You are on page 1of 21

Introduction to Atomistic Simulation through

Density Functional Theory


Nico de Koker
Bayerisches Geoinstitut, Universität Bayreuth, Germany

1
2

SUMMARY

Atomistic simulation has become a valuable tool in the study of mate-


rials, including those relevant to the Earth sciences. It provides insight
into the physical properties of minerals at the atomic level, and access to
pressure and temperature conditions not readily achieved experimentally,
thus enabling a fundamental understanding of processes associated with the
structure and evolution of planets. With ongoing advances in computational
technology, numerically intensive simulations have become feasible. The
current state of the art employs Density Functional Theory (DFT) to com-
pute the electronic structure of the material of interest self-consistently,
providing a robust representation of chemical bonding and thus of elec-
tronic, structural and thermodynamic properties. Fundamental to DFT is
the Hohenberg-Kohn theorem, which states that any property of a system
of interacting particles can be represented as a functional of the ground
state charge density. Application of this theorem is made possible through
the Kohn-Sham approach, by which ground state functionals of many elec-
tron systems are obtained. The remaining challenge in determining the
electronic structure is the description of the potential due to exchange and
correlation of electrons in the system, which is only known for the ho-
mogeneous electron gas. To account for these two factors in interacting
systems the Local Density Approximation (LDA) and Generalized Gradi-
ent Approximation (GGA), both of which follow from the exchange and
correlation potential of the homogeneous electron gas, have been com-
monly used with great success. Questions relevant to the deep interior of
the Earth to which DFT has been applied include the structure, thermody-
namics and elasticity of minerals and liquids representative of the mantle
and core, by means of structural optimization, lattice dynamics, molecular
dynamics and Monte-Carlo methods.
3

1. ATOMISTIC SIMULATION

Atomistic simulation involves the computational investigation of mate-


rials at the atomic level by applying the principles of condensed matter the-
ory. It has become a valuable tool in the study of materials, including those
associated with geological processes, and provides deep insight into the
properties of minerals at the atomic and electronic levels. Computational
techniques enable investigation of materials and conditions not amenable
to experimental probing, such as hazardous radioactive compounds and
extreme pressure-temperature conditions.
A large variety of techniques exist, each suited to investigating a spe-
cific subset of physical properties. Important considerations in choosing a
suitable technique of simulation include the nature of the property of inter-
est, whether thermal and/or electronic effects need to be characterized, the
desired accuracy of the result, the quality of constraining parameters, and
the available computational resources. Commonly applied simulation tech-
niques include static lattice energy calculations, lattice dynamics, Monte
Carlo simulations, molecular dynamics, and electronic structure determina-
tions. All these methods require as input the chemical composition, initial
structure, and information on interatomic potentials/electronic structure.
The classical approach involves specification of the potential contribu-
tions due to various types interaction types explicitly

Vtotal = Vcoul + Vvdw + Vstretch + Vangle + Vtorsion (1)

where Vcoul is the contribution due to Coulomb forces, Vvdw due to van
der Waals forces, and Vstretch , Vangle , Vtorsion arise due to two-, three-,
and four body bond deformation. These terms each contain variable pa-
rameters which are obtained by fitting to physical properties for which the
values are well known. The requirement that parameters be predetermined
4

explicitly by fitting to existing data limits the predictive power of the ap-
proach, especially at conditions that are experimentally difficult/imposible
to achieve, which is precisely where computational approaches can be of
greatest value.
First Principles, or ab initio1 calculations apply the fundamental theory
describing condensed matter physics to characterize material properties.
The potential energy of the material responsible for bonding is obtained in
situ by computing the electronic structure of the material. Density Func-
tional Theory is the state of the art of such calculations, and has been widely
applied with great success in the chemical and physical sciences.

2. THEORITICAL BACKGROUND

2.1. Modern Physics


Quantum physics arose as a result of the inability of classical electro-
magnetic theory to account for a number of phenomena. In 1899 Rayleigh
and Jeans showed that a serious discrepancy existed between the classically
predicted radiation from a material at a given temperature and that mea-
sured experimentally (often referred to as the "ultraviolet catastrophe").
Soon thereafter, in 1900, Max Planck proposed that this discrepancy can be
resolved by requiring the energy of electromagnetic radiation to be discrete
instead of continuous. He postulated that the energy is discritized such that
it must satisfy the relation

E = nhν n = 0, 1, 2, . . . , (2)

1
The term ab initio is latin for "from the beginning", and indicates that the calculation is
from first principles and that no empirical data is used.
5

where ν is the frequency of the electromagnetic radiation and h is a universal


constant, now known as the Planck constant (h = 6.626 × 10−34 J s). The
postulate may be applied to any osscilatory physical entity.
Following Planck’s original postulate, in 1905 Albert Einstein success-
fully explained the minimum frequency required for incident electromag-
ntic radiation to induce the photoelectric effect, by invoking Planck’s re-
quirement of quantized energy values. He further proposed that instead of
continuous electromagnetic waves, light should be viewed as discrete par-
ticles called photons. Similarly, by requiring that the energy of electrons
within atoms take on discrete values, Niels Bohr constructed a model of the
atom in 1913 which was able to account for the atomic spectra of hydrogen.
The particle-like nature of light was further confirmed in 1923 by Arthur
Compton, who showed that the two sharply defined intensity peaks he had
found for X-rays scattered in graphite, can only be explained by invoking
photons.
The fact that light, long viewed as wave-like, also exibited particle-like
behaviour, promted Louis de Broglie to postulate in 1924 that particles,
such as electrons and protons, must also have wave-like character. He
showed that if this is the case, any particle with momentum p will posess a
corresponding wavelength of

λ = h/p. (3)

A number of subsequent experiments for a variety of particle types, in-


cluding those by Thomson, by Davisson and Germer, by Estermann, Stern
and Frisch, and by Fermi, Marshall and Zinn, confirmed this particle-wave
duality of matter postulated by de Broglie. The wavelike nature of particles
may be interpreted as a probability function governing the localization of a
quantum of energy. From this view arose the uncertainty principle formu-
lated by Werner Heisenberg in 1927, that the position and momentum of a
6

particle cannot be exactly constrained simultaneously. Heisenberg showed


that

ΔpΔr ≥ h/4π, (4)

where Δp and Δr are the uncertainties in momentum and position, respec-


tively.
In 1926 Schrödinger put forth a differential equation of which the eigen-
functions describe the wave-like character of a system, and the eigenvalues
describe the permissable energy states. The time independent form of this
equation is 2

h̄2 2
− ∇ Ψ(r) + V (r)Ψ(r) = EΨ(r). (5)
2m
where h̄ = h/2π, m is the mass of the particle, V is the potential energy, and
Ψ is the wavefunction. Recasting the equation in terms of the Hamiltonian
operator, we have

ĤΨ = EΨ, (6)

where Ψ and E is respectively the eigenfunction and eigenvalue of the


Hamiltonian operator

h̄2 2
Ĥ = − ∇ + V (r). (7)
2m
By applying his equation to an electron within a Coulomb potential,
Schrödinger matched the Bohr model of the hydrogen atom, from which
follows the well known s, p, d and f orbitals. Similarly, in an N electron
atom, orbitals are represented (approximately) by the individual eigenval-
ues i occupied by the electrons each corresponds to an eigenfunction ψi .
Electrons populate these orbitals such that their energy is minimized, one
2
with the Laplacian ∇2 = ∂ 2 /∂ 2 x + ∂ 2 /∂ 2 y + ∂ 2 /∂ 2 z
7

FIGURE 1. Correlation energy of a homogeneous electron gas as a func-


tion of the density parameter, rs , which is the radius of a sphere containing
an average of one electron. These results are determined by quantum Monte
Carlo calculations, with the values of Ceperley and Alder considered to be
essentially exact.

consequence of which is that degenerate electrons will arrange such that the
total spin is maximized. Further conditions on the wave function demand
that it be antisymmetric under the interchange of any two electrons, from
which follows the well known exclusion principle due to Wolfgang Pauli.
The exclusion principle requires each electron in a many electron system
to occupy a unique energy state. In condensed matter, this results in the
formation of a band structure in the electronic density of states.
Two non-classical electronic interaction effects arise as a result of the
wavelike nature of particles, namely exchange and correlation. Exchange
arises because electrons of the same spin avoid each other, and thus the
8

interaction energy between them is smaller than that of electrons of opposite


spins. Correlation results from the redistribution of electons that occurs due
to the contrast between delocalized wave and point charges descriptions.
The exchange energy can be written as an exact function of the single
particle orbitals, but the form of the correlation energy is unknown. For
a homogeneous electron gas of density ρ, the exchange energy has been
shown to be given by
 1/3
3 6
heg
x =− ρ , (8)
4 π

and the correllation energy has been determined using quantum Monte
Carlo calculations (Fig. 1).
For a system of electrons that is non-interacting except that they satisfy
the exclusion principle, the total wave function may be constructed by
means of a Slater determinant,

1
Ψ = √ det[ψ1 , ψ2 , . . . , ψN ]. (9)
N!
However, constructing the total wave function for fully interacting electrons
is not straightforward, and prohibitively difficult for more than a few tens
of electrons.
The connection between the wave function and the behaviour of the
associated particle was realized by Born, who postulated that the probability
of finding the particle at r is given by

p(r) = Ψ∗ (r)Ψ(r), (10)

where Ψ∗ is the complex conjugate of Ψ. From this folows that the expec-
tation value of a given observable is given by
 ∞
f (r) = Ψ∗ (r)f (r)Ψ(r)d3 r, (11)
−∞
9

or more concisely, using the Dirac bra-ket notation

f (r) = Ψ|f (r)|Ψ. (12)

2.2. Structure of Crystalline Solids

FIGURE 2. An infinite number of unit cells, each defined by a set of


lattice parameters (a and b in this case) can be chosen for a given lattice.

Crystallographers had hypothesised crystalline solids to consist of a pe-


riodic lattice of a repeating basis well before William Bragg and Lawrence
Bragg obtained direct experimental evidence for this through X-ray diffrac-
tion in 1913. The structure of a crystalline solid is completely specified by
the types and positions of nuclei within the unit cell, and the basis which
describes the unit cell itself. This unit cell is repeated quasi-infinitely in
each of the crystallographic axial directions on a lattice (Fig. 2). Augustus
Bravais pointed out that only 14 unique lattice types can be constructed in
10

three dimensions (Fig. 3). If the basis for the unit cell is described by ai ,
any point R in the Bravais lattice follows as

R = n 1 a1 + n 2 a2 + n 3 a3 , (13)

where ni are integers. While an infinite number of unique unit cells exist
for any given Bravais lattice, a unique primitive cell with the full symmetry
of the lattice exists for each type. This unit cell, known as the Wigner-Seitz
cell, is the region of space about a single lattice point that is closer to that
point than to any other point in the lattice (Fig. 4).
Mathematically, the periodic nature of crystal lattices is represented suc-
cinctly in reciprocal space, the lattice of which is defined by the set of
all wave vectors K of planewaves that have the periodicity of the Bravais
lattice

K = m1 b1 + m2 b2 + m3 b3 . (14)

Allowed instances of K must therefore satify the relation eiR·K = 1.


The reciprocal lattice is itself a Bravais lattice, the Wigner-Seitz unit cell
of which is termed the Brillouin zone (BZ). Computation of various quan-
tities, including energy and electronic density of states, require integration
over the BZ. In practice this is done by using the symmetry of the rele-
vant Wigner-Seitz cell to reduce the volume over which integration is to
be performed to the irreducable Brillouin zone (IBZ), and discretizing the
IBZ into a grid of well chosen points at which the quantities are explicitly
calculated (k-points), and then numerically integrated.
11

FIGURE 3. The 14 unique space lattice types, as set out by Bravais.


12

FIGURE 4. (top) Wigner-Seitz cell illustrated for a 2-dimensional lattice,


and (bottom) a for body-centered cubic Bravais lattice

2.3. Hamiltonian for a System of Interacting Particles


The Hamiltonian for a system of electrons and nuclei interacting through
Coulombic forces only is given by

 h̄2  h̄2 1  ZI ZJ e2
Ĥ = − ∇2i − ∇2I +
i
2me I
2MI 2 I=J |RI − RJ |
 ZI e2 1 e2
− + , (15)
i,I
|ri − RI | 2 i=j |ri − rj |

where h̄ is Planck’s constant divided by 2π, e is the electronic charge, me


is the mass of an electron, MI is the mass of nucleus I, ZI is the atomic
13

number of nucleus I, ri is the position of electron i and RI is the position


of nucleus I.
Let ρ(r) be the charge density at r, normalized such that the total number
of electrons is

ne = ρ(r)d3 r. (16)

In many electron systems, exchange and correlation effects modify the


electron-electron interaction potential. Separating this potential into the
classical and exchange correlation contributions, we may write

Vint = VHartree + Vxc , (17)

where VHartree is the classical Coulomb potential,


 
1 ρ(r)ρ(r ) 3 3 
VHartree = d rd r . (18)
2 |r − r |
Due to the relatively large value of MI , the term describing the nuclear
kinetic energy is small and will be ignored (Born-Oppenheimer or adiabatic
approximation). Adopting atomic units (h̄ = me = e = 4π/0 ≡ 1), we
may then rewrite Eq. 14 as

1  2 1  ZI ZJ  ZI
Ĥ = − ∇i + − d3 r
2 i 2 I=J |RI − RJ | i
|r − RI |
 
1 ρ(r)ρ(r ) 3 3 
+ d rd r + Vxc , (19)
2 |r − r |
which we simplify as

Ĥ = T̂ + VIJ + Vext + VHartree + Vxc , (20)

with T̂ the kinetic energy operator, Vext the potential due to interaction of
electrons with nuclei.
Given the charge density ρ(r) and the positions of the nuclei, the prop-
erties of the system can be obtained exactly from Eq. 20, to the extent that
14

the individual terms are known. The form for the classical Coulomb term
is known, and it can be evaluated exactly. If the wavefunction Ψ of the
system is known exactly, the kinetic energy term can also be constrained.
However, as mentioned above, the exact form of the exchange-correlation
term is not known, and must be approximated. Herein lies the key challenge
in the accurate determination of the electronic structure of an interacting
many-body system.

3. DENSITY FUNCTIONAL THEORY

Constructing the wavefunction for a system of fully interacting electrons


is prohibitively difficult for more than a few tens of electrons. This presents
a major obstacle in solving Eq. 20 even for relatively simple materials.
Density functional theory proposes a solution to this dillema by recasting
the problem in terms of a single scalar function of position, namely the
charge density ρ(r). The foundation for the theory was put in place by
Hohenberg and Kohn (1964) and Kohn and Sham (1965), work for which
Walter Kohn was awarded the 1998 Nobel prize in chemistry.
The underlying idea of the method is to express the problem as a func-
tional of the charge density ρ(r)3 , and had been around in the somewhat
rudementary form of the Thomas-Fermi model prior to the formal devel-
opment of DFT. It assumes the charge density in a material to be uniformly
ρ, ignores the effects of exchange and correlation, and replaces the kinetic
energy term in Eq. 20 by that of a homogeneous electron gas of the relevant
density,

3
TT F (ρ) = (3π 2 )2/3 ρ5/3 dr. (21)
10
3
Where a function maps a variable x to a result f (x), a functional maps an entire function
f onto a resulting number F (f )
15

FIGURE 5. The Thomas-Fermi model fails to represent the density


profiles of the interiors of Jupiter (Z ≈ 1) and Earth (Z ≈ 10 in mantle;
Z ≈ 26 in core). Taken from Stixrude et al. (1998).

The error in the kinetic energy in Eq. 21 is very large, and the assumption
of uniform charge density precludes the existence of chemical bonding,
so that the Thomas-Fermi model is not useful as a tool for detailed char-
acterization. For example, it fails to account for the known properties of
terrestrial planets, and the gas giants, strongly overestimating the density
profiles of these planets (Fig. 5). The method is nonetheless insightful
as an early solution to the many body Hamiltonian of a system described
in terms of the charge density, and pointed the way for the formulation of
Density Functional Theory.
16

The two Hohenberg-Kohn theorems form the foundation of Density


Functional Theory. The first theorem states that the ground-state density
ρ(r) of a bound system of interacting electrons in some external potential
V (r) determines this potential uniquely, and therefore guarantees that there
is a one to one mapping between the ground state charge density and the
observable properties. The second theorem holds that, for a trial density ρ̃,
the ground state total energy of a system is variational with respect to the
electron density

E0 ≤ Ĥ(ρ̃), (22)

in other words, the exact ρ(r) provides the minumum possible energy for
the ground state. The theorem provides the theoretical underpinning by
which the Hamiltonian may be solved, in principle exactly.
The Kohn-Sham approach is to replace the difficult interacting many body
system with a non-interacting system which can be more readily solved, and
from which the exact solution of the interacting system can ultimately be
obtained. The kinetic energy term is replaced by that of the non-interacting
electron gas
1  KS 2 KS
T KS (ρ) = − ψi |∇ |ψi , (23)
2
where ψiKS are the individual Kohn-Sham orbitals, from which the total
wavefunction ΨKS is constructed as a Slater determinant. Keep in mind
that the Kohn-Sham orbitals (ψiKS ) and their corresponding eigenvalues
(i ) are introduced as a mathematical trick, and are not physical properties
of the interacting system. The difference between the non-interacting and
true kinetic energy values is small, and is added to the exchange-correlation
energy term.
The form of the Hamiltonian for the Kohn-Sham system is then

Ĥ KS = T̂ KS + VIJ + Vext + VHartree + Vxc , (24)


17

from which the Schrödinger-like Kohn-Sham equations follows after ap-


plying the variational principle

(Ĥ KS − KS
i )ψi (r) = 0,
KS
(25)

1
Ĥ KS (r) = − ∇2 + V KS (r), (26)
2
V KS (r) = Vext (r) + VHartree (r) + Vxc (r). (27)

Solution of the Kohn-Sham equations involves arriving at the exact so-


lution by way of an iterative set of computations under the constraint of
selfconsistency between the charge density ρ(r) and the potential Kohn-
Sham potential V KS (Fig. 6). Once the self consistent charge density has
been obtained, the various output quantities such as the energy, stress tensor
and electronic eigenvalues can be determined.
The Kohn-Sham equations are exact, but the functional form of the po-
tential due to exchange and correlation is not known. Noting this difficulty,
Kohn and Sham (1965) proposed making the local density approximation
(LDA), in which the exchange and correlation energy at a given point is
taken as that of the homogenous electron gas (Eq. 8) with the same density

3
LDA
Exc (ρ) = x (ρ) + c (ρ)]d r.
ρ(r)[heg heg
(28)

Given its simplicity, the LDA performs remarkably well for a wide variety
of materials, though there are numerous instances in which it does not
reproduce known thermodynamic properties. For example, the LDA is
known to overpredict densities (overbinding). In some cases (though not
always), these problems are resolved by applying the generalized gradient
approximation (GGA), in which the exchange correlation term is a function
of the charge density gradient |∇ρ| as well

3
GGA
Exc (ρ) = x (ρ)Fxc (ρ, |∇ρ|)]d r,
ρ(r)[heg (29)
18

FIGURE 6. The self-consistency cycle by which the Kohn-Sham equa-


tions are solved for the ground state properties.
19

FIGURE 7. A pseudopotential is designed to match the true functions


beyond cutoff radius rc , while accounting for the core electrons only ap-
proximately. Modified from Payne et al. (1992).

where Fxc (ρ, |∇ρ|) is a dimensionless enhancement factor, for which a


number of forms have been proposed, the most commonly used being those
by Perdew and Wang (PW91), and Perdew, Burke and Enzerhof (PBE).
In performing the DFT calculation, the wavefunctions are represented
through basis functions, most commonly atomic orbitals (gaussians) or
plane waves. Plane waves are the natural choice for crystals, while Gaus-
sians work well for molecules. In the case of planewaves, basis sets of
progressively larger wave vectors are added together to represent the func-
tion, and the highest wave vector used is generally expressed in terms of the
corresponding energy (cutoff energy). A very large number of planewaves
are needed to reproduce the rapidly varying wavefunctions of core elec-
trons, which poses a major computational obstacle. A number of solutions
to this problem exist, including the use of Augmented Plane Waves (APW),
20

Linearized Augmented Plane Waves (LAPW), Linear Muffin Tin Orbitals


(LMTO), Projector Augmented Waves (PAW) and Pseudopotentials. In
making these various solutions, the form of the potential due to the nucleus
changes (Vext ), and must also be specified.
The plane wave pseudopotentiall method is very common in DFT appli-
cations. The underlying philosophy of pseudopotentials is that for many
applications of DFT, only the valence electrons are important. The pseu-
dopotential method replaces the nucleus and core electrons by an effective
nuclear potential, by substituting the true wave- and potential functions by
more smoothely varying functions (Fig. 7). By demanding that the pseu-
dopotentials satisfy a number of strict criteria, potentials can be constructed
for atoms of the various elements and then applied in a wide variety of
chemical compounds. The transferability of such pseudopotentials makes
codes that implement the plane wave pseudopotential method very power-
ful tools for scientific investigation, although the problems and weaknesses
inherent in the various approximations that are applied to make the exact
Density Functional Theory viable for computation of material properties
must always be kept in mind.
Common codes that apply DFT in first principles atomistic simulation
include:

- Abinit (http://www.abinit.org)
- Castep (http://www.tcm.phy.cam.ac.uk/castep)
- Espresso (http://www.quantum-espresso.org)
- Crystal (http://www.crystal.unito.it)
- Gaussian (http://www.gaussian.com)
- Siesta (http://www.uam.es/siesta)
- Vasp (http://cms.mpi.univie.ac.at/vasp/vasp)
- Wien (http://www.wien2k.at)
21

4. FURTHER READING

Ashcroft, N. W. and N. D. Mermin, Solid State Physics, international ed., Philadel-


phia: Saunders College, 1976.
Eisberg, R. and R. Resnick, Quantum Physics of Atoms, Molecules, Solids, Nuclei,
and Particles, New York: John Wiley & Sons, 1974.
Harrison, W. A., Electronic Structure and the Properties of Solids, San Francisco:
W.H. Freeman and Co., 1980.
Hohenberg, P. and W. Kohn, “Inhomogeneous Electron Gas,” Physical Review B.,
1964, 136, B864.
Kohn, W., “Nobel Lecture: Electronic structure of matter - wave functions and
density functionals,” Reviews of Modern Physics, 1999, 71, 1253–1266.
and L. J. Sham, “Self-Consistent Equations Including Exchange and Correla-
tion Effects,” Physical Review, 1965, 140, 1133.
Martin, R. M., Electronic Structure, Cambridge: Cambridge University Press, 2004.
Parr, R. G. and W. Yang, Density-Functional Theory of Atoms and Molecules,
Oxford: Oxford University Press, 1989.
Payne, M. C., M. P. Teter, D. C. Allan, T. A. Arias, and J. D. Joannopoulos,
“Iterative Minimization Techniques for Ab initio Total-Energy Calculations -
Molecular-Dynamics and Conjugate Gradients,” Reviews of Modern Physics,
1992, 64, 1045–1097.
Stixrude, L., “First principles theory of mantle and core phases.,” in R. T. Cy-
gan and J. D. Kubicki, eds., Molecular Modeling Theory: Applications in the
Geosciences, Vol. 42 of Review in Mineralogy and Geochemistry, Washing-
ton, D.C.: Geochemical Society and Mineralogical Society of America, 2001,
pp. 319–343.
, R. E. Cohen, and R. J. Hemley, “Theory of minerals at high pressure,” in R. J.
Hemley, ed., Ultrahigh-Pressure Mineraolgy, Vol. 37 of Review in Mineralogy
and Geochemistry, Washington, D.C.: Geochemical Society and Mineralogical
Society of America, 1998, pp. 639–671.

You might also like