Professional Documents
Culture Documents
2008 Chapter BasicAtomisticModeling
2008 Chapter BasicAtomisticModeling
2.1 Introduction
Molecular dynamics simulation techniques are very widely applicable, for
a range of materials and states, including gases, liquids, and solids. The
first molecular dynamics studies were focused on modeling thermodynamical
behavior of gases and liquids in the 1950s and 1960s [81–84]. It was not until
the early 1980s when the first molecular dynamics modeling works were pub-
lished that were applied to the mechanical behavior of solids. As a consequence
of the general applicability of molecular dynamics, many of the methods and
approaches described in this book can also be useful for the study of gases and
liquids, as well as the interaction of those with solids in systems that contain
both solids and liquids.
The outline of this chapter is as follows. We begin with a presentation of
the basic formulation of molecular dynamics (sometimes also referred to as
“MD”). After discussing of the numerical strategies associated with molec-
ular dynamics, we introduce interatomic potentials that mimic the energy
landscape predicted by quantum mechanics (we emphasize here that quan-
tum mechanics will not be discussed explicitly in this book, and the reader
32 Atomistic Modeling of Materials Failure
Fig. 2.1 Molecular dynamics can be used to study material properties at the inter-
section of various scientific disciplines. This is because the notion of a “chemical
bond” as explicitly considered in molecular dynamics provides a common ground as
it enables the cross-interaction between concepts used in different disciplines (here
exemplified for the disciplines of biology, mechanics, materials science and physics)
The significance of the atomistic viewpoint for failure processes and the enor-
mous computational burden associated with such problems makes modeling
and simulation of failure a promising and exciting area of research. In this
section we discuss some fundamental concepts associated with model build-
ing and the solution of the particular numerical problems to be computed in
molecular dynamics simulations.
The terms modeling and simulation are often used in conjunction with
the numerical solution of physical problems. However, it is important to note
that the two words have quite distinct meanings. The term modeling refers
2 Basic Atomistic Modeling 33
Fig. 2.2 This figure illustrates the concept of model building. Panel (a) on the left
shows the physical situation of a map of the subway lines. This representation makes
it quite difficult to determine a strategy to use the subway system to travel from
the cities of Braintree to Revere, for instance. The model representation depicted
in panel (b) on the right enables one to determine quite easily which subway line
to take, where to change the subway line, and how many subway stops there are
in between. This example illustrates that even though the model representation on
the right misrepresents the actual distances and directions, it elegantly displays the
connectivity. This figure was created based on a snapshot from the Massachusetts
Bay Transportation Authority (MBTA) web site (URL: http://www.mbta.com/),
reprinted with permission from the the Massachusetts Bay Transportation Authority
along with a definition of how atoms interact with each other (for a discussion
of Newton’s law, see Sect. 3.1). Despite this quite simple basis, very complex
phenomena can be simulated. Unlike many continuum mechanics approaches,
atomistic techniques require no a priori assumption on the defect dynamics
or its behavior.
Once the atomic interactions are chosen, the entire material behavior is
determined. This aspect of atomistic modeling provides a terrific opportunity
to build insightful models, that is, models that capture the essentials, elucidate
fundamental mechanisms, and thereby provide an elegant representation of
the key principles that underline the key observations. While in some cases it
is difficult to find an appropriate interatomic potential for a material, atomic
interactions can also be chosen such that generic properties common to a large
class of materials are incorporated (e.g., describing a general class of brittle or
ductile materials). This approach refers to the design of “model materials” to
study specific materials phenomena. Despite the fact that such model building
has been in practice in fluid mechanics for many years, the concept of “model
materials” in materials science is relatively new. On the other hand, atomic
interactions can also be chosen very accurately for a specific atomic interaction
using quantum mechanical methods such as the density functional theory
(DFT) [87], which enables one to approximate the solutions to Schrodinger’s
equation for a particular atomistic model.
Richard Feynman has also emphasized the importance of the atomistic
viewpoint in his famous Feynman’s Lectures in Physics [88], where he stated
that the atomic hypothesis (or the atomic fact, whatever you wish to call it)
that all things are made of atoms – little particles that move around in per-
petual motion, attracting each other when they are a little distance apart,
but repelling upon being squeezed into one another [...] provides an enormous
amount of information about the world, if just a little imagination and thinking
are applied.”
This underlines atomistic simulations as a natural choice to study materials
at a fundamental level. This is particularly true for studies of materials failure!
The atomistic level provides the most fundamental, sometimes referred to as
the ab initio, description of the failure processes. Many materials phenom-
ena are multiscale phenomena. For a fundamental understanding, simulations
should ideally capture the elementary physics of single atoms and reach length
scales of thousands of atomic layers at the same time. This can be achieved
by implementation of numerical models of atomistic models on very large
computational facilities.
In atomistic simulations, the goal is to predict the motion of each atom in the
material, characterized by the atomic positions ri (t), the atomic velocities
vi (t), and their accelerations ai (t) (see Fig. 2.3). Each atom is considered
as a classical particle that obeys Newton’s laws of mechanics. The collective
behavior of the atoms allows one to understand how the material undergoes
deformation, phase changes, or other phenomena by providing links between
the atomic scale to meso- or macroscale phenomena. Extraction of information
from atomistic dynamics can be challenging and typically involves methods
rooted in statistical mechanics.
The total energy of such as system is
H = K + U, (2.6)
where K is the kinetic energy of the entire system and U the potential energy.
The kinetic energy is a function of the kinetic energy of all N particles,
38 Atomistic Modeling of Materials Failure
1
N
K= mi v2i , (2.7)
2 i=1
and the total potential energy is the sum of the potential energy of all particles:
N
U (r) = Ui (r), (2.8)
i=1
noting that the potential energy of each particle Ui depends on the position
of itself and all other particles in the system, expressed by r = {ri } as defined
above. For now we leave this expression as an unknown.
The total energy H is also referred to as the Hamiltonian. We note that
K = K(p) and U = U (r), that is, the kinetic energy depends only on the
velocities or the linear momenta of the particles and the potential energy is a
function only of the position vectors.
To satisfy Newton’s law Fi = mi ai for each particle i in the system, the
following equation governs the dynamics of the system:
d2 ri dU (r)
mi 2
=− . (2.9)
dt dri
The right-hand side corresponds to the gradient of the potential energy, which
is the force (note that the potential energy of the system depends on the
positions of all particles r). Equation (2.9) represents a system of coupled
second-order nonlinear partial differential equations, corresponding to a cou-
pled system N -body problem for which no exact solution exists when N > 2.
2 Basic Atomistic Modeling 39
Equation (2.13) provides a direct link between new positions (at t0 + ∆t) and
the old positions and accelerations (at t0 ). The accelerations can be obtained
from the forces by considering Newton’s law,
Fi
ai = . (2.14)
m
This updating scheme is referred to as the Verlet central difference method.
There exist many other integration schemes that are frequently used
in molecular dynamics implementations. In the following few sections we
summarize a few additional popular algorithms.
Leap-Frog Algorithm
When the equations reviewed in the previous section are integrated, the
resulting thermodynamical ensemble is N V E, which means that the parti-
cle number N , the system volume V , and the total energy of the system E
remain constant throughout the simulation. Other thermodynamical ensem-
bles can be realized by modifying the equations of motion in an appropriate
way, leading to N V T or N P T ensembles. Table 2.1 shows an overview over
various thermodynamical ensembles.
Fig. 2.4 Schematic of the atomic displacement field as a function of time. The
atomic displacement field consists of a low-frequency (“coarse”) and high frequency
part (“fine”)
42 Atomistic Modeling of Materials Failure
A critical step in solving the dynamical equations (for instance, using the
Velocity Verlet scheme) is to consider the size of the time step. Figure 2.4
depicts a schematic of the atomic displacement field as a function of time.
As can be seen, the displacement history consists of low- and high-frequency
contributions, where the total displacements can be written as
with u (t) as the fine contribution and u(t) as the coarse part. To solve the
equations of motion, the fine part needs to be discretized, which results in
a significant computational burden as most systems require time steps of
approximately 1 fs or 10−15 s to discretize the rapid oscillations of u (t). Inter-
atomic bonds that involve relatively light hydrogen atoms sometimes require
even smaller time steps on the order of 0.1 fs.
There are also adaptive techniques that are based on the idea to dynami-
cally change the time step in a simulation, depending on the maximum atomic
velocities [18]. Such approaches may help to increase the efficiency of molecular
dynamics studies without adversely affecting the results.
Fig. 2.5 Example of harmonic oscillator with spring constant k = φ (r = r0 ), used
to extract information about the time step required for integration of the equations
of motion. The dashed line shows the (nonlinear) realistic potential function between
a pair of atoms, of which the harmonic oscillator is the second-order approximation.
To estimate the time step for a particular system, one can estimate the
oscillation frequency of a harmonic oscillator:
1 k
ν= (2.22)
2π m
2 Basic Atomistic Modeling 43
where k is the spring constant, given by the second derivative of the potential
function with respect to the diameter (k = φ (r = r0 )), evaluated at r = r0 .
The function φ describes how the energy of the bond changes as a function of
radius r. This is schematically shown in Fig. 2.5. The time step should then
be chosen
1
∆tmin (2.23)
ν
In summary, the time step ∆t needs to be small enough to model the vibra-
tions of atomic bonds correctly. The vibration frequencies may be extremely
high, in particular for light atoms as
√
∆tmin ∼ m (2.24)
The stiffer the bond is around its equilibrium position, the lower the critical
time step as
1
∆tmin ∼ √ . (2.25)
k
The fact that the time step is on the order of several femtoseconds has major
implications on the time scale molecular dynamics can reach. For example,
approximately 1,000,000 integration steps are needed to calculate a trajectory
that covers 1 ns, providing a severe computational burden. Further, the time
step typically cannot be varied during simulation. The total time scale that can
be reached by molecular dynamics is typically limited to a few nanoseconds.
Some exceptionally long simulations have been reported that cover up to a few
microseconds. However, such simulations typically take up to several months.
This aspect of molecular dynamics is sometimes referred to as the time
scale dilemma. Even though the number of atoms in a simulation can be
easily increased by adding more processors (e.g., using parallel computing),
time cannot easily be parallelized. As can be see in the updating scheme (2.13),
an atomistic system is generally not independent in time: The behavior at t0
influences the state at t1 > t0 , and the time stepping cannot be carried out
independently, on multiple processors.
Several researchers are currently developing techniques such as the tem-
perature accelerated dynamics method, parallel replica method, and many
others to overcome this limitation, and make use of massively parallelized
computing to expand the accessible time scales [89]. Some of these techniques
will be discussed later.
statics. Such approaches have been used to study dislocation nucleation from
crack tips or the deformation of carbon nanotubes. It mimics a quasistatic
experiment, albeit neglecting the effect of finite temperature.
A variety of algorithms exist to perform energy minimization, most notably
conjugate gradient methods or steepest descent methods. Figure 2.6 depicts an
example result of an energy minimization, showing how the potential energy
of the system decreases systematically with the number of iteration steps and
finally converges to a finite value.
Fig. 2.8 Summary of the Metropolis–Hastings Monte Carlo algorithm. Please see
Figure 2.7 for an illustration of how state B is generated based on a random pertur-
bation from state A. The procedure is repeated NA times, the number of iterations.
The number of steps is chosen so that convergence of the desired property is achieved
Fig. 2.9 Schematic of the typical characteristic of a chemical bond, showing repul-
sion at small distances below the equilibrium separation r0 , and attraction at larger
distances
featuring many different chemical bonds with varying strength, such as exem-
plified in materials including cement, proteins, or ceramics, or at interfaces
between metallic systems and organic components.
Despite the differences between different chemical bonds, many atom–atom
interactions show similar characteristic featyres. Figure 2.9 depicts the typical
characteristic of a chemical bond, showing repulsion at small distances below
the equilibrium separation r0 of a pair of atoms, and attraction at larger
distances.
In general, for any material we must consider the interplay of chemical
interactions that include, ordered by their approximate strength:
• Covalent bonds (due to overlap of electron orbitals, e.g., found in carbon
nanotubes, C–C bond, organic molecules such as proteins)
• Metallic bonds (found in all metals, e.g., copper, gold, nickel, silver)
• Electrostatic (ionic) interactions (Coulombic interactions, e.g., found in
ceramics such as Al2 O3 or in SiO2 )
• Hydrogen-bonds (e.g., found in polymers, proteins), as well as
• Weak or dispersive van der Waals (vdW) interactions (e.g., found in wax).
Electrostatic interactions can be significantly weakened by screening due
to electrolytes, which can lead to interactions that are weaker than vdW
interactions.
48 Atomistic Modeling of Materials Failure
Fig. 2.10 Atoms are composed of electrons, protons, and neutrons. Electrons and
protons are negative and positive charges of the same magnitude. In classical molec-
ular dynamics, the three-dimensional atom structure is replaced by a single mass
point
Fig. 2.11 Overview over different simulation tools and associated lengthscale and
timescale
There is no single approach that is suitable for all materials and for all
materials phenomena. The choice of the interatomic potential depends very
strongly on both the application and the material. Popular choices in par-
ticular for modeling mechanical properties of materials are semiempirical
or empirical methods, which typically allow one to simulate large systems
with many thousands to billions of particles. However, to address different
aspects of the mechanical behavior of a specific material typically requires the
application of a range of simulation approaches.
An overview over the most prominent materials simulation techniques
is shown in Fig. 2.11. In the plot we also indicate which lengthscale and
timescale can be reached with the various methods. The methods included
in the figure refer to quantum mechanics based methods, classical molecular
50 Atomistic Modeling of Materials Failure
1
N N
Utotal = φ(rij ), (2.27)
2 j=1
i=j=1
where rij is the distance between particles i and j. Note the factor 1/2 to
account for the double counting of atomic bonds. The procedure of summing
up the energies is shown in a schematic in Fig. 2.12.
The term φ(rij ) describes the potential energy of a bond formed between
two atoms, as a function of its distance rij . How can one obtain expressions
for φ? A possible approach is sketched in Fig. 2.13, showing how a full-electron
representation of the pair of atoms is reduced to a case where two point par-
ticles interact. The energy–distance relationship must be identical in both
cases. An approach often used is to carry out quantum mechanical calcula-
tions that provides the relationship between distance and energy of a pair of
2 Basic Atomistic Modeling 51
1
2 5
1
2 5
4
Fig. 2.12 Pair interaction approximation. The upper part shows all pair interactions
of atom 1 with its neighbors, atoms 2, 3, 4, and 5. When the bonds to atom 2 are
considered, the energy of the bond between atoms 1 and 2 is counted again (bond
marked with thicker line). This is accounted for by adding a factor 1/2 in (2.27)
atoms, which is then used to determine the parameters of the pair potential
expression. Pair potentials must capture the repulsion at short distances due
to the increasing overlap of electrons in the same orbitals, leading to high
energies due to the Pauli exclusion principle. At large distances, the poten-
tial must capture the effect that atoms attract each other to form a bond. In
many pair potentials, two separate terms are used to describe repulsion and
attraction, and the sum of these repulsive and attractive interactions yield
the total energy dependence on the radius:
φ = φRepulsion + φAttraction . (2.28)
A pair of atoms is in the equilibrium position at a balance between the
attractive and repulsive terms.
Pair potentials are the simplest choice for describing atomic interactions.
Even so, in some materials the interatomic interactions are best described
by pair potentials (because the underlying quantum mechanical governing
equations actually predict such a behavior). Prominent examples include the
noble gases (e.g., argon, neon) [83] as well as Coulomb interactions due to
partial charges. Pair potentials have also proven to be a reasonable model
for more complex materials such as SiO2 [102]. The potential energy of an
individual atom is given by
52 Atomistic Modeling of Materials Failure
Ni
Ui = φ(rij ) (2.29)
j=1
The LJ potential can be fitted to the elastic constants and lattice spacing
of metals (however, this model has some shortcomings with respect to the
stacking fault energy and the elasticity of metals). The term with power 12
represents atomic repulsion, and the term with power 6 represents attractive
interactions. The parameter σ scales the length and 0 the energy of atomic
bonds. Often, pair potentials are cutoff smoothly with a spline cutoff function
(see for instance [104] or [105]).
For the LJ potential, the equilibrium distance between atoms (denoted as
r0 ) is given by √
r0 = σ 6 2. (2.31)
The maximum force between two atoms is
2.3940
Fmax,LJ = . (2.32)
σ
Figure 2.14 shows a plot of the LJ potential (dotted line) and its deriva-
tive (continuous line, describing the interatomic forces), in a parametrization
2 Basic Atomistic Modeling 53
Fig. 2.14 Plot of the LJ potential and its derivative (for interatomic forces) in a
parametrization for copper as reported in [9]
1
φ(rij ) = a0 + k(rij − r0 )2 (2.34)
2
54 Atomistic Modeling of Materials Failure
The concept that the total energy of the system is simply a sum over the energy
contributions between all pairs of atoms in a system is a great simplification
that leads to great challenges. For example, at a surface of a crystal, the
atomic bonds may have different properties than in the bulk. Pair potentials
are not capable of capturing this effect. The limitation of pair potentials to
model more complex situation, in particular the dependence of the properties
of chemical bonds between pairs of atoms on the environment, is sketched
in Fig. 2.15. This behavior is particularly important for metals, because of
quantum mechanical effects that describe the influence of the electron gas.
Fig. 2.15 Difference in bond properties at a surface. Pair potentials (left panel) are
not able to distinguish bonds in different environments, as all bonds are equal. To
accurately represent the change in bond properties at a surface, one needs to adapt
a description that considers the environment of an atom to determine the bond
strength, as shown in the right panel. The bond energy between two particles is then
no longer simply a function of its distance, but instead a function of the positions
of all other particles in the vicinity (that way, changes in the bond strength, for
instance at surfaces, can be captured). Multibody potentials (e.g., EAM) provide
such a description
longer simply a function of its distance, but instead a function of the posi-
tions of all other particles in the vicinity. This behavior can be captured
in multibody potentials. The idea behind multibody potentials is to incor-
porate more specific information on the bonds between atoms than simply
the distance between two neighbors. In such potentials, the energy of bonds
therefore depends not only on the distance of atoms, but also on its local
environment, that is, on the positions of neighboring atoms. In the case of
metals, the interactions of atoms can be quite accurately described using
potentials based on the embedded atom method (EAM), or other so-called
n-body potentials (e.g., [94, 107]. Several variations of the classical EAM
potentials exist [108, 109]). Another, similar approach is based on effective
medium theory (EMT) [110, 111]. Particularly appropriate models have been
reported for metals such as copper and nickel. Other metals (e.g., aluminum)
are more difficult to model with such approaches [112, 113].
An EAM potential for metals is typically given in the form
Ni
Ui = φ(rij ) + f (ρi ), (2.35)
j=1
where ρi is the local electron density and f is the embedding function. The
electron density ρi depends on the local environment of the atom i, and
the embedding function f describes how the energy of an atom depends on
the local energy density. The electron density itself is typically calculated
based on a simple pair potential that maps the distance between atoms to
the corresponding contribution to the local electron density. The potential
features a contribution by a two-body term φ to capture the basic repulsion
and attraction of atoms (just like in LJ or Morse potentials), in conjunction
with a multibody term that accounts for the local electronic environment of
the atom.
Overall multibody potentials allow a much better reproduction of the
elastic properties of metals than pair potentials (e.g., [109]). For instance,
most real materials violate the Cauchy relation (that is, the condition that
c1122 = c1212 ). Any pair potential predicts an agreement with the Cauchy rela-
tion [109]. Multibody potentials are capable of reproducing the appropriate
elastic behavior. Figure 2.16 illustrates how an EAM-type multibody potential
can represent different effective pair interactions between bonds at a surface
and in the bulk.
However, most conventional multibody potentials are not capable of mod-
eling any effect of directional bonding. Whereas this is not important for
metals such as Ni or Cu, this becomes quite significant in materials with a
more covalent character of the interatomic bonding. To address these effects,
modified embedded atom potentials (MEAM) have been proposed that can
be parameterized, for instance, for silicon [114].
56 Atomistic Modeling of Materials Failure
Fig. 2.16 This plot illustrates how an EAM-type multibody potential can represent
different effective pair interactions between bonds at a surface and in the bulk
The bases for simulations of polymers, organic substances, or proteins are force
fields that describe the various chemical interactions based on a combination
of energy terms. This is required since in these materials, the set of char-
acteristic chemical bonds is much more diverse than in metals, for instance,
which requires the explicit consideration of ionic, covalent, and vdW interac-
tions. Figure 2.17 illustrates this chemical complexity, exemplified in a small
alpha-helical coiled coil protein domain.
A prominent example for this approach is the classical force field CHARMM
[115]. The CHARMM force field is widely used in the protein and biophysics
community, and provides a reasonable description of the behavior of proteins.
This force field is based on harmonic and anharmonic terms describing
covalent interactions, in addition to long-range contributions describing vdW
interactions, ionic (Coulomb) interactions, as well as hydrogen bonding. Since
the bonds between atoms are modeled by harmonic springs or its variations,
bonds (other than H-bonds) between atoms cannot be broken, and new bonds
cannot be formed. Also, the charges are fixed and cannot change, and the
equilibrium angles do not change depending on stretch. The CHARMM force
field belongs to a class of models with similar descriptions of the interatomic
forces; other models include the DREIDING force field [116], the UFF force
field [117], the GROMOS force field, or the AMBER model (see [118], for
instance, for a review of various force fields for biological systems).
In the CHARMM model, the mathematical formulation for the empirical
energy function that contains terms for both internal and external interactions
2 Basic Atomistic Modeling 57
Fig. 2.18 Schematic of the contributions of the different terms in the potential
expressions given in (2.36), illustrating the contributions of bond stretching, angle
bending, bond rotations, electrostatic interactions, and vdW interactions
58 Atomistic Modeling of Materials Failure
and
Utorsion = φtorsion . (2.41)
quadruples
Due to the harmonic approximation, these expressions are only valid for
small deformation from the equilibrium configuration of the bond. Large
deformation or fracture of these bonds cannot be described. Figure 2.18
schematically illustrates the energy contributions provided in (2.36).
The Coulomb energies are evaluated between pairs of atoms and are
described as
qi qj
φCoulomb = (2.42)
1 rij
where qi and qj are the partial charges of atoms i and j, and 1 is the effective
dielectric constant (1 = 1.602 × 10−19 C for vacuum). The total Coulomb
energy is then given by
2 Basic Atomistic Modeling 59
N N
qi qj
UCoulomb = (2.43)
r
j=1 1 ij
j=i=1
Table 2.2 Overview over force fields suitable for organic substances
atoms is not constant, but depends on the local environment, similar to the
EAM approach (please see [119] for a discussion). However, specific terms
are included to specify the directional dependence of the bonding. The idea
is based on the concept of mapping bond distances to bond orders, which
enables one to determine specific quantum chemical states of a molecular
structure. The concept of bond order was initially introduced by Pauling
[120]. Very well-known models in this class of potentials are the reactive
bond order potentials (REBO) [121, 122], the Tersoff potential [10, 123], the
Stillinger–Weber potential [124], Brenner’s force fields [125, 126], the Stu-
art reactive potential [127], and more recently, the ReaxFF reactive force
field [93].
Abell–Tersoff Approach
The basic concept of bond order potentials is simple to explain. The key idea
is to modulate the bond strength based on the atomic environment, taking
advantage of some theoretical chemistry principles.
Instead of expressing φ(rij ) as a harmonic function or an LJ function (see
above), in the Abell–Tersoff approach the interaction between two atoms is
expressed as
Fig. 2.19 The plot shows the cohesive energy per atom (upper plot, in eV) and
the bond length (lower plot, in Å), for several real and hypothetical polytypes of
carbon, comparing the predictions from the Tersoff potential [10] for C with experi-
mental and other computational results. The structures include a C2 dimer molecule,
graphite, diamond, simple cubic, BCC, and FCC structures. The squares correspond
to experimental values for these phases and calculations for hypothetical phases [11].
The circles are the results of Tersoff’s model [10]. The continuous lines are spline
fits to guide the eye. Reprinted from: J. Tersoff, Empirical interatomic potentials
for carbon, with applications to amorphous carbon, Physical Review Letters, Vol.
61(25), 1988, pp. 2879–2883. Copyright c 1988 by the American Physical Society
the Pauling relationship between bond length and bond order. Abell suggested
that
Mij ∼ Z −δ (2.45)
where δ depends on the particular system and Z is the coordination number of
atom i that depends on the bond radius. For pair potentials, Mij ∼ Z, which
is not true for many real materials. The Abell–Tersoff approach provides a
realistic model for these effects.
In the Tersoff potential [10, 123], the many-body term depends explicitly
on the angles of triplets of atoms, in addition to considering the effect of
coordination. Thus, this and related potentials are also referred to as three-
body potentials. The explicit angular dependence also illustrates the difference
to EAM potentials: Here, the multibody term solely depends on the elec-
tron density, and not on any directional information. Tersoff has successfully
parametrized his potential approach for carbon, silicon, and other semicon-
ductors. Figure 2.19 shows the cohesive energy per atom for several real and
62 Atomistic Modeling of Materials Failure
Fig. 2.20 An example to demonstrate the basic concept of the ReaxFF potential.
It has been developed to accurately describe transition states in addition to ground
states
2 Basic Atomistic Modeling 63
torsion are included in Utors . The term Uconj describes energetic contributions
of resonance effects. A maximum contribution of the conjugation energy is
obtained when successive bonds have bond order values of 1.5, as it is the
case in benzene, for instance. H-bonds are treated in the term UH-bonds , and
its interactions are calculated between groups X–H and Y, where X and Y
are atoms that can form H-bonds (for instance, N, O). Nonbonded two-body
interactions are included in UvdW and in UCoulomb . They are included for all
atom pairs whether they are bonded or nonbonded. This is important to avoid
energy discontinuities when chemical reactions occur. To enable the calcula-
tion for these interactions for all atom pairs, a shielded Coulomb potential of
the form
Ni Ni
qi qj
φi = 3 + γ −3 )1/3
(2.48)
i=1 j=1
1 (rij ij
This correction refers to the fact that carbon, for instance, cannot have more
than four bonds, or hydrogen cannot have more than one bond. This expres-
sion illustrates that the bonding term is a multibody expression, as it depends
on all j neighbors of an atom i.
Once the bond orders are known, the energy contributions can be calcu-
lated. For the term considered here,
φbond, ij = −De BOij exp pbe,1 1 − BOpbe,1 ij , (2.51)
where De and pbe,i are additional bond parameters. The total bond energy is
then given by a summation over all bonds in the system,
2 Basic Atomistic Modeling 65
Fig. 2.21 Illustration of basic concept of bond order potentials. Subplot (a) shows
how the bond order potential allows for a more general description of chemistry, since
all energy terms are expressed dependent on bond order. In contrast, conventional
potentials (such as LJ, Morse) express the energy directly as a function of the bond
distance as shown in subplot (b). Subplot (c) illustrates the concept for a C–C single,
double, and triple bond, showing how the bond distance is used to map to the bond
order, serving as the basis for all energy contributions in the potential formulation
defined in (2.47)
Ubond = φbond,ij (2.52)
all bonds ij
Equation (2.51) also illustrates that the energy contributions vanish when the
bond order goes to zero, which corresponds to a broken chemical bond.
All other terms are also expressed as a function of bond orders. For
instance, the angle contributions are given as
Fig. 2.22 Results of a ReaxFF study of water formation, comparing the production
rate with and without a Pt catalyst. The presence of the Pt catalyst significantly
increases the water production rate (results taken from [12])
An example simulation is shown in Fig. 2.22 [12]. This plot depicts the
results of a ReaxFF study of water formation, comparing the production rate
with and without a Pt catalyst (same pressure and temperature, the only
difference is the catalyst). It is evident that the presence of the Pt catalyst
significantly increases the water production rate.
2 Basic Atomistic Modeling 67
Fig. 2.23 Water production at varying temperature, for constant pressure. Subplot
(a) depicts the water production rate. Subplot (b) shows an Arrhenius analysis,
enabling us to extract the activation barrier for the elementary chemical process of
12 kcal/mol. This result is close to DFT level calculations of the energy barrier [12]
68 Atomistic Modeling of Materials Failure
Figure 2.23 depicts an analysis of the system with Pt catalyst, for vari-
ations of the temperature. These simulations involves thousands of reactive
atoms, a computational task that cannot be achieved using DFT or simi-
lar approaches. Thus, the ReaxFF approach provides a very useful bridge
between quantum mechanical methods and empirical potentials, as illus-
trated in Fig. 2.24. ReaxFF simulations have been reported with system sizes
approaching millions of atoms, as recently reported in the ReaxFF parallelized
algorithm [133].
Fig. 2.24 The ReaxFF force field fills a gap between quantum mechanical methods
(e.g., DFT) and empirical molecular dynamics
Fig. 2.25 Schematic of the numerical scheme in carrying out molecular dynamics
simulations
Figure 2.25 depicts the basic numerical scheme of carrying out a molecular
dynamics study. The basic steps of a molecular dynamics simulation are
• Define initial conditions and boundary conditions (including positions and
velocities at t = t0 ); typically the velocities of particles are drawn from a
Maxwell–Boltzmann distribution.
70 Atomistic Modeling of Materials Failure
Fig. 2.26 Schematic of the numerical scheme in carrying out molecular dynamics
simulations
We recall that at each integration step, forces are required to obtain acceler-
ations. Forces are calculated from the positions of all atoms, by considering
the atomic potential energy surface U (r) that depends on the positions of all
atoms. According to (2.9), the force vector is given by taking partial deriva-
tives of the potential energy surface with respect to the atomic coordinates of
the atom considered,
dU (r)
Fi = − . (2.54)
dri
Fig. 2.27 Schematic of force calculation scheme in molecular dynamics, for a pair
potential. To obtain the force vector F one takes projections of the magnitude of
the force vector F into the three axial directions xi (this is done for all atoms in the
system)
For the special case where the interatomic forces are described by a poten-
tial φ that depends only on the distance between pairs of atoms, here denoted
by r, the contributions to the total force vector due this particular interaction
can be obtained by taking projections of the magnitude of the force vector F
into the three axial projections ri of the vector between the pair of particles.
The magnitude of the force vector due to interactions of pairs of atoms is then
given by
dφ(r)
F = , (2.55)
dr
where F =| F |. The ith component of the force vector is then given by
ri
Fi = −F . (2.56)
r
Figure 2.27 depicts this approach schematically.
In principle, all atoms in the system interact with all other atoms, which
requires a nested loop for calculation of the force vectors of all atoms. This
renders the total computational time requirement to solve the problem second
order with respect to the number of particles N ,
72 Atomistic Modeling of Materials Failure
Fig. 2.28 Use of neighbor lists and bins to achieve linear scaling ∼N in molecular
dynamics. Panel (a): Example of how neighbor lists are used. The four neighbors
of the central atom (in the circle) are stored in a list so that force calculation can
be done directly based on this information. This changes the numerical problem to
a linear scaling effort. Panel (b): The computational domain is divided into bins
according to the physical position of atoms. Then, atomic interactions must only be
considered within the atom’s own bin and atoms in the neighboring bins
ttot ∼ N 2 . (2.57)
The first loop goes over all atoms, and the second loop goes through all
possible neighbors of each atom. The following pseudocode illustrates this
process:
for i=1 to N # loop over all atoms i
for j=1 to N (i not equal to j) # second loop over all atoms neighboring
atom i
r=distance(i,j) # calculate distance between atoms i and j
F=f(r) # calculate force depending on radius
that are the nearest, second nearest, and so forth neighbors of each particle.
This method is used to save time from checking every particle in the system
every time a force calculation is made. The list can be used for several time
steps before updating. Each update is expensive since it involves N × N oper-
ations for an N -particle system. In low-temperature solids where the particles
do not move very much, it is possible to perform an entire simulation without
or with only a few updating, whereas in simulation of liquids, updating every
5 or 10 steps is quite common. Figure 2.28a shows a schematic of how neighbor
lists are used. We note that neighbor lists can only be implemented if particles
interact only up to a certain cutoff radius; for very long range interactions,
the definition of neighbor lists may not be feasible.
An alternative to generation of neighbor lists is the decomposition of
the computational domain into bins. The size of the bin is chosen compa-
rable to the cutoff radius of the potential, so that atomic interactions must
only be considered within the atom’s own bin and the neighboring bins (see
Fig. 2.28b).
2 K
T = . (2.58)
3 N kB
Note that the numerical value for the Boltzmann constant kB = 1.3806503 ×
10−23 m2 kgs−2 K−1 , relating energy and temperature at the level of individual
atoms or particles (its units are energy per absolute temperature). Please see
also Table 4.7 for other units and their conversion to SI units.
74 Atomistic Modeling of Materials Failure
The first term stems from the kinetic contributions of particles hitting the
wall of the container, and the second term stems from the interatomic forces,
expressed as the force vector contribution multiplied by the distance vector
component between particles i and j.
r+ r r+ 2
r
2
N (r ± ∆r
2 )
g(r) = , (2.61)
Ω(r ± 2 )ρ0
∆r
where N (r ± ∆r
2 ) is the number of particles in the volume shell with volume
Ω(r ± ∆r
2 ).
2 Basic Atomistic Modeling 75
Fig. 2.30 Radial distribution function g(r) for various atomistic configurations,
including a solid (crystal), a liquid and a gas
1
N
∆r2 = (ri (t) − ri (t = 0))2 . (2.62)
N i=1
If averaged over all particles, the mean square displacement function pro-
vides the square distance that particles have moved during time t. The mean
square displacement function is zero at t = 0, and it grows with increasing
time. In a solid, it is expected that the mean square displacement function
grows to a characteristic value and then saturates at a constant value. In
a liquid, all atoms move continuously through the material, as in Brownian
motion. Diffusivity in liquids is related to the linear variation of the mean
square displacement function with time t. The diffusivity D can be calculated
as
1 d 2
D= lim ∆r , (2.63)
2d t→∞ dt
where d = 2 in 2D and d = 3 in 3D (the parameter d describes the
dimensionality).
76 Atomistic Modeling of Materials Failure
1 1
N N
v(0)v(t) = vi (t0 )vi (t0 + t), (2.64)
N i=1 N j=1
VAF
Ideal gas
1
Dense gas
Liquid
T
Solid
Fig. 2.31 Velocity autocorrelation function (VAF) for a gas, liquid, and solid
Fig. 2.32 Relating the continuum stress with the atomistic stress. The left shows
a continuum system in which σij (r) is defined at any point r. In contrast, in the
atomistic system the stress tensor is only defined at discrete points where atoms are
located
Note that the pressure P as discussed in Sect. 2.8.2 is a special case of the
stress tensor definition. The virial stress needs to be averaged over space and
time to converge to the Cauchy stress tensor. For further discussion on the
virial stress and other definitions of the Cauchy stress tensor (e.g., the Hardy
stress) we refer the reader to a review article [136].
F1
1 2 3
F2
r21 r23
Fig. 2.33 Example of how to calculate the stress tensor in a 1D system
Figure 2.33 depicts a schematic that shows how the stress tensor can be
calculated in a 1D system, for a nearest neighbor pair potential. Neglecting
the kinetic contribution (v ≈ 0) and with
dφ(r) ri
F =− (2.68)
dr r
as the force between two particles, the stress tensor coefficient σ11 is
11
σ11 = (F1 r21 + F23 r23 ) . (2.69)
V 2
huge challenges for data analysis and visualization, or just for data handling
and storage. However, for many properties of materials it is not necessary to
consider 1023 atoms. Many thermodynamical properties of solids or liquids
can be accurately described in systems of thousands of atoms or less, in some
instances. Many mechanical properties require larger system sizes, containing
at least tens of thousands to billions of atoms.
The need for military applications has strongly driven the development of
supercomputers. (This includes computers operated by the U.S. Department
of Energy, DoE, or the U.S. National Science Foundation, which are among the
most powerful supercomputers of the world. DoE computers are for instance
being used to maintain the U.S. stockpile of nuclear weapons.) Many other
large computational centers have been established in recent years in Europe
and Asia, including Japan’s Earth Simulator (which, for some time, was the
most powerful computer on Earth).
Early supercomputers, such as the Almaden Spark that was used in the
1960s were capable of simulating 100s of atoms. GFLOPS computing enabled
the simulation of millions of particles in the mid 1990s. TFLOPS systems led
to first simulations of systems that exceed one billion. The current state-of-the
art in atomistic simulation reaches system sizes of tens to hundreds of bil-
lions of atoms, corresponding to several micrometer-sized physical dimensions
80 Atomistic Modeling of Materials Failure
[133, 137–141] (see Fig. 2.34). However, only few computational groups or
scientists routinely carry out such large simulations.
Fig. 2.34 Increase in computer power over the last decades and possible system sizes
for classical molecular dynamics modeling. The availability of PFLOPS computers
is expected by the end of the current decade, which should enable simulations with
hundreds of billions of atoms
Fig. 2.35 Summary of top 10 of the TOP500 supercomputer list, as of Spring 2008
with a given particle, only the cell in which it is located and the next-nearest
neighbors have to considered. Since placing the particles in the correct cells
scales linearly with N , the problem originally scaling with N 2 can therefore be
reduced to N . With a parallel computer whose number of processors increases
with the number of cells (the number of particles per cell does not change),
the computational burden remains constant.
The speedup factor S is defined as the ratio of execution time on one
processor (Ts ) over the execution time on p processors (Tp ):
Ts
S= . (2.70)
Tp
The perfectly efficient parallel computer would exhibit linear speedup. This
would mean that the computation time for p processors is 1/p times the
execution time on one processor. However, the speedup depends strongly on
the fraction of the work done in parallel. We refer the reader to Plimpton’s
algorithms for molecular dynamics with short-range forces [143].
Figure 2.36 depicts a schematic and scaling results of a modern paralleliza-
tion scheme, referred to as the tunable hierarchical cellular decomposition
scheme (THCD) (see Fig. 2.36a) [133]. The THCD involves a more com-
plex domain decomposition that includes a hierarchy of parameterized cell
data/computation structures, and adaptive load balancing through wavelet-
based computational-space decomposition. The THCD was applied to paral-
lelize the ReaxFF reactive force field approach (referred to as F-ReaxFF). This
parallelization scheme enabled the simulation of 0.56 billion-atom F-ReaxFF
systems. Figure 2.36b depicts the total execution (circles) and communica-
tion (squares) times per molecular dynamics time step as a function of the
number of processors, illustrating that the parallel speedup is almost perfect
82 Atomistic Modeling of Materials Failure
Fig. 2.36 Modern parallelization scheme. Subplot (a) depicts the schematic of the
tunable hierarchical cellular decomposition scheme (THCD). The physical volume
is subdivided into process groups, PGγ , each of which is spatially decomposed into
processes, Pγπ . Each process consists of a number of computational cells (e.g., linked-
list cells in molecular dynamics). Subplot (b) shows the total execution (circles) and
communication (squares) times per molecular dynamics time step as a function of the
number of processors for the F-ReaxFF molecular dynamics algorithm with scaled
workloads (in a 36,288P atom RDX systems on P processors (P = 1, . . . , 1920)
of Columbia [Columbia is a supercomputer at NASA]). Reprinted from Computa-
tional Materials Science, Vol 38(4), A. Nakano, R. Kalia, K. Nomura, A. Sharma, P.
Vashishta, F. Shimojo, A. van Duin, W.A. Goddard III, R. Biswas and D. Srivastava,
A divide-and-conquer/cellular-decomposition framework for million-to-billion atom
simulations of chemical reactions, pp. 642–652, copyright c 2007, with permission
from Elsevier
2.9.3 Discussion
We emphasize, however, that the “size” of the simulations does not determine
how “useful” a simulation is by itself. Instead, the most important issue and
measure for a successful simulation is always the physics that can be extracted
from the simulation. This objective should dictate the system size. In many
cases, such as for dislocation–dislocation interaction, system sizes on the order
of micrometers are needed (dislocation interaction is associated with a char-
acteristic length scale of micrometers). This example illustrates that there is
still a need for the development of simulation techniques and more computer
power.
Future development using cheap off-the-shelve technology based on LINUX
clusters to build supercomputers (instead of using very expensive UNIX-based
supercomputers) is promising, as indicated by recent publications [68,69] and
the analysis of data in the TOP500 list.
2 Basic Atomistic Modeling 83
real-time interaction of users with particles in the simulation. For example, the
biophysics group around Klaus Schulten has set up a system where scientists
can interact with the simulation by using a tool to manipulate molecules [145].
The researchers implemented a system called Interactive Molecular Dynamics
(IMD). This system allows manipulation of molecules in molecular dynamics
simulations with real-time force feedback, as well as graphical display. Com-
munication is achieved through an efficient socket connection between the
visualization program (VMD) and a molecular dynamics program (NAMD)
running on single or multiple machines. In this method, a natural force
feedback interface for molecular steering is provided by a haptic device [145].
For the analysis of atomistic simulations focused on mechanical properties,
measures like strain, stress, or potential energy of atoms are important quan-
tities that can be used to analyze atomistic data, in particular with respect
to continuum mechanics theories. However, it is often advantageous to post-
process the data and derive new quantities providing more information of the
defect structure. Here we discuss a few examples for the analysis of crystal
defects in metals that will become particularly important in the third part of
this thesis.
Richard Hamming’s statement the purpose of scientific computing is
insight, not numbers underlines the importance of processing the raw
simulation data appearing as “useless” numbers to gain understanding.
2 Basic Atomistic Modeling 85
Fig. 2.38 Analysis of a dislocation network using the energy filtering method in
nickel with 150,000,000 atoms [13, 14]. Subplot (a) shows the whole simulation cell
with two cracks at the surfaces serving as sources for dislocations, and subplot (b)
shows a zoom into a small subvolume. Partial dislocations appear as wiggly lines,
and sessile defects appear as straight lines with slightly higher potential energy
To visualize crystal defects, the easiest approach is to use the energy method.
This method has frequently been used to “see” into the interior of the solid
(e.g., [138, 146]). In this method, only those atoms with potential energy
greater than or equal to a critical energy φcr above the bulk energy φb are
shown. The energy method has been very effective for displaying disloca-
tions, microcracks, and other imperfections in crystal packing. This method
reduces the number of atoms being displayed by approximately two orders of
magnitude in three dimensions [138].
An example of an analysis of a dislocation network using the energy
method is shown in Fig. 2.38. Figure 2.38a shows the whole simulation cell
with two cracks at the surfaces serving as sources for dislocations, and
Figure 2.38b shows a zoom into a small subvolume of the material revealing
a complex dislocation microstructure. The data are taken from a simula-
tion of work-hardening in nickel that comprises of approximately 150,000,000
atoms [13].
Assuming a crystal defect is identified as a dislocation, it can be studied
in more detail based on a geometric analysis of the lattice close to the dislo-
cation core allowing to determine the Burgers vector and the slip plane. For
that purpose, one can rotate the atomic lattice such that one is looking onto a
86 Atomistic Modeling of Materials Failure
{111}-plane, with the horizontal (x) axis oriented into a 110 direction, and
the vertical (y) axis aligned with a 111 direction. To help visualizing disloca-
tions, stretching the atomic lattice by a factor of 5 to 10 in the 110 direction
is helpful. A systematic rotation of the atomic lattice to investigate all possible
Burgers vectors is then necessary. Instead of analyzing a part of the atomic
lattice containing many dislocations, one can choose a domain of the atomic
lattice which contains only one dislocation. This approach requires a very
detailed understanding of the lattice and dislocations [38, 60]. This method
of analysis is similar to the analysis of TEM images from “real” laboratory
experiments.
Fig. 2.40 The figure shows a close view on the defect structure in a simulation of
work-hardening in nickel analyzed using the centrosymmetry technique [13,14]. The
plot shows the same subvolume as in Figure 2.38b
its nearest neighbors. During deformation, bonds will change direction and/or
length, but they remain equal and opposite within the same pair. This rule
breaks down when a defect is close to an atom under consideration. The
centrosymmetry method is particularly helpful to separate different types of
defects from one another, and to display stacking faults (in contrast, using the
energy method it is difficult to observe stacking faults). The centrosymmetry
parameter for an atom is defined as [36]
3
6
ci = | rk,j + rk,j+6 |2 , (2.71)
j=1 k=1
where rk,j is the kth component of the bond vector (here, k = 1, 2, 3 corre-
sponding to the directions x, y, and z) of atom i with its neighbor atom j,
and rk,j+6 is the same quantity with respect to the opposite neighbor in a
FCC crystal. We summarize the interpretation of ci in Table 2.4 (assuming
that the nearest neighbor distance does not change near a defect). For the
analysis, it is reasonable to display ranges of these parameters. The method
can also be applied at elevated temperature, which is not possible using the
energy method due to the thermal fluctuation of atoms.
Fig. 2.41 Analysis of a dislocation using the slip vector approach. From the result of
the numerical analysis, direct information about the Burgers vector can be obtained.
The slip vector s is drawn at each atom as a small arrow. The Burgers vector b is
drawn at the dislocation (its actual length is exaggerated to make it better visible).
The dislocation line is approximated by discrete, straight dislocation segments. A
line element between “a” and “b” is considered
from an energy analysis, and the line direction of a segment between point
“a” and “b” of the dislocation line is indicated by the vector l. The Burgers
vector b is given by the slip vector s directly. The analysis reveals that the
dislocation has Burgers vector b = 16 [112]. The unit vector of line direction of
the segment is l ≈ [−0.3618 0.8148 − 0.5530]. The length of the line segment
is approximately 9 nearest neighbor distances in the [110] direction. The slip
plane normal is given by the cross product ns = l × b ∼ [111], and the
dislocation thus glides in the (111) plane.
Table 2.5 Distinguishing modeling and simulation, for tasks associated with
classical molecular dynamics
Modeling Simulation
Mathematical model of physical system ×
Numerical solution of problem ×
Choice of model geometry ×
Choice of numerical integrator (e.g., Verlet) ×
Choice of interatomic potential and parameters ×
Force calculation ×
Choice of boundary conditions ×
Implementation of boundary conditions ×
Choice of system size ×
Choice of thermodynamical ensemble ×
Algorithm to specify the thermodynamic ensemble
(e.g., temperature control, pressure control) ×
Fig. 2.43 Simulation method of domain decomposition via the method of virtual
atom types. The atoms in region 2 do not move according to the physical equations
of motion, but are displaced according to a prescribed displacement history. An
initial velocity gradient as shown in the right half of the plot is used to provide
smooth initical conditions
By monitoring the applied force (F ) and the position of the atoms that are
pulled over the simulation time, it is possible to obtain force-vs.-displacement
data that can be used to derive the mechanical properties such as bending stiff-
ness or the Young’s modulus (or other mechanical properties). SMD studies
are typically carried out with a spring constant k = 10 kcal mol−1 Å−2 , albeit
this value can be varied depending on the particular situation considered.
Figure 2.44 depicts a schematic that illustrates how load is applied with
SMD, comparing an AFM experiment with the numerical scheme. One of
the first applications of the SMD technique was to study protein unfolding.
Fig. 2.45 shows steered molecular dynamics simulations of I27 extensibility
under constant force, illustrating the molecular geometry under different force
levels (compare with Fig. 1.16a) as well as the force–extension curve (compare
with Fig. 1.16b).
2 Basic Atomistic Modeling 93
Fig. 2.44 Schematic to illustrate the use of steered molecular dynamics to apply
mechanical load to small-scale structures (subplot (a): AFM experiment; subplot
(b) Steered Molecular Dynamics model)
2.13 Summary
We summarize the main points presented in this section. We discussed analy-
sis techniques, to extract useful information from molecular dynamics results,
including the velocity autocorrelation function, the atomic stress, and the
radial distribution function. These are useful analysis methods since they pro-
vide quantitative information about molecular structure in the simulation, for
example during phase transformations, to study how atoms diffuse, for elastic
(mechanical) properties, and others. We introduced several interatomic poten-
tials that describe the atomic interactions. The basic approach in developing
such models is to condense out electronic degrees of freedom and to model
atoms as point particles.
Properties accessible to molecular dynamics can be classified into these
broad categories [86]:
• Structural - crystal structure, g(r), defects such as vacancies and intersti-
tials, dislocations, grain boundaries, precipitates
• Thermodynamic – equation of state, heat capacities, thermal expansion,
free energies
• Mechanical – elastic constants, cohesive and shear strength, elastic and
plastic deformation, fracture toughness
• Vibrational – phonon dispersion curves, vibrational frequency spectrum,
molecular spectroscopy
• Transport – diffusion, viscous flow, thermal conduction
94 Atomistic Modeling of Materials Failure
Fig. 2.45 Steered molecular dynamics simulations of I27 extensibility under con-
stant force. Subplot (a) shows snapshots of the structure of the I27 module simulated
at a force of 50 pN (I, at 1 ns) and 150 pN (II, at 1 ns). At 50 pN, the hydrogen
bonds between strands A and B are maintained, whereas at 150 pN they are broken.
Subplot (b) displays the corresponding force–extension relationship obtained from
the simulations. The discontinuity observed between 50 and 100 pN corresponds
to an abrupt extension of the module by 4–7 Å caused by the rupture of the AB
hydrogen bonds, and the subsequent extension of the partially freed polypeptide
segment. Reprinted with permission from Macmillan Publishers Ltd., Nature [6] c
1999