Beruflich Dokumente
Kultur Dokumente
Fredrik Zetterling
Stockholm 2003
Doctoral Dissertation
Royal Institute of Technology
ISBN 91-7283-547-8
TRITA-NA-0313
ISSN 0348-2952
ISRN KTH/NA/R--03/13--SE
The current work has been done using molecular dynamics computer simulations
of a monatomic system using a simple pair-potential for the interactions. Two
new pair-potentials has been developed, the Zetterling-1(Z1) and Zetterling-2(Z2)
potentials. They are specically modeled to impose icosahedral order. The basis for
the development of the potentials was the old Dzugutov potential which is known to
freeze into a dodecagonal quasicrystal. The new Zetterling potentials have a longer
interaction range and a narrower rst minimum. The more narrow rst minimum
will enhance the local icosahedral ordering and the longer interaction range was
introduced to incorporate a second maximum in the potential mimicing the Friedel
oscillations found in metallic systems. These Friedel oscillations are due to the
singularity which arises at the Fermi surface due to the screening of the positive
charge by the electron gas.
Five papers are included in the study. The rst two papers are studies of
icosahedral clustering in the liquid and supercooled liquid. The simulations in
Paper I was done using the old Dzugutov potential while the new potentials were
used in Paper II using both molecular dynamics and the Basin Hopping algorithm
presented in Chapter 5. Paper III considers the concept of dynamical ergodicity in
the context of the super-cooled liquid behaviour. The simulations were made using
the old Dzugutov potential. Paper IV reports a molecular dynamics simulation
using the Dzugutov potential under super-cooling. A formation of icosahedrally
structured domains with distinctly slow diusion which grows with cooling in a
low-dimensional manner and percolate around Tc , the critical temperature of the
mode-coupling theory. A sharp slowing down of the structural relaxation relative to
diusion is observed. It is concluded that this eect cannot be accounted for by the
spatial variation in atomic mobility. The low-dimensional clustering is discussed as a
possible mechanism of fragility. Paper V investigates the crystallization of a simple
monatomic liquid model which utilizes the Zetterling-1 potential. The system forms
a thermodynamically stable solid phase exhibiting cubic symmetry. Its diraction
pattern is identied as that of -brass, a tetrahedrally packed crystalline structure
with 52 atoms in the unit cell.
Keywords: simple liquids, molecular dynamics, pair potential, icosahedral cluster.
iii
iv
Contents
1 Introduction 1
1.1 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Icosahedral Ordering . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Essential New Results . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.4 Presentation of papers . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.4.1 Paper I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4.2 Paper II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4.3 Paper III . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4.4 Paper IV . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4.5 Paper V . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3 Computer Simulations 9
3.1 Why Do We Need Simulations in Condensed Matter Physics? . . . 9
3.2 Statistical Mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.2.1 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.3 Periodic Boundary Conditions . . . . . . . . . . . . . . . . . . . . . 11
3.4 Time Stepping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.5 The Verlet Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.6 Implementation and Parallellization . . . . . . . . . . . . . . . . . 14
3.7 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.7.1 Main steps of an MD algorithm . . . . . . . . . . . . . . . . 15
3.7.2 Force calculation . . . . . . . . . . . . . . . . . . . . . . . . 16
3.7.3 Preparation of the Data Structure . . . . . . . . . . . . . . 16
3.7.4 Usage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.8 Parallelization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4 Pair Potential 19
4.1 History . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.2 The Pair Potentials Used in this Study . . . . . . . . . . . . . . . . 20
4.2.1 The Zetterling Potential . . . . . . . . . . . . . . . . . . . . 20
v
vi Contents
5 Energy Landscapes 23
5.0.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . 23
5.1 The Method of Basin Hopping . . . . . . . . . . . . . . . . . . . . 25
5.1.1 Metropolis Monte Carlo . . . . . . . . . . . . . . . . . . . . 26
5.2 Analysis of Potential Energy Minima Clusters . . . . . . . . . . . . 27
5.2.1 Results for the Lennard-Jones Potential . . . . . . . . . . . 28
5.2.2 Results for the modied Dzugutov (m-Dz) Potential . . . . 28
5.2.3 Results for the Dzugutov Potential . . . . . . . . . . . . . . 28
5.2.4 Results for the Z1 and Z2 potential . . . . . . . . . . . . . . 29
References 39
Papers 42
I Formation of large-scale icosahedral clusters in a simple liquid
approaching glass transition 43
I would like to thank a bunch of people for making this work possible. But I will
do it in swedish. Frst vill jag tacka min handledare, Mikhail Dzugutov, fr all
hjlp, alla samtal om allt man kan tnka sig och lite till, och fr att du r en
lysande vetenskapsman. Jag vill ocks tacka kollegor i branschen, framfrallt Sergei
Simdyankin fr hjlp och rumssllskap. Tack alla andra som har sttt mig nr
det har gtt dligt, hjlpt mig nr jag har behvt hjlp(Erik, tack fr hjlpen
med bilden som jag aldrig anvnde), spelat ipper med mig(Maria, Magnus, Mns,
Olle E), kt och badat med mig(Anna, Dennis, Olle B, Olof), sett till att jag
inte haft trkigt(Erik, Matti, Anders), lackat onmnbara frsiga metallsaker svarta
i Lill-Jans skogen med mig(Calle) eller lagat min rygg(Klas). Just det, delar av
min doktorandtjnst var sponsrad av NTM, ntverk inom tillmpad matematik,
tacktack.
vii
viii
Chapter 1
Introduction
The history of numerical simulations of atomic systems started fty years ago. The
rst atomistic computer simulation of a liquid was carried out in 1953[29] using a
method called Monte Carlo (MC). It is so called because it explicitly uses a random
number generator.
A dierent simulation method called Molecular Dynamics (MD) was rst im-
plemented in 1959[1][2]. In this method, the dynamical evolution of a system of
interacting particles is followed by solving Newtonian equations of motion. Since
then, both of these methods have been elaborated and widely used for simulation
of various kinds of systems.
The rst MD simulation was carried out with a system of hard spheres. In
this idealized model, particles interact only at the contact distance. Several years
later, Rahman[34] performed a successful MD simulation in which the equations
of motion were solved for a system of particles interacting via a continuous
nite-range Lennard-Jones pair-potential, see equation 3.9.
1.1 Methods
Investigating the inter-atomic interactions in metals can be undertaken by
electronic-structure calculations but are very hard computationally. Quantum-
mechanically based computations are well suited to the problem but becomes very
1
2 Chapter 1. Introduction
vast and complicated quickly. A third approach is the use of a simple pair-potential
reproducing the interactions within the metals on a pair-wise level. This will reduce
the calculations enormously and it will be even better if using a monatomic system.
Monatomic simulations using a simple pair-potential has been very successfully his-
torically with the famous Lennard-Jones potential reproducing the properties of the
noble gases accurately. Later, many dierent potentials have been proposed and
used for a large number of dierent materials.
1.4.1 Paper I
The paper presents formation of extended icosahedral congurations in a molecu-
lar dynamics simulation of the IC[13] liquid under super-cooling. Geometrical
frustration limit bulk growth of connected icosahedral congurations - for most
one-component systems icosahedral clusters contain very few particles. Due to the
shape of the IC potential, the inter-penetrating and face-sharing 13-atom icosa-
hedra in our simulations have been found to form clusters with signicant tendency
for low-dimensional growth, thus circumventing the constraints imposed by frus-
tration. It has been observed that the size of the largest icosahedral cluster grows
under cooling and eventually forms a percolating network. These results directly
demonstrate that the super-Arrhenius slowing down of the liquid dynamics under
super-cooling and the eventual glass transition are accompanied by an appreciable
structural transformation.
1.4.2 Paper II
The paper examines the favoured cluster structures for the two Zetterling poten-
tials discussed in chapter 5. Both potentials behave as monatomic glass-formers
in bulk. The oscillations in the potential are found to lead to global minima that
are non-compact arrangements of linked icosahedra. The structural properties of
the clusters correlate with the glass-forming propensities of the potentials. The
fragilities of the corresponding super-cooled liquids are also found to correlate.
The existing theoretical model are only successful in describing the early, relat-
ively high-temperature stages of super-cooled dynamics. Further cooling results in
a fundamental transformation of the liquid state.
1.4.4 Paper IV
The paper reports a molecular dynamics simulation of a simple monatomic glass-
forming liquid. The transition to deeper minima in the energy landscape under
super-cooling are shown to result in the formation of icosahedrally structured do-
mains with distinctly slow diusion. The domains grow with cooling in a low-
dimensional manner and percolate around Tc , the critical temperature of the mode-
coupling theory. Simultaneously, a sharp slowing down of the structural relaxation
relative to diusion is observed. It is concluded that this eect cannot be accounted
for by the spatial variation in atomic mobility. the low-dimensional clustering is
discussed as a possible mechanism of fragility.
1.4.5 Paper V
The paper present a molecular dynamics simulation investigating the crystallization
of a simple monatomic liquid model which utilizes an inter-atomic interaction poten-
tial. The shape of the potential imitates eective inter-ionic interaction in metallic
systems. The potential is constructed to favour icosahedral arrangement in the
rst coordination shell. We observe that, having been cooled from the equilibrium
liquid state to below its apparent freezing point, the system forms a thermodynam-
ically stable solid phase exhibiting cubic symmetry. Its diraction pattern can be
identied as that of -brass phase[5], a tetrahedrally packed crystalline structure
with 52 atoms in the unit cell. We conclude that the monatomic version of -brass
structure generated in this simulation can be regarded as a new archetype for the
-brass structures observed in many inter-metallic compounds.
Chapter 2
Analysis of liquid metals, metallic clusters and metallic glasses will show the local
order to be predominately tetrahedral[39][16][7].
tetrahedra
5-fold bipyramid
5
6 Chapter 2. Tetrahedrally Close-Packed Phases in Intermetallic Systems
icosahedron
There are three known ways of close-packing atoms in 3-D, FCC face centered
cubic, HCP hexagonally close packed, and icosahedrally packed. But only the
former two will pack periodically containing a mixture of tetrahedra and octahedra.
The icosahedra can also be distorted by inserting more atoms surrounding the
central atom. This is useful if the tetrahedra are only slightly distorted as in some
complex alloys of atoms similar in size. This will allow the formation of bulk phases
with polytetrahedral order. Frank and Kasper[17][18] analyzed the structure of
such alloys and found 3 other coordination polyhedra beside the Z12 which closely
7
resembles an icosahedron, called Z14, Z15 and Z16, with approximately equilateral
triangles as faces, see gure 2.1.
They all have 12 vertices with a coordination of 5, and two, three and four
vertices with a coordination of 6. The vertices with a coordination of 6 are never
adjacent.
Under cooling, these non-icosahedral defects can be linearly connected forming
four-fold and six-fold disclination lines. Under further cooling, the four-fold and
six-fold defects cancel each other, but there will always be remaining six-fold defects
in the form of disclination lines in the system left.
The disclination lines can further be connected forming a network throughout
the bulk of the system and thus connecting all of the coordination polyhedra besides
the Z12.
Under suciently slow cooling, the disclination lines can align and order them-
selves in space, forming a regular periodic pattern thus leading to the crystal form-
ation. Alternatively, if the cooling rate is suciently fast, the disclination lines
will not be able to form a globally ordered periodic pattern; instead, they can be
entangled in an aperiodic network, which leads to the formation of a glass.
Frank and Kasper dened a class of crystals with a skeleton of disclination lines
connecting the atoms with 6 common neighbours. The Z14, Z15 and Z16 are two-,
three- and fourfold nodes in this structure. The periodic crystalline structures
8 Chapter 2. Tetrahedrally Close-Packed Phases in Intermetallic Systems
which are formed in the described way constitute the family of Frank-Kasper
phases. Several such Frank-Kasper phases have been identied. See [32][33][36] for
further references.
The Zetterling-1 potential was shown in Paper V to crystallize into the -brass
structure containing several pentagonal bipyramids in the nearest-neighbour en-
vironment. Paper II discuss the similarities and the dierence between the two
Zetterling potentials in their icosahedral local order.
The identication of icosahedra and icosahedral clusters are further discussed
in Chapter 6.
Chapter 3
Computer Simulations
9
10 Chapter 3. Computer Simulations
coecient D, shear viscosity and the structure factor S(k) may be derived al-
though they describe macroscopic properties. The structure factor describes the
Fourier components of the density uctuations in a liquid, k is the magnitude of
the wave-vector.
An isolated macroscopic system of N identical particles each of which has three
translational degrees of freedom is completely specied by 3N coordinates:
q = (q1 , q2 , . . . , qN ) (3.1)
and 3N momenta:
p = (p1 , p2 , . . . , pN ) (3.2)
of the particles. These values dene a phase-point in the 6N -dimensional phase-
space of the system. The motion of the phase-point in phase-space along its phase-
trajectory is determined by Hamiltons equations
HN
q i = ;i = 1...N (3.3)
pi
HN
p i = ;i = 1...N (3.4)
qi
Also 6N initial values of the coordinates and momenta are required for a simulation.
3.2.1 Model
The Hamiltonian H of a system of N molecules can be written as a sum of kinetic
and potential energy functions of the set of coordinates qi and momenta pi of each
molecule. With equation 3.1, and equation 3.2, we have
H(q, p) = K(p) + V(q). (3.5)
The kinetic energy K is of the form
N
K= p2i /2mi (3.6)
i=1
where mi is the molecular mass and the index runs over the dierent momentum
coordinates.
The potential energy may be divided into terms depending on the coordinates
of individual atoms, pairs, triplets etc, accounting for pair, triplet etc interactions.
V= v1 (qi ) + v2 (qi , qj ) + v3 (qi , qj , qk ) + . . . (3.7)
i i j>i i j>i k>j>i
The potential energy function of spherical particles then takes the form
ef f
V= v1 (qi ) + v2 (rij ). (3.8)
i i j>i
where rij is the distance between atom i and atom j. The rst term represents the
eect of the container walls and an external eld. The second term represents the
eects of the interactions between every pair of molecules in the system. All higher
terms such as three-body contributions are neglected.
The rst MD simulation was carried out with a system of hard spheres. In
this idealized model, particles interact only at the contact distance. Several years
later, Rahman[34] performed a successful MD simulation in which the equations of
motion were solved for a system of particles interacting via a continuous nite-range
Lennard-Jones pair-potential, equation 3.9.
In order to solve the equations of motion for a system with continuously changing
interaction potential, a step-wise algorithm had to be employed. After these early
studies, the eld of computer simulation of liquids and solids has grown rapidly.
The eective pair potential is usually the Lennard-Jones[46] 12-6 potential, g-
ure 4.2, of the form
v LJ (r) = 4((/r)12 (/r)6 ). (3.9)
The parameters and should be chosen appropriately. It has a negative well of
depth and a steeply rising repulsive wall at distances less than . The parameter
is usually expressed in units of temperature and divided with kB , where kB is
Boltzmanns constant. The Lennard-Jones potential is valid for e.g. liquid noble
gases.
Other highly idealized pair-potentials are sometimes used in testing the prop-
erties of a liquid. These are unrealistic but very simple in their form and to use in
computer simulations. The hard-spheres potential is of the form
, r <
v HS (r) = . (3.10)
0, r
Other examples of pair-potentials can be seen in 5.3 and [3].
The simple pair-potential is usually used in simulations of noble gases, but can
also be used in simulations of liquid metals. Liquid metals are two component
systems but can within the Born-Oppenheimer approximation be treated as a one
component system of ions, where the movements of the electrons and ions are
separated in time-scale. The average distribution of electrons is included in the
eective inter-ionic pair-potential.
If, for instance a liquid drop is subject to simulation, that is not a problem since the
suface tension will keep the liquid together, otherwise one has to implement some
kind of container wall to hold the atoms together. This is not always satisfactory
since it will inuence a major part of the particles, therefore periodic boundary
conditions are usually introduced.
Periodic boundary conditions mean that when an atom leaves the central cube
subject to simulation, its periodic image enters at the opposite face of the cube.
That is, the cube is repeated in all directions so that no boundary conditions apply.
In gure 3.1 the atom 1 leaves the central box and enters box B, at the same time
its copy from box F enters at the bottom of the box.
A B C
1
2
H D
G F E
Figure 3.1. A two-dimensional periodic system. Atoms can enter and leave each
box across each of the four edges.
Of course, only one set of atoms is needed during the simulation. The particles
should not be able to sense the symmetry of the cube, and therefore the potential
has to be cut o at a certain distance, and the cube must be at least 2rcutof f wide,
gure 3.2.
The introduction of a cut o potential leads to the introduction of neighbour
lists of the atoms. A neighbour list is a list containing the neighbours of each atom
of the system within the cuto distance, it is continuously updated. Since every
atom only interacts with the atoms within a certain distance, the calculation of the
force acting on the atom is reduced. Otherwise, every atom must be checked with
every other atom in every timestep of the simulation.
A B C
1
2
H 3 D
G F E
acceleration a(t) for each particle will be the vector sum of the force acting on each
particle imposed by the pair-potential used in the simulation. The equations are
solved in a step-by-step basis and the time t is chosen appropriate to the system.
It is usually much smaller than the time it takes for a particle or molecule to travel
its own length.
t2 a(t)
r(t + t) = r(t) + tv(t) + + ... (3.12)
2
t2 a(t)
r(t t) = r(t) tv(t) + ... (3.13)
2
and addition of the two equations 3.12 and 3.13 the equation for advancing the
positions is
The velocities do not enter the equation at all, although they are useful in cal-
culating the kinetic energy. The velocities are obtained by subtraction of equation
3.12 with 3.13 to get
r(t + t) r(t t)
v(t) = + O(t2 ) (3.15)
2t
The Verlet algorithm advances all of the particle positions in one step instead
of two as in other algorithms which rst predict the positions and velocities of the
particles in one step, and then add a correction to the positions and the velocities
in another following step.
The Verlet algorithm is also symplectic. This means that the algorithm correctly
holds the phase-space volume constant during the simulation. This implies time
reversibility. It is also possible to prove that an algorithm with the property of
phase-space conservation can be rewritten as an exact solution for a Hamiltonian
H which diers from the Hamiltonian of the simulated system H by terms of
order of the time-step .
Predictor-corrector schemes are non-symplectic. This can lead to unphysical
results[30] and care has to be taken when using such algorithms.
A modication to the Verlet algorithm with better round-o properties is the
half-step leap-frog scheme. The algorithm is
1
r(t + t) = r(t) + tv(t + t) (3.16)
2
1 1
v(t + t) = v(t t) + ta(t). (3.17)
2 2
The equation 3.17 is evaluated rst to get the mid velocities and then the new
positions are calculated via equation 3.16. The current velocities may be evaluated
between the two steps by
1 1 1
v(t) = (v(t + t) + v(t t)). (3.18)
2 2 2
This is necessary in order to calculate the total energy (H = K + V) of the system
in each step of the simulation. It is nescessary to monitor the total energy during
the simulation since a deviation from an expected constant value may be a sign of
the timestep being to long.
simulation carried out for several millions of timesteps is still only of the order of
nanoseconds in real time. This is in practice not really a problem since we only
need to consider systems with a size suciently large compared to the correlation
length observed for a time suciently large compared to the correlation time .
These can be estimated from space- and time-correlation functions [24, 20].
The nite-dierence integration method used in the simulation is not critical in
speeding up the calculations. The time-consuming part of molecular dynamics is
the calculation of the force acting on every particle for each step. Using a pair-
potential, the force can be calculated as:
N
dv(rij ) rij
Fi = (3.19)
drij rij
j=i
where ri is the position vector of particle i and v(r) is the pair potential. The
identication of all of the interacting pairs of particles is of major importance in
the design of an ecient algorithm.
The size of the time-step used should be as large as possible enabling the sim-
ulation to cover the desired observation time in as few as possible time-steps. The
Verlet algorithm is of second order and the error will of course be larger for a large
time-step which may interfere with energy conservation of the system. This is true
for all simulations and you need to chose a time-step giving an acceptable accuracy
running as fast as possible.
3.7 Implementation
Considering only simple monatomic systems interacting via a pair-potential min-
imizes the load of the computer in the simulation and maximizes the observation
time. This is often necessary to study the extreme conditions in condensed phases
at very low temperatures and studying phase transitions. This also simplies the
analysis of the output.
y
x
Figure 3.3. Data structure, the shaded cells belong to the simulation box. The
curved arrows indicate the translations of the periodic images which are represented
by transparent cells. Figure courtesy of S. Simdyankin.
The particle sites are lled by using the map between the simulation box and
the extended box.
3.7.4 Usage
For the force calculation it now suces to loop over the columns from the simulation
box. The contributing forces from all pairs of particles from the same column and
from the neighboring columns are accumulated.
The particle sites will introduce a very high locality in the data so that particles
that are close in space will be close to each other in memory.
The potential and its derivative, the force, is usually precalculated using a table
in which they are tabulated for specic distances between particles to save time in
the computation. For intermediate separations of particles interpolation is used.
18 Chapter 3. Computer Simulations
3.8 Parallelization
The described algorithm and data-structure can very easily be parallelized by the
use of a divide-and-conquer strategy. The approach is very similar whether it is a
shared-memory or a distributed memory computer. The cells are divided equally
among the processors and to prevent accidental overlap which may lead to incorrect
results, there is a need for synchronization at a barrier. Parallelization will only give
a speed-up if the system simulated is large enough. For shared-memory computers
the speed-up can be somewhat hindered by heavy memory-bus trac.
Chapter 4
Pair Potential
4.1 History
19
20 Chapter 4. Pair Potential
1.2
u/
1.2
1 3
r/
Figure 4.1. A generic metallic pair potential, featuring Friedel oscillations, courtesy
of Sergei Simdyankin.
1.5
0.5
U
0.5
1
0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6
r
and a slightly dierent cuto distance. The functional form of this potential is
expressed by 4.1:
A exp(r) cos(2KF r) P
V (r) = + B (r) + V0 (4.1)
r3
It utilizes a repulsive core followed by the periodic decaying tail imitating the
Friedel oscillations. The parameter V0 is chosen so that the potential and its deriv-
ative, the force, vanish at rcut , the cuto distance, which is chosen to be the point
at the third minimum of the potential.
The potential consists of a short-range repulsive core and a longer-range oscil-
latory part. The latter is assumed to represent the Friedel oscillations which are
characteristic of the eective inter-ionic potentials in simple metals. The period of
these oscillations is determined by the Fermi wave-vector KF which is the function
of the density of valent electrons n.
n = KF 3 /3 2 (4.2)
22 Chapter 4. Pair Potential
It is thus possible to calculate the theoretical valency of the model metal from
the KF used in the simulation. For the current potentials it is between two and
three, consistent with the materials it is supposed to crystallize into.
Using a proper choice of parameters the potential was constructed to encour-
age icosahedral arrangement in the rst coordination shell, consisting of all of the
nearest-neighbour atoms. The values of parameters dening the Zetterling-1 pair-
potential are presented in Table 4.2.1, and the Zetterling-2 potential is tabulated
in Table 4.2.1.
A B KF P
1.58 0.95533 4.12 -0.22 -18
Table 4.1. The values of parameters used in the present study for dening the
Zetterling-1 pair potential described by Eq. 1.
A B KF P
1.04 0.94656 4.139 0.33 -14.5
Table 4.2. The values of parameters used in the present study for dening the
Zetterling-2 pair potential described by Eq. 1.
Energy Landscapes
5.0.2 Introduction
The topography of the potential energy hyper-surface, the potential energy land-
scape, see g 5.1, provides useful information about materials. This schematic
picture makes clear why it is sometimes referred to as an energy landscape.
The set of points following a steepest descent potential energy minimization
that would end up in a specic minimum are referred to as a basin of attrac-
tion for the minimum. These basins are connected by the saddle-points of the
hyper-surface. The more complex the potential is, the more complex is the energy
landscape, with more saddle points and a more complex connectivity.
23
24 Chapter 5. Energy Landscapes
Figure 5.1. Schematic diagram illustrating the energy landscape, open circles are
maxima, lled circles are minima, crosses are saddle points. The solid curves are
equipotential curves. Dierent basins are indicated by dashed curves. Picture from
[37].
Protein folding, the complex physics of glasses, and the structure and dynamics
of molecular and atomic clusters used to be studied separately. Energy landscape
theory unies the experimentalists and theorists in the apparently dierent dis-
ciplines as a way to describe the structural formations and the dynamics in these
diverse systems.
These three systems, proteins, glasses and clusters, can all be seen as having a
energy landscape with a low-dimensional free energy surface.
In order to study the behaviour of varying materials and the correlation between
the potential, or pseudo-potential, and the structure in computer simulated systems,
a method called the basin-hopping [43] global optimization algorithm has been
developed. It explores the energy landscape of a potential in a system, revealing
clues to the clustering tendencies and the underlying structure.
The method can generate clusters of increasing sizes for a given potential.
These can be directly compared to clusters in a molecular dynamics simulation.
The local order of the clusters should be very well comparable to the smaller
clusters of the liquid and the larger cluster sizes should resemble the network of
clusters found in the super-cooled liquid of the molecular dynamics simulation.
5.1. The Method of Basin Hopping 25
The idea of the potential energy surface as a key to determining the prop-
erties of glasses was recognized by Goldstein [19]. There was much progress
in the 1980s in calculating the thermodynamical properties directly from the
knowledge of local minima. An approach that has been extended to dynam-
ical properties and used on a large variety of atomic and molecular clusters.
Also, global features of the energy landscape can explain dierent properties of
a variety of glasses. It has recently been suggested [47] that glassy liquids are
composed of large numbers of dierent clusters, each with its own energy landscape.
Following Goldstein, Stillinger and Frank [37] introduced the inherent structure
for liquids, a multidimensional potential-energy minima that can be reached by
steepest-descent minimization of the liquid structure. The inherent structure gives
insight to the complexity of the liquid-solid transition and the solid structure of
the system, and stays remarkably constant at varying temperatures in the liquid
domain, undergoing a distinct change as the system crystallizes and leaving the
area sampled at the liquid state for a minimum at a lower potential energy.
The theory of the inherent structure is a help in the understanding of the
thermodynamics and dynamics by the energy landscape for any system of particles.
Previous work [6] on energy landscapes has identied features enabling the sys-
tem to nd its global minimum eciently. The analysis of model energy landscapes
has been particularly helpful with the use of a master equation approach for the
dynamics, providing evidence of a large potential energy gradient and the lowest
possible transition state energies or rearrangement barriers for such a surface. A
simple way of transforming the potential energy surface which does not change the
global minimum nor the relative energies of the local minima was introduced by
Wales and Doye [43].
To every point X they assign the transformed energy E dened by
E(X) = min{E(X)} (5.1)
26 Chapter 5. Energy Landscapes
Figure 5.2. Schematic diagram illustrating the energy transformation, The solid
line is the energy of the original surface, the dashed line is the transformed energy.
The convergence criterion for the conjugate gradient method used in the Monte
Carlo simulation does not need to be very tight for the method to function properly
[43].
Using the basin-hopping algorithm, a xed number of atoms interacting via the
chosen potential are placed in a container. The combination of the Monte Carlo
simulation and the minimization at each step will eventually produce the potential
energy minimum cluster for that number of atoms. The algorithm is applied to an
increasing number of atoms in the system to study the behaviour and structural
growth of clusters at increasing sizes. The method is not guaranteed to produce
the global minimum cluster for every chosen size, but the main interest is in the
growth pattern and the local structure of the atoms in the cluster produced.
Pair Potential
2
Zetterling1 Potential
Dzugutov Potential
1.5 Zetterling2 Potential
Modified Dzugutov Potential
1
0.5
U
0.5
1
1 1.5 2 2.5
r
Some of the potentials to which the basin-hopping algorithm has been success-
fully applied are displayed in gure 5.3. These are further discussed in the following
sections.
28 Chapter 5. Energy Landscapes
Figure 5.4. Previously unreported global minima for the Lennard-Jones potential.
Figure from [43]
17 57
26 76
55 129
78 166
Figure 5.5. Representative structures of the global minima for the modied
Dzugutov potential. On the right is the complete structure, in the middle the dis-
clination network and to the left the center of the cluster. Figure from [12]
ations for increasing sizes. One consequence of the maximum in the potential is the
favouring of polytetrahedral structures, and the expected structures are found for
the smallest sizes of the clusters. The 13-atom global minimum is the icosahedron
and the smaller sizes is the growth sequence for an icosahedron. The further growth
exhibit a linear character, as shown in gure 5.6.
At larger sizes the global minima become more two-dimensional because of the
strain associated with polytetrahedral growth. The conclusion is that Dzugutov
clusters tend to form non-compact polytetrahedral structures based on a local ico-
sahedral structure.
13 19 23 25 29
31 33 35 37 38
42 46 52 58
60 66 67 86
Figure 5.6. Representative structures of the global minima for the Dzugutov po-
tential. Figure from [11]
33
34 Chapter 6. Icosahedrally Ordered Solid Phases
4
g(r)
0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
r
where vectors rk denote the particle positions. First we consider the spatial de-
composition of the main maximum of S(Q). For this purpose, S(Q) is calculated
on the sphere the radius of which corresponds to the position of the maximum, Qo .
In this way, the symmetry of the structure can be established. After detecting the
symmetry axis, the diraction pattern can be calculated for comparison with the
diraction patterns for known structures. See gure 6.3.
0
0 5 10 15 20 25 30
Figure 6.2. The static structure factor computed from the radial distribution
function in gure 6.1.
several millions of time-steps and the simulation will be stopped when it is determ-
ined by dynamical properties that the system is statistically in an equilibrium state.
This must be determined by investigating the course-grained potential energy and
pressure of the system which must be stable. The mean square displacement is
empirically known to be at a value of at least 1 for a normal liquid, before it is
in equilibrium, and for super-cooled liquids the mean square displacement must be
even higher before it can be considered to be in an equilibrium state. Any phase
transition will show itself as discontinuities in the potential energy and as a change
in the slope of the mean square displacement, and the system must be equilibrated
at the new phase.
After identifying the clusters, the stability of the clusters can be monitored by
looking at the trajectories of the individual atoms in the cluster, and the diusion
rate can be computed for atoms within and outside clusters.
At the higher density, having been cooled from the equilibrium liquid state to
below its apparent freezing point, the system forms a thermodynamically stable
solid phase exhibiting cubic symmetry. This was identied as that of the -brass
phase, a tetrahedrally packed crystalline structure with 52 atoms in its unit cell.
This is reported in Paper V.
At the lower reduced density = 0.84, the system forms another thermody-
namically stable solid phase with cubic symmetry after a stepwise cooling from the
liquid state. The system was thoroughly equilibrated at each step in the cooling
6.5. Crystallization Patterns for the Zetterling-1 Potential 37
procedure. A liquid-solid phase transition occured as the system was cooled below
the reduced temperature T = 0.50. This was observed as a discontinuous drop
in the systems potential energy. The onset of an anisotropic long-range order is
signaled by the appearance of the sets of sharp peaks in S(Q) gure 6.2.
Figure 6.4. The diraction pattern for a perfect model -manganese structure.
The perfect -manganese structure was also relaxed at a nite temperature and
was stabilized by the potential used thus conrming that it is at least in a metastable
equilibrium. The nature of the apparent 8-fold symmetry of the diraction pattern
in 6.3 is revealed by a closer inspection of the structure. A case of twinning is
present and in gure 6.5 three separate domains are tiled and outlined, showing the
angle between domains to be 45 degrees and 90 degrees.
38 Chapter 6. Icosahedrally Ordered Solid Phases
Figure 6.5. A layer of the -manganese structure cut perpendicular to the axis
corresponding to the diraction pattern in gure 6.3, showing the twinning of the
crystal.
The tiling is performed by identifying rings of atoms with one or two atoms in
between the rings enclosed in a rhombus. The chosen layer is quite thin to ease the
guiding of the eye in the tiling process. This result is still to be further analyzed
and published.
The apparent dierence between the structures formed from the same pair po-
tential at dierent densities opens up for studies of possible liquid-liquid and solid-
solid phase transitions in the simulated metallic-like model system.
References
[9] J. P. K. Doye and D. J. Wales, J. Phys. B: At. Mol. Opt. Phys. 29, 4859 (1996)
[12] J. P. K. Doye and D. J. Wales, Polytetrahedral clusters, Phys. Rev. Lett. 86,
5719-5722 (2001)
39
40 References
[16] J. Farges, B. Raoult and G. Torchet, J. Chem. Phys. 59, 3454 (1973)
[18] F. C. Frank and J. S. Kasper, Complex alloy structures regarded as sphere pack-
ings. ii. analysis and classification of representative structures, Acta Cryst., 12
(1959), pp. 483499.
[20] J.-P. Hansen and I. R. McDonald, Theory of Simple Liquids, Academic Press,
London, 2nd ed., 1986.
[26] M. H. McAddon and W. A. Goddard, Phys. Rev. Lett. 55, 2563 (1985)
[28] N. Metropolis and S. Ulam, The Monte Carlo Method, J. Am. Stat. Ass. 44,
335-41 (1949)
[32] D. R. Nelson, Order, frustration, and defects in liquids and glasses, Phys. Rev.
B, 28 (1983), pp. 55155535.
Non-ergodic dynamics in
supercooled liquids