Castro Thesis

Modeling of Carbon Nanotube
Field-Effect Transistors
by
Leonardo de Camargo e Castro
B.A.Sc. (Electrical Engineering), The University of British Columbia, 2001
A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF
THE REQUIREMENTS FOR THE DEGREE OF
Doctor of Philosophy
in
The Faculty of Graduate Studies
(Electrical and Computer Engineering)
The University Of British Columbia

July 2006
c Leonardo de Camargo e Castro 2006
Abstract
In this thesis, models are presented for the design and analysis of carbon nanotube field-effect
transistors (CNFETs). Such transistors are being seriously considered for applications in the
emerging field of nanotechnology. Because of the small size of these devices, and the nearone-dimensional nature of charge transport within them, CNFET modeling demands a rigorous quantum-mechanical basis. This is achieved in this thesis by using the effective-mass
Schrodinger Equation (SE) to compute the electron and hole charges in the nanotube, and by
using the Landauer Equation to compute the drain current. A Schr
odinger-Poisson (SP) solver
is developed to arrive at a self-consistent potential distribution within the device. Normalization of the wavefunction in SE is achieved by equating the probability density current with the
current predicted by the Landauer Equation. The scattering matrix solution is employed to
compute the wavefunction, and an adaptive integration scheme to obtain the charge. Overall
convergence is sought via the Picard or Gummel iterative schemes. An AC small-signal circuit
model, employing the DC results from the SP solver, is also constructed to obtain estimates of
the high-frequency capabilities of the transistors.
The DC results predict the unusual ambipolar behaviour of CNFETs reported in the literature,
and explore the possibilities of using work-function engineering to tailor I-V characteristics for
different device applications. The model qualitatively agrees with some experimental results in
the literature, and gives confidence that the performance of coaxial devices, when they become
available, will be well predicted by the models. In the AC regime, it was found that under
somewhat ideal operating conditions the operating limit of these devices might just reach into
the 1-10 THz regime.
In addition to the development of rigorous modeling procedures for CNFETs, a preliminary
compact model is developed, in which some of the essence of the detailed model is distilled into
a set of simpler equations, which may prove useful in guiding device design towards CNFETs
for applications in nanoelectronics.
ii
Table of Contents
Abstract
ii
Table of Contents
iii
List of Figures
vi
List of Symbols
viii
Acknowledgments
ix
Co-Authorship Statement
Chapter 1. Introduction
1
1.1 Carbon Nanotubes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.1 Crystal Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.2 Electronic Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.1.3 Transport . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.2 Carbon Nanotube Field-Effect Transistors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.2.1 Planar Devices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.2.2 Coaxial Devices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.3 Modeling Coaxial CNFETs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.3.1 Electrostatics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.3.2 Charge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.3.3 Transport . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
1.4 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
1.5 Specific Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
Chapter 2. Towards a Compact Model for Schottky-Barrier CNFETs
36
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.2 The Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.3 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
Chapter 3. Electrostatics of Coaxial Schottky-Barrier CNFETs
44
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.2 Coaxial Nanotube Electrostatics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
iii
Table of Contents
Chapter 4. An Evaluation of CNFET DC Performance

58
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.2 Fabrication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.3 Theoretical Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.4 DC Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
4.5.1 Ambipolarity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.5.2 Conductance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.5.3 Subthreshold Slope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.5.4 ON Current . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.5.5 Transconductance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
Chapter 5. A Schr
odinger-Poisson Solver for Modeling CNFETs
75
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.2 Solution Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5.2.1 Potential . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5.2.2 Charge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
5.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
5.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
Chapter 6. An Improved Evaluation of the DC Performance of CNFETs
85
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
6.2 Correspondence of the Compact and Quantum Models . . . . . . . . . . . . . . . . . . . . . . . . . 86
6.3 Quantum-Mechanical Reflection for the Thermionic Case . . . . . . . . . . . . . . . . . . . . . . . 89
6.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
Chapter 7. Quantum Capacitance in Nanoscale Device Modeling
96
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
7.2 Equilibrium Quantum Capacitance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
7.2.1 Two Dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
7.2.2 One Dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
7.3 General Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
7.4 Application: CNFETs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
7.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
Chapter 8. Method for Predicting ft for CNFETs
110
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
8.2 The Small-Signal Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
8.2.1 Equivalent circuit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
8.2.2 Model parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
iv
Table of Contents
8.3
8.4
Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114

Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
Chapter 9. High-frequency Capability of Schottky-Barrier CNFETs

123
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
9.2 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
9.3 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
9.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
Chapter 10. Extrapolated fmax for CNFETs
129
10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
10.2 Modeling Procedures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
10.3 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
10.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
Chapter 11. Conclusion and Recommendations for Future Work
140
Appendix A. Complete Schr

odinger-Poisson Model
145
A.1 Sample Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
List of Figures
1.1
1.2
1.3
1.4
1.5
1.6
1.7
1.8
1.9
1.10
1.11
Hybridized carbon atom and graphene lattice structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

Carbon nanotube lattice structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
Graphene energy dispersion relation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
Nanotube energy dispersion relation, density of states, and carrier velocity . . . . . . . . . . 8
Nanotube properties for various tube radii . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
Examples of planar CNFETs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
Examples of electrolyte-gated CNFETs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
Coaxial CNFET structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
Solution to Laplaces equation in 2D cylindrical coordinates . . . . . . . . . . . . . . . . . . . . . . . 21
Electric field on nanotube surface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
CNFET band diagrams for Laplace solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.1
2.2
2.3
2.4
2.5
Model geometry: closed cylindrical structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Conduction energy band diagram for various bias conditions . . . . . . . . . . . . . . . . . . . . . . .
Electron distribution in the mid-length region of nanotube . . . . . . . . . . . . . . . . . . . . . . . . .
Drain current-voltage characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Comparison of drain I-V characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
37
40
40
41
41
3.1
3.2
3.3
3.4
3.5
3.6
Equilibrium energy band diagram: (16,0) tube, tins = 5.6 nm and ins = 3.9 . . . . . . . .
Equilibrium energy band diagram: (16,0) tube, tins = 30.9 nm and ins = 3.9 . . . . . . .
Equilibrium energy band diagram: (16,0) tube, tins = 5.6 nm and ins = 19.5 . . . . . . .
Sub-threshold current: (16,0) tube, tins = 5.6 nm and ins = 3.9 . . . . . . . . . . . . . . . . . . . .
Non-equilibrium energy band diagram: (16,0) tube, tins = 5.6 nm and ins = 3.9 . . . .
Drain I-V characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
48
49
50
52
53
54
4.1
4.2
4.3
4.4
4.5
4.6
4.7
4.8
Coaxial CNFET model geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Band diagram illustrating CNFET ambipolarity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
ID -VGS for various contact work functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Band diagrams for various contact work functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Ratio of equilibrium quantum capacitance to insulator capacitance . . . . . . . . . . . . . . . . .
ID and (b) gm as a function of gate-source voltage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
ID -VDS for various contact work functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Band diagrams for various gate work functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
61
65
66
66
67
68
69
70
5.1
5.2
5.3
5.4
CNFET net carrier density as a function of position and VDS . . . . . . . . . . . . . . . . . . . . . .

Conduction band edges for VDS = 0 and 0.4 V . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Conduction band edges and transmission probabilities for electrons . . . . . . . . . . . . . . . .
Energy distribution of mid-channel, source-originated electron concentration . . . . . . .
80
80
81
82
6.1
6.2
6.3
6.4
Drain current versus gate-source voltage for various models . . . . . . . . . . . . . . . . . . . . . . . .

Drain current and transconductance for various models . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Transmission probabilities of source-injected electrons . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Drain current versus drain-source voltage for various models . . . . . . . . . . . . . . . . . . . . . . .
91
92
92
93
vi
List of Figures
7.1
7.2
7.3
7.4
7.5
Equilibrium 1D quantum capacitance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Non-equilibrium 1D quantum capacitance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Energy band diagram comparison for Laplace and Schr
odinger-Poisson solutions . .
CNFET quantum capacitance as a function of gate- and drain-source voltages . . . .
CNFET transconductance and its constituent elements . . . . . . . . . . . . . . . . . . . . . . . . . . .
101
102
103
105
106
8.1
8.2
8.3
8.4
8.5
8.6
Small-signal equivalent circuit for ft calculation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Charge supply to and through the source and drain electrodes . . . . . . . . . . . . . . . . . . . .
Charge density in CNFET as a function of position and energy . . . . . . . . . . . . . . . . . . .
Components of the source and drain capacitances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
ft and its components. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Bias and temperature dependences of capacitance and transconductance . . . . . . . . . .
112
113
115
116
117
118
9.1
9.2
9.3
9.4
9.5
Structure of the modeled CNFET . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Small-signal equivalent circuit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Capacitances and transconductance for the model device . . . . . . . . . . . . . . . . . . . . . . . . .
Charge density versus energy and position . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Extrapolated figures of merit for various contact resistances . . . . . . . . . . . . . . . . . . . . . .
124
125
126
126
127
10.1 Unilateral power-gain for Device 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

10.2 eff estimates for Device 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.3 fmax estimates for Device 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.4 Error in fmax prediction for Device 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.5 fmax for improved Device 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
134
135
136
137
137
A.1
A.2
A.3
A.4
A.5
147
148
149
154
155
Simulation space for CNFET with metal contacts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Complex bands in semiconducting nanotube . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Dispersion relation and density of states in 1D metal contact . . . . . . . . . . . . . . . . . . . . .
Agreement with experimental data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Agreement with atomistic simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
vii
List of Symbols
k
E
V
m
T
z
Eg
Ec
Ev
EF ,
E0
Evac
Lt
Lu
Lg
Lc
Rt
Rg
tins
tg
tc
t
ins
Q
n, p
Cins
CQ
f
T
g
G
L
H
q
kB
h
0
Wavevector
Energy
Electric potential
Carrier effective mass
Temperature
Position along nanotube axis of symmetry
Radial distance from nanotube axis of symmetry
Electron affinity of semiconducting nanotube
Workfunction
Energy bandgap
Conduction band edge
Valence band edge
Fermi energy level
Charge neutrality level
Distance between sub-band bottom and charge neutrality level
Vacuum energy level
Nanotube length
Gate underlap
Gate length
Source/drain contact length (when assumed equal)
Nanotube radius
Cylindrical gate radius
Insulator thickness
Gate thickness
Source/drain contact thickness (when assumed equal)
Nanotube relative permittivity
Insulator relative permittivity
Charge, charge density
Electron and hole carrier densities
Insulator capacitance
Quantum capacitance (also known as semiconductor capacitance)
Wavefunction
Fermi function
Transmission function
Density of states
Local density of states
Linear operator in Poissons equation
Hamiltonian operator in Schr
odingers equation
Magnitude of electronic charge
Boltzmanns constant
h
is sometimes labeled Diracs constant)
Plancks constant (~ = 2
Permittivity of free space
viii
Acknowledgments
I would like to acknowledge my colleagues Dylan McGuire and Sasa Ristic, and the members
of my supervisory committee Anthony Peirce and John Madden, for many fruitful discussions.
I would also like to thank David John of the UBC Nanoelectronics Group as our collaboration
was invaluable in my learning about and understanding carbon nanotube transistors.
It is with great pleasure that I thank my mentor Dave Pulfrey, who provided me with many
opportunities over the last few years of graduate studies, was always a source of inspiration,
support and experience, and turned out to be much more than just a thesis supervisor.
Finally, this thesis is dedicated to my family, Erika, Fernando and Olavo, as it could not have
been realized without their unwavering support and encouragement.
ix
Co-Authorship Statement
The core chapters of this thesis are publications based on work performed with other members
of the UBC Nanoelectronics Group during the period 2002-2005. This statement acknowledges
the role of this thesis author in their creation. Most of the derivations in Chap. 3 and Chap. 7
were the work of David L. John. Chap. 5 and Chap. 6 had equal contributions from this author
and David L. John. The author of this thesis was responsible for all the remaining research
presented herein, with the exception of a few equations present in Chap. 4 and Chap. 8, which
are due to co-authors. As far as manuscript preparation is concerned, all the publications
included here were collaborative efforts by their co-authors, and the author of this thesis had
a major role in the production of all but Chap. 3 and Chap. 7, wherein he was involved mostly
in editing. Nearly all the data and figures in this thesis were prepared by its author.
Chapter 1
Introduction
The last few years witnessed a dramatic increase in nanotechnology research. Among others,
one of the most exciting fields to emerge is nanoelectronics, where a myriad of possibilities are
appearing in the form of sensors, actuators, and transistors, each characterized by feature sizes
of the order of a few nanometres.
All this innovation has been fueled by the discovery of new materials and the invention of manufacturing methods that allow design and development at such a minute scale. Carbon nanotubes
are at the forefront of these new materials, due to the unique mechanical and electronic properties that give them, for example, exceptional strength and conductivity. One exciting possibility
is the creation of nanometre-scale transistors, perhaps to be embedded, in the future, inside
complex and miniscule electronic circuits that will make todays chips seem enormous in comparison. Moreover, these nanotubes exhibit a tremendous current-carrying ability, potentially
allowing for increased miniaturization of high-speed and high-power circuits. Although some
devices have already been produced, the technology is still in its infancy when compared to, for
instance, that of bulk-silicon MOSFETs.
This thesis is a report on studies performed during the period 2002-2005 with the UBC Nanoelectronics Group, with the aim of understanding and obtaining performance predictions for
carbon nanotube field-effect transistors (CNFETs). During the time this research was being
conducted, there were few published works illustrating certain phenomena predicted herein.
During the course of this work, and the writing of this manuscript, several researchers from
many institutions, both private and academic, reported results from experimental devices that
1
could be explained by the models presented herein.

The remainder of this chapter will cover some of the CN properties that are relevant to this
work on transistor modeling, will introduce the Schottky-barrier carbon nanotube field-effect
transistor (SB-CNFET), and will describe the problem to be modeled. Finally, an outline of
the thesis will be presented.
1.1 Carbon Nanotubes

Carbon nanotube (CN) molecules have a cylindrical structure and are formed by one or more
concentric, crystalline layers of carbon atoms. These atoms are assembled in hexagonal-lattice
graphene sheets, which are rolled up into seamless tubes and named according to the number
of concentric sheets as being either multi- or single-wall nanotubes. Both kinds were originally
observed experimentally, via transmission electron microscopy, decades ago and work in the
field has subsequently increased dramatically [1, 2].
These molecules exhibit unique physical properties and, while this thesis mostly focuses on employing their electronic characteristics for nanoelectronic applications, it is important to note
that CNs are being studied in a variety of fields that make use of other properties. Nanotubes
appear to be paving the way for myriad possibilities in the growing nanotechnology and emerging biotechnology industries, particularly in nanoelectronics in the form of sensors, transistors,
and nanowires. An in-depth description of CNs and their properties may be found, for example,
in Ref. [3] and Ref. [4]. It is also noteworthy that carbon is not the only chemical element to
form nanotubesfor instance, inorganic BN, WS2 , and V2 O5 nanotubes have been reported in
the literature [5].
Since we will be dealing, in subsequent chapters, with carrier density and transport in carbon
nanotubes, it is necessary to understand the basic electronic properties of the material. In
particular, we are concerned with the energy dispersion (-k) relation, density of states (DOS)
calculations, and scattering mechanisms.
1.1.1 Crystal Structure

Carbon is found in nature most commonly as graphene or diamond, which are crystal formations
of covalently-bonded carbon atoms. In its ground state, carbon has an electron configuration
of 1s2 2s2 2p2 (6 electrons in 3 orbitals). Covalent bonding occurs via hybridization of the two
outermost shells (4 electrons), and in graphene this takes shape via sp2 orbitals, as illustrated
in Fig. 1.1A.
(B)
(A)
Figure 1.1: Pictorial representation of (A) sp2 -hybridized carbon and (B) graphene lattice
structure. Note that the orbitals are for illustration only, and are neither rigorously-derived
probability densities nor drawn to scale.
A carbon atom in graphene will assemble in a single-sheet hexagonal lattice resembling the
surface of a honeycomb, as illustrated in Fig. 1.1B. This is also known as a trigonal-planar
-bond framework, with an inter-atomic spacing, acc , of approximately 1.42
A along the bonds
that are separated by 120 degrees. The 2p electrons from all the atoms on the sheet constitute
a cloud of delocalized -orbitals surrounding the carbon cores, and these valence electrons,
once excited, are responsible for conduction in graphene. Note that in Fig. 1.1B, the delocalized orbitals are illustrated as individual lobes connected by hexagonal rings above and below
the sheeta proper derivation of probability clouds would show that in reality the electrons
form thicker, donut-shaped rings above the hexagonal lattice, and that adjacent donuts merge,
thereby allowing electrons to move about the entire two-dimensional plane. Furthermore, multiple sheets of graphene may assemble in stacks, whereby two adjacent sheets are held together
weakly by dispersion forces1 and have an inter-layer spacing of about 3.35

A. While the strength
of the -bonds is responsible for some incredible mechanical properties of carbon nanotubes,
the weak dispersion forces are the reason sheets of graphene readily slide off each other, giving
the softness of graphite in pencils.
air
ch
arm
chiral
(n1,n2)=(4,1)
zigzag
(A)
(B)
Figure 1.2: Pictorial representation of (A) unrolled and (B) rolled carbon nanotube lattice
structures. The latter shows a (16,0) tube.
Single-wall nanotubes are characterized by a wrapping (or chiral) vector Ch = n1 a

1 + n2 a
2 ,
where [
a1 , a
2 ] are lattice unit vectors separated by 60 degrees and the indices (n1 , n2 ) are
positive integers (0 n2 n1 ) that specify the chirality of the tube [6], as shown in Fig. 1.2A.
The chiral vector begins and ends at equivalent lattice points, so that the particular (n1 , n2 )
tube is formed by rolling up the vector so that its head and tail join, forming a ring around
the tube. The length of Ch is thus the circumference of the tube, and the radius is given by
p
the formula Rt = |Ch |/(2) = acc 3(n21 + n22 + n1 n2 )/(2), where, for example, Rt 0.63 nm
for a (16,0) nanotube (see Fig. 1.2B). The smallest possible radius, a limit imposed by the
requirement that the energy of the system in tube form be lower than that of the unrolled
graphene equivalent, is thought to be 2
A [7], whereas the upper limit on radius is in the
several nanometre range. Depending on their (n1 , n2 ) indices, nanotubes are placed in one of
three groups, which are named according to the shape of the cross-section established by the
1
Dispersion forces, also known as London forces, are intermolecular attractive forces caused by instantaneous dipoles created by electron motion about the nuclei; they are also a subset of van der Waals
forces.
chiral vector slicing across the hexagonal pattern: armchair (n1 = n2 and = 30 ), zig-zag
(n2 = 0 and = 0 ), and chiral (all other cases), where is the angle between Ch and a
1 [6].
The indices also serve to quickly identify the conduction properties of a nanotubewhen (n1
n2 ) is a multiple of 3, the nanotube is metallic, otherwise it is semiconducting.
Finally, to appreciate the size of these molecules, it is convenient to keep in mind the number
of atoms composing a given tube. The number of atoms per nanometre-length on a single-wall
nanotube can be estimated by the formula
Natoms 2
Acyl 1
8Rt Lt 1
=
240Rt ,
Ahex Lt
3 3a2cc Lt
where A denotes area, and Rt and Lt are, respectively, the tube radius and length in nanometres.
Since a typical tube used in the devices examined in this thesis will have dimensions of Rt
0.63 nm, we expect to have an atom density of roughly 150 nm1 contributing to the conducting
cloud.
1.1.2 Electronic Properties

Owing to the small diameter of carbon nanotubes, it is necessary to account for the quantization of wavevectors in the circumferential direction. Moreover, the thinness of the nanotubes
cylindrical shell obviously yields an even shorter length of confinement in the radial direction,
thus making the material virtually one-dimensional as far as electron transport is concerned.
Many published works to date have corroborated this claim with experimental evidence from
device studies [8].
In order to examine the band structure and conductivity properties of the nanotube, it is necessary to derive its -k relation. This is done by starting from the equivalent relation of a twodimensional graphene lattice (a function of two wavevectors), and introducing a quantization
of wavevectors in the direction of the chiral vector Ch via the imposition of a periodic boundary condition. A detailed derivation of the energy dispersion relations using a tight-binding
approximation is presented in Ref. [6], and we only outline its major points here. It should
be noted that this approach ceases to be valid at the onset of curvature effects, i.e., for tubes
5
of radius smaller than 0.5 nm [7, 9, 10]. These effects are a manifestation of re-hybridization
(mixing) of the and orbitals due to their proximity in small-diameter tubes. This impacts
the determination of bandgap (and thus the conduction properties) and density of states, and,
as such, the tubes in this work are kept to radii that try to avoid significant contribution of
these second-order effects.
The dispersion relation for graphene, obtained by the Slater-Koster tight-binding scheme, considering only -orbitals, and following the lattice vector conventions given in Ref. [6] is:
"
Egraphene (kx , ky ) = t 1 + 4 cos
3kx acc
2
!
!#1/2
3ky acc
3ky acc
2
cos
+ 4 cos
, (1.1)
2
2
where the positive and negative terms correspond to the symmetrical bonding and anti-bonding
energy bands, respectively, the ks are wavevectors, and t is the transfer integral (or nearestneighbour parameter), the value of which is typically taken to have magnitude 2.8 eV [11]. A
plot of Eq. (1.1), representing a surface of allowed energies for the two-dimensional wavevector,
is illustrated in Fig. 1.3, where the high-symmetry points are indicated by capital letters (K, M,
and ). The K-points are degenerate near the Fermi energy (E = 0 in the plot) and indicate the
zero-gap characteristic of conducting graphene. Near these K-points, as k 0, the dispersion
relation is approximately cone-shaped and from the constant slope the Fermi velocity is given
by [6]
vF =
3
acc |t| .
2~
We now seek an expression for the nanotube dispersion relation, which is obtained by taking
slices of the surface above, each cut being determined by the circumferential quantization.
Again following the lattice vector definitions in Ref. [6], we switch the wavevector notation
(kx , ky ) (kz , p), where the subscript z denotes the direction of transport, and p is an integer
representing the quantized modes obtained by imposing a periodic boundary condition in the
circumferential (perpendicular to transport) direction. The nanotube dispersion relation is then
given by
1/2
Et (kz , p) = t 1 + 4 cos 1 cos 2 + 4 cos2 2
,
Figure 1.3: Energy dispersion relation for graphene, from nearest-neighbour tight-binding calculation, using Slater-Koster approximation. , M, and K are high-symmetry points, where
the K-points lie on the plane of E = 0 (the Fermi level).
where the cosine arguments are given by
3acc n2 n1
kz +
4
3acc n2 + n1
2 =
kz +
4
1 =
n1 + 3n2
p
2
2
3n1 n2
p,
2
2
in which 2 = n21 + n22 + n1 n2 , < (3acc /dg )kz < +, p = 0...(2/dg 1), and dg =
gcd(2n1 n2 , 2n2 n1 ).
The one-dimensional density of states, g(E), is obtained via a summation over all wavevectors.
This summation can be split between transverse and longitudinal components, and we convert
the latter summation into an integral over energy [12, p.52] (note that quantization due to the
tube length is neglected here):
X Z
Z
XX
X
2
kz
1
dE ,
g(E) =
dkz =
(2) =
2
2
Et
p
p
p
(1.2)
kz
where the factors of 2 in the numerator account for spin degeneracy and include k-states
(Et /k is an even function). Fig. 1.4 shows the band structure and density of states for the
lowest, doubly-degenerate [6], bands of two different nanotubes. Note also that a simpler alternative to the tight-binding calculations was derived analytically, under some approximations,
by Mintmire et al. [13] and allows one to calculate the density of states with a good fit of the
7
bandgap for most cases, but a progressively worse fit to the DOS as one moves up in energy
and includes more bands.
Energy (eV)
1.5
1
0.5
0
0
Wavevector
Density of States
5
10
5
Velocity (m/s) x 10
Figure 1.4: Energy dispersion relation, density of states, and carrier velocity plots for: (16,0)
(solid) and (10,0) (dotted lines) carbon nanotubes, illustrating the lowest, doubly-degenerate [6],
bands. The energies are referenced to the Fermi level (E = 0) and are thus symmetrical about
the x-axis. The carrier velocity is computed from v(E) = 1/~(E/k).
Knowledge of the carbon nanotube dispersion relation allows one to identify whether a tube
produces metallic or semiconducting behaviour, and in the latter case, determine the conduction
and valence band edges and thus the bandgap. Typical values for the bandgap are in the range
of tenths of an electron-volt to a few electron-volts, and for the previously mentioned (16,0) tube
it is 0.62 eV (see Fig. 1.4). For transistors, it is desirable to have a bandgap large enough to
suppress excessive thermal generation of carriers, i.e., beyond what the gate is able to control.
However, as will be shown later, a smaller bandgap allows for greater carrier densities on the
tube under certain conditions, and thus higher currents. It is noteworthy that the bandgap can
be found to fit the expression [6]:
Eg =
|t|acc
.
2Rt
The bandgap is thus a geometrically-tunable property, and given that we can make devices by
choosing nanotubes by their approximate radius (presently via scanning tunneling microscopy),
we may be able to exploit this tunability in nanoelectronics. The bandgap and other parameters
of some tubes are illustrated in Fig. 1.5. Also shown is the intrinsic electron concentration, ni ,
typically employed in describing bulk semiconductors; although it may not be as useful of a
8
measure in quantum devices involving quasi-1D carbon nanotubes, it is worthwhile mentioning.
(a)
0.6
0.4
0.2
10
0.5
1
1.5
Nanotube radius (nm)
Effective mass (m0)
Energy gap (eV)
0.8
(16,0)
log (n ) (m1)
0.1
(b)
0.08
0.06
0.04
0.02
0
0.5
1
1.5
2
Nanotube radius (nm)
Figure 1.5: Nanotube properties for various tube radii (larger chiral numbers correspond to
larger radii): (a) intrinsic electron concentration (at T =300 K) and energy gap; (b) carrier
effective mass of lowest subband, normalized to the free electron mass m0 = 9.11 1031 kg.
During the early stages of research in nanotube transistor fabrication, the question of doping was
one of much controversy. Although many experimental works have shown nanotubes to behave
as p-type semiconductors, i.e., with holes as majority carriers, it is now thought that such
behaviour is caused by inadvertent doping of the contacts or surrounding materials. However,
a more recent theoretical explanation for the effect of hole self-doping has been proposed by
Rakitin et al., where it is claimed that - hybridization due to curvature in small-radius
tubes causes a charge transfer from the portion of the -orbitals lying inside the tube to the
-orbitals lying in-plane with the nanotube surface [14]. However, given that this study has not
been corroborated experimentally, and that effects from external dopants (e.g., due to charge
trapped in oxide layers) and contacts are more significant, herein we treat all tubes as being
intrinsic.
Finally, we deal with the elusive issue of obtaining the work function and electron affinity of
carbon nanotubes. Given the uncertainty in determining these parameters, we have employed
a value of 4.5 eV in most chapters in this thesis, in agreement with Ref. [15]. As will be
made clear throughout this work, since both the nanotube and metal-contact work functions
9
are difficult to measure, they are simply taken as inputs to the model. Still, here we note
that subsequent to performing the work in the aforementioned chapters, we became aware of
experimental results published by other research groups: Suzuki et al. found the work function
to be 4.8 eV, which is about 0.1-0.2 eV larger than graphene [16], and is in agreement with
measurements by Chen et al. [17]; Kazaoui et al. gave an experiment-based estimate for the
electron affinity and first ionization energy to be, respectively, 4.8 eV and 5.4 eV [18]. Moreover,
a recent work by Chen et al. on carbon nanotube transistors correlated the Schottky barrier
height at the metal-nanotube interfaces with the tube diameter [19], and from this we inferred
values for the work functions in Chap. 10.
1.1.3 Transport
From early carbon nanotube experiments, owing to their molecular uniformity and quasi-onedimensional nature, it was expected that they would exhibit ballistic transport properties,
and early theoretical predictions stated they would be ballistic for most radii encountered in
experiments [20]. Research of such characteristics delves into the realm of mesoscopic systems,
those being materials or devices that are small enough (between 1 and 100 nm) to have their
behaviour governed by interactions at the scale of the electron wavefunction, yet large enough
that we need not deal with microscopic (at the atomic scale) phenomena. Carbon nanotubes
(even samples that are a few microns in length), modern short-channel MOSFETs with twodimensional electron gas (2DEG) inversion channels, and quantum-well lasers are all examples
of devices that may be labeled as mesoscopic. Such systems have been the focus of much study
in recent history, driven by an attempt to bridge the knowledge gap between bulk semiconductor
technologies and nanoscale devices. The reader may refer to the material in Refs. [12, 21] for
in-depth information on the subject. Also, it must be noted that although CNs exhibit lowtemperature transport effects such as Coulomb blockade, these are not dealt with in this work
as they are not pertinent to room-temperature transistor operation.
A mesoscopic system may be subject to a variety of scattering mechanisms. Electron-electron
interactions have been studied in metallic carbon nanotubes, showing that a transition from a
10
Fermi Liquid to a Luttinger Liquid2 , as expected for a one-dimensional conductor, can play a
role in the transport properties of CNs [22]. However, further research is still required to determine experimentally their role in nanotube devices. Backscattering due to electron-phonon
interactions is another phenomenon that has been demonstrated in single-wall carbon nanotube
samples at biases of several volts [23]. Still, this is only manifested in relatively low-energy electrons, and recent experimental CN-based transistor studies have indicated that no backscattering occurs for operation under a bias of about one volt and device length of several-hundred
nanometres [24].
A variety of studies have also reported values for mobility, mostly derived from conductance
experiments in transistors, typically in the range of 103 104 cm2 /Vs [25, 26]. Theoretical
predictions, accounting for electron-phonon scattering and using Monte Carlo techniques for
electron transport simulations, have also yielded a mobility of 104 cm2 /Vs in semiconducting
tubes of radii up to 2 nm [27]. Yet another report yielded conductance measurements indicating that metallic nanotubes are indeed ballistic conductors up to several microns in length,
while semiconducting ones have strong barriers along the tube impeding transport [28]. Further
studies have focused on the dependence of transport properties on mechanical deformations and
defects [2931].
Recent experiments with CN transistors, devised for the purpose of studying transport properties, have concluded that for these devices transport appears to be ballistic in nature [8, 24].
As such, we assume in this work that, under the simulated conditions, inelastic scattering processes are negligible and thus we are dealing with an effectively ballistic transport regime. In
this situation, we adopt the single-particle approach to transport modeling, which is based on
the Landauer-B
uttiker formalism [21], and still allow for elastic scattering from macroscopic
potentials.
As a final note, we point out that the current-carrying capacity of multi-wall nanotubes has
2
A Fermi liquid is a population of electrons whose interactions do not significantly alter their energy
distribution near the Fermi level, and thus they are governed by the Fermi function. A Luttinger Liquid
is one in which electron-electron interactions give rise to exotic properties.
11
been demonstrated to be more than 109 A/cm2 , without degradation (such as that due to
electromigration) after several weeks or during operation well above room-temperature [32].
This is a promising characteristic for fabricating devices designed for high-power circuits. In a
device context, subsequent chapters will show that a CNFET with a 1 nm-diameter single-wall
nanotube can reach currents of at least 10 A, which normalized to its circular cross-section
corresponds to a current density of order 108 A/cm2 .
1.2 Carbon Nanotube Field-Effect Transistors

Following the discussion on the properties of carbon nanotubes, we now give an overview of
an important application and the topic of this thesis: the carbon nanotube field-effect transistor. This three-terminal device consists of a semiconducting nanotube bridging two contacts
(source and drain) and acting as a carrier channel, which is turned on or off electrostatically
via the third contact (gate). Presently, there are various groups pursuing the fabrication of
such devices in several variations, achieving increasing success in pushing performance limits,
while encountering myriad problems, as expected for any technology in its infancy. While the
ease of manufacturing has improved significantly since their first conception in 1998, CNFETs
still have a long way to go before large-scale integration and commercial use become viable.
Furthermore, as these transistors evolve at every research step, the specifics of their workings
become clearer, and given that the aim of this thesis is to present a working model of CNFETs,
it is reassuring to see some of the findings presented herein being proven by recently released
experimental data.
As regards the CNFETs principle of operation, we briefly introduce two distinct methods
by which the behaviour of these devices can be explained. Primarily, the typical CNFET
is a Schottky-barrier device, i.e., one whose performance is determined by contact resistance
rather than channel conductance, owing to the presence of tunneling barriers at both or either
of the source and drain contacts. These barriers occur due to Fermi-level alignment at the
metal-semiconductor junction, and are further modulated by any band bending imposed by the
gate electrostatics. Moreover, in some devices, the work-function-induced barriers at the end
12
contacts can be made virtually transparent either by selecting an appropriate metallization or by

electrostatically forcing via a separate virtual-gate terminal (see, for example, Ref. [24]). These
devices, sometimes labeled as bulk-modulated transistors, operate differently in that a thicker
(non-tunneling) barrier, between the source contact and the mid-length region of the device, is
modulated by the gate-source voltage. This operation is akin to that of a ballistic MOSFET,
and effectively amounts to a channel modulation, by the gate, of a barrier to thermionicallyemitted carriers, injected ballistically from the end contacts.
We now provide a brief description of typical CNFET geometries, which are grouped in two
major categories, planar and coaxial. The specifics of nanotube growth and transistor fabrication issues, albeit of tremendous importance for this emerging field of nanoscale transistors,
are beyond the scope of this work. The reader may refer to numerous journal papers on the
subject for more information, or for a fairly current summary, to Refs. [33, 34].
1.2.1 Planar Devices

Planar CNFETs constitute the majority of devices fabricated to date, mostly due to their
relative simplicity and moderate compatibility with existing manufacturing technologies. The
nanotube and the metallic source/drain contacts are arranged on an insulated substrate, with
either the nanotube being draped over the pre-patterned contacts, or with the contacts being
patterned over the nanotube. In the latter case, the nanotubes are usually dispersed in a solution
and transferred to a substrate containing pre-arranged electrodes; transistors are formed by trial
and error. Manipulation of an individual nanotube has also been achieved by using the tip of
an atomic force microscope (AFM) to nudge it around the substrate; due to its strong, but
flexible, covalent bonds, this is possible to do without damaging the molecule. In the case
where the electrodes are placed over the tube, manipulation of the CN is not required and
alignment markers, pre-arranged on the substrate, allow accurate positioning of the contacts
once the nanotube is located via examination by a scanning tunneling microscope (STM). The
gate electrode is almost always on the back side of the insulated substrate, or alternatively is
patterned on top of an oxide-covered nanotube.
13
Figure 1.6: Examples of planar CNFETs: (A) Ref. [15], (B) Ref. [35], (C) Ref. [36], (D) Ref. [8],
and (E) Ref. [24].
The first CNFET devices were reported in 1998, and involved the simplest possible fabrication. They consisted of highly-doped Si back gates, coated with thick SiO2 , and patterned
source/drain metal contacts, either using Au or Pt, as shown in Fig. 1.6A [15, 37]. Experimentations with different metals such as Ti, Ni, Al, and Pd have since been carried out by several
groups, primarily to manipulate the work function difference between the end contacts and the
nanotube3 . Subsequent work also produced a device that replaced the back gate with an electrode placed over the substrate, perpendicular to the source and drain contacts, as illustrated
in Fig. 1.6B [35]. Here, the nanotube was separated from this gate electrode by a thin insulating
layer of Al2 O3 , with the source/drain electrode strips placed over the tube ends for reduced
3
Recent ab initio theoretical studies comparing the interfaces between different bulk metals and metallic
nanotubes, studying both end- and side-contacted tubes, concluded that Ti contacts yield superior
conductance over their Au and Al counterparts [38, 39].
14
contact resistance.
Fig. 1.6C shows a further improvement in CNFETs through the placement of the gate electrode
over the semiconducting nanotube, thus improving the channel electrostatics via the thin gate
oxide [36,40]. Moreover, the Ti source/drain metalizations in this device form titanium carbide
abrupt junctions with the nanotube, yielding increased conductance [41]. Another attempt
to obtain better gate electrostatics involved materials with high dielectric constants, such as
zirconia (ZnO2 ) and hafnia (HfO2 ), being used as gate insulators [42]. Fig. 1.6D illustrates a
device built with Pd source/drain contacts in order to exploit the sensitivity of this materials
work function to hydrogen [8]. A multi-gate device, as shown in Fig. 1.6E, has recently been
reported, whereby parallel top gates are used to independently control the electrostatics of
different sections of the channel, thus facilitating a study of the transport characteristics of the
nanotube channel [24]. Most recently, a device with excellent DC characteristics was fabricated
with Pd end contacts, Al gate, and hafnia for the insulator [43].
1.2.2 Coaxial Devices

Although yet to be fabricated in its ideal form, coaxial devices are of special interest because
their geometry allows for better electrostatics than their planar counterparts. Capitalizing on
the inherent cylindrical shape of nanotubes, these devices would exhibit wrap-around gates that
maximize capacitive coupling between the gate electrode and the nanotube channel.
Figure 1.7: Examples of electrolyte-gated CNFETs: (A) Ref. [44], (B) Ref. [45].
Presently, the closest approximation to this geometry has been the development of electrolytegated devices. Kruger et al. reported the first such device, shown in Fig. 1.7A, using a multi-wall
15
nanotube for the channel [44]. Two gates can be activated: a highly-doped Si back gate similar
to planar devices; and an electrolyte gate, formed by a droplet of LiClO4 electrolyte contacted
by a thin platinum wire. Fig. 1.7B illustrates an improved version of this device, this time
using single-wall carbon nanotubes and NaCl for the electrolyte, and yielding current-voltage
characteristics that match those of modern Si MOSFETs [45].
Alternative structures for CN devices that place the tube vertically with respect to the substrate have already been used for field-emission applications.Coaxial CNFETs could perhaps
be fashioned by placing nanotubes inside the cavities of a porous material such as alumina,
surrounding them by an electrolyte solution for gating of individual devices.
Carbon nanotube transistors are not, however, the only devices in which an increased channel
coupling is being sought. Other Si technologies, such as the FinFET and the tri-gate MOSFET
are presently attempting to do this, and wrap-gated InAs-nanowire transistors have already
been successfully prototyped [46].
1.3 Modeling Coaxial CNFETs

In this section we will describe the problem being modeled, creating the framework for the
CNFET models presented in later chapters. To begin with, modeling of the CNFET requires
an understanding of the electrostatics in the device, i.e., the relationship between the potential
and charge therein. Furthermore, appropriate treatment of carrier transport in the nanotube
is necessary for the determination of the current-voltage characteristics. Evidently due to the
nanometre scale dealt with here, we must look at quantum phenomena and their influence on
device performance. While this work does not delve into all quantum physical issues, it attempts
to balance the incorporation of key phenomena with some simplicity of implementation. This is
done primarily because quantum phenomena in carbon nanotube transistors are still a freshly
debated topic, but also for the sake of avoiding elaborate ab initio calculations, achieving
reasonable simulation times, and arriving at some conclusions regarding performance trends.
As in other field-effect transistors, the CNFET relies on one of its three terminals, the gate,
16
to modulate the carrier concentration in the device channel by applying a field perpendicular
to the charge flow between the other two contacts, the source and the drain. In a typical
MOSFET, for example, the gate lies squarely on top of the substrate, wherein a very thin layer
of mobile charge is induced by the gate contact, via capacitive coupling through a thin oxide.
The volume occupied by this charge then constitutes the channel, enabling charge flow from the
source terminal to the drain terminal. A CNFET, whether planar or coaxial, relies on similar
principles, while being governed by additional phenomena such as 1D density of states (DOS)
and ballistic transport, which we have already presented in Sect. 1.1 and must now deal with
in the device model context.
The coaxial geometry maximizes the capacitive coupling between the gate electrode and the
nanotube surface, thereby inducing more channel charge at a given bias than other geometries.
This improved coupling is desirable in mitigating the short-channel effects that plague technologies like CMOS as they downsize device features. It is also of importance to low-voltage
applicationsa dominating trend in the semiconductor industryand to allow, potentially, for
easier integration with modern implementations of existing technologies such as CMOS. Besides
the wraparound gate, special attention must also be paid to the geometry of the end contacts,
since these play a role in determining the dimensions of the Schottky barriers that are present
in the channel near the device ends and have a direct effect on current modulation. We hereafter deal specifically with the coaxial geometry of the CNFET shown in Fig. 1.8, but we note
that most concepts and results discussed in this work are still transferable to planar devices,
at least in a qualitative sense. The key device dimensions are: the gate inner radius, Rg , and
thickness, tg ; the nanotube radius, Rt , and length Lt ; the insulator thickness tins = Rg Rt ; the
end-contact radius, tc (the source and drain may sometimes be of different sizes), and length,
Lc ; and the gate-underlap Lu . Note that in the following section, and in Chapters 2 to 7
a closed-cylinder structure was employed for simplicity in treating the electrostatics, wherein
Lc = tg = Lu = 0, tc = Rg and Lg = Lt .
17
Figure 1.8: Coaxial CNFET structure. The insulator fills the entire simulation space not
occupied by metal or the nanotube.
1.3.1 Electrostatics
In the system of Fig. 1.8, we solve Poissons equation, restricted to just two polar-coordinate
dimensions due to the device symmetry in the azimuthal direction, as given by
2V
2V
1 1 V
Qv
+
+
,
+
=
2

z 2
(1.3)
where V (, z) is the potential within the device cylinder, Qv is the volumetric charge density,
is the permittivity, and the / term allows for continuous changes in the dielectric constant.
In the case of abrupt dielectric heterointerfaces, as outlined in subsequent chapters, a jump
condition on the electric flux is employed. Under certain conditions, Poissons equation can be
solved analytically using, for example, a Greens function approach [47], or more generally and
with limited accuracy using, for example, a finite-difference numerical algorithm.
The Dirichlet boundary conditions presented by the three terminals, referenced to the source
potential, are given by
VS = S /q
VD = VDS D /q
VG = VGS G /q ,
18
where the terms represent the work function of each electrode, and VGS and VDS are, respectively, the gate- and drain-source voltages. These conditions hold in the absence of Fermi-level
pinning [48]. We also do not include any series resistance in the contacts4 , i.e., the voltages
VGS and VDS are measured at the contact surfaces in Fig. 1.8 and not in the external circuit. The remaining (open) boundaries of the system are assigned a null-Neumann condition;
this assumption is valid with appropriate choices of geometry, which have been ensured in the
simulations herein [49, 50].
Aside from the electric potential boundary conditions imposed by the contacts and required for
solving Poissons equation inside the closed cylindrical gate, we must also deal with discontinuities in the electrical properties at each interface. The CNFET is composed of at least three
different materials: metal contacts, semiconducting nanotube, and insulating oxide. At the
interface between any two different materials there will likely be a discontinuity in one or more
of the properties, thus affecting the potential profile in the device. Alternative prototype devices, such as the electrolyte-gated CNFET, may introduce different types of heterointerfaces,
but the device performance would nonetheless be sensitive to the aforementioned or similar
discontinuities.
The major material properties that need to be considered at the heterointerfaces are electron affinity, work function, and relative permittivity. In the coaxial device, the interfaces
we are concerned with are grouped in three categories: (1) metal-semiconductor (source-tube,
drain-tube); (2) metal-insulator (source-oxide, drain-oxide, gate-oxide); and (3) semiconductorinsulator (tube-oxide).
In a bias-free, intrinsic-nanotube CNFET, the first category would correspond to the presence
of Schottky barriers at the endpoints of the tube, near the source and drain contacts. The
potential at either endpoint depends on the work function of the metal contact, , and on
the electron affinity, , of the nanotube itself.5 Since the Fermi level of an intrinsic nanotube
4
This statement pertains to DC modelingin the AC analysis described in Chapters 9 and 10 of this
work, the contact series resistances are included in the small-signal model of the CNFET.
5
In a semiconductor, the work function = + (Ec EF ).
19
lies exactly at midgap6 , and in the cases where the metal and nanotube work functions are
not identical, a Schottky barrier will be present near the contacts to maintain a flat Fermi
level throughout the bias-free device (see, for example, Ref. [51]). Depending on the relative
magnitude of the material properties, this barrier may be deemed positive or negative with
regards to the carrier flow, and as will be shown below, the shape of a positive Schottkybarrier is critical in modulating the drain current. The presence of Schottky barriers in carbon
nanotube transistors has already been explored thoroughly in various works [8, 48, 5255].
The remaining heterointerfaces concern oxide-mismatch surfaces and are of importance to the
solution of Poissons equation, since they directly influence the permittivity-sensitive terms of
Eq. (1.3). The CNFET insulator is typically homogenous, and previously fabricated devices
have typically used SiO2 (r 3.9) [56], ZrO2 (r 25) [42] or HfO2 (r 16) [57]. Highpermittivity dielectrics are preferable, since they allow for better electrostatic coupling of the
gate to the channel; employment of non-SiO2 materials with Si technologies typically yields
surface roughness problems that are detrimental to carrier mobility. The relative permittivity
of the nanotube is taken to be unity [58], and the tube-insulator interface creates field-fringing
effects and a slight variation in the potential profile on the tube surface that cannot be neglected.
It must be noted that, although the solution of Eq. (1.3) encompasses the entire volume of the
device, we are primarily concerned with the longitudinal potential profile on the surface of the
tube, VCS (z), since knowledge of band bending there is required for determination of the Schottky barriers and subsequent transport calculations. Furthermore, the key challenge in modeling
the electrostatics lies in stipulating the charge distribution on the nanotube, as required on the
right-hand-side of Poissons equation. Prior to doing so, we further explore the electrostatics of
the CNFET by analyzing solutions to Laplaces equation, i.e., in the absence of carriers on the
tube (Qv = 0 in Eq. (1.3)). This is equivalent to observing the potential within the cylindrical
system in the absence of the carbon nanotube. Since knowledge of local electrostatic potential
allows us to determine any band bending imposed by the contacts, we subsequently derive
6
This is under the assumption of symmetrical density of states, as presented in Section 1.1.2.
20
qualitative band diagrams that serve to illustrate some important behaviours of the devices.
We begin by taking an arbitrary pair of voltages, and Fig. 1.9A illustrates the solution for
the bias conditions of VGS = 0.3 V and VDS = 0.5 V, and device dimensions Rg = 1 nm and
Lt = 50 nm. Note that the potential rises towards the drain end of the channel (z = Lt ), and
that in the mid-length region it is held at V VGS , i.e., the mid-length region of the device is
entirely controlled by the gate electrode, tracking its voltage identically.
Figure 1.9: Complete solution (A) and cross-section of solution (B) to Laplaces equation in
two-dimensional cylindrical coordinates for a cylinder 50 nm in length, 1 nm in diameter, and
biased at VGS = 0.3 V and VDS = 0.5 V.
Fig. 1.9B, a cross-section of the solution in Fig. 1.9A, taken at = Rt = 0.5 nm, shows the
nanotube surface potential, VCS (z), i.e., the potential where the nanotube surface would be.
The exact mathematical behaviour of the decaying end-regions of the plot is determined by the
particular geometry, and in the cylindrical case it is specified by Bessel functions [47]. Moreover,
in the short-channel limit these regions will interfere with one another and we cannot make
the previous assumption of constant VCS (Lt /2), as this only applies to device lengths greater
than the sum of the decay lengths of both barriers. The mutually-perpendicular electric field
components, i.e., radial and longitudinal, are shown in Fig. 1.10, for the same conditions given
in Fig. 1.9.
Alternate contact geometries will undoubtedly affect these profiles. For instance, decreasing the
21
6
x 10
0
Electric Field (V/m)
Electric Field (V/m)
x 10
0
2
2
4
6
Radial Component
4
0
20
z (nm)
Longitudinal Component
8
0
40
20
z (nm)
40
Figure 1.10: Electric field correspondent to the structure and potential described in in Fig. 1.9B.
gate length, for the same device length, would reduce the field strength near the end electrodes,
creating thicker potential barriers. Moreover, reducing the source and drain planar contacts to
point (needle) contacts yields field focusing at the end regions and significantly thinner barriers.
As previously stated, barring a choice of extreme geometries that are unfavourable to device
performance, most of the results discussed here are readily generalized, qualitatively, to other
CNFET structures.
From the potential profile of Fig. 1.9B, we continue with an illustration of how the CNFET
energy band diagram is constructed. Under the approximations of perturbation theory [59, 60],
which is valid for slow-changing potentials, we assume that relatively small variations in the
local electrostatic potentialdue to the application of a bias to the CNFETcause the bands
to rigidly shift with the local potential. This implies that the band structure obtained from
the -k relation does not get significantly distorted and that the density of states calculation
does not need to be re-computed at each point, but rather just shifted in energy. As such, the
conduction and valence band edges, Ec and Ev respectively, can be considered to be pinned to
VCS . This approximation is expected to be valid for transverse electric fields (radial component)
Er < 8 108 V/m and tubes with radii Rt 1.5 nm, worsening with larger radii and stronger
fields [61].
Moreover, in the absence of mid-channel Ec discontinuities, the solution to Poissons equation,
22
which is always continuous, is here directly related to the amount of band bending of the
vacuum level, such that Evac (z) = qVCS (z). This band bending also implies that there is an
equal amount of bending in the conduction and valence bands, but the exact position of these
will also be determined by discontinuities in work functions at the metal-nanotube interfaces.
The band diagram is thus obtained via the relations
Ec (z) = Evac (z)
Ev (z) = Ec (z) Eg ,
where Ec and Ev are, respective the conduction and valence band edges. Note that the potential
at the points z = 0 and z = Lt are determined by both the voltage of the nearest electrode
and also the work function difference there, either condition being sufficient to give rise to a
Schottky barrier.
We now examine the band diagrams derived from Laplace solutions, as shown in Fig. 1.11. The
cases illustrated assume, for simplicity, that there are no work function differences between any
of the electrodes, and they show four progressive stages, first stepping VGS and then VDS . As
noted before, the extent in the z-direction of the source barrier is independent of VDS , while the
drain barrier length is a function of both bias voltages. The dependence of the band diagram
on the work function difference between the nanotube and metal contacts will be shown later,
in Fig. 4.4.
Further analysis of these band profiles, in conjunction with the concept of ballistic transport
introduced in Sect. 1.1.3, allow us to infer the nature of the nanotube carrier distribution. Due
to the small thickness of the barriers, both tunneling and thermionically emitted components
are considered. Observing Fig. 1.11C, we note that the asymmetrical source and drain barriers
will give rise to different transmission probabilities, at any given energy, thus creating carrier
distribution functions that are distorted from their equilibrium (Fermi) forms. At high VGS ,
where the conduction band edge at mid-length, Ec (Lt /2), dips significantly below the source
Fermi level, the effect of hot carrier injection into the channel will dominate transport. It should
be evident that these non-equilibrium distributions impose difficulties in the computation of
23
0.5
(B)
VGS = 0, VDS = 0
(C)
0.5
Energy (eV)
Energy (eV)
0
10
20 30
z (nm)
40
VGS = 0.5 V, VDS = 0.4 V
10
20 30
z (nm)
40
=0
20 30
z (nm)
40
DS
0.5
(D)
0.5
0
10
50
VGS = 0.5 V, VDS = 0.9 V
0.5
1
1.5
50
= 0.5 V, V
50
0.5
GS
0.5
0.5
Energy (eV)
Energy (eV)
(A)
10
20 30
z (nm)
40
50
Figure 1.11: CNFET band diagrams for Laplace solutions. There is no work function difference
between the nanotube and any of the metal contacts.
charge in the channel under the conditions of non-zero drain-source voltages.
1.3.2 Charge
With the inclusion of the nanotube inside the empty gate cylinder, as in a real device, the
potential profile will be affected by the presence of any electron- and hole-charges on the tube
surface. Here we present a method for calculating the nanotube carrier concentration, necessary
for the charge term in Eq. (1.3), via a ballistic-transport, flux-derived distribution point of view.
In the case of short-channel devices, Schr
odingers equation is employed in order to account
for quantum-mechanical reflection in the device; for long-channel devices, a compact model for
expressing the charge is used by assuming that the mid-length region of the device is characterized by incoherent transport. Herein we discuss the latter case only, leaving Schr
odingers
equation to subsequent chapters.
Under equilibrium conditions, i.e., when VDS = 0 and there is no net charge flow, the carrier
24
distributions are given by the Fermi function. Application of a bias to the gate electrode is
manifested as band-bending in the mid-length region of the device, as illustrated in Fig. 1.11B,
with Ec dropping relative to the fixed EF for VGS > 0, or rising for VGS < 0. In this situation, a
self-consistent solution for the system is obtained by solving Poissons equation in conjunction
with the 1D carrier density equation, which accounting for both carrier types is
Z
Qv (z) = q(p n) = q
g(E) {f [E + qVCS (z)] f [E qVCS (z)]} dE ,
(1.4)
where g(E) is the tight-binding density of states from Sect. 1.1.2, f (E) is the Fermi function,
and VCS (z) is the amount of band-bending.
Charge modulation is brought about by shifting the bands relative to the Fermi level, and for
example, as the conduction band approaches the Fermi level (dotted line), there is an increase
in the local 1D electron concentration, with a dependence on E = Ec EF . Under a positive
gate-source bias regime, this accounts for the accumulation of negative charges at the centre
of the tube, attracted by the more positive gate and repelled by the more negative source.
Note also that the application of a positive gate-source bias also causes the Schottky barriers,
present near the source and drain contacts, to become taller and thinner relative to the midlength conduction band edge, thereby reducing the restriction on tunneling currents from those
terminals; note that these fluxes are balanced under equilibrium conditions.
Self-consistency is important because application of Eq. (1.4) with an arbitrary potential profile
will yield a net charge Qv (z) that is not consistent with that required by Eq. (1.3). As an
example, if we computed the net charge with Eq. (1.4) on a Laplace band diagram under the
conditions of VGS > 0 and VDS = 0, and subsequently solved Eq. (1.3) with this charge, we would
find an excess of electrons in the device that would yield a lower electrostatic potential observed
in the mid-length region, thus raising Ec there. Recomputing the charge with the new potential
profile we would find a discrepancy with the previously determined amount. Similarly, a large
hole concentration, induced by the application of a negative gate potential, would cause the
local electrostatic potential to increase, pulling the mid-length region of Fig. 1.9B downwards.
Thus, it is important to iterate between both equations, using, for example, Newtons method,
25
until a self-consistent solution is attained.

Since the operation of the CNFET requires the application of biases to both the gate and drain
contacts, with the aims of, respectively, inducing mobile charge in the channel and transporting
that charge through the device, we now turn to the non-equilibrium situation. Under such
conditions, the Fermi level only has meaning in the metal contacts, and the difference between
EF in the end contacts is responsible for the presence of a drain current. Under such conditions,
one approach to charge calculation is to apply an equation similar to Eq. (1.4), except that
the nanotube distribution functions are postulated to be shifted Fermi functions, having their
equiprobable point, labeled as the quasi-Fermi level, lying somewhere in the energy range
between the source and drain potentials. While this method yields satisfactory results, it fails
to properly account for hot carriers in the channel.
In this work, we calculate the nanotube non-equilibrium distributions based on incoming flux
from both the source and drain contacts. The electrodes are highly-conductive regions assumed
to be in local thermodynamic equilibrium, and thus electrons there are distributed according to
the respective Fermi functions. Given that the device operates in the ballistic regime, electrons
with sufficient energy to thermionically emit over or tunnel through the Schottky barrier at
either end, populate the channel at the same energy as they entered. Thus the bands in
the nanotube are populated such that the forward-traveling electron states, +k-states, are
primarily occupied by source-injected electrons, while drain-injected electrons primarily fill the
k-states. This method therefore amounts to a tunneling-barrier modulation of the Fermi
functions present in the source and drain metals, while accounting for backscattering due to
macroscopic electrostatic potentials.
Finally, CNFETs are different from most other semiconductor devices in that they can demonstrate, when properly engineered, bipolar and ambipolar behavioura single device can exhibit
transport by either electrons or holes depending on the bias conditions, as well as simultaneous transport by both carrier types under other conditions. The possibility of ambipolarity is
clear in Fig. 1.11D, which reveals how source electrons and drain holes face similar potential
26
barriers to injection into the channel. This phenomenon yields unusual current-voltage characteristics, which will be explored in later chapters. Recently, experimental observations of
ambipolarity have been reported in the literature, manifested via either photoemission [62] or
photoabsorption [63] in the infrared spectrum7 , and opening up possibilities for applications as
light detectors or emitters.
1.3.3 Transport
As was described in Sect. 1.1.3, transport is expected to be ballistic for the device lengths
and operating bias ranges chosen in this work [8]. Moreover, under the Landauer-B
uttiker
formalism, we treat the device as being composed of two carrier reservoirs, separated by a
1D scattering region described by an energy-dependent transmission probability T (E). Each
contact region is described by its own equilibrium carrier statistics f (E), and is assumed to
also be one-dimensional. The connection of this contact region to the actual macroscopic
electrode, wherein we have a 3D electron density, produces a quantized conductance, sometimes described as a mode constriction resistance, and this phenomenon has been verified
experimentally [12, 21]. Under this formalism, we modify the standard electron current expression I=qnv (where the I replaces the usual J because of the 1D nature of the system) to
account for the energy dependence of its constituents8 , and using the relations for electron density n(E)=f (E)g(E)T (E), density-of-states g(E)=(1/)(k/E), and carrier group velocity
v(E)=(1/~)(E/k), we find the net current between the source and drain to be
2q
I=
h
Z
[f (E S ) f (E D )] T (E) dE ,
where S and D are, respectively, the source and drain Fermi levels. Note that for each contact
we are only concerned with injection into the device, so we consider either +k or k modes;
thus we only use half of the g(E) given in Eq. (1.2), which accounted for both spin-degeneracy
and k.
7
Due to the bandgap of 0.7 eV being employed in typical devices, photoemission produces light of
wavelength hc/Eg 1.8 m, lying in the infrared range of the spectrum.
8
Recall that, under the elastic conditions here, each energy level of transport is independent of all
others.
27
The transmission coefficient T (E) is a function of virtually all device parameters, but particularly the gate- and drain-source voltages. Because of this, it will be shown in subsequent
chapters that it is important to employ a Schr
odinger-Poisson model to properly estimate the
transmission, and that its behaviour is strongly dependent on the self-consistent computation
of the device electrostatics. Conversely, the Fermi functions are solely functions of the contact
Fermi energies, which are related to the drain-source voltage only. For example, applying a
drain-source voltage lowers D in energy and depletes the backward flux, while the forward flux
remains constant. This is responsible for the saturation of the current as VDS increases.
1.4 Thesis Outline

The remainder of this thesis, prior to the concluding chapter, is a collection of manuscripts
published in journals or conference proceedings while this work was being carried out. Chap. 2
presents a preliminary unipolar compact model, which employs simplified, long-channel electrostatics and a crude estimate of the transmission function. Chap. 3 shows a detailed study of
the equilibrium electrostatics of the device, and, under non-equilibrium, employs the compact
model in its bipolar implementation. Chap. 4 gives an overview of the compact model results,
with the aim of evaluating the maximum attainable DC performance. Chap. 5 presents the earliest version of a Schrodinger-Poisson model and its results. Chap. 6 improves on the original
compact model, particularly on the transmission coefficient, and provides a comparison with the
more sophisticated model of Chap. 5. Chap. 7 is a study on the concept of quantum capacitance
in nanoscale transistors, and its particular application to 1D CNFET devices. Finally, a smallsignal model, in conjunction with an improved Schr
odinger-Poisson model (outlined in detail
in Appendix A), is employed in Chapters 8, 9, and 10 to obtain AC performance predictions
for CNFETs, namely the fT and fmax transistor figures-of-merit.
1.5 Specific Contributions

This work has contributed several models of varying detail to the growing body of knowledge in
CNFETs. In particular, a novel Schr
odinger-Poisson (SP) model gave results that were faithful
28
to the trends observed in prototype devices and served as the foundation for developing a smallsignal methodology from which high-frequency figures-of-merit were estimated. The inclusion
of parasitics, previously not considered in the literature, proved to have a substantial influence
in these predictions of small-signal AC performance. Also developed was a DC compact model,
and its relation to the SP model was mathematically explained. Both showed good agreement
in modeling high-performance negative-Schottky-barrier devices, and revealed that earlier DC
performance predictions in the literature may have been too optimistic. A detailed analysis
of quasi-bound states has also been provided, and this was found to be crucial in explaining
device behaviour, particularly as regards quantum capacitance and transconductance.
References
[1] A. Oberlin, M. Endo, and A. T. Koyama, Filamentous growth of carbon through benzene
decomposition, J. Cryst. Grow., 32(3), 335349 (1976).
[2] Sumio Iijima, Helical microtubules of graphitic carbon, Nature, 354, 5658 (1991).
[3] M. S. Dresselhaus, G. Dresselhaus, and P. C. Eklund, Science of Fullerenes and Carbon
Nanotubes (Academic Press, Toronto, 1996).
[4] Mildred S. Dresselhaus, Gene Dresselhaus, and Phaedon Avouris, eds., Carbon Nanotubes:
Synthesis, Structure, Properties, and Applications, vol. 80 of Topics in Applied Physics
(Springer, Berlin, 2001).
[5] Reshef Tenne and Alex K. Zettl, Nanotubes from inorganic materials, Mildred S. Dresselhaus, Gene Dresselhaus, and Phaedon Avouris, eds., Carbon Nanotubes: Synthesis, Structure, Properties, and Applications, vol. 80 of Topics in Applied Physics, 81112 (Springer,
Berlin, 2001).
[6] R. Saito, G. Dresselhaus, and M. S. Dresselhaus, Physical Properties of Carbon Nanotubes
(Imperial College Press, London, 1998), 1st ed.
29
[7] X. Blase, Lorin X. Benedict, Eric L. Shirley, and Steven G. Louie, Hybridization effects
and metallicity in small radius carbon nanotubes, Phys. Rev. Lett., 72(12), 18781881
(1994).
[8] Ali Javey, Jing Guo, Qian Wang, Mark Lundstrom, and Hongjie Dai, Ballistic carbon
nanotube field-effect transistors, Nature, 424, 654657 (2003).
[9] Alex Kleiner and Sebastian Eggert, Curvature, hybridization, and STM images of carbon
nanotubes, Phys. Rev. B , 64, 11340211134024 (2001).
[10] I. Cabria, J. W. Mintmire, and C. T. White, Metallic and semiconducting narrow carbon
nanotubes, Phys. Rev. B , 67, 12140611214064 (2003).
[11] Jeroen W. G. Wildoer, Liesbeth C. Venema, Andrew G. Rinzler, Richard E. Smalley, and
Cees Dekker, Electronic structure of atomically resolved carbon nanotubes, Nature, 391,
5962 (1998).
[12] Supriyo Datta, Electronic Transport in Mesoscopic Systems, vol. 3 of Cambridge Studies
in Semiconductor Physics and Microelectronic Engineering (Cambridge University Press,
New York, 1995).
[13] J. W. Mintmire and C. T. White, Universal density of states for carbon nanotubes, Phys.
Rev. Lett., 81(12), 25062509 (1998).
[14] A. Rakitin, C. Papadopoulos, and J. M. Xu, Carbon nanotube self-doping: Calculation
of the hole carrier concentration, Phys. Rev. B , 67, 03341110334114 (2003).
[15] Sander J. Tans, Alwin R. M. Verschueren, and Cees Dekker, Room-temperature transistor
based on a single carbon nanotube, Nature, 393, 4952 (1998).
[16] Satoru Suzuki, Chris Bower, Yoshio Watanabe, and Otto Zhou, Work functions and
valence band states of pristine and Cs-intercalated single-walled carbon nanotube bundles,
Appl. Phys. Lett., 76(26), 40074009 (2000).
30
[17] J. Chen, M. A. Reed, A. M. Rawlett, and J. M. Tour, Large on-off ratios and negative
differential resistance in a molecular electronic device, Science, 286, 15501552 (1999).
[18] S. Kazaoui, , N. Minami, N. Matsuda, H. Kataura, and Y. Achiba, Electrochemical
tuning of electronic states in single-wall carbon nanotubes studied by in situ absorption
spectroscopy and ac resistance, Appl. Phys. Lett., 78(22), 34333435 (2001).
[19] Zhihong Chen, Joerg Appenzeller, Joachim Knoch, Yu-Ming Lin, and Phaedon Avouris,
The role of metal-nanotube contact in the performance of carbon nanotube field-effect
transistors, Nano Lett., 5(7), 14971502 (2005).
[20] C. T. White and T. N. Todorov, Carbon nanotubes as long ballistic conductors, Nature,
393, 240242 (1998).
[21] David K. Ferry and Stephen M. Goodnick, Transport in Nanostructures (Cambridge University Press, New York, 1997).
[22] Mark Bockrath, David H. Cobden, Jia Lu, Andrew G. Rinzler, Richard E. Smalley, Leon
Balents, and Paul L. McEuen, Luttinger-liquid behaviour in carbon nanotubes, Nature,
397, 598601 (1999).
[23] Zhen Yao, Charles L. Kane, and Cees Dekker, High-field electrical transport in single-wall
carbon nanotubes, Phys. Rev. Lett., 84(13), 29412944 (2000).
[24] S. J. Wind, J. Appenzeller, and Ph. Avouris, Lateral scaling in carbon-nanotube fieldeffect transistors, Phys. Rev. Lett., 91(5), 05830110583014 (2003).
[25] M. S. Fuhrer, B. M. Kim, T. D
urkop, and T. Brintlinger, High-mobility nanotube transistor memory, Nano Lett., 2(7), 755759 (2002).
[26] Paul L. McEuen, Michael S. Fuhrer, and Hongkun Park, Single-walled carbon nanotube
electronics, IEEE Trans. Nanotechnol., 1(1), 7885 (2002).
31
[27] G. Pennington and N. Goldsman, Semiclassical transport and phonon scattering of

electrons in semiconducting carbon nanotubes, Phys. Rev. B , 68, 045426104542611
(2003).
[28] A. Bachtold, M. S. Fuhrer, S. Plyasunov, M. Forero, Erik H. Anderson, A. Zettl, and Paul L.
McEuen, Scanned probe microscopy of electronic transport in carbon nanotubes, Phys.
Rev. Lett., 84(26), 60826085 (2000).
[29] Marco Buongiorno Nardelli and J. Bernholc, Mechanical deformations and coherent transport in carbon nanotubes, Phys. Rev. B , 60(24), R16338R16341 (1999).
[30] Amitesh Maiti, Alexei Svizhenko, and M. P. Anantram, Electronic transport through
carbon nanotubes: Effects of structural deformation and tube chirality, Phys. Rev. Lett.,
88(12), 12680511268054 (2002).
[31] Marcus Freitag, A. T. Johnson, Sergei V. Kalinin, and Dawn A. Bonnell, Role of single
defects in electronic transport through carbon nanotube field-effect transistors, Phys. Rev.
Lett., 89(21), 21680112168014 (2002).
[32] B. Q. Wei, R. Vajtai, and P. M. Ajayan, Reliability and current carrying capacity of
carbon nanotubes, Appl. Phys. Lett., 79(8), 11721174 (2001).
[33] Hongjie Dai, Nanotube growth and characterization, Mildred S. Dresselhaus, Gene Dresselhaus, and Phaedon Avouris, eds., Carbon Nanotubes: Synthesis, Structure, Properties,
and Applications, vol. 80 of Topics in Applied Physics, 2953 (Springer, Berlin, 2001).
[34] Jean-Christophe Charlier and Sumio Iijima, Growth mechanisms of carbon nanotubes,
Mildred S. Dresselhaus, Gene Dresselhaus, and Phaedon Avouris, eds., Carbon Nanotubes:
Synthesis, Structure, Properties, and Applications, vol. 80 of Topics in Applied Physics,
5580 (Springer, Berlin, 2001).
[35] Adrian Bachtold, Peter Hadley, Takeshi Nakanishi, and Cees Dekker, Logic circuits with
carbon nanotube transistors, Science, 294, 13171320 (2001).
32
[36] S. J. Wind, J. Appenzeller, R. Martel, V. Derycke, and Ph. Avouris, Vertical scaling
of carbon nanotube field-effect transistors using top gate electrodes, Appl. Phys. Lett.,
80(20), 38173819 (2002).
[37] R. Martel, T. Schmidt, H. R. Shea, T. Hertel, and Ph. Avouris, Single- and multi-wall
carbon nanotube field-effect transistors, Appl. Phys. Lett., 73(17), 24472449 (1998).
[38] Chih-Kai Yang, Jijun Zhao, and Jian Ping Lu, Binding energies and electronic structures
of adsorbed titanium chains on carbon nanotubes, Phys. Rev. B , 66, 04140310414034
(2002).
[39] J. J. Palacios, A. J. Perez-Jimenez, E. Louis, E. San Fabi
an, and J. A. Verges, Firstprinciples phase-coherent transport in metallic nanotubes with realistic contacts, Phys.
Rev. Lett., 90(10), 10680111068014 (2003).
[40] R. Martel, V. Derycke, J. Appenzeller, S. Wind, and Ph. Avouris, Carbon nanotube
field-effect transistors and logic circuits, Proc. Design Automation Conf., 9498 (2002).
[41] R. Martel, V. Derycke, C. Lavoie, J. Appenzeller, K. K. Chan, J. Tersoff, and Ph. Avouris,
Ambipolar electrical transport in semiconducting single-wall carbon nanotubes, Phys.
Rev. Lett., 87(25), 25680512568054 (2001).
[42] Ali Javey, Hyoungsub Kim, Markus Brink, Qian Wang, Ant Ural, Jing Guo, Paul McIntyre,
Paul McEuen, Mark Lundstrom, and Hongjie Dai, High- dielectrics for advanced carbonnanotube transistors and logic gates, Nature Mater., 1, 241246 (2002).
[43] Ali Javey, Jing Guo, Damon B. Farmer, Qian Wang, Erhan Yenilmez, Roy G. Gordon,
Mark Lundstrom, and Hongjie Dai, Self-aligned ballistic molecular transistors and electrically parallel nanotube arrays, Nano Lett., 4(7), 13191322 (2004).
[44] M. Kr
uger, M. R. Buitelaar, T. Nussbaumer, C. Sch
onenberger, and L. Forr
o, Electrochemical carbon nanotube field-effect transistor, Appl. Phys. Lett., 78(9), 12911293
(2001).
33
[45] Sami Rosenblatt, Yuval Yaish, Jiwoong Park, Jeff Gore, Vera Sazonova, and Paul L.
McEuen, High performance electrolyte gated carbon nanotube transistors, Nano Lett.,
2(8), 869872 (2002).
[46] Lars-Erik Wernersson, Tomas Bryllert, Erik Lund, and Lars Samuelson, Wrap-gated InAs
nanowire field-effect transistor, IEDM Tech. Digest, 273276 (2005).
[47] D. L. John, Leonardo C. Castro, Jason Clifford, and David L. Pulfrey, Electrostatics
of coaxial Schottky-barrier nanotube field-effect transistors, IEEE Trans. Nanotechnol.,
2(3), 175180 (2003).
[48] Francois Leonard and J. Tersoff, Role of Fermi-level pinning in nanotube Schottky
diodes, Phys. Rev. Lett., 84(20), 46934696 (2000).
[49] J. P. Clifford, D. L. John, L. C. Castro, and D. L. Pulfrey, Electrostatics of partially
gated carbon nanotube FETs, IEEE Trans. Nanotechnol., 3(2), 281286 (2004).
[50] D. L. McGuire and D. L. Pulfrey, Error analysis of boundary condition approximations in
the modeling of coaxially-gated carbon nanotube field-effect transistors, Phys. Stat. Sol.
(a) (2006). Accepted February 21, 2006.
[51] Robert F. Pierret, Semiconductor Device Fundamentals (Addison-Wesley, New York,
1996).
[52] M. S. Fuhrer, J. Nyg
ard, L. Shih, M. Forero, Young-Gui Yoon, M. S. C. Mazzoni, Hyoung Joon Choi, Jisoon Ihm, Steven G. Louie, A. Zettl, and Paul L. McEuen, Crossed
nanotube junctions, Science, 288, 494497 (2000).
[53] Arkadi A. Odintsov, Schottky barriers in carbon nanotube heterojunctions, Phys. Rev.
Lett., 85(1), 150153 (2000).
[54] J. Appenzeller, J. Knoch, V. Derycke, R. Martel, S. Wind, and Ph. Avouris, Fieldmodulated carrier transport in carbon nanotube transistors, Phys. Rev. Lett., 89(12),
12680111268014 (2002).
34
[55] S. Heinze, J. Tersoff, R. Martel, V. Derycke, J. Appenzeller, and Ph. Avouris, Carbon
nanotubes as Schottky barrier transistors, Phys. Rev. Lett., 89(10), 10680111068014
(2002).
[56] S. Heinze, M. Radosavljevic, J. Tersoff, and Ph. Avouris, Unexpected scaling of the
performance of carbon nanotube Schottky-barrier transistors, Phys. Rev. B , 68, 235418
12354185 (2003).
[57] Ali Javey, Ryan Tu, Damon B. Farmer, Jing Guo, Roy G. Gordon, and Hongjie Dai,
High performance n-type carbon nanotube field-effect transistors with chemically doped
contacts, Nano Lett., 5(2), 345348 (2005).
[58] Francois Leonard and J. Tersoff, Dielectric response of semiconducting carbon nanotubes, Appl. Phys. Lett., 81(25), 48354837 (2002).
[59] Neil W. Ashcroft and N. David Mermin, Solid State Physics (Harcourt College Publishers,
New York, 1976), 1st ed.
[60] Supriyo Datta, Quantum Phenomena, vol. VIII of Modular Series on Solid State Devices
(Addison-Wesley, New York, 1989).
[61] Yong-Hyun Kim and K. J. Chang, Subband mixing rules in circumferentially perturbed
carbon nanotubes: Effects of transverse electric fields, Phys. Rev. B , 64, 1534041
1534044 (2001).
[62] J. A. Misewich, R. Martel, Ph. Avouris, J. C. Tsang, S. Heinze, and J. Tersoff, Electrically
induced optical emission from a carbon nanotube FET, Science, 300, 783786 (2003).
[63] M. Freitag, Y. Martin, J. A. Misewich, R. Martel, and Ph. Avouris, Photoconductivity
of single carbon nanotubes, Nano Lett. (2003).
35
Chapter 2
Towards a Compact Model for
Schottky-Barrier CNFETs
2.1 Introduction
Nanoscale transistors fashioned from carbon nanotubes are an exciting possibility [1, 2]. Transport is essentially one-dimensional, and there is very little carrier-phonon interaction [3]. Thus,
the drain current tends to be controlled by modulation of the Schottky-barrier potential profiles
at the source and drain ends of the nanotube [4, 5]. Present experimental devices are planar in
nature [4], but coaxial structures offer better opportunities for modulating the Schottky-barrier
properties via capacitative coupling between the gate and the contacts [6, 7]. In the present
work we concentrate on coaxial, Schottky-barrier devices formed with intrinsic nanotubes, and
seek to develop a compact model for the prediction of the drain I-V characteristics. The basic
transistor structure is shown in Fig. 2.1.
Under equilibrium conditions, which in these devices means no drain current, i.e., the drainsource voltage VDS = 0, but the gate-source voltage VGS is not necessarily zero, it is straightforward to compute the charge and potential profile by solving Poissons equation consistently
with the equilibrium charge density on the nanotube [8]. For VGS > 0, as considered in this
work, the charge on the tube is negative, and is due to a surfeit of electrons over holes. Here
c [2002] IEEE. Reprinted, with permission, from L. C. Castro, D. L. John, and D. L. Pulfrey, Towards
a Compact Model For Schottky-Barrier Nanotube FETs, Proc. IEEE Conf. on Optoelectronic and
Microelectronic Materials and Devices, 303306 (Sydney, Australia, 2002).
36
Gate
Oxide
CNT
Gate
Oxide
CNT
Source
Rg
Drain
Lt
Figure 2.1: Schottky-barrier carbon nanotube FET model geometry. The gate forms the curved
surface of the outer cylinder, and the source and drain form the two ends. The semiconducting
nanotube is placed coaxially with the outer cylinder.
we take the electronic charge to dominate and we neglect the contribution of the holes. Away
from equilibrium, when VDS > 0, the induced electron distribution on the tube will deviate
considerably from a Maxwellian or Fermi-Diracian form, on account of hot electron injection
from the contacts and, in the ballistic case considered here, the lack of opportunity for thermalizing collisions. This precludes the calculation of the non-equilibrium charge using simple,
quasi-Fermi-Dirac statistics.
In the present work, we obtain an estimate of the non-equilibrium charge QC in the mid-length
region of the tube, i.e., away from the source and drain potential barriers, by an extension of the
method of Guo et al. [9, 10]. QC is related to the mid-length potential on the tube VCS , which
connects, and affects, the potential profiles of the Schottky barriers. These, in turn, affect
the tunneling probabilities for electrons entering the tube from the reservoirs of equilibrium
charge at the source and drain metallizations. The energy-dependent tunneling probabilities
serve to distort the electron distribution in the tube from an equilibrium form. Here, we take
the potential profiles at the barriers to have an exponential form, and then solve for VCS by
equating the values of QC computed by the non-equilibrium-flux approach and an infinite-tube
approach [9, 10]. This gives a solution to the complete potential profile, and to the tunneling
probabilities, which then allows computation of the drain current from Landauers expression.
Our inclusion of the Schottky-barrier nature of the contacts leads to a significantly different
saturation current than predicted by the earlier model [9], in which the tunneling barriers were
not considered.
37
2.2 The Model

At equilibrium, i.e., when VDS = 0, simple electrostatics gives:
C = Cins (VGS VCS ) ,
Q
(2.1)
where VCS is the equilibrium potential, with respect to the source, of the carbon nanotube
at its mid-length, i.e., away from the influence of the source and drain contacts, and Cins is
the insulator capacitance for an infinitely long coaxial system. Out of equilibrium, i.e., when
VDS 6= 0, VCS is influenced by VDS :
VCS = VCS + VDS
QC
= Cins (VGS VCS ) ,
(2.2)
where is a parameter that needs to be determined in order to specify VCS . For VGS > 0, as
considered in this work, QC is a negative, electronic charge.
An alternative method for calculating QC follows from the flux approach, in which electrons
in the forward- and backward-directed fluxes are summed [9]. Here, we do not restrict the
fluxes within the tube to be hemi-Maxwellian or hemi-Fermi-Diracian in nature, but, instead,
we allow the actions of tunneling and repeated reflections between the potential barriers to
modify the electron distributions from the equilibrium form that they possess outside the tube,
at the actual source and drain metallic contacts. In the following, fS /2 denotes the positive- or
forward-directed part of the Fermi-Dirac distribution outside the tube at the source, whereas
fD /2 represents the negative- or backward-directed part of the distribution outside the tube at
the drain. Thus:
2
fS (E)
1 T ? dE
= q
g(E)
2
T
D
E
Z c
fD (E, VDS ) 2
?
1 T dE ,
+
g(E)
2
TS
Ec
Z
QC
(2.3)
where g(E) is the 1-D density of states computed from the tight-binding approximation [11],
the conduction band edge Ec in the mid-length region of the tube is dependent on VDS , and
the tunneling probabilities at the source and drain, TS and TD respectively, are computed using
38
the JWKB approximation and are dependent on E, VGS and VDS . The overall transmission
probability T ? = TS TD /(TS + TD TS TD ). In our method, we equate Eq. (2.2) and Eq. (2.3)
and solve for , thereby determining VCS . An iterative procedure is necessary because of the
dependence on VCS of TS and TD , via the potential profile, which are represented by
i
VCS h (za)
1
0za
V
CS
e 1
V (z) =
VCS
az La
h
i
DS
VCS CS
e a (Laz) 1 L a z L ,
e 1
(2.4)
where z is the distance from the source and L is the tube length. This representation is based on
the observation of the trends in barrier shape under those circumstane for which exact solutions
are presently possible, namely: Poissons equation at equilibrium, and Laplaces equation out
of equilibrium. The trends are: a barrier basewidth of a 2RG , where RG is the radius of the
gate (see Fig. 2.1); a barrier height at the source of Vpk = VCS (see Fig. 2.2); a barrier height at
the drain that varies from Vpk = VCS VDS when a spike is present, through Vpk = 0, to a
negative value when VDS > VCS (see Fig. 2.2); a barrier concavity that is captured by 3.6
for the device considered here. The barrier profiles are correct inasmuch as they prescribe values
for the tunneling probabilities TS and TD that yield a mid-length charge that is consistent with
that predicted by Eq. (2.2).
To complete the calculation of the drain current, the Landauer expression is used:
X 2q Z
[fS (E) fD (E, VDS )] T dE ,
ID =
~ Ec,i
i
where the sum is over i doubly-degenerate conduction bands, the edges Ec,i of which are functions of , VGS and VDS .
2.3 Results and Discussion

Results are presented for a (16,0) tube, which has a radius of 0.63 nm, and a gate/tube-radius
ratio of 10; the source and drain work functions are taken to be equal to that of the intrinsic
nanotube. The tube is sufficiently long that there is a mid-length region where the tube
potential VCS is flat. In this region, we have found that there is essentially perfect agreement
39
between the values of VCS calculated by the infinite-tube method (2.1), and by an exact 2-D
solution [8], at least under the equilibrium conditions tested thus far. The energy band diagram
for a variety of bias conditions is shown in Fig. 2.2. Note the VDS -dependence of VCS and of
0.1
Source
Contact
0.3
EF
Energy (eV)
0.2
~
q VCS
q VCS
Drain
Contact
the barrier shapes.
VGS = 0
V =0
DS
0.1
0.2
VGS > 0
0 < VDS < VGS
0.3
VGS > 0
V >V
0.4
DS
0.5
0
GS
20
40
60
qV
DS
VGS > 0
VDS = 0
80
100
Distance from source contact (nm)
Kinetic energy at z=0.5Lt (eV)
Figure 2.2: Conduction energy band diagram for various bias conditions.
0.3
VGS = 0.5 V
0.25
0.2
0.15
Backward Channel
Electrons travelling from
Drain to Source
Forward Channel
Electrons travelling from
Source to Drain
0.1
0.05
0
0.2
0.1
0.1
Electron distribution kstates
0.2
0.2
0.4
Drain
source voltage
Figure 2.3: Electron distribution in the mid-length of the tube as a function of VDS for
VGS = 0.5 V.
Electrons from the source and drain reservoirs are drawn into the tube by tunneling through,
and thermionic emission over, the potential barriers at the contacts. This distorts the injected
electron distributions from their equilibrium forms. The total distribution within the tube is
determined by the action of reflections at the Schottky barriers on the injected distributions.
When VDS = 0 this action produces an equilibrium, Fermi-Dirac distribution, as can be seen
in Fig. 2.3. As VDS increases, there is less injection from the drain, and less reflection from the
40
diminishing spike at the drain. Thus, the backward-directed part of the distribution starts
to disappear, and the forward part assumes a definitely non-equilibrium shape, with a bulge at
the kinetic energy corresponding to that of the maximum tunneling flux.
Drain current, I D (A)
0.6
VGS = 0.5 V
0.5
0.4
0.3
0.2
VGS = 0.4 V
0.1
GS
0.1
= 0.3 V
0.2
0.3
0.4
Drainsource voltage, VDS
0.5
Figure 2.4: Drain current-voltage characteristics.
Drain current, I
(A)
4
V
GS
= 0.5 V
This work
Reference 9
0
0
0.1
0.2
0.3
0.4
Drainsource voltage, VDS
0.5
Figure 2.5: Comparison of drain current-voltage characteristics at VGS = 0.5 V. The solid line
is this work; the dashed line is using the model of Ref. [9].
The drain I-V characteristics are shown in Fig. 2.4. The saturation current at VGS = 0.5 V is
around 1 A, which is not inconsistent with values emerging from prototype devices [12]. A
revealing comparison with earlier predictions is shown in Fig. 2.5. Note how the present model
indicates a considerably larger saturation voltage VDS,sat . This is because the reflecting action
of the potential spike at the tube-drain interface delays the realization of the full saturation
current. Also, note how the new model predicts an ID,sat that is about one-order of magnitude
less than that of the model of Guo et al. [9]. This is indicative of the importance of accounting
for the restrictive action that the Schottky barriers at the source and drain impose on the
current.
41
2.4 Conclusions
From this work on the modeling of coaxial, carbon nanotube FETs, it can be concluded that
the Schottky barriers at the source and drain contacts play a dominant role in determining the
I-V characteristics of the transistors. The model presented here represents a significant step
towards producing a compact model for these promising new nano-devices.
References
[1] V. Derycke, R. Martel, J. Appenzeller, and Ph. Avouris, Carbon nanotube inter- and
intramolecular logic gates, Nano Lett., 1(9), 453456 (2001).
12680111268014 (2002).
(2002).
[6] Christopher P. Auth and James D. Plummer, Scaling theory for cylindrical, fully-depleted,
surrounding-gate MOSFETs, IEEE Electron Device Lett., 18(2), 7476 (1997).
[7] Brian Winstead and Umberto Ravaioli, Simulation of Schottky barrier MOSFETs with a
coupled quantum injection/Monte Carlo technique, IEEE Trans. Electron Devices, 47(6),
12411246 (2000).
42
2(3), 175180 (2003).
[9] Jing Guo, Mark Lundstrom, and Supriyo Datta, Performance projections for ballistic
[10] Jing Guo, Sebastien Goasguen, Mark Lundstrom, and Supriyo Datta, Metal-insulatorsemiconductor electrostatics of carbon nanotubes, Appl. Phys. Lett., 81(8), 14861488
(2002).
[11] Keivan Esfarjani, Amir A. Farajian, Yuichi Hashi, and Yoshiyuki Kawazoe, Electronic,
transport and mechanical properties of carbon nanotubes, Y. Kawazoe, T. Kondow, and
K. Ohno, eds., Clusters and NanomaterialsTheory and Experiment, Springer Series in
Cluster Physics, 187220 (Springer-Verlag, Berlin, 2002).
2(8), 869872 (2002).
43
Chapter 3
Electrostatics of Coaxial Schottky-Barrier
CNFETs
3.1
Introduction
Carbon nanotubes (CNs) are being intensively investigated as possible structures from which
nanoscale transistors and logic gates might be fabricated [1, 2]. In devices where the gate
electrode covers the entire length of the nanotube, transistor action is achieved by the modulation, by the gate, of the potential profile at the Schottky-barrier contact appearing at the
source-tube interface [3, 4], rather than by the modulation of the channel properties, as in a
traditional, silicon-like field-effect transistor (FET) [5, 6]. Here we concentrate on the coaxial
Schottky-barrier carbon nanotube FET (SB-CNFET), in which the cylindrical gate surrounds
the tube and is insulated from it by a dielectric. The basic structure was shown previously
in Fig. 2.1. Whereas planar structures are currently being used experimentally [4], coaxial
structures, although much more difficult to fabricate, are likely to exhibit better short-channel
performance [7], and, as regards modulating the Schottky-barrier thickness via capacitative
coupling between the gate and the contact [8], are likely to prove more efficient.
As a first step towards providing a model for these new devices, we examine the electrostatics
c [2003] IEEE. Reprinted, with permission, from D. L. John, L. C. Castro, J. Clifford, and D. L.
Pulfrey, Electrostatics of Coaxial Schottky-Barrier Nanotube Field-Effect Transistors, IEEE Trans.

Nanotechnol., 2 (3), 175180 (2003). Note: the Appendix (An Analytical Solution for the Potential)
present in the published version is omitted here, as that was prepared solely by the journal papers first
author.
44
of coaxial SB-CNFETs, using both analytical and numerical procedures to obtain the potential
profile. Of course, the solution must be consistent with the electron and hole charge induced on
the surface of the nanotube, and any inherent charge, such as that due to dopants. Although
the latter are easily accommodated, they are not considered here in view of the findings that
procedures previously thought to lead to doping of a nanotube are more probably serving to
affect the work functions at the metal contacts to the CN [4]. Thus the nanotubes here are
considered to be intrinsic. The electron and hole charge densities can be computed using the
nearest-neighbour tight-binding approximation for the nanotube density of states (DOS) [9].
Results are presented here for the equilibrium situation, i.e., the drain-source voltage, VDS , is
zero, as this is presently the only case for which the carrier distribution functions are known
with certainty. The dependence of the potential profile along the tube on the work functions
of the source-, drain- and gate-metallizations, and of the thickness and permittivity of the gate
dielectric, is reported. Outside of equilibrium, i.e., for VDS 6= 0, the distribution functions
are likely to be highly distorted from their equilibrium shape [10]. This is due to the absence
of thermalizing collisions in this one-dimensional (1-D) system, for which there is very little
carrier-phonon interaction [11]. Presently, the only way to obtain exact results for VDS 6= 0 is
to solve Laplaces equation. This may be appropriate for studying sub-threshold conduction.
Such results are presented here and they indicate a sub-threshold slope which depends on
dielectric thickness in a different manner from that recently reported for planar-geometry SBCNFETs [12].
Although, as stressed above, a procedure for obtaining a fully self-consistent solution for the
above-threshold case is not yet available, the results presented here for the equilibrium potential
profiles can be used to infer the general form of the drain current-voltage (I-V) characteristic.
On doing this, the interesting spectre of having simultaneous injection of electrons and holes into
the nanotube is raised. The drain characteristics for such a situation are briefly examined using
a rudimentary, non-equilibrium, compact model [10], in which the source and drain potential
profiles are approximated by exponential expressions that have their basis in the electrostatic
solutions presented herein.
45
3.2
Coaxial Nanotube Electrostatics
The electrostatic problem reduces to that of a bounded cylinder of length Lt and radius Rg ,
as shown in Fig. 2.1. In cylindrical coordinates, using the source electrode as reference, the
appropriate boundary conditions for the potential V (, , z) are:
V (Rg , , z) = VGS G /q
V (, , 0) = S /q
V (, , L) = VDS D /q
V (, , z) = V (, + 2, z) ,
where G , S and D are the work functions of the gate-, source- and drain-metallizations,
respectively, VGS is the gate-source voltage, and q is the magnitude of the electronic charge.
The boundary conditions at z = 0 and z = L are appropriate in the absence of Fermi-level
pinning [13]. Note, too, that V (0, , z) is assumed to be finite.
An analytical solution, at least for the case of a homogeneous permittivity within the metallized
enclosure, is possible following the methods of [14] and [15]. For the inhomogeneous case of
different permittivities for the dielectric and the nanotube, numerical techniques are easier to
implement. We have used a standard finite-element package for this purpose1 .
The net charge density, comprising electrons and holes, is taken to reside on the surface of the
CN, and is given by
Q(r) =
1
( Rt )Qz (z) ,
2
where Qz (z) is the net 1-D charge concentration, Rt is the CN radius, and (x) is the Dirac
delta function. The nanotube charge needs to be computed self-consistently with the potential
on the nanotube but, as mentioned in Sect. 3.1, a difficulty arises under non-equilibrium conditions because of the present inability to rigorously specify the distribution function for the
hot carriers within the tube. However, for equilibrium conditions, there is no such problem and
the carrier concentrations are found by allowing the local electrostatic potential to rigidly shift
1
FEMLAB, see http://www.comsol.com
46
the CN DOS [14, 15]. Using the nearest-neighbour tight-binding approximation [9], the DOS
is symmetrical about EF , so the net 1-D carrier density at some point along the intrinsic tube
may be computed as
Z
Qz (z) = q
g(E) [f (E + qV (z) + t ) f (E qV (z) t )] dE ,
where g(E) is the 1-D tube DOS, the degeneracy in the energy bands is as considered in [16],
f (E) is the Fermi-Dirac distribution function, EF is taken to be zero, and t is the work function
of the intrinsic carbon nanotube.
For the non-equilibrium case, the only exact solution that can be given presently is for the case
of no charge on the nanotube, namely: Qz (z) = 0. A complete solution awaits the formulation
of an appropriate solver, likely of the Poisson-Schr
odinger variety.
3.3
Results and Discussion
Results are presented for (16, 0) tubes having a radius of 0.63 nm, a length of 100 nm and a gate
work function of 4.5 eV. Various ratios of gate radius to tube radius, relative permittivity of
the dielectric, ins , and source- and drain-work functions, are considered. The electron affinity
for the carbon nanotube is taken to be 4.18 eV, based on a work function of 4.5 eV [17], and
an intrinsic-tube band gap of 0.64 eV. Unless otherwise stated, the relative permittivity of the
nanotube, t , is taken to be the same as that of the gate dielectric. The temperature is taken
to be 300 K.
At equilibrium conditions, and when S = D , the potential profile along the tube will be
symmetrical. Thus, only profiles near one contact need be shown. Fig. 3.1 shows the energy
band diagrams near the source for Rg /Rt = 10, ins = 3.9 and for various S = D with
VDS = 0 V and VGS = 0.2 and 0.5 V. In Fig. 3.1(a), S = 4.5 eV and corresponds to the case of
equal work functions for the metal and the nanotube, whereas S = 4.33 eV [Fig. 3.1(b)] and
S = 4.63 eV [Fig. 3.1(c)] refer to low- and high-metal work functions, respectively [13]. The
potential in the body of the tube, distal from the contacts, depends directly on VGS , leading
to potential spikes in the tube at the source and drain of height determined by both S,D and
47
(a)
Energy (eV)
0.4
(b)
0.4
0.2
0.2
0.2
0.2
0.2
0.2
0.4
0.4
0.4
0.6
0.6
0.6
0.8
20
40
0.8
20
(c)
0.4
40
0.8
20
40
Distance from source (nm)
Figure 3.1: Self-consistent equilibrium energy band diagram near the source for a (16,0) tube
with a 5.6 nm gate dielectric thickness (Rg /Rt = 10), and S = D set to (a) 4.5 eV, (b) 4.33 eV,
and (c) 4.63 eV. Data are for VDS = 0 V and VGS = 0.2 (dashed line) and 0.5 V (solid line).
Energies are with respect to the Fermi level (dotted line).
VGS . Only in the low-S case at low VGS is thermionic emission likely to make a significant
contribution to the source current. In all other cases shown in Fig. 3.1, tunneling of electrons
through the spike will dominate.
The band diagrams for the same work function cases as used in Fig. 3.1, but for Rg /Rt = 50,
are shown in Fig. 3.2. The reduced band bending in the tube at the contacts, due to poorer
coupling between the gate and the nanotube, is very evident and will lead to a dramatic decrease
in current, except in the low work function case at low VGS where, as mentioned previously, the
electron current will be due to thermionic emission and will be determined by the height, and
not the shape, of the barrier. The present state of the art as regards gate-dielectric thinness
is 2 nm [12]. Regarding the permittivity of the dielectric, recent work has reported the use of
48
(a)
Energy (eV)
0.4
(b)
0.4
0.2
0.2
0.2
0.2
0.2
0.2
0.4
0.4
0.4
0.6
0.6
0.6
0.8
20
40
0.8
20
(c)
0.4
40
0.8
20
40
with a 30.9 nm gate dielectric thickness (Rg /Rt = 50), and the same sequence of work functions
as in Fig. 3.1, namely: S = D set to (a) 4.5 eV, (b) 4.33 eV, and (c) 4.63 eV. Data are for
VDS = 0 V and VGS = 0.2 (dashed line) and 0.5 V (solid line). Energies are with respect to the
Fermi level (dotted line).
zirconia [6], for which ins is around five times higher than that used in obtaining the above
figures. The effect of such a change in ins can be seen by comparing Figs. 3.1 and 3.3. At
VGS = 0.5 V, the increased capacitative coupling between the gate and the tube drives the
mid-tube potential energy to lower values, yet does not change significantly the width of the
source barrier at its base. Thus, obviously, an increased current for a given bias will result from
using a higher ins . At lower VGS , e.g., 0.2 V, the increased ins makes essentially no difference
to the potential profile because, at least for Rg /Rt = 10, there is virtually no charge induced on
the tube. From Figs. 3.1, 3.2 and 3.3, it appears that the width of the potential barrier at its
base depends strongly on the radius of the gate Rg for the contact geometry considered here,
as has been remarked upon elsewhere [18], and here we indicate that it also depends barely at
49
(a)
Energy (eV)
0.4
(b)
0.4
0.2
0.2
0.2
0.2
0.2
0.2
0.4
0.4
0.4
0.6
0.6
0.6
0.8
20
40
0.8
20
(c)
0.4
40
0.8
20
40
with a 5.6 nm gate dielectric thickness (Rg /Rt = 10), a gate dielectric with permittivity five
times higher than used in Fig. 3.1, and the same sequence of work functions as in Fig. 3.1,
namely: S = D set to (a) 4.5 eV, (b) 4.33 eV, and (c) 4.63 eV. Data are for VDS = 0 V and
VGS = 0.2 (dashed line) and 0.5 V (solid line). Energies are with respect to the Fermi level
(dotted line).
all on VGS .
Note that the effect of changing the gate work function from the value of 4.5 eV used here can
be readily appreciated from the foregoing figures as an increase in G of 0.1 eV, for example,
has the same effect as a corresponding decrease in VGS .
Turning now to the non-equilibrium case, as mentioned above, only the case of zero charge on the
nanotube can be solved for exactly, pending the determination of the appropriate distribution
functions or wave-functions for the hot carriers. In sub-threshold, the charge accumulation
is not significant in the electrostatics solution, thus, we solve Laplaces equation in order to
study the behaviour in this regime [12]. The current is computed from the Landauer formula,
50
assuming ballistic transport and incorporating the effect of carrier transmission and reflection at
the internal source/tube- and drain/tube-barriers in the transmission coefficient, T , computed
under the WKB approximation,
4q
I=
h
Z
T (E)[f (E) f (E + qVDS )] dE .
Some results for various gate/tube dimensions are shown in Fig. 3.4: note that the dielectric
thickness tins = Rg Rt . The sub-threshold current can be due to either mainly electrons or
mainly holes, depending on the bias conditions. For VDS = 0.1 V, as used in obtaining Fig. 3.4,
at VGS = 0 there is no band bending at the source end of the tube, but there is a spike in
the valence band at the drain end of the tube, which permits a hole tunneling current. This
current increases as VGS becomes more negative. For positive values of VGS , a spike appears
in the conduction band at the source, thereby allowing electron tunneling, but the spike in the
valence band at the drain thickens, and so the hole current is reduced. These changes with VGS
lead to a minimum in current, which occurs at VGS = VDS /2, when the tunneling barriers for
source electrons and drain holes are of equal thickness. The magnitudes of the sub-threshold
slopes, |S| = |(d log10 I/d VGS )1 |, for both mainly electron conduction (VGS > VDS /2), and
mainly hole conduction (VGS < VDS /2), are identical, owing to the use here of a symmetrical
DOS function for the conduction- and valence-bands, and equal work functions for the source
and drain contacts.
In analyzing their planar SB-CNFETs, Heinze et al. [12] noted that changing the dielectric
thickness is equivalent to rescaling the gate voltage. Thus, they found that S and the tube
potential scaled with dielectric thickness in the same manner, i.e., as tins . This suggests
that we seek a scaling relationship for S in our coaxial devices by examining how the VGS dependent part of the potential in the vicinity of the source contact varies with tins . This can
be accomplished by expanding the first term of the first mode of an analytical Laplace solution
(Ref. [19], Appendix), i.e.,
4(qVGS G )I0 Rt L1
zL1
t
t
V '
,
qI0 Rg L1
t
51

RG/RT = 2 (S 76 mV/dec)
5
RG/RT = 5 (S 80 mV/dec
10
Drain current (A)
10
10
(16,0) tube
VDS = 0.1 V
t = 1
ins = 3.9
10
10
10
10
0.4
Hole
Majority
0.3
Electron
Majority
0.2
0.1
0
0.1
0.2
Gatesource voltage (V)
0.3
0.4
Figure 3.4: Sub-threshold current for the same tube properties as in Fig. 3.1(a), but with various
ratios of gate radius to tube radius, for VDS = 0.1 V. Note: the dielectric thickness is Rg Rt .
where I0 is the zeroth-order Modified Bessel Function of the First Kind. Thus
1
V

VGS
I0 Rg L1
t
and, therefore, we can expect
S ' I0 (tins ) ,
where and are fitting parameters. It is found that a reasonable fit to S, from the data
included in Fig. 3.4, results with 79 mV/decade and 0.15 nm1 . The fact that a fit
can be obtained confirms that the dependence of S on tins is related to the specific geometry
of the transistor, with a Bessel function being involved in this case because of the cylindrical
structure. Electrically, tins is related, of course, to the gate capacitance, through which VGS is
coupled to the CN potential.
Considering now the case of above-threshold conduction, the solution to Laplaces equation for
various values of VDS is worth examining as it gives an idea of the evolution of the barrier
profiles with drain-source voltage. The exact solutions for the drain end of the tube show that
the conduction-band spike diminishes with VDS , and that, on further increasing VDS , a spike
occurs in the valence band. The base widths of the potential profiles at both source and drain
are found, as in the equilibrium case, to be of the order of Rg . This fact has been used, in
52
conjunction with a self-consistent procedure for solving for the potential and the charge in the
mid-length, field-free region of the tube, to generate compact expressions which describe the
potential along the entire length of the tube as a function of VGS and VDS [10].
0.4
Energy (eV)
0.2
0
0.2
0.4
0.6
0.8
0
//
10
90
100
Figure 3.5: Energy band diagrams for the same tube properties as in Fig. 3.1(a) for VGS = 0.3 V
and various VDS : 0 V (dotted), 0.2 V (dashed), 0.5 V (solid). For the equilibrium case the profile
is the exact solution of Poissons equation; for VDS 6= 0 the solutions are from Castros compact
model [10], with the base widths of the potential profiles at the source and drain being taken
as 2Rg .
Results are shown in Fig. 3.5 for the SB-CNFET for which the equilibrium band diagram is
shown in Fig. 3.1(a). As mentioned, the barrier at the drain for electron flow from the nanotube
into the drain diminishes as VDS is increased. This will lead to reduced reflection of sourceinjected electrons and an increase, and eventual saturation, of the electron current. Further
increase in VDS causes a spike to appear in the valence band profile at the drain. This will
allow hole tunneling to occur, and raises the interesting prospect of a hole current issuing from
the drain contact and adding to the electron current, thereby leading to a significant increase
in the total drain current. When the electron current is small due to the absence of a tunneling
barrier at the source, as will occur at low VGS , the drain current will be due almost entirely to
holes, and the drain I-V characteristic will appear near-exponential in shape.
An illustrative drain I-V characteristic is shown in Fig. 3.6. This data is generated from the
compact model of Castro et al. [10], as used to produce the potential profiles in Fig. 3.5, with
the base width of the potential barriers being taken as equal to 2Rg . Castros model is based
on that of Guo et al. [5], and improves upon it by: 1) introducing Schottky barrier contacts at
53

1.0
V
GS
=0V
GS
= 0.2 V
Drain current (A)
0.8
0.6
0.4
VGS = 0.5 V
0.2
GS
0.1
0.2
0.3
0.4
0.5
0.6
Drainsource voltage (V)
= 0.4 V
0.7
0.8
Figure 3.6: Drain I-V characteristics, as calculated by the method of Ref. [10], for a (16,0)
tube, Rg /Rt = 10, S = D = 4.5 eV and various values of VGS .
the source and drain; 2) accounting for reflection of carriers in the tube between the source and
drain barriers; 3) allowing for simultaneous electron and hole flows; and 4) not demanding that
the charge in the mid-length tube region remain at its equilibrium value. It is an approximate
model due to the estimated shape of the barrier profiles, so the current magnitudes given
in Fig. 3.6 are also only approximate. However, they do confirm the evolution of the drain
characteristics, as inferred from the above discussion of the exact potential profiles displayed
in the present work. Furthermore, our predictions are consistent with very recent experimental
results that show near-exponential drain I-V characteristics for VGS < VDS [20].
As a final comment on the simultaneous presence of electrons and holes in the nanotube, some
recombination is to be expected, in which case the drain current will be less than the sum of
two noninteracting particle flows, and in practice may not increase as dramatically as indicated
here. Experimentally, evidence of recombination within the nanotube of a SB-CNFET has been
demonstrated via the measurement of light emission under bias conditions appropriate for the
simultaneous injection of holes and electrons [20].
54
3.4
Conclusions
From this work on the electrostatics of coaxial Schottky-barrier carbon nanotube FETs, the
following can be concluded.
1. The potential barriers at the source/nanotube and drain/nanotube interfaces are strongly
affected by the work functions of the source, drain and gate, and by the thickness and
permittivity of the dielectric that surrounds the nanotube.
2. An analytical solution for the potential distribution in the case of equal permittivities of
the gate dielectric and the nanotube gives a good approximation to the numerical solution
for the case when the difference in permittivities of the dielectric and tube is taken into
account. In other words, the radial field inside the nanotube, for this particular geometry,
is not of great importance.
3. The sub-threshold slope approaches the thermionic limit of 60 mV/decade as the dielectric thickness is reduced, in a manner consistent with the cylindrical geometry of the
device.
4. From the results presented here, trends in the above-threshold drain I-V characteristics
can be inferred. The possibility of contributions to the drain current from both electron
and hole flow is indicated.
References
[1] V. Derycke, R. Martel, J. Appenzeller, and Ph. Avouris, Carbon nanotube inter- and
intramolecular logic gates, Nano Lett., 1(9), 453456 (2001).
(2002).
55
12680111268014 (2002).
[7] Sang-Hyun Oh, Don Monroe, and J. M. Hergenrother, Analytic description of shortchannel effects in fully-depleted double-gate and cylindrical, surrounding-gate MOSFETs,
IEEE Electron Device Lett., 21(9), 445447 (2000).
[8] Brian Winstead and Umberto Ravaioli, Simulation of Schottky barrier MOSFETs with a
coupled quantum injection/Monte Carlo technique, IEEE Trans. Electron Devices, 47(6),
12411246 (2000).
[9] Keivan Esfarjani, Amir A. Farajian, Yuichi Hashi, and Yoshiyuki Kawazoe, Electronic,
transport and mechanical properties of carbon nanotubes, Y. Kawazoe, T. Kondow, and
K. Ohno, eds., Clusters and NanomaterialsTheory and Experiment, Springer Series in
Cluster Physics, 187220 (Springer-Verlag, Berlin, 2002).
[10] L. C. Castro, D. L. John, and D. L. Pulfrey, Towards a compact model for Schottky-barrier
nanotube FETs, Proc. IEEE Conf. on Optoelectronic and Microelectronic Materials and
Devices, 303306 (Sydney, Australia, 2002).
12354185 (2003).
56
[14] Francois Leonard and J. Tersoff, Novel length scales in nanotube devices, Phys. Rev.
Lett., 83(24), 51745177 (1999).
[15] Arkadi A. Odintsov, Schottky barriers in carbon nanotube heterojunctions, Phys. Rev.
Lett., 85(1), 150153 (2000).
[16] M. S. Dresselhaus, G. Dresselhaus, and P. C. Eklund, Science of Fullerenes and Carbon
Nanotubes (Academic Press, Toronto, 1996).
[18] Takeshi Nakanishi, Adrian Bachtold, and Cees Dekker, Transport through the interface
between a semiconducting carbon nanotube and a metal electrode, Phys. Rev. B , 66,
07330710733074 (2002).
2(3), 175180 (2003).
57
Chapter 4
An Evaluation of CNFET DC Performance
4.1 Introduction
Carbon nanotube molecules can be either metallic or semiconducting, which raises the fascinating spectre of filamentary integrated circuits fashioned from nanoscale transistors and
interconnects [1]. Such circuits would be of a different form from that of todays silicon circuits,
and would surely lead to a host of new applications. The atoms within a carbon nanotube
molecule bond covalently in hexagonal rings, and this graphite-like structure has great strength
and stability. Electrically, this helps in significantly reducing electromigration, thereby allowing high-current operation. Furthermore, carbon nanotubes conduct heat nearly as well as
diamond, so extremely high device-packing densities should be possible. When one adds to
these advantageous properties the fact that the bandgap of carbon nanotubes can be tuned by
varying the tube diameter, then it is clear that carbon nanotubes are worth investigating as a
candidate molecule for applications in nanotechnology.
The integrity of carbon nanotube molecules may preclude substitutional doping, but adsorbed
material [2], and ions within the nanotube [3], can alter the tubes conductivity from its natural
intrinsic state. Thus, bipolar junction devices are possible, perhaps operating in single-electron
mode with a p-type quantum well between two n-type regions [2]. In this paper, we focus on
A version of this chapter has been published. L. C. Castro, D. L. John, and D. L. Pulfrey, Carbon
Nanotube Transistors: An Evaluation, Proc. SPIE Conf. Device and Process Technologies for MEMS,
Microelectronics, and Photonics III , vol. 5276, 110 (Perth, Australia, 2004).
58
transistors made from intrinsic nanotubes, in which the conductivity is modulated by a gate
electrode, i.e., carbon nanotube field-effect transistors (CNFETs). These devices have been
under experimental investigation since 1998 [4], and engineering models to aid in their design
and analysis are starting to appear [511]. Thus, it is timely to assess the performance characteristics of CNFETs, at least under DC operation, for which there is considerable experimental
data.
4.2 Fabrication
Metallic and semiconducting nanotubes are often sorted on the basis of tube diameter, as
measured by atomic force microscopy [12], but recent reports of covalent chemical functionalization should make this task more practical [13]. Successful attempts to merge metallic and
semiconducting nanotubes have been reported [14], but this desirable union has not yet been
incorporated into experimental CNFETs. Instead, these devices presently rely for their fabrication on the traditional microelectronic techniques of metal deposition and electron-beam
lithography to make and define metal contacts to the nanotube, which forms the channel of the
transistor. The nanotubes themselves have either been grown in-situ by CVD from catalytic
island sites [15], or dispersed from ropes produced by laser ablation of a catalyst-containing
carbon target, and positioned on-chip by atomic force microscopy [16]. Clearly, if there is to be
any hope of mass-production of CNFETs, some method of organized growth, or self-assembly
is necessary. Progress in this direction includes aligned growth in an electric field [17], and
growth of arrays of vertical tubes within alumina nanopores [18].
4.3 Theoretical Considerations

Carbon nanotubes are cylinders of graphene, with a wall thickness of one atomic layer. For small
diameter tubes, quantization of the wave vectors in the circumferential (transverse) direction
leads to a quasi-one-dimensional energy versus wave vector (-k) relation, with a continuum of
states only in the longitudinal direction. In many cases, this entails the opening of a bandgap,
which for a 1 nm-diameter tube, for example, is about 0.8 eV. The band structure is usually
59
determined from a tight-binding Hamiltonian, with one p-orbital per carbon atom and a nearestneighbour matrix element of about 2.8 eV [19]. This approach renders the conduction- and
valence-band structures symmetrical. Note that each of the transverse wave vectors corresponds
to an allowed mode of propagation, of which there are two in the lowest sub-band, the most
relevant in transport calculations.
The small number of propagating modes is in contrast to the very large number of modes
in bulk terminals to which a device or circuit must ultimately be connected. This modeconstriction leads to a quantized, interfacial conductance, which will exist even if conduction
down the nanotube itself is ballistic [20]. In the modeling of carbon nanotube diodes and
field-effect transistors, it is becoming customary to locate such interfaces at the end contacts
to a semiconducting nanotube, and to regard the nanotube as an object with a characteristic
transmission probability T [7, 11, 21]. The end contacts are viewed as reservoirs of charge,
maintained under equilibrium conditions, from which carriers are injected into the nanotube,
depending on T and the applied bias, e.g., the drain-source voltage VDS of a CNFET. This
Landauer formalism, as described in, e.g., Ref. [20], allows the net electron current for two
modes to be expressed as
4q
Ie =
h
Emid
Te (E)[f (E) f (E + qVDS )] dE ,
(4.1)
where q is the magnitude of the electronic charge, h is Plancks constant, Emid is the energy
of the conduction band edge in the mid-section of the nanotube, f (E) is the Fermi-Dirac
distribution function, and Te refers to electrons. Clearly, the lower limit of integration would
need to be changed if source-drain tunneling were an issue. A similar expression can be written
for holes, using a transmission probability Th . Thus, once the T s are known, the problem of
the DC drain current is solved.
4.4 DC Modeling
Present CNFETs are usually planar devices, with a single gate situated either above [22], or
below [16], the gate dielectric and the nanotube. Double-gated devices [23], and electrolyt60
ically gated devices [24], more closely resemble the coaxial structure with wrap-around gate
that would be ideal, given the cylindrical form of the nanotube. As we are interested in
ultimate-performance assessment, we take a coaxial geometry for modeling purposes, as illustrated in Fig. 4.1. If the source and drain end contacts are the same radius as the gate, a
closed metallic cylinder results, for which there is an analytical solution for the device electrostatics [8]. For smaller diameter end contacts, or for partially gated devices in which the gate
electrode does not cover the entire length of the nanotube, an open-boundary problem exists,
and numerical methods need to be employed [10, 25].
Figure 4.1: Coaxial CNFET model geometry.

The transmission probabilities T are functions of energy and local potential V (, , z). The
latter is related to charges within the system, such as: electrons, holes and adsorbed ions on
the nanotube; charges trapped within the gate dielectric or the nanotube; charges residing
in so-called metal-induced-gap-states (MIGS), or evanescent states within the semiconductor
bandgap near the junctions. Here, we consider only electrons and holes, which are taken to
reside on the surface of the tube, i.e.,
Q(r) =
1
( Rt )Qz (z) ,
2
where Qz (z) is the net 1-D carrier concentration, Rt is the nanotube radius, and (x) is the
Dirac delta function.
If we allow for a spatially varying permittivity in the radial direction, Gauss Law implies that:
2V
1 1 V
2V
Q(r)
+
+
+
=
,
2

z 2
61
where the potential within the device reduces to a function of just two cylindrical coordinates
V (, z), due to the symmetry of the device in the angular direction. The charge Qz (z) is
distributed within the allowed 1-D density-of-states (DOS), which is taken to be rigidly shifted
by V (, z).
The solution for the potential is relatively straightforward under equilibrium conditions because
Fermi-Dirac statistics apply not just in the metal contacts, but everywhere along the tube, and
can be used to compute the electron and hole concentrations [8]. In the absence of gate leakage,
equilibrium occurs when VDS = 0.
Out of equilibrium in the sub-threshold region of operation or even, as it has been claimed,
in the turn-on mode [26], Qz (z) may be so small that it can be set to zero, and a solution to
Laplaces equation can be used for V (, z).
In the fully-on state, Qz (z) cannot be ignored. In the earliest CNFET model, this charge
was taken to maintain its equilibrium value [5], as would be appropriate for a FET working
in the traditional, charge-control mode [27]. Because of the large insulator capacitance, and
small quantum capacitance, that can be obtained in CNFETs (see next section), a voltagecontrolled mode of operation is more relevant to these devices [27], at least when tunneling
is not important. In this mode, it is the mid-tube potential energy Emid that is important
because it controls the height of the barriers to thermionic emission flow at the end contacts.
This potential energy is solved for self-consistently with the mid-length charge in a recent
model [7], which has been extended to account for both electron and hole charge and transport.
In this model, Emid is connected to the end-contact barrier heights using a compact expression
for the potential profiles based on solutions to Laplaces equation. Obviously, this will only
give approximate results in cases where the actual shape of the barrier is important, but it
will give good results in devices where it is only the height of the barrier that is important,
as in, for example, metal/nanotube junctions with a negative barrier height for electrons, or
doped/intrinsic nanotube junctions. This method was used to obtain the results presented here.
Another approach is to employ quasi-equilibrium statistics, with a separate quasi-Fermi level
62
for each sub-band. In the case of ballistic transport, the quasi-Fermi levels are flat along the
length of the tube, and they split at the end contacts. A flux-balancing approach can then
be used to include the current, as well as the charge and the potential, in a self-consistent
solution [9].
The shortcoming of any semi-classical model is that, while the non-equilibrium electron and
hole concentrations within the allowed bands may be computed, allowance is not made for
charge in evanescent states. However, if this charge does not lead to a change in Emid , then the
neglect of this charge will not affect the computation of the salient thermionic emission current.
A full, quantum-mechanical, approach is necessary to achieve a complete solution in cases where
the charge in evanescent states is important, and where tunneling Schottky barriers are present,
and in all devices where resonance and coherency are issues, and also where source-to-drain
tunneling is a possibility [11].
Irrespective of the independent method used to compute Qz (z), Poissons equation needs to
be solved subject to the appropriate boundary conditions. Conformal mapping is effective
when open boundaries are present [25]. For the metallic electrodes, a simple phenomenological
representation, assuming no Fermi-level pinning [21], is
V (Rg , z) = VGS G /q
V (, 0) = S /q
V (, Lt ) = VDS D /q ,
where G , S and D are the work functions of the gate, source and drain metallizations, respectively, and VGS is the gate-source voltage. Recent experimental results, which show a strong
dependence of device characteristics on metal work function, indicate that this phenomenological representation of the contacts may be appropriate [28].
63

In this section we wish to draw attention to the following features of CNFETs: ambipolarity,
high conductance, geometry-dependent sub-threshold slope, high ON-current and transconductance.
A closed, coaxial geometry is used, with a gate dielectric of thickness 2.5 nm and a relative
permittivity of 25, as is appropriate for zirconia, which has been employed in some CNFETs [22].
An intrinsic (16,0) nanotube is used, for which the radius, bandgap and electron affinity are
0.63 nm, 0.62 eV and 4.2 eV, respectively [4]. The tube length is taken to be 20 nm, which
should ensure that transport is ballistic over the bias range considered [29]. The effect of
changing the work functions for all three electrodes is examined. The values chosen are 4.5, 4.2
and 3.9 eV, corresponding, in the case of the source/drain electrodes, to barriers for electrons
at the metal/nanotube interfaces that are positive, zero, and negative, respectively. These
electron barrier heights are given by the difference between the metal work function and the
nanotube electron affinity, i.e., Bn = M t . All simulations are performed for a temperature
T = 300 K.
4.5.1 Ambipolarity
Ambipolarity refers to the fact that, under certain bias conditions, channel conduction in CNFETs is due to either electrons or holes [30]. Thus, depending on the bias, a particular CNFET
may be either n-type or p-type. The situation is illustrated in Fig. 4.2 for a CNFET with
positive barrier end contacts. In this case Bn = Eg /2, where Eg is the nanotube bandgap. The
figure shows that, for a constant VDS , the conduction can be due to either electrons or holes,
depending on the gate bias VGS . In this case, at VGS + = VDS /2, where is the difference
in work function between the nanotube and the gate metal, the electron and hole currents are
equal, and the total current attains its minimum value [8], as can be seen on the ID -VGS plot
in Fig. 4.3. Ambipolarity is an undesirable feature in FETs because it leads to an unwanted
OFF current at VGS = 0. Fig. 4.3a shows that, while a negative barrier end contact can reduce
the minimum current, it offers no advantage over a zero-barrier end contact as regards reducing
64
IOF F . This is because at zero gate bias Emid is determined only by G , and so the same barrier
height is presented to the thermionic currents in the two cases (see Fig. 4.4). The OFF current
is smaller in the positive-barrier case because of tunneling. Work function engineering of the
gate metal can be used to laterally shift the I-V curves so that the minimum current occurs at
VGS = 0 (see Fig. 4.3b). Even though the higher Emid that brings this about also reduces the
ON current, the ON/OFF ratio is improved. Clearly, there is opportunity for creative work
function engineering here, and an ON/OFF ratio of around 103 would appear to be possible.
0.4
Energy (eV)
0.2
0
0.2
0.4
0.6
0.8
10
z (nm)
15
20
Figure 4.2: Band diagram illustrating ambipolarity in a device with S,D,G = 4.5 eV and
VDS = 0.4 V. Hole injection at VGS = 0.05 V (dotted line and arrow); electron and hole injection
at VGS = 0.2 V (solid line and arrows); electron injection at VGS = 0.35 V (dashed line and
arrow).
4.5.2 Conductance
The quantized interfacial conductance, as mentioned earlier, has a maximum value Gmax =
4q 2 /h for a nanotube with two transverse modes [20]. Measurements of conductance, G =
ID /VDS , can only be expected to approach Gmax for ballistic transport in nanotubes with
end contacts that have neither ohmic- nor tunneling-resistance. Ballistic transport demands
measurement at low VDS to avoid exciting optical photons at higher biases, and a nanotube
length less than the mean-free-path for acoustical phonon scattering (about 300 nm) [29]. Using
devices of about this length with Pd end contacts, which yield low-resistance contacts with nearzero barrier height for holes, impressive values of G 0.4Gmax have been reported already [28].
65
10
10
(a)
(b)
10
Drain Current (A)
Drain Current (A)
10
10
10
10
10
10
10
0.5
0.25
0
0.25
0.5
0.5
0.25
0
0.25
0.5
Gatesource voltage (eV)
Figure 4.3: ID -VGS at VDS = 0.4 V. (a) G = 4.2 eV and various S,D : 3.9 (solid line); 4.2
(dotted line); 4.5 eV (dashed line). (b) S,D = 3.9 eV and various G : 3.9 (solid line); and
4.37 eV (dashed line).
0.4
0.2
Energy (eV)
0
0.2
0.4
0.6
0.8
1
1.2
0
10
z (nm)
15
20
Figure 4.4: Band diagrams at VGS = 0 and VDS = 0.4 V for the three devices used in Fig. 4.3a.
G = 4.2 eV and various S,D : 3.9 (dotted line); 4.2 (solid line); and 4.5 eV (dashed line).
4.5.3 Subthreshold Slope

The limiting value for the sub-threshold slope S, in situations where the sub-threshold current is thermionically determined, is about 60mQ mV/decade, where mQ = 1 + CQ /Cins is the
quantum capacitance coefficient, with CQ being the quantum capacitance [5], and Cins the
insulator capacitance. The use in CNFETs of high-permittivity dielectrics, such as zirconia [22]
66
or aqueous solutions [24], opens up the possibility of attaining values of mQ approaching unity.
Thus, near-minimum values of S would be expected to be approached in CNFETs with negative
barrier heights at the end contacts. Higher values can be expected for positive barrier heights
due to the presence of tunneling barriers. However, if these barriers are rendered essentially
transparent by a suitable gate bias, then values of S in these devices should also approach the
thermionic limit. In all cases, because the injecting barrier is modulated by the gate voltage
via capacitive coupling, S will show a dependence on the gate/channel geometry [8, 23]. For
CNFETS with positive barrier-height end contacts and insulator thickness tins = 2 nm, for example, S = 110 mV/decade has been measured for planar CNFETs [23], and S 80 mV/decade
has been predicted for coaxial devices, in which the capacitive coupling is superior. For a SiO2
gate oxide with tins = 67 nm, and tubes contacted with palladium, planar devices have been
reported with S 150 mv/decade [28]. With a thinner insulator, it is likely that Pd-contacted
CNFETs will attain values of S close to the theoretical limit.
0
10
CQ/Cins
10
10
0.4
0.3
0.2
0.1
0
0.1
0.2
0.3
0.4
Figure 4.5: Ratio of equilibrium quantum capacitance to insulator capacitance for an insulator
relative permittivity and thickness of 25 and 2.5 nm, respectively, and for all work functions
equal to 4.5 eV.
4.5.4 ON Current
For maximizing the ON current, tunneling barriers must be avoided, and electron injection from
the drain must be suppressed. This situation can be simulated by using a high-enough VDS in
Eq. (4.1) and setting T = 1, which also implies neglecting quantum-mechanical reflection at the
67
60
140
(a)
Transconductance ( S)
Drain current ( A)
50
160
40
30
20
10
0
(b)
120
100
80
60
40
20
0
0 0.1 0.2 0.3 0.4

0 0.1 0.2 0.3 0.4

Figure 4.6: (a) ID , and (b) gm , as a function of gate-source voltage. The solid lines are for
the quantum-capacitance limit from Eqs. (4.2), (4.3) and (4.4). The dashed lines are for a
device with S,D = 3.9 eV and G = 4.37 eV, i.e., the device that gave the lowest OFF current
in Fig. 4.3.
metal/nanotube interfaces. Integrating Eq. (4.1) then leads to
Ie, max =
4q
kB T ln(1 + eEmid /kB T ) ,
h
(4.2)
where kB is Boltzmanns constant. This equation would be more useful if Emid were converted
to an independent parameter, such as VGS . The relationship between these two quantities for
an intrinsic nanotube is
Emid =
Eg
qVGS +
.
2
mQ
(4.3)
Note that the relationship would be essentially linear if CQ Cins , i.e., in the quantumcapacitance limit of mQ 1. This inequality can be readily examined in the mid-tube region,
where CQ = q dQz /dEmid and is easily computed at equilibrium. The result is shown in
Fig. 4.5 from which it is clear that CQ Cins only at low bias, i.e., when there is very little
charge in the nanotube. However, if one allows this inequality to hold to higher bias [27],
then Eq. (4.2)) reduces to a linear form in which the control by VGS is obvious. Such a result is
plotted from Eq. (4.2) and Eq. (4.3) in Fig. 4.6a, where the solid line sets an upper limit to the
unipolar current. As a reality check, the maximum drain current that has been measured so
68
far in a CNFET is about 25 A [28]. Much higher values should be possible with appropriate
work function engineering, as shown by the simulation results of Fig. 4.7. Low electron barriers
at the end contacts give high currents (see Fig. 4.7a), which can be further enhanced at a given
gate bias by reducing the gate work function (see Fig. 4.7b).
The effect of G is examined in the band diagrams of Fig. 4.8, which are for the case of VGS =
VDS = 0.4 V. At G = 3.9 eV, Emid is depressed to the extent that a tunneling barrier forms
for the low energy electrons in the source. With low G , clearly a larger VDS is required to
suppress the drain-injected electron current, leading to a higher saturation voltage VDS,sat . In
the example shown in Fig. 4.7b this could be an issue if CNFET logic circuits were constrained to
operate at 0.4 V, which is the power-supply voltage specified for 10 nm-scale Si MOSFETs [31].
Note that the hole barriers in all cases are too high for ambipolar effects to be seen at the
bias values considered here. Thus, the drain characteristics are of the traditional saturating
variety, with no rapid rise due to hole injection, as has been observed in some experimental
data [32], presumably due to the use of positive-barrier end contacts.
90
80
90
(a)
80
70
Drain current (A)
Drain current (A)
70
60
50
40
30
60
50
40
30
20
20
10
10
0
0
(b)
0
0
0.2
0.4
0.6
0.2
0.4
0.6
Figure 4.7: Drain characteristics at VGS = 0.4 V. (a) G = 4.2 eV and various S,D : 3.9 (dashed
line); 4.2 (solid line); and 4.5 eV (dotted line). (b) S,D = 3.9 eV and various G : 3.9 (dashed
line); 4.2 eV (solid line); and 4.5 eV (dotted line).
In view of the fact that a CNFET with high ON current should also have a low OFF current,
we compute ID for the case of G = 4.37 eV and S,D = 3.9 eV, which led to the low OFF
current shown in Fig. 4.4b. The result is shown in Fig. 4.6a. It can be seen that at VDS = 0.4 V,
69
0.2
0
Energy (eV)
0.2
0.4
0.6
0.8
1
1.2
0
10
z (nm)
15
20
Figure 4.8: Band diagrams for S,D = 3.9 eV, VGS = VDS = 0.4 V and various G : 3.9 (dashed
line); 4.2 (solid line); and 4.5 eV (dotted line).
ID is about 75% of the ultimate value. Finally, it is noteworthy that the highest drain current
shown in Fig. 4.7b) is equivalent to a current density (ID,max /2Rt ) of about 70 mA/m!
4.5.5 Transconductance
Turning now to the maximum attainable transconductance, this limit can be obtained from the
differentiation of Eq. (4.1), which, in the quantum-capacitance limit, yields
Eg /2 qVGS 1
4q 2
1 + exp
gm,max =
.
h
kB T
(4.4)
As Fig. 4.6b reveals, at high VGS , gm, max attains its limiting value of 4q 2 /h, which, interestingly,
is the same value attainable by Gmax , as noted elsewhere [33]. Taking once more the device
with G = 4.37 eV and S,D = 3.9 eV as an example, at VDS = 0.4 V, a value of gm close to
80% of the ultimate value is indicated.
4.6 Conclusions
From this evaluation of the DC performance of carbon nanotube field-effect transistors, it can
be concluded that in n-type devices, for example, the use of negative barrier-height source
and drain contacts, and low work function gate metallization, should allow attainment of subthreshold slopes, conductances, transconductances and ON currents close to the ultimate limits.
70
These features, allied to the excellent thermal and mechanical properties of carbon nanotubes,
make these molecules strong contenders for implementation in nanoscale integrated circuits.
References
[1] Philip G. Collins and Phaedon Avouris, Nanotubes for electronics, Sci. Am., 6269
(2000).
[2] Jing Kong, Jien Cao, Hongjie Dai, and Erik Anderson, Chemical profiling of single nanotubes: Intramolecular p-n-p junctions and on-tube single-electron transistors, Appl.
Phys. Lett., 80(1), 7375 (2002).
[3] Yoshiyuki Miyamoto, Angel Rubio, X. Blase, Marvin L. Cohen, and Steven G. Louie,
Ionic cohesion and electron doping of thin carbon tubules with alkali atoms, Phys. Rev.
Lett., 74(15), 29932996 (1995).
(2002).
2(3), 175180 (2003).
71
[9] Jason Clifford, D. L. John, and David L. Pulfrey, Bipolar conduction and drain-induced
barrier thinning in carbon nanotube FETs, IEEE Trans. Nanotechnol., 2(3), 181185
(2003).
[10] Jing Guo, Jing Wang, Eric Polizzi, Supriyo Datta, and Mark Lundstrom, Electrostatics
of nanowire transistors, IEEE Trans. Nanotechnol., 2(4), 329334 (2003).
[11] Jing Guo, Ali Javey, Hongjie Dai, Supriyo Datta, and Mark Lundstrom, Predicted performance advantages of carbon nanotube transistors with doped nanotubes as source/drain,
(2003). [Online.] Available: http://arxiv.org/pdf/cond-mat/0309039.
[12] Ralph Krupke, Frank Hennrich, Hilbert v. Lohneysen, and Manfred M. Kappes, Separation of metallic from semiconducting single-walled carbon nanotubes, Science, 301,
344347 (2003).
[13] Michael S. Strano, Christopher A. Dyke, Monica L. Usrey, Paul W. Barone, Matthew J.
Allen, Hongwei Shan, Carter Kittrell, Robert H. Hauge, James M. Tour, and Richard E.
Smalley, Electronic structure control of single-walled carbon nanotube functionalization,
Science, 301, 15191522 (2003).
[14] Cees Dekker, Carbon nanotubes as molecular quantum wires, Phys. Today, 52(5), 2228
(1999).
[15] Hyongsok T. Soh, Calvin F. Quate, Alberto F. Morpurgo, Charles M. Marcus, Jing Kong,
and Hongjie Dai, Integrated nanotube circuits: Controlled growth and ohmic contacting
of single-walled carbon nanotubes, Appl. Phys. Lett., 75(5), 627629 (1999).
[16] R. Martel, T. Schmidt, H. R. Shea, T. Hertel, and Ph. Avouris, Single- and multi-wall
[17] Ant Ural, Yiming Li, and Hongjie Dai, Electric-field-aligned growth of single-walled carbon nanotubes on surfaces, Appl. Phys. Lett., 81(18), 34643466 (2002).
72
[18] Won Bong Choi, Jae Uk Chu, Kwang Seok Jeong, Eun Ju Bae, Jo-Won Lee, Ju-Jin Kim,
and Jeong-O Lee, Ultrahigh-density nanotransistors by using selectively grown vertical
carbon nanotubes, Appl. Phys. Lett., 79(22), 36963698 (2001).
[19] Jeroen W. G. Wildoer, Liesbeth C. Venema, Andrew G. Rinzler, Richard E. Smalley, and
Cees Dekker, Electronic structure of atomically resolved carbon nanotubes, Nature, 391,
5962 (1998).
New York, 1995).
12354185 (2003).
2(8), 869872 (2002).
[26] M. Radosavljevic, S. Heinze, J. Tersoff, and Ph. Avouris, Drain voltage scaling in carbon
nanotube transistors, Appl. Phys. Lett., 83(12), 24352437 (2003).
[27] Anisur Rahman, Jing Guo, Supriyo Datta, and Mark S. Lundstrom, Theory of ballistic
nanotransistors, IEEE Trans. Electron Devices, 50(9), 18531864 (2003).
73
[29] Ali Javey, Jing Guo, Magnus Paulsson, Qian Wang, David Mann, Mark Lundstrom, and
Hongjie Dai, High-field, quasi-ballistic transport in short carbon nanotubes, Phys. Rev.
Lett., 92(10), 10680411068044 (2004).
[30] Richard Martel, Hon-Sum Philip Wong, Kevin Chan, and Phaedon Avouris, Carbon nanotube field effect transistors for logic applications, IEDM Tech. Digest, 159162 (2001).
[31] Semiconductor Industry Association, International Technology Roadmap for Semiconductors (2001). [Online.] Available: http://www.itrs.net.
[33] Jing Guo, Supriyo Datta, and Mark Lundstrom, Assessment of silicon MOS and carbon
nanotube FET performance limits using a general theory of ballistic transistors, IEDM
Tech. Digest, 711714 (2002).
74
Chapter 5
A Schr
odinger-Poisson Solver for Modeling
CNFETs
5.1 Introduction
Carbon nanotubes [1] are attracting great interest for their use in nanoscale electronic devices.
Recent modeling efforts of carbon nanotube field-effect transistors (CNFETs) have been successful in examining the subthreshold behaviour of these devices through a simple solution to
Laplaces equation [2, 3], while the above-threshold behaviour has been modeled using bulk
device concepts [4, 5]. Accurate CNFET modeling requires a self-consistent solution of the
charge and local electrostatic potential. In order to properly treat such quantum phenomena
as tunneling and resonance, the charge is computed via Schr
odingers equation. Owing to the
presence of metal-semiconductor interfaces, we also account for the penetration of evanescent
wavefunctions from the metal into the energy gap of the nanotube.
We deal specifically with the coaxial geometry of the CNFET shown previously in Fig. 4.1.
The device consists of a semiconducting carbon nanotube surrounded by insulating material
(relative permittivity ins ) and a cylindrical, wrap-around gate contact. The source and drain
contacts terminate the ends of the device. The device dimensions of note are the gate radius,
c
2004
NSTI http://nsti.org. Reprinted and revised, with permission, from D. L. John, L. C. Castro,
P. J. S. Pereira, and D. L. Pulfrey, A Schrodinger-Poisson Solver for Modeling Carbon Nanotube FETs,
Tech. Proc. of the 2004 NSTI Nanotechnology Conf. and Trade Show , vol. 3, pp. 6568, March 711,
Boston, U.S.A.
75
Chapter 5. A Schrodinger-Poisson Solver for Modeling CNFETs
Rg , the nanotube radius, Rt , the insulator thickness tins = Rg Rt , and the device length, Lt .
In this closed, metallic cylinder system, Poissons equation, restricted to just two dimensions
by azimuthal symmetry, is
2V
1 V
2V
Q
+
+
= ,
2
2
(5.1)
where V (, z) is the potential within the outer cylinder, Q is the charge density, and is the
permittivity. It must be noted that, although the solution of Eq. (5.1) encompasses the entire
volume of the device, we are primarily concerned with the longitudinal potential profile along
the surface of the tube, hereafter labeled VCS (z) V (Rt , z), since knowledge of this potential
is required for carrier transport calculations.
We treat the nanotube as a quasi-one-dimensional conductor, and the linear carrier density is
then computed via the time-independent Schr
odinger equation given by
2
2m
= 2 (E Epot ) ,
2
z
~
(5.2)
where (z, E) is the wavefunction of a carrier with total energy E and effective mass m, traveling
in a region with local effective potential U (z). While Q may include sources such as trapped
charge within the dielectric, we neglect any charge other than that of electrons and holes on
the nanotube.
5.2 Solution Method

We require a solution to Eq. (5.1) with Q = Q(V ). Convergence for this non-linear system is
achieved with the Picard iterative scheme, whereby iteration k + 1 is given by
Vk+1 = Vk L1 rk ,
rk = LVk Q(Vk ) ,
where rk is the residual of the k-th iteration, 0 < 1 is a damping parameter, and L
represents the linear, differential operator allowing Eq. (5.1) to be written as LV = Q.
76
5.2.1 Potential
The boundary conditions for V are given by
V (Rg , z) = VGS G /q ,
V (, 0) = S /q ,
V (, Lt ) = VDS D /q ,
V
(0, z) = 0 ,
(5.3)
where G,S,D represent the work functions of the gate, source, and drain metallizations, respectively, and VGS and VDS are the gate- and drain-source voltages. Due to the discontinuity in
across the nanotube surface, we must also apply the usual matching condition
V
V
q(p n)
ins
t
=
,
R+
R
2Rt 0
t
where p and n are the one-dimensional hole and electron carrier densities, t is the nanotube
relative permittivity, and 0 is the permittivity of free-space.
The solution to Eq. (5.1) was obtained via the finite difference technique, implemented by
discretizing the spatial domain and using central differencing to generate a linear system of
equations, for some known Q, and subject to the boundary conditions specified by Eq. (5.3).
Finite differencing was chosen over an FFT-Greens function approach due to its flexibility
in modeling more complex structures. The singularity at = 0 was addressed by applying
lHopitals Rule to the offending term, yielding
2V
1 V
'
.

2
The amount of energy band bending in the vacuum level, along the length of the nanotube,
is given by Evac (z) = qVCS (z), since we assume that the local electrostatic potential rigidly
shifts the nanotube band structure. The potential energies seen by electrons and holes in the
nanotube are
Epot,e (z) = Ec (z) = Evac (z) t ,
Epot,h (z) = Ev (z) = (Ec (z) Eg ) ,
where Eg and t are, respectively, the nanotube band-gap and electron affinity.
77
(5.4)
5.2.2 Charge
Having established a solution for the potential and its relation to the energy band structure,
we now determine the carrier concentration. In our system, the charge density is given by
Q = q(p n)
( Rt )
,
2
where is the Dirac delta function in cylindrical coordinates, and p(z) and n(z) are computed via
Eq. (5.2), where the nanotube effective mass is obtained from the tight-binding approximation
of the band structure, and is the same for both electrons and holes due to symmetry [1]. Only
the first, doubly-degenerate band is included in the calculations presented herein. The potential
energy, Epot , for each carrier type is specified by Eq. (5.4), given a potential profile VCS (z).
We solve Eq. (5.2) using the scattering-matrix method in which a numerical solution is propagated by cascading 22 matrices [6]. We find that the use of piecewise constant potentials
(plane-wave solutions) are preferable to piecewise linear potentials (Airy function solutions) due
to the considerable reduction of simulation time without an appreciable increase in the error.
Matching of the wavefunction and its derivative on the boundary between intervals n and n + 1,
assuming a constant effective mass, is performed via the usual relations
n = n+1 ,
n+1
n
=
.
z
z
In order to completely specify the wavefunction, we require two boundary conditions. In the
contacts, the wavefunction at a given energy is of the form
AS eikS z + BS eikS z
, z < 0,
=
A eikD z + B eikD z , z > L ,
t
D
D
where kS and kD are the wavevectors in the source and drain contacts, respectively, and AS ,
BS , AD , and BD are constants. As an example, noting that an analogous calculation may be
performed for the drain by exchange of variables, we now illustrate source injection. For this
case, BD = 0 for all energies. In addition, we expect that the Landauer equation [7] will hold
78
for the flux, and must be equal to the probability current. For the transmitted wave, this yields
2q
q~
fS T =
kD |AD |2 ,
~
m
(5.5)
where the pre-factor of 2 accounts for the aforementioned band degeneracy, fS is the FermiDirac carrier distribution in the source, and T is the transmission probability specified by
T =
kD |AD |2
.
kS |AS |2
Simple manipulation yields the normalization condition

|AS |2 =
2m fS
.
~2 kS
(5.6)
At any given energy, multiplication of the unnormalized wavefunction by a constant satisfies

Eq. (5.6).
Including source and drain injection components, the normalized wavefunctions yield the total
carrier densities in the system,
Z
n(z) =
Eref,e
p(z) =
Eref,h
|e,S |2 + |e,D |2
|h,S |2 + |h,D |2
dE ,
dE ,
where the Eref terms are taken to be the bottom of the band, for either electrons or holes,
in the appropriate metallic contact, and correspond to the bottom of the band in the metal.
In practice, the integrals are performed using adaptive Romberg integration, where repeated
Richardson extrapolations are performed until a predefined tolerance is reached [8]. We find
that an adaptive integration method is a necessity for convergence, in order to properly capture
, which is typically highly-peaked in energy for propagating modes. Alternatively, one could
employ a very fine discretization in energy, however the Romberg method allows for the mesh
size to change based on the requirements of the integrand, and results in a much improved
simulation time.
79
5.3 Results
We now present results for a CNFET with a (16, 0) nanotube (Rt 0.63 nm; Eg 0.62 eV),
Lt = 20 nm, tins = 2.5 nm, and ins = 25. All work functions are taken to be 4.5 eV unless
otherwise noted, and t = 4.2 eV. The nanotube is presumed to have a free-space relative
permittivity t = 1 [9], and Eref was taken to be 5.5 eV below the metal Fermi level, as a rough
estimate [10].
Figure 5.1: Net carrier density, p(z) n(z), for the model device as a function of position and
VDS .
Conduction band edge (eV)
0.3
0.2
0.1
0
0.1
0.2
0
5
10
15
20
Figure 5.2: Conduction band edges for the model device with VGS = 0.5 V, and VDS = 0
(dashed) and 0.4 V (solid). Energies are with respect to the source Fermi level.
80
In equilibrium, i.e., for VDS = 0, we obtain reasonable agreement for the carrier concentrations
away from the contacts with that computed using equilibrium statistics [2]. Out of equilibrium,
however, interference effects influence the carrier distributions throughout the device. Fig. 5.1
shows the carrier distributions for VGS = 0.5 V as a function of position and VDS , and Fig. 5.2
shows the corresponding band edges for VDS = 0 and 0.4 eV.
Under a positive gate bias, the band bending results in an increase in the electron concentration
throughout the device as more propagating modes are allowed in the channel. As VDS is
increased, this concentration is considerably reduced in the mid-length region. Evanescent
modes dominate the carrier concentrations near the end contacts, thus impacting on the local
potential. Due to the exponential dependence of the transmission probability on the barrier
shapes, the flux is significantly modified if these modes are neglected.
(a)
Energy (eV)
0.2
0
0.2
0.4
0.6
(b)
Energy (eV)
0.3
0.2
0.1
0
0.1
0.2
0
10
z (nm)
20
0.5
Probability
Figure 5.3: Conduction band edges and transmission probabilities for electrons at VGS = 0.5 V
and VDS = 0.4 V: (a) S = D = 3.9 eV and (b) S = D = 4.5 eV. Energies are with respect
to the source Fermi level.
We note, also, that it is important to allow for the full inclusion of quantum mechanical reflection
for the thermionic component of the flux. Often, carriers above the barrier are assumed to have
a transmission probability near unity. However, this approximation does not hold in general,
81
as Fig. 5.3 shows, wherein the significant reflection is due to Eref being much lower than the
conduction band edge in the nanotube. The effect is most important for devices where the
metal-nanotube work function difference yields a negative barrier, shown in Fig. 5.3(a). Here,
a classical treatment would considerably overestimate the Landauer flux, a function of T , for
energies in the vicinity of the Fermi level.
Finally, the present Schrodinger-Poisson method allows for explicit calculation of the carrier
distribution functions, as shown in Fig. 5.4. The result is in marked contrast to a previous
self-consistent model [5] that utilized quasi-equilibrium distribution functions to calculate the
non-equilibrium carrier concentrations. Moreover, while the model provided in Ref. [4] yields
similar non-equilibrium carrier distributions, it is not equipped to account for the resonant
peaks illustrated here.
Energy (eV)
0.1
0.1
0.2
0
0.2
0.4
0.6
Normalized n(E)
0.8
Figure 5.4: Source-originated electron concentration at Lt /2, normalized to its maximum value.
VGS = 0.5 V, VDS = 0.4 V, and S = D = 3.9 eV. Energies are with respect to the source Fermi
level.
5.4 Conclusions
From this work on the modeling of CNFETs with a coupled Schr
odinger-Poisson solver, we
conclude that:
1. equilibrium statistics are not adequate in describing the carrier distributions in energy;
82
2. consideration of the evanescent modes is crucial for the accurate simulation of devices
where transport is dominated by tunneling through the interfacial barriers;
3. for devices dominated by thermionic emission, a full solution of Schr
odingers equation is
still required in order to account for significant reflection above the barriers.
References
[1] R. Saito, T. Takeya, T. Kimura, G. Dresselhaus, and M. S. Dresselhaus, Raman intensity
of single-wall carbon nanotubes, Phys. Rev. B , 57(7), 41454153 (1998).
2(3), 175180 (2003).
12354185 (2003).
[5] Jason Clifford, D. L. John, and David L. Pulfrey, Bipolar conduction and drain-induced
barrier thinning in carbon nanotube FETs, IEEE Trans. Nanotechnol., 2(3), 181185
(2003).
[6] David Yuk Kei Ko and J. C. Inkson, Matrix method for tunneling in heterostructures:
Resonant tunneling in multilayer systems, Phys. Rev. B , 38(14), 99459951 (1988).
[8] Lee W. Johnson and R. Dean Riess, Numerical Analysis (Addision-Wesley, Don Mills,
Ontario, 1977).
83
84
Chapter 6
An Improved Evaluation of the DC
Performance of CNFETs
6.1 Introduction
In a recent evaluation of carbon nanotube field-effect transistors (CNFETs), devices were specified that yielded simulated drain currents and transconductances approaching the ultimate
limits of a one-dimensional (1-D) ballistic transistor [1]. These devices were coaxial in geometry and had a thin, high-permittivity, gate dielectric. The source and drain metallizations
to the ends of the nanotube imposed negative Schottky barriers to electron flow, and as such,
were predicted to perform much better than CNFETs with positive-barrier end contacts. The
designation of barriers as either negative or positive is used here with respect to electrons.
With appropriate changes in work functions it applies equally to holes, which are the dominant carrier in many experimental devices [2, 3]. Recently fabricated short-channel CNFETs
have shown that, with suitable design, excellent performance is attainable with negative-barrier
metal contacts [4].
The compact non-equilibrium model, from which the results were obtained, allows for quantummechanical tunnelling of electrons and holes at appropriate interfaces, using a simplified expression obtained from the JWKB approximation. Consideration of tunnelling is important
A version of this chapter has been published. L. C. Castro, D. L. John, and D. L. Pulfrey, An Improved
Evaluation of the DC Performance of Carbon Nanotube Field-Effect Transistors, Smart Mater. Struct.,
15, S9S13 (2006). Online at http://www.iop.org/journals/sms.
85
for studying positive-barrier devices, as in the original presentation of the model [5], but, in
negative-barrier devices, attention must be paid to thermionic emission of electrons, and to
their quantum-mechanical reflection at energies above the barrier height. By assigning a transmission probability of unity to all carriers of energy above the barrier, the original compact
model (CM1) is likely to severely overestimate the current.
In this paper, the need for a more detailed treatment of quantum-mechanical reflection by
the compact model is confirmed by examining the correspondence of the latters method of
computing the nanotube charge with that of a solution from Schr
odingers wave equation.
Then, we present a derivation of a tractable expression for quantum-mechanical reflection that
is incorporated into a new version, CM2, of the compact model. The predictions of the latter for
the ON/OFF current ratio, ON current, and transconductance, match more closely the results
from a self-consistent Schrodinger-Poisson solver (SP) [6], and indicate that Schottky-barrier
CNFETs are likely to operate further from the ultimate limit than previously thought.
6.2 Correspondence of the Compact and Quantum Models

In this work, SP is used not only to obtain a better evaluation of the performance of CNFETs,
but also to indicate how CM1 may be improved to achieve the same end. To accomplish the
latter, it is first necessary to establish that the quantum and compact models correspond at
a fundamental level. The basic premise is that there exists a region in the mid-length of the
tube in which the potential energy, Emid , is relatively flat, and serves to connect the regions of
rapidly varying potential energy near to the end contacts. This condition is primarily dependent
on device length, insulator thickness and contact geometries [7]. The decay length for the end
potential is of the order of the gate radius [8], which is about 3 nm in the example used here. A
constant Emid is commensurate with a constant charge in the mid-length region of the nanotube,
and it is under these conditions that we seek to prove the correspondence of the CM and SP
approaches. It is also assumed that ballistic transport applies, which should not be unreasonable
for the tubes of length 20 nm that are considered here [9].
86
In our compact models, the mid-length charge is estimated from a calculation of Emid via a
simple capacitance expression, thereby obviating the use of Poissons equation. Self-consistency
is achieved by reconciling the resulting mid-length charge with that computed from the fluxes of
electrons and holes through and over the interfacial barriers. In the mid-length region, transport
is taken to be phase-incoherent, i.e., the phase of the electron wavefunctions is ignored. At the
interface regions, back-scattering is allowed, leading to a composite transmission probability for
the device,
T? =
TS TD
,
TS + TD TS TD
(6.1)
where TS,D are, respectively, the source and drain transmission probabilities. The mid-length
electron density is given by
Z
nmid =
Emid
g1D T ?
fS
2
2
fD
2
1 +
1 dE,
TD
2 TS
(6.2)
where g1D is the 1-D nanotube density-of-states, numerically computed from a tight-binding
method, f is the Fermi-Dirac distribution (as is relevant for carriers in the metallized regions),
and the subscripts S and D refer, respectively, to source and drain injection. The two terms
in the square brackets can be viewed as the components of nmid arising from electrons injected
from the source (first term) and the drain (second term).
If we now consider just the lowest band, which is doubly degenerate, and introduce into Eq. (6.2)
the effective-mass approximation via
2
4 dk
g1D (E) =
dE
~
2m
,
E Emid
and consider, for brevity, just the component originating at the source, we obtain
Z
1
2m
TS (2 TD )
nmid,S =
fS
dE.
~ Emid E Emid
TS + TD TS TD
(6.3)
Now, from a quantum-mechanical perspective, let the amplitude of a unity-input wavefunction immediately after crossing the source barrier region be given by P . Then, under phaseincoherent transport conditions, and allowing for multiple reflections of these carriers between
87
the source and drain barriers, the total probability density in the mid-length region of the channel due to source injection, |mid,S |2 , can be found from the infinite geometric series relation,
and is given by
|mid,S |2 =
|P |2 + |P |2 (1 TD )
.
TS + TD TS TD
(6.4)
However, the source transmission probability can also be written as

s
jtrans
kmid 2
E Emid
TS
=
|P | =
|P |2 ,
jinc
kS
E Eref,S
where the js are probability density currents, the ks are wavevectors, and Eref,S is the energy
of the conduction band edge of the source metal. Therefore, from Eq. (6.4),
E Emid
TS (2 TD )
=p
|mid,S |2 .
TS + TD TS TD
E Eref,S
Substituting into Eq. (6.3), and including the analogous term for injection from the drain, we
get
nmid
Z
fS |mid,S |2
fD |mid,D |2
2m
p
+p
dE .
=
~ Emid
E Eref,S
E Eref,D
(6.5)
This is precisely the expression for the electron contribution to the mid-length charge that
results from a self-consistent Schrodinger-Poisson solution under the conditions of: a single,
doubly degenerate band; a constant effective mass for both nanotube and end-contact metallization; a nanotube length that is sufficient for the contribution to the charge at mid-length due
to evanescent states to be neglected, and for the transport to be considered phase-incoherent;
and a normalization of the carrier density using the Landauer equation, as in Ref. [6]. The
effective-mass representation of the band structure is employed in CM2; thus, the correspondence of Eq. (6.3) and Eq. (6.5) proves the fundamental equivalence of CM2 and SP under the
stated assumptions.
As regards the actual numerical equivalence of SP and CM2, it can now be appreciated that
this will depend totally on how the transmission probabilities are estimated. Concerning the
numerical equivalence of SP and CM1, this will depend also on the agreement between the
effective-mass- and density-of-states-representations of the band structure. The agreement is
sufficiently good for the single-band case considered here that the numerical difference between
88
SP and CM1 is due almost entirely to the difference in estimating the transmission probabilities
(T s). In CM1 and CM2, the T s are computed for phase-incoherent transport using the JWKB
approximation in the case of tunnelling, and in CM1 are set equal to unity in the case of
thermionic emission. In SP, a full Schr
odinger calculation, under phase-coherent transport
conditions, yields exact values for the T s at all energies. It is not reasonable to expect that
there is a compact expression for T in the phase-coherent case, but one may well exist for
phase-incoherent transport of thermionically emitted carriers, in which case its incorporation
into the compact model should yield a significant improvement. Such an expression, which is
derived in the next section, is incorporated into CM2.
6.3 Quantum-Mechanical Reflection for the Thermionic Case

Typically, the JWKB approximation is used to compute the tunnelling probability for carriers
through a barrier; however, it may also be used to compute the reflection of carriers above
the barrier. For this thermionic current component, we assume the usual JWKB form for the
wavefunction in three regions:
Aeikcon z + Beikcon z
, z < 0,
Rz
Rz
1
(z) =
, 0 < z < w,
Cei 0 kbar (z )dz + Dei 0 kbar (z )dz
k
(z)
bar
F eikmid (zw) + Geikmid (zw)

, z > w,
where A to G are constants, w is the barrier width, and kcon , kbar (z), and kmid are the wavevectors in the contact, in the region of the nanotube close to the contact where the potential may
change significantly, and in the mid-length region of the nanotube where the potential is relatively constant, respectively. Note that only kbar is a function of z. Taking source injection as
an example, we set G = 0, and assume an abrupt change in the band edge when crossing from
the source metal into the nanotube. This permits the usual continuity condition for and its
derivative.
For phase-incoherent transport, we can compute the transmission through the regions close to
the source and drain contacts separately. If we consider the source barrier, for example, we
89

0 (w) = 0, where the primed notation denotes a derivative with
note that kbar (w) = kmid and kbar
respect to z. This provides a compact expression for the source transmission probability,
TS =
3
16kcon kbar0
,
0
2
(kbar0
)2 + 4(kbar0
+ kcon kbar0 )2
(6.6)
where the zero subscript indicates that the quantity is evaluated at z = 0.

An analogous expression holds for the drain transmission probability. The new compact model
(CM2) incorporates Eq. (6.1) and Eq. (6.6), whereas in the original compact model (CM1) T ? =1.

We model the coaxial geometry CNFET illustrated previously in Fig. 4.1. The device consists of
a semiconducting carbon nanotube surrounded by insulating material of relative permittivity
ins , and a cylindrical, wrap-around gate contact. The source and drain contacts terminate
the ends of the device. The device dimensions of note are the device length, Lt , the gate
radius, Rg , and the nanotube radius, Rt . Here we take Rg /Rt = 5, and we consider a (16,0)
tube with Rt = 0.63 nm and Lt = 20 nm. For the relative permittivities, ins = 25, and t ,
which is not relevant to CM1 and CM2, is set to unity in SP [10]. For the work functions,
4.5 eV is taken for the nanotube and the gate, and 3.9 eV is taken for the source and drain.
This arrangement leads to a negative barrier height of approximately one-half of the bandgap,
as used elsewhere in simulations of high-performance CNFETs [1, 11]. All simulations are
performed for a temperature of 300 K.
The gate characteristics are shown in Fig. 6.1. It can be seen that the improved models do not
alter the previous conclusion of Ref. [1] that ON/OFF ratios of around 103 appear possible.
Higher values could result from operating at lower VDS [11]. This is because, with a saturating
ID -VDS characteristic, selection of VDS at the onset of saturation ensures the highest ON
current, yet the low value of VDS delays the onset of hole conduction when VGS is reduced, thus
allowing a lower OFF current to be attained. In practical circuitry it would be desirable to use
a single power supply, so it is to be hoped that metals of suitable work function exist to give a
flat-band voltage such that the minimum in drain current can be engineered to occur at a gate
90
bias of VGS = 0 [1, 11].

2
Drain current (A)
10
10
10
10
10
0.2
0
V
0.2
(V)
0.4
0.6
GS
Figure 6.1: Drain current versus gate-source voltage at VDS = 0.4 V for the various models: SP
(circles), CM1 (dashed), and CM2 (solid).
The results for the ON current and transconductance are shown in Fig. 6.2. The dotted lines
are for the ultimate limit, as defined previously [1, 12]. The shortfall predicted by CM1 is an
indication of how far below this limit CNFETs would perform, even if the transmission probability for all thermionically injected carriers were unity. The further reduction in performance
predicted by SP is due mainly to a more realistic representation of this transmission probability,
T ? (E). The effect is severe and suggests that CNFETs, even with negative barrier heights as
extreme as one-half of the bandgap, are unlikely to come close to performing at the ultimate
limit.
Another revelation of the improved models used in this work is their prediction of a decline in
the transconductance, gm , at high gate bias. This is due to the complicated interaction of the
charge and VGS with Emid [13]. In fact, CM1 predicts a similar decline, but at a much higher
gate bias than CM2 and SP due to its overestimation of the charge on the nanotube.
The actual form of T ? (E) is illustrated in Fig. 6.3. Obviously, CM1 does not capture the
interference phenomena exhibited in the results of SP by virtue of the latters consideration of
phase-coherent transport.
Equally clear is that, unless all the carriers are grouped together at an energy for which T ? (E)
91

60
1
Transconductance (4q2/h)
(a)
Drain current (A)
50
40
30
20
10
0
0
0.3
VGS (V)
(b)
0.8
0.6
0.4
0.2
0
0
0.6
0.3
VGS (V)
0.6
Figure 6.2: (a) Drain current and (b) transconductance, as a function of gate-source voltage at
VDS = 0.4 V. The dotted lines are for the ultimate limit (see text). Other curves illustrate SP
(circles), CM1 (dashed), and CM2 (solid).
Energy (eV)
0.3
0.25
0.2
0.15
0.1
0.05
0
0
0.2
0.4
0.6
Probability
0.8
Figure 6.3: Transmission probabilities, above Emid , of source-injected electrons for the SP
(solid) and CM2 (dashed) models, at VGS = 0.4 V and VDS = 0.4 V.
shows a peak close to unity, then CM1s employment of an energy-independent value of T ? = 1
will lead to a substantial overestimate of the charge and the current. Evidently, this is happening
in the results shown in Fig. 6.2. The employment in CM2 of Eq. (6.6), and its analogue for drain
injection, should lead to some improvement because, even though the expression is derived for
phase-incoherent transport, it does allow for T ? (E) to take on values of less than unity. The
ensuing, greatly improved correspondence in the predictions of the current between the compact
model and SP is demonstrated in Figs. 6.1, 6.2 and 6.4.
It may appear unreasonable to expect that the still-large difference in T ? (E) between CM2 and
92
Drain current (A)
10
8
6
4
2
0
0
0.1
0.2
0.3
0.4
VDS (V)
0.5
0.6
Figure 6.4: Drain current versus drain-source voltage for various CNFET models at VGS =
0.4 V: SP (circles), CM1 (dashed), and CM2 (solid).
SP should allow such an improved concordance in current. However, as quantities of interest,
such as the charge and the current, are computed by performing an integral over energy, some
averaging occurs, and, evidently, leads to a mean value for the SP case that is close to that
predicted by the phase-incoherent analysis. The structure in the SP results for T ? (E) is due to
phase coherence, which will become less important for longer devices, so we would expect the
new compact model to give even better results for tubes with Lt > 20 nm. The converse applies
to shorter tubes, when, additionally, issues due to evanescent charge and direct tunnelling
between source and drain will need to be considered. Operation at lower gate bias may also
lead to the appearance of larger phase-coherence effects, due to the increase in height of the
potential barriers at the end contacts. These phenomena will need to be taken into account in
further compact modeling of CNFETs.
6.5 Conclusions
From this re-evaluation of the DC performance of coaxial carbon nanotube field-effect transistors
with negative-barrier contacts, it can be concluded that:
1. ascribing a value of unity to the transmission probability of thermionically injected carriers
leads to a significant overestimate of the current and transconductance;
2. inclusion of a short expression for quantum-mechanical reflection into a compact model
93
yields much-improved predictions for the current and transconductance, inasmuch as they
are in excellent agreement with results from a comprehensive Schr
odinger-Poisson solver;
3. accounting for quantum-mechanical reflection indicates that these transistors with metallized end contacts may not be capable of operating as close to the ultimate limit as
previously thought.
References
[1] L. C. Castro, D. L. John, and D. L. Pulfrey, Carbon nanotube transistors: An evaluation, Proc. SPIE Conf. Device and Process Technologies for MEMS, Microelectronics,
and Photonics III , vol. 5276, 110 (Perth, Australia, 2004).
[2] J. Appenzeller, J. Knoch, and Ph. Avouris, Carbon nanotube field-effect transistorsan
example of an ultra-thin body Schottky barrier device, Proc. TMS 61st Annual Device
Research Conference, 167170 (Salt Lake City, U.S.A., 2003).
[6] D. L. John, L. C. Castro, P. J. S. Pereira, and D. L. Pulfrey, A Schr
odinger-Poisson solver
for modeling carbon nanotube FETs, Tech. Proc. of the 2004 NSTI Nanotechnology Conf.
and Trade Show , vol. 3, 6568 (Boston, U.S.A., 2004).
94
[8] Sang-Hyun Oh, Don Monroe, and J. M. Hergenrother, Analytic description of shortchannel effects in fully-depleted double-gate and cylindrical, surrounding-gate MOSFETs,
IEEE Electron Device Lett., 21(9), 445447 (2000).
[9] Ali Javey, Jing Guo, Magnus Paulsson, Qian Wang, David Mann, Mark Lundstrom, and
Hongjie Dai, High-field, quasi-ballistic transport in short carbon nanotubes, Phys. Rev.
Lett., 92(10), 10680411068044 (2004).
[11] Jing Guo, Supriyo Datta, and Mark Lundstrom, A numerical study of scaling issues
for Schottky-barrier carbon nanotube transistors, IEEE Trans. Electron Devices, 51(2),
172177 (2004).
[13] D. L. John, L. C. Castro, and D. L. Pulfrey, Quantum capacitance in nanoscale device
modeling, J. Appl. Phys., 96(9), 51805184 (2004).
95
Chapter 7
Quantum Capacitance in Nanoscale Device
Modeling
7.1 Introduction
The concept of quantum capacitance was used by Luryi [1] in order to develop an equivalent
circuit model for devices that incorporate a highly conducting two-dimensional (2D) electron
gas. Recently, this term has also been used in the modeling of one-dimensional (1D) systems,
such as carbon nanotube (CN) devices [2, 3]. Here, we derive expressions for this capacitance
in one- and two-dimensions, showing the degree to which it is quantized in each case.
Our discussion focuses primarily on the 1D case, for which we use the carbon nanotube fieldeffect transistor (CNFET) as the model device, although the results apply equally well to other
types of 1D semiconductors. The 2D case has been discussed in Ref. [1], and is included here
only to illustrate key differences.
Equilibrium expressions are derived, and these are extended to cover two extremes in the
non-equilibrium characteristic, namely: phase-coherent and phase-incoherent transport. In the
former case, the wavefunction is allowed to interfere with itself, and may produce resonances
depending on the structure of the device. This results in the charge, and the quantum cac 2004 American Institute of Physics. Reprinted, with permission, from D. L. John, L. C. Castro, and
D. L. Pulfrey, Quantum Capacitance in Nanoscale Device Modeling, J. Appl. Phys., 96(9), 51805184
(2004).
96
pacitance, becoming strong functions of the length of the semiconductor. In the latter case,
this type of resonance is not allowed, and the quantum capacitance is more uniform. Finally,
we show how the quantum capacitance affects the transconductance of a CNFET, where the
Landauer expression can be used to compute the current [4].
7.2 Equilibrium Quantum Capacitance

In order to derive analytical expressions, it is assumed that our device is in quasi-equilibrium,
and that the carrier distribution functions are rigidly shifted by the local electrostatic potential.
If the density of states (DOS) is symmetric with respect to the Fermi level, EF , as in graphene,
then we can write the charge density, Q, due to electrons and holes in the semiconductor, as
Z
Q=q
0
g(E) [f (E + Eg /2 + qVa ) f (E + Eg /2 qVa )] dE ,
(7.1)
where q is the magnitude of the electronic charge, E is the energy, g(E) is the 1D or 2D DOS,
f (E) is the Fermi-Dirac distribution function, Va is the local electrostatic potential, Eg is the
bandgap, and EF is taken to be mid-gap when Va = 0. The quantum capacitance, CQ , is
defined as
CQ =
Q
,
Va
and has units of F/m2 and F/m in the 2D and 1D cases, respectively.
7.2.1 Two Dimensions

In the two-dimensional case, if we employ the effective-mass approximation with parabolic
bands, the DOS is given by
g(E) =
m
(E),
~2
where (E) is the number of contributing bands at a given energy, m is the effective mass, and
~ is Diracs constant. If we combine this with Eqs. (7.1) and (7.2), and exchange the order of
differentiation and integration, we get
mq 2
CQ =
4~2 kB T
Z
0
2 E + Eg /2 qVa
2 E + Eg /2 + qVa
(E) sech
+ sech
dE ,
2kB T
2kB T
97
where kB is Boltzmanns constant, and T is temperature. If is a constant, we can perform

the integration to get
mq 2
2~2
CQ =
cosh
sinh (Eg /2kB T )
,
Eg /2 + qVa
Eg /2 qVa
cosh
2kB T
2kB T
which reduces to
CQ =
mq 2
,
~2
(7.2)
when Eg = 0 in agreement with Ref. [1], where metallic properties were assumed. Note that
this function is quantized in the metallic case, but continuous for a semiconductor. For Eg
greater than about 15kB T , however, the function makes a rapid transition from a small value
to that given by Eq. (7.2) when Va crosses Eg /2, and is thus effectively quantized.
7.2.2 One Dimension

In the one-dimensional, effective-mass case, we have
r
(E) 2m
g(E) =
.
~
E
The explicit energy dependence of this DOS complicates the evaluation of our integral for CQ .
The approach suggested in Ref. [2], i.e., using the fact that the derivative of f (E) is peaked
about EF in order to approximate this integral using a Sommerfeld expansion [5], cannot be
done in general, due to the presence of singularities in the 1D DOS.
The capacitance is given by
q2
CQ =
2kB T h
m
2
Z
0
(E)
2 E + Eg /2 qVa
2 E + Eg /2 + qVa
sech
+ sech
dE ,
2kB T
2kB T
E
(7.3)
where h is Plancks constant. For sufficiently large |Va |, we can completely neglect one of
the sech2 () terms. As a simple example, if Va = 0.1 V for a material with Eg ' 1 eV, the
contribution to the integral from the first term is roughly four orders of magnitude greater
than the second. This approximation is equivalent to neglecting hole charge for positive Va ,
and electron charge for negative Va . The solid line in Fig. 7.1 shows the equilibrium CQ as a
98
function of Va for a semiconductor with two valence and conduction bands: at 0.2 and 0.6 eV
away from the Fermi level. An effective mass of 0.06m0 is assumed, where m0 is the free-electron
mass. The van Hove singularities, at each band edge, result in corresponding peaks in CQ .
For a linear energy-wavevector relation, such as that near the Fermi level in graphene or a
metallic CN, the DOS is constant. This is the case considered by Burke [3], and is valid when
Va is such that f (E) is approximately zero before the first van Hove singularity is encountered
in the integral. Since the higher energy bands are not relevant to the integration under such a
condition, is constant, and the DOS is given by
g(E) =
2
,
hvF
where vF is the Fermi velocity. The result is

CQ =
2q 2
,
hvF
which agrees with the expression quoted in Ref. [3].

Note that in Eq. (7.3) CQ does not manifest itself as a multiple of some discrete amount, so
quantum capacitance is not an appropriate description for a 1D semiconductor, unlike in the
metallic 2D and metallic 1D CN cases, where the capacitance is truly quantized.
7.3 General Considerations

We can now extend our discussion to include the non-equilibrium behaviour for a general, 1D,
intrinsic semiconductor. All of the numerical results are based on the methods described in
Refs. [6] and [7], which consider the cases when transport in the 1D semiconductor is either
coherent or incoherent, respectively. While these methods were developed in order to describe
CNFETs, their use of the effective-mass approximation allows them to be used for any device,
and bias, where the semiconductor is described well by this approximation.
For phase-incoherent transport, we utilize a flux-balancing approach [7,8] to describe the charge
in an end-contacted semiconductor. If we consider only the electrons that are far away from
99
the contacts, i.e., in the mid-length region, Eq. (7.1) becomes [8]
q
Q=
2
2
g(E)T (E) f (E + Eg /2 qVa )
1
TR (E)
0
2
+f (E + Eg /2 + q (Vbias Va ))
1 dE ,
TL (E)
(7.4)
where Vbias is the potential difference between the end contacts, TL (E) and TR (E) are the
transmission probabilities at the left and right contacts respectively, T ? (E) = TL TR /(TL +
TR TL TR ) is the composite transmission probability for the entire system, and Va is evaluated
in the mid-length region. A similar expression holds for holes.
The first term in Eq. (7.4) resembles the equilibrium case, so we expect a similar form for that
contribution to CQ . The peak for each contributing band will occur at the same Va , but the
overall magnitude will be smaller due to the multiplication by the transmission function. The
second term is also similar except that these peaks will now be shifted by Vbias . This is depicted
by the dashed curve in Fig. 7.1, where the case illustrated by the solid curve has been driven
from equilibrium by Vbias = 0.2 V. Note the splitting of each large peak into two smaller peaks:
one at the same point, and the other shifted by Vbias . Of course, the numerical value of the
non-equilibrium capacitance depends on the exact geometry considered, as it will influence both
Va and the transmission probabilities in Eq. (7.4), but the trends shown here are general and
geometry-independent.
For the coherent, non-equilibrium case, it is instructive to consider a metal-contacted device,
in which the band discontinuities at the metal-semiconductor interfaces are sufficient to allow
significant quantum-mechanical reflection of carriers even above the barrier. Further, we restrict
our attention to short devices since the importance of coherence effects is diminshed as the device
length is increased. Due to the phase-coherence, then, we have a structure very much like a
quantum well, even for devices where tunneling through the contact barriers is not important.
For our device, we expect quasi-bound states to emerge at the approximate energies
En '
n2 2 ~2
,
2mL2
where L is the semiconductor length. For m ' 0.06m0 , such as in a (16, 0) CN, En '
100

0.45
0.4
0.35
CQ (nF/m)
0.3
0.25
0.2
0.15
0.1
0.05
0
1
0.5
0
Va (V)
0.5
Figure 7.1: 1D quantum capacitance as a function of the local electrostatic potential at equilibrium (solid), and for the mid-length region of an end-contacted semiconductor with a bias
voltage of 0.2 V (dashed) between the end contacts. The effective mass is taken to be 0.06m0 ,
and energy bands are situated at 0.2 and 0.6 eV on either side of the Fermi level.
6.3(n/L)2 eV, where L is in nanometres. This may be compared with the result for metallic
CNs, where the linear energy-wavevector relationship yields a 1/L dependence [3,9]. Fig. 7.2(a)
displays CQ as a function of position and Va (in the mid-length region) for this choice of m. The
maxima, indicated as brighter patches, show a dependence on Va that reveals the population
of quasi-bound states. Moreover, the maxima in position clearly show the characteristic modes
expected from our simple square-well analogy. Note that the peak-splitting occurs for coherent
transport as well, as shown in Fig. 7.2(b), where the peaks have been split by Vbias = 0.1 V.
The main difference, between the coherent and incoherent cases, is the presence of the quasibound states. These serve to increase the number of CQ peaks, since each quasi-bound state
behaves like an energy band, and they also give rise to a strong spatial dependence. While
Fig. 7.2 shows only a single-band, coherent result, inclusion of multiple bands would cause CQ
to exhibit peaks corresponding to each band, and to each quasi-bound state.
101
Figure 7.2: 1D quantum capacitance, in arbitrary units, for a short-channel, phase-coherent

semiconductor as a function of position and the local electostatic potential in the mid-length
region for applied bias voltages of (a) 0 and (b) 0.1 V between the end contacts. The bright
areas indicate higher capacitance.
7.4 Application: CNFETs

We now elaborate on the above in the context of CNFET modeling. In particular, for the
purpose of developing compact models, it would be useful to ascertain the conditions under
which the quantum capacitance is small in comparison with that due to the insulator geometry,
a regime previously described as the quantum capacitance limit [2, 10]. To this end, we
examine a coaxial CNFET, and treat Va as the potential, with respect to the source contact,
on the surface of the CN in the mid-length region. CQ can be considered to be in series with
the insulator capacitance Cins [11], however, the ratio of these capacitances is related not to
Va and the gate-source voltage, VGS , but to VGS /Va . If the charge accumulation were linear
over some bias range, as might be deemed appropriate at the local extrema of CQ , we could
relate this ratio directly to the potentials.
Knowledge of the CQ limit is beneficial since a relatively low CQ implies that changes in Va will
closely track changes in VGS , obviating the need to calculate CQ when computing the energy
band diagram. Note, however, that CQ cannot be neglected when considering performance
102
metrics that depend on the total capacitance, i.e., the propagation delay may be dominated
by CQ in this limit. We find that CQ Cins only when Q is small enough as to allow for the
employment of a Laplace solution [12] for the position-dependent potential: eliminating the
need for a cumbersome self-consistent Schr
odinger-Poisson solution. The difference between
these solutions is illustrated in Fig. 7.3 for a coaxial CNFET with an insulator thickness and
CN radius of 2.5 and 0.6 nm, respectively, and an end contact work function that is 0.6 eV less
than that of the CN. Figs. 7.3(a) and (b) correspond to the off and turn-on states, respectively.
While equilibrium band diagrams are shown for simplicity, similar trends prevail with the
application of a drain-source voltage. For a device dominated by thermionic emission, such as
the one depicted here, the discrepancy shown in (a), close to the contacts, will not significantly
affect the current calculation, while in (b), the error would clearly be much greater. For a
device dominated by tunneling, i.e., if the energy bands had the opposite curvature, a similar
discrepancy would result in a large error in the current calculations due to the exponential
dependence of the tunneling probability on the barrier shape.
Energy (eV)
(a)
0.2
0.1
0
0.1
0.2
0.3
0
Energy (eV)
(b)
10
15
20
10
15
20
0.1
0
0.1
0.2
0.3
0
Figure 7.3: Comparison of the equilibrium energy band diagrams, for a model CNFET, at
gate-source voltages of (a) 0.2 and (b) 0.32 V, computed via the solutions to a self-consistent
Schrodinger-Poisson system (solid), and to Laplaces equation (dashed). The Fermi energy is
at 0 eV.
Now, we seek to theoretically quantify the condition under which CQ Cins . From Fig. 7.1, we
see that the first local maxima is on the order of 0.3 nF/m. For this peak to be insignificant,
103
we would require Cins to be orders of magnitude higher than this. For a 2 nm-thick dielectric in
a coaxial device, we would require a relative permittivity of 530 in order to give two orders
of magnitude difference between Cins and CQ . Reports of solid, high-permittivity dielectrics
for CN devices [1316] have quoted values only as high as 175 for the relative permittivity [16],
so we conclude that, for realistic dielectrics, we can expect to only marginally enter the CQ
limit, and that the first CQ peak will be significant. If we consider an electrolytically-gated
CNFET [17], we could perhaps achieve a relative permittivity of 80, and an effective thickness
of 1 nm, as considered in Ref. [2], but this would yield Cins ' 25CQ , and would, again, only
marginally be entering this limit.
For a short-channel, phase-coherent device, the requirement for negligible CQ is that the Fermi
levels for the injecting contacts should be far away from E1 ' 6.3/L2 eV. If we consider positive
applied voltages to the gate and drain, this would imply that qVa should be more than about
5kB T below Eg /2+E1 . For the long-channel or phase-incoherent cases, this condition is given by
E1 = 0, corresponding to the conduction band edge. The relative importance of CQ , computed
in the mid-length region of the device, is depicted in Fig. 7.4 for a phase-incoherent device as
a function of VGS and the drain-source voltage, VDS , where we note that VDS corresponds to
Vbias , and that Va is influenced by both VDS and VGS . Here, the aforementioned peak-splitting
for non-zero VDS is clearly evident in the diverging bright lines. Only for low bias voltages can
CQ be neglected, as shown by the black region in the centre of the figure. However, this figure
also reveals the regions where it becomes approximately constant, i.e., the bias ranges where
the series capacitance relationship can be used to estimate Va from VGS [11]. Note, though,
that this is a single-band calculation, and these regions may not be as prevalent when higher
transverse modes are considered.
Finally, we consider the influence of CQ on the transconductance for our model device, which
has a doubly-degenerate lowest band. If we employ the Landauer equation [4] for transport in
104
Figure 7.4: Quantum capacitance for a long-channel CNFET as a function of the gate- and
drain-source voltages. Numerical values are displayed as a fraction of the insulator capacitance.
two conducting channels, the current is
Z
4q
Tn (E) [f (E) f (E + qVDS )] dE
I=
h
Ec
Z Ev
Tp (E) [f (E) f (E qVDS )] dE ,
where Ec = Eg /2 qVa is the spatially constant conduction band edge in the mid-length region
of a long-channel device, Ev = Ec Eg is the valence band edge, and Tn (E) and Tp (E) are
the transmission probabilities for electrons and holes, respectively, from one end contact to the
other. The transconductance is defined as
gm =
I
,
VGS
which yields
gm
Cins
4q 2
{Tn (Ec ) [f (Ec ) f (Ec + qVDS )]
=
h CQ + Cins
Tp (Ev ) [f (Ev ) f (Ev qVDS )]} .
(7.5)
Note that, if we assume only electron transport with CQ Cins , low temperature, and high
VDS , this expression reduces to the classic Landauer result [4] for two conducting channels
gm =
4q 2
Tn ,
h
105
which is the ultimate transconductance in this case [7, 10]. Fig. 7.5(a) shows the theoretical
transconductance, from Eq. (7.5), for our model device, while (b) and (c) show the energydistribution term (in curly braces), and the capacitance ratio term (in square brackets), respectively. The decrease in gm , at high VGS , is due primarily to the decreasing difference in
the contact distribution functions as, for example, Ec becomes closer to qVDS . However, the
exact magnitude of gm is dependent on the capacitance ratio. Further, CQ will be responsible
for additional oscillations in gm , if higher bands or quasibound states are considered in the
calculation. Such transconductance features have been observed experimentally in Ref. [18],
and have also been described in Ref. [19].
0.8
gm (4q2/h)
(a)
0
0
0.8
Capacitances
Distributions
(b)
(c)
0.4
0.1
0.2
0.3
0.4
0.5
0.6
0.1
0.2
0.3
0.4
0.5
0.6
0.1
0.2
0.3
VGS (V)
0.4
0.5
0.6
0.4
0
0
1.1
0.8
0.5
0
Figure 7.5: (a) Electron transconductance for a model CNFET as a function of the gate-source
voltage for drain-source voltages of 0.2 (solid) and 0.4 V (dashed). Constituent elements of
the theoretical transconductance from Eq. (7.5) are (b) the energy-distribution term (in curly
braces), and (c) the capacitance-ratio term (in square brackets).
106
7.5 Conclusions
From this theoretical study on the charge-voltage relationship in one- and two-dimensional
systems, it can be concluded that:
1. the quantum capacitance occurs in discrete quanta for 2D and 1D metals if Va is such
that the Fermi level falls in a linear portion of the energy-wavevector relationship;
2. for 2D semiconductors, this capacitance is approximately quantized if the bandgap is
greater than about 15kB T , and varies continuously otherwise;
3. for long, 1D systems with parabolic bands, and with Va such that these bands contribute
to the charge density, the equilibrium capacitance exhibits maxima that are related to
the number of contributing bands at a given energy;
4. application of a bias to a 1D semiconductor causes each equilibrium capacitance peak
to split into two smaller peaks, with one remaining at the equilibrium position, and the
other shifting by the applied bias;
5. the potential in the mid-length region of a 1D semiconductor cannot be computed, in
general, from potential division due to two capacitors in series due to the nonlinearity of
CQ ;
6. for short, phase-coherent structures, the quasi-bound states cause the capacitance peaks
to occur at higher local electrostatic potentials, with additional maxima corresponding to
the occupation of these states;
7. for a CNFET, it is unlikely that the insulator capacitance can become high enough to
allow the quantum capacitance to be neglected in energy band calculations, except in
cases where the accumulated charge is low enough that the solution to Laplaces equation
is sufficient for the calculation, or if extremely high permittivity dielectrics are used;
8. the quantum capacitance has a significant effect on the transconductance, and should be
considered when modeling CNFETs.
107
References
[1] Serge Luryi, Quantum capacitance devices, Appl. Phys. Lett., 52(6), 501503 (1988).
[3] P. J. Burke, An RF circuit model for carbon nanotubes, IEEE Trans. Nanotechnol.,
2(1), 5558 (2003).
[7] L. C. Castro, D. L. John, and D. L. Pulfrey, An improved evaluation of the DC performance of carbon nanotube field-effect transistors, Smart Mater. Struct., 15(1), S9S13
(2006).
[9] Zhen Yao, Cees Dekker, and Phaedon Avouris, Electrical transport through single-wall
carbon nanotubes, Mildred S. Dresselhaus, Gene Dresselhaus, and Phaedon Avouris, eds.,
Carbon Nanotubes, vol. 80 of Topics Appl. Phys., 147171 (Springer-Verlag, Berlin, 2001).
108
(2002).
12354185 (2003).
12680111268014 (2002).
[16] B. M. Kim, T. Brintlinger, E. Cobas, M. S. Fuhrer, Haimei Zheng, Z. Yu, R. Droopad,
J. Ramdani, and K. Eisenbeiser, High-performance carbon nanotube transistors on
SrTiO3 /Si substrates, Appl. Phys. Lett., 84(11), 19461948 (2004).
2(8), 869872 (2002).
[18] Minkyu Je, Sangyeon Han, Ilgweon Kim, and Hyungcheol Shin, A silicon quantum wire
transistor with one-dimensional subband effects, Solid-State Electron., 44, 22072212
(2000).
[19] D. Jimenez, J. J. Saenz, B. I
nguez, J. Su
ne, L. F. Marsal, and J. Pallar`es, Modeling of
nanoscale gate-all-around MOSFETs, IEEE Electron Device Lett., 25(5), 314316 (2004).
109
Chapter 8
Method for Predicting ft for CNFETs
8.1 Introduction
Carbon nanotube field-effect transistors (CNFETs) are being seriously considered for meeting
the requirements of the 11-nm technology node [1]. Their DC performance is predicted to
be superior to that of ultimately scaled silicon MOSFETs [2, 3], and impressive values for
drain current and transconductance have already been reported in prototype devices [4]. The
AC capabilities of CNFETs are not yet so obvious. Thus far, measurements on laboratory
devices have been limited by experimental difficulties and parasitics [57]. Thus, the highest
reported frequency of 580 MHz, for operation without signal degradation, cannot be viewed
as a representative value for an intrinsic device1 . One way to investigate the AC capabilities
of CNFETs would be to perform AC simulations with the same rigour that has characterized
earlier DC simulations [9, 10]. This means using a self-consistent Schr
odinger-Poisson solver to
compute values for the parameters appearing in, for example, a small-signal equivalent circuit,
from which a useful metric, such as ft , could be obtained. Through the use of this self-consistent
procedure, we expect a more accurate result than that predicted in Ref. [11], where capacitances
in the equivalent circuit model were computed assuming a metallic nanotube and electrostatics
for an infinitely long coaxial system.
c [2005] IEEE. Reprinted, with permission, from L. C. Castro, D. L. John, D. L. Pulfrey, M. Pourfath,
A. Gehring, and H. Kosina, Method for Predicting fT for Carbon Nanotube FETs, IEEE Trans.
Nanotechnol., 4(6), 699704 (2005).
1
Very recently, operation up to 10 GHz has been reported [8], albeit with considerable signal attenuation.
110
In this paper, we perform a rigorous calculation of the gate-voltage dependencies of both

the transconductance [12], and the capacitances, including the so-called quantum capacitance [13], in order to compute the small-signal, equivalent-circuit parameters, from which
our improved estimates of ft for CNFETs are obtained. This analysis reveals a bias dependence that is quite unusual, and which may prove useful in voltage-controlled, high-frequency
circuitry.
8.2 The Small-Signal Model

8.2.1 Equivalent circuit
Starting from Maxwells first two equations, and considering a system with three electrodes,
through which charge can enter or leave the system, it follows that
Qs Qd Qg
Q
=
+
+
=0,
t
t
t
t
(8.1)
where Q is the total charge within the system and Qs , Qd , and Qg refer to charges associated with each of the devices three terminals, namely: source, drain, and gate, respectively.
Labelling displacement currents with a superscript d, it follows from Eq. (8.1) that
ids + idd + idg = 0 .
The standard notations are used for the voltages, e.g., the total gate voltage vG comprises a
DC voltage VG and an AC small-signal vg . If we suppose that Qs , Qd , and Qg are functions
of time through the application of time-dependent voltages vS (t), vD (t), and vG (t), then each
displacement current will comprise three terms, e.g.,
idg = Cgs
vD
vG
vS
Cgd
+ Cgg
,
t
t
t
where the capacitances come from the set

Cij =
Qi
,
vj
i, j = s, d, g ,
(8.2)
where the minus sign is taken when i 6= j, and the plus sign when i = j [14]. Not all the
capacitances in this set are independent, and it can be easily shown that, for example,
Cgg = (Cgs + Cgd ) = (Csg + Cdg ) .
111
Thus, if we now consider vS and vD to be held constant, as is appropriate for an estimation of

ft in the absence of parasitic source and drain resistances, then it follows that
idg = (Csg + Cdg )
vGS
.
t
This displacement current, which is also the total gate current ig , can be computed from an
equivalent circuit, such as is shown in Fig. 8.1.
ig
G
Cgs
Cgd
gmvgs
is
S
id
D
Figure 8.1: Small-signal, equivalent circuit for the total AC currents, under the condition that
only the potential on the gate is time-dependent.
Turning now to the conduction currents, we employ the standard quasi-static approach for a
current that depends on two potential differences, i.e., vGS and vDS in this case. In its linear
implementation, this leads to a drain conduction current of
icd = gm vgs + gds vds .
Here, as we are keeping vDS constant, the term involving the output conductance gds need not
be considered.
This completes the specification of the small-signal equivalent circuit. It is a well-founded
circuit, with the only approximation being the use of quasi-statics to obtain linear expressions for the conduction currents. Note that the interfacial conductance of 4q 2 T /h, due to
transverse-mode reduction on passing from a large, many-mode electrode to a two-mode, quasi1-D nanotube with a transmission probability T [15], is not shown explicitly in Fig. 8.1, as it is
implicit in the transconductance gm .
On the basis of the circuit shown in Fig. 8.1, the common-source, short-circuit, unity-currentgain frequency, as extrapolated from a frequency at which the square of the gain rolls off at
112
20 dB/decade, is given by
ft =
gm
.
2(Csg + Cdg )
8.2.2 Model parameters

In order to relate Csg and Cdg to meaningful physical quantities, firstly we split the charges Qs
and Qd into two components:
Qs = Qse + Qst
Qd = Qde + Qdt
where Qse and Qde are charges on the actual source and drain electrodes, respectively; and Qst
and Qdt are charges on the nanotube that enter via the source and drain electrodes, respectively.
Each of these charges is supplied by the appropriate displacement current, as illustrated in
Fig. 8.2. A capacitance can now be related to each of the charge components. For the source
capacitance, for example, we have
Csg =
Qs
vG
Qse Qst
+
vG
vG
= Cse + Cst .
These capacitive components are readily calculated because their related charges can be comLg
Source
is
Qse
Lu
Qst
Wrap-around gate
tg
Semiconducting nanotube
Qdt
Insulator
Drain
Qde
id
tins
Lt
Figure 8.2: Charge supply to and through the the source and drain electrodes in a coaxial
CNFET.
puted from a recently described self-consistent DC Schr

odinger-Poisson solver [9]. This solver
113
has been adapted to employ Neumann boundary conditions at the non-metallic bounding surfaces in the structure depicted in Fig. 8.2. In our solver, the source- and drain-related nanotube
charges, Qst and Qdt , are computed from integrations of the line charges that are related to
the wavefunctions associated with carriers communicating with the system via the source and
drain, respectively. Wavefunctions are computed via the effective-mass Schr
odinger equation
with plane-wave solutions assumed in the metal contacts. A phenomenological band discontinuity is used to model the electrode-nanotube heterointerfaces as simple Schottky barriers.
Normalization of the wavefunctions is achieved by setting the probability density current equal
to the current expected from the Landauer equation for ballistic transport. The small-signal
parameters Cst and Cdt are computed for a given drain bias via numerical derivatives with
a perturbation in the gate voltage on the order of 0.1 mV. Similarly, gm is computed from
Landauers equation.
Cse is associated with the change in the charge that resides on the actual source electrode Qse .
Similarly, Cde is related to a change in Qde . These charges are computed from appropriate
applications of Gauss Law in integral form.

Results are presented for a coaxial transistor structure, as shown in Fig. 8.2, with Schottkybarrier contacts at the source/tube and drain/tube interfaces. This embodiment, which avoids
the need to dope the nanotube, and which employs the ultimate multi-gate to combat shortchannel effects, is being actively pursued experimentally [16]. Here, by way of an example,
we consider a (16,0) carbon nanotube with a radius of 0.63 nm, a length Lt = 20 nm, and a
relative permittivity of 1 [17]. The insulator has a thickness tins = 2.5 nm, and its relative
permittivity is taken to be 25, as is appropriate for zirconia, which is used in some highperformance CNFETs [18]. The gate electrode is separated from the source and drain electrodes
by Lu = 4 nm, and has a thickness tg = 3 nm. The work function of the gate is taken to be the
same as that of the intrinsic nanotube (4.5 eV), whereas the source and drain metallizations
have a work function of 3.9 eV. Although this yields an n-type device, inasmuch as the dominant
114
carriers are electrons, p-type operation is directly analogous, and, owing to the symmetry of
the nanotubes band-structure about the midgap energy, can be obtained through the use of a
higher work function metal. Thus, the CNFET considered here can be classed as a negativeSchottky-barrier device, such as has been predicted to give DC characteristics that are superior
to those of devices with either zero- or positive-Schottky barriers at the end contacts [2]. In
negative-barrier CNFETs the conduction current is due to thermionic emission at the sourcetube and drain-tube interfaces. Quantum-mechanical reflection at these interfaces, due to the
band discontinuities mentioned earlier, leads to resonances, and the appearance of quasi-bound
states, at least in nanotubes of the short length considered here. This plays a significant role
in determining the values for the model parameters discussed below. An illustrative example
of the charged quasi-bound states, and the conduction-band profile in the device, is presented
in Fig. 8.3. The example shows charge in the first and second quasi-bound states, and the
appearance of charge in an additional quasi-bound state in the potential well at the drain end
of the device.
0.1
0
Energy (eV)
0.1
0.2
0.3
0.4
0.5
5
10
z (nm)
15
20
25
Figure 8.3: Grey-scale representation of the energy- and position-dependence of the electronic
charge in the nanotube at VGS = 0.4 V and VDS = 0.5 V. The uniform columns to the left and
right represent the energy range below the Fermi level in the source and drain, respectively.
The conduction-band edge is superimposed (solid line).
The results presented here are intended to illustrate the ability of the proposed method to
provide meaningful estimates of ft for CNFETs. An optimization of the CNFET structure to
115
suggest an upper bound for ft is not attempted at this stage, but some comments are offered
after the discussion of the present results as to the factors that might be important in this
regard. All the results presented below are for operation at VDS = 0.5 V.
The various components of the capacitance are shown in Fig. 8.4. Considering, firstly, the tube
Capacitance (aF)
6
5
CSE
CDE
CST
CDT
2
1
0
0
0.2
0.4
0.6
Gate voltage (V)
0.8
Figure 8.4: Components of the source and drain capacitances.
capacitance Cst due to charge injected from the source, the peak at around VGS = 0.35 V
corresponds to alignment in energy of the source Fermi level and the first quasi-bound state
for electrons in the nanotube [13]. The peak in Cdt is displaced from the peak in Cst by
approximately VDS [13], and corresponds to alignment in energy of the drain Fermi level and
the first quasi-bound state.
Considering now the capacitances associated with changes in charge on the actual end contacts,
it can be seen that these are relatively bias-independent. In fact, this is due to Cse and Cde being
dominated by the regions of overlap of the end contacts with the edges of the gate electrode.
Obviously, this capacitance could be reduced by increasing the separation between the end
contacts and the edges of the gate, or by making the gate electrode thinner, or by making
the end contacts more needle-like [19]. The latter could be achieved by utilizing metallic
116
nanotubes for the source and drain. The total capacitances associated with each electrode are
shown in Fig. 8.5(a).
Frequency (THz)
Capacitance (aF)
12
(a)
CSG
10
CDG
8
6
0.4
(b)
0.3
0.2
0.1
0
0.6
(c)
0.4
0.2
0
0
0.2
0.4
0.6
Gate voltage (V)
0.8
Figure 8.5: ft and its components. (a) The total capacitances associated with the source, Csg ,
and with the drain, Cdg ; (b) transconductance; (c) ft .
We now discuss the transconductance, as shown in Fig. 8.5(b). Firstly, note that the choice of
end-contact work function renders the device unipolar, except at very low bias. Thus, the hole
contribution to the transconductance at moderate and high VGS is negligible. Secondly, it can
be seen that gm reaches a maximum, and then decreases as VGS increases. This phenomenon has
been reported elsewhere [12,13]. The overall reduction in gm at high VGS relates to the increasing
electron injection from the drain as the potential energy in the mid-length region of the tube is
reduced. The considerable structure in the transconductance plot is due to the presence of the
117
quasi-bound states referred to earlier. As VGS increases, the conduction band edge is pushed
below the source Fermi level, and as the quasi-bound states cross this level, gm increases. Thus,
the situation is not dissimilar to that which gives rise to the peaks in capacitance. Indeed, at low
temperatures, our simulations reveal that the peaks in transconductance and capacitance do
occur at the same biases (see Fig. 8.6). Evidently, in going from T=4 K to T=300 K, thermal
broadening causes peaks that are close together to merge, with the taller one dominating.
Thus, the second peak in transconductance dominates the first, while the opposite is true in
the capacitance case.
16
(a)
CSG T=300K
Capacitance (aF)
14
CDG T=300K
CSG T=4K
12
CDG T=4K
10
8
0.6
T=300K
T=4K
(b)
0.4
0.2
0
0
0.2
0.4
0.6
Gate voltage (V)
0.8
Figure 8.6: Bias dependence, at two temperatures, for (a) capacitance, and (b) transconductance. Note that the peaks in transconductance coincide with peaks in capacitance at the lower
temperature.
The changes in capacitance and transconductance discussed above lead to a very interesting
and unusual bias dependence in the cut-off frequency ft , as illustrated in Fig. 8.5(c). For the
118
example of a 20 nm tube, as used here, ft peaks at about 600 GHz. This is a long way from the
value of 4 THz, which can be inferred from a recent model that ignored the bias dependence
of the transconductance and capacitances, and attributed the device capacitance to that of an
infinite coaxial system, with the quantum capacitance given by that of a metallic, rather than
a semiconducting, nanotube [11]. In the finite coaxial system considered here, the mid-tube
quantum capacitance is not explicitly identified, as it does not relate to the terminals on which
the equivalent circuit is based. It is contained within Cst and Cdt , the peak values of which turn
out to have comparable magnitudes to the electrostatic electrode capacitances, Cse and Cde , in
this particular example; thus, the overall capacitance shows significant bias dependence.
In future studies, we will attempt to optimize the Schottky-barrier CNFET as regards highfrequency performance. However, in concluding this paper, we can make a few comments
regarding the parameters that are likely to be of importance. Clearly, the magnitude of the band
discontinuity at the end contacts is significant. We have used a value of -5.5 eV for the depth
of the metal conduction band below the Fermi level [9]. Higher values may be appropriate for
noble metals of the type that appear suited to end contacts for CNFETs, in which case one can
expect more quantum mechanical reflection and a lower transconductance, leading to a reduced
ft . Increasing the barrier height by increasing the work function of the end-contact metal (in
the case of n-type devices) will significantly degrade performance because of the appearance of
a thick tunneling barrier in the ungated portion of the nanotube. Changing the nanotube to
one of larger bandgap, yet maintaining the barrier height at about Eg /2, may also degrade
transconductance, at a given bias, because the ON current can be expected to be smaller, at
least at low gate bias. Further, the peaks in transconductance and capacitance will be displaced
to higher VGS as more depressing of the conduction-band edge under the gate will be required
to align the quasi-bound states with the source Fermi energy. Increasing the ungated regions
Lu should be advantageous in a negative-barrier device because the electrostatic electrode
capacitance will be reduced without a degradation in gm . We have been quite aggressive in
the vertical scaling of our device as we have used a high permittivity and a small thickness
for the gate insulator. Relaxing these values will not change the resonances, but will shift the
119
peaks in capacitance and transconductance to higher biases, due to the poorer electrostatic
coupling between gate and nanotube. Finally, we should mention that in non-Schottky barrier
CNFETs, in which the source and drain regions are formed by doping the ungated portions
of the nanotube [20, 21], potential wells will form between the end contacts and the intrinsic,
gated part of the nanotube, and could lead to resonances somewhat similar to those described
here if the doped regions are short enough.
8.4 Conclusions
From this work on AC small-signal simulations of Schottky-barrier carbon nanotube field-effect
transistors, it can be concluded that:
1. the generic small-signal, equivalent-circuit model for FETs is appropriate for studying the
quasi-static AC performance of CNFETs, provided the model parameters are rigorously
derived;
2. in the case of short nanotubes with Schottky-barrier end contacts, a resonant structure
is formed, leading to the appearance of quasi-bound states;
3. the quasi-bound states lead to gate-bias dependencies of the capacitances and transconductance, which, in turn, give rise to a short-circuit, unity-current-gain frequency ft ,
which displays a dependence on VGS that is unusual in its oscillatory nature.
References
[1] Ken
David,
Silicon
research
at
Intel,
(2004).
[Online.]
Available:
ftp://download.intel.com/research/silicon/Ken David GSF 030604.pdf.

120
[5] David J. Frank and Joerg Appenzeller, High-frequency response in carbon nanotube fieldeffect transistors, IEEE Electron Device Lett., 25(1), 3436 (2004).
[6] J. Appenzeller and D. J. Frank, Frequency dependent characterization of transport properties in carbon nanotube transistors, Appl. Phys. Lett., 84(10), 17711773 (2004).
[7] Dinkar V. Singh, Keith A. Jenkins, J. Appenzeller, D. Neumayer, Alfred Gill, and H.S. Philip Wong, Frequency response of top-gated carbon nanotube field-effect transistors,
IEEE Trans. Nanotechnol., 3(3), 383387 (2004).
[8] X. Huo, M. Zhang, Philip C. H. Chan, Q. Liang, and Z. K. Tang, High frequency S
parameters characterization of back-gate carbon nanotube field-effect transistors, IEDM
Tech. Digest, 691694 (2004).
[10] Jing Guo, Supriyo Datta, and Mark Lundstrom, A numerical study of scaling issues
for Schottky-barrier carbon nanotube transistors, IEEE Trans. Electron Devices, 51(2),
172177 (2004).
[11] Peter J. Burke, AC performance of nanoelectronics: Towards a ballistic THz nanotube
transistor, Solid-State Electron., 48, 19811986 (2004).
(2006).
121
[14] Yannis P. Tsividis, Operation and Modeling of the MOS Transistor (McGraw-Hill, Toronto,
1987).
New York, 1995).
[16] Wolfgang Hoenlein, Franz Kreupl, Georg Stefan Duesberg, Andrew Peter Graham, Maik
Liebau, Robert Viktor Seidel, and Eugen Unger, Carbon nanotube applications in microelectronics, IEEE Trans. Compon. Pack. T., 27(4), 629634 (2004).
[19] Enzo Ungersboeck, Mahdi Pourfath, Hans Kosina, Andreas Gehring, Byoung-Ho Cheong,
Wan-Jun Park, and Siegfried Selberherr, Optimization of single-gate carbon-nanotube
field-effect transistors, IEEE Trans. Nanotechnol., 4(5), 533538 (2005).
[21] Yu-Ming Lin, Joerg Appenzeller, and Phaedon Avouris, Novel carbon nanotube FET
design with tunable polarity, IEDM Tech. Digest, 687690 (2004).
122
Chapter 9
High-frequency Capability of
Schottky-Barrier CNFETs
9.1 Introduction
Carbon nanotube field-effect transistors (CNFETs) are predicted to have superior DC characteristics to those of foreseeable silicon MOSFETs [1, 2]. Experimentally, impressive values
for drain current and transconductance have already been reported in prototype devices [3].
However, the AC performance of CNFETs has not yet received much attention [4], and the
likelihood of high-frequency operation needs to be established. Here, we take a step in this
direction by predicting two useful figures of merit for high-frequency transistors, namely: the
unity current-gain and unity power-gain frequencies, fT and fmax , respectively. We employ the
standard small-signal method [5], abetted by a self-consistent Schr
odinger-Poisson solver [6],
in order to obtain preliminary estimates of these valuable figures of merit. In particular, we
consider the cylindrically gated device depicted in Fig. 9.1. The CNFET geometry consists of
a nanotube with radius Rt , coated by an insulator of thickness tins , wrapped by a cylindrical
gate of thickness tg , and capped at the ends by planar source and drain contacts of radius
tins +tg +Rt . These end contacts are separated longitudinally from the gate by a distance Lu .
A version of this chapter has been accepted for publication. L. C. Castro, D. L. Pulfrey, and D. L.
John, High-Frequency Capability of Schottky-Barrier Carbon Nanotube FETs, Solid-State Phenomena
(2005). Accepted December 22, 2005.
123
Lg
tg
Wrap-around gate
tins
Lu
Source
Drain
Semiconducting nanotube
Rt
Insulator
Lt
Figure 9.1: Structure of the modeled CNFET.
9.2 Model
Since the CNFET is much like a traditional field-effect transistor, we may employ the standard
equivalent circuit model for this device [5], where the small-signal parameters themselves, such
as the transconductance and the various transcapacitances, are computed by taking numerical
derivatives based on the results of self-consistent DC charge-voltage calculations [6,7]. Through
our use of a Schrodinger-Poisson solver for these DC results, we are able to include the effects
of geometry, quantum capacitance, and spatial non-uniformity of the charge in our calculations. While we have previously presented results for the intrinsic ft with only three circuit
elements [7], here we consider the additional effects of series resistances associated with each of
the terminals, and we include the remaining capacitance components. This allows us to compute the extrinsic fT , and also to consider fmax . For clarity, we include the standard equivalent
circuit in Fig. 9.2, where we note that the subscripts s, d and g refer to the source, drain and
gate, respectively. In the usual way, we consider the fT and fmax expressions, which may be
arrived at by extrapolating the characteristic decay in gain, from its value at an appropriate
low frequency, to the 0 dB point.

The model device in this study employs a (16,0) nanotube of radius 0.63 nm, length 30 nm,
bandgap 0.62 eV, unity relative permittivity, and effective mass 0.06 m0 , where m0 is the free
124
g
Rg
Cgs
Cgd
gds
Cmvgs
gmvgs
Csd
D
Rd
s
Rs
S
Figure 9.2: Full small-signal equivalent circuit for the CNFET. The transcapacitance Cm =
Cdg Cgd .
electron rest mass. Moreover, tins is 2.5 nm, tg is 3 nm, the insulator permittivity is 25 as is
appropriate for zirconia [8], and Lu is 5 nm for this initial investigation. The work function of
the nanotube is taken to be 4.5 eV, while that of the end-contacts is 3.9 eV, yielding negativeSchottky barrier, n-type transistor operation. All results are for a drain-source voltage of
0.5 V. Fig. 9.3 illustrates the intrinsic parameters for our device. The oscillatory peaks in the
capacitances and transconductance are related to the formation of quasi-bound states in the
short channel [7, 9]. With the modulation of the applied voltage, the states, indicated by the
bright patches in Fig. 9.4, are shifted in energy by band bending in the channel. As they cross
the source or drain Fermi level, they become populated, and we obtain peaks in the charge
accumulation and, consequently, in the capacitances. Since the charge accumulation affects the
amount of band bending in the channel through our self-consistent DC calculations, we also see
peaks in the transconductance gm . The gm behaviour is complicated due to its dependence both
on the charge and on the Fermi distributions at the injecting contacts. In Fig. 9.5, we present
our main results, the predictions of fT and fmax for our model CNFET. Fig. 9.5(a) shows fT for
Rs = Rd set to 1 k, 10 k, and 100 k, and we recall that Rg has no influence on this figure
of merit. Note that the values of the parasitic resistors Rs and Rd are chosen to be comparable
to the contact resistance that results from mode constriction when carriers pass from a manymoded material to a material with only a few modes [10]. In the carbon nanotube case, we
consider the lowest two degenerate modes in energy for an equivalent contact resistance of
around 6-7 k. Note that the contact resistance is automatically included in our self-consistent
125
0.4
Transconductance (4q /h)
Cgs
20
Capacitance (aF)
Cgd
C
15
sd
Cm
10
5
0
0
0.2
0.4
0.6
0.8
Gate voltage (V)
0.3
0.2
0.1
0
0
0.2
0.4
0.6
0.8
Gate voltage (V)
Figure 9.3: Capacitances and transconductance for the model device.
0.02
Energy (eV)
0
0.02
0.04
0.06
0.08
0.1
0
10
15
z (nm)
20
25
30
Figure 9.4: Charge density for the model device subject to a gate-source voltage of 0.38 V.
Brighter patches indicate higher charge density, while a portion of the conduction band edge is
shown by the white line. The energy values are referenced to the source Fermi level.
DC calculations, so an explicit resistor is not needed to represent it in the equivalent circuit..
Thus, the resistors shown in Fig. 9.2 are solely parasitic. It is evident that the maximum value
of fT occurs at the first peak in gm . Turning now to fmax , shown in Fig. 9.5(b), we focus on
the effect of Rg , and show results for Rg set to 100 , 1 k, and 10 k, with Rs = Rd = 10 k.
Again, a pronounced peak coincides with the first peak in gm .
126
1 kOhm
10 kOhm
100 kOhm
(a)
600
800
fmax (GHz)
fT (GHz)
800
400
200
0
0
100 Ohm
1 kOhm
10 kOhm
(b)
600
400
200
0.2
0.4
0.6
0.8
Gate voltage (V)
0
0
0.2
0.4
0.6
0.8
Gate voltage (V)
Figure 9.5: Extrapolated figures of merit: (a) fT with Rs = Rd set to 1 k (solid), 10 k

(dashed), and 100 k (dotted), and (b) fmax with Rs = Rd = 10 k for Rg set to 100 (solid),
1 k (dashed), and 10 k (dotted).
9.4 Conclusions
From this simulation of the high-frequency performance of CNFETs, it can be concluded that,
in short-channel devices, with negative-barrier Schottky contacts for the source and drain, and
with very short ungated regions, and where coherent transport is possible, resonance effects
can lead to a strong bias dependence of the high-frequency figures of merit, fT and fmax . At
the resonance peaks, these frequencies may reach the THz level.
References
(2006).
127
[4] Dinkar V. Singh, Keith A. Jenkins, J. Appenzeller, D. Neumayer, Alfred Gill, and H.S. Philip Wong, Frequency response of top-gated carbon nanotube field-effect transistors,
IEEE Trans. Nanotechnol., 3(3), 383387 (2004).
1987).
[7] L. C. Castro, D. L. John, D. L. Pulfrey, M. Pourfath, A. Gehring, and H. Kosina, Method
for predicting fT for carbon nanotube FETs, IEEE Trans. Nanotechnol., 4(6), 699704
(2005).
New York, 1995).
128
Chapter 10
Extrapolated fmax for CNFETs
10.1 Introduction
The frequency fmax , at which the extrapolated power gain becomes unity, is a well-established
figure-of-merit for characterizing the high-frequency performance of transistors. A useful, compact expression for fmax is available for heterojunction bipolar transistors [1], but, in Si metaloxide-semiconductor field-effect transistors, the need to consider the electrical properties of
the substrate makes for a more complicated situation. However, in carbon nanotube field-effect
transistors (CNFETs), the substrate is not an active part of the device, so the traditional, smallsignal equivalent circuit, in which there are no elements representing the substrate [2, p.441],
can be used as a basis for deriving a useful expression for fmax . Moreover, because of the small
size of CNFETs, the quasi-static approximation should be valid up to very high frequencies.
Here, starting from the small-signal parameters of the equivalent circuit, we systematically
make a series of approximations that lead to compact expressions for the extrapolated fmax .
These expressions are shown to be applicable over a wide range of conditions, and to be useful
in guiding the design of high-frequency devices.
A version of this chapter has been published. L. C. Castro and D. L. Pulfrey, Extrapolated fmax
for Carbon Nanotube Field-Effect Transistors, Nanotechnology, 17(1), 300304 (2006). Online at
http://www.iop.org/journals/nano.
129
10.2 Modeling Procedures

The small-signal, extrinsic z-parameters for the equivalent circuit previously shown in Fig. 9.2
are given by the standard expressions [2, p.440]:
z11e = y22 /Y + Rsg
z12e = y12 /Y + Rs
z21e = y21 /Y + Rs
(10.1)
z22e = y11 /Y + Rsd

Y
= y11 y22 y12 y21 ,
where Rsg = Rs + Rg , Rsd = Rs + Rd , and the intrinsic y-parameters are [3, p.378]:
y11 = j(Cgs + Cgd )
y12 = jCgd
(10.2)
y21 = gm j(Cm + Cgd )

y22 = gds + j(Csd + Cgd ) .
The transcapacitance Cm relates non-reciprocal capacitance pairs, and is given by, for example, Cdg Cgd . From the components in Eq. (10.1), an expression for the radian frequency
T at which the short-circuit, common-source, current gain reaches unity, when extrapolated
from some lower frequency at which the square of the gain rolls-off at -20 dB/decade, follows,
namely [2, p.441]:
1
1
= [1 + gds Rsd ] + Rsd Cgd ,
T
t
(10.3)
where the intrinsic cut-off frequency is given by

t =
gm
.
(Cgs + Cgd )
(10.4)
The assumptions made in arriving at the expressions for the extrapolated unity-current-gain
frequencies are that the frequency at which the extrapolation can properly begin is subject to
130
the following restrictions:

2
2 gm
/(Cm + Cgd )2
(10.5)
2 gm /(ARs )
(10.6)
2
2 gm
/(Cdg + BRs )2
(10.7)
2 (Cgg + BRsd )2 /(ARsd )2 ,
(10.8)
where
A = Cdg Cgd Cgg Cdd
B = Cgg gds + Cgd gm
Cgg = Cgs + Cgd
Cdd = Csd + Cgd .
The above limitations should easily be satisfied by CNFETs intended for operation at frequencies of several hundreds of GHz . For the power gain, we use Masons unilateral gain [4]:
U=
|z21e z12e |2
.
4 [<(z11e )<(z22e ) <(z12e )<(z21e )]
(10.9)
One further restriction on the extrapolation frequency is required to obtain an expression for
U with the desired dependence on 2 :
2 B 2 /A2 .
(10.10)
In fact, Eq. (10.10) is a more stringent restriction than the fourth assumption of Eq. (10.8),
which is, therefore, rendered redundant.
It follows, after lengthy algebraic manipulation, that U can be written in the form that is
familiar for bipolar transistors [1]:
U=
T
2
4 (RC)eff
(10.11)
where (RC)eff is an effective time constant. Because the assumptions we wish to make in
order to simplify the expression for fmax affect both T and (RC)eff we elect not to isolate the
131
expression for the latter [1], but, instead, to work on the expression for the reciprocal cyclic
frequency eff , where
2
eff
=
(RC)eff
,
T
(10.12)
1
.
4eff
(10.13)
and
fmax =
With the assumptions made so far, eff is given by

2
eff,1
=
B
2
gm
{Rg (Cgg + 2Rc B) + Rc [Cgg Cm
A
+2Csd + Rc B + B
(gm + 2gds ) ,
(10.14)
where, for convenience, we have assumed similar source and drain contacts, and set Rs = Rd =
Rc .
To make progress in simplifying Eq. (10.14), one has to compare component values, which,
because of their bias- and device-dependence, cannot be expected to result in relations that
are as generally applicable as the frequency limitations stated earlier in Eqs. (10.5)-(10.8) and
(10.10). We start by asserting
Cgs = Cgd .
(10.15)
The motivations for doing this are the small size and longitudinal symmetry of CNFETs: the
electrodes are inevitably very close together, so the extrinsic contributions to Cgs and Cgd will
be significant; and the symmetry would make them equal. We can anticipate this equality
breaking down at low- and high-gate bias, when the electrode-dependent quantum-capacitance
contribution to Cgs and Cgd , respectively, is particularly significant [5]. Using Eq. (10.15) in
Eq. (10.14) leads to a considerable simplification:
2
Cgd
2gds
2
1+
(2Rg + Rc )[1 + Rc (gm + 2gds )] .
eff,2 =
gm
gm
(10.16)
Finally, in the interests of further simplification, we suggest:

gm 2gds .
(10.17)
This inequality may break down in CNFETs with small-diameter (large-bandgap) nanotubes,
for which the transconductance is generally less than in those with large-diameter tubes. The
132
result of this additional assumption is a very compact expression:

2
eff,3
=
2
Cgd
gm
(2Rg + Rc )(1 + gm Rc ) .
(10.18)
In the next section we evaluate the validity of Eqs. (10.14), (10.16) and (10.18) for several
Schottky-barrier CNFETs. The component values are evaluated as described previously [5],
using a Schrodinger-Poisson solver [6], with the inclusion of the complex band structure of the
nanotube [7]. We found that it was not necessary to consider more than the lowest, doublydegenerate band for the tubes and bias ranges considered in this work.

In seeking ultimate performance limits we examine devices of the coaxial structure shown
previously in Fig. 1.8, but we base values for the physical properties on those of presently
realizable planar structures, such as a recent, high-DC-performance device [8]. All the devices
considered here have a gate of length Lg = 50 nm and of thickness tg = 20 nm, an insulator
relative permittivity of 16 (HfO2 ), and Pd end-contacts of radius tc = 4 nm. Unless otherwise
stated, the contact length is Lc = 100 nm, the gate underlaps are Lus = Lud = 5 nm, and the
contact resistances are computed from a Pd resistivity of 0.48 knm, which can be inferred from
Ref. [8]. The data of Ref. [9] was used for the tube-dependent, end-contact barrier heights, while
the work function of the gate was set equal to that of the nanotube [10]. The latter assignment
is arbitrary in view of the lack of information on other factors, such as oxide charge, that will
affect the threshold voltage in practice, and serves only to change the effective gate potential.
The gate resistance can be expected to have a large effect on fmax [11], but, in the present
absence of knowledge about practical gate connection configurations, we take, unless otherwise
stated, Rg = 1 k. With the dimensions listed above, for example, Rc 0.9 k. The trends
in the capacitances illustrated in Fig. 9.2 have been discussed previously [5, 12], and for the
particular devices described here Cgs and Cgd are on the order of 10 aF.
Firstly, we consider Device 1, which has a nanotube diameter dt = 1.7 nm (taken to correspond to a tube of chirality (22,0)), for which the Pd contacts produce a negative barrier for
133
U (dB)
Unilateral power gain (dB)
100
10
0
10
11
50
10
12
10
Frequency (Hz)
0
8
10
10
10
11
10
10
Frequency (Hz)
12
10
13
10
Figure 10.1: Unilateral power-gain for Device 1 using Eq. (10.9) (solid), and Eq. (10.11) with
Eq. (10.14) (dots), Eq. (10.16) (circles), and Eq. (10.18) (crosses). The inset magnifies the curves
near 0 dB. VGS = VDS = 0.5 V.
holes of 0.04 eV [9]. The combination of low barrier height and an insulator thickness of
tins = 2.5 nm should produce a device of high transconductance. Masons power gain U is
shown in Fig. 10.1, from which it is clear that, for this particular device at the given biases of
VGS = VDS = 0.5 V, all the assumptions leading to Eqs. (10.14), (10.16) and (10.18) are reasonable. The effects of the assumptions do appear, however, at different VGS , as illustrated in
Fig. 10.2. It can be seen that the lowest eff is 0.16 ps, which corresponds to fmax 500 GHz.
At high, negative, gate bias, injection of holes from the drain is facilitated [11], leading to
an increase in the quantum-capacitance contribution to Cgd . Thus, assumption Eq. (10.15)
overestimates Cgs , leading to Eq. (10.16) overestimating the true eff at the most negative bias
considered. The effect of assumption Eq. (10.17) is more severe at high bias because gm falls
off considerably. Again, this is due to holes being injected into the nanotube from the drain:
the resulting hole flow bucks that issuing from the source, reducing gm . Moreover, gds rises in
that bias range, ultimately yielding a ratio 2gds /gm 1 near VGS = 0.8 V and invalidating
assumption Eq. (10.17).
We turn now to Device 2, which has a nanotube diameter dt = 0.8 nm (taken to correspond
to a tube of chirality (10,0)), a positive hole-barrier of 0.3 eV at the Pd end-contacts [9], an
increased insulator thickness tins of 8 nm and shorter contacts Lc of 30 nm. The higher barriers
and thicker insulator will reduce gm below that of Device 1. These features should lead to a
lower fmax than predicted for Device 1. However, this should be mitigated somewhat by lower
134
0.6
eff (ps)
0.5
0.4
0.3
0.2
0.1
0.8
0.7
0.6
0.5
0.4
Gate voltage (V)
0.3
Figure 10.2: eff estimates for Device 1 using Eq. (10.9) (solid), and Eq. (10.13) with Eq. (10.14)
(dots), Eq. (10.16) (circles), and Eq. (10.18) (crosses). VDS = 0.5 V.
capacitances Cgs and Cgd , due to the larger tins and smaller Lc . The results shown in Fig. 10.3
show that fmax is, indeed, significantly lower than for Device 1. Interestingly, eff ,2 is a better
approximation to the true eff in this case, which is perhaps unexpected, given that the shorter
Lc and thicker tins should reduce the inter-electrode capacitances that would otherwise help
to equalize Cgs and Cgd . The reason lies in the positive barrier heights and larger bandgap,
which restrain charge injection into the nanotube (see inset to Fig. 10.3), thereby reducing the
quantum capacitance contributions to Cgs and Cgd even more than the above physical changes
reduce the inter-electrode contributions. The higher barrier at the source generally reduces
the drain current, so both gm and gds are affected, and eff ,3 is no worse an approximation,
relatively speaking, than it was for Device 1.
One of the reasons for the low fmax shown for Device 2 in Fig. 10.3 is that the effective gate
bias is lower than for Device 1 because of the higher threshold voltage due to the thicker gate
insulator. While this could be ameliorated by application of a higher negative bias to the gate, or
by using a higher work function for the gate metal, Fig. 10.3 is useful because it illustrates that
our equations are reasonable over an effectively different bias range than applies to Device 1.
So far, we have used resistances of Rc 0.9 k and 0.3 k for Devices 1 and 2, respectively,
and Rg = 1 k. To examine the effect of parametrically changing these values, results are
135
Energy (eV)
200
100
10
z (nm)
20
max
(GHz)
150
0.3
0.2
0.1
0
0.1
0.2
0.3
50
0
0.8
0.7
0.6
0.5
0.4
Gate voltage (V)
0.3
Figure 10.3: fmax estimates for Device 2, at VDS = 0.5 V, using Eq. (10.9) (solid) and
Eq. (10.13) with Eq. (10.14) (dots), Eq. (10.16) (circles), and Eq. (10.18) (crosses). The inset illustrates the valence band edge profiles near the source contact for Devices 1 (dotted) and
2 (solid) at VGS = VDS = 0.5 V. Energies are referenced to the source Fermi level.
presented in Fig. 10.4 for Device 1. The error in the estimation of the prediction of fmax
was examined, after making each of the assumptions leading to the three expressions for eff .
For the first two, the error in fmax is less than 1 % over the range of resistances shown in
Fig. 10.4 (a). Fig. 10.4 (b) depicts the case after, additonally, making assumption Eq. (10.17),
and shows that the error is greatest at large Rc . This is because the approximated term,
(gm + 2gds ) gm in simplifying Eq. (10.16), is multiplied by the square of Rc , whereas Rg
appears without exponentiation. Fig. 10.4 indicates that the compact expressions are useful
over a wide range of resistance values.
Finally, we demonstrate the utility of the compact expressions in guiding design towards CNFETs that should lead to improved fmax . Obviously, reducing Cgd would be helpful because of
its domination of the output admittance. Eqs. (10.16) and (10.18) highlight this by elucidating
the direct dependence of eff on Cgd . By contrast, eff has a lesser dependence on transconductance. One way to trade-off gm against Cgd would be to increase tins . Ways to reduce Cgd
directly would be to shorten the drain contact Lc , and to increase the gate-drain underlap Lud .
Although the functional dependencies of gm and Cgd on tins and Lud are not readily attainable,
the beneficial effect to Device 1 of making these changes is illustrated in Fig. 10.5, where the
136
(a)
(b)
14
Error (%)
Error (%)
0.5
12
10
8
0
5
10
10
Rc (Ohms)
10 10
10
10
10
Rg (Ohms)
10
Rc (Ohms)
10 10
10
10
Rg (Ohms)
Figure 10.4: Error in fmax prediction for Device 1, incurred by the use of: (a) Eq. (10.16); (b)
Eq. (10.18). VDS = 0.5 V.
600
fmax (GHz)
500
400
300
200
100
0.8
0.7
0.6
0.5
0.4
Gate voltage (V)
0.3
Figure 10.5: fmax for Device 1 and fmax improvement for Device 1 computed from Eq. (10.9)
without assumptions, both at VDS = 0.5 V. Solid line: Device 1 as originally specified; dotted
line: Device 1 with tins = 8 nm, Lc = 30 nm, Lus = 5 nm, and Lud = 15 nm.
peak value of fmax is raised by about 15 % to 580 GHz.
10.4 Conclusions
From this study of the extrapolated fmax in Schottky-barrier CNFETs it can be concluded that:
1. compact expressions for fmax can be derived that are useful over wide ranges of physical
properties, parasitic resistances and gate biases;
2. the compact expressions provide a useful guide to the design of high-frequency devices;
3. fmax values in excess of 0.5 THz should be realizable.
137
References
[1] M. Vaidyanathan and D. L. Pulfrey, Extrapolated fmax of heterojunction bipolar transistors, IEEE Trans. Electron Devices, 46, 301309 (1999).
[2] W. Liu, Fundamentals of III-V Devices (John Wiley, Toronto, 1999).
1987).
[4] S. J. Mason, Power gain in feedback amplifier, IRE Trans. on Circuit Theory, 1, 2025
(1954).
[5] L. C. Castro, D. L. John, D. L. Pulfrey, M. Pourfath, A. Gehring, and H. Kosina, Method
for predicting fT for carbon nanotube FETs, IEEE Trans. Nanotechnol., 4(6), 699704
(2005).
[7] H. Flietner, The E(k) relation for a two-band scheme of semiconductors and the application to the metal-semiconductor contact, Phys. Stat. Sol. (b), 54, 201208 (1972).
[9] Zhihong Chen, Joerg Appenzeller, Joachim Knoch, Yu-Ming Lin, and Phaedon Avouris,
The role of metal-nanotube contact in the performance of carbon nanotube field-effect
transistors, Nano Lett., 5(7), 14971502 (2005).
[10] Jijun Zhao, Jie Han, and Jian Ping Lu, Work functions of pristine and alkali-metal intercalated carbon nanotubes and bundles, Phys. Rev. B , 65, 19340111934014 (2002).
138
[11] L. C. Castro, D. L. Pulfrey, and D. L. John, High-frequency capability of Schottky-barrier

carbon nanotube FETs, Solid-State Phenomena (2005). Accepted December 22, 2005.
139
Chapter 11
Conclusion and Recommendations for
Future Work
This thesis describes a body of work that yielded insightful results regarding the DC and
AC performance of carbon nanotube field-effect transistors (CNFETs). Various models were
devised and applied to several types of devices, with the aim of explaining device behaviour and
related physical phenomena, as well as ultimately obtaining predictive performance metrics for
guiding device design. Given that various papers were published from this work, the thesis was
composed by their incorporation as chapters herein, arranged in a roughly chronological order.
The preliminary version of a CNFET compact model (Chap. 2) introduced a self-consistent,
non-equilibrium compact model for Schottky-barrier CNFETs, incorporating a crude JWKB
tunneling expression for positive barriers at the contacts. It represented an improvement on
concurrent models in the literature in that it included the effects of the Schottky barriers
assuming transparent contacts overestimates both the channel charge and current. However, the
model was still based on compact electrostatics, i.e., it implicitly assumed the behaviour of an
infinite-length tube, and relied on other equilibrium models to estimate the barrier shapes out
of equilibrium. As discussed in Chap. 3, an equilibrium electrostatics model served as a guide
to the behaviour of the potential profile in the channel under non-equilibrium conditions, and
was the basis for the fitting parameters in the aforementioned compact model. Finally, the nonequilibrium charge did not account for metal-induced gap states nor include any short-channel
effects, both of which require a Schr
odinger-Poisson (SP) solver.
The compact model was then employed in a preliminary evaluation of CNFET DC performance,
140
as outlined in Chap. 4. With the incorporation of hole transport, the ambipolar feature of
CNFETs was revealed, whereby under certain bias- and work-function conditions it is possible
to have injection of electrons and holes in different ranges of the devices I-V -characteristics.
In particular, at high positive VDS applied to an n-type transistor, hole injection from the
drain increases the CNFET current even though it has already reached the saturation regime
as typically defined for silicon devices. A logarithmic I-VGS plot for a given device and nonzero VDS reveals a v-shaped characteristic that indicates the electron and hole branches of the
current. From these, it was possible to examine the expected performance of various metal
contacts, given that their particular work function would yield a different-polarity Schottky
barrier for electrons or holes. Such work function engineering allows for negative barriers to
facilitate majority carrier transport and stave off injection of the supposedly minority carrier.
The suppression of ambipolar behaviour is particularly important for digital applications, as it
severely affects the ON/OFF ratio and subthreshold slope of the CNFET. The evaluation of
CNFETs further studied their limits for ON current, conductance and transconductance.
With the aim of improving the non-equilibrium charge computation, a SP solver was elaborated,
as illustrated in Chap. 5. This model was based on the effective mass approximation, the parameters for which were drawn from a tight-binding nearest-neighbour calculation. The results
were self-consistent solutions of the phase-coherent charge with the local electrostatic potential,
allowing for a more accurate study of short-channel ballistic devices. The inclusion of evanescent modes injected from the metal contacts into the forbidden energy gap of the nanotube,
also known as metal-induced gap states, revealed the importance of these for positive-barrier
devices, since the exponential dependence of the tunneling currents on the barrier shape makes
these CNFETs more susceptible to slight variations in the near-contact carrier density. The
model also served to demonstrate the importance of including phase-coherent charge in short
devices, because the carrier distributions in energy were highly modified from their equilibrium
form, owing to the formation of quasi-bound states in the channel.
Subsequent to the implementation of the SP solver, its equivalence to the compact model
under certain conditions was proven and described in Chap. 6. An improvement for the trans141
mission function was sought, with the aim of obtaining a more accurate estimate of the I-Vcharacteristics in negative barrier devices, i.e., those that had yielded the best DC performance.
It was also found that this kind of CNFET would not, as was previously thought possible, come
as close to the ultimate DC current limits of a ballistic carbon nanotube channel. This is due to
the fact that the quantum mechanical transmission of above-barrier carriers is less than unity
in much of the current-carrying energy range. The formation of quasi-bound states, between
two large potential steps in close spatial proximity, produces a weak localization of the current
in energy, i.e., restricts which energy ranges transmit or not.
Another study, presented in Chap. 7, delved further into the concept of quantum capacitance
(CQ ) in nanoscale transistors, and illustrated its application to CNFET devices. In particular,
the quantum capacitance limit being mentioned in the literature was methodically examined and its range of validity specified. This limit is represented by the insulator capacitance
overwhelming that of the semiconductor, and is thus desirable because it allows the gate to
exercise a stronger control of the channel over a wider range of biases. This limit was claimed
to be reachable with CNFETs, but was found here to require unviable material properties.
In CNFETs, unlike in a transistor with a 2DEG channel, the 1D density of states was found
to produce a CQ that was neither quantized nor readily computed from a simple formula.
Moreover, with the employment of the aforementioned compact model, the bias dependence of
CQ was also studied, and found to reveal highly nonlinear features, owing particularly to the
non-equilibrium nature of the carrier distributions.
Having studied the DC performance of the CNFETs, a small-signal model, in conjunction with
an improved Schrodinger-Poisson solver, was employed to compute their AC figures-of-merit
fT and fmax , a first in the CNFET literature. In its final version, the SP solver included the
effects of interband tunneling and an energy-dependent correction to the effective mass. The
small-signal model also proved useful in studying the effects of parasitic capacitances and resistances brought about by the employment of various device dimensions and contact geometries.
Several trends were studied, particularly the oscillations of the capacitances and transconductance with gate bias, and their effect on fT and fmax . These oscillations were correlated to
142
the interaction between the Fermi energy levels of the injecting contacts and the quasi-bound
states in the channel, the position in energy of the latter being controlled by the gate. This
behaviour was demonstrated clearly by analysing the low-temperature operation of CNFETs.
Further illustrated by the small-signal analyses was the dependence of the figures-of-merit on
the nanotube bandgap. It was found that larger tubes, with smaller bandgaps, benefited from
higher transconductances but these gains were often offset by increased capacitances, brought
about by the larger charge densities present in the devices. Finally, compact expressions for
fmax were derived, from which the importance of the gate-drain capacitance and the transconductance were ascertained. Extrinsic fT and fmax values in the order of tenths of THz were
computed for representative devices.
However, the high-frequency models employed herein have been quasi-static in nature and the
validity of this must be established if they are to be employed at the frequencies that are being expected of these devices as potential replacements or supplements of current technologies.
Various works examining this issue in other nanoscale devices can be found in the recent literature, and many could be employed in modeling CNFETs. Another avenue to be pursued is
to study the role of various scattering mechanisms at the frequency and bias ranges of interest
here. At these dimensions, and under certain conditions, scattering could play an important
role in nanoscale devices. Moreover, as the device length approaches the nanotube diameter,
the effective-mass, mode-space approach employed here should give way to a more detailed
quantum model in order to account for variations at the atomistic scale of quantities such as
the density-of-states. Finally, the relevance and applicability to carbon nanotube transistors of
the concept of kinetic inductance, which has emerged in the recent literature, merits further
investigation.
Based on the current efforts of the semiconductor industry, and barring improvements in manufacturability and prototype performance, research on carbon nanotube transistors may be
supplanted by that based on other nanowires, given they are easier to manufacture repeatably.
Moreover, existing technologies that are already well-developed, such as those based on silicon,
may resolve their impending scalability problems and prove to be more viable solutions. Still,
143
carbon nanotubes remain an attractive material for employment in state of the art semiconducting devices, owing to their high current capability, near-ballistic transport, and minute
scale.
144
Appendix A
Complete Schr
odinger-Poisson Model
Here we describe a numerical model for the CNFET, an improvement on that of Chap. 5 and
presented in more detail. The objective is to describe the final version of the Schr
odingerPoisson solver developed in this work. The key changes are the inclusion of interband tunneling
and the handling of both metallic and doped-semiconductor contacts.
Throughout the ensuing description of the model it is assumed that the contacts are metallic (as
treated in the preceding chapters), unless otherwise stated. We seek a self-consistent solution
of the potential V , charge Q, and current I for the CNFET. This is accomplished via the
employment of the Schrodinger, Poisson and Landauer equations, which are useful inasmuch
as they yield an accurate portrait of the electronic charge in a ballistic device where quantummechanical effects (e.g. tunneling, resonance) are present. The problem being modeled here is
specified by the system
LV
= Q+B
Z
X
Q = q
Db
Emin
I =
(A.1)
Emax
GS,b [1 u fS ] + GD,b [1 u fD ] dE
Z Emax
2q X
Db
(fS fD )T dE
h
Emin
(A.2)
(A.3)
H = E ,
(A.4)
where L and B are, respectively, a linear differential operator and a boundary condition vector
for Poissons equation, Db is the degeneracy factor of sub-band b, Gc,b = Gc,b (z, E) = ( )c,b
is the local density of states associated with injection from contact c, is the wavefunction
of a carrier with energy E, fc = f (E c ) is the contact Fermi function with Fermi level c ,
145
Appendix A. Complete Schrodinger-Poisson Model
T = T (E) is the transmission function, and H is the effective-mass Hamiltonian in Schr

odingers
equation. The unit step function, referenced to the midgap energy E0 (also known as the charge
neutrality level), is specified by
1 , E > E0
u = u(E E0 ) =
0 ,E < E .
0
To address Eq. (A.1), we deal specifically with the coaxial geometry of the CNFET as shown in
Fig. 1.8 and recall the discussion in Sect. 1.3.1, noting that care must be taken with discretization
and derivative discontinuities. Given the charge vector Q, which is non-zero only for elements
on the tube surface, and boundary condition vector B, we compute the potential vector V
by inversion, i.e., V = L1 (Q + B). In building the operator L and vector B, we employ
three equations for the interior points of the simulation space (all points excluding the contact
interfaces and open boundaries): (1) since we only consider charge on the tube surface, a jump
discontinuity for the normal-component of the electric flux is enforced there (all points in set
St , solid line at 0.6 nm in Fig. A.1); (2) a homogeneous Neumann condition is applied on
the axis of revolution by symmetry arguments (all points in set Sa , dash-dot line in Fig. A.1);
(3) a Laplace solution is carried out for all other interior points (in set Si ). Thus,
V
Q
V
(, z) St
t
=
ins
R+
R
2Rt 0
t
t
V
=0
(, z) Sa
2V
1 V
2V
+
+
=0
(, z) Si ,
z 2

2
where Rt is the tube radius, ins and t are, respectively, the relative permittivities of the insulator and nanotube, 0 is the permittivity of free-space, and the 1D charge density Q has
been uniformly smeared around the tube circumference. Open boundaries (dotted lines in
Fig. A.1) are treated with a homogeneous Neumann condition, equivalent to assuming that
the normal displacement is zero, and this is typically valid with appropriate choices of contact
dimensions [1]. Points lying on a contact surface are assigned fixed (Dirichlet) boundary conditions, specified by the bias voltages. Note that in the case of doped-semiconductor contacts, no
Dirichlet conditions are specified; homogeneous Neumann conditions are enforced in the entire
146
20
(nm)
15
10
5
0
0
10
20
30
40
50
z (nm)
Figure A.1: Simulation space for CNFET with metal contacts. Neumann boundary conditions
are specified on dotted perimeter, and Dirichlet conditions on edges of contacts. Linear charge
specified on nanotube surface (solid line at 0.6 nm).
perimeter of the simulation space.
The solution of Eq. (A.1) is effected via the finite-difference scheme, chosen primarily for simulation speed. This involves discretizing the two-dimensional solution space of Fig. A.1 into a
grid and solving difference equations for every grid point [2].
We now turn to the computation of the linear charge density Q from the 1D time-independent
Schrodingers equation (Eq. (A.4)). The general 1D effective-mass Hamiltonian is given by
H=
~ 2
+ Epot ,
2mb z 2
where mb is the carrier effective mass in sub-band b (equal for electrons and holes, and constant
across the device), ~ is Diracs constant, and Epot is the potential energy of the electron or
hole. Employing a plane wave solution, Schr
odingers equation reduces to
2
= k 2 ,
z 2
where (z, E) is the electron wavefunction. The wavevector k = k(z, E) in the nanotube is
given by
, |E E0 | > b
2mb |E E0 | b
2
k = 2
(A.5)
2
2
|E E0 | b
~
, |E E0 | < b ,
2b
where b is the energy distance between the bottom of sub-band b and E0 , and the two energy
ranges distinguish, respectively, between carriers outside and inside the bandgap (also defined
147
as propagating and evanescent modes). The intra-gap wavevector employs a correction for the
non-parabolicity of the -k relation there, via an energy-dependent effective mass [3]. This is
consistent with the so-called complex band picture [4] that provides a smooth transition of
the -k relation between the conduction and valence bands, as illustrated in Fig. A.2. Note that
for the lowest sub-band (and omitting b subscript) the condition |E E0 | > = Eg /2 may be
rewritten as a double requirement E E0 > and E E0 < , or E > E0 + = Ec and
E < E0 = Ev .
Energy relative to E0 (eV)
0.8
0.6
0.4
electron
0.2
0
hole
0.2
0.4
0.6
0.8
imaginary k
real k
Wavevector
Figure A.2: Complex bands in semiconducting nanotube: comparison between nearestneighbour tight binding (solid), effective mass (dashed) and energy-dependent effective mass
within bandgap, as per Ref. [3] (circles). The agreement between the energy-dependent effective
mass and the tight-binding case is clear.
The distinction between electrons and holes is of concern insofar as the carriers react differently
to an applied field. This is reflected in the wavefunction computation by labeling a carrier
according to its energy-position in the -k relation, i.e., at a given position in the device
channel, the solution is counted as an electron if E > E0 or as a hole otherwise. This
book-keeping does not keep track of the actual electron and hole concentrations, but rather
models deviations from charge neutrality. Under this scheme, it is possible that a carrier
traversing a region of strong band-bending, while maintaining a constant energy under the
elastic conditions assumed here, may be labeled differently at different points of its path. This
allows for the possibility of interband tunneling, as in the case, for example, of a reverse-biased
148
p-n junction, where one may describe the situation by two equivalent pictures: that a valenceband hole on the p-side tunnels through the bandgap into the conduction-band on the n-side, or
that a conduction-band electron on the n-side tunnels into the p-side valence-band. Note that
the implementation herein only allows for uncoupled interband tunneling, i.e., it only accounts
for transitions within a corresponding pair of valence and conduction sub-bands. With regards
to the current computation, the existence of open states on the opposite side of the device is
ensured in the derivation of the Landauer equation, which takes the Fermi level of both contacts
into account and applies identically to an electron or hole picture.
In the metal contacts the -k relation is given by
k2 =
2mb
(Eref |E c |)
~2
, |E c | < Eref ,
(A.6)
where mb is the effective mass of the nanotube sub-band into which the carriers are being
injected, and Eref is a fitting parameter related to the type of metal being used. A half-full
metal band is assumed, such that Eref = Etop c = c Ebottom and c = E0,c (the contact
charge neutrality level). A typical value used in the simulations was 5.5 eV, as estimated for
various metals commonly used in CNFETs from the data in Ref. [5]. The metal -k relation
Energy (eV)
and density of states are illustrated in Fig. A.3.
6
0
1
2
3
1
Wavevector (nm )
6
0
1
2
3
1
Density of states (nm )
Figure A.3: Dispersion relation and density of states in 1D metal contact.
149
Given that the wavefunction amplitude is arbitrary to any multiplying factor in Eq. (A.4), we
must normalize it to compute the correct carrier density in the device. In many quantum
physics problems, such as the near-infinite quantum well, the electron is confined in space,
and normalization may be effected by a spatial restriction on the electrons location, via the
requirement that the integral of the positional probability density in the confinement region is
unity. In our problem, however, this is not viable, as we have open boundary conditions at the
contacts (infinite-length contact regions) and we cannot assume the entire carrier-wavefunction
is confined within the device or its near vicinity. We thus resort to a flux normalization, whereby
the amplitude of the wavefunction is related to the injecting carrier flux, emanating from the
metallic contact regions. In essence, this is achieved by equating, at each energy level being
considered, the particle current to that in Landauers equation (Eq. (A.3)).
We illustrate the normalization for carriers injected from the source contact; injection from the
drain contact may be derived in analogous fashion. With a wavefunction of the form
(z, E) = Aeik(z,E)z ,
where A = A(z, E) is a constant to be determined from the normalization, we may write the
particle current with effective mass m at some energy E as
"
#

q~k 2
q~
() =
|A| ,
()
IP (E) = i
2m
z
z
m
(A.7)
and the Landauer current as

IL (E) =
q
f (E S )T (E) ,
~
(A.8)
where the transmission function, defined as a ratio of outbound to inbound particle currents, is
T (E) =
k(Lt , E) |A(Lt , E)|2

.
k(0, E) |A(0, E)|2
Equating Eq. (A.8) with Eq. (A.7) in the drain region (z = Lt ), i.e., at the output for a sourceinjected wave, we obtain the requirement for normalization
|A(0, E)|2 =
m
~2 k(0, E)
150
f (E S ) .
(A.9)
Note that this normalization condition is related to the 1D density of states in the metal contact
(assuming parabolic bands):
g(E) =
2m
2 k
=
,
E
~2 k
(A.10)
where the apparent discrepancy by a factor of 2 is due to the fact that g(E) includes both
k-states, while |A|2 is related to a single-direction flux. Substitution of either Eq. (A.5) or
Eq. (A.6) into Eq. (A.9) yields the normalization condition for, respectively, semiconductor or
metal contacts.
Note that the effective-mass Hamiltonian is the simplest, yet quickest to simulate, implementation of Eq. (A.4). More detailed models were not pursued here, as we did not expect large
discrepancies, for the devices simulated in this work, between the current and more elaborate
models. A major reason for this is that the 1D density of states of CNs (see Fig. 1.4) should
lend itself well to the employment of the parabolic approximation, given that most carriers
reside near the sub-band edge, where the approximation is at its best. An agreement in charge
profiles would yield commensurate accuracy in the computation of band profiles and currents.
Very recently, some works have begun to address the validity of the effective mass approximation in CNFET performance computations [6, 7], with Ref. [7] supporting the use of this
approximation, for the sake of computational savings, under most practical bias conditions.
Regarding the computation of the wavefunction (i.e., solving Eq. (A.4) for ), a scattering matrix method was employed in lieu of the simpler transfer matrix method so as to circumvent numerical issues related to rounding errors [8]. The implementation of the algorithm for obtaining
the wavefunction profile along the channel, for a particle at a given energy, is rather straightforward and will not be detailed here. However, it should be noted that in solving Eq. (A.2)
for a given Gc,b (computed from the appropriate s), an adaptive integration scheme should be
used, given that the integrand can be highly peaked, particularly in short-channel devices where
quasi-bound states are prominent. Here we employ the Adaptive Simpson Quadrature method,
modified to process various energy ranges in parallel, a procedure that speeds up computation
by reducing the number of function calls to the systems computational bottleneck, Eq. (A.4).
151
We normally solve the non-linear Schr

odinger-Poisson system for the potential and charge
density using the Picard iterative scheme. The (k+1)-th iteration is given by
Vk+1 = Vk L1 rk
(A.11)
rk = LVk [Q(Vk ) + B] ,
where rk is the residual of the k-th iteration, and 0 < 1 is a damping parameter. Note
that if = 1, the scheme reduces to back-substitution. The complete algorithm for achieving
a self-consistent solution under a given set of simulation parameters is as follows:
1. build matrix L and vector B, and initialize k to 1;
2. compute V0 by solving Laplaces equation (Q = 0 in Eq. (A.1));
3. solve Schrodingers equation for all relevant E to get corresponding s;
4. integrate s to obtain carrier density Qk (Vk1 );
5. compute new Vk by solving Poissons equation with new Qk in Eq. (A.11);
6. check that ||rk || has reached tolerance and go to step 3 if not.
Typically, a solution is deemed to have converged if the absolute potential reached a tolerance
of 1 V. In the actual implementation, a further check is performed on the current, as computed
with Eq. (A.3), such that the error was at most 1 nA. The choice of has a strong influence
on convergence and unfortunately, the appropriate value depends on too many simulation parameters to be ascribed a simple formula. In general, 0.2 virtually guarantees convergence,
but often at prohibitive simulation times; in most cases, a value of 0.7 gave a better
performance trade-off.
An alternative solution method, the Gummel iterative scheme, is employed for simulations that
have difficulty converging with the above method. Under this scheme, we employ a quasi-Fermi
level (EqF (z)) formulation that effectively localizes the chargethis is important because the
charge density obtained with Schrodingers equation is highly non-local. The delocalized charge
152
Qapprox is specified by an equation identical to Eq. (A.2), except the functions Gc,b are replaced
by the position-independent 1D density of states (akin to that in Eq. (A.10)). The outer loop
is solved by undamped fixed-point iterations, while an inner loop to solve for the potential as
a function of Qapprox employs a Newtons method, where the derivative Qapprox /V required
for the Jacobian operator is analytic. The algorithm is given by the following steps:
1. build matrix L and vector B; make initial guess for EqF ;
2. inner loop to find V (Qapprox (EqF )), employing Newtons method: compute Qapprox (EqF )
and analytic Qapprox /VCS , update potential with
Vupdate = (L Qapprox /V )1 (LV Qapprox B), and loop until tolerance reached;
3. compute new Q(V ) using potential from step 2 and Eq. (A.2);
4. check if V = L1 (Q + B) has reached desired tolerance relative to last iteration of outer
loop; if not, find new EqF by inverting Q(EqF ) and go to step 2.
In simulations with doped-semiconductor contacts, this method produces much better convergence than Picard. It is also found that the initial guess for the quasi-Fermi level plays a
significant role in whether convergence is reached. A typical initial guess is to make the source
and drain Fermi levels reach into a quarter of the (intrinsic) channel region and make the centre
of the channel region take an average of the contact Fermi levels. The best results in using an
initial guess were obtained by employing previous solutions at close values of bias for the same
device geometry.
Once a converged solution has been attained, we have a self-consistent description of the charge
and potential profiles in the device. The drain current may then be computed from Eq. (A.3).
A.1 Sample Results

The model described above and variations of it were used to obtain much of the data in this
thesis. Here, in the interest of establishing its validity, we directly compare this model with
experimental data from the literature and with a more detailed, atomistic model. We present
153
results for two distinct devices. Firstly, we show an approximate fit to the I-V characteristics
of the high-performance SB-CNFET device of Ref. [9], based on the data of Figure 1(c,d)
of that work. Secondly, we compare the results of this model to that of the more detailed,
atomistic approach in Ref. [10], as applied to a p-i-n tunneling device. In the comparisons, the
-k relation of the nanotube is used in three permutations, defined by the simulation flag F,
which takes on the value F=0 when the -k relation is parabolic (top component of Eq. (A.5))
for all energies, F=1 when the -k relation is described exactly by Eq. (A.5), and F=2 when the
-k relation has an energy-dependent effective mass (bottom component of Eq. (A.5)) for all
energies, i.e., for both |E E0 | < b and |E E0 | > b . Two sub-bands were included for all
simulations in this section, although omitting the second sub-band would likely have a minor
effect in the comparison given the lowest sub-band accounts for nearly all charge and current.
0
5
10
ID (A)
ID (A)
10
10
10
12
16
(a)
1
0.5
0
Gate voltage (V)
0.5
(b)
0.4 0.3 0.2 0.1

Drain voltage (V)
Figure A.4: Agreement between Schr

odinger-Poisson model and data from experimental device
of Ref. [9]: (a) I-VGS plot for VDS = 0.3 V, and (b) I-VDS plot for (from top to bottom of
figure) VGS = 0.43, 0.73 and 1.03 V (these values correspond to VGS = 0.1, 0.4 and
0.7 V in Ref. [9]; see text for explanation). Solid lines are experimental results, while symbols
are simulated values. Simulation parameters are given in text.
In fitting the curves of Ref. [9], as illustrated in Fig. A.4, we employ a (20,0) nanotube
(Eg =0.52 eV), a self-aligned gate with Lg = 50 nm (underlap Lu = 0) and thickness tg = 4 nm,
an insulator of thickness tins = 8 nm and relative permittivity of 15 (HfO2 ), and Pd end-contacts
(=5 eV) of radius tc = 6 nm and length Lc = 4 nm. We set Eref = 1 eV for the metal contacts,
154
and F=2. Note also that all the simulation curves in Fig. A.4 are offset in the gate bias by
0.33 V in order to better match the experimental curvesthis can be attributed to the potential presence of trapped oxide charges in the actual device, causing a shift in the transistor
threshold voltage Vt . No other scaling or manipulation was employed.
5
Drain current (A)
10
10
F=0
F=1
F=2
NEGF
10
0.2
0.2
0.4
0.6
0.8
Gate Voltage (V)
Figure A.5: Comparison of the solver described in this section to the atomistic simulations of
Ref. [10], in modeling the I-V characteristics of a p-i-n device. Refer to text for parameters
employed in simulation and an explanation of figure caption.
In another comparison, we illustrate in Fig. A.5 the simulated I-V characteristics for a p-i-n
device, as given by this model and one of more detail from Ref. [10]. The device parameters
are: (16,0) tube, Lg = 7 nm, Lu = 0, Lc = 10 nm, tg = 0.1 nm, tins = 2 nm, ins = 3.9, and
VDS = 0.4 V. The p-type source and n-type drain nanotube contacts are doped, respectively
with +0.39 and 0.39 carriers per nanometre. The solid circles in Fig. A.5 are the results
of the model in Ref. [10], and the remaining points are obtained by this work, in the three
permutations of the -k relation as described above. Note that because the p-i-n transistor is
a tunneling device, and correctly estimating the interband tunneling current is more difficult
than estimating currents dominated by propagating modes, the discrepancy between this model
and the atomistic simulation of Ref. [10] likely indicates an upper bound on the error of this
model.
155
References
[1] D. L. McGuire and D. L. Pulfrey, Error analysis of boundary condition approximations in
the modeling of coaxially-gated carbon nanotube field-effect transistors, Phys. Stat. Sol.
(a) (2006). Accepted February 21, 2006.
[2] G. D. Smith, Numerical Solutions of Partial Differential Equations: Finite Difference
Methods (Clarendon Press, Oxford, 1985), 3rd ed.
[3] H. Flietner, The E(k) relation for a two-band scheme of semiconductors and the application to the metal-semiconductor contact, Phys. Stat. Sol. (b), 54, 201208 (1972).
[4] T.-S. Xia, L. F. Register, and S. K. Banerjee, Quantum transport in carbon nanotube
transistors: Complex band structure effects, J. Appl. Phys., 95(3), 15971599 (2004).
[6] Andres Godoy, Zhicheng Yang, Umberto Ravaioli, and Francisco Gamiz, Effects of nonparabolic bands in quantum wires, J. Appl. Phys., 98, 01370210137025 (2005).
[7] S. O. Koswatta, N. Neophytou, D. Kienle, and M. S. Lundstrom, Dependence of DC
characteristics of CNT MOSFETs on bandstructure models, IEEE Trans. Nanotechnol.
(2006). To be published.
[8] David Yuk Kei Ko and J. C. Inkson, Matrix method for tunneling in heterostructures:
Resonant tunneling in multilayer systems, Phys. Rev. B , 38(14), 99459951 (1988).
[10] David Llewellyn John, Simulation studies of carbon nanotube field-effect transistors, Ph.D.
thesis, University of British Columbia, Vancouver, Canada (2006).
156

Castro Thesis

Caricato da

Informazioni sul documento

Copyright

Formati disponibili

Condividi questo documento

Condividi o incorpora il documento

Opzioni di condivisione

Hai trovato utile questo documento?

Questo contenuto è inappropriato?

Copyright:

Formati disponibili

Castro Thesis

Caricato da

Copyright:

Formati disponibili

Modeling of Carbon Nanotube

The University Of British Columbia

Chapter 4. An Evaluation of CNFET DC Performance

Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114

Chapter 9. High-frequency Capability of Schottky-Barrier CNFETs

Appendix A. Complete Schr

Hybridized carbon atom and graphene lattice structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

Model geometry: closed cylindrical structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Coaxial CNFET model geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

CNFET net carrier density as a function of position and VDS . . . . . . . . . . . . . . . . . . . . . .

Drain current versus gate-source voltage for various models . . . . . . . . . . . . . . . . . . . . . . . .

Equilibrium 1D quantum capacitance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Small-signal equivalent circuit for ft calculation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Structure of the modeled CNFET . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

10.1 Unilateral power-gain for Device 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Simulation space for CNFET with metal contacts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

could be explained by the models presented herein.

1.1 Carbon Nanotubes

1.1.1 Crystal Structure

weakly by dispersion forces1 and have an inter-layer spacing of about 3.35

Single-wall nanotubes are characterized by a wrapping (or chiral) vector Ch = n1 a

1.1.2 Electronic Properties

measure in quantum devices involving quasi-1D carbon nanotubes, it is worthwhile mentioning.

Effective mass (m0)

Energy gap (eV)

1.2 Carbon Nanotube Field-Effect Transistors

contacts can be made virtually transparent either by selecting an appropriate metallization or by

1.2.1 Planar Devices

1.2.2 Coaxial Devices

1.3 Modeling Coaxial CNFETs

Electric Field (V/m)

Electric Field (V/m)

VGS = 0.5 V, VDS = 0.4 V

VGS = 0.5 V, VDS = 0.9 V

g(E) {f [E + qVCS (z)] f [E qVCS (z)]} dE ,

until a self-consistent solution is attained.

1.4 Thesis Outline

1.5 Specific Contributions

[27] G. Pennington and N. Goldsman, Semiclassical transport and phonon scattering of

Chapter 2. Towards a Compact Model for Schottky-Barrier CNFETs

Chapter 2. Towards a Compact Model for Schottky-Barrier CNFETs

2.2 The Model

= Cins (VGS VCS ) ,

Chapter 2. Towards a Compact Model for Schottky-Barrier CNFETs

2.3 Results and Discussion

Chapter 2. Towards a Compact Model for Schottky-Barrier CNFETs

the barrier shapes.

Distance from source contact (nm)

Kinetic energy at z=0.5Lt (eV)

Electron distribution kstates

Chapter 2. Towards a Compact Model for Schottky-Barrier CNFETs

Drain current, I D (A)

Figure 2.4: Drain current-voltage characteristics.

Chapter 2. Towards a Compact Model for Schottky-Barrier CNFETs

Chapter 2. Towards a Compact Model for Schottky-Barrier CNFETs

Pulfrey, Electrostatics of Coaxial Schottky-Barrier Nanotube Field-Effect Transistors, IEEE Trans.

Chapter 3. Electrostatics of Coaxial Schottky-Barrier CNFETs

Chapter 3. Electrostatics of Coaxial Schottky-Barrier CNFETs

Coaxial Nanotube Electrostatics

FEMLAB, see http://www.comsol.com

Chapter 3. Electrostatics of Coaxial Schottky-Barrier CNFETs

g(E) [f (E + qV (z) + t ) f (E qV (z) t )] dE ,

Results and Discussion

Chapter 3. Electrostatics of Coaxial Schottky-Barrier CNFETs