Molecular Simulation Course:

Exercises

Thijs J.H. Vlugt, Maddalena Venturoli, Daan Frenkel and Berend Smit

Department of Chemical Engineering, University of Amsterdam

Nieuwe Achtergracht 166, 1018 WV Amsterdam, The Netherlands

This document describes the exercises for the Molecular Simulation Course by Berend Smit and
Daan Frenkel. This document and the accompanying programs are available on the web [1].

Version 4.0, July 27, 2000.

Contents

1 Introduction 1

2 Statistical Mechanics 3
2.1 Distribution of particles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Boltzmann distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.3 Coupled harmonic oscillators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.4 Random Walk on a 1D lattice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.5 Random Walk on a 2D lattice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

3 Basic Monte Carlo techniques 7

3.1 Calculation of  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.2 The photon gas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.3 Hard disks in a square . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.4 MC of a Lennard-Jones system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.5 Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.6 Ewald summation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.7 Parallel tempering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

4 Basic Molecular Dynamics techniques 13

4.1 MD of a Lennard-Jones system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
4.2 Parallel MD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

5 MD and MC in various ensembles 17

5.1 Barrier crossing (Part 1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
5.2 MC in the NPT ensemble . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
5.3 Ising model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

6 Phase equilibrium and free energy calculations 21

6.1 Vapor-liquid equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
6.2 Umbrella sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

7 Rare events 23
7.1 Barrier crossing (Part 2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
7.2 Transition Path Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
7.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
7.2.2 MC sampling from the distribution F (x0 ; T ) . . . . . . . . . . . . . . . . . 26
7.2.3 Model system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
7.2.4 Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
iv Contents

8 Configurational-Bias Monte Carlo 29

8.1 CBMC of a single chain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
8.2 CBMC of a simple system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
8.3 Overlapping distribution for polymers . . . . . . . . . . . . . . . . . . . . . . . . . 32

9 Appendix A: Introduction to GNU/Linux 35

10 Appendix B: Introduction to FORTRAN77 39

11 Appendix C: Introduction to MOLMOL 43

12 Appendix D: Thermodynamic equations 45

13 Appendix E: Equations of motion from the Lagrangian or Hamiltonian 47

13.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
13.2 Lagrangian . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
13.3 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
13.4 Hamiltonian . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

14 Appendix F: Questions about the book “Understanding ..” [2] 53

14.1 Introduction to FORTRAN77 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
14.2 Chapter 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
14.3 Chapter 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
14.4 Chapter 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
14.5 Chapter 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
14.6 Chapter 6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
14.7 Chapter 7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
14.8 Chapter 8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
14.9 Chapter 11 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
14.10Chapter 13 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

15 Appendix G: Possible research projects 63

15.1 Adsorption in porous media . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
15.2 Transport properties in liquids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
15.3 Parallel Molecular Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
15.4 Phase diagram of ethane . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
15.5 Diffusion in a porous media . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
15.6 Multiple time step integrators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
15.7 Thermodynamic integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
15.8 Hints for programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

16 Appendix H: Software 71

Acknowledgements 73

Bibliography 75
Chapter 1

Introduction

This documents describes the exercises for the Molecular Simulation Course by Daan Frenkel
and Berend Smit. To do the exercises, it is essential to have a copy of the book “Understanding
Molecular Simulation” by Daan Frenkel and Berend Smit [2]. Some of the exercises in this doc-
ument were taken from several standard textbooks [2–6].

In most of the exercises, some programming has to be done. One does not have to write a
program from scratch, but rather one has to make some small modifications to an existing code
 . We have programmed the code in FORTRAN77 because it is quite easy to learn. However,

the random number generator that is used in some of the programs is written in C [7]. See,
for example, ref. [8] for more information about random number generators. All programs are
provided with a makefile that has been prepared for GNU/Linux using the g77/gcc compil-
ers [9, 10]. Compiling the programs on other Unix systems is straightforward. Solutions of the
programming exercises as well as the LATEX source of this document are available on request.

The course and exercises are divided into several parts:

 Introduction to Unix and FORTRAN77 (0.5 day)

 Basic Statistical Mechanics (1.5 day)

 Basic Monte Carlo (MC) techniques (2 days)

 Basic Molecular Dynamics (MD) techniques (2 days)

 MD and MC in various ensembles (1 day)

 Phase equilibrium and free energy calculations (1 day)

 Rare events (1 day)

 Configurational-Bias Monte Carlo (1 day)

We have setup the course in such a way that lectures are given every morning (3 hours),
while the exercises have to be done in the afternoon (5 hours); we do not expect that all students
have enough time to complete all exercises. After all exercises have been completed, a special
research project has to be done (duration: 2 weeks) to pass the exam. In Appendix G, we have
given some suggestions for research projects. We have also included a list of questions about
 Although the text may suggest otherwise, all programs do not have to be written from scratch. For more infor-
mation, see the README file from the distribution.
2 Introduction

the book of Daan Frenkel and Berend Smit [2] (Appendix F).

We have done our best to remove all errors from this document and the programs. However,
we have to make the following statement:

“We make no warranties that this document or the programs accompanied by this document
are free of error, or that they will meet your requirements for any particular application. The
authors disclaim all liability for direct and consequential damages resulting from the use of this
document or the programs accompanied by this document. It is not allowed to distribute this
document and the accompanying programs.”

We are interested in what you think about this course. Any comment would be appreciated,
even severe criticism. Comment can be send to Berend Smit.

Contact address:

Berend Smit
Department of Chemical Engineering
University of Amsterdam
Nieuwe Achtergracht 166
1018 WV Amsterdam, The Netherlands
email: [email protected]
web: http://molsim.chem.uva.nl
Chapter 2

Statistical Mechanics

2.1 Distribution of particles

Consider an ideal gas of N particles in a constant volume at constant energy. Let us divide the
volume in p identical compartments. Every compartment contains ni molecules such that

N=
X
i=p
ni (2.1)
i=1

An interesting quantity is the distribution of molecules over the p compartments. Because the
energy is constant, every possible eigenstate of the system will be equally likely. This means
that in principle it is possible that one of the compartments is empty.

1. Enclosed is a program that calculates the distributions of molecules along the p compart-
ments. Run the program for different numbers of compartments (p) and total number of
gas molecules (N). Note that the code has to be completed first (see the file distribution.f).
The output of the program is the probability to find x particles in a particular compartment
as a function of x. This is printed in the file output.dat, which can be plotted using xmgr
-nxy output.dat (the option -nxy means that more than one column is plotted).

2. Why does it (almost) never happen that one of the compartments is empty ?

3. Consider the case of p = 2. The probability of finding n 1 molecules in compartment 1 and

n2 = N - n1 molecules in compartment 2 is given by
N!
P (n1 ) = (2.2)
n1 !  (N - n1 ) !  2N
Compare your numerical results results with the analytical solution for different values of
N. Show that this distribution is a Gaussian for small n1 . Hint: For x > 10, it might be
useful to use Stirling’s approximation:

x!  (2) 2 xx+ 2 exp [-x℄

1 1
(2.3)

2.2 Boltzmann distribution

Consider a system of N energy levels with energies 0; ; 2;


; (N - 1)   and  > 0.
4 Statistical Mechanics

Questions:

1. Calculate, using the given program, the occupancy of each level for different values of the
temperature. What happens at high temperatures ?

2. Change the program in such a way that the degeneracy of energy level i equals i + 1. What
do you see ?

3. Modify the program in such a way that the occupation of the energy levels as well as the
partition function (q) is calculated for a linear rotor with moment of inertia I. Compare
your result with the approximate result
2I
q= (2.4)
h̄2
for different temperatures. Note that the energy levels of a linear rotor are

h̄2
U = J (J + 1) (2.5)
2I
with J = 0; 1; 2;


; 1. The degeneracy of level J equals 2J + 1.

2.3 Coupled harmonic oscillators

Consider a system of N harmonic oscillators with a total energy U. A single harmonic oscillator
1
has energy levels 0; ; 2;


; ( > 0). All harmonic oscillators in the system can exchange
energy.

Questions:

1. Invent a computational scheme for the update of the system at constant total energy (U).
Compare your scheme with the scheme that is incorporated into the given computer code
(see the file harmonic.f).

2. Make a plot of the energy distribution (output.dat) of the first oscillator as a function of the
number of oscillators for a constant value of U=N. Which distribution is recovered when
N becomes large ? What is the function of the other N - 1 harmonic oscillators ? Explain.
3. Compare this distribution which the canonical distribution of a single oscillator at the
same average energy.

4. How does this exercise relate to the derivation of the Boltzmann distribution for a system
at temperature T on page 12 of ref. [2] ?

2.4 Random Walk on a 1D lattice

Consider the random walk of a single particle on a line. The probability that after N jumps the
net distance of the particle equals n, is:
 
1 2 n2
ln (P (n; N))  ln - (2.6)
2 N 2N
2.5 Random Walk on a 2D lattice 5

Questions:

1. Derive this equation. The probability to jump to one direction equals the probability to
jump to the other direction. For large x, x! can be approximated by:

x!  (2) 2 xx+ 2 exp [-x℄

1 1
(2.7)

2. Compare this theoretical result with the computed root mean square displacement and the
computed function P (n; N) (see the file output.dat). What is the diffusivity of this system ?

3. Modify the program in such a way that the probability to jump in one direction equals 0:8.
What happens ?

2.5 Random Walk on a 2D lattice

Consider the random walk of N particles on a M  M lattice. Two particles cannot occupy the
same lattice site. On this lattice, periodic boundaries are used. This means that when a particle
leaves the lattices it returns on the opposite side of the lattice.

Questions:

1. What is the fraction occupied sites () of the lattice as a function of M and N ?

2. Make a plot of the diffusivity D as a function of  for M = 32. For low values of , the
diffusivity can be approximated by

D  D0 (1 - ) (2.8)

Derive this equation. Why is this equation not exact ?

3. Modify the program in such a way that the probability to jump in one direction is larger
than the probability to jump in the other direction. Explain the results.

4. Modify the program in such a way that periodic boundary conditions are used in one
direction only. What happens ?

5. Modify the program in such a way that a certain fraction of the particles are “frozen”.
Investigate the influence of the fraction of frozen particles on the diffusivity.
Chapter 3

Basic Monte Carlo techniques

3.1 Calculation of 
Consider a circle of diameter d surrounded by a square of length l (l  d). Random coordinates
within the square are generated. The value of  can be calculated from the fraction of points
that fall within the circle.

Questions:

1. How can  be calculated from the fraction of points that fall in the circle ? Remark: the
“exact” value of  can be computed numerically using  = 4  arctan (1).

2. Complete the small Monte Carlo program to calculate  using this method.

3. How does the accuracy of the result depend on the ratio l=d and the number of generated
coordinates ? Derive a formula to calculate the relative standard deviation of the estimate
of .

4. Is it a good idea to calculate many decimals of  using this method ?

3.2 The photon gas

The average occupancy number of state j of the photon gas (hn j i) can be calculated analytically;
see equation 14.13. However, it is also possible to compute this quantity using a Monte Carlo
scheme. In this exercise, we will use the following procedure to calculate hn j i:

1. Start with an arbitrary nj

2. Decide at random to perform a trial move to increase or decrease n j by 1.

3. Accept the trial move with probability

acc (o !n)= min (1; exp [- ( U (n) - U (o))℄) (3.1)

Of course, nj cannot become negative !

8 Basic Monte Carlo techniques

Questions:

1. How can this scheme obey detailed balance when nj = 0? After all, n
j can not become
negative !

2. Is the algorithm still correct when trial moves are performed that change n j with a random
integer from the interval [-5; 5℄ ? What happens when only trial moves are performed that
change nj with either -3 or +3 ?

3. Assume that N = 1 and j = . Write a small Monte Carlo program to calculate hnj i as a
function of . Compare your result with the analytical solution.

4. Modify the program in such a way that the averages are updated only after an accepted
trial move. Why does this lead to erroneous results ? At which values of does this error
become more pronounced ?

5. Modify the program in such a way that the distribution of n j is calculated as well. Com-
pare this distribution with the analytical distribution.

3.3 Hard disks in a square

Consider a system of N identical hard disks with diameter 1 in a square of size 10. No periodic
boundary conditions are applied. The energy of this system (U) is either U = 0 (no overlaps) or
1
U = (at least one overlap). We would like to calculate the radial distribution function of the
disks (g (r) ; r < 5) in the canonical ensemble. There are 2 methods of calculating this:

1. Dynamic scheme.
The disks are initially placed on a lattice and trial moves are performed to translate the
disks. A trial move is accepted when there are no overlaps and rejected when there are
one or more overlaps.

2. Static Scheme.
In every cycle, all disks are put at a completely random position in the system. The radial
distribution function of this configuration is calculated and multiplied by the Boltzmann
factor of the system (0 when there is at least one overlap and 1 when there are no overlaps).

Questions:

1. Write a program to calculate g (r) using both methods (see sample.f). Make sure that both
methods give the same result when N is small. Use the MOLMOL program to visualize
the Monte Carlo simulation.

2. At which N does the second method starts to fail ? Why ?

3. For which method will there be a larger correlation between successive elements of the
Markov chain ? Explain.
3.4 MC of a Lennard-Jones system 9

3.4 MC of a Lennard-Jones system

In this exercise, we will study a 3D Lennard-Jones system. See also Case Study 1 from ref. [2].

Questions:

1. In the present code, the pressure of the system is not calculated. Modify the code in such
a way that the average pressure can be calculated. You will only have to make some
modifications in the subroutine ener.f.

2. Perform a simulation at T = 2:0 and various densities. Up to which density does the ideal
gas law

p= (3.2)

hold ?

3. The program produces a sequence of snapshots of the state of the system. Try to visualize
these snapshots using the program MOLMOL.

4. For the heat capacity at constant volume one can derive

U2 - hUi2
Cv = (3.3)
kB T 2
in which U is the total energy of the system. Derive a formula for the dimensionless heat
capacity. Modify the program (only in mc nvt.f) in such a way that Cv is calculated.

5. Instead of performing a trial move in which only one particle is displaced, one can do a
trial move in which all particles are displaced. Compare the maximum displacements of
these moves when 50% of all displacements are accepted.

6. Instead of using a uniformly distributed displacement, one can also use a Gaussian dis-
placement. Does this increase the efficiency of the simulation ?

3.5 Scaling
Consider a system in which the energy is a function of one variable (x) only:

exp [- U (x)℄ =  (x)  (1 - x) (3.4)

in which  (x) is the Heaviside step function:  (x < 0) = 0 and  (x > 0) = 1. We would
like to calculate the distribution of x in the canonical ensemble. We will consider two possible
algorithms (we will use Æ > 0):

1. Generate a random change in x between [-Æ; Æ℄. Accept or reject the new x according to its
energy.

2. Generate a random number  between [1; 1 + Æ℄. With a probability of 0:5, decide to invert
 or not. The new value of x is obtained by multiplying x with .
10 Basic Monte Carlo techniques

Questions:

1. Derive the correct acceptance/rejection rules for both schemes.

2. Complete the computer code to calculate the probability density of x. The program writes
this distribution to distri.dat.

3. What happens when the acceptance rule of method 1 is used in the algorithm of method 2
? Why ?

3.6 Ewald summation

An example of long-range interactions is the coulombic potential between point charges:
qi qj
U (rij ) = (3.5)
40 rij
Special techniques like the Ewald summation are required to compute the total potential energy
correctly. When the Ewald summation is used, the total potential energy (U t ) consists of three
contributions: the fourier space part (U f ), the real space part (Ur ), and the self energy (U s ):

Ut = Uf + Ur + Us
1 X 1 exp -k2=4 2 iX=N 2
Uf = exp [-ik
ri ℄
2V0 k=6 0 k2 i=1
1 X qi qj erf ( rij )
Ur =
40 i<j rij
=N q2
1 iX i
Us = - (3.6)
40 i=1 
p

in which N is the number of ions and V is the volume of the (rectangular) unit-cell. The positions
are indicated by ri . The complementary error function (erf (x) =
p2  x1 dt exp -t2 ) falls to zero with increasing x. Details of the Ewald summation can be
of the Rions in the 
unit-cell


found in ref. [2].

Consider a simple cubic lattice (for example, NaCl) in which the kations are located at
[0; 0; 0℄ , [1; 1; 0℄, [1; 0; 1℄, and [0; 1; 1℄ and the anions are located at [0; 0; 1℄, [1; 1; 1℄ , [1; 0; 0℄, and
[0; 1; 0℄ . The total energy of an infinitely large NaCl crystal equals:

nq2 M
U=- (3.7)
40 r0
in which r0 is the closest distance between ion pairs, n is the number of ion pairs and M is the
Madelung constant.

Questions:

1. What is the relation between Ut and the coulombic virial ?

2. How many ion pairs are present in a unit-cell of NaCl ?

3.7 Parallel tempering 11

3. Compute the Madelung constant of NaCl using the given program. Make a plot of the
total energy as a function of k for different values of . What is a good choice for and
k ? You will have to program the real space part of the Ewald summation yourself (in
realspace.f). Note: We will use r0 = 1, 40 = 1 and qNa+ = -qCl- = 1.

4. Rewrite the program in such a way that each ion is given a random displacement from
the interval [-0:3; 0:3℄ (this may correspond to a liquid structure). Make a plot of the total
energy as a function of k for different values of .

5. Consider a system of dipolar molecules AB (qA = -qB 6= 0). What will happen with
the radial distribution function when a truncated and shifted coulombic potential is used
instead of the Ewald summation ?

3.7 Parallel tempering

The method of parallel tempering [11–14] is a Monte Carlo scheme that has been derived to
achieve good sampling of systems that have a free energy landscape with many local minima.
In parallel tempering we consider N systems. In each of these systems we perform a simula-
tion in the canonical ensemble, but each system is in a different thermodynamic state. Usually,
but not necessarily, these states differ in temperature. In what follows we assume that this is the
case. Systems with a sufficiently high temperature pass all barriers in the system. The low tem-
perature systems, on the other hand, mainly probe the local energy minima. The idea of parallel
tempering is to include MC trial moves that attempt to “swap” systems that belong to different
thermodynamic states, e.g., to swap a high temperature system with a low temperature system.
If the temperature difference between the two systems is very large, such a swap has a very low
probability of being accepted. This is very similar to particle displacement in ordinary Monte
Carlo. If one uses a very large maximum displacement a move has a very low probability of be-
ing accepted. The solution to this problem is to use many small steps. In parallel tempering we
use intermediate temperatures in a similar way. Instead of making attempts to swap between
a low and a high temperature, we swap between ensembles with a small temperature difference.

The total partition function of a system with N canonical subsystems (Q) equals

Q=
YQ
i=N
(3.8)
i
i=1

in which Qi is the canonical partition function of the individual system i

Qi =
X exp [- i U (xi )℄ (3.9)
xi

in which i = 1= (kB Ti). For each of these systems, individual trial moves are performed. After
a randomly selected number of trial moves, an attempt is made to exchange configurations.
Two systems (i and j, ji - jj = 1) are selected at random, the systems are exchanged by choosing
xi (n) = xj (o) and xj (n) = xi (o). The ratio of acceptance probabilities equals
acc (o !n )
 (U (xi (o)) - U (xj (o)))℄
acc (n !o )
= exp [( i- j) (3.10)

Such trial moves will be accepted when there is enough overlap between the energies of systems
i and j. To demonstrate this technique, consider a two dimensional system of 9 particles that are
12 Basic Monte Carlo techniques

confined in a square ( = 2:5). Each particle pair interact with a soft repulsive potential:

 (r - 1)2 r  1
U (r) = (3.11)
0 r>1
We will use  = 1. At low temperatures, the particles are not able to pass each other and there-
fore a particle is confined at its original position. However, in principle the distribution of the
position of a particle should be symmetrical.

Questions:

1. Derive equation 3.10.

2. Compute the distribution of the position of the first particle as well as the distribution of
the total energy for T = 0:001 when no exchange moves are allowed (N = 1). At which
temperature are the particles allowed to diffuse ?

3. Complete the code for the exchange moves and find out how many systems are needed
to ensure that for T = 0:001 the particle is allowed to diffuse. You will have to check the
distribution of the total energy to ensure that there is enough overlap between the systems.

4. At high temperature, confirm that the equilibrium distributions with and without ex-
change moves are identical.

Book PN 78
Chapter 4

Basic Molecular Dynamics techniques

4.1 MD of a Lennard-Jones system

Enclosed is a Molecular Dynamics (MD) program for a Lennard-Jones fluid in the NVE ensem-
ble. Unfortunately, the program does not conserve the total energy because it contains three
errors.

Questions:

1. Find the three errors in the code. The person that spots the errors first will receive a bottle
of French wine from Berend Smit (people from Berend’s group are excluded from this
contest). Hint: there are two errors in integrate.f and one in force.f. See the file system.inc for
documentation about some of the variables used in this code.

2. How is one able to control the temperature in this program ? After all, the total energy of
the system should be constant (not the temperature).

3. To test the energy drift
U of the numerical integration algorithm for a given time step
t
after N integration steps, one usually computes [15]

U (
t) = N1
X
i=N
U (0) - U (i
t) (4.1)
i=1
U (0)
In this equation, U (x) is the total energy (kinetic+potential) of the system at time x. Change
the program (only in mdloop.f) in such a way that
U is computed and make a plot of
U
as a function of the time step. How does the time step for a given energy drift change with
the temperature and density ?

4. One of the most time consuming parts of the program is the calculation of the nearest
image of two particles. In the present program, this calculation is performed using an if-
then-else-endif construction. This works only when the distance between two particles is
smaller than 1:5 and larger than -1:5 times the size of the periodic box. A way to overcome
this problem is to use a function that calculates the nearest integer nint

x = x - box  nint (x  ibox) (4.2)

in which ibox = 1:0=box. Which expression is faster ? (Hint: You only have to make some
modifications in force.f) Which expression will be faster on a vector computer like a Cray
14 Basic Molecular Dynamics techniques

C90 ? Because the nint function is usually slow, you can write your own nint function.
For example, when x < -998, we can use

nint (x) = int (x + 999:5) - 999 (4.3)

What happens with the speed of the program when you replace the standard nint function
? Do you have an explanation for this  ?

5. In equation 4.2, ibox is used instead of 1=box. Why ?

6. An important quantity of a liquid or gas is the so called self diffusivity D. There are two
methods to calculate D:

(a) by integrating the velocity autocorrelation function:

D =
1 Z 1 Dv (t) v

t + t0
E
dt 0
3 0

R1 P = D
i N
v (i; t)
v

i; t + t 0
E
dt 0
0 i=1
= (4.4)
3N
in which N is the number of particles and v (i; t) is the velocity of particle i at time
t. One should choose t in such a way that independent time origins are taken, i.e.
t = ia
t, i = 1; 2; ; 1 and v (t) v (t + a
t) 0 (why ?).



h
i 

(b) by calculating the mean square displacement:

   
2
x t + t 0 - x ( t)
D= lim
6t 0 (4.5)
t0 !1
One should be very careful with calculation of the mean square displacement when
particles are always transformed to the central box (why ?).

Modify the program in such a way that the self diffusivity can be calculated using both
methods. Only modifications in subroutine sample diff.f are needed. Why is it important
to use only independent time origins for the calculation of the means square displacement
and the velocity autocorrelation function ? What is the unit of D in SI units ? How can one
transform D into dimensionless units ?

7. For Lennard-Jones liquids, Naghizadeh and Rice report the following equation for the self
diffusivity (dimensionless units, T  < 1:0 and p < 3:0) [16]

10
log (D ) = 0:05 + 0:07p -
1:04 + 0:1p (4.6)
T
Try to confirm this equation with simulations. How can one translate D  to a diffusivity
in SI units ?

8. Instead of calculating the average energy hUi directly, one can use the radial distribution
function g (r). Derive an expression for hUi using g (r). Compare this calculation with a
direct calculation of the average energy. A similar method can be used to compute the
average pressure.
 The result will strongly depend on the computer/compiler that is used.
4.2 Parallel MD 15

9. In the current version of the code, the equation of motion are integrated by the Verlet
algorithm. Make a plot of the energy drift
U for the following integration algorithms [2]:

 Verlet
 Velocity Verlet [17]
 Euler (never use this one except here !!!)

4.2 Parallel MD
Please find enclosed a parallel version of the code of the previous exercise. This code is written
in FORTRAN77 using an MPI library. We will use LAM [18] for this. To run the program in
parallel, several things have to be done:

1. Copy the file mpif.h from $fLAMHOMEg/h. This file is strongly dependent on the version
of MPI.

2. Compile the code using LAM. The commands hf77 and hcc will use the GNU compilers
g77 and gcc to compile the source.

3. Find out on how many computers you would like to run the code. The user must be able
to remotely execute on the machine with rsh. Remote host permission must be provided in
either /etc/hosts.equiv or the remote user’s /.rhosts file. The remote user’s shell must have
a search path that will locate LAM executables and the remote shell’s startup file must not
print anything to standard error when invoked non-interactively.

4. Type recon -v hostnames in which the file hostnames should contain the hostnames of all
machines that will be used.

5. When no errors occur, you can start the LAM daemon by typing lamboot -v hostnames.
When this is successful too, you can start the run-script. Beware that the command mpirun
uses an absolute path !

6. After the execution is completed, the LAM daemon can be killed by typing wipe -v host-
names.

Questions:

1. Try to find out which part of the calculation is performed in parallel.

2. In the file force.f, there is a line:

do I=(Iirank+1),(Npart-1),Iisize
In principle, the program also works correct when this line is replaced by:
do I=((Iirank*(Npart-1)/Iisize) + 1),((Iirank+1)*(Npart-1)/Iisize)
What is the difference between these two methods and how is this reflected in the scaling
of the code ? Hint: Npart is the number of particles, Iisize is the number of processors and
Iirank is the index of a particular processor (Iirank = 0; 1; 2;


; Iisize - 1).

3. Run the code on different numbers of nodes and record the execution time as a function
of the number of nodes. What happens when using a large number of nodes ?
16 Basic Molecular Dynamics techniques

4. Perform this calculation again for a much larger system at the same density. What do you
see ?

5. Can you come up with a better strategy to parallelize this code ?

6. Can the same strategy be used for a MC simulation of the same system in the canonical
ensemble ?
Chapter 5

MD and MC in various ensembles

5.1 Barrier crossing (Part 1)

Consider the movement of a single particle that moves on a 1D potential energy surface with
the following functional form:
8 Bx2
< x<0
U (x) =  (1 - cos (2x)) 0  x  1
: B (x - 1)2 x>1
(5.1)

The energy, force and the derivative of the force are continuous functions of the position x and
 > 0.

Questions:

1. Derive an expression for B. Make a sketch of the energy landscape.

2. A program is provided that integrates the equation of motion of the particle starting at
x (t = 0) = 0 using several methods:
(a) No thermostat (NVE ensemble). What do you expect the phase space trajectories to
look like ?
(b) Andersen thermostat. In this method, the velocity of the particle is coupled to a
stochastic heat bath which leads to a canonical distribution.
(c) A Nosé-Hoover chain [15, 19]. In this method, the motion of the particle is coupled
to a chain of thermostats. The equations of motion are integrated using an explicit
time-reversible algorithm that might look a little-bit complicated at first sight [15],
see integrate res.f. One can prove that this method yields a canonical distribution
provided that the system is ergodic.
(d) No molecular dynamics, but a simple Monte Carlo scheme.

The Andersen thermostat and the NVE integration algorithm are not implemented yet, so
you will have to do this yourself (see integrate nve.f and integrate and.f). Try to use all
methods for a low temperature, T = 0:05, for which the system behaves like a harmonic
oscillator. Pay special attention to the following:

(a) Why does the phase space distribution of the MC scheme look so much different at
low temperatures ?
18 MD and MC in various ensembles

(b) Why does the phase space distribution of the NVE scheme look like a circle ?
(c) Compare the phase space distributions of the Nosé-Hoover chain method with distri-
bution generated by the Andersen thermostat. How long has the Nosé-Hoover chain
to be to obtain a canonical distribution ?

3. Investigate at which temperature the particle is able to cross the energy barrier.

4. Another widely used algorithm is the “temperature coupling” of Berendsen [20]. It is

important to note that this method does not produce a canonical ensemble and therefore
we should never use it. In this algorithm, the temperature of the system is controlled by
scaling the velocities every time step with a factor 

t  T  21
 = 1 +  T0 - 1 (5.2)
T

in which T0 is the desired temperature, T is the actual temperature,
t is the time step of
the integration algorithm and T is a constant. The temperature coupling algorithm can
be used in combination with a Leap-Frog algorithm
     
t  F (t) 
v t +
t
2 =  t -
t 2  v t - 2 + m
t
x (t +
t) = x (t) + v t +
t
2
t (5.3)

Compare the distributions of the Berendsen temperature bath with the canonical distribu-
tions.

5. Modify the program in such a way that the potential energy function

U =  (1 - cos (2x)) (5.4)

is used. Calculate the diffusion coefficient as a function of the temperature. Why is it

impossible to calculate the diffusivity at low temperatures using ordinary molecular dy-
namics ? Why is the diffusion coefficient obtained by using the Andersen thermostat a
function of the collision frequency ?

5.2 MC in the NPT ensemble

Enclosed is a program to simulate hard spheres (diameter 1) in the NPT ensemble using MC.

Questions:

1. Why is it impossible to calculate the virial for this system directly ?

2. In the current code, a random walk is performed in ln (V ) instead of V . Change the code
in such a way that a random walk in V is performed. Check that the average densities
calculated by both algorithms are equal.

3. Make a plot of the acceptance ratio for volume displacements as a function of the maxi-
mum volume displacement for both algorithms.
5.3 Ising model 19

5.3 Ising model

In this exercise we consider a 2D Ising model. In this model, N spins s (1) are arranged on a
lattice. Every spin (i) has 4 neighbors (j = 1; 2; 3; 4). The total energy of the system equals

U=-
XX
 i=N j=4
ss
2 i=1 j=1 i j
(5.5)

in which si = 1 and  > 0. The second summation is a summation over all spin pairs of spin
i. The total magnetization M equals the sum over all spins:

M=
Xs
i=N
i (5.6)
i=1
The 2D Ising model has a critical point close to  0:44.
Questions:

1. Complete the given simulation code for this system (see ising.f).

2. Calculate the distribution of M for N = 32  32 and = 0:5 in the canonical ensemble. In

principle, this distribution should be symmetrical:

p (M) = p (-M) (5.7)

Why does this not seem to be the case ?

3. Instead of a simulation in the canonical ensemble, one can perform the simulation in the
ensemble :

 / exp [- U + W (M)℄ (5.8)

The average value of an observable O in the canonical ensemble equals:

O exp [-W (M)℄i

Oi =
h
(5.9)
hexp [-W (M)℄i
h

in which h


i is an ensemble average in the ensemble . Derive this equation.

4. Perform simulations with some given distributions W (M) (w.type1.dat and w.type2.dat).
Explain your results. How should one choose the function W (M) to obtain the optimal
efficiency ?

5. What happens when W (M) is a Gaussian,

"  2#
M
W (M) = A exp - (5.10)

with A > 0 ?

6. What happens when W (M) = W (U) = - U?

Chapter 6

Phase equilibrium and free energy

calculations

6.1 Vapor-liquid equilibrium

In this exercise, we will use Widom’s test particle method to locate a vapor-liquid equilibrium.
This is compared with a Gibbs-ensemble simulation.

Questions:

1. Modify the Monte Carlo program of Lennard-Jones particles in the NVT ensemble (only in
the file mc nvt.f) in such a way that the chemical potential can be calculated using Widom’s
test particle method:


ln -1 hexp [-
U+ ℄i
 = 0 - (6.1)

in which  is the number of particles per volume, U+ is the energy of a test particle and


0 =
- ln 3 (6.2)

 Make a plot of the chemical potential and pressure as a function of the density for
T = 0:8.
 Why is it more difficult to calculate the chemical potential at high densities than at
low densities ?
 How can you locate the vapor-liquid coexistence densities ?

2. Perform a Gibbs-ensemble simulation of the system at T = 0:8. In the Gibbs ensemble, the
chemical potential of box i is equal to [2, 21]
D  E
ln Vi
ni + 1 exp -
U+i
i = 0 - (6.3)

in which ni is the number of particles in box i and Vi is the volume of box i. Do the
vapor/liquid density and chemical potential agree with your previous results ?
22 Phase equilibrium and free energy calculations

6.2 Umbrella sampling

Consider a single particle on a one dimensional energy landscape. The energy as a function of
the position is given by

U (x) = x2 (6.4)

One would like to calculate the probability distribution of finding the particle at a position x
(p (x)) by using a Monte-Carlo scheme (we will use =  = 1).

Questions:

1. Why is this distribution difficult to calculate for large values of x when a conventional MC
scheme is used ?

2. Alternatively, one can divide the x axis in overlapping slices and calculate the distribution
p (x; i) in all slices i. This scheme is often referred to as umbrella sampling. Show that
p (x; i) / p (x; j) when i 6= j (This also means that ln (P (x; i)) = ln (P (x; j)) + C in which C
is a constant).

3. Compare the results from the conventional MC algorithm with the results from the um-
brella sampling simulations, especially at large values of x. Different slices can be com-
bined by using xmgr.

BOOK PN 181
Chapter 7

Rare events

7.1 Barrier crossing (Part 2)

In one of the previous exercises, we have seen that sometimes there may be energy barriers is a
system that are so high, that the crossing rate can not be calculated using conventional molecular
dynamics techniques. A method that can be used to calculate the transition rate is the Chandler-
Bennett approach [2, 4, 22, 23]. We will use this approach to calculate the crossing rate of a single
particle over an energy barrier
8> 1
>< 0 x < -5
-5  x < 0
U (x) =
>> 0 1
( - cos (2x)) 0x1 (7.1)
:1 1<x5
x>5

We will call the region left from the barrier the reactant side A and the region right from the
energy barrier the product side B. The time dependent rate constant k A!B (t) can be written
as [2]

q̇ (0) Æ (q? - q (0))  (q (t) - q? )i Æ (q? - q)i

kA!B (t) =
h h
?

?
(7.2)
hÆ (q - q (0))i h (q - q)i

We are mainly interested at the plateau value of kA!B (t), which is the hopping rate. At first
sight, this equation might look pretty horrible. First of all, we need to define our symbols:

 q is the reaction coordinate. In this case, q = x is a sensible choice. q (0) is the reaction
coordinate at t = 0.

 q? is the position of the dividing surface. We will choose 0 < q ? < 1. Note that the
dividing surface does not always have to be on top of the energy barrier !

denotes an ensemble average

R1
 h


i

 Æ (x) is Dirac’s delta function: Æ (x = 0) = 1, Æ x ( 6= 0) = 0, -1 Æ (x) dx = 1 and

Æ (s - s0 )
Æ (h (s)) = (7.3)
jh (s)j
0

when h (s0 ) = 0.
24 Rare events

  (x) is the Heaviside step function:  (x < 0) = 0 and  (x > 0) = 1. Note that the defini-
tion of  in ref. [2] on page 251 is incorrect.

Now it is a lot easier to understand what the terms in equation 7.2 mean:

 The first term on the r.h.s. is the conditional average of the product [q̇ (0)  (q (t) - q ? )℄
given that the initial position is at the top of the barrier (q (0) = q ? ). The term  (q (t) - q? )
is equal to 1 when the particle is in state B at time t and equal to 0 otherwise. q̇ (0) is the
initial velocity of the particle at the top of the barrier.

 The second term on the r.h.s is the probability of finding the system on top of the barrier
divided by the probability that the system is on the reactant side of the barrier.

Questions:

1. What is the unit of kA!B (t) in SI units and in dimensionless units ?

2. Is the force on a particle a continuous function of x ?

3. A program is provided to perform a Molecular Dynamics simulation for a single particle

provided that q (0) = q? . The sampling of the crossing rate has still to be programmed
(see mdloop.f and sample.f).

4. Calculate the crossing rate for different temperatures and positions of q ? . Make a plot of
ln (kA!B ) versus T -1.

5. Check that the crossing rate is independent of the location of the dividing surface.

7.2 Transition Path Sampling

In this section, we will study the Transition Path Sampling method for the calculation of rate
constants of rare events. This method is able to compute rate constants without having to as-
sume a transition state. For more details about this method, the reader is referred to ref. [24, 25],
on which this exercise is based.

7.2.1 Introduction
Consider a dynamical system with two stable states, A and B, in which transitions from A to
B are rare. The transition rate, k, from A to B can be calculated from the time derivative of an
autocorrelation function C (t),

dC (t)
k = tmol < t  trxn (7.4)
dt
hhA (x0 ) hB (xt )i
C (t) = ; (7.5)
hhA (x0 )i

provided that the reaction time t rxn of the system [A; B℄ is much larger than the molecular relax-
ation time tmol of the system in region A or B. In equation 7.5, x t represents the momenta p and
positions q of the system at time t. We will only consider deterministic trajectories for which
xt is completely determined by the initial conditions x 0 , i.e. xt = xt (x0 ). The functions hA and
7.2 Transition Path Sampling 25

hB characterize the regions A and B; hA;B (x) = 1 when x 2 A; B, respectively, and hA;B (x) = 0
otherwise. Note that A and B must be chosen in such a way that A \ B = ?.
Since the function xt is fully determined by the initial condition x0 , the ensemble averages
in equation 7.5 can be written as an integration over the initial conditions weighted with the
equilibrium distribution N (x0 ),
R
dx0 NR (x0 ) hA (x0 ) hB (xt (x0 ))
C (t ) = :
dx0 N (x0) hA (x0 )
(7.6)

We can also look at this equation as the ensemble average of h B (xt ) weighted with the equi-
librium distribution N (x0 )  hA (x0 ). In other words, C (t) is the fraction of trajectories that
start in A with distribution N (x0 ) and reach B after time t. Since we are sampling over paths
this ensemble is called the path ensemble. A procedure to sample this ensemble would be to
perform a MD simulation to generate a new path of length t and subsequently use a MC proce-
dure to decide whether to accept or reject this new path. In this way, we generate an ensemble of
paths which we can use to compute ensemble averages. We will only consider a micro-canonical
ensemble of initial conditions x0 , i.e.

N (x0) = Æ (N (x0) - E) (7.7)

in which E is the total energy. In principle we could compute C (t) from an “ordinary” path
ensemble simulation. This would imply that we generate an ensemble of paths of length t that
start at A and we would count all paths that are at time t in B. However, since the transition
from A to B is a rare event, the number of paths that ends in B is so small that such an approach
would require very long simulations. Therefore, we need to help the system explore the regions
of interest.
Suppose that region B can be defined by the value of an order parameter ; x t 2 B if min 
 (xt )  max . For equation 7.5, we may write
R Z max
dx0 exp [- H (x0 )℄ hA (x0 ) hB (xt (x0 ))
C (t) = R dP (; t) ;
dx0 exp [- H (x0 )℄ hA (x0 )
= (7.8)
min

in which
R
dx0 expR[- H (x0 )℄ hA (x0 ) Æ [ -  (xt (x0 ))℄
P (; t)
dx0 exp [- H (x0 )℄ hA (x0 )
=
R
dxo f (xR0; t) Æ [ -  (xt (x0))℄
= : (7.9)
dxo f (x0 ; t)
P (; t) can be interpreted as the probability for the system to be in a state with a certain  after
time t given that the system is in A at time 0. Because P (; t) is quite small in B (i.e. transitions
from A to B are rare), special techniques such as umbrella sampling [2,3] are required to compute
P (; t). As shown in refs. [24, 25], it is advantageous to rewrite C (t) as

hhB (t)iF(x0;T )
C (t) = C t 0 
hhB (t 0)iF(x0;T ) ; (7.10)

in which t; t 0 2 [0; T ℄ and

F (x0 ; T ) = exp [- H (x0)℄ hA (x0) HB (x0; T )
HB (x0 ; T ) = max hB (xt (x0 )) : (7.11)
0tT
26 Rare events

The distribution F (x0 ; T ) can be interpreted as the ensemble of trajectories starting in A and
visiting B at least once in the time interval [0; T ℄. In this way, one has to perform only a single
transition path sampling calculation of C (t 0 ) when calculating the time derivative of C (t),
h i
d h h t i
k = dCdt(t) = hh (Ct ()ti )
B ( ) F(x0;T )

0

dt ; (7.12)
B F(x0 ;T )
0

while the functions hhB (t)iF(x0 ;T ) and hhB (t 0 )iF(x0 ;T ) can be calculated from a (single) separate
simulation.

7.2.2 MC sampling from the distribution F (x0 ; T )

In the transition path sampling method, transition pathways are harvested by sampling the path
ensemble F (x0 ; T ) with a MC procedure. In this subsection, we will present two types of MC
trial moves to generate a new path from an existing one to sample the distribution F (x 0 ; T ).
Sampling of the distribution f (x0 ; t) is similar. We will use the symbols n and o for the new and
old configuration respectively.

Shooting

In a shooting move, first one picks a time t 0 randomly from the interval [0; T ℄ and one makes
an attempt to rotate the old momenta vector p in such a way that the total energy E is constant
(see equation 7.7). The rotation angle is chosen at random from a uniform distribution in a
finite interval [-
;
℄. Second, one has to construct a new path by integrating backward and
forward to obtain the new path. To obey detailed balance, the new path has to be accepted with
a probability
 
acc (o !n )= min 1; FF ((xx0 ((no)) ;; TT)) = hA (x0 (n)) HB (x0 (n) ; T ) (7.13)
0

Shifting

In a shifting move, one translates the initial conditions in time by an amount
t:

x0 (n) = x
t (x0 (o)) : (7.14)

To simplify the acceptance rule we choose a symmetric generation probability for
t,

Pg (
t) = Pg (-
t) : (7.15)

Due to energy conservation along a trajectory the acceptance rule for this trial move equals

acc (o !n )= hA (x0 (n)) HB (x0 (n) ; T ) : (7.16)

Although shifting trial moves do not sample the phase space ergodically because the energy of
the path is not changed, they greatly improve statistics.
7.2 Transition Path Sampling 27

7.2.3 Model system

To illustrate this method, consider a two-dimensional system consisting of 15 WCA particles:
 

 1 + 4 r-12 - r-6 r  rWCA
uWCA (r) = (7.17)
0 r > rWCA

in which r is the distance between two particles and r WCA = 21=6 . We will use  = 1. However,
particles 1 and 2 interact via a double well potential:
" #
2 2
( - w - rWCA )
udw () = h 1- (7.18)
w2

This potential has stable minima at  = rWCA and  = rWCA + 2w that are separated by an energy
barrier with height h. When h is large compared to the total energy E the transitions between
these minima are rare.
In all simulations, we have used w = 0:25 and E = 9. We have confined this system of 15
particles in a circle of diameter 6, resulting in a density of 0:53. We have defined region A as
all configurations for which  < 1:30. We have defined region B as all configurations for which
 > 1:45.

7.2.4 Questions
1. What is the value of C (t) for t = 0 and t ! 1 ? Why ?

2. Perform conventional MD simulations for different values of the barrier height h and com-
pare the equilibrium distribution of . How is C (t) related to this distribution ?

3. Calculate the function C (t) for h = 2 by transition path sampling and compare the result
with a conventional MD simulation. To perform the calculation of P (; t), you can divide
the phase space in five overlapping regions:

 0:00 < 1 < 1:22

 1:20 < 2 < 1:26
 1:24 < 3 < 1:30
 1:28 < 4 < 1:45
 1:40 < 5 < 1

How can one match these distributions ?

BOOK PN 219
Chapter 8

Configurational-Bias Monte Carlo

8.1 CBMC of a single chain

In this exercise, we will look at the properties of a single chain molecule. We will compare
various sampling schemes. Suppose that we have a chain molecule of length n in which there
are the following interactions between beads:
 Two successive beads have a fixed bond-length l. We will use l = 1.
 Three successive beads have a bond-bending interaction
1
U = kt ( - 0 )2 (8.1)
2
in which  is the bond-angle, 0 is the equilibrium bond angle and kt a constant. We will
use 0 = 2:0 rad ( 114:6Æ ) and kt = 2:0.
 Every pair of beads that is separated by more than two bonds has a soft repulsive interac-
tion
Æ A(r-rrcut )2
U (r) = rcut 2
r  rcut (8.2)
0 r > rcut
in which rcut is the cut-off radius (we will use rcut = 1:0 and A > 0).
An interesting property of a chain molecule is the distribution of the end-to-end distance,
which is the distance between the first and the last segment of the chain. There are several
possible schemes to study this property (see also section 3.3):

1. Dynamic schemes.
In a dynamic scheme, a Markov chain of states is generated. The average of a property B
is the average of B over the elements of the Markov chain
Pi N B =
i
hBi  i 1 =
(8.3)
N
When N !1 the expression is exact. Every new configuration is accepted or rejected
using an acceptance/rejection rule:
 When unbiased chains are generated
acc (o ! n) = min (1; exp [- (U (n) - U (o))℄) (8.4)
in which U is the total energy (soft repulsion and bond-bending) of a chain.
30 Configurational-Bias Monte Carlo

 When Configurational-Bias Monte Carlo (CBMC) [26–28] is used

 
W (n)
acc (o !n )= min 1;
W (o)
(8.5)

Q= P=
in which
i n
i=2
j k
j=1 exp [- U (i; j)℄
W= n-1
(8.6)
k
In this equation, k is the number of trial positions and U (i; j) is the energy of the j-th
trial position of the i-th chain segment. The term U (i; j) does not contain the bond-
bending potential, because that potential has already been used for the generation of
the trial positions.
2. Static schemes.
In a static scheme, all generated configurations contribute to the average. To obtain a
canonical distribution, we will have to use a weight factor R
P = B R
i N
h Bi = P= = R
i 1 i
i N
i
(8.7)
i=1 i
For Ri we can write

 When random chains are generated

Ri = exp [- Ui ℄ (8.8)
Here, Ui is the total energy of the chain.
 When CBMC is used
Ri = W (8.9)

These equations are derived in ref. [2].

Questions:

1. The program is provided to calculate chain properties using these four methods. However,
some additional programming has to be done in the file grow.f, which is a subroutine to
grow a new chain using either CBMC or random insertion.
2. Compare the end-to-end distance distributions of the four methods. Which method will
have the best performance ? Investigate how the efficiency of CBMC depends on the
number of trial directions (k).
3. Investigate the influence of chain-length on the end-to-end distance distribution. For
which chain-lengths do the four methods start to fail ?
4. For high temperatures (and for low k t and A), the end-to-end distance distribution looks
like the distribution of a non-self-avoiding random walk. This means that the chain seg-
ments are oriented completely random and the segments are allowed to overlap. For the
mean square end-to-end distance, we can write

r2
=
Xx ! X
* i=n =
y2
!
X
= !+
z+
i n
2
+
i n
2
(8.10)
l2 i=1
i
i=1
i
i=1
i
8.2 CBMC of a simple system 31

in which (xi ; yi ; zi ) are the projection of each segment on the (x; y; z) axis
xi = sin (i ) cos (i )
yi = sin (i ) sin (i )
zi = cos (i ) (8.11)
This set of equations can be reduced to
r2
= n (8.12)
l2
Questions:

 Derive equation 8.12. Hint: the following equations will be very useful:
cos2 (i ) + sin2 (i ) = 1
cos (i - j ) = cos (i ) cos (j ) + sin (i ) sin (j )
h cos (i - j )i = 0 (8.13)
The last equation holds because  i - j is uniformly distributed.
 Modify the program in such a way that r2 is calculated for a non-self-avoiding
random walk. Compare your results with the analytical solution.
 Does
D E
r2 / n (8.14)

hold for a chain with a potential energy function described in this exercise ? Investi-
gate the influence of A on the end-to-end distance distribution.

8.2 CBMC of a simple system

Consider a system with three coordinates (x 1 ; x2 ; x3 ) and phase space density
h  i h i
 (x1 ; x2 ; x3 ) = exp - x21 + x22 + x23 = exp -r
2
(8.15)

We would like to calculate the average r2 ,

D E RRR
dx dx dx r2 
r 2
= RRR 1 2 3 (8.16)
dx1 dx2 dx3 
by using the CBMC algorithm of Falcioni and Deem [12]:

 Generate k sets of new coordinates B 1 ;


; Bk by adding a random vector to the old con-
figuration (A1 ).
 Select one set (i) with a probability proportional to its Boltzmann factor,
pi = exp [-rBi ℄ (8.17)
This leads to a Rosenbluth factor of
j= k
X h i
2
W (n) = exp -rBj (8.18)
j=1
32 Configurational-Bias Monte Carlo

 Starting from the selected configuration B i , k - 1 configurations (A2 ;


; Ak ) are generated

by adding a uniform vector to Bi . A1 is the old configuration. This leads to the Rosenbluth
factor of the old configuration

X
j k
= h
2
i
W (o) = exp -rAj (8.19)
j=1

 The new configuration Bi is accepted with a probability

 
W (n)
acc (o ! n) = min 1; (8.20)
W (o)

Questions:

1. Make a schematic sketch of the configurations A 1 ;


; Ak and B1 ;


; Bk . Show that for

k = 1, this algorithm reduces to the standard Metropolis algorithm for particle displace-
ments.

2. Proof that this algorithm obeys detailed balance.

3. Is it possible to calculate r2 analytically ? Hint: for a > 0,

Z1 h i p

2 2
exp -a x dx = (8.21)
0 2a

4. Enclosed is a computer program for this CBMC sampling scheme. Unfortunately, the
program has to be completed by you (see the file cbmc.f) ! Make sure that your estimate of
r2 is independent of the number of trial directions (k).
5. What happens with the fraction of accepted trial moves when the number of trial direc-
tions (k) is increased ? Make of plot of the fraction of accepted trial moves as a function of
k for various maximum displacements. Explain your results.
6. Why is this CBMC method very useful when the system is far from equilibrium ?

8.3 Overlapping distribution for polymers

In this exercise, we would like to calculate the chemical potential of a chain of length n in a
solvent of monomers. We will use the following potential:

 Two successive beads have a fixed bond-length of 1.

 Every pair of beads of the chain that is separated by more than one bond has a soft repul-
sive interaction
Æ A(r-rrcut )2
U (r) = rcut 2
r  rcut (8.22)
0 r > rcut
in which rcut is the cut-off radius (we will use rcut = 1:0 and A > 0). The same interactions
are used for monomer-monomer and polymer-monomer pair interactions.
8.3 Overlapping distribution for polymers 33

The chemical potential () of the chain in the solvent can be calculated from the average
Rosenbluth factor when test chains are grown using CBMC:

 = - ln hW i (8.23)

However the same thing can be achieved by removing a real chain from the simulation box. The
principle of the overlapping distribution method is to calculate a histogram of the Rosenbluth
weight when adding or removing the chain. If these two distributions overlap a reliable esti-
mate of the chemical potential can be obtained. If these distributions do not overlap sampling
problems may be expected [2, 29].
By constructing the following functions we can calculated the excess chemical potential 

f (- ln (W )) = ln (p0 (- ln (W ))) + 12 ln (W )
(8.24)
g (- ln (W )) = ln (p1 (- ln (W ))) - 2 ln (W )
1

In which p0 is the probability density of - ln (W ) in the case of adding a chain and p1 of remov-
ing one. The chemical potential is found by subtracting these two functions,

 = g (- ln (W )) - f (- ln (W )) (8.25)

Questions:

1. Complete the given program. Additions have to be made in subroutine overdist.f.

2. Why is the difference between the functions g and f not exactly constant ?

3. Why do we still need two simulations when the length of the chain equals 1 ?

4. Compare the two methods to compute the chemical potential for different values of the
density and the chain length. When does the test particle method fail ?

 In the article of Mooij and Frenkel [29] the functions f and g contain some typographic errors and should read
as equation 8.24.
Chapter 9

Appendix A: Introduction to
GNU/Linux

GNU/Linux [30] is a Unix Operating System (OS) for Intel PC’s and other architectures. There
are several advantages to use GNU/Linux instead of Windows 98/NT:

 It is free software, including almost all applications such as compilers and editors. Popular
GNU/Linux distributions, like Redhat, can be downloaded directly from the web [31].

 It is multiprocessor, multitasking and multiuser.

 There is a strict distinction between OS and user files.

 Unix is available on many different computer platforms, ranging from simple PC’s to su-
percomputers.

There are however also some disadvantages:

 It is a little-bit more difficult to learn than Windows 98/NT.

 The commands can be rather cryptic.

 It is not very suitable if you use it only once in a while.

 It might sometimes be difficult to get immediate support.

To use a Unix computer, one always has to login first. To login, one has to provide a user
name (let us assume that the user name is tampert) and user-specific password. One will end up
in the directory /home/tampert. This is called the home-directory of user tampert.

Several GNU/Linux commands or applications are:

acroread: displays an Adobe .pdf file [32]

bzip2: an alternative for gzip

cat <file>: display a file

36 Appendix A: Introduction to GNU/Linux

cd directory: changes the directory to directory. The current directory is called . and the previous
one is called .. cd tampert will change the current directory to the directory tampert. The com-
mand cd will change the current directory to the home directory /home/tampert

compress and uncompress: zip or unzip a .Z file

diff file1 file2: displays the differences between file1 and file2

emacs: a very nice text editor, especially for writing C/Fortran code

ftnchek: a program to detect logical errors in a FORTRAN77 code that the compiler does not
detect [33]. To use more than 80 character on a single line, use the options -columns=131 -
portability=all -nopretty.

g77: GNU FORTRAN77 compiler [9]. Four important compiler options are:
 -O2 produces optimized code
 -C -g makes code suitable to the debugger and checks for array-bound over/underflow.
 -ffixed-line-length-132 allows the use most than 80 character on a single line. Note that this
particular option is strongly machine dependent.
 -Wall prints all warnings to the screen
gcc: GNU C compiler [10]

g++: GNU C++ compiler [10]

gdb: GNU debugger. Very useful to analyze core dumps.

ghostview: displays a postscript (.ps) file

gnuplot: traditional program to make graphs. We recommend you to use xmgr however.

grep text file: find the word text in the file file

gzip and gunzip: zip or unzip a .gz file

joe: a simple text editor to manipulate text files. joe tampert will edit the text-file tampert. Several
useful commands are: Ctrl-K-H provides a help screen, Ctrl-K-X exits with saving the file, Ctrl-C
exits without saving.

kedit and nedit: nice text editors for GNU/Linux

less: an alternative for more

logout: logout

ls: shows the content of a directory. ls -l shows more information. ls t* will only show files or
directories starting with the character t.
 37

make: used to compile programs when a Makefile is provided. To compile a program, simply
type make.

man command: looks for the manual page of command. man -k keyword will browse all manual
pages for keyword. man man will show the manual of man.

mkdir directory: makes a new directory directory

more filename: displays the content of a text-file filename

netscape: especially http://www.garfield.com is interesting

pwd: displays the current directory

rm filename: removes the file filename. There is no undelete command for Unix, so be careful !!!
Use rm -rf to delete a complete directory structure. Be very careful !!

rmdir directory: removes directory directory

startx: start the X (graphical) server of GNU/Linux

tar: tape archive utility. tar cvf <file> will archive and tar xvf <file> will unpack

vi: this is the standard Unix editor which is available on all Unix systems. There are still people
that use it.

xmgr: a nice program to make graphs [34]. To plot more than one graph from a single file, use
xmgr -nxy <file>. Recently, xmgr has evoluted into xmgrace.

xv: a nice program to display and to convert all kinds of graphic files

Beware of the following:

1. Unix is case sensitive, this means that TEST is different from test.

2. Unix uses the no-news-is-good-news principle. If a command is executed and the action is
successful, there is usually no information provided about the result of the action. Exam-
ple: rm t* will remove all files in a directory that start with the character t. The command
rm t < space > * will first try to remove the file t and then all files (* means all files). This
means that everything in the directory is lost !!!. Again, this means lost forever !!!

3. One can use the standard redirection of input/output > < j. For example, cat < file j more
will display file and redirect the output to the command more. Alternatively, one could use
more file.

More information about GNU/Linux can be found on the webpage of this course.
Chapter 10

Appendix B: Introduction to
FORTRAN77

FORTRAN77 is a very old language (1977). It is relatively simple to use, and the syntax is
quite trivial. In this course we will use ANSI FORTRAN77 with a few extensions. We will only
mention some of the very basics of FORTRAN. For more information, there are several sources:

 The Unix man command. For example, man nint will give information about the FOR-
TRAN nint command. Unfortunately, the man pages of FORTRAN77 commands and func-
tions are not always available.

 The book: Interactive Fortran 77: A Hands on Approach, by Ian Chivers and Jane Sleightholme.
It is available on the website of this course.

 The Fortran Market on the web [35].

Some FORTRAN77 keywords/hints:

 All programs start with PROGRAM program-name, SUBROUTINE subroutine-name or

FUNCTION function-name and end with END.

 The first five characters on a line are reserved for line numbers, the sixth character is re-
served for a continuation character. If the first character is a C, the line is just comment.

 FORTRAN77 does not use >; ; <; ; ==; = =, but :gt:; :ge:; :lt:; :le:; :eq:; :ne: instead. How-
ever, on most compilers >; >=; <; <= will work.

 Variable types: integer (-1 0 1 2 3 etc.), double precision (1.0d0 -1.0d0 5.0d0 etc., 1.0d7 means
1:0  107 ), logical (.true. or .false.). These variables are declared at the beginning of the
program. If the statement implicit none is used, all variables must be declared. We strongly
recommend the use of implicit none.

 All arrays start at 1 instead of 0 in C/C++. So integer qq(5) has the elements 1; 2;


; 5,
while integer qq(a:b) has the elements a


b (b > a). Negative values of a,b are allowed.

 Logical operators: :and:; :or:and :not:

 if-then-else-endif, do-enddo and do while-enddo construction

 goto linenumber: jumps to linenumber. We recommend you to avoid the use of many goto
statements.
40 Appendix B: Introduction to FORTRAN77

 stop terminates a program

 Subroutines: call of a subroutine: call integrate(tampert) will call the subroutine integrate
with the argument tampert passed. If the function modifies tampert its value is returned.
The use of functions is almost similar.

 Functions: a b equals ab , sqrt (square-root), sin (sine), asin (arcsine), cos (cosine), acos (ar-
ccosine), tan (tangent), atan (arctangent), int (conversion to integer using truncation), nint
(conversion to nearest integer), dble (conversion to a double precision), exp (exponential),
log (natural logarithm), log10 (common logarithm), max (maximum of two ore more num-
bers), min (minimum of two or more numbers), sinh (hyperbolic sine), cosh (hyperbolic
cosine), tanh (hyperbolic tangent), mod (integer remainder of its first argument divided by
its second argument), abs (absolute value).

 common block. used for global variables.

 include ’filename’ include a file. The include statement is replaced by the content of filename.
We strongly recommend you to put all common blocks in include files.

 save variable-name: stores the value of variable variable-name when the execution of the
subroutine is completed.

 The programs ftnchek is very useful to detect logical errors in your FORTRAN77 code that
are not detected by the compiler [33].

When you use FORTRAN77, there is nearly always an emotional discussion with people
that use C/C++. We completely agree that C/C++ is a much nicer language than FORTRAN77.
The reasons that we (and with us many other people) still use FORTRAN77 are

 FORTRAN77 is very simple to learn and to use. It will generally take more time to learn
C/C++.

 Only a few people that read this manual will be a computer scientist. As physicist or
chemist, we want a program that works and not a program that is programmed in a nice
way. Physics and chemistry are complicated enough !

 FORTRAN77 compilers still produce faster code than C/C++ compilers. This is due to the
longer evolution of the compiler.

Other languages like Pascal, Delphi or Java are either too slow or not available on all Unix
platforms so it is quite obvious why almost nobody uses them in the field of molecular simu-
lations. Although FORTRAN77 has already evoluted to FORTRAN90, FORTRAN95 and HPF,
these compilers are not (yet) part of the GNU project [36] so we have chosen to use an older
version.
 41

program test1 this program will calculate

implicit none sqrt(1.0d0) + ... + sqrt(5.0d0)

integer i declare variables

double precision ee,dd

ee = 0.0d0
do i=1,5 loop over i=1,2,3,4,5
dd = dble(i) convert i to double precision
call tampert(dd) call subroutine tampert
ee = ee + dd
enddo
write(*,*) ee print final result
end end of the program

subroutine tampert(input) subroutine tampert

implicit none

double precision input

input = sqrt(input) call sqrt(input)
return return to main program
end

program test2 this program will calculate

implicit none the sine of a given input
all variables have to be declared
integer i
logical OK
double precision value

1 write(*,*) ’Give a value’ print header

read(*,*) value read in a value
if (value.le.0.0d0.or. test if range is all right
+ value.ge.1.0d0) OK=.false.
if(.not.ok) stop if not all right, then terminate
write(*,*) sin(value) write result
goto 1 return to line 1
end
Chapter 11

Appendix C: Introduction to MOLMOL

Many of the programs used in this course produce .pdb files. A .pdb file contains a snap-
shot/movie of the system, which can be visualized using the MOLMOL program. To view
the movie, one should start molmol first. Then choose File, ReadMol, PDB, system.pdb. Click Select
all (this is on the right panel), Ball/Stick (right panel as well), Options, Animation, Start. Mouse
actions: Left button (rotating the system), Middle button (translating the system), right button
(zoom in/out). One can also zoom in/out by using the menus. More information about MOL-
MOL can be found on the MOLMOL website [37].
Chapter 12

Appendix D: Thermodynamic equations

This section summarizes some basic thermodynamic equations that might be useful while doing
the exercises.

Fundamental Functions

Energy U (12.1)
Enthalpy H =U+pV
Helmholtz free energy F =U- TS
Gibbs free energy G =H- TS
Sometimes the symbol A is used for the Helmholtz free energy.

Differentials of Fundamental Functions

dU = TdS-pdV+
X  dn (12.2)
i i

X
dH = TdS+ Vdp+  dn
i

i i

X
dF = -SdT-pdV+  dn
i

i i

X
dG = -SdT+ Vdp+  dn
i

i i

Maxwell Relations
For example, from dU = TdS - pdV +
P  dn one can derive
i i

 T   p 
i

V
= S (12.3)
S;ni V;ni

and also
 U 
T= (12.4)
S V;ni
46 Appendix D: Thermodynamic equations

Gibbs-Helmholtz equations
 F=T 
T
= - TU2 (12.5)
V;ni

 G=T 
T
= - TH2 (12.6)
p;ni

Specific heat
 
U
Cv = T
(12.7)
V;ni

 U 
Cp = T
(12.8)
p;ni

 p   V 
Cp - Cv =T (12.9)
T V;ni T p;ni
Chapter 13

Appendix E: Equations of motion from

the Lagrangian or Hamiltonian

13.1 Introduction
To demonstrate the basics of molecular dynamics it is sufficient to start with Newton’s equations
of motion. The more advanced techniques use the Lagrangian and the Hamiltonian to derive
the equations of motion. Here we demonstrate the relation between these two approaches. We
outline the principles (based on the famous Feynman lectures on physics [38]). For a more
detailed and formal description of classical mechanics, the reader is referred to the book of
Goldstein [39].
Consider the following statement: the equation of motion gives the path for which the ac-
tion, S, is minimum. The action is defined as the integral over the difference in kinetic U K and
potential UK energy:
Z te
S= dt [UK - UP ℄ : (13.1)
tb

For any other path S is bigger. Let us see whether this statement is reasonable for a few simple
cases.
The first case is a single particle in a zero potential UP = 0. Let us write the velocity of the
particle as the sum of the average velocity vav and the deviation from it (t):

v(t) = vav + (t) (13.2)

where the average velocity is defined in such a way that:

Z Z
dtv (t) = dt vav (13.3)

Since we have only kinetic energy, we obtain for equation 13.1:

Z Z
1 1
S= m dt [vav + (t)℄2 = Sav + m dt 2 (t) (13.4)
2 2
Since the last term is always greater than zero, the action has its minimum if (t) = 0. This
implies that if there is no force acting on a particle, the particle is moving with a constant velocity.
This is equivalent to Newton’s first law.
48 Appendix E: Equations of motion from the Lagrangian or Hamiltonian

We now let our particle move in a one-dimensional potential U(x). The action for this case
is:
" #
Z te  2
1 dx(t)
S= dt m - U(x) : (13.5)
tb 2 dt

An arbitrary path, x(t), can be written as the true path, x̄(t), plus a small deviation from the true
path (t):

x(t) = x̄(t) + (t) (13.6)

We assume that the boundary conditions are chosen such that the beginning and the end of the
paths are fixed, or, (t b ) = (te ) = 0. Furthermore, we assume that the deviation of the true
path is small. The action is minimal if the difference of the action along path x(t), S and the
action along the true path, S̄, is minimal. This is a problem that can be solved using calculus of
variation. Since (t) is small, we can make an expansion of the action around the action of the
true path:
Z te  2
1 dx̄(t) d(t)
S = dt m + - U [x̄(t) + (t)℄ (13.7)
tb 2 dt dt
" #
Z te 2  
1 dx̄(t) dx̄(t) d(t) dU(x̄)
= dt m +2 - U(x̄(t)) + (t)
tb 2 dt dt dt dx
Z te  
dx̄(t) d(t) dU(x̄)
= S̄ + dt m - (t)
tb dt dt dx
te Z te  
dx̄(t) d2 x̄(t) dU(x̄)
= S̄ + m (t) - dt m + (t)
dt tb tb dt 2 dx

The last step of this equation has been obtained via partial integration. Since by definition
(t) = 0 at the boundaries, the second term on the right hand side is zero. The action has its
minimum if the term under the integration is zero for all paths, i.e., for all values of (t):

d2 x̄(t) dU(x̄)
m =- (13.8)
dt 2 dx
which is exactly Newton’s second law. This shows that Newton’s equations of motion can be
derived from our statement that a particle follows a path for which the action is the minimum.

13.2 Lagrangian
One may wonder whether we can use this path formulation of the equations of motion for
something more useful than an elegant derivation of something we learned in high-school. The
answer becomes clear if we realize that this alternative formulation is not limited to Cartesian
coordinates but can be written in any sets of coordinates. Suppose that we would like to use
some generalized coordinates q instead of the Cartesian x coordinate. For example, if we have a
pendulum hanging on the ceiling we could use the angle with the vertical to describe the motion
of the pendulum. Since the true path should be independent of the coordinates which we use to
describe the path, the action should be the same:
Z Z
S = dt L(x; ẋ) = dt L(q; q̇) (13.9)
13.2 Lagrangian 49

where L is called the Lagrangian. The Lagrangian is defined as the kinetic energy minus the
potential energy :

L  UK(q̇) - UP(q) (13.10)

We again introduce our ideal path q̄(t) and the deviation (t) from it:

q(t) = q̄(t) + (t) (13.11)

q̇(t) = q̄˙ (t) + ˙ (t) (13.12)

We can write for the Lagrangian:

L(q; q̇) = L(q̄; q̄˙ ) + L(q̄;q̇ q̄) ˙ (t) + L(q

q̄; q̄)
˙ ˙
(t) (13.13)

Like in the previous section, we use calculus of variation to derive an expression for the true
path. We substitute this Lagrangian in the expression for the action (equation 13.9). Next we
write the actual path as the sum of the true path plus a correction and separate the action of
the true path. Then we use partial integration, and use the fact that at the boundaries of the
integration the deviation from the true path is zero. Finally, we find that the action has its
minimum if:
Z  
d L(q̄; q̄˙ ) L(q̄; q̄˙ )
dt - + (t) = 0 (13.14)
dt q̇ q
To derive the equation of motion we need to introduce a momentum associate to the generalized
coordinate q:

L(q; q̇)
pq  (13.15)
q̇
Substitution of this expression into equation 13.14 gives:

L(q; q̇)
ṗq = (13.16)
q
which is the equation of motion in terms of the generalized coordinates (q; p q ). If the above for-
mulation is valid for any coordinate system, it should certainly hold for Cartesian coordinates.
In these coordinates the Lagrangian reads:

L(x; ẋ) = 21 mẋ2 - U(x) (13.17)

The momentum associated to x is:

L(x; ẋ)
px = = mẋ (13.18)
ẋ
and the equation of motion is:

U(x)
mdotx = - (13.19)
x
which is indeed the result we would obtain from Newton’s equation of motion.
 The true definition is more restrictive; see ref. [39] for more details.
50 Appendix E: Equations of motion from the Lagrangian or Hamiltonian

13.3 Example
Consider the simple pendulum of length l with mass m hanging on the ceiling. The gravitational
force is acting on the pendulum and the potential energy is a simple function of the angle with
the vertical :

U() = mgl [1 - cos ()℄ (13.20)

We would like to write down the equations of motion in terms of the generalized coordinate
. It is an exercise for the reader to do the same using the Cartesian coordinates x and y. The
Lagrangian is:
 
L = UK - UP = 21 m ẋ2 (t) + ẏ2 (t) - U() (13.21)
ml2 ˙ 2
=  - U()
2
The generalized impulse is defined as:

L
p = = ml2 ˙ (13.22)
q̇

and the equation of motion follows from equation 13.16

U()
ṗ = - (13.23)

or

1 U()
˙ = - (13.24)
ml2 

13.4 Hamiltonian
Using the Lagrangian, we have derived the equation of motions in terms of q and q̇. In some
applications one would like to write the equations of motion in terms of q and the conjugate
momentum pq . To do this we can perform a Legendre transformation: y

H(q; pq )  pqq̇ - L(q; q̇; t) (13.27)

y
In thermodynamics the Legendre transformation is used to derive the auxiliary functions. For example, the
energy U is a function of the entropy S and volume V : U = U (S; V ). In some practical application it is more
convenient to work with the temperature T instead of the volume. Since the temperature is the conjugate variable to
the entropy, we can make a Legendre transformation to remove the S dependence:

A = U - TS (13.25)

giving

dA = dU - d(TS) = -SdT - pdV (13.26)

13.4 Hamiltonian 51

This equation is defining the Hamiltonian of the system. For the differential we can write

d H(q; pq) = d(pq q̇) - d L(q; q̇ ) 

(13.28)

= pq dq̇ + q̇dpq -

dq +
L dq̇ +
L 
dt
L
q q̇ t

= pq dq̇ + q̇dpq - ṗq dq - pq dq̇ -


dt
L
t
= q̇dpq - ṗq dq -

dt
L
t
=
H dpq +
 H
dq +

dt
H
pq q t

in which we have used the definitions of p q and ṗq , equations 13.15 and 13.16, respectively.
From this equation we read:

q̇ =
 H (13.29)
pq

ṗq = -
H (13.30)
q

Which are the desired equations of motion in terms of q; p q . For most systems the Lagrangian
does not explicitly depend on time. If this is the case the Hamiltonian is equal to the total energy,
which gives a conservation law.
Also the Hamiltonian formulation is independent of the coordinate system. Of course, it
should be valid for Cartesian coordinates, for which the Hamiltonian reads:

H(x; px) = ẋpx - L(x; ẋ) (13.31)

1
= mẋ2 - mẋ2 + U(x)
2
1 2
= p + U(x)
2m x
The equations of motion are:

ẋ =
H =
px
(13.32)
px m
ṗx = -
H

=-
U(x)
(13.33)
x x
Which are the familiar Newton’s equation of motion again. Comparison with the Lagrangian
formalism shows that with the Hamiltonian we obtain two first order differential equations,
while the Lagrangian gives one second order equation. Of course, in both formalisms the
physics is the same.
Chapter 14

Appendix F: Questions about the book

“Understanding ..” [2]

14.1 Introduction to FORTRAN77

Question 1 (Standard deviation) Write a short FORTRAN77 program that computes the average
N X
and standard deviation of a sequence of numbers. The average of a variable is defined as
P= X i N
hX =
N
i = i 1 i
(14.1)

and the standard deviation as

sP =
=
i N
X X )2
i=1 ( i - h i

N (14.2)

Why is this expression not useful to use for a very large N ? Rewrite the expression for the standard
deviation in such a way that Xi does not have to be stored.

Question 2 (Matrix Multiplication) Write a FORTRAN77 program that multiplies a [a b℄ matrix 

with a [b  ℄ matrix. The result is a [a

 ℄ matrix. By the way, what is the matrix [40] ?

14.2 Chapter 2
Question 3 (Number of configurations)
1. Consider a system A A A
consisting of subsystems 1 and 2 , for which 1 = 10
20 and
2 =
1022 . What is the number of configurations available to the combined system ? Also, compute the

SS S
entropies , 1 , and 2 .
2. By what factor does the number of available configurations increase when 10m3 of air at 1:0atm
and 300K is allowed to expand by 0:001 per cent at constant temperature ?
3. By what factor does the number of available configurations increase when 150kJ is added to a
system containing 2:0mol of particles at constant volume and = T 300K
?
5
4. A sample consisting of five molecules has a total energy . Each molecule is able to occupy states
j
of energy , with =j 0;1;2; ; 1



. Draw up a table with columns by the energy of the states and
write beneath them all configurations that are consistent with the total energy. Identify the most
probable configurations.
54 Appendix F: Questions about the book “Understanding ..” [2]

Question 4 (Thermodynamic variables in the canonical ensemble) If one has an expression for
the Helmholtz free energy (F) as a function of N; V; T

- ln (Q (N; V; T))
F= (14.3)

one can derive all thermodynamic properties. Show this by deriving equations for U, p, and S. You might
want to take a look at Appendix D to refresh your memory.

Question 5 (Ideal Gas (Part 1)) The canonical partition function of an ideal mono-atomic gas is equal

Z
to

1 VN
Q (N; V; T) = 3N d exp [- H℄ = 3N (14.4)
h̄ N!  N!
p
in which  = h̄= 2m= and d = dq1


dqN dp1


dpN . Derive expressions for the following
thermodynamic properties:

 F (N; V; T) (hint: ln (N!)  N ln (N) - N)

 p (N; V; T) (which leads to the ideal gas law !!!)

  (N; V; T) (which leads to  = 0 + RT ln )

 U (N; V; T)

 S (N; V; T)

 Cv (heat capacity at constant volume)

 Cp (heat capacity at constant pressure)

Question 6 (Ising model) Consider a system of N spins arranged on a lattice. In the presence of a
magnetic field, H, the energy of the system is

U=-
XN
Hsi - J
X si sj (14.5)
i=1 i>j

in which J is called the coupling constant (J > 0) and si = 1. The second summation is a summation
over all pairs (D  N for an infinitely large system, D is the dimensionality of the system). This system
is called the Ising model.

Questions:

1. Show that for positive J, and H = 0, the lowest energy of the Ising model is equal to

U0 = -DNJ (14.6)

in which D is the dimensionality of the system.

14.2 Chapter 2 55

2. Show that the free energy per spin of a 1D Ising model with zero field is equal to
F ( ; N) ln (2 cosh ( J))
=- (14.7)
N
when N ! 1. The function cosh (x) is defined as:
exp [-x℄ + exp [x℄
cosh (x) = (14.8)
2
3. Derive equations for the energy and heat capacity of this system.

Question 7 (The photon gas) A photon gas is an electromagnetic field in thermal equilibrium with its
container. From the quantum theory of the electromagnetic field, it is found that the total energy of the
system (U) can be written as the sum of energies of harmonic oscillators:

U=
X n ! h̄ X n 
N
=
N
(14.9)
j j j j
j= 1 j=1

in which j is the characteristic energy of oscillator j, nj = 0; 1; 2;


; 1 is the so called occupancy

number of oscillator j and N is the number of oscillators.

Questions:

1. Show that the canonical partition function of the system can be written as

Q=
YN
1
(14.10)
j=1
1 - exp [- j ℄

Hint: you will have to use the following identity for j x j< 1

X1 x
i=
i
=
1
(14.11)
i=0
1-x
For the product of partition functions of two independent systems A and B we can write

QA  QB = QAB (14.12)

when A \ B = and A [ B = AB.

2. Show that the average occupancy number of state j, hnj i, is equal to

 ln Q 1
nj i = = (14.13)
 (- j ) exp [ j ℄ - 1
h

3. What happens with hnj i when T ! 1 and T ! 0 ?

Question 8 (Ideal Gas (Part 2)) An ideal gas is placed into a gravitational field  (z) = mgz. The
temperature in the system is uniform and the system infinitely large. The magnitude of the field, g is
constant. We assume that the system is locally in equilibrium, so we are allowed to use a local partition
function.
56 Appendix F: Questions about the book “Understanding ..” [2]

Questions:

Z
1. Show that the grand-canonical partition function at height z is equal to

Q (; V; T; z) =
X1 exp [ N℄
d exp [- ( H0 + mgz)℄ (14.14)
N=0
h̄3N N!

in which H0 is the Hamiltonian of the system at z = 0.

2. Explain that a change in z is equivalent to a change in chemical potential, . Use this to show that
the pressure of the gas at height z is equal to

p (z) = p (z = 0)  exp [- mgz℄ (14.15)

(Hint: you will need the formula for the chemical potential of an ideal gas).

3. How does this formula change when the gravitational force g is not constant but a function of z ?
Use the symbol g0 for the gravitation at the surface of the Earth and R for the radius of the Earth.

Question 9 (Chemical Reaction) Consider the chemical reaction

A B (14.16)

in the (ideal) gas phase.

1. The equilibrium ratio of the A and B populations is given by

hnA i qA gA
exp [-
℄
hnB i =
qB
=
gB
(14.17)

in which qi is the partition function and gi the degeneracy of component i and
 the energy
difference between A and B. Shown how the same result follows from the condition of chemical
equilibrium, A = B .

2. The partition function Q of this system can be written as

1 N
Q= (q + qB ) (14.18)
N! A
in which N = nA + nB . Show that the condition of chemical equilibrium is identical to finding the
partitioning of A and B that minimizes the Helmholtz free energy
F F
= = 0 (14.19)
nA nB
Hint: Use ln n!  n ln n - n and remember that F = -kB T ln Q.

3. Show that
D E hnA i hnA i
[ nA - hna i℄2 = (14.20)
N
Hint: Use
  ln Q  NqA
hnA i = qA
qA
=
qA + qB
(14.21)
qB ;N
14.3 Chapter 3 57

14.3 Chapter 3
Question 10 (Reduced units) A typical set of Lennard-Jones parameters for Argon and Krypton is
Ar = 3:41Å; Ar =kB = 119:8K and Kr = 3:38Å; Kr =kB = 164:0K [3].
1. At the reduced temperature T = 2:0, what is the temperature of Argon and Krypton ?

2. A typical time step for MD is
t =0:001. What is this in SI units for Argon and Krypton ?
3. If we simulate Argon at T = 278K and density  = 2000 kg/m3 with a Lennard-Jones potential, for
which conditions of Krypton can we use the same data ? If we assume ideal gas behavior, compute
the pressure in reduced and normal units.

4. What are the main reasons to use reduced units anyway ?

Question 11 (Heat capacity)

1. The heat capacity can be calculated from differentiating the total energy of a system with respect to
the temperature. Is it a good idea to calculate the heat capacity from energies of different simulations
at different temperatures ? Explain.

2. The heat capacity can also be calculated from fluctuations in the total energy in the canonical
ensemble:
U2 - hUi2
Cv = kB T 2 (14.22)

Derive this equation.

3. In a MC-NVT simulation, one does not calculate fluctuations in the total energy but in the potential
energy. Is it then still possible to calculate the heat capacity ? Explain.

Question 12 (A new potential) The authors were getting bored with the Lennard-Jones (12 - 6) po-
tential and decided to introduce the 10 - 5 potential
    5 
U(r) = 5  10 - (14.23)
r r
But the authors are also lazy and it is therefore up to you to derive the tail corrections for the energy,
pressure, and chemical potential (see Chapter 7 of [2]). If we use this potential in an MD simulation in
the truncated and shifted form we still have a discontinuity in the force. Why ? How should the potential
be modified to remove this discontinuity ? If you compare this potential with the Lennard-Jones potential,
will there be any difference in efficiency of the simulation ? (Hint: there are two effects !)

14.4 Chapter 4
Question 13 (Integrating the equations of motion)
1. If you do an MD simulation of the Lennard-Jones potential with a time step that is too big you will
find an energy drift. This drift is towards a higher energy. Why ?

2. Why don’t we use Runga-Kutta methods to integrate the equations of motion of particles in MD ?

3. Which of the following quantities are conserved in the MD simulation of Case Study 4: potential
energy, total impulse, center of mass of the system and angular momentum ?
58 Appendix F: Questions about the book “Understanding ..” [2]

4. Show that the Verlet and velocity Verlet algorithms have identical trajectories.



5. Derive the Leap-Frog algorithm by using Taylor expansions for v t +
t2 , v t -
t2 , x (t +
t)
and x (t).

Question 14 (Correlation functions)

1. To which quantity is the velocity autocorrelation function (vacf) related for t = 0 ?

2. Calculate the limit of the vacf for t ! 1.

3. What is the physical significance if vacf < 0 ?

4. When you calculate the mean square displacement for particles in a system in which periodic bound-
ary conditions are used and in which particles are placed back in the box, you should be very careful
in calculating the displacement. Why ?

5. What is more difficult to calculate accurately: the self-diffusion coefficient or the viscosity ? Ex-
plain.

14.5 Chapter 5
Question 15 (Trial moves)
1. Explain why a large fraction of particle swap trial moves in the grand-canonical ensemble is bad for
the equilibration of the system.

2. Which trial move in Chapter 5 of ref. [2] will be computationally most expensive ? Why ?

3. In a simulation of a molecule that consists of more than one interaction site, a trial move that rotates
the molecule around its center of mass is usually included. Why ? What is the acceptance/rejection
rule for this trial move ?

4. When a particle is added in the grand-canonical ensemble, there might be an energy change due
to a change in the tail-corrections. Why ? Derive an expression for this energy change when a
Lennard-Jones potential is used.

Question 16 (Multicomponent simulation) We consider a grand-canonical scheme for a mixture of

two components. Assume the temperature is T and the chemical potential of the components 1 and 2 .

1. To add or remove particles the following scheme is used:

 Select at random to add or remove a particle.

 Select at random a component.
 Add or remove a particle of this component.

Derive the acceptance rules for these trial moves.

2. An alternative scheme would be:

 Select at random to add or remove a particle.

 Select at random a particle (either in the system or in the reservoir), independent of its identity.

Is this scheme obeying detailed balance if the previous acceptance rules are used ? If not can this be
corrected ? Hint: you might want to see ref. [41].
14.6 Chapter 6 59

14.6 Chapter 6
Question 17 (Andersen thermostat)

1. Why do static properties calculated by NVT-MD using the Anderson thermostat do not depend on
?
2. Suppose we simulate a single particle using the Andersen thermostat. How do we have to choose
the frequency to obtain exactly the same temperature fluctuations as in the canonical ensemble ?

3. Why does the self-diffusivity decrease with increasing  ?

Question 18 (Nosé-Hoover thermostat)

1. Explain when we have to use g = 3N + 1 and g = 3N in the Nosé-Hoover thermostat.

2. Often, the friction term in equation 6.1.25 of [2] is calculated iteratively. Are there some disadvan-
tages when you do this ?

3. Instead of a single Nosé-Hoover thermostat, one can also use a chain of thermostats. Does this
increase the total CPU time of the simulation for a big system ? Explain.

4. Another widely used thermostat is the “temperature coupling” of Berendsen [20]. It is important
to note that this method does not produce a canonical ensemble and therefore we should never use
it. In this algorithm, the temperature of the system is controlled by scaling the velocities every time
step with a factor 

t  T  21
 = 1 +  T0 - 1 (14.24)
T

in which T0 is the desired temperature, T is the actual temperature,
t is the time step of the
integration algorithm and T is a constant.

 Show that this scaling is equivalent with a temperature coupling of the system with a heat
bath at T = T0

J= T
( 0- T) (14.25)

in which J is the heat flux and is the heat transfer coefficient.

 How do and T relate ?

14.7 Chapter 7
Question 19 (Free energy)

1. Why does equation 7.2.11 of ref. [2] (although in principle correct) does not work for hard spheres
? Is there something wrong with the theory ?

2. Derive an expression for the error in estimate of the chemical potential obtained by Widom’s test
particle method for a system of hard spheres. The probability of generating a position with at least
one overlap is equal to p.
60 Appendix F: Questions about the book “Understanding ..” [2]

3. An alternative method to calculate the free energy difference between state A and state B is the use
of the difference in Hamiltonian:
 
- ln hexp [- (HA - HB )℄iN;V;T;B
FA - FB = (14.26)

Derive this equation. What are the limitations of this method ? Show that the Widom’s test particle
method is just a special case of this equation.

Question 20 (Ghost volume) The virial equation is not very convenient to compute the pressure of a
hard sphere fluid. Why ? It is much more convenient to perform a constant pressure simulation and
compute the density. An alternative way to compute the pressure of a hard sphere fluid directly is to
use a ghost-volume change. In this method, a virtual displacement of the volume is performed and the
probability that such a (virtual) move it is accepted has to be computed. Derive that this is indeed a correct
way of calculating the pressure. (Hint: consider the analogy of Widom’s test particle method).

14.8 Chapter 8
Question 21 (Gibbs ensemble)

1. When one of the boxes in the Gibbs ensemble in infinitely large and the molecules in this box do not
have intermolecular interactions, the acceptance/rejection rule for particle swap becomes identical
to the acceptance/rejection rule for particle swap in the grand-canonical ensemble. Derive this
result.

2. From this, it is also possible to derive an equation for the particle exchange in the grand-canonical
ensemble when the gas-phase is not ideal. Proof that in this case, we have to replace the pressure p
by the fugacity f. How is the fugacity defined anyway ?

3. Why does the Gibbs ensemble not work for solid-liquid equilibria ?

Question 22 (Scaling of the potential) When an attempt is made to change the volume in the Gibbs
ensemble, for some systems the energy of the new configuration can be calculated efficiently when the
scaling properties of the potential are used. Consider a system of Lennard-Jones particles. The total
energy U of the system is equal to

U=
X 4 "

12
-


6 #
(14.27)
i<j
rij rij

Suppose that the box-size of the system is changed from L to L0 and s = L 0 =L.

1. Why is this scheme so efficient ?

2. What is the expression for the total virial of this system ?

3. Why does this method only work when the cut-off radius is scaled as well ?

4. Derive expressions for the new energy U0 and new virial V 0 as a function of s, the old energy (U)
and virial (V ).
14.9 Chapter 11 61

14.9 Chapter 11
Question 23 (Barrier crossing) Consider the barrier crossing problem of exercise 7.1.

1. Why does kA!B (t) always decreases as a function of time when the dividing surface is not located
at the top of the barrier ?

2. How will kA!B (t) depend on  when NVT-MD with an Andersen thermostat is used instead of
NVE-MD to compute trajectories ? In this question, you can assume that the dividing surface is
on top of the energy barrier.

14.10 Chapter 13
Question 24 (Biased CBMC) In CBMC, trial positions are selected with a probability that is propor-
tional to the Boltzmann factor of each trial segment. However, in principle one can use another probability
function [42] to select a trial segment. Suppose that the probability of selecting a trial segment i is pro-
portional to

pi / exp [- ? ui ℄ (14.28)

in which 6
? = .

1. Derive the correct acceptance/rejection rule for this situation.

2. Derive an expression for the excess chemical potential when this modified CBMC method is used to
generate configurations of test particles.

3. What will happen when ? ! 1 and ! 0 ?

?
Chapter 15

Appendix G: Possible research projects

15.1 Adsorption in porous media

In this project we will investigate the adsorption behavior in porous media. As a model we use
a slit-like pore. The interactions with the pore are given by

0 0<z<L
U (z ) =
1 otherwise
(15.1)

where L is the width of the slit. We will investigate the adsorption of methane which we model
with a Lennard-Jones potential. The starting point is Case Study 9 of ref. [2] which includes a
program to simulate the Lennard-Jones fluid in the grand-canonical ensemble. The project is to
develop a program the simulate an adsorption isotherm of methane in the slit-like pore. Before
starting to program you may want to think about the following points:

1. What is the geometry of the system and how should one apply the periodic boundary
conditions?

2. What is the type of Lennard-Jones potential (truncated, truncated and shifted, with or
without tail corrections) ? And what are the parameters to model methane ?

3. How is an adsorption isotherm defined thermodynamically ?

With your program, try to answer the following questions.

1. Compute the density profile (density as function of the distance z between the plates) for
L = 1; 2; 5, and 10 for  = 0:7 and T = 2:0. The total number of particles should be of the
order of 100 to 500 for the widest slit.

2. Compute the excess chemical potential and chemical potential as a function of the distance
between the plates for L = 1; 2; 5, and 10 for  = 0:6 and T = 2:0 Try to explain the
differences.

3. Compute the adsorption isotherms for L = 2 and L = 5 for T = 2:0 and T = 0:8. Try to
explain the results.

15.2 Transport properties in liquids

Molecular dynamics simulations can be used to compute the transport properties of liquids.
Examples of these transport properties are the (self) diffusion coefficient and the viscosity. In
64 Appendix G: Possible research projects

Case Study 5 the results for the diffusion coefficient are shown. In this project we extend these
results to mixtures of Lennard-Jones (LJ) molecules. Before starting to program you may want
to think about the following points:

1. The generalization of the diffusion coefficient from a pure component to a mixture is not
trivial. Try to find in the literature how one should define a diffusion coefficient of a
mixture and how can one compute this in a simulation. See, for example, refs. [43, 44].

2. What is the type of LJ potential (truncated, truncated and shifted, with or without tail
corrections) ? And what are the parameters to model Argon and Krypton ?

With your program, try to answer the following questions:

1. Compute the pressure, viscosity, and diffusion coefficient of the LJ fluid at T = 1:0; 1:5,
and 2:0 for  = 0:7.

2. Compute the diffusion coefficients D11 , D12 , and D22 for a mixture of 50% -50% LJ parti-
cles in which the components 1 and 2 have the same interactions (12 = 11 = 22 and
12 = 11 = 12 ) but carry a different color. This means that the particles are labelled.
Experimentally this could be done by radioactive labelling.

3. Compute the pressure, viscosity, and diffusion coefficient of the LJ fluid at T = 1:0; 1:5,
and 2:0 for  = 0:7. Compute the diffusion coefficients D 11 , D12 , and D22 for a mixture
of 50-50% LJ particles, but now for a system of 50% Argon and 50% Krypton (use the
parameters of Argon to compute the reduced temperatures). Does the Einstein equation
for the diffusivity hold for this system ?

15.3 Parallel Molecular Dynamics

In this project, we will test the scaling of a parallel Molecular Dynamics program on a Beowulf
PC cluster [45, 46]. It is interesting to compare the performance of such a cluster with off-the-
shelf components with an expensive supercomputer. For simplicity, we will do MD in only two
dimensions. We will use a soft short-ranged repulsive potential of equation 8.2. Successively,
the following things have to be done:

 Write the sequential MD code using a linked-cell list (see, for example, refs. [2, 3]). How
does your code scale with the number of particles ? Make sure that your code calculates
the average energy drift, the pressure and the radial distribution function.

 Make yourself familiar with MPI [47]. This will take you a few days. We will strongly ad-
vice you to use MPI instead of other parallel libraries like, for example, PVM [48]. Popular
MPI implementations are LAM [18] and MPI-CH [49].

 The program will be parallelized using domain decomposition, which means that every
processor controls a fixed part of the simulation box. For simplicity, we will use a domain
decomposition in slabs. In every domain, the forces are calculated using a linked-cell
method. In a simulation, the following steps are executed:

– Particles are assigned to a processor.

– Every processor has to know the positions and velocities of the particles in its domain,
and also the positions and velocities of the particles in a cell of length r cut next to its
domain (This is the so-called halo region).
15.4 Phase diagram of ethane 65

– Every time step, the halo region has to be swapped between neighboring slabs.
– To calculate quantities like pressure and temperature a global summation over all
processors is necessary.

 Make a detailed work-plan about how you are going to program this. Make a list of all
key variables and arrays. This is necessary because writing this program won’t be easy.

 Discuss this work-plan with your supervisor before you start programming.

 Make a detailed investigation of the scaling of this code with the number of processors.
How does the scaling change with the number of processors ?

 Compare your results with a replicated data algorithm.

15.4 Phase diagram of ethane

Ethane can be modeled as two Lennard-Jones particles that are connected with a spring. In lit-
erature, Gibbs ensemble simulations have been performed of such a model. In this exercise you
will have to modify Case Study 16 for using this model. Because an ethane molecule is quite
small, we can still use conventional Monte Carlo techniques to insert or remove molecules.
Try to find LJ parameters that correctly describe the vapor-liquid part of the phase diagram of
ethane. You might want to take a look at refs. [50, 51] for some details about a model for ethane.

Questions:

1. What kind of potential are you going to use (truncated, truncated and shifted, with or
without tail corrections) ?

2. Is it a good idea to include a new type of trial move (rotation of the molecule around its
center of mass) ? Explain.

3. Do you take a fixed C - C bond-length or do you use a harmonic potential ? Is there any
difference in the algorithm between a fixed or flexible C - C bond ? Explain !

4. What is the bond-length of ethane in dimensionless units ?

5. How can you make an accurate estimate of the critical temperature and density ?

6. How does the fraction of accepted swap moves change with the temperature ?

15.5 Diffusion in a porous media

The behavior of liquid in confined geometries is different from the behavior in the bulk liquid.
In the project we investigate the diffusion coefficient in a cylindrical pore. The starting point
is Case Study 4 of [2] in which the dynamic properties of a Lennard-Jones fluid are simulated.
This Case Study can be used as a starting point for our study. The interactions with the walls
are described with a repulsive potential:
Æ
10
U ( r) =  L-r 0rL
1 r>L
(15.2)
66 Appendix G: Possible research projects

where L is a radius characterizing the size of the pore and the center of the cylinder is located at
r = 0 and  > 0,  > 0. Some questions that one should answer before one starts programming:

1. Is the potential for the interactions with the walls appropriate for a Molecular Dynamics
simulation ?

2. What is the volume of the pore as a function of the parameter L ?

3. What is the dimension of our problem ? Do we have diffusion in 1, 2, or 3 dimensions ?

In the first part of the project we study the diffusion in a smooth pore as defined by the above
potential as a function of the pore diameter.

1. Compute the diffusion coefficient of a bulk Lennard-Jones liquid for  = 0:6 and T = 2:0
and T = 1:5. Since the program uses an NVE ensemble it is not possible to simulate at
exactly the requested temperature. However, one can ensure to be close to this tempera-
ture by an appropriate equilibration of the system (this is also the case for the following 2
questions) during the first part of the MD simulation.

2. Compute the density as a function of the distance from the center of the pore for  = 0:6
and T = 2:0 and T = 1:5 and L = 5 L = 2. Interpret the results.

3. Compute the diffusion coefficient for  = 0:6 and T = 2:0 and T = 1:5 and L = 5, L = 2,
and L = 1. Interpret the results. The interpretation is not trivial.

4. The above calculations have been performed using the NVE ensemble. This implies that
there is no coupling with the atoms of the walls. In a real system the walls are not smooth
and can exchange heat with the adsorbed molecules. A possible way of modeling this
is to assume that we have a Andersen thermostat in the boundary layer with the wall.
Investigate how the results depend on the thickness of the boundary layer and the constant
 of the Andersen algorithm.

The next step is to model the corrugation caused by the atoms. This corrugation could be a
term:
"  #
r-L 2
U (z; r) = A sin2 (z=w ) exp - (15.3)
L0

where z is the w is a term characterizing the size of the atoms of the wall and A the strength of
the interaction. The exponential is added to ensure that the potential is localized closed to the
walls of cylinder. Investigate the diffusion coefficient as a function of the parameters  w and A
both in the NVE and in the Anderson thermostat case.

15.6 Multiple time step integrators

The time step in a Molecular Dynamics simulation strongly depends on the steepness of the
potential energy surface. However, most potentials like the Lennard-Jones potential are steep at
short distances. As short-range interactions can be computed very fast, it would be interesting
to use a multiple time step integration algorithm, in which short-range (computationally cheap)
interactions are computed every time step and in which long-range (computationally expensive)
interactions are evaluated every n time steps (n > 1). Recently, there has been a considerable
effort to construct time-reversible multiple time step algorithms [15, 52].
15.7 Thermodynamic integration 67

1. Why is it important to use time-reversible integration schemes in MD ?

2. Modify Case Study 4 in such a way that pairwise interactions are calculated using a Verlet
neighbor list [2, 3]. For every particle, a list is made of neighboring particles within a
distance of rcut +
. All lists only have to be updated only when the displacement of a
single particle is larger than
=2. Hint: The algorithm on micro-fiche F.19 of the book of
Allen and Tildesley [3, 53] is a good starting point.

3. Investigate how the CPU time per time step depends on the size of
for various system
sizes. Compare your results with table 5.1 from ref. [3].

4. Modify the code in such a way that the NVE multiple time step algorithm of ref. [15] is
used to integrate the equations of motion. You will have to use separate neighbor-lists for
the short-range and the long-range part of the potential.

5. Why does one have to use a switching function in this algorithm ? Why is it a good idea
to use a linear interpolation scheme to compute the switching function from ref. [15] ?

6. Make a detailed comparison between this algorithm and the standard Leap-frog integrator
(with the use of a neighbor-list) at the same energy drift.

15.7 Thermodynamic integration

The difference in free energy between state A and state B can be calculated by thermodynamic
integration [2]
Z =1  
FA - FB = d U() (15.4)
=0 

in which  = 1 in state A and  = 0 in state B. In order to calculate the excess chemical potential
of a Lennard-Jones system, we might use the following modified potential [54]
  12   6 

U (r; ) = 4  r5 - r
3 (15.5)

Recall that the excess chemical potential is the difference in chemical potential between a real
gas ( = 1) and an ideal gas ( = 0).

Questions:

1. Make a plot of the modified LJ potential for various values of .

2. Show that
 
2 F < 0 (15.6)
2
when

U = U0  + U1 (1 - ) (15.7)

3. Derive equation 15.4.

68 Appendix G: Possible research projects

4. Modify the code of Case Study 1 for this modified potential.

5. Perform the thermodynamic integration and compare your results with the conventional
particle insertion method.

6. Calculate the chemical potential as a function of the density by scaling .

15.8 Hints for programming

The official rules for writing a program are that it should involve three persons. One for the
design, one for the implementation, and one for the testing of a program. In our case you are
doing all steps. The reason why there is a need for three persons is that if you do not realize a
certain aspect in the design phase, it is very likely that you do not think about this either while
implementing or testing a program. The following tips may be useful to develop a program:

 First try to understand every line of the starting code. If you do not understand the starting
point it is impossible to make modifications.

 Try to develop a modification plan consisting of the following steps:

1. Make a copy of the program in a new directory.

2. Describe in words what and why a certain subroutine needs to be modified.
3. Start the programming by writing comment lines what the modifications are and
make the modifications according to your written instructions. If you find during the
implementations that the original ideas were not correct describe this in your notes.
This is to ensure that you think before you do. This may sound obvious and it is, but
as you may find out it is a very difficult rule to stick to.
4. Do not start modifying the entire program. Try to do it in steps. Try to test the mod-
ifications as soon as possible. For example, if the program is a Gibbs ensemble sim-
ulation you probably have to modify, the particle displacement routine, the volume
change, and the particle exchange. However, there is no need to do this in all subrou-
tines at the same time. First modify the particle displacement and make all the test
you can do (for example, energy conservation) without use the volume displacement
and particle exchange subroutines, before you continue to the next subroutine.
5. The first time you run the modified code, you might want to use array-bound check-
ing (usually, this option is not used by default). This will slow down the code, but
it is very useful to detect errors. It is also very useful to use code checkers like for
example ftnchek [33].
6. Try to find limiting cases for which your program gives known results. For example,
if you have written a program for a mixture the results should be identical to the pure
component results if all interaction parameters are taken to be identical. Or does the
program give in the ideal gas limit the same results as the ideal gas results that can
be obtained from theory. This is, of course, not alway possible, but if it is possible it
is a good test case.
7. For a Molecular Dynamics program, it is a good idea to check the conservation of the
total energy. In most systems, the total impulse should also be conserved (usually
zero).
15.8 Hints for programming 69

8. In a Monte Carlo program, one has to calculate energy differences for each trial move.
It is a good idea to write a subroutine that rigorously calculates the total energy of
the system. In this way, the sum of the initial energy and all energy differences of all
accepted trial moves should equal the total energy, any difference should only be due
to the limited accuracy of the computer.
9. Especially for lattice simulations, the result of a MC simulation might be dependent
on the random number generator that is used. In principle, all random number gen-
erators are bad, some are less worse than others. It is always a good idea to do a MC
simulation with two different random number generators. See, for example, ref. [55]
for a discussion about this topic.

 If you see numbers try to understand them and check whether they are reasonable. Since
we use reduced units most properties are in the range [-1; 1℄ this implies that if we find
-2:4 it is probably all right, but if we find 2:4  1018 you should be suspicious.
 If you are writing a parallel code, we strongly recommend you to make a detailed work-
plan before you start programming. A common mistake is that you forget to pass system
variables (for example, temperature or masses of the particles) to all processors.
Chapter 16

Appendix H: Software

barrier1 5.1 MD of a single particle in various ensembles

barrier2 7.1 Barrier crossing of a single particle using rare events
boltzmann 2.2 Boltzmann distribution
cbmc 8.2 CBMC program for a simple system
distribution 2.1 Calculate the distribution of particles among compartments
ewald 3.6 The Ewald summation
gibbs 6.1 Gibbs ensemble of LJ particles
harddisks 3.3 Simulation of hard disks in a square
ising 5.3 2D Ising Model using multi-canonical simulations
mc-lj 3.4/6.1 MC simulation of LJ particles in the NVT ensemble
mc-npt 5.2 MC simulation of hard spheres in the NPT ensemble
md-lj 4.1 MD program of LJ particles
md-mpi 4.2 Parallel MD program of LJ particles
oscillators 2.3 System of coupled harmonic oscillators
overlap 8.3 Overlapping distribution for polymers
photon 3.2 MC simulation of a photon gas
pi 3.1 Calculation of PI using brute force MC
pt 3.7 Parallel Tempering
polymer 8.1 CBMC of a single chain molecule
random1d 2.4 1D random walk of a single particle
random2d 2.5 2D random walk of multiple particles
scaling 3.5 MC algorithm in which coordinates are scaled
tps 7.2 Transition Path Sampling
umbrella 6.2 Umbrella sampling for a particle over an energy surface
Acknowledgements

We would like to thank Roland van Vliet for a critical reading of the manuscript and all students
from the March 1999 and January 2000 courses in Amsterdam for their useful comment. Thanks
to Marieke Kranenburg for her contribution to the January 2000 course.
Bibliography

[1] http://molsim.chem.uva.nl/course
[2] Frenkel, D.; Smit, B., Understanding Molecular Simulations: from Algorithms to Applications
Academic Press; San Diego, 1996.
[3] Allen, M.P.; Tildesley, D.J., Computer Simulation of Liquids Clarendon Press; Oxford, 1987.
[4] Chandler, D., An Introduction to Modern Statistical Mechanics Oxford University Press; New
York, 1987.
[5] Rapaport, D.C., The art of molecular dynamics simulation Cambridge University Press; Cam-
bridge, 1995.
[6] Atkins, P.W., Physical Chemistry Oxford University Press; New York, 1998.
[7] Matsumoto, M.; Nishimura, T. ACM Trans. on Modeling and Computer Simulation 1998, 8,
3–30.
[8] http://random.mat.sbg.ac.at
[9] http://egcs.cygnus.com
[10] http://www.gnu.org/software/gcc/gcc.html
[11] Geyer, C.J.; Thompson, E.A. J. Am. Stat. Assoc. 1995, 90, 909–920.
[12] Falcioni, M.; Deem, M.W. J. Chem. Phys. 1999, 110, 1754–1766.
[13] Wu, M.G.; Deem, M.W. Mol. Phys. 1999, 97, 559–580.
[14] Yan, Q.L.; Pablo, J.J.de J. Chem. Phys. 1999, 111, 9509–9516.
[15] Martyna, G.J.; Tuckerman, M.; Tobias, D.J.; Klein, M.L. Mol. Phys. 1996, 87, 1117–1157.
[16] Naghizadeh, J.; Rice, S.A. J. Chem. Phys. 1962, 36, 2710–2720.
[17] Swope, W.C.; Andersen, H.C.; Berens, P.H.; Wilson, K.R. J. Chem. Phys. 1982, 76, 637–649.
[18] http://www.mpi.nd.edu/lam/
[19] Martyna, G.J.; Klein, M.L.; Tuckerman, M. J. Chem. Phys. 1992, 97, 2635–2645.
[20] Berendsen, H.J.C.; Postma, J.P.M.; Gunsteren, W.F. van; DiNola, A.; Haak, J.R. J. Chem. Phys.
1984, 81, 3684–3690.
[21] Smit, B.; Frenkel, D. Mol. Phys. 1989, 68, 951–958.
[22] Bennett, C.H., in Diffusion in Solids: Recent Developments, edited by Nowick, A.S.; Burton,
J.J. Academic Press; New York, 1975, pp. 73–113.
[23] Chandler, D. J. Chem. Phys. 1978, 68, 2959–2970.
[24] Bolhuis, P.G.; Dellago, C.; Chandler, D. Faraday Discuss. 1998, 110, 421–436.
[25] Dellago, C.; Bolhuis, P.G.; Chandler, D. J. Chem. Phys. 1999, 110, 6617–6625.
[26] Siepmann, J.I.; Frenkel, D. Mol. Phys. 1992, 75, 59–70.
[27] Laso, M.; Pablo, J.J.de ; Suter, U.W. J. Chem. Phys. 1992, 97, 2817–2819.
[28] Mooij, G.C.A.M.; Frenkel, D.; Smit, B. J. Phys.: Condens. Matter 1992, 4, L255–L259.
[29] Mooij, G.C.A.M.; Frenkel, D. J. Phys.: Condens. Matter 1994, 6, 3879–3888.
[30] http://www.linux.org
76 Bibliography

[31] http://www.redhat.com
[32] http://www.adobe.com/acrobat
[33] http://www.dsm.fordham.edu/ ftnchek/
[34] http://plasma-gate.weizmann.ac.il/Xmgr/
[35] http://www.fortran.com/fortran/market.html
[36] http://www.gnu.org
[37] http://www.mol.biol.ethz.ch/wuthrich/software/molmol/
[38] Feynman, R.P.; Leighton, R.B.; Sands, M., The Feynmann lectures on physics Addison-Wesley;
Reading, 1965.
[39] Goldstein, H., Classical Mechanics, 2nd ed. Addison-Wesley; Reading, 1980.
[40] http://www.whatisthematrix.com
[41] Rull, L.F.; Jackson, G.; Smit, B. Mol. Phys. 1995, 85, 435–447.
[42] Vlugt, T.J.H.; Martin, M.G.; Smit, B.; Siepmann, J.I.; Krishna, R. Mol. Phys. 1998, 94, 727–733.
[43] Hansen, J.P.; McDonald, I.R., Theory of Simple Liquids, 2nd ed. Academic Press; London,
1986.
[44] Ven-Lucassen, I.M.J.J.van de ; Vlugt, T.J.H.; Zanden, A.J.J.van der ; Kerkhof, P.J.A.M. Mol.
Phys. 1998, 94, 495–503.
[45] http://www.beowulf.org
[46] http://molsim.chem.uva.nl/cluster
[47] MPI: A Message-Passing Interface Standard (http://www.mpi-forum.org)
[48] http://www.epm.ornl.gov/pvm/
[49] http://www-unix.mcs.anl.gov/mpi/mpich/
[50] Smit, B.; Karaborni, S.; Siepmann, J.I. J. Chem. Phys. 1995, 102, 2126–2140, erratum: J. Chem.
Phys. 1998, 109, 352.
[51] Martin, M.G.; Siepmann, J.I. J. Phys. Chem. B 1998, 102, 2569–2577.
[52] Tuckerman, M.; Berne, B.J.; Martyna, G.J. J. Chem. Phys. 1992, 97, 1990–2001.
[53] http://www.dl.ac.uk/CCP/CCP5/main.html
[54] Ilario, G.; Tironi, G.; Gunsteren, W.F. van Mol. Phys. 1994, 83, 381–403.
[55] Hellekalek, P. Mathematics and Computers in Simulation 1998, 46, 485–505.