Research Collection
Doctoral Thesis
Ab-initio quantum transport simulations with tight-binding-like Hamiltonians
Author(s): Stieger, Christian
Publication Date: 2019
Permanent Link: https://doi.org/10.3929/ethz-b-000418185
Rights / License: In Copyright - Non-Commercial Use Permitted
This page was generated automatically upon download from the ETH Zurich Research Collection. For more information please consult the Terms of use.
ETH Library Ab-initio quantum transport simulations with tight- binding-like Hamiltonians
Diss. ETH No. 26312
Ab-initio quantum transport simulations with tight-binding-like Hamiltonians
A dissertation submitted to ETH ZURICH
for the degree of Doctor of Sciences
presented by CHRISTIAN STIEGER Master of Physics, ETH born June 3rd, 1983 citizen of Wallisellen ZH, Switzerland
accepted on the recommendation of Prof. Dr. Mathieu Luisier, examiner Prof. Dr. Jim Greer, co-examiner
2019
Acknowledgements
First of all, I’d like to express my sincere gratitude to Prof. Dr. Math- ieu Luisier for the opportunity to work in his group at the Integrated Systems Laboratory (IIS) and for his support and guidance. I am especially thankful for his thorough revision of this thesis. I would also like to thank Prof. Dr. Jim Greer for accepting co-examination. I am thankful to all the members at the Nano-TCAD group for their company and support. Special thanks go to Dr. Petr A. Khomyakov for his valuable advice, Dr. Hamilton Carrillo Nunez and Cedric Klinkert for testing Winterface and supplying feedback, and to Hansj¨org Gisler for keeping the coffee machine in working order at all times.
v
Abstract
In this thesis a framework for quantum transport simulation from first principles is introduced, focusing on coherent electronic transport, but discussing the importance of electron-phonon interactions for transi- tion metal dichalcogenides (TMDs) as well. The transport model is based on the non-equilibrium Green’s function (NEGF) formalism which requires a localized basis set. To obtain a suitable Hamilto- nian matrix for a given device structure, a representative unit cell is first identified and its electronic properties calculated with density- functional theory (DFT) expressed in a plane-wave basis. These re- sults must be transformed into a localized basis set using so-called maximally localized Wannier functions (MLWFs). From the MLWF representation of the original unit cell, the device Hamiltonian can be constructed with the help of properly designed upscaling techniques. The main focus of this thesis lies in the interfacing part of the MLWF representation with a quantum transport code, OMEN in our case. To this end, a code called Winterface was developed. Elaborations of its concepts, algorithms, and general functionality are presented on the basis of a molybdenum disulfide (MoS2) monolayer structure, as well as heterostructures involving tungsten disulfide (WS2). The process of Hamiltonian upscaling was simplified through the application of a two-stage process, where the first step consists of a conversion of the MLWF Hamiltonian data into blocks corresponding to interactions along atomic bonds. This is the key innovation of this work. This does not only make the actual upscaling procedure, which represents the second step, more transparent, as opposed to an approach rely- ing on raw MLWF data, but also allows for an investigation of the local properties of super structures. The latter can help shed light
vii viii ABSTRACT onto the influence of local features such as interface regions. On the basis of this, more complex device geometries can be generated and the corresponding Hamiltonians be approximately constructed. Ex- amples of such super structures are provided with MoS2 on top of WS2 or both 2-D materials next to each other. The developed ap- proach can be considered completely general, restricted only by the capability of the user to accurately model the developed physics in DFT and to ”wannierize” plane-wave results, as well as the imposed computational intensity.
Finally, the importance of the inclusion of electron-phonon inter- actions into the transport model is demonstrated on the basis of self- heating effects in MoS2, WS2, and black phosphorous devices. Due to the versatiliy of the NEGF formalism, this can be done conveniently through additional scattering self-energy terms. The interactions for the phonon part can also be derived from first-principles using density functional perturbation theory (DFPT). The electronic part results from multiple wannierizations using the concepts already introduced for the coherent case. Dissipative calculations present themselves as natural extensions of ballistic ones, but with a significant increase in computational burden. Zusammenfassung
In dieser Arbeit wird eine Umgebung f¨urTranportsimulationen auf quantenmechanischer Ebene pr¨asentiert, wobei der Fokus auf elektro- nischem Transport liegt. Allerdings wird auch auf die Wichtigkeit von Elektron-Phonon Wechselwirkung in Ubergangsmetall-Dichalcogenen¨ (TMDs) eingegangen. Das Transportmodell ist dabei basierend auf dem nichtgleichgewichts Greens Funktionen (NEGF) Formalismus, der eine lokalisierte Basis ben¨otigt. Um eine Hamiltonianmatrix f¨ur eine spezifische Kanalregion eines Transistors zu bekommen, muss zuerst eine repr¨asentative Einheitszelle gew¨ahltwerden, deren elek- tronische Eigenschaften anschliessend mittels Dichtefunktionaltheorie (DFT) berechnet werden k¨onnen.Da daf¨uraber eine Basis aus ebe- nen Wellen verwendet wird, welche nicht lokalisiert ist, muss zuerst eine Transformation zu einer Basis so genannter maximallokalisierten Wannier Funktionen (MLWFs) vorgenommen werden. Von dieser aus- gehend, kann auf Grundlage der primitiven Einheitszelle und speziell daf¨urentwickelten Hochskalierungstechniken eine Hamiltonianmatrix generiert werden, die die gew¨unschte Struktur beschreibt. Der Fokus dieser Arbeit liegt daher auf der Schnittstelle zwischen der Repr¨asen- tation durch MLWF und einem Transportsimulator, in unserem Fall OMEN. Zu diesem Zweck wurde ein Programm namens Winterface entwickelt, dessen Konzepte, Algorithmen und generelle Funktion- alit¨atam Beispiel einer Einzelschicht von Molybd¨andisulfid (MoS2), sowie Heterostrukturen des letzteren mit Wolframdisulfid (WS2), erkl¨art werden. Der Prozess der Hochskalierung von Hamiltonianmatrizen konnte mittels einer Zerlegung in zwei Schritte vereinfacht werden. Im ersten Schritt wird die Konvertierung einer Hamiltonianmatrix in MLWF Repr¨asentation in eine andere Repr¨asentation basierend auf
ix x ZUSAMMENFASSUNG
Wechselwirkungen entlang Bindungen zwischen Atomen vorgenom- men. Hierbei handelt es sich um die haupts¨achliche Innovation dieser Arbeit, da dadurch nicht nur der zweite Schritt, die Hochskalierung von Hamiltonianmatrizen, massiv vereinfacht wird, sondern zus¨atzlich auch Untersuchungen von lokalen Eigenschaften gr¨oererStrukturen erm¨oglicht werden, womit der Einfluss lokaler Eigenschaften auf ihre Umgebung abgesch¨atztwerden kann. Auf dieser Grundlage k¨onnen komplexe Geometrien verwirklicht und entsprechende Hamiltonian- matrizen approximativ generiert werden. Beispiele werden in Form von Heterostrukturen von einer MoS2 Schicht auf einer aus WS2, sowie mit beiden Materialen in derselben Schicht, pr¨asentiert. Die entwickelte Methodik kann als allgemein anwendbar betrachtet wer- den, eingeschr¨anktnur durch die N¨aherungenin DFT, sowie der da- rauffolgenden Wannierisierung und der resultierenden rechnerischen Belastung.
Schlussendlich wird die Wichtigkeit der Ber¨ucksichtigung von Elek- tron-Phonon Wechselwirkungen in das Transportmodell anhand von Beispielen bestehend aus MoS2, WS2 und schwarzem Phosphor, disku- tiert. Aufgrund der Erweiterbarkeit des NEGF Formalismusses, k¨onnen diese Effekte leicht ¨uber zus¨atzliche Streuungsterme einbezogen wer- den. Die phononischen Wechselwirkungen k¨onnen¨uber Dichtefunk- tionalst¨orungstheorie(DFPT) erlangt werden und der elektronische Teil aus mehreren Wannierisierungen, wozu die bisher entwickelten Werkzeuge ausreichend sind. Dissipative Berechnungen stellen daher eine nat¨urliche Erweiterung des koherenten Modells dar, allerdings auf Kosten einer signifikant erh¨ohten Anforderung an die Rechenleistung. Contents
Acknowledgements v
Abstract vii
Zusammenfassung ix
1 Introduction 1
2 Quantum Transport From First-Principles 5 2.1 Density-Functional Theory ...... 5 2.2 Maximally Localized Wannier Functions ...... 9 2.3 Transport in the Wavefunction Formalism ...... 12 2.4 Transport in the NEGF Formalism ...... 15 2.5 Description of the Toolchain ...... 17
3 Hamiltonian Data in Terms of Bonds 23 3.1 Atomic Lattices and Unit Cells ...... 23 3.1.1 Unit Cell, Definition and Example ...... 23 3.1.2 Basis Expansions ...... 25 3.1.3 Metric in Periodic Space ...... 28 3.1.4 Automatic Detection of Basis Expansions . . . 29 3.2 Upscaling Technique for an Ideal Case ...... 31 3.3 Generation of Bond Interactions ...... 33 3.3.1 Matching Wannier Centers to Atomic Positions 33 3.3.2 Generating Interaction Data Along Atomic Bonds 39
xi xii CONTENTS
4 Generating Hamiltonian Matrices 49 4.1 Exact Upscaling Technique and Bandstructure Calcu- lations ...... 49 4.2 Interface with OMEN ...... 58 4.2.1 Wannierization Process ...... 61 4.2.2 Error Estimation ...... 63 4.3 Approximate Upscaling Technique and Local Bandstruc- tures ...... 67 4.3.1 Bond Index Substitution ...... 69 4.3.2 Combining Multiple Wannierizations ...... 74 4.4 Results ...... 83
5 Dissipative Transport 91 5.1 Transport Model ...... 91 5.1.1 Hamiltonian and Dynamical Matrix ...... 91 5.1.2 Electron and Phonon Coupling ...... 96 5.1.3 Calculation of Observables ...... 98 5.2 Device Results ...... 100
6 Structure of the Code 111 6.1 Libmat ...... 111 6.2 Liblat ...... 114
7 Conclusion and Outlook 117
A 2-D vs. 3-D k-point Grid in Monolayer MoS2 121
B Basis Set Reduction after Wannierization 125
List of Publications 127
Curriculum Vitae 141 Chapter 1
Introduction
The inevitable end of Moore’s scaling law [1] calls for novel transistor concepts that can deliver reliable logic performance in future ultra- scaled technology nodes. Although the semiconductor industry has already moved to three-dimensional FinFETs [2], innovations at the architecture and material levels will be required for next-generation devices. Besides the widely studied gate-all-around (GAA) nanowire (NW) [3, 4, 5] and ultra-thin-body [6] field-effect transistors, recent years have seen the emergence of a new class of two-dimensional (2- D) materials consisting of atomically thin layers connected by van der Waals forces. Its first and probably most famous member is graphene, which was discovered by Novoselov et. al. in 2005 [7]. Despite im- pressive carrier mobility values (> 100,000 V/ms), graphene does not lend itself to logic applications due to the lack of a band gap. How- ever, the available design space for 2-D materials is huge, according to recent theoretical investigations [8]. Currently, a strong accent is set on transition metal dichalcogenides (TMD), such as MoS2, which appear more promising as future channel materials than graphene. A transistor made of a single monolayer of MoS2 was experimentally re- alized in 2011 [9]. Applications involving few-layer of heterostructures of 2-D materials have already been demonstrated, e.g. light-emitting diodes [10], photodetectors [11, 12, 13], memory cells [14] or memris- tors [15]. Among the 2-D materials that have received wide attention, black phosphorous (BP) stands out [16, 17]. As compared to TMDs,
1 2 CHAPTER 1. INTRODUCTION
BPs exhibit highly anisotropic electrical and thermal properties [18], which could pave the way for other original applications.
At this point, it is not clear whether 2-D materials can compete with other technologies and if so, which component or heterostruc- ture is the most suitable at performing a given task. Whilst such a large number of possible configurations offers exciting opportunities in terms of novel device concepts, it also requires improved solutions, both experimental and theoretical, to explore the available design space. Technology computer aided design (TCAD) represents a pow- erful and well-established approach to address this challenge. Thanks to its cost- and time-effectiveness, TCAD can help experimentalists to rapidly converge towards the most promising contenders. However, at the current nanometer scale of the transistor dimensions, classical and semi-classical simulation methods such as the drift-diffusion or Boltzmann transport equations should be replaced by a quantum me- chanical treatment of the device properties. To describe the electronic structure of 2-D materials, different methods exist, from the most fun- damental ones, such as density functional theory (DFT) [19], up to empirical ones, for example tight-binding (TB) [20]. When coupled with a quantum transport solver, both approaches suffer from their own limitations, such as the size of the system that can be handled for DFT, or the need to create physically meaningful parameters sets for TB. As a compromise between them, a maximally localized Wan- nier function (MLWF) representation of the DFT results can be used [21, 22]. This can be seen as a first step towards ab initio device in- vestigations.
The idea consists of identifying a primitive unit cell that is repre- sentative of the system of interest, perform a DFT calculation of it, convert the plane-wave results into a set of MLWF, and finally scale up the obtained TB-like Hamiltonian matrix to the size of the considered device. Once the general concept is established, what remains to be developed is a technique capable of providing the necessary through- put to screen the large design space of 2-D materials. This thesis aims at providing a toolbox, called Winterface, to automatically up- scale the MLWF representation of comparatively small unit cells to desired, sometimes complex, geometries, e.g. heterostructures. Since 3 the problem to be solved depends on numerical outputs and cannot be directly formulated in terms of equations, practical examples will be provided to demonstrate the principle of the proposed approach, identify its shortcomings, and give evidence of its utility to simulate nano-devices based on 2-D materials. The atomic structures selected as testbeds will ideally illustrate the key concepts that have been implemented. Although the physics of these examples might be in- teresting, this is not the selection criterion that was applied. Here, what matters is the robustness and versatility of the created Win- terface code. So far, hundreds of material configurations have been constructed and the transport properties of all of them could be suc- cessfully investigated, thus demonstrating the quasi-universality of the methodology. However, at this point, a complete automation was not possible, because critical components are missing, in particular an initial guess projection for the MLWF [23]. Additionally, acceptable approximations depend on a compromise between computational bur- den and physical accuracy, which have to be prioritized by the user of the upscaling algorithms available in Winterface.
While the main focus of this work is on coherent transport, it is expected that thermally-induced phenomena are considered as well: they are expected to play a significant role in ultra-scaled structures due to the close proximity of electrons and phonons and their increased coupling [24]. Few studies mainly concerned with graphene [25, 26, 27], but also molybdenum disulfide [28, 29] and black phosphorous [30] have discussed these issues. Still, it can be generally said that the thermal behavior of 2-D materials is not completely understood yet. Another reason for the inclusion of a dissipative scattering mecha- nism is the presence of a non-physical negative differential resistance in the ballistic simulation of several TMDs [31]. Such a behaviour has never been experimentally observed at room temperature [32]. The NDR originates from the TMD bandstructure, which exhibits several narrow energy bands that cannot carry any current if the elec- trostatic potential undergoes large variations between the source and drain contacts [33]. It is therefore an artifact of the ballistic approx- imation. Accounting for electron-phonon scattering helps get rid of NDR by connecting bands that would otherwise remain independent from each other [34]. Depending on the material under investigation, 4 CHAPTER 1. INTRODUCTION scattering is therefore required for accurate modeling of the current vs. voltage characteristics.
The outline of the thesis is as follows: in Chapter 2, the toolchain mentioned above, and its constituents are introduced. Then, in Chap- ter 3 the notion of upscaling is explained on the basis of an ideal ex- ample. Subsequently, the first step of a general approach to upscaling, the generation of interactions along atomic bonds, is presented. This preliminary work greatly simplifies the actual process of interfacing plane-wave DFT results to the quantum transport code OMEN [35], and allows for additional approximate upscaling techniques dedicated to heterostructures, as discussed in Chapter 4. In Chapter 5 the afore- mentioned importance of self-heating effects is elaborated in detail for single-gate devices with a molydenum disulfide monolayer, tungsten disulfide, and black phosphorous as channel material. In Chapter 6, the code structure behind Winterface is briefly discussed and finally, the thesis is concluded and an outlook is provided in Chapter 7. Chapter 2
Quantum Transport From First-Principles
2.1 Density-Functional Theory
Semiconductor and other solid-state devices have long been modeled using classical theories such as the drift-diffusion equations. However, as the channel length of modern transistors is reaching 10nm and be- low, a full quantum mechanical treatment has become unavoidable. Condensed matter (solid bodies) consists of atomic nuclei (ions), usu- ally arranged in an elastic lattice, and of electrons. To capture their properties, the Schr¨odingerequation must be solved. Restricting our- selves to the stationary problem of N charged particles, we must deal with a 3N dimensional eigenvalue problem of the form
N h 2 X i − ~ ∇2 +V (r , .., r ) Ψ (r , ..., r ) = EΨ (r , ..., r ), (2.1) 2m i 1 N e 1 N e 1 N 0 i
where Ψe(r1, ..., rN ) is the many-particle electronic wavefunction of N electrons situated at positions r1 to rN , m0 the rest mass of the electron, E the total energy, and V (r1, .., rN ) is a many-body term including all electron-electron and electron-ion interactions. Solving
5 6 CHAPTER 2. QUANTUM TRANSPORT this problem numerically is unfeasible, even for small systems, due to the exponential scaling of its complexity with the number of particles. For example, a system of just 3 particles solved on a grid of 10 points along each direction results in a matrix of size (10 · 10 · 10)(3·3) = 1027 to be diagonalized. Without approximations, any attempt at dealing with this problem is hopeless. The first step is the Born-Oppenheimer approximation [36], where the assumption is made that the motion of atomic nuclei and electrons can be separated. This simplification was already applied to Eq. (2.1), the nucleus wavefunction Ψn(R1, ..., RN ) having been removed. This allows for the overall numerical problem to be solved in two less complicated steps. Usually, the motion of atomic nuclei, along with core electrons in closed shells, is handled partly classically, while the remaining electrons receive a full quantum- mechanical treatment. Even with this approximation, the complexity of Eq. (2.1) needs to be further reduced by introducing single-particle wavefunctions φi(rj), which interact only with the mean-field of their surroundings. Here, each electron moves in a slightly different electro- static potential, resulting from all charges in the system. According to the Hartree-Fock method [37], the many-electron wavefunction can be recovered from the single-electron wavefunctions by means of the Slater determinant
φ1(r1) . . . φn(r1) 1 . . Ψe(r1, ..., rn) = √ det . . . (2.2) N! . . φ1(rn) . . . φn(rn) A popular alternative is a formalism called density-functional the- ory [19] (DFT), where the many-electron problem is formulated in terms of the spatial electron density
Z Z ∗ n(r) = N dr2... drN Ψe(r, r2, ..., rn)Ψe(r, r2, ..., rn), (2.3)
which is given as a functional of the many-electron wavefunction Ψe(r1, ..., rn). Similar to the Hartree-Fock method, a system of fic- titious single-particle wavefunctions is introduced in such a way that the charge density of this simplified system given as 2.1. DENSITY-FUNCTIONAL THEORY 7
n X 2 n(r) = |φi(r)| (2.4) i is equivalent to that of the original many-body problem. It can be shown that such a system is governed by the Kohn-Sham equation:
2 − ~ ∇2 + V (r) φ (r) = ε φ (r). (2.5) 2m s i i i
In Eq. (2.5), the φi(r)’s are the independent particle Kohn-Sham orbitals and Vs(r) the single-particle effective potential given by
Z 0 0 n(r ) Vs(r) = V (r) + dr 0 + Vxc(r), (2.6) |{z} |r − r | | {z } (1) | {z } (3) (2) which is the sum of the external (1), the Hartree (2), and the ex- change (3) potentials. The latter is defined as the functional derivative of the exchange correlation energy
δE [n(r)] V (r) = xc . (2.7) xc δn(r) By definition, the exchange correlation energy is the difference between the total energy and the quantities that can be computed exactly. Contributions from the Pauli exclusion principle and many- body Coulomb interactions must be included, but the exact form of the exchange-correlation functional is not known. It can be approxi- mated for the case of slowly varying densities as Z LDA Exc [n↑(r), n↓(r))] = drn(r)xc[n↑(r), n↓(r)] (2.8)
with the total electron density n(r) = n↑(r) + n↓(r) as the sum of the electron spin densities and xc[n↑(r), n↓(r)] the exchange-correlation energy per particle of a homogeneous electron gas of charge densitiy n(r). This form of Exc is known as the local density approximation (LDA), which has been successfully applied to many quantum chem- istry problems, but it is known to underestimate the band gap of semiconductors by as much as 50%. Improvements have been made 8 CHAPTER 2. QUANTUM TRANSPORT by taking into account the gradient of the electron density in the so-called generalized gradient approximation (GGA) whose most es- tablished form was parametrized by Perdew, Burke, and Ernzerhof [38] (PBE). In GGA, the exchange-correlation functional is written as
Z GGA Exc [n↑(r), n↓(r)] = drf(n↑(r), n↓(r), ∇n↑(r), ∇n↓(r)), (2.9)
∇n↑ being the derivative of the charge density at position r. Fur- ther improvements have been made in the form of a hybrid approach combining the GGA and the Hartree-Fock exchange correlation by Heyd, Scuseria, and Ernzerhof [39]. Using it, it was shown that the lattice constant and band gap of most semiconductors can be repro- duced with a high degree of accuracy [40], but at the cost of a signif- icant increase in computational demand.
Since the single-particle effective potential in Eq. (2.6) depends on the charge density, Eqs. (2.5-2.7) form a self-consistent system of equations, which must be solved iteratively until convergence is achieved. Finally, the total energy E[n] of the system is given by
Z 1 Z Z n(r)n(r0) E[n] = drV (r)n(r) + dr dr0 + T [n] + E [n], 2 |r − r0| s xc (2.10)
where Ts is the kinetic part of the non-interacting system.
Density-Functional Theory is a very efficient formalism scaling as O(N 3). Approximations enter the formalism only through the exchange-correlation energy. Simulations of systems up to 1000 atoms are nowadays pretty much standard. Furthermore, as DFT does not depend on empirical parameters, it can be used to generate mean- ingful predictions for novel materials from first-principles. As such, it is better suited than empirical models such as tight-binding [41], when it comes to the investigation of heterostructures [42, 43], metal- insulator-metal junctions [44], or novel 2-D materials [8]. 2.2. MLWF 9
2.2 Maximally Localized Wannier Func- tions
The physics of a (perfect) lattice can be encapsulated entirely within one unit cell, usually, but not necessarily a primitive cell. The Hamil- tonian operator for such a system is then composed of a kinetic part and a periodic potential. This is the situation present in Eq. (2.5), and thus the Kohn-Sham orbitals form a system of orthonormal eigen- functions known as Bloch waves. They take the form:
ikr ψn,k(r) = un,k(r) · e , (2.11)
where un,k(r) is a periodic function in real-space with band in- dex n and wave vector k. A Bloch wave is periodic in real-space up to a phase factor eikR where R is a vector pointing to a neighbor- ing unit cell. This representation in terms of energy eigenfunctions is very convenient for many applications because the Hamiltonian ma- ∗ trix elements obey Hmn(k) = hψn,k|H|ψm,ki = mnδmn. However, if we want to find a Hamiltonian operator localized in a super cell, this plane-wave representation is no more convenient as Bloch waves extend over all space.
An alternative representation that is centered and localized in real- space and that can be translated along periodic images of the unit cell is better suited for many applications. Bloch waves can be trans- formed into such a representation through a process called wannieriza- tion [21] by which periodic wave functions are transformed into a set of orthonormal basis functions called Wannier functions, which have the desired localized properties. The process is essentially a Fourier transform where the fact that Bloch waves are defined only upto a phase factor eiφ(k) offers a large degree of freedom, which can be cast (k) into a unitary matrix Umn. Wannier functions are defined as:
V Z X w (r) = d3k U (k)ψ (r) e−ikR. (2.12) nR (2π)3 mn mk BZ m In Eq. (2.12), the set of momentum(k)-dependent wave functions ψmk(r) of band index m are replaced through a unitary transform by 10 CHAPTER 2. QUANTUM TRANSPORT
a set of Wannier functions wnR(r) with Wannier index n and assigned to the unit cell situated at vector R with respect to the origin. The integration is performed over the entire Brillouin zone and V is the volume of the atomic unit cell in real space. For the case where the (k) unitary matrix Umn is chosen such that the spread functional
X 2 2 Ω = hwn0(r)|r |wn0(r)i−|hwn0(r)|r|wn0(r)i| (2.13) n
is minimal, we speak of Maximally Localized Wannier Functions (MLWF). In this configuration the Wannier functions themselves as well as the Hamiltonian matrix elements can be proved to be real. A comparison of Bloch waves and MLWF is presented in Fig. 2.1.
The translation symmetry of the lattice is expressed through R vectors, each of them corresponds to a translated (with respect to a reference unit cell), but otherwise identical Wannier function placed at position R. Due to the spacial localization, pairwise interactions among Wannier functions extend only over a finite subset of R vectors, where the origin may be set arbitrarily. This localization in space is achieved at the cost of localization in energy, i.e. Wannier functions form an orthonormal basis set, but they are not eigenfunctions of the Hamiltonian operator. Therefore, the Wannier index n is not a band index and in general the whole set of Wannier functions contributes to each band. The Hamiltonian operator can be expressed in terms of pairwise interactions among Wannier functions among a finite range of unit cells described by R vectors, i.e.
Hnm(R) = hwn0|H|wmRi . (2.14)
Each H(R) thus describes the interactions of the atoms in the unit cell shifted according to R with the atoms in the home cell at R = 0. This spatially localized representation has ”tight-binding-like” characteristics, as needed for upscaling a set of Hamiltonian matrices H(R) defined on a cluster of unit cells to a set of Hamiltonian matrices H˜(R˜ ) describing interactions among a cluster of super cells. Note that what we call upscaling is the transformation that allows us to go from 2.2. MLWF 11
(a) (b) (c)
Figure 2.1: Illustration of plane-wave and Wannier representations of a quantum mechanical ground state on a one-dimensional grid of five unit cells. The parameter R gives the position of each cell relative to the center and the black dots represent atomic positions. (a) Cell- periodic part of a Bloch wave un,k(r). (b) Corresponding Bloch wave ikr ψn,k(r) = un,k(r)·e . (c) Example of a maximally localized Wannier function (MLWF). This is a real function localized in real space and centered on a specific site. Here, the solid line represents the MLWF centered at R = 0, whereas the dashed lines refer to equivalent images centered at R 6= 0, thus illustrating the periodicity of the crystal in the Wannier picture. 12 CHAPTER 2. QUANTUM TRANSPORT the DFT calculation of a small unit cell to the production of a large- scale Hamiltonian matrix corresponding to a device structure. This concept will be explained in detail in Chapters 3 and 4.
2.3 Transport in the Wavefunction For- malism
Density-functional theory, when expressed in a plane-wave basis, lends itself perfectly to electronic structure calculations of periodic struc- tures or small molecules. The purpose of this thesis is however to eval- uate the ”current vs. voltage” characteristics of nano-devices. To do that, the atomic system of interest must be driven out-of-equilibrium by an external voltage source. Such situations can be realized by at- taching reservoirs to the simulation domain, from which electrons can be injected and collected. Furthermore, to allow for the definition of the required open boundary conditions (OBCs) [45, 46], the Hamil- tonian that describes the electronic properties of the device must be expressed in a localized basis. Here, the choice is that of MLWF, as explained in the previous Section. The Schr¨odingerequation with OBCs takes the form
h RB i C C E − HMLW F (kt) − Σ (E, kt) Ψ (E, kt) = S (E, kt). (2.15)
In the resulting Wavefunction formalism, also called Quantum Transmitting Boundary Method (QTBM), E is a diagonal matrix con- taining the energy of the injected electrons, HMLW F (kt) describes the Hamiltonian matrix of the isolated system in a MLWF basis, and kt is the electron momentum modeling the directions assumed periodic. RB Σ (E, kt) is the boundary self-energy matrix, S(E, kt) the corre- C sponding injection vector, and Ψ (E, kt) the wavefunction injected from contact C ∈ {L, R} at energy E and momentum kt. It contains entries for each atomic location and Wannier function. For devices with two contacts, refered to as L (left) and R (right), injections of electrons from the left (right) contact correspond to the state ΨL R L R RB (Ψ ) and the injection vector S (S ). Note that Σ (E, kt) can be computed iteratively [47], from (generalized) eigenvalue problems 2.3. WAVEFUNCTION FORMALISM 13
[48, 49, 50] or from contour integral techniques [51].
Eq. (2.15) is no longer an eigenvalue problem, but a sparse linear system of equations that must be solved for each injection energy E and transverse momentum kt. For the calculation of observables, both quantities must be discretized in such a way that integrations over them capture the energy- and momentum-dependent behavior C C,σ of Ψ (E, kt). Assuming that Ψ (E, kt, ri) is the wavefunction at positions ri for Wannier function σ, the density-of-states of contact C is defined as
1 X X dE −1 gC (r ,E, k ) = |ΨC,σ(E, k , r )|2 , (2.16) i t 2π n t i dkC σ n n
C where kn refers to the momentum along the transport direction of the nth state injected from contact C at energy E and transverse momentum kt. The electron density n(ri) can then be calculated as
Z ∞ X X C C n(ri) = g (ri,E, kt) · f(E,EF ), (2.17) CBMC C kt
where the integral is performed from the conduction band mini- C mum (CBM) of each contact C and f(E,EF ) is the Fermi distribution function of the electrons in equilibrium in each contact. The energy C EF refers to the Fermi level of contact C.
The Schr¨odingerequation in Eq. (2.15) contains an external po- tential, whose Hartree component VH (r) must be self-consistently cal- culated with the Poisson equation since the charge in the device gives rise to an electrostatic potential:
ρ(r) ∇2V (r) = − , (2.18) H ε(r) where the position-dependent charge density may include several components such as the acceptor and donor concentrations as well as the electrons and hole densities. In the following, for simplicity, we 14 CHAPTER 2. QUANTUM TRANSPORT restrict ourselves to the presence of electrons only (n(r)) and donors (ND(r)). Since the charge density n(r) depends on VH (r) through the Schr¨odingerequation and VH (r) on n(r) through the Poisson equa- tion, the circular dependencies must be resolved self-consistently until convergence is reached. Once the out-of-equilibrium state of the sys- tem has been determined in this way, the famous Landauer-B¨uttiker formula [52, 53] can be used to produce the net ballistic current flow- ing through the device from the left to the right contact or vice-versa:
Z ∞ e X dE σ L R I = − T (E, kt) · f(E,EF ) − f(E,EF ) . (2.19) ~ CBM L 2π kt,σ
σ Here kt is the transverse momentum, σ the spin, and T (E, kt) the transmission function per spin, defined as
σ,L 2 |Ψ L (E, kt, xout)| −1 σ X km dE dE T (E, kt) = , (2.20) σ,L 2 R L |Ψ L (E, kt, xin)| dkn dkm n,m km where the indices run over all propagating modes in each contact σ,L and Ψ L (E, kt, xout/in) is the spin-dependent wavefunction injected km from contact L and evaluated at the first (xin) and last (xout) point of the device along the transport direction. Coherent current through a L R device from left to right is observed if f(E,EF ) 6= f(E,EF ), i.e. if a bias is applied, and propagating modes are available in both contacts.
If the mean free path of electrons exceeds the dimensions of the device, the ballistic approach of single-particle wavefunctions is suffi- cient. However, for cases where scattering mechanisms such as electron- phonon interactions must be included for accurate physical model- ing, a different approach relying on Green’s functions and called the non-equilibrium Green’s function (NEGF) [54, 55] formalism is called for. The great advantage of NEGF over the Wavefunction formalism is that many-body effects can be introduced in a perturbative way through scattering self-energy terms [56, 57]. However, for the mo- ment we will focus on the ballistic NEGF case in the next Section and introduce the dissipative case later in Chapter 5. 2.4. NEGF FORMALISM 15
2.4 Transport in the NEGF Formalism
The Green’s function associated with the linear operator L(r) is de- fined as
L(r)G(r, r0) = δ(r − r0), (2.21) where δ(r−r0) is the Dirac-delta function. Once the Green’s func- tion is known, inhomogeneous differential equations of the form
L(r)f(r) = s(r) (2.22) can be solved as Z f(r) = dr0G(r, r0)s(r0). (2.23)
Eq. (2.21) for the operator in Eq. (2.15) then becomes