THEORY OF ELECTRONIC AND STRUCTURAL PROPERTIES OF ...
Transcript of THEORY OF ELECTRONIC AND STRUCTURAL PROPERTIES OF ...
The Pennsylvania State University
The Graduate School
Eberly College of Science
THEORY OF ELECTRONIC AND
STRUCTURAL PROPERTIES OF MATERIALS:
NOVEL GROUP-IV MATERIALS AND
REAL SPACE METHODS
A Thesis in
Physics
by
Peihong Zhang
c© 2001 Peihong Zhang
Submitted in Partial Fulfillment
of the Requirements
for the Degree of
Doctor of Philosophy
August 2001
We approve the thesis of Peihong Zhang.
Date of Signature
Vincent H. Crespi
Assistant Professor of PhysicsDownsbrough Professor
Thesis Advisor, Chair of Committee
Renee Diehl
Professor of Physics
Nitin Samarth
Associate Professor of Physics
Thomas E. MalloukProfessor of Chemistry
DuPont Professor of Materials Chemistry
James B. Anderson
Evan Pugh Professor of Chemistry
Jayanth R. Banavar
Professor of Physics
Head of the Department of Physics
Abstract
This dissertation consists of two parts. The first part employs existing computa-
tional techniques to study the electronic and structural properties of novel group-IV
and related materials, namely, carbon nanotubes, boron nitride nanotubes, and crys-
talline group-IV alloys. In the second part, we develop a new electronic structure
calculation technique based on finite element methods with multigrid acceleration.
The computation time of our new technique scales quadratically with the number of
atoms in the system i.e., O(N2), as opposed to the unfavorable cubic scaling (O(N3))
for most existing ab initio methods.
Chapter 1 gives an overview of theoretical methods involved in this dissertation.
Topics covered in this chapter are orthogonal and nonorthogonal tight-binding to-
tal energy models, density functional theory, pesudopotential planewave methods of
electronic structure, and the GW approximation for calculating accurate bandgaps
in semiconductors.
Chapter 2 focuses on the structural properties of carbon and boron nitride nan-
otubes. First, a new nucleation model for carbon nanotubes is proposed. Detailed
atomistic simulations indicate that this model agrees in several aspects with experi-
mental observations. The new model does not resort to the formation of energetically
unfavorable pentagonal rings, as other nucleation models do, to produce the required
curvature for a nanotube nucleus. Second, plastic deformation of carbon nanotube
under high tensile stress is studied using a tight-binding total energy model. The
elastic limits of carbon nanotubes are found to be higher than any other known ma-
terials and very sensitive to the so-called wrapping angle of nanotubes. Elastic limits
of boron nitride nanotubes are also studied and compared against those of carbon
nanotubes.
Chapter 3 is devoted to the electronic and structural properties of novel group-
iii
IV alloys formed from CVD precursor. Group-IV alloys have attracted considerable
research interest recently. Using the newly developed UltraHigh Vacuum Chemical
Vapor Deposition (UHV CVD) technique, group-IV alloys such as Si4C and Ge4C,
which contain 20 atomic % carbon, have been realized. In this chapter, ab initio
results on the electronic and structural properties of these high carbon concentration
group-IV materials are presented. General trends of the effects of substitutional
carbon are understood. More importantly, two molecular precursors are proposed
for synthesizing the group-IV alloys Si2Sn2C and Ge3SnC, which have direct energy
band gaps and lattice match silicon to better than 1%. This result might open a new
way of integrating silicon-based microelectronics with optoelectronics. Finally, a new
class of molecular precursors, X6C2H18 (X=Si, Ge), is proposed for group-IV alloys
containg 25 atomic % carbon.
The computational time of traditional ab initio techniques such as pseudopotential
planewave methods scales at least as O(N3), where N is the number of atoms in the
system. This unfavorable scaling limits the number of atoms one can study using these
methods to several hundreds, even with the most powerful supercomputers available
today. In chapter 4, we develop an O(N2) ab initio electronic structure calculation
technique based on the finite element methods with multigrid acceleration. O(N2)
scaling is achieved by avoiding explicit re-orthogonalization between eigenvectors,
which is made possible by a multigrid algorithm. With these new techniques, we can
perform ab initio calculations for systems containing more than 32 atoms on a single
workstation (Compaq alpha DS10).
iv
Contents
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii
List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xii
List of Acronyms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiv
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xv
Chapters
1 Theoretical Methods 1
1.1 Tight-binding total energy models . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Orthogonal tight-binding formalism . . . . . . . . . . . . . . . 2
1.1.2 Nonorthogonal tight-binding models . . . . . . . . . . . . . . 4
1.2 Density functional theory . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2.1 The density functional formalism . . . . . . . . . . . . . . . . 7
1.2.2 Local density approximation . . . . . . . . . . . . . . . . . . . 9
1.3 Pseudopotential plane-wave methods . . . . . . . . . . . . . . . . . . 11
1.3.1 The Bloch theorem and plane-wave expansion . . . . . . . . . 12
1.3.2 Pseudopotential concepts . . . . . . . . . . . . . . . . . . . . . 13
1.3.3 Pseudopotential construction . . . . . . . . . . . . . . . . . . 15
1.3.4 The pseudopotential plane-wave method . . . . . . . . . . . . 18
1.4 Beyond LDA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.4.1 The GW approximation . . . . . . . . . . . . . . . . . . . . . 20
v
1.4.2 Generalized Kohn-Sham schemes . . . . . . . . . . . . . . . . 25
2 Carbon and Boron Nitride Nanotubes 27
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.1.1 Specification of carbon nanotubes . . . . . . . . . . . . . . . . 28
2.1.2 Novel electronic and structural properties . . . . . . . . . . . . 30
2.2 Nucleation of carbon nanotubes without pentagonal rings . . . . . . . 34
2.3 Plastic deformation of carbon nanotubes . . . . . . . . . . . . . . . . 43
2.4 Plastic deformation of carbon boron-nitride nanotubes: an unexpected
weakness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3 Electronic and Structural Properties of Novel Group-IV Compounds
Formed from CVD Precursors 60
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.2 Computational details . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.3 Si4C, Ge4C and Sn4C . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.4 Si2Sn2C and Ge3SnC . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.5 Si6C2 and Ge6C2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
3.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4 O(N2) real-space electronic structure methods 87
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.1.1 Momentum-space versus real-space formalism . . . . . . . . . 88
4.1.2 O(N) methods . . . . . . . . . . . . . . . . . . . . . . . . . . 90
4.1.3 O(N2) electronic structure methods: an intermediate step . . 93
4.2 The FE method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
4.2.1 The Ritz-Galerkin Method . . . . . . . . . . . . . . . . . . . . 95
4.2.2 FE triangulation and nodal basis functions . . . . . . . . . . . 96
vi
4.3 FE discretization of the Kohn-Sham equations . . . . . . . . . . . . . 99
4.4 Construction of the ab initio potential matrices . . . . . . . . . . . . 99
4.5 Multigrid Poisson equation solver . . . . . . . . . . . . . . . . . . . . 103
4.5.1 Standard iterative methods for solving linear systems . . . . . 105
4.5.2 Multigrid acceleration . . . . . . . . . . . . . . . . . . . . . . 106
4.5.3 Performance of the MG Poisson equation solver . . . . . . . . 108
4.6 Full multigrid Kohn-Sham equation solver . . . . . . . . . . . . . . . 110
4.6.1 Brief description . . . . . . . . . . . . . . . . . . . . . . . . . 110
4.6.2 Coarse grid solution . . . . . . . . . . . . . . . . . . . . . . . 112
4.6.3 O(N2) iterative diagonalization methods . . . . . . . . . . . . 115
4.6.4 Simultaneous convergence of potential and eigenvalues . . . . 120
4.7 Performance of the FMG electronic structure code . . . . . . . . . . . 121
4.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
vii
List of Figures
1.1 Ionic pseudopotentials of Si. . . . . . . . . . . . . . . . . . . . . . . . 14
2.1 Rolling up a graphene sheet into a tube. . . . . . . . . . . . . . . . . 28
2.2 Specification of single-walled carbon nanotubes. The circumferential
vector shown in the figure corresponds to that of a (5,2) nanotube.
The chiral angle θ is also shown. . . . . . . . . . . . . . . . . . . . . . 29
2.3 Near-Fermi surface electronic structure of graphene. The Fermi surface
is reduced to two distinct (Fermi) points. . . . . . . . . . . . . . . . . 30
2.4 Allowed lines of wavevectors in carbon nanotubes: (a) the allowed lines
miss the Fermi points and (b) the allowed lines pass through the Fermi
points. The hexagon is the Brillouin zone of graphene. . . . . . . . . 31
2.5 Electronic properties of carbon nanotubes. Solid circles are metallic;
triangles are small bandgap semiconductors and the rest are semicon-
ductors with relatively large bandgaps. . . . . . . . . . . . . . . . . . 32
2.6 A schematic diagram of the nucleation model in which a nanotube
forms via edge mediated opening of a double-layered graphitic patch. 35
2.7 The relaxed structures from T=0 molecular dynamics for two-layered
graphitic patches of sizes necessary to form nanotube nuclei of diam-
eters (a) 13.7 A and (b) 27.4 A. The smaller nucleus spontaneously
opens into a circular cross-section which defines a strong preferential
growth axis whereas the larger nucleus remains flattened. . . . . . . . 37
viii
2.8 The energy as a function of a reaction coordinate (i.e., the distance
between opposite inner surfaces as the tube nucleus opens) for nucle-
ation patches of varying diameter when opened and of length 1 nm.
The smallest patches have no barrier to opening. For sufficiently large
patches the opened state is energetically disfavored. For intermediate
sizes, thermal excitation can activate opening on experimental time-
scales. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.9 Two tubes with bond rotation defects. . . . . . . . . . . . . . . . . . 44
2.10 Energy versus length for pure (solid) and defective (dashed) (8,0) nan-
otubes. The cell contains 96 atoms per bond rotation defect with E = 0
for the pure undistorted tube. . . . . . . . . . . . . . . . . . . . . . . 46
2.11 Transition tensions to plastic deformation for a family of (n, 0) and one
of nearly equal-radius (n,m) tubes. The radial dependence is quite
weak except for the smallest tubes, whereas the wrapping angle de-
pendence is very strong. . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.12 Bond rotation defects in (a) (n, 0), (b) (n, n) BN nanotubes, and (c)
(n, 0) and (n, n) carbon nanotubes. There are two distinct bond ro-
tation patterns in a generic (m,n) boron nitride nanotube. For an
(8,0) BN nanotube, the energy difference between the two patterns is
about 0.2 eV/defect at equilibrium (i.e., zero tension), while for (n, n)
tubes the two states are degenerate. Various bond lengths marked (not
shown to scale) correspond to tubes under about 12% strain. . . . . . 53
3.1 Molecular precursors and the corresponding relaxed crystalline phases
of Si4C. Silicon is yellow, carbon is grey, and the terminal group is red.
Ge4C and Sn4C have similar structures. . . . . . . . . . . . . . . . . . 64
3.2 Irreducible Brillouin zone of the bct lattice. . . . . . . . . . . . . . . 66
ix
3.3 LDA band structures of Si4C, Ge4C. . . . . . . . . . . . . . . . . . . 68
3.4 LDA band structures of Sn4C and silicon calculated on a 10-atom unit
cell. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.5 Band structure of Ge4C upon lattice expansion to match that of silicon. 70
3.6 Molecular precursors and the corresponding relaxed crystalline phases
of Si2Sn2C (top) and Ge3SnC (bottom). Silicon is yellow, carbon is
grey, tin is magenta, germanium is green and the terminal group is
red. Ge4C and Sn4C have similar structures. . . . . . . . . . . . . . . 72
3.7 Band structures (in blue) of Si2Sn2C (top) and Ge3SnC (bottom). The
ionicity opens a bandgap deep in the valence band. Symmetry breaking
in Si2Sn2C removes the degeneracy at the X point, pushing one band
up significantly so that it becomes the conduction band maximum and
produces in a direct band gap at X (red arrow). In the case of Ge3SnC,
the conduction band minimum occurs at Γ, which produces a direct
gap at Γ (red arrow). The band structure of silicon (orange) calculated
in the same unit cell is also plotted for comparison. . . . . . . . . . . 77
3.8 Charge density isosurfaces for the top of the valence band at Γ (top)
and X (bottom) in Si2Sn2C. Electrons are localized around Si–Si bonds
(in yellow) for the Γ state, whereas for the X state they are localized
around Sn–Sn bonds (in magenta). . . . . . . . . . . . . . . . . . . . 78
3.9 Proposed molecular precursor for synthesizing Si6C2 and the corre-
sponding relaxed crystalline structure. The molecular precursor and
crystal structure of Ge6C2 are similar to those shown. . . . . . . . . . 80
3.10 LDA band structure of Si6C2 (top) and Ge6C2( bottom). . . . . . . . 83
3.11 Density of states of Si6C2 (top) and Ge6C2( bottom). . . . . . . . . . 84
4.1 n-simplices: 1, 2, and 3-simplices. . . . . . . . . . . . . . . . . . . . . 97
x
4.2 Finite-element triangulation examples in 2D: (a) uniform triangulation
and (b) adaptive (nonuniform) triangulation. . . . . . . . . . . . . . . 97
4.3 A 3D triangulation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
4.4 A linear nodal basis function. The function is defined corresponding
to nodal point Ni. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
4.5 Typical convergence behavior of the GS method . . . . . . . . . . . . 107
4.6 Convergence properties of the MG Poisson equation solver. The solid
and the dotted lines correspond to fine-grid matrix sizes of 32,768×32,768and 131,072×131,072, respectively. . . . . . . . . . . . . . . . . . . . 109
4.7 Comparison of the convergence properties of the standard Davidson
iteration and the MG relaxation scheme we developed. The relaxations
were performed on the finest grid, using the intermediate-grid solution
as a initial guess. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
xi
List of Tables
2.1 Elastic strength of single-walled carbon nanotubes, (a) assuming the
effective cross-sectional area is πr2, where r is the radius of the tube,
and (b) the cross-section is the area per nanotube in a nanotube bundle
with inter-tube distance 3.0 A. . . . . . . . . . . . . . . . . . . . . . 33
2.2 Comparison of the elastic limits of boron nitride and carbon nanotubes,
showing both full and reduced transition tensions (as defined in the
text). The reduced transition tensions for (n, n) carbon and boron
nitride nanotubes are similar, but these quantities differ greatly for
(n, 0) carbon and boron nitride nanotubes. . . . . . . . . . . . . . . 56
3.1 Structural properties of X4C in the local density approximation. The
bulk moduli in parentheses are for silicon, germanium, and gray tin
respectively. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.2 Structural properties of Si2Sn2C and Ge3SnC. All cited values are LDA
results. The results for silicon calculated with a 10-atom supercell are
also listed for comparison. . . . . . . . . . . . . . . . . . . . . . . . . 75
3.3 Structural properties of Si6C2 and Ge6C2. The lattice notation of the
bct structure is used since both structures are only slightly distorted
from bct. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.1 O(N) scaling properties of the MG Poisson equation solver. All calcu-
lations were done on an alpha DS10 workstation. . . . . . . . . . . . 109
xii
4.2 Scaling and convergence properties of the FMG real-space electronic
structure code. All calculations were done on an alpha DS10 worksta-
tion. Note that the errors refer to the differences in the correspond-
ing values between two consecutive iterations, not the absolute errors
with respect to the converged values. CPU time for the corresponding
plane-wave calculation is shown in the parenthesis. . . . . . . . . . . 122
xiii
List of Acronyms
TB Tight-BindingDFT Density Functional TheoryLCAO Linear Combination of Atomic OrbitalsPBC Periodic Boundary ConditionsKS Kohn-ShamLDA Local Density ApproximationSCTB Self-Consistent Tight-BindingHF Hartree-FockCI Configuration InteractionOPW Orthogonalized Plane WaveKB Kleinman-BylanderRPA Random Phase ApproximationGKS Generalized Kohn-ShamBZ Brillouin ZoneDOS Density-of-StatesCVD Chemical Vapor DepositionBN Boron NitrideUHV UltraHigh-VacuumFFT Fast Fourier TransformLNV Li, Nunes, and VanderbiltVSEPR Valence Shell Electron Pair RepulsionFD Finite DifferenceFE Finite ElementMG MultiGridGS Gauss-SiedelCG Conjugate GradientAMG Algebraic MultiGridSOR Successive Over RelaxationFMG Full MultiGridSC Self-Consistent
xiv
Acknowledgments
First, I would like to thank my advisor, Prof. Vincent Crespi, for his constant
support and encouragement. Without his enthusiasm and guidance, the research
projects reported in this dissertation would never have been completed. I would also
like to thank Prof. Steven Louie, Prof. Marvin Cohen and Dr. Eric Chang for the
fruitful collaboration and for hosting me during my visit to UC, Berkeley in Summer,
1999. I thank Dr. Paul Lammert for the enjoyable collaboration and discussion,
and for solving most of my problems with UNIX. Prof. Thomas Mallouk and Prof.
John Kouvetakis are acknowledged for the useful discussion and experimental inputs.
Prof. Jinchao Xu and Prof. Jesse Barlow are acknowledged for their help on numerical
algorithms. Finally, I thank my family for their love and support.
xv
Chapter 1
Theoretical Methods
The work presented in this thesis involves a number of theoretical/computational
techniques, ranging from the simple and fast empirical tight-binding (TB) models to
the fairly sophisticated and computationally intensive GW method. In this chapter,
we give an overview of these methods: orthogonal and nonorthogonal TB total energy
models, the density functional theory (DFT), pseudopotential plane-wave methods,
and the GW approximation.
1.1 Tight-binding total energy models
Empirical classical potential models (known as force-field models in the chem-
istry and biochemistry communities) have been applied extensively to model systems
containing thousands to millions of atoms with great success. However, in situa-
tions where quantum mechanical effects are important, classical potentials often fail
to produce meaningful results. For example, pair potentials fail to stabilize tetra-
hedral structures such as diamond and zincblende, since the directional covalent
bonding in these systems is primarily determined by quantum mechanical effects.
As a result, complicated three-body potentials have been proposed to describe such
systems[1, 2, 3, 4, 5, 6]. DFT-based ab initio techniques (to be discussed in more
detail later), on the other hand, are very accurate and take the quantum mechanical
1
effects into full consideration. However, ab initio methods are also very computation-
ally demanding. Consequently, even with the most powerful supercomputers available
today, the largest system one can handle using ab initio methods is limited to the
order of 1,000 atoms. The TB method[7] takes an intermediate step towards mod-
elling meterials that takes into account quantum mechanical effects without too much
computational effort.
1.1.1 Orthogonal tight-binding formalism
The idea of using a linear combination of atomic orbitals (LCAO) to represent the
wavefunctions of aggregates of atoms (e.g., solids) was due to Bloch[8]. However, the
success of modern orthogonal TB schemes should be credited to Slater and Koster[7],
who presented the first detailed TB formalism for band structure calculations. In the
following, we shall assume periodic boundary conditions (PBC). (Extension of the
TB formalism to finite-size systems is straightforward.)
Denoting an atomic orbital n = (τ, α) at unit cell i as φn(r − rτ −Ri), we have
the Bloch sum
Φnk =1√N
N∑i=1
eik·(rτ+Ri)φn(r− rτ −Ri), (1.1)
where the sum is over the atoms in equivalent positions in all unit cells in the crystal.
τ is the atomic index in the unit cell and α specifies the atomic orbitals of atom τ . For
example, in the diamond structure, τ = 1, 2 and α = s, px, py, pz. We may express
the wavefunctions of the system as linear combinations of Bloch sums. However, the
Bloch sum formed from atomic orbitals is not an ideal basis, since atomic orbitals
at different sites may not be orthogonal to each other. It is more convenient to
work in a orthogonal basis, the so-called Lowdin orbitals ψn(r), which are formed
by the reorthogonalization of atomic orbitals φn(r)[9]. We shall assume that such a
2
reorthogonalization[9] is done and rewrite the Bloch sum as
Ψnk =1√N
N∑i=1
eik·(rτ+Ri)ψn(r− rτ −Ri). (1.2)
The Hamiltonian matrix element between two Bloch sums with the same k is
Hn,n′(k) =N∑i=1
eik·(rτ+Ri−rτ ′)〈ψn(r− rτ )|H|ψn′(r− rτ ′)〉. (1.3)
In the above equation, the sum has to be carried out over all unit cells in principle.
In practice, one always assumes that the matrix element 〈ψn|H|ψn′〉 is short-rangedand decays to zero beyond some cut-off distance[10, 11, 12]. The matrix elements
〈ψn|H|ψn′〉 in the TB formalism are modelled by some empirical parameters which
are either obtained by fitting experimental data or from more accurate ab initio cal-
culations. Eigenvalues and corresponding wavefunctions are then calculated by diag-
onalizing the Hamiltonian matrix H. This is the original Slater-Koster TB model[7].
The Slater-Koster TB formalism computes only the band energies. For atomistic
simulations, the total energy of the system must also be derived. Chadi[10] proposed
that the total energy of a system in the TB framework can be written as band
structure energy plus a sum of pair-wise repulsive potentials
Etot = Ebs + Erep =∑
i∈occupiedεi +
∑i,j
U(|ri − rj|). (1.4)
This is consistent with the local density approximation (LDA) expression of the total
energy, which is as a sum of the band energy and some potential contributions (see
Eqn. (1.27)). The idea that the terms in the total energy which are not included
in the single-electron band energy can be approximated by a sum of pair terms is of
great importance and is almost universally applied[13]. (However, the justification
of doing so is still questionable.) The force acting on an atom in the TB formalism
contains two terms, one arising from the band structure energy Ebs and the other
from the repulsive potential energy[14]. In chapter 2, the TB parameters for carbon
3
systems by Xu, et al.[15] are adopted for studying the structural properties of carbon
nanotubes.
1.1.2 Nonorthogonal tight-binding models
The orthogonal TB model described in section 1.1.1 is conceptually simple and
computationally efficient. However, it shares with other empirical potential methods
a serious drawback, the lack of transferability[13]. The assumption that an orthgo-
nal set of orbitals can be constructed from the nonorthogonal atomic orbitals by the
Lowdin procedure is not universally valid, i.e., an orthogonal transformation is valid
only for a single configuration of the system. Furthermore, the Lowdin orbitals are
usually more extended than the nonorthogonal atomic orbitals, which requires that
the Hamiltonian matrix element 〈ψn|H|ψn′〉 be longer-ranged. The usually short-
ranged TB parametrization is therefore not expected to work well across different en-
vironments. Failure to properly account for the overlap effects is probably the main
reason for the lack of transferability of orthogonal TB models[13], although other
issues such as the two-center approximation of the Hamiltonian matrix element, the
pairwise repulsive potentials, and lack of self-consistency are also important factors.
The transferability can be improved somewhat by including a wider range of empir-
ical data; however, the improvement is limited and not systematic. Dorantes-Davila
et al.[16] and Dorantes-Davila and Pastor[17] have proposed an iterative scheme to
incorporate in an orthogonal TB framework the effects of the overlaps of atomic or-
bitals. A different approach is the so-called nonorthogonal TB model. As its name
tells, this approach eliminates the assumption of orthogonality so that most of the
drawbacks associated with this assumption can be overcome. (The effect of neglecting
the nonorthogonality of the basis has been investigated by Mirabella et al.[18] and
Mckinnon and Choy[19].)
In the nonorthogonal TB model, one has to solve the generalized eigenvalue prob-
4
lem
(H− EiS)Ψi = 0, (1.5)
where S is the overlap matrix between TB orbitals,
Sn,n′ =∫
φn(r)φn′(r)dr, (1.6)
and φ are nonorthogonal atomic TB orbitals as opposed to the orthogonalized Lowdin
orbitals ψ. The repulsive part of the total energy is usually assumed to be pairwise
as in the orthogonal TB models. The on-site (diagonal) Hamiltonian matrix elements
are usually taken as the atomic ionization potentials of the corresponding orbitals, i.e.,
Hn,n = εn and the off-diagonal elements can be determined using Mulliken analysis[20,
21]:
Hn,n′ =1
2K(εn + εn′)Sn,n′, (1.7)
where K is an adjustable parameter, or by other means[22, 23]. We employ a
nonorthogonal TB scheme by Porezag et al.[24] to study the plastic deformation
of boron-nitride nanotubes in Chapter 2.
1.2 Density functional theory
Seeking an accurate yet tractable description of many particle systems has been
an important goal of physicists and chemists during the past few decades. The TB
methods described in the previous section represent the early efforts to incorporate
realistic quantum mechanics effects into a model Hamiltonian. Although TB mod-
els have been successfully applied to a number of systems and have advantages over
empirical potentials, there are limitations. For example, it is very difficult to obtain
TB parameters for systems containing more than two species. Also, there are only a
few accurate and transferrable TB parameters available. TB parameters constructed
by fitting data from one phase might not be readily applied to the study of other
5
phases or the transitions among them, i.e., TB parameters are not very transferable.
Although nonorthogonal TB schemes improve somewhat the transferability of TB by
relaxing the orthogonality assumption, the improvements are limited by their empir-
ical nature. Finally, the assumption that the Hamiltonian matrix does not depend
upon the distribution of electrons brings about severe difficulties when dealing with
systems where charge transfer is important (this problem is partially addressed by
the so-called self-consistent tight binding (SCTB) methods.). Therefore, theoretical
methods that require no empirical parameters, the so-called first principles or ab ini-
tio methods, have attracted intensive research interest since the availability of digital
computers.
The most natural approach to ultimately describe a many-electron system is
to calculate its many-electron wavefunctions. The earliest such attempt is due to
Hartree[25]. In the Hartree method, the many-body wavefunction is approximated
by a product of single-electron wavefunctions.
Ψ(r1, r2, · · · rN) = ψ1(r1)ψ1(r2) · · ·ψN(rN). (1.8)
Each of the wavefunctions ψi satisfies a one-electron Schrodinger equation with the
potential term replaced by an effective self-consistent Coulomb (or Hartree) potential.
Better approximations can be achieved by considering the antisymmetric property of
the many-electron wavefunction as required by the Pauli principle, resulting in a
scheme called Hartree-Fock (HF) approximation[26, 27]. In the HF approximation,
the many-electron wavefunction is represented by an antisymmetrized wavefunction,
i.e., a Slater determinant. The effective potential in the HF approximation consists of
two terms, a local Hartree (or Coulomb) potential and a nonlocal exchange potential
which arises from the antisymmetrization of the many-electron wavefunctions. The
HF approximation has had tremendous success and remains an indispensable tool in
molecular physics and quantum chemistry[28]. However, the nonlocal exchange po-
6
tential introduced by the HF approximation drastically increases the computational
cost. Furthermore, correlation effects are not included in the HF scheme. Systematic
improvement over the HF approximation can be further achieved by the configuration
interaction (CI) scheme, in which the many-electron wavefunction is expanded as a
superposition of many Slater determinants (as opposed to the single Slater determi-
nant used in the HF approximation). CI is a very attractive approach since it yields,
in principle, an exact solution to the many-body problem. However, the explosive
increase in the computer resources required in CI calculations with increasing number
of electrons means that only very small systems can be treated with high accuracy in
this method.
In an effort to avoid the computationally demanding exchange term in the HF ap-
proximation while retaining the overall exchange effects, Slater[29, 30] later replaced
the nonlocal exchange potential with a very simple potential, known as Slater’sXα po-
tential, that depends only on the local charge density, i.e., −α 32π(3π2n(r))1/3. Slater’s
Xα potential can be regarded as an early version of the so-called local density approxi-
mation (LDA) to the exchange-correlation potential in the DFT framework. Actually,
the Thomas-Fermi approximation was the first attempt to describe a many-electron
system using the local charge density as the basic variable[31, 32]. However, it was
not fully realized that the local charge density could indeed serve as the fundamen-
tal variable for solving the many-electron problem until the work of Hohenberg and
Kohn[33] and Kohn and Sham[34].
1.2.1 The density functional formalism
In a paper published in 1964, Hohenberg and Kohn proved that the ground state
properties of a many-electron system can be expressed as functionals of its ground
state charge density[33]. The proof of this remarkable theorem, which was later
generalized by Levy[35], turns out to be rather simple. We follow here the derivation
7
in a review article by Jones and Gunnarsson[28].
Let’s denote n(r) as “N-representable” densities, i.e., those which can be obtained
from some antisymmetric many-electron wavefunction Ψ
n(r) = N∫|Ψ(r, r2, r3, · · · , rN)|2dr2 · · · drN , (1.9)
assuming that the wavefunction Ψ is normalized to one. The Hamiltonian for N
electrons moving in an external potential V ext is
H = T + V ee +N∑i=1
V ext(ri), (1.10)
where T is the kinetic energy operator and V ee is the electron-electron interaction.
Levy defined a functional
F [n] = minΨ→n〈Ψ|T + V ee|Ψ〉, (1.11)
where the minimization is taken over all wavefunctions Ψ that give the charge density
n(r). For all N -representable n(r), the two basic theorems of DFT are
E[n] ≡∫
V ext(r)n(r)dr+ F [n] ≥ E0, (1.12)
where E0 is the ground state energy, and∫V ext(r)n0(r)dr+ F [n0] = E0, (1.13)
where n0 is the ground state charge density.
The variational principle (1.12) is obvious since, for any N -representable density
n, the wavefunction that minimizes the functional F [n], denote it as Ψnmin, is a super-
position of eigenfunctions of the system. Therefore, the expectation value of Ψnmin,
namely, E[n], should be higher or equal to the ground state energy.
If Ψ0 is the ground state wavefunction, n0 is the ground state charge density,
and Ψn0min is the wavefunction that minimizes the functional F [n0], we have, from the
definition of ground state,
E0 ≡ 〈Ψ0|V ext + T + V ee|Ψ0〉 ≤ 〈Ψn0min|V ext + T + V ee|Ψn0min〉. (1.14)
8
Since
〈Ψ0|V ext|Ψ0〉 = 〈Ψn0min|V ext|Ψn0min〉 =∫
V ext(r)n0(r)dr, (1.15)
we have
〈Ψ0|T + V ee|Ψ0〉 ≤ 〈Ψn0min|T + V ee|Ψn0min〉. (1.16)
On the other hand, the definition of functional F [n] yields
〈Ψn0min|T + V ee|Ψn0min〉 ≤ 〈Ψ0|T + V ee|Ψ0〉. (1.17)
This can be true only if
〈Ψ0|T + V ee|Ψ0〉 = 〈Ψn0min|T + V ee|Ψn0min〉. (1.18)
Therefore, we have
E0 ≡ 〈Ψ0|V ext + T + V ee|Ψ0〉=
∫V ext(r)n0(r)dr+ 〈Ψ0|T + V ee|Ψ0〉
=∫V ext(r)n0(r)dr+ 〈Ψn0min|T + V ee|Ψn0min〉
=∫V ext(r)n0(r)dr+ F [n0].
(1.19)
This completes the proof of the two basic theorems of DFT.
1.2.2 Local density approximation
The DFT of Hohenberg and Kohn states that knowledge of the ground state charge
density n0(r) of a system is, in principle, sufficient to calculate any of its ground
state properties, provided that the exact functionals are known. For example, to
calculate the ground state energy, one simply needs to know the functional form of
F [n]. However, obtaining an exact functional F [n] is obviously impossible. Kohn and
Sham[34] proposed the simplest, and surprisingly, the most successful approximation
to the energy functional, which is now known as the local density approximation
(LDA).
Kohn and Sham first partitioned the functional F [n] into three parts:
F [n] = T 0[n] + Eh[n] + Exc[n], (1.20)
9
where T 0[n] is the kinetic energy of a system of noninteracting electrons with density
n(r), Eh[n] is the Hartree energy and Exc[n] is the exchange-correlation energy func-
tional, which remains to be approximated. They assumed a slowly varying charge
density n(r) and approximated the exchange-correlation energy functional as
Exc[n] =∫
n(r)εxc(n(r))dr, (1.21)
where εxc is the exchange-correlation energy per electron of a uniform electron gas
of density n(r)[34]. This expression can be regarded as the leading term of a more
general Taylor expansion of Exc[n]:
Exc[n] =∫
n(r)εxc(n(r))dr+∫
εxc2 (n)|∇n|2dr+ · · · . (1.22)
Therefore, in the slowly varying charge density limit, expression (1.21) becomes exact.
By expanding the charge density n(r) as a sum of single-particle-like densities
n(r) =N∑i=1
|φi(r)|2, (1.23)
and applying the variational principle to the energy functional, they reached a set of
one-particle Schrodinger equations, known as Kohn-Sham equations (in atomic unit)
[−12∇2 + V H(r) + V xc(r) + V ext(r)]φi(r) = εiφi(r), (1.24)
where
V xc(r) =d(nεxc(n))
dn(1.25)
is the exchange-correlation potential and
V H(r) =∫ n(r′)|r− r′|dr
′ (1.26)
is the Hartee potential. The total energy of the system is
E =N∑i=1
εi − 1
2
∫n(r)n(r′)|r− r′| drdr
′ +∫
n(r)[εxc(n(r))− V xc(n(r))]dr. (1.27)
10
The Kohn-Sham equations (1.24) have to be solved self-consistently. Furthermore,
the exchange-correlation energy εxc(n) needs to be approximated.
In their original work, Kohn and Sham did not expect this simple approximation,
namely LDA, to work well for real systems, commenting “we do not expect an accurate
description of the chemical bonding”[34] (by LDA). Therefore, it is a rather surprising
fact that LDA has been applied to a wide variety of systems with great success. In
some cases, the charge densities are not at all slowly varying. Apart from the fact
that LDA (somewhat unexpectedly) produces rather accurate ground state properties
of materials, the Kohn-Sham orbitals φi, which are introduced as a variational basis
and have no obvious connection with the quasiparticle wavefunctions, turn out to be
surprisingly close to the latter[36]. Also, the Kohn-Sham eigenvalues εi are often very
close to quasiparticle energies. Therefore, although DFT was formally established for
computing ground state properties, it is now almost universally applied to calculate
and interpret the electronic structure of materials. However, one should bear in
mind that calculations of the excited state properties are actually out of the scope
of the DFT, and neither LDA nor DFT should be blamed for any inconsistency
between the DFT (LDA) eigenvalues and the experimentally measured quasiparticle
excitation energies. In fact, there are well known problems. For example, LDA
usually underestimates the band gap of semiconductors and insulators. For simple
metals such as lithium and sodium, LDA gives free-electron-like valence band widths
which are typically 20∼30% wider than the experimental ones. We will return to this
point in section 1.4.
1.3 Pseudopotential plane-wave methods
The density functional theory, particularly the Kohn-Sham LDA form, greatly
simplifies the calculation of materials properties and significantly reduces the compu-
tational effort. Many of the modern computational methods in solid state physics are
11
based on DFT. In this section, we give a brief review of pseudopotential plane-wave
methods[37].
1.3.1 The Bloch theorem and plane-wave expansion
The Bloch theorem is one of the most fundamental theorems in solid state physics.
Considering an electron moving in a periodic potential V (r), the Schrodinger equation
is
(−12∇2 + V (r))φi(r) = εiφi(r), (1.28)
where
V (r) = V (r+Rn), (1.29)
and Rn is the lattice vector. Bloch showed that Eqn. (1.28) has solutions
φi,k(r) = eik·rui,k(r), (1.30)
where
ui,k(r) = ui,k(r+Rn). (1.31)
Since uk(r) has the periodicity of the lattice, it can be expanded in a Fourier sum
ui,k(r) =∑G
Ci,G(k)eiG·r, (1.32)
where the sum, in principle, should be carried out over an infinite number of reciprocal
lattice G’s. In practice, beyond a maximum cutoff Gmax, the coefficients can be
assumed to be zero. Substituting the Fourier expansion (1.32) into Eqn. (1.28), we
have the algebraic eigenvalue equations
∑G′
HGG′(k)Ci,G′ = εi,kCi,G, (1.33)
where the Hamiltonian matrix elements are
HGG′(k) ≡ 〈k+G| − 1
2∇2 + V |k+G′〉 = 1
2|k+G|2δG,G′ +VG,G′(k). (1.34)
12
The Hamiltonian matrix can be diagonalized to obtain the band structure and elec-
tronic wavefunctions of the system, provided that the potential matrix elements
VG,G′(k) are known. In the following sections, we shall discuss the concept of pseu-
dopotentials and how to evaluate theVG,G′(k) within the pseudopotential framework.
1.3.2 Pseudopotential concepts
Plane-waves are a very natural basis for describing electronic wavefunctions of pe-
riodic systems. However, if we need to represent highly localized wavefunctions such
as those of core electrons, an enormous number of plane-waves is needed. Fortunately,
the pseudopotential concepts introduced by Fermi[38] and Hellmann[39] make it pos-
sible to accurately describe the properties of materials without dealing explictly with
the core electrons.
The cancellation theorem derived from the orthogonalized plane wave (OPW)
formalism by Phillips and Kleinman is very helpful in understanding the pseudopo-
tential concepts[40]. This theorem demonstrates that the requirement that valence
states should be orthogonal to the core states effectively provides a repulsive potential
for the valence states. If this effective repulsive potential is added to the attractive
potential in the core region, they almost cancel out, leaving a relatively weak net po-
tential or pseudopotential[40]. In the OPW methods, the valence state wavefunctions
are expressed as the sum of a smooth wavefunction φ and a sum over core states φc,
|Ψ〉 = |φ〉+∑c
αc|φc〉, (1.35)
with the requirement that the valence states Ψ are orthogonal to the core states φc,
which yields
|Ψ〉 = |φ〉 −∑c
|φc〉〈φc|φ〉. (1.36)
We now find an effective Schrodinger equation that φ satisfies, assuming that Ψ
13
-16
-14
-12
-10
-8
-6
-4
-2
0
2
0 1 2 3 4 5r (a.u.)
3s
V
3p3d
Figure 1.1: Ionic pseudopotentials of Si.
satisfies the following Schrodinger equation
H|Ψ〉 = (T + Vc)|Ψ〉 = E|Ψ〉, (1.37)
where T is the kinetic energy operator and Vc denotes the attractive core potential.
Substituting (1.36) into the above equation, we have
H|φ〉+∑c
(E −Ec)|φc〉〈φc|φ〉 = E|φ〉, (1.38)
where Ec is the energy of the core state φc, i.e.,
H|φc〉 = Ec|φc〉. (1.39)
Eqn. (1.38) can be rewritten as
(T + V ps)|φ〉 = E|φ〉, (1.40)
where the psuedopotential V ps is
V ps = Vc +∑c
(E − Ec)|φc〉〈φc|. (1.41)
14
It can be shown that the second term of V ps is a strongly repulsive potential
which almost cancels the strongly attractive core potential Vc, leaving a weak net
pseudopotetial. Fig. 1.1 shows the ionic pseudopotentials of silicon. Compared with
the −1/r behavior of the Coulomb potential, the potentials shown in the figure aremuch softer.
1.3.3 Pseudopotential construction
To generate the ionic pseudopotentials, we start by solving the all-electron Kohn-
Sham equations of the atom of interest
[−12∇2 + V H(r) + V xc(r)− Z ion
r]φnlm(r) = εnlφnlm(r). (1.42)
Having obtained the all-electron potential, the valence electron orbital energies and
wavefunctions, there are several proposed schemes to construct the intermediate, so-
called screened pseudopotentials[41, 42, 43, 44, 45]; these will not be discussed in
detail here. We assume that we have constructed the screened pseudopotentials (the
meaning of “screened” will be clear later as we proceed our discussion). The valence
pseudo-wavefunctions satisfy
[−12∇2 + V ps,src
l (r)]φpslm(r) = εpsl φpslm(r). (1.43)
Writing
φpslm(r) =upsl (r)
rYlm(Ωr), (1.44)
we have
[−12
d2
dr2+
l(l + 1)
2r2+ V ps,src
l (r)]upsl (r) = εpsl upsl (r). (1.45)
Note that different angular momentum components have different pseudopotentials.
Most of the modern norm-conserving (see definition below) pseudopotentials share
the following characteristics.
15
(i) The eigenvalues calculated with the pseudopotential are identical to all-electron
eigevalues, i.e.,
εpsl = εl. (1.46)
(ii) The logarithmic derivatives of pseudo-wavefunctions and those of the correspond-
ing all-electron ones agree beyond some cut-off radii rcutl , i.e.,
d
drln upsl (r) =
d
drln ul(r) for r > rcutl . (1.47)
(iii) Pseudo-wavefunctions satisfy the norm conservation condition
upsl (r) = ul(r) for r > rcutl , and∫ ∞
0|upsl (r)|2dr =
∫ ∞
0|ul(r)|2dr = 1. (1.48)
The corresponding pseudopotentials are called norm-conserving pseudopotentials.
(iv) The pseudo-wavefunction contains no radial nodes.
After obtaining the screened pseudopotentials V ps,srcl and the valence pseudo-
wavefunctions upsl , we are ready to generate the so-called ionic, or unscreened pseu-
dopotentials by subtracting the electrostatic and exchange-correlation screening con-
tributions due to the valence electrons from the screened pseudopotentials,
V psl (r) = V ps,src
l (r)− V H [ρpsv ; r]− V xc[ρpsv ; r], (1.49)
where ρpsv is the valence pseudo-charge density. The total ionic pseudopotential is
〈r|V ps|r′〉 = V ps(r, r′) =lmax∑l=0
m=l∑m=−l
Y ∗lm(Ωr)V
psl (r)
δ(r − r′)r2
Ylm(Ωr′), (1.50)
which is obviously nonlocal (or semi-nonlocal since only the angular part is nonlocal).
Applying such nonlocal pseudopotentials on the wavefunctions φ(r) for large systems
is formidable, since the integration
V psφ(r) =lmax∑l=0
m=l∑m=−l
∫Y ∗lm(Ωr)V
psl (r)
δ(r − r′)r2
Ylm(Ωr′)φ(r′)dr′, (1.51)
16
has to be carried out for every distinct point r in the space, which involves O(N3)
operations, where N is the number of atoms in the system. The O(N3) scaling is due
to: first, the number of points in the space scales as O(N); the number of projectors
in Eqn. (1.50) scales as O(N); and finally, a single integration in Eqn. (1.51) requires
O(N) operations.
Note that all V psl (r) at large r (r > rcutl ) reduce to the ionic Coulomb potential,
−Z ion/r, indenpendent of l. Therefore, it is expedient to split the semi-nonlocal
pseudopotential into a multiplicative local term plus some short-ranged corrections
V ps(r, r′) = V pslocδ(r− r′) +
lmax∑l=0
m=l∑m=−l
Y ∗lm(Ωr)δV
psl (r)
δ(r − r′)r2
Ylm(Ωr′), (1.52)
where
V psloc(r) = V ps
l−loc(r), (1.53)
and
δV psl (r) = V ps
l (r)− V psloc(r), (1.54)
which vanishes beyond max(rcutl , rcutloc ). Applying the above nonlocal pseudopotential
on wavefunctions requires O(N3) operations in momentum-space based methods (e.g.,
the pseudopotential plane-wave method, which is discussed in the following section.)
and O(N2) operations in the real-space formalism, since in real-space calculations,
the integration in Eqn. (1.51) can be carried out locally. Real-space methods will be
discussed in Chapter 4.
Kleinman and Bylander (KB)[46] proposed an elegant prescription that transforms
the semi-nonlocal pseudopotential into a fully separable, truly nonlocal pseudopoten-
tial
V psKB =
lmax∑l=0
m=l∑m=−l
|δV psl φpslm〉〈φpslmδV ps
l |〈φpslm|δV ps
l |φpslm〉, (1.55)
where φpslm are the atomic valence pseudo-wavefunctions in Eqn. (1.43) and δV psl is
defined in Eqn. (1.54). We can readily verify that the KB form of the nonlocal
17
pseudopotential reproduces the orbital energy for the reference electronic states, i.e.,
states which we use to construct the pseudopotential:
V psKB|φpsl′m′〉 =
lmax∑l=0
m=l∑m=−l
|δV psl φpslm〉〈φpslmδV ps
l |〈φpslm|δV ps
l |φpslm〉|φpsl′m′〉 = δV ps
l |φpsl′m′〉. (1.56)
Applying the KB nonlocal pseudopotential on the wavefunctions requires O(N) op-
erations in the real-space formalism and O(N2) operation in the momentum-space
formalism. (We will return to this point in Chapter 4.) The KB form of pseudopo-
tential considerably reduces the computational effort associated with the operations
of the semi-nonlocal pseudopotential on the wavefunctions and is now almost univer-
sally used in pseudopotential calculations. However, there is also a minor drawback of
the KB nonlocal pseudopotentials: they sometimes lead to unphysical “ghost” states
at energies below or near those of the physical states. These “ghosts” are artifacts of
the KB potential in which the nodeless pseudo-wavefunctions need not be the lowest
possible eigenstates, unlike in the case of the semi-local form. Usually, ghost states
can be eliminated by properly chosing the local pseudopotential component and/or
varing the cutoff radii rcutl of the pseudopotentials.
1.3.4 The pseudopotential plane-wave method
Most of today’s large-scale electronic structure and total energy calculations for
solids employ pseudopotentials. The plane-wave expansion described in Section (1.3.1),
when combined with pseudopotentials to eliminate the highly localized core electrons,
can be very efficient. In this section, we give a brief review of the pseudopotential
plane-wave formalism[37].
Let’s start with the Kohn-Sham equation (1.24), with the external potential V ext
now replaced by ionic pseudopotential V ps
[−12∇2 + V H(r) + V xc(r) + V ps]φi,k(r) = εi,kφi,k(r). (1.57)
18
We have introduced an additional index k to specify the crystal momentum. V ps is
a nonlocal ionic pseudopotential (in the KB form)
V ps(r, r′) =∑τ,i
V psτ (r− rτ −Ri, r
′ − rτ −Ri), (1.58)
where rτ is the position of τ -atom in unit cell i. The operation of the nonlocal
pseudopotential on the wavefunctions should be understood as
V psφi,k(r) =∫
V ps(r, r′)φi,k(r′)dr′. (1.59)
In a plane-wave basis the wavefunction φ is expanded as
φi,k(r) = eik·r∑G
Ci,G(k)eiG·r, (1.60)
and the Hamiltonian matrix elements
HGG′(k) =1
2|k+G|2δG,G′+V H(G−G′)+V xc(G−G′)+V ps(k+G,k+G′), (1.61)
where the Fourier transform of the pseudopotential in (1.58) can be written
V ps(k+G,k +G′) =∑τ
V psτ (k +G,k+G′)ei(G−G′)·rτ . (1.62)
Note that only the Fourier components of the nonlocal part of the pseudopotential
depend on k.
The eigenvalue problem can then be solved by either traditional direct or newer
iterative diagonalization schemes to obtain the eigenvalues and corresponding wave-
functions. Total energy and other quantities of interest are subsequently evaluated[37].
1.4 Beyond LDA
The Kohn-Sham version of the DFT provides a simple and yet powerful means
to calculate the ground state properties of materials. However, applications of DFT
(particularly in its LDA form) have gone far beyond its original concept. LDA results
19
are now routinely used to interpret electronic band structures of solids with tremen-
dous success. Strictly speaking, there is no clear connection between the Kohn-Sham
eigenvalues and single-particle excitation energies. (In contrast, Koopman’s theorem
provides a formal interpretation of the Hartree-Fock orbitals and their corresponding
eigenvalues for atomic and molecular systems.) Therefore, it is rather surprising that
LDA band structures in general agree well with experimental measurements. De-
spite all of its success, there are serious discrepancies between LDA band structures
and quasiparticle energies measured in, e.g., photoemission experiments: (i) the LDA
bandgaps in semiconductors are usually underestimated by ∼ 30 to 50%; (ii) the
LDA bandwidth in some simple metals such as Li and Na are ∼ 10 to 30% too large.
For the transition metal Ni, the LDA bandwidth is 4.5 eV as compared to 3.3 eV
experimentally; and (iii) the ionization energies of atoms, molecules or clusters are
usually underestimated.
Sham and Schluter[53, 54] and Perdew and Levy[55] have shown formally that
there is an explicit correction to the Kohn-Sham bandgap even when the exact
exchange-correlation functional is used, and therefore LDA is not responsible for
the above-mentioned difficulties. Methods to improve the LDA description of quasi-
particle energies have thus attracted considerable research interest. The GW ap-
proximation, generalized Kohn-Sham schemes (GKS)[49, 50, 51], and time-dependent
LDA[47, 48] represent recent research efforts along this direction. In the following,
we give an brief introduction to the GW approximation.
1.4.1 The GW approximation
As we have mentioned in the previous section, the Kohn-Sham eigenvalues do
not in general agree well with experimental excitation energies, which, in principle,
can be obtained from many-body perturbation theory by solving Dyson’s equation.
Excitations of a system of strongly interacting particles (e.g., electrons) can often be
20
described in terms of weakly interacting quasiparticles, which are composed of bare
particles plus polarization clouds surrounding them. The interaction between quasi-
particles is thus a screened, rather than bare, Coulomb potential. There are funda-
mental differences between a non-interacting single particle picture in the Kohn-Sham
scheme and a quasiparticle description. For example, a non-interacting particle has
infinite lifetime, whereas the lifetime of a quasiparticle is finite. The energy differ-
ence between a non-interacting particle and a quasiparticle is usually described by the
self-energy operator Σ, which must account for all exchange-correlation effects beyond
the Hartree approximation. The GW approximation of Hedin and Lundqvist[52] is
derived systematically from many-body perturbation theory with the self-energy ap-
proximated by a product of the one-particle Green function G and the dynamically
screened interaction W .
Green functions are powerful tools in many-body theory. Quasiparticle properties
such as energies, lifetimes, etc., can be determined by single-particle Green functions.
The one-particle Green function is defined as
iG(x, x′) = 〈N |T [Ψ(x)Ψ(x′)†]|N〉=
〈N |Ψ(x)Ψ(x′)†|N〉 for t > t′ (electron)−〈N |Ψ(x′)†Ψ(x)|N〉 for t′ > t (hole),
(1.63)
where x = (r, t), |N〉 is the N -particle ground state, Ψ(x) is the field operator and Tis the time-order operator. The Fourier transform of the one-particle Green function
G(x, x′), G(r, r′;ω), is related to the spectral function A(r, r′;ω) by
G(r, r′;ω) =∑i
φi(r)φ∗i (r
′)ω − ωi
=∫
A(r, r′;ω′)ω − ω′ dω′, (1.64)
and
A(r, r′;ω) =1
π|ImG(r, r′;ω)|, (1.65)
where φi are quasiparticle wavefunctions. The above expression is also called the
spectral representation of the Green function. The equation of motion of the one-
21
particle Green function is
[i∂
∂t−T−V ext]G(x, x′)+i
∫v(r, r′′)〈N |T [Ψ(r′′, t)†Ψ(r′′, t)Ψ(r, t)Ψ(r′, t′)†]|N〉dr′′ = δ(x, x′).
(1.66)
In the above equation, T is the kinetic energy operator, V ext is the external potential
and v(r, r′) is the interaction between particles, for example, the Coulomb interac-
tion for electrons. The term with four field operators contains a two-particle Green
function and describes the two-body correlations in the system. In principle, Eqn.
(1.66) can be taken as a starting point from which an infinite chain of equations is
generated to introduce successive corrections[52]. Another approach is to introduce
a nonlocal time-dependent self-energy operator Σ, defined by the following equation
[i∂
∂t− T − V (x)]G(x, x′)−
∫Σ(x, x′′)G(x′′, x′)dx′′ = δ(x, x′). (1.67)
V (x) is now the total average potential in the system[52], i.e., the external potential
plus the Hartree potential:
V (x) = V ext(x) + V h(r). (1.68)
The solution of Eqn. (1.67) can be formally written as
G(x, x′) = Go(x, x′) +∫
Go(x, x′′)Σ(x′′, x′′′)G(x′′′, x′)dx′′dx′′′, (1.69)
where G0(x, x′′) is the non-interacting one-particle Green function, which satisfies
[i∂
∂t− T − V (x)]G0(x, x′′) = δ(x, x′′). (1.70)
Eqn. (1.69) is called Dyson equation, which may be written in a more compact and
familiar form
G(k, ω) =Go(k, ω)
1− Σ(k, ω)Go(k, ω)(1.71)
in its Fourier space representation. For a time independent external field, the Fourier
transform of Eqn. (1.67) is
[ω − T − V (r)]G(r, r′;ω)−∫Σ(r, r′′;ω)G(r′′, r′;ω)dr′′ = δ(r, r′), (1.72)
22
and the corresponding quasiparticle wavefunctions satisfy the following equation
[T + V (r)]Ψnk(r) +∫Σ(r, r′;Enk)Ψnk(r′)dr′ = EnkΨnk(r). (1.73)
The complicated many-body character of the above equations is due to the fact that
the self-energy operator Σ is time (i.e., energy) dependent. Hartree and Hartree-Fock
approximations are two examples in which Σ is independent of energy:
ΣH = 0;ΣHF = V ex(r, r′),
(1.74)
where V ex is the exchange potential.
From Eqn. (1.73), our task to calculate the quasiparticle energy Enk and the
corresponding wavefunctions is reduced that of evaluating the self-energy operator Σ
and solving Eqn. (1.73). In general, the Enk are complex; the imaginary part gives
the quasiparticle lifetime. It can be shown that the self-energy Σ and the one-particle
Green function G are related by four integral equations, which are to be solved with
an iterative process[52]:
Σ(1, 2) = i∫
G(1, 3)Γ(3, 2; 4)W (4, 1+)d(34), (1.75)
W (1, 2) = v(1, 2) +∫
v(1, 3)P (3, 4)W (4, 2)d(34), (1.76)
P (1, 2) = −i∫
G(1, 3)G(4, 1+)Γ(3, 4; 2)d(34), (1.77)
and
Γ(1, 2; 3) = δ(1, 2)δ(1, 3) +∫
δΣ(1, 2)
δG(4, 5)G(4, 6)G(7, 5)Γ(6, 7; 3)d(4567). (1.78)
Here we adopt the abbreviated notation (1) ≡ x1 = (r1, t1), etc., and 1+ means that
t1 → t1 + δ where δ is a positive infinitesimal. We have introduced three more quan-
tities, namely the (dynamically) screened Coulomb interaction W , the irreducible
polarization propagator P , and the vertex function Γ. The screened Coulomb inter-
action W is related to the dielectric function ε(x, x′) by
W (1, 2) =∫
v(1, 3)ε−1(3, 2)d(3), (1.79)
23
where the dielectric function ε(x, x′) is
ε(1, 2) = δ(1, 2)−∫
v(1, 3)P (3, 2)d(3). (1.80)
Eqns. (1.75—1.78) can be solved iteratively. For example, we start with Σ = 0 in
Eqn. (1.78), which gives the first approximation of Γ
Γ(1, 2; 3) = δ(1, 2)δ(1, 3), (1.81)
from which we can evaluate Σ
Σ(1, 2) = iG(1, 2)W (2, 1+). (1.82)
The above expression of the self-energy Σ is called the GW approximation. Sometimes
it is more convenient to express the above quantities in their Fourier transforms (for
example, when working with pseudopotential plane-wave methods):
Σ(r, r′, ω) =i
2π
∫G(r, r′, ω + ω′)W (r, r′, ω′)dω′. (1.83)
The one-electron Green function can be calculated by
G(r, r′, ω) =∑n,k
φnk(r)φ∗nk(r
′)ω − εnk − iδnk
=∑k
∑G,G′
ei(k+G)·rGG,G′(k, ω)ei(k+G′)·r′, (1.84)
whereG andG′ are reciprocal lattice vectors, k is the wavevector in the first Brillouin
zone and n is the band index. εnk and φnk are the one-eletron eigenvalues and
the corresponding wavefunctions calculated by, for example, LDA. Similarily, for the
screened interaction W , we have
W (r, r′, ω) =∑k
∑G,G′
ei(k+G)·rWG,G′(k, ω)ei(k+G′)·r′, (1.85)
where
WG,G′(k, ω) = ε−1G,G′(k, ω)v(k+G′). (1.86)
24
The dielectric function ε is related to the irreducible polarizability P by
εG,G′(k, ω) = δG,G′ − v(k+G)PG,G′(k, ω). (1.87)
The real task in GW calculation is to evaluate the dynamic dielectric function
εG,G′(k, ω) or the irreducible polarizability PG,G′(k, ω) in Eqn. (1.87). Note that if
the dielectric function ε (and therefore W ) is calculated using the approximate ir-
reducible polarization propagator P ≈ P 0 = −iG(1, 2)G(2, 1), it corresponds to therandom-phase approximation (RPA)[36]. In most modern GW calculations, the dy-
namic dielectric function is approximated by extending the static dielectric function
calculated from LDA wavefunctions to finite frequencies using the so-called general-
ized plasmon-pole model[36].
Although the GW approximation for computing the quasiparticle energies was
proposed in the 1960’s[56, 52], it was not applied to real materials until the mid
1980s, starting with the work of Hybertsen and Louie[57, 58, 36]. Since then, the
GW approximation has been successfully applied to calculate quasiparticle energies
for various systems. For a review of the GW approximation and its applications, see,
for example, Aulbur, et al.[59].
1.4.2 Generalized Kohn-Sham schemes
The GW approximation has been shown to resolve many of the discrepancies be-
tween LDA band structures and experiments for a variety of materials, ranging from
metals to wide-gap insulators. Unfortunately, GW methods are usually computation-
ally intensive, making it impossible (or very difficult) to perform GW calculations for
large systems. Moreover, most of the GW methods calculate only the quasi-particle
excitation energies. In many cases, it will be very appealing if we can calculate both
the excitation energies and the total energy of a system at the same time with rea-
sonable accuracy. Although it is possible to calculate the total energy within the GW
approximation[60], the accuracy of such calculations has yet to be verified[49].
25
Recently, there has been some interest in the so-called generalized Kohn-Sham
(GKS) schemes[49, 50, 51], which attempt to develop KS-like total energy models
with improved quasi-particle description, while keeping the extra computational effort
to a minimum. For details of recently proposed GKS models, see refs [49, 50, 51].
26
Chapter 2
Carbon and Boron NitrideNanotubes
2.1 Introduction
Carbon is a rather special element which can form very flexible bonding networks,
for example, one dimensional (sp1) carbon chains or rings, two dimensional sp2 lay-
ered structures such as graphite, and three dimensional (sp3) diamond. In contrast,
no sp2 silicon phases are known to exist. It is remarkable that both graphite and dia-
mond are very stable. (Under normal conditions, graphite is just slightly more stable
than diamond). Whereas graphite is a zero-gap semimetal and has the highest in-
plane stiffness, diamond is a wide-gap insulator and is the hardest three-dimensional
material available so far. Diamond and graphite also have the highest thermal con-
ductivity of known materials in 3-D and 2-D. However, the story of carbon seems
to not have ended yet. The discovery of the nearly spherical molecule C60[61] and
related fullerene structures, and the subsequent synthesis of carbon tubules[62], now
referred to as carbon nanotubes, have inspired another wave of research into carbon-
based materials[63]. For an excellent review on various form of carbon materials (for
example, carbon fibers, glassy carbon, carbon black, amorphous carbon, etc.) see
Dresselhaus, et al.[63].
27
Figure 2.1: Rolling up a graphene sheet into a tube.
In a 1991 paper, Iijima reported the synthesis of a new type of carbon structure,
“needle-like tubes”, using an arc-discharge vaporization method similar to that used
for fullerene synthesis[62]. The original carbon tubules consisted of several (ranging
from 2 to about 50) concentric graphitic shells, and had diameters of a few nanome-
ters. These multi-shelled carbon microtubules are now called multi-walled carbon
nanotubes. Single-walled carbon nanotubes were later synthesized by arc-discharge
vaporization [64, 65, 66], laser ablation of carbon targets[67] and chemical vapor de-
position (CVD)[68, 69, 70] in the presence of a catalyst (Fe, Ni, or Co).
2.1.1 Specification of carbon nanotubes
From a purely geometrical perspective, a single-walled carbon nanotube is formed
by rolling up a graphene (single-layer graphite) sheet into a cylinder (see Fig. 2.1).
(Note that it is not likely that carbon nanotubes are formed this way in real ex-
periments. In fact, the nucleation and growth mechanisms of carbon nanotubes are
still an active research topic. Recently, we have proposed a new nucleation model
for carbon nanotubes[71], which is presented in Section 2.2.) A single-walled carbon
28
a1
a2
C
5a1
2a2
θθθθ
Figure 2.2: Specification of single-walled carbon nanotubes. The circumferentialvector shown in the figure corresponds to that of a (5,2) nanotube. The chiral angleθ is also shown.
nanotube is uniquely specified by its circumferential vector C = na1 + ma2[72] as
shown in Fig. 2.2, where a1 and a2 are the unit vectors of graphite, and n and m are
integers. A single-walled carbon nanotube with C = na1 + ma2 is called an (n,m)
nanotube.
Two particular classes of single-walled nanotubes are worth mentioning. One is
(n,0), also called zig-zag nanotubes, and the other is (n,n) or armchair nanotubes.
Generic (n,m) nanotubes are called chiral nanotubes. In contrast, (n,0) and (n,n)
nanotubes are achiral (non-chiral). The diameter d of an (m,n) nanotube can be
calculated by[63]
d =√3aC−C(m2 +mn + n2)1/2/π, (2.1)
where aC−C is the C–C bond length. The chiral angle (or wrapping angle) θ is
θ = tan−1[√3m/(m+ 2n)]. (2.2)
29
Kx
Ky
E
Figure 2.3: Near-Fermi surface electronic structure of graphene. The Fermi surfaceis reduced to two distinct (Fermi) points.
For (n,0) nanotubes, θ = 0o. However, the wrapping angle of (n,n) nanotubes is not
zero but θ = 30o (although they are also non-chiral). A double walled nanotube can
be specified by indexing each of the walls, for example, (5,5)@(10,10).
2.1.2 Novel electronic and structural properties
Carbon nanotubes have been the focus of intensive research for the past decade
due to their unique properties that could have impact on broad areas of science and
technology, ranging from strong and ultra-light composites[73] to nanoelectronics[74,
75, 76]. The novel electronic properties[77, 78, 79] of carbon nanotubes are related
30
K
Γ
K
Γ
(a) (b)
Figure 2.4: Allowed lines of wavevectors in carbon nanotubes: (a) the allowed linesmiss the Fermi points and (b) the allowed lines pass through the Fermi points. Thehexagon is the Brillouin zone of graphene.
to the unique Fermi surface structrure of graphene, combined with the quasi one-
dimensional nature of the tube.
Fig. 2.3 shows the 2-D tight-binding band structure of graphene, considering only
the pz orbital of carbon. The Fermi surface of graphene is reduced to two distinct
points in the Brillouin zone( BZ). Because of the periodic boundary condition of the
nanotube along the circumferential direction, only certain lines of wavevectors in the
BZ are allowed. If the allowed lines pass the Fermi points (corresponding to (n,m)
nanotubes with n−m = 3k, where k is an integer) the nanotube is metallic, otherwise
it is semiconducting, see Fig. 2.4. Therefore, from this purely geometrical argument,
we conclude that, depending on its wrapping indices, a carbon nanotube is either
metallic or semiconducting. Furthermore, since the dispersion of bonding and anti-
bonding pπ bands around the Fermi-points (see Fig. 2.3) is almost linear, the band
31
(8,8)(5,5) (7,7)(6,6)
(6,0)
(9,0)
(12,0)
(4,4)
Figure 2.5: Electronic properties of carbon nanotubes. Solid circles are metallic;triangles are small bandgap semiconductors and the rest are semiconductors withrelatively large bandgaps.
gaps of the semiconducting nanotubes must have an envelope function proportional
to 1/r, where r is the tube diameter.
The discussion thus far does not consider the effect of curvature on the electronic
structure of nanotubes. It turns out that the metallic nature of an (n,m) nanotube
with n−m = 3k and k = 0 is not stable against the perturbation induced by the cur-
vature, which opens up a small bandgap for these nanotubes. As a result, only (n,m)
nanotubes with n = m remain metallic[77] and are free from the Peierls distortion
at room-temperature[78]. (The Peierls distortion is strongly suppressed due to the
rigidity of the sp2 bonding network and the low density-of-states (DOS) of graphene
near the Fermi level, therefore can be easily destroyed by thermal effects.) Therefore,
we have the following general rules for predicting the band gaps of a carbon nanotube:
32
Nanotube Young’s Modulus (TPa)(6,6) 1.20(9,3) 1.15(10,2) 1.11(12,0) 1.03(10,0) 1.18(8,0) 1.52(6,0) 2.12
(a)
Nanotube Young’s Modulus (GPa)(6,6) 554(9,3) 525(10,2) 522(12,0) 500(10,0) 518(8,0) 577(6,0) 628
(b)
Table 2.1: Elastic strength of single-walled carbon nanotubes, (a) assuming the ef-fective cross-sectional area is πr2, where r is the radius of the tube, and (b) thecross-section is the area per nanotube in a nanotube bundle with inter-tube distance3.0 A.
(1) if n = m, metallic;
(2) if n−m = 3k, where k is an integer and k = 0, small band gap semiconductor;
(3) others, semiconducting with band gaps proportional to 1/r.
The exceptionally high elastic strength[80, 81, 82, 83, 84, 85] of carbon nanotubes
is due to the extremely strong sp2 bonding network of carbon. Table 2.1a lists the
Young’s moduli of several carbon nanotubes we calculated with tight-binding param-
eters by Xu, et al.[15], assuming that the effective cross-section is πr2, where r is
the radius of the nanotube (measured from the center of the tube to the carbon nu-
clei). However, a more reasonable estimate of the effective area might be the area per
tube in a nanotube bundle (or nanotube crystal). Rescaled results according to this
estimate are given in Table 2.1b.
In the following sections, we first propose a new nucleation model for carbon
nanotubes[71]. Detailed simulation using a hybrid TB/classical potential technique
shows that this model qualitatively agrees with experimental observations in several
aspects. Next, we address the issue of plastic deformation of carbon nanotubes under
high axial stress[86]. The elastic limits of carbon nanotubes are found to be higher
33
than any other known material and depend strongly on the chiral angle. Finally,
elastic limits of boron nitride nanotubes are studied and compared with those of
carbon nanotubes[87].
2.2 Nucleation of carbon nanotubes without pen-
tagonal rings
The unique one-dimensional structure of carbon nanotubes depends critically upon
a nucleation mechanism which somehow restricts graphitic growth to a single axis.
The transient and extreme growth conditions of carbon nanotubes have obscured
the mechanism of nucleation. Most hypothetical nucleation mechanisms invoke pen-
tagon formation to produce a hemispherical graphitic cap, with the cap’s edges either
open[88, 89, 67], attached to a substrate via a tubular segment[90, 91, 92], or mated
with a second cap into a closed fullerene[93, 94, 95]. These models all require an
abrupt transition from a regime favoring pentagon formation (when the cap is form-
ing) to one that favors exclusively hexagon formation (when the tube is lengthening)
once exactly six pentagons have formed. The last two models are also inconsistent
with experimental evidence favoring open-ended growth[96, 97]. Other models nucle-
ate a nanotube from a ∼40-atom polyyne ring[96, 98]. Unfortunately, such a large
carbon ring is unlikely: carbon clusters change from simple rings to double rings and
cages above ∼20 atoms[99, 100, 101].Very recently, Bandow et al. have discovered that the temperature during synthe-
sis controls the diameter of the nanotubes[102]. Since the diameter of a nanotube is
fixed by its size at nucleation, these results provide rare experimental clues into nu-
cleation itself. Here we propose a new nucleation model for carbon nanotubes which:
(1) contains only hexagonal rings within the tube nucleus and therefore does not re-
quire a transition from pentagon to exclusively hexagon formation, (2) can explain
the temperature dependence of the nanotube diameter distribution, (3) accounts for
34
Figure 2.6: A schematic diagram of the nucleation model in which a nanotube formsvia edge mediated opening of a double-layered graphitic patch.
the narrow diameter distribution of single-walled nanotubes, and (4) can explain the
wider diameter distribution of multi-walled nanotubes.
Carbon nanotube synthesis apparently requires a surface on either an electrode
or a metallic particle. In contrast, smaller closed-shell fullerenes typically form in a
low-density, dangling-bond-dominated gas phase. Surfaces favor the growth of small
flat graphitic patches. (Such graphitic patches can form even in much higher-density
environments[103]). It may be difficult to imagine a means by which a single graphitic
patch could curl into the nucleus of a nanotube without the incorporation of pentagons
to produce the needed curvature, as shown schematically in fig 2.1. One such nucle-
ation mechanism is for a single graphitic sheet to curl around a small particle with
the particle/graphite attraction overcoming the mean curvature energy of graphite
at a characteristic size of ∼ 1 nm. We herein describe a natural kinetic pathway by
which a two-layered graphitic patch can transform into the nucleus of a nanotube.
35
Fig. 2.6 illustrates schematically such a nucleation model. A double-layered
graphitic patch nucleates on a surface, with edges on adjacent layers interconnected
via bridge atoms[104]. When edges are bridged on opposite sides of the patch, such
a two-layered patch looks similar to a very short flattened nanotube[105, 106]. A
small-diameter flattened nanotube is unstable towards popping open, since the at-
traction between the inner surfaces cannot counteract the excess curvature energy on
the edges. If a double-layered patch pops open, then growth in the bridged direction
is arrested and further carbon addition lengthens the nucleus axially to form a nan-
otube. At least one experimental image directly supports this model by capturing a
possible intermediate state of nucleation: High resolution transmission electron mi-
croscopy images show a small-diameter flattened boron nitride nanotube extending
outward in one direction from a metallic particle and simultaneously opening into a
circular cross-section[107]. (However, not too much should be drawn from any single
TEM image of a rare intermediate state.) Since no pentagons are involved, the mech-
anism need not appeal to a fortuitous transition from pentagon formation to exclusive
hexagon formation during growth. The curvature necessary to form a tube nucleus
arises not from pentagons, but from the extreme curvature induced by bridging across
the interface between two adjacent parallel graphitic patches.
Although this model is qualitatively attractive, detailed numerical modelling is
necessary to determine if the energetic balance actually favors this route to nucle-
ation under realistic conditions. We use a tight binding total energy model[15] to
simulate the edge-mediated opening of a double-layered graphitic patch into a single-
walled nanotube nucleus. The tight binding parameterization accounts for the co-
valent electronic energetics and the nearest-neighbor repulsion of atomic cores. The
weak interlayer attraction is incorporated through an empirical Lennard-Jones po-
tential which yields the correct equilibrium interlayer distance and interlayer binding
energy (0.035± 0.015 eV/atom[106]) for graphite.
36
(a) (b)
Figure 2.7: The relaxed structures from T=0 molecular dynamics for two-layeredgraphitic patches of sizes necessary to form nanotube nuclei of diameters (a) 13.7 Aand (b) 27.4 A. The smaller nucleus spontaneously opens into a circular cross-sectionwhich defines a strong preferential growth axis whereas the larger nucleus remainsflattened.
We model two parallel layers of graphite 3.4 A apart with periodic boundary
conditions along one direction and bridge atoms connecting the exposed transverse
edges in the perpendicular dirction. For simplicity we set T = 0 in the simulation; the
primary effects of non-zero temperature are easily incorporated as will be explained
later. Within a few femtoseconds, bulbs form at the bridged edges as the bridge
atoms incorporate themselves into an energetically favorable continuous sp2 bonding
network. If the energetic cost of curvature in the bulbs overcomes the interlayer
attraction between the two layers, then the incipient tube pops open to a circular
cross-section, which then defines a preferential axis for one-dimensional growth. Fig.
2.7a shows this opening process during nucleation of a (10,10) tube. In contrast, the
37
nucleus of a (20,20) tube (Fig. 2.7b) does not pop open, as the interlayer attraction
dominates the energetics.
0
1
2
3
4
5
6
7
8
5 10 15 20 25 30
Energy (eV)
Antipodal Distance (Å)
D=20Å
D=27Å
D=33Å
Figure 2.8: The energy as a function of a reaction coordinate (i.e., the distance be-tween opposite inner surfaces as the tube nucleus opens) for nucleation patches ofvarying diameter when opened and of length 1 nm. The smallest patches have nobarrier to opening. For sufficiently large patches the opened state is energeticallydisfavored. For intermediate sizes, thermal excitation can activate opening on exper-imental time-scales.
Fig. 2.8 shows the total energy during the opening process for incipient nuclei
of (n, n) nanotubes with radii 20, 27, and 33 A (similar results are expected for
other wrapping indices (n,m)). Nuclei smaller than the cross-section of a D = 23 A
diameter tube have no energetic barrier to popping open. The size of the barrier scales
with the length of the tube segment; we choose a length of one nm, which produces
a roughly square patch with a plausible aspect ratio (we return to this point later).
In larger tube nuclei, the energetic barrier against assuming the circular cross-section
38
increases roughly linearly with size as
EB = ε(D − 23A) eV (2.3)
with ε ≈ 0.2 eV/A. The linearity follows from the linear dependence of the interlayer
“glue” area on the diameter D. Above D = 34 A the circular cross-section is en-
ergetically disfavored. Since the curvature energy is more important in multilayered
structures while the interlayer binding energy of the inner surfaces is constant[106],
multiwalled tubes could nucleate at larger diameters; double-walled nanotubes with
inner diameters of ∼45A have been observed[62].
At finite temperature, the system can thermally surmount the barrier towards
popping open. Experimentally, the yield of the smallest-diameter nanotubes (D ≤ 1
nm) is only weakly dependent on temperature[102], in accord with our model, as
these small nanotube nuclei have no energetic barrier to opening. In contrast, the
yield of larger-diameter nanotubes is very sensitive to synthesis temperature with
higher temperatures yielding larger-diameter tubes[102]. Given the energetic barrier
towards opening as a function of the incipient tube diameter, simple kinetic arguments
can then estimate the theoretically expected temperature dependence of the largest
possible nucleation patch produced on experimentally relevant timescales.
These kinetics are dominated by the size dependence of the activation barrier. The
thermally activated opening rate for an incipient nanotube nucleus is approximately
rate = f(R, T )ωvibe−EB/kBT , (2.4)
where f(R, T ) gives the temperature-dependent size distribution of appropriately-
shaped graphitic patches, EB is the calculated energetic barrier and ωvib is the attempt
rate (comparable to the characteristic phonon frequency for transverse vibrations of a
graphitic sheet at a wavevector comparable to the inverse tube radius, i.e., ωvib ∼ 1012
Hz). Several vibrational degrees of freedom might also facilitate opening, as could
39
be revealed by direct finite-temperature molecular dynamics. For small-radius tubes
the exponential factor is absent, hence the number of small-radius tubes produced
reflects directly f(R < R∗, T ) where R∗ is the largest diameter without an activation
barrier to opening. In fact, the number of smallest-radius tubes (R < 1 nm) produced
is nearly independent of temperature from T = 780 K to T = 1000 K[102]. Therefore
f(R < 1nm, T ) is a weak function of T for 780 K < T < 1000 K within the model.
Since the emergence of an energetic barrier at larger radii has no direct influence on
f(R, T ), f(R, T ) should have a relatively weak temperature dependence at somewhat
larger R as well, a regime in which the temperature dependence of the nucleation
rate is then dominated by the rapidly varying exponential term. At sufficiently high
temperatures other physical processes could preemptively cut off f(R, T ) for large R,
regardless of the exponential factor. Recent experiments above 1200 K (well above
the temperatures considered here) show a reduction in typical tube diameter[108].
The number of large-radius tubes falls off rapidly as a function of diameter[102],
consistent with a rapid falloff in the Boltzmann factor. Therefore we focus on the
exponential factor in analyzing the nanotube diameter distribution in the temperature
range from 780 to 1000 K. For example, at 1000 K, the single-site opening rate for
a barrier of 1.5 eV is roughly 104 sec−1. For a barrier of 2.0 eV the opening rate is
∼ 101 sec−1, too slow for significant tube production, while at EB = 1 eV the opening
rate is very fast, ∼ 107 sec−1. We take EB ≈ 1.5 ± 0.3 eV as the maximum barrier
beyond which nuclei do not form on the experimental timescale in the laser ablation
cloud or the dynamic plasma arc discharge.
As tube diameter (and energetic barrier) increases, maintenance of a constant
nucleation rate implies a higher temperature, i.e.,
EB,2/kBT2 = EB,1/kBT1. (2.5)
In other words, at higher temperatures a fixed nucleation rate occurs at a larger
40
radius. The overall diameter distribution at large diameters should shift upwards as
the temperature is increased. Examining the data of Bandow et al.[102] this upward
shift is (1.0−1.3)× 10−2A/K for both the peak and the downslope of the distribution.
Since the maximal surmountable barrier is much larger than kBT , a small change
in temperature produces a large change in the diameter of the largest tube which
can nucleate. As mentioned above, the difference in opening barrier for two tubes
depends linearly on their diameter difference,
∆EB = ε∆D. (2.6)
Where ε ≈ 0.2 eV/A. Therefore a change in temperature ∆T yields a typical diameter
change of∆D
∆T=
EBεT≈ 0.8× 10−2A/K, (2.7)
reasonably consistent with the experimental result of (1.0 − 1.3) × 10−2A/K for the
temperature variation in the most probable nanotube diameter[102]. This pop-open
nucleation mechanism sets a temperature-dependent upper limit on the nanotube
diameter distribution through suppression of thermally activated nucleation; the rel-
atively temperature-independent lower limit at D ≈ 0.8 nm[102, 109, 64, 65] might
then be set by the reduced chemical stability of the highly strained smallest-diameter
tubes[109].
Within this simplified model, the predicted diameter distribution is shifted to-
wards larger diameters than those observed (i.e., ∼ 29 ± 9 A versus the ∼15 A ob-
served at 1000 K; the error bar arises primarily from the uncertainty in the strength
of the interlayer attraction). This discrepancy is expected, for three reasons. First,
an attractive external surface adjacent to one graphitic layer reduces the propensity
for popping open. To quantify this effect, we introduced a rigid external surface with
the same attractive potential as graphite itself; this additional stabilizing force re-
duces the characteristic nucleus diameter by ∼ 3 A. Second, the incipient nuclei of
41
larger-diameter tubes would tend to be longer axially and therefore would have rela-
tively larger barriers to opening than the fixed-length nuclei herein considered. This
effect would also increase ε slightly but would not change the overall consistency with
experiment. Third, the calculations simulate an ideal nanotube nucleus wherein pe-
riodic boundary conditions suppress interlayer bridging along one direction. In more
realistic configurations, the irregular shape and registry of the graphitic patches and
the possible competition between bridging on different pairs of edges reduces the typ-
ical size at which popping open occurs. This competition does not prevent popping
open, since the first edge to bridge induces local bulging, which partially separates
the adjacent edges and leaves the opposite edge as the most likely to bridge next. The
differing reactivities of the zigzag and armchair graphitic edges might favor wrapping
indices around the armchair (n, n) configuration[67, 102, 110, 111]. Finally, patches
must grow from a small size. to pop open at a radius R, one must have failed to
pop open at smaller sizes. This reduces the chance of forming nuclei of large radius
nanotubes.
The nucleation mechanism for carbon nanotubes is one of the most important
open questions in nanotube synthesis. Although definitive determination on this
point awaits further experiments, we have described a plausible and qualitatively new
nucleation mechanism. This mechanism, the spontaneous opening of double-layered
graphitic patches, is quantitatively consistent with the recently discovered tempera-
ture dependence of the nanotube diameter and also produces the required curvature
without the necessity for the formation of exactly six pentagonal rings within the
incipient tube. Independent of the actual nucleation mechanism, the nucleation rate
is likely much higher for single-walled compared to multi-walled tube synthesis. Tube
nucleation and growth occur at temperatures (∼1000 K) where carbon atoms will
likely stick to exposed graphitic surfaces. If the nucleation rate is low, then the con-
centration of tube nuclei is low compared to the concentration of carbon atoms/dimers
42
and outer walls can accrete, particularly when the tubes are long and atoms cannot
quickly diffuse along the outer surface to the reactive ends. If the nucleation rate is
high, then the incipient tube density is high, and the tube nuclei are more likely to
collide with each other rather than with ambient carbon atoms/dimers, so the system
would tend to form bundles of single-walled tubes, as observed.
2.3 Plastic deformation of carbon nanotubes
Carbon nanotubes have remarkable elastic properties. On one hand, they are rel-
atively flexible under twisting[112], bending and squashing[113, 114, 115, 116, 117].
On the other hand, the axial elastic strength (measured by the Young’s modulus)
of carbon nanotubes is exceptionally high[80, 81, 82, 83, 84], of the order of 1 TPa
and is nearly independent of wrapping indices (see Table 2.1). As a comparison,
the Young’s moduli of titanium, stainless steel, and tungsten are 117 GPa, 193 GPa
and 400 GPa, respectively. The extreme elastic strength of carbon nanotubes is not
surprising since graphite is the stiffest planar material known. In fact, most struc-
tural applications proposed for carbon nanotubes, for example, ultra strong fibers
[73] and robust atomic force microscopy tips[118, 119], rely mainly upon their ex-
treme elastic strength. However, elastic response alone provides a limited picture
of mechanical properties. Beyond a certain stress, the elastic limit, a material de-
forms irreversibly via changes in chemical bonding topology. As we have mentioned
in Section 2.1, a nanotube is defined by its radius and chiral angle (or wrapping an-
gle), i.e., the wrapping index. Since the high symmetry of a graphite sheet implies
isotropic elastic properties, a nanotube’s radius alone dominates its elastic response.
Symmetry-breaking plastic distortions, by contrast, are quite sensitive to wrapping
angles.
A π/2 bond rotation in a graphitic network transforms four hexagons into two
pentagons and two heptagons, which is known as a Stone-Wales transformation[120].
43
(9,0) (6,6)
Figure 2.9: Two tubes with bond rotation defects.
The structure elongates along the axis connecting the pentagons and shrinks along
the perpendicular direction, as shown in Fig. 2.9. Thus, in a nanotube, rotation
of a bond from a predominately circumferential direction to a predominately axial
orientation lengthens the tube. If the tube is under tension, such defects relieve
strain and constitute a fundamental mode of plastic deformation. The efficiency of
this strain release depends sensitively on the alignment between the defect and the
tube axis. For example, essentially no strain is relieved by rotation of a bond initially
oriented π/4 from the circumference. One third of the bonds in an (n, n) tube are
aligned precisely with the circumference, the optimal orientation from which to rotate
for the release of strain. In contrast, the most favorable bonds in an (n, 0) nanotube
are misaligned by π/6 from the circumference, as shown in Fig. 2.9.
The occurence of a structural transformation such as plastic deformation is af-
fected by both thermodynamics and kinetics. Whereas the energy barrier controls
the detailed kinetics (i.e., transformation rate, intermediate states, and so on), the
44
phase diagram (i.e., the transition tension) is determined by the relative Gibbs free
energies of the elastically and plastically distorted structures. The energetic barrier
for the Stone-Wales transformation has been discussed extensively in the literature
for the carbon system[121, 122, 123, 124]. As our primary goal is to determine the
phase diagram of carbon nanotubes under tension, we focus on the relative Gibbs free
energies of the elastically and plastically distorted structures.
We use quantum-mechanical atomistic calculations to investigate the plastic de-
formation of carbon nanotubes under strain. A tight binding total energy scheme[15]
is employed. To separate the effects of tube radius from those of wrapping indices,
we study two families of tubes, a series of (n, 0) tubes of different radii and a series
of (n,m) tubes of nearly equal radius. In all cases we first consider transitions at
constant tension, later addressing the transitions at constant length.
Fig. 2.10 shows the energy versus length for the pristine and plastically deformed
(8,0) tube. Since a single defect would produce an inconvenient kinked unit cell,
the deformed tubes we studied contain a pair of well-separated rotated bonds on
opposite sides (but all quoted results are normalized to a single defect). The common
tangent of the energy-versus-length curves gives the critical force (Fc) at which the
Gibbs free energies G = E − FL of the two configurations are equal; Fc = δE/δL,
where δE and δL are the energy difference and elongation across the transition. (This
construction is familiar in three dimensions as yielding the transition pressure between
two structures.) This structural phase transition to a defective phase is eventually
arrested by a repulsive defect-defect interaction at high defect density. In our study
the defect density is low enough that this interaction does not significantly disturb the
onset of plastic deformation (doubling the defect separation changes the transition
tension by less than 1%). Since one expects Fc roughly proportional to the radius,
R, of a tube, we define the reduced transition tension fc ≡ Fc/R, which should vary
weakly amongst nanotubes of identical wrapping angle.
45
0
10
20
30
40
50
60
12 12.5 13 13.5 14 14.5 15
Energy (eV)
Length (A)˚
(8,0)
Figure 2.10: Energy versus length for pure (solid) and defective (dashed) (8,0) nan-otubes. The cell contains 96 atoms per bond rotation defect with E = 0 for the pureundistorted tube.
We define the wrapping angle χ as the angle between the circumferential direction
and the best bonds to rotate. χ ranges from zero for an (n, n) tube to π/6 for an
(n, 0) tube. Note that χ is different from the wrapping angle we defined in Section 2.1
but relates to the latter by χ = 30o− θ. The C2v symmetry of a bond rotation defect
implies that physical quantities are even functions of χ with period π. Hence, to low-
est order in angular dependence, they are well approximated by a(R) + b(R) cos 2χ,
with radius-dependent coefficients a(R) and b(R). Fig. 2.11 reveals a nearly linear
dependence of transition tension on cos 2χ for the family of nearly equal-radius tubes
(12,0), (10,2), (9,3), (8,4), (6,6). The elastic limit varies from about 100 nanoNew-
tons per nanometer for the (6,6) tube to about 180 nN/nm for the (12,0) nanotube.
The radial dependence of the reduced transition tension is signficant only for the
46
6
7
8
9
10
11
12
0.4 0.5 0.6 0.7 0.8 0.9 1
ReducedTransitionTension(eV/A )˚2
Cos(2χ )
(12,0)
(10,2)
(9,3)
(8,4)
(6,6)
(10,0)
(8,0)
(7,0)
(6,0) (nN/nm)
192
176
160
144
128
112
96
Figure 2.11: Transition tensions to plastic deformation for a family of (n, 0) and oneof nearly equal-radius (n,m) tubes. The radial dependence is quite weak except forthe smallest tubes, whereas the wrapping angle dependence is very strong.
smallest tubes.
The radial and angular dependences of the critical tension can be understood by
treating the defect as an extended strain in an elastic continuum with parameters
extracted from the atomistic calculations. For the tension Fc at the transition, the
defect density is vanishingly small and the energy (δE) and length (δL) changes
upon introduction of one defect are close to their values at zero tension. Therefore
the reduced transition tension is well approximated by
FcR=1
R
δE
δL
∣∣∣F=0
. (2.8)
Since the length change δL depends upon only R and the integrated strain of the
defect (with units of L2), δL scales as R−1 with angular dependence in the form
a+ b cos 2χ. Fitting to the atomistic data gives
δL ≈ 3.4A2
2πR(0.05 + cos 2χ), (2.9)
47
which differs by only 15% from a very simple geometrical analysis assuming regular
polygons. The 1/R dependence of the elongation cancels the normalization prefactor
of Eqn. (2.8) and therefore implies that the radial dependence of the critical tension
fc arises only through the defect energy δE. The nanotube at zero tension can freely
relax the defect strain both axially and circumferentially, so this defect energy in
tubes of moderate radius is a weak function of the wrapping angle (4.0 ± 0.1 eV for
the family (12,0), (10,2), (9,3), (8,4), (6,6)). Therefore the angular dependence
of the plastic limit arises mainly from the angular dependence of the defect-induced
elongation and the radial dependence of the plastic limit arises mainly from the radial
dependence of the defect energy.
The detailed radial dependence of the defect energy is a complex combination of
bare defect energy, the elastic energy induced in the surrounding medium and the
curvature-induced rehybridization of the defect. The bare energy is of order 5 eV.
The elastic term is small, of order µ(2πR)−2(∫ε2 d2x) ≈ 0.03 eVnm2/R2, where µ
is the in-plane elastic modulus of graphene and ε is the strain. The rehybridization
energy dominates the radial dependence; narrower nanotubes of the same χ typically
have lower defect formation energies since (i) the additional distortions due to bond
rotation are more easily accomodated into the curved surface of a narrower tube and
(ii) the defective region has a slightly lower mean curvature modulus than a corre-
sponding pristine region. This reduction in defect formation energy with decreasing
radius explains the lower plastic threshold in the smallest (n, 0) tubes. Note that the
radial dependence of the plastic threshold asymptotes (on increasing R) at a relatively
small radius whereas the pronounced χ dependence remains at all sizes.
Beyond the threshold for plastic deformation, the defect density is more easily
controlled by fixing the length rather than the tension. At constant length, the linear
defect density 9 is proportional to the fractional extension beyond the critical length:
9 = (L−Lc)/L δL ≈ (ε−εc)/((1+εc)δL), where ε = (L−L0)/L0 is the axial strain with
48
εc its critical value. Just as the critical force is very sensitive to wrapping angle, so the
critical strain varies from about 6% extension for a (6,6) tube to 12% extension for
a (12,0) tube. These strains greatly exceed those typical of the plastic deformation
in bulk solids, in accord with expectations for the extremely robust bonding in a
defect-free carbon nanotube at T=0. Above this critical extension the defect density
increases rapidly with length, varying as 24(ε − 0.06) nm−1 for the (6,6) tube and
as 43(ε − 0.12) nm−1 for the (12,0) tube. The defect concentration increases most
rapidly in the (n, 0) tubes since in these tubes each individual defect is least efficient
in relieving strain. Since δL varies as 1/R, larger tubes will contain more defects than
smaller tubes of the same wrapping angle at fixed relative strain, although the areal
density of defects remains constant (except for the smallest tubes where the defect
formation energy is a strong function of the radius).
Under an external force above the plastic threshold, proliferation of bond-rotation
defects is eventually halted by repulsive defect-defect interaction. The elastic inter-
action of two defects a distance D apart varies roughly as µ/D2 times the product of
their integrated strains. The strength of this interaction is on the order of 0.08nm2/D2
eV. This simple estimate is consistent with tight binding total energy calculations on
the (6,0) tube, which yield a 0.13 eV increase in the repulsive defect-defect interac-
tion for a reduction in D from 7 to 13 A. A uniform density of defects then has an
interaction per defect quadratic in the linear density λ (neglecting alteration of the
tube’s spring constant by the defects), so that λ ∝ (F − Fc)1/2 at constant force.
These defect densities as a function of strain are interesting not only for the me-
chanical response: the bond rotation defect is also a plausible route towards modu-
lating the electronic properties of carbon nanotubes. An extension of only 1% beyond
the critical length induces enough bond rotation defects to signficantly alter the elec-
tronic structure, increasing or decreasing bandgaps and modulating the density of
states[125]. The repulsive defect-defect interaction may tend to favor periodic defect
49
arrangements and thereby minimize the random component of the defect-induced
potential felt by the electronic states.
The discussion thus far has been in a purely equilibrium context, ignoring the
kinetics of defect formation. We have also calculated the barrier to defect formation
at critical external tension by calculating energy as a function of bond orientation,
considering both in-plane and out-of-plane rotations. Of the kinetic pathways exam-
ined, the lowest barrier for bond rotation in a (6,6) tube occurs when the bond is
tilted roughly 15o out of plane, which yields an upper bound of approximately 4.0
eV for the barrier at the transition, decreasing with increasing strain. This compares
with the 5.4 eV barrier for the related Stone-Wales transformation in C60[123] and
the 10.4 eV barrier for bond rotation in flat graphite[124]. The barrier height could
be further lowered dynamically by thermal fluctuations in the neighborhood (distinct
from thermal excitation of the reaction coordinate itself). One can ascribe the lower
barrier in the strained nanotube to the external tension. The bond rotation barrier
at zero external tension is not symmetric; it is about 8.5eV/defect from the defect
free side and 5.5eV/defect from the defective side. This relatively large barrier pre-
vents both the creation of bond rotation defects and annealing of the defects after
the external strain is released at room temperature. Bond rotation defects comprise
a pair of pentagon-heptagon defects (5-7 and 7-5) which can be separated by ad-
ditional bond rotations. These pairs represent crystal dislocations; the long-range
strain fields favor the gliding of these pairs in opposite directions under an external
stress[126]. However, the local curvature energy of the double pentagon-heptagon
structure is minimized when the oppositely directed curvature dipoles are adjacent.
In a flat unstrained graphitic sheet, tight binding total energy calculations reveal a
very strong short range attraction between a 5-7 and 7-5 defect. Separating them
by one intervening hexagon costs 4.5 eV, while separation by two hexagons costs an
additional 3.5 eV, although in certain situations this barrier may be lower[127]. This
50
tall and wide energetic barrier will help suppress fission of the bond rotation defect.
It will be very interesting to test our predictions experimentally. We here propose
an experimental geometry in which these defects could be created and measured in a
controlled fashion. A nanotube could be securely fixed at each end (e.g., embedded
within blobs of solid material) and then pulled taut and pressed transversely in the
middle of the free section, s s? . This geometry would create a very large mechanical
advantage of order δH/L where L is the suspended length and δH is the transverse
deflection. In addition, the change in tube length upon bond rotation would yield
a greatly amplified transverse deflection. If tube extension decreased the external
force, then the defects would appear one at a time, giving a characteristic signture of
successive quantized defections of the external probe. Note that the defects must be
created via tension, not compression, as a nanotube buckles under compression.
Although the nanotube wrapping angle controls the electronic properties through
the overlay of periodic boundary conditions on the graphitic semimetallic band struc-
ture, the nanotube elastic properties are insensitive to structural details. However,
the wrapping angle reasserts its importance under symmetry breaking plastic dis-
tortions. This strong sensitivity to wrapping angle will affect high-load structural
applications of carbon nanotubes. The electronic consequences of these plastic defor-
mations also open a possible new route towards modulating electronic properties of
carbon nanotubes. Our theoretical prediction on the exceptionally high elastic limits
of carbon nanotubes are consistent with several recent experiments[128, 129, 130].
2.4 Plastic deformation of carbon boron-nitride nan-
otubes: an unexpected weakness
As far as structural properties are concerned, boron nitride systems resemble car-
bon systems in several respects. Both boron nitride and carbon can form cubic and
layered structures. Cubic boron nitride materials are also well-known as hard mate-
51
rials, comparable in hardness to diamond. Therefore, it is not surprising that boron
nitride can also form nanotubes. Unlike carbon nanotubes, which were first observed
in experiment[62], boron nitride (BN) nanotubes were proposed by theorists[131, 132]
and subsequently synthesized in experiments[133]. Indeed, nanotubes based on other
layered compounds have also been successfully synthesized[134].
The elastic strength of BN nanotubes has been shown to be comparable to that of
carbon nanotubes[80, 81, 135, 136]. Therefore, it would be very interesting to compare
the elastic limits of boron nitride nanotubes against those of carbon nanotubes. Since
the boron nitride structure breaks the symmetry of the bipartite graphitic lattice, the
mechanical behavior of BN tubes beyond the elastic limit could differ significantly
from that of their carbon counterparts[86, 121]. In addition, BN tubes have uniformly
larger bandgaps than carbon tubes[131, 132]. This contrast in mechanical response
is potentially important for situations which might require, for example, insulating
behavior coupled with high mechanical toughness. Here we demonstrate that there are
qualitative differences between carbon and boron nitride nanotubes in their response
to high tensile stress. Whereas an (n, n) BN nanotube has an elastic limit against
bond rotation which is comparable to that of an (n, n) carbon nanotube, that of
an (n, 0) BN nanotubes is as much as 40% lower than that of the corresponding
(n, 0) carbon nanotubes In addition, the bond-rotation defects induced by external
tension can induce bond breaking in (n, 0) BN nanotubes even at zero temperature,
whereas carbon nanotubes can sustain greater strains without breaking bonds. These
interesting differences in the structural properties under high tensile stress arise due
to the presence of B–B and N–N bonds during the plastic deformation which we
consider.
The Stone-Wales transformation we discussed in the previous section can be read-
ily applied to BN nanotubes as shown in Fig. 2.12 with subtle differences (discussed in
detail later). (Note that the intermediate states in the rotation of the bond need not
52
1.83(Å)
(Å)1.92
B
N
N
B
tube
axi
s
(a)
1.70 (Å)
1.68 (Å)
B
N
N
B
tube
axi
s
(b)
C
CC
C(Å)1.60
(Å)1.60
tube
axi
s
(c)
Figure 2.12: Bond rotation defects in (a) (n, 0), (b) (n, n) BN nanotubes, and (c)(n, 0) and (n, n) carbon nanotubes. There are two distinct bond rotation patternsin a generic (m,n) boron nitride nanotube. For an (8,0) BN nanotube, the energydifference between the two patterns is about 0.2 eV/defect at equilibrium (i.e., zerotension), while for (n, n) tubes the two states are degenerate. Various bond lengthsmarked (not shown to scale) correspond to tubes under about 12% strain.
53
be in-plane[86, 122], although under tension the intermediate states tend to become
more planar in character). As we have shown in the previous section, such bond rota-
tions in a nanotube can relieve external strain at a moderate formation energy with
an energy barrier that decreases significantly with increasing tensile stress (also see
ref. [86, 121]). For example, under about 10% strain, the energy barrier against bond
rotation in an (n, n) carbon nanotube decreases by 30 – 50% from the value in a C60
fullerene[122], with reasonable agreement between tight binding total energy[86] and
classical Tersoff-Brenner potential[121] models. Since the character of the directional
bonding in planar boron nitride is similar to that in graphitic systems, we expect this
bond rotation deformation mode to be important also for BN nanotubes.
Since bond rotation defects in a BN sp2 network create energetically unfavorable
B–B and N–N bonds, even membered rings (for example, 4- and 8-membered rings,
which require no B–B and N–N bonds) might be more stable[133, 137]. We have ex-
amined this possibility and found that the kinetic route to the formation of such 4884
defects within a hexagonal network is multi-stage and complex, with the final state
involving very large distortions of the sp2 network. An alternative 4875 distortion
is kinetically accessible, but only by passing through the 5775 distortion considered
here as an intermediate state. Therefore, while even-membered rings can be favored
in the nanotube cap, local defects induced by external stress are less likely to be com-
prised exclusively of even membered rings. For (n, 0) nanotubes, due to the particular
arrangement of the bonding network, direct bond breaking along the axial direction
might also be an important deformation mode; we will discuss this briefly later.
We employ a non-orthorgonal tight binding total energy scheme as developed
by Porezag and co-workers[24]. Instead of fitting the tight binding parameters to
functional forms, the repulsive potentials and hopping elements, which are obtained
by fitting ab initio results, are tabulated and interpolated. In relaxed structures of
BN nanotubes, the nitrogen atoms buckle outwards. In other words, a single-walled
54
BN tube has two concentric cylinders, one of solely boron, the other nitrogen. The
difference in the radii of the two cylinders decreases from about 0.2 atomic units (1
atomic unit=0.529 A) for a (6,0) tube to 0.1 atomic units for a (12,0) tube, which is
in good agreement with a previous ab initio calculation[132]. The equilibrium B–N
bond length is about 1.46 A, very close to the bond length of planar hexagonal boron
nitride. This good agreement shows the reliablity of the tight binding parameters
under these conditions.
Fig. 2.12a and 2.12b shows the bond rotation defects in (n, 0) and (n, n) boron
nitride nanotubes. Note that whereas there is only one favorable bond rotation pat-
tern for carbon nanotubes under strain, the situation is a bit more complicated for
BN nanotubes due to the sublattice symmetry breaking. For a generic (n,m) tube
with n=m, clockwise and counter-clockwise bond rotations are distinct. The extremecases are the (n, n) and (n, 0) tubes. For an (n, 0) tube, the difference between the
two choices is maximal: B–B and N–N bonds after the bond rotation align nearly
along either axial or circumferential directions. Tight binding total energy calcula-
tions show that the former is energetically favorable by 0.2 eV/defect. Therefore we
focus on this lowest-energy bond rotation defect. For an (n, n) BN tube, the two
bond rotation patterns are degenerate. The bond rotation defects in (n, 0) and (n, n)
carbon nanotubes are also shown in Fig. 2.12c for comparison. The existence of a
lower-energy bond rotation pattern for (n, 0) tubes is one of the reasons that (n, 0) BN
nanotubes have significantly lower elastic limits than do corresponding (n, 0) carbon
nanotubes, as will be discussed further below.
Table 2.2 compares the elastic limits (i.e., the transition tension) Fc and the
reduced transition tensions fc of carbon and BN nanotubes from the pristine structure
into a plastically deformed bond-rotated phase. (For definitions of Fc and fc, see the
previous section.) Both carbon and BN (n, 0) tubes have higher elastic limits to this
particular failure mode than do (n, n) tubes of similar radii, the reason for which has
55
(8,0) (12,0) (6,6)Fc fc Fc fc Fc fc
(eV/A) (eV/A2) (eV/A) (eV/A2) (eV/A) (eV/A2)BN 20.5 6.4 32.4 6.8 21.0 5.1C 32.7 10.4 51.2 10.9 25.8 6.3
Table 2.2: Comparison of the elastic limits of boron nitride and carbon nanotubes,showing both full and reduced transition tensions (as defined in the text). The reducedtransition tensions for (n, n) carbon and boron nitride nanotubes are similar, but thesequantities differ greatly for (n, 0) carbon and boron nitride nanotubes.
been discussed in the previous section.
As shown in Table 2.2, an (n, n) BN nanotube has an elastic limit only slightly
lower than that of an (n, n) carbon nanotube. This difference can be understood by
considering the approximation for the transition tension[86] we have discussed in the
previous section, namely Eqn. (2.8). The defect formation energy δE in a (6,6) BN
nanotube is smaller than that of a (6,6) carbon nanotube, 3.13 eV versus 3.95 eV
from our tight binding calculations, while the elongation δL is slightly shorter, 0.13
A versus 0.14 A. Taking into account these two factors, this simple model predicts
the elastic limit of a (6,6) BN nanotube to be about 83% of that for a (6,6) carbon
nanotube, which is very close to the actual result (81%) from the detailed simulations.
The elastic limit of an (n, 0) BN nanotube is, however, only about 60% of that for
the corresponding (n, 0) carbon nanotube. The reason that (n, 0) BN nanotubes have
much lower elastic limits is twofold. First, as mentioned above, the lower symmetry
of BN nanotubes results in two distinct bond rotation patterns with different defect
formation energies. If the elastic limit of the (8,0) BN nanotube were calculated
with the higher formation energy defects, the transition tension would have been 22.7
eV/A, as compared to the actual value of 20.5 eV/A. This is, however, still much
smaller than the transition tension of an (8,0) carbon nanotube, which is 32.7 eV/A.
Therefore, the lower defect formation energy alone cannot explain the discrepancy.
56
The remaining discrepancy can be explained by carefully examining the structural
properties in the vicinity of the defect in a BN nanotube. We find that whereas bond
rotation defects in all carbon nanotubes are rather robust and can sustain considerable
tension beyond the elastic limit (at low temperature) without breaking bonds, the
defective area in an (n, 0) BN nanotube is very vulnerable to bond breaking under
tension. Although both carbon and BN nanotubes have an sp2 bonding network, the
BN bonding is partially ionic and less robust under deformation. Rotating a bond
creates unfavorable B–B and N–N neighbors. These homopolar bonds weaken the
entire defect area: The local strain in the defect area is then considerably larger than
in other parts of the tube. Due to the particular arrangement of bonds in an (n, 0)
tube, two bonds align well with the axial direction in the defect area (the shaded
bonds in Fig. 2.12a). These two bonds stretch the most under stress and are then
susceptible to subsequent breakage. For example, for an (8,0) BN nanotube under
12% overall strain, about 30% stretch is observed for those particular bonds (Fig.
2.12a). The most-stretched bond is clearly in the process of breaking. In contrast,
no particularly over-stretched bonds are found in a carbon (8,0) nanotube under the
same amount of strain (Fig. 2.12c). For (n, n) tubes, there are no bonds that align
exactly along axial direction. As a result four bonds (shaded bonds in Fig. 2.12b)
share the stress around the defect area. No bond breaking is observed in an (n, n)
BN nanotube up to 12% strain at zero temperature. Therefore, we conclude that the
weakness of the defect area in a BN nanotube, whose effect is enhanced in an (n, 0)
nanotube, combined with the lower defect formation energy, implies that an (n, 0)
BN nanotube has a much lower elastic limit against this deformation than does a
corresponding (n, 0) carbon nanotube. We show here only the results for (n, 0) and
(n, n) tubes, which represent two extreme cases. For generic (n,m) tubes, the elastic
limit should lie in between that of (n, 0) and (n, n) tubes of similar radius. Note that
kinetic barriers could increase the observed transition tension under certain thermal
57
conditions, whereas preexisting imperfections in the tube (e.g., vacancies) could allow
alternative deformation modes with lower thermodynamic transition tensions.
We now consider one alternative mode of plastic deformation which is relevant
even in ideal systems, namely direct bond breaking. Such a deformation is expected
to be most favorable for (n, 0) tubes wherein certain bonds point directly along the
axial direction. We analyze this possibility by breaking a small fraction of the axially
aligned bonds (i.e., extending them to 2.1 times their normal length) in an (8,0) tube
under the same 9% strain at which bond rotation deformations are thermodynamically
favored. We then relax the structure in two steps, under a constant 9% strain. In
the first step the two atoms comprising the broken bond are fixed. This prevents
the broken bonds from reforming before the neighboring atoms relax into a favorable
configuration. Next, we relax the entire structure without any constraint, i.e., every
atom is allowed to move. After this two-step relaxation, the system spontaneously
heals the broken bonds. These results suggest that the bond-broken state is not a
local minimum of the Gibbs free energy at the strain at which bond rotations are
thermodynamically favored. Therefore direct bond breaking is likely not a favorable
mode of plastic deformation at low temperature thermodynamic under conditions of
constant strain.
In summary, we used a nonorthogonal tight binding total energy model to study
the bond rotation mode of plastic deformation of BN nanotubes. Whereas boron
nitride nanotubes have elastic properties similar to those of their carbon counter-
parts, they have a very different response to plastic deformation. Apart from the
already-understood wrapping angle dependence of this deformation, we found that
BN nanotubes with wrapping indices n m have unusually low-tension failure modes
which contribute to brittle failure via bond breaking. Therefore, although (n, n) boron
nitride nanotubes have comparable plastic limits to those of carbon tubes with sim-
ilar wrapping angle, (n, 0) boron nitride nanotubes have much smaller elastic limits
58
against bond rotations. In (n, 0) tubes, bond rotation defects induced by plastic de-
formation quickly lead to bond breaking and thereby cause tubes in this range of
wrapping angles to be substantially weaker than would be expected from their purely
elastic response. These results could have relevance to the choice of materials for
future high-strength nanomechanical applications of carbon and boron-nitride based
systems.
59
Chapter 3
Electronic and StructuralProperties of Novel Group-IVCompounds Formed from CVDPrecursors
3.1 Introduction
In the previous chapter, we explored some of the interesting properties of car-
bon nanotubes. Actually, most group-IV elements possess fascinating and unique
properties. Moving downward in the periodic table, silicon is undoubtedly the most
technologically important element; germanium has very high hole mobility. Whereas
grey tin (α-tin, diamond structure) is a semimetal, white tin (β-tin, double body
center cubic structure) is a metal. The transition from α-tin to β-tin happens just
slightly below room temperature at atomospheric pressure. Also, the band gaps of
elemental group-IV (C, Si, Ge, and Sn) materials vary from zero (graphite, α-Sn) to
5.5 eV (diamond).
Recently, group-IV alloys have attracted considerable research interest due to their
wide tunability in both structural and electronic properties and the hope that such
alloys might combine the above mentioned properties in new ways[138, 139, 140, 141,
142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152]. Incorporation of substitutional
60
carbon into other group-IV semiconductors is of particular interest. On one hand,
the small radius of carbon might help reduce the lattice mismatch between bandgap
engineered Si/Ge materials and the silicon substrate; on the other hand, the presence
of strongly electronegative carbon atoms, along with the strain introduced by them,
offers a unique way to tailor the electronic structure with a compelling possibility of
obtaining direct gap group-IV semiconductors that lattice match silicon[153].
Although there has long been speculation that alloying carbon with other group-
IV materials might produce interesting properties, little progress was made towards
this goal until recently. The major obstacle for synthesizing group-IV compounds with
high concentrations of carbon is that the solubility of carbon in silicon is very low: less
than 10−4 at. % at the silicon melting point; in germanium the solubility is negligible.
Fortunately, recent advances in synthesis have demonstrated that tetrahedral group-
IV compounds containing high concentrations of carbon are possible[138, 139, 140].
More recently, Kouvetakis et al. reported a synthesis[147, 148] based on UltraHigh-
Vacuum Chemical Vapor Deposition (UHV-CVD) that allows the production of semi-
conductor alloys with well-defined (and thermodynamically inacccessible) 1:4 C:Si and
C:Ge stoichiometries through the use of precursor molecules which build in the re-
quired interatomic bonding, for example, C(SiH3)4. The terminal groups (e.g. H)
are eliminated to produce the pure C/Si (or C/Ge) ordered alloys, such as Si4C or
Ge4C. Such techniques allow one to produce bulk amounts of metastable alloys of con-
trolled stoichiometry which are not accessible by standard near-equilibrium synthesis
methods.
In this chapter, we present systematic ab initio results on the electronic and struc-
tural properties of Si4C, Ge4C, Sn4C and other related group-IV compounds that ei-
ther have already been synthesized or are designed to be accessible to the UHV-CVD
technique using precursor molecules as building blocks. For Si4C and Ge4C, the band
gaps remain indirect while for Sn4C, the the band gap is direct but slightly negative
61
(-0.1 eV). Since the local density approximation (LDA) underestimates the band gap,
we expect Sn4C to have a moderate positive direct band gap (of the order of 0.5 eV)
if quasiparticle corrections are taken into account. In Section 3.3, we will study the
properties of Si4C, Ge4C and Sn4C.
Another subject of interest is silicon-based optically active semiconductors. For all
of their exciting properties, the group-IV (both elemental and compound) semicon-
ductors known so far have indirect energy gaps (not including zero gap materials).
Integration of silicon-based electronics with optical components requires optically
active materials that can be grown on silicon with high-quality interfaces. Unfortu-
nately, no direct-bandgap semiconductors have been produced which lattice-match sil-
icon (and previously suggested structures pose formidable synthetic challenges[154]).
The resulting lattice mismatch generally causes cracks and poor interface properties
whenever the mismatched overlayer exceeds a very thin critical thickness. Much re-
cent work has focused on introducing compliant transition layers between mismatched
components[155, 156, 157]. It might also be possible to make optically active group-
IV materials that are compatible with silicon technologies. Most previous efforts
along this direction have resorted to band folding, surface states and quantum con-
finement. Examples are Si-Ge superlattices[158, 159, 160, 161], porous silicon and
silicon nanowires[162, 163, 164, 165]. Si-Ge superlattices usually have poor optical
properties[159, 160]. The mechanism of light emission of porous silicon is often as-
cribed to the quantum confinement, surface states and/or band folding[166, 167, 168],
although the physical origin is still not fully settled.
We have suggested a more direct solution to this problem[153], proposing two novel
group-IV compounds, namely, Si2Sn2C and Ge3SnC, which have direct energy gaps
and lattice match silicon to within 1% along low-index lattice planes. The bandgaps
of these two group-IV compounds lie within the window of minimal absorption in
current optical fibers. Detailed ab initio studies of Si2Sn2C and Ge3SnC are presented
62
in Section 3.4.
One can build in even higher concentrations of carbon using C2(SiX3)6 and C2(GeX3)6
as precursors for solid Si6C2 or Ge6C2. Although these two compounds contain
C–C bonds, which have long been regarded as energetically unfavorable in group-IV
alloys[142], our detailed ab initio calculations show that they are energetically compa-
rable to the already-synthesized group-IV compounds, e.g., Ge4C. Due to the relative
high strain in these two materials, Si6C2 and Ge6C2 are semi-metallic within LDA.
(Strain induced gap closing has been observed previously in Si-C random alloys[141].)
The highly anisotropic C–C bonds in Si6C2 and Ge6C2 should have significant effects
on the electronic and structural properties, as discussed in Section 3.5.
3.2 Computational details
We use the ab initio pseudopotential plane-wave method[37] in the local density ap-
proximation with norm-conserving pseudopotentials[43] in a nonlocal form suggested
by Kleinman and Bylander[46], as described in chapter 1. The parametrization of
the Ceperley-Alder results[169] by Perdew and Zunger[170] is used for the exchange-
correlation. For germanium and tin, scalar relativistic effects[171] are taken into ac-
count. In all of the calculations, we include the nonlinear partial core correction[172]
to account for the nonlinearity of core-valence exchange-correlation energy. The en-
ergy cutoff of the plane-wave basis is set at 60 Ryd to ensure the convergence of the
calculations, since all of the systems studied include carbon atoms. The k-point set is
generated by the Monkhorst-Pack scheme[173] with a density of 3×3×3, which wouldbe equivalent to a k point density of about 6×6×6 in a more familiar 2-atom diamond
unit cell.
63
3.3 Si4C, Ge4C and Sn4C
Fig. 3.1 shows the precursor molecule C(AX3)4 (A=Si, Ge or Sn) and a correspond-
ing ordered crystalline structure. (Although Sn4C has not been synthesized yet, we
extend our study to this compound for comparison purposes and to gain a better
understanding of the general trends arising from differences in electronegativity and
core radii.) The smallest possible unit cell of these compounds contains 10 atoms and
has a body-centered tetragonal (bct) lattice. In these systems, there are no nearest
or second nearest carbon–carbon pairs, which are normally energetically unfavorable
in group-IV alloys[142].
Figure 3.1: Molecular precursors and the corresponding relaxed crystalline phases ofSi4C. Silicon is yellow, carbon is grey, and the terminal group is red. Ge4C and Sn4Chave similar structures.
Since all of the group-IV compunds studied here are metastable, a measure of their
64
stability is very important. We define the excess energy δE as the negative binding
energy with respect to the constituent elemental materials in the diamond structure.
For group-IV compounds containing Sn, comparing the energy with respect to gray
tin may not be the best way to study their stability, since gray tin is not stable at
room temperature. However, the energy difference between gray and white tin is small
compared to the excess energy. Another quantity of interest regarding the stability of
materials is the energetic barrier against the structural phase transitions. However, a
study of the transition barriers, which are expected to be relatively large for covalent
group-IV compounds, would involve extensive molecular dynamics simulations. Since
all of the structures investigated are locally stable with no indications of soft-mode
behavior, we do not examine the energy barriers in detail.
LatticeConstant(A)
Bond Length(A)
δE(eV/atom)
BulkModulus(GPa)
Si4C a=b=7.82c=4.97
Si–Si 2.30, 2.38Si–C 1.89
0.05 137(100)
Ge4C a=b=8.17c=5.22
Ge–Ge 2.40, 2.48Ge–C 2.00
0.28 108(76)
Sn4C a=b=9.19c=5.93
Sn–Sn 2.77, 2.86Sn–C 2.19
0.42 74(47)
Table 3.1: Structural properties of X4C in the local density approximation. The bulkmoduli in parentheses are for silicon, germanium, and gray tin respectively.
Table 3.1 shows the structural properties of X4C (X=Si, Ge, Sn) compounds. The
excess energy δE for Si4C is very small, 0.05 eV/atom, indicating the relative stability
of this particular compound. The lattice constant of Si4C differs from that of silicon by
about 8.1%. This result is very close to a previous study[142]. The very large lattice
mismatch between Si4C and silicon implies that pseudomorphic growth of Si4C on a
silicon substrates would quickly lead to cracks or other strain relieving defects. The
65
lattice of Ge4C is closer to that of silicon, with a 3.7% mismatch. Another interesting
observation (which will be discussed in more detail later) is that Ge4C becomes a
direct gap semiconductor upon lattice expansion to match that of silicon.
N
Γ
X
P
M
Σ
W
∆
Q
Figure 3.2: Irreducible Brillouin zone of the bct lattice.
Due to the symmetry breaking in these systems, there are two types of X–X (X=Si,
Ge, or Sn) bonds whose bond lengths differ by about 3%, with twice as many shorter
bonds as longer bonds. The weighted average lengths of the X–X bonds are very close
to those in the corresponding elemental bulk materials. For example, the Si–Si bond
length as calculated by LDA is 2.332 A, while the weighted average Si–Si bond length
in Si4C is 2.327 A. Si–C bonds, however, are slightly stretched compared to the bond
length in binary silicon carbide (SiC), 1.89 A vs. 1.87 A. The Ge–C bond length is
2.00 A, which lies between those in C(GeBr3)4 (2.05 A) and C(GeH3)4 (1.97 A)[147].
Although no stable GeC phase exists at ambient conditions, as a privilege for theorists,
66
we can still do the calculation for this idealized compound and we find that compared
to that of GeC, the Ge–C bond length in Ge4C is stretched by 2%, similar to Si–C.
The over-stretched Si–C and Ge–C bonds as compared to their binary counterparts
is expected in these strained systems[147]. The bulk moduli of these materials show
a systematic increase from that of the dominant elemental constituent: 37% for Si4C
and 57% for Sn4C (compared to gray tin).
Fig. 3.3 and 3.4 shows the LDA band structures of Si4C, Ge4C and Sn4C. (A
schematic plot of the irreducible Brillouin zone with high-symmetry k points is shown
in Fig. 3.2.) The LDA band structure of silicon, calculated on the same 10-atom unit
cell, is also plotted for comparison. (The specific points X, etc. mentioned below
refer to this 10-atom bct cell and not the standard 2-atom silicon Brillouin zone).
Although the band gaps of both Si4C and Ge4C remain indirect, the effects of the
substitutional carbon on the band structures of these systems are easily seen. First,
the lowest conduction band at the Γ point is pulled down significantly. The difference
between the direct energy gap at Γ and the minimum indirect gap of silicon is about
2.1 eV, this difference is reduced to about 1.5 eV for Si4C and 0.4 eV for Ge4C. We
believe this trend results from the competition between the overall volume contraction
and charge transfer. Contraction in volume due to the incorporation of carbon tends
to lift the energy of the conduction band at Γ (relatively), while charge transfer tends
to lower it. For Si4C, charge transfer seems dominant whereas for Ge4C, the two effects
are comparable. The LDA band gap of Sn4C is direct and slightly negative (-0.1 eV).
Since LDA calculations usually underestimate the band gap of semiconductors, we
expect the real band gap of Sn4C to be moderately positive.
As we briefly mentioned above, although Ge4C has a indirect energy gap, upon
volume expansion it acquires a direct gap. This high sensitivity arises from the
volume variation in the local potential energy of the conduction band[174]. Such a
lattice expansion could be achieved by deposition of a thin layer of Ge4C on a silicon
67
M Σ Γ ∆ X W P Γ
4Si C
N Q P
ener
gy (e
V)
−15
−10
−5
0
5
M Σ Γ ∆ X W P Γ
4Ge C
N Q P
wavevector K
ener
gy (e
V)
−15
−10
−5
0
5
Figure 3.3: LDA band structures of Si4C, Ge4C.
68
M Σ Γ ∆
4Sn C
X W P Γ N Q P
ener
gy (e
V)
−15
−10
−5
0
5
M Σ Γ ∆ X W P Γ
Si
N Q P
wavevector K
ener
gy (e
V)
−15
−10
−5
0
5
Figure 3.4: LDA band structures of Sn4C and silicon calculated on a 10-atom unitcell.
69
M Σ Γ ∆ X W P Γ N Q P
wavevector K
ener
gy (e
V)
−15
−10
−5
0
5
Figure 3.5: Band structure of Ge4C upon lattice expansion to match that of silicon.
substrate. Fig. 3.5 shows the band structure of Ge4C calculated with the silicon lattice
constant, which shows the direct band gap of Ge4C under expansion. Actually, less
than 2% lattice expansion makes the band gap of Ge4C direct.
3.4 Si2Sn2C and Ge3SnC
Although A4C (A=Si, Ge, Sn) group-IV alloys possess interesting structural and
electronic properties, they might not be good candidates for certain areas such as
optoelectronic applications. Si4C has an indirect gap and the lattice constant differs
from that of silicon by as much as 8%. Ge4C would have a direct band gap upon
lattice expansion to match that of silicon. However, a 3.7% lattice expansion is too
large for most practical applications. Sn4C is expected to have a moderate direct
70
bandgap but again does not lattice match silicon. One may ask, is it possible to
make direct bandgap group-IV alloys that lattice match silicon?
Although bulk silicon is an indirect bandgap semiconductor, it is nearly sur-
rounded by direct-bandgap elemental (unary) and compound semiconductors in the
periodic table. Moving downwards from Si, the unary group-IV materials acquire
larger cores containing d-electrons; these states affect the conduction band states
through orthogonality requirements and changes in the overall volume. On moving
from silicon to germanium to tin, the Γ2′ conduction band (as labelled in silicon) at
k = 0 drops in energy until, in grey tin, the material acquires a direct (and vanishing)
bandgap at k = 0. Moving horizontally on the periodic table, GaAs, a prototypical
direct-bandgap semiconductor, differs from Si or Ge in that the constituents have
different electronegativities. The resulting antisymmetric component in the crystal
potential flattens the bands and opens the bandgap. In addition, the Γ6 point of the
conduction band drops relative to the other points in the band until it becomes the
bottom of the conduction band in most of the well-known direct bandgap III-V and
II-VI semiconductors.
These well-known results suggest that perhaps one can combine these mechanisms:
larger cores and partial ionicity, within alloys containing only group-IV elements
wherein the difference in rows provides the contrast in electronegativity. Such a
system is likely to require the presence of tin to obtain a direct bandgap[175, 176, 177,
178]; the combination of tin (which is substantially larger than silicon) with carbon
(which is substantially smaller than silicon) not only affords the greatest contrast in
electronegativity (which is needed to open the zero-bandgap of elemental Sn), but also
opens the possibility of tuning the stoichiometry to obtain a lattice match to silicon.
Also, the countervailing effects of the larger Sn atoms might compensate the strain
introduced by carbon within a Si/Ge lattice and therefore enhance the solubility of
carbon within a silicon and germanium matrix.
71
Figure 3.6: Molecular precursors and the corresponding relaxed crystalline phasesof Si2Sn2C (top) and Ge3SnC (bottom). Silicon is yellow, carbon is grey, tin ismagenta, germanium is green and the terminal group is red. Ge4C and Sn4C havesimilar structures.
72
To be practical, such an alloy must be accessible to a well-defined synthetic tech-
nique. Ternary group-IV compounds containing Sn, for example, SiySn4−yC and
GeySn4−yC, could be accessible to the aforementioned UHV CVD synthesis techniques
by replacing silicon or germanium atoms with tin in the corresponding molecular
precursors. One interesting previous study[154] in computational design of direct-
bandgap silicon lattice matching materials, in constrast, imposes a formidable syn-
thetic challenge (the material requires the sequential deposition of 8 distinct non-
interdiffusing monolayers of As/Zn/As/Si/As/Zn/P/Si).
Of particular relevance to the current study are several additional synthetic de-
velopments. First, the Sn–D bond has recently been shown to be stable for long
periods (at 0o C) in molecules which incorporate Sn–D3 moieties (the Sn–H3 moiety,
in contrast, is known to be unstable)[179]. Second, techniques have been developed
to synthesize tetrahedral precursors where a central carbon atom is attached to four
tins[180]. Finally, the synthetic yields of, e.g., CSi4H12 and CGe4H12 have improved
significantly since their discovery, to the range where multi-gram quantities can be
regularly prepared[181, 182]. Keeping these important advances in mind, we focus
on novel group-IV alloys which can be built from C(Si,Ge,Sn)4X12 precursors. In
particular, we demonstrate that two alloys, Si2Sn2C and Ge3SnC, have low-energy al-
lotropes which lattice match silicon to better than 0.5-1% and have direct bandgaps.
To obtain the energy gaps more accuratly , we also calculate the quasi-particle energy
gap using the GW approximation[36].
Fig. 3.6 shows the proposed molecular precursors for Si2Sn2C and Ge3SnC, namely,
C(SiX3)2(SnX3)2 and C(GeX3)3(SnX3) (where X is a terminal group), and the cor-
responding ordered crystalline phases. These particular crystal structures are the
lowest-energy (and highest-symmetry) among several allotropes with small unit cells
built from these precursor molecules. The structures shown are fully relaxed locally
and all phonon modes are stable, indicating local structural stability. The absence
73
of strong C-X (e.g. C-H) bonds allows a low growth temperature which is crucial for
synthesizing metastable materials, particularly with tetrahedrally coordinated Sn.
The Si2Sn2C material has a body-centered tetragonal (bct) lattice with 10 atoms
per unit cell like that of the A4C structure. This relatively high-symmetry structure
can minimize the anisotropic strains associated with deposition on a low-index silicon
surface. In this structure, there are four Sn–Sn, four Sn–Si and four Si–Si bonds in a
unit cell, achieving the highest possible rotational symmetry. The fully relaxed den-
sity functional results reveal that the volume per atom in Si2Sn2C is only 0.6% less
than that of bulk silicon (we compare the theoretical LDA volumes to minimize sys-
tematic errors). This is only a slight deviation from Vegard’s law, which predicts that
stoichiometries near Si2.22Sn1.78C lattice-match silicon. The body centered tetragonal
structure of Si2Sn2C provides several lattice planes which can be well-matched to
low-index planes of bulk silicon. For example, the (110) and (111) planes of silicon
match to Si2Sn2C better than 0.5% in all directions.
The 3:1 ratio of noncarbon atomic constituents in Ge3SnC yields a slightly lower-
symmetry structure which is 1.7% lower in volume than bulk silicon. Due to the lower
crystal symmetry, the lattice is slightly distorted from bct. However, the deviation
from bct is very small, about ±1%. Therefore, we use the notation of the special (highsymmetry) k points of bct structure to label the band structure of Ge3SnC. The lower
symmetry of Ge3SnC results in a slightly worse lattice matching to silicon as com-
pared to Si2Sn2C, about 1% in linear dimension (with a combination of compression
and dilation along different axes). Mismatches of this order should yield a critical
thickness of hundreds to thousands of Angstroms[183]. In addition, a synthesis based
on these molecular precursors could allow the addition of a small concentration of a
second precursor that helps to precisely tune the lattice match. Another important
issue regarding lattice compatability is the thermal expansion coefficient. Although
we did not study the thermal expansion compatibility between these two compounds
74
LatticeConstant(A)
Bond Length(A)
δE(eV/atom)
BulkModulus(GPa)
Si2Sn2C a=b=8.43c=5.46
Si–Si 2.36Si–Sn 2.62Sn–Sn 2.72Si–C 1.88Sn–C 2.22
0.30 104
Ge3SnC a=8.50b=8.34c=5.41β=89.6o
Ge–Ge 2.43, 2.49Ge–C 2.00Ge–Sn 2.58, 2.66Sn–C 2.20
0.32 80
Silicon a=b=8.51c=5.38
Si–Si 2.33 0.0 100
Table 3.2: Structural properties of Si2Sn2C and Ge3SnC. All cited values are LDAresults. The results for silicon calculated with a 10-atom supercell are also listed forcomparison.
and silicon, we would expect the difference in thermal expansion to be moderate,
considering that, at room temperature, the thermal expansion coefficient of diamond
(1×10−6/K) is smaller than that of silicon (about 2.5×10−6/K) while both germa-
nium (about 5×10−6/K) and α-tin (about 6×10−6/K) have larger thermal expansion
coefficients.
Table 3.2 shows the structural properties of Si2Sn2C and Ge3SnC. The A–C (A=Si,
Ge, or Sn) bond lengths change little as compared to those in A4C alloys, indicating
the robustness and stiffness of these bonds. However, both Si–Si and Ge–Ge bonds are
longer than those in Si4C and Ge4C, due to the presence of Sn. The bulk modulus of
Si2Sn2C is comparable to that of silicon while the bulk modulus of Ge3SnC is similar
to that of germanium. The excess energies of both Si2Sn2C and Ge3SnC are moderate
and comparable to that of Ge4C, which has been successfully synthesized, indicating
that these two compounds might well be accessible to current CVD techniques so long
75
as the molecular cores of the precursors are stable under the synthesis conditions.
Fig. 3.7. compares the pseudopotential density functional bandstructures of Si2Sn2C
and Ge3SnC to the corresponding Si bandstructure within the same (10-atom) unit
cell. Si2Sn2C has a direct bandgap in an unusual position, the X point. The con-
duction band minimum at the X point is rather close to the folded X point in Si.
Symmetry breaking due to the alloying splits a degeneracy at the X point. In the
uppermost valence band, this splitting raises the higher-energy state substantially
until it passes the Γ point state and thereby becomes the new valence band maxi-
mum. This change in the valence-band maximum compared to Si is facilitated by
the smaller bandwidths in this more ionic semiconductor alloy. Meanwhile, the con-
duction band at Γ follows a well-known pattern: the introduction of Sn pulls down
the lowest conduction band at Γ and produces a local (but in this case not global)
conduction band minimum at Γ.
Although the characteristics of the conduction band of Si2Sn2C are not unex-
pected, having a valence band maximum at X, rarely seen in common semiconductors,
is an intriguing feature. To understand this somewhat abnormal band structure, i.e.,
a direct energy gap at a point other than Γ, we analyze the charge density at the top
valence band for the X and Γ points. Fig. 3.8 shows the three dimensional charge
density isosurfaces for the top valence band at Γ and X for Si2Sn2C. The charge den-
sity at Γ preferentially concentrates around Si–Si bonds, whereas the X-point charge
density is more localized around Sn–Sn bonds. The greater Sn character then ac-
counts for the higher energy of X state. Measuring the bond charge contained within
fixed-width cylinders, we obtain (QΓ(Si–Si)/QΓ(Sn–Sn)):(QX(Si–Si)/QX(Sn–Sn)) ∼4:1. (We define the bond charge as the total charge enclosed by a cylinder connect-
ing the two atoms. Although the absolute value of the bond charge scales with the
volume of the cylinder, the ratio of the bond charges is insensitive to the cylinder
diameter.) Since the valence atomic orbital energies of Sn are higher than those of
76
M Σ Γ ∆ X W P Γ N Q P−15
−10
−5
0
ener
gy (
eV)
M Σ Γ ∆ X W P Γ N Q P
wavevector K
−15
−10
−5
0
ener
gy (
eV)
Figure 3.7: Band structures (in blue) of Si2Sn2C (top) and Ge3SnC (bottom). Theionicity opens a bandgap deep in the valence band. Symmetry breaking in Si2Sn2Cremoves the degeneracy at the X point, pushing one band up significantly so thatit becomes the conduction band maximum and produces in a direct band gap at X(red arrow). In the case of Ge3SnC, the conduction band minimum occurs at Γ,which produces a direct gap at Γ (red arrow). The band structure of silicon (orange)calculated in the same unit cell is also plotted for comparison.
77
Γ
X
Figure 3.8: Charge density isosurfaces for the top of the valence band at Γ (top) andX (bottom) in Si2Sn2C. Electrons are localized around Si–Si bonds (in yellow) for theΓ state, whereas for the X state they are localized around Sn–Sn bonds (in magenta).
78
Si, the greater Sn character of the X point accounts for the unusual valence band
structure of Si2Sn2C. The fundamental LDA band gap of Si2Sn2C at the X point is
0.59 eV. The GW correction increases the band gap to 0.9 eV.
The mechanisms which produce a direct bandgap in Ge3SnC have a more familiar
interpretation. Due to the increased fraction of the lower-row elements Ge and Sn,
the drop in the conduction band at Γ is more pronounced than that in Si2Sn2C. This
Γ-point state becomes the new conduction band minimum, directly above the valence
band maximum, which remains at Γ. Therefore, the direct band gap of Ge3SnC can
be understood as due to a large difference in electronegativity between carbon and
germanium/tin as well as the large ionic core which tends to pull the conduction
band at Γ down more rapidly than at the zone edge. The LDA band gap at Γ is 0.30
eV, increasing to 0.71 eV under the GW correction. The fundamental band gap at
Γ is very sensitive to external pressure, dropping by 0.1 eV for a fractional change
in volume of δV/V = 0.01. This high sensitivity arises from the volume variation
in the local potential energy of the conduction band[174] and provides a potentially
important mechanism for tuning the size of the direct bandgap. For consistency, the
band structures are calculated at the experimental volume per atom of the posited
silicon substrate. Since the volume differences between these materials and bulk
silicon are very small, the band structures are very similar to those calculated at
their equilibrium volumes. Si2Sn2C shows no significant changes in bandstructure
under such a volume change. The high volume sensitivity of Ge3SnC means that
the small expansion to match the experimental volume per atom of silicon actually
drives the bandgap direct. (At the relaxed bulk Ge3SnC volume the bandgap is just
slightly indirect). For Ge3SnC the optical matrix elements at the direct transition are
comparable to those in GaAs. For Si2Sn2C, the optical matrix elements are somewhat
smaller but the transition is still clearly dipole-allowed.
In summary, we have demonstrated in this section how materials design can yield
79
multiple candidate materials which simultaneously possess three desirable materials
properties: a very close lattice match to silicon, a direct bandgap in the 0.7-1.0
electron volt range, and accessibility to well-defined synthetic strategies that can
potentially produce bulk quantities of material.
3.5 Si6C2 and Ge6C2
Figure 3.9: Proposed molecular precursor for synthesizing Si6C2 and the correspond-ing relaxed crystalline structure. The molecular precursor and crystal structure ofGe6C2 are similar to those shown.
In previous sections, we studied group-IV compounds which are isostructural
with already synthesized compounds that use the propotype molecular precursor
80
C(AH3)4, where A = (Si, Ge, Sn). These systems achieve the highest carbon concen-
tration (20 at. %) possible without creating C–C bonds or second nearest neighbor
carbon pairs. An interesting question remains, however. Using specific molecular pre-
cursors, can group-IV compounds with higher carbon concentrations be synthesized?
Such systems would inevitably contain either C–C bonds or second nearest neighbor
carbon pairs. Although C–C bonds are thermodynamically unfavorable in group-IV
alloys, if C–C bonds are built into the molecular cores of precursor molecules and are
stable under film deposition, then we could expect C–C bonds to persist in the result-
ing alloys. We propose here two new molecular precursors, C2(SiH3)6 and C2(GeH3)6,
for synthesizing group-IV compunds with C–C bonds and 25 at. % carbon.
Fig. 3.9 shows the proposed molecular precursor and sample crystalline structure,
fully relaxed from LDA calculations of Si6C2. Both Si6C2 and Ge6C2 have a mono-
clinic lattice which is a slight distortion from bct. For Si6C2, the distortion is very
small, less than 0.1% while for Ge6C2 it is about 1.0%. To a good approximation,
we can still use the more familiar k point notation of the bct structure for these two
compounds.
LatticeConstant(A)
Bond Length(A)
δE(eV/atom)
BulkModulus(GPa)
Si6C2 a=b≈4.91c≈9.46
Si–Si 2.29, 2.322.27, 2.48
C–C 1.55Si–C 1.90, 1.99
0.20 146
Ge6C2 a=b≈5.18c≈9.84
Ge–Ge 2.40, 2.442.37, 2.58
C–C 1.50Ge–C 2.01, 2.12
0.40 111
Table 3.3: Structural properties of Si6C2 and Ge6C2. The lattice notation of the bctstructure is used since both structures are only slightly distorted from bct.
81
Table 3.3 shows the structural properties of these two group-IV compounds. Con-
sidering that the fraction of C–C bonds is large, 1/16, the excess energy of these two
compound is lower than expected. For Si6C2, the excess energy is 0.2 eV/atom, which
is lower than that of Ge4C. One curious result is that the C–C bond in Si6C2 is longer
than the normal bond length in diamond: 1.55 A vs. 1.53 A (both are LDA values),
whereas for Ge6C2, the C–C bond is shorter than the normal one, only 1.52 A. The
strong Si–C bonds, which effectively pull the two carbons apart to minimize the strain
energy, might explain the slightly longer C–C bond length in Si6C2. For Ge6C2, the
relatively weak Ge–C bonds, coupled with stronger C–C bonds due to charge transfer
from Ge atoms to C atoms, imply that shorter C–C bonds are the optimal strain-relief
configuration. Due to the lowering in symmetry, there are two types of A–C (A=Si or
Ge) bonds. The weighted average Si–C and Ge–C bond lengths are 1.93 A and 2.05 A
repectively, substantial longer than the normal Si–C and Ge–C bonds. The existence
of C–C bonds defines a particular direction along which the bond length of both Si–Si
and Ge–Ge bonds vary significantly. For example, the length of Si–Si bonds along
the direction defined by the C–C dimer varies widely from 2.27 to 2.48 A. (However,
the average Si–Si bond length does not deviate significantly from the bulk value.)
The Si–Si bonds more perpendicular to the C–C dimers are closer to the bulk value,
varying from 2.29 to 2.32 A. The presence of over-stretched Si–Si bonds along the
C–C bond direction are consistent with substantial strain on the C–C bonds. Similar
results are seen for Ge6C2.
Fig. 3.10 shows the LDA band structures of Si6C2 and Ge6C2. Both alloys are
metallic within the local density approximation. Strain-induced reduction of the
band gap has been reported before[141]. However, we are not convinced that Si6C2
is indeed metallic since LDA usually underestimates the band gap and in some cases,
semiconductors become metallic in LDA calculations. Since the band overlap for
Si6C2 is small, less than 0.5 eV, Si6C2 might have a small direct gap if quasiparticle
82
M Λ Γ ∆ X W P Q N Γ
ener
gy (e
V)
−15
−10
−5
0
5
M Λ Γ ∆ X W P Q N Γ
−15
wavevector K
ener
gy (e
V)
−10
−5
0
5
Figure 3.10: LDA band structure of Si6C2 (top) and Ge6C2( bottom).
83
energy (eV)
DO
S (
1/eV
)
−16 −12 −8 −4 0 40
1
2
3
4
energy (eV)
DO
S (
1/eV
)
−16 −12 −8 −4 0 40
1
2
3
4
Figure 3.11: Density of states of Si6C2 (top) and Ge6C2( bottom).
84
calculations are carried out. For Ge6C2, the band overlap is large and we do not
expect quasiparticle corrections to fully open the gap. Notice that although the
conduction and valence bands overlap at Γ, we still can see a “quasi-direct” gap,
which is evidence of increasing ionicity, i.e., charge transfer between Si (Ge) and C
in these group-IV compounds. Another consequence of the large ionicity, possibly
along with symmetry breaking, is that the lowest valence band is completely split
off from the rest of valence bands and is very flat, indicating that this band is very
localized (on carbon). Density-of-states plots of Si6C2 and Ge6C2 are shown in Fig.
3.11. The density of states at the Fermi level is very low. However it increases
rapidly below or above the Fermi level. A small amount of doping might change the
electronic properties of these two materials greatly. The anisotropic bonding network
of both materials might produce an anisotropic phonon dispersion, while the covalent
bonding and relatively strong volume sensitivity of the bandstructure in this class
of materials could yield relatively large electron-phonon matrix elements. The large
phonon frequencies of C-associated modes suggests the possibility for a moderately
large superconducting transition temperature under doping heavy enough to climb
out of the pseudogap.
3.6 Conclusion
We have carried out pseudopotential DFT studies on the electronic and structural
properties of novel group-IV alloys that either have been synthesized or are designed
to be synthesizable by UHV-CVD techniques using precursor molecules which build
in the desired bonding. All of the alloys studied are metastable with moderate ex-
cess energies. While the bandgaps of Si4C and Ge4C remain indirect, the effect of
substitional carbon on the band structure is prominent. For Ge4C, a 2% lattice ex-
pansion brings the bandgap direct. Sn4C has a direct and slightly negative LDA gap.
More importantly, we have demonstrated two silicon lattice matching, direct-bandgap
85
group-IV semiconductors, namely, Si2Sn2C and Ge3SnC. Finally, we have proposed a
new class of precursor molecules, X6C2 (X=Si, Ge or Sn), which incorporates 25 at.%
carbon. These high carbon concentration group-IV compounds exhibit interesting
structural and electronic properties which might stimulate experimental search for
this new class of group-IV alloys. The Kouvetakis team at Arizona State University
has recently made some progress in synthesizing X6C2 alloys[184].
The idea of using CVD precursors which build in the desired chemical bonding
to synthesize metastable alloys can be extended to systems beyond group-IV alloys.
This technique opens great possibilities of synthesizing materials that are inaccessible
to conventional thermal equilibrium synthesis techniques. With the help of modern
powerful computers and well-developed ab initio computational techniques, one might
be able to design other specific precursors for novel materials with desired properties.
86
Chapter 4
O(N2) real-space electronicstructure methods
4.1 Introduction
The past few decades have witnessed dramatic progress in our ability to explain
and predict various properties of materials knowing only the identities of their con-
stituent atoms. This progress has been driven mainly by two important factors,
namely, the availability of ever-increasing computer speed and advances in theoret-
ical/computational methods, which have made computer simulation another branch
of scientific research, in addition to the traditional theoretical and experimental ap-
proaches to doing science.
DFT-based ab initio methods, such as pseudopotential plane-wave methods[37] as
we have discussed in Chapter 1, have been very popular and successfully applied to
calculate the electronic, optical, and structural properties of a vast array of materials.
Unfortunately, these traditional electronic structure methods also suffer from some
significant drawbacks which are intrinsic to the formalism, for example, the cubic
scaling of computational time with respect to the number of atoms (or electrons)
in the system, the so-called O(N3) problem. Due to this unfavorable scaling, the
number of atoms one can study using these methods is limited to a few hundred even
87
with the most powerful computers available today. Consequently, an alternative, the
so-called real-space approach, has attracted much research interest recently. (The
plane-wave formalism is also called momentum-space formalism for obvious reasons.)
The real-space methods have some advantages over the momentum-space methods
which we shall discuss in more detail later.
The reason that traditional electronic structure methods, such as pseudopotential
plane-wave methods, must scale at least as O(N3) is the following. The electronic
wavefunctions generally extend over the entire volume of the system, computing an in-
dividual wavefunction takes at least O(N) operations, the number of electrons (hence
the number of wavefunctions needed) grows linearly as the number of atoms, and
finally, most iterative diagonalization algorithms require explicit reorthogonalization
between eigenfunctions which brings in an additional multiplicative O(N) scaling. It
can easily be seen that, if the explicit requirement of wavefunction orthogonalization
is eliminated, an O(N2) scaling can be achieved. Developing such an O(N2) electronic
method is the main subject of this chapter.
Although this chapter involves a considerable amount of mathematical formulae
and numerical algorithms, we shall focus more on the physics, not the completeness
and rigor of the mathematical formalism.
4.1.1 Momentum-space versus real-space formalism
Plane-waves are natural choices for describing the electronic wavefunctions in peri-
odic systems. Plane-wave expansion, when coupled with pseudopotentials to eliminate
the need to explicitly consider the highly localized core electrons (see Chapter 1), can
be very advantageous: (i) the plane-wave basis set is well structured, which greatly
simplifies code development, (ii) the accuracy of the calculation can be improved
systematically by a single parameter, i.e., the highest kinetic energy of the plane-
waves included in the wavefunction expansion, (iii) plane-waves do not depend on the
88
atomic positions and therefore simplify the calculation of forces on atoms and atomic
geometry, (iv) the kinetic energy operator is strictly diagonal in momentum space,
whereas the local pseudopotential is diagonal in real space. Fast Fourier transforms
(FFT) can be carried out between the two spaces.
However, some of the advantages of momentum space formalism may become dis-
advantages in certain circumstances. For example, the fact that plane-waves do not
depend on the atomic positions is problematic when dealing with systems such as
molecules, clusters and surfaces, in which a large number of plane-waves is needed
to decribe the nearly vacuum space. For very localized systems such as the first row
elements or transition metals containing partially filled d or f shells, a plane-wave ex-
pansion becomes very inefficient even if the system contains only one such atom. Also,
the FFT algorithm cannot be efficiently parallelized on modern distributed memory
computers. Furthermore, the computational cost of the traditional momentum-space
methods scales at least as O(N3), where N is the number of atoms in the system, as
we discussed at the beginning of this chapter.
Recently, there has been considerable research interest in real-space methods [185-
205], in which the wavefunctions or other quantities of interest, for example, charge
densities, potentials, etc., are described directly on real space grid(s). In fact, tra-
ditional quantum chemistry calculations are done solely in real space using either
analytical bases (for example, Gaussian-type or Slater-type) or numerical localized
atomic orbitals. These methods usually cannot take advantage of multigrid conver-
gence acceleration[206, 207], so we will not discuss them in detail here. Instead, we
will focus on grid-based real-space methods. Grid-based calculations usually require
large amounts of computer memory (of the order of 10 GB for a system containing a
few hundred atoms) and thus could not be done until recently.
There are a number of reasons for carrying out calculations directly in real space
grid for isolated systems and in some cases, even for periodic systems. First, unlike
89
plane-waves, which extend over the entire system with uniform resolution, the reso-
lution of wavefunctions in real-space can be adjusted to reflect the detailed atomic
arrangement of the system, resulting in an adaptive description of wavefunctions.
Second, the Hamiltonian matrix in real space is usually very sparse. This signifi-
cantly reduces the computation cost for iterative diagonalization algorithms. In con-
trast, the potential matrix V (G,G′) = V (G−G′) = V H(G−G′) + V xc(G−G′) +
V ps(G−G′) in Eqn. (1.61) is a full dense matrix. (Note that we need not to store
the entire potential matrix since V depends only on (G−G′).) Third, parallelization
of real-space based methods is usually more efficient than that of momentum-space
based methods since one can assign specific regions of space to particular proces-
sors and there are no communciation-intensive FFT operations involved. Fourth,
multi-scale[206, 207] (also known as multi-level or multi-grid) convergence accelera-
tion techniques can be readily applied to real space calculations. Finally and very
importantly, a real-space formalism is essential for implementing the so-called O(N)
(or linear scaling) algorithms[208, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218,
219, 220, 221, 222, 195], which promise to achieve O(N) scaling with respect to the
number of atoms in the system, as opposed to O(N3) for the traditional techniques.
4.1.2 O(N) methods
The pioneer work of Yang[208],Galli, et al.[209], Li, et al.[210] and Daw[211] in the
early 90’s showed that, by examing the localization properties of quantum systems,
electronic structure calculations are possible with computational cost that scales lin-
early with respect to the number of atoms in the system. Common features of all
proposed O(N) algoritms are (i) the use of a localization approximation, and (ii)
no computation of the individual wavefunctions and eigenenergies. The key idea of
O(N) methods is to evaluate the total energy of the system directly by exploiting the
localization properties of either a Wannier-like orbital basis[209, 212, 213, 214, 195]
90
(the wavefunction approach) or the density matrix ρ(r, r′)[210, 211, 215, 216, 217,
218, 219, 220] (the density matrix approach). There are other O(N) approaches
as well, for example, the projection methods[221, 222] and the divide-and-conquer
methods[208, 224].
It is remarkable that, although the one-particle wavefunctions, which are com-
monly used for describing many particle systems, usually extend over the entire sys-
tem, the physical properties in a given region of a material are affected mainly by its
local environment. This insensitivity of local properties to the long-distance pertur-
bation is called, in Kohn’s word[225], the “nearsightedness” of equilibrium quantum
systems. Actually, locality in quantum systems has been appreciated for a long time
in traditional chemistry. (However, tranditional electronic structure calculations do
not fully exploit this interesting property.) One obvious such example is the cova-
lent bonding in molecules and covalence solids. The bonding nature of these system
is mostly determined by the local atomic arrangement and is described in terms of
electron pairs shared by nearby atoms. The famous Valence Shell Electron Pair Re-
pulsion (VSEPR) theory of Gillespie and Nyholm[226, 227], which is widely used
for determing the molecular geometry, is based on such observations. Surprisingly,
although the conventional bonding concept is no longer valid in metallic systems,
locality still exists. This is supported by the well-known fact that the charge den-
sity in a metal can be reasonably described by the superposition of its consituent
atomic charge densities. The “nearsightedness” of quantum systems is best mani-
fested in the exponentially decaying property of the one-particle charge density ma-
trix ρ(r, r′): ρ(r, r′) ∼ exp(−γ|r− r′|), where γ > 0. It was shown recently that
γ ∝ Eg for semiconductors and γ ∝ T for metals, where Eg is the bandgap and T is
the temperature[228]. The O(N) method of Li, Nunes and Vanderbilt (LNV) is based
on such “nearsightedness” of the density matrix, which is briefly discussed below.
Recall that the traditional way to solve for the band energy Ebs is to diagonalize
91
the Hamiltonian:
HΨi = εiΨi, (4.1)
where Ψi and εi are eigenfunctions and the corresponding eigenvalues. The band
energy is the sum of all occupied eigenvalues:
Ebs =∑i∈occ
εi. (4.2)
As we have discussed in Section 4.1, the above process requires O(N3) operations.
An alternative is to solve for the ground state density-matrix and formulate the band
energy in terms of the density matrix, i.e.,
Ebs = minρtr[ρH ], (4.3)
where ρ satisfies two constraints,
tr[ρ] = Ne, and ρ2 = ρ. (4.4)
The first constraint can be eliminated by working with the grand potential Ω
Ω = Ebs − µNe, (4.5)
where µ is the chemical potential. The second constraint, the idempotency con-
straint, can also be removed by introducing the so-called purification transformation
of Mcweeny[229, 210]: If ρ is a trial density-matrix which is nearly idempotent, then
ρ = 3ρ2 − 2ρ3 (4.6)
is a purified density-matrix which is closer to idempotent than the original ρ.
By taking advantage of the fact that the density-matrix is local in real space, i.e.,
ρ(r, r′) → 0 as |r− r′| → ∞, LNV developed a variational scheme which minimizes
the grand potential Ω in Eqn. (4.5) with the density-matrix ρ replaced by the purified
version ρ. Since the number of non-zero elements of ρ is O(N), the minimization
92
process should take nearly O(N) operations to be completed. The LNV algorithm
is best formulated within the tight-binding (TB) or other empirical models[210, 215,
214, 219], although LNV-like DFT-based O(N) methods using localized Gaussian
orbitals has also been reported[220].
4.1.3 O(N2) electronic structure methods: an intermediatestep
The O(N) methods represent a group of aggressive approaches to ultimately reduce
the computational effort. Whereas the TB O(N) methods[221, 214, 216, 217, 218]
are relatively well developed and widely used, most of the proposed DFT-based
O(N) methods suffer some difficulties[223]. We propose here an intermediate step
towards DFT-based large-scale ab initio calculations by developing an O(N2) real-
space electronic structure method without introducing any localization assumptions.
One important feature of our method is that all occupied electronic wavefunctions and
eigenvalues are computed explicitly, whereas most O(N) methods calculate only the
total energy of the system. Our scheme is based on the finite-element method with
multigrid acceleration, which is one of several proposed grid-based real-space elec-
tronic structure approaches. The O(N2) scaling is achieved by avoiding the explicit
reorthogonalization between eigenfunctions through use of a multigrid algorithm.
Grid-based real-space electronic approaches include finite-difference (FD)[185, 186,
187, 188, 189, 190, 191, 192, 193, 194, 195], finite-element (FE)[196, 197, 198, 199, 200,
201] and wavelet[202, 203, 204, 205] methods. Although the FE method has a long and
successful history in engineering applications, there are only a few quantum mechani-
cal calculations using FE methods. The reasons for this lack of interest in FE methods
among the physics community deserves further discussion, which is beyond the scope
of this thesis. Electronic structure calculations for atoms and diatomic molecules
using FE methods started in the 70’s and 80’s[230, 231, 232, 233, 234, 235, 236].
93
These early calculations used symmetry properties to reduce the dimensionality of
the system. White, et al. carried out the first full 3D FE calculations for atoms and
molecules[197].
Not surprisingly, there have been even fewer solid-state calculations based on FE
methods. Hermansson and Yevick[196] first applied the FE method to 3D solid-state
electronic structure calculations but reached a negative conclusion that “the standard
FEM formalism is far less efficient than standard (plane-wave) techniques”. Recently,
there has been revived interest in FE electronic structure calculations in solids start-
ing from the work of Tsuchida and Tsukada[198, 199, 200] and a recent paper of Pask,
et al.[201]. Unfortunately, none of the aforementioned work implemented the multi-
grid technique[206, 207], which is critical for improving the convergence rate. Here
we present a self-consistent (SC) FE electronic structure calculation with multigrid
acceleration. Furthermore, our method scales nearly as O(N2). In the following, we
give a brief introduction to the FE method.
4.2 The FE method
The mathematical treatment of the FE method is based on the variational for-
mulation of differential equations. Consider the Poisson equation in some domain Ω
with, for example, Dirichlet boundary conditions:
−u(x) = f(x) in Ω,
u(x) = 0 on ∂Ω.(4.7)
Before proceeding with our discussion, we need to introduce some mathematical no-
tations. We use L2(Ω) to denote the space of square-integrable functions over domain
Ω, Hm(Ω) is the Sobolev space of L2(Ω) with square-integrable derivatives up to
order m in Ω and Hm0 (Ω) the subspace of H
m(Ω) of functions with zero boundary
conditions[237]. The above boundary value problem is equivalent to the variational
94
formulation: find u ∈ H10 (Ω) such that
a(u, v) = (f, v) ∀v ∈ H10 (Ω), (4.8)
where
a(u, v) =∫Ω∇u · ∇v, and (f, v) =
∫Ωfv. (4.9)
Actually, the solution of Eqn. (4.8) is the solution of the following variational problem:
find u ∈ H10 (Ω) such that
J(u) =1
2a(u, u)− (f, u) (4.10)
attains its minimum. For the Kohn-Sham equation (1.24), we have the following
variational formulation: find ψi ∈ H1(Ω) such that
1
2
∫Ω∇φ∗ · ∇ψi +
∫Ωφ∗[V eff − εi]ψi = 0, ∀φ ∈ H1(Ω), (4.11)
where V eff = V H + V xc + V ext is the effective potential. Here H1 denotes the space
of functions that satisfies certain boundary conditions. For periodic systems, we may
write ψi,k = exp(ik · r)ui,k and the corresponding variational formulation for ui,k
becomes[201]
1
2
∫Ω∇v∗ ·∇ui,k+
∫Ωv∗[−2ik·∇ui,k+(V eff−k2−εi,k)ui,k] = 0, ∀v ∈ H1(Ω). (4.12)
The function space H1 include all smooth functions satisfying periodic boundary
conditions.
4.2.1 The Ritz-Galerkin Method
Solving the variational form of partial diferential equations, for example, Eqn.
(4.8), (4.11), or (4.12) exactly in the functional space H1 is a formidable challenge
and usually not necessary. Instead, we solve them in some approximated finite-
dimensional subspace Sh, where h measures the discretization parameter. The solu-
tions should converge to the exact ones as h → 0. This is called the Ritz-Galerkin
95
method. The problem is therefore reduced to, in the case of the Poisson equation
(4.8): find uh ∈ Sh such that
a(uh, vh) = (f, vh), ∀vh ∈ Sh(Ω). (4.13)
Suppose φ1, φ2, · · · , φn is a basis of Sh, then Eqn. (4.13) is equivalent to: finduh ∈ Sh such that
a(uh, φi) = (f, φi), i = 1, 2, · · · , n. (4.14)
Furthermore, we can expand uh as a superposition of φi, namely, uh = ∑ciφi, and
reach the following system of algebraic equations
AC = b, (4.15)
where
Aij = a(φi, φj), bj = (f, φj), (4.16)
and C is the column vector of ci, i.e., (c1, c2, · · · , cn)T . The procedure thus far seemsvery familiar. The idea of FE analysis is to approximate the subspace Sh by a piece-
wise polynomial basis with respect to a partition of a domain, the so-called FE space.
The partition of a domain is called the FE triangulation for historical reasons.
4.2.2 FE triangulation and nodal basis functions
In the FE analysis, the domain of the system is divided into subdomains called
“elements”. (The term “element” also refers to the FE space associated with a partic-
ular triangulation, for example, quadratic triangular element, bilinear quadrilateral
element, etc..) For problems in 2D, it is natural to partition the domain into trian-
gles (2-simplices, see Fig. 4.1), as shown in Fig. 4.2. This is probably the reason that
the FE partition is called a triangulation. FE discretization is far more flexible than
FD discretization. For 3D problems, a partition into either tetrahedrons or hexahe-
drons (parallelpipeds, cubes, etc.) are commonly used. Fig. 4.3 shows a tetrahedron
partition example of a cubic domain.
96
(a) (b) (c)
Figure 4.1: n-simplices: 1, 2, and 3-simplices.
(a) (b)
Figure 4.2: Finite-element triangulation examples in 2D: (a) uniform triangulationand (b) adaptive (nonuniform) triangulation.
Figure 4.3: A 3D triangulation.
97
φi
iN
Figure 4.4: A linear nodal basis function. The function is defined corresponding tonodal point Ni.
After partitioning the domain, we need to decide on the polynomial order of the
basis functions. These functions are called nodal basis functions since each is as-
sociated with an individual node. The simplest finite elements are linear elements,
the nodal basis of which consists of linear functions. However, linear elements are
usually not very accurate. In our calculations, we use a quadric tetrahedral element.
Implemention of higher order elements is also possible. Fig. 4.4 shows a linear nodal
basis function φi defined with respect to node i. The nodal basis functions have the
following properties, regardless of their order: (i) φi(xj) = δij, (ii) φi is strictly local-
ized (locally supported), (iii) the nodal basis functions forms a basis φ1, φ2, · · · , φn,over which any function u ∈ Sh can be expanded: u(x) =
∑nj=1 ujφj(x). It can be
easily seen from (i) that the coefficients of expansion are simply the corresponding
functional values, namely, uj = u(xj).
98
4.3 FE discretization of the Kohn-Sham equations
Having given a brief introduction to the FE method, we now return to our orig-
inal goal: to solve the Kohn-Sham equations in real-space with FE approximations.
We will assume periodic boundary conditions. Extension to non-periodic boundary
condition is straightforward. For simplicity, we will also assume k = 0 in Eqn (4.12),
which is then reduced to Eqn. (4.11).
The FE discretization of Eqn (4.11) results in a generalized eigenvalue problem:
HΨi = εiSΨi, (4.17)
where
Hij =∫Ω[1
2∇φi · ∇φj + V effφiφj] (4.18)
is the Hamiltonian matrix and
Sij =∫Ωφiφj (4.19)
is the overlap matrix. Note that the wavefunctions are written as superpositions on
the FE basis, i.e., Ψi(r) =∑
Cijφj(r), and again, φi are the FE basis functions.
Therefore, our task is to compute the relevant matrix elements and solve the general-
ized eigenvalue problem (4.17). Both the kinetic energy matrix Tij =12
∫Ω∇φi · ∇φj
and the overlap matrix Sij can be calculated analytically; we will not discuss this in
detail here. The construction of the potential matrix Vij =∫Ω V effφiφj, however, is
not trivial.
4.4 Construction of the ab initio potential matrices
In the psuedopotential formalism, the effective potential consists of four terms,
the exchange-correlation potential V xc, the Hatree potential V H , the local ionic pseu-
dopotential V ps,loc and the nonlocal pseudopotential, which is usually written in the
Kleinman-Bylander (KB) form[46], V psKB (see Section 1.3.3).
99
First, the charge density ρ is calculated in the usual way:
ρ(r) = 2∑i
Ψi(r)∗Ψ(r), (4.20)
where i sums over all occupied orbitals. (Initially, the charge density is taken as the
superposition of atomic charge densities. We assume no spin polarization.) Since
V xc(r) is a local function of charge density ρ(r) in LDA, its matrix elements are
simple to compute
Vxcij =
∫ΩV xc(r′)φi(r′)φj(r′) =
∑k
V xck
∫Ωφk(r
′)φi(r′)φj(r′). (4.21)
Note that in the above equations, Vxcij are the matrix elements of V
xc(r) while V xcj is
the FE discretization of V xc(r), i.e.,
V xc(r) =∑i
V xci φi(r). (4.22)
We introduce here another notation Fijk =∫Ω φi(r)φj(r)φk(r), which can be computed
analytically in the FE framework. Eqn. (4.21) can be rewritten as
Vxcij =
∑k
V xck Fijk. (4.23)
Similarily, matrix elements of the Hartree and the ionic pseudopotentials are calcu-
lated by
VHij =
∑k
V Hk Fijk, and (4.24)
Vps,locij =
∑k
V ps,lock Fijk, (4.25)
where V Hj and V ps,loc
j are the FE discretizations of the corresponding potentials, i.e.,
V H(r) =∑
V Hj φj(r) and V ps,loc(r) =
∑V ps,locj φj(r).
To calculate the Hartree potential in real space, instead of using the standard
formula
V H(r) =∫Ω
ρ(r′)|r− r′| , (4.26)
100
which requires O(N2) operations, we solve the Poisson equation
−V H(r) = 4πρ(r) (4.27)
with proper boundary conditions. The Poisson equation can be solved in O(N)
operations with multigrid techniques, as will be discussed in detail later.
The local part of the ionic pseudopotential is
V ps,loc(r) =∑i
∑τ
V psτ (r− rτ −Ri), (4.28)
where V psτ (r− rτ −Ri) is the pseudopotential due to the τth atom in the ith unit
cell. Since the ionic pseudopotential is long-ranged, the above expression involves an
infinite sum over all unit cells, making it impossible to carry out the calculation in
real space directly. There are several ways to get around this problem. For example,
we may Fourier transform the pseudopotential to momentum space, do the sum, (see
Eqn. (1.62)) then transform it back to the real space. However, this involves FFT,
which is what we want to avoid (see Section 4.1.1). We could also employ the elegant
procedure, known as the Ewald’s method[238], which would involve O(N2) operations.
Here we describe an alternative scheme that calculates the ionic pseudopotential in
O(N) operations.
First, we notice that the ionic pseudopotential can be regarded as the potential
due to some ficticious charge density, which satisfies
ρfict(r) =1
4πr
d2
dr2(rV ps(r)). (4.29)
This charge density has no obvious connection with the electronic charge density of
the core electrons, except for the fact that the total ficticious charge inside the core
region equals the number of valence electrons, i.e,
∫r<rc
4πρfict(r)r2dr = Zval. (4.30)
101
Since the pseudopotential decays like −Zval/r outside the core cut-off radius rc, theficticious charge density vanishes for r > rc. Therefore, although the pseudopotential
is long-ranged, the ficticious charge density is strictly localized inside the ionic core.
The total ficticious charge density in the unit cell can be evaluate in O(N) operations:
ρfict(r) =∑τ
ρfictτ (|r− rτ |). (4.31)
where τ sum over atoms in the unit cell. The advantage of working with the ficti-
cious charge density is readily seen: there is no infinite sum over all unit cells (see
Eqn. (4.28)). The full ionic pseudopotential is then recovered by solving the Poisson
equation
−V ps,loc(r) = 4πρfict(r) (4.32)
with proper boundary conditions. This is essentially the same procedure as that for
computing the Hartree potential in Eqn. (4.27) and therefore can be done in O(N)
operations.
Finally, we need to evaluate the matrix elements of the nonlocal pseudopotential
(1.55), which can be written as
V ps,nlocKB =
∑τ
lτmax∑l=0
m=l∑m=−l
|χτ,lm(r− rτ )〉〈χτ,lm(r− rτ )|. (4.33)
where the projector function
|χτ,lm〉 = |φps,τlm δV ps,τl 〉√
〈φps,τlm |δV ps,τl |φps,τlm 〉
(4.34)
is due to the contribution from the τth atom in the unit cell. See also Eqn. (1.55).
The corresponding matrix elements of the nonlocal pseudopotential are
Vps,nlocij = 〈φi|V ps|φj〉 =
∑τ
lτmax∑l=0
m=l∑m=−l〈φi|χτ,lm(r− rτ )〉〈χτ,lm(r− rτ )|φj〉. (4.35)
We have suppressed the subscript KB. The advantage of the KB form of nonlocal
pseudopotential is obvious: there is no need to evaluate the full matrix, instead, we
102
only need to calculate the projection vectors
∆τ,lmi = 〈φi|χτ,lm(r− rτ )〉 =
∫Ωφi(r)χ
τ,lm(r− rτ ). (4.36)
The full nonlocal pseudopotential matrix is simply the sum of the direct product of
the projection vectors:
Vps,nloc =∑τ
lτmax∑l=0
m=l∑m=−l
∆τ,lm(∆τ,lm)†. (4.37)
Eqn. (4.36) can be evaluated by a high-order numerical integration
∆lmi =
∫Ωφi(r)χ
lm(r) =∑k
ωkφi(rk)χlm(rk), (4.38)
where ωk is the weight associated with integration point k. We have suppressed
superscript τ for simplicity. An alternative is to discretize χ as
χlm(r) =∑k
χlmk φk(r), (4.39)
and evaluate ∆lmi via
∆lmi =
∑k
Sikχlmk , (4.40)
where Sik is the overlap matrix defined in Eqn. (4.19). We have implemented both
and see no significant differences between the two methods, although the numerical
integration scheme is slightly more accurate.
The speed of applying the nonlocal pseudopotential to the wavefunctions is an-
other advantage of carrying out calculations in real space since it involves only O(N)
operations, compared to O(N2) operations required in the momentum formalism.
4.5 Multigrid Poisson equation solver
As we have discussed in Section 4.2.1, FE discretiation of the The Poisson equation,
e.g., Eqn. (4.27) or Eqn. (4.32), results in a system of linear equations
AV = b, (4.41)
103
where Aij =∫Ω∇φi · ∇φj , V is the FE discretization of the potential, i.e., V (r) =∑
i Viφi(r) and bi = 4π∫Ω ρ(r)φi(r). Again, matrix A can be calculated analytically,
and the right-hand side b can be evaluated by either of the two schemes discussed in
the previous section for calculating ∆lm.
There are subtle issues that need to be addressed when solving the Poisson equa-
tion with periodic boundary conditions. First, solving the equation (4.27) or (4.32)
separately violates the so-called well-posedness condition for periodic systems or sys-
tems with Neumann boundary conditions, which requires∫Ωρ(r) = 0. (4.42)
The physical meaning of this condition is that if the charge in the unit-cell is not zero,
i.e.,∫Ω ρ(r) = 0, then the total charge in the system is infinite. Therefore the potential
diverges. This problem can be easily solved by subtracting the average charge density
ρ from ρ(r), i.e.,
ρ(r)← ρ(r)− ρ. (4.43)
There is another problem, however. Even if the well-posedness condition is satisfied,
the solution V is arbitrary up to a constant since V = 0 if V = const. This
arbitrariness can be avoided by imposing the following constraint on V :∫ΩV (r) = 0. (4.44)
The above two treatments are equivalent to discarding the G = 0 Fourier component
of the potential terms in the plane-wave method[37]. If the system is strictly Coulom-
bic, the G = 0 component of the ionic potential and that of the electronic Hartree
potential cancel out exactly. However, since the ionic pseudopotential (the local part)
is not Coulombic inside the ionic core, the two terms do not cancel each other. There-
fore, the calculated total energy (as well as the band structure) contains a constant
shift. This is a well-known proplem of the pseudopotential plane-wave calculations
and can be remedied by introducing a correction term to the total energy[37].
104
Solving the above linear system by any direct method, e.g., Gaussian elimination
or LU decomposition, is impractical since (i) the dimension of the matrix A is usually
very large, of the order of 106×106, making it impossible to store the entire matrix inthe computer memory, as required for most direct methods, and (ii) direct methods
usually cannot take full advantage of the sparseness of the matrix. Iterative methods
such as Gauss-Siedel (GS) or Conjugate-Gradient (CG)[239], on the other hand, are
very suitable for this type of problem. CG methods are usually stable and very fast,
and are among the best choices for pure algebraic problems. However, the computa-
tional effort of CG methods scales nearly as O(N2) even for sparse matrices. Since
the linear system we are about to solve is not a pure algebraic problem but the result
of the FE discretization of the Poisson equation, numerical mulitgrid methods[207]
can be easily applied. (There are also so-called Algebraic Multigrid (AMG) methods
for pure algebraic problems[207]). MG methods have been the most efficient tech-
niques in solving algebraic systems arising from discretization of partial differential
equations and are well known for their O(N) scaling. Standard iterative methods
such as GS and Jacobi methods, which are briefly discussed below, are often essential
components of MG methods.
4.5.1 Standard iterative methods for solving linear systems
A typical cycle of the standard iterative methods consists of three steps:
(1) form residual r = b−AV old,
(2) solve the residual equation Ae = r approximately, i.e., find B ≈ A−1 and solve
for the error function e = Br, and
(3) update the solution V new = V old + e and go to step (1) until convergence criteria
are met.
Different choices ofB result in different iterative methods. If we writeA = D− L−U,
where D is the diagonal part of A, −L and −U are the lower and upper triangu-
105
lar parts of A, respectively, we have the following classification of standard iterative
methods:
B =
ω, (ω > 0) ,Richardson,D−1 , Jacobi,ωD−1 ,Damped Jacobi,(D− L)−1 ,GS,ω(D− ωL)−1 , Successive Over Relaxation (SOR).
(4.45)
These standard iterative methods usually converge very quickly in the first few itera-
tions, but then the convergence slows down quickly, see Fig. 4.5. The reason for this
behavior is that the standard iterative methods have exceptional smoothing proper-
ties but are very inefficient in correcting the slowly-varying errors. In other words,
most of the high-spatial-frequency errors are smoothed out quickly in the first few
(less than 5) iterations. The low-frequency errors, on the other hand, persist after
many iterations. The quick drop in the error shown in Fig. 4.5 at the very begining re-
flects the fast decrease in high-frequency errors while the tail is due to the long-lasting
low-frequency errors.
4.5.2 Multigrid acceleration
The standard iterative methods are very effective in correcting the high frequency
errors, leaving mainly low frequency errors after a few iterations. The question there-
fore is, how can we speed-up the convergence of the low-frequency errors? Note that
the low-frequency errors are relatively smooth (this is what low-frequency means)
thus can be represented on a coarser grid (e.g., by doubling the grid size). Solving
the problem excatly on a coarser grid might be possible since the dimension of the
matrix A decreases as the grid size increases. Based upon these observations, we can
conceive the following two-grid method:
(1) Smoothing: Do a few (2 or 3, for example) GS iterations on the fine grid h to
smooth out the high-frequency errors, obtaining an approximate solution V h.
(2) Restriction: Form residual rh = bh − AhV h and transfer the residual rh to grid
106
0.01
0.1
1
0 5 10 15 20 25 30 35 40 45 50
Number of iterations
error
||e||
||V||
Figure 4.5: Typical convergence behavior of the GS method
2h by some interpolation scheme, i.e., r2h = I2hh rh.
(3) Coarse grid solution: Formulate the residual equation on grid 2h, namely,A2he2h =
r2h, and solve for the error function e2h.
(4) Prolongation: transfer the error function from grid 2h back to the original fine
grid h and update the solution, i.e., V h ← V h + Ih2he2h.
We have introduced superscript h and 2h to indicate the grid size. Ih2h and I2hh are a
restriction operator and prolongation operator, respectively, which represent interpo-
lation between grid h and grid 2h.
In this two-grid algorithm, we need to solve the residual equation on the coarse
grid 2h. However, if the problem on this grid is still too large, which is often so, exact
solution might still be very costly or even impossible. Fortunately, this problem can be
easily overcome by applying the above two-grid algorithm recursively until a coarsest
possible level is reached. In other words, instead of solving the residual equation on
the coarse grid, we may simply do a few GS iterations again and then transfer the
107
current residual (which now becomes the residual of the residual) equation onto an
even coarser grid. The process is repeated until an exact solution of the residual
equation becomes feasible. This is the basic idea of the multigrid (MG) method.
There are several flavors of MG algoritms[207] which we will not discuss in detail here.
Actually, the MG procedure we have just described is called back-slash MG algorithm.
We adopted a widely used and more powerful MG scheme, the so-called V-cycle MG
algorithm[207] to solve the Poisson equation. The pre-smoothing and post-smoothing
on all grids other than the coarsest one are done with two GS relaxations, while the
residual equation on the coarsest grid is solved nearly exactly.
4.5.3 Performance of the MG Poisson equation solver
Fig. 4.6 show the typical convergence behavior of the MG Poisson equation solver.
Uniform exponential decrease in error can be easily seen, regardless of the matrix
size. Although we only show the 2-norm error, we find the convergence behavior to
be independent of the choice of the functional norm. All calculations start from zero
initial guesses for the potential as indicated by the 100% starting errors. Table 4.1
shows the O(N) scaling properties of this algorithm. The slight deviation from the
exact O(N) scaling might due to other issues that are not related to the algorithm
itself, for example, data transfer latency between the main memory and the cache
tends to increase as the matrix size increases.
108
Number of MG iterations
error
||e||
||V||2
2
1e-14
1e-12
1e-10
1e-8
1e-6
1e-4
1e-2
1
0 1 2 3 4 5 6
N=32,768
N=131,072
Figure 4.6: Convergence properties of the MG Poisson equation solver. The solidand the dotted lines correspond to fine-grid matrix sizes of 32,768×32,768 and131,072×131,072, respectively.
# atoms Matrix size # MG iterations error: ‖e‖2‖V ‖2 CPU time (sec)
2 32,768×32,786 6 1.02E-13 1.654 65,536×65,536 6 7.62E-14 3.598 131,072×131,072 6 7.78E-14 7.7416 262,144×262,144 6 9.46E-14 17.5332 524,288×524,288 6 1.00E-13 34.29
Table 4.1: O(N) scaling properties of the MG Poisson equation solver. All calcula-tions were done on an alpha DS10 workstation.
109
4.6 Full multigrid Kohn-Sham equation solver
Once the Hartree and ionic pseudopotentials are calculated and all necessary ma-
trix elements are evaluated, we are ready to solve the generalized eigenvalue problem
(4.17) arising from discretization of the Kohn-Sham equations. Solving the (general-
ized) eigenvalue problem is, however, far more complicated than solving the Poisson
equation. In this section, we develop a full multigrid (FMG) algorithm for solving the
Kohn-Sham equations self-consistently. (The meaning of FMG will be clear later.)
The FMG algorithm for solving eigenvalue problems was first developed by A. Brandt,
et al.[206]. Costiner, et al. [243, 244] later implemented an FMG algorithm to solve
nonlinear model Schrodinger equations. Our FMG algorithm differs from the previous
implementations in several respects. First, we do not adopt the full approximation
scheme (FAS). (Comments on the drawbacks of applying the FAS to solve the Kohn-
Sham equations for real systems can be found in Ref. [189].) Instead, we employ
the residual correction MG approach similar to that for solving Poisson’s equation
discussed in the previous section, which can be regarded as the MG implementation
of the Davidson method[242] for eigenvalue problems. Second, while the previous
work was on a standard eigenvalue problems, our method deals with the generalized
eigenvalue probem. Third, instead of working on some model problems, we intend
to develop an ab initio electronic structure method for real systems based on the FE
method. Finally, our approach achieves potential self-consistency and convergence in
eigenpairs (eigenvalues and corresponding eigenfunctions) simultaneously in as few as
15 iterations on the finest level.
4.6.1 Brief description
Here we describe a 3-level (i.e., coarse, intermediate and fine levels) FMG scheme
for the Kohm-Sham problem. The FMG eigenvalue solver starts on the coarse level Lc,
110
on which the generalized eigenvalue problem is solved with a CG algorithm[240, 241],
as will be discussed in detail later. After performing a few self-consistent iterations
on level Lc, we obtain a set of initial guesses for the KS eigenvalues and eigenfunc-
tions. The eigenfunctions are then interpolated onto an intermediate level Li and
renormalized. The interpolations are done with the standard FE prolongation pro-
cedure. At the intermediate level, both the self-consistency of the potential and the
accuracy of the eigenpairs are improved simultaneously through an MG algorithm
similar to that for solving the Poisson equations. The results obtained on level Li are
further transferred to the fine level Lf ; subsequent refinement of the eigenpairs and
improvement of the self-consistency are carried out, again, simultaneously, through
an MG algorithm.
The FMG scheme for eigenvalue problems differs from the MG Poisson equation
solver described in the previous section in several respects:(i) Whereas the MG Pois-
son equation solver starts on the fine grid, the FMG eigenvalue solver begin with
the coarse grid solution. The reason is that the MG Poisson equation solver usually
converges very quickly for almost any initial guess satisfying proper boundary condi-
tions (e.g., a zero initial guess). Solving the eigenvalue problem, on the other hand,
is very time consuming and the convergence is usually slow and depends strongly
on the initial guess. Therefore, obtaining a reasonable initial guess is crucial to the
algorithm’s performance. In self-consistent electronic structure calculations, we need
initial guesses for both potentials and eigenpairs. The solutions on the coarse grid,
after being transferred onto the intermediate level Li as initial guesses, could signif-
icantly reduce the computational effort on the level Li. Similarly, results obtained
on level Li serve as good initial guesses for the fine level solutions. (ii) In the FMG
scheme, calculations on all levels other than the coarse one involve an MG loop. This
explains why the algorithm is called full multigrid. Actually, each MG loop in the
FMG eigenvalue solver is similar to the MG algorithm for solving the Poisson equa-
111
tion. (iii) In order to obtain a meaningful initial guess, the coarse grid must be fine
enough that the essential characteristics of the wavefunctions can be represented. In
contrast, the lowest (coarsest) level in the MG Poisson equation solver could be much
coarser than Lc. Since the coarse grid in the FMG scheme can not be too coarse,
we believe that two levels of subsequent refinements are enough for obtaining a fairly
accurate solution, although more intermediate grid levels can be easily implemented.
Note that the coarse grid solution not only provides an initial guess for the finer level
problem but also separates the clusters of eigenvectors so that no explicit reorthogo-
nalization between clusters of eigenvectors is necessary. However, eigenvectors within
a cluster (including degenerate eigenvectors) need to be separated explicitly. We will
return to this point later.
4.6.2 Coarse grid solution
CG algorithms[240, 241] are among the best for finding the lowest eigenpair for
large sparse symmetric (or Hermitian) matrices without any good initial guesses.
Therefore, they are very suitable for solving the coarse grid problem. We adopt
the CG algorithm of Yang[240], which is based on the minimization of the Rayleigh
quotient for generalized eigenvalue problems. Although Yang only proved the con-
vergence properties of the algorithm for the lowest eigenpair[240], we find that the
algorithm is capable of calculating the lowest few eigenpairs with some modifications.
Rayleigh quotient
The Rayleigh Quotient R of the generalized eigenvalue problem
Hx = εSx (4.46)
is defined by
R(x) =x†Hx
x†Sx, (4.47)
112
where x ∈ Cn, x = 0. Obviously, R is bounded by the two extreme eigenvalues of
Eqn. (4.46), i.e., εmin ≤ R(x) ≤ εmax. Since the lower bound of R is the smallest
eigenvalue of the system, minimization of R(x) with respect to x should give λmin.
Minimization of the Rayleigh quotient by the conjugate gradient method
A typical gradient-based minimization algorithm consists of three steps:
(i) Choose a search direction pk,
(ii) Perform one-dimentional minimization along direction pk, i.e.,
minαk
R(xk−1 + αkpk), (4.48)
(iii) Update xk = xk−1 + αkpk.
As we can see, the two basic ingredients of the minimization algorithm are computing
search directions and performing line minimizations.
In the CG method, the search direction p is related to the gradient g by
pk = −gk + βkpk−1, (4.49)
where the gradient g(x) of Rayleigh quotient R(x) is given by
g(x) =2
x†Sx(Hx− R(x)Sx), (4.50)
and has the property
x†g(x) = 0, ∀x = 0. (4.51)
At the kth iteration, βk is given by
βk =
0 if k = 0,
g†kgk/g†k−1gk−1 otherwise.
(4.52)
In practical calculations, βk is usually reset to zero after a fixed number of iterations.
The other parameter (αk) is determined by line minimization of the Rayleigh quotient
113
(4.48). By setting the derivative of R(xk + αkpk) to zero, we have the following
equation for αk:
aα2k + bαk + c = 0, (4.53)
where
a = (x†kSpk)(p
†kHpk)− (x†
kHpk)(p†kSpk), (4.54)
b = (x†kSxk)(p
†kHpk)− (x†
kHxk)(p†kSpk), (4.55)
and
c = (x†kSxk)(x
†kHpk)− (x†
kHxk)(x†kSpk). (4.56)
Solving the Eqn. (4.53) gives
α±k =
−b±√b2 − 4ac2a
. (4.57)
α± corresponds to the minimum and maximum of R along the direction p and we
should choose α+. Numerically more stable formulas for calculating α+ are
α+k =
−2c/(b+√b2 − 4ac) if b > 0,
(−b+√b2 − 4ac)/2a if b ≤ 0 and a = 0.(4.58)
Special care must be taken for the case when a = 0 and b ≤ 0[240]. However, this
almost never happens in numerical calculations.
The CG method for generalized eigenvalue problems described above is rather
stable. In most cases, the minimization converges to the lowest eigenpair with almost
any initial guess. By imposing orthgonalization constrains, we have generalized the
algorithm for computing the lowest few eigenpairs. Unfortunately, this algorithm
scales like O(N3), making it unsuitable for large-scale problems. However, since
the matrix size on the coarse level is relatively small, the computational effort is
insignificant for reasonable system size (say, up to a few hundred). For a system
containing 32 atoms, only about 3% of the total CPU time is spent on this level.
114
4.6.3 O(N2) iterative diagonalization methods
After solving the Kohn-Sham equations on the coarse grid, we obtain a set of
eigenpairs, which, upon being transferred, serves as a reasonable initial guess for the
finer level problem. There are several iterative diagonalization schemes, for example,
Rayleigh-quotient [239] and Davidson iteration[242], which do not require explicit
re-orthogonalization between eigenfunctions if a good initial guess for the eigenvalues
and the eigenvectors is given. As we have discussed at the begining of this chapter,
if no explicit re-orthogonalization is needed, O(N2) scaling can be achieved.
Rayleigh-quotient iteration
Rayleigh-quotient iteration is closely related to the inverse iteration technique,
based on the fact that if λ is close to a distinct eigenvalue, say, λi, of the matrix
H, then the following iteration converges to the eigenvector corresponding to λi with
some initial guess x:
do while (not converged)y = (H− λI)−1xx = y/‖y‖2
end do
(4.59)
The convergence property of the above iteration can be understood by expanding x
as a superposition of eigenvectors xi
x =n∑i
cixi, (4.60)
so that, after one iteration,
y =n∑i
ciλi − λ
xi. (4.61)
It can be easily seen that y has a larger component in the direction of xi than the
original vector x does if λ is close to λi.
The Rayleigh-quotient iteration replaces the fixed parameter λ by the Rayleigh-
115
quotient (x†Hx) (assuming x is normalized):
do while (not converged)λ = x†Hxsolve (H− λI)y = x for yx = y/‖y‖2
end do
(4.62)
For generalized eigenvalue problem, we have
do while (not converged)λ = x†Hxsolve (H− λS)y = Sx for yx = y/‖y‖S
end do
(4.63)
where ‖y‖S = (y†Sy)1/2. The Rayleigh-quotient iteration almost always converges
and when it does, the rate of convergence is cubic[239]. A serious drawback of the
Rayleigh-quotient iteration is that it involves solving the linear equation (H−λS)y =
Sx, which is not as easy as solving the Poisson equation discussed in Section 4.5 due
to the positive-indefiniteness of the matrix (H− λS). (Note that the matrix size of a
typical real-space calculation is of the order of 106× 106). Also, the matrix (H− λS)
gets close to singular as λ approaches to the exact eigenvalue, making it even more
difficult to solve the linear equation. Relaxation techniques, on the other hand, do
not suffer from these drawbacks.
Improving the eigenvectors by relaxation
Recall that the ultimate goal of solving the eigenvalue problem is to find λ ∈ C
and x ∈ Cn such that
Hx = λx, (4.64)
or
(H− λI)x = 0. (4.65)
116
(For simplicity, we return to the standard eigenvalue problem.) This is essentially
a linear equation if λ is known. Given some reasonable initial guess x0 to x, we
may approximate λ by the Rayleigh-quotient λ0 = (x†0Hx0)/(x
†0x0) and improve the
eigenvector x = x0 + δ by relaxing the following equation
(H− λ0I)(δ + x0) = 0, (4.66)
or equivalently,
(H− λ0I)δ = r, (4.67)
where the residual r is
r = −(H− λ0I)x0. (4.68)
Note that we shall not solve Eqn. (4.67) exactly. In fact, the exact solution gives a
trivial result, namely, δ = −x0, or x = δ + x0 = 0. However, we may employ some
relaxation techniques, e.g., GS or Jacobi methods (see Section 4.5.1), to obtain an
approximate solution δ. A Jacobi-type iteration gives
δ = (D− λ0I)−1r, (4.69)
where D is the diagonal of H. This is the so-called Davidson iteration[242]. At this
point, we may update the eigenvector
x0 ← x0 + δ (4.70)
and the Rayleigh-quotient
λ0 ← x†Hx
x†x. (4.71)
The above process, namely, Eqn. (4.69–4.71), is repeated until a certain convergence
criterion is satisfied. Unfortunately, Davidson iteration usually converges very slowly.
This slow convergence rate of Davidson iteration is not unexpected based on our
discussion in Section 4.5, since the Davidson method is similar to the Jacobi method
117
for pure linear systems. The similarity between the two problems, on the other hand,
inspires us to formulate an MG algorithm for the eigenvalue problem.
Multigrid acceleration
1e-12
1e-10
1e-08
1e-06
1e-4
1e-2
1
0 1 2 3 4 5 6 7 8 9
MG relaxation
Davidson relaxation
Eigenvalueerror
Number of iterations
Figure 4.7: Comparison of the convergence properties of the standard Davidson iter-ation and the MG relaxation scheme we developed. The relaxations were performedon the finest grid, using the intermediate-grid solution as a initial guess.
The basic idea of our MG method for the generalized eigenvalue problem (4.17) is
very similar to what we have discussed in Section 4.5.2, which we shall not repeat here.
However, we shall point out several important differences between the two schemes.
First, as we have mentioned previously, we shall not solve Eqn (4.67) exactly. In-
stead, the equation is relaxed (in a MG fashion) and the eigenvalue is updated by
the Rayleigh-quotient (4.71). Second, unlike the Poisson operator, which is positive-
definite, the matrix (H− λI) (or (H− λS) for the generalized eigenvalue problem) is
positive indefinite. Third, whereas the Poisson operator is uniquely defined on each
118
grid level, multigrid representations of the Kohn-Sham Hamiltonian matrix (the po-
tential part) on grids other than the finest grid are somewhat arbitrary. For example,
the ionic potential on the coarse grid may be either calculated by solving the Poisson
equation directly or by interpolating the results from the finer grid. Nevertheless, the
final converged results do not depend on these subtle differences.
Fig. 4.7 compares the typical convergence behavior of traditional Davidson iter-
ation and our MG relaxation algorithm. Both relaxations use the solution on the
coarser grid as a initial guess. Whereas the convergence of the Davidson method
resembles that of standard iteration methods for linear systems (see Fig. 4.5), the
MG relaxation schems for the generalized eigenvalue problem has the similar con-
vergence properties to that of the MG Poisson equation solver (see Fig. 4.6). The
error in eigenvalue is reduced to about 10−12 in less than 10 MG relaxations. (The
absolute value of the eigenvalue is of the order of 1.) Note that there have been
conjectures in the literature that it is not possible to formulate a residual-correction
MG scheme for nonlinear (eigenvalue) problems[245]. Our results demonstrate that
the residual-correction MG scheme is also suitable for solving the eigenvalue problem.
Rayleigh-Ritz Projection
The algorithm described in the above works well if all eigenvalues are distinct, and
since individual eigenpairs can be improved independently, we expect this algorithm
to scale like O(N2). However, if there are clustered or degenerate eigenvalues, this
algorithm is not able to separate the corresponding eigenvectors well. The result is
that the eigenvectors with the cluster might not be orthogonal to each other after
the MG relaxation. Fortunately, this problem can be easily solved by the so-called
Rayleigh-Ritz projection. The Rayleigh-Ritz projection is essentially a subspace diag-
onalization, which can be used to separate the eigenvectors and find the correspond-
ing eigenvalues if only linear combinations of eigenvectors are known. For example,
119
suppose we have m vectors, φ1, φ2, · · ·φm, which are linear combinations of eigenvec-tors corresponding to m-fold degenerate (or clustered) eigenvalues of the generalized
eigenvalue problem (4.46). The eigenvectors and the corresponding eigenvalues can
be obtained by solving the following m×m eigenvalue problem:
HC = εSC, (4.72)
where
Hij = φ†iHφj, and Sij = φ†
iSφj. (4.73)
The separated eigenvector corresponding to eigenvalue εi is
ψi =∑j
φjCji. (4.74)
Therefore, whereas the separation of clusters of eigenvectors is accomplished through
the coarse grid solutions, the eigenvectors within the cluster are separated by Rayleigh-
Ritz projection. We have implemented a simple algorithm to identify clustered eigen-
values and perform the Rayleigh-Ritz projection after every MG relaxation. Although
the Rayleigh-Ritz projection procedure brings about some extra computational effort
to the algorithm, the overhead is reasonably small.
4.6.4 Simultaneous convergence of potential and eigenvalues
In ab initio electronic structure calculations, the electronic potential and wavefunc-
tions have to be calculated in a self-consistent manner. A typical SC loop consists of
4 steps:
120
do while (not converged)
(1) construct the potential V = V (ρold)
(2) solve the Kohn-Sham equations (through MG relaxation),
(3) calculate the new charge density ρnew
(4) mix the old charge density with the new one, e.g.ρnew ← αρold + (1− α)ρnew; ρold = ρnew
end do
Solving the Kohn-Sham equations is obviously the most time consuming part of the
problem if we require the wavefunctions (eigenvalues) to be well converged at each SC
iteration. As Fig. 4.7 shows, it takes about 10 MG iterations to converge the eigen-
value to about 10−12 with a fixed potential. However, before the potential converges,
it might not be neccessary to insist that the wavefunctions be well converged. If we
update the potential after, say, 2 or 3 MG relaxation, we might be able to save 60%
− 80% of the CPU time on this step and hopefully, we will also be able to obtain a
simultaneous convergence in eigenvalues and the potential at the end. In fact, we do
achieve this simultaneous convergence in as few as 15 SC iterations.
4.7 Performance of the FMG electronic structure
code
Table 4.2 shows the scaling and convergence properties of our 3-level FMG code
for silicon systems. (Note that the Poisson equation solver might involve more than
3 levels). The calculations were done on an alpha DS10 workstation equipped with
1GB memory. Although we have only done Γ-point sampling, implementing more
general k-point sampling is straightforward.
The volume of the 2-atom unit cell is uniformly partitioned into 24,576 tetrahedral
elements on a 32×32×32 grid. (Adaptive nonuniform FE triangulation is intended
121
as future work.) The grid sizes of the 4-atom, 8-atom, 16-atom and 32-atom cells are
32×32×64, 32×64×64, 64×64×64 and 64×64×128, respectively.
# atoms CPU time # SC charge density max eigenvalue
(sec) iterations error∫
|δρ|d3r∫ρd3r
error (Ryd)
2 85 (23) 15 1.8E-5 3.3E-74 255 (85) 15 1.4E-5 2.4E-78 895 (349) 15 1.4E-5 2.0E-716 3873 (3676) 15 1.7E-5 6.4E-732 14967 (16987) 15 3.7E-5 3.3E-6
Table 4.2: Scaling and convergence properties of the FMG real-space electronic struc-ture code. All calculations were done on an alpha DS10 workstation. Note that theerrors refer to the differences in the corresponding values between two consecutiveiterations, not the absolute errors with respect to the converged values. CPU timefor the corresponding plane-wave calculation is shown in the parenthesis.
For all of the calculations, we carry out 8 SC iterations on the coarse and inter-
mediate levels before moving onto the fine level, on which we perform an additional
15 SC iterations. The generalized eigenvalue problem on the coarse level is solve with
the CG algorithm described in Section 4.6.2. Actually, the CG algorithm is only used
at the very beginning of the SC iteration. Further improvement of the eigenvectors
and the potential is done with the relaxation methods describe in Section 4.6.3, since
the eigenvalues and eigenvectors from the previous SC step provide reasonable initial
guesses for the relaxation i scheme. The eigenvalue problem is solved by the MG
relaxation algorithm on the intermediate and the fine levels. Within each SC cycle,
we perform only one MG relaxation on the eigenvectors and subsequently update the
eigenvalues and the charge density (potential). The charge density mixing is done
with a linear mixing scheme:
ρnew ← αρold + (1− α)ρnew, (4.75)
122
with α = 0.65 for all calculations. As it is shown in Table 4.2, the CPU time scales
approximately as O(N2), where N is the number of atoms in the system. Simultane-
ous convergence of the charge density (or equivalently, potential) and the eigenvalues
(eigenfunctions) are achieved with as few as 15 iterations on the finest level. CPU
times for the corresponding plane-wave calculation with comparable accuracy are
shown in parentheses for comparison. Our real-space code has potentially better per-
formance than the traditional plane-wave code for system containing more than 32
atoms.
Now we briefly discuss the storage (memory) requirement of the code. Since
we calculate all of the occupied wavefunctions, the memory usage scales as O(N2)
asymptotically for large N . For small systems, the dominant storage requirement is
for the relevant matrices, i.e., the Hamiltonian, kinetic energy, and potential, and
scales linearly with the system size. For large systems, however, the storage for the
wavefunctions becomes more demanding, and scales quadratically with system size.
For example, the grid size of a 32-atom silicon system is 64×64×128=524,288. Sincethe number of occupied bands is 64, the memory required for storing the wavefunc-
tions is 64 × 524, 288 × 8 = 268, 435, 456 Bytes = 256 MB. A supercomputer node
nowadays is usually equipped with 4 − 8 GB of memory (1 GB= 1,024 MB), which
should be capable of handling systems containing up to 150 atoms. One way to get
around the memory bottleneck is to store the wavefunctions temporarily on the hard
disk. However, the performance of the code will be degraded significantly.
4.8 Summary
In summary, we have developed a FMG real-space electronic structure method
based on the FE method. The computational cost of this method scales nearly like
O(N2), as compared to the O(N3) scaling of the traditional methods such as pseu-
dopotential plane-wave methods[37]. The O(N2) scaling is acheived by avoiding the
123
explicit reorthogonalization of wavefunctions. Whereas the separation of clusters
wavefunction is accomplished through the FMG algorithm, wavefunctions within a
cluster are separated by the Rayleigh-Ritz projection.
124
Bibliography
[1] F. H. Stillinger and T. A. Weber, Phys. Rev. B 31, 5262 (1985).
[2] J. Tersoff, Phys. Rev. Lett. 56, 632 (1986).
[3] J. Tersoff, Phys. Rev. B 37, 6991 (1988)
[4] R. Biswas and D. R. Hamann, Phys. Rev. Lett. 55, 2001 (1985).
[5] B. W. Dodson, Phys. Rev. B 35, 2795 (1987).
[6] D. W. Brenner and B. J. Garrison, Phys. Rev. B 34, 1304 (1986).
[7] J.C. Slater and G. F. Koster, Phys. Rev. 94, 1489 (1954).
[8] F. Bloch, Z. Physik 52, 555 (1928).
[9] P. O. Lowdin, J. Chem. Phys. 18 365 (1950).
[10] D. J. Chadi, Phys. Rev. B 19, 2074 (1979).
[11] S. Froyen and W. A. Harrison, Phys. Rev. B 20, 2420 (1979).
[12] L. Goodwin, A. J. Skinner, and D. G. Pettifor, Europhys. Lett. 9, 701 (1989).
[13] See, for example, C. M. Goringe, D. R. Bowler, and E. Hernandez, Rep. Prog.
Phys. 60, 1447 (1997).
125
[14] See, for example, C. Z. Wang, C. T. Chan, and K. M. Ho, Phys. Rev. B. —bf
39, 8586 (1989).
[15] C. H. Xu, C. Z. Wang, C. T. Chan, and K. M. Ho, J. Phys. Condens. Matter.
4, 6047 (1992).
[16] J. Dorantes-Davila, A. Vega, and G. M. Pastor, Phys. Rev. B 47, 12995 (1993).
[17] J. Dorantes-Davila and G. M. Pastor, Phys. Rev. B 51, 16627 (1995).
[18] D. A. Mirabella, C. M. Aldao, and R. R. Deza, Phys. Rev. B 50, 12152 (1994).
[19] B. A. Mckinnon and T. C. Choy, Phys. Rev. B 52, 14531 (1995).
[20] R. S. Mulliken, J. Chem. Phys. 46, 497 (1949).
[21] R. S. Mulliken, J. Chem. Phys. 46, 675 (1949).
[22] M. van Schilfgaarde and W. A. Harrison, J. Phys. Chem Solids 46, 1093 (1985).
[23] M. van Schilfgaarde and W. A. Harrison, Phys. Rev. B 33, 2653 (1986).
[24] D. Porezag, Th. Frauenheim, and Th. Kohler, Phys. Rev. B 51, 12947 (1995).
[25] D. R. Hartree, Proc. Cambridge Philos. Soc. 24, 89 (1928).
[26] V. Fock, Z. Phys. 61, 126 (1930).
[27] J. C. Slater, Phys. Rev. 35, 210 (1930).
[28] R. O. Jones and O. Gunnarsson, Rev. Mod. Phys. 61, 689 (1989).
[29] J. C. Slater, Phys. Rev. 81, 285 (1951).
[30] J. C. Slater, Phys. Rev. 82, 538 (1951).
[31] L.H Thomas, Proc. Cambridge Philos. Soc. 23, 542 (1927).
126
[32] E. Fermi, Z. Phys. 48, 73 (1928).
[33] P. Hohenberg and W. Kohn, Phys. Rev. 136, B 864, (1964).
[34] W. Kohn L. J. Sham, Phys. Rev. 140, A 1133, (1965).
[35] M. Levy, Proc. Natl. Acad. Sci. (USA) 76, 6062 (1979).
[36] M. S. Hybertsen and S. G. Louie, Phys. Rev. B 34, 5390 (1986).
[37] J. Ihm, A. Zunger, and M. L. Cohen, J. Phys. C 12, 4409 (1979).
[38] E. Fermi, Nuovo Cimento 11, 157 (1934).
[39] H. J. Hellmann, J. Chem. Phys. 3, 61 (1935).
[40] M. L Cohen and J. R. Chelikowsky, in Electronic Structure and Optical Prop-
erties of Semiconductors, (Springer-Verlag Berlin Heidelberg, 1989).
[41] D. R. Hamann, M. Schluter, and C. Chiang, Phys. Rev. Lett. 43, 1494 (1979).
[42] D. R. Hamann, Phys. Rev. B 40, 2980 (1989).
[43] N. Troullier and J. L. Martins, Phys. Rev. B 43, 1993 (1990).
[44] A. M. Rappe, K. M. Rabe, E. Kaxiras, and J. D. Joannopoulos, Phys. Rev. B
41, 1227 (1990).
[45] D. Vanderbilt, Phys. Rev. B 41, 7892 (1990).
[46] L. Kleinman and D. M. Bylander, Phys. Rev. Lett. 48, 1425 (1982).
[47] See, for example, M. E. Casida, in Recent Advances in Density-Functional Meth-
ods, edited by D. P. Chong (World Scientific, Singapore, 1995), Pt. 1, p. 155.
127
[48] See, for example, E. K. U. Gross, J. F. Dobson, and M. Petersilka, in Density
Functional Theory, edited by R. F. Nalewajski (Springer-Verlag, Berlin, 1996),
p. 81.
[49] G. E. Engel and W. E. Pickett, Phys. Rev. B 54, 8420 (1996).
[50] A. Seidl, A. Gorling, P. Vogl, J. A. Majewski, and M. Levy, Phys. Rev. B 53,
3764 (1996).
[51] Paula Sanchez-Friera and R. W. Godby, Phys. Rev. Lett. 85, 5611 (2000).
[52] L. Hedin and S. Lundqvist, Solid State Phys. 23 1969, eds. H. Ehrenreich, F.
Seitz and D. Turnbull (New York Academic).
[53] L. J. Sham and M. Schluter, Phys. Rev. Lett. 51 1888 (1983).
[54] L. J. Sham and M. Schluter, Phys. Rev. B 32 3883 (1985).
[55] J. P. Perdew and M. Levy, Phys. Rev. Lett. 51 1884 (1983).
[56] L. Hedin, Phys. Rev. 139 A796 (1965).
[57] M. S. Hybertsen and S. G. Louie, Phys. Rev. Lett. 55, 1418 (1985).
[58] M. S. Hybertsen and S. G. Louie, Phys. Rev. B. 32, 7005 (1985).
[59] W. G. Aulbur, L.Jonsson, and J. W. Wilkins, Solid State Phys. 54, 1 (2000).
[60] B. Farid, R. W. Godby, and R. J. Needs, in 20th International Conference
on the Physics of Semiconductors, E. M. Anastasskis and J. D. Joannopoulos
(World Scientific, eds. Singapore, 1990), Vol. 3, p. 1759.
[61] H. W. Kroto, J. R. Health, S. C. O’Brien, R. F. Curl, and R. E. Smalley, Nature
318, 162 (1985).
128
[62] S. Iijima, Nature 354, 56 (1991).
[63] See, for example, M. S. Dresselhaus, G. Dresselhaus, and P. C. Eklund, Science
of Fullerenes and Carbon Nanotubes, (Academic Pressg, 1996).
[64] S. Iijima and T. Ichihashi, Nature 363, 603 (1993).
[65] D. S. Bethune, C. H. Kiang, M. S. de Vries, G. Gorman, R. Savoy, J. Vazques,
and R. Beyers, Nature 363, 605 (1993).
[66] C. Journet, W. K. Master, P. Bernier, A. Loiseau, M. Lamy de la Chapelle, S.
Lefrant, P. Deniard, R. Lee, and J. E. Fischer, Nature 388, 756 (1997).
[67] A. Thess, R. Lee, P. Nikolaev, H. Dai, P. Petit, J. Robert, C. Xu, Y. H. Lee,
S. G. Kim, A. G. Rinzler, D. T. Colbert, and R. E. Smalley, Science 273, 483
(1996).
[68] J. Kong, A. M. Cassell, and H. Dai, Chem. Phys. Lett. 292, 567 (1998).
[69] J. Kong, H. Soh, A. M. Cassell, C. F. Quate, and H. Dai, Nature 395, 878
(1998).
[70] A. M. Cassell, J. A. Raymakers, J. Kong, and H. Dai, J. Phys. Chem. B 103
6484 (1999).
[71] P. Zhang and V. H. Crespi, Phys. Rev. Lett. 83, 1791 (1999).
[72] R. Saito, M. Fujita, G. Dresselhaus, and M. S. Dresselhaus, Appl. Phys. Lett.
60, 2204 (1992).
[73] T. W. Ebbeson, Phys. Today, 6, 26 (1996).
[74] S. J. Tans, A. R. M. Verschueren, and C. Dekker, Nature 393, 49 (1998).
129
[75] R. Martel, T. Schmidt, H. R. Shea, T. Hertel, and Ph. Avouris, Appl. Phys.
Lett. 73, 2447 (1998).
[76] C. Dekker, Phys. Today 5, 22 (1999).
[77] N. Hamada, S. Sawada, and A.Oshiyama, Phys. Rev. Lett. 68, 1579 (1992).
[78] J. W. Mintmire, B. I. Dunlap, and C. T. White, Phys. Rev. Lett. 68, 631 (1992).
[79] R. Saito, M. Fujita, G. Dresselhaus, and M. S. Dresselhaus, Appl. Phys. Lett.
60, 2204 (1992).
[80] J. Tersoff and R. S. Ruoff, Phys. Rev. Lett. 73, 676 (1994).
[81] M. M. J. Treacy, T. W. Ebbeson and J. M. Gibson, Nature,381, 678 (1996).
[82] E. W. Wong, P. E. Sheehan, C. M. Lieber, Science, 277, 1971 (1997).
[83] B. I. Yakobson, C. J. Brabec, and J. Bernholc, Phys. Rev. Lett. 76, 2511 (1995).
[84] C. F. Cornwell and L. T. Wille, Solid State Comm. 101, 555 (1996).
[85] J. P. Lu, Phys. Rev. Lett. 79, 1297 (1997).
[86] P. Zhang, P. E. Lammert, and V. H. Crespi, Phys. Rev. Lett. 81, 5346 (1998).
[87] P. Zhang and V. H. Crespi, Phys. Rev. B. 62, 11050 (2000).
[88] X. F. Zhang, X. B. Zhang, G. Van Tendeloo, J. Cryst. Growth 130, 368 (1993).
[89] S. Amelinckx, D. Bernaerts, X. B. Zhang, G. Van Tendeloo, and J. Van Landuy,
Science 267, 1334 (1995).
[90] M. Yudasaka, R. Kikuchi, T. Matsui, Y. Ohki, S. Yoshimura, and E. Ota, Appl.
Phys. Lett. 67, 2477 (1995).
130
[91] A. Maiti, C. J. Brabec, and J. Bernholc, Phys. Rev. B 55, R6097 (1997).
[92] M. Yudasaka, R. Kikuchi, Y. Ohki, E. Ota, and S. Yoshimura, Appl. Phys.
Lett. 70, 1817 (1997).
[93] M. Endo and H. W. Kroto, Solid State Comm. 84, 201(1992).
[94] M. S. Dresselhaus, G. Dresselhaus, and R. Saito, J. Phys. Chem. 96 6941 (1992).
[95] R. Saito, M. Fujita, G. Dresselhaus, and M. S. Dresselhaus, Mat. Sci. Eng. 19,
185 (1993).
[96] C. H. Kiang and W. A. Goddard III, Phys. Rev. Lett. 76, 2515 (1996).
[97] S. Iijima and T. Ichihashi, and Y. Ando, Nature 356, 776 (1992).
[98] N. Hatta and K. Murata, Chem. Phys. Lett. 217, 398 (1994).
[99] G. von Helden, M. T. Hsu, P. R. Kemper, and M. T. Bowers, J. Chem. Phys
95, 3835 (1991).
[100] G. von Helden, M. T. Hsu, N. Gotts, and M. T. Bowers, J. Phys. Chem. 97,
8182 (1993).
[101] H. Handschuh, P. S. B. G. Ganteor, B. Kessler, and W. Eberhardt, Phys. Rev.
Lett. 74, 1095 (1995).
[102] S. Bandow, S. Asaka, Y. Saito, A. M. Rao, L. Grigorian, and P. C. Eklund,
Phys. Rev. Lett. 80, 3779 (1998).
[103] C. Z. Wang, K. M. Ho, and C. T. Chan, Phys. Rev. Lett. 70, 611 (1993).
[104] J.-C. Charlier, A. De Vita, X. Blase, R. Car, Science 275, 646 (1997); Y.-K.
Kwon, Y. H. Lee, S.-G. Kim, P. Jund, D. Tomanek, and R. E. Smalley, Phys.
Re. Lett. 79, 2065 (1997).
131
[105] N. G. Chopra, L. X. Benedict, V. H. Crespi, M. L. Cohen, S. G. Louie, and A.
Zettl, Nature 377, 135 (1995);
[106] L. X. Benedict, N. G. Chopra, M. L. Cohen, A. Zettl, S. G. Louie, and V. H.
Crespi, Chem. Phys. Lett. 286, 490(1998).
[107] H. Kroto, personal communication.
[108] S. Suzuki, et al., Bull-Am. Phys. Soc. 44, No. 1, Part II, 1433 (1999).
[109] P. M. Ajayan and S. Iijima, Nature 358, 23 (1992).
[110] A. M. Rao, E. Richter, and M. S. Dresselhaus, Science 275, 187 (1997).
[111] L. Qin, S. Iijima, H. Kataura, Y. Maniwa, S. Suzuki, and Y. Achiba, Chem.
Phys. Lett. 268, 101, (1997). Nanotubes distributed across all wrapping angles
have been seen, e.g. L. Qin, S. Iijima, Chem. Phys. Lett. 269, 65 (1997).
[112] C. L. Kane and E. J. Mele, Phy. Rev. Lett. 78, 1932 (1997).
[113] A. Rochefort, D. R. Salahub, and Ph. Avouris, Chem. Phys. Lett. 297, 45
(1998).
[114] M. R. Falvo, G. J. Clary, R. M. Taylor II, V. Chi, F. P. Brooks, Jr., S. Washburn,
and R. Superfine, Nature 389, 582 (1997).
[115] A. Bezryadin, A. R. M. Verschueren, S. J. Tans, and C. Dekker, Phys. Rev.
Lett. 80, 4036 (1998).
[116] L. Liu, C. S. Jayanthi, M. Tang, S. Y. Wu, T. W. Tombler, C. Zhou, L.
Alexseyev, J. Kong, and H. Dai, Phys. Rev. Lett. 84, 4950 (2000).
[117] P. E. Lammert, P. Zhang, and V. H. Crespi, Phys. Rev. Lett. 84, 2453 (2000).
132
[118] S. S. Wong, E. Joselevich, A. T. Woolley, C. L. Cheung, and C. M. Lieber,
Nature 394, 52 (1998).
[119] J. H. Hafner, C. L. Cheung, C. M. Lieber, Nature 398, 761 (1999).
[120] A. J. Stone and D. J. Wales, Chem. Phys. Lett. 128, 501 (1986).
[121] M. B. Nardelli, B. I. Yakobson, and J. Bernholc, Phys. Rev. B 57, R4277 (1998).
[122] R. L. Murry, D. L. Strout, G. K. Odom, and G. E. Scuseria, Nature 366, 665
(1993).
[123] J. Y. Yi and J. Bernholc, J. Chem. Phys. 96, 8634 (1992).
[124] T. Kaxiras and K. C. Pandey, Phys. Rev. Lett. 61, 2693 (1988)).
[125] V. H. Crespi, A. Rubio, and M. L. Cohen, Phys. Rev. Lett. 79 2093 (1998).
[126] B. I. Yacobson, Appl. Phys. Lett. 72, 918 (1998).
[127] M. B. Nardelli, B. I. Yakobson, J. Bernholc, Phys. Rev. B 57 , 4277 (1998).
[128] M. F. Yu, O. Lourier, M. J. Dyer, K. Moloni, T. F. Kelly, and R. S. Ruoff,
Science 287, 637 (2000).
[129] C. M. Lieber, personal communication.
[130] O. Lourie, D. M. Cox, and H. D. Wagner, Phys. Rev. Lett. 81, 1638 (1998).
[131] A. Rubio, J. L. Corkill, and M. L. Cohen, Phys. Rev. B 49, 5081 (1994).
[132] X. Blase, A. Rubio, S. G. Louie, and M. L. Cohen Europhys. Lett. 28, 335
(1994).
[133] N. G. Chopra, R. L. Luyken, K. Cherrey, V. H. Crespi, M. L. Cohen, S. G.
Louie, and A. Zettl Science 269, 966 (1995).
133
[134] Z. Weng-Sieh, K. Cherrey, N. G. Chopra, X. Blase, Y. Miyamoto, A. Rubio, M.
L. Cohen, S. G. Louie, A. Zettl, V. H. Crespi, M. L. Cohen, S. G. Louie, and
R. Gronsky, Phys. Rev. B 51, 11229 (1995).
[135] E. Hernandez, C. Goze, P. Bernier, and A. Rubio, Phys. Rev. Lett 80, 4502
(1998).
[136] N. G. Chopra and A. Zettle, Solid State Commun. 105, 297 (1998).
[137] M. Terrones, W. K. Hsu, H. W. Kroto, and D. R. M. Walton, Top. Curr. Chem
199, 189 (1999).
[138] J. B. Posthill, R. A. Rudder, S. V. Hattangady, G. G. Fountain, and R. J.
Markunas, Appl. Phys. Lett. 56, 734 (1990).
[139] S. S. Iyer, K. Eberl, M. S. Goorsky, F. K. LeGoues, J. C. Tsang, and F. Cardone,
Appl. Phys. Lett. 60, 356 (1992).
[140] J. W. Strane, H. J. Stein, S. R. Lee, S. T. Picraux, J. K. Watanabe, and J. W.
Mayer, J. Appl. Phys. 76, 3656 (1994).
[141] A. A. Demkov and O. F. Sankey, Phys. Rev. B 48, 2207 (1993).
[142] H. Rucher, M. Metthfessel, E. Bugiel, and H.J. Osten, Phys. Rev. Lett. 72,
3578 (1994).
[143] L. Hoffmann, J.C. Bach, and B. B. Nielsen, Phys. Rev. B 55, 11167 (1997).
[144] B. K. Yang, M. Krishnamurthy, and W. H. Weber, J. Appl. Phys. 82, 3287
(1997)
[145] B. K. Yang, M. Krishnamurthy, and W. H. Weber, J. Appl. Phys. 84, 2011
(1998).
134
[146] M. Ohfuti, M. Ikeda, Y. Awano, and N. Yokoyama, Phys. Rev. B60, 13547
(1999).
[147] J. Kouvetakis, D. Nesting, and D. J. Smith, Chem. Mater. 10, 2935 (1998).
[148] J. Kouvetakis, D. Chandrasekhar, and D. J. Smith, Appl. Phys. Lett. 72, 930
(1998).
[149] G. He, M. D. Savellano, and H. A. Atwater, Appl. Phys. Lett. 65, 1159 (1994).
[150] H. Iwano, K. Yoshikawa, A. Kojima, K. Hayashi, S. Zaima, and Y. Yasuda,
Appl. Surf. Sci. 100, 487 (1996).
[151] M. A. Berding, A. Sher, and M. van Schilfgaarde, Phys. Rev. B 56, 3885 (1997).
[152] S. Krishnamurthy, M. A. Berding, A. Sher, M. van Schilfgaarde, and A. B.
Chen, Appl. Phys. Lett. 75, 3153 (1999)
[153] P. Zhang, V. H. Crespi, E. Chang, S. G. Louie, and M. L. Cohen, Nature 409,
69 (2001).
[154] T. Wang, N. Moll, K. Cho, and J. D. Joannopoulos, Phys. Rev. Lett. 82, 3304
(1999).
[155] F. E. Ejeckam, Y. H. Lo, S. Subramanian, H. Q. Hou, and B. E. Hammons,
Appl. Phys. Lett. 70, 1685 (1997).
[156] Y. H. Lo, Appl. Phys. Lett. 59, 2311 (1991).
[157] A. R. Powell,S. S. Iyer, and F. K. LeGoues, Appl. Phys. Lett. 64, 1856 (1994).
[158] T. P. Pearsall, J. Bevk, L. C. Feldman, J. M. Mannaerts, and A. Ourmazd,
Phys. Rev. Lett. 58 729 (1987).
135
[159] M. S. Hybertsen and M. Schluter, Phys. Rev. B 36, 9683 (1987).
[160] S. Froyen, D. M. Wood, and A. Zunger, Phys. Rev. B. 37, 6893 (1988).
[161] E. Kasper, Phys. Scripta T35, 232 (1991).
[162] L. T. Canham, Appl. Phys. Lett. 57, 1046 (1990).
[163] A. G. Cullis and L. T. Canham, Nature 353, 335 (1991).
[164] A. M. Morales and C. M. Lieber, Science 279, 208 (1998)
[165] Y. F. Zhang, Y. H. Tang, N. Wang, D. P. Yu, C. S. Lee, I. Bello, and S. T. Lee,
Appl. Phys. Lett. 72 1835 (1998).
[166] A. J. Read, R. J. Needs, K. J. Nash, L. T. Canhan, P. D. J. Calcott, and A.
Qteish, Phys. Rev. Lett. 69, 1232 (1992).
[167] F. Buda, J. Kohanoff, and M. Parrinello, Phys. Rev. Lett. 69, 1272 (1992).
[168] G. D. Sanders and Y. C. Chang, Phys. Rev. B 45, 9202 (1992).
[169] D. M. Ceperley and B. J. Alder, Phys. Rev. Lett. 45, 566 (1980).
[170] J. P. Perdew and A. Zunger, Phys. Rev. B 23, 5048(1981).
[171] D. R. Hamann, Phys. Rev. B 40, 2980 (1989).
[172] S. G. Louie, S. Froyen, and M. L Cohen, Phys. Rev. B 26, 1738 (1982).
[173] H. J. Monkhorst and J. D. Pack, Phys. Rev. B 13, 5188 (1976).
[174] J. L. Corkill and M. L Cohen, Phys. Rev. B 47, 10304 (1993).
[175] D. W. Jenkins and J. D. Dow, Phys. Rev. B 36, 7994 (1987).
136
[176] K. A. Mader, A. Baldereschi, and H. von Kanel, Solid State Commun. 69, 1123
(1989).
[177] G. He, and H. A. Atwater, Phys. Rev. Lett. 79, 1937 (1997).
[178] K. S. Min and H. A. Atwater, Appl. Phys. Lett. 72, 1884 (1998).
[179] J. Kouvetakis, personal communication.
[180] K. W. Klinkhammer, S. Kuhner, B. Regelmann, and J. Weidlein, Organomet.
Chem. 486, 241 (1995).
[181] H. Schmidbaur and J. Zech, Eur. J. Solid State Inorganic Chem. 29, 5 (1992).
[182] J. Kouvetakis, A. Haaland, D. J. Shokorov, H. V. Volder, G. G. Sirichev, and
P. T. Matsunaga, J. Am. Chem. Soc. 120, 6738 (1998).
[183] For example, H. A. Herman and H. Sitter, in Molecular Beam Epitaxy: Funda-
mentals and Current Status, (Springer-Verlag Berlin, Heidelberg, 1989).
[184] J. Kouvetakis, personal communication.
[185] J. R. Chelikowsky, N. Troullier, and Y. Saad, Phys. Rev. Lett, 72, 1240 (1994).
[186] E. L. Briggs, D. J. Sullivan and J. Bernholc, Phys. Rev. B 52, 5471 (1995).
[187] A. P. Seitsonen, M. J. Puska, and R. M. Nieminen, Phys. Rev. B 51, 14057
(1995).
[188] F. Gygi and G. Galli, Phys. Rev. B 52, R2229 (1995).
[189] E. L. Briggs, D. J. Sullivan and J. Bernholc, Phys. Rev. B 54, 14362 (1996).
[190] G. Zumbach, N. A. Modine, and E. Kaxiras, Solid State Comm. 99, 57 (1996).
[191] N. A. Modine, G. Zumbach, and E. Kaxiras, Phys. Rev. B 55, 10289 (1997).
137
[192] F. Ancilotto, P. Blandin, and F. Toigo, Phys. Rev. B 59, 7868 (1999).
[193] J. Wang and T. L. Beck, J. of Chem. Phys. 112, 9223 (2000).
[194] I. H. Lee, Y. H. Kim and R. M. Martin, Phys. Rev. B 61, 4397 (2000).
[195] J. L. Fattebert, J. Bernholc, Phys. Rev. B 62, 1713 (2000).
[196] B. Hermansson and D. Yevick, Phys. Rev. B 33, 7241 (1986).
[197] S. R. White, J. W. Wilkins, and M. P. Teter, Phys. Rev. B 39, 5819 (1989).
[198] E. Tsuchida and M. Tsukada, Solid State Comm. 94, 5 (1995).
[199] E. Tsuchida and M. Tsukada, Phys. Rev. B 52, 5573 (1995).
[200] E. Tsuchida and M. Tsukada, Phys. Rev. B 54, 7602 (1996).
[201] J. E. Pask, B. M. Klein, C. Y. Fong, and P. A. Sterne, Phys. Rev. B 59, 12352
(1999).
[202] K. Cho, T. Arias, J. Joannospoulos, and P. Lam, Phys. Rev. Lett. 71, 1808
(1993).
[203] S. Wei and M. Y. Chou, Phys. Rev. Lett. 76, 2650 (1996).
[204] C. J. Tymczak and X. O. Wang, Phys. Rev. Lett. 78, 3654 (1997).
[205] R. A. Lippert, T. A. Arias, abd A. Edelman, J. Comput. Phys. 140, 278 (1998).
[206] A. Brandt, S. McCormick, and J. Ruge, J. Sci. Stat. Comput. 4, 244 (1983).
[207] See, for example, S. F. McCormick, ed., Multigrid Methods (Society for Indus-
trial and Applied Mathematics, Philadelphia, 1987).
[208] W. Yang, Phys. Rev. Lett. 66, 1438 (1991).
138
[209] G. Galli and M. Parrinello, Phys. Rev. Lett. 69, 3547 (1992).
[210] X. P. Li, R. W. Nunes, and D. Vanderbilt, Phys. Rev. B 47, 10891 (1993).
[211] M. S. Daw, Phys. Rev. B 47, 10895 (1993).
[212] F. Mauri, G. Galli, and R. Car, Phys. Rev. B 47, 9973 (1993).
[213] F. Mauri and G. Galli, Phys. Rev. B 50, 4316 (1994).
[214] S. Y. Qiu, C. Z. Wang, K. M. Ho, and C. T. Chan, J. Phys.: Condens. Matt.
6, 9153 (1994).
[215] R. W. Nunes and D. Vanderbilt, Phys. Rev. B 51, 17611 (1994).
[216] P. Ordejon, D. Drabold, R. M. Martin, and M. P Grambach, Phys. Rev. B 51,
1456 (1995)
[217] S. Itoh, P. Ordejon, and R. M. Martin, Comput. Phys. Comm. 88, 173 (1995).
[218] A. Cannning, G. Galli, F. Mauri, A. D. Vita, and R. Car, Comput. Phys.
Comm. 94, 89 (1996).
[219] A. D. Daniels, J. M. Millam, and G. E. Scuseria, J. Chem. Phys. 107, 425
(1997).
[220] J. M. Millam and G. E. Scuseria, J. Chem. Phys. 106, 5569 (1997).
[221] S. Goedecker and L. Colombo, Phys. Rev. Lett. 73, 122 (1994).
[222] U. Stephan and D. A. Drabold, Phys. Rev. B 57, 6391 (1998).
[223] S. Goedecker, Rev. Mod. Phys. 71, 1085 (1999).
[224] W. Yang and T. S. Lee, J. Chem. Phys. 103, 5674 (1995).
139
[225] W. Kohn, Phys. Rev. Lett. 76, 3186 (1996).
[226] R. J. Gillespie and R. S. Nyholm, Q. Rev. Chem. Soc. 11, 339 (1957).
[227] R. J. Gillespie, Molecular Geometry, Van Nostrand Reinghold, London, 1972.
[228] S. Ismail-Beigi and T. A. Arias, Phys. Rev. Lett. 82, 2127 (1999).
[229] R. McWeeny, Rev. Mod. Phys. 32, 335 (1960).
[230] J. L. Gazquez and H. J. Silverstone, J. Chem. Phys. 67, 1887 (1977).
[231] M. Friedman, Y. Rosenfeld, A. Rabinovitch, and R. Theiberger, J. Comput.
Phys. 26, 169 (1978).
[232] D. P. Carroll, H. J. Silverstone, and R. M. Metzger, J. Chem. Phys. 71, 4142
(1979).
[233] W. K. Ford and F. S. Levin, Phys. Rev. A 29, 43 (1984).
[234] F. S. Leven and J. Shertzer, Phys. Rev. A 32, 3285 (1985).
[235] W. Schulze and D. Kolb, Chem. Phys. Lett. 122, 271 (1985).
[236] D. Heinemannm D. Kolb, and B. Fricke, Chem. Phys. Lett. 137, 180 (1987).
[237] see, for example, D. Braess, Finite Elements: Theory, Fast Solvers, and Appli-
cations in Solid Mechanics (Cambridge University Press, Cambridge, England,
1997).
[238] See, for example, J. M. Ziman, Principles of The Theory of Solids (Cambridge
University Press, Cambridge, England, 1972).
[239] See, for example, G. H. Golub and C. F. Van Loan, Matrix Computations (The
Johns Hopkins University Press, Baltimore, 1996).
140
[240] H. Yang, Computing 51, 79 (1993).
[241] W. W. Bradbury and R. Fletcher, Numer. Math. 9, 259 (1966).
[242] E. R. Davidson, J. Comp. Phys. 17, 87 (1975).
[243] S. Costiner and S. Ta’asan, Phys. Rev. E 51, 3704 (1995).
[244] S. Costiner and S. Ta’asan, Phys. Rev. E 52, 1181 (1995).
[245] T. L. Beck, Rev. Mod. Phys. 72 1041 (2000).
141
Vita
EducationSummer, 2001 : Ph. D., Condensed Matter Physics
Ph. D. Minor, High Performance ComputingThe Pennsylvania State University
Spring, 1996 : M.S., Atomic & Molecular Physics, Institute of Physics,Chinese Academy of Sciences, China
Spring, 1993 : B.S., Physics, Xiamen (Amoy) University, China
Awards1998— 1999 : Braddocks fellowship, The Pennsylvania State University
2000— 2001 : Braddocks fellowship, The Pennsylvania State University
2001 : The Xerox award for outstanding Ph. D. thesis at PSU in the area ofmaterials
Selected Publications
1 P. Zhang, V. H. Crespi, E. Chang, S. G. Louie, and M. L. Cohen, ComputationalDesign of Direct Bandgap Semiconductors That Lattice-Match Silicon, Nature 409,69 (2001).
2 P. Zhang and V.H. Crespi, Plastic Deformation of Boron Nitride Nanotubes: AnUnexpected Weakness, Phys. Rev. B. 62, 11050 (2000).
3 P. Lammert, P. Zhang and V.H. Crespi, Gapping by Squashing: Metal-InsulatorTransitions in Collapsed Carbon Nanotubes, Phys. Rev. Lett. 84, 2453 (2000).
4 P. Zhang and V.H. Crespi, Nucleation of Carbon Nanotubes without PentagonalRings, Phys. Rev. Lett. 83, 1791 (1999).
5 P. Zhang, P. Lammert, and V.H. Crespi, Plastic Deformations of Carbon Nan-otubes, Phys. Rev. Lett. 81, 5346 (1998).
6 J.-M. Li, P. Zhang, Y. Yang, and L. Lei, Theoretical Study of Adatom Self-diffussion on Metallic fcc001 Surfaces, Chinese Phys. Lett.14, 768 (1997).
7 Y. Zhang, P. Zhang, and J.-M. Li, Near-threshold Structure in Inner-shell Pho-toabsorption Process of N2 and CO, Phys. Rev. A 56, 1819 (1997).
8 P. Zhang and J.-M. Li, Geometry and Electronic Structure of Na3, Acta Phys.Sin. 46, 870 (1997).
9 P. Zhang and J.-M Li, Theoretical Studies of Electronic Excited States for Na3,Phys. Rev. A 54, 665 (1996).
10 J. Yan, P. Zhang, and J.-M Li, Fine Structure Inversion in f Channel of AlkaliAtoms, Acta phys. Sin. (1996).