountered in experiments, the sensor challenge, the
sampling challenge and the multiscale challenge.

3.2.2.1.1 The sensor challenge

For data-oriented simulations, accurate force fields / sensors are needed to evaluate precisely
energetic interactions. Study of highly polar systems, reliable and relevant extrapolation of carefully
set force field parameters, absence of temperature dependency of these parameters are key
improvements of molecular simulation models over existing macroscopic models.
The model system is usually a parallelepiped box filled with particles which energetic interactions are
described by a force field enabling to compute the system internal energy. In order to mimic a
homogeneous phase, the box is usually replicated in 3-D by applying periodic boundary conditions.
Typical size ranges from 20 to 1000 Å and may vary during simulations. Edge effects are to be
envisaged and can be attenuated by increase the box size.
The development of a force field requires a strong collaboration with theoretical chemists and
physicists. Indeed, different kinds of force fields can arise: some based on quantum chemistry
concepts, some based on molecular mechanics (Sandler, 2003).
Quantum-based models are used in static modeling and naturally in computer theoretical chemistry
calculations. Solving the Schrödinger equation, they provide the nuclear and electronic properties
system and consequently the true energy of the system (e.g. the energy of ionization) physically
measurable. Molecular mechanics models are used in molecular simulation to calculate intensive
properties (T, P) and extensive ones among which the internal energy of the system which is not
measurable directly by an experiment but enables to calculate other thermodynamic properties by
using thermodynamic laws. Properties like vaporization enthalpy connected to differences in internal
energy are computed and can be compared to experiments.
Quantum (QM) models are practical on a few tens of atoms at best and more and more used in
combination with molecular mechanics models for some part of the system where accurate electronic
distribution are needed, e.g. a reactive zone or to provide a description of the electronic distribution.
Molecular mechanics (MM) models are the most used and are based on a springs and beads
mechanistic description of the intermolecular interactions and of intramolecular bonds. They allow
calculations on several hundreds of particles which enable to model real systems in a satisfactory
way. They contain physical parameters evaluated from quantum calculation but also empirical
parameters which must be regressed from experimental data. However, this empiricism is attenuated
by some physical significance attributed to the parameters. Moreover, MM force fields show amazing
properties: valid over a large pressure and temperature range, they can be used to compute many
properties and all molecules can be described from a small set of parameters if careful
parameterization is conducted, which constitutes the first challenge.

3.2.2.1.2 The sampling challenge

The second challenge requires a strong involvement of process engineers: novel and smart methods
must be developed to sample specific states of the model system which are of great interest for
Process Engineering: transition states that set the reaction energetic barrier, azeotropes that affect
strongly the distillation process feasibility and design, dew points …. Usually, existing molecular
simulation methods sample non specific states like a vapor – liquid equilibrium point. Unlike
measurement time, its experimental equivalent, numerical sampling can be advantageously biased to
sample the specific state of interest but it requires expertise to comply with statistical thermodynamics
principles which enable to bridge the microscopic and the macroscopic scales.

3
Furthermore, for existing methods based on molecular dynamic or Monte Carlo methods, sampling
efficiency should be improved in particular for complex molecules like macromolecules even if the
alternate solution of running more simulations is still the leading choice as computer power increases.
With this second challenge, Process Engineering finds a new use for molecular modeling: it cannot be
solely data-oriented but also discovery-oriented and assumes its status of numerical experiment.

3.2.2.1.3 Molecular modeling in a multiscale approach

The integration of molecular modeling in applicable models for the study of macroscopic systems and
their properties is of the utmost importance for process engineering. Indeed, often considered as
decisive, phenomena related to energetic interactions have often been left aside during a process
study because of a lack of suitable tools or incorporated into parameters. Thermodynamic models
used in phase equilibrium calculations are a good example: binary interaction parameters must be
found empirically despite their solid physical meaning. The first illustrative example addresses the
issue of calculating binary parameters by molecular modeling methods.
Process Engineering models are knowledge-based models. In most domains, process study requires
a multiscale approach. As a technique of experimentation, molecular modeling makes it possible to
visualize on a molecular scale physicochemical phenomena. It can thus be used to develop or revisit
theories, models or parameters of models and therefore improve our knowledge of processes and
increase the capacity of predictions and extrapolation of existing models.

Process Unit operation Interface Molecule

Process Engineering Molecular Modeling

MACRO MICRO MESO NANO

Figure 1. Process Engineering and Molecular Modeling.

3.2.3. STATISTICAL THERMODYNAMIC BACKGROUND

Suggested readings:
D.A. Mc Quarry, Statistical thermodynamics, Harper and Collins Publishers, New York, USA, 1976.
M.P. Allen and D.J. Tildesley, Computer Simulation of Liquids, Oxford University Publications, UK,
1987. [new edition in 2000].
D. Frenkel, B. Smit , Understanding Molecular Simulation. From Algorithms to Applications, Academic
Press, San Diego, USA, 1996. [new edition in 2002].

3.2.3.1 A microscopic description of macroscopic properties

Traditional thermodynamics and statistical thermodynamics are interested in the same problems but
differ in their approach: thermodynamics provides general relations without any consideration of the
intrinsic constitution of the matter while statistical thermodynamics supposes the existence of atoms,
molecules, of particles, to calculate and interpret thermodynamic properties at the molecular level.
The objective of statistical thermodynamics is to describe the behavior of a macroscopic system in
terms of microscopic properties of a system of molecular entities.

4
 Ti
Tobs

-3 -8
Measurement time (10 s) >> fluctuations (10 s)
Figure 2. Measurement of a ‘mean’ temperature and its relation with instantaneous temperature
The main idea is to evaluate an average property value and its standard deviation from a statistically
significant number of configurations, much like a real experiment. Indeed, the temperature reading on
a thermometer appears falsely constant. At the molecular level, a positive temperature is the resultant
of atomic vibrations and collisions occurring at a time scale (e.g. 10-8 s.) much lower than the sampling
period of the experimental sensor (e.g. 10-3 s.) (Figure 2). Using statistical thermodynamic concepts,
molecular simulation will do the same and perform a numerical experiment. Each instantaneous
configuration (atomic positions and moments) of the system exists according to a probability
distribution. The most probable will have the largest contribution to the computed average value. For
the experimental system the macroscopic property X value is a time average over a set of
configurations (t) sampled during the measurement time t meas :

X macro  Xt time  0meas Xt dt

1 t
lim [1]
t meas   meas
t

But knowing all configurations (t) is impractical because the number of particles (6.023x1023 for a
mole) and thus the number of positions and moments are incommensurable. Statistical
thermodynamics was developed to solve this problem statistically.
The first postulate of statistical thermodynamics is that "the value of any macroscopic property is equal
to its average value over a sample of the model system configurations".

 total
Xt time  X  X
1
 First postulate [2]
ensemble 
total   1
where  total is the number of sampled configurations. The notation  ensemble refers to a statistical
ensemble. By definition it consists in a significant number of sub-ensembles having the same
macroscopic properties. The thermodynamic state of a macroscopic system is perfectly specified by a
few parameters, for example the number of mole N, the pressure P and the temperature T. From
them, one can derive a great number of properties (density, chemical potential, heat capacity, diffusion
coefficient, viscosity coefficient …) through equations of state and other thermodynamic relations.
Reproducing conditions occurring in experiments, the « canonical » NVT, and the « isobar -
isothermal » NPT are quite useful. Notations, NVT and NPT means respectively that the number of
moles N + volume V + the temperature T; the number of moles N + the pressure P + the temperature
T are kept constant for each system configuration during simulations run in those ensembles.
One considers that the postulate of statistical thermodynamics apply during simulations in a statistical
ensemble on systems with a few thousands of particles replicated by periodic boundary conditions and
that averages are made on a few million configurations. Sampling size and quality are often the
Achilles' heels of molecular simulations.

3.2.3.2 Probability density

Equation 2 states that configurations have the same weight in the average, the same probability of
existence. This is the second postulate of statistical thermodynamics: "All the accessible and distinct
quantum states from a closed system of fixed energy (« microcanonic » NVE) are equiprobable".
Then equation 2 is rewritten:

5
 Xt time  X
ensemble
  X   ensemble  Second postulate [3]


where  ensemble () is the probability density, that is the probability of finding a configuration with
positions and moments (). In the NVT ensemble, any configuration probability density is connected
to its energy E and to Q NVT the total partition function, namely the sum over all configurations, by the
Boltzmann formula:

1  E 
NVT  exp 
 [4]
QNVT  kBT 

Two points are noteworthy:

1. The knowledge of the partition function would allow calculating all thermodynamic properties.
But this can never be done fully but rather imperfectly through the generation of a statistically
representative number of configurations.
2. A model is required to evaluate any configuration energy in order to calculate the partition
function. This is done through a force field.

3.2.3.3 Average, fluctuations and correlation functions

Equation 2 is the usual mean formula to calculate an average value (molar fractions, …). Other
properties (heat capacity, …) are calculated from the variance expressing the fluctuations around the
mean:

 total
2
X
1 
 total   1 
 2
  X  X ensemble   X2  X

2
 Variance [5]

Correlation coefficients give access to properties describing the dynamic state of the system. The non
normalized form of the correlation coefficient over  configurations is:

 total
correl   X  X0  
1
 X   X
0 0   Correlation coefficient [6]
XX total
 0 1

The integration of the non normalized correlation coefficients enables to calculate directly macroscopic
transfer coefficients (diffusion, viscosity or thermal diffusivity coefficient). Their Fourier transform can
be compared with experimental spectra.

3.2.3.4 Statistical error

Molecular simulation is a numerical experiment. Consequently, the results are prone to systematic and
statistical errors. The systematic errors must be evaluated then eliminated. They are caused by size
effects, bad random number generation, insufficient equilibration period (see further). Statistical errors
are inversely proportional to sampling and are thus null for infinite sampling.
On the assumption that the Gauss law applies, the statistical error is the variance (equation 5).
However, sampling a large but finite number of configurations induces a correlation between the  total
configurations which persists during a certain number of successive configurations. A statistical factor
of inefficiency s is introduced to evaluates the number of correlated successive configurations. The
 total configurations are cut into n b blocs of  b configurations upon which the average X b and its
variance 2(X b are computed. By selecting several increasing values of  b , the statistical inefficiency
s and the statistical error 2( X total ) is evaluated:

6
 b   2 X
b
 
s lim [7]
 b    X 
2
total


 X total
 s
 total
 x total [8]

3.2.4. NUMERICAL SAMPLING TECHNIQUES

Trajectory

t t t
 F  m dt
dVi
i
i
?
Molecular Dynamics

?
  Ei  
Monte Carlo accept  min1, exp 


?   k BT  

Random configurations

Figure 3. Basic concepts of the Monte Carlo method and Molecular Dynamics
The generation of a statistically representative sample of the model system configurations is mainly
done by two techniques, molecular dynamics and the Monte Carlo method. They obey the principles
summarized on figure 3. Both methods differ in their applications: the Monte Carlo method is adapted
for the study of static phenomena (equilibrium, static interface) while molecular dynamics is adapted to
the study of dynamic phenomena (shear induced flow): a phase equilibrium easily computed with
Monte Carlo methods would be extremely difficult to reach in molecular dynamics because of the time
needed and of boundary effects near the interface.

3.2.4.1 Molecular Dynamics

Molecular dynamics generates a trajectory by integrating the classical equation of motion over time
steps t starting from an initial configuration whose particle positions and velocity are known (figure 3).
The nth configuration can be traced down the initial one by reverse integration. In the equation of
motion (figure 3), the forces F i acting on the particle of mass m i are equal to the derivative of the V i (r)
potential describing the interactions of the particle i with its surrounding
The integration of the differential set of equations is carried out mainly by Verlet-like algorithms rather
than by Gear-like algorithms which are widespread in process engineering. The Verlet algorithm
calculates the new particle positions r(t) using a 3rd order taylor expansion and replacing the second
derivative by the forces thanks to the equation of motion, one gets a formula with no velocity term:

r t  t   2r t   r t  t  
F( t ) t 2
m

2!
 O t 4   [9]

Velocities are computed afterwards:

dr t  r t  t   r t  t 
v t  
dt

2t
 O t 2   [10]

This algorithm shows several interesting characteristics. (i) It is symmetrical versus t, which makes
the trajectory reversible over the time. (ii) It preserves the total energy of the system over long periods
of integration, a key point to get long trajectories and deduce with accuracy some correlation

7
functions. In particular, it is more precise than the Gear-like algorithms for large t (the reverse is true
for small t), what makes it suitable to simulate long trajectories, which is our goal. (iii) It requires less
data storage than Gear-like algorithms.
Transport coefficients (self diffusion, thermal diffusivity and viscosity) are computed from auto-
correlation coefficients, the ‘green-kubo' formulas, for instance the coefficient of self diffusion D i is
related to the relative particle velocities:

v i   v i 0  d
1
Di 
3 
O
[11]

Similarly, viscosity is obtained from the shear stress tensor auto-correlation coefficient related to the
pressure exerted on the particle and the thermal diffusivity is obtained from the energy flow auto-
correlation coefficient.
A challenge in molecular dynamics run in a statistical ensemble where the temperature is set constant
is to keep it constant when moving and interacting particles inevitably heat the system. A solution is to
place the system in a large thermostated bath periodically set in contact with the model system
through techniques like Andersen or Nose-Hoover methods.

3.2.4.2 Monte Carlo Method

The Monte Carlo method generates system configurations randomly. The nth configuration is related to
its preceding one but it is impossible to go back to the initial configuration.
First of all, randomness is particularly critical and has given its name to the method in reference to the
Monte Carlo casino. The advice is to always use a published robust random number generator and
never try to build one or use the falsely random precompiled computer Ran function. Systematic
deviation, repetitive sequence can be checked by running simple tests.

probability probability

configurations configurations
Uniform sampling Preferential sampling

Figure 4. Uniform and preferential sampling

The second key issue is sampling (figure 4). Uniform sampling allows a good estimate of the partition
function needed to compute all macroscopic properties, but at the expense of sampling high energy
and thus improbable configurations. Preferential, or Metropolis, sampling samples the configurations
with the largest contribution in the calculation of the partition function and of averages.
The disadvantage of the metropolis sampling is that the partition function (equivalent to the surface
under the curve) is no longer correctly evaluated. Thus, the question arise of finding how to generate
the configurations with a correct probability distribution without having to calculate the function of
partition which occurs in the definition of the probability density (equation 4)? The solution is to obey
the microscopic law of reversibility:
Given an old (o) and a new (n) configuration, their probability densities  are proportional to exp(-
E (0) /k B T) and exp(-E (n) /k B T) in the NVT ensemble (equation 4). Defining the transition probability
M(on) of going from (o) to (n), the microscopic reversibility states that at equilibrium the number of
transitions from (o) to (n) and from (n) to (o) corrected by the probability densities must be equal.

 NVT
(0)
 Mo  n   NVT
(n )
 Mn  o  microscopic reversibility [12]

8
In addition an acceptance criterion is introduced acc(on) along with (on) an "a priori" probability
of trying to go from (o) to (n) who is supposed to be symmetrical ((on) = (no)):
Mo  n  o  n  acc o  n [13]

Then by exploiting the symmetry of , equation 12 becomes:

acc o  n NVT
(n )

  E (n )  E ( 0 )
 ( o )  exp
  [14]
acc n  o  NVT  kBT 

In this equation, the partition function so difficult to calculate no longer appears.
Metropolis idea is to choose acc(on) asymmetrically. As indicated on figure 5:
 if the new configuration energy is lower than the old one, the transition is always accepted.
 if the new configuration energy is higher than the old one, one picks a random number 
between zero and one:
o if  =  1  exp(-(E (n) -E (0) )/k B T), the transition is accepted,
o if  =  1 > exp(-(E (n) -E (0) )/k B T), the transition is rejected.

exp (-E/kBT)

1
reject 2
always accept
accept
1
0 E(n)-E(0) E

Figure 5. Metropolis preferential sampling. Criterion of acceptance

Applied to equation 14, this acceptance criterion enables to define the acceptance probability of a
random displacement in general and in the NVT ensemble:
 ( n )     E  
Pacc,displaceme nt  min1, ( o )    min1, exp   [15]
 
  
in NVT
  kBT  
Last, a symmetrical (on) in chosen to allow a sampling effective in terms of acceptance and
efficient of the configuration space. Usually one defines a maximum value associated with the
transition, like a maximum displacement d max that is fixed in order to satisfy a rate of 50% of accepted
transitions. If several movements are possible (e.g. translation, rotation, volume), the type of
movement will be chosen randomly from a predetermined statistical distribution. Again, one insists on
the randomness of the choices of particle, of the type of movement… in order to respect the
microscopic reversibility.
In ensembles other than NVT, probability densities are corrected with respect of the microscopic
reversibility law.

3.2.4.3 Phase equilibrium calculations using Gibbs Ensemble Monte Carlo

The Gibbs Ensemble was imagined by Panagiotopoulos in 1987 to simulate vapor – liquid equilibrium.
Simulations are carried out in a NVT ensemble on two microscopic boxes located within two
homogeneous phases far from any interface. Each box is simulated with periodic boundary conditions.
Constant total volume V and total N particles are divided between the two phases V 1 , N 1 and V 2 , N 2 .
The temperature is set constant in the simulations and random movements are performed to satisfy
the phase equilibrium conditions as described in figure 6:
 Displacements (translation, rotation) within each phase to ensure minimal internal energy.

9
  Change of volume proportional between the phases: V 1 = -V 2 so that the total volume is
constant. This should satisfy the pressure equality.
 Transfer of particle from one box to the other to equalize the chemical potentials.

Real bulk Initial Gibbs Ensemble Monte Carlo

phases state

T
Phase I

Phase II

Internal Particle
Movement type V1 = -V2
displacements transfer

Equilibrium
Ei minimal P1 = P2 i,1 = i,2
conditions

Figure 6. Principles of phase equilibrium simulations in the Gibbs Ensemble.

The acceptance probabilities of the various movements in the case of a single component system are:
For the translation in each area:
   E  
Pacc,translation  min1, exp   [16]
 
  k BT 
For the change of volume; V 1 being increased by V and V 2 being decreased by as much:
   E 1 E 2 V  V V  V  
Pacc,volume  min1, exp   N1 ln 1  N 2 ln 2   [17]
 
  kBT kBT V1 V2 
with V chosen by generating a uniform random number  bertween 0 and 1 ; V max being the change
of maximum volume adjusted to obtain a fixed percentage (e.g. 50%) of acceptance of the move:
V    Vmax  minV1, V2  [18]
For the transfer of a particle of area 2 to area 1:
 N 2  V1   E 1 E 2  
Pacc,transfert  min1, exp    [19]
 N  1  V 
 1 2  k B T k B T 
One of the main difficulties of the Gibbs Ensemble Monte Carlo method resides in the transfer of
particles to satisfy the chemical potentials equality because of the difficulty to insert polyatomic
molecules in the dense phase. An alternative is to seek open spaces where insertion is eased the
particle. This affects randomness and introduces a statistical bias like the configurational bias method
which consists in inserting segment by segment a molecule in a phase. The probability of acceptance
of the transfer of the particle of equation 19 is then modified by introducing the energy differences E i
into weighting factors W i that represent the total energy of interaction with its surrounding of the
inserted molecule. For a L segment molecule inserted in m possible directions:
 N 2  V1  W1 
Pacc,transfert  min1,  [20]
 N1  1  V2  W 2 

10
 L  m   E1i ,j  
W1   

exp
 kBT 
 [21]
j1  i1  

More generally, the introduction of a bias consists in defining a “a priori” probability (on) which is
no longer symmetrical. Equation 14 and 15 become:

acc o  n n  o   
(n )

 [22]
acc n  o  o  n   ( o )

 (n  o)   (n ) 
Pacc,mouvement  min1,  [23]
 (o  n)   ( o ) 
 
To conclude this section, both molecular dynamics and Monte Carlo methods require the calculation of
the interaction energy; molecular dynamics to derive forces exerted on the system particles; Monte
Carlo method to calculate the acceptance criterion. The following section reviews the main features of
the force fields enabling to calculate intra and intermolecular interaction energies.

3.2.5. INTERACTION ENERGY

Suggested readings:
A.R.Leach, Molecular Modelling, Principles and Applications, Longmann, Harlow, UK, 1996.
M. Karplus and R.N. Porter, Atoms and Molecules, Benjamin, New York, USA, 1970.

3.2.5.1 Quantum chemistry models

Quantum chemistry models are never used alone in molecular simulation because of the still
prohibitive computation time. However, they must be considered as they can provide for less
sophisticated molecular mechanics models with partial electronic charges and various dipoles usefull
to compute Coulombic and dipolar interactions as well as with accurate values of the strain constants
describing the bonding intramolecular interactions associated with the various oscillatory modes within
the molecules (stretching, bending, torsion).
In quantum chemistry, only atomic nuclei surrounded by revolving electrons are considered.
Calculations provide the nuclear and electronic properties system and the true total energy of the
system. Total energy is related to the general time dependent wavefunction (r,t) by means of the
generalized equation of Schrödinger:

H  E [24]

where H is the Hamiltonian, a mathematical operator with kinetic and potential energetic contributions.
Apart from an analytical solution for the sole hydrogen atom, the Schrödinger equation solutions are
always approximative to various degree that are compromise between computation time and
accuracy. Three levels of approximation are considered, namely ab-initio methods, mean field like
DFT (density functionnal theory) methods and semi-empirical methods.
Among ab-initio methods, CI (configuration interaction) methods are the most accurate but the
slowest. Calculated energy values have a precision comparable with experimental ones (0.001 eV).
The CI solutions are obtained by minimizing a linear combination of the wavefunctions associated with
the system fundamental state and all excited states.
The Self Consist Field Molecular Orbital concept considers atomic orbitals which represent
wavefunctions of electrons moving within a potential generated by the nucleus and by an average
effective potential generated by the other electrons. The best such wavefunctions are Hartree-Fock
ones and solve the Schrödinger equation for a given electronic configuration (e.g. the fundamental

11
state) without any empirical parameter. They can be used for CI calculations. Atomic orbitals
wavefunctions are approximated using Gaussian functions, which leads to peculiar denominations like
‘STO-3G hartree-fock calculations’ (use of a basis set of 3 gaussian functions). The larger the basis
set, the longer and the more accurate the calculation.
The semi-empirical methods are the most approximated quantum methods: Hückel calculations can
be done on a sheet of paper; finer semi-empirical models enable to obtain with a good precision
ionization energy, optimal conformations, electronic surface potential. However, they present the
disadvantage to calculate approximately the wavefunctions by replacing various integrals by fitted
empirical parameters.
Between the two levels of approximation, one finds the methods of a mean field of the popular Density
Functionnal Theory. The idea is "rather than to seek to solve the exact problem Hartree-Fock in an
approximate way, one could seek to solve an approximate problem in an exact way". That consists in
modifying the Hamiltonian operator and replacing the term of exchange of correlation accounting for
multi atomic orbital interactions by the electronic density  i . The results are obtained with a satisfying
accuracy and much faster than hartree-fock calculations, enabling to even study periodic systems of
interesting size.
Nevertheless, all quantum mechanical calculations are performed for a static configuration of the
system under 0 K conditions. But as provider of key properties like electronic distribution, they should
be systematically used in any molecular simulation aiming to be quantitative.

3.2.5.2 Molecular mechanics models

Molecular simulation uses molecular mechanics models to calculate the internal energy of the system.
It consists in considering that the molecules can be represented by centers of forces like beads and
bonds by springs (figure 7). As figure 7 shows it, the total internal energy is the sum of intramolecular
or bonding interactions and of intermolecular or non-bonding interactions. The set of molecular
mechanics parameters is called a force field.

1-stretching 2-bending 3-torsion

2

3    k2 l  l
V rN  l
0 2  
k
2
  0 2  
WN
2
1  cosn   
1 bonds angles torsions

    12    6  
N N
   ij  ij   qi q j 
 
i 1
 
 ij   
ji1
4
  rij 
  
 rij   40rij 
  
4,5   

4-Van der Waals 5-Coulomb

Figure 7. Typical molecular mechanic force field

Intramolecular energy takes into account vibrational phenomena between bonded centers of forces.
As the beads and spring model suggests, they are described by harmonic functions and handle
stretching, bending, torsion as well as improper rotation if needed. Average parameters l 0 ,  0 and
harmonic constants k i are usually fitted to accurate vibrationnal energy calculations made with
quantum mechanic methods.
Intermolecular energy takes into account the 2-body interactions between the centers of forces. 3-
body interactions are rarely included. Short range interactions can be described by a Van der Waals
potential modeled by a 12-6 Lennard Jones function. The 1/r12 term represents the repulsive
contribution which becomes significant below 3 Å. The 1/r6 term represents the attractive contribution
related to the dispersive effect of induced dipoles. More rigorous forms may include 1/r8 or 1/r10 terms
or other functional forms (Buckingham potential, “exponential-6” potential....)

12
Electrostatic interactions are a major contribution to intermolecular energy as they are long range
interactions felt up to 25 Å for multicharged ions. Permanent dipole and multipole are rarely include
but coulombic interaction related to partial atomic charges q i and q j is a must be. All electronic
parameters (dipoles, partial charges) should be fitted to electronic surface potential computed by
quantum mechanics to improve quantitative predictions of molecular simulations.
Hydrogen bonding interactions are either modeled explicitely by a 12-10 Lennard Jones function or
assumed to be implicitely taken into account in the Van der Waals interaction.
The functional form of the molecular mechanics energy shows that it is not a true energy which could
be measured experimentally. Rather, for a single molecule, it is zero at its most stable conformation
whereas true zero energy corresponds to the protons, neutrons and electrons infinitely split apart.
Molecular mechanics predictions of conformations are in excellent agreement with experimental ones.
Nevertheless, the practical use of molecular mechanics is great because for a system of several
molecules, it computes the thermodynamic internal energy from which many interesting properties can
be derived.
Force fields can be of AA ' all atoms' type as in figure 7 in which there is a center of force on each
atom. Their names are Dreiding, Universal Force Field, Compass, OPLS,…. But other types exists
where atoms are grouped (e.g. -CH 3 ) under a single center of force in order to reduce the computing
time of short range interactions. This lead to UA ' United atoms' force fields.
In all cases long range electrostatic interaction is split in as many center as possible, usually on all
atoms and sometimes on virtual centers. A similar idea is at the origin of polarizable force field like the
AUA ' Anisotropic United Atoms' which intends to take into account the electronic clouds shift when
two particles approach: the charged center is displaced along the resultant of the nearby bonds.
Since all intramolecular parameters and electrostatic parameters are systematically derived from
quantum mechanical calculations, molecular simulations using molecular mechanics force fields have
greatly improved their accuracies. However, even if for a particle i, Lennard Jones parameters  i et  i
are respectively associated with the colllision diameter (the distance for which energy is null) and with
the potential well, they must still be fitted to some extend as will be shown later using experimental
data (enthalpies, formation energies, densities…).
For multi component systems, the diameter  ij and the energy parameter  ij are obtained from pure
substances ones by using traditional mixing rules like those of Lorentz-Berthelot:

i   j
 ij  and  ij   i   j [25]
2
These very simple rules have rarely been questionned, another proof of the strong physical basis of
molecular simulation. Furthermore, they highlight that the study of a system with M different centers of
forces requires only the knowledge of 2M parameters whereas a traditional approach with a
thermodynamic model with binary interaction parameters would require M(M+1)/2 such parameters.
Even if the main functional forms of the potentials (stretching, bending, torsion, Van der Waals,
Coulomb) are present in all quoted force fields, the choice must be made by knowing the type of
experimental data which were used to regress the Lennard Jones parameters and the way
electrostatic interaction are described. Mixing Lennard Jones parameters from several force field
without any confrontation to experimental data is acceptable only if qualitative results are sought.

3.2.6. RUNNNING THE SIMULATIONS

How to represent the behavior of macroscopic systems when the model system contains typically only
a few thousands of particles? The problem is solved by adopting periodic boundary conditions which
duplicate in all directions identical images of the model system. In molecular dynamics, care should be

13
taken that any particle moving through one wall of the main image will reenter at the opposite wall with
the same velocity.
Notice that the interaction energy of a particle must include interaction with all including replicated
particles. However, for long range interactions this would require too much computer effort and limiting
techniques are implemented: rough ones such as a 'cutoff' distance beyond which the interaction is
supposed to be null; or more accurate ones like the Ewald summation. In the case of a cutoff, it is
necessary to include long range corrections.
The initial particles in the box are usually set along a periodic network to avoid overlaps that would
result in an infinite energy. Then, a statistical ensemble and a sampling technique are chosen. Force
field parameters are associated to all center of forces and for molecular dynamics simulation a
statistical distribution of initial velocities is set. At last, simulation is launched. It consists of a phase of
equilibration and a phase of production. The purpose of the phase of equilibration is to bring the
system from an initial configuration to a configuration representative of the system: random distribution
of the molecules and the velocities within a system with imposed thermodynamic conditions (that of
the chosen statistical ensemble). In molecular dynamics under fixed temperature T, the system in
gradually heated to the T set value.
The phase of production starts when key properties like potential energy, pressure and density
fluctuate over mean values. Each configuration then generated is kept to calculate the macroscopic
properties from averages, of fluctuations or coefficients of correlation. As statistical error decreases
when the number of configuration increases, at least 106 configurations should be generated.

3.2.7. APPLICATIONS
Vapor – liquid equilibrium calculations are a major field of investigation because of the importance of
processes like distillation. Too often, data are missing. We present two approaches that use molecular
modeling to obtain such data. The first example aims at computing binary interaction parameters
occurring in the UNIQUAC activity coefficient model. The second example computes directly the
equilibrium compositions using a Gibbs ensemble Monte Carlo method.

3.2.7.1 example n°1 : validation of the UNIQUAC theory

3.2.7.1.1 Overview of UNIQUAC model

The practical calculation of vapor – liquid equilibrium (equation 26) involves an activity coefficient  i to
describe the non ideality of the liquid phase due to energetic interactions.

 i (T,P, x i )  fi0L (T,P)

yi   xi activity coefficient approach of phase equilibrium [26]
iV (T,P, y i )  P

By applying the thermodynamic relation of Gibbs-Duhem, one connects the individual coefficients of
activity  i with the excess gibbs energy GE:
 GE 
RT ln  i    [27]
 n 
 i  T,P,n j ,i  j

The UNIQUAC model proposes an expression for the GE with two contributions: a combinatorial part
which describes the dominant entropic contribution, a residual part which is due mainly to the
intermolecular forces which are responsible for the mixing enthalpy. The combinatorial part is related
to the composition x i and to the molecule shape and size. It requires only pure component data. The
residual part depends in addition to the interaction forces embedded into two binary interaction
parameters A ij and A ji .

14
 GE  GE   GE 
  
RT  RT   RT 
 residual
combinator ial [28]
GE
 fcombinator ial ( x i , ri , qi )  frésidual ( x i , qi ' , A ij , A ji )
RT
Parameters r i , q i et q i ’ are molecular constants for each pure component i, related respectively to its
size, its external geometrical surface and its interaction surface. q’ can be different from q, in particular
for polar molecules. The model system upon which the UNIQUAC theory was developped consider
interacting molecules. Then, the two binary interaction parameters A ij et A ji can be expressed in terms
of interaction energies U ij between dissimilar molecules i and j, and U ii between similar molecules i.

  1 Uij  U jj  N A 
A ij  exp 2 

 RT 
 
[29]
A ji  exp 2
1 
  U ji  Uii  N A 


 RT 
 
where N A is the Avogadro number.
The WILSON activity coefficient model also proposes to use two binary interaction parameters. It is a
simplification of the UNIQUAC model in which parameters r, q et q’ are all set to unity, interaction
surfaces are not taken into account and molar volumes are eliminated in the equation.
The relation between the WILSON and UNIQUAC parameters is as follows:
Vj
A ijWilson  A UNIQUAC
ji qi  T  ln [30]
Vi
where V i et V j are the components i and j molar volumes.
The traditional approach consists in regressing A ij et A ji from experimental data, with all drawbacks
associated with this approach: data specific parameters, poor extrapolation capacity, temperature and
pressure dependency, need of experimental data. A few years ago, an attempt to calculate directly the
binary interaction parameters has been made and is reported below.

3.2.7.1.2 Calculation using molecular mechanics UNIQUAC binary

interaction parameters

In 1994, Jonsdottir et al. [JON 94] computed interaction energies between isolated couples of
molecules. They use molecular mechanics models not in a molecular simulation perspective but rather
like a quantum mechanical approach: For a given orientation of the two molecules, a energy
minimization is run to reach a stable conformation. Many orientations are selected and the mean
interaction energy U ii and U ij is evaluated by weighting each value using its Boltzmann factor exp(-
U ij /k B T).
This corresponds to a rought sampling obviously not statistically representative as a few hundred
couples are investigated. This questions the validity of the first statistical thermodynamic postulate that
equals ensemble average and macroscopic time average value. Rightfully, the authors claim to
perform a molecular static approach in between quantum mechanics and molecular simulation
approaches. The ‘Consistent Force Field’ parameters are optimized for the alkanes and ketones that
are the molecules of interest but no value in particular no partial atomic charges values are provided.
But alkane conformers are taken into account, an advantage of molecular modeling approaches over
classical parameter fitting.
The U ii and U ij interaction energies are computed as the difference between the molecules couple
energy and the energy of each molecule isolated:

U ij  E i... j  E i  E j  [31]

15
Simulation results are used with the UNIQUAC equation to predict vapor – liquid equilibrium data
(equation 26) which are compared with experimental ones:
 For the alkane/alkane systems (n-butane/n-pentane ; n-hexane/cyclohexane (figure 8) and n-
pentane/n-hexane), the relative error ranges from 1.1 to 4% for the pressure and the absolute
error tandis ranges from 0,011 to 0,042 for the molar fractions.
 For the alkane/ketones (n-pentane/acétone ; acétone/cyclohexane (figure 8); cyclohexane /
cyclohexanone), the relative error ranges from 4.3 to 17.6% for the pressure and the absolute
error ranges from 0,016 to 0,042 for the molar fractions.

n-hexane / cyclohexane
22.0 1.0

20.0 0.8
P [kPa]

18.0 0.6

16.0
y1
0.4

14.0 0.2

12.0 0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
x1 x1
acetone / cyclohexane
40.0 1.0

0.8

30.0
P [kPa]

0.6
y1

0.4
20.0

0.2

10.0 0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
x1 x1

Figure 8. Bubble curve at 298.15 K for the n-hexane / cyclohexane (top figures) and
acetone/cyclohexane (bottom figures) systems. Full line: simulation. Stars: experimental data
(Reprinted from Fluid Phase Equil., Vol. 100, 1994, Jonsdottir S.O., Rasmussen K. Fredenslund A.
‘UNIQUAC parameters determined by Molecular Mechanics’. 121-138. Copyright 1994, with
permission from Elsevier).
In conclusion, the error increases along with the molecule polarity. One may question the force field
ability to handled electrostatic interaction in addition to likely insufficient sampling of the system
configurations. Finally, the authors tested the WILSON model and found errors 4 times greater for the
system N-pentane/acetone. They concluded that the UNIQUAC equation has better physical basis
that that of Wilson. That was foreseeable since the WILSON model is a simplified form of the
UNIQUAC model.

3.2.7.1.3 ab-initio calculation of UNIQUAC binary interaction

parameters

Compared to Jonsdottir’s et al. work [JON 94], Sum and Sandler [SUM 99] uses quantum mechanics
models so as to improve the representation of electrostatic interactions. As often in such calculations,
no rigorous sampling is performed, the accent being made on minimizing the system total energy.

16
For each binary system, 8 molecules are considered, 4 of each and a stable system conformation is
found by minimization using semi-empirical methods. Then the energy is minimized using an ab-initio
methods (Hartree-Fock method with an extended basis set 6-311**G(3d,2p)). Then couples of
molecules are isolated and their interaction energy is computed according to equation 31. The
average interaction energy is computed at best on 10 pairs of molecules. Then binary interaction
parameters are derived for the UNIQUAC and WILSON equation which enable to compute vapor –
liquid equilibrium data from equation 26. Systems studied are higly polar : water – methanol; water -
ethanol; water – formic acid; eau – acetic acid and water - acetone. Simulation data are then
compared with experimental data and with prediction using the activity coefficient group contribution
method UNIFAC.
The results obtained for the WILSON model are never quantitative and even qualitative wrong: as it
does not manage to reproduce the azeotropic behavior of the water - ethanol mixture (figure 9). On
the other hand, despite the poor sampling, simulations with the UNIQUAC model give good
quantitative results, comparable with experimental data and UNIFAC predicted data. Two points are
significant: no experimental data where used at any stage, no temperature or pressure conditions
where set, an advantage over regressed binary parameters. Indeed, the same set of water - acetone
parameters is used to generate accurate data over a large temperature and pressure range (figure 9).

water / ethanol water / acetone

Figure 9. vapor – liquid equilbrium at 298.15 K for the water – ethanol and water – acetone systems
(Reprinted with permission from Sum A.K. and Sandler S.I. ‘A novel approach to phase equilibria
predictions using ab-initio methods’. Ind. Chem. Eng. Res. 38, 2849-2855, 1999. Copyright 1999.
American Chemical Society).

In conclusion of this example, when no experimental data are available, vapor – liquid equilibrium data
can be predicted using UNIQUAC binary interaction parameters directly computed with molecular
modeling methods. If the sampling issue is not yet settled, quantum mechanics methods which
descirbe accurately electronic distribution have demonstrated their interest while force field
approaches did not for polar systems.
The next example shows that accurate predictions can be made with carefully set force field
approaches using efficient sampling of phase equibrium systems thanks to the Gibbs Ensemble Monte
Carlo method.

17
 3.2.7.2 Example n°2: direct prediction of nitrile vapor – liquid equilibria
As highlighted before, the development of molecular simulation as a systematic provider of accurate
physico-chemical data is impeded by the availability of accurate force fields. Force fields of the 80’s
were derived to reproduce physicochemical data of monophasic systems or devoted to
macromolecules of biological interest (amino acids, proteins…). But the simulation of multiphasic
systems was neglected until the Gibbs Ensemble Monte Carlo method and the active development of
new force fields of the type AA, AUA et UA : OPLS of Jorgensen [JOR 84], Trappe of Martin and
Siepmann [MAR 98], NERD of Nath, Escobedo and de Pablo [NAT 98], Exp6 of Errington and
Panagiotopoulos [ERR 99] and AUA of Toxvaerd [TOX 90, TOX 97]) participate to this worldly effort.
As stated before, force field development consists in deriving short range Van der Waals interaction
parameters like the  and  Lennard Jones parameters. But the challenge is to obtain generic values
which can be used for many molecules much like in a group contribution approach and for many
properties with various sampling techniques: phase equilibrium data (Gibbs Ensemble Monte Carlo,
transport coefficients (Molecular dynamics), absorption isotherm (Monte Carlo). No existing
macroscopic model could compute such a wide variety of properties using so few parameters.
In the AUA4 model, generic parameters have been derived for linear, branched and cyclic alkanes,
aromatics, hydroxyl, carboxyl and thiol groups [DEL 99, UNG 00, DEL 00, BOU 02a, BOU 02b]. For
the nitrile group -CN, we proceeded as follow [HAD 03]:
1. quantum mechanics calculations using DFT for the acetonitrile molecule, for which many
experimental data are available, to find a stable conformation, determine harmonic constants
for the intramolecular contribution of the force field potential, determine discrete partial atomic
charges from quantum electrostatic surface potential.
2. setting up the acetonitrile (CH 3 CN) force field for which CH 3 Lennard Jones parameters are
taken from the generic databank of the AUA4 force field. The same general expression than in
figure 7 is used.
3. running Gibbs Ensemble Monte Carlo simulations to identify missing ( N , N ) and ( C , C )
Lennard Jones parameters of the nitrile group. The acetonitrile molecule is fully flexible and
long range electrostatic interactions where evaluated with a cut-off and tail corrections.
Reference experimental data [FRA 75; CHA 84; KRA 85; WAR 94] are the saturated vapor
pressure ln(Psat) at 433.15 and 453,15K the vaporization enthalpy H vap and the liquid density
 liq at 273.15, 298.15, 433.15 and 453,15K. The optimization method is a simple gradient
method and the objective function used is a square mean root function with uncertainty values
set equal to 0,1for ln(Psat), 1 kJ.mol-1 for H vap and 10 kg.m-3 for  liq .
4. Once ( N , N ) and ( C , C ) values reproduce acetonitrile data accurately, their genericity is
evaluated by predicting with no further parameter adjustment vapor – liquid equilibrium data of
other linear nitriles (propionitrile, butyronitrile). In these molecules ( N , N ) et ( C , C ) are taken
equal to the values obtained for acetonitrile whereas CH 2 and CH 3 Lennard Jones parameters
are extracted from the AUA4 databank. Harmonic constant, partial charges and stable
conformation are obtained from quantum DFT calculations.
DFT electrostatic surface potential are fitted to partial atomic charges using the simple Mulliken
population analysis which splits equally the electronic distribution according to the Van der Waals
radius or using the MEP method which mimic the electrostatic potential surface with a least square
method. As shown below, the MEP analysis gives the best results but does not pass the genericity
test. Each atom bears a partial charge.
All quantum calculated conformations and dipolar moment agree with experimental data [GOL 96] and
harmonic constants agree with literature reference values [GOL 96, UNG 00].

18
Parameter regression requires for each optimization cycle (two are used) 16 Gibbs Ensemble Monte
Carlo simulations to compute the gradients varying [( N + N , N ) ; ( C , C )], [( N , N + N ) ; ( C , C )],
[( N , N ) ; ( C , C + C )], [( N , N ) ; ( C + C , C )] for each of the 4 temperatures considered.
Each simulation takes 20 hours on a Linux Pentium IV, 1.9 GHz with 512 Mo RDRAM. Equilibration
period requires 106 configurations and the production period ranges from 2.3 to 4.5.106 configurations.
Results of the optimization of the Lennard Jones parameters are:

 C /k B (K)  N /k B (K)  C (A°)  N (A°)

Charges MEP optimization 50.677 65.470 3.5043 3.3077

Charges Mulliken optimization 95.52 162.41 3.2183 3.5638

Table 1. Optimal ( N , N ) and ( C , C ) parameter values for the nitrile group. Mulliken Charges.
As indicated below (figure10), the set of MEP parameters gives the best results for the acetonitrile with
a mean standard deviation over all reference values of 1.9% and a very good estimate of critical point
With an underestimation of the vapor densities, an overestimation of liquid densities at elevated
temperature and a poor estimation of the critical point, the Mulliken set gives an error of 3.1%.
But MEP generic character is poor. Whereas Mulliken one is excellent for propionitrile and butyronitrile
vapor pressure predictions (figure 11).
Francesconi 75
Warowny 94
simulations (MEP charges)
600 estimated critical point (MEP charges)
simulations (Mulliken charges)
550 estimated critical point (Mulliken charges)

500
Temperature (K)

450

400

350

300

250
0 200 400 600 800
Density (kg.m-3)

Figure 10. Acetonitrile experimental and simulated vapor – liquid equilibrium.

A possible explanation in the poor MEP predictions lies in the charge values computed for the
propionitrile and n-butyronitrile [HAD 04]: the least square fitting of the quantum calculated
electrostatic potential surface has in that case led to unphysical values with positive nitrogen atomic
charge, in contradiction with the well known electronegativity character of this atom.Also, the MEP
nitrile  C parameter value which is too elevated compared to other  C values associated with other
carbonated chemical groups of the AUA4 force field (table 3). Correctly, the  C value with the Mulliken
distribution follows the decreasing trend as the carbonated chemical group size decreases.
So, two criteria for generic set of Lennard Jones parameters able to model the Van der Waals
interaction: (i) physically meaningful values of the Lennard-Jones and  parameters, (ii) physically
meaningful set of atomic charges representing the electrostatic potential surface of the molecule. The
importance of representation of the electrostatic potential surface has also been acknowledged in
COSMO approaches for the computation of physical properties which had recently won the first
industrial fluid properties simulation challenge [SAN, 04; CAS 04].

19
 n-butyronitrile - Chakhmuradov 84 10000
n-butyronitrile (Mulliken charges)
propionitrile - Chakhmuradov 84
propionitrile (Mulliken charges)

Pressure (kPa)
1000

100
1/380 1/570
0,00175
-1
1/Temperature [K]

Figure 11. Predicted saturated vapor pressure – propionitrile and butyronitrile. Mulliken charges.

AUA4 Chemical -C MEP -C Mulliken

-CH 3 =CH 2 ≡CH
group charges charges

 C (Å) 3.6072 3.4612 3.3625 3.5043 3.2183

Table 2. Comparison of  C parameters for various chemical groups in the AUA4 force field.
Critical points in figure 10 are obtained using Ising method [FRE 96]. Direct simulations near the
critical point are difficult to converge as shown on figure 12 for the densities versus configurations for a
simple Lennard Jones fluid: fluctuations increases and boxes interchange as the reduced temperature
nears 1. Experimental observation of a similar phenomenon is well known and demonstrate that the
molecular simulation is indeed a numerical experiment that can not only compute accurate physico-
chemical data but also behave as an efficient sensor of system behavior at the molecular scale.

Figure 12. Density versus configuration number in the vicinity of the critical point for a Lennard Jones
fluid.

3.2.8. CONCLUSIONS
Molecular modeling is an emerging discipline for the study of energetic interaction phenomena. A
molecular simulation performs numerical experiments that enable to get accurate physico-chemical
data provided sampling and energy force field issues are addressed carefully. Still computer
demanding, molecular modeling tools will likely not be used “on line” or be incorporated in process
simulators. But, rather like computer fluid dynamics tools, they should be used in parallel with existing
efficient simulation tools in order to provide information at the molecular scale on energetic interaction

20
phenomena and increase the knowledge of processes that must manufacture ever more demanding
end products.

REFERENCES

[ALL 87] M.P. Allen and D.J. Tildesley, Computer Simulation of Liquids, Oxford University
Publications, UK, 1987. [new edition in 2000].
[BOU 02a] BOURRASSEAU E., P. UNGERER, A. BOUTIN, A. FUCHS. Monte Carlo simulation of branched
alkanes and long chain n-alkanes with ansotropic united atoms intermolecular potential. Mol. Sim.,
2002, 28(4), 317.
[BOU 02b] BOURRASSEAU E., UNGERER P., BOUTIN A., Prediction of Equilibrium Properties of Cyclic
Alkanes by Monte Carlo Simulations – New Anisotropic United Atoms Intermolecular Potential – New
Transfer Bias Method. J. Phys. Chem. B, 2002, 106, 5483.
[CAS 04] CASE F., CHAKA A., FRIEND D.G., FRURIP D., GOLAB J., JOHNSON R., MOORE J., MOUNTAIN R.D.,
OLSON J., SCHILLER M. AND STORER J., The first industrial fluid properties simulation challenge, Fluid
Phase Equilibria, 2004, 217(1), 1. and http://www.cstl.nist.gov/FluidSimulationChallenge/
[CHA 84] CHAKHMURADOV C.G., GUSEINOV S.O., Iz. Vys. Uc. Zav., 65-69, 1984 (in russian)
[CHE 02] Chen C.C., P.M. Mathias, Applied Thermodynamics for Process Modelling, AIChE Journal,
2002, 48 (2), 194.
[DEL 00] DELHOMMELLE J., C. TSCHIRWITZ, P. UNGERER, G. GRANUCCI, P. MILLIE, D. PATTOU, A.H.
FUCHS, Derivation of an Optimized Potential Model for Phase Equilibria (OPPE) for Sulfides and
Thiols, J. Phys. Chem. B, 2000, 104, 4745.
[DEL 99] DELHOMMELLE J., G. GRANUCCI, V. BRENNER, P. MILLIE, A. BOUTIN, A.H. FUCHS, A new method
for deriving atomic charges and dipoles for alkanes: investigation of transferability and geometry
dependence, Mol. Phys., 1999, 97(10), 1117.
[DEP 02] DE PABLO J.J., F.A. ESCOBEDO, Perspective: Molecular Simulations in Chemical Engineering:
Present and Future, AIChE Journal, 2002, 48 (12), 2716.
[ERR 99] ERRINGTON J. R., A. Z. PANAGIOTOPOULOS, A New Potential Model for the n-Alkanes
Homologous Series, J. Phys. Chem. B, 1999, 103, 6314.
[FRA 75] FRANCESCONI A.Z., FRANCK E. U., LENTZ H., Die PVT-daten des Acetonitrils bis 450°C und
2500 bar, Ber. Bunsen-Ges. Phys. Chem. 1975, 79, 897 (in German).
[FRE 96] D. Frenkel, B. Smit , Understanding Molecular Simulation. From Algorithms to Applications,
Academic Press, San Diego, USA, 1996. [new edition in 2002].
[GOL 96] GOLDSTEIN E., BUYONG MA, LII J.H., ALLINGER N.L., Molecular Mechanics Calculations (MM3)
on Nitriles and Alkynes. J. Phys. Org. Chem., 1996, 9, 191.
[HAD 03] HADJ-KALI M.K., GERBAUD V., JOULIA X., LAGACHE M., BOUTIN A., UNGERER P.,
MIJOULE C., DUFAURE C. Prediction of liquid-vapor equilibrium by molecular simulation in the Gibbs
ensemble: application to nitriles. Comput. Aided Chem. Engin., 14, 653-658, proceedings of ICheap 6
Ed.: Pierucci S. ISBN 88-900775-2-2 (Pise, Italy, 06/08-11/03), 2003
[HAD 04] HADJ-KALI M.K., Application de la simulation moléculaire pour le calcul des équilibres
liquide vapeur des nitriles et pour la prédiction des azéotropes, Thèse de Doctorat en Génie des
Procédés, Institut National Polytechnique de Toulouse, Toulouse, France.
[HLO 00] HLOUCHA M., SUM A.K., SANDLER S.I., Computer Simulation of Acetonitrile and Methanol with
ab-initio Based Pair Potential.J. Chem. Phys., 2000, 113 (13), 5401.
[JON 94] JONSDOTTIR S.O., RASMUSSEN K. FREDENSLUND A. UNIQUAC parameters determined by
Molecular Mechanics. Fluid Phase Equilibria , 1994,100, 121.

21
[JOR 84] JORGENSEN W. L., J. D. MADURA AND C. J. SWENSON. Optimized Intermolecular Potential
Functions for Liquid Hydrocarbons. J. Am. Chem. Soc., 1984, 106, 6638.
[KAR 70] M. Karplus and R.N. Porter, Atoms and Molecules, Benjamin, New York, USA, 1970.
[KRA 85] KRATZKE H., MULLER S. Thermodynamic properties of acetonitrile 2. (P, , T) of saturated and
compressed liquid acetonitrile. J. Chem. Thermodynamics, 1985, 17, 151.
[LEA 96] A.R.Leach, Molecular Modelling, Principles and Applications, Longmann, Harlow, UK, 1996.
[MAR 98] MARTIN M.G., AND J.I. SIEPMANN. Transferable Potentials for Phase Equilibria. 1. United Atom
Description of n-Alkanes. J. Phys. Chem. B, 1998, 102, 2569.
[MCQ 76] D.A. Mc Quarry, Statistical thermodynamics, Harper and Collins Publishers, New York,
USA, 1976.
[NAT 98] NATH S.K., F.A. ESCOBEDO AND J.J. DE PABLO. On the Simulation of Vapor-Liquid Equilibrium
for Alkanes. J. Chem. Phys. 1998, 108, 9905.
[SAN 03] SANDLER S.I.. Quantum mechanics: a new tool for engineering thermodynamics, Fluid Phase
Equilibria, 2003, 210 (2), 147.
[SUM 99] SUM A.K. and SANDLER S.I. A novel approach to phase equilibria predictions using ab-initio
methods. Ind. Chem. Eng. Res. , 1999, 38, 2849.
[TOX 90] TOXVAERD S.. Equations of state of alkanes I. J. Chem. Phys. 1990, 93, 4290.
[TOX 97] TOXVAERD S.. Equations of state of alkanes II. J. Chem. Phys. 1997, 107, 5197.
[UNG 00] UNGERER P., BEAUVAIS C., DELHOMMELLE J., BOUTIN A., ROUSSEAU B. and FUCHS A.H...
Optimisation of the anisotropic united atoms intermolecular potentiel for n-alkanes. J. Chem. Phys.,
2000, 112 (12), 5499.
[WAR 94] WAROWNY W., Volumetric and phase behavior of acetonitrile et temperatures from 363K to
463K, J. Chem. Eng. Data, 1994, 39, 275.

22