STRUCTURAL PROPERTIES OF GLASSES:

AN INTRODUCTION

Dr. Giorgio F. SIGNORINI, PhD

Dipartimento di Chimica, Università di Firenze
via della Lastruccia, 3
I-50019 Sesto F. (Firenze), Italy
[email protected]

- last update: July 7, 2021

Contents

1 Introduction 3
1.1 A glass has a structure -but not only . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Description of glass structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2.1 order in a disordered structure? . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2.2 the pair distribution function, g (r) . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2.3 properties of g (r) and the radial distribution function . . . . . . . . . . . . . . . 9
1.2.4 extension to binary systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.2.5 information contained in g (r) . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.2.6 geometry of the atom surroundings (Voronoi polyhedra) . . . . . . . . . . . . . 14
1.2.7 more structural parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.2.7.1 atom connectivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.2.7.2 bond angles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.2.7.3 medium-range structures (rings) . . . . . . . . . . . . . . . . . . . . . 20
1.3 reference texts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

2 Methods for the investigation of glass structure 21

2.1 Experimental methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.1.1 methods based on diffraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.1.1.1 X-ray scattering (XRS) . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.1.1.2 neutron scattering (NS) . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.1.1.3 Electron scattering (ES) . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.1.1.4 Comparison of different scattering methods . . . . . . . . . . . . . . 30
2.1.1.5 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.1.2 NMR and vibrational spectroscopy . . . . . . . . . . . . . . . . . . . . . . . . 35
2.1.2.1 NMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.1.2.2 Vibrational spectroscopy . . . . . . . . . . . . . . . . . . . . . . . . 36
2.1.3 Reference texts for section §2.1 . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.2 Computational methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.2.1 Empirical force fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.2.2 Monte Carlo simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.2.2.1 Thermodynamical averages . . . . . . . . . . . . . . . . . . . . . . . 43
2.2.2.2 How MC works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.2.2.3 Energy barriers and the cooling process . . . . . . . . . . . . . . . . . 45
2.2.2.4 computation of g (r) . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.2.3 Molecular Dynamics simulation . . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.2.3.1 Phase space averages . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.2.3.2 How MD works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

1
CONTENTS 2

2.2.3.3 Non-ergodic systems . . . . . . . . . . . . . . . . . . . . . . . . . . 49

2.2.3.4 Time dependent quantities . . . . . . . . . . . . . . . . . . . . . . . . 49
2.2.4 Comparison between MC and MD . . . . . . . . . . . . . . . . . . . . . . . . . 50
2.2.4.1 Computer simulation of glasses . . . . . . . . . . . . . . . . . . . . . 51
2.2.5 Examples of computational studies . . . . . . . . . . . . . . . . . . . . . . . . 51
2.2.5.1 Computer simulation of the CKN glass . . . . . . . . . . . . . . . . . 56
2.2.6 Reverse Monte Carlo . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
2.2.6.1 How RMC works . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
2.2.6.2 RMC Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
2.2.7 Reference texts for section §2.2 . . . . . . . . . . . . . . . . . . . . . . . . . . 65

3 Models of glass structure 66

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.1.1 Bonds in solids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.1.2 Classification of crystals according to bond type . . . . . . . . . . . . . . . . . 68
3.1.3 Structural models of glasses . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
3.2 Continuous Random Network (CRN) . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
3.2.1 General properties of the model . . . . . . . . . . . . . . . . . . . . . . . . . . 73
3.2.1.1 Open or closed structure? . . . . . . . . . . . . . . . . . . . . . . . . 74
3.2.1.2 Number of constraints . . . . . . . . . . . . . . . . . . . . . . . . . . 75
3.2.1.3 Coordination numbers . . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.2.1.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.2.1.5 Comparison to experimental structures . . . . . . . . . . . . . . . . . 77
3.2.2 Polyatomic CRN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3.2.2.1 Oxide glasses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3.2.2.2 Structural theories of oxide glass formation; Zachariasen’s rules . . . . 84
3.2.2.3 Glasses with no chemical order . . . . . . . . . . . . . . . . . . . . . 86
3.2.3 Coordination number and rigidity (Phillips-Thorpe theory) . . . . . . . . . . . . 87
3.2.4 Bond models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
3.3 Random Close Packing (RCP) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
3.3.1 Comparison of RCP to crystal close-packed structures . . . . . . . . . . . . . . 98
3.3.2 Icosahedral geometry and local versus long-range packing . . . . . . . . . . . . 99
3.3.3 More on comparing RCP to crystal close-packed structures . . . . . . . . . . . . 103
3.3.3.1 lattice polyhedra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
3.3.3.2 Voronoi polyhedra . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
3.3.4 close-packed structures of binary systems . . . . . . . . . . . . . . . . . . . . . 107

4 Glass structures 109

4.0.1 Oxide glasses: simple . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
4.0.2 Oxide glasses: multi-component . . . . . . . . . . . . . . . . . . . . . . . . . . 109
4.0.3 Non-oxide glasses with CRN structure . . . . . . . . . . . . . . . . . . . . . . . 109
4.0.4 Other non-oxide glasses: metal, ionic . . . . . . . . . . . . . . . . . . . . . . . 109
Chapter 1

Introduction

1.1 A glass has a structure -but not only

When dealing with glasses from a physico-chemical perspective, we first need to clarify what scientists
mean with the term “glass”.
In a broad sense, a glass is often defined as a “solid with no long-range order”, sometimes with some
additional requirement. See for example Varshneya (1994):

This view comes from the observation that the glassy state closely resembles the liquid in terms of
their generic response to diffraction (see figure 1.1):

3
CHAPTER 1. INTRODUCTION 4

Figure 1.1: schematic of structure and diffraction from a monoatomic substance in different states

It should be noted, however, that the definition we just introduced also encompasses systems that have
no direct relation to the liquid state, a point that, instead, many physicist consider central. In their view, a
glass has a “liquid-like” structure because it originates from a liquid: it is a supercooled liquid that is no
longer capable of rearranging in the laboratory time scale. In fact, a more generally accepted definition
is that of Shelby Shelby (2005):
CHAPTER 1. INTRODUCTION 5

The dynamical (as opposed to structural) properties of a glass manifest themselves in time-dependent
phenomena. In the transition from liquid to glass, the substance’s structure loses its ability to relax or
de-correlate (figure 1.2); the diffusion of the particles comes to an arrest (figure 1.3); and the viscosity
approaches infinity (figure 1.4).

Figure 1.2: Time self-correlation function of particles in a model system (the Lennard-Jones liquid), at
different temperatures around the liquid-glass transition. At low T, decorrelation times are of the order
of days.

Figure 1.3: mean squared displacements of particles in the Lennard-Jones liquid, at different tempera-
tures around the liquid-glass transition.
CHAPTER 1. INTRODUCTION 6

Figure 1.4: viscosity (η) of various substances near the liquid-glass transition. By convention, the transi-
tion temperature is where η = 1013 P oise

We will not be treating the dynamical properties of glasses here; however, it should be kept in mind
that they are an essential feature of the glassy state.
Lastly, note that structure and dynamics are not completely independent in a system. This can be
readily understood by considering that in order for an atom to move to a different environment, that is to
diffuse, there are structural requirements that have to be realized, at least instantaneously (see figure 1.5).

More specifically, the ability to diffuse is determined by collective moves and by their time scale as
compared to single particle velocities; see Royall and Williams (2015).

1.2 Description of glass structure

Limiting ourselves to the structure, the central point is how it can be described; that is, which structural
parameters or functions are best suited to describe the glassy state.
The structure of systems with long range order, such as crystals, can be defined exactly by few phys-
ical quantities: three vectors to define the unit cell, and a limited number of atom positions within the
cell. On the contrary, there cannot be any exact description of atom positions in a glass; only a statistical
description is possible, perhaps through some order parameters.

1.2.1 order in a disordered structure?

One could wonder if it even makes any sense to look for some order in a disordered system.
CHAPTER 1. INTRODUCTION 7

Figure 1.5: The free volume available for movement allows atom A, but not atoms B and C, to diffuse

However, as in the liquid state, glasses do not exhibit a completely disordered structure, that is the
condition where atom positions are totally random, as in a perfect gas. Indeed, in liquids and glasses
there are some preferred atom arrangements, due to the particles impenetrability, on one side, and to their
tendency to be in close or near-close contact, on the other. Thus, moving from a certain atom in the
origin, there is a zero probability of finding a second atom at distances less than the contact distance (the
sum of the atom radii, 2σ), while there is a higher probability of finding it at 2σ than at a longer distance.
In systems where directional bonds impose additional structural constraints, there may be an increased
probability of finding a third atom at a distance dictated by bond length and angle. Correlation with atoms
further apart are expected to vanish rapidly.
In summary, while they have no long range order, glasses exhibit partial short-range order.

1.2.2 the pair distribution function, g (r)

The most useful parameter, and by far the most used one, containing structural information of a glass is
the pair distribution function g (r)1 .
The pair distribution function (in short PDF) measures the relative probability of finding an atom at
distance r from an other atom, scaled by the same probability if atom positions were totally random, or
uncorrelated:
1
many authors give g (r) the name radial distribution function; we will use this term for a different function, closely related
to g (r) -see below
CHAPTER 1. INTRODUCTION 8

hn (r)i
dr
g (r) =
hn (r)iunc
r

where A
hn (r)i dr = (conditioned) probability of finding an
atom in a small volume dr around
position r with respect to a given atom
(A)
hn (r)iunc dr = the same, in an uncorrelated system (such
as an ideal gas), and
hi = represents a thermodynamic average
Thus, g (r) tells us how much, on average, the presence of one atom influences the presence of another
atom at a certain distance, with g (r) = 1 for no influence.
Also note that hn (r)i is a probability density: it has dimensions of numbervolume
of particles
. If integrated over
some spatial domain D, it gives the number of atoms in that domain:
Z
hn (r)i dr = ND (1.1)
D

The integral over the total volume V of the systems equals the total number N of particles2 :
Z
hn (r)i dr = N
V

We now consider two important simplifications on g (r).

In principle, in a disordered system, the arrangement of atoms around one origin atom -which is the
property represented by g (r)- may be different when measured in different points of the system; think,
for example, of a material that has a density gradient along some direction. This is not the case in a glass,
which is homogeneous: that is, its properties are (on average) the same in every point. In such a system,
g (r) does not depend on the position of the origin atom, but only on the vector distance r from it.
In a homogeneous system the uncorrelated probability density hn (r)iunc is constant, and equals the
number density ρ0 :
hn (r)i
g (r) =
ρ0
with
N
ρ0 =
V
N = total number of atoms
V = total volume

Second, we note that, unlike crystals, glasses are also isotropic: their physical properties do not
change with sample orientation. As a consequence, g (r) does not depend on the direction of r, but only
on its absolute value.
2
more precisely, it yields N − 1, since one particle lies in the origin; but in macroscopic samples, where N  1, this
distinction is irrelevant
CHAPTER 1. INTRODUCTION 9

In conclusion, the pair distribution function of glasses is a function of a scalar value:

g (r) → g (r)

and
hn(r)i
g (r) = ρ0
(1.2)

1.2.3 properties of g (r) and the radial distribution function

In agreement with its definition, the pair distribution function of a glass features a first peak at the first
neighbours shell, then a few rapidly fading peaks at higher values, before getting to its limiting value (see
figure 1.6).
As the interparticle distance r increases, correlations vanish, and accordingly

lim g(r) = 1
r→∞

The space differential in equation (1.1) is best expressed in polar coordinates {r, θ, φ}

dr = r² sin θdrdθdφ

-which is the volume of an element of spherical shell- and the expression 1.1 for the number of atoms in
domain D in this reference Z
ND = hn (r)i r² sin θdrdθdφ
D

can be integrated over all θ and φ values, as hn (r)i is independent from angular coordinates, giving
Z
ND = 4πr² hn (r)i dr
ZD

= 4πr²ρ0 g (r) dr
D

As the last equation suggests, the function 4πr²ρ0 g (r), which is called by many authors the radial
distribution function or RDF, is of some interest in itself: its differential

ρ0 4πr2 g (r) dr

represents the number of atoms contained in a spherical shell of thickness dr (figure 1.8), while the
integral Z R
ρ0 4πr2 g (r) dr (1.3)
0
is the number of atoms in a sphere of radius R; for example, the average number of nearest neighbours,
or coordination number, of an atom can be estimated from the area under the first peak of the RDF (see
figure 1.7).
Note also that the limit of the RDF for no correlation, at longer distances, is not a constant value as
in g (r) but the parabola 4πr²ρ0 (compare figure 1.9 to figure 1.1).
CHAPTER 1. INTRODUCTION 10

Figure 1.6: (From Varshneya (1994)) schematic illustration of the configuration of a glass around an
origin atom (in black) with the corresponding g (r)
CHAPTER 1. INTRODUCTION 11

Figure 1.7: radial distribution function 4πr²ρ0 g (r). The number of nearest neighbours n can be evaluated
from the area under the first peak

Figure 1.8: number of atoms in a spherical shell of radius dr

1.2.4 extension to binary systems

In a system with two components, A and B, partial distribution functions can be defined. If

probability of finding a B atom in volume dr

hnAB (r)i dr =
placed at distance r from a given A atom
hnAB (r)inonc dr = the same, in an uncorrelated system

then the partial distribution function gAB (r) is

hnAB (r)i hnAB (r)i

gAB (r) ≡ = (1.4)
hnAB (r)inonc ρB
CHAPTER 1. INTRODUCTION 12

Figure 1.9: sketch of RDF in different systems

Note that

gAB (r) = gBA (r) (1.5)

In fact, the number of A-B and B-A contacts with a distance value between r e r + dr must be the same:

ρA hnAB (r)i 4πr2 dr = ρB hnBA (r)i 4πr2 dr

From this equality, 1.5 can be recovered through the definition of gij (r).
Coordination numbers of B atoms around A atoms and vice versa are proportional to their respective
densities. For example, the number of B atoms whose distance from an A atom is less than R is (compare
1.3) Z R
NAB = ρB 4πr2 gAB (r) dr
0
and similarly Z R
NBA = ρA 4πr2 gBA (r) dr
0
Now recalling that gBA = gAB we verify that
NBA NAB
=
ρA ρB

1.2.5 information contained in g (r)

The function g (r) carries information about distances between two atoms; it doesn’t say anything about
the mutual positions of three atoms, such as the angle they form.
This can be illustrated by an example (see figure 1.10). Suppose we have a binary AB compound and
we know, from g AB (r), that atom A’s first and second neighbors, B and B 0 , are placed at distances r1AB
and r2AB ; note that B 0 may be anywhere on a spherical surface of radius r2AB .
We know, additionally, that there is a second B atom also at a distance rBB from B, where rBB is the
first peak in g BB (r). However, in no way can we know whether this atom is the same as B 0 ! it could
as well be in position B III while A’s second neighbour is in B II , see the figure. Only if B III e B II are
the same, triangle ABB’ is univocally determined, including the angle ABB \0 .
Careful analysis of the peaks in the PDF will help here:
CHAPTER 1. INTRODUCTION 13

B'

r2
rBB
B II r2 BIII
r1 rBB
A B
Figure 1.10: mutual positions of three atoms A, B and B I (or B II , or B III ) with first neighbour distances
r1 , r2 , and rBB

• if B II 6= B III , there must be a third peak in the g AB (r) at a distance between r2 and r1 + rBB

-not at a shorter distance than r2 , since in this case it would have appeared before!- and a second
peak between rBB e (r1 + r2 ) in the g BB (r) (both of which with coordination number 1).

• if the three atoms are equal and B III e B II coincide in B 0 , then A,B and B 0 form an isosceles
triangle with sides r1 , r1 , r2 where the “bonding” angle BAB
\0 = θ is given by

θ r2
sin =
2 2r1
(Elliott (1984) p.63, table at p.74; Rao et al. (1998))

• similarly, if r2 = r1 , with NAB = 2 (such as in an AB2 molecule), the bonding angle BAB
\0 = φ is
given by:
φ rBB
sin =
2 2r1
CHAPTER 1. INTRODUCTION 14

1.2.6 geometry of the atom surroundings (Voronoi polyhedra)

The g (r) gives information about the average distances from an origin atom, but, due to the assumption
of (average) spherical symmetry, it does not say anything about the shape of the atom surroundings.
However, even in a disordered material, neighbouring atoms may be arranged following preferential
geometries; for instance, in many metal alloys it has been observed that the first contacts tend to assume
an “approximately” icosahedral shape.
There are different ways of partitioning the space in a disordered arrangement of atoms (figure 1.11,
left), into geometrical figures (polyhedra) that add up to the whole space. A natural choice is considering
the polyhedra whose vertices are the atoms themselves and whose edges are the connection lines between
atoms; this is what is commonly done when representing crystal structures. If we add the requirement
that these elemental volumes contain no atoms, we obtain a “lattice” like the one illustrated by red lines
in the right panel of the same figure 1.11

Figure 1.11: Left: a disordered array of points. Right: the same points connected in a lattice (red lines),
and their Voronoi polyhedra (blue dashed lines)

Note that the choice of these minimal polyhedra is a purely geometrical abstraction and does not
imply the existence of real (chemical) bonds between connected atoms. Indeed, the choice is to some
extent arbitrary, like the choice of a primitive cell in crystallography.
There is an alternative method of partitioning the space, in which the atoms are placed at the center
rather than at the vertices of the polyhedra. This partitioning is illustrated in two dimensions by the
dashed blue lines in the figure. The elemental units of space delimited by the blue lines are called
Voronoi polyhedra and are defined as follows:

the Voronoi polyhedron of an atom is the locus of all points that are closer to that atom
than to any other

The recipe for drawing the Voronoi polyhedron of atom A is

• draw all the lines connecting A to its neighbours

• cut each of these lines in the middle point with the normal plane
CHAPTER 1. INTRODUCTION 15

• the planes delimit a convex figure around A that is its Voronoi polyhedron3

The Voronoi subdivision of space finds an ever-increasing number of applications. Among the most
obvious: associating each point in a map to the nearest fire station!
As a direct consequence of its definition, the Voronoi partition, or tessellation, of an ensemble of
points is unique. As we have seen, this is not so for the lattice that connects atoms; however, a unique
lattice can be derived from the Voronoi partition4 with the so-called Delauney triangulation:

in the Delauney lattice two atoms are connected if they share one face of their Voronoi
polyhedra

The red network in figure 1.11 (right) is the Delauney lattice of the points.
Clearly, the Delauney and Voronoi partitioning are duals of each other.
It can be shown that in 2D the cells of the Delauney lattice are triangles (hence the term “Delauney
triangulation”), while in 3D the cells are tetrahedra; Voronoi cells, on the contrary, can be any kind of
polygons and polyhedra, respectively. The vertices of the Voronoi polygons (polyhedra) are the circum-
centers of the Delauney triangle (tetrahedron) that contains them; thus, each Delauney cell contains only
one Voronoi vertex, in the same way as each Voronoi cell contains only one vertex of the Delauney lattice
(one atom).
Given a certain array of atoms, Voronoi cells describe the geometry of the space around atoms;
Delauney cells describe the dimension of the empty spaces, or holes, between atoms.
If the array has translational symmetry, as a crystal lattice, Voronoi polyhedra are termed Wigner-
Seitz cells: they are primitive cells whose symmetry properties are the same as the non-translational
properties of the lattice (for example, the dashed hexagons in the regular triangular lattice in figure 1.12).

Figure 1.12: (from Zallen (1998)) A regular hexagonal 2D lattice of atoms (the circles, or the nodes of
the heavy line network). The dashed hexagons are Voronoi cells, called in this case Wigner-Seitz cells

In a disordered array, instead, there is a whole ensemble of different Voronoi polyhedra. They can be
grouped according to several properties, such as their volume, number of faces, number of edges per face,
3
to understand this, consider that all points on the same side of the plane as A are closer to it than to the neighbour
4
in fact, in special cases, called degenerate, the lattice is not uniquely determined
CHAPTER 1. INTRODUCTION 16

etc. In a totally disordered system (figure 1.13, left) these properties are distributed over a wide range of
values; partially disordered systems (figure 1.13, right) exhibit less diversity. In figure 1.14 distributions
of Voronoi cell volumes for a simulated solid, liquid and gas are reported.

Figure 1.13: Voronoi polyhedra of a highly disordered system (left) and in a less disordered system (right)

Figure 1.14: (from Montoro and Abascal (1993))

The next figure (figure 1.15) shows the statistics of the number of faces and number of edges per face
of Voronoi polyhedra in a random close-packed structure (which will be treated in detail in section 3.3).
CHAPTER 1. INTRODUCTION 17

The number of faces of the Voronoi cell of an atom represents the number of adjacent cells, which can
be taken as definition of the number of neighbours of the atom. Note that in real aggregates where atoms
have non-zero radius, these geometrical neighbours (the atoms whose Voronoi polyhedra are in contact)
may be not the same as the first neighbours (atoms that are themselves in direct contact).

Figure 1.15: Frequency distribution of number of faces and number of edges per face of Voronoi polyhe-
dra of a simulated random close-packed structure

1.2.7 more structural parameters

In addition to the g (r) and the Voronoi analysis and pair correlation quantities, such as the coordination
numbers, that can be derived from them, there are other quantities, useful in the description of glass
structure, that can be obtained from computer simulations and -at least in principle- from experiments.
We briefly review them here, in order of atom correlations involved (2, 3, and more atoms); some will be
discussed in more detail below.

1.2.7.1 atom connectivity

In many glassy substances, there is a network of chemical (covalent) bonds connecting neighbouring
atoms. In some cases (silicate glasses, chalcogenide glasses) the number of bonds that one atomic species
can form is variable. For example, in silicate glasses, oxygen atoms can link to two silicon atoms, thus
acting as bridges between them, or can carry a negative charge and be bound to one silicon atom only:
this is illustrated in figure 1.16, a 2-D reduction of a sodium silicate glass . In real space silicate glasses,
of the four oxygens surrounding one silicon atom, the number of “bridging” atoms ranges from zero to
four.
CHAPTER 1. INTRODUCTION 18

Figure 1.16: (from Varshneya (1994)) 2-D example of bridging and non-bridging atoms in a sodium
silicate glass
CHAPTER 1. INTRODUCTION 19

1.2.7.2 bond angles

As we have seen, structure models that respect observed atom distances can differ as to bond angles.
Figure 1.17 shows two models for silica glass that predict a median Si − O − Si angle of ∼ −180◦ and
∼ −140◦ , respectively.

Figure 1.17: (from Keen (1998)) Two different Si − O − Si bond angle distributions (circled in red) in
silica glass
CHAPTER 1. INTRODUCTION 20

1.2.7.3 medium-range structures (rings)

Chemical bonds in a covalent glass form ring structures of varying sizes. The relative abundance of rings
composed of n atoms (see figure 1.18) is a property that can be evaluated with computer simulations and
in some cases be inferred from experimental data.

Figure 1.18: ring size statistics (lower left histogram) in a 2-D covalent glass (from Zallen (1998))

1.3 reference texts

Two texts, Varshneya (1994) and Shelby (2005), although a little outdated, are still considered standard
textbooks in the general subject of glasses from an experimentalist’s point of view; specific chapters are
devoted to structural properties and to methods for the description of glass structure. A more theoretical
description, not limited to glasses but covering the broader subject of disordered matter, is in Zallen
(1998).
Chapter 2

Methods for the investigation of glass structure

In this section we will briefly review the methods for analyzing the structure of a glass. They are di-
vided into experimental and computational methods; to each one will be devoted one of the following
subsections.

2.1 Experimental methods

Traditionally, the techniques of choice for investigating the atomic structure of substances have been
techniques based on diffraction. There is a one-to-one correspondence between the prevalence of this
kind of measurements and the prevalence of g (r) as structure descriptor; in fact, diffraction methods
probe interactions between pairs of scattering centers, so the main output is a pair distribution function.
Other experimental techniques are based on the response produced by three or more atoms, so they
provide information about more complex atomic environments; we will touch vibrational and NMR
spectroscopy.

2.1.1 methods based on diffraction

In a typical diffraction experiment, the incident beam of particles (the wave) is scattered by individual
centers of the sample with different phases; the scattered rays interfere with each other, either positively
or negatively, and the resulting wave is collected at some angle with the incident beam by a distant
detector. The signal at a certain angle depends on the phase difference between scattered waves, which
in turn depend on the angle and the atomic distances.
First, a review of essential quantities.
A plane wave travelling in the k direction has the equation:

A = A0 sin (k·r − ωt)

or, more generally,

A = A0 ei(k·r−ωt)

21
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 22

where

ω = frequency
k = wave vector
2π
|k| =
λ
ω
= velocity
|k|

Energy and momentum of an electromagnetic wave are

E = }ω
p = }k

and are connected by the velocity expression ω

|k|
= c.

Interference of waves of equal frequency. The interference of two waves of equal frequency is con-
veniently evaluated with the rotating vectors method (see for example Alonso and Finn (1982)):

Two waves of equal amplitude, A0 , and frequency, ω, but different spatial phase kr1 , kr2 , with

δ = k· (r1 − r2 )

superimpose giving rise to a wave of amplitude

p
ξ0 = A0 2 (1 + cosδ)

The amplitude has a maximum (constructive interference) for cos δ = +1, that is for

δ = 2nπ (2.1)

with n = 0, ±1, ±2, ...

CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 23

Geometry of the scattering experiment. We want to study the interference between the wave scat-
tered by an origin center O and a generic point P, when collected at an angle θ.
Let
}ki ,}kf = momentum of incident and final (scattered) wave
The phase difference between the wave scattered by O and P is
k i · RP − k f · RP = Q · RP (2.2)
with
Q ≡ ki − kf

For better clarity, in the following figure, the two spatial phases ki · RP , kf · RP are marked by
colors:

If scattering is elastic, kinetic energy of the incident wave is conserved, and so is the modulus of its
momentum (since |k|ω
= c; see above)
|ki | = |kf | = k
so that for a given k the modulus of Q depends only on the scattering angle:
 
θ
|Q| = Q = 2k sin
2
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 24

and substituting λ for k (recall k = 2π

λ
):
 
4π θ
Q= sin (2.3)
λ 2

The latter is the modulus of the scattering vector at angle θ.

The intensity will be maximum at angles that realize the constructive interference condition on the
phase difference (equation (2.1)), that is,

Q · RP = 2nπ

It is useful to write Q as its modulus times its direction:

 
4π θ
Q= sin uQ
λ 2

(where uQ is a unit vector in the direction of Q); now the condition becomes
 
4π θ
sin uQ · RP = 2nπ
λ 2
or
nλ
uQ · RP = (2.4)
2 sin 2θ

Maximum constructive interference with radiation scattered by O is given by points whose distance
from O, projected in the direction of Q, equals 2 sinnλ θ
(2)
The maximum interference condition is illustrated in the following figure:
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 25

kf

Q
θ
P’

d2 P"

d1

Figure 2.1: scattering planes are a manifold of parallel planes spaced d = λ

2 sin( θ2 )

All points (P , P 0 , P ” ) lying on the series of planes ⊥ Q and whose distance from O is dn = nλ
2 sin( θ2 )
contribute to the maximum interference.
A few comments on the important relation of 2.4.

1. Bragg’s law. Equation (2.4) is just a different form of the well-known Bragg’s law of crystal
scattering
 
θ
2d sin = nλ
2
which is illustrated in standard textbooks with sketches like the following1 :

Figure 2.2: Bragg’s law (from Elliott (1984))

1
note that in this figure θ
2 →θ
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 26

However, similar representations are somehow misleading, since they do not show that positive inter-
ference with the point on the lower plane is given by any point on the plane above, not just the one that
is on the perpendicular: compare our figure 2.1.

2. Measurement resolution. The smallest distances that can be measured with a diffraction method
are of the order of the spacing between planes
λ
d1 = θ


2 sin 2

The point on the first plane that is closest to O is the one for which RP is closest to the direction of
Q, so that uQ ·RP ∼ |RP |; if the scattering angle is not too small, that is sin 2 ∼ 1, we have
θ

|RP | ∼ λ

which shows that the wavelength λ of the radiation involved must be of the order of |RP |, the min-
imum distance between two centers that we want to detect (of the order of Å -like X-rays- for the short
range atomic distances).

3. Angle dependence. Note also that at constant λ the angle θ of the first diffraction peak and the
distance |RP | between the two centers are inversely proportional:
 
θ λ
sin ∼
2 2 |RP |
that is, interference between centers at a larger distance than λ can be seen at small angles
In Small Angle X-ray Scattering (SAXS), for example, X-rays (λ ∼ 10−10 m) are used to detect
particles of dimension scale of nm at angles of the order of 1◦
We now consider the cumulative effect of interference between many (pairs of) centers.
When there is an ensemble of scattering points P, each one interferes with the radiation scattered
by O, and the total radiation is the sum of the single superimposed waves. The resulting intensity is
proportional to the square of this amplitude:
2
X
I (Q) = A fP (Q) exp (iQ · RP )
P

where

fP (Q) = scattering length of a single site

(and A is some proportionality constant).

The intensity depends on

• the type of incident wave

• the type of scattering center

• the direction of scattering

CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 27

2.1.1.1 X-ray scattering (XRS)

X-rays are scattered by electrons in matter. Within a good approximation, in a polyatomic sample (made
up of heavy atoms) the electronic density is concentrated in the close vicinity of the atomic nuclei. This
allows us to first consider scattering from an isolated atom, then sum up all single atom contributions.
The derivation we will follow is, to some extent, general.
The intensity of radiation scattered by one atom, expressed as a function of the radiation scattered by
one electron (“in electron units”), is
Z 2
Ieu (Q) ∝ ne (r) exp (i Q· r) dr

with

ne (r) dr = fraction of electron in volume dr

In a system of N atoms, contributions from each atom must be cumulated. If we define:

Ri = position of atom i (nucleus)

r0 = distance within the atom (from nucleus)

it can be shown (Cusack (1987)) that (for a collection of equal atoms) the resulting intensity can be
written

N 2
X
Ieu (Q) = exp (i Q· Ri ) · |fa (Q)|2 (2.5)
i=1

with
Z
fa (Q) ≡ ne (r 0 ) exp (i Q· r 0 ) dr 0 (2.6)
atom

representing the scattering from a single atom.

N.B.: By means of equation (2.5), which is general, Ieu (Q) is partitioned into two factors, of
which one depends on atoms and one on their positions.
Accordingly, one can consider each factor separately.
The factor dependent on atoms, or atomic form (or scattering) factor fa (Q), expresses the total
interference between infinitesimal elements of the space around one atom; it is known and tabulated for
each atomic species and for each type of incident wave.
In the X-ray case, electronic density of the atom has an approximately spherical symmetry, and thus
the atomic form factor does not depend on the direction of the incident wave with respect to the atom,
but only on the scattering angle:
 
sinθ
fa (Q) → fa (Q) = fa
λ
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 28

Clearly, fa is larger for heavier atoms (which

contain more electrons), while the hydrogen atom
1
H gives a negligible scattering. At θ = 0 there
is total scattering and fa is simply the number of
electrons of the atom(recall equation (2.6) with
Q = 0)
We may note that fa (Q) has a roughly Gaus-
sian shape; this is not unexpected, since equa-
tion (2.6) defines it as the spatial Fourier transform
of electron density which is approximately Gaus-
sian itself2 .
As for the factor that contains atom positions,
it is called the structure factor S (Q)

N 2
X
N S (Q) ≡ exp (i Q· Ri ) (2.7)
i=1

It depends on the distribution of interatomic distances (Ri − Rj ).

It is therefore reasonable to expect it to be connected to g (r). Indeed, it can be proved (Cusack
(1987)) that its thermodynamic average is
Z
hS (Q)i = 1 + ρ0 (g (r) − 1) exp (i Q· r) dr

In an isotropic medium, such as a liquid or a glass, g (r) → g (r) so S (Q) → S (Q). Omitting
brackets hi for clarity: Z ∞
sin Qr
S (Q) = 1 + ρ0 (g (r) − 1) 4πr2 dr
0 Qr
The last formula shows that S (Q) is directly connected to the Fourier transform of g (r), or that of
the radial distribution function ρ0 4πr2 g (r)

2.1.1.2 neutron scattering (NS)

Neutron beams from a nuclear reactor have a wavelength of 0.1 ÷ 1 Å, which is suitable for atomic
structure investigation, like X-rays.
Neutron scattering (NS) can be treated in much the same way as XRS (Cusack (1987)), with some
important differences.
First, neutrons are scattered by nuclei (as opposed to electrons); so it is the atom positions that will
be detected, rather than the (centre of) electronic density.
Differently from XRS, the atomic form factor for NS, named b, is isotropic and does not depend on
Q (see figure 2.3). The atomic form factor, indeed, is the Fourier transform of a Dirac δ function in this
case, and consequently it has a constant value (compare the corresponding argument for X-rays).
Also, b does not increase with the atomic number Z of the element: light nuclei scatter as efficiently
as heavier ones. This is a major difference from XRD, and explains why when we are interested in the
position of small atoms, such as hydrogen atoms, which are not seen in XRS, we must resort to NS.
2
see for example Bracewell (1986)
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 29

Different isotopes of the same element can have different form factors; this fact is exploited in the
technique of isotopic substitution, by which binary pair distribution functions, gAB (r) , are obtained.

Figure 2.3: (from Cusack (1987))

Inelastic neutron scattering, where some energy and momentum are transferred to the substance, can
be used for studying its vibrational states. In this case the experimental quantity of interest is

S (Q) → S (Q, ω)
where
~ω = energy lost to the system
From S (Q, ω) the vibrational density of states can be measured. Also, it gives information about the
time relaxation of the structure, which is an important topic in the glass transition.

2.1.1.3 Electron scattering (ES)

Electrons from an electron microscope, for which λmin = 0.05Å, are also employed for scattering.
Electrons interact with both nuclei and electrons of the sample. One way to see this is that they
are scattered by the screened Coulomb field of the atoms, which is reflected in the electronic atomic
scattering factor fe (Q) being connected to the X-ray atomic form factor fa (Q) (Cusack (1987)):

Z − fa (Q)
fe (Q) ∝
Q2
Because of the strong interactions with the charge density of the sample, fe (Q) > fa (Q) (see fig-
ure 2.3 again)
Strong Coulombic interactions also prevent electrons from reaching the inner parts of a substance.
Thin samples must be prepared to be examined with ES.
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 30

2.1.1.4 Comparison of different scattering methods

A useful comparison of the different scattering techniques we have just mentioned is reported in table 2.1
below:

Table 2.1: Advantages and disadvantages of different scattering techniques (from Rao (2002))
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 31

2.1.1.5 Examples
X-Ray Scattering is the method of choice for studying monoatomic liquids or glasses:

Figure 2.4: (from Cusack (1987))

With Neutron Scattering, partial structure factors of binary compounds can be obtained, using isotopic
substitution (figure 2.5)
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 32

Figure 2.5: (from Cusack (1987)) Neutron scattering of liquid NaCl with isotopic substitution. (a) from
top: Na35 Cl,NaCl,Na37 Cl; (b) partial structure factors

Partial structure factors give the binary PDFs (figure 2.6)

Note that layers of positive and negative charge alternate; and that g (r)Na−Na superimposes with
g (r)Cl−Cl

Figure 2.6: (from Cusack (1987)) PDF of liquid NaCl from NS with isotopic substitution. Solid line:
g (r)Na−Cl . Broken: g (r)Na−Na . Dotted: g (r)Cl−Cl
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 33

In the case of BaCl2 , instead, the Ba − Ba curve differs from Cl − Cl3 :

Figure 2.7: (from Cusack (1987)) PDF of liquid BaCl2 .

As we have seen, coordination numbers can be estimated from the Radial Distribution Function
RDF= ρ0 4πr2 g (r), equation (1.3). The number of atoms contained in a sphere of radius D is given by
RD
ND = ρ0 0 g (r) 4πr2 dr. In vitreous silica (figure 2.8), the area under the first peak gives NSi−O ≈ 4.
3
The different first-neighbour peaks might well be due only to the different Cl and Ba densities. Assuming complete
disorder, since ρCl = 2ρBa the volume that contains 1 neighbour is vCl = 21 vBa and consequently the corresponding sphere
1
radius is rrBa
Cl
= 2 3 = 1.26, and this is the ratio of average distances. From the g (r) curves one can compute rrBa
Cl
≈ 45 = 1.25.
Does this make any sense to the reader?
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 34

Figure 2.8: (from Cusack (1987)) RDF ρ0 4πr2 g (r) of vitreous SiO2 from NS.

The RDF can be extracted also from Electron Scattering. Like XRS, and contrary to NS, ES gives no
access to partial structure factors (figure 2.9)

Figure 2.9: (from Rao (2002)) RDF ρ0 4πr2 g (r) of vitreous B2 O3 from ES.
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 35

2.1.2 NMR and vibrational spectroscopy

We mention here, for completeness, two other basic experimental methods that have been traditionally
used in the investigation of the structure of glasses: NMR and optical vibrational (IR) spectroscopy. The
physical principles and experimental setup of apparatus of each are treated in standard physical chemistry
textbooks. Here we will only show some examples of how they are exploited for deriving other glass
structural parameters than g (r), such as bond angles or rings.

2.1.2.1 NMR
In a nutshell:

• Nuclei that are active in NMR (13 C, 31

P, 29
Si, 77
Se, ...) have a magnetic momentum

• Interaction of this momentum with an external magnetic field, and with momentum of neighbouring
atoms, causes a splitting of energy levels which is probed with electromagnetic radiation

• Chemical shifts, which represent the perturbation due to neighbouring atoms, give information
about the molecular environment of active nuclei.

• In a liquid or disordered solid an isotropic shift, an average scalar value, is observed

E. g. in silica the hexacoordinate group SiO6 → −180 ÷ −220 ppm; while the tetracoordinate SiO4 →
−65 ÷ −120 ppm (Source: Leonova (2009)).
The following slides have been borrowed from a seminar by Micoulaut (2013)
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 36

2.1.2.2 Vibrational spectroscopy

There are two main types of optical vibrational spectroscopy:

• InfraRed (IR) is due to the interaction between electromagnetic radiation and the change in the
electric dipole of a molecule associated to a vibration

• Raman is due to interaction between electromagnetic radiation and the change of molecular polar-
izability associated to a vibration

Both techniques reveal the nature and geometry of chemical bonds in the system. Molecular vibrations
commonly involve a small group of neighbouring atoms, the arrangement of which can be inferred from
the spectra.
Again, we borrow a few examples from Micoulaut (2013):
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 37
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 38

2.1.3 Reference texts for section §2.1

This section is based on Cusack (1987), chapter 3 “Investigation of disordered structures”, with addi-
tions from Rao (2002), chapter 4 “Structural Techniques”. Examples are taken from the same books;
from a lesson by Matthieu Micoulaut (Micoulaut (2013), lesson 2); and from Leonova (2009). A good
description of diffraction geometry -though one limited to crystals- is also in Alonso and Finn (1982),
section 23.8.
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 39

2.2 Computational methods

2.2.1 Empirical force fields
The simulation methods whose main features we will describe in the next two sessions4 share a common
foundation: a reasonably accurate, and computationally accessible, representation of the forces acting on
particles.
When modelling the interaction field in a solid, two approaches can be taken:

• The system can be treated within the framework of quantum mechanics (QM), using ab initio
or semiempirical methods. Although these are the most accurate methods available, they are ex-
tremely demanding in terms of computational resources; the more so, if the simulated system is
made up a large number (tens of thousands at least) of atoms, as is necessarily the case with an
intrinsically disordered system like a liquid or glass.

• Or, the interactions can be represented in a simple analytical form, which is a function of the par-
ticles’ classical dynamical variables, coordinates and velocities, and contains partially adjustable
parameters that can be fitted to ab initio results and/or to measured properties. In addition to
being less resource-consuming, this empirical field approach is also more readily understood by
researchers, who can picture the meaning of the functions and parameters used and possibly adjust
them.

It is useful to review some aspects of the latter approach, the empirical field.
If we consider two rare gas atoms, their interaction can be modelled with a function depending on
the (scalar) interatomic separation r, that approaches infinity at small distance, and whose limit is zero
at high values of r (figure 2.10). There is an energy minimum at the atom-atom contact distance rAB ∗
;
for r > rAB the force is attractive, while it is increasing repulsive for r < rAB . There are a number of
∗ ∗

analytical forms for this, for example the Lennard-Jones function

  
σ 12  σ 6
V (r) = 4 − (2.8)
r r

where  is the minimum depth and σ corresponds to the zero of the potential energy function (see fig-
ure 2.10). (Also sketched in the figure is the oversimplified hard sphere model, mainly employed in
theoretical models, which includes no attraction term).
The parameters in equation (2.8) depend on the atomic species involved. To each couple of atom
species (i, j) a set of parameters is associated: in the Lennard-Jones expression (ij , σij ).
4
but not the Reverse Monte Carlo method that will be presented at the end of the chapter
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 40

Figure 2.10: atom-atom interaction potential (Cramer (2004))

In systems made up of molecules, single atoms interact in different manners. It is convenient to

consider interactions between bonded and non-bonded atoms separately: the former include internal
bond distances, angles, and torsions (dihedrals); the latter act between atoms either belonging to different
molecules, or belonging to the same molecule but being placed more than 3 bonds apart. Non-bonded
interactions are both of dispersive-repulsive (Van der Waals), and electrostatic nature.
The analytical expression may be, for example

where the first three terms account for bond distances, angles and torsions; the double sum in the second
line is extended to all “non-bonded” atoms, and contains the Van der Waals potential (compare equa-
tion (2.8)), and the electrostatic potential.
Bond distance and angle potentials have a harmonic form. The torsional term has the following
behaviour:
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 41

Many glass-forming materials (e.g. AX2 with A = Si, Ge e X = O, S, Se, Te) do not feature separate
molecules, but instead an extended network of A − X bonds with mixed ionic-covalent character.
In purely ionic compounds only non-bonded interactions are in effect; the potential energy contains
only the three terms electrostatic, repulsive and dispersive, where the repulsive term can have an exponent
n < 12:
N X N  
X qi qj A C
V = + n− 6
i=1 j=i+1
4π 0 rij rij rij

In a purely covalent solid, bonded interactions dominate; however in order to model such a solid
with bonded interactions only, one must know a priori the complete, detailed map of interatomic bonds,
which obviously is not available in a disordered system.
Thus, in mixed ionic-covalent glasses a non-bonded potential is used, that allows for bond forming
and breaking; it may include, in addition to the pair term, a three-body term which represents the non-
harmonic angle interaction:
N X
X N N X
X N N
X
V = Vij (r) + Vijk (r, r0 , θ)
i=1 j=i+1 i=1 j=i+1 k=j+1
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 42

Example: simulations of SiO2 in different crystal phases and in the glass phase (Vashishta et al. 1990)
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 43

As you can see from the example above, the potential parameters like qi , σij , ij , A, C, etc., are not
specific of every single atom but, in order to be physically meaningful, they are a property of an atomic
species of species pair (Si, O, etc.):
As a rule, such parameters cannot be safely exported from a substance to another; in different sub-
stances they have similar, but not identical, values. These values are fitted to ab initio calculations and
measured properties.

2.2.2 Monte Carlo simulation

In a Monte Carlo (MC) simulation, a statistical sample of the system’s microscopic configurations is
collected in accordance with the thermodynamical state of the system, in order to compute average prop-
erties.

2.2.2.1 Thermodynamical averages

The basic principle of statistical mechanics is that the observed system is an average over the statistical
ensemble of all microscopically defined copies of the system that are compatible with the macroscopical
variables, such as volume and temperature, that define its thermodynamical state.
For example, a closed system at constant V and T is represented by the so-called canonical ensemble
(or “(N, V, T )” ensemble):

Figure 2.11: The canonical, or (N, V, T ), ensemble); (β = 1

kB T
)

In these thermodynamical conditions, the total energy of the system is not fixed, but is exchanged
between the system and the thermostat. The probability of seeing microscopic state i with energy Ei is
 
Ei
pi ∝ exp − (2.9)
kB T

and the average of a macroscopical quantity A whose value in state i is Ai is the weighted average
 
Ei
P
i Ai exp − kB T
P
i pi Ai
hAi = P = P   (2.10)
i pi exp − Ei
i kB T

For example, in a perfect gas the total energy hEi = 32 N kB T

In general, hAi can be computed by correctly sampling the ensemble of states i.
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 44

2.2.2.2 How MC works

The steps in a MC simulation can be summarized as follows

1. setup:

• prepare a conveniently sized system of N particles in a container of volume V (typically,

N ∼ 104 atoms), by assigning atom positions in an acceptable configuration
• select a force field expression and parameters (for example 2.2.1)
• compute the initial energy E0

2. randomly generate a new configuration (new atom positions);

3. compute the energy Ei of the new configuration and compare it to the previous value of energy

4. add the new configuration to the statistical sample only if it meets a predefined criterion (usually,
the Metropolis criterion), chosen so that the sample will build up the correct statistics 2.9;

5. repeat from step 2 until the desired number of steps is completed. The sample is now ready for
computing averages.

A critical point in this scheme is how the new configurations are generated. Although this may be
not immediately apparent to the reader, only an extremely small fraction of the theoretically possible
configurations give a nonnegligible contribution to the statistics. To give a feeling of what this means:
it can be shown (Frenkel and Smit (2002)) that in a system of 100 hard spheres near the freezing point
only one out of 10260 configurations has a finite (that is, not infinite) energy and has a nonzero weight.
Clearly, a device is needed to avoid wasting all our time in generating instances that will be discarded
anyway. The technique of sampling an ensemble more efficiently in those regions that are statistically
more important is called importance sampling.
The basic idea in MC is that, due to the fact that the potential energy is a well-behaved function of
the particle positions, if one starts with an acceptable configuration, not far from the energy minimum
which is the probability maximum, and then moves by a “small” step (see figure 2.12), the resulting
state will probably have a finite probability too. The step should be small enough to keep the system in
an acceptable zone, yet it should be large enough to ensure an efficient sampling of the configurational
space.
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 45

Figure 2.12: A Monte Carlo move (from Frenkel and Smit (2002))

A well-known, and very effective, representation of such random walk sampling technique is depicted
in figure 2.13: if one needs to measure the average -say- pollutant content of river waters, it is best to
sample it along the river, not on a regular grid!

Figure 2.13: (from Frenkel and Smit (2002))

2.2.2.3 Energy barriers and the cooling process

The above scheme works reasonably well as long as the system is capable of exploring all the important
configurations. A critical situation, however, is when the energy profile of the system features two minima
separated by a high barrier (figure 2.14): in this case, if started from the local, as opposed to the global,
minimum (the one to the left in the figure), it may not have the chance of reaching the global minimum
(right), as this would involve passing through intermediate steps at high energy and low probability
which will be rejected by the algorithm. Note that both the length scale of the move and the value of
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 46

temperature are key elements here; the lower the T , the more unlikely it will be for the simulation to
traverse the barrier (recall the probability 2.9).

Figure 2.14: Potential energy with two wells separated by a barrier

Consider the cooling process of a liquid. One can think of the potential energy “landscape” as a
roughly shaped profile crowded with local minima, with the crystal corresponding to a narrow and deep
global minimum (see figure 2.15). At high temperature (left panel in the figure), the system’s kinetic
energy allows it to explore the whole configurational space; obviously, “high” temperature means that
kB T is large with respect to the depth of the minimum. On cooling, the system is gradually confined to
the lower-energy portions of the space; until, at the melting temperature, only the crystal configuration
remains accessible. If cooling is instantaneous, however, the system will get “trapped” in one local
minimum close to its momentary configuration (any one of the partially filled wells in the right panel): it
will then exhibit solid properties (restricted displacement), but be freezed in a liquid-like structure - that
we now know is exactly the state of a glass.

Figure 2.15: a system with a “ragged” potential profile and the regions of visited configurations at high
and low temperature (from Royall and Williams (2015))
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 47

In MC, due to the dimensional limits of the simulation, configurational trapping is routine, and special
techniques have been devised in order to constantly keep the system capable of exploring all of the
configurational space (see for example Rauscher et al. (2009)). The bottom line is that when simulating
the cooling of a liquid substance with MC it is extremely likely that it will transition to a glassy solid,
although not necessarily the experimental glass.

2.2.2.4 computation of g (r)

As an example of how an observable quantity is computed in MC, let us consider the cross pair distribu-
tion function of a binary substance 1.4 :
hnAB (r)i
gAB (r) =
ρB
The two quantities are evaluated from the MC ensemble averages:

T N
PA
1 1
NitAB (r, ∆)
P
AB T NA
n (r, ∆) t=1 i=1
gAB (r) = =
nAB (r, ∆)unc ρB 4πr2 ∆
with
NitAB (r, ∆) = number of type B atoms whose distance from
A-type atom i is between r and r + ∆, at simulation
step t of T

NA = total number of type A atoms

ρB = number density of type B atoms

Some typical results are shown in figure 2.16

Figure 2.16: g AB (r) in amorphous GeS2 at 300 K, from MC simulation (ISAACS 2013)

2.2.3 Molecular Dynamics simulation

As the name suggests, with Molecular Dynamics methods the dynamical evolution of the system is
simulated; this is done by numerically solving the system’s equations of motion.
We will only consider MD in classical dynamics.
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 48

2.2.3.1 Phase space averages

In classical mechanics, the state of a many-body system is defined by the value of continuous variables
momentum (p) and position(q) of each particle. The domain {p, q} of these variables is a hyperspace
called phase space. Energy, that in a quantum mechanical framework would take discrete values Ei , is
classically a continuous function of the phase space variables:

Ei → E (p, q)

Accordingly, the average of some observable A (which depends on the dynamical state) in the
(N, V, T ) ensemble is  
R E(p,q)
A (p, q) exp − kB T dpdq
hAi =  
exp − E(p,q)
R
kB T
dpdq
Compare the last equation to equation (2.10). More generally, if the probability of phase space point is
P (p, q), the phase space average of A is
R
A (p, q) P (p, q) dpdq
hAi = R (2.11)
P (p, q) dpdq
In MD, the aim is to follow the system’s evolution, or trajectory in phase space, for a long enough
time to allow it to explore all the accessible regions of phase space with the correct probability. System
“snapshots” (positions and velocities) are recorded at some defined interval, and phase space averages
(2.11) are computed from the recorded trajectory.

2.2.3.2 How MD works

In classical physics, positions and momenta (or velocities) of an isolated collection of particles can be
known at any time t if we know their values at the initial time {pi (0) , q i (0)}and the force acting on each
particle i, which in turn, in conservative systems, is the gradient of the potential energy:

F i = −∇i V
In a Molecular Dynamics simulation the potential energy V is the sum of particle interactions, ex-
pressed with a formula like (2.10). The equations of motion, which are differential equations, are solved
numerically, with the help of approximate methods in which the time differential is replaced by a small
time step ∆t; for example (Euler’s approximation):
pi (t)
q i (t + ∆t) ≈ q i (t) + · ∆t
mi
pi (t + ∆t) ≈ pi (t) + mi ai (t) · ∆t

where
F i (t) = mi ai (t)
It must be noted that in a Molecular Dynamics simulation the total energy E of the system is con-
served. Thus, it may be viewed as a “Monte Carlo” simulation in which the importance sampling method
is the dynamical trajectory of the system and the statistical ensemble is the (N, V, E), or microcanonical,
ensemble.
Typical dimensions of a MD simulation are:
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 49

number of atoms ∼ 1000

time step (∆t) ∼ 10fs
total simulation length 1ns ÷ 1µs

2.2.3.3 Non-ergodic systems

If the simulation is carried for a sufficiently long time, the system can be assumed to have sampled all
phase space. In this case

time average = statistical average

where the time average is

M
1 X
A= A (ti )
M 1
or, in the limit of continuous sampling of variable t:

1 T
Z
A= A (t) dt
T 0
Such a system is called ergodic.
Not all systems may be safely assumed to be ergodic. As a typical non-ergodic model, let us consider
the double-well potential of figure 2.14 again. The simulation time may be not long enough for the
system started from the local minimum to the left to meet a state which is capable of crossing the barrier
and reach the global minimum to the right.
Glasses are non-ergodic materials.
In the complex energy landscape depicted in figure 2.15, a system possessing sufficient kinetic energy
(which is to say, a high enough value of kB T ) will explore all the phase space in the laboratory time scale,
but if cooled suddenly may get freezed in a local energy minimum, that corresponds to a liquid structure,
thus becoming a glass. Again, this is what normally happens in a simulation, where cooling speeds are
of the order of 1013 K/s. A MD liquid, when cooled, will not crystallize, but will invariably become a
glassy solid, unless specialized methods are used.

2.2.3.4 Time dependent quantities

Since a MD simulation follows the real time evolution of the system, time-dependent quantities may be
computed. This is a major difference from MC simulations.

For example, the diffusion coefficient D is defined by Fick’s first law D as the linear coefficient between
the particles’ flux j and the concentration gradient∇c:

j = −D∇c
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 50

Figure 2.17: time dependence of the single particle mean squared displacement in a MD simulation (from
Frenkel and Smit 2002)

From statistical mechanics we know that D is connected to the mean squared displacement of a single
particle:
∂ h∆2 r (t)i
6D =
∂t
The function ∆ r (t) can be computed from a MD trajectory, and from its limiting slope (the time deriva-
2

tive) the diffusion coefficient is obtained.

A slightly more complicated example is offered from the time evolution of the PDF. A generalized
g (r) is the particle self-correlation (van Hove) function:

Gs (r, t) = probability of finding a particle in r at

time t, if the same particle was in r = 0
at time 0
4πr2 Gs (r, t) dr = probability that a particle travels a
distance between r e r + dr in time t

• In liquids, for t → ∞, Gs (r; t) tends to a gaussian of width Dt (that is, it broadens tending to the
constant value 0)

• In solids, , for t → ∞, Gs (r; t) tends to some fixed, nonzero distribution.

2.2.4 Comparison between MC and MD

Let us now summarize the key points.
In its original definition, a Monte Carlo computation is performed at constant T and samples the
(N, V, T ) ensemble, while a Molecular Dynamics simulation is performed at constant E and samples the
(N, V, E) ensemble. In principle, this is somehow a limitation of MD, because in a laboratory it is easier
to control T than E.
However, both methods have later been extended and refined. MD simulations schemes in the
(N, V, T ) e (N, p, T ) ensembles are quite popular now.
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 51

A major difference between the two techniques is that, as we have already remarked, time-dependent
quantities can be computed from MD only.
As for structural properties, when the same force field is employed MD and MC, in the limit of infinite
simulations, they should give the same results.

2.2.4.1 Computer simulation of glasses

To simulate a glassy substance, the technique of choice is Molecular Dynamics in the (N, p, T ) ensemble.
Through MD one can investigate all the dynamical behaviour of the system, which -as we have seen in
the introduction- is a fundamental feature of liquid-to-glass transition.
A MD simulation of a glass follows this scheme:

• a high T (e.g. 3000K) trajectory is performed for a standard amount of system time (∼ 100ps);
the substance is liquid

• starting from the last configuration, a new trajectory is started at a lower T (e.g. 2000K)

• this cycle is repeated at increasingly low temperatures

With this procedure the effective cooling velocity is of the order of 10

1000K
−10 s = 10
13 K
/s, much faster than
any attainable in the laboratory (≤ 100 /s).
K

The transition of the simulated system to glassy state is marked by a structural “arrest” that can
be followed through time-dependent quantities such as time correlation of particle positions, diffusion
coefficient, etc.
As noted earlier, almost all simulated liquids form a glass when cooled, even though only few of them
do so in the laboratory; they are “computer glasses”. Conversely, a substance that undergoes the glass
transition in experimental conditions (for example, SiO2 ) will also do so in simulation.

2.2.5 Examples of computational studies

Complete superposition of g (r) computed with MC and MD, and excellent agreement of both with
experimental data of liquid Ar, is seen in this early study (from Cusack 1987)
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 52

Figure 2.18: observed and computed g (r) of liquid Ar

The following figures (from slides by Micoulaut (2013)) show how PDFs from simulations can (or
cannot) reveal structure changes with temperature or pressure:
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 53

Pair distribution function : examples

Visual inspection allows to distinguish between a crystalline and an

amorphous structure

CCP Ni lattice with EAM potential

[email protected] Atomic modeling of glass – LECTURE5 MD CALCULATING

Figure 2.19: Structure of Ni from computations (from Micoulaut (2013))

CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 54

Effect of thermodynamic variables : temperature

The integral of g(r) allows to

Amorphous Se determine the number of neighbors
around a central atom.
Remember

The integral to the first minimum

gives the coordination number.

= 4
rm
Running coordination number N(r)

Caprion, Schober, PRB 2000

( )= ′ 4 ′ ′

[email protected] Atomic modeling of glass – LECTURE5 MD CALCULATING

Figure 2.20: temperature dependence of g (r) and second-shell coordination number in amorphous Se
(from Micoulaut (2013))
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 55

Effect of thermodynamic variables : pressure

d-GeO2
� Direct comparison with experiments can fail

� Simple force fields can not account for

pressure-induced changes (metallization)

� Additional structural insight is provided by

partial correlation functions : Ge-Ge, Ge-O,
O-O

Expt. (neutron) Salmon, JPCM 2011

MD: 256 GeO2 using Oeffner-Elliot FF
Micoulaut, JPCM 2004

[email protected] Atomic modeling of glass – LECTURE5 MD CALCULATING

Figure 2.21: Pressure dependence of g (r) and failure to reproduce experimental structure at high p (from
Micoulaut (2013))
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 56

2.2.5.1 Computer simulation of the CKN glass

The mixture [Ca (NO3 )2 ]0.4 [K (NO3 )]0.6 is a well-known glass former (“CKN glass”), the glass transition
being around 350 − 400K (see viscosity and structural relaxation in figure 2.22):

Figure 2.22: viscosity and density correlation in the CKN glass

CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 57

In the simulation, the Ca − Ca peak in g (r) splits under a certain temperature (figure 2.23): this
indicates two distinct spatial arrangements of the succession Ca2+ · · · NO−
3 · · · Ca
2+
and is a mark of
glassy phase onset. Ca lies very close to the NO3 plane (consider the almost complete splitting of
2+ −

peaks relative to Ca − O1 , Ca − N e Ca − O2 )

Figure 2.23: g (r)of CKN glass (from Signorini et al. (1990))

CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 58

∂ h∆2 r(t)i
The diffusion coefficient of Ca2+ (recall 6D = ∂t
) vanishes between 450K and 350K (fig-
ure 2.24)

Figure 2.24: (from Signorini et al. (1990))

The transition is also evident in the time-dependent PDF Gs (r, t) at various temperatures, figure 2.25.
At T = 600K the distribution maximum shifts to larger distances and the distribution broadens; at
T = 350K the ion is virtually still in this time window.
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 59

Figure 2.25: Distance travelled by Ca2+ at times t = 5, 15, 30, 45ps for various temperatures
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 60

2.2.6 Reverse Monte Carlo

Reverse Monte Carlo (RMC) is a computational technique that is widely used in the investigation of
disordered structures. However, it differs from the ones described above in that it is not really a simulation
technique.
In MC and MD simulations, the elements of the model are the particles and the force field through
which they mutually interact. These allow us to build a hopefully accurate replica of the real substance,
from which observable quantities can be computed and compared to their experimental counterpart. A
successful comparison has a twofold implication: it is a confirmation that the model is appropriate, and
it gives us some confidence about the accuracy of the microscopic details (e.g. single particle positions)
which cannot be measured.
However, our model may be biased, in that it may contain structural features -for example, a too stiff
bond angle (see expression 2.2.1)- that is not really required to reproduce the experiments; while it may
exclude other features that are equally consistent with the data.
So the question is: what structural information do measurements provide by themselves? Can we
learn anything from them about the microscopic structure, with minimal to no assumptions about the
model?
In RMC, no potential model is postulated. From randomly generated configurations (this is the
“Monte Carlo” aspect) observable structural properties such as the structure factor (2.7) are computed,
then compared to experimental values, and a goodness-of-fit measure such as the squared difference be-
tween the two is evaluated. The one structure that best fits the experiments is selected as the “observed”
structure.
Instead of using random techniques for optimizing a model, then comparing computed properties to
experimental ones, in RMC one starts from experimental measurements, and uses random generation of
hypothetical structures to optimize agreement between computed and experimental properties (this is the
“reverse” aspect).

2.2.6.1 How RMC works

In practice:

• one starts with a “reasonable” configuration and computes the observed quantity, say S (Q)

• the squared deviation χ2 between computed and experimental values is evaluated:

N
X [Sexp (Qi ) − Scomp (Qi )]2
χ2 =
i
σi2

(where N is the number of measured data, exp = experimental, comp = computed, and σi is the
error on measurement i)

• the structure of the model is perturbed with a small move (perhaps by moving one atom) and χ2 is
evaluated again; the new structure is accepted with a Metropolis-like criterion, exactly like MC

• with this procedure, a statistical sample is built with a Gaussian distribution of the deviation:
χ2
P (χ) = e− 2

which is what is expected for a random error in the measurement.

CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 61

2.2.6.2 RMC Examples

Figure 2.26 reports structure factors of SiO2 glass computed for a system of 3000 atoms with a RMC
algorithm where coordination numbers of Si and O were constrained at the known values (4 and 2)

Figure 2.26: SiO2 glass. Experimental and RMC fitted structure factors (from Keen 1998)

Experimental data are reproduced rather accurately; however, local structure (which is responsible for
the peaks at medium to high Q) is slightly more disordered than expected. For example, the “tetrahedral”
angle O − Si − O (dashed line in figure 2.27) is more distorted than expected (continuous line).
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 62

Figure 2.27: SiO2 glass (from Keen 1998)

The RMC procedure can be refined with the additional constraint that the tetrahedron must keep a
regular geometry. With this requirement, while the accordance with data is maintained, the local structure
is obviously respected (figure 2.27, full line), even though there is still some inaccuracy about medium
to long range structure, as evidenced by the distribution of the angle between tetrahedra, which shows a
maximum at 180◦ instead of the expected value of ∼ 140◦ (figure 2.28)
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 63

Figure 2.28: SiO2 glass (from Keen 1998)

There is a simple way to test the quality of the RMC method, which is to apply RMC to a system
whose structure is known with perfect accuracy: an MD simulated system!
CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 64

Figure 2.29: Li2 O · SiO2 (Muller et al. 2010) Neutron Scattering simulated with MD, along with RMC
fitting with two starting structures

Figure 2.30: (Muller et al. 2010) Partial g (r) of Li2 O · SiO2

CHAPTER 2. METHODS FOR THE INVESTIGATION OF GLASS STRUCTURE 65

Figure 2.31: (Muller et al. 2010) Ring size distribution in Li2 O · SiO2 . See caption to figure 2.29

In a study on Li2 O · SiO2 , partial PDFs (figure 2.30) and properties on small lengths scales, such as
coordination numbers and bond angles (not shown), are well reproduced by RMC, while properties in
the medium range order (for example, ring size distributions, figure 2.31) are not. This may be due to
the fact that the property whose value the RMC tries to fit is not so sensitive to variations in medium
range structure; recall that S (Q) depends on pair interactions, while medium range structure depends on
multi-body interactions.

2.2.7 Reference texts for section §2.2

The subject material of this section is treated -in much more detail than here- in Cramer (2004), chapter
3 (sections 3.1-3.5). The essential reference texts for the simulation of molecular systems are Frenkel
and Smit (2002) (chapters 3 and 4) and Allen and Tildesley (1987) (chapters 3 and 4). The interested
reader can find a more in-depth coverage, particularly of the Monte Carlo method at Chapter 13, in the
lesson notes I wrote for the “Struttura e dinamica di biomolecole” university course (Signorini (2013)).
Examples are taken from Micoulaut (2013) and Signorini et al. (1990). RMC is explained rather clearly
in Keen (1998).
Chapter 3

Models of glass structure

3.1 Introduction
3.1.1 Bonds in solids
The appropriate structural model for glasses is strictly correlated to the type of bonds that exist between
atoms, namely whether they are true chemical bonds (that is, covalent, electron-pair bonds) or not.
Crystalline solids also can contain both covalent and non-covalent bonds, however the same structural
model can be applied to all. Indeed, in most network-type crystal structure depictions, lines between
atoms are meant to just point to geometrical coordination and do not necessarily imply the existence of
covalent bonds. For example, in 3.1, of the representative compounds of the two structures, ZnS has a
mixed covalent-ionic character while CaF2 is purely ionic.

66
CHAPTER 3. MODELS OF GLASS STRUCTURE 67

Figure 3.1: In this figure (from O’Keeffe and Hyde 2020) lines between atoms represent chemical bonds
in (mainly covalent) ZnS, but not in ionic CaF2 .
CHAPTER 3. MODELS OF GLASS STRUCTURE 68

3.1.2 Classification of crystals according to bond type

In this section we briefly review how crystals can be grouped according to the main type of bonds:
1. covalent crystals (examples: C, Si; SiO2 , GaAs; PbO2 )

• a network of bonds extends to the entire crystal: crystal is like a huge “macromolecule”
• low coordination numbers (CN) (typically CN=4; sometimes CN=3 or 6); strong, directional
bonds.
• properties: hard; high melting T

C(diamond) (CN=4)

β − PbO2 (CN=6)
CHAPTER 3. MODELS OF GLASS STRUCTURE 69

2. ionic crystals

• formed by positive and negative ions, held together by electrostatic forces, which are spheri-
cally symmetric and long range
• CN (anions around a cation) in the medium range; typically CN=6; also 4,8,9. (Smaller
cation is often in holes of a compact structure of bigger anions).
• properties: brittle, medium to high melting T

NaCl (CN=6)

CsCl (CN=8)
CHAPTER 3. MODELS OF GLASS STRUCTURE 70

3. metals (examples: Cu[fcc,12], Zn[hcp,12], Fe[bcc,8])

• metallic bond, non-directional and delocalized

• high CN (typically, CN=12; also CN=8,...).
• properties: ductile, low melting T

Fe (bcc, CN=8)

Cu, Zn (fcc,CN=12)
CHAPTER 3. MODELS OF GLASS STRUCTURE 71

4. molecular crystals

• isolated molecules held together by Van der Waals forces

• CN: low within the molecule, medium to high among molecular units

CO2 crystal
CHAPTER 3. MODELS OF GLASS STRUCTURE 72

3.1.3 Structural models of glasses

While the structural model for crystals is always based on geometry and symmetry, for glasses a differ-
ent model is used depending on the bond type:

• if covalent bonds dominate, the model is that of a Continuous Random Network (CRN), in which
the solid is viewed as one infinite covalent compound with values for bond lengths and angles that
are either fixed or restricted to a narrow range. The model involves a number of aspects, among
which

– the geometrical and topological properties of the model in itself (coordination numbers, exis-
tence and dimensions of rings, etc.)
– structural theory of glass formation (Zachariasen rules)
– mechanical properties: average coordination and rigidity (Phillips-Thorpe theory)
– chemical order and concentration of partially bonded species (bond models)

• in metallic and ionic compounds the Random Close Packing (RCP) model is appropriate, where
the solid is represented as a collection of hard spheres tending to be in the closest possible contact
and fill the space

With the advent of computational methods (in which the “model” is the force field) traditional models
have begun to lose importance; however they still have value as limiting structures whose properties are
known

3.2 Continuous Random Network (CRN)

In the Continuous Random Network (CNR) model

• the system is assumed to be composed of “nodes” (the atoms) connected by node-to-node “links”
(the chemical bonds).

• each node forms a fixed number of links

• the links are rigid

• the angle between two links sharing a node (the bond angles) are also rigid

• links and angles can take a limited number of given values

The last condition is not strictly required by the model, but reflects the physical fact that in a real chemical
system only a few of different bond length and angle values are found. Also, the bond distances are of
the order of ∼ 0.1 nm, and angles are always in the range 90◦ ÷ 180◦ .
CHAPTER 3. MODELS OF GLASS STRUCTURE 73

3.2.1 General properties of the model

The first realizations of the CRN were hand built models of a few tens or hundreds of atoms such as
those of Bell and Dean for SiO2 (1966) (figure 3.2), Polk for amorphous Si (1971) (figure 3.3), Greaves
and Davis for amorphous As (1974), etc., who used plastic balls and aluminum small rods, or similar
materials, and simple devices to make angles rigid. These works not only proved that the model was
physically plausible, but also that the histogram of its measured atom-atom distances was remarkably
consistent with the observed radial distribution function (figure 3.2).

Figure 3.2:

Figure 3.3:

Yet, a number of issues need to be addressed, that are best treated in a more formal approach:
CHAPTER 3. MODELS OF GLASS STRUCTURE 74

1. does the network of bonds necessarily include rings or can it also be an open structure?

2. is there a limit to the number of constraints (prescribed values for bonds and angles)?

3. is there a limit to the number of bonds one atom can form (coordination numbers)?

To answer these questions we will consider the simplest case of a network made of N identical atoms
each with a fixed number of bonds, m.

3.2.1.1 Open or closed structure?

A little analysis shows that, although in principle it is always possible to imagine an infinite structure that
is open (with no rings), such structure would not be physically observable1 .
Clearly, in a totally open structure, which takes the form of a tree branching from an origin (the so-
called “Bethe lattice”), the number of terminal atoms increases exponentially with the branching order k
(see figure 3.4).

Figure 3.4:

It can be shown2 that moving toward the exterior shell, the ratio of the number of terminal atoms Nk
to that of the bulk atoms tends to a simple value
Nk
lim →m−2
k→∞ N − Nk

that is, for the number of terminal atoms to be negligible with respect to the bulk atoms, as it should be
in a real structure, we must have m = 2, which is a chain!
1
Note that finite chemical structures with this topology exist; they are called dendrimers and typically have a maximum
branching order k ≤ 10 and dimensions of a few nanometers (Sapra et al. 2019)
2
Signorini, 2019, unpublished results
CHAPTER 3. MODELS OF GLASS STRUCTURE 75

Also, consider that, since angles are in the range 90◦ ÷ 180◦ , the structure will curl over itself already
at the first few branching orders, and rapidly become infinitely crowded.
So, in the real world only closed structures (with rings) are observed

3.2.1.2 Number of constraints

We next observe that in closed structures with m ≥ 3, it is not possible to assign a chosen value to
each distance and each angle.
Indeed, the number of constraints (assignable values for distances and angles) exceeds the number of
degrees of freedom (3 per atom).
We will quantify this statement below. For now, consider the example of m = 3 (figure 3.5): if one
chooses, for each atom, 2 distances and 2 angles (which amounts to 3 constraints per atom, because each
distance is shared with the neighbour atom and counts for 12 ), the rest of the constraints (1 distance and 1
angle) are determined by geometry, cannot be assigned arbitrarily.

Figure 3.5: In the irregular structure of the left picture, only 2 distances (in green) and 2 angles (in red)
for each atom can be assigned; the rest will be determined by geometry. The structure to the right is an
example of how in some regular arrangements geometry allows bonds and angles to be all equal. Think
of these pictures as in 3 dimensions.

As a special case, it is generally not possible to assign one chosen value to all bond lengths and
to all angles. Such constraints will be only compatible with particular regular arrangements, the crystal
lattices. Think for example of graphite (m = 3) and diamond (m = 4).
Important consequences of constraints being in excess is that (a) the values imposed by geometry in a
random network will in general be considerably different from those those required by chemistry; and (b)
that they will assume a whole spectrum of ranges instead of being fixed to the few chemically accepted
values. The deviation from equilibrium values creates an excess energy, or stress, which will be higher
for bond distances than for bond angles, since force constants of the latter are smaller. Consequently,
bond angles values will be more broadly distributed than bond lengths.
CHAPTER 3. MODELS OF GLASS STRUCTURE 76

3.2.1.3 Coordination numbers

An interesting observation is that the number of excess constraints over degrees of freedom increases
with m.
The higher the coordination number, the higher the fraction of distances and angles that cannot be
assigned arbitrarily. Again, we will quantify this statement later. For now, we note that since chemistry
does require bond lengths and angles to assume certain values (that is, all constraints to be respected)
CRNs with increasingly high coordination numbers will be increasingly difficult to form. We will see
that coordination numbers 3 or 4 are most frequently found in random networks, while m = 5 or 6 are
only compatible with regular or quasi-regular latticesfigure 3.6.

Figure 3.6: Quasi-crystal (a somewhat ordered lattice, which however lacks translational symmetry) with
coordination numbers=5, 6

3.2.1.4 Summary
The last considerations, related both to topology and to physico-chemical requirements, can be summa-
rized as follows.
In real random networks we expect to find

1. closed structures. with rings

2. not very high coordination numbers, with typical values of 3 or 4

3. bond lengths and (particularly) angles distributed over a range of values

It should be kept in mind that this model does not take into account other constraints that exist in chemical
systems, although with less energy involved: torsions, nonbonded interactions, etc.
CHAPTER 3. MODELS OF GLASS STRUCTURE 77

3.2.1.5 Comparison to experimental structures

Before moving on to the general case of polyatomic systems, it is interesting to compare the structural
features that are obtained with the monoatomic CRN to experimental data. We have already noted that
the well-known ball-and-stick model of silica glass by Bell and Dean (figure 3.2) reproduces the observed
g (r) to a remarkable degree. The same is quite not true for the Polk model of a − Si (figure 3.3), as evi-
denced by the following figure (figure 3.7): the fact that RDF peaks are too narrow clearly indicates that
the ball-and-stick model is too rigid. The reason for the different agreement with experiments between
the two models will be clearer later on, but we can anticipate that the CRN model of a − Si implies much
more excess constraints than that of SiO2 .
The agreement of the Polk model would undoubtedly improve if some degree of flexibility in con-
straints is introduced, something that could be implemented with simple adaptations to sticks and joints;
note, in any case, that a broadening of the first peak (the first neighbours’ shell) can be obtained only
by allowing some tolerance in the bond length, not in the bond angle, although the force constant of the
latter is certainly softer than that of the former.

Figure 3.7: comparison of a − Si radial distribution function of the Polk ball-and-stick model, and exper-
imental RDF (from Varshneya 1994)

More flexibility is allowed by simulating the CRN model in a computer program. The early (Stein-
hardt et al. 1974) study reported in figure 3.8, regarding a − Ge (isomorphous to a − Si), used what can be
considered a simplified version of a Monte Carlo simulation, consisting in a minimization of the potential
energy, with no Boltzmann sampling, but with the inclusion of “structure relaxation” procedures taking
into account thermal oscillations, experimental broadening and other effects. In spite of the physical
inaccuracy of the computations the results appear quite satisfactory.
CHAPTER 3. MODELS OF GLASS STRUCTURE 78

Figure 3.8: Radial distribution function of amorphous Ge (from Steinhardt et al. 1974)

3.2.2 Polyatomic CRN

Many covalent glasses contain atoms of two or more atomic species with different coordination numbers.
The corresponding CRN has the same basic features as a monoatomic CRN, the main difference being
that in the analysis outline above an average hmi must be considered instead of m.
Polyatomic CRN can be divided in two main classes

1. Networks with chemical order, where only mixed bonds A − B are allowed; the typical case are
oxide glasses, like SiO2 , where the metal (or nonmetal) atoms are connected by bivalent oxygen
atoms acting as bridges between two metal atoms (see figure 3.9).

2. Networks without chemical order, where all types of bonds, A − A, A − B and B − B, are possi-
ble: for example, chalcogenide glasses like As2 S3 ; or fluorides, like BeF2 ; in many cases one can
have non-stoichiometric compound of general formula Ax B1−x (figure 3.10)

3.2.2.1 Oxide glasses

Oxides Ax Oy occupy a special place in glass theory, both because they are the glassformers which have
been known for the longest time, (SiO2 , B2 O3 , ...), and because it is for oxide glasses that the CRN
model was first developed and tested.
The network of oxide glasses is made of A atoms linked through bridging O atoms. One way to see
this is that the elemental units of this CRN are, instead of atoms, identical AOm (m = 3 or 4) polyhedral
units sharing O vertices (figure 3.11). This is justified by the observation that (usually) the geometry
of AOm is fixed: both A − O distances and O − A − O “internal” angles are rigid. In this approach,
the Ax Oy network can be mapped into an isomorphous monoatomic network whose nodes represent
CHAPTER 3. MODELS OF GLASS STRUCTURE 79

Figure 3.9: Chemically ordered AB network

Figure 3.10: Non-chemically ordered network with tri- and bi- coordinated atoms
CHAPTER 3. MODELS OF GLASS STRUCTURE 80

the polyhedron units (figure 3.12). Put differently, the polyatomic network can be obtained from the
monoatomic one by substituting (“decorating”) each A − A link with an A − O − A bridge.

Figure 3.11: 3D vision of silica glass network as formed by SiO4 tetrahedra

CHAPTER 3. MODELS OF GLASS STRUCTURE 81

Figure 3.12: Isomorphism between a monoatomic network and a biatomic network with chemical order.
Yellow tetrahedra and red bridging atoms in the bottom panel correspond to nodes and links, respectively,
in the top panel

Note, however, that even if there is topological equivalence between the two networks, they will
be geometrically, and mechanically, different if A − O − A (inter-polyhedra) angles are flexible. The
Si − O − Si angle flexibility is at the origin of the broader Si − Si peak observed in the histogram of the
SiO2 ball-and-stick model as compared to the first peak in the a − Si model (figure 3.13), and can be
observed in XRD (figure 3.14).
CHAPTER 3. MODELS OF GLASS STRUCTURE 82

Figure 3.13: Si − Si peak in the histograms of the ball-and-stick models of SiO2 glass (left) and amor-
phous Si (right)

Figure 3.14: Distribution of Si − O − Si angle in fused silica, obtained with XRD (from Varshneya
(1994))

One important property of oxide glasses is that unit polyhedra share only vertices, not edges or faces
(figure 3.15, left).
Sharing an edge amounts to forming a 4-atom ring (figure 3.15, right), that would imply a high stress
either of the intra-polyhedron O − A − O angle or of the inter-polyhedra A − O − A angle, or of both,
at least some of which would have to assume values ≤ 90◦ .
Also, edge sharing would reduce the amount of branching in the network, thus making the structure
less rigid 3
3
rather unexpectedly: see Lucas (2014), Wilson (2012))
CHAPTER 3. MODELS OF GLASS STRUCTURE 83

Figure 3.15: SiO4 tetrahedra share only vertices, not edges or faces
CHAPTER 3. MODELS OF GLASS STRUCTURE 84

In summary:

The CRN model of an oxide Ax Oy obeys the following rules:

• number of bonds (coordination number)

– O=2
– A = 3 or 4

• flexibility, in decreasing order:

– A − O − A angles (external to polyhedron) >

– O − A − O angles (internal) >
– A − O bonds

• angles are always > 90◦ ; no 4-atom ring, polyhedra share vertices only

3.2.2.2 Structural theories of oxide glass formation; Zachariasen’s rules

The above rules are topological criteria for covalent oxide glass formation.
In fact, a structural theory for glass formation -now in contrast with the kinetic theory (any substance
can form a glass, provided it is cooled fast enough)- has been advanced early in the study of glasses, and
has enjoyed a considerable success since.
In 1932, Zachariasen made an attempt to explain why some oxides can form glasses rather easily
while others cannot, by considering local structures (coordination geometries) found in corresponding
crystals. He took the observation that glasses exhibit mechanical properties, density, enthalpy of forma-
tion, etc., that are very similar to the crystalline form of the same substance, as an indication that local
structures of crystals are conserved in the glass phase.
On this basis, he stated a set of rules for the formation of oxide glasses (Varshneya (1994); Rao
(2002)), that were bound to have great influence on glass science.
These rules, in essence, correspond to the rules for the formation of a continuous random network
such as the ones summarized above. In fact, Zachariasen described oxides as made up of ions (partly, for
continuity with a previous model by Goldschmidt which considered the ratio of ionic radii, as in Pauling’s
analysis of crystal structures); however, his theory is based on bonds (or “coordinations”) between atom
pairs, which is more consistent with a covalent nature of interactions4 .
4
For a more detailed discussion of the relationship between the nature of bonding and the glass-forming ability see Rao
(2002), p. 36
CHAPTER 3. MODELS OF GLASS STRUCTURE 85

Here are Zachariasen’s rules Rao (2002):

In a glass forming oxide of the formula Am On

1. Oxygen atom may be linked to no more than two A atoms

2. The number of oxygen atoms surrounding A atoms must be small (3 or 4)

3. The oxygen polyhedra share only corners with each other, neither edges nor faces

4. At least three comers in each polyhedron must be shared.

The first three rules are consistent with the CRN rules we have found above. Rule 4 is simply the
statement that for a simple network to not reduce to a chain the average number of links of each node (in
this case, the polyhedron) must be m ≥3.
For an illustration of Zachariasen’s rules with examples involving oxides of different formula, see
http://www1.chim.unifi.it/u/signo/did/vetri/3-modelli-strut.pdf (in Italian).
The CRN structure of oxide glasses was first observed in real space in 2012 (Lichtenstein et al.
(2012); Huang et al. (2012)), with different electron microscopy techniques, in a thin layer of glassy
SiO2 adsorbed on a surface. An ordered domain (figure 3.16, right) and an amorphous domain can be
seen, with a striking resemblance to the original drawings of 2D structures by Zachariasen.
CHAPTER 3. MODELS OF GLASS STRUCTURE 86

Figure 3.16: STM image of a thin layer of glassy SiO2 adsorbed on a Ru surface. In (b) O atoms are
marked (in red) and the position of Si atoms is calculated to be in the centre of O triangles

3.2.2.3 Glasses with no chemical order

As we have seen, a binary CRN can also be formed with two elements not including O.
The reference compounds for this class are chalcogenides, which are formed by

• one element of group 16 (S,Se,Te) heavier than O (“chalcogens”)

• and one element of groups 14 and 15 (Si,Ge; P,As)

In chalcogenides there needs to be no chemical order: homopolar bonds (between same-species atoms)
are allowed. This opens the possibility for non-stoichiometric compounds such as Six Se1−x (fig-
ure 3.17):
CHAPTER 3. MODELS OF GLASS STRUCTURE 87

Figure 3.17: a fragment of network with homopolar bonds

In these substances Zachariasen’s rules, which are based on the exclusive presence of ABm units
linked by their vertices, do not hold. Indeed, the ABm groups, while not being the only groups that are
present in the network, are also found to form edge-sharing dimers (figure 3.18)

Figure 3.18: dimer structures found in chalcogenide glasses (CS, corner sharing; ES, edge sharing; ET,
ethane-like)

3.2.3 Coordination number and rigidity (Phillips-Thorpe theory)

In the CRN model, the rigidity of a glass is connected to the degree of connectivity between atoms. The
more bonds an atom forms, and the more rigid these bond are, the more rigid the lattice will be.
Rigidity properties depend on the difference between the number of constraints and the number
of degrees of freedom of atoms.
CHAPTER 3. MODELS OF GLASS STRUCTURE 88

We have seen above that in a monoatomic CRN like a − Si, with a high coordination number (CN=4),
not all constraints on bond lengths and angles can be satisfied, and the lattice can be built only if the
constraints are relaxed, even partially.
If atoms with lower coordination (for example, O or S, with CN=2) are introduced, a less constrained
CRN results. In these binary systems, the higher the fraction of low CN atoms, the lower the rigidity
is.

Figure 3.19: A high coordination, rigid network (left) and a low coordination, less rigid one (right)

Phillips and Thorpe (Thorpe 1983) have developed a simple theory about the rigidity of a CRN as a
function of the average coordination.

Figure 3.20: degrees of freedom (d) and constraints in a CRN with coordination r
CHAPTER 3. MODELS OF GLASS STRUCTURE 89

In a CRN of atoms with coordination number r, for each atom we define (figure 3.20)
r
nbond =
2
bond distances (since each bond is shared by two atoms), and

nangle = 2r − 3

bond angles (the computation of nangle takes into account the fact that angles are not all independent). If
distances and angles are fixed, the total number n of constraints is therefore5
r
n= + 2r − 3
2
In a mixed lattice like As2 S3 an average coordination number may be defined as
X
hri = ri xi
species i

where ri is the CN of della species i, and xi is its molar fraction. For example,
2·3+3·2
As2 S3 : hri = = 2.40
2+3
1·4+2·2
SiO2 : hri = = 2.67
1+2
So the number of constraints per atom in a mixed lattice is

hri
n= + 2 hri − 3 (3.1)
2
To estimate the system’s rigidity, one compares n to the number d of degrees of freedom per atom; in
tri-dimensional space it is obviouslyd = 3 (the 3 cartesian coordinates)

• If n < d → less constraints than degrees of freedom; the system is underdetermined or floppy

• If n > d → more constraints than degrees of freedom; the system is overdetermined or stressed
rigid

• If n = d → the constraints equal all degrees of freedom; the system is optimally constrained and
isostatic

– for example, density has a maximum for n = d; more constraints would prevent the system
from reaching the minimum energy configuration (Lucas (2014))

In tridimensional space the value of hri for which the critical value n = d is reached is given by substi-
tuting n = d = 3 in equation (3.1)

hri
3 = + 2 hri − 3
2
hri = 2.40 (3.2)
CHAPTER 3. MODELS OF GLASS STRUCTURE 90

Figure 3.21: Rigidity percolation (from Lucas 2014). f is the fraction of free modes (f ∼ (d − n) or
f = 0 if (d − n) < 0)

Now, in non-stoichiometric systems hri depends on the relative concentration of the species with
different CN. By increasing the fraction of the species with higher CN, hri is increased: starting from an
underdetermined system composed of a floppy (F) matrix with isolated rigid (R) domains, when these
domains eventually interconnect the system transitions to a rigid matrix with isolated floppy regions; this
transition, which occurs in a narrow interval of composition, is called rigidity percolation (figure 3.21)
The rigidity of the disordered network is found to be also connected to the difficulty in forming a
glass. For example, in the system Gex Se1−x one has

hri = 4x + 2 (1 − x)
= 2x + 2

so the critical value hri = 2.40 (equation (3.2)) is reached for x = 0.20. Indeed, this composition
corresponds to minimum difficulty in forming the glass, as measured by the cooling rate required for
glass formation.

The main merit of Phillips-Thorpe theory is in linking the mechanical properties of a series of glasses
to a single parameter, the average coordination number hri.
This is clearly illustrated in the data for Ge − As − Se mixtures with varying composition(Halfpap
and Lindsay 1986). Two series were prepared with variable hri, one at high- and one at low-As content
(squares and circles in the ternary phase diagram). In spite of the large difference in the composition (and
5
it is assumed that rings are formed by 6 or more atoms, otherwise there are additional dependencies between distances
and angles (see Thorpe (1983),Lucas (2014))
CHAPTER 3. MODELS OF GLASS STRUCTURE 91

Figure 3.22: Glass forming difficulty as a function of composition in Gex Se1−x (from Varshneya 1994 -
including caption)
CHAPTER 3. MODELS OF GLASS STRUCTURE 92

Figure 3.23: Composition (left) and elastic constant C11 as a function of average coordination number
hri (right) in two series of Ge − As − Se mixtures

so in the microscopic structure) between the two series, it is found that the elastic constant C11 begins to
rise around hri ∼ 2.40 for both series (figure 3.23).
Obviously the original approach of Phillips and Thorpe, that we have presented here, suffers from
oversimplification. Points that have been addressed by later research include: accounting for rings with
< 6 elements and for systems with dangling bonds or atoms with varying r (such as silicon Q(n) species
in alkali silicate glasses).

3.2.4 Bond models

Another topological approach to the structure of network glasses investigates the role of the nature and
number of bonds formed by an atom. We will delineate the approach here while leaving a more detailed
discussion to specific sections below.
One issue regards the relative amount of heteropolar (AB) and homopolar (AA, BB) bonds in a binary
non-stoichiometric compound Ax B1−x . Can this relative amount be predicted from composition only?
Chalcogenide glasses (As − Se, Ge − Se), figure 3.24, are an example of such systems.
CHAPTER 3. MODELS OF GLASS STRUCTURE 93

Figure 3.24: A possible configuration of As (red)-S (yellow) system

Chalcogenides can be studied with a topological model where

• each atom species forms a fixed number of bonds

• no dangling bonds are present

This topological model is then integrated with considerations about bond energies and their influence on
bond type distributions. The subject is treated in a later section.
A second field of investigation is systems where the same unit can have different number of bonds,
with the total number of bonds being constrained by the system’s overall composition.
In mixed oxides containing a “network former” (Si or B) and a “network modifier” (alkali metal )
such as
(M2 O)x · (SiO2 )1−x
anionic groups are formed with Non-Bridging Oxygen, NBO). Formally, they can be thought as derived
from a pure SiO2 network upon addition of the alkali oxide:
CHAPTER 3. MODELS OF GLASS STRUCTURE 94

At high M2 O concentrations several groups can be formed with different number of NBO:

Clearly, there is no a priori way of knowing if (say) one M2 O unit produces two Q(3) units, as in the
reaction above, or one Q(2) unit.
A similar scenario is encountered in alkali-borate glasses
(M2 O)x · (B2 O3 )1−x
where different boron units exist:

Note that in this case, in contrast to the previous one, besides all the tri-coordinate boron groups with
all the possible number of NBO, B 2 , B 1 e B 0 , at low M+ concentration also the tetracoordinate group
B 4 carrying a negative charge is found.
CHAPTER 3. MODELS OF GLASS STRUCTURE 95

In systems like alkali silicate or alkali borate glasses the relative abundance of the various groups (or
species) is determined

• by quantitative relations originating from the fixed molar fractions of atoms (composition con-
straints)
In general, these constraints are not sufficient to determine the abundance of the various species.
For example in (M2 O)x · (SiO2 )1−x we have the following system of equations
X
Qi = 1 − x
i
(4Q0 + 3.5Q1 + 3Q2 + 2.5Q3 + 2Q4 ) = 2 − x

that is, for each composition x, 2 equations with 5 unknowns (Q0 , . . . , Q4 ).

• from simplified models based on some additional hypotheses and/or parameters

• from thermodynamical data such as equilibrium constants

As anticipated, these bond models will be treated in more detail when discussing specific applications
below.

3.3 Random Close Packing (RCP)

In systems where atom pair attractive forces are long-range and non-directional (like metallic, Coulom-
bic or van der Waals interactions) we expect the structure to be driven only by the tendency toward
minimization of atom-atom distances (and thus, of total energy), with no role being played by the direc-
tion of atom-atom links. Obviously, distances cannot go to zero: there is some minimum value, the sum
of neighbour atoms’ radii, below which they cannot be reduced.
In fact, most elemental metals in the solid state assume dense packing crystalline structures, like
Hexagonal Close Packing (HCP) or Face Centered Cubic (FCC), suggesting that their potential energy
is lowest in these arrangements. Note that some metals, and especially alloys, can form also amorphous
solids.
The simplest way to model isotropic force systems is thinking of them as a collection of hard spheres
that are compressed to maximum density.
Historically, researchers have started with physical models. It is an experimentally reproducible result
that if a container (preferably one with irregular shape and rough inner surfaces) is filled with hard spheres
of equal dimensions, and shaken until the volume occupied by the spheres cannot be further reduced, then
the density reaches a well defined limiting value: if expressed as the fraction of the total volume occupied
by the spheres this value is
Vspheres
φRCP = ' 0.637 (3.3)
Vtotal
Note that the spheres, although packed to the maximum, do not assume a regular arrangement. This
type of packing is therefore called Random Close Packing (RCP). Other names are Dense Random
Packing (DRP), or, more specifically, Dense Random Packing of Hard Spheres (DRPHS).
Other procedures aimed at reducing the total space occupied by the spheres, such as putting hard
balls in a rubber balloon which is then kneaded, or settling them in oil, also give the limit value of
equation (3.3). The same result is obtained by simulating the system in a computer.
CHAPTER 3. MODELS OF GLASS STRUCTURE 96

Figure 3.25: Random Close Packing of equal hard spheres

This reproducibility suggests that such systems have a rather well-defined structure. Indeed, structural
elements of RCP, such as coordination numbers, have a characteristic distribution; figure 3.26 shows the
g (r) of a steel ball model and of a computer simulated system.

Figure 3.26: Pair distribution function g (r) of RCP in a physical model and in a computer simulation
CHAPTER 3. MODELS OF GLASS STRUCTURE 97

There has long been a general consensus about the view that RCP “provides the most satisfactory
model for the structure of amorphous metals” (Zallen 1998) and, to some extent, also of “metallic glasses”
(which in fact may contain also nonmetals, like Ni76 P24 ; see table in figure 3.27).

Figure 3.27:

In figure 3.28 the g (r) of RCP is compared to the observed (XRD) g (r) of the metallic glass Ni76 P24 .
Single-sized balls RCP is only partially appropriate for this system, as the atomic radii of Ni and P are
r(P)
similar, but not equal ( r(Ni) = 0.786 ).
6
Sheng et al. (2006)
CHAPTER 3. MODELS OF GLASS STRUCTURE 98

Figure 3.28: Comparison of atom-atom contacts distribution of RCP and of Ni76 P24 (from Varshneya
1994); the quantity plotted on the y axis is the “reduced RDF” 4πr [ρ (r) − ρ0 ] = 4πrρ0 [g (r) − 1]

The definition of RCP that we have given above is, clearly, totally empiric. In a more formal way,
RCP can be expressed as “the densest, non-regular arrangement of spheres, such that no empty space
remains where another sphere could fit”. However, RCP is still not well-defined mathematically. One
reason for this is that one can always make a closer-packed structure that is less random (Torquato et al.
2000) - since, as we will see shortly, there are some regular structures that have a higher packing fraction
φ than RCP. In any case, there are a number of theoretical frameworks about the subject of the so-called
“jammed matter” where RCP appears as a special case. In one of these theories (Song et al. 2008) a
value is found for the packing fraction of RCP that is very close to the empirically determined value
(equation (3.3)):
6
φRCP = √ = 0.63397
6+2 3

3.3.1 Comparison of RCP to crystal close-packed structures

It should be stressed that RCP is not the most efficient packing of spheres in space.
The problem of finding the densest packing of spheres has been addressed in the past, motivated by
practical purposes, such as how to efficiently store cannonballs in a ship. Back in 1611, Kepler hypothe-
sized that the densest packing was a regular one, namely that of the regular close-packed structures with
hexagonal symmetry, HCP and FCC, which is
π
φHCP = √ = 0.74048
18
CHAPTER 3. MODELS OF GLASS STRUCTURE 99

Vspheres
packing fraction φ = Vtotal
FCC/HCP 0.740
RCP ' 0.64
Table 3.1: Packing fraction of crystal close-packed structures and RCP

Figure 3.29: Crystal close packed structures: left, Hexagonal Close Packing (HCP); right, Face Centered
Cubic (FCC)

It was later proved that HCP/FCC are the densest regular packings, while the mathematical proof of
Kepler’s conjecture, that this is the highest absolute packing, either regular or irregular, was first found
only in 1998 (Hales 2005).
As is well known, HCP and FCC can be generated by superimposing layers of hexagonally arranged
spheres with the spheres of the next layer occupying positions above the center of the triangles formed
by the spheres of the previous one; the third layer is placed in the same position as the first one, with an
ABABAB... sequence, in HCP, or in a third possible position with an ABCABC... sequence, in FCC
(see figure 3.29). Although this is not usually noted, any non-periodic alternation of planes A, B, C (such
as ABCBABACB...), called “Barlow packing”, gives a quasi-lattice with the same packing fraction.
Packing fractions of regular and random close packing are summarized in Table 3.1.

3.3.2 Icosahedral geometry and local versus long-range packing

What, then, makes disordered metal structures possible, if packing is more efficient in ordered ones?
There are a number of possible explanations for this, including the role of differently sized atoms and
the fact that the atom-atom potential is not well reproduced by a hard sphere approximation.
However, the main reason is of geometrical nature, and lies in what can been seen as a competition
between local and long-range packing.
To understand this, let us first consider dense packing of hard disks in 2D. One way to find the
CHAPTER 3. MODELS OF GLASS STRUCTURE 100

arrangement of maximum density is to start with one disk, then add one disk at a time in the closest
position to the other disks. Obviously, the closest that two disks can go is to be in contact; the third disk
can be in contact to both the first two, with centers forming an equilateral triangle. We can now place a
new disk in contact with two existing ones, forming a triangle that shares one side with the first triangle.
If this procedure is iterated, a hexagonal arrangement is built, that fills up the whole plane. We can
describe this by saying that the densest-packing elementary figure in the plane, the equilateral triangle,
can tile the plane exactly, without leaving empty areas.

Figure 3.30:

Now let us turn to 3D. Analogous to 2D, three spheres are best packed in a triangular arrangement.
The fourth sphere can be laid just over the center of the three, thus forming a regular tetrahedron, where
each sphere is in contact with the other three. The regular tetrahedron is the densest elementary packing
figure in space; however, it cannot completely fill the space without leaving empty holes. We see this
in figure 3.31: up to five tetrahedra can be stacked sharing one face, leaving a small dihedral of δ5 = 7.4◦
empty. This means that if we follow the same procedure as in 2D, adding one sphere upon the face of an
existing tetrahedron, and we iterate, we will end up with empty cavities (and, additionally, with an atom
pattern that has no symmetry).

Figure 3.31: Five adjacent regular tetrahedra leave an empty segment

A more efficient filling of space can be accomplished by distorting tetrahedra slightly: for example,
if the three spheres on one face are moved away from each other by about 5%, still being in contact with
the fourth one, such modified tetrahedra can be joined to form a regular icosahedron around a center
sphere (figure 3.32) where all spheres on the vertices are in contact with the center sphere.
CHAPTER 3. MODELS OF GLASS STRUCTURE 101

Figure 3.32: Slightly distorted tetrahedra form a regular icosahedron

With the icosahedral geometry a good local packing is accomplished: there are 12 spheres in contact
with a central sphere and almost in contract with each other.
This however turns out to be only a partial success on a larger scale, because, again, icosahedra
cannot fill space.
In the HCP structure (see for example O’Keeffe and Hyde 2020) each sphere, like in the icosahedron,
is surrounded by 12 atoms (or spheres) at distance d. Here, tetrahedra are regular, but share edges,
not faces; and the coordination polyhedron has also square faces, where only two atoms are in contact
(figure 3.33). As a result, the average distance between outer atoms is larger than in the icosahedron. The
same holds for FCC.

Figure 3.33: HCP structure contains both regular tetrahedra and square pyramids
CHAPTER 3. MODELS OF GLASS STRUCTURE 102

Figure 3.34: Direct comparison of icosahedral (left) and HCP (right) arrangements of 12 atoms around a
central atom

The 13-atom cluster in the icosahedron has a higher (local) packing fraction (φicos = 0.7667 ) than
HCP, and has a lower energy (De Graef and McHenry 2007). But, unlike the icosahedron, the 13-atom
cluster in HCP can be repeated indefinitely, because it can fill up space 8 .
Thus, by sequentially adding up tetrahedra to an icosahedral seed, the resulting structure will in-
creasingly include empty regions, until (at around 5000 atoms, (Cheng and Ma 2011, p 404)) its density
becomes lower than that of HCP.
One way to put this is that FCC/HCP structures, as compared to icosahedron-based arrangements,
renounce optimal short-range packing in favour of optimal long-range packing.
The conceptual map of building the densest packing at each range level, with “frustration” steps
where the geometrical element cannot be propagated, is illustrated in figure 3.35.
7
Hermann et al. (2007)
8
together with regular octahedra https://mathworld.wolfram.com/TriangularOrthobicupola.html
CHAPTER 3. MODELS OF GLASS STRUCTURE 103

Figure 3.35:

Icosahedral-type (nonregular) clusters are found in RCP, in monoatomic liquids or in metallic glasses
of similar-sized atoms, although in these contexts clusters with higher CN (e.g. 13 or 14, see below)
prevail.
In this perspective, one may say that a glass is formed, in place of a crystal, when an isotropic liquid is
cooled so fast that there is not enough time (probability) for the locally disfavoured FCC/HCP structures
to form and grow in competition with the locally favoured clusters.

3.3.3 More on comparing RCP to crystal close-packed structures

A more detailed comparison of close-packed structures, disordered (RCP) and regular (HCP/FCC), can
be made by looking at either their lattice polyhedra or Voronoi polyhedra.

3.3.3.1 lattice polyhedra

In crystal close-packed structures, we can consider those lattice polyhedra9 whose vertices are atoms in
close contact and such that they can host one smaller sphere; these are often called “holes” or “cavities”.
There are two types of holes in a crystal of N atoms:
9
we do not refer here to Delauney lattice, which is formed only by tetrahedra
CHAPTER 3. MODELS OF GLASS STRUCTURE 104

• regular tetrahedra (2N = 32 )

• regular octahedra (N = 13 )

These two types of polyhedra make up a complete tiling of space. Note that euclidean space cannot be
completely tiled with either tetrahedra alone, or octahedra alone.

Figure 3.36: Tetrahedral and octahedral holes in FCC/HCP

In the RCP, one can similarly define polyhedra or holes, including, besides atoms in “close contact”,
also atoms in “near-close contact”, separated by a distance less than 5% longer than one diameter (com-
pare the geometry of the icosahedron, above). In his seminal work of the 1960s, J. D. Bernal (see for
example Bernal 1960; a complete review of his work is in found in Finney 2013) showed that the RCP
structure can be completely tiled with only five different, slightly distorted polyhedra whose faces are
equilateral triangles.
CHAPTER 3. MODELS OF GLASS STRUCTURE 105

Figure 3.37: Bernal canonical holes

Figure 3.38: Another representation of Bernal holes

CHAPTER 3. MODELS OF GLASS STRUCTURE 106

These “Bernal canonical holes” have different occurrences in RCP than their counterparts in FCC/HCP:
tetrahedra (a) are ∼ 80%; octahedra (b) ∼ 10%10

3.3.3.2 Voronoi polyhedra

In crystal close-packed structures the Voronoi polyhedra are

• HCP: rhombic dodecahedron, (a) in figure 3.39 11 ;

• FCC: Trapezo-rhombic dodecahedron (coloured in figure 3.39)

Figure 3.39:

In the RCP, on the other hand, there many types of VP, e.g. (b) e (c) in figure 3.39, with number of
faces = 12 ÷ 17
The distribution of Voronoi polyhedra in RCP is reported in figure 1.15. For comparison, in FCC/HCP
structures

• the number of faces per cell is 12

• the number of edges per face is 4 in FCC, and 4 or 5 in HCP

The following table summarizes properties of crystal and random close-packed structures:
FCC HCP RCP
packing fraction φ 0.74048 0.637
2
tetrahedra 3
∼ 80%
lattice polyhedra (holes) octahedra 1
3
∼ 10%
other (Bernal) 0 ∼ 10%
faces 12 12 ÷ 17 (max at 14)
Voronoi polyhedra
edges / face 4 4 or 5 3 ÷ 8 (max at 5)
10
Bernal (1964); (Chaudhari and Turnbull (1978) has 84% and 5%)
11
this is the VP corresponding to the CN = 12 arrangement of neighbours of figure 3.34 (right)
CHAPTER 3. MODELS OF GLASS STRUCTURE 107

3.3.4 close-packed structures of binary systems

As already mentioned, most metal glasses are formed by binary mixtures of metals, or metal/metalloid
pairs. More precisely, by the term “metallic glass” we mean (Elliott 1984) an amorphous material that
(a) contains (although not exclusively) metal elements, and (b) exhibits metallic properties (electrical,
magnetic, optical). This definition excludes, for example, a − Sb o a − As, that are semiconductors, not
metallic; or a − Si : H which has metallic properties but does not contain metal elements.
The structure of binary systems equivalent to RCP is determined by the atomic radii ratio R∗ . If
R∗ ≈ 1, as in Ni63 Nb37 the structure is similar to monoatomic RCP, with a prevalence of CN> 12. If
R∗ is lower, and the smaller atom has a lower molar fraction than the bigger one (“solute” and “solvent”,
respectively), clusters of solvent atoms around one solute atom tend to assume lower CNs. In a study
that makes use of experimental data and calculations, many local packing clusters were found with a
distribution that depends on R∗
CHAPTER 3. MODELS OF GLASS STRUCTURE 108

Figure 3.40: (from Sheng et al. (2006))

Chapter 4

Glass structures

4.0.1 Oxide glasses: simple

4.0.2 Oxide glasses: multi-component
4.0.3 Non-oxide glasses with CRN structure
4.0.4 Other non-oxide glasses: metal, ionic

109
Bibliography

M. P. Allen and D. J. Tildesley. Computer simulation of liquids. Oxford University Press, New York,
1987. ISBN 0-19-855645-4.

M. Alonso and E. J. Finn. Elementi di fisica per l’universita, vol II: campi e onde. Masson-Addison
Wesley, 1982.

JD Bernal. Geometry of the structure of monatomic liquids. Nature, 185(4706):68–70, 1960.

John Desmond Bernal. The bakerian lecture, 1962. the structure of liquids. Proceedings of the Royal
Society of London. Series A, Mathematical and Physical Sciences, 280(1382):299–322, 1964.

Ronald Newbold Bracewell. The Fourier transform and its applications, volume 31999. McGraw-Hill
New York, 1986.

P Chaudhari and D Turnbull. Structure and properties of metallic glasses. Science, 199(4324):11–21,
1978.

Y.Q. Cheng and E. Ma. Atomic-level structure and structure-property relationship in metallic glasses.
Progress in Materials Science, 56(4):379 – 473, 2011. ISSN 0079-6425. doi: https://doi.org/
10.1016/j.pmatsci.2010.12.002. URL http://www.sciencedirect.com/science/article/pii/
S0079642510000691.
C. J. Cramer. Essentials of Computational Chemistry: Theories and Models, 2nd ed. Wiley, 2004. ISBN
0-470-09182-7.

N. E. Cusack. The Physics of Structurally Disordered Matter: An Introduction. IOP Publishing Ltd,
1987. ISBN 0-85274-591-5.

Marc De Graef and Michael E McHenry. Structure of materials: an introduction to crystallography,

diffraction and symmetry. Cambridge University Press, 2007.

S. R. Elliott. Physics of amorphous materials. Longman London ; New York, 1984. ISBN 0582446368.

John L. Finney. Bernal’s road to random packing and the structure of liquids. Philosophical Magazine,
93(31-33):3940–3969, 2013. doi: 10.1080/14786435.2013.770179. URL http://dx.doi.org/10.
1080/14786435.2013.770179.
D. Frenkel and B. Smit. Understanding molecular simulation. Academic Press, London and San Diego,
California, 2002. ISBN 0-12-267351-4.

Thomas C Hales. A proof of the kepler conjecture. Annals of mathematics, pages 1065–1185, 2005.

110
BIBLIOGRAPHY 111

B L Halfpap and S M Lindsay. Rigidity percolation in the germanium-arsenic-selenium alloy system.

Physical Review Letters, 57(7):847, 1986.
Helmut Hermann, Antje Elsner, and Dietrich Stoyan. Behavior of icosahedral clusters in computer sim-
ulated hard sphere systems. Journal of non-crystalline solids, 353(32-40):3693–3697, 2007.
Pei-Yin Huang, Robert Hovden, Qingyun Mao, D. Muller, S. Kurasch, U. Kaiser, J. Kotakoski, A.V.
Krasheninnikov, Anchal Srivastava, Viera Skákalová, J. Smet, and J.C. Meyer. Quantitative atomic-
resolution imaging and spectroscopy of a 2d silica glass. Microscopy and Microanalysis, 18:340–341,
07 2012. doi: 10.1017/S1431927612003558.
ISAACS. The radial distribution functions, 2013. URL http://people.cst.cmich.edu/petko1vg/
isaacs/phys/rdfs.html.
DA Keen. Reverse Monte Carlo refinement of disordered silica phases. In Billinge, SJL and Thorpe, MF,
editor, LOCAL STRUCTURE FROM DIFFRACTION, FUNDAMENTAL MATERIALS RESEARCH,
pages 101–119. PLENUM PRESS DIV PLENUM PUBLISHING CORP, 1998. ISBN 0-306-45827-6.
E. Leonova. Structural investigation of Complex Glasses by Solid-state NMR. PhD thesis, Stockholm
University, 2009.
Leonid Lichtenstein, Christin Buechner, Bing Yang, Shamil Shaikhutdinov, Markus Heyde, Marek
Sierka, Radoslaw Wlodarczyk, Joachim Sauer, and Hans-Joachim Freund. The atomic structure of
a metal-supported vitreous thin silica film. Angewandte Chemie International Edition, 51(2):404–407,
2012. ISSN 1521-3773. doi: 10.1002/anie.201107097. URL http://dx.doi.org/10.1002/anie.
201107097.
P. Lucas. Mean coordination and topological constraints in chalcogenide network glasses. In Jean-Luc
Adam and Xianghua Zhang, editors, Chalcogenide Glasses, pages 58 – 81. Woodhead Publishing,
2014. ISBN 978-0-85709-345-5. doi: http://dx.doi.org/10.1533/9780857093561.1.58. URL http:
//www.sciencedirect.com/science/article/pii/B9780857093455500039.
M. Micoulaut. "Atomistic Modeling of Glass Structure and Glass Properties", 2013. URL http://www.
lptl.jussieu.fr/user/mmi/IMI.html.
Jcg Montoro and Jlf Abascal. The Voronoi Polyhedra as tools for structure determination in simple
disordered-systems. JOURNAL OF PHYSICAL CHEMISTRY, 97(16):4211–4215, APR 22 1993. ISSN
0022-3654. doi: {10.1021/j100118a044}.
Christian R. Muller, Vindu Kathriarachchi, Michael Schuch, Philipp Maass, and Valeri G. Petkov. Re-
verse monte carlo modeling of ion conducting network glasses: An evaluation based on molecular
dynamics simulations. Phys. Chem. Chem. Phys., 12:10444–10451, 2010. doi: 10.1039/C003472J.
URL http://dx.doi.org/10.1039/C003472J.
Michael O’Keeffe and Bruce G Hyde. Crystal structures. Courier Dover Publications, 2020.
K. J. Rao. Structural Chemistry of Glasses. Elsevier Science and Technology Books, 2002. ISBN
0080439586.
N Ramesh Rao, PSR Krishna, Saibal Basu, BA Dasannacharya, KS Sangunni, and ESR Gopal. Structural
correlations in gexse1- x glasses–a neutron diffraction study. Journal of non-crystalline solids, 240(1-
3):221–231, 1998.
BIBLIOGRAPHY 112

Sarah Rauscher, Chris Neale, and Régis Pomes. Simulated tempering distributed replica sampling, virtual
replica exchange, and other generalized-ensemble methods for conformational sampling. Journal of
chemical theory and computation, 5(10):2640–2662, 2009.

C Patrick Royall and Stephen R Williams. The role of local structure in dynamical arrest. Physics
Reports, 560:1–75, 2015.

Rachit Sapra, Ram P Verma, Govind P Maurya, Sameer Dhawan, Jisha Babu, and V Haridas. Designer
peptide and protein dendrimers: A cross-sectional analysis. Chemical reviews, 119(21):11391–11441,
2019.

J. E. Shelby. Introduction to Glass Science and Technology, 2nd edition. The Royal Society of Chemistry,
2005. ISBN 0-85404-639-9.

HW Sheng, WK Luo, FM Alamgir, JM Bai, and E Ma. Atomic packing and short-to-medium-range
order in metallic glasses. NATURE, 439(7075):419–425, JAN 26 2006. ISSN 0028-0836. doi: {10.
1038/nature04421}.

Giorgio F. Signorini. "Simulazione di molecole di interesse biologico", 2013. URL http://www1.chim.

unifi.it/u/signo/did/biomol/lezioni.pdf.
Giorgio F. Signorini, J.L. Barrat, and M.L. Klein. Structural relaxation and dynamical correlations
in a molten salt near the liquid-glass transition: a molecular dynamics study. J. Chem. Phys.,
92:1294–1303, 1990. URL http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.
1050.1201&rep=rep1&type=pdf.
C. Song, P. Wang, and H. A. Makse. A phase diagram for jammed matter. Nature, 453:629–632, May
2008. doi: 10.1038/nature06981.

P Steinhardt, R Alben, and D Weaire. Relaxed continuous random network models:(i). structural charac-
teristics. Journal of Non-Crystalline Solids, 15(2):199–214, 1974.

M.F. Thorpe. Continuous deformations in random networks. Journal of Non-Crystalline Solids, 57(3):
355 – 370, 1983. ISSN 0022-3093. doi: http://dx.doi.org/10.1016/0022-3093(83)90424-6. URL
http://www.sciencedirect.com/science/article/pii/0022309383904246.
Salvatore Torquato, Thomas M Truskett, and Pablo G Debenedetti. Is random close packing of spheres
well defined? Physical review letters, 84(10):2064, 2000.

A. K. Varshneya. Fundamentals of Inorganic Glasses. Academic Press, 1994. ISBN 0-12-714970-8.

P. Vashishta, R. K. Kalia, J. P. Rino, and I. Ebbsjö. Interaction potential for SiO2 : A molecular-dynamics
study of structural correlations. Phys. Rev. B, 41:12197–12209, Jun 1990. doi: 10.1103/PhysRevB.41.
12197. URL http://link.aps.org/doi/10.1103/PhysRevB.41.12197.

Mark Wilson. Model investigations of network-forming materials. Phys. Chem. Chem. Phys., 14:12701–
12714, 2012. doi: 10.1039/C2CP41644A. URL http://dx.doi.org/10.1039/C2CP41644A.

R. Zallen. The Physics of Amorphous Solids. A Wiley-Interscience publication. Wiley, 1998. ISBN
9780471299417. URL https://books.google.it/books?id=V0jr74rPdUIC.