Master Thesis
Quantum transport calculations using wave function propagation and the Kubo formula
Troels Markussen (s001477) [email protected]
Supervisors: Mads Brandbyge and Antti-Pekka Jauho
MIC – Department of Micro and Nanotechnology Technical University of Denmark
9th March 2006
Preface
This thesis is submitted in candidacy for the Master of Science degree in Engi- neering Physics from the Technical University of Denmark (DTU). The project has been carried out at MIC - Department of Micro and Nanotechnology in the period from February 1st 2005 to February 1st 2006 with Mads Brandbyge and Antti-Pekka Jauho as supervisors. I would like to thank my supervisors for their enthusiasm and constructive criticism. I have enjoyed the many fruitful discussions and in general working in the Theoretical Nanotechnology group. In particular I thank Magnus Paulson for help with obscure Linux problems and Python programming. I would like to thank Oticon Fonden for their generous financial support. This enabled me to visit N. Lorente and R. Rurali in Toulouse, France, to discuss collaborative work on the metallic silicon nanowires and learn about the siesta- program. The grant also made it possible to participate in the conference Elec- Mol’05 in Grenoble, France, where some of the results from the project was presented on a poster. I thank the people at MIC for making it a pleasant place to spend a lot of time. I also thank the guys in room 119 for fruitful and indispensable discussions concerning anything but my project. Finally, I thank my girlfriend Wiebke for her patience and support during the project, and for reminding me that there are other things in life than physics.
Troels Markussen MIC – Department of Micro and Nanotechnology Technical University of Denmark 9th March 2006
Abstract
This project is concerned with modelling the electrical properties of silicon nanowires. The fabrication of nanowires can be controlled with high accuracy, making them very promising candidates for future electronic components with the capability of exceeding conventional technologies. Also, it has been demon- strated that nanowires as chemical and biological sensors can be very sensitive with capability of single virus detection. Calculations on nanowires with randomly placed dopants or defects, require an atomistic model since quantum effects are important due to the small diameter, which can be down to 2 nm. The length of the wires can, however, be several µm and the numerical method should be able to treat more than 104 atoms, implying that (N) scaling is required. O Two (N) methods, both using a tight-binding model based on ab initio calcu- lations,O have been applied in this project. The first approach, derived from the Kubo-Greenwood formula, uses time propagation of wave packets to estimate the conductance. This method has until now only been applied by a single group and primary to model carbon nanotubes. One of the main objectives in this work has been to understand, implement and test this relatively new method. The second approach is based on the Landauer formula and the conductance is found by recursive calculations of Green’s functions. Comparison of the two approaches revealed that the Green’s function method is the preferred choice for modelling thin nanowires. It is more rigorous than the Kubo method and does not suffer from having many convergence parameters. The transport properties in both pure silicon wires and wires passivated with hydrogen is shown to be strongly affected by surface disorder. Randomly missing H atoms causes the electrons to localize, and both the elastic mean free paths and localization lengths scales linearly with the average distance between H vacancies. Comparison with estimated electron-phonon interactions indicates that impurities might be the dominant scattering mechanism for thin wires even at room temperature.
Resum´e
Dette projekt omhandler modellering af elektriske egenskaber i silicium nano- tr˚ade. Fremstillingen af nanotr˚ade kan kontrolleres med stor nøjagtighed, hvilket gør nanotr˚ade til lovende kandidater for fremtidige elektroniske komponenter med mulighed for at overg˚akonventionelle teknologier. Det er ligeledes blevet demonstreret, at nanotr˚ade som kemiske og biologiske sensorer kan være meget følsomme med mulighed for at detektere enkelte virusmolekyler. Modellering af nanotr˚ade med tilfældigt fordelte doteringer eller defekter kræver en atomar model, da kvanteeffekter er væsentlige p˚agrund af den lille diameter, som kan være ned til 2 nm. Længden af nanotr˚adene kan til gengæld være adskillige µm, og den numeriske metode skal derfor være i stand til a behandle mere end 104 atomer, hvilket betyder, at (N) skalering er nødvendigt. O Der er i projektet blevet anvendt to (N) metoder, der begge benytter en tight-binding model baseret p˚a ab initioO beregninger. Den første metode, der er udledt fra Kubo-Greenwood formlen, bruger tidsudvikling af bølgepakker til at estimere konduktansen. Denne metode er indtil nu kun blevet anvendt af en enkelt gruppe til at studere kulstof nanorør, og et af hovedform˚alene i dette arbejde har derfor været at forst˚a, implementere og teste denne relativt nye metode. Den anden fremgangsm˚ade er baseret p˚aLandauer formlen og konduktansen findes ved rekursive beregninger af Greens funktioner. En sammenligning af de to metoder gav, at Greens funktions metoden er det foretrukne valg til at modellere tynde nanotr˚ade. Den er mere stringent end Kubo-metoden, og har ingen konvergens parametre. Det er vist, at transportegenskaberne i b˚ade rene siliciumtr˚ade og hydrogen passiverede tr˚ade, er stærkt p˚avirket af overfladedefekter. Tilfældigt fjernede H atomer bevirker, at elektronerne lokaliserer, og b˚ade den elastiske fri mid- delvejlængde og lokaliseringslængden skalerer lineært med den gennemsnitlige afstand mellem H vakancerne. Sammenligning med estimerede elektron-phonon vekselvirkninger indikerer, at defekter kan være den væsentligste sprednings- mekanisme selv ved stuetemperatur.
Contents
List of symbols xi
1 Introduction 1 1.1 Semiconducting nanowires ...... 1 1.1.1 Growth of silicon nanowires ...... 2 1.1.2 Electronic applications ...... 3 1.1.3 SiNWsassensors...... 4 1.1.4 Critical issues and possible theoretical help ...... 5 1.2 Theoretical techniques ...... 6 1.2.1 Atomic and electronic structure ...... 6 1.2.2 Transport calculations ...... 7 1.2.3 Real-space Kubo-method ...... 7 1.2.4 Applications of the real-space Kubo method ...... 8 1.3 Motivation and outline ...... 10 1.3.1 Outline ...... 11
2 The real-space Kubo formalism 13 2.1 Derivation of the real-space Kubo formula ...... 13 2.1.1 Preliminary notations ...... 13 2.1.2 Rewriting the Kubo-Greenwood formula ...... 15 2.1.3 Time evolution ...... 17 viii CONTENTS
2.2 Three regimes of transport ...... 17 2.2.1 Ballistic regime ...... 18 2.2.2 Diffusiveregime ...... 18 2.2.3 Localization...... 19 2.2.4 Simple model for diffusion and localization ...... 20
3 Numerical methods 22 3.1 Exact solution to the time-dependent Schr¨odinger equation ...... 22 3.2 The Chebyshev method ...... 23 3.2.1 Basic properties of Chebyshev polynomials ...... 23 3.2.2 Expansion of Uˆ(t) in Chebyshev polynomials ...... 23 3.2.3 Convergence properties ...... 24 3.2.4 Convergence of the coefficients ...... 25 3.2.5 Comparison of Chebyshev and Taylor ...... 26 3.3 Continued fraction technique ...... 26 3.4 Overview of the computer codes ...... 28 3.4.1 Important parameters ...... 28 3.4.2 Dataflow ...... 29
4 The one dimensional chain: Testing of the method 31 4.1 Local Density of States ...... 31 4.1.1 The infinite chain ...... 32 4.1.2 Comparison of numerical and analytical LDOS ...... 32 4.1.3 Comparing the velocities ...... 34 4.2 Disorderedchain ...... 35 4.2.1 Random initial states ...... 36 4.2.2 Time dependent diffusion coefficient (E,t)...... 36 D 4.2.3 Meanfreepath ...... 39 4.2.4 Conductance of disordered wire ...... 40 4.3 Carbon nanotubes: Comparison with published results ...... 43 4.4 ParallelChains ...... 46 CONTENTS ix
4.4.1 Calculation of the conductance ...... 47
5 The Landauer formula and recursive Green’s function method 50 5.1 Recursive Green’s function method ...... 50 5.1.1 Introduction to the Landauer formalism ...... 50 5.1.2 Recursive (N)growthprocess...... 52 O
6 Building a Tight-binding model 55 6.1 Introduction and ideas behind the model ...... 55 6.1.1 Finding the tight-binding parameters ...... 56 6.2 From atomic basis set to orthogonal tight-binding ...... 57 6.2.1 Hamiltonian for the full wire ...... 58 6.2.2 L¨owdin transformation ...... 59 6.2.3 Approximate orthogonalization ...... 60 6.2.4 Example - The one-dimensional chain ...... 61
7 Modelling of silicon nanowires 64 7.1 Metallic silicon nanowires ...... 64 7.1.1 Truncating the Hamiltonian ...... 65 7.1.2 Conductance of a pristine wire ...... 66 7.1.3 Anderson disorder ...... 67 7.1.4 Sub-conclusion ...... 69 7.2 Hydrogen passivated Si-wire ...... 71 7.2.1 Setting up the model ...... 71 7.2.2 Length and energy dependent conductance ...... 73 7.2.3 Distribution of conductances ...... 74 7.2.4 Resistance vs. length ...... 76
7.2.5 Mean free path and peaks of σG(L)...... 77 7.2.6 Comparison with analytical models ...... 78
7.2.7 Scaling of le and ξ ...... 80 7.2.8 Sub-conclusion ...... 81 x CONTENTS
8 Summary and outlook 82 8.1 Evaluation of the real-space Kubo method ...... 82 8.2 Modelling of silicon nanowires ...... 85 8.3 Outlook ...... 86
Appendix 87
A L¨owdin transformation 88 A.1 The generalized eigenvalue problem ...... 88 A.2 Formal orthogonalization procedure ...... 90 A.3 How to do the orthogonalization ...... 91
B Numerical Methods 93 B.1 The Continued Fraction Technique ...... 93 B.1.1 Truncation of the continued fraction - self-energy . . . . . 94 B.2 Tridiagonalization procedure ...... 96 B.3 Exact solution to the time-dependent Schr¨odinger equation . . . 97 B.3.1 Time propagation using Taylor expansion ...... 98 B.4 Recursive calculation of self-energies ...... 99
C Miscellaneous 101 C.1 Total and Local Density of States ...... 101 C.1.1 Green’s function and local density of states ...... 102 C.2 Ohm’s law and mean free paths ...... 104 List of common symbols
Symbol Description Unit
2 1 (E,t) Time- and energy depentent diffusion m s− D coefficient 1 di(E) Local density of state eV− E Energy eV 1 G Green’s function matrix eV− 1 Gˆ Green’s function operator eV− 1 G Conductance Ω− H Hamiltonian matrix eV Hˆ Hamiltonian operator eV h On-site hamiltonian matrix eV 1 k Wavevector nm− L Length nm le Elastic mean free path nm N Number of conducting channels Ni Number of non-zero elements in a ran- dom phase state Ntri Size of tri-diagonal matrix, Htri Ntot Size of Hamiltonian matrix, H 1 n(E) Total density of state eV− R Position m R Resistance Ω R0 Contact resistance Ω S Overlap matrix xii LIST OF SYMBOLS
Symbol Description Unit
T (E) Transmission Tn(x) Chebyshev polynomial of order n t Time fs Uˆ(t) Time evolution operator 1 Vˆx x-component of velocity operator, m s− Schr¨odinger picture 1 Vˆx(t) x-component of velocity operator, Heis- m s− senberg picture V Coupling matrix eV v Coupling matrix eV 1 v(E) Velocity at energy E m s− Xˆ Position operator, Schr¨odinger picture m Xˆ(t) Position operator, Heissenberg picture m 2(E,t) Squared, average displacement of wave m2 X packets with energy E γ Hopping energy eV δ Cutoff energy eV δi Random on-site noise eV ∆ε Noise amplitude (disorder strength) eV ε0 On-site energy eV η Infinitesimal energy eV 1 ρ Resistivity Ω m− 1 σDC DC conductivity Ω− m σ Overlap integral 1 σG Standard deviation of conductance Ω− Σ Self-energy eV ξ Localization length nm ψ Random phase state | ri ψi General state |n,iR n’th orbital located at position R |Φ i n’th Chebysheb state | ni
Constant Description Value
34 h Planch’s constant 6.63 10− J s × 34 · ~ Planch’s constant h/2π 1.05 10− J s × 19 · e Electron charge 1.602 10− C × Chapter 1
Introduction
In this chapter we begin with an introduction to the field of nanowires focusing on the fabrication process and the demonstrated applications. Next, we discuss some of the theoretical techniques that have been applied to model nanowires, and pay special attention to a new real-space method, which we will apply in the project. We end up discussing the motivations for this project and give an outline of the report.
1.1 Semiconducting nanowires
Over the last decade increasing effort has been put into the research of semicon- ducting nanowires, experimentally as well as theoretically. The experimental work is pioneered by the groups of C. M. Lieber (Harvard) and L. Samuelson (Lund). The growth of the nanowires can be controlled to a very high degree: The length of the wires can be several µm and is controlled by the growth time, while the location and diameter are determined by the position and size of a seed particle [1, 2, 3]. Various heterostructures with atomically sharp transitions have been produced, opening up for a wide range of device applications such as single electron transistors (SET), resonant tunneling diodes and field effect transistors (FET) [2, 3, 4, 5]. Silicon nanowires (SiNWs) have also been applied as chemical and biological sensors with very high sensitivity and capability of single virus detection [6, 7]. In the following subsections we will go into more details about the fabrication techniques and applications. Since the calculations in this work have been con- cerned with silicon nanowires we will focus on these, and pay less attention to the III-V nanowires such as InAs, InP and GaAs. We end up discussing critical issues where modeling might be a helpful tool to gain understanding. 2 1.1 Semiconducting nanowires
1.1.1 Growth of silicon nanowires
Most of the SiNWs reported in the literature are produced by a vapuor-liquid- solid (VLS) growth method [8]. The basic principles of the method are the following. Gold nanoclusters are placed on an oxidized silicon wafer inside a reaction chamber where pressure, temperature and chemical environment can o be controlled. The chamber is heated to 400 C and SiH4 is added [1, 9]. The growth initiates at the Au nanocluster where∼ Si crystallizes and pushes the gold particle upwards forming a wire underneath. The diameter of the grown SiNWs is controlled within a few nm by the size of the Au nanoparticle [1] and ranges from 1.3 nm [10] to 100 nm [11]. Most typical sizes are, however, 5 30 nm [1, 4, 6, 12, 13]. Note,∼ that the very thin SiNWs reported in Ref. [10] were− made in a different way using an oxide-assisted growth method, where a powder of SiO is heated in an alumina tube to form wires. The growth directions are primarily [111] for larger wires while smaller wires tend to grow in the [110] direction [1]. The length of the SiNWs is controlled by the growth time.
Figure 1.1: Left: SiNW with a Si core and a relative thick sheath of SiOx. Scale bar, 10 nm. The image is from ref. [9]. Right: STM image of a hydrogen passivated SiNW. The SiOx has been removed by a HF etch. The image is from ref. [10].
After the growth process the SiNWs are covered by a thin sheath ( 1 5 nm) of ∼ − amorphous SiOx [1]. This is clearly seen in figure 1.1 (left), showing a diffraction contrast TEM image of a SiNW with a relative thick SiOx sheath [9]. The scale bar is 10 nm. The SiOx can be removed using a hydroflouric (HF) etch resulting in a hydrogen passivated surface [10]. An STM image of such a wire is shown to the right [10]. It is possible during growth to add dopants to the SiNW thereby making them either n- or p-type semiconductors. Adding B2H6 to the reaction chamber results in boron-doped SiNWs (p-type), while phosphorus doping (n-type) can be obtained by shining high intensity laser pulses on an Au-P nanocluster during 1.1 Semiconducting nanowires 3 growth [11]. Importantly, the doping can be altered during growth thereby creating heterostructures such as p-n junctions[13]. Recently metallic/semiconducting heterostructures with nm sharp transitions have been fabricated using NiSi as the metallic part [12]. First a normal, doped SiNW is grown and subsequently lithographically defined regions of the wire are covered by Ni. Heating the wire to 550oC and afterwards removing the residual Ni produces a metallic NiSi wire with a slightly larger diameter than the original SiNW. The junctions between the metallic NiSi part and the semiconducting Si part forms Ohmic contacts, and FETs were produced with the metallic NiSi regions working as source and drain electrodes [12].
1.1.2 Electronic applications
The ability to create nanowire heterostructures with very high accuracy makes them very promising candidates for future electronic devices. Moreover, the excellent growth control is a major advantage compared to carbon nanotubes, since it is not (yet) possible to control whether a single wall nanotube is metallic or semiconducting. High performance field effect transistors (FETs) have been demonstrated using boron doped SiNWs contacted to Ti/Gold electrodes. Measured key parameters such as transconductance and mobility showed that the SiNWs have the poten- tial to exceed conventional MOSFET devices [5]. The principal setup of the
Figure 1.2: Left: Schematic setup of a SiNW FET. The inset shows a TEM image of a SiNW. Notice the amorphous SiOx at the edges. Right: Conductance vs. gate voltage before (green) and after (red) passivation of surface defects. The polarized surface defects screen the gate field and effectively lower the gate capacitance. Both images are from ref. [5].
SiNW FET is shown in figure 1.2 (left). To the right is shown the conductance 4 1.1 Semiconducting nanowires vs. gate voltages before (green) and after (red) passivation of surface defects. We shall return to this below. Diodes have been fabricated using modulation doped SiNWs forming p-n junc- tions [13] and more complex structures such as complementary inverters and bipolar transistors have been produced using systems of crossed p- and n-type wires [4]. Nanowires of different III-V semiconductor materials, such as InAs-InP het- erostructures have been applied to make diodes, single electron transistors (SET) and resonant tunneling diodes [2, 3]. The various devices demonstrated and the metal-semiconductor wire-junctions all together make it possible to integrate very dense nanosystems that include both active device areas and high-performance interconnects [14].
1.1.3 SiNWs as sensors
In the recent years new applications of SiNWs as chemical and biological sensors have been demonstrated [6, 7, 15]. Zhou et al. [15] showed that a bundle of SiNWs, where the SiOx sheath was removed by a HF etch, was highly sensitive to NH3 but not to N2. With the SiOx sheath still present no significant response were observed. The sensitivity of the resistance to exposure of NH3 could either be due to changes in the surface resistance of individual wires or a changed inter-wire connection. Lieber et al. [6, 7] have produced several sensors all based on a field effect: A boron-doped SiNW is coated with anti-bodies specific for a certain species, e.g. a virus. Since the virus is electrically charged it will work as a local gate when it binds to the anti-body and the resistance of the wire changes. Figure 1.3 (left) shows the principle in such a sensor. The one-dimensional nature of the nanowires is important for the sensitivity, since the binding of a molecule to a wire surface will alter the carrier concentration in the whole cross section of the wire, whereas only the surface region of a planar device would be changed [6]. The right part of figure 1.3 shows simultaneous electrical and optical data. The upper panel shows the conductance of the wire as function of time. The numbers correspond to the optical images of the flourescently labeled influenza virus in the lower panel. Remarkably, the conductance drops precisely when the virus binds to the wire (points 2 and 4) and returns to the base level when the virus unbinds. These data show that SiNW sensors are very sensitive and that single virus detection is possible. Using different antibody receptors on different wires Patolsky et al. [7] were able to detect and distinguish two different viruses, thereby opening up for massive parallel detection of various species. 1.1 Semiconducting nanowires 5
Figure 1.3: A: Principle sketch of a sensor. As a virus binds to the surface of a SiNW device modified with antibody receptors, the conductance drops. When the virus unbinds, the conductance returns to the base value. B: Simultaneous conductance and optical data recorded for a SiNW device. The images corre- spond to the two binding/unbinding events marked by points 1-3 and 4-6 in the conductance data, with the flourescently labeled virus appearing as a red dot in the images. The images are from Ref. [6].
1.1.4 Critical issues and possible theoretical help
Several issues concerning SiNWs are still unanswered, and theoretical work and computer modelling might be helpful together with experiments to gain better understanding. Here we shall discuss some questions that might be addressed using transport calculations. It is evident from the above-mentioned that doping of SiNWs is a crucial point when making devices. Since the nanowires are quasi one-dimensional structures with diameters down to 10 atoms, the dopant atoms act as strong scattering centers that might affect∼ the performance significantly. From experiments it is difficult to determine the nature of the scattering events, i.e. whether defects, vacancies, dopants or maybe phonons govern the resistance. Transport calcula- tions, on the other hand, can address the different scattering events selectively. A realistic model should preferably be fully ab initio or at least based on parame- ters obtained from ab initio calculations. Interesting studies could for instance be to find out whether some dopants scatter less than others and thereby lead to better performance. Such a knowledge could possibly help experimentalist to make more efficient devices. 6 1.2 Theoretical techniques
In bulk materials, electron-phonon interactions are often the dominant scat- tering mechanism at room temperature. However, the phonon scattering is suppressed in small diameter nanowires [16], and recent experiments indicated ballistic transport in undoped Si/Ge core-shell wires at room temperature with an estimated phonon scattering mean free path lph = 540 nm [17]. This might imply that even at room temperature, defects could be the most important scattering source. Moreover, as shown in ref. [5] the conductance of a SiNW in- creased by approximately one order of magnitude after thermal annealing. The reason to the enhanced conductance is believed to be better metal-SiNW con- tacts and passivation of defects in the Si-SiOx interface. Furthermore, surface modification by reaction with 4-nitrophenyl octadecanoate lead to an increase in transconductance and mobility by approximately one order of magnitude, see figure 1.2. The explanation is probably that polar surface sites are passivated and therefore no longer screen the gate field [5]. Modelling the effect of defects and vacancies could possibly support and elaborate on the present explanations. It might also be possible to model some of the sensor applications. Calculations on e.g. NH3 and N2 on SiNW surfaces would be an interesting study that could help to understand why NH3 affects the resistance significantly while N2 does not [15]. It might also be possible to model more complex molecules and their influence on the conductance.
1.2 Theoretical techniques
Theoretical modelling of nanowires can be divided into two fields: First, deter- mination of the atomic and electronic structure, and second, calculation of the transport properties of the wire. The transport properties will often depend critically on the structure of the wire and both aspects are therefore important for a complete description.
1.2.1 Atomic and electronic structure
The atomic and electronic structure is primarily determined from density func- tional theory (DFT) calculations [18, 19, 20, 21] or from tight binding (TB) models [20, 22]. While DFT calculations can be very accurate they suffer from an (N 3) scaling, where N is the number of atoms. This means that an increase in computerO power by a factor of 1000 only allows to study systems increased by a factor of 10. Today, the system sizes are limited to 1000 atoms. TB calcula- tions can, on the other hand be of (N) [23]. The drawback∼ of TB methods is that they rely on accurate input parametersO that are often specific for a given system. An example of structure calculations on pure silicon wires (i.e. with no surface passivation of O or H) is given in ref. [20, 21], where both DFT and TB 1.2 Theoretical techniques 7 methods were applied to study wires grown in the [100]-direction. It was shown that small changes in the surface reconstruction had pronounced effects on the electronic structure, changing the wire from being semi-metallic to metallic. In chapter 7 we will apply this specific structure in conductance calculations, thus highlighting the important connection between structural and transport properties.
1.2.2 Transport calculations
Transport calculations in nanoscale systems range from full ab initio meth- ods based on non-equilibrium Green’s functions [24, 25, 26] over tight-binding models [27] to the Boltzmann equation and effective mass theory [28]. As the accuracy goes down (from ab initio to Boltzmann) the possible system sizes go up. Das et al. [28] used the Boltzmann equation to calculate the carrier mobility in relative thick (d = 10 90 nm) GaAs wires focusing on the diameter dependence. Sundaram et al. [29]− also used the Boltzmann equation to study surface effects on the transport in large diameter wires. Y. Zheng et al. [27] applied a tight-binding model of a hydrogen passivated wire and studied the effect of wire thickness on the bandgap, effective masses and transmission. The band gap was found to increase for decreasing diameters, in agreement with experiments [10] and ab initio calculations [30]. In chapter 7 we will use the same wire structure as ref. [27] to calculate transport properties of the H-passivated wires with randomly removed H atoms. Very recently X. Blase and coworkers [26] used DFT calculations to study the effect of dopants on the transmission through both passivated an un-passivates SiNWs. It was shown that B- and P-doping lead to significantly different con- ductance properties in the wires as compared to bulk material. To our knowledge, no theoretical works concerning nanowires, based on ab initio methods and including many scattering events, have been published.
1.2.3 Real-space Kubo-method
One of the primary objectives in this work has been to implement and test a rel- atively new real-space method to calculate electronic transport properties such as conductance and mean free paths. The method is developed by the group of S. Roche and D. Mayou and has been applied in a number of publications dis- cussed below [31, 32, 33, 34, 35, 36, 37, 38, 39, 40]. The method is based on the Kubo-Greenwood formalism [41, 42] rewritten in a real-space framework, and we will therefore refer to it as the ’real-space Kubo-method’. The fundamental philosophy behind the method is that the transport properties are governed by the movement of electrons in a given sample as time evolves. If the sample 8 1.2 Theoretical techniques is a perfect crystal, the electrons can travel through the sample without any backscattering and the resistance of the sample is zero (except for the contact resistance, of course). This is called the ballistic regime since the motion re- sembles classical, ballistic particles. If defects are present, e.g. in the form of different atoms (dopants) or missing atoms (vacancies), the electrons will even- tually hit a defect and scatter, and the motion through the sample resembles more a random walk, leading to a diffusive behaviour and Ohmic resistance. In the real-space Kubo method one calculates the time-dependent broadening of initially localized wave packets. This measures the ease with which electrons can propagate in the sample and using this, the conductivity, mean free path or resistance can be calculated. Importantly, the method scales as (N) which means the computation time doubles if the system size doubles. O In the following subsection we briefly review the applications and results ob- tained with the method. A more throughout discussion of the details is given in chapter 2.
1.2.4 Applications of the real-space Kubo method
Most of the published work using the real-space Kubo method have been con- cerned with carbon nanotubes (CNTs). S. Roche and R. Saito showed that the magnetoresistance of small diameter CNTs could be tuned from positive to negative by changing the chemical potential and the orientation of the magnetic field [35].
F. Triozon et al. [36] calculated the elastic mean free path, le, in different single walled carbon nanotubes (SWNTs) as function of energy and disorder strength. Figure 1.4 shows large variations of le of three different armchair SWNTs as
Figure 1.4: Energy dependent mean free path in different armchair SWNTs with random on-site disorder. The inset shows the scaling of the mean free path with disorder strength. The figure is from ref. [36]. 1.2 Theoretical techniques 9
function of energy. The inset shows the scaling of le vs. disorder strength in a (5,5) SWNT. In chapter 4 we validate our numerical implementation by reproducing the scaling of le and compare with analytical calculations. S. Latil et al. [37] calculated transport properties of chemically doped CNTs. The electronic structure of a graphene layer with one boron or nitrogen sub- stitution was calculated using DFT, and renormalized tight-binding parameters were obtained through fitting of the band structures. Figure 1.5 (left) shows the super cell of the graphene layer, with a boron atom located in the corner. As an effect of the doping atom, the on-site energies were changed up to third nearest neighbours, while the hopping energies were unchanged. The changed tight-binding parameters calculated for graphene were assumed to be valid for a CNT as well. Figure 1.5 (right) shows the energy dependent conductance for
Figure 1.5: Left: A DFT calculation on a graphene layer with a single boron substitution (in the corner) is used to obtain renormalized on-site energies up to third nearest neighbours. Right: Energy dependent conductance for a (10,10) CNT with 0.1% boron atoms calculated for different lengths.
different tube lengths. The boron concentration is 0.1%, and the boron atoms are randomly distributed on the tube. For short tubes the conductance at pos- itive energies is ballistic with well defined conductance plateaus, whereas for longer tubes the conductance is strongly affected by the dopants. Notice that the largest tube is more than 1 µm long, thus containing more than 105 atoms. In a related study S. Latil et al. [40] recently calculated transport properties of CNTs with random coverage of physisorbed molecules. Tight-binding pa- rameters were again determined from band structure fitting, and the analysis showed that the impact of physisorption on the transport strongly depends on the HOMO-LUMO gap of the attached molecules. 10 1.3 Motivation and outline
1.3 Motivation and outline
The motivation for the present work is that computer modelling is becoming increasingly important in understanding physical properties of nanoscale sys- tems. As discussed above, there are several issues concerning nanowires where a combination of theoretical and experimental work could benefit our knowledge. A calculation of the conductance of, say a SiNW with randomly positioned doping atoms or defects, puts strong requirements on the method. The quasi one-dimensional nature of the nanowires requires on the one hand an atomistic model taking quantum effects into account and, on the other hand the method should be able to treat 104 atoms and include many scattering events due to the µm length of the wires.∼ In this work, two (N) methods based on DFT calculations are used to study the effect of manyO random defects in long nanowires. Since the real-space Kubo method has been successfully applied to CNT systems and since it scales as (N), it is appealing to understand the method in detail, implement it and use itO to study SiNWs. Because only one group has used it in the literature and applied it primary on carbon nanotubes, there are many unanswered questions that will be addressed in the following chapters:
How does one derive the formulas used in the articles, and are the deriva- • tions rigorous?
How should the method be implemented? •
Does the method yield the same results as other methods? •
Is it applicable to other, more difficult systems than carbon nanotubes? •
The second method we apply is more well-known. It is based on the Landauer formula, and the transmission through a sample is found from recursive calcu- lations of Green’s functions (GFs). The aim of the work is not to make very accurate calculations reproducing exper- imental data exactly. To that the model is too simple: We use a single electron model, neglecting electron-electron interactions as well as electron-phonon in- teractions. Also, any leads connected to the samples will be assumed to be perfectly conducting with reflectionless contacts to two reservoirs [43]. Any real contact resistance will thus not be included. The objective is therefore rather to make qualitative estimates and answer questions like: What kind of defects are important? Is surface- or bulk disorder the most important? Does some dopants scatter more than other? Is phonon or defect scattering the most important? 1.3 Motivation and outline 11
1.3.1 Outline
The outline of the report is the following: In chapter 2 we review the details in the real-space Kubo formula, and show that the transport properties are governed by the time dependent diffusion coefficient. We end the chapter by discussing three regimes of transport: ballistic, diffusive and localization. A key element in the numerical procedure is the time evolution of wave packets. In chapter 3 we show how this efficiently can be calculated using Chebyshev polynomials. The chapter also includes an introduction to the continued fraction technique which is used to calculate densities of states, together with a brief description of the numerical implementation. In chapter 4 we test the numerical implementation of the Kubo method on a simple one-dimensional chain. We focus on the parameters governing the accuracy and compare numerical result with analytical expressions. We further validate our implementation by reproducing previously published results. We end the chapter with a discussion of an inherent problem in the method leading to incorrect conductances near band edges. The recursive Green’s function method is explained in chapter 5. Starting with a semi-infinite lead, a device area is ’grown’ by adding unit cells one at a time. At each growth step the transmission through the wire is calculated thus giving a length dependent resistance. As discussed above, a reliable model of e.g. a SiNW should be based on first principles calculations. In chapter 6 we show how to construct a tight-binding Hamiltonian from a DFT calculation. In order to perform the time evolution in the Kubo method, it is preferable to have an orthogonal basis set. However, the DFT program siesta uses a non-orthogonal atomic basis set, and it is there- fore necessary to orthogonalize the output from siesta by a so-called L¨owdin transformation. In chapter 7 we apply both the real-space Kubo method and the recursive GF method to study silicon nanowires. First, an un-passivated, metallic wire with random on-site disorder is considered. Is is shown that disorder in the bulk has little influence on the conductance at the Fermi level, whereas disorder on the surface atoms affects the transport significantly. Comparison of the two methods show that they yield approximately the same results. The GF method is, however, found to be faster than the Kubo method, and to some extent also more reliable, and the rest of the results are obtained using the GF approach. All wires reported in the literature are passivated by either SiOx or hydrogen. We therefore model an H passivated SiNW with randomly missing H atoms. Resistance scalings and mean free paths are calculated together with localization lengths. In chapter 8 we summarize the comparison of the two numerical methods. The real-space Kubo method suffers from inherent problems near band edges, which 12 1.3 Motivation and outline has significant influences for the studied SiNWs. Another drawback is that dif- ferent accuracy parameters need to be determined manually. The GF method is more rigorous and does not contain any adjustable parameters. Also, it is possible to use a non-orthogonal basis and thus avoid the L¨owdin transforma- tion. On the other hand, the Kubo method is able to find the conductance at all energies in one calculation, whereas a full calculation is needed for each energy when using the recursive GF method. We finish the report with an outlook. Chapter 2
The real-space Kubo formalism
In this chapter the real-space Kubo formulas for conductivity and conductance are derived. This is not a standard textbook procedure and has only partly been presented in the literature [33]. We will therefore present the derivations in detail. In section 2.2 we will discuss three different regimes of transport: Ballistic, diffusive and localization, all of which are captured by the method.
2.1 Derivation of the real-space Kubo formula
We start by introducing the notation and continue to rewrite the Kubo-Greenwood formula. The following derivations are primarily based on the (French) Ph.D. thesis of F. Triozon [44].
2.1.1 Preliminary notations
The mean value of an arbitrary operator Aˆ over states having energy E is written as:
1 N Tr[δ(E Hˆ )Aˆ] Aˆ = ψ(i) Aˆ ψ(i) = − , (2.1) h iE N h E | | E i ˆ i Tr[δ(E H)] X − 14 2.1 Derivation of the real-space Kubo formula where ψ(i) are N degenerate eigenstates of Hˆ , all having energy E. The last | E i equality sign can be verified by using the eigenstates, ψ of Hˆ in the trace: | En i Tr[δ(E Hˆ )Aˆ] ψ δ(E Hˆ ) Aˆ ψ − = nh En | − | En i Tr[δ(E Hˆ )] ψ δ(E Hˆ ) ψ − P nh En | − | En i ψ δ(E E )Aˆ ψ = Pnh En | − n | En i ψ δ(E E ) ψ P nh En | − n | En i 1 = P ψ(i) Aˆ ψ(i) N h E | | E i i X The mean value of the spreading in the x-direction of states having energy E is:
2 2(E,t) = Xˆ(t) Xˆ(0) , (2.2) X − E i Htˆ i Htˆ where Xˆ(t) = e ~ Xˆ e− ~ is the x-component of the position operator in the Heisenberg picture. The velocity autocorrelation function of states with energy E is in the same way defined as:
C(E,t) = Vˆx(t), Vˆx(0) = Vˆx(t) Vˆx(0) + Vˆx(0) Vˆx(t) . (2.3) { } E E D E D E Differentiation of (2.2) leads to
d 2 (E,t) = Vˆx(t) Xˆ(t) Xˆ(0) + Xˆ(t) Xˆ(0) Vˆx(t) , (2.4) dtX − − E D E where the velocity operator is defined as d i Vˆ (t) = Xˆ(t) = H,ˆ Xˆ(t) . (2.5) x dt ~ h i Changing the time arguments in the trace (2.4) allows us to write
d 2 (E,t) = Vˆx(0) Xˆ(0) Xˆ( t) + Xˆ(0) Xˆ( t) Vˆx(0) , dtX − − − − E D E and the second derivative of 2(E,t) can thus be written as: X 2 d 2 (E,t) = Vˆx(0) Vˆx( t) + Vˆx( t)Vˆx(0) . dt2 X − − E D E Changing the time arguments again and using (2.3) we finally get the relation between the velocity autocorrelation function and the average spread:
d2 2(E,t) = C(E,t). (2.6) dt2 X 2.1 Derivation of the real-space Kubo formula 15
2.1.2 Rewriting the Kubo-Greenwood formula
The Kubo-Greenwood formula for the DC-conductivity, σDC , at zero tempera- ture can be written as [33]
2 ~ e2 π σ (E) = Tr Vˆ δ(E Hˆ ) Vˆ δ(E Hˆ ) , (2.7) DC Ω x − x − h i where Ω is the volume of the system. The last delta-function is rewritten as an integral
1 ∞ i(E Hˆ )t/~ δ(E Hˆ ) = dt e − − 2π~ Z−∞ and inserted into (2.7):
2 1 ∞ i(E Hˆ )t/~ σ (E) = 2 ~ e π dt Tr Vˆ δ(E Hˆ ) Vˆ e − DC 2π~ x − x Z−∞ h i 2 ∞ iEt/~ iHt/ˆ ~ = e dt Tr Vˆ δ(E Hˆ ) e Vˆ e− . x − x Z−∞ h i We have incorporated Ω in the states in the trace, thus normalizing them to the volume. Due to the delta-function we can write eiEt/~ = ei Hˆ t/~ and thereby get ∞ σ (E) = e2 dt Tr Vˆ (0) δ(E Hˆ ) Vˆ (t) , (2.8) DC x − x Z−∞ h i where Vˆx(t) is the x-component of the velocity operator, written in the Heisen- berg notation. Using the notation from (2.1) of an energy average we get
Tr Vˆx(0) δ(E Hˆ ) Vˆx(t) = Tr δ(E Hˆ ) Vˆx(t) Vˆx(0) , − − E h i h i D E which is inserted into (2.8):
2 ∞ σDC (E) = e dt Tr [δ(E Hˆ )] Vˆx(t) Vˆx(0) . (2.9) − E Z−∞ D E Using (2.3) and (2.6) we can rewrite the integral and evaluate it:
2 ∞ σDC (E) = e dt Tr [ δ(E Hˆ ) ] C(E,t) 0 − Z t= d ∞ = e2 Tr [ δ(E Hˆ ) ] 2(E,t) − dtX t=0 d = e2 Tr[ δ(E Hˆ ) ] lim 2(E,t), (2.10) t − →∞ dtX 16 2.1 Derivation of the real-space Kubo formula
d 2 2 since dt (E,t) t=0 = 0. Using the definition of (E,t) in equation (2.2) and the generalX equation (2.1) of the energy mean, weX finally get:
2 2 d σDC (E) = e Tr[ δ(E Hˆ ) ] lim Xˆ(t) Xˆ(0) t − dt − E →∞ ˆ ˆ ˆ 2 d Tr δ(E H) (X(t) X(0)) = e2 Tr[ δ(E Hˆ ) ] lim − − − t dt h Tr[ δ(E Hˆ )] i →∞ − d = e2 Tr[ δ(E Hˆ ) ] lim (t (E,t)) , (2.11) t − →∞ dt D where we have defined the time and energy dependent diffusion coefficient, (E,t), by D 2 Tr Xˆ(t) Xˆ(0) δ(E Hˆ ) 2(E,t) 1 − − (E,t) = X = . (2.12) D t t Tr δ(E Hˆ ) − h i In the diffusive regime, (E,t) = (E) and (2.11) reduces to D D0 2 σDC (E) = e n(E) lim (E,t). (2.13) t →∞ D where the total density of states Tr[ δ(E Hˆ ) ] = n(E). − The conductance of a wire of length L is found from (2.13) as [35] e2 G(E, L) = n(E) (E,τ) (diffusive), (2.14) L D where the time τ is defined by the condition L = 2(E,τ), (2.15) X i.e. the time for the electrons at energyp E to spread out by an amount equal to L. This relation between τ and L seems rather arbitrary and could just as well have been L = a 2(E,τ) with a being of the order of 1. X Strictly speaking,p the formula (2.14) is only correct in the diffusive regime. It yields, however the correct physics in the ballistic regime also, since there 2 d (E,t) = v(E) t and dt (t (E,t)) = 2 (E,t), resulting in a factor 2 differ- enceD compared to (2.14): D D e2 G(E, L) = 2 n(E) (E,τ) (ballistic). (2.16) L D
In spite of the difference between the diffusive and ballistic formula, equation (2.14) is used in ref. [35, 36, 37, 33, 32] and applied to both regimes, not commenting on the factor 2 difference. In the section 2.2 we will look into three regimes of transport using (2.14). The discussion is mostly qualitative and we shall not care about the factor 2 difference. 2.2 Three regimes of transport 17
2.1.3 Time evolution
In order to calculate (E,t) we need to move the time-dependence from the operator Xˆ(t) to the statesD in the trace. Consider therefore the numerator in (2.12) which we name nx(E,t) and rewrite as
2 n (E,t) = Tr Xˆ(t) Xˆ(0) δ(E Hˆ ) x − − iHtˆ iHtˆ iHtˆ iHtˆ = Tr e Xeˆ − Xˆ δ(E Hˆ ) e Xeˆ − Xˆ , − − − h i where we have used that the trace is permutation invariant. Inserting Iˆ = iHtˆ iHtˆ iHtˆ e e− and using that [H,ˆ e ] = 0 leads to
iHtˆ iHtˆ iHtˆ iHtˆ n (E,t) = Tr e Xˆ Xeˆ − δ(E Hˆ ) Xeˆ − e− Xˆ x − − − h i = Tr [X,ˆ Uˆ(t)]† δ(E Hˆ ) [X,ˆ Uˆ(t)] , (2.17) − h i iHtˆ where Uˆ(t) = e− is the time evolution operator (see chapter 3). The effect of operating with the commutator, [X,ˆ Uˆ(t)], on one of the states, ψi , in the trace is: | i [X,ˆ Uˆ(t)] ψ = Xˆ ψ (t) Uˆ(t) Xˆ ψ . (2.18) | ii | i i − | ii ˆ (x) ˆ X operates on the time propagated state ψi(t) , and the state ψi = X ψi is (x) | i | i (x| ) i propagated to time t, ψi (t) . Denoting the state Ψi(t) = ψi(t) ψi (t) , the numerator in (2.12)| is compactlyi written as a sum| of locali | densitiesi − | of statei terms: n (E,t) = Ψ (t) δ(E Hˆ ) Ψ (t) . (2.19) x h i | − | i i i X In chapter 3 we show how to calculate Uˆ(t) using Chebyshev polynomials and explain how the local densities of states are calculated with a continued fraction method.
2.2 Three regimes of transport
It is evident from the above mentioned that the transport properties to a large extent are governed by the time and energy dependent diffusion coefficient (E,t). Figure 2.1 illustrates a typical time evolution of (E,t). The bal- Dlistic regime is characterized by an initial linear increase, withD a slope given by the velocity. This is followed by the diffusive regime with constant diffusion coefficient (E,t) = 0(E). Finally the electrons start to localize resulting in a decreasingD diffusionD coefficient. In the following we shall briefly discuss some of the properties characterizing the three regimes. 18 2.2 Three regimes of transport
D(t)
Diffusive D0
Localization Ballistic
v2
t
Figure 2.1: Illustration of the time development of the diffusion coefficient (E,t). From an initial ballistic regime the transport becomes diffusive before enteringD the localization regime.
2.2.1 Ballistic regime
For ballistic propagation 2(E,t) = v(E)2 t2 the diffusion coefficient grows linearly with time (E,t)X = v(E)2 t. Inserting this into (4.14) and writing L = v(E) τ we get: D
G(E) = e2 n(E) v(E). (2.20)
The energy dependence of the one-dimensional density of states and the velocity cancels, thus resulting in a length independent conductance, proportional to the number of conducting channels, as expected from the Landauer formula [43]. The elastic mean free path can be estimated from the diffusion coefficient as
max (E,t), t > 0 (E) l (E) = {D } = D0 , (2.21) e v(E) v(E) where the velocity v(E) is determined from the initial slope, α(E), of the diffu- sion coefficient, as v(E) = α(E) - see figure 2.1. p 2.2.2 Diffusive regime
In the diffusive regime, where (E,t) is constant the resistance increases linearly with length: D 1 L R(E, L) = = , (2.22) G(E) e2 n(E) (E) D0 2.2 Three regimes of transport 19
2 1 in agreement with Ohm’s law. Using (2.20) to write R0 = (e n(E) v(E))− (2.22) can be rewritten as
R(E, L) = R0 L/le(E). (2.23)
The Ohmic slope is therefore determined by the mean free path, a relation we shall use in chapter 7, when analyzing numerical data. The conductivity (2.13) becomes
σ (E) = e2 n(E) (E), (2.24) DC D0 which is the Einstein relation for conductivity [43].
2.2.3 Localization
Localization in disordered systems has been studied during the last 50 years since Andersons work in 1958 [45]. The field is extensively reviewed by Lee and Ramakrishnan [46] and by Kramer and MacKinnon [47], while Al’tshuler and Lee [48] give a more introducing review. A consequence of disorder in one- dimensional systems is that electrons always localize with a resistance increasing exponentially with length [46]
R(L) = R (eL/ξ 1), (2.25) 0 − where ξ is called the localization length. The formula was originally derived by Anderson et al. [49] using a subtle argumentation: One seeks a quantity, f(L) that has an additive mean which implies that the averaged resistance should not depend on how you add pieces of the sample together, i.e. if a piece of length L0 results in a factor f(L0) another piece with length 2L0 should give a factor 2 f(L0). Anderson et al. showed that the quantity ln(1+r) (r is the resistance in units of h/2e2) fulfilled this requirement meaningh that oni averaging over the distribution of two scatters with resistances r1 and r2 one have
ln(1 + r) = ln(1 + r ) + ln(1 + r ) . h i h 1 i h 2 i 1 This leads to a scaling behaviour ln(1+r) = αL, where α = ξ− is interpreted as the inverse localization length,h and (2.25)i follows. Notice, that for L ξ ≪ (2.25) gives R(L ξ) R0L/ξ. Comparing with (2.23) we see that for the one- dimensional system≪ the≈ mean free path is of the same order as the localization length, ξ le, in accordance with the general relation ξ Nle, where N is the number of∼ conducting channels [46, 50]. ∼ Although the study of disordered one-dimensional systems dates back many years it is still an active field of research. The emergence of new quasi one- dimensional structures such as CNTs and nanowires has opened up for a vari- ety of experimental systems where localization can be studied. Very recently G´omez-Navarro et al. [51] were able to tune the localization length in SWNT by 20 2.2 Three regimes of transport ion irradiation. An exponentially increasing resistance was observed in excellent agreement with the theory (2.25). In the framework of the real-space Kubo method, localization is seen for long time propagations, where the spreading of the wave packet reaches a constant level 2(E,t) L for t . X → 0 → ∞ It follows from (2.12) that the diffusion coefficient will decrease for increasing times leading to a decreasing conductance.
2.2.4 Simple model for diffusion and localization
We now introduce a simple, classical model for diffusion and localization. The model closely follows Todorov [50] and the reason for using it, is that it fits numerical calculations very well (see chapter 7.2).
Left Device Right jin jout jref
0 L x
Figure 2.2: A device is coupled to two perfect conducting leads (Left and Right). Within the device area the propagation is diffusive. The incoming, reflected and outgoing currents (jin, jref , jout) constitute the boundary conditions.
We consider the model in figure 2.2 with left and right leads being perfect conductors coupled to a device area. We assume N conducting channels each having transmission probability T , and wish to calculate T = jout , where j jin in and jout is the incoming and outgoing currents, respectively. First, we assume that the electrons diffuse through the device area. The number of conducting electrons per unit length is given by the steady-state diffusion equation:
n′′(x) = 0 n(x) = A + Bx. (2.26) ⇒ The constants A and B are determined by the boundary conditions. To the left, n(0)v = j + j = 2j j yielding A = jin (2 jout ), where v is the x in ref in out vx jin x − − 1 velocity in the x-direction. Using that j = Dn′(L), where D = v l is the out − 2 x e diffusion coefficient and le is the mean free path, the right boundary condition 2.2 Three regimes of transport 21
2A n(L)vx = jout yields B = . The transmission can now be found as − le+2L j v n(L) l T = out = x = e (2.27) jin jin le + L
2 2e 1 and the resistance R = ( h NT )− becomes
R0 R = R0 + L, (2.28) le h where R0 = 2e2N . This is simply Ohms law, where the resistance increases linearly with the length L of the sample, and the Ohmic slope is determined by the contact resistance, R0, and the mean free path, le (cf. equation (2.23). This result can also be found by calculating the transmission probability of a series of scatterers, as shown in appendix C.2. In order to treat localization we extend the model by adding a random distribu- tion of traps in the device area, which occasionally captures the electrons. The traps are modeled by an effective sink, η2n(x), term added to the diffusion equation (2.26): −
2 ηx ηx n′′(x) η n(x) = 0 n(X) = Ae + Be− . (2.29) − ⇒ Again, the constants are found from the boundary conditions and the resistance becomes [50]: h R(L) = [c sinh(L/ξ) + cosh(L/ξ)] , (2.30) 2e2N
le 2 where c = ξ(1 + ( 2L ) )/le and ξ = 1/η is the localization length. For L ξ (2.30) reduces to Ohms law (2.28). Furthermore, for systems with one conduct-≪ L/ξ ing channel where le ξ, the resistance increases for L ξ as R(L) R0e , in accordance with (2.25).∼ With a few conducting channels≫ (2.25) gives≈ a faster deviation from the linear region than (2.30). Equation (2.30) fits numerical data remarkably well, as will be evident in chapter 7.2. Studies of the localization regime should, however, be done with great care. Since localization is caused by interference effects, phase randomization due to inelastic scattering will ’destroy’ the localization. Any inelastic scattering event such as electron-phonon or electron-electron scattering will result in a finite phase coherence length, lφ. If lφ < ξ localization will not be seen, which is the reason why everyday life metallic wires are conducting. Normally, we should therefore not expect to see localization effects at room temperature, and the localization length in the numerical data should probably mostly be an indicator of the length range in which we can apply the model, without inelastic effects. However, Lu et al. [17] recently reported ballistic transport in nanowires at room temperature, with estimated lφ = 540 nm. By introducing defects purposely as done in ref. [51] it might be possible to see localization effects in nanowires. In any case, the quantum mechanical nature of localization is interesting and we will return to it in chapter 7. Chapter 3
Numerical methods
In the previous chapter we saw that the transport properties of a system is determined by the time dependent diffusion coefficient. In order to calculate this we need to know the time evolution of the states involved in the traces (2.12). In section 3.1 and 3.2 we discuss time evolution and present in detail the efficient Chebyshev method, where the time-evolution operator is expanded in a set of orthogonal polynomials - the Chebyshev polynomials. In section 3.3 we show how local densities of states are calculated using a continued fraction technique. We finish the chapter by a short description of the numerical implementation focusing on important parameters and the data flow.
3.1 Exact solution to the time-dependent Schr¨odinger equation
The time evolution of a quantum mechanical state is governed by the time- dependent Schr¨odinger equation (TDSE), which, together with a boundary con- dition, constitutes a first order differential equation
∂ Hˆ ψ(t) = i~ ψ(t) , ψ(t = 0) = ψ , (3.1) | i ∂t| i | i | 0i with Hˆ being the Hamiltonian operator, which we shall assume to be time independent. The solution to (3.1) is found by simple integration:
i Htˆ ψ(t) = Uˆ(t) ψ = e− ~ ψ , (3.2) | i | 0i | 0i i Htˆ where we have introduced the time-evolution operator Uˆ(t) = e− ~ . Although the formal solution in (3.2) is very simple, it does not help us, actually finding the state ψ(t) at time t. In order to proceed, we project the Hamiltonian | i 3.2 The Chebyshev method 23 operator Hˆ onto a known basis set yielding a Hamiltonian matrix, H, and assume to know the eigenstates: H E = E E . (3.3) | ni n | ni As shown in appendix B.3 the time-evolution matrix can be written as
iE1t iE2t iEnt U(t) = V diag e− , e− ,...,e− ,... V†, (3.4) where the matrix V with the eigenvectors E as columns, diagonalizes H: | ni V† H V = diag(E1, E2,...,En,...). However, in general we do not know the eigenstates of the system, and finding them involves the diagonalization of H, an operation which scales as (N 3). For a long Si-wire containing 100,000 atoms this is an impossible taskO. We therefore have to find the time-evolution operator in an other way.
3.2 The Chebyshev method
In this section we describe in detail how the time-evolution operator Uˆ(t) can be expanded in the orthogonal set of Chebyshev polynomials. We first discuss some basic properties of the Chebyshev polynomials and then describe how the expansion is implemented and discuss the convergence properties. Finally, we compare the Chebyshev method with a simple Taylor expansion.
3.2.1 Basic properties of Chebyshev polynomials
The Chebyshev polynomials, Tn(x) are defined by the weight function w(x) = 1 through the orthogonality relation[52]: √1 x2 − 1 1 π π Tn(x) Tm(x) δnm , n, m > 0 Tn(x) w(x) Tm(x)dx = dx = 2 2 1 2 1 √1 x 2 , n = m = 0 Z− Z− − with the first Chebyshev polynomial T0(t) = 1. The other polynomials can be determined from the orthogonality relation through a Gram-Schmidt procedure. The Chebyshev polynomials obey the simple and very useful recurrence relation:
Tn+1(x) = 2xTn(x) Tn 1(x) , n 1, (3.5) − − ≥ which we shall use extensively in the following.
3.2.2 Expansion of Uˆ(t) in Chebyshev polynomials
We now show how to expand the time-evolution operator, Uˆ(t), in the Cheby- shev polynomials. If the spectrum of the Hamiltonian does not lie within the 24 3.2 The Chebyshev method interval [ 1; 1], the Hamiltonian is mapped to the new, normalized operator [53]: −
2Hˆ (Emax + Emin)Iˆ Hˆ EminIˆ Hˆ ′ = − = 2 − I,ˆ (3.6) E E E E − max − min max − min with eigenvalues in the range [ 1; 1]. In the following we shall assume that the Hamiltonian is properly scaled.− Given an initial state, ψ = ψ(t = 0) and a final time, t, we have: | 0i | i iHtˆ ψ(t) = Uˆ(t) ψ = e− ψ | i | 0i | 0i ∞ ∞ = c (t) T (Hˆ ) ψ = c (t) Φ , (3.7) n n | 0i n | ni n=0 n=0 X X where we have defined the Chebyshev states Φ T (Hˆ ) ψ . The coefficients | ni ≡ n | 0i c(t) are calculated in section 3.2.4. Using the recurrence relation (3.5) the Chebyshev states can by calculated recursively as
Φ = T (Hˆ ) ψ = ψ | 0i 0 | 0i | 0i Φ = T (Hˆ ) ψ = Hˆ ψ | 1i 1 | 0i | 0i Φ = T (Hˆ ) ψ = 2HTˆ (Hˆ ) ψ T (Hˆ ) ψ | 2i 2 | 0i 1 | 0i − 0 | 0i = 2Hˆ Φ Φ | 1i − | 0i . .
Φn+1 = Tn+1(Hˆ ) ψ0 = 2Hˆ Φn Φn 1 . | i | i | i − | − i The (n + 1)’th Chebyshev state, is simply found from the two previous and one operation with the Hamiltonian. This utilization of the recurrence relation makes it computationally very efficient to calculate the terms in the sum (3.7).
3.2.3 Convergence properties
In order to investigate the convergence properties of the expansion, we first recall the operator norm:
Let Oˆ : H H , Oˆ B(H, H) be a bounded operator from H to H. We define→ the operator∈ norm by:
Oˆ B = sup Oxˆ H , x H 1 k k {k k k k ≤ }
Since the Chebyshev polynomials constitute an orthogonal set, we have
N ∞ 1 = Uˆ(H,tˆ ) = c T c T . k k k n nk ≥ k n nk n=0 n=0 X X 3.2 The Chebyshev method 25
This means that the expansion is converged, when it holds that N cn(t) Φn 1 < ǫ, (3.8) k | ik − n=0 X for some predefined tolerance ǫ. For any ǫ we can always find a N such that the inequality (3.8) is satisfied. This has the great advantage, that one does not need to have any knowledge about the system (besides an estimate of the spectrum of the Hamiltonian) before doing the expansion. One simply keeps adding terms to the sum (3.7) until the approximated operator becomes unitary, i.e. (3.8) is satisfied.
3.2.4 Convergence of the coefficients
We have just seen that the expansion (3.7) converges, but we have not yet said anything about how fast this convergence is. This is determined by the behavior of the coefficients, cn(t). These appear as the scalar product (defined by the weight function w(t)) between the Chebyshev polynomials and the operator to expand: ′ 1 ix t 2 e− Tn(x′) cn(t) = dx′. 2 π 1 √1 x′ Z− − Applying the scaling procedure (3.6) and defining W = E E , β = max − min Emax + Emin leads to W ′ β 1 i( 2 x + 2 )t 2 e− Tn(x′) cn(t) = dx′ 2 π 1 √1 x′ − Z −W ′ 1 i 2 x t 2 i β t e− Tn(x′) = e− 2 dx′ 2 π 1 √1 x′ Z− − i β t n W = 2 e− 2 ( i) J t , (3.9) − n 2~ where Jn(x) is the Bessel function of the first kind and order n, and where we have reinserted ~. For large n the Bessel function behaves as 1 x n J (x) . n ≈ n! 2 n+1 n and by using Stirlings formula, n! √2πn 2 e− the asymptotic behavior of the coefficients become ≈ exp(1) W t n c (t) 0 for n . | n | ∼ 4 ~ n → → ∞ exp(1) W t Thus, when n > 4 ~ the coefficients go to zero exponentially. This decay is especially fast when t is large, because in that case n also has to be large, exp(1) W t and and when finally n > 4 ~ the coefficients drop to almost zero since (1 δ)n 0 for n 1, even for small δ. We therefore expect a very sudden decay− of→ the coefficients,≫ when a certain expansion order is reached, and the expansion (3.7) will converge fast. 26 3.3 Continued fraction technique
3.2.5 Comparison of Chebyshev and Taylor
In order to test the Chebyshev method, we consider a tight-binding description of a simple one-dimensional chain with only nearest neighbour (n.n.) interac- tions. The non-zero elements of the Hamiltonian matrix are:
Hii = ε0 , Hi,i+1 = Hi+1,i = γ. (3.10) The electron is initially located at one site and the time evolution is solved in three different ways: Exact solution by diagonalizing the Hamiltonian and using (3.4), by Taylor expansion of Uˆ(t) for different expansion orders and different time steps, ∆t (see appendix B.3), and finally ψ(t) is found with the Chebyshev method. As a measure of error we use the maximum| i (or sup) norm: exact Error = max ψi (t) ψi(t) , i {| − |} where ψ (t) is the i’th element in the vector ψ(t) . i | i Figure 3.1 shows the error of both the Taylor and Chebyshev method as function of operations with the Hamiltonian. The hopping energy γ = 3 eV while the on- site energies were set to zero, ε0 = 0, and the total propagation time T = 100 fs. The size of the chain were N = 4000 sites. It is evident that the Chebyshev
0 10
−5 10
Error 3rd order Taylor −10 10 4th order Taylor 5th order Taylor Chebyshev
−15 10 2 3 4 5 6 10 10 10 10 10 Operations with H
Figure 3.1: Comparison between Taylor and Chebyshev expansions method is much more efficient giving a very high accuracy at more than two orders of magnitude fewer Hamiltonian operations. Also, the very abrupt change in error for the Chebyshev method shows the point when the coefficients in the expansion suddenly decay, as discussed in the previous section.
3.3 Continued fraction technique
The local density of states (LDOS) terms like φ δ(E H) φ , where φ is an arbitrary vector, are evaluated using the continuedh | fraction− | i technique.| Thei 3.3 Continued fraction technique 27 numerical details are described in appendix B.1 and B.2. Here we summarize the basic steps in the computation:
1. The Hamiltonian H is mapped to a smaller tridiagonal matrix called Htri:
M ( φ , N ) : H H . (3.11) | i tri → tri The mapping depends on the vector φ and on the number N which | i tri determines the size of Htri. The energy resolution is partly determined by Ntri but also depends on the size of H. The technical details in how the mapping is performed are described in appendix B.2.
2. The local density of the state φ , can be calculated as (see appendix C.1.1): | i
1 dφ(E) = φ δ(E H) φ lim Im G11(E iη) , (3.12) h | − | i ≈ η 0 −π − → 1 where the Greens function G(E) = (E Htri)− . The small imaginary energy η acts as a self-energy term and is− needed to make a smooth LDOS. For η 0 the LDOS for any finite system will be a series of delta peaks → 1 separated by an energy ∆E N − . In all calculations we therefore use a 1 ∝ η N − . ∝ tri 3. The first diagonal element of the Greens function is calculated by the continued fraction:
1 G11(E) = , (3.13) β2 E α 1 − 1 − β2 E α 2 − 2− . .. 2 E αN β Σ(E) − tri − Ntri
where αi and βi are the diagonal and off-diagonal elements of Htri. Usu- ally, the coefficients converge and the remaining terms of the continued fraction can therefore be analytically summed up to give a self-energy, Σ(E).
The most time-consuming step in the calculation is the construction of Htri, since it involves Ntri operations with H. This means that the time it takes to calculate dφ(E) at many energies does not differ much from the time it takes to calculate it for a single energy. This is a major advantage of the Kubo method, since e.g. the conductance can be found for the whole energy spectrum in one calculation. 28 3.4 Overview of the computer codes
3.4 Overview of the computer codes
All computer code used in calculations with the Kubo method is written by myself. Much initial testing and prototyping was carried out in matlab but all larger calculations was performed using the program language Python. This is a relatively easy language, since many standard libraries with build-in functions are readily imported. Also, one should not care about allocating memory for the variables. Moreover, with the extension MPI Python, it is relatively simple to make the programs run in parallel, a feature we have used extensively. In this section we briefly run through the program focusing on the data flow and numerically important parameters.
3.4.1 Important parameters
The important parameters determining the speed and accuracy of the calcula- tions are the following:
Size of the Hamiltonian, Ntot: The total system size determines the maximum propagation time, since no reflection from the end of the wire is allowed. Also, it partly determines the energy resolution (together with Ntri). A large Ntot results in a long computation time, since the matrix operations with the Hamil- tonian depend on the total system size.
Size of the tridiagonal matrix, Ntri: Partly determines the energy resolution. The computation time depends strongly on Ntri, since the construction of Htri requires Ntri operations with the full Hamiltonian, and the subsequent eval- uation of the continued fractions also involves Ntri steps. For systems with closely lying bands such as the SiNWs considered in chapter 7, a large value of Ntri & 3000 is needed. Width of the delta peaks in the density of states, η: The imaginary energy broadens the delta peaks in the density of states making the LDOS smooth, which is intended in order to simulate an infinite system. But it also broadens the van-Hove peaks, and should be as small as possible, still making the LDOS smooth. As mentioned above, we use η Ntri, which generally works satisfying. The computation time does not depend∝η. Random phase state: We use a limited number of random phase states to estimate the traces in (2.12), and they are discussed further in the next chapter. Briefly, they are normalized vectors of length Ntot, with Ni random complex numbers (of the same absolute value) in the middle of the vector. The ends of the vector consist of zeros. Generally Ni should be chosen large to ensure fast convergence of the LDOS, but a large Ni requires also a large Ntot in order for the electron not to ’hit’ the edge as time evolves. The number of random phase states is another important parameter. It is prior to a calculation difficult to say, how many random phase states are needed to give a good estimate of the 3.4 Overview of the computer codes 29 trace, and we have checked for convergence manually. Note, that the estimation of traces is ideal for parallel computation, since each node can calculate parts of the trace individually.
Time vector: A vector tvec = [t1, t2,..., tn] determines the times at which to evaluate the spread of the initial state. The time evolution is performed in steps such that φ(t ) = Uˆ(t t ) φ(t ) . | i+1 i i+1 − i | i i File ID: In order to subsequently reload the results for analysis, each calculation is characterized by a file ID. At the end on the calculations, each node in the parallel program saves the results (with a filename specific for each node) using the python package ’pickle’.
3.4.2 Data flow
The data flow in the program in summarized in figure 3.2. The program uses MPI Python which enables parallel computation on N nodes (node0, node1 ,..., nodeN 1). The input parameters described above, are specified in a main file. − The Hamiltonian is constructed by node0 and broadcasted to the other nodes. We use the package ’spmatrix’ to represent the Hamiltonian in a sparse format speeding up the calculations dramatically and saving a lot of memory. Each of the N nodes constructs a random phase state φ and calculates the | ni local density of state φn δ(E Hˆ ) φn . In a loop over the times t1,..., tN the time evolution of theh | initial− state| togetheri with the local density of states (x) 2 dn (E,ti) = φn(ti) (Xˆ(t) Xˆ(0)) δ(E Hˆ ) φn(ti) is calculated for each time. At the end ofh the calculation,| − each node− saves| thei results in a file specific for that node. If specified, each node can construct several random phase states and repeat the calculation. The results are subsequently reloaded and the total densities of states are esti- mated as the sum of all the local densities, calculated by each node:
n(E) = φ δ(E Hˆ ) φ h n| − | ni n X (x) and in the same way, nx(E,ti) is found as the sum of different dn (E,ti):
n (E,t ) = φ (t ) (Xˆ(t) Xˆ(0))2 δ(E Hˆ ) φ (t ) . x i h n i | − − | n i i n X The time dependent diffusion coefficient is then
n (E,t ) (E,t ) = x i , D i n(E, ) and the results can be extracted from this. 30 3.4 Overview of the computer codes
node_0
Construct Hamilton, H Make ID−files Distribute H and ID−file names to the other nodes Receive H and ID−file names node_1 node_2 node_3
φ Make initial state i φ δ φ d i = ι (E−H) i
Propagate φι(t) (x) φ 2 δ φ d i = i [X(t)−X(0)] (E−H) i Save data (using ID−file names) ...... Load saved data and analyze results
Σ φι δ(E−H)φ Calculate traces n(E) = ι i Σ φ 2 δ φ nx (E) = ι i [X(t)−X(0)] (E−H) i
n (E) x Calculate diffusion coefficient, D(E,t) D(E,t) = n(E) D
Plot results
t
Figure 3.2: Principal data flow in a parallel computation. Chapter 4
The one dimensional chain: Testing of the method
In this chapter we evaluate the numerical implementation, primary using the infinite one-dimensional chain as a test case. There are several reasons for studying this system: First of all it is the simplest one-dimensional system and therefore a natural starting point for a theoretical as well as a numerical analysis. Due to the simplicity, analytical expressions can be found, and thus serve to validate the numerical methods. Secondly, many properties of the one- dimensional chain are also found in more complicated systems. The chapter is organized as follows: In section 4.1 we calculate the local den- sity of state (LDOS) using the continued fraction technique and compare with analytical results. This is followed by section 4.2, where we discuss how to estimate traces using random phase states and show how to calculate various transport properties of the one-dimensional chain with Anderson disorder. To further validate the numerical method, we calculate the mean free path in a dis- ordered carbon nanotube and compare with published results. In section 4.4 we use a model system consisting of two parallel chains to explain some apparent problems with the real-space Kubo-formula. The problems arise for degenerate energies near the band edges.
4.1 Local Density of States
In this section we derive an analytical result for the local density of state in an infinite chain. We use this to test the numerical result, which we obtain using the continued fraction technique (section 3.3), focusing on the parameter Ntri, determining the accuracy. 32 4.1 Local Density of States
4.1.1 The infinite chain
We consider the infinite one-dimensional chain shown in figure 4.1. If we assume only nearest neighbour interaction, the non-zero elements of the Hamiltonian matrix are: Hii = ε0 , Hi,i+1 = Hi+1,i = γ, (4.1) where N /2