Modeling of Leakage Currents in High-Κ Dielectrics

TECHNISCHE UNIVERSITÄT MÜNCHEN Institut für Nanoelektronik

Modeling of Leakage Currents in High-κ Dielectrics

Gunther Christian Jegert

Vollständiger Abdruck der von der Fakultät für Elektrotechnik und Informationstechnik der Technischen Universität München zur Erlangung des akademischen Grades eines Doktors der Naturwissenschaften genehmigten Dissertation.

Vorsitzender: Univ.-Prof. Dr. Th. Hamacher Prüfer der Dissertation: 1. Univ.-Prof. P. Lugli, Ph.D. 2. Univ.-Prof. Dr. P. Vogl

Die Dissertation wurde am 29.09.2011 bei der Technischen Universität München eingereicht und durch die Fakultät für Elektrotechnik und Informationstechnik am 20.12.2011 angenommen.

1. Auflage März 2012

Verein zur Förderung des Walter Schottky Instituts der Technischen Universität München e.V., Am Coulombwall 4, 85748 Garching. Alle Rechte vorbehalten. Dieses Werk ist urheberrechtlich geschützt. Die Vervielfältigung des Buches oder von Teilen daraus ist nur in den Grenzen der geltenden gesetzlichen Bestimmungen zulässig und grundsätzlich vergütungspflichtig.

Titelbild: Electronic transport mechanisms in a high-k thin film capacitor structure

Druck: Printy Digitaldruck, München (http://www.printy.de)

ISBN: 978-3-941650-37-4

Contents

Abstract 1

1 Introduction 3

2 DRAM Technology 11 2.1 DRAM Basics: Operation Principle ...... 13 2.2 Introduction of High-κ Oxides...... 14 2.3 ExperimentalMethods ...... 21 2.3.1 Atomic Layer Deposition ...... 22 2.3.2 Electrical Characterization ...... 25

3 Models for Charge Transport 27 3.1 DirectTunneling ...... 29 3.1.1 Transmission Coeﬃcient and Tunneling Eﬀective Mass . . . . 31 3.1.2 Thermionic Emission and Field Emission ...... 34 3.1.3 Sensitivity Analysis ...... 37 3.2 TunnelingIntoandOutofDefects ...... 41 3.2.1 TheElasticCase ...... 41 3.2.2 Sensitivity Analysis ...... 43 3.2.3 TheInelasticCase ...... 46 3.2.4 Sensitivity Analysis ...... 48 3.3 Defect-DefectTunneling ...... 51 3.3.1 TheElasticCase ...... 52 3.3.2 TheInelasticCase ...... 52 3.3.3 Sensitivity Analysis ...... 54 3.4 Poole-FrenkelEmission...... 57 3.4.1 Sensitivity Analysis ...... 60

i ii CONTENTS

4 Methods for Transport Simulations 63 4.1 KineticMonteCarlo(kMC) ...... 63 4.1.1 Modeling of Leakage Currents:State-of-the-Art vs. kMC . . . 64 4.1.2 History and Development of kMC ...... 65 4.1.3 Basics of kinetic Monte Carlo ...... 68 4.1.4 ASimplekMCAlgorithm ...... 73 4.2 Leakage Current Simulations via kMC ...... 80 4.2.1 Algorithm ...... 82 4.2.2 TimeScalingProblem ...... 83

5 Leakage Current Simulations 85

5.1 High-κ Thin Film TiN/ZrO2/TiN Capacitors ...... 86 5.1.1 Measurements...... 87 5.1.2 kMCSimulations ...... 87 5.1.3 TransportModel ...... 89 5.1.4 Contributions from Diﬀerent Defect Positions ...... 91 5.1.5 Sensitivity Analysis ...... 93 5.1.6 Summary ...... 95 5.2 ElectrodeRoughnessEﬀects ...... 95 5.2.1 FractalGeometry ...... 97 5.2.2 Structural Characterization of the TiN Electrodes ...... 101 5.2.3 Generation of Random Rough Electrodes ...... 101 5.2.4 Mesh Generation and Solution of Poisson’s Equation . . . . . 105 5.2.5 Results and Discussion ...... 106 5.2.6 Summary ...... 114 5.3 Structural Relaxation of Defects ...... 115 5.3.1 Introduction...... 116 5.3.2 GWCalculations ...... 117 5.3.3 DefectModel ...... 119 5.3.4 Kinetic Monte Carlo Framework ...... 122 5.3.5 First-principles kMC Leakage Current Simulations ...... 124 5.3.6 Trap-assisted Tunneling Across Oxygen Vacancies ...... 125

5.3.7 TiN/ZrO2/Al2O3/ZrO2/TiN Capacitors ...... 128 5.3.8 Summary ...... 129

Conclusions and Outlook 131 CONTENTS iii

Bibliography 135

List of Own Publications and Contributions 154

List of Abbreviations 156

Acknowledgements 159

Abstract 1

Abstract

Leakage currents are one of the major bottlenecks impeding the downscaling efforts of the semiconductor industry. Two core devices of integrated circuits, the transistor and, especially, the DRAM storage capacitor, suffer from the increasing loss currents. In this perspective a fundamental understanding of the physical origin of these leakage currents is highly desirable. However, the complexity of the involved transport phenomena so far has prevented the development of microscopic models. Instead, the analysis of transport through the ultra-thin layers of high-permittivity (high-κ) dielectrics, which are employed as insulating layers, was carried out at an empirical level using simple compact models. Unfortunately, these offer only limited insight into the physics involved on the microscale. In this context the present work was initialized in order to establish a framework of microscopic physical models that allow a fundamental description of the transport processes relevant in high-κ thin films. A simulation tool that makes use of kinetic Monte Carlo techniques was developed for this purpose embedding the above models in an environment that allows qualitative and quantitative analyses of the electronic transport in such films. Existing continuum approaches, which tend to conceal the important physics behind phenomenological fitting parameters, were replaced by three-dimensional transport simulations at the level of single charge carriers. Spatially localized phenomena, such as percolation of charge carriers across point- like defects, being subject to structural relaxation processes, or electrode roughness effects, could be investigated in this simulation scheme. Stepwise a self-consistent, closed transport model for the TiN/ZrO2 material system, which is of outmost importance for the semiconductor industry, was developed. Based on this model viable strategies for the optimization of TiN/ZrO2/TiN capacitor structures were suggested and problem areas that may arise in the future could be timely identified. Clearly, such a simulation environment should also be of use for the study of follow- up material systems which, already in the front-end research, show very promising performance, however, combined with great challenges especially in gaining control of the leakage currents. 2

Zusammenfassung

Leckströme in High-k Dielektrika sind eines der wesentlichen Hindernisse für die weitere Miniaturisierung integrierter Schaltkreise. Ein fundamentales, physikalis- ches Verständnis ihres Ursprungs ist daher sehr erstrebenswert. In der vorliegenden Arbeit werden mikroskopische, physikalische Modelle für die relevanten Leckstrommechanismen entwickelt und in einen Simulationsalgorithmus, basierend auf kinetischem Monte Carlo, integriert. Transportsimulationen werden mit einer Auﬂösung einzelner Ladungsträger in drei Dimensionen durchge- führt. Dies erlaubt das Studium räumlich lokalisierter Phänomene wie Perkolation von Ladungsträgern über punktförmige Defekte oder des Eﬀekts von Elektroden- rauigkeiten. Schrittweise wird ein selbstkonsistentes, abgeschlossenes Transport- modell für das technologisch bedeutsame TiN/ZrO2-Materialsystem entwickelt und aus diesem werden Optimierungsstrategien abgeleitet. Chapter 1

Introduction

Introducing high-κ (κ: dielectric permittivity) materials has turned out to be one of the most challenging tasks the semiconductor industry ever had to master. Many problems have to be solved in order to successfully integrate these novel materials. As this is at the very core of the present thesis, the ﬁrst chapter is intended to provide the reader with a sound overview of the topic. Major problems arising upon the replacement of SiO2 with high-κ oxides are highlighted, and the topic of the thesis is put in a wider context.

Referring to the last 40 years, people involved in the integrated circuit (IC) business nostalgically talk of the "days of happy scaling". Indeed, semiconductor industry has been able to keep pace with Moore’s lawI up to now. This strong urge to shrink has been driven by both cost reduction due to more eﬃcient use of substrate area and improved device performance, e.g. switching speed, coming hand in hand as circuit dimensions are decreased. In large part this success story is owed to the unique properties of SiO2. SiO2 can be easily grown in very high quality, i.e. with very low defect densities, via thermal oxidation of silicon and additionally forms a smooth, nearly defect-free interface with the underlying silicon. In fact, this is widely thought to be the major reason why we have a silicon industry today. However, miniaturization lead to a point where the semiconductor industry hit some fundamental physical boundaries, the most serious problems being associated with the transistor gate stack, which consists of the gate electrode (polycrystalline-Si) and the dielectric layer (SiO2) sitting on the silicon channel. Eventually, the thickness

IIn 1965 Gordon Moore, co-founder of Intel, observing the industry of integrated circuits at that time, stated his famous law: "The number of transistors on integrated circuits doubles every two years."

3 4 1. Introduction

of the dielectric film was scaled down to 1.4 nm, a value so thin that intrinsic ∼ quantum mechanical tunneling currents through the dielectric exceeded a tolerable level. Due to these currents in combination with the exponentially increasing density of transistors on a chip power dissipation drastically increased. But, especially nowadays, energy efficiency is extremely relevant, since low-power electronics gain more and more importance, e.g. mobile phones, laptops etc. As a result, after more than 40 years SiO2, at least for some applications, had to be abandonned. New materials, enabling further device scaling, were searched for. The most urgent problem to be addressed was the leakage current. The aim was clear: a reduction of the leakage current without loss in device performance; and the way to go was predetermined. Field-effect transistors are capacitance-operated devices, i.e. the source-drain current depends on the gate capacitance C, calculated from

A C = ǫ κ . (1.1) 0 d Here, ǫ0 is the permittivity of free space, κ is the dielectric permittivity, A is the capacitor area, and d is the oxide thickness. Since tunneling currents exponentially decrease with ﬁlm thickness, one had to replace SiO2 with a physically thicker layer of a new material with a higher dielectric permittivity, a so called "high-κ" oxide, to retain the capacitance and thus the transistor performance. In the year 2007, Intel introduced its 45 nm technology for their new processor generation, which uses a completely new material combination for the transistor gate stack - a hafnium-based high-κ gate dielectric and metal gate electrodes (e.g. TiN)[1]. The above equally applies to the DRAMII industry. Here, due to the ever shrinking chip size, the available area for the cell capacitor, where information is stored in form of an electrical charge, is reduced for each new product generation. However, a certain cell capacitance has to be maintained. Compared to the transistor gate stack, discussed above, which has a more lenient leakage criterion ( 10 2 A/cm2 @ 1 V), ∼ − the requirements are much stricter for the DRAM capacitor to ensure that a suf- ﬁciently large amount of the stored charge is preserved between two refresh cycles ( 64 ms). Leakage currents therefore have to stay below 10 8-10 7 A/cm2 @ 1 V. ∼ ∼ − − Again, high-κ oxides may resolve this issue, since they allow to increase the capacitance. Eventually, high-κ dielectrics have found their way to the market and, at least in certain areas, have replaced SiO2. Today, most of the DRAM producers like

IIDRAM (Dynamic random access memory) is the primary working medium for volatile information storage right across the range of electronic systems. 5

Samsung, Hynix, Micron, and Elpida have shifted to zirconium- and hafnium-based dielectrics[2, 3]. This new class of materials, in principle, provides extendibility of Moore’s law over the next technology generations. But, in contrast to the well-known and well-behaving SiO2, these new materials are mostly unexplored. Meanwhile, the rising commercial interest has triggered huge research efforts dealing with the problems inherently associated with the introduction of high-κ oxides. Above, two major fields of application for high-κ oxides in large-scale industrial production have been identified, the transistor gate stack and the DRAM storage capacitor. Their integration in the transistor implies a lot of problems which are addressed by current research, like the growth of unwanted interfacial layers[4], carrier mobility degradation in the channel due to remote phonon scattering[5], and, not to forget, reliability issues of the dielectric and Fermi level pinning effects[6], which indicate that for the introduction of high-κ oxides in the gate stack the replacement of the polycrystalline-Si electrode by a metal electrode will be a prerequisite. How- ever, in transistor applications, having quite lenient leakage current requirements, the key aspect is the electronic transport in the channel and not the transport across the high-κ layer. This of course will change when, due to further device scaling, gate oxide thicknesses again reach critical values. In the following, since it is at the very core of this thesis, we comment on the application of high-κ oxides in DRAM storage capacitors. Here, due to the rigid leakage current criterion, the focus clearly lies on the insulating properties of the dielectric. Shrinking the DRAM cell dimensions enforces higher capacitance densities to ensure a sufficient cell capacitance. This is achieved either by decreasing the thickness of the dielectric film or by utilizing high-κ materials. The ultimate goal is a metal- insulator-metal (MIM) structure that provides a sufficient capacitance density and concurrently fulfills the leakage current criterion. Besides this, issues arising due to integration and reliability have to be controlled. Since the majority of the high-κ candidate materials are not well-known, no estal- ished recipes for large-scale production of high-quality films are known. To guarantee sufficient insulation, the electrode material has to be chosen in accordance with the dielectric. The electrode work function has to ensure that the potential barrier at each band, conduction band (CB) and valence band (VB), is larger than 1 eV in ∼ order to inhibit conduction via Schottky emission of electrons or holes into the oxide bands. Some oxides additionally need the correct crystalline structure to reach a high permittivity - e.g. TiO2 needs the rutile phase[7]. Typically, the crystal phase has to be induced by the bottom electrode acting as growth substrate. In the effort 6 1. Introduction

to minimize chip area, besides downscaling feature sizes, DRAM cells are more and more designed vertically. Nowadays, storage capacitors with aspect ratios as high as 80:1 are realized in a stack or trench geometry[8], imposing strong requirements on the conformity of ﬁlm growth. Atomic layer deposition (ALD)[9], a method of cyclic deposition and oxidation, allows the most conformal growth of ultra-thin ﬁlms, even in these three-dimensional(3D) structures, and thus is the growth method of choice. The aim of the current ALD research is, starting from precursor chemistry, to develop well-characterized ALD processes and transfer these to the industrial scale. This has resulted in several new processes for materials of interest to the DRAM industry, e.g. ZrO2, HfO2, Sr(Ba)TiO3, TiN, Ru, and Pt[10, 11, 12]. Of these, ZrO2,

HfO2, and TiN have subsequently been studied worldwide and have found their way into the newest chip generations. Unfortunately, the high-κ oxides are materials with a high intrinsic defect concentration. Due to their higher coordination number and their more ionic bonding compared to SiO2 their chemical bonds cannot relax easily. In addition, ALD is known to generally introduce large amounts of impurities, such as C, H, or Cl, depending on the precursor, which further worsen the ﬁlm quality.

Figure 1.1: Band gap Eg of high-κ oxides vs. permittivity κ, showing an inverse scaling behaviour, i.e. E 1/κ. Currently tetragonal ZrO (t-ZrO ) is employed g ∼ 2 2 featuring a permittivity close to 40 and a band gap of 5.5 eV. ∼

Much of the present-day research of the high-κ oxides thus consists of pragmatic 7

strategies to reduce defect densities by processing control and annealing[13, 14, 15]. The high intrinsic and process-induced defect density comes together with the fact that the κ-value of the candidate oxides tends to vary inversely with the band gap, see Fig. 1.1, meaning highest-κ materials are rather weak insulators. Against this backdrop the present work was initialized to identify immediate needs for action. Viable strategies for leakage current reduction were searched for. A profound understanding of electronic transport in high-κ ﬁlms will be a prerequisite for the realization of future DRAM cell capacitor generations, and it will also be very useful for the transistor gate stack where leakage currents will become an issue in the medium term.

Modeling of Leakage Currents: The State-of-the-art

To gain deeper insights into the important transport mechanisms, a huge amount of experimental leakage current data has been collected for the electrode/high-κ oxide systems of interest. However, the analysis of the data, so far, typically remained on the level of simplified compact-models that can be evaluated very fast[16, 17, 18, 19]. Often the functional form of the leakage currents could be reproduced quite well, but only if unrealistic or even unphysical fitting parameters were used[20, 21]. This is due to the fact that these models, at best, only grasp certain aspects of the observed transport and neglect essential parts, describing for instance only one single step in a multistep transport process. A prominent example is the standard Poole-Frenkel (PF) treatment[22], which only accounts for the field-enhanced thermal emission of charge carriers localized at defect states, but does not describe the carrier injection step. At worst, these models are misused in the wrong context, e.g. the often cited Richardson-Dushman equation was originally intended to describe thermal emission of electrons from metals into the vacuum[23, 24]. Parameters extracted from such an erroneous analysis can be misleading. More sophisticated transport models were proposed, e.g. the "dead layer" model, suggesting bulk limitation of the electronic transport in SrTiO3 and (Ba,Sr)TiO3[25]. The drawback of this model is that good fitting of experimental results is only achieved at elevated temperature ( 400 K) and low defect densities, since trans- ∼ port across defect states is not considered. A similar model considers tunneling through the interfacial layers to dominate the leakage current[26, 27]. For hafnium- and zirconium-based dielectrics a tunneling-assisted PF conduction mechanism has been suggested[28] that expands the standard PF treatment. Also a first effort to embed a hopping transport model via multiple traps in a Monte Carlo algorithm 8 1. Introduction

was undertaken[29]. To summarize, existing approaches usually are tailored to describe certain material systems and thus focus on certain (isolated) aspects of transport. They fail at providing a closed model for the transport. Often diﬀerent transport mechanisms are present at the same time, and these turn out to be mutually interdependent, e.g. in case of a charge carrier localized in a defect state which might either escape via tunneling or thermal emission - competing with each other, the stronger one mechanism is, the weaker is the other.

Towards the modeling of leakage currents via kinetic Monte Carlo

It was a major goal of this thesis to provide closed models for electronic transport through high-κ thin films and to incorporate these into simulation tools. A global picture encompassing all relevant transport mechanisms and their interdependence was developed. Effects arising in systems with ultra-thin high-κ oxide films, e.g. the influence of rough electrode/dielectric interfaces leading to local electric field enhancement or film thinning, or structural relaxation of defects that are involved in defect-assisted tunneling, were addressed. Modeling and simulation, in contrast to most of the existing approaches, was carried out in three dimensions to avoid ar- tifacts due to lower dimensionality. This allowed for the investigation of phenomena that are localized in all three dimensions like transient degradation of the dielectric film, i.e. the accumulation of point-like defects, which opens a percolation path for charge carriers shortening the electrodes. A kinetic Monte Carlo (kMC) approach was developed, which provides fascinating possibilities, since kMC simulations take place at the level of individual charge carriers. Complex phenomena like carrier- carrier interaction and its implications for the transport or transient phenomena like charge trapping can be investigated in a straightforward manner. The ultimate goal was to provide the theoretical background that allows for a deep understanding of the electronic transport in high-κ films and thus supports the experimental efforts in target-oriented system optimization and the identification of future device scaling potentials.

Focus was laid on the study of the state-of-the-art TiN/ZrO2 material system for the DRAM storage capacitor, which is employed in current chip generations. As future material systems, based on TiO2 and SrTiO3, are still far away from being ripe for their introduction into the production process, optimization of the TiN/ZrO2 system will be the key enabler for further scaling in the next years. The structure of the present thesis is as follows. To get the reader, who is maybe 9

not a specialist in DRAM technology, started, chapter 2 provides the basics of DRAM technology, beginning with the functional principle of a DRAM cell and a general survey of nowadays DRAM business. Implications of the downsizing efforts of the semiconductor industry are discussed and the topic of the present thesis is put in a wider context, showing its significance for this technology field. At this point, experimental details on the fabrication and characterization of the investigated high-κ capacitor structures are supplied. The following chapter 3 sets the physical basis for the simulations carried out in the framework of this thesis by introducing the models for the charge carrier transport mechanisms that are relevant on the nanoscale. Complementary, sensitivity analyses are carried out to establish a deeper understanding of the system parameters and their significance for the device performance. Chapter 4 introduces the kinetic Monte Carlo techniques. Providing some historical background, the generic kinetic Monte Carlo algorithm is derived, its main advantages but also its limitations are discussed. Subsequently, its implementation and application for leakage current simulations in high-κ dielectrics is presented. Chapter 5 may be regarded as the core of this thesis, the traditional "Results and Discussion" part. Kinetic Monte Carlo leakage current simulations of different kinds of capacitor structures are presented. They are contrasted with measurement data. The whole process of developing a transport model for TiN/ZrO2/TiN and

TiN/ZrO2/Al2O3/ZrO2/TiN capacitors is sketched. Higher order effects, such as electrode roughness, are systematically taken into account. Finally, a brief conclusion is given, intended to sum up the most significant findings of the present thesis. Possible fields and trends of future research are highlighted in a concise outlook, encompassing the application of the simulation framework, established in this thesis, to future DRAM material systems as well as the adaption of the simulator to other interesting devices, such as organic solar cells. 10 1. Introduction Chapter 2

DRAM Technology

DRAM production is a highly competitive business and todays global market is shared among a handful of huge companies, which are capable to raise the gigantic amounts of capital needed to built up production facilities and to buy the processing equipment. DRAM industry traditionally is very cyclic. In good times with high market prices the top-tier companies can pile up huge profits. In the fourth quarter of 2010, e.g., market leader Samsung controlled roughly 40% of the entire market, which amounts to a business volume of 2.5 billion euros. In such phases the companies heavily invest into additional production capacities in order to expand their market share. Eventually, the market starts to suffer from the resulting oversupply of DRAM chips, causing a massive plunge of the prizes. In such a situation an intrinsic problem of the DRAM market emerges. The amount to which the companies can react to the market situation by reducing their chip production is very limited. Firstly, decreasing the chip production does not lead to a significant reduction of the overall production costs, and, secondly, beyond a certain point it is simply not possible, as shutting down a production line would irreversibly lead to the destruction of the equipment. Thus, the situation of oversupply remains until the market gains enough momentum to generate a sufficient request for the chips, which drives the prizes up again. In short words, DRAM production is one of the harshest business environments one can think of. As the functionality that has to be provided by the memory chips is strictly standard- ized in nearly every detail, the companies do not have the possibility to differentiate much from each other by providing a superior chip functionality. Thus, to gain an advantage about the competitors, one has to manufacture the same product at lower cost. Grinding down the suppliers of the raw materials, i.e. especially the silicon wafers,

11 12 2. DRAM Technology

200

180

DRAM nodes

ITRS Forecast 2010

160

140

120

100

60 half-pitch (nm)

0 2000 2002 2004 2006 2008 2010 2012 2014 2016 2018 2020 2022 2024

year of production

Figure 2.1: Evolution of DRAM node half-pitches shown for chip generations since 2000. Predicted values until 2024 are based on the ITRS. and the manufacturers of the production tools is no viable strategy to pursue, as the small set of these companies is shared by all the competitors. They would just switch over to another company. So, the cost advantage has to be designed in. A rough measure for this, neglecting differences arising to process complexity, is, how efficiently the silicon wafer area is used. Since a wafer costs the same for all companies, the one wins, who produces the most chips from one wafer. This urge keeps the miniaturization process alive and its speed at a high level, as can be seen in Fig. 2.1. Here, the evolution of the minimum feature size, the so called half-pitch, over the past eleven years and a forecast based on estimations of the International Technology Roadmap for Semiconductors (ITRS) are shown. In the following, we will first, in 2.1, outline the basic operation principle of a single DRAM cell, and then, in 2.2, we will comment on how this downscaling of DRAM chips is realized via technology and design improvements. 2.1. DRAM Basics: Operation Principle 13

2.1 DRAM Basics: Operation Principle

Whenever information has to be stored in a way that allows quick access, Dynamic Random Access Memory (DRAM) is the medium of choice. While the miniaturization has driven the DRAM production process to an extreme level of sophistication, the working principle of the DRAM cell, being very simple, has not changed during the last decades.

Figure 2.2: Schematic picture of a DRAM cell consisting of a capacitor and a transistor. Via the wordline a voltage can be applied to switch on the transistor and allow access to the storage capacitor. Readout and writing of the capacitor is done via the bitline.

As depicted in Fig. 2.2, a single DRAM cell consists of a capacitor and a transistor, connected to a grid of perpendicular, conducting lines. The information is stored in form of an electrical charge on the capacitor, the so called storage node or cell node. Either the capacitor is charged ("0") or not ("1"). Thus, every cell can store one bit of digital information. Access to the capacitor is controlled by the transistor. Its gate is connected to the wordline (WL). By changing the bias applied to the WL, access to the storage node via the bitline (BL) can be controlled by opening or blocking the transistor channel. A readout process is carried out by applying a voltage to the WL, switching on the transistor channel. The stored information is sensed as a voltage change in the BL, caused by the charge ﬂowing from the storage node onto it. To write to the memory, the BL is forced to the desired high or low voltage state, and the capacitor is charged or discharged. In reality, due to the non-ideality of the dielectric ﬁlm, acting as insulating layer 14 2. DRAM Technology

between the capacitor electrodes, the charge stored on the capacitor, i.e. the information, is lost over time. Therefore, the information has to be refreshed in regular intervals. The JEDEC (foundation for developing semiconductor standards) has deﬁned a refresh cycle length of 64 ms or less. This is the reason why this type of memory is called "Dynamic" Random Access Memory. It is one of the central topics of this thesis to investigate the loss mechanisms involved in the transport of charge across the dielectric layer. These so called leakage currents are studied in detail, invoking Monte Carlo techniques to simulate the transport phenomena. A better understanding of the underlying physics is intended to foster the leakage current minimization and thus to extend miniaturization potentials.

2.2 Introduction of High-κ Oxides

While the principle of a DRAM cell is very simple, as discussed in 2.1, its realization is very complex due to the urge of minimizing the wafer area per cell. Over the last years, this has lead to a growing degree of vertical integration. Exemplarily, Fig. 2.3 shows a scanning electron microscopy (SEM) picture of a cut through a modern 90 nm DRAM cell with deep trench capacitor. In order to save wafer surface area the capacitor is realized as a trench, which is etched deep down into the Si wafer. The capacitor electrode, on which the charge is stored, is marked blue. It is separated by an ultra-thin dielectric layer (typical thicknesses nowadays are below 10 nm) from the plate, the grounded counter-electrode of the capacitor, which is shared among the individual cells. Here, the Figure 2.3: SEM cut through dielectric is only visible as a thin, white line. a 90 nm deep trench DRAM cell Also the WL, running perpendicular to the (source: Qimonda AG). cutting plane, and the transistor channel directly below it can be seen (marked red). In the upper part of the picture, BL and BL contact to the cell are marked green. The ingenuity of engineers has allowed 2.2. Introduction of High-κ Oxides 15

Figure 2.4: SEM image of 90 nm deep trench capacitors and zoom on the top part of a DRAM cell (source: Qimonda AG). aspect ratios as high as 1:80 for the trench capacitor, as can be seen in Fig. 2.4. Due to the shrinking DRAM cell dimensions the size of every single cell component has to be reduced. A rigid criterion for the cell capacitor is set by the desired minimum cell capacitance. In DRAMs the digital information is read by detecting the change in the BL voltage ∆VBL = αVddCcell/(CBL +Ccell) that is caused by the charge flow from the storage node onto the BL. Here, α 0.5, V is the operational ∼ dd voltage, and Ccell and CBL are the capacitances of the cell capacitor and the BL, respectively. To ensure a reliable information readout, a voltage change ∆V BL ≥ 0.1 V is necessary to differentiate between "0" and "1". This can be translated to a minimum capacitance that has to be maintained. So, every loss of capacitor area has to be counterbalanced by an increase of capacitance per area (capacitance density). The situation is worsened by the fact that soon a very drastic change in the capacitor design will inevitably take place. Currently, the cell capacitor is realized as a hollow cylinder, as shown in Fig. 2.5, allowing to use both the inner and the outer surface as capacitor area. Within the upcoming node generations the shrinking lateral cell dimensions will enforce the switch to a block capacitor design. This means a massive loss of area which will make drastic increases of the capacitance density necessary. Let us quantify this. Capacitancy density usually is measured in terms of equivalent oxide thickness (EOT). The EOT gives the thickness of a SiO2 dielectric film required 16 2. DRAM Technology

Figure 2.5: Hollow cylinder and block design for the cell capacitor.

ITRS forecast 0.6

0.4

0.2 EOT (nm)

0.0 2010 2012 2014 2016 2018 2020 2022 2024

year of production

Figure 2.6: Prediction of required EOTs for the storage capacitor in the next years according to the ITRS.

to reach a certain capacitance density. With κ(SiO2)= 3.9 the EOT of a capacitor using a dielectric with permittivity κ and thickness d, can be calculated from

3.9 EOT = d. (2.1) κ · The currently best capacitor concepts, suited for large-scale production, feature

EOTs around 0.7 nm[30] and employ a trilaminate ZrO2/Al2O3/ZrO2 dielectric. Fig. 2.6 shows the EOT requirements for the next years according to ITRS estimates, one year being roughly equivalent to one chip generation. As can be seen, the downscaling of the capacitor has already fallen two years behind the schedule. Only a few years after successful integration of high-κ dielectrics the semiconductor industry 2.2. Introduction of High-κ Oxides 17

seems to have reached a point where further scaling proves to be very hard. Via clever cell architectures, up to now, the requirements could be mitigated.

But why is that? To understand what makes it so hard to increase the capacitance density and to fulﬁll the leakage current criterion at the same time, let us have another glance on the formula for the planar plate capacitor, given in equation (1.1). In principle, one has two options.

First, one can shrink the thickness of the dielectric layer. But, as the operational voltage is downscaled only very weakly between two node generations, this approach leads to higher electric fields being present in the dielectric. These growing fields, in turn, cause an increase of leakage currents, as the most transport mechanisms strongly scale with the electric field. Additionally, higher electric fields may cause early breakdown of the dielectric.

The alternative approach is to increase the permittivity κ, inevitably requiring the introduction of a new dielectric material, which typically comes along with many problems connected with, e.g., the development of suitable industry-scale growth processes and their integration into the overall production ﬂow.

In the past, the latter has proven to be a Herculean challenge, one only meets if it cannot be avoided. Therefore, SiO2 was scaled down until the limit of quantum mechanical tunneling was reached. As has been mentioned earlier in chapter 1, there is an unfortunate correlation between the insulating properties of a material and its dielectric properties. The higher the permittivity of a material is, the lower is its band gap, making it a worse insulator. To get a rough feeling of the requirements on the materials, one can make a simple estimate.

Let us assume that we have introduced a new dielectric material with permittivity κ and a conduction band offset (CBO) Φ with respect to the Fermi level of the employed electrodes. The CBO defines the height of the tunneling barrier, as hole transport usually is many orders of magnitude lower due to the much larger VB offset. We further assume that the film is perfect, i.e. free of any kind of defects. As long as we are above the limit of quantum mechanical tunneling, ensured by a sufficient film thickness, in this scenario leakage currents will be dominated by Schottky emission, i.e. thermionic emission of charge carriers from the electrode into the CB of the dielectric. This transport mechanism will be discussed in detail in 3.1.2. For now, we simply estimate the amount of leakage current via the Richardson-Dushman equation for the current density jSE due to Schottky emission: 18 2. DRAM Technology

Figure 2.7: Current density due to Schottky emission as a function of the CBO, calculated from eqn. (2.2). The dashed green line marks the current criterion for the storage capacitor.

emk2 e eF j = B T 2 exp Φ . (2.2) SE 2 3   v  2π h¯ · −kBT − u4πǫ0ǫopt u   t  Here, m is taken as the free-electron mass, F is the applied electric ﬁeld, and ǫopt is the optical permittivity of the dielectric material. The requirement on the leakage 7 2 current for the storage capacitor is j < 10− A/cm for an applied electric ﬁeld of

F = 1 MV/cm and a temperature of 125◦C. ǫopt is exemplarily set to 5.6[31], the value of t-ZrO2, currently the state-of-the-art dielectric in the storage capacitor. With these values, a minimum CBO of 1.3 V is necessary, as can be seen in Fig. ∼ 2.7. This value of Φ has to be understood as a minimum requirement. As will be seen, meeting this criterion alone is by far not sufficient. But, just by regarding the evolution of the CBO with the introduction of high-κ oxides, one gets a first impression of the problems, which may arise. In Fig. 2.8 the most promising electrode/dielectric material combinations are listed together with an estimate of the CBO, which is to be expected. The displayed values are calculated from the difference of the electrode work function Φm and the electron affinity Ξa of the dielectric, i.e. it is implicitly assumed that the interfaces are ideal. This case is called the Schottky-limit[32] and the listed values thus only constitute an upper limit for the CBO. Realized CBOs 2.2. Introduction of High-κ Oxides 19

Figure 2.8: CBO values for electrode/dielectric material combinations in volts. Ideal interfaces, i.e. without Fermi-level pinning or charge transfer are assumed. Thus, the CBO is calculated as Φ = Φ Ξ , Φ being the electrode work function m − a m and Ξa the electron aﬃnity of the dielectric. can be signiﬁcantly reduced due to Fermi level pinning caused by interface states in the band gap or charge transport across the metal/insulator interface leading to the formation of an interface dipole.

As can be seen, for the currently employed TiN/ZrO2 material, with a CBO of 2.2 V in the Schottky-limit, no noteworthy leakage currents due to Schottky emission are to be expected. In contrast, for the next generation of high-κ materials, i.e. TiO2 and SrTiO3, problems may arise. While Pt promises to be the best suited electrode for these materials in terms of CBO, its integration into the DRAM production process simply would be too costly and would also rise a lot of issues connected with the processing of Pt. And, experimentally, the actual CBO, e.g. at the Pt/TiO2 interface, was found to be only 1.05 V[33]. TiO2 in its rutile high-κ phase has a band gap of 3 eV[34]. The fact that the Fermi level of the metal electrode is shifted ∼ towards the CB of TiO2 is caused by the n-type character of TiO2 due to self-doping with oxygen vacancies[35]. Thus, p-doping is investigated as a measure to increase the CBO for TiO2 and SrTiO3. For TiO2 it has been demonstrated that doping with Al increases the CBO by 0.5 V to 1.54 V[36]. This is close to the optimum value of this material, since a larger CBO would correspond to a too small VB oﬀset giving rise to hole injection. Besides shrinking CBOs with increasing κ, other severe problems have to be addressed, e.g. to ensure growth in the correct crystalline phase. While rutile TiO2 has a permittivity larger than 100, in its anatase phase the permittivity is only 20 2. DRAM Technology

around 40[37]. So far, the only successful approach to reach the rutile crystalline phase for TiO2 thin films was to induce the correct crystalline structure via the substrate, i.e. the bottom electrode of the capacitor structure, for which Ru was used. On top of the Ru a thin, rutile RuO2 layer was formed prior to the deposition of the dielectric. And, when extremely high values for the permittivity are discussed, one usually refers to the bulk value of the materials. Unfortunately, however, especially for the ultra-high-κ materials discussed above, the permittivity has been found to be thickness dependent. At the metal/dielectric interfaces thin layers with a siginificantly reduced dielectric constant are formed, so called "dead layers". Obviously, this is detrimental for the scaling of the capacitor, as it reduces the effective dielectric constant of the whole film. Consequently, many groups have studied this phenomenon and tried to track down its origin. Both intrinsic and extrinsic effects have been discussed, e.g. a distortion of the dielectric response of the insulator due to imper- fections, secondary phases or interdiffusion at the interface, or the finite penetration length of the electric field into the metal electrode[38]. A detailed overview of the current knowledge is given in ref. [33]. The multitude of the above-discussed aspects, which have to be considered concerning leakage currents, gives a first impression of the complexity of the task to downscale the capacitor, while ensuring sufficiently low leakage currents. However, the above is still just a part of the problems that one encounters on the path of downscaling. So far, we emanated from the assumption of perfect, defect-free dielectric thin films. Thus, the above-mentioned problem areas cover only the intrinsic problems associated with the introduction of high κ materials, i.e. low CBO, dead − layers, stabilization of the desired crystalline structure. Now we drop the assumption of perfect, defect-free dielectrics, since it is far from reality. Indeed, all high-κ candidate materials are known to suffer from high defect densities[39]. This is due to their high coordination number, hindering the relaxation and rebonding of broken bonds at a defect site. Often, defects create localized electronic states in the band gap, so called defect states. These give rise to another class of transport mechanisms, subsumed under defect-assisted transport (DAT). Charge carriers, here, are thought to tunnel/hop across defect states, which, acting as stop-overs, lead to drastically increased transmission probabilities of particles through the barrier set up by the dielectric. A detailed introduction to the zoo of transport mechanisms relevant in high-κ thin films will be given in chapter 3. For now, it is only important to be aware of these additional transport pathways. 2.3. Experimental Methods 21

The above in mind, the steps in optimizing a material system, i.e. a metal/high-κ dielectric combination, are clear. First, it has to be ensured that intrinsic currents are sufficiently low by providing the minimum CBO of Φ = 1.3 V. Then, extrinsic transport has to be dealt with. The research effort, here, is focused on the reduction of defect densities, e.g. via suited annealing strategies[40] or their passivation via doping with other elements. Doping with La, Sc, or Al, for example, has been demonstrated to passivate oxygen vacancies in HfO2 by repelling the defect states out of the band gap[41] and also doping with F has been shown via ab initio calculations to have an advantageous effect[42]. But, in order to guide this effort one first has to identify the defects, which open up the transport channels, i.e. a transport model has to be set up for a material system. In the present work, we intend to present a scheme how such a model can be systematically developed. We lay emphasis on the constant exchange between simulations and experiment. After an elaborate discussion of the basic transport mechanisms in chapter 3, we will, in chapter 4, outline a novel, kMC-based simulation approach to the complex transport phenomena in high-κ films. Having established the routines, a closed transport model for the currently employed state-of-the-art material system

TiN/ZrO2 will be presented. Such a model is very useful in pointing out strategies for the minimization of leakage currents.

2.3 Experimental Methods

At this point, although the focus of this work is not on the experimental side, we will give a concise overview of the process ﬂow for fabrication of the TiN/ZrO2/TiN capacitors, which are investigated here with respect to their electronic properties. Also characterization methods will be described. For further details, the reader should refer to the dissertation of Wenke Weinreich on the "Fabrication and characterization of ultra-thin ZrO2-based dielectrics in metal-insulator-metal capacitors"[43]. There, a very detailed overview over experimental details is given, also for exactly those capacitor structures which are investigated here from a theoretical point of view. All samples investigated in this thesis were fabricated by the Qimonda AG in collaboration with the Fraunhofer Center Nanoelectronic Technology in Dresden. 22 2. DRAM Technology

Figure 2.9: (a) TEM picture of 2.5 µm deep trenches, etched into a Si substrate and subsequently coated via ALD with ZrO2 (taken from ref. [43]). (b) TEM picture of a ZrO2/Al2O3 multi-layer structure, formed via ALD.

2.3.1 Atomic Layer Deposition

Atomic layer deposition is a method for the chemical growth of thin films from the vapour phase. It is based on alternating, saturating surface reactions[44]. ALD guarantees film deposition with a high degree of conformity and homogeneity. Accurate film thickness control and high reproducibility is ensured via a cyclic growth procedure. The self-limitation of the involved chemical surface reactions allows conformal coating of three-dimensional structures, as they are found in the case of the DRAM storage capacitor, which is either realized as a trench, etched down in the Si-wafer as shown in Fig. 2.9 (a), or in a stack geometry. Composition of a multi-component material can be controlled with a resolution down to monolayers of atoms (Fig. 2.9 (b)). As said, ALD is a cyclic process, whereas one growth cycle can be subdivided into four steps, as sketched in Fig. 2.10. (1) Pulse of educt or precursor A until the surface is saturated with precursor molecules. (2) Purging of the growth reactor with an inert gas to remove excessive educts and volatile byproducts. (3) Pulse of educt or precursor B until the surface is saturated with precursor molecules. (4) Purging of the growth reactor with an inert gas to remove excessive educts and volatile byproducts. Steps (1)-(4) are repeated until the desired film thickness is deposited. ALD is a strictly surface-controlled process. Thus, other parameters besides reac- 2.3. Experimental Methods 23

Figure 2.10: ALD ﬂow scheme. (top) substrate surface before ALD growth, (1) pulse of educt A, (2-1) purging of growth reactor with inert gas, (3) pulse of educt B, (4-1) purging of growth reactor with inert gas, (2-2) repetition of step (2-1), (4-2) repetition of step (4-1). 24 2. DRAM Technology

tants, substrate, and temperature have no or only negligible influence on the film growth. Of course, a real ALD growth process suffers from non-idealities, such as incomplete chemical reactions, decomposition of the precursor materials, and especially incorporation of impurities in the growing film. The latter are in the case of high-κ dielectrics comparably high. Electrically active defect states, leading to large leakage currents, can be the consequence. Therefore, a huge effort is undertaken in order to optimize the ALD growth recipes (e.g. via the choice/development of suitable precursors[44]) and minimize impurities.

Fabrication of the Capacitor Structures

For the fabrication of the TiN/ZrO2/TiN (TZT) capacitors, which are investigated in this thesis, (001)-silicon wafers were coated with a bottom electrode consisting of

10 nm of chemical vapour-deposited TiN, with TiCl4 and NH3 as precursors, grown at 550◦C[45]. Subsequently, the ZrO2 ﬁlm was grown by ALD at 275◦C and 0.1 Torr vapour pressure with a Jusung Eureka 3000, a single-wafer (300 mm) processing equipment of Jusung Engineering (Korea). Here, Tetrakis(ethylmethylamino)- zirconium (TEMAZr) was employed as metal precursor and ozone (O3) as oxi- dizing agent. A high growth rate of 0.1 nm/ALD cycle was reported for this ∼ process[46, 47, 48]. For ALD growth of Al2O3 ﬁlms Trimethylaluminum (TMA) was used as metal precursor. The top TiN electrode was deposited at 450◦C by chemical vapour deposition.

TiN/ZrO2/Al2O3/ZrO2/TiN (TZAZT) capacitors, studied in 5.3, were additionally subjected to a post-deposition anneal at 650◦C in N2 prior to TiN top electrode deposition in order to crystallize the ZrO2 layers. As grown, the thinner layers of ZrO ( 3.5-4.5 nm) employed in the TZAZT capacitors are amorphous, which is not 2 ∼ surprising, since crystallization of ZrO2 ﬁlms is known to be thickness dependent[3]. 2.3. Experimental Methods 25

2.3.2 Electrical Characterization

For the electrical characterization of the TZT capacitors, temperature-dependent current measurements were carried out on a heatable wafer holder with a Keithley 4200 parameter analyzer, delay times of 2 s, and a voltage ramp rate of 0.05 V/s. Additionally, capacitance measurements were carried out using an automated setup PA300 from SüssMicroTec and an Agilent E4980A, whereas the standard measuring frequency was 1 kHz, and the voltage amplitude was set to 50 mV. From capacitance measurements the eﬀective permittivity of a dielectric ﬁlm can be easily extracted via

Cd κ = , (2.3) ǫ0A whereas C is the measured capacitance, d the thickness of the dielectric ﬁlm, and A the capacitor area. 26 2. DRAM Technology Chapter 3

Models for Charge Transport

In the preceding chapter the multitude of possible mechanisms by which charge carriers can overcome the barrier for transport set up by the dielectric layer was already mentioned and divided into two categories, intrinsic mechanims, i.e. those which are always present, even if a dielectric film of perfect quality is assumed, and extrinsic mechanisms, which are related to the existence of defect states. As will be discussed in detail in chapter 4, if one plans to carry out kMC transport simulations, rate models for every relevant transport process are crucial input. kMC, as will be described, enables the study of the statistical interplay between the different transport mechanisms. Thus, in this chapter, a detailed discussion of all important transport mechanisms and their sensitivity towards system parameters and external parameters is given. This is intended to form the basis for the following investigations, carried out on the currently most important material system TiN/ZrO2. In the following we will limit the discussion to electron transport, since hole transport in the relevant material systems is negligible due to the comparably large VB offset. The set of transport mechanisms, which are discussed in the remainder of this chapter and are implemented in the kMC simulation tool, is sketched in Fig. 3.1. These encompass both, intrinsic mechanisms, such as (i) direct tunneling (which merges into Fowler-Nordheim tunneling for high voltages) and (ii) Schottky emission, as well as extrinsic mechanisms, such as (iii) elastic/inelastic tunneling of electrons into and (iv) out of defects, (v) tunneling between defects, and (vi) field-enhanced thermal emission of electrons from defects, so called PF emission. In an extra section we will address the role of structural relaxation of defects upon a change of their charge state, inducing a shift of the defect level energy (not shown).

27 28 3. Models for Charge Transport

Figure 3.1: Schematic band diagram of a MIM structure with applied voltage U. Transport mechanisms, implemented in the kMC simulator: (i) direct/Fowler Nord- heim tunneling, (ii) Schottky emission, (iii) elastic and inelastic tunneling (p = number of phonons with energy hω¯ ) into and (iv) out of defects, (v) defect-defect tunneling, (vi) PF emission. 3.1. Direct Tunneling 29

3.1 Direct Tunneling

Referring to direct tunneling, one usually means coherent tunneling through an energetically forbidden barrier region, which, in our situation, is set up by the band gap of the dielectric material. Thus, the barrier height is determined by the CBO of the high-κ oxide, and its width is given by the layer thickness. In the following we will derive the standard expression for direct tunneling, the so called Tsu-Esaki tunneling formula[49], since it is a very introductive exercise in order to get familiar with the relevant physics. For an in-depth study the reader is referred to ref. [50]. We consider elastic, energy conserving tunneling through a barrier that varies only in transport direction and is constant and of infinite extent in the two lateral dimensions. Conservation of the lateral momentum of the tunneling electron is assumed. Effects of inhomogeneities, impurities or interface roughness only induce minor quantitative deviations but do not lead to a qualitatively different behaviour. Thus, we neglect them.

Figure 3.2: Band diagram for a MIM The general problem we are dealing with, is structure setting up a tunneling bar- illustrated in Fig. 3.2. Here, a MIM struc- rier for electrons. Charge ﬂow is indi- ture with an applied bias U is shown, set- cated. After tunneling thermalization ting up a generic tunneling barrier for elec- takes place. trons. The two metal contacts, or, more general, charge carrier reservoirs, are assumed to be in equilibrium. A Fermi-Dirac distribution of the electrons in the electrodes is assumed. In case of current ﬂow, this is not exactly true, but again, we neglect minor deviations arising from this fact. The CB bottom of the left contact is chosen as zero-reference of the potential energy. Assuming a single, parabolic, and isotropic CB minimum, we write the energy of the tunneling electron to the left (l) and to the right (r) of the barrier as

2 2 2 2 h¯ kx,l h¯ ky,z,l El = Ex,l + Ey,z,l = + (3.1) 2m∗ 2m∗ and 30 3. Models for Charge Transport

2 2 2 2 h¯ kx,r h¯ ky,z,r Er = Ex,r + Ey,z,r = + + Ec,r. (3.2) 2m∗ 2m∗

Here, kx and ky,z are the momenta in the transport direction and the two lateral directions. Ec,r denotes the CB minimum of the right contact. Since we assume that the lateral momenta are conserved, i.e. ky,z,l = ky,z,r, we arrive at

2 2 2 2 h¯ kx,l h¯ kx,r Ex,l = = + Ec,r. (3.3) 2m∗ 2m∗ After tunneling, the electron is assumed to thermalize. Via inelastic collisions excess energy is tranferred to the reservoir and the electron loses memory of its previous state. The overall electron current is calculated, in this picture, as net diﬀerence between the electron current from the left to the right jl and the current from the right to the left jr. We consider the current ﬂow in x direction for a given energy E with a component

Ex,l in transport direction. Particles in an inﬁnitesimal volume element of momentum space dkl around kl cause a current density ji,l impinging from the left onto the barrier, calculated as

j = eN(k )f (k )v (k )dk . (3.4) i,l − l l l x l l 3 fl denotes the Fermi-Dirac distribution in the left contact and N = 1/(4π ) the density of states in k-space. The velocity perpendicular to the barrier from the left is

1 ∂E(kl) hk¯ x,l vx(kl)= = . (3.5) h¯ ∂kx,l m∗ To calculate the current density transmitted to the right, equation (3.4) is multiplied with the transmission probability T (kx,l), which, for the above simpliﬁcations, is only a function of the momentum in transport direction. Thus, we have

eh¯ jl = 3 T (kx,l)fl(kl)kx,ldkx,ldky,z,l. (3.6) −4π m∗ In an analogous way one arrives for the current transmitted to the left for given energies E and Ex at:

eh¯ jr = 3 T (kx,r)fr(kr)kx,rdkx,rdky,z,r. (3.7) −4π m∗ 3.1. Direct Tunneling 31

2 From equation (3.3) it follows that kx,ldkx,l = kx,rdkx,r = m∗dEx/h¯ . If we additionally use the symmetry of the transmission coeﬃcient with respect to the tunneling direction, we can calculate the total net current ﬂow from

2π e ∞ ∞ jDT = dEx dk′k′ dφT (Ex) fl(Ex,k′) fr(Ex,k′) . (3.8) 4π3h¯ 0 0 0 − Z Z Z h i As mentioned above, we assume that the electrons in the contacts are distributed according to the Fermi-Dirac function given by

1 f (Ex,Ey,z)= , (3.9) l/r (Ex+Ey,z E )/(k T ) e − f,l/r B whereas Ef,l = Ef,r +eU. Again using the assumption of parabolic bands, we arrive at

em∗ ∞ ∞ jDT = 3 dExT (Ex) dE′E′ fl(Ex,E′) fr(Ex,E′) (3.10) 2π2h¯ 0 0 − Z Z h i (Ef,l Ex)/(kBT ) em∗kBT ∞ 1+ e − = dExT (Ex)ln , (3.11) 2 3 (Ef,l eU Ex)/(kBT ) 2π h¯ Z0 1+ e − − ! the Tsu-Esaki formula for tunneling currents. All information on the contacts is contained in the effective mass and the logarithmic term, while information on the tunneling barrier is contained in the transmission coefficient. As said, some uncritical assumptions were made concerning the electron distribution within the contacts and the band structure - cf. page 29. Dropping these does not lead to major modifications of the tunneling current. The crucial point in carrying out studies of tunneling currents is how one deals with the transmission coefficient.

3.1.1 Transmission Coeﬃcient and Tunneling Eﬀective Mass

The transmission coeﬃcient T (E) is deﬁned as the ratio of the quantum mechanical (qm.) current density due to a wave impinging on a potential barrier V (x) and the transmitted current density. It is a measure for the probability of a particle, described by a wave package, to penetrate a potential barrier, whereas E is the energy perpendicular to the barrier. Tsu and Esaki developed the transfer-matrix method[49], which allows the straightforward calculation of T (E). As sketched in Fig. 3.3, calculation of the transmission through an arbitrarily shaped barrier is carried out by approximating the barrier as a series of piecewise constant rectangular barriers, for which the transmitted 32 3. Models for Charge Transport

Figure 3.3: Tunneling of a particle with energy E through an arbitrarily shaped barrier V (x). For numerical computation of the transmission coefficient the barrier is approximated by a series of rectangular barriers with constant potential. qm. current density can be easily determined. The integration over this series of infinitesimal barriers is also reflected in the well-known Wentzel-Kramers-Brillouin (WKB) approximation (e.g. ref. [51]):

∗ x1 2m 2 dx 2 (V (x) E) ∗ x0 h¯ x1 2m e− − 2 dx 2 (V (x) E) T (E) q e− x0 h¯ − . (3.12) R ∗ 2 ≈ x1 2m ≈ q 2 dx 2 (V (x) E) R 1 − x0 h¯ − 1+ 4 e  R q    Here, the integration is performed between the classical turning points x0 and x1, at which E = V (x). The huge advantage of the WKB transmission coefficient is that its numerical evaluation involves only small computational effort. However, a point that is often not addressed is the importance of the tunneling mass m∗, which has frequently been misused as an adjustable parameter to fit experimentally observed tunneling currents (e.g. ref. [52]). The correct treatment of the tunneling mass, in contrast, leads to an energy-dependent mass that, just as the band masses, can be calculated for every material, as will be outlined below. A particle, say an electron, that tunnels through the band gap of an insulator, is described by a wavefunction with complex wave vector k, whereas the imaginary part of k describes the dampening of the evanescent wave within the tunneling barrier. It enters the above formula for the transmission coefficient, equation (3.12), 2m∗ and is given by k = 2 (V (x) E). Thus, T (E) strongly depends on the relation h¯ − k(E), which can beq extracted from the dispersion relation in the band gap, i.e. the imaginary band structure (IBS) of the insulator. The IBS is the complex extension 3.1. Direct Tunneling 33

Figure 3.4: Energy dependence of the eﬀective tunneling mass of cubic ZrO2 in the upper part of the band gap, as it is derived from the dispersion models in equations (3.13) and (3.14) and tight binding (TB) calculations. The ﬁgure is taken from ref. [53]. of the lowest real conduction band of the insulator and determines the decay of the tunneling wave function[53]. The simplest possible assumption, i.e. a parabolic (pb) dispersion in the band gap leads to

2mpbE′ k (E′)= , (3.13) pb q h¯ with a constant tunneling mass mpb throughout the whole band gap, as can be seen in Fig. 3.4. Here, E′ is the negative diﬀerence between electron energy and the CB bottom. Assuming a Franz-type dispersion[54], one arrives at[55]

√2m E k = F √E 1 ′ . (3.14) F ′v h¯ u − Eg u t Equation (3.14) reduces to a parabolic dispersion close to the CB edge (E 0) ′ → and the VB edge (E E , E being the size of the band gap) with the same ′ → g g 34 3. Models for Charge Transport

effective mass (see Fig. 3.4). Since VB and CB masses are different in most cases, Dressendorfer proposed a modified dispersion

E′ √ 1 2mc √ − Eg kD = E′v ′ . (3.15) h¯ u mc E u1 1 m E u − − v g t This dispersion reduces to a parabolic one with the appropriate masses mc close to the CB and mv close to the VB. The accuracy can be further increased by performing theoretical calculations of the IBS, as it has been done for SiO2[56] and recently also for ZrO2[53], unfortunately in the cubic crystalline phase and not in the tetragonal phase, which features the highest permittivity. Since the dispersion of the wavevector in the band gap, equivalent to an effective tunneling mass, enters the exponents in equation (3.12), the tunneling mass has a strong impact on the transmission coefficient. This issue will again be referred to in the following sections, where we introduce rate models, which are necessary to perform kMC simulations and frequently involve the calculation of transmission coefficients.

3.1.2 Thermionic Emission and Field Emission

As CBOs of high-κ materials are comparably low (see Fig. 2.8), emission of charge carriers over the tunneling barrier into the CB of the high-κ film, called thermionic emission, plays an important role. This process has been studied in detail[57]. In the metal electrode a free-electron gas is assumed, whereas the electrons follow the Fermi-Dirac statistics. The major task in order to determine the current due to thermionic emission is to calculate the transmission probability for the tunneling barrier depicted in Fig. 3.5. Here, a rectangular barrier V0 (grey) with a CBO of EB with respect to the electrode Fermi level at EF is shown. Application of an electric field F tilts the barrier by introducing the additional potential V (x)= eF x. Then F − the barrier has a slope eF (dash-dotted line). The study can be extended to take − image charge effects into account. An electron which is emitted from the metal electrode induces a positive image charge within the metal electrode that attracts the emitted electron. The attractive force follows a screened Coulomb law, whereas a traveling distance x of the electron from the metal surface corresponds to a distance

2x between the electron and its image charge. With ǫopt we denote the optical permittivity of the barrier material. Thus, the image force can be written as 3.1. Direct Tunneling 35

Figure 3.5: Eﬀective barrier shape (black) at a metal-insulator interface. The rectangular barrier V0 is tilted, if an electric ﬁeld F is applied (VF ), and further reduced due to the image potential Vim.

e2 Fim(x)= − 2 (3.16) 4πǫ0ǫopt(2x) and the image potential Vim (dashed line) is then

x e2 Vim(x)= Fim(x′)dx′ = − . (3.17) − 16πǫ0ǫoptx Z∞ Addition of the diﬀerent contributions to the total potential energy leads to an eﬀective tunneling barrier with a reduced CBO (drawn in black). We can write the total potential as

e2 V (x)= V0(x)+ VF (x)+ Vim(x)= EB eF x . (3.18) − − 16πǫ0ǫoptx

From dV (x)/dx = 0 for x = xm we conclude that V (x) has its maximum at xm = e/(16πǫ0ǫoptF ) with a value of q e3F e3F e3F V (x )= E = E . (3.19) m B v v B v − u16πǫ0ǫopt − u16πǫ0ǫopt − u4πǫ0ǫopt u u u t t t 3 Thus, the barrier is reduced by ∆E = e F . SE 4πǫ0ǫopt r To accurately calculate the current through the barrier, equation (3.11) has to be solved for the given barrier shape. While this, in general, is only possible numerically, for some limiting cases well-known analytical approximations exist[58, 59]. 36 3. Models for Charge Transport

Figure 3.6: Comparison of analytical formulas for SE, equation (3.20), and FN tunneling, equation (3.21), with an accurate numerical solution. A CBO of 1 eV and a barrier thickness of 10 nm was assumed. SE describes the limit of low electric ﬁelds and high temperatures and FN the limit of high electric ﬁelds and low temperatures.

For high temperatures and low electric ﬁelds emission of electrons over the tunneling barrier delivers the dominant current contribution, while for low temperatures and high electric ﬁelds tunneling of electrons close to the Fermi levels dominates. The former case is called Schottky emission (SE) - if barrier lowering due to the image potential is taken into account. This limit is described by the Richardson-Dushman equation[57]

em k2 T 2 1 e3F j = ∗ B exp E . (3.20) SE 2 3 B v 2π h¯ −kBT  − u4πǫ0ǫopt  u   t  The latter case is known as Fowler-Nordheim (FN) tunneling. An approximate formula is found, if one calculates the tunneling of electrons close to the Fermi level at zero Kelvin through a triangular barrier, which is realized at high electric ﬁelds. Then, the current is described by[60]

3 2 3 e F 4 2m∗EB jFN = exp q . (3.21) 8πhEB − 3heF¯    In Fig. 3.6 a comparison of the approximations given in equations (3.20) and (3.21) 3.1. Direct Tunneling 37

with a numerical solution of equation (3.11), as it is implemented in the kMC simulator, is shown. Calculations were carried out for diﬀerent temperatures and a ﬁxed

CBO of EB = 1 eV. As can be seen, in this case equation (3.21) is a reasonable approximation for electrical fields larger than 3 MV/cm and temperatures of 200 K and lower. For higher temperatures, the approximation fails, as it does not account for temperature scaling of the current at all. Equation (3.20), on the other hand, is a good approximation for temperatures above 350 K and electrical fields below 1.5 MV/cm. In the intermediate transition regime only the numerical solution ∼ delivers accurate results. There is often the misunderstanding that direct tunneling, FN tunneling, and SE describe qualitatively different transport mechanisms. Thus, to conclude, we point out that they are all contained in the general treatment of electron emission through a tunneling barrier leading to equation (3.11). Sometimes it is convenient to study them separately. This can be done, e.g. by splitting the integral over Ex, the electron energy perpendicular to the barrier, in equation (3.11), as discussed in [59]. SE, for example, denotes emission over the tunneling barrier, and thus one only integrates over energies Ex close to the maximum energy of the barrier. However, if not stated otherwise, in the remainder of this thesis we will refer to all these three regimes as direct tunneling.

3.1.3 Sensitivity Analysis

If one aims at developing transport models based on simulations, it is important to know how sensitive the simulation results are towards changes of the adjustable parameters. In general, changes in simulation parameters should be carried out with outmost care, and the plausibility of the parameter values should always be scrutinized, especially, if the simulations react very sensitively towards the change of a certain parameter. In the case of kMC the simulation algorithm is a purely stochastical procedure in the sense that it works correctly from a mathematical point of view. Physics enter this stochastic framework via rate models, which are discussed in this chapter. The validity of the models itself and of the input parameters which enter the simulations with the rate models has to be ensured by the scientist carrying out the simulations, cf. chapter 4. A sensitivity analysis of the model parameters has proven to be very useful in this context and thus is provided for all models implemented in the kMC simulator, beginning with the above-discussed model for direct tunneling. We regard formula (3.11), stated again below for convenience: 38 3. Models for Charge Transport

l (Ef Ex)/(kBT ) em∗kBT ∞ 1+ e − jDT = dExT (Ex)ln l . 2 3  (E eU Ex)/(k T )  2π h¯ Z0 1+ e f − − B   Apart from the effective electron mass in the contacts, which we keep fixed during all simulations, there is no further explicit dependence on input parameters. All information on the investigated structure is implicitly included in the transmission coefficient T (Ex). T (Ex) contains information on the barrier shape, i.e. the barrier thickness, its height (CBO), and the barrier material, determining, e.g., the tunneling effective mass or ǫopt. In Fig. 3.7 the effect of a change of the CBO between electrode and dielectric material on the current flow is illustrated. For a barrier of 10 nm thickness EB was varied between 0.8 eV and 1.6 eV. The tunneling effective mass was treated according to equation (3.15), whereas for m and m values of 1.16 m and 2.5 m were used, c v · 0 · 0 corresponding to t-ZrO2[61], and the transmission coefficient was calculated in the

WKB approximation, given in equation (3.12). ǫopt in equation (3.18) was set to the literature value 5.6 for t-ZrO2[31].

At 125◦C and for lower CBOs also at 25◦C SE dominates in the whole field region studied. For higher barriers (E 1.6 eV) FN tunneling becomes dominant for F > B → 2.5 MV/cm, visible as an increase of the slope of the current-voltage(IV)-curves. The higher the CBO, the more important becomes FN tunneling. As to the sensitivity of the tunneling current, a reduction of ∆E = 0.1 eV in the CBO corresponds to an increase of the current by one and a half orders of magnitude. If FN tunneling dominates, a slightly weaker current increase by one order of magnitude is observed. Both SE and FN tunneling do not depend on the barrier thickness, provided that for different thicknesses identical electric fields, and not identical voltages, are compared. This can be understood regarding Fig. 3.8. SE (Fig. 3.8 (a)) only depends on the effective barrier height and the electron distribution within the contact, since the electrons are emitted over the barrier. In the case of FN tunneling (Fig. 3.8 (b)), electrons penetrate through a triangular barrier and are injected into the CB of the dielectric. The tunneling distance is determined by the classical turning points and thus does not depend on the total thickness of the dielectric. Any thickness dependence observed for direct tunneling has its origin in the electrons at lower energies Ex, which tunnel through a trapezoidal barrier, as shown in Fig. 3.8 (c). In this case an exponential dependence on the barrier thickness is to be expected. However, this contribution to the total direct tunneling current only becomes dominant for extremely thin films (e.g. 3 nm for t-ZrO on TiN electrodes) and very ∼ 2 3.1. Direct Tunneling 39

Figure 3.7: Direct tunneling current was calculated at temperatures of (a) 25◦C and (b) 125◦C for a 10 nm thick tunneling barrier consisting of t-ZrO2. EB was varied between 0.8 eV and 1.6 eV. 40 3. Models for Charge Transport

Figure 3.8: Different mechanisms subsumed under the general term direct tunneling. (a) SE. Electrons with large thermal energy are emitted over the tunneling barrier. (b) FN tunneling. At high electric fields electrons with a lower energy tunnel through a triangular barrier (marked orange). (c) At low electric fields, electrons with energies around the electrode Fermi level face a trapezoidal tunneling barrier (marked orange). 3.2. Tunneling Into and Out of Defects 41

low temperatures, i.e. for conditions which have not been part of this work.

3.2 Tunneling Into and Out of Defects

Besides intrinsic conduction mechanisms, as discussed in 3.1, especially in high-κ oxides with large defect densities extrinsic conduction mechanisms are important. A ﬁrst step in DAT is, naturally, always the injection of electrons from the electrode into a defect state. First studies of such processes appeared in the early seventies[62, 63].

Figure 3.9: (a) Injection and (b) emission of an electron with energy Ee + phω¯ into/out of a defect state with depth ED. Both elastic tunneling (red) and inelastic, multiphonon-assisted tunneling (black) under emission or absorption of p phonons are shown.

As shown in Fig. 3.9, tunneling can be both an elastic process or an inelastic one, in the latter case under emission or absorption of p lattice phonons of energy hω¯ (single-mode approximation), whereas ω is a typical phonon frequency. Both kinds of electron injection into and emission out of defects will be discussed in the following.

3.2.1 The Elastic Case

Lundström et al. determined the probability for a metal electron to tunnel into a defect state[62] using Bardeen’s method[64], according to which the transition matrix element can be calculated as

TED = ψD∗ (H Ee)ψEdr, (3.22) Z − where ψE and ψD are the wave functions of the electron in the metal electrode and of the trapped electron, respectively. H is the complete Hamiltonian of the 42 3. Models for Charge Transport

system and Ee the total electron energy. The transition probability per unit time p 2 2 for a constant energy E = h¯ k /(2mE) in the lateral dimensions then follows from || || Fermi’s golden rule, i.e.

2π dN p = T 2 E . (3.23) h¯ | ED| dE !E=Ee

Here, NE(E) is the electron density of states and mE the electron mass in the electrode. Integration over k delivers the total transition probability per unit time || P . Assuming a three-dimensional delta-like defect potential, Lundström arrived at[62]

3 2 4hm¯ E Kk0 2 x0k0 P = 2 2 2 1 exp( 2Kd). (3.24) 3mi K + k0 − 5 K ! − 2 2 Here, mi is the electron mass in the insulator, K = 2miED/h¯ , with the defect 2 2 depth ED (see Fig. 3.9), k0 = 2mEEe/h¯ , and the distance x0 of the defect from the electrode. One can also derive an alternative expression for P , where P is reformulated as a product of a time constant and a transmission coeﬃcient[65], which can be easily computed, e.g. in the WKB approximation. Then, we obtain

3 5/2 3/2 4 hm¯ k m 8Ee P = E 0 T (E )= e T (E ), (3.25) 2 e   e 3 mi K mi 3h¯√ED   an expression which, for convenience, was used in later works, e.g. to study the enhanced conductivity of nitrided SiO2 ﬁlms[66, 67]. To arrive at the ﬁnal rate for electron injection into the defect, P has to be weighted with the probability that the electronic states in the electrode at energy Ee are occupied, i.e. f(Ee), given in equation (3.9). Thus, we arrive at

5/2 3/2 m 8Ee R = e f(E )T (E )= C f(E )T (E ), (3.26) ED   e e e e mi 3h¯√ED ·   5/2 3/2 with C = me 8Ee . In an analogues manner, release of trapped electrons mi 3h¯√ED to the electrode is calculated via

5/2 3/2 m 8Ee R = e (1 f(E ))T (E )= C (1 f(E ))T (E ), (3.27) DE   e e e e mi 3h¯√ED − · −   whereas (1 f(E )) is the probability that the electronic states in the electrode at − e energy Ee are empty. 3.2. Tunneling Into and Out of Defects 43

3.2.2 Sensitivity Analysis

7 )

10 -1 (s

6 ED

10 R

(a)

0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0

E (eV)

Figure 3.10: (a) Rate RED for injection of electrons into defects as a function of defect depth ED. The defect was placed in a distance of 2 nm from the electrode.

EB was set to 1 eV, and temperature was set to 25◦C. (b) Sketch of electron injection into a defect for varying defect depth.

Since C is a slowly varying function of Ee and ED, the product f(Ee)T (Ee) is in the focus of the sensitivity analysis. The interplay of the two factors governs the behaviour of the injection rate RED. We calculated RED for a barrier with a CBO 44 3. Models for Charge Transport

of EB = 1 eV and a defect in a distance of 2 nm from the electrode. The defect depth ED was varied between 0.3 eV and 1.7 eV, as can be seen in Fig. 3.10 (b). A maximum rate is observed, for a defect aligned with the electrode Fermi level. The rate strongly decreases for shallower defects and weakly decreases for deeper defects.

For shallow defects the tunneling barrier is small, thus T (Ee) is large. But this is overcompensated by the low occupation probability f(Ee) for energies that are far above the Fermi level. For defects below the Fermi level and defects aligned with the Fermi level the number of electrons impinging on the tunnel barrier is nearly equal. Thus, a moderate reduction of the injection rate is observed, which is due to the higher tunneling barrier for the deeper defects. The higher the temperature, the more the maximum rate will shift towards shallower defects. It is important to keep in mind that, if an electric field is applied, the bands of the insulator are tilted. Consequently, for growing fields, shallow defect will move closer to the Fermi level, while deeper defects will be pushed further down, as shown in Fig. 3.11. In Fig. 3.11 (a), calculated injection rates under an applied electric field of up to 4 MV/cm are shown for a shallow defect with ED = 0.5 eV and a deeper one with E = 1.0 eV. For F 2.5 MV/cm we observe a maximum injection for D ∼ the shallow defect. At a distance of x = 2 nm from the electrode the potential drop due to F amounts to ∆V = eF x = 0.5 eV. Thus, the defect is perfectly aligned with the Fermi level. For higher fields, as can be seen in Fig. 3.11 (b), the tunneling barrier increases and the injection rate is slightly reduced. The deeper defect with E = 1 eV, accordingly, has its maximum injection rate at F 0 and R slightly D ∼ ED decreases for higher fields. 3.2. Tunneling Into and Out of Defects 45

0.5 eV

1.0 eV

) -1

(s 10

ED R

(a)

0 1 2 3 4

F (MV/cm)

Figure 3.11: (a) Rate RED as a function of the electric ﬁeld F for two defect depths

ED = 0.5 eV and ED = 1.0 eV. The defect was placed in a distance of 2 nm from the electrode. EB was set to 1 eV, and the temperature was set to 25◦C. (b) Change of the barrier shape due to the applied ﬁeld. Orange colour marks the barrier for

ED = 0.5 eV and green colour for ED = 1.0 eV. 46 3. Models for Charge Transport

3.2.3 The Inelastic Case

We describe inelastic tunneling from an electrode contact into a defect and the reverse process of emission of trapped electrons into the electrode as a multi-phonon- assisted tunneling process. A ﬁrst analysis of tunnel-assisted release of trapped electrons into the electrode was carried out by Kiveris et al.[68], which was later on augmented by Dalidchik[69], who reﬁned the multi-phonon transition theory and coupled it to WKB tunneling expressions. Herrmann further expanded the model by calculating Shockley-Read-Hall lifetimes, assuming that electron capture/emission is a multi-phonon-assisted transition between a free-electron state in the electrode and a localized state at the defect site[70]. We followed his treatment, adapting it to our needs.

As depicted in Fig. 3.9, we consider a defect state at energy Ee(x), which, in presence of an electrical field F , depends on the x coordinate of the defect. Coupling to the phonons of the insulator is described in the single-mode approximation, i.e. we replace the coupling to the various phonon modes of the insulator by a coupling to only one effective phonon mode, characterized by the effective phonon energy hω¯ and two coupling constants, the lattice relaxation energy Elattice and the Huang- Rhys factor S. Consequently, we obtain a series of virtual states of the coupled electron-phonon system at energies Ep = Ee + phω¯ . Each of these states can be initial or final state of a multi-phonon-assisted tunnel process. MP We express the rates for multi-phonon-assisted tunneling into (RED ) and out of MP defects (RDE ) via characteristic times, needed to occupy or to empty a defect, i.e.[70]

MP 1 RED = MP , (3.28) τED MP 1 RDE = MP . (3.29) τDE Unlike in former treatments where the interest was focused on the steady-state conditions of the system, i.e. the equilibrium occupation probability of the defect[70], we intend to explicitly study the dynamical behaviour of the system. Besides, a steady-state analysis in the framework of our kMC simulations would also have to encompass other transport mechanisms that are present at the same time and, competing with multi-phonon-assisted tunneling, inﬂuence the steady-state occupation probability of the defect. The time constants given in equations (3.28) and (3.29) are calculated from[70] 3.2. Tunneling Into and Out of Defects 47

1 = ∞ N (E)f(E)T (E,x)c (E,x)dE (3.30) τ MP E ED ED ED Z−∞ 1 = ∞ N (E)[1 f(E)]T (E,x)c (E,x)dE. (3.31) τ MP E − DE DE DE Z−∞

Here, NE(E) is the density of states in the electrode,

1 2m 3/2 N (E)= E E E Θ(E E ) (free-electron gas), (3.32) E 2π2 2 − c · − c h¯ q where Θ is the Heaviside function and Ec is the CB edge of the electrode. TED(E,x)= T (E,x) T (E,x) is the transmission coeﬃcient through the tunnel barrier be- DE ≡ tween the electrode and the defect at position x for electron energy E. cED and cDE denote the characteristic electron-phonon interaction rates,

Ee(x) E exp − , if Ee(x) E> 0, c (E,x)= c ∞ L (z)δ (E E (x)) − kBT − ED 0 p p  p= ,p=0 − 1, otherwise, −∞X 6  (3.33)  and

E Ee(x) exp − , if E Ee(x) > 0, c (E,x)= c ∞ L (z)δ (E E (x)) − kBT − DE 0 p p  p= ,p=0 − 1, otherwise. −∞X 6  (3.34)  For a 3D, delta-like defect potential c0 has been estimated to[71, 72]

2 3 (4π) rD 3 c0 = (hE¯ 0) . (3.35) hE¯ g

Here, rD = h/¯ √2miED is the localization radius of the electron in the defect with depth ED. hE¯ 0 denotes the electro-optical energy in the insulator,

1/3 e2h¯2F 2 hE¯ 0 = , (3.36) 2mi ! and Eg is the band gap of the insulator material.

Apparently, only energies E = Ep = Ee(x)+ phω¯ , where p is the integer phonon number, are allowed for the multi-phonon-assisted processes. This is an immediate consequence of the single-mode approximation. In this picture the multi-phonon transition probability Lp(z), introduced in equation (3.33), can be calculated as 48 3. Models for Charge Transport

p/2 fBE + 1 Lp(z)= exp[ S(2fBE + 1)]Ip(z), (3.37) fBE ! − where Ip(z) denotes the modiﬁed Bessel function of order p with the argument z = 2S fBE(fBE +1), and fBE is the Bose-Einstein distribution, i.e. q 1 fBE = . (3.38) exp hω¯ 1 kBT − Inserting equations (3.33) and (3.34) into equations (3.30) and (3.31) delivers

phω¯ MP ∞ k T RED = c0NE(Ep)f(Ep)T (Ep,x)Lp(z) e B p<0 · X (3.39) ∞ + c0NE(Ep)f(Ep)T (Ep,x)Lp(z) p>X0

RMP = ∞ c N (E )[1 f(E )]T (E ,x)L (z) DE 0 E p − p p p pX0

3.2.4 Sensitivity Analysis

In general, the above-presented model for inelastic, multiphonon-assisted tunneling introduces a multitude of parameters. However, most of these can be chosen to match experimentally extracted values, as stated below. Compared to elastic injection of electrons into defects, the inelastic injection is usually suppressed by some orders of magnitude. This is at least the case for the above-chosen approximation, i.e. especially the chosen defect potential shape which determines the magnitude of MP c0. To study the sensitivity towards the model parameters, we calculated RED for a defect in a distance of x = 2 nm from the electrode, a CBO of 1 eV, and a defect depth of ED = 1 eV. MP In Fig. 3.12, the sensitivity of RED on the most important variables, i.e. (a) temperature, (b) Huang-Rhys factor, and (c) phonon energy is illustrated, whereas the contributions from individual phonon numbers, calculated from

Ee(x) E exp − , if p> 0, c N (E )f(E )T (E ,x)L (z) − kBT (3.41) 0 E p p p p ·  1, otherwise,

 3.2. Tunneling Into and Out of Defects 49

6 10

4 10

) ) -1 -1 (s (s 0

10 MP ED MP ED R R

-2

-2 T S 10

0.5 25°C

1.0 -4 125°C

-4

10 10

10.0 225°C

(a) (b)

-6 -6

10 10

-10 -5 0 5 10 -10 -5 0 5 10

p p

) -1

0 (s

MP ED R

-2

0.02 eV

-4

10 0.04 eV

0.06 eV

(c)

-6

-20 -15 -10 -5 0 5 10 15 20

E/0.04 (eV)

MP Figure 3.12: Rate RED for inelastic, multi-phonon-assisted injection into a defect in a distance of x = 2 nm from the electrode. A CBO of 1 eV, a defect depth

ED = 1 eV , and an electrical ﬁeld F = 1 MV/cm were used. The dependence of the rate on (a) temperature, (b) Huang-Rhys factor, and (c) phonon energy is depicted. are resolved. For electron injection negative phonon numbers correspond to phonon absorption and positive numbers to phonon emission (and vice versa for electron emission from the defect).

Increasing the temperature leads to an increased injection rate (see Fig. 3.12 (a)), as intuitively expected, since the number of available phonons increases. An increase in MP temperature of 100◦C leads to an increase of RED by 1-2 orders of magnitude. Addi- tionally, the maximum contribution to the total rate slightly shifts to higher phonon numbers. While for T = 25 C it lies at p = 1, it shifts to p = 1 for T = 225 C. This ◦ − ◦ shift is due to the broadening of the Fermi-distribution, i.e. the increased probability to ﬁnd occupied electron states with higher energy in the electrode, which can 50 3. Models for Charge Transport

tunnel under emission of phonons into the defect. Coupling to oxide phonons, here represented by an effective phonon energy hω¯ , leads to a thermal broadening of the defect level, whereas the Huang-Rhys factor S is a measure for the coupling strength of the electron-phonon interaction. If one would set S to zero, thus eliminating the electron-phonon interaction, one would observe a delta-like peak in Fig. 3.12 (b). The higher S, the more this peak is broadened, i.e. the probability for transitions involving higher numbers of phonons increases. In principle, S can be calculated for a certain material, if its phonon modes and the electron-phonon coupling potential are known[73, 74]. Experimentally, via studies of temperature-dependent photoluminescence of zirconia films, a Huang-Rhys factor of up to 40 was found[75], indicating very strong electron-phonon coupling in these films. In most common binary semiconductor materials like GaAs, GaP, GaSb, InAs, InP, or InSb S ranges from 1 to 3[76]. In our kMC simulations we used values between 1 and 10 for S. MP The dependence of RED on the assumed effective phonon mode is shown in Fig. 3.12 (c), MP where RED is depicted as a function of the energy change ∆E during an inelastic transition. This energy ∆E has to be either absorbed by or emitted from the electron via phonon absorption or emission. For a given ∆E the transition probability, MP i.e. RED , decreases with decreasing phonon energy hω¯ , since this means that a higher number of phonons is necessary to supply or to dissipate this energy, thus making it a less likely process. In agreement with theoretically and experimentally determined values for ZrO hω¯ was kept in the interval 30 50 meV[75, 31] for all 2 − kMC simulations. 3.3. Defect-Defect Tunneling 51

3.3 Defect-Defect Tunneling

In the preceding section electron injection into defects was investigated, forming the ﬁrst step in transport across defect states. In a subsequent step trapped electrons can be 1) reemitted to the electrode, they came from, or tunnel to the counter- MP electrode, described by RDE in equation (3.40), 2) reach the CB of the insulator via thermal emission, as discussed in 3.4, or 3) tunnel to another defect in their vicinity. For the latter kind of transition, unlike trapping and detrapping processes involving delocalized states in the electrode, which are dealt with in 3.2, both initial and ﬁnal state, have to be described by strongly localized wave-functions.

Figure 3.13: Schematic of (a) elastic and (b) inelastic tunneling between two defects i and j in the band gap of an insulator. In (a) no electric ﬁeld is applied, thus assuming two identical defects, the defects have the same energy level. The exponentially decaying wave functions of the trap states are sketched. In (b) an electrical ﬁeld is applied, which tilts the bands and shifts the defect level energies with respect to each other. The electron transition from defect i to j can only occur under phonon emission.

As sketched in Fig. 3.13, defect-defect tunneling can be (a) an elastic process, e.g. between two defects with the same energy without an applied electric field or for transport perpendicular to an applied electric field, or (b) an inelastic, multi- phonon-assisted process, e.g. between defects energetically shifted with respect to each other due to an applied electric field. 52 3. Models for Charge Transport

3.3.1 The Elastic Case

Elastic defect-defect transitions are modeled as a hopping process between isolated electron states. For elastic hopping from defect i to defect j Mott derived the transition rate[77]

2rij Rij = ν exp − . (3.42) · rD 13 1 Here, ν is a typical phonon frequency, i.e. of the order of 10 s− , rij is the distance between the two defects, and rD is deﬁned by the wave function of the localized states, decreasing as exp( r/r ). For a delta-like defect potential the normalized − D wave funtion of the trapped electron is given by[71]

r e r/rD ψ = D − (r> 0), (3.43) r 2π r as sketched in Fig. 3.13, with the localization radius

h¯ rD = , (3.44) √2miED which has already been introduced in 3.2. Here, ED is the depth of the defect, i.e. the deeper a defect level, the stronger the electron localization.

3.3.2 The Inelastic Case

Inelastic defect-defect tunneling, depicted in Fig.3.13 (b), analogous to the case of inelastic electron injection into a defect, described in 3.2, is a phonon-assisted process, where the electron exchanges energy with the crystal lattice via electron- phonon interaction. Phenomenologically, the inelasticity can be incorporated into the rate expression, given in equation (3.42) by multiplication with an exponential weighting factor. Then the rate for inelastic tunneling from defect i to defect j is[78]

2r exp ∆E , if ∆E> 0 R = ν exp − ij − kBT ij · r ·  D 1, otherwise, where ∆E is the energy diﬀerence between ﬁnal and initial defect state. The above expression for Rij corresponds to the Miller-Abrahms hopping model, which was developed to describe impurity conduction in doped semiconductors[79]. The inelastic defect-defect transitions can also be regarded as a two-step process: 1) elastic tunneling from defect i into a virtual state at the position of defect j and 2) subsequent 3.3. Defect-Defect Tunneling 53

energetic relaxation to Ej, the energy of defect j, moderated by the electron-phonon interaction. Then, one arrives at

exp ∆E , if ∆E> 0 R = ν T − kBT (3.45) ij · ij ·  1, otherwise, whereas Tij is the transmission coeﬃcient through the tunneling barrier separating the two defects.

Alternatively, we can calculate the rate Rij from expressions analogue to equations (3.30) and (3.31). Adopted to the current situation, one can write

∞ ∞ R = N (E)f N (E′)[1 f ]T (E,r )c (E,E′)dEdE′. (3.46) ij i i j − j ij ij ij Z−∞ Z−∞ Here, N (E)= δ(E E ) and N (E)= δ(E E ) are the densities of states for the i − i j − j two defects at energies Ei and Ej, and fi and fj are there occupations, i.e. either zero or one, if multiple occupation is forbidden. A transition from defect i to defect j, in this picture, is only possible, if defect i is occupied and defect j is empty. Thus, we arrive at

Rij = Tij(Ei,rij)cij(Ei,Ej). (3.47)

Ridley carried out quantum-mechanical calculations of the rate cij for multi-phonon- assisted, non-radiative defect-defect transitions[76]. Adopting the single-mode approximation, he arrived at a widely applicable, simple expression for cij, given by

cem = R (f + 1)pexp( 2f S) (3.48) ij 0 BE − BE for a transition under emission of p phonons and

cabs = R f p exp( 2f S) (3.49) ij 0 BE − BE for a transition under absorption of p phonons. The prefactor R0 can be approximately written as

p 1 S 2 π S − e− 2 ∆ R0 = 2 0.26 V + 0.18| | (p 1) , (3.50) h¯ ω (p 1)! " | | S − # − with

2 2 V = Vq (3.51) | | q | | X 54 3. Models for Charge Transport

and

2 2 V ∆ = Vq∆q . (3.52) | | | q | X Here, Vq and ∆q are deﬁned as

Vq =<ψ Uq ψ > (3.53) j| | i and

∆q =<ψ Uq ψ > <ψ Uq ψ > . (3.54) j| | j − i| | i Uq denotes the electron-phonon coupling potential, which may depend on the phonon wave vector q. ψi and ψj are the wave functions of the initial defect i and the ﬁnal defect j. So far, no assumptions have been made on the localized electron states or the electron-phonon interaction. If detailed models for Uq, ψi and ψj are supplied, the matrix elements V 2 and V ∆ 2 can be evaluated and, thus, also R in equation | | | | DD (3.48), a task which has been carried out in the past for various models[80, 81, 82]. For our purposes the calculation of the above matrix elements was performed numerically by the simulator assuming for the defect wave function a constant probability density in a cubic volume centered around the defect position and a vanishing probability density outside the cube, i.e. for a defect at r = 0,

1/2 4π r3 − , if R x,y,z R ψ = 3 D − 2 ≤ ≤ 2 (3.55) h i 0, otherwise.

The edge length of the cube was chosen to be

R = (4π/3 r )1/3, (3.56) · D and for the electron-phonon interaction an optical deformation potential

Uq = U exp(iq r) (3.57) D · · was adopted.

3.3.3 Sensitivity Analysis

In equations (3.45) and (3.47) two models for inelastic, multi-phonon-assisted tunneling between two defect states were introduced, whereas the ﬁrst can be evaluated 3.3. Defect-Defect Tunneling 55

straightforwardly, while the second involves much more mathematics and also more assumptions, which are reﬂected in additional model parameters. As the phonon 13 1 vibration frequency ν was kept ﬁxed to 10 s− in all our simulations, it is not discussed here. Anyway, the dependence of Rij on it is trivial.

Figure 3.14: Sketch of two defect levels i and j in the band gap of an insulator. (a)

If U1 > 0, then Ei >Ej. Thus, tunneling from i to j can only occur under emission of p1 =∆E1/hω¯ phonons. (b) If U2 > U1, then ∆E2 > ∆E1. Thus, the number of emitted phonons increases, i.e. p2 >p1. The rate Rij for non-radiative, multi- phonon-assisted transitions from i to j was calculated for rij = 2 nm, ED = 1 eV, hω¯ = 40 meV and T = 125◦C.

Instead, we compare the two models introduced in the preceding subsection with respect to their scaling with phonon number p. To this end, we investigated the following situation, shown in Fig. 3.14. We regard two defect states i and j in the bulk of an insulator separated rij = 2 nm from each other. The defect depth is chosen to be ED = 1 eV. Electric fields between 0 MV/cm and 4 MV/cm are applied in order to shift the defect levels with respect to each other. The rate for non-radiative, multi-phonon-assisted tunneling Rij from defect i to j was computed. The higher the applied bias, the larger ∆E, the energy difference between the defect levels. Thus, for higher fields higher phonon numbers are desired for the transition to take place. 56 3. Models for Charge Transport

eqn. (3.45)

eqn. (3.47)

) -1 (s ij

0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0

F (MV/cm)

Figure 3.15: Rij was computed for the defect-defect transition sketched in Fig. 3.14 for varying electric ﬁelds. The two models given in equations (3.45) and (3.47) are compared.

As Fig. 3.15 shows, rates Rij calculated from equations (3.45) and (3.47) are in close agreement, both, in quantitative terms and in their scaling behaviour with the electric ﬁeld, or, equivalently, the phonon number. The minor diﬀerences, which are observed, have their origin in the assumptions made for the defect potential and the localized electron wave functions, which enter equation (3.47).

In Fig. 3.16 Rij is shown for different defect depths. ED was varied between 0.5 eV and 1.5 eV. In general, for all investigated defect depths the Miller-Abrahams- like (MA) model (equation (3.45)) scales stronger with the electric field than the model derived from Ridley’s considerations (equation (3.47)), as already observed in Fig. 3.15. The rate calculated according to the MA model also scales stronger with the defect depth. For both models, the scaling behaviour is dominated by the scaling of the transmission coefficient, thus it is very similar. However, in equation

(3.47) also cij depends on ED. The matrix elements in equation (3.50) depend on the shape of the wave functions which again depends on ED. In the chosen approximation the distance, to which the localized states extend into the insulating material, is determined by the edge length R, which is calculated from (3.56). The dependence of cij on ED enters the model via the localization radius rD. 3.4. Poole-Frenkel Emission 57

E = 0.5 eV eqn. (3.45) 8

E = 0.5 eV eqn. (3.47)

E = 1.0 eV eqn. (3.45)

7 D

E = 1.0 eV eqn. (3.47)

E = 1.5 eV eqn. (3.45)

10 E = 1.5 eV eqn. (3.47)

) -1

(s 10 ij

0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5

F (MV/cm)

Figure 3.16: Rij was computed for the defect-defect transition sketched in Fig. 3.14 for diﬀerent defect depths ED between 0.5 eV and 1.5 eV for varying electric ﬁelds. The two models given in equations (3.45) and (3.47) are compared.

3.4 Poole-Frenkel Emission

Discussing DAT, we so far only considered tunneling processes, where the electrons never reach the CB of the insulator. We regard an electron trapped in a defect state. Besides tunneling to other nearby defects or to the contacts, it can also leave the defect via thermal emission to the CB. In the late 1930s, Frenkel developed a model for this kind of process, motivated by the observation that in insulators and semiconductors the conductivity increases, if large electric fields are applied[22]. A field-dependent mobility of the charge carriers could be ruled out, since the same increase of conductivity was also found upon illumination of the materials, whereas the applied electric field was left constant. Thus, the reason for this enhancement was assumed to be an increase of the charge carrier density. Frenkel envisioned a field-assisted detrapping, freeing immobile charge carriers, which subsequently contribute to the conduction. In a medium with high defect densities, as for instance the high-κ oxides, defect states are assumed to be the sources of the additional charge carriers. The increased conductivity, in this picture, is explained by a field-assisted barrier lowering, whereas for an electron trapped in a defect the barrier height equals the defect depth. 58 3. Models for Charge Transport

A simple model, describing the field-enhanced detrapping of electrons from defect states, can be obtained from the following considerations: We regard a defect state, which, if occupied with an electron, is neutral. Thus, as soon as the electron leaves the defect, the latter becomes positively charged. The resulting Coulomb attraction between electron and charged defect is sufficient to explain a barrier lowering in presence of an electric field, as can be understood from Fig. 3.17.

Figure 3.17: Effective electron potential (black) in the vicinity of a positively charged defect under an applied electric field F . The constant crystal potential V0 is tilted by the electrostatic potential VF . Combined with the screened Coulomb potential VD of the defect, this leads to a field-dependent barrier reduction ∆EPF for emission of trapped electrons.

A constant crystal potential

VC(x)= V0, (3.58) describing, e.g., the CB edge of the insulator, is assumed. We add an electric ﬁeld F , which gives rise to an electrostatic potential

V (x)= eF x. (3.59) F − The attractive Coulomb interaction between the positively charged defect at position x0 and the emitted electron at position x is described by a screened Coulomb 3.4. Poole-Frenkel Emission 59

potential

e2 V (x)= . (3.60) D −4πǫ ǫ x x 0 opt| − 0| ǫopt is the optical permittivity of the insulator material. Its usage at this place has been subject to discussion[83]. The overall potential V , shown as continuous line in

Fig. 3.17, is calculated by superimposing VC, VF , and VD:

e2 V (x)= V (x)+ V (x)+ V (x)= V eF x . (3.61) C F D 0 − − 4πǫ ǫ x x 0 opt| − 0| The combination of the band tilting due to the electric ﬁeld and the attractive Coulomb potential leads to the barrier reduction ∆E, which can be easily determined. From the condition

dV (x) = 0 (3.62) dx !x=xm we extract the coordinate of the maximum of the barrier for electron emission xm,

e xm = x0 + . (3.63) s4πǫ0ǫoptF Back in equation (3.61) we obtain

e3F V (x )= V eF x (3.64) m 0 0 v − − uπǫ0ǫopt u and thus an eﬀective barrier height of t

e3F e3F E ∆E = V eF x (V E eF x )= E (3.65) D PF 0 0 v 0 D 0 D v − − − uπǫ0ǫopt − − − − uπǫ0ǫopt u u t t 3 with a ﬁeld-dependent barrier reduction ∆E = e F . This expression is quite PF πǫ0ǫopt r similar to ∆ESE, the barrier reduction for thermionic emission, calculated on page 35, which resulted from the interaction of the electron with its image charge in the electrode, see 3.1.2.

∆EPF = 2∆ESE, (3.66) whereas the factor 2 is due to the fact that, in the case of PF emission, the positive charge is spatially ﬁxed at the defect site, while, in case of thermionic emission, if the electron moves a certain distance in the x-direction, its image charge moves the same distance in the x-direction. − 60 3. Models for Charge Transport

Emission of the trapped electron into the CB can be modeled as a thermally excited process, i.e. one can calculate the probability that, at a certain point in time, a suﬃcient amount of thermal energy is transferred to the electron so that it can overcome the barrier. The corresponding transition rate is

1 e3F R = ν exp E . (3.67) PF D v · −kBT  − uπǫ0ǫopt  u   t  ν denotes the lattice vibration frequency ( 1013 s 1), which can be physically ∼ − interpreted as number of times per second the electron interacts with the surrounding crystal lattice[84]. If the above one-dimensional treatment, which only considers electron emission parallel to the applied electric ﬁeld, is expanded to three dimensions, allowing electron emission under every angle with respect to the electric ﬁeld, one arrives at[85]

2 ED kBT β√F β√F 1 RPF = ν exp 1+ 1 exp + , (3.68) · −k T  β√F ! " k T − ! k T !# 2 B  B B 

3   with β = e /(πǫ0ǫopt). q 3.4.1 Sensitivity Analysis

The major, and in fact the only, parameter governing the rate for field-enhanced thermal emission of trapped electrons from defects into the insulator CB is the trap 13 1 depth ED, since we adopted throughout all simulations a value of 10 s− for ν. In Fig. 3.18 the rate for PF emission has been calculated for electrical fields F up to 4 MV/cm varying ED. ǫopt = 5.6, the value for t-ZrO2[31], was adopted. The comparison between the 1D model, given in equation (3.67), and the 3D model, given in equation (3.68), shows that both have the same scaling behaviour with respect to ED. Rates calculated according to the 3D model are slightly smaller, which is due to the fact that in this case the net rate was obtained by averaging over the rates for electron emission in all angles with respect to the applied field, whereas the 1D model only treats emission parallel to the field. In the latter case the barrier reduction takes on its maximum value, resulting in a higher emission rate.

Variation of ǫopt has a huge impact on the calculated emission rates, as shown in

Fig. 3.19. Smaller values of ǫopt lead to a stronger increase of RPF for increasing electric ﬁelds. Unfortunately, ǫopt has been frequently misused as a free ﬁtting 3.4. Poole-Frenkel Emission 61

14 D

0.6 eV

12 0.8 eV

1.0 eV

1.2 eV

1.4 eV

10 ) -1

6 (s

PF R

-2

-4

400 600 800 1000 1200 1400 1600 1800 2000

0.5 0.5

F (V/cm)

Figure 3.18: RPF is calculated varying ED between 0.6 eV and 1.4 eV. Continuous lines correspond to the 3D model given in equation (3.68), dashed lines to the 1D model given in equation (3.67).

opt

4.6

5.6

10 6.6

5 )

10 -1 (s

4 R 10

400 600 800 1000 1200 1400 1600 1800 2000

0.5 0.5

F (V/cm)

Figure 3.19: RPF is calculated for ED = 1 eV varying ǫopt. The 3D model from equation (3.68) was used. 62 3. Models for Charge Transport

parameter in order to match experimental current-voltage data, e.g. in refs. [86, 87].

In contrast to this, we ﬁx ǫopt to literature values for the respective insulator material. Chapter 4

Methods for Transport Simulations

4.1 Kinetic Monte Carlo (kMC)

As outlined in 2.2, the downscaling of semiconductor devices inevitably lead to the introduction of high-κ materials, e.g. as gate dielectric in transistors or as insulating film in the DRAM storage capacitor. In general, they allow higher capacitance densities and thus enable further downscaling. Unfortunately, these high-κ thin films typically suffer from high defect densities, giving rise to defect states that open up pathways for undesirable leakage currents. Especially the DRAM storage capacitor poses high requirements to the insulating properties of the dielectric. Hence, further scaling requires a detailed understanding of the transport across these films. It has been the major task of this thesis to develop a deeper insight into the transport in high-κ thin film capacitors for the usage as DRAM storage capacitors. To this end, models, describing the individual charge transport mechanisms, see chapter 3, were developed and, in a next step, integrated in a simulation algorithm, which allows us to mimic the charge carrier flow in the capacitors, i.e. to simulate the movement of charge carriers through the insulating high-κ layer, which separates the two capacitor electrodes. This framework was used to develop and validate transport models for the state-of-the-art material systems employed in the storage capacitors of the most recent DRAM chip generations. Since transport through high-κ films is very complex and involves numerous different microscopic processes, the algorithm had to be very versatile, in order to be able to unite the individual processes in one simulator. Especially DAT so far was described by continuum approaches. However, in the face of the peculiarities of the system

63 64 4. Methods for Transport Simulations

under investigation, i.e. the high permittivity, the high defect density, and the small dielectric film thickness, such treatments are questionable. Many phenomena connected with high-κ thin films require a simulation of charge transport at the level of single charge carriers that resolves their point-like nature, as well as point- like defects and other spatially localized phenomena, such as localized electric field enhancement due to electrode roughness, cf. 5.2, or transient degradation of the dielectric via defect accumulation[88, 89]. Therefore, we decided to carry out our simulations in a kinetic Monte Carlo (kMC) framework. The advantages, but also the difficulties, that come along with this choice will be outlined in this chapter. First, we will give a concise survey of the state-of-the-art as what concerns modeling of leakage currents in high-κ films in order to show up the deficiencies in the established analysis techniques and in how far kMC promises to bring an improvement. The reader will be introduced to Monte Carlo techniques in general and especially into the specialties of the subtype kMC. We will briefly discuss the origins of kMC, as far as it impacts the design of the algorithm, and we will outline the basic ideas of a kMC simulation, before we move on to the practical implementation of the algorithm for our purposes.

4.1.1 Modeling of Leakage Currents: State-of-the-Art vs. kMC

Up to now transport in the most established high-κ materials, ZrO2 and HfO2, has been studied using simpliﬁed compact-models, e.g. refs. [90] and [86]. In ZrO2 transport was assumed to be dominated by PF emission from defects. According to the standard-PF model the current density j is proportional to[91]

1 e3F j F exp E , (4.1) D v ∝ · −kBT  − uπǫ0ǫopt  u   t  where ED is the defect depth, F is the electric ﬁeld, and ǫopt the optical permittivity of the dielectric. However, good ﬁts to experiments could only be achieved by using unrealistic values for material parameters, especially for ǫopt. This is due to the fact that this picture of charge transport is incomplete. Only charge carrier emission from traps is described, while the injection from the electrodes into the traps is neglected. A tunneling-assisted Poole-Frenkel (TAPF) conduction mechanism has been suggested for HfO2[28] that incorporates the injection step. However, the TAPF model only allows for homogeneous defect distributions and it does not take 4.1. Kinetic Monte Carlo (kMC) 65

into account that, by applying an electric field, the defect energy becomes a function of the defect position. Furthermore, competing transport channels like detrapping of electrons via quantum mechanical tunneling, cf. 3.2.1 and 3.2.3, are not considered. In general, a closed model for charge transport in these high-κ films is missing. These restrictions may be lifted by employing Monte Carlo techniques, which have proven to be a versatile tool for transport simulations[92, 93, 94]. kMC provides a rigorous stochastic framework[95, 96] to describe correctly the statistical dynamics of the charge carriers in high-κ dielectrics, which are very complicated due to the diversity of mutually interdependent transport processes, which play a role. Sim- ulations take place at the level of single charge carriers, whose pathway through the dielectric is tracked. In this way all relevant transport channels are accounted for concurrently, whereas the well-defined relation between kMC simulation time and real time, see 4.1.4, allows one to study the real-time dynamics of the charge carriers. Local phenomena, as described above, are resolved and can be taken into account in a straightforward manner, e.g. carrier-carrier interactions by coupling the simulations to a Poisson solver.

4.1.2 History and Development of kMC

The term "Monte Carlo" subsums a class of computational algorithms that rely on the usage of random numbers. First algorithms of this kind were developed in the late 1940’s at the Los Alamos Scientific Laboratory to study neutron diffusion in fissionable material[97]. These algorithms were developed to study static or equilibrium properties of physical model systems. For instance the conforma- tion of polymers[98] was investigated using the so-called Metropolis algorithm, which was developed in 1953[99]. The idea behind this is quite simple and sketched briefly in the following: It shall be the aim of the simulation to find the Figure 4.1: Sketch of a polymer lowest energy configuration for a given poly- configuration. mer. To this end a random walk in the con- formation space of the polymer is performed, i.e. the polymer is transferred from 66 4. Methods for Transport Simulations

configuration to configuration, or from state to state, respectively, in order to find the one with the lowest energy. Let state A of the polymer, sketched in Fig. 4.1, be defined by the coordinates of its N atoms, i.e. by the tuple

A =(r1, .. ri, .. rN ). (4.2)

We now consider a transition from state A to state B, e.g. given by the tuple

B = r , .. r′, .. r , (r ′ = r ). (4.3) 1 i N i 6 i The energies EA and EB for both states are computed and the change of the system energy due to the transition ∆E = E E is calculated. In order to direct the B − A random walk towards low energies, a transition A B is only accepted with a → probability

1, if ∆E 0, P (A B)=  ≤ (4.4) → ∆E exp , otherwise. − kBT This favors transitions for which ∆E 0, but stills allows for uphill steps so that ≤ the system does not get stuck in local minima of the system’s total energy surface. In this manner the polymer is iteratively driven into its state of minimum energy. However, we do not learn anything about the system dynamics, since there is no link between the probability P to accept a transition and the time needed for this transition to happen. A ﬁrst algorithm, aiming at a simulation of system dynamics, appeared in the 1960’s with Beeler’s simulation of radiation damage annealing[100]. In the following decades this development gained momentum and numerous applications of such algorithms, studying surface processes, crystal growth etc., appeared. An informative review of these developments can be found in ref. [101]. While this type of approach originally was called "dynamic Monte Carlo", in the 1990’s, the term "kinetic Monte Carlo" settled in. What makes it so attractive is the prospect that, in principle, it delivers the exact dynamical behaviour of a system. Due to practical reasons, as will be discussed below, this ideal is virtually never achieved. For simulations of the dynamical behaviour of ensembles of atoms "molecular dynamics" (MD) were developed and, since then, have been the workhorse in this ﬁeld. In this simulation scheme, interatomic interactions and interactions with external forces are described by a model potential, which is determined in advance and fed 4.1. Kinetic Monte Carlo (kMC) 67

into the simulations. Then the dynamics of the atoms emerge from forward prop- agation in time of Newton’s classical equations of motion. No further assumptions or user input are required. While the first application in the late 1950’s assumed a hard sphere potential[102] and was very successful in studying simple liquids, more accurate model potentials appeared a few years later with the first use of Lennard-Jones potentials in MD[103] for liquid argon. The development of faster computers enabled the usage of more sophisticated approaches, which are combined with high computational costs. At the very core of these novel methods stands the idea that forces, acting on atoms/molecules at a certain point in time, are calculated on-the-fly using the results from ab initio electronic structure calculations. This type of MD is known as Car-Parrinello MD or ab initio MD[104]. In contrast to the early approaches in this case the treatment of the interatomic interactions adapts to the current situation, i.e. also systems in which the dynamics lead to a qualitative change of the interatomic interactions during a simulation can be investigated. Unfortunately, all MD simulations suffer from an intrinsic problem of this approach. In order to accurately compute the atomic movements, the time steps ∆t, with which 15 the simulation advances, have to be chosen sufficiently short, i.e. ∆t< 10− s, in order to resolve the vibrational motions of the atoms. Thus, typical MD simulations only reach overall simulation times of less than a microsecond. The same applies to the length scale. Unfavorable scaling of the computational effort with the number of simulated atoms also sets strict limitations to the system size, which is accessible to simulations. At present, one can simulate ensembles up to some hundred particles, which translates to some hundred nanometers extension in every spatial dimension. Often, however, processes one would like to study take place on much longer time and length scales, e.g. macroscopic surface diffusion processes or phase transitions. The ultimate goal is to unravel the link between macroscopic phenomena, which can be directly observed, and the underlying mechanisms on the molecular scale. To this end one has to bridge many orders of magnitude both in time and in space, setting up large obstacles for computational approaches. This problem is known as time- scale/length-scale problem and, in fact, was the major trigger for the development of the kMC method. A lot of research has been dedicated to exactly this problem and advances have been made. A good summary of the state-of-the-art can be found in ref. [105] A comprehensive model of such a system inevitably requires a multi-scale approach, which closes the gap between the molecular scale, accessible to MD, and the macro- 68 4. Methods for Transport Simulations

Figure 4.2: Multi-scale modeling. kMC is thought to bridge the gap between ab initio based approaches like ab initio MD and large-scale continuum descriptions. scopic scale, treated via continuum approaches. kMC algorithms have been devised for this purpose, see Fig. 4.2. Employing kMC brings along a boost in simulation speed, which is achieved by describing the system dynamics in a coarse-grained fash- ion, as will be explained in 4.1.3, neglecting in comparison to ab initio molecular dynamics e.g. the atomic vibrations.

4.1.3 Basics of kinetic Monte Carlo

Before diving into the implementation of kMC for our specific task, the modeling of electron transport through high-κ dielectrics, we will first give a general introduction to the basic ideas of kMC and apply these on a simple example system. We aim at an accurate description of the dynamics of a system. To this end the system is regarded as an entity, which, for every instant of time, can be characterized by its macroscopic state. Different system states are connected with each other via state-to-state transitions, as sketched in Fig. 4.3 (a). All transitions are identified and subsequently reduced to a set of relevant transitions, i.e. transitions which are regarded as of minor importance for the overall system dynamics are neglected. This coarse-graining of the system dynamics is responsible for the speed up of kMC compared to MD. Obviously, one has to pay particular attention not to lose essential parts of the dynamics. A transition rate R is associated with every transition, whereas R is proportional to the probability of the transition. Usually, it is assumed that the rate Rij for a transition from state i to state j does not depend one the 4.1. Kinetic Monte Carlo (kMC) 69

Figure 4.3: (a) Macroscopic states of a system interconnected via transitions. To each transition i j belongs a transition rate R , corresponding to the transition → ij probability. (b) Conﬁguration space of a sample system. Each box corresponds to a macroscopic state. States are connected by a complex network of transitions (arrows). system’s prehistory. This means that the system has no memory on how it reached state i. If this is the case, we have to deal with so called Markov processes[106]. The time evolution of such a system is governed by the master equation

dPi(t) = j=iRijPi(t)+ j=iRjiPj(t), (4.5) dt − 6 6 X X whereas Pi(t) is the probability to ﬁnd the system at time t in state i. Essentially, kMC models the time evolution by solving the system’s master equation in a stochastic framework. Often the number of possible states is huge and their interrelation complex, as sketched in Fig. 4.3 (b), and an analytical solution of equation (4.5) is not feasible. kMC oﬀers a rigorous procedure to correctly propagate the system from state to state. Such a trajectory (red in Fig. 4.3 (b)) is called Markov chain. If one computes a large number of Markov chains and averages over these, one obtains a numerical solution of equation (4.5) and thus the system dynamics (for the reduced set of transitions). The kMC algorithm itself, which will be presented in 4.2.1, is nothing but a stochastic procedure. It is up to the researcher to incorporate the physics describing the system under investigation via the transition rates. 70 4. Methods for Transport Simulations

Figure 4.4: (left) Trajectory of an atom adsorbed on a solid surface. Vibrational motion of the atom around adsorption sites is interrupted by occasional hopping to a neighbouring adsorption site. (right) Coarse-grained picture of the dynamics. Vibrational motion is neglected.

Simple Showcase

Let us now regard a simple showcase: a single atom chemisorbed on a surface[105]. We want to study the atom’s diffusive motions between different adsorption sites, drawn in Fig. 4.5 on the left side. In this case the macroscopic system (solid surface + adatom) state is defined by the location of the adatom. The surface atoms are assumed to be immobile. Thus, the system dynamics reduce to the vibrational motion of the adatom and to the hopping between adsorption sites. We further reduce the dynamics by allowing only nearest-neighbour hopping, since we expect this to be much more probable than longer jumps, and by neglecting the vibrational motion of the adatom. After this coarse-graining procedure only the hopping events are left, as depicted in Fig. 4.5 on the right. How can we justify these simplifications? To this end we regard the potential energy surface (PES) of our system, sketched in Fig. 4.5 (a). At a certain instant of time the adatom is found at a certain adsorption site, corresponding to a local energy minimum. In this basin it vibrates, whereas this vibrational motion is orders of magnitude faster than the hopping events to another adsorption site, for which a comparably large energy barrier has to be overcome. If we would carry out a MD simulation, which would correspond to the pink trajectory in Fig. 4.5, even for such a simple system, we would be stuck, as we would need a time step size of 1 fs in ∼ order to resolve the vibrations of the adatom. This would mean that we would have to carry out 109 MD iterations to observe one single hopping event. This is, at ∼ 4.1. Kinetic Monte Carlo (kMC) 71

Figure 4.5: (a) PES of the solid surface + adatom system. Each energy basin corresponds to an adsorption site. The MD trajectory (pink) and the coarse-grained kMC trajectory (green) are shown. (b) Lines of constant potential energy and adatom trajectories are shown[101] . best, ineﬃcient and often simply not feasible. For the kMC calculation we neglect these vibrations saying that a vibrating atom does not leave a certain state in the sense that, if, at a certain instant of time, we would freeze all the atoms and reduce the system temperature to zero, the adatom would relax to the bottom of the basin, it is currently found in. In short words, we identify a macroscopic state with the energy basin on the PES, in which the adatom is found, and we associate all vibrational motions in this basin to the very same state. In this picture the dynamics are those of a so called infrequent-event system (IES), where long periods of inactivity are interrupted by fast transitions. kMC is predestined for such systems. Performing a kMC simulation for this simple showcase, we arrive at the green trajectory in Fig. 4.5. Compared to a MD simulation we generate the same sequence of adsorption positions, to which the atom moves. In this sense the trajectories of MD and kMC simulations can be regarded as identical, whereas the kMC simulation, in this case, would speed up the simulation by orders of magnitude. In general, of course, the investigated systems are of much higher complexity. While, as said, for MD the system dynamics emerge naturally, in the case of a kMC simulation the researcher’s intelligence and his understanding of the system is asked for. All possible events have to be foreseen. The kMC algorithm will carry out only those transitions speciﬁed in advance. If processes that are important for the system dynamics are overlooked, the simulation results will be completely incorrect. And 72 4. Methods for Transport Simulations

Figure 4.6: Sketch of the PES between two energy basins i and j. Barrier shapes and heights are often extracted from ab initio calculations or they are estimated from the experiment. often the transitions in a system will not be so well-separated in terms of their rates, i.e. there will be more or less a continuum of transition rates spanning many orders of magnitude in time, which will decrease the eﬃciency gain of kMC. Actually, this is one of the biggest problems of kMC, and it has to be addressed in an adequate way that depends on the system under study. We will come back to this point in 4.2.2. Back to the showcase. Via coarse-graining we have reduced the dynamics to the hopping between neighbouring energy basins, corresponding to diﬀusion of the adatom to a neighbouring adsorption site. In kMC the transition rates are user input, i.e. they have to be supplied in advance. In the present case, sketched in Fig. 4.6, the rate Rij for hopping from basin i to basin j depends on the shapes of the potential basins i and j, and the shape of the ridge-top connecting them. It can be calculated via ab initio methods[105] or it is derived from transition state theory[107, 108]. Then, it reads

b Ei→j R = ν e− kBT , (4.6) ij 0 ·

b with Ei j being the height of the barrier which separates the two basins. Accord- → ingly, one can calculate the rate for the back transition from j to i as

b Ej→i R = ν e− kBT . (4.7) ji 0 ·

As can be seen in Fig. 4.6, one would obtain for the ratio of the two rates 4.1. Kinetic Monte Carlo (kMC) 73

E −E Rij j i = e− kBT . (4.8) Rji Thus, the property of detailed balance is satisﬁed in this framework.

Figure 4.7: The adatom (red) has four neighbouring adsorption sites to which it can hop.

In a next step we will conveniently assume that all transition rates in the system are known. After the above-introduced reductions of the problem we are left with the situation depicted in Fig. 4.7. An adatom located at a certain adsorption site can carry out four diﬀerent transitions. In order to correctly compute the reduced dynamics, we have to answer two questions:

1. Which transition will be carried out?

2. When will the transition be carried out?

These questions were ﬁrst answered by Bortz, Kalos, and Lebowitz (BKL) and Gille- spie, who derived stochastic algorithms that randomly choose transitions and transition times under appropriate consideration of the transition probabilities/rates[96, 109]. Nowadays, these are known as BKL algorithm and Gillespie algorithm. In the next section we will introduce these kMC techniques together with their rigorous derivation.

4.1.4 A Simple kMC Algorithm

We summarize how, according to 4.1.3, a basic kMC algorithm should look like. A ﬂow chart is shown in Fig. 4.8. Every kMC simulation starts by deﬁning the initial macroscopic state of the system under investigation. In a next step a list of all 74 4. Methods for Transport Simulations

Figure 4.8: Flow chart containing the basic building blocks of a kMC algorithm. transitions from this state to other system states is compiled, and the corresponding transition rates are calculated according to the supplied models. This step also encompasses coarse-graining of the dynamics, if desired. We have already studied this on the showcase in subsection 4.1.3. The next step takes us back to the two fundamental questions on page 73. The algorithm has to provide procedures to, in a statistical sense, correctly choose the next transition and to compute the time ∆t, after which the transition should be carried out. After this is done, the system is propagated to the next state and the system time is incremented by ∆t. Then it is checked, whether or not the kMC loop, framed blue in Fig. 4.8, should be continued. If so, the next kMC iteration is started by recompiling the list of possible transitions. This list may have changed due to the last transition, since the system is now in another state. From this point one proceeds as described above until the user- speciﬁed abortion conditions are fulﬁlled and the kMC simulation is terminated.

Choice of Transition and Time Step Size

One kMC iteration allows us to perform one single step in a Markovian random walk. Via multiple iterations we obtain a Markov chain. As described above, the calculation and subsequent averaging over a large number of Markov chains, sampling the conﬁguration space, is equivalent to a numerical solution of the master 4.1. Kinetic Monte Carlo (kMC) 75

equation (equation (4.5)). To this end we have to determine by which transition µ and after which time τ the system will pass into another state. In the following, we will closely follow the ideas of Gillespie[96] and adapt them to our needs[110]. The problem we are dealing with may be outlined as follows: We are given a system in a certain macroscopic state with a set of M possible transitions T (1 µ M). µ ≤ ≤ The set T will usually comprise a multitude of, in part, very diﬀerent physical { µ} processes. We aim at simulating the time evolution of the system, knowing only the initial state, the forms of possible transitions Tµ, and the values of the associated transition rates Rµ.

We start by defining Rµdt as the probability, to first order in dt, that a particular transition Tµ will take place in the next time interval dt. Here, Rµ is the transition rate associated with transition Tµ. The time step dt is required to be infinitesimal to ensure that the probability of another transition T , ν = µ, to occur in the time ν 6 interval dt is negligible. In order to derive a rigorous algorithm for simulating the time evolution of the system, we define the probability density function (PDF) P (τ,µ) as

P (τ,µ)dτ = probability at time t for the system to undergo a transition Tµ within the diﬀerential time interval (t + τ,t + τ + dτ).

P (τ,µ) is a joint PDF on the space of the continuous variable τ (0 τ < ) and ≤ ∞ the discrete variable µ (µ is a positive integer; 1 µ M). To correctly compute ≤ ≤ the stochastic dynamics of the system, we have to randomly choose a pair (τ, µ) according to P (τ,µ)dτ, thus specifying the next step in the Markov chain. To derive an exact analytical expression for P (τ,µ), we calculate P (τ,µ)dτ as the product of

P0(τ), the probability at time t that no transition will occur in the time interval

(t,t + τ) and Rµdτ:

P (τ,µ)dτ = P0(τ)Rµdτ. (4.9)

To calculate P0(τ), we devide the interval (t,t + τ) into K subintervals of length

ǫ = τ/K. Then the probability that none of the transitions Tµ occurs in the ﬁrst subinterval (t,t + ǫ) is, following the deﬁnition of Rµdt and the multiplication theorem for probabilities,

M M [1 R ǫ] = 1 (R ǫ), (4.10) − µ − µ µY=1 µX=1 whereas terms of higher order in ǫ have been neglected. This is also the subsequent probability that no transition occurs in (t + ǫ,t + 2ǫ) and then in (t + 2ǫ,t + 3ǫ), etc. 76 4. Methods for Transport Simulations

Then P0(τ) reads

K M P (τ)= 1 (R ǫ) (4.11) 0  − µ  µX=1   K M = 1 (R τ/K) . (4.12)  − µ  µX=1   In particular we are interested in the limit of inﬁnitely large K. Thus,

M K µ=1 (Rµτ) P0 (τ) = lim 1 (4.13) K  − K  →∞ P  M  Rτ =exp R τ = e− (4.14) −  µ  µX=1     Here, R M R is the cumulative transition rate. Equation (4.14) simply de- ≡ µ=1 µ scribes theP exponential decay of the existing state through all available transition channels. Inserting equation (4.14) into equation (4.9) delivers

Rτ Rµe− , if τ [0, ), 1 µ M, P (τ,µ)=  ∈ ∞ ≤ ≤ (4.15) 0, otherwise.

This PDF is properly normalized, as can be seen by evaluating

M M ∞ ∞ dτ P (τ,µ)= Rµ dτ exp( Rτ) (4.16) 0 0 × − Z µX=1 µX=1 Z 1 =R − exp( Rτ) ∞ (4.17) R − 0 = (0 1) = 1 (4.18) − − We continue by writing the two-variable PDF as product of two one-variable PDFs:

P (τ,µ)= P (τ) P (µ τ), (4.19) 1 × 2 | where P1 (τ)dτ is the probability that the next transition will occur between t + τ and t+τ +dτ, irrespective of which transition it might be. P (µ τ) is the probability 2 | that it will be a Tµ transition, given that the transition occurs between t + τ and t + τ + dτ. Using equations (4.15), (4.19) and the addition theorem for probabilities[96], we obtain: M Rτ P (τ)= P (τ,µ)= R e− , (4.20) 1 · µX=1 4.1. Kinetic Monte Carlo (kMC) 77

P (τ,µ) R e Rτ R P (µ τ)= = µ · − = µ . (4.21) 2 | P (τ) R e Rτ R 1 · − The idea for the MC step is to ﬁrst generate a random number value τ according to the distributionP1 (τ), which determines the time after which the next transition will occur, and then to generate a random integer µ according to the distribution P (µ τ) 2 | to select the transition. The resulting pair (τ,µ) will be distributed according to P (τ,µ).

Inversion Method

The so called inversion method[96] allows us to use random numbers from the uniform distribution in the unit interval to construct random numbers distributed according to any prescribed PDF. Suppose one wishes to generate a random number according to the PDF P (x). Then P (x′)dx′ is the probability that x will lie between x′ and x′ + dx′. One can deﬁne the function

x F (x)= P x′ dx′. (4.22) Z−∞ Then F (x ) is the probability that x < x . Note that F ( )=0 and F ( ) = 1. 0 0 −∞ ∞ F (x) rises monotonously, since P (x′) > 0 for all x′. The inversion method can be described as follows. To generate x according to a given PDF, one simply has to draw a random number r from the uniform distribution in the unit interval and take for x that value which satisﬁes F (x)= r; in other words 1 1 take x = F − (r), where F − is the inverse of F , whose existence is ensured by the monotonicity of F . To prove the correctness of this approach, we calculate the probability that x lies between x′ and x′ + dx′. By construction this probability is the same probability that r will lie between F (x′) and F (x′ + dx′), which is

F x′ + dx′ F x′ = F ′ x′ dx′ = P x′ dx′. (4.23) − Thus, x is indeed distributed according to P (x).

Determination of Time Step Size τ

We apply the inversion method to the desired PDF for the size of the time step τ, given in equation (4.20),

R exp( Rτ), 0 τ < , P1 (τ)=  × − ≤ ∞ (4.24) 0, else

 78 4. Methods for Transport Simulations

and obtain

τ τ F (τ)= P1 τ ′ dτ ′ = R exp Rτ ′ (4.25) 0 Z−∞ Z R τ = exp Rτ ′ = [ exp( Rτ) + 1] (4.26) R − − 0 − − h i F (τ) = 1 exp( Rτ) =! r (4.27) ⇔ − − ln(1 r) τ = − − (4.28) ⇔ R ln(r) τ = − . (4.29) ⇔ R In the last step of the calculation (1 r) was replaced by the statistically equivalent − random variable r.

Choice of the Transition Tµ

First, we brieﬂy demonstrate, how the inversion method can be applied to discrete variables. The task shall be to generate a random integer number i distributed according to P (i). P (i′) is the probability that i = i′. The corresponding distribution function is

i F (i) P i′ , (4.30) ≡ i′= X−∞ and F (i ) is the probability that i i . In this case, one has to draw a random 0 ≤ 0 number r and take i, so that

F (i 1) < r F (i). (4.31) − ≤ Proof: i i 1 − F (i) F (i 1) = P i′ P i′ = P (i). (4.32) − − i′= − i′= X−∞ X−∞ The desired PDF for the choice of the transition Tµ was given in equation (4.21),

R P (µ τ)= µ . (4.33) 2 | R Thus, we have to choose µ such that

µ 1 µ − P2 µ′ τ < r P2 µ′ τ (4.34) ′ | ≤ ′ | µX=1 µX=1 or 4.1. Kinetic Monte Carlo (kMC) 79

µ 1 µ ′− Rµ′ ′ Rµ′ µ =1 < r µ =1 (4.35) P R ≤ P R µ 1 µ − wµ′ < r wµ′ . (4.36) ⇔ ′ ≤ ′ µX=1 µX=1

Here, we have introduced the transition probabilities wµ, which are easily obtained by normalization of the corresponding transition rates Rµ to the cumulative transi- M tion rate. Of course, then µ=1 wµ = 1. P

Figure 4.9: Flow chart of the kMC algorithm, cf. Fig. 4.8.

In summary, one kMC step, i.e. the task of generating a random pair (τ,µ), distributed according to P (τ,µ), is cut down to the task of generating two random numbers r1, r2 from the uniform distribution in the unit interval. τ and µ then follow from eqns. (4.29) and (4.36). Accordingly, we adapt the ﬂow scheme of the kMC algorithm, see Fig. 4.9. 80 4. Methods for Transport Simulations

4.2 Leakage Current Simulations via kMC

As described in the previous section, kMC provides a rigorous stochastic framework[95, 96] to simulate the dynamical behaviour of a physical system, provided that valid rate models for all relevant physical processes are supplied. The well-deﬁned relation between kMC simulation time and real time, for instance, allows one to study the real-time dynamics of the charge carriers, whereas kMC captures the statistical interplay between diﬀerent intertwined microscopic processes. This makes it a perfect candidate for our purposes. In the remainder of this section, we will closely follow the general framework introduced in 4 and will shed light on its implementation for our transport simulations.

Problem Deﬁnition

In the present work, transport through high-κ dielectrics was modeled in a MIM capacitor configuration, whereas the high-κ film works as an insulating material in between the two capacitor electrodes. The problem, we are dealing with, may be outlined as follows. As simulation region we choose a cuboid-shaped volume V of dielectric material with the two metal electrodes attached to opposing faces. As said earlier, charge carriers are assumed to flow preferentially across defect states in the band gap of the high-κ material. Thus, by default, N point-like defects are randomly allocated in V , with N = n V , D D D × where nD is the user-specified defect density. Each defect site can accommodate at most one charge carrier, as long as we exclude multiple occupation. Within the electrodes a continuous Fermi-Dirac distribution of the charge carriers is assumed.

Since our studies focus on the TiN/ZrO2 system, where hole transport is negligible due to the large VB oﬀset[111], we will limit our discussion to electrons. Nonethe- less, the method is equally applicable to holes. We neglect the time electrons in the CB need to reach the anode, as it is small compared to the time needed by the tunneling processes[111]. Upon entering V , the electrons are discretized and thenceforth treated as point-like particles with a deﬁned energy until they leave the dielectric to one of the electrodes. Their individual trajectories are simulated. The number of electrons, N, in V may change upon injection from or emission into the electrodes. The leakage current through the dielectric then simply equals the net number of charge carriers that pass the dielectric in a given time interval. 4.2. Leakage Current Simulations via kMC 81

Following the remarks of the previous section, we first have to define what, in this context, is to be understood as a macroscopic state. We define the state of the system at hand, i.e. the MIM structure, via the electron distribution, i.e. the positions of all N electrons in the dielectric. We regard the movement of a single electron as a transition of the whole system from one macroscopic state to another. Further following the recipe presented in 4, we provide the full Figure 4.10: Schematic band diagram set of possible transitions Tµ . They { } of a MIM structure showing the trans- are illustrated in Fig. 3.1 and, for conve- port mechanisms, which are implemented nience, on the right in a smaller version in the kMC simulator. Larger version de- of the original figure. The list of possi- picted in Fig. 3.1. ble transitions Tµ, which are taken into account, reads:

(i) direct/FN tunneling

(ii) thermionic/SE

(iii) elastic/phonon-assisted tunneling into a defect

(iv) elastic or phonon-assisted tunneling out of a defect

(v) elastic or phonon-assisted tunneling between defects

(vi) PF emission out of a defect into the conduction band

(vii) defect relaxation upon electron occupation

Defect relaxation plays a special role, which will be revisited in 5.3. For now, we will neglect it. Details on the rate models for the transitions (i) to (vi) are provided in chapter 3. In short, we now have deﬁned our physical system and its macroscopic states, and we have provided a set T of transitions together with tested rate models. At this { µ} point, the kMC procedure, introduced in 4.1.4, allows us to perform the single steps in the Markovian random walk, thus mimicking the electron dynamics. Given an electron distribution at time t, we can determine which of the N electrons will move 82 4. Methods for Transport Simulations

in the next kMC iteration (also injection from the electrodes is possible), where it moves, i.e. also by which type of transition, and when it moves. The algorithm determines by which transition Tµ and after which time τ the system will pass into the next state.

4.2.1 Algorithm

Although some problems arise, to which we will dedicate special attention, in principle the practical implementation of the kMC algorithm is straightforward. It can be outlined as follows:

1) Set time variable t = 0 and deﬁne the initial state, i.e. the electron distribution,

by assigning N electrons to the ND defect states.

2) Solve Poisson’s equation on volume V with appropriate boundary conditions,

and calculate rates Rµ for all possible transitions. They collectively determine the PDF. Specify sampling times at which the current ﬂow is evaluated and a

stopping time tstop at which the simulation ends. Independence of the initial conditions has to be assured.

3) Generate a pair of random numbers (r1,r2), and, using equations (4.29) and (4.36), calculate the corresponding pair (τ,µ).

4) Advance t by τ, and carry out electron transition Tµ, i.e. adapt the electron distribution accordingly.

5) If t has been advanced through a sampling point, read out current. Check, if the current has reached a constant level. If t t , terminate calculation, ≥ stop otherwise return to step 2).

Let us briefly comment on the indivudual steps of the algorithm. 1) Setting up the system means that we first specify the simulation region, i.e. the region on which e.g. Poisson’s equation has to be solved and in which electrons are discretized. For a MIM structure, assuming continuous electron distributions within the electrodes, the dimensions of the simulation volume are those of the dielectric film. In transport direction these are given by the film thickness, i.e. of the order of 10 nm, and in the two lateral dimensions they are chosen as a compromise between fast simulation runs, sufficient statistics, and negligible edge effects - typically some ten nanometers. Defects are placed within the dielectric volume according to the 4.2. Leakage Current Simulations via kMC 83

desired distribution, e.g. homogeneously or with a concentration gradient from one electrode to the other. All physical properties that are necessary to characterize the dielectric material and the defects, such as permittivity, band gap, dispersion of the tunneling mass, defect depth and charge etc., are specified. 2) Many of the rate models, introduced in chapter 3, to describe electron transitions require precise knowledge of the electrostatic potential distribution within the dielectric. Transmission coefficients for instance depend on the shape of the tunneling barrier between initial and final electron position. Thus, we have to solve Poisson’s equation and, doing so, take into account electron-electron interactions as well as charged defects and image charge effects due to the presence of the electrodes. By default, we assume ideally flat electrode/dielectric interfaces. If we want to study electrode roughness effects, advanced techniques to generate the rough interfaces and the mesh (adaptive mesh refinement, cf. 5.2.4), on which Poisson’s equation is solved with a finite elements formalism, are applied. These allow us, e.g., to accurately determine the electric field distribution in the vicinity of convex protrusions. More details to that issue will be given in 5.2. In general, the boundary conditions for our simulation volume are of mixed Dirichlet/von Neumann type. At the faces to which the metal electrodes are attached the potential is given by the applied voltage, while on the other faces the normal derivative of the electric field was assumed to vanish. 3) to 5) As discussed in 4.1.4, the task of propagating the investigated system from one state to another can be reduced to picking two random numbers from the uniform distribution in the unit interval. After having chosen a transition, it is carried out, i.e. all necessary modifications to the system are made, mainly by manipulating matrices in which information on the current system state is stored. As said earlier, kMC is well-suited to study transient effects, e.g. charging of the dielectric by applying defined voltage pulses. However, the focus in this work was laid on the leakage current, i.e. the current which is observed after transient contributions have faded away. Thus, simulation times have to be chosen significantly longer than the slowest transient current contributions need to decay.

4.2.2 Time Scaling Problem

While the algorithm sketched in the preceding subsection is fairly straightforward, in practice problems arise due to the fact that the transition rates range over many orders of magnitude. It may happen that the major part of the computing time is used for the simulation of the fast events, while the rare events, associated with slow rates, are not captured by the simulation. Often these rare events are of particularly 84 4. Methods for Transport Simulations

high interest[112]. Thus, the algorithm should include a strategy to bridge the time scales, allowing one to access all of them. Different approaches to tackle this problem have been proposed, among these weighted kinetic Monte Carlo[112]. In our case, a rate equation Monte Carlo approach turned out to be the best-suited. The basic idea is to treat certain parts of the dynamics via rate equations and couple these to the kMC algorithm as follows: A simulation run starts with the full set of transitions T , i.e. all possible tran- { µ} sitions are allowed. After having executed a sufficient number of kMC iterations, the algorithm identifies the most frequent events, i.e. transitions Tξ with Rξ larger than a specified cutoff rate Rcutoff , and removes them from the simulation. Their effect on both current flow and electron distribution in the dielectric is, in the following, taken into account in the form of averaged values, extracted from the previous kMC iterations. A simple example for such a situation is a defect state close to an electrode. It will be frequently populated by electrons being injected from the electrode, and it will be depopulated by the same electrons being reemitted to the electrode (transitions of type (iii) and (iv) in Fig. 3.1). To save computation time these individual transitions are removed from the set T . Instead of resolving the { µ} fast fluctuations of the defect occupation between 0 and 1, we calculate the occupation probability f of the defect as ratio of time for which the defect has been occupied to the time for which it has been unoccupied. Before each kMC iteration, the defect occupation will be set either to 1 with a probability f, or it will be set to 0 with a probability (1 f). f is subsequently used as a weighting factor for − other transitions involving this defect. The size of the time step ∆t, which is calculated according to equation (4.29), is mainly determined by the largest rates, as these deliver the main contribution to the cumulative rate R. Now, following the above-described scheme, ∆t can be increased stepwise by decreasing Rcutoff . The inaccuracies resulting from this procedure were usually found to be negligible, as the Coulomb interaction between the charge carriers is screened very efficiently in the high-κ material. However, outmost care is necessary. Again, we are confronted with the task to increase the simulation speed by coarse-graining the dynamics without the loss of essential information which would deteriorate the real system dynamics. Chapter 5

Leakage Current Simulations

On downscaling of the DRAM storage capacitor, as discussed in chapter 2, the replacement of SiO2 with high-κ dielectrics became necessary to decrease ﬂowing leakage currents while achieving higher capacitance densities. Many diﬀerent materials have been studied as potential candidates for the dielectric, e.g. Al2O3,

HfO2, ZrO2, TiO2, and SrTiO3[36, 113, 114]. Among these, Zr-based dielectrics are currently favored[3, 115, 116], since ZrO2 in its tetragonal phase has a very high permittivity of κ 38 and a sufficiently large band gap of 5.5 eV, which suppresses ∼ quantum-mechanical tunneling currents. Since direct tunneling is not an issue, at least for the next chip generations, the focus is now on the reduction of the intrinsically high defect density in the aforementioned materials. The electronic states associated with these defects often lie within the band gap and thus open up pathways for undesirable leakage currents. One strategy to reduce the defect density that has been pursued is defect passivation via doping, e.g. with F or La[13, 41, 42]. Defect-assisted tunneling has proven to be hard to model. Although the leakage current in the aforementioned materials has been studied extensively, the analysis has mostly remained on the level of simplified compact models[86, 90, 117]. These typically account only for certain aspects of the observed transport and neglect other essential parts. The widely used model for PF conduction, cf. equation (3) in ref. [91], for example, describes the field-enhanced thermal emission of trapped electrons from defect states into the CB. However, electron injection into the defects is not considered, i.e., transport is implicitly assumed to be purely bulk limited. For HfO2, a TAPF conduction mechanism has been proposed, which accounts for the injection step but does not include the position-dependent alignment of defect levels with respect to the electrode Fermi level upon an applied bias[28].

85 86 5. Leakage Current Simulations

In summary, existing models usually fail to provide a complete picture of transport. Neither do they take into account that transport is mostly a multistep process, nor do they include the competition between parallel conduction channels, e.g. diﬀerent transport mechanisms, leading to their mutual interdependence.

This motivated the development of our kMC simulator, which is described in chapter 4. It can handle tunneling transport within the band gap, including transport across defect states, which is often predominant in high-κ dielectrics. Thus, our approach promises to remove the aforementioned restrictions.

In the following we will present the results of our study of leakage currents through DRAM storage capacitors with Zr-based high.κ dielectrics. By comparison of simulation and experiment transport models are derived. In a first step, models are developed under the assumption of idealized conditions. These assumptions are subsequently dropped stepwise. Electrode roughness is introduced, and its impact on the electronic transport is evaluated. Finally, the role of structural relaxation of defects upon charging is investigated with respect to its influence on the leakage current. Following such a hierarchical appraoch, the transport models are iteratively refined.

Identiﬁcation of the most important transport channels is intended to support the undergoing experimental eﬀorts toward system optimization and to outline potential for further scaling of the DRAM storage capacitor.

5.1 High-κ Thin Film TiN/ZrO2/TiN Capacitors

In the following, an elaborate evaluation of the TZT capacitor architecture with respect to its electrical performance is provided. Measurements of TZT capacitors are presented and subsequently compared to kMC simulation results. The transport model derived from this analysis is presented and its implications for further downscaling of the DRAM capacitor are pondered. Temperature and thickness scaling of the leakage current are discussed as well as contributions from diﬀerent defect positions to the overall leakage current. We analyze the mutual interdependence of the competing conduction mechanisms and test our model with respect to its sensitivity toward changes of the crucial model parameters. A concise overview of our analysis is also given in refs. [111] and [110]. 5.1. High-κ Thin Film TiN/ZrO2/TiN Capacitors 87

parameter symbol value

optical permittivity of ZrO2 ǫopt 5.6[31]

band gap of ZrO2 Eg 5.4 eV Huang-Rhys factor S 1

eﬀective phonon energy hω¯ 40 meV[31]

CB mass of ZrO2 mcb 1.2me[61]

VB mass of ZrO2 mvb 2.5me[61]

Table 5.1: Fixed model parameters used in the kMC simulations. In chapter 3 the sensitivity of the employed transport models with respect to changes of these parameters is discussed.

5.1.1 Measurements

TZT capacitors with 9 nm (TZ9T) and 7 nm (TZ7T) thick ZrO2 layers were investigated. Measured κ-values of 38 for both capacitors evidence that the ZrO2 is fully crystallized in the tetragonal phase. IV characteristics of both capacitors for temperatures between 25◦C and 120◦C are shown in Fig. 5.1. Dielectric breakdown of the samples occurred at voltages of 3.2 V and 2.6 V, respectively (not shown in Fig. 5.1). While a strong temperature dependence of the current was observed for voltages below 2.5 V and 2 V, respectively, above these values the leakage current scales weaker with temperature. Especially for lower temperatures, the slope of the curves increases. For voltages below 1 V constant, temperature-independent ∼ leakage currents are observed.

5.1.2 kMC Simulations kMC simulations for the TZ9T and TZ7T capacitors, whose leakage current curves are shown in Fig. 5.1, were carried out using best-estimate physical parameters from experiment and theory. The full set of employed model parameters is given in Tab. 5.1 together with the corresponding literature sources. Throughout all kMC simulations these values were kept ﬁxed. To avoid ambiguities in the interpretation of the experimental data, we have chosen the simplest model that describes the essential features of the measured currents. A single defect level and a uniform defect distribution in the dielectric were assumed. The model was calibrated on the 88 5. Leakage Current Simulations

voltage U (V)

0 1 2 3 -3

-4

(a) T9ZT 10

-5

-6

10 ) ) -7 transient currents 2

10 25°C

50°C -8

90°C

-9

120°C

simulation

-10

-2

-3

(b) T7ZT

-4

-5

-6 leakage current density j (A/cm

25°C 10

50°C

-7

90°C

-8

120°C

simulation

-9

1.0 1.5 2.0 2.5

voltage U (V)

Figure 5.1: Leakage current data (symbols) of (a) TZ9T and (b) TZ7T capacitor compared to kMC simulation results (lines). In the low-voltage region (exemplarily depicted in (a)) no comparison between experiment and simulation is possible, as here the measured current seems to be dominated by transient currents while the kMC results describe the steady state. Dielectric breakdown was observed for U ≥ 3.2 V and U 2.6 V, respectively (not shown). ≥

TZ9T capacitor. Only the CBO EB between TiN and ZrO2 and the defect depth

ED with respect to the ZrO2 CB were varied. Very good agreement between simulation (lines) and experiment (symbols) was achieved, especially in terms of the temperature dependence of the leakage current (Fig. 5.1 (a)). The same parameter set was used to simulate the TZ7T capacitor. As shown in Fig. 5.1 (b), thickness scaling of the leakage current is modeled very accurately. The experimental data is closely ﬁtted, and especially the more pronounced increase of the slope of the current curves (here at 2 V), compared to ∼ 5.1. High-κ Thin Film TiN/ZrO2/TiN Capacitors 89

the TZ9T capacitor, is reproduced.

5.1.3 Transport Model

The transport model, as it emerges from kMC simulations, is sketched in Fig. 5.2.

Figure 5.2: Schematic band structure of a TZT capacitor for low (left side) and high (right side) voltage. Arrows illustrate electron ﬂow, whereas a dashed arrow indicates that this electron emission path is slower than the other one. Barriers for tunnel emission are highlighted with color.

For the TZ9T capacitor, for voltages below 2.5 V the leakage current was found to be dominated by PF emission of electrons from positively charged defects with a depth E = 1.15 eV with respect to the ZrO CB and a density of n = 3 1018 cm-3. Elec- D 2 D · trons are injected from the TiN cathode into the defects via elastic and multiphonon- assisted tunneling. Detrapping occurs in a next step via PF emission into the ZrO2 CB, where the electrons drift to the anode. Thermally activated detrapping of electrons is the transport limiting step in this multistep process. Therefore, it governs the temperature scaling of the current, which is found to be strong in this voltage range. As the kMC simulations, in contrast to standard treatments, account for the injection step, it is possible to extract information on the CBO EB between TiN and

ZrO2, which was found to be EB = 1.74 eV. According to the metal induced gap states (MIGS) model[32], assuming a work function of 4.7 eV for TiN[118], an electron aﬃnity of 2.5 eV and a charge neutrality level of 3.6 eV for ZrO2[32], a barrier of 1.45 eV is predicted. Since the MIGS model tends to overestimate the barrier reduction due to Fermi level pinning for wide gap oxides[119], the extracted value seems reasonable (An upper limit for EB would be 2.2 eV in the Schottky limit, i.e. without any Fermi level pinning.) 90 5. Leakage Current Simulations

The defect level lies well above the electrode Fermi level, suggesting a limitation of the current ﬂow due to the low occupation probability of the defect states. In this sense, the current is electrode-limited, explaining the failure of the standard PF formula in describing the leakage current[111], since the latter implies a bulk limitation of transport. This point will be further discussed below. For U > 2.5 V, at lower temperatures T 50 C, an increase of the slope of the ≤ ◦ current-voltage curves was observed together with a reduced temperature scaling. Here, the dominant conduction mechanism changes from PF emission to tunnel emission, also known as trap-assisted tunneling (TAT), i.e. electrons predominantly escape from the defect states via tunneling through the potential barrier. At high voltages, when the barrier shape changes from trapezoidal to triangular (cf. Fig. 5.2), the rate for this mechanism strongly increases with the applied voltage, so that it ﬁnally surpasses the rate for PF emission. The stronger voltage dependence of the only weakly temperature-dependent TAT explains the increased slope. For T 90 C the rate for PF emission is higher than the one for TAT over the whole ≥ ◦ voltage range so that PF emission dominates the leakage current. We assume that below 1.3 V transient currents, which arise due to the applied ∼ voltage ramp (0.05 V/s), dominate over the steady-state leakage. These transient currents are thought to consist of two contributions, (1) dielectric relaxation currents and (2) charging/discharging phenomena like charge trapping/detrapping involving defect states in the ZrO2 band gap.

In ZrSixOy films, as investigated in ref. [120], enormous transient currents, depending on the voltage ramp rate, were found. To get a rough feeling, one could compare the measured currents at an applied electric field of 1 MV/cm. This corresponds to a bias of 5 V for the 5 nm thick film investigated in ref. [120] compared to 0.9 V for the 9 nm film of our investigation. In ref. [120] currents of 2 10 7 A/cm2, 10 7 A/cm2, × − − 4 10 8 A/cm2 and 3 10 8 A/cm2 were found for voltage ramp rates of 1 V/s, × − × − 9 2 0.33 V/s, 0.2 V/s and 0.14 V/s, respectively. We measured 10− A/cm for a voltage ramp rate of 0.05 V/s. Thus, the order of magnitude of the measured currents seems indeed to suggest a strong transient contribution. Additionally, these transient currents typically only weakly depend on the temperature[120], as it is observed here (see Fig. 5.1 (a)). However, the possibility cannot be excluded that a transport channel not captured by our model is present, e.g. one involving a second defect type. The analysis of transient leakage data, as suggested in ref. [121], should shed light on this in the future. In general, the modeling of leakage current is a complex problem with 5.1. High-κ Thin Film TiN/ZrO2/TiN Capacitors 91

many unknowns, especially in case of DAT. Thus, further thorough experimental investigation is needed in order to complement the present investigation and put the transport model to the test, e.g. deep-level transient spectroscopy [122] to study the spatial and energetic defect distribution.

5.1.4 Contributions from Diﬀerent Defect Positions

In the presence of an electric ﬁeld the energetic alignment of the defect states with respect to the cathode Fermi level Ef,l, see Fig. 5.2, depends on their distance from the electrode. Simulations show that, varying U, diﬀerent defect positions deliver the largest current contribution (see Fig. 5.3 and Fig. 5.4). In Fig. 5.3 the simulated

-2 defect

10 PF TAT

position ) ) 2

1.5 nm

-4

10 2.0 nm

2.5 nm

-6

3.0 nm

3.5 nm

-8

4.0 nm

sum

-10

-12

-14

10 leakage current density j (A/cm

-16

1 2 3

voltage U ( V)

Figure 5.3: Leakage current of the TZ9T capacitor at 25◦C. Current contributions of traps, lying at distances x = 1.5-4 nm of the cathode, are shown. Filled symbols indicate transport via PF emission, empty symbols transport via trap-assisted tunneling.

leakage current for the TZ9T capacitor at 25◦C is shown (continuous line). Con- tributions of defects to the leakage current have been resolved with respect to their position x. Additionally, Fig. 5.4 exemplarily depicts the occupation probability f of defects at x =2 nm and x = 4 nm as a function of U. 92 5. Leakage Current Simulations

voltage U ( V)

1 2 3 0 0

10 10

-2

(a) defect at 4 nm

-4 10

) ) 10 2

-4

-8

10 -6

-12

-8

-10 -16

10 10

(b) defect at 2 nm -2

-4

-8 PF (f=1)

-6 10 occupation probability f

TAT (f=1) 10

leakage current density j (A/cm -12 -8

TAT 10 10

-10

-16

1 2 3

voltage U ( V)

Figure 5.4: Occupation probability vs. voltage of defects at (a) x = 2 nm and (b) x = 4 nm is shown. Lines show upper limits for their current contributions, assuming f = 1 for all voltages. Symbols indicate simulated currents.

For U 1 V defects deeper in the dielectric, with x = 3.5-4.0 nm, deliver the high- ∼ est current contribution, as they are nearly level with Ef,l. Nonetheless, f stays 3 below 10− , since electron injection is quite slow, in particular slower than electron extraction via PF emission, which depopulates the defects. Contributions from defects lying even deeper in the dielectric are suppressed due to the very slow electron injection, while contributions from interface-near defects are low, since these lie well above Ef,l. Tunneling into these defects is very fast, however, reemission into the cathode is even faster, leading for defects at x = 2 nm to a low occupation f 10-6. ∼ For U 2 V defects closer to the cathode are gradually aligned with E , thus, their ∼ f,l occupation rises, as can be seen in Fig. 5.4 (b). Here, defects at 2.5-3.0 nm give the highest contribution to the current. At 2.5 V the dominant transport mechanism changes from PF emission to TAT, i.e. tunnel emission. In this case, defects between 2.0 and 2.5 nm, being nearly fully occupied, deliver the highest current contribution. Figure 5.4 also shows the theoretical maxima for the current contributions (lines) of the defects. Full occupation is assumed, i.e. f = 1, independent of U. The real contributions (symbols), 5.1. High-κ Thin Film TiN/ZrO2/TiN Capacitors 93

as discussed, are roughly obtained by multiplying the theoretical maxima with f. Especially for low voltages, the current, in this sense, is injection-limited. Furthermore, Fig. 5.3 illustrates the mutual interdependence of the parallel transport channels. At voltages where TAT dominates the leakage current, a drop of the PF current can be observed, although the PF emission rate still increases. In this voltage region, electrons are very eﬃciently extracted from the defects via tunneling and thus are no longer available for PF emission. The defect occupation decreases (see Fig. 5.4 (a)). Clearly the competition between the two transport channels alters their individual characteristics. As transport across interface-near defects is suppressed, for further scaling of these

ZrO2 films, the defect density deeper in the bulk has to be minimized. Concerning electrode roughness, the leakage current is expected to be rather insensitive to the local electric field enhancement near convex apertures. For extremely thin films, however, the more important consequence of electrode roughness is local thinning of the dielectric film. Then, TAT, depending exponentially on the tunneling barrier thickness, is expected to strongly increase and to become the dominant transport mechanism. Thus, also the sensitivity of the leakage current towards electrode roughness is expected to increase. This will be discussed in detail in 5.2.

5.1.5 Sensitivity Analysis

The most crucial simulation parameters, which were attuned in order to reproduce the leakage current data of the TZT capacitors, are the CBO EB between TiN and ZrO2 and ED, the defect depth. They were varied systematically in order to investigate the sensitivity of our model.

In Fig. 5.5 (a) simulation results for a ﬁxed EB = 1.74 eV and varying ED =

0.95-1.25 eV are shown. For 25◦C the leakage current is maximum for ED = 1.05 eV, while it is lower going to shallower or deeper defects. The shallower a defect is, the higher the rate for PF emission of trapped electrons. The occupation probability, however, is lower due to the higher energy diﬀerence to Ef,l. For deeper defects the occupation probability is high, while the PF emission rate is low. Thus, the existence of a maximum is expected. Temperature scaling of the leakage current is found to be stronger for deeper defects, as expected for a thermally activated process like PF emission. Increasing ED, the transition from PF conduction to TAT, visible as change in the slope, occurs at lower voltages, for the depicted set of defect depths best seen comparing ED = 1.25 eV and ED = 1.15 eV. Thus, the temperature scaling of the leakage current and the change of the dominant conduction mechanism can 94 5. Leakage Current Simulations

-31.0 1.5 2.0 2.5 3.0

(a) E 25°C 120°C E = 1.74 eV

-4 D B

0.95 eV

-5

1.05 eV

1.15 eV

-6 ) ) 10 1.25 eV 2

-7

-8

-9 25°C

120°C

-10

-3

(b)

E = 1.15 eV 25°C

-4 D

10 120°C

-5

-6

-7

E 25°C 120°C

10 B leakage current density j (A/cm

1.54 eV

-8

10 1.64 eV

-9

1.74 eV

1.82 eV

-10

1.0 1.5 2.0 2.5 3.0

voltage U ( V)

Figure 5.5: Leakage current data of the TZ9T capacitor for 25◦C and 120◦C

(symbols). Simulated currents for (a) ﬁxed EB = 1.74 eV and varying ED and (b)

ﬁxed ED = 1.15 eV and varying EB are shown. As in Fig. 5.1, the low-voltage region, dominated by transient currents, is not depicted. be used to determine ED, given EB. In Fig. 5.5 (b) simulation results for a ﬁxed

ED = 1.15 eV and varying EB = 1.54-1.84 eV are shown. For lower EB, especially at lower voltages, where none of the defects are aligned with Ef,l, the current rises, since the defects are pushed closer to Ef,l, and their occupation probability thus rises. The eﬀect is more pronounced at 120◦C, where the leakage current is mainly limited by the low occupation of the defects, while at 25◦C it is limited by the slow PF emission. Consequently, changing the CBO also changes the temperature scaling of the leakage current.

Without further information on EB, ED or the defect density nD, the triplet (nD,EB,ED), extracted from the kMC simulations is not completely unambiguous. However, as soon as one of these values is fixed, the others can be determined with high accuracy. We have chosen to fix n to 3 1018 cm-3, resulting in E = 1.74 eV, a reasonable D · B value for the CBO[111], and ED = 1.15 eV, a defect level that might be ascribed to 5.2. Electrode Roughness Effects 95

oxygen vacancies, cf. ref. [123].

5.1.6 Summary

Based on kMC simulations, a self-consistent leakage current model for TiN/ZrO2/TiN capacitors was developed. kMC proved to be a useful technique for studying the mutual interdependence of parallel transport channels. Most importantly, electron injection into defect states is accounted for. In contrast to the pure bulk-limitation of the transport suggested by the standard PF model, the CBO between TiN and

ZrO2 turns out to strongly limit the conduction. Bulk defects at a distance of 2-4 nm from the injecting electrode are found to deliver the main current contribution. These results strongly suggest an optimization of the electrode work function and a reduction of the bulk defect density as a viable strategy for decreasing the leakage currents.

5.2 Electrode Roughness Eﬀects

By introducing high-κ materials like HfO2 and ZrO2 as dielectrics in the DRAM storage capacitor[2, 3, 115, 116, 117] the undesired leakage currents could be reduced. However, the ITRS requires a downscaling of the storage capacitor to an EOT of 0.4 nm within the next two years. The necessary thickness reduction of the dielectric film will cause leakage currents to grow again. Sensitivity of these thin-film capacitors on interface effects is expected to increase. Thus, in the face of the stringent leakage current specifications for the DRAM storage capacitor, electrode roughness (ER) is one of the major concerns for the realization of future chip generations. ER, leading to local electric field enhancement (LEFE) close to convex protrusions of electrode material into the dielectric due to the aperture effect[124], was found to cause an increase of the leakage currents in thin-film capacitors [125, 126] as well as in the gate stack[127]. Standard one-dimensional models, usually employed to interpret experimental leakage data[90, 128, 129], cannot describe these roughness effects. To arrive at a comprehensive picture of transport in the TiN/ZrO2 material system, it has turned out to be imperative to take ER effects into account. Such a model, allowing to quantify the effect of ER on the leakage current of TZT capacitors, is still missing. So far, ER effects in thin-film capacitors have only been quantified based on a mean- field point of view. Via a perturbation solution of Poisson’s equation Zhao et al. 96 5. Leakage Current Simulations

studied the effects of ER on the electric field and the leakage current[130]. Rough electrode surfaces were found to increase the average electric field F in the dielec- h i tric. As many transport mechanisms are strongly field-dependent, such an increase of the electric field is expected to cause an increase of the leakage current, as it has been discussed for SE and PF emission. In the case of transport across defect states the mean-field approach implies the assumption that all defects in the dielectric are subject to an identical electric field F and that they equally contribute to the h i leakage current. Our investigations, cf. 5.1 and ref. [110], suggest that transport in TZT capacitors occurs across defect states in the dielectric via PF conduction for low and medium electric fields[111]. At high electric fields, and especially in extremely thin films, TAT may become dominant. Thus, every transport channel that can be used by charge carriers to cross the dielectric is linked to a point-like defect. The individual contribution of each channel to the conduction has been found to depend on both the position of the defect and the electric field it is subjected to[110], which, in presence of ER, can be strongly inhomogeneous. Thus, the mean-field point of view is not applicable to TZT capacitors. Instead, an approach that considers the local nature of the transport channels has to be chosen. So does our kMC simulation technique and thus it seems to be well- suited for this purpose. Here, tunneling transport of charge carriers in the band gap of the dielectric, including DAT mechanisms, is treated fully self-consistently. Transport is simulated at the level of single charge carriers, the trajectory of each being computed as it crosses the dielectric. The simulator is coupled to a finite- elements-based Poisson solver, working on an adaptively refined grid, cf. 5.2.4. In this way, the dielectric, bounded by rough electrodes, can be modeled accurately, and the local changes in the electric field, e.g. in the vicinity of surface features, are resolved. This allows to correctly take the individual current contributions of the transport channels into account. In section 5.1 and in refs. [110, 111] a transport model for TZT capacitors was developed by comparing kMC simulations to experimental leakage current data of TZT capacitors with thicknesses d of the dielectric film between 10 nm and 7 nm, corresponding to EOTs between 1.0 nm and 0.7 nm (assuming a permittivity of

40 for high-κ tetragonal ZrO2). Thickness scaling of the leakage current as well as temperature scaling between 25◦C and 120◦C was accurately reproduced by the simulations. Exemplary leakage current data of a 9 nm TZT capacitor (EOT of 0.9 nm) is shown in Fig. 5.1. 5.2. Electrode Roughness Eﬀects 97

Therefore, based on the closed microscopic transport model for TZT thin-film capacitors, outlined in 5.1, as a step of further refinement, we investigated the influence of ER on the leakage current. As will be seen, for thin dielectric films (<10 nm), these roughnesses at the interfaces between electrode material and dielectric play an important role. Besides leading to quantitative changes in the simulated leakage currents, our simulations suggest that ER roughness also induces qualitative changes in the transport for ever shrinking dielectric film thicknesses, as will be commented on in 5.2.5. We assessed the ultimate scaling potential of TZT capacitors in terms of EOT, film smoothness, thickness fluctuations, defect density and distribution, and CBO in a fictive scenario. Based on experimental data for an EOT of 0.9 nm, we extrapolated via kMC simulations down to EOTs of 0.4 nm. ER was described in the framework of fractal geometry, which will be introduced in 5.2.1. A concise overview of these ER studies has been published in ref. [131].

5.2.1 Fractal Geometry

Aiming for a comparative study of ER effects, the ability to classify and describe rough surfaces in quantitative terms is essential. Frequently, rough surfaces are well-described by the concepts of fractal geometry[132], as they often exhibit self- affinity[133]. Fractal geometry does not describe the roughness, but it allows to describe the way roughness changes upon alteration of the observation scale. The symmetry obeyed by self-affine objects on rescaling is used to classify them. Often rough surfaces, as formed during a growth process, belong to the class of self-affine random fractals, just like, e.g., a coastline, as shown in Fig. 5.6. Despite their randomness they are self-affine in a statistical sense. A mathematical defini- tion of self-affinity will be given below. A more descriptive explanation of what is understood under "self-affinity" is that if you regard a self-affine surface and zoom in/out, i.e. change the scale of observation, the surface will qualitatively still look the same. Many approaches exist to characterize such systems. Ref. [135] provides an excellent overview of these. Here, however, we will only introduce the approach, which fits best to our needs, namely, for random rough surfaces, the height difference correlation functions, defined as

q 1/q Cq(∆r)= h(x) h(x + r) . (5.1) | − | r =∆r D E| | Here, h(x) is the height of the rough surface at position x above a smooth refer- 98 5. Leakage Current Simulations

Figure 5.6: Section of the coastline of South Britain (source: GoogleEarth). In 1967 Mandelbrot published his famous paper "How Long Is the Coast of Britain? Statistical Self-Similarity and Fractional Dimension"[134], where he examined the coastline paradox, i.e. the observation that the length of the coast of Britain depends on the scale of measurement and theoretically is inﬁnite.

Figure 5.7: One-dimensional random rough surface, parametrized by h(x), above a flat reference plane. w measures the vertical surface fluctuations and will be used below as a parameter to classify random rough surfaces. 5.2. Electrode Roughness Effects 99

ence surface, in general providing a coarse-grained, smooth approximation to the rough surface, as shown in Fig. 5.7. stands for the average over all directions hi in the reference plane. The problem of describing the surface morphology is then transferred to describing the shapes of the correlation functions. The latter can be advantageous. Often the correlation functions have the form

C (∆r) ∆rαq (5.2) q ∼ over a substantial range of length scales.

In most simple cases the exponent αq is independent of q, at least for q > 0 (If

αq depends on q, one speaks of multi-affine structures.). Then, the rough surface is called a self-affine fractal with Hurst exponent α over the corresponding range of length scales. The mean height difference ∆h C(∆r) scales with horizontal ≡ distance ∆r as

∆h ∆rα. (5.3) ∼ Within the scaling regime the surface is invariant to a change of length scales by a factor of λ in lateral directions combined with a change of length scales by λα in vertical direction (cf. equation 5.7). The scaling regime is bounded by the upper and lower correlation lengths ξ+ and ξ in both the horizontal ( ) and the vertical − || ( ) direction, i.e. ⊥

+ ξ− < ∆r < ξ (5.4) k k + ξ− < ∆h < ξ , (5.5) ⊥ ⊥ related by

+ + α (ξ /ξ−)=(ξ /ξ−)− . (5.6) ⊥ ⊥ k k Mathematically, self-affinity then can be defined as follows: A surface is called self- affine within a certain scaling regime, bounded by equations (5.4) and (5.5), if its statistical properties are invariant to an anisotropical rescaling, i.e.

α h(x) λ− h(λx). (5.7) ∼ Here, λ is the horizontal scaling factor, and α is the above-introduced Hurst exponent, in the following referred to as surface smoothness (0 <α< 1). Why this term is appropriate can be understood regarding Fig. 5.8. There, self-aﬃne random rough 100 5. Leakage Current Simulations

0.1

0.5

0.9

0 vertical coordinate h (nm) vertical -2

-4

0 20 40 60 80 100

lateral coordinate x (nm)

Figure 5.8: Self-affine random rough surfaces generated with the algorithm introduced on page 103 by using identical random seeds but different values for the smoothness α. For better visibility the surfaces with α = 0.1 and α = 0.9 are offset by -2 nm and 2 nm, respectively. surfaces with different α, generated with the algorithm that will be introduced on page 103 by using identical random seeds, are shown. A low smoothness α = 0.1 corresponds to a very jagged surface, for which the amplitude of vertical fluctuations stays nearly constant down to smallest lateral length scales (short-range roughness). α 0 would mean ∆h = const., i.e. ∆h is completely independent of ∆r. The → higher α is, the smoother is the surface, since the amplitude of vertical fluctuations decreases with decreasing lateral length scale. In other words, α describes, how fast the size of vertical fluctuations changes upon a change of the horizontal length scale.

For a given surface α can be determined, e.g., from double-logarithmic plots of

C2(∆r),

C (∆r) ∆rα, (5.8) 2 ∼ the height difference correlation function for q = 2, vs. ∆r. In the self-affine scaling regime this should yield a linear curve with slope α. 5.2. Electrode Roughness Effects 101

5.2.2 Structural Characterization of the TiN Electrodes

The TiN electrodes, employed in the investigated TZT capacitor structures, were investigated with respect to their self-aﬃnity. Bare, planar TiN electrode surfaces, characterized via atomic force microscopy (AFM), were found to exhibit self-aﬃne scaling up to ξ = 20 nm, as can be seen in Fig. 5.9.

20 nm

0.6 (nm) 2 C

0.1

fit

1 10

lateral separation r (nm)

Figure 5.9: Height diﬀerence correlation function of a bare TiN electrode, extracted from AFM data, and ﬁt according to equation (5.8).

This allows a systematic description of ER in the framework of fractal geometry. Typically found TiN surface parameters are a smoothness of α = 0.6 and a thickness ﬂuctuation (or vertical surface width) of w 1.2 nm. w was introduced in Fig. 5.7. ∼ According to the AFM data this can be translated to a root-mean-square roughness of ∆r 0.3 nm via the empirical relation w 4 ∆r. ∼ ∼ ×

5.2.3 Generation of Random Rough Electrodes

For simulation purposes self-affine random rough surface profiles h(x) - characterized by the parameters α, ξ, and w, as introduced in 5.2.1 - were generated by adapting a modified midpoint displacement algorithm[136]. Fractional Brownian motion is at the very core of this algorithm, a famous special case of it being Brownian motion. Let us start with the well-known Brownian motion. We consider a one-dimensional 102 5. Leakage Current Simulations

random walker, that start at the coordinate x = 0. It can perform steps xi in +x- direction or in -x-direction. Then, its position after n steps is given by the sum

Xn = x1 + x2 + ... + xn. (5.9)

The steps are taken from a Gaussian distribution with x =0 and x2 = σ2, h ii i whereas denotes the average over the ensemble. We deﬁne the correlationD functionE hi as

F (p) x x . (5.10) ≡h i i+pi

If F (p) decays fast enough, such that ∞ F (p) < , then, for large p, the mean p=0 ∞ square displacement scales with p as[137]P

x2 p. (5.11) p ∼ D E This is the case for the well-known ordinary Brownian motion. If the above sum diverges, one arrives at

x2 p2α, (5.12) p ∼ D E where 0 <α< 1 is the Hurst exponent. This type of random walk is called fractional Brownian motion[138]. Obviously, then

1/2 x2 pα, (5.13) p ∼ D E reminding us of the height diﬀerence correlation functions, which were discussed above, cf. equations (5.1) and (5.8). This property can be exploited to generate fractal random rough surfaces with every desired smoothness α. 1 In general one can divide fractional Brownian motion into three categories: α< 2 , 1 1 1 α = 2 , and α> 2 . As said, α = 2 is the case of ordinary Brownian motion, where 1 the increments are independent from each other, i.e. uncorrelated. For α> 2 the increments are positively correlated, i.e., if xp0 > 0, then the probability for xp0+1 > 0 is larger than 0.5. Drawing the curve, which connects the points (p,Xp), one will 1 obtain a rather smooth curve. For α< 2 the opposite is true. Subsequent increments are negatively correlated, leading to a very jagged curve. This can immediately be transferred to the generation of rough surfaces, as they were depicted for varying α in Fig. 5.8. The simple algorithm used for the task of surface generation will be presented in the following. 5.2. Electrode Roughness Eﬀects 103

Midpoint displacement method

The aim is to generate fractal random rough surfaces (with a variance σ), satisfying the height diﬀerence correlation function

α C2(∆r)= σ∆r . (5.14)

Figure 5.10: Points (red crosses) generated by the midpoint displacement algorithm after i iterations. Vertical displacements decrease according to equation (5.22), here for α = 0.3.

Let us assume, we want to generate a surface h(x) on the interval 0 x L by ≤ ≤ generating a set of points (x,h(x)), with the desired resolution on [0;L], which will subsequently be connected by a continuous line (Straight lines are used in Fig. 5.10, while a continuous connection of the points is realized in the kMC simulator via cubic splines.). The midpoint displacement method is an iterative approach, from which one can stepwise obtain random rough surfaces, as illustrated in Fig. 5.10. We start by setting h(0) = 0 and selecting h(L) as a sample of a Gaussian random variable with mean 0 and variance L2ασ2. Then,

var(h(L) h(0)) = L2ασ2, (5.15) − and thus we expect for 0 < x1 < x2 < L 104 5. Leakage Current Simulations

var(h(x ) h(x ))=(x x )2ασ2. (5.16) 2 − 1 2 − 1 L We set h( 2 ) to be the average value of h(0) and h(L) plus a random Gaussian oﬀset 2 D1 with mean 0 and variance ∆1. Then, we arrive at

L 1 h( ) h(0) = (h(L) h(0))+ D . (5.17) 2 − 2 − 1 Consequently, h( L ) h(0) has a mean value 0 and so does h(L) h( L ). In order to 2 − − 2 fulﬁll equation (5.16) we must require

2α L 1 2 ! L 2 var(h( ) h(0)) = var(h(L) h(0))+∆1 = σ , (5.18) 2 − 4 − 2 which leads to

2 2α 2 σ L 2α 2 ∆ = 1 2 − . (5.19) 1 22α − L For the next iteration of the algorithm, we apply the above idea to h( 4 ), which gives

2α L 1 L 2 ! L 2 var(h( ) h(0)) = var(h( ) h(0))+∆2 = σ , (5.20) 4 − 4 2 − 4 from which follows

2 2α 2 σ L 2α 2 ∆ = 1 2 − , (5.21) 2 2 2α − (2 ) 3L and equally to h( 4 ). Increasing the resolution via further iterations, we obtain

2 2α 2 σ L 2α 2 ∆n = 1 2 − , (5.22) (2n)2α − n as the variance of displacement Dn, corresponding to lateral distances L/2 . Again, we see that the surface smoothness α, or the Hurst exponent, respectively, describes, how vertical fluctuations scale upon a change of the horizontal scale. The surface profiles generated in this way were subsequently fed into the kMC simulator. By connecting several surface segments of length ξ in series, larger surfaces with an approximate correlation length ξ and with the desired values for α and w can be provided. Throughout the following analysis ξ was fixed to 20 nm, the value extracted for the TiN electrodes via AFM measurement, and α and w were varied. The influence of the electrode topology on the electrical performance, i.e. the transport properties of the TZT capacitor devices was investigated. 5.2. Electrode Roughness Effects 105

5.2.4 Mesh Generation and Solution of Poisson’s Equation

The task at hand is the following. Poisson’s equation has to be solved within the simulation volume, in this case a cube with one or two rough faces, depending on whether one wants to study one or two rough electrodes at the same time. Dis- cretization of Poisson’s equation in the finite elements framework induces errors. Via posteriori error estimation one has to gain control of these, whereas the common approach is to increase the number of degrees-of-freedom in regions, where the initial finite elements model was not adequate. This is by nature an iterative process. In a first step a mesh is generated, on which the differential equation is solved. Then, errors are estimated, and, based on this estimate, the mesh has to be refined. The brute force approach would be to globally refine the mesh, i.e. to subdivide all the finite elements. However, to save computational resources, it is more convenient to refine only in those regions with larger errors, i.e. adaptively. After error estimation and mesh refinement, again, the differential equation has to be solved and the errors have to be estimated. Convergency of the finite elements solution to the correct results is monitored by suited error criteria, which are imposed, and the refinement loop is to be repeated until the desired accuracy of the solution is achieved. Kelly et al.[139] developed a means to locally estimate the error of the finite elements solution for every grid cell. This error estimator approximates the error per cell by integration of the jump of the gradient of the solution along the faces of each cell. It was specially designed in order to estimate the error for the generalized Poisson equation

ρ(r) (κ(r) Φ(r)) = , (5.23) ∇ ∇ − ǫ0 with either Dirichlet or generalized Neumann boundary conditions, i.e. exactly for our needs. Here, κ(r) and Φ(r) denote the position-dependent permittivity and electrostatic potential, respectively, and ρ(r) is the charge density. Implementation of the error estimator is straighforward. The error ηK for a cell K is calculated from

2 2 h ∂Φ(r) ηK = κ(r) dS. (5.24) 24 Z∂K " ∂n # The reason why we address this topic when discussing ER effects is that we expect greater variations in the electrostatic potential close to roughness features. These cause a LEFE, i.e. larger gradients of Φ, leading to higher errors of the finite elements solution in these regions. The simulation grid has to be fine enough to resolve these 106 5. Leakage Current Simulations

features. Therefore, we adaptively refine the grid, wherever it is necessary. This leaves us with grids, in which neighbouring cells may have a different degree of refinement. As a constraint, we allow neighbouring cells to differ from each other in their degree of refinement only by one. The framework for the generation of such grids, taking advantage of the above-mentioned error estimation, is provided by the open-source finite elements library "deal II", specially designed for the analysis of differential equations[140]. A typical iterative generation process of locally refined grids, as used in our simulations, can be seen in Fig. 5.11. First, for a given simulation volume, say a cube of 20 10 nm2 with one rough face, Poisson’s equation is solved on a globally refined × grid, as drawn in Fig. 5.11 (a), in order to identify regions of great variations in the electrostatic potential, which reduce the quality of the finite elements solution. Based on equation (5.24), the error for each grid cell is estimated and, e.g., the 30% of all cells with the largest error or those with an error above a critical value, are flagged. In a next step, the flagged cells, marking the critical regions, are refined, see Fig. 5.11 (b) . Again, for the new grid the errors are estimate, and, if necessary, subsequent refinement steps (Fig. 5.11 (c)-(d)) are carried out until one arrives at a final mesh that provides a sufficiently accurate solution of Poisson’s equation in the whole simulation volume. In Fig. 5.11 (e) a grid with a typical degree of refinement, as used for our kMC simulations, is depicted, and in Fig. 5.11 (f) the electric field distribution resulting from the solution of Poisson’s equation on this grid is shown. For an applied voltage of 1 V across a distance of 10 nm (from top to bottom face) one expects a mean electric field of 1 MV/cm. As can be seen, convex protrusions lead to a local enhancement of the electric field by nearly a factor of two. In these regions the electrostatic potential changes faster than in other regions. There, the grid has to be refined to a very high degree.

5.2.5 Results and Discussion

In the preceding paragraphs we have introduced all modifications to our kMC simulator that are necessary, if one intends to study the effect of ER on the leakage currents in TZT capacitor structures. We introduced fractal geometry as a means to classify and describe self-affine random rough surfaces, such as as-grown TiN electrode surfaces, and we presented the algorithm that allows to generate as-desired random rough surfaces, which are fed into the kMC simulator. Issues concerning grid generation, adaptive refinement, and the solution of Poisson’s equation in a simulation volume with rough boundaries were discussed. The framework for the 5.2. Electrode Roughness Effects 107

Figure 5.11: Grids with different degree of refinement, generated by the kMC simulator, for a simulation volume of 10 20 nm2 with one rough surface. (a) three × times refined globally, (b) three times refined globally and once adaptively(c) three times refined globally and twice adaptively, (d) three times refined globally and three times adaptively, (e) five times refined globally and three times adaptively. (f) Electric field distribution for a bias of 1 V applied between top an bottom surface. 108 5. Leakage Current Simulations

kMC simulations is set. For the sake of simplicity, in the following, we only investigate capacitor structures with one rough electrode. The second electrode is assumed to be perfectly flat. In reality of course both electrodes are rough. Often the large scale undulations of the top electrode roughly follow those of the bottom electrode, since the dielectric films in between the two electrodes, the substrate for the growth of the second electrode, are grown via ALD, which guarantees very homogeneous growth. The kMC simulator can handle a second rough electrode without further modifications. However, while adding a second rough electrode may lead to quantitative changes of the ER effects, it complicates the analysis, and it does not introduce additional phenomena besides those discussed below.

Small thickness ﬂuctuations

Figure 5.12: Electric field distribution in a capacitor with one rough (α = 0.5, w = 2 nm) and one flat electrode. A voltage of 0.9 V was applied across the dielectric with a mean thickness of d = 9 nm, leading to F = 1 MV/cm. White lines mark h i a distance of x = 2 nm and x = 3 nm from the rough electrode. In black, lines of constant electric field are drawn for a field enhancement of 10-50 %. The inset shows leakage current data of a 9 nm TZT capacitor at 120◦C and kMC simulation results assuming flat and rough (w = 2 nm) electrodes for 0.3 α 0.9. ≤ ≤

Figure 5.12 shows the simulated electric field distribution in a TZT capacitor with one rough (α = 0.5, w =2 nm, ξ = 20 nm) and one flat electrode. The ER leads to 5.2. Electrode Roughness Effects 109

a LEFE near convex protrusions of the electrode material into the dielectric due to the aperture effect. As can be seen in the inset of Fig. 5.12, the leakage current, although being dominated by PF conduction, which is strongly field-dependent, remains unaffected for small thickness fluctuations, up to w = 2.0 nm for the capacitor with a 9 nm thick ZrO2 dielectric. As discussed in 5.1, transport in the 9 nm TZT capacitor mainly occurs across defects in the interior of the film, 2-3 nm away from the injecting electrode (This region is marked in Fig. 5.12). These defects are subjected only to a weak LEFE, usually smaller than 10 %. The expected increase of the leakage current due to ER can be roughly estimated as follows. The rate for PF emission of electrons from defect states RPF is roughly given by equation (3.67)[85]

E e3F/(πǫ ǫ ) D − 0 opt RPF = ν exp q , × − kBT    where F is the electric field at the defect, ǫ is the optical permittivity and ν opt ∼ 13 1 10 s− a typical phonon frequency. We define the field enhancement factor γ as δF γ , (5.25) ≡ F h i where δF denotes a local change in the electric field with respect to the average field F . Then one can approximate (for δF << F ): h i 3 R ( F + γ F ) e F /(πǫ0ǫopt) PF h i h i exp h i γ . (5.26) R ( F ) ≈ q 2k T ·  PF h i B   For T = 393 K, F =1 MV/cm, and γ = 0.1 this yields h i R ( F + γ F )/R ( F ) 1.6, (5.27) PF h i h i PF h i ≈ i.e. only a negligible increase in the emission rate and, in turn, in the current is expected. Defects located near the electrode, which are subject to the strongest LEFE, do not significantly contribute to the transport because of their low electron occupation probability (Fig. 5.4). These findings retrospectively justify the assumption of flat electrodes made in 5.1 and in ref. [111], since for a 9 nm TZT capacitor, simulated there, the thickness fluctuations of the employed electrodes (w 1.2 nm) do not ∼ influence the leakage current. We also observe that the leakage current is, at least for an EOT of 0.9 nm, insensitive on changes of the smoothness α (see Fig. 5.13), in contrast to the predictions of the mean-field approach[130]. In Fig. 5.14 the electric field distributions in two capacitors are compared, whose electrodes only differ in the smoothness α, governing 110 5. Leakage Current Simulations

0 0

10 10

= 0.3 = 0.3

-2 -2

= 0.5 = 0.5

10 10

= 0.7 = 0.7

= 0.9 = 0.9

-4 -4

10 10 flat electr. flat electr. ) ) 2

data data

-6 -6

10 10

-8 -8

10 10

w = 3 nm (a) w = 4 nm (b)

-10 -10

10 10

1 2 3 1 2 3

0 0

10 10

= 0.3 = 0.3

-2 -2 = 0.5 = 0.5

10 10

= 0.7 = 0.7

= 0.9 = 0.9

-4 -4

10 10 flat electr. flat electr.

data data

-6 -6

10 10 leakage current density j (A/cm

-8 -8

10 10

w = 5 nm (c) w = 6 nm (d)

-10 -10

10 10

1 2 3 1 2 3

electric field F( MV/cm)

Figure 5.13: Leakage current data of a 9 nm TZT capacitor at 120◦C and kMC simulation result assuming ﬂat electrodes, taken from ref. [111]. Leakage currents were simulated for 0.3 α 0.9. Assumed thickness ﬂuctuations w were (a) 3 nm, ≤ ≤ (b) 4 nm, (c) 5 nm, (d) 6 nm.

the short-range roughness. Small surface features are more defined for the surface with the lower α. On a larger length scale the surface topologies, i.e. the surface undulations, are very similar. The field enhancement due to the short-range roughness is highly localized near the surface. Deeper in the bulk the field distribution is insensitive to such roughness. Here, the large-scale undulations determine the field distribution which is consequently nearly identical for α = 0.3 and α = 0.7. As the defects which deliver the highest current contributions lie in the bulk region, the leakage current does not depend on α. 5.2. Electrode Roughness Effects 111

Figure 5.14: Electric field distributions in a capacitor with one flat and one rough electrode (w = 3 nm). Identical surface profiles of the rough electrode were assumed in (a) and (b), differing only in the smoothness, taken as (a) α = 0.3 and (b) α = 0.7, respectively. A voltage of 0.9 V was applied across the dielectric with a mean thickness of 9 nm, leading to F = 1 MV/cm. In black, lines of constant electric h i field are drawn.

Large thickness ﬂuctuations

For larger thickness fluctuations convex protrusions of electrode material extend further into the dielectric. Consequently, besides a LEFE confined to the direct surrounding of sharp surface features, a second effect becomes important: the local reduction of the dielectric film thickness, e.g. from d to dmin in Fig. 5.14 (a). At such constrictions a long-range increase γc of the electric field, which extends through the whole film, is observed. Assuming d d w/2, we can estimate γ in terms of min ≈ − c relative thickness fluctuations w/d as

Fc F U/dmin 1 γc(w/d)= −h i = 1 , (5.28) F U/d − ≈ 2 1 h i w/d − where Fc is the average electric ﬁeld in the region of the constriction. In this case, those defects which deliver the dominant contributions to the leakage current are subject to an average electric ﬁeld enhancement γc, e.g. γc(3 nm/9 nm) = 20% for the capacitor in Fig. 5.14 (a). Thus, the PF conduction slightly increases, as seen in Fig 5.13 (a). 112 5. Leakage Current Simulations

However, up to w = 5 nm (Fig. 5.13 (b)-(d)) the 9 nm TZT capacitor fulfills the leakage criterion. Experimentally achievable thickness fluctuations are far below this value. Thus, for an EOT of 0.9 nm the effects of ER on the leakage current are uncritical for the operation of the DRAM storage capacitor.

Downscaling of the capacitor

2 2

10 10

data

w = 0 nm

0 w = 0 nm 0

w = 1 nm

10 10

w = 3 nm

w = 2 nm

w = 4 nm

-2 w = 3 nm -2

10 10 w = 5 nm

w = 4 nm

w = 6 nm ) )

-4 -4 2

w = 7 nm

10 10

-6 -6

10 10

-8 -8

10 10

EOT = 0.9 nm (a) EOT = 0.6 nm (b)

-10 -10

10 10

1 2 3 1 2 3 2 2

10 10

w = 0 nm w = 0 nm

0 0

w = 1 nm w = 1 nm

10 10

w = 2 nm w = 2 nm

-2 -2 w = 3 nm w = 3 nm

10 10

-4 -4

10 10

-6 -6 leakage current density j (A/cm

10 10

-8 -8

10 10

-10 -10

10 10

1 2 3 1 2 3

electric field F( MV/cm)

Figure 5.15: Leakage current simulations for EOTs of (a) 0.9 nm, (b) 0.6 nm, (c) 0.5 nm, (d) 0.4 nm. As the leakage current does not depend on α, shown currents are averaged over 0.3 α 0.9. ≤ ≤ To study how ER effects on the leakage current behave on further downscaling of the capacitor, we carried out leakage current simulations for capacitor structures with film thicknesses of 6 nm, 5 nm, and 4 nm, as shown in Fig. 5.15 (b) - (d). For flat electrodes (w = 0 nm) PF conduction remains the dominant transport mechanism down to 4 nm. However, in comparison to the 9 nm film (Fig. 5.15 (a)), a drastically 5.2. Electrode Roughness Effects 113

-2

10 ) ) 2

-4

w( nm)

-6 5

-8

10 1

-10 current j at 1.3 MV/cm (A/cm MV/cm 1.3 at j current 10

0.5 1.0

EOT ( nm)

Figure 5.16: Simulated leakage currents at F =1.3 MV/cm for different EOTs and h i thickness fluctuations w. Upon downscaling the sensitivity of the leakage current toward ER drastically increases. increased sensitivity of the leakage current on ER is observed, as illustrated in Fig. 5.16. In fact, when comparing equal thickness fluctuations w for different film thicknesses d, the relative thickness fluctuations, i.e. w/d, increase going to thinner dielectrics. Thus, the electric field enhancement γc(w/d) at the dielectric constrictions is stronger. However, the resulting increase of PF conduction is not sufficient to explain the large increase of the leakage current observed for EOTs 0.6 nm. Most ≤ important for scaling, as it turns out, is the transformation of the dominant transport mechanism from PF conduction to TAT, evidenced by the kMC simulations. TAT scales much stronger with the film thickness than PF conduction, since it depends exponentially on the thickness of the tunneling barriers. Thus, for extremely thin films and large enough thickness fluctuations TAT across local constrictions of the dielectric becomes the dominant transport mechanism and completely controls the leakage current once w is further increased. Accordingly, for the upcoming chip generations the requirements on the uniformity of the fully 3D-integrated electrodes drastically increase, e.g. to w 1 nm for an EOT ∼ of 0.5 nm (see Fig. 5.16). The feasibility of an EOT of 0.4 nm seems questionable. Strategies to achieve the smallest possible EOT values are needed. Based on the 114 5. Leakage Current Simulations

presented transport model, and considering ER eﬀects, two technological options exist.

The ﬁrst is the reduction of the defect density nD. As discussed, the dominant transport mechanisms PF conduction and TAT both involve bulk defects in the

ZrO2. Usually each single defect opens up one transport channel. Thus, the leakage current is approximately proportional to nD, whereas only the contribution of bulk defects in a certain distance from the electrodes, e.g. 2-3 nm for a 9 nm TZT capacitor, should be considered. Both PF conduction and TAT would be aﬀected by a reduction of nD in the same way. In our simulations a homogeneous defect density n = 3 1018 cm 3 has been assumed. How much this defect density can D × − be reduced in the future is not clear. Further reﬁnement of the transport model by integration of structural relaxation of defects, discussed in the following section, suggests the possibility of higher defect densities what would make this strategy appear more promising. The second option is to increase the CBO, e.g. by introducing an Al-doping at the

TiN/ZrO2 interface[36, 118, 141]. This option is a unique result of our model, as standard leakage current models for TZT capacitors describe the current as purely bulk-limited PF conduction and do not include the electron injection and thus the influence of the electrode work function. Figure 5.17 depicts a possible scaling scenario for TZT capacitors to realize an EOT of 0.4 nm. An increase of the CBO by 0.1-0.2 eV with respect to 1.74 eV, extracted from kMC simulations, is sufficient to fulfill the leakage current criterion for 0.4 nm EOT.

5.2.6 Summary

Performing 3D kMC simulations, we have carried out a quantitative analysis of the influence of electrode roughness on the leakage current in TiN/ZrO2/TiN capacitors for application in DRAM cells. While the leakage current is found to be insensitive on the electrode smoothness (α), thickness fluctuations (w) turn out to have a huge impact. We have found that, upon downscaling the EOT, the sensitivity of the leakage current on thickness fluctuations drastically increases due to a change of the dominant transport mechanism from PF conduction to TAT. For a feasible thickness fluctuation of w 1 nm the leakage current criterion could not be fulfilled ∼ for a desirable EOT of 0.4 nm. According to the new insights into the transport mechanisms in TiN/ZrO2/TiN capacitors an increase of the CBO by 0.1-0.2 eV is suggested in order to enable an EOT of 0.4 nm. 5.3. Structural Relaxation of Defects 115

) ) 10

E E = 1.15 eV

B D

1.64 eV

-2

1.74 eV 10

1.84 eV

1.94 eV

-4

-6

-8

EOT = 0.4 nm

-10 leakage current density j (A/cm j density currentleakage

1 2 3 4

electric field ( MV/cm)

Figure 5.17: Leakage current simulations of a 4 nm TZT capacitor (EOT=0.4 nm) at 120◦C, varying the CBO EB between TiN and ZrO2.

5.3 Structural Relaxation of Defects

So far, all studies of leakage currents in high-κ dielectrics did not take into account structural relaxation of defects upon a change of their charge state (e.g. refs. [90, 129]). Also the transport model, developed in 5.1, was set up under the assumption that structural relaxation is negligible. However, recently, M. Häufel carried out accurate ﬁrst principles calculations of defect energy levels for diﬀerent charge states in t-ZrO2[142], which, in contrast, demonstrate that structural relaxation is very pronounced in t-ZrO2, leading to shifts of the defect energy levels larger than 1 eV. Such large energy shifts suggest that in case of charge transport across defect states, as it is envisaged for TZT capacitors, cf. 5.1 and 5.2, structural relaxation may play a major role.

Consequently, a defect model was derived from Häufel’s ﬁrst principles calculations and subsequently coupled to our kMC charge transport simulations. ZAZ thin ﬁlms, which were shown to have slightly better insulating properties than pure

ZrO2 ﬁlms[3], were taken as an exemplary system to assess the impact of structural relaxation on the electron transport. 116 5. Leakage Current Simulations

5.3.1 Introduction

For decades, semiconductor bulk defects have been subject of intensive studies[143] and it is common knowledge that the energy of defect levels depends on the defect charge state[144]. Electrostatic interactions between charges localized at defect sites and the surrounding lattice atoms lead to structural relaxation. The atoms in the vicinity of the defect undergo a rearrangement in order to minimize the total energy of the system. Remarkable shifts of atoms, and thus defect levels, can be observed, a prominent example being the bistability of the oxygen vacancy in SiO2

(E′-center)[145]. Although structural relaxation of defects caused by a change of their charge state has been known for a long time, its influence on charge transport across defect states has been scarcely addressed[146]. Especially in modern nanoscale devices, featuring film thicknesses below ten nanometers, unanticipated effects may arise. Especially for the high-κ oxides this is to be expected. It is well-known that the high-κ oxides, due to their large coordination, suffer from high defect densities[39]. As the main fields of application of these materials, usage as dielectrics in the state- of-the-art transistor gate stack[147], in future carbon-nanotube transistors[148], and in the storage capacitor of DRAMs[2], impose high requirements on their insulating properties, numerous studies of defect levels in materials like HfO2, ZrO2, TiO2, and SrTiO3[123, 149, 150, 151] were conducted, which incidentally demonstrated the enormous effect of structural relaxation on the defect energy levels in these materials. As has been discussed in the previous sections, transport mechanisms involving charge flow across defect levels in the oxide band gap dominate the conduction in high-κ films currently employed as dielectrics[110, 111]. Obviously, as soon as charges flow across defects, causing frequent changes of the defect charge states, defect relaxation will strongly influence the electronic transport. Considering the current stand of the semiconductor industry, which, only a few years after successful integration of high-κ dielectrics, is pushing these materials to their scaling limits[152], a better understanding of the charge transport is crucial. Envis- aging the current downscaling speed, a value of 0.4 nm for the EOT will be required within the next two years (according to the ITRS). Yet no feasible concept is known that fulfills the strict leakage current criteria for the DRAM storage capacitor and, at the same time, enables such low EOTs. Future material systems are investigated, whereas the focus is laid on SrTiO3[153, 154, 155] and rutile TiO2[36] as dielectrics in combination with Ru-based electrodes, since these materials provide permittivi- 5.3. Structural Relaxation of Defects 117

ties of more than 100. However, these material systems are far from being ripe for integration into the large-scale production process. Thus, other approaches have been suggested, under these laminate dielectric stacks, such as trilaminate ZAZ dielectrics[3, 116, 89], which are used in the latest chip generations. Here, two polycrystalline layers of t-ZrO2, featuring a permittivity close to 40, surround a thin ( 0.5 nm) amorphous Al O interlayer. It was found ∼ 2 3 that these ZAZ ﬁlms have insulating properties slightly superior compared to pure

ZrO2 films[3], which were discussed in 5.1 and 5.2. The actual reason for this improvement is still a subject of discussion. It was suggested that the interlayer prevents ZrO2 grains from extending over the whole dielectric and causing excessive surface roughness[156], which is known to have a detrimental effect on the capacitor performance[131]. It was also argued qualitatively that the large band gap of Al2O3 sets up a barrier for charge flow[116]. However, a quantitatively predictive model for the leakage currents in TZAZT capacitors is still missing. The fact that structural relaxation of defects easily causes energetic shifts of the defect levels larger than one electron volt[157] strongly suggests that relaxation processes have to be accounted for in this context. To take structural relaxation of defects into account, we augmented the established kMC framework for transport simulations by coupling our kMC algorithm to a defect model derived from first principles GW calculations. The impact of structural relaxation of defects and the accompanying change of the defect energy levels on charge carrier transport is investigated and a complete transport model for TZAZT capacitors is developed. The first principles GW calculations, described in the following were carried out by Martin Häufel at the chair for Theoretical Semiconductor Physics, Walter Schottky Institute, Technical University of Munich[142].

5.3.2 GW Calculations

As said, many elaborate studies of defect levels in the high-κ materials are available. To save computation time, usually defect levels are calculated in the local density approximation (LDA), i.e. the results suffer from the systematically underestimated band gap. While precise rules for rescaling the band gap are known, due to the rescaling process, the energy of the defect levels within the band gap is afflicted with some uncertainty. For an unambiguous identification of structural defects which might constitute the dominant leakage current channels a more accurate analysis is needed. Alternative calculation schemes exist, which are on the one hand much more costly, but on the other hand deliver very precise results for defect level energies. 118 5. Leakage Current Simulations

One of these methods is the GW approximation [158, 159], in which dynamical screening effects are taken into account, leading to better results for the band gap and excited states of many semiconducting and insulating materials [159, 160]. In the GW approach crystal electrons are viewed to be surrounded by a polarization cloud. Thus, a screened interaction with the other electrons takes the place of bare Coulomb interaction. The excitations of the many-particle system can therefore be described as those of weakly interacting quasiparticles. The GW approximation can formally be written as an expansion of the self-energy Σ, defined as the energy difference between the quasiparticle and the bare electron, in terms of the screened interaction W ,

Σ GW, (5.29) ≈ where G denotes the single-particle Green’s function. In Feynman diagram repre- sentation this reads

(5.30)

The equation of motion is given by the quasiparticle equation

3 hΨi(r)+ d r′ Σ(r,r′;Ei)Ψi(r′)= EiΨi(r), (5.31) Z where h is the one-particle Hamiltonian in the Hartree approximation, and Ψi and

Ei are the quasiparticle wavefunction and energy, respectively. In this deﬁnition the self-energy contains all the exchange and correlation contributions beyond the Hartree approximation. Note that, for simplicity, spin degrees of freedom are neglected in this presentation, whereas the actual calculations are carried out in a spin-polarized way. Comparing the quasiparticle equation to the Kohn-Sham equations

h + vxc(r) ϕi(r)= εiϕi(r), (5.32) with ϕi and εi denoting the Kohn-Sham orbitals and energies, respectively, one can see that the self-energy term takes the place of the exchange-correlation potential vxc. As a starting point the Kohn-Sham energies and wavefunction were determined self- consistently within the LDA. Then the quasiparticle eﬀects enter the calculations as energy corrections, obtained using ﬁrst order perturbation theory to evaluate the operator (Σ v ) at the Kohn-Sham energies and for the Kohn-Sham wavefunctions. − xc 5.3. Structural Relaxation of Defects 119

All defect energy level calculations were carried out with the ABINIT [161, 162, 163, 164] software package. The density functional theory simulations were performed for the crystal at zero temperature. Vanderbilt ultrasoft pseudopotentials (USPP) [165] were used, as, according to our calculations, the band gap energy differs only by 0.1 eV to the value obtained with norm-conserving pseudopotentials. For the generation of the pseudopotentials the following reference configurations were used. For oxygen all six electrons in the n = 2 shell (2s22p4) were included. For zirconium, all electrons in the n =4 shell plus the 5s electrons (4s24p64d25s2 = 12 electrons) were treated as valence electrons. Exchange and correlation were described using the Perdew-Zunger-Ceperley-Alder potential [166, 167]. Quasiparticle corrections for the highest occupied valence band state and the lowest lying conduction band state were calculated using the GW approximation. The one-shot G0W0 approximation was applied using the standard Plasmon-Pole model of Godby and Needs[168] at two frequencies: the static limit (0 eV) and another frequency, imaginary, of the order of the plasmon frequency (the peak in the electron energy loss spectrum). We used the volume plasmon pole energy of 13.8 eV given in ref. [169]. First, the energy gap of t-ZrO was determined using the same 2 2 2 supercell 2 × × consisting of 96 atoms, that was used for the defect calculations. Note that the cell (as well as all defect supercells used in the calculations below) was relaxed using a different ultrasoft pseudopotential (the volume per unit formula differs by 1%). The ∼ relative error caused by this is assumed to cancel in the final results, as we are only interested in the positions of the defect states within the gap. The valence pseudo- wavefunctions are expanded in a projector-augmented-wave (PAW) [170] basis set up to a kinetic energy of 15 ha. In the screening calculation and in the calculation of the self-energy matrix elements the summation was performed over 730 and 700 electronic bands, respectively. That is, in order to obtain converged GW results, about 350 unoccupied buffer bands were included in the summation. The calculation was carried out at the Γ-point only. All parameters were converged within 0.1 eV. As a result a LDA energy gap of 3.52 eV and a GW gap of 5.52 eV were obtained, which, in the following, are used as a reference for the defect levels.

5.3.3 Defect Model

Oxygen vacancies in t-ZrO2, a common defect in this material[171] and the most probable candidate for causing the leakage currents in this dielectric, were studied. Calculations were carried out using a 2 2 2 supercell consisting of 95 atoms, as × × shown in Fig. 5.18 (top). 120 5. Leakage Current Simulations

Figure 5.18: (top) 95-atom supercell of t-ZrO2 used to perform the GW calculations. O atoms are shown in light red, Zr atoms in dark blue. An oxygen vacancy was formed by removing one oxygen atom (green). Subsequently the atomic structure was relaxed. (bottom) Zoom on the direct vicinity of the oxygen vacancy (missing 0 O atom colored green). The relaxed structure for the neutral vacancy VO is pic- tured. Semitransparent, dark blue balls show the positions of the nearest-neighbor Zr atoms for the positively charged vacancy. Upon removal of an electron, the Zr atoms are pushed away from the vacancy by the localized positive charge. For better visibility the displacement is upscaled by a factor of 5. 5.3. Structural Relaxation of Defects 121

charge state This thesis Theor.

LDA GW t-ZrO2[172] m-ZrO2[173] c-ZrO2[174]

-1 2.58 (0.94) 2.59 (2.94) 2.40 (4.20)

-1 3.49 (0.03) 5.00 (0.53) 4.40 (2.20)

neutral 0.96 (2.56) 3.06 (2.44) 3.35 (3.55) 2.20 3.45 (2.1)

+1 (α) 0.68 (2.84) 3.08 (2.43) 4.20 (2.70)

+1 (β) 3.41 (0.11) 4.41 (1.09) 6.10 (0.80)

+2 3.48 (0.04) 5.25 (0.27) 7.00 (-0.10)

Table 5.2: Defect energy levels of the neutral and of charged oxygen vacancies in t-ZrO2 (in eV) in comparison to published theoretical results. In braces we give the energy diﬀerence to the CB minimum.

In Tab. 5.2 the results of the defect level calculations are presented along with published theoretical values. According to defect formation energies, calculated in ref. [149] for the chemically similar HfO2, the prevailing charge states for oxygen vacancies (VO) are neutral 0 + ++ (VO), positive (VO ), and twofold positive (VO ). The position of the corresponding defect energy levels of the relaxed oxygen vacancy within the band gap of t-ZrO2 is shown in Fig. 5.19. 0 For VO, two degenerate, ﬁlled defect levels are found 2.44 eV below the ZrO2 CB, corresponding to two electrons that are strongly localized at the vacancy site. Upon + removal of an electron (VO ), the now unoccupied spin down state (β) moves up to 1.09 eV, while the occupied spin up state (α) remains energetically unaﬀected. For the neutral oxygen vacancy, screened Coulomb repulsion between the two localized electrons, as described by equation (5.33) below, leads to a potential energy contribution of the order of 0.3-0.5 eV. Thus, upon removal of one electron one would expect the defect energy level to shift downwards by this amount. However, upon positive charging of the defect a large structural relaxation takes place. Figure 5.18 (bottom) shows a close up view of the relaxed atomic structure in the vicinity of a 0 neutral oxygen vacancy VO. If one electron is removed, the remaining positive charge localized at the vacancy site forces the positive Zr ions to relax 0.11 Å away from the vacancy (shown as semitransparent blue balls). Thus, the attractive potential energy between the remaining electron and the Zr ions decreases, causing the (α) 122 5. Leakage Current Simulations

Figure 5.19: Energy level diagram showing the positions of the oxygen vacancy states within the t-ZrO2 band gap. All defect energy levels were obtained using + the GW approximation; all values are given in electron volts (eV). For VO the non- degenerate spin up and down states are denoted (α) and (β), respectively. Open (closed) circles denote unoccupied (occupied) states. state to more or less remain energetically unchanged and the unoccupied (β) state to shift 1.35 eV upwards. A similar value of 0.10 Å for the relaxation distance was found in refs. [123] and [172]. However, in ref. [123] the exchange correlation was described by the generalized gradient approximation leading to a defect level shift of only 0.6 eV, while embedded cluster calculations in ref. [172], overestimating the ++ ZrO2 band gap, found a shift of 2.75 eV. By removing another electron (VO ), the two unoccupied states move up close to the CB.

5.3.4 Kinetic Monte Carlo Framework

The defect model, derived from the GW calculations (Fig. 5.19), was incorporated into our kMC algorithm to study the eﬀect of the structural relaxation on the electronic transport. To get a correct picture of the electron transport and, especially, the statistical interplay between diﬀerent transport mechanisms, structural relaxation was added to the list of state-to-state transitions, which have to be taken into account concurrently in the kMC algorithm. The full set of transitions is sketched in Fig. 5.20; for comparison to the earlier model framework the reader should refer to chapter 3 5.3. Structural Relaxation of Defects 123

and Fig. 3.1 therein. These transitions encompass both, electrode-limited mechanisms, such as (i) direct tunneling and (ii) Schottky emission, as well as bulk-limited mechanisms, such as (iii) elastic/inelastic tunneling of electrons into and (iv) out of defects, (v) ﬁeld-enhanced thermal emission of electrons from defects (PF emission), and (vi) tunneling between defects. In addition, we integrated self-consistently (vii) structural relaxation of defects upon a change of their charge state, causing a shift of the defect energy level.

Figure 5.20: Schematic band diagram of a MIM structure with applied voltage U. Transport mechanisms, implemented in the kMC simulator: (i) direct/FN tunneling, (ii) SE, (iii) elastic and inelastic tunneling (p = number of phonons with energy hω¯ ) into and (iv) out of defects, (v) PF emission, (vi) defect-defect tunneling, and (vii) structural relaxation of a defect upon a change of its charge state (waved arrows).

The physical models for the diﬀerent transport mechanisms are an essential input of our kMC simulations. They have been introduced in chapter 3; a concise overview 124 5. Leakage Current Simulations

layer thicknesses (nm) layer permittivities

bott. top ZAZ bott. top ZAZ

samples ZrO2 Al2O3 ZrO2 stack ZrO2 Al2O3 ZrO2 stack

T7ZAZT 3.4 0.5 3.4 7.3 (7.3) 38 9 38 31.1 (31.4)

T9ZAZT 4.5 0.4 4.5 9.4 (9.4) 38 9 38 33.4 (35.2)

Table 5.3: Model parameters for the individual layers of the ZAZ dielectric used in the kMC simulations. For comparison, measured values, only available for the whole dielectric stack, are given in braces. can be found in ref. [111]. In order to account for the structural relaxation of the defects due to a change of their charge state the defect model is coupled to the kMC simulations by setting the defect energies for diﬀerent charge states to the GW values. It turns out that the residence time of electrons in the defects is orders of magnitude larger than the time needed for structural relaxation (in the picosecond regime[175]) around the defect after a change of the charge state. Thus, we neglected the relaxation time in our simulations, i.e. for transitions of type (vii) we assumed a rate R . relax → ∞ The electrostatic potential in the vicinity of charged defects was modeled as a screened Coulomb potential. For defect i with electrical charge Ze, located at position ~ri, it can be written as

Ze V (~r)= . (5.33) 4πǫ ǫ ~r ~r opt 0 | − i| Here, Z is an integer number, giving the charge state of the defect, and ǫopt is the optical permittivity of the surrounding dielectric material, taken as 5.6 in t-ZrO2[31]. In the above-described first principles kMC modeling framework charge transport in planar TZAZT capacitors, fabricated as described in chapter 2.3, was simulated. Total film thicknesses of the ZAZ stacks, determined via spectroscopic ellipsometry, and their effective permittivities, extracted from capacitance measurements, given in Tab. 5.3, are in excellent agreement with our model assumptions.

5.3.5 First-principles kMC Leakage Current Simulations

Figure 5.21 shows the leakage current data of the two capacitor structures for temperatures between 25 C and 150 C as a function of the mean electric ﬁeld F = U/d, ◦ ◦ h i 5.3. Structural Relaxation of Defects 125

with the applied bias U and the total dielectric stack thickness d. In order to study the effect of structural relaxation on electron transport in these thin film capacitors by means of kMC some assumptions had to be made, whose justification will be commented on later. For the sake of simplicity it was assumed that one single type of defect, homogeneously distributed within the dielectric, causes the dominant leakage current contributions. Best estimate physical parameters, entering the kMC transition rates, were taken to characterize the dielectric stack and kept fixed during the simulations. They are listed in Tab. 5.1. Assumed thicknesses of the individual ZrO2 and Al2O3 layers, given in Tab. 5.3, were chosen so as to reproduce the measured total film thicknesses and permittivities of the ZAZ laminates. The only fitting parameters are the CBO

EB at the TiN/ZrO2 interface and the defect density. In this model framework we found a single set of parameters that accurately re- produces both the temperature and the thickness scaling of the leakage current, as can be seen in Fig. 5.21 (a) and (b). The closest fit to the experimental data was achieved for a barrier height of 1.9 eV, a value which compares well to the earlier simulation result of 1.74 eV for TZT capacitors, studied in 5.1 and ref. [111], and to theoretical estimates given on page 89. For the best possible fit of the kMC simulations to the experimental data the defect level energies for the different charge states were slightly changed with respect to + the GW values (given in braces) to Ed,(β) = 1.1 eV (1.09 eV) for the unoccupied (β) 0 state of the positively charged vacancy and Ed = 2.3 eV (2.44 eV) for the degenerate states of the neutral defect with respect to the CB. The energies of the unoccupied states of the twofold positively charged defect and the occupied (α) state of the positively charged defect were left unchanged. These minor changes lie well within the error margin of the GW results. This suggests that, indeed, oxygen vacancies are responsible for the dominant leakage current contribution in TZAZT capacitors.

5.3.6 Trap-assisted Tunneling Across Oxygen Vacancies

From the kMC simulations the following picture of electron transport across the dielectric, as shown in Fig. 5.22, emerges. An electron is injected from the TiN electrode into the unoccupied electron state of a positive vacancy 1.1 eV below the ZrO CB (Fig. 5.22 (a)) which subsequently relaxes by E = E0 E+ = 1.2 eV 2 relax d − d,(β) (Fig. 5.22 (b)). The trapped electron then moves by phonon-assisted tunneling, cf. 3.3.2, into a neighboring positively charged vacancy (Fig. 5.22 (c)) which again relaxes. In this way, percolating along a chain of oxygen vacancies, the electron crosses 126 5. Leakage Current Simulations

- 2

10 ) 2

- 3 data sim

25°C

50°C - 4

100°C

- 5 150°C

- 6

- 7

- 8

- 9

relaxation currents 10

- 10

(a) leakage current density j (A/cm j density current leakage

- 11

0 1 2 3 4

mean electric field (MV/cm)

- 2

10 ) 2

- 3 data sim

25°C

50°C - 4

90°C

- 5 120°C

- 6

- 7

- 8

- 9

relaxation currents

- 10

(b) leakage current density j (A/cm j density current leakage

- 11

0 1 2 3 4

mean electric field (MV/cm)

Figure 5.21: Leakage current data (symbols) of TZAZT capacitors with dielectrics of (a) 7 nm and (b) 9 nm thickness compared to kMC simulation results (lines). In the low-ﬁeld region the measured current is dominated by transient relaxation currents, while the kMC results describe the steady state. Thus, no comparison is possible[110]. Dielectric breakdown was observed for high ﬁelds. 5.3. Structural Relaxation of Defects 127

Figure 5.22: The different steps of an electron crossing the dielectric film are sketched. (a) First the electron is injected into a positiv oxygen vacancy which (b) subsequently undergoes structural relaxation causing the defect energy level to shift downwards. (c) In a next step the electron moves via an inelastic, multiphonon- assisted transition, cf. 3.3.2, to a neighbouring positively charged oxygen vacancy. (d) By repeating the last two steps, the electron can percolate along a chain of defect states until it is finally emitted to the counter-electrode. Here, p is the number of phonons with energy hω¯ . the dielectric (Fig. 5.22 (d)). The number of percolation sites varies between one and three, depending on film thickness and applied bias (The thinner the dielectric and the higher the applied bias, the fewer sites are needed.). Earlier models which neglected structural relaxation effects came to the conclusion that transport is dominated by PF conduction[90, 129, 110] via a defect level located, e.g., 1.15 eV[110] below the CB. The analysis of the ab initio results shows that this corresponds almost exactly to the relaxation energy Erelax of the oxygen vacancy. Such coincidence explains why in both models a similar temperature scaling of the leakage current is obtained. Clearly, only the model including structural relaxation 128 5. Leakage Current Simulations

effects is consistent with first principles defect level calculations, as shown here and in refs. [123] and [172]. Best quantitative matching of the simulations with the experimental data was achieved for a defect concentration of 1 1020 cm3. Such a high concentration might be due to × the presence of the high work function TiN electrodes, which are known to increase the concentration of oxygen vacancies in such high-κ thin films[41]. Via soft x-ray photoelectron spectroscopy defect densities in the range of 2 1019 to 1 1020 cm3 × × were estimated[176]. The investigation of (ZrO2)0.8(Al2O3)0.2 films via conductive atomic force microscopy[177] suggests that transport predominantly occurs along grain boundaries in the polycrystalline ZrO2. Since oxygen vacancies are known to cluster at grain boundaries[176, 178], such experimental evidence strongly suggests that they are responsible for opening the loss paths.

5.3.7 TiN/ZrO2/Al2O3/ZrO2/TiN Capacitors

In Fig. 5.23 (a) leakage current data of capacitors with pure ZrO2 dielectrics, discussed in 5.1, and of the present TZAZT capacitors is plotted as a function of F . h i As can be seen, introduction of the amorphous Al2O3 interlayer leads to a reduction of the leakage currents. So far, the Al2O3 was thought to act as a barrier for electron ﬂow due to its large band gap of around 8.9 eV. Additionally, it was suggested that the interlayer interrupts the grain boundaries of the polycrystalline material. While these eﬀects may contribute to the reduction of the leakage current in ZAZ

films compared to pure ZrO2 films, see Fig. 5.23 (a), our simulations suggest that their influence is negligible. Instead, the electron transport turns out to be mainly ZAZ controlled by F , the electric field in the ZrO2 layers of the ZAZ stack. Upon introduction of the low-κ Al2O3 interlayer (κ = 9), the electric field in the ZrO2 layers is reduced with respect to F ZrO2 , the value it would have in a dielectric film of pure ZrO . Assuming that the electric displacement D = κ F is constant throughout the 2 · dielectric film, F ZAZ can be calculated from

U F ZAZ = . (5.34) κ(ZrO2) d(ZrO2)+ κ(Al2O3) d(Al2O3)

Here, U is the applied voltage, κ(Al2O3) and κ(ZrO2) are the permittivities in

Al2O3 and ZrO2, respectively, and d(Al2O3) and d(ZrO2) are the total thicknesses of the materials. In Fig. 5.23 (b) we replot the data from Fig. 5.23 (a) as a function of F ZAZ for the ZrO2 ZAZ films and F for the ZrO2 films, so as to eliminate the field reduction effect 5.3. Structural Relaxation of Defects 129

of the Al2O3 interlayer. Apparently, the major effect of the interlayer is the reduction of the electric field in the ZrO2 layers. While this might also be advantageous in terms of reliability[89], it comes with the cost of an increased EOT. The fact that, for an identical F ZAZ, the current does not scale with the film thickness, at least in the considered thickness range, is consistent with our transport model. As electrons pass the dielectric in a multistep process, the rate for this to happen is governed by the slowest single step, i.e. phonon-assisted tunneling of an electron between two defects. Thus, relevant for the scaling of the current is not the total film thickness but the defect-defect distance, i.e. the defect concentration.

5.3.8 Summary

In summary, multistep, inelastic, phonon-assisted tunneling across oxygen vacancies has been identified as the dominant conduction mechanism in TZAZT capacitors. A closed model, taking structural relaxation of the involved defects upon a change of their charge state into account, was presented. Consideration of defect relaxation leads to a qualitative modification of our picture of electron transport. Instead of PF conduction, TAT across oxygen vacancies is identified as the dominant transport mechanism. 130 5. Leakage Current Simulations

- 1

10 ) 2

- 2 9 nm ZAZ

1 10

9 nm ZrO

- 3

10 7 nm ZAZ

7 nm ZrO - 4

- 5

- 6

- 7

- 8

- 9

(a) leakage current density j (A/cm j density current leakage

- 10

0 1 2 3 4

(MV/cm)

- 1

10 ) 2

9 nm ZAZ - 2

9 nm ZrO

- 3

10 7 nm ZAZ

7 nm ZrO

- 4

- 5

- 6

- 7

- 8

- 9

(b) leakage current density j (A/cm j density current leakage

- 10

0 1 2 3 4

ZAZ ZrO F , F (MV/cm)

Figure 5.23: Experimental leakage current data of pure ZrO2 dielectrics and of

ZAZ stacks for a temperature of 120◦C is compared. In (a) the current density j is plotted vs. F , the mean electric field in the dielectric and in (b) vs. the field in h i ZAZ ZrO2 the ZrO2 layers, i.e. F for the ZAZ films and F for the pure ZrO2 films, thus accounting for the field reduction in the ZAZ stack caused by the low-κ Al2O3 interlayer. Conclusions and Outlook

The common practice in analyzing leakage currents in thin-film capacitor structures is to fit experimental results with simplified compact models. Their big advantage is their low complexity and their speed, as they usually provide analytical expressions, reducing the task of fitting to the adjustment of a small set of parameters. However, often these models oversimplify the situation, i.e. conclusions, which can be drawn based on the fitting of experimental data, are limited or, worse, might even be misleading. In general, an in-depth description of electronic transport in such structures is a very ambitious task, facing alone the large number of unknown quantities in the typical systems of interest. It was the goal of the present work to go beyond the so far existing approaches and to establish a higher-level simulation framework for leakage current studies. Therefore, a modeling environment was set up, where microscopic models for every relevant physical transport process were united in a comprehensive simulation framework with a kinetic Monte Carlo algorithm at its heart. The unique properties of the kinetic Monte Carlo formalism were exploited, such as modeling at the microscopic scale, e.g. to resolve single charge carrier motions, straightforward coupling to finite elements based differential equation solvers allowing to take into account, e.g., carrier-carrier interactions, the study of phenomena spatially localized in all three dimensions (electrode roughness effects, structural relaxation of defects, defect-assisted transport), and the investigation of the statistical interplay of different microscopic processes, e.g. the concurrent simulation of multiple transport mechanisms which allows to investigate their mutual interdependence. Having such a modeling environment at hand, we carried out elaborate studies on the TiN/ZrO2 material system, which is of outmost importance for the semiconductor industry, where it is employed in the DRAM storage capacitor. Closed transport models were developed, which shed light on the microscopic transport in high-performance TiN/ZrO2/TiN capacitors. Again, making use of the kinetic Monte Carlo framework and its modular architecture, which allows to augment

131 132 5. Leakage Current Simulations

transport models with only little effort, we iteratively improved our transport models by taking more and more physical effects into account. At the end, we could provide a self-consistent transport model for the TiN/ZrO2 material system. The detailed modeling on the microscopic scale, in contrast to the compact models, enabled us to propose target-oriented strategies how to optimize the performance of the capacitor structures. Problems which may arise in the future due to further downscaling of the capacitors could be identified in advance, and viable approaches how to tackle these could be provided. In general, the flexibility of the developed modeling framework makes it easily trans- ferable allowing to study electronic transport in a variety of other, even very different, systems, as will be commented on in the outlook. 5.3. Structural Relaxation of Defects 133

While in this work TiN/ZrO2, the currently employed material system for the DRAM storage capacitor, was in the focus of our investigations, it will, in the medium term be replaced by other materials, such as StTiO3 and TiO2, which feature even higher permittivities. These, unfortunately are weak insulators; their band gaps are around 3.5 eV. Currently ongoing investigations, thus, not unsurpris- ingly, measure very high leakage currents for capacitor structures employing these dielectrics. The conduction band offsets provided by these materials are roughly 0.5 eV too low to surpress intrinsic leakage currents due to, mostly, Schottky emission. Strategies to increase these offsets via acceptor doping are on pursued. How- ever, these systems are far from being ripe for introduction into the DRAM production process. The simulation framework proposed in the present work will, in the future, be used to the study of these systems in order to provide theoretical support for the experimental efforts by enraveling the dominant microscopic transport mechanisms in these systems. But, besides DRAM applications, the developed simulation tool, due to the variety of the implemented models and the versatility of the kinetic Monte Carlo algorithm and the program architecture itself, is well-suited to investigate transport in other metal-insulator-metal structures. For example, tunneling diodes for the rectification of high-frequency ac currents, which consist of an aluminum and a gold electrode separated by a ultrathin ( 3.6 nm) Al O film, were investigated[179, 180]. Cur- ∼ 2 3 rents were found to be dominated by direct and Fowler-Nordheim tunneling and electrode roughness, for such ultrathin dielectric films, was suggested to play an important role[181, 182]. Further optimization of the fabrication process of these devices is on its way and will be supported by transport simulations. Another field of rising interest, for which kinetic Monte Carlo transport simulations seem to be a very promising tool for theoretical studies, is the fabrication and optimization of organic solar cells, especially those based on the bulk-heterojunction concept[183]. The simulation framework established in the present work has been adapted in order to simulate the complex microscopic processes involved in the photocurrent generation in these devices, namely exciton generation, difussion, separation, decay and charge carrier injection, transport, recombination, extraction etc. The current functionality of the simulator corresponds to the state-of-the-art in this field, subsumed in refs. [184, 185, 186, 187, 188]. Thus, the foundations are laid to study experimental data of such devices in the future. 134 5. Leakage Current Simulations Bibliography

[1] R. Chau, J. Brask, S. Datta, G. Dewey, M. Doczy, B. Doyle, J. Kavalieros, B. Jin, M. Metz, A. Majumdar, and M. Radosavljevic, “Application of high- κ gate dielectrics and metal gate electrodes to enable silicon and non-silicon logic nanotechnology,” Microelectronic Engineering, vol. 80, pp. 1–6, 2005.

[2] S. Monaghan, K. Cherkaoui, E. O’Connor, V. Djara, P. Hurley, L. Oberbeck,

E. Tois, L. Wilde, and S. Teichert, “TiN/ZrO2/Ti/Al Metal-Insulator-Metal

Capacitors With Subnanometer CET Using ALD-Deposited ZrO2 for DRAM Applications,” IEEE Electron Device Letters, vol. 30, no. 3, pp. 219 –221, 2009.

[3] H. Cho, Y. Kim, D. Park, E. Lee, C. Park, J. Jang, K. Lee, H. Kim, Y. Ki,

and I. Han, “New TIT capacitor with ZrO2/Al2O3/ZrO2 dielectrics for 60 nm and below DRAMs,” Solid-State Electronics, vol. 51, no. 11-12, pp. 1529–1533, 2007.

[4] V. Narayanan, S. Guha, M. Copel, N. Bojarczuk, P. Flaitz, and M. Gribelyuk, “Interfacial oxide formation and oxygen diﬀusion in rare earth oxide–silicon epitaxial heterostructures,” Applied Physics Letters, vol. 81, p. 4183, 2002.

[5] M. Fischetti, D. Neumayer, and E. Cartier, “Eﬀective electron mobility in Si inversion layers in metal-oxide-semiconductor systems with a high-κ insulator: The role of remote phonon scattering,” Journal of Applied Physics, vol. 90, p. 4587, 2001.

[6] K. Shiraishi, K. Yamada, K. Torii, Y. Akasaka, K. Nakajima, M. Kohno, T. Chikyo, H. Kitajima, and T. Arikado, “Physics in fermi level pinning at the polysi/hf-based high-k oxide interface,” in Symposium on VLSI Technology, 2004. Digest of Technical Papers, 2004, pp. 108 – 109.

[7] K. Frohlich, J. Aarik, M. Tapajna, A. Rosova, A. Aidla, E. Dobrocka, and

K. Huskova, “Epitaxial growth of high-κ TiO2 rutile ﬁlms on RuO2 electrodes,”

135 136 BIBLIOGRAPHY

Journal of Vacuum Science & Technology B: Microelectronics and Nanometer Structures, vol. 27, no. 1, pp. 266–270, 2009.

[8] G. Aichmayr, A. Avellán, G. S. Duesberg, F. Kreupl, S. Kudelka, and M. Liebau, “Carbon/high-k Trench Capacitor for the 40 nm DRAM Gen- eration,” in VLSI Technology, 2007 IEEE Symposium on. IEEE, 2007, pp. 186–187.

[9] M. Ritala, K. Kukli, A. Rahtu, P. I. Räisänen, M. Leskelä, T. Sajavaara, and J. Keinonen, “Atomic layer deposition of oxide thin ﬁlms with metal alkoxides as oxygen sources,” Science, vol. 288, no. 5464, p. 319, 2000.

[10] Leskelä, M. and Ritala, M., “Atomic layer deposition: from precursors to thin ﬁlm structures,” Thin Solid Films, vol. 409, no. 1, pp. 138–146, 2002.

[11] Ritala, M. and Leskelä M., Handbook of Thin Film Materials, H. S. Nalwa, Ed. Academic Press, 2001, vol. 1.

[12] J.-S. Min, Y.-W. Son, W.-G. Kang, S.-S. Chun, and S.-W. Kang, “Atomic Layer Deposition of TiN Films by Alternate Supply of Tetrakis(ethylmethylamino)-Titanium and Ammonia,” Japanese Journal of Applied Physics, vol. 37, no. Part 1, No. 9A, pp. 4999–5004, 1998.

[13] K. Tse and J. Robertson, “Defect passivation in HfO2 gate oxide by ﬂuorine,” Applied Physics Letters, vol. 89, p. 142914, 2006.

[14] G. D. Wilk and D. A. Muller, “Correlation of annealing eﬀects on local elec-

tronic structure and macroscopic electrical properties for HfO2 deposited by atomic layer deposition,” Applied Physics Letters, vol. 83, p. 3984, 2003.

[15] J. Robertson, “High dielectric constant gate oxides for metal oxide Si transistors,” Reports on Progress in Physics, vol. 69, p. 327, 2006.

[16] M.-T. Wang, T.-H. Wang, and J. Y. min Lee, “Electrical Conduction Mech-

anism in Metal-ZrO2-Silicon Capacitor Structures,” Journal of The Electro- chemical Society, vol. 152, no. 3, pp. G182–G185, 2005.

[17] M.-T. Wang, T.-H. Wang, B. Y.-Y. Cheng, and J. Y. min Lee, “Tempera-

ture Dependence of the Tunnel-Emission Conduction Current in Metal-ZrO2- Silicon Capacitor Structures,” Journal of The Electrochemical Society, vol. 153, no. 1, pp. F8–F11, 2006. BIBLIOGRAPHY 137

[18] Z. Xu, M. Houssa, S. De Gendt, and M. Heyns, “Polarity eﬀect on the tem-

perature dependence of leakage current through HfO2/SiO2 gate dielectric stacks,” Applied Physics Letters, vol. 80, no. 11, pp. 1975–1977, 2009.

[19] W. Chim, T. Ng, B. Koh, W. Choi, J. Zheng, C. Tung, and A. Du, “Inter- facial and bulk properties of zirconium dioxide as a gate dielectric in metal– insulator–semiconductor structures and current transport mechanisms,” Jour- nal of Applied Physics, vol. 93, p. 4788, 2003.

[20] G. W. Dietz, M. Schumacher, R. Waser, S. K. Streiﬀer, C. Basceri, and A. I.

Kingon, “Leakage currents in Ba0.7Sr0.3TiO3 thin ﬁlms for ultrahigh-density dynamic random access memories,” Journal of Applied Physics, vol. 82, no. 5, pp. 2359–2364, 1997.

[21] S. Zafar, R. E. Jones, B. Jiang, B. White, V. Kaushik, and S. Gillespie, “The electronic conduction mechanism in barium strontium titanate thin ﬁlms,” Applied Physics Letters, vol. 73, no. 24, pp. 3533–3535, 1998.

[22] J. Frenkel, “On pre-breakdown phenomena in insulators and electronic semiconductors,” Physical Review., vol. 54, no. 8, pp. 647–648, 1938.

[23] L. Nordheim, “Zur Theorie der thermischen Emission und der Reﬂexion von Elektronen an Metallen,” Zeitschrift für Physik A Hadrons and Nuclei, vol. 46, no. 11, pp. 833–855, 1928.

[24] A. Sommerfeld, “Zur Elektronentheorie der Metalle auf Grund der Fermischen Statistik,” Zeitschrift für Physik A Hadrons and Nuclei, vol. 47, no. 1, pp. 1– 32, 1928.

[25] H. Schroeder, S. Schmitz, and P. Meuﬀels, “Leakage currents in high- permittivity thin ﬁlms,” Applied Physics Letters, vol. 82, p. 781, 2003.

[26] J. D. Baniecki, R. B. Laibowitz, T. M. Shaw, C. Parks, J. Lian, H. Xu, and

Q. Y. Ma, “Hydrogen induced tunnel emission in Pt/(BaxSr1 xTi1+yO3+z/Pt − thin ﬁlm capacitors,” Journal of Applied Physics, vol. 89, no. 5, pp. 2873–2885, 2001.

[27] J. D. Baniecki, T. Shioga, K. Kurihara, and N. Kamehara, “A study of current

transport in BaxSr1 xTi1+yO3+z thin-ﬁlm capacitors containing a voltage- − dependent interface state charge distribution,” Journal of Applied Physics, vol. 97, no. 11, p. 114101, 2005. 138 BIBLIOGRAPHY

[28] D. S. Jeong and C. S. Hwang, “Tunneling-assisted Poole-Frenkel conduction

mechanism in HfO2 thin ﬁlms,” Journal of Applied Physics, vol. 98, p. 113701, 2005.

[29] Y. Okuyama, S. Kamohara, Y. Manabe, K. Okuyama, Y. Kubota, T. Kobayashi, and K. Kimura, “Monte Carlo simulation of stress-induced leakage current by hopping conduction via multi-traps in oxide,” in International Electron Devices Meeting, 1998. IEDM’98 Technical Digest. IEEE, 1998, pp. 905–908.

[30] K. H. Kuesters, M. F. Beug, U. Schroeder, N. Nagel, U. Bewersdorff, G. Dallmann, S. Jakschik, R. Knoefler, S. Kudelka, C. Ludwig, D. Manger, W. Mueller, and A. Tilke, “New materials in memory development sub 50 nm: Trends in flash and dram,” Advanced Engineering Materials, vol. 11, no. 4, pp. 241–248, 2009.

[31] G.-M. Rignanese, F. Detraux, X. Gonze, and A. Pasquarello, “First-principles study of dynamical and dielectric properties of tetragonal zirconia,” Physical Review B, vol. 64, no. 13, p. 134301, Aug 2001.

[32] J. Robertson, “Band oﬀsets of wide-band-gap oxides and implications for future electronic devices,” Journal of Vacuum Science & Technology B: Micro- electronics and Nanometer Structures, vol. 18, p. 1785, 2000.

[33] S. K. Kim, S. W. Lee, J. H. Han, B. Lee, S. Han, and C. S. Hwang, “Capacitors with an Equivalent Oxide Thickness of < 0.5 nm for Nanoscale Electronic Semiconductor Memory,” Advanced Functional Materials, vol. 20, no. 18, pp. 2989–3003, 2010.

[34] Y. Jin, G. Li, Y. Zhang, Y. Zhang, and L. Zhang, “Photoluminescence of

anatase TiO2 thin ﬁlms achieved by the addition of ZnFe2O4,” Journal of Physics: Condensed Matter, vol. 13, p. L913, 2001.

[35] D. M. Smyth, The Defect Chemistry of Metal Oxides. New York: Oxford University Press, 2000.

[36] S. K. Kim, G. J. Choi, S. Y. Lee, M. Seo, S. W. Lee, J. H. Han, H. S. Ahn,

S. Han, and C. S. Hwang, “Al-doped TiO2 ﬁlms with ultralow leakage currents for next generation DRAM capacitors,” Advanced Materials, vol. 20, no. 8, pp. 1429–1435, 2008. BIBLIOGRAPHY 139

[37] S. K. Kim, W.-D. Kim, K.-M. Kim, C. S. Hwang, and J. Jeong, “High dielectric

constant TiO2 thin ﬁlms on a Ru electrode grown at 250◦C by atomic-layer deposition,” Applied Physics Letters, vol. 85, no. 18, pp. 4112–4114, 2004.

[38] C. Black and J. Welser, “Electric-ﬁeld penetration into metals: Consequences for high-dielectric-constant capacitors,” IEEE Transactions on Electron De- vices, vol. 46, no. 4, pp. 776–780, 1999.

[39] J. Robertson, “High dielectric constant oxides,” The European Physical Jour- nal Applied Physics, vol. 28, no. 3, pp. 265–291, 2004.

[40] M. Y. Li, B. S. Tsai, P. C. Jiang, H. C. Wu, Y. H. Wu, and Y. J. Lin,

“Structure and property changes of ZrO2/Al2O3/ZrO2 laminate induced by

low-temperature NH3 annealing applicable to metal-insulator-metal capacitor,” Thin Solid Films, vol. 518, no. 18, pp. 5272–5277, 2010.

[41] D. Liu and J. Robertson, “Passivation of oxygen vacancy states and suppres-

sion of Fermi pinning in HfO2 by La addition,” Applied Physics Letters, vol. 94, no. 4, p. 042904, 2009.

[42] T. Schimizu and M. Koyama, “Control of electronic properties of HfO2 with ﬂuorine doping from ﬁrst-principles,” Applied Surface Science, vol. 254, no. 19, pp. 6109–6111, 2008.

[43] W. Weinreich, “Herstellung und Charakterisierung ultradünner ZrO2 basierter Schichten als Isolatoren in Metall - Isolator - Metall Kondensatoren,” Ph.D. dissertation, Der Technischen Fakultät der Friedrich-Alexander-Universität, Erlangen-Nürnberg, 2011.

[44] M. Ritala and M. Leskelae, Handbook of Thin Film Materials: Volume 1: Depostion and Processing of Thin Films. Academic Press, San Diego, CA, 2002.

[45] W. Weinreich, R. Reiche, M. Lemberger, G. Jegert, J. Müller, L. Wilde, S. Te- ichert, J. Heitmann, E. Erben, L. Oberbeck et al., “Impact of interface variations on JV and CV polarity asymmetry of MIM capacitors with amorphous

and crystalline Zr(1 x)AlxO2 ﬁlms,” Microelectronic Engineering, vol. 86, no. − 7-9, pp. 1826–1829, 2009. 140 BIBLIOGRAPHY

[46] D. M. Hausmann, E. Kim, J. Becker, and R. G. Gordon, “Atomic layer deposition of hafnium and zirconium oxides using metal amide precursors,” Chem- istry of Materials, vol. 14, no. 10, pp. 4350–4358, 2002.

[47] J. Niinisto, K. Kukli, M. Kariniemi, M. Ritala, M. Leskela, N. Blasco, A. Pin- chart, C. Lachaud, N. Laaroussi, and Z. e. a. Wang, “Novel mixed alkylamido-

cyclopentadienyl precursors for ALD of ZrO2 thin ﬁlms,” Journal of Materials Chemistry, vol. 18, no. 43, pp. 5243–5247, 2008.

[48] B. Lee, K. J. Choi, A. Hande, M. J. Kim, R. M. Wallace, J. Kim, Y. Senzaki, D. Shenai, H. Li, M. Rousseau et al., “A novel thermally-stable zirconium

amidinate ALD precursor for ZrO2 thin ﬁlms,” Microelectronic Engineering, vol. 86, no. 3, pp. 272–276, 2009.

[49] R. Tsu and L. Esaki, “Tunneling in a ﬁnite superlattice,” Applied Physics Letters, vol. 22, pp. 562–564, 1973.

[50] D. Ferry and S. M. Goodnick, Transport in Nanostructures, H. Ahmed, M. Pepper, and A. Broers, Eds. Cambridge University Press, 1997.

[51] D. Bohm, Quantum Theory. Englewood Cliﬀs, New Jersey: Prentice-Hall, Inc., 1951.

[52] C. L. Hinkle, C. Fulton, R. J. Nemanich, and G. Lucovsky, “A novel approach

for determining the eﬀective tunneling mass of electrons in HfO2 and other high-K alternative gate dielectrics for advanced CMOS devices,” Microelec- tronic Engineering, vol. 72, no. 1-4, pp. 257–262, 2004.

[53] F. Sacconi, J. Jancu, M. Povolotskyi, and A. Di Carlo, “Full-Band Tunneling in High-κ Oxide MOS Structures,” IEEE Transactions on Electron Devices, vol. 54, no. 12, pp. 3168–3176, 2007.

[54] W. Franz, Handbuch der Physik, G. S. Flügge, Ed. Berlin, Ed. Springer- Verlag, 1956, vol. 17.

[55] J. Maserjian, “Tunneling in thin MOS structures,” Journal of Vacuum Science and Technology, vol. 11, no. 6, pp. 996–1003, 1974.

[56] M. Städele and K. Tuttle, B. R. Hess, “Tunneling through ultrathin SiO2 gate oxides from microscopic models,” Journal of Applied Physics, vol. 89, no. 1, pp. 348–363, 2001. BIBLIOGRAPHY 141

[57] O. W. Richardson, The Emission of Electricity from Hot Bodies. Longmans Green and Company, 1921.

[58] E. Murphy and R. H. Good Jr, “Thermionic emission, ﬁeld emission, and the transition region,” Physical Review, vol. 102, no. 6, p. 1464, 1956.

[59] S. G. Christov, “General theory of electron emission from metals,” Physica Status Solidi (b), vol. 17, no. 1, pp. 11–26, 1966.

[60] M. Lenzlinger and E. H. Snow, “Fowler-Nordheim Tunneling into Thermally

Grown SiO2,” Journal of Applied Physics, vol. 40, no. 1, pp. 278–283, 1969.

[61] J. E. Medvedeva, A. J. Freeman, C. B. Geller, and D. M. Rishel, “Screened- exchange determination of the electronic properties of monoclinic, tetragonal, and cubic zirconia,” Physical Review B, vol. 76, no. 23, p. 235115, 2007.

[62] I. Lundström and C. Svensson, “Tunneling to traps in insulators,” Journal of Applied Physics, vol. 43, no. 12, pp. 5045–5047, 1972.

[63] C. Svensson and I. Lundström, “Trap-assisted charge injection in mnos structures,” Journal of Applied Physics, vol. 44, no. 10, pp. 4657–4663, 1973.

[64] J. Bardeen, “Tunnelling from a many-particle point of view,” Physical Review Letters, vol. 6, no. 2, pp. 57–59, 1961.

[65] L. Lundkvist, I. Lundstrom, and C. Svensson, “Discharge of MNOS structures,” Solid-State Electronics, vol. 16, no. 7, pp. 811–823, 1973.

[66] X. R. Cheng, Y. C. Cheng, and B. Y. Liu, “Nitridation-enhanced conductivity

behavior and current transport mechanism in thin thermally nitrided SiO2,” Journal of Applied Physics, vol. 63, no. 3, pp. 797–802, 1988.

[67] S. Fleischer, P. T. Lai, and Y. C. Cheng, “Simpliﬁed closed-form trap-assisted tunneling model applied to nitrided oxide dielectric capacitors,” Journal of Applied Physics, vol. 72, p. 5711, 1992.

[68] A. Kiveris, A. Kudzmauskas, and P. Pipinys, “Release of electrons from traps by an electric ﬁeld with phonon participation,” physica status solidi (a), vol. 37, no. 1, pp. 321–327, 1976.

[69] F. I. Dalidchik, Eks. Theo. Fiz., vol. 74, p. 472, 1978. 142 BIBLIOGRAPHY

[70] M. Herrmann and A. Schenk, “Field and high-temperature dependence of the long term charge loss in erasable programmable read only memories: Mea- surements and modeling,” Journal of Applied Physics, vol. 77, p. 4522, 1995.

[71] G. Lucovsky, “On the photoionization of deep impurity centers in semiconductors,” Solid State Communications, vol. 3, no. 9, pp. 299 – 302, 1965.

[72] A. Schenk, Ph.D. dissertation, Humboldt University, Berlin, 1986.

[73] D. Światła and W. Bartczak, “Calculation of the Huang-Rhys factor for electron capture by a neutral impurity,” Physical Review B, vol. 43, no. 8, p. 6776, 1991.

[74] A. M. Stoneham, “Phonon coupling and photoionisation cross-sections in semiconductors,” Journal of Physics C: Solid State Physics, vol. 12, p. 891, 1979.

[75] S. E. Paje and J. Llopis, “Photoluminescence-spectra study of yttria-stabilized zirconia,” Applied Physics A: Materials Science and Processing, vol. 57, pp. 225–228, 1993, 10.1007/BF00332593.

[76] B. Ridley, “On the multiphonon capture rate in semiconductors,” Solid-State Electronics, vol. 21, no. 11-12, pp. 1319–1323, 1978.

[77] N. F. Mott, Festkörperprobleme - Advances in Solid State Physics, O. Madelung, Ed. Akademie-Verlag, Berlin, 1969, vol. 9.

[78] H. Krause, “Tunnel hopping current and trap ﬁlling in insulating layers,” Phys. Stat. Sol. (a), vol. 62, p. 565, 1979.

[79] A. Miller and E. Abrahams, “Impurity conduction at low concentrations,” Physical Review, vol. 120, no. 3, pp. 745–755, 1960.

[80] K. Huang and A. Rhys, “Theory of light absorption and non-radiative transitions in f-centres,” Proceedings of the Royal Society of London. Series A, Mathematical and Physical Sciences, vol. 204, no. 1078, pp. pp. 406–423, 1950.

[81] H. Gummel and M. Lax, “Thermal capture of electrons in silicon,” Annals of Physics, vol. 2, no. 1, pp. 28–56, 1957.

[82] R. Kubo and Y. Toyozawa, “Applicaiton of the Method of Generating Func- tion to Radiative and Non-Radiative Transitions of a Trapped Electron in a Crystal,” Progress of Theoretical Physics, vol. 13, no. 2, pp. 160–182, 1955. BIBLIOGRAPHY 143

[83] D. S. Jeong, H. B. Park, and C. S. Hwang, “Reasons for obtaining an optical dielectric constant from the Poole-Frenkel conduction behavior of atomic-

layer-deposited HfO2 ﬁlms,” Applied Physics Letters, vol. 86, p. 072903, 2005.

[84] J. W. McPherson and H. C. Mogul, “Underlying physics of the thermochemical

E model in describing low-ﬁeld time-dependent dielectric breakdown in SiO2 thin ﬁlms,” Journal of Applied Physics, vol. 84, p. 1513, 1998.

[85] J. L. Hartke, “The three-dimensional Poole-Frenkel eﬀect,” Journal of Applied Physics, vol. 39, p. 4871, 1968.

[86] S. Chakraborty, M. K. Bera, G. K. Dalapati, D. Paramanik, S. Varma, P. K. Bose, S. Bhattacharya, and C. K. Maiti, “Leakage current characteristics and

the energy band diagram of Al/ZrO2/Si0.3Ge0.7 hetero-MIS structures,” Semi- conductor science and technology, vol. 21, p. 467, 2006.

[87] P. V. Aleskandrova, V. K. Gueorguiev, T. E. Ivanov, and J. B. Koprinarova,

“Poole-Frenkel conduction in Al/ZrO2/SiO2/Si structures,” The European Physical Journal B-Condensed Matter and Complex Systems, vol. 52, no. 4, pp. 453–457, 2006.

[88] D. Zhou, U. Schröder, G. Jegert, M. Kerber, S. Uppal, R. Agaiby, M. Reinicke, J. Heitmann, and L. Oberbeck, “Time dependent dielectric breakdown of

amorphous ZrAlxOy high-k dielectric used in dynamic random access memory metal-insulator-metal capacitor,” Journal of Applied Physics, vol. 106, no. 4, p. 044104, 2009.

[89] D. Zhou, U. Schroeder, J. Xu, J. Heitmann, G. Jegert, W. Weinreich, M. Ker-

ber, S. Knebel, E. Erben, and T. Mikolajick, “Reliability of Al2O3-doped ZrO2 high-k dielectrics in three-dimensional stacked metal-insulator-metal capacitors,” Journal of Applied Physics, vol. 108, no. 12, p. 124104, 2010.

[90] Y. Seo, S. Lee, I. An, C. Song, and H. Jeong, “Conduction mechanism of

leakage current due to the traps in ZrO2 thin ﬁlm,” Semiconductor Science and Technology, vol. 24, p. 115016, 2009.

[91] J. R. Yeargan and H. L. Taylor, “The Poole-Frenkel eﬀect with compensation present,” Journal of Applied Physics, vol. 39, p. 5600, 1968. 144 BIBLIOGRAPHY

[92] M. Zebarjadi, C. Bulutay, K. Esfarjani, and A. Shakouri, “Monte carlo simulation of electron transport in degenerate and inhomogeneous semiconductors,” Applied Physics Letters, vol. 90, no. 9, p. 092111, 2007.

[93] C. Jirauschek and P. Lugli, “Monte-Carlo-based spectral gain analysis for terahertz quantum cascade lasers,” Journal of Applied Physics, vol. 105, no. 12, pp. 123 102–123 102, 2009.

[94] C. Jacoboni and P. Lugli, The Monte Carlo Method for Semiconductor Device Simulation. Springer, New York, 1989.

[95] K. A. Fichthorn and W. H. Weinberg, “Theoretical foundations of dynamical Monte Carlo simulations,” Journal of Chemical Physics, vol. 95, p. 1090, 1991.

[96] D. T. Gillespie, “A general method for numerically simulating the stochastic time evolution of coupled chemical reactions,” Journal of Computational Physics, vol. 22, pp. 403–434, 1976.

[97] N. Metropolis and S. Ulam, “The Monte Carlo method,” Journal of the Amer- ican Statistical Association, vol. 44, no. 247, pp. 335–341, 1949.

[98] S. D. Stellman and P. J. Gans, “Eﬃcient computer simulation of polymer con- formation. I. Geometric properties of the hard-sphere model,” Macromolecules, vol. 5, no. 4, pp. 516–526, 1972.

[99] N. Metropolis, A. W. Rosenbluth, M. N. Rosenbluth, A. H. Teller, and E. Teller, “Equation of state calculations by fast computing machines,” The Journal of Chemical Physics, vol. 21, no. 6, p. 1087, 1953.

[100] J. R. Beeler Jr and R. A. Johnson, “Vacancy clusters in α-iron,” Physical Review, vol. 156, no. 3, p. 677, 1967.

[101] A. F. Voter, Radiation eﬀects in solids. Springer, NATO Publishing Unit, Dordrecht, The Netherlands, 2005, ch. Introduction to the Kinetic Monte Carlo Method.

[102] B. J. Alder and T. E. Wainwright, “Phase transition for a hard sphere system,” The Journal of Chemical Physics, vol. 27, p. 1208, 1957.

[103] A. Rahman, “Correlations in the motion of atoms in liquid argon,” Physical Review, vol. 136, no. 2A, pp. 405–411, 1964. BIBLIOGRAPHY 145

[104] R. Car and M. Parrinello, “Uniﬁed approach for molecular dynamics and density-functional theory,” Physical Review Letters, vol. 55, no. 22, pp. 2471– 2474, 1985.

[105] K. Reuter, “First-principles kinetic Monte Carlo simulations for heterogeneous catalysis: Concepts, status and frontiers,” Modeling Heterogeneous Catalytic Reactions: From the Molecular Process to the Technical System. Wiley-VCH, Weinheim, 2009.

[106] R. Norris, Markov Chains. Cambridge University Press, Cambridge, UK, 1997.

[107] K. A. Fichthorn and W. H. Weinberg, “Inﬂuence of time-dependent rates of mass transfer on the kinetics of domain growth,” Physical Review Letters, vol. 68, no. 5, pp. 604–607, 1992.

[108] K. A. Fichthorn and M. Scheﬄer, “Island Nucleation in Thin-Film Epitaxy: A First-Principles Investigation,” Physical Review Letters, vol. 84, no. 23, pp. 5371–5374, 2000.

[109] A. B. Bortz, M. H. Kalos, and J. L. Lebowitz, “A new algorithm for Monte Carlo simulation of Ising spin systems,” Journal of Computational Physics, vol. 17, no. 1, pp. 10–18, 1975.

[110] G. Jegert, A. Kersch, W. Weinreich, and P. Lugli, “Monte Carlo Simulation

of Leakage Currents in TiN/ZrO2/TiN Capacitors,” IEEE Transactions on Electron Devices, vol. 58, pp. 327–334, 2011.

[111] G. Jegert, A. Kersch, W. Weinreich, U. Schroeder, and P. Lugli, “Modeling of leakage currents in high-kappa dielectrics: Three-dimensional approach via kinetic monte carlo,” Applied Physics Letters, vol. 96, no. 6, p. 062113, 2010.

[112] F. Rossi, P. Poli, and C. Jacoboni, “Weighted Monte Carlo approach to electron transport in semiconductors,” Semiconductor Science and Technology, vol. 7, p. 1017, 1992.

[113] T. M. Pan, C. I. Hsieh, T. Y. Huang, J. R. Yang, and P. S. Kuo, “Good

high-temperature stability of TiN/Al2O3/WN/TiN capacitors,” IEEE Elec- tron Device Letters, vol. 28, no. 11, pp. 954–956, 2007. 146 BIBLIOGRAPHY

[114] N. J. Son, Y. Oh, W. Kim, S. M. Jang, W. Yang, G. Jin, D. Park, and K. Kim, “A unique dual-poly gate technology for 1.2-v mobile dram with simple in situ n+-doped polysilicon,” IEEE Transactions on Electron Devices, vol. 51, no. 10, pp. 1644–1652, 2004.

[115] A. Berthelot, C. Caillat, V. Huard, S. Barnola, B. Boeck, H. Del-Puppo,

N. Emonet, and F. Lalanne, “Highly reliable TiN/ZrO2/TiN 3D stacked capacitors for 45 nm embedded DRAM technologies,” in Solid-State Device Re- search Conference, 2006.

[116] Y.-H. Wu, C.-K. Kao, B.-Y. Chen, Y.-S. Lin, M.-Y. Li, and H.-C. Wu, “High

density metal-insulator-metal capacitor based on ZrO2/Al2O3/ZrO2 laminate dielectric,” Applied Physics Letters, vol. 93, no. 3, p. 033511, 2008.

[117] S. H. Lin, K. C. Chiang, A. Chin, and F. S. Yeh, “High-Density and Low- Leakage-Current MIM Capacitor Using Stacked Insulators,” IEEE Electron Device Letters, vol. 30, no. 7, pp. 715 –717, july 2009.

[118] K. Xiong, J. Robertson, G. Pourtois, J. Pétry, and M. Müller, “Impact of

incorporated Al on the TiN/HfO2 interface eﬀective work function,” Journal of Applied Physics, vol. 104, no. 7, p. 074501, 2008.

[119] J. Robertson, “Band offsets and work function control in field effect transistors,” Journal of Vacuum Science & Technology B: Microelectronics and Nanometer Structures, vol. 27, p. 277, 2009.

[120] V. K. Gueorguiev, P. V. Aleksandrova, T. E. Ivanov, and J. B. Koprinarova, “Hysteresis in metal insulator semiconductor structures with high temperature

annealed ZrO2/SiOx layers,” Thin Solid Films, vol. 517, no. 5, pp. 1815–1820, 2009.

[121] G. W. Dietz and R. Waser, “How to analyse relaxation and leakage currents of dielectric thin ﬁlms: Simulation of voltage-step and voltage-ramp techniques,” Integrated Ferroelectrics, vol. 8, no. 3, pp. 317–332, 1995.

[122] D. V. Lang, “Deep-level transient spectroscopy: A new method to characterize traps in semiconductors,” Journal of Applied Physics, vol. 45, no. 7, pp. 3023– 3032, 1974. BIBLIOGRAPHY 147

[123] J. Robertson, K. Xiong, and B. Falabretti, “Point defects in ZrO2 high-κ gate oxide,” IEEE Transactions on Device and Materials Reliability, vol. 5, no. 1, pp. 84–89, 2005.

[124] M. C. V. Lopes, S. G. dos Santos Fo., C. M. Hasenack, and V. Baranauskas,

“Si-SiO2 Electronic Interface Roughness as a Consequence of Si-SiO2 Topo- graphic Interface Roughness,” Journal of The Electrochemical Society, vol. 143, no. 3, pp. 1021–1025, 1996.

[125] N. Gaillard, L. Pinzelli, M. Gros-Jean, and A. Bsiesy, “In situ ﬁeld simulation in metal/insulator/metal capacitors,” Applied Physics Letters, vol. 89, p. 133506, 2006.

[126] M. C. King, C. F. Chang, H. J. Lin, and A. Chin, “Comparison of mim performance with various electrodes and dieletric in cu dual damascene of cmos ms/rf technology,” Journal of The Electrochemical Society, vol. 153, no. 12, pp. G1032–G1034, 2006.

[127] J. Son, W. Maeng, W.-H. Kim, Y. H. Shin, and H. Kim, “Interface roughness eﬀect between gate oxide and metal gate on dielectric property,” Thin Solid Films, vol. 517, no. 14, pp. 3892 – 3895, 2009, the proceedings of the 1st International Conference on Microelectronics and Plasma Technology (ICMAP 2008).

[128] F.-C. Chiu, Z.-H. Lin, C.-W. Chang, C.-C. Wang, K.-F. Chuang, C.-Y. Huang, J. Y. Lee, and H.-L. Hwang, “Electron conduction mechanism and band dia-

gram of sputter-deposited Al/ZrO2/Si structure,” Journal of Applied Physics, vol. 97, no. 3, p. 034506, 2005.

[129] S. Ramanathan, C. M. Park, and P. C. McIntyre, “Electrical properties of thin ﬁlm zirconia grown by ultraviolet ozone oxidation,” Journal of Applied Physics, vol. 91, no. 7, pp. 4521–4527, 2002.

[130] Y.-P. Zhao, G.-C. Wang, T.-M. Lu, G. Palasantzas, and J. T. M. De Hosson, “Surface-roughness eﬀect on capacitance and leakage current of an insulating ﬁlm,” Physical Review B, vol. 60, no. 12, pp. 9157–9164, Sep 1999.

[131] G. Jegert, A. Kersch, W. Weinreich, and P. Lugli, “Ultimate scaling of

TiN/ZrO2/TiN capacitors: Leakage currents and limitations due to electrode roughness,” Journal of Applied Physics, vol. 109, p. 014504, 2011. 148 BIBLIOGRAPHY

[132] B. B. Mandelbrot, The fractal geometry of nature. Freeman, 1983.

[133] P. Meakin, “The growth of rough surfaces and interfaces,” Physics Reports, vol. 235, no. 4-5, pp. 189–289, 1993.

[134] B. Mandelbrot, “How long is the coast of Britain? Statistical self-similarity and fractional dimension,” Science, vol. 156, no. 3775, p. 636, 1967.

[135] A. L. Barabási and H. E. Stanley, Fractal concepts in surface growth. Cam- bridge Univ. Pr., 1995.

[136] H.-O. Peitgen, D. Saupe, and M. F. Barnsley, The Science of Fractal Images. Springer, New York, 1988.

[137] B. B. Mandelbrot and J. R. Wallis, “Noah, Joseph, and operational hydrol- ogy,” Water Resources Research, vol. 4, no. 5, pp. 909–918, 1968.

[138] M. Ding and W. Yang, “Distribution of the ﬁrst return time in fractional Brownian motion and its application to the study of on-oﬀ intermittency,” Physical Review E, vol. 52, no. 1, p. 207, 1995.

[139] D. W. Kelly, J. P. d. S. R. Gago, O. C. Zienkiewicz, and I. Babuska, “A posteriori error analysis and adaptive processes in the ﬁnite element method: Part i-error analysis,” International Journal for Numerical Methods in Engineering, vol. 19, no. 11, pp. 1593–1619, 1983.

[140] W. Bangerth, R. Hartmann, and G. Kanschat, “deal. II-A general-purpose object-oriented ﬁnite element library,” ACM Transactions on Mathematical Software, vol. 33, no. 4, pp. 24–es, 2007.

[141] O. Sharia, A. A. Demkov, G. Bersuker, and B. H. Lee, “Eﬀects of aluminum

incorporation on band alignment at the SiO2/HfO2 interface,” Physical Review B, vol. 77, no. 8, p. 085326, 2008.

[142] M. Häufel, “Defect energy levels in high-k transition metal oxides,” Ph.D. dissertation, Technical University of Munich, 2011.

[143] E. G. Seebauer and M. C. Kratzer, “Trends in semiconductor defect engineering at the nanoscale,” Materials Science and Engineering: R: Reports, vol. 55, p. 57, 2006. BIBLIOGRAPHY 149

[144] M. Levinson, M. Stavola, J. L. Benton, and L. C. Kimerling, “Metastable m center in inp: Defect-charge-state-controlled structural relaxation,” Physical Review B, vol. 28, no. 10, p. 5848, 1983.

[145] P. E. Blöchl, “First-principles calculations of defects in oxygen-deﬁcient silica exposed to hydrogen,” Phys. Rev. Lett., vol. 62, p. 6158, 2000.

[146] P. Broqvist and A. Pasquarello, “Oxygen vacancy in monoclinic hfo2: A consistent interpretation of trap assisted conduction, direct electron injection, and optical absorption experiments,” Applied Physics Letters, vol. 89, p. 262904, 2006.

[147] J. Kavalieros, B. Doyle, S. Datta, G. Dewey, M. Doczy, B. Jin, D. Lionberger, M. Metz, W. Rachmady, M. Radosavljevic et al., “Tri-gate transistor architecture with high-k gate dielectrics, metal gates and strain engineering,” in Symposium on VLSI Technology, 2006. Digest of Technical Papers. IEEE, 2006, pp. 50–51.

[148] A. Javey, H. Kim, M. Brink, Q. Wang, A. Ural, J. Guo, P. McIntyre, P. McEuen, M. Lundstrom, and H. Dai, “High-κ dielectrics for advanced carbon-nanotube transistors and logic gates,” Nature Materials, vol. 1, no. 4, pp. 241–246, 2002.

[149] K. Xiong and J. Robertson, “Defect energy levels in HfO2 high-dielectric- constant gate oxide,” Applied Physics Letters, vol. 87, p. 183505, 2005.

[150] D. Ricci, G. Bano, G. Pacchioni, and F. Illas, “Electronic structure of a neutral

oxygen vacancy in SrTiO3,” Physical Review B, vol. 68, no. 22, p. 224105, 2003.

[151] T. Bjørheim, S. Stølen, and T. Norby, “Ab initio studies of hydrogen and

acceptor defects in rutile TiO2,” Physical Chemistry Chemical Physics, vol. 12, no. 25, pp. 6817–6825, 2010.

[152] R. Chau, B. Doyle, S. Datta, J. Kavalieros, and K. Zhang, “Integrated nanoelectronics for the future,” Nature Materials, vol. 6, no. 11, pp. 810–812, 2007.

[153] S. Schmelzer, D. Bräuhaus, S. Hoﬀmann-Eifert, P. Meuﬀels, U. Böttger,

L. Oberbeck, P. Reinig, U. Schröder, and R. Waser, “SrTiO3 thin ﬁlm capacitors on silicon substrates with insigniﬁcant interfacial passive layers,” Applied Physics Letters, vol. 97, no. 13, p. 132907, 2010. 150 BIBLIOGRAPHY

[154] K. M. Rabe, “Nanoelectronics: New life for the ’dead layer’,” Nature Nan- otechnology, vol. 1, no. 3, pp. 171–172, 2006.

[155] C. J. Först, C. R. Ashman, K. Schwarz, and P. E. Blöchl, “The interface between silicon and a high-k oxide,” Nature, vol. 427, pp. 53–36, 2004.

[156] D. S. Kil, H. S. Song, K. J. Lee, K. Hong, J. H. Kim, K. S. Park, S. J. Yeom, J. S. Roh, N. J. Kwak, H. C. Sohn et al., “Development of New

TiN/ZrO2/Al2O3/ZrO2/TiN Capacitors Extendable to 45 nm Generation

DRAMs Replacing HfO2 Based Dielectrics,” in Symposium on VLSI Tech- nology, 2006. Digest of Technical Papers. IEEE, 2006, pp. 38–39.

[157] K. Xiong, Y. Du, K. Tse, and J. Robertson, “Defect states in the high- dielectric-constant gate oxide HfSiO,” Journal of Applied Physics, vol. 101, p. 024101, 2007.

[158] L. Hedin and S. Lundqvist, “Eﬀects of electron-electron and electron-phonon interactions on the one-electron states of solids,” Solid State Physics, vol. 23, pp. 1–181, 1970.

[159] W. G. Aulbur, L. J "onsson, and J. W. Wilkins, “Quasiparticle calculations in solids,” Solid State Physics, vol. 54, pp. 1–218, 1999.

[160] M. Shishkin and G. Kresse, “Self-consistent gw calculations for semiconductors and insulators,” Physical Review B, vol. 75, no. 23, p. 235102, 2007.

[161] M. Torrent, F. Jollet, F. Bottin, G. Zérah, and X. Gonze, “Implementation of the projector augmented-wave method in the abinit code: Application to the study of iron under pressure,” Computational Materials Science, vol. 42, no. 2, pp. 337–351, 2008.

[162] X. Gonze, G. M. Rignanese, and R. Caracas, “First-principle studies of the lattice dynamics of crystals, and related properties,” Zeitschrift für Kristallo- graphie, vol. 220, no. 5/6/2005, pp. 458–472, 2005.

[163] X. Gonze, B. Amadon, P. M. Anglade, J. M. Beuken, F. Bottin, P. Boulanger, F. Bruneval, D. Caliste, R. Caracas, M. Cote et al., “Abinit: First-principles approach to material and nanosystem properties,” Computer Physics Com- munications, vol. 180, no. 12, pp. 2582–2615, 2009. BIBLIOGRAPHY 151

[164] X. Gonze, J. M. Beuken, R. Caracas, F. Detraux, M. Fuchs, G. M. Rignanese, L. Sindic, M. Verstraete, G. Zerah, F. Jollet et al., “First-principles computation of material properties: the abinit software project,” Computational Materials Science, vol. 25, no. 3, pp. 478–492, 2002.

[165] D. Vanderbilt, “Soft self-consistent pseudopotentials in a generalized eigen- value formalism,” Physical Review B, vol. 41, no. 11, p. 7892, 1990.

[166] J. P. Perdew and A. Zunger, “Self-interaction correction to density-functional approximations for many-electron systems,” Physical Review B, vol. 23, no. 10, p. 5048, 1981.

[167] D. Ceperley and B. Alder, “Ground state of the electron gas by a stochastic method,” Physical Review Letters, vol. 45, no. 7, pp. 566–569, 1980.

[168] R. W. Godby and R. J. Needs, “Metal-insulator transition in kohn-sham theory and quasiparticle theory,” Physical Review Letters, vol. 62, no. 10, pp. 1169– 1172, 1989.

[169] K. O. Axelsson, K. E. Keck, and B. Kasemo, “AES and EEL spectra from

Zr, Zr + O2 and ZrO2; peak energy shifts and intensity variations at various 6 3 stages of oxidation in the pressure range 10− -10 Torr,” Surface Science, vol. 164, no. 1, pp. 109–126, 1985.

[170] P. E. Blöchl, “Projector augmented-wave method,” Physical Review B, vol. 50, no. 24, p. 17953, 1994.

[171] J. S. Thorp, A. Aypar, and J. S. Ross, “Electron spin resonance in single crystal yttria stabilized zirconia,” Journal of Materials Science, vol. 7, no. 7, pp. 729–734, 1972.

[172] D. M. Ramo, P. V. Sushko, J. L. Gavartin, and A. L. Shluger, “Oxygen va-

cancies in cubic zro2 nanocrystals studied by an ab initio embedded cluster method,” Physical Review B, vol. 78, no. 23, p. 235432, 2008.

[173] A. S. Foster, V. B. Sulimov, F. L. Gejo, A. L. Shluger, and R. M. Niemi- nen, “Structure and electrical levels of point defects in monoclinic zirconia,” Physical Review B, vol. 64, no. 22, p. 224108, 2001.

[174] B. Králik, E. K. Chang, and S. G. Louie, “Structural properties and quasiparticle band structure of zirconia,” Physical Review B, vol. 57, no. 12, p. 7027, 1998. 152 BIBLIOGRAPHY

[175] X. Urli, C. L. Dias, L. J. Lewis, and S. Roorda, “Point defects in pure amorphous silicon and their role in structural relaxation: A tight-binding molecular- dynamics study,” Physical Review B, vol. 77, no. 15, p. 155204, 2008.

[176] G. Lucovsky, D. M. Fleetwood, S. Lee, H. Seo, R. D. Schrimpf, J. A. Felix, J. Lning, L. B. Fleming, M. Ulrich, and D. E. Aspnes, “Diﬀerences between

charge trapping states in irradiated nano-crystalline HfO2 and non-crystalline Hf silicates,” IEEE Transactions on Nuclear Science, vol. 53, no. 6, pp. 3644– 3648, 2006.

[177] D. Martin, M. Grube, W. M. Weber, J. Rustig, O. Bierwagen, L. Geelhaar, and H. Riechert, “Local charge transport in nanoscale amorphous and crys-

talline regions of high-k (Zr02)0.8(Al2O3)0.2 thin ﬁlms,” Applied Physics Let- ters, vol. 95, no. 14, pp. 142 906–142 906, 2009.

[178] G. Lucovsky, C. C. Fulton, Y. Zhang, Y. Zou, J. Luning, L. F. Edge, J. L. Whitten, R. J. Nemanich, H. Ade, D. G. Schlom et al., “Conduction band-edge states associated with the removal of d-state degeneracies by the jahn-teller eﬀect,” IEEE Transactions on Device and Materials Reliability, vol. 5, no. 1, pp. 65–83, 2005.

[179] M. Bareiss, B. Tiwari, A. Hochmeister, G. Jegert, U. Zschieschang, H. Klauk, B. Fabel, G. Scarpa, G.and Koblmüller, G. Bernstein, W. Porod, and P. Lugli, “Nano antenna array for terahertz detection,” IEEE Transactions on Mi- crowave Theory and Techniques, accepted for publication, 2011.

[180] M. Bareiss, A. Hochmeister, G. Jegert, U. Zschieschang, H. Klauk, R. Huber, D. Grundler, W. Porod, B. Fabel, G. Scarpa, and P. Paolo, “Printed array of thin-dielectric mom tunneling diodes,” Jour. Appl, under review, 2011.

[181] M. Bareiss, A. Hochmeister, G. Jegert, U. Zschieschang, H. Klauk, B. Fabel, G. Scarpa, W. Porod, and P. Lugli, “Quantum carrier dynamics in ultra- thin mim tunneling diodes,” in 17th International Conference on Electron Dynamics in Semiconductors, Optoelectronics and Nanostructures, EDISON 17, 2011.

[182] M. Bareiss, A. Hochmeister, G. Jegert, Z. U. Koblmüller, G., H. Klauk, B. Fa- bel, G. Scarpa, W. Porod, and P. Lugli, “Energy harvesting using nano antenna arrays,” in 11th International Conference on Nanotechnology, NANO, 2011. BIBLIOGRAPHY 153

[183] P. Peumans, S. Uchida, and S. R. Forrest, “Eﬃcient bulk heterojunction photovoltaic cells using small-molecular-weight organic thin ﬁlms,” Nature, vol. 425, no. 6954, pp. 158–162, 2003.

[184] M. Casalegno, G. Raos, and R. Po, “Methodological assessment of kinetic Monte Carlo simulations of organic photovoltaic devices: The treatment of electrostatic interactions,” Journal of Chemical Physics, vol. 132, p. 094705, 2010.

[185] F. Yang, K. Sun, and S. R. Forrest, “Eﬃcient solar cells using all-organic nanocrystalline networks,” Advanced Materials, vol. 19, no. 23, pp. 4166–4171, 2007.

[186] F. Yang and S. R. Forrest, “Photocurrent generation in nanostructured organic solar cells,” Nano, vol. 2, p. 1022, 2008.

[187] R. A. Marsh, C. Groves, and N. C. Greenham, “A microscopic model for the behavior of nanostructured photovoltaic devices,” Journal of Applied Physics, vol. 101, p. 083509, 2007.

[188] P. K. Watkins, A. B. Walker, and G. L. B. Verschoor, “Dynamical monte carlo modelling of organic solar cells: The dependence of internal quantum eﬃciency on morphology,” Nano Letters, vol. 5, no. 9, pp. 1814–1818, 2005. 154 List of Own Publications and Contributions

List of Own Publications and Contributions

G. Jegert, M. J. Haeufel, D. Popescu, W. Weinreich, A. Kersch, P. Lugli, The role of defect relaxation for trap-assisted tunneling in high-k thin ﬁlms: A ﬁrst principles kinetic Monte Carlo study, Physical Review B (under review)

G. Jegert, A. Kersch, W. Weinreich, P. Lugli,

Monte Carlo Simulation of Leakage Currents in TiN/ZrO2/TiN Capacitors, IEEE Transactions on Electron Devices 58, 327 (2011)

G. Jegert, A. Kersch, W. Weinreich, P. Lugli,

Ultimate scaling of TiN/ZrO2/TiN capacitors: Leakage currents and limitations due to electrode roughness, Journal of Applied Physics 10, 014504 (2011)

M. Bareiss, B. Tiwari, A. Hochmeister, G. Jegert, U. Zschieschang, H. Klauk, B. Fabel, G. Scarpa, G. Koblmüller, G. Bernstein, W. Porod, P. Lugli, Nano Antenna Array for Terahertz Detection, IEEE Transactions on Microwave Theory and Techniques, accepted for publication (2011)

M. Bareiss, A. Hochmeister, G. Jegert, G. Koblmüller, U. Zschieschang, H. Klauk, B. Fabel, G. Scarpa, W. Porod, P. Lugli, Energy Harvesting using Nano Antenna Arrays, 11th International Conference on Nanotechnology, NANO (2011)

M. Bareiss, A. Hochmeister, G. Jegert, U. Zschieschang, H. Klauk, B. Fabel, G. Scarpa, W. Porod, P. Lugli, Quantum carrier dynamics in ultra-thin MIM tunneling diodes 17th International Conference on Electron Dynamics in Semiconductors, Optoelectronics and Nanostructures, EDISON 17 (2011) List of Own Publications and Contributions 155

M. Bareiss, A. Hochmeister, G. Jegert, U. Zschieschang, H. Klauk, R. Huber, D. Grundler, W. Porod, B. Fabel, G. Scarpa, P. Lugli, Printed Array of Thin-Dielectric MOM Tunneling Diodes, Jour. Appl. Phys., 110, 044316 (2011).

D. Zhou, U. Schroeder, J. Xu, J. Heitmann, G. Jegert, W. Weinreich, M. Kerber, S. Knebel, E. Erben, T. Mikolajick,

Reliability of Al2O3-doped ZrO2 high-k dielectrics in three-dimensional stacked metal- insulator-metal capacitors, Journal of Applied Physics 108, 124104 (2010)

G. Jegert, A. Kersch, W. Weinreich, U. Schroeder, P. Lugli, Modeling of leakage currents in high-k dielectrics: Three-dimensional approach via kinetic Monte Carlo, Applied Physics Letters 96, 062113 (2010)

D. Zhou, U. Schroeder, G. Jegert, M. Kerber, S. Uppal, R. Agaiby, M. Reinicke, J. Heitmann, L. Oberbeck,

Time dependent dielectric breakdown of amorphous ZrAlxOy high-k dielectric used in dynamic random access memory metal-insulator-metal capacitor, Journal of Applied Physics 106, 044104 (2009)

W. Weinreich, R. Reiche, M. Lemberger, G. Jegert, J. Müller, L. Wilde, S. Teichert, J. Heitmann, E. Erben, L. Oberbeck, U. Schroeder, A.J. Bauer, H. Ryssel, Impact of interface variations on J-V and C-V polarity asymmetry of MIM capacitors with amorphous and crystalline Zr(1 x)AlxO2 ﬁlms, Microelectronic Engineer- − ing 86, 1826 (2009)

U. Schroeder, W. Weinreich, E. Erben, J. Mueller, L. Wilde, J. Heitmann, R. Agaiby, D. Zhou, G. Jegert, A. Kersch, Detailed Correlation of Electrical and Breakdown Characteristics to the Structural

Properties of ALD Grown HfO2- and ZrO2-based Capacitor Dielectrics, The Electrochemical Society Transactions 25, 357 (2009) 156 List of Abbreviations List of Abbreviations

3D Three-Dimensional ALD Atomic Layer Deposition BKL Bortz Kalos Lebowitz BL Bitline CB Conduction Band CBO Conduction Band Oﬀset DAT Defect-Assisted Transport DRAM Dynamic Random Access Memory EOT Equivalent Oxide Thickness ER Electrode Roughness FN Fowler-Nordheim IBS Imaginary Band Structure IC Integrated Circuit IES Infrequent Event System ITRS International Technology Roadmap for Semiconductors IV Current-Voltage JEDEC Joint Electron Device Engineering Council kMC kinetic Monte Carlo LDA Local Density Approximation LEFE Local Electric Field Enhancement MA Miller-Abrahams MD Molecular Dynamics MIGS Metal-Induced Gap States MIM Metal-Insulator-Metal PAW Projector-Augmented Wave pb parabolic PDF Probability Density Function PES Potential Energy Surface

157 158 List of Abbreviations

PF Poole-Frenkel SE Schottky Emission SEM Scanning Electron Microscopy t-ZrO2 tetragonal ZrO2 TAPF Tunneling-Assisted Poole-Frenkel TAT Trap-Assisted Tunneling TB Tight Binding TxZAZT TZAZT capacitor with x nm thick dielectric TxZT TZT capacitor with x nm thick dielectric

TZAZT TiN/ZrO2/Al2O3/ZrO2/TiN

TZT TiN/ZrO2/TiN USPP Ultra-Soft Pseudopotentials VB Valence Band WKB Wentzel Kramers Brillouin WL Wordline Acknowledgements 159

Acknowledgements

At this point, I would like to take the opportunity to mention all those people I owe a debt of gratitude for their support, their participation and their contributions which helped me to complete this work: First, and most of all, I would like to thank my supervisor Prof. Dr. Alfred Kersch for his unlimited support and for sharing his vast knowledge in the field of the semiconductor device simulations and many other fields in countless discussions. He has been the best possible guide and advisor on this long journey. I thank Prof. Dr. Paolo Lugli, for his support, for his huge efforts which allowed me to complete this work after the bankruptcy of Qimonda, for many fruitful discussions and for providing me an excellent work environment. The present work would not have been possible without several scientific collab- orations which I enjoyed very much. I would like to thank all the people which were involved in these. In particular, I am indebted to Wenke Weinreich, Dr. Uwe Schröder, Dr. Dayu Zhou, Mario Bareißand Martin Johannes Häufel. Moreover, I thank all my colleagues at the Institute for Nanoelectronics and the members of the Qimonda simulation group for scientic support as well as for creating an excellent atmosphere in which it was a pleasure to work. I owe special thanks to Dr. Wolfgang Hönlein for being a fantastic group leader during my time at Qimonda and for his outstanding support during the turbulent times following the bankruptcy of Qimonda. Last, but most importantly, I thank Angela Link for the unconditioned and loving support throughout the thesis, which helped me to endure also the stressful phases during the last years. In der Schriftenreihe des Walter Schottky Instituts der Technischen Universität München sind bisher folgende Bände erschienen:

Vol. 1 Vol. 7 Cornelia Engel Markus Sexl Si/SiGe basierende Phototransistoren Verspannte und gitterrelaxierte 131 Seiten In(GaAl)As Heterostrukturen ISBN 3-932749-01-4 144 Seiten ISBN 3-932749-07-3

Vol. 2 Vol. 8 Peter Schittenhelm Christian Obermüller Selbst-Organisation und Selbst-Ordnung Photolumineszenszspektroskopie mit in Si/SiGe-Heterostrukturen optischen Nahfeldmethoden an GaAs- 151 Seiten Nanostrukturen ISBN 3-932749-02-2 140 Seiten ISBN 3-932749-08-1

Vol. 3 Vol. 9 Andreas Nutsch Edilson Silveira Selektive Epitaxie von (GaIn)(AsP) Inelastische Lichtstreuung an niedrig- Schichtstrukturen dimensionalen Halbleiterstrukturen 129 Seiten 104 Seiten ISBN 3-932749-03-0 ISBN 3-932749-09-X

Vol. 4 Vol. 10 Peter Baumgartner Eberhard Christian Rohrer Optische und elektronische Eigenschaften Photoleitungs-Spektroskopie von lasergeschriebener GaAs-Nanostrukturen Diamant 180 Seiten 153 Seiten ISBN 3-932749-04-9 ISBN 3-932749-10-03

Vol. 5 Vol. 11 Walter Franz Rieger Thomas Wimbauer Untersuchung der elektronischen und Magnetische Resonanz-Untersuchungen an strukturellen Eigenschaften von modernen Halbleitermaterialien GaNAlN und deren Legierungen 125 Seiten 158 Seiten ISBN 3-932749-11-1 ISBN 3-932749-05-7

Vol. 6 Vol. 12 Markus Hauser Herbert Verhoeven Oberflächenemittierende Laserdioden Thermische Eigenschaften von mit Mehrfachepitaxie CVD-Diamantschichten 148 Seiten 154 Seiten ISBN 3-932749-06-5 ISBN 3-932749-12-X

Vol. 13 Vol. 20 Hans-Christoph Ostendorf Christoph Martin Engelhardt Trennung von Volumen- und Zyklotronresonanz zweidimensionaler Oberflächenrekombination in Silizium Ladungsträgersysteme in Halbleitern, 128 Seiten Effekte der Elektron-Elektron-Wechsel- ISBN 3-932749-13-8 wirkung und Lokalisierung 317 Seiten Vol. 14 ISBN 3-932749-20-0 Martin Städele Dichtefunktionaltheorie mit exaktem Vol. 21 Austausch für Halbleiter Eduard Neufeld 202 Seiten Erbium-dotierte Si/SiGe-Lichtemitter und ISBN 3-932749-14-6 -Wellenleiter 136 Seiten Vol. 15 ISBN 3-932749-21-9 Helmut Angerer Herstellung von Gruppe III-Nitriden mit Vol. 22 Molekularstrahlepitaxie Gert Schedelbeck 144 Seiten Optische Eigenschaften von Halbleiter- ISBN 3-932749-15-4 nanostrukturen hergestellt durch Über- wachsen von Spaltflächen Vol. 16 154 Seiten Wolfgang Heller ISBN 3-932749-22-7 Spektroskopie einzelner Quantenpunkte in magnetischen und elektrischen Feldern Vol. 23 128 Seiten Jürgen Zimmer ISBN 3-932749-16-2 Optoelektronisches Verhalten von Dünn- schichtbauelementen aus amorphem und Vol. 17 mikrokristallinem Silizium Molela Moukara 171 Seiten Pseudopotentiale mit exaktem Austausch ISBN 3-932749-23-5 117 Seiten ISBN 3-932749-17-0 Vol. 24 Berthold Schmidt Vol. 18 Leistungsoptimierung abstimmbarer Ralph Oberhuber InGaAsP/InP Halbleiterlaser Elektronische Struktur und Transport in 85 Seiten verspannten Halbleiterschichtsystemen ISBN 3-932749-24-3 110 Seiten ISBN 3-932749-18-9 Vol. 25 Jianhong Zhu Vol. 19 Ordering of self-assembled Ge and SiGe Reiner Pech nanostructures on vicinal Si surfaces High-Energy Boron-Implantation into 120 Seiten Different Silicon Substrates ISBN 3-932749-25-1 158 Seiten ISBN 3-932749-19-7 Vol. 26 Vol. 33 Gerhard Groos Martin Rother Herstellung und Charakterisierung von Elektronische Eigenschaften von Silizium-Nanostrukturen Halbleiternanostrukturen hergestellt durch 168 Seiten Überwachsen von Spaltflächen ISBN 3-932749-26-X 196 Seiten ISBN 3-932749-33-2

Vol. 27 Vol. 34 Uwe Hansen Frank Findeis Theorie der Reaktionskinetik an Optical spectroscopy on single self- Festkörperoberflächen assembled quantum dots 119 Seiten 156 Seiten ISBN 3-932749-27-8 ISBN 3-932749-34-0

Vol. 28 Vol. 35 Roman Dimitrov Markus Ortsiefer Herstellung und Charakterisierung Langwellige Vertikalresonator-Laser- von AlGaN/GaN-Transistoren dioden im Materialsystem InGaAlAs/InP 196 Seiten 152 Seiten ISBN 3-932749-28-6 ISBN 3-932749-35-9

Vol. 29 Vol. 36 Martin Eickhoff Roland Zeisel Piezowiderstandsmechanismen in Optoelectronic properties of defects in Halbleitern mit großer Bandlücke diamond and AlGaN alloys 151 Seiten 140 Seiten ISBN 3-932749-29-4 ISBN 3-932749-36-7

Vol. 30 Vol. 37 Nikolai Wieser Liwen Chu Ramanspektroskopie an Gruppe III- Inter- und Intraband Spektroskopie an Nitriden selbstorganisierten In(Ga)As/GaAs 161 Seiten Quantenpunkten ISBN 3-932749-30-8 124 Seiten ISBN 3-932749-37-5

Vol. 31 Vol. 38 Rainer Janssen Christian Alexander Miesner Strukturelle und elektronische Intra-Valenzbandspektroskopie an SiGe- Eigenschaften amorpher Silizium-Suboxide Nanostrukturen in Si 275 Seiten 100 Seiten ISBN 3-932749-31-6 ISBN 3-932749-38-3

Vol. 32 Vol. 39 Martin W. Bayerl Szabolcs Kátai Magnetic resonance investigations Investigation of the nucleation process of of group III-nitrides chemical vapour deposited diamond films 155 Seiten 178 Seiten ISBN 3-932749-32-4 ISBN 3-932749-39-1

Vol. 40 Vol. 46 Markus Arzberger Jan Schalwig Wachstum, Eigenschaften und An- Feldeffekt-Gassensoren und ihre wendungen selbstorganisierter InAs- Anwendung in Abgas- Quantenpunkte nachbehandlungssystemen 236 Seiten 125 Seiten ISBN 3-932749-40-5 ISBN 3-932749-46-4

Vol. 41 Vol. 47 Markus Oliver Markmann Christopher Eisele Optische Eigenschaften von Erbium in Novel absorber structures for Si-based Si/Si1-xCx, Si/Si1-xGex und Si/SiOx thin film solar cells Heterostrukturen 126 Seiten 182 Seiten ISBN 3-932749-47-2 ISBN 3-932749-41-3

Vol. 42 Vol. 48 Rainer Alexander Deutschmann Stefan Hackenbuchner Two dimensional electron systems in Elektronische Struktur von Halbleiter- atomically precise periodic potential Nanobauelementen im thermodynamischen 210 Seiten Nichtgleichgewicht ISBN 3-932749-42-1 213 Seiten ISBN 3-932749-48-0

Vol. 43 Vol. 49 Uwe Karrer Andreas Sticht Schottky-Dioden auf Galliumnitrid: Herstellung und Charakterisierung Eigenschaften und Anwendungen in der von dünnen Silizium/Siliziumoxid- Sensorik Schichtsystemen 182 Seiten 166 Seiten ISBN 3-932749-43-X ISBN 3-932749-49-9

Vol. 44 Vol. 50 Günther Anton Johann Vogg Giuseppe Scarpa Epitaxial thin films of Si and Ge based Design and fabrication of Quantum Zintl phases and sheet polymers Cascade Lasers 169 Seiten 193 Seiten ISBN 3-932749-44-8 ISBN 3-932749-50-2

Vol. 45 Vol. 51 Christian Strahberger Jörg Frankenberger Vertikaler Transport und extreme Optische Untersuchungen an zwei- Magnetfelder in Halbleitern dimensionalen Ladungsträgersystemen 167 Seiten 158 Seiten ISBN 3-932749-45-6 ISBN 3-932749-51-0

Vol. 52 Vol. 58 Doris Heinrich Po-Wen Chiu Wavelength selective optically induced Towards carbon nanotube-based molecular charge storage in self-assembled electronics semiconductor quantum dots 116 Seiten 144 Seiten ISBN 3-932749-58-8 ISBN 3-932749-52-9

Vol. 53 Vol. 59 Nicolaus Ulbrich Tobias Graf Entwurf und Charakterisierung von Spin-spin interactions of localized Quanten-Kaskadenlasern und electronic states in semiconductors Quantenpunktkaskaden 194 Seiten 133 Seiten ISBN 3-932749-59-6 ISBN 3-932749-53-7

Vol. 54 Vol. 60 Lutz Carsten Görgens Stefan Klein Analyse stickstoffhaltiger III-V Halbleiter- Microcrystalline silicon prepared by hot Heterosysteme mit hochenergetischen wire CVD: preparation and characteri- schweren Ionen sation of material and solar cells 116 Seiten 157 Seiten ISBN 3-932749-54-5 ISBN 3-932749-60-X

Vol. 55 Vol. 61 Andreas Janotta Markus Krach Doping, light-induced modification and Frequenzverdreifacher mit Anti-Seriellem biocompatibility of amorphous silicon Schottky-Varaktor für den Terahertz- suboxides bereich 180 Seiten 156 Seiten ISBN 3-932749-55-3 ISBN 3-932749-61-8

Vol. 56 Vol. 62 Sebastian Tobias Benedikt Gönnenwein Ralph Thomas Neuberger Two-dimensional electron gases and AlGaN/GaN-Heterostrukturen als ferromagnetic semiconductors: chemische Sensoren in korrosiven materials for spintronics Medien 198 Seiten 153 Seiten ISBN 3-932749-56-1 ISBN 3-932749-62-6

Vol. 57 Vol. 63 Evelin Beham Sonia Perna Photostromspektroskopie an einzelnen Wasserstoff-Passivierung von tri- Quantenpunkten kristallinem Silizium durch hydro- 186 Seiten genisiertes Siliziumnitrid ISBN 3-932749-57-X 136 Seiten ISBN 3-932749-63-4

Vol. 64 Vol. 71 Oliver Schumann Andreas Florian Kreß Einfluss von Stickstoff auf das Manipulation of the Light-Matter-Inter- Wachstum und die Eigenschaften action in Photonic Crystal Nanocavities von InAs-Quantenpunkten 185 Seiten 148 Seiten ISBN 3-932749-71-5 ISBN 3-932749-64-2

Vol. 65 Vol. 72 Gerhard Rösel Markus Grau Entwicklung und Charakterisierung von Molekularstrahlepitaktische Herstellung Typ-II-Heterostrukturen für die Abstimm- von antimonidischen Laserdioden für region in abstimmbaren Laserdioden die Gassensorik 101 Seiten 138 Seiten ISBN 3-932749-65-0 ISBN 3-932749-72-3

Vol. 66 Vol. 73 Angela Link Karin Buchholz Zweidimensionale Elektronen- und Löcher- Microprocessing of silicon on insulator Gase in GaN/AlGaN Heterostrukturen substrates and biofunctionalisation of 156 Seiten silicon dioxide surfaces for sensing ISBN 3-932749-66-9 applications in fluids 170 Seiten Vol. 67 ISBN 3-932749-73-1 Matthias Sabathil Opto-electronic and quantum transport Vol. 74 properties of semiconductor nanostructures Dominique Bougeard 156 Seiten Spektroskopische Charakterisierung von ISBN 3-932749-67-7 Germanium-Quantenpunkten in Silizium 154 Seiten Vol. 68 ISBN 3-932749-74-X Frank Fischer Growth and electronic properties of two- Vol. 75 dimensional systems on (110) oriented GaAs Jochen Bauer 139 Seiten Untersuchungen zum kontrollierten ISBN 3-932749-68-5 Wachstum von InAs-Nanostrukturen auf Spaltflächen Vol. 69 140 Seiten Robert Shau ISBN 3-932749-75-8 Langwellige oberflächenemittierende Laser- dioden mit hoher Ausgangsleistung und Vol. 76 Modulationsbandbreite Ingo Bormann 198 Seiten Intersubband Spektroskopie an Silizium- ISBN 3-932749-69-3 Germanium Quantenkaskadenstrukturen 124 Seiten Vol. 70 ISBN 3-932749-76-6 Andrea Baumer Structural and electronic properties of hydrosilylated silicon surfaces 163 Seiten ISBN 3-932749-70-7 Vol. 77 Vol. 84 Hubert Johannes Krenner Claudio Ronald Miskys Coherent quantum coupling of excitons New substrates for epitaxy of group III in single quantum dots and quantum nitride semiconductors: challenges and dot molecules potential 160 Seiten 207 Seiten ISBN 3-932749-77-4 ISBN 978-3-932749-84-1

Vol. 78 Vol. 85 Ulrich Rant Sebastian Friedrich Roth Electrical manipulation of DNA-layers n- and p-type transport in (110) GaAs on gold surfaces substrates, single- and double-cleave 249 Seiten structures ISBN 3-932749-78-2 138 Seiten ISBN 978-3-932749-85-8 Vol. 79 René Todt Vol. 86 Widely tunable laser diodes with Mario Gjukic distributed feedback Metal-induced crystallization of 152 Seiten silicon-germanium alloys ISBN 3-932749-79-0 309 Seiten ISBN 978-3-932749-86-5 Vol. 80 Miroslav Kroutvar Vol. 87 Charge and spin storage in quantum dots Tobias Zibold 150 Seiten Semiconductor based quantum ISBN 3-932749-80-4 information devices: Theory and simulations Vol. 81 151 Seiten Markus Maute ISBN 978-3-932749-87-2 Mikromechanisch abstimmbare Laser-Dioden mit Vertikalresonator Vol. 88 170 Seiten Thomas Jacke ISBN 3-932749-81-2 Weit abstimmbare Laserdiode mit vertikal integriertem Mach-Zehnder- Vol. 82 Interferometer Frank Ertl 165 Seiten Anisotrope Quanten-Hall-Systeme, Vertikale ISBN 978-3-932749-88-9 Ultrakurzkanal- und Tunneltransistoren 170 Seiten Vol. 89 ISBN 3-932749-82-0 Nenad Ocelić Quantitative near-field phonon- Vol. 83 polariton spectroscopy Sebastian M. Luber 174 Seiten III-V semiconductor structures for biosensor ISBN 978-3-932749-89-6 and molecular electronics applications 212 Seiten ISBN 978-3-932749-83-4

Vol. 90 Vol. 96 Kenji Arinaga Stefan Ahlers Control and manipulation of DNA on Magnetic and electrical properties gold and its application for biosensing of epitaxial GeMn 111 Seiten 184 Seiten ISBN 978-3-932749-90-2 ISBN 978-3-932749-96-0

Vol. 91 Vol. 97 Hans-Gregor Hübl Emanuele Uccelli Coherent manipulation and electrical Guided self-assembly of InAs quantum detection of phosphorus donor spins dots arrays on (110) surfaces in silicon 172 Seiten 162 Seiten ISBN 978-3-932749-97-1 ISBN 978-3-932749-91-9

Vol. 92 Vol. 98 Andrea Friedrich Shavaji Dasgupta Quanten-Kaskaden-Laser ohne Growth optimization and characteri- Injektorbereiche zation of high mobility two-dimensional 140 Seiten electron systems in AlAs quantum wells ISBN 978-3-932749-92-6 152 Seiten ISBN 978-3-932749-98-8

Vol. 93 Vol. 99 Oliver Dier Werner Hofmann Das Materialsystem (AlGaIn) (AsSb): Eigen- InP-based long-wavelength VCSELs schaften und Eignung für GaSb-basierte and VCSEL arrays for high-speed Vertikalresonator-Laserdioden optical communication 174 Seiten 142 Seiten ISBN 978-3-932749-93-3 ISBN 978-3-932749-99-5

Vol. 94 Vol. 100 Georg Steinhoff Robert Lechner Group III-nitrides for bio- and electro- Silicon nanocrystal films for chemical sensors electronic applications 197 Seiten 227 Seiten ISBN 978-3-932749-94-0 ISBN 978-3-941650-00-8

Vol. 95 Vol. 101 Stefan Harrer Nebile Işık Next-generation nanoimprint lithography: Investigation of Landau level spin Innovative approaches towards improving reversal in (110) oriented p-type flexibility and resolution of nanofabrication GaAs quantum wells in the sub-15-nm region 114 Seiten 161 Seiten ISBN 978-3-941650-01-5 ISBN 978-3-932749-95-7

Vol. 102 Vol. 109 Andreas Florian Härtl Sebastian Strobel Novel concepts for biosensors using Nanoscale contacts to organic molecules diamond-based field effect transistors based on layered semiconductor 255 Seiten substrates ISBN 978-3-941650-02-2 140 Seiten ISBN 978-3-941650-09-1 Vol. 103 Felix Florian Georg Hofbauer Realization of electrically tunable single Vol. 110 quantum dot nanocavities Ying Xiang 160 Seiten Semiconductor nanowires and ISBN 978-3-941650-03-9 templates for electronic applications 152 Seiten Vol. 104 ISBN 978-3-941650-10-7 Dominic F. Dorfner Novel photonic biosensing based on Vol. 111 silicon nanostructures Michael Kaniber 169 Seiten Non-classical light generation in ISBN 978-3-941650-04-6 photonic crystal nanostructures 177 Seiten Vol. 105 ISBN 978-3-941650-11-4 Till Andlauer Optoelectronic and spin-related properties Vol. 112 of semiconductor nanostructures in Martin Hermann magnetic fields Epitaktische AlN-Schichten auf 157 Seiten Saphir und Diamant ISBN 978-3-941650-05-3 216 Seiten ISBN 978-3-941650-12-1 Vol. 106 Christoph Bihler Vol. 113 Magnetic semiconductors Dominik Heiss 190 Seiten Spin storage in quantum dot ensembles ISBN 978-3-941650-06-0 and single quantum dots 196 Seiten Vol. 107 ISBN 978-3-941650-13-8 Michael Huber Tunnel-Spektroskopie im Vol. 114 Quanten-Hall-Regime Tillmann Christoph Kubis 164 Seiten Quantum transport in semiconductor ISBN 978-3-941650-07-7 nanostructures 253 Seiten Vol. 108 ISBN 978-3-941650-14-5 Philipp Achatz Metal-insulator transition and super- Conductivity in heavily boron-doped diamond and related materials 151 Seiten ISBN 978-3-941650-08-4

Vol. 115 Vol. 122 Lucia Steinke Ilaria Zardo Magnetotransport of coupled quantum Growth and raman spectroscopy Hall edges in a bent quantum well studies of gold-free catalyzed semi- 194 Seiten conductor nanowires ISBN 978-3-941650-15-2 184 Seiten ISBN 978-3-941650-22-0 Vol. 116 Christian Lauer Vol. 123 Antimonid-basierte Vertikalresonator- Andre Rainer Stegner Laserdioden für Wellenlängen oberhalb 2 µm Shallow dopants in nanostructured and 180 Seiten in isotopically engineered silicon ISBN 978-3-941650-16-9 185 Seiten ISBN 978-3-941650-23-7 Vol. 117 Simone Maria Kaniber Vol. 124 Optoelektronische Phänomene in hybriden Andreas J. Huber Schaltkreisen aus Kohlenstoffnanoröhren und Nanoscale surface-polariton spectros- dem Photosystem I copy by mid- and far-infrared near- 136 Seiten field microscopy ISBN 978-3-941650-17-6 144 Seiten ISBN 978-3-941650-24-4 Vol. 118 Martin Heiß Vol. 125 Growth and properties of low-dimensional Marco Andreas Höb III-V semiconductor nanowire hetero- Funktionalisierung von Gruppe structures IV-Halbleitern 172 Seiten 186 Seiten ISBN 978-3-941650-18-3 ISBN 978-3-941650-25-1

Vol. 119 Vol. 126 Sandro Francesco Tedde Daniel Claudio Pedone Design, fabrication and characterization of Nanopore analytics – electro-optical organic photodiodes for industrial and studies on single molecules medical applications 114 Seiten 277 Seiten ISBN 978-3-941650-26-8 ISBN 978-3-941650-19-0

Vol. 120 Vol. 127 Danche Spirkoska Jovanov Casimir Richard Simeon Katz Fundamental properties of self-catalyzed Multi-alloy structures for injectorless GaAs nanowires and related heterostructures Quantum Cascade Lasers 200 Seiten 131 Seiten ISBN 978-3-941650-20-6 ISBN 978-3-941650-27-5

Vol. 121 Vol. 128 Jürgen Sailer Barbara Annemarie Kathrin Baur Materials and devices for quantum Functionalization of group III-nitrides Information processing in Si/SiGe for biosensor applications 158 Seiten 215 Seiten ISBN 978-3-941650-21-3 ISBN 978-3-941650-28-2 Vol. 129 Vol. 136 Arne Laucht Christian Claus Jäger Semiconductor quantum optics with Polycrystalline Silicon Thin Films for tailored photonic nanostructures Electronic Applications 232 Seiten 202 Seiten ISBN 978-3-941650-29-9 ISBN 978-3-941650-36-7

Vol. 130 Vol. 137 Jia Chen Gunther Christian Jegert Compact laser-spectroscopic gas sensors Modeling of Leakage Currents in using Vertical-Cavity Surface-Emitting High-к Dielectrics Lasers 170 Seiten 150 Seiten ISBN 978-3-941650-37-4 ISBN 978-3-941650-30-5

Vol. 131 Hans Leonhard Prechtel Ultrafast Photocurrents and Terahertz Radiation in Gallium Arsenide and Carbon Based Nanostructures 160 Seiten ISBN 978-3-941650-31-2

Vol. 132 Roland Enzmann Technologieentwicklung für eine Einzelphotonenquelle 142 Seiten ISBN 978-3-941650-32-9

Vol. 133 Alexander Bachmann Antimonide-Based Vertical-Cavity Surface-Emitting Lasers 144 Seiten ISBN 978-3-941650-33-6

Vol. 134 Markus Andreas Mangold Two-Dimensional Gold Nanoparticle Arrays – A Platform for Molecular Optoelectronics 140 Seiten ISBN 978-3-941650-34-3

Vol. 135 Stefan Birner Modeling of semiconductor nanostructures and semiconductor–electrolyte interfaces 227 Seiten ISBN 978-3-941650-35-0