See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/231026257

Compound Semiconductor Device Modelling

Article  in  Reports on Progress in Physics · January 1999

DOI: 10.1088/0034-4885/48/2/002

CITATIONS READS

125 1,174

1 author:

C.M. Snowden
University of Southampton
234 PUBLICATIONS   2,126 CITATIONS   

SEE PROFILE

All content following this page was uploaded by C.M. Snowden on 02 April 2014.

The user has requested enhancement of the downloaded file.

Rep. Prog. Phys., Vol 48, p p 223-275, 1985. Printed in Great Britain

Semiconductor device modelling

Christopher M Snowden
Solid State Materials, Devices and Applications Group, Department of Electrical and Electronic Engineering,
University of Leeds, Leeds LS2 9JT, U K

Abstract

An overview of semiconductor device modelling is presented which describes the

principal methods of representing and analysing modern solid-state devices. The review
deals with classical, semiclassical, particle and quantum transport methodologies and
compares the relative merits of each approach. The background behind each modelling
technique is briefly summarised and recent developments in each area are described.
The importance of accounting for non-stationary and quantum effects in small geometry
devices is emphasised.

This review was received in August 1984.

0034-4885/85/020223 + 53$09.00 @ 1985 The Institute of Physics 223

224 C M Snowden

Contents
Page
1. Introduction 225
2. Modelling techniques based on semiconductor equations 221
2.1. The classical semiconductor equations 221
2.2. Methods of solution for the semiconductor equations 23 1
2.3. Semiclassical transport equations 248
3. Monte Carlo particle modelling of semiconductor devices 256
3.1. The Monte Carlo method 251
3.2. Application of Monte Carlo techniques to device modelling 264
4. Quantum transport theory 265
4.1. Quantum transport theory 266
4.2. Application of quantum transport theory to semiconductor
modelling 269
5. Conclusions 270
References 27 1
 Semiconductor device modelling 225

1. Introduction

The rapid developments in semiconductor technology over the past 15 years have led
to a dramatic increase in interest in device modelling. The need to understand the
detailed operation and optimise the design of silicon very large-scale integrated devices
( V L S I ) , very high-speed integrated circuits (VHSIC) and compound semiconductor
devices has meant that device modelling now plays a crucial role in modern technology.
As the scale of the individual semiconductor devices decreases and the complexity of
the physical structure increases, the nature of the device characteristics depart from
those obtained from many of the classically held modelling concepts. Furthermore,
the difficulty encountered in performing measurements on these devices means that
greater emphasis must be put on results obtained from theoretical characteristics.
Modelling also allows new device structures to be rigorously investigated prior to
fabrication.
Semiconductor device models can be considered in two broad categories-physical
device models and equivalent circuit models. Physical device models attempt to
incorporate the physics o f device operation whilst equivalent circuit models are based
on electrical circuit analogies representing the electrical behaviour. The latter approach
is generally limited in its applicability because of the DC bias and frequency dependence
and the non-linear behaviour of most devices with respect to signal level. The principal
advantage of this technique is that it is easy to implement and analyse. In contrast,
physical device models can provide greater insight into the detailed operation of
semiconductor devices, encompassing a wide range of operating conditions, but usually
needing a lengthy analysis. The analysis requirements for physical models are normally
satisfied using numerical techniques implemented on computers. With the advent of
fast and powerful computing resources, this makes physical modelling techniques a
very attractive proposition for both the physicist and engineer.
Physical device models are solved using either bulk carrier transport equations (the
semiconductor equations), solutions to the Boltzmann transport equation or quantum
transport concepts. Historically, bulk transport solutions have satisfied most device
models, whilst Boltzmann and quantum transport solutions have provided a strong
insight into the detailed device physics. However, the trend towards very small geometry
devices, operating in the hot electron range, means that non-equilibrium transport
conditions must be accounted for and the importance of the latter two techniques has
become more significant. The understanding of material properties, physical boundary
conditions (such as surface physics, contact properties and device geometry) and
device-circuit interaction are steadily improving, allowing more intricate models to be
developed.
This review will concentrate on physical device modelling techniques. In particular,
the application of numerical solutions in the design, characterisation and optimisation
of semiconductor devices will be discussed. Interest in the numerical simulation of
semiconductor devices began nearly 20 years ago, when it became clear that this
approach could be successfully used to explain the operation of many solid-state
devices. Since that time, this field of research has continued to grow and has become
firmly established as a major area of interest in semiconductor device technology,
226 C M Snowden

receiving considerable publicity with entire conferences dedicated to the subject

(for example, NASCODE I, 11, I11 (1979, 1981, 1983) and IEEE-SIAM Boston
(1982)).
Prior to the widespread availability of powerful modern digital computers, semicon-
ductor problems were solved using analytical techniques with closed-form solutions.
This technique, which was pioneered by workers such as Shockley (1952), approached
the problem by dividing the device into areas over which linear approximations applied
(‘segmentation’), joined by appropriate boundary conditions (Grebene and Ghandi
1969, Puce1 et a1 1975). This method of device analysis has been extended to include
carrier velocity saturation (Dacey and Ross 1955, Hauser 1967) and two-dimensional
effects in silicon (Lehovec and Zuleeg 1970) and gallium arsenide (GaAs) devices
(Turner and Wilson (1968) and, for example, Shur and Eastman (1978)). However,
although this approach allows a rapid analysis and provides a basic insight into device
behaviour, it cannot produce the generalised and accurate solutions required for device
design and optimisation, and is unsuitable for modelling sub-micron devices such as
the field-effect transistors (FET) found in the developing range of VHSIC and GaAs
microwave monolithic integrated circuits ( M M I C ) . These types of application demand
the type of rigorous solutions obtained only from numerical simulations.
Early numerical device simulations were centred around one-dimensional models.
McCumber and Chynoweth (1966) demonstrated Gunn instabilities in a GaAs sample
using a one-dimensional electron temperature model. A steady-state transistor simula-
tion was described by Gummel as early as 1964. Limitations in computing power
meant that numerical device simulations at the time were limited to one-dimensional
models based on a set of ‘phenomenological equations’ (see, for example, De Mari
1968, Scharfetter and Gummel 1969).
In order to obtain a more realistic representation of the characteristics of semicon-
ductor devices, two-dimensional simulations were developed. Slotboom’s model ( 1969)
for npn bipolar junction transistors (BJT) was one of the first two-dimensional models
used to investigate spreading and surface effects. Two-dimensional models also allow
other important phenomena such as current crowding and high-level injection in BJT,
and the effects of short and narrow channels in FET to be quantified, which is not
possible using one-dimensional approximations. This type of model allows non-planar
devices to be rigorously investigated. Detailed descriptions of two-dimensional simula-
tions for silicon BJT and FET began to emerge in the early 1970s (for example, Kennedy
and O’Brien 1970, Slotboom 1973). GaAs and InP FET were subsequently investigated
(Himsworth 1972, 1973), although the doping levels in these early simulations were
restricted to low levels (typically 102’-1022m-3) in line with early devices and because
of numerical stability problems. Later simulations have modelled highly doped Si,
GaAs and InP devices and accounted for more complex material properties.
Although two-dimensional models can accurately represent short channel effects
in FET and non-uniform current flow in BJT, they cannot account for three-dimensional
effects such as those encountered in small devices with narrow widths or for current
spreading due to the extrinsic base region in BJT. A particularly important area where
three-dimensional models must be used is in the analysis of small geometry VLSI
metal-oxide semiconductor field-effect transistors ( MOSFET), where the channel width
approaches the thickness of the depletion layer. Three-dimensional models also allow
non-uniformities in the channel, breakdown voltage and threshold voltage variations
to be investigated. Fringing field effects in small geometry devices (typical of digital
FET) can also be determined (Husain and Chamberlain 1982). Yoshii et a1 (1982) have
 Semiconductor device modelling 227

applied three-dimensional simulations to a variety of solid-state devices and have

shown that three-dimensional effects are significant in many types of device.
The majority of advanced semiconductor devices are subject to regions of high
electric fields, carrier gradients and current densities which give rise to hot electron
transport. The basic classical transport equations which consider only the field-
dependent mobility of carriers do not account for carrier heating, and hence are not
strictly suitable for modelling most VLSI and high-frequency devices. Current interest
in low-dimensional, ballistic and high mobility semiconductor devices ( MOSFET, MES-
FET (metal semiconductor field-eff ect transistor), HBT, HEMT (high electron mobility
transistor)) is providing a strong incentive to develop models capable of accounting
for high field effects, non-equilibrium transport and relaxation mechanisms. This
problem is being addressed using modified classical models (accounting for energy
and momentum relaxation mechanisms), Monte Carlo techniques and quantum-
mechanical models. Frequently, the results obtained from Monte Carlo investigations
of the basic material parameters are used with the modified classical approach to
produce an easily evaluated model of complete semiconductor devices.
Since Kurosawa ( 1966) first introduced Monte Carlo techniques into semiconductor
modelling, interest in using this method for analysing both material properties and
device operation has grown, its increasing popularity being mainly due to the continuing
search for a detailed understanding of semiconductor properties and because of the
steady improvement in computer memory and speed. At present Monte Carlo tech-
niques are applied to the study of a wide range of semiconductors (including Si, GaAs
and InP) and to the direct analysis of device operation.
The trend towards very small devices has meant that the area of quantum transport
theory is finding an important application in the study of semiconductor devices as
well as in the more established regime of solid-state theory. Traditionally, quantum
transport theory has been used in semiconductor modelling to authenticate and deter-
mine the range of validity of Boltzmann transport models. It has not been extensively
used in device modelling because of the difficulties in implementation. However, more
recently quantum transport theory is being used to analyse novel device structures
where genuine quantum transport phenomena occur. Barker (1980a, b) has described
a quantum-mechanical model for very small-scale devices (<O. 1 pm). Another topical
area of interest where quantum transport theory will play an essential role is in
understanding device operation in the optical frequency regime, where it has been
shown that Boltzmann transport theory is not valid (Barker 1980b), even for linear
field distributions.
The following sections provide an overview of semiconductor device modelling
techniques and results, spanning classical, semiclassical, particle and quantum transport
theories. The range of application of each type of model is outlined and the importance
of considering non-equilibrium transport in small geometry devices is emphasised.
This review attempts to indicate the trends and significant contributions of each
modelling technique.

2. Modelling techniques based on semiconductor equations

2.1. The classical semiconductor equations

The classical semiconductor equations, which form the basis of the majority of physical
device models, may be obtained from an approximate solution for the first two moments
228 C M Snowden

of the Boltzmann transport equation. In general terms, these models require the
self-consistent solution of the following equations.
Poisson's equation:

where $ is the electrostatic potential, q is the electronic charge, E ~ + E is

, the permittivity,
ND is the donor doping density, N A is the acceptor doping density, n is the electron
concentration and p is the hole concentration.
Continuity equations:

_
dn - -1
- V.J,+G for electrons (2.2)
at 4
_
ap-- --v1 .J,- G for holes (2.3)
at

where J, and Jpare the electron and hole current densities, respectively, and G is the
generation-recombination rate.
Current densities:

J , = qnv, + qD,V n for electrons (2.4)

J p= qnv, - qD,Vp for holes (2.5)

where U, and up are the electron and hole velocities, and D, and D, are the carrier
diffusion coefficients. Majority carrier devices such as MESFET, Schottky varactor
diodes and GaAs transferred-electron devices (TED) are usually analysed using a single
species set of the basic transport equations (2.1)-(2.5).
The semiconductor transport equations are often normalised to simplify the
expressions and save on computer time (for example, De Mari 1968, Slotboom 1969,
1973):
V 2 $ =n - p - N (2.6)
an
-=V. J,+G (2.7)
at

J,= r,'(-nV$+Vn) (2.9)

J p= r , ' ( - p V $ -- Vp) (2.10)
where 7;' and 7;' are the normalised carrier mobilities for electron and hole, respec-
tively. The expressions for current densities J , and Jpassume the Einstein relationship
for the diffusion coefficients. Hence normalisation cannot be used in simulations which
adopt other relationships between the diffusion coefficient and electric field (e.g. Wada
and Frey 1979a, b).
A further variation on the form of semiconductor equation used in simulations is
to work in terms of the electron and hole quasi-Fermi levels $, and $, rather than
 Semiconductor device modelling 229

directly calculating the electron and hole concentrations:

I)n=I)-ln(n) (2.1 1)
(cl,= $+In (PI. (2.12)
These expressions are then used to replace the carrier concentrations n and p in either
the original set of semiconductor equations (2.1)-(2.5) or in the normalised equations
(2.6)-(2.10) (Slotboom 1969, 1973, Sutherland 1980a). By substituting two new vari-
ables, 'pn and p' , where
(2.13)
(2.14)
and assuming a steady-state analysis, the equations (2.6), (2.9) and (2.10) reduce to
three coupled elliptic partial differential equations in I),'p, and p
' ,. This set of transport
equations is generally easier to solve than those which contain the time-dependent
continuity equations which are non-linear parabolic partial differential equations.
The quasi-Fermi level approach has been used in both finite-difference (De Mari
1968, Slotboom 1969, 1973, Sutherland 1980a) and in finite-element (Adachi et al 1979)
numerical solution methods. Quasi-Fermi level models are often used for simulating
devices where both types of carrier are accounted for, and in semiconductor devices
which have a complex energy band structure (pn junction diodes, bipolar junction
and heterojunction bipolar transistors).
Equations (2.4) and (2.5) are often referred to as the semiconductor mobility
equations when the carrier velocities are replaced by expressions incorporating the
carrier mobilities pn and pup:
On = P n E (2.15)
up = p p E . (2.16)
The electric field E is given by
E = -V$. (2.17)
In these circumstances the diffusion coefficients D, and D, may be replaced by the
well-known Einstein relationships:

Dn=-PnkT (2.18)
4

(2.19)

where k is the Boltzmann constant and T is the lattice temperature.

The transport model described by these equations assumes that the velocities
respond instantaneously to changes in the electric field and that the mobility and
diffusion coefficients are functions of the electric field alone. In practice the carriers
do not respond instantly to stimuli and the mobility and diffusion coefficients of carriers
in semiconductors are tensor quantities, dependent on several parameters in addition
to the electric field (for example, temperature and carrier concentration). Furthermore,
the derivation of the continuity and current density equations assumes that the semicon-
ductor has a single-valley, parabolic band structure. In the case of compound semicon-
ductors such as GaAs and InP, which have multi-valley band structures, the analysis
230 C M Snowden

is more complicated since each valley is described by its own Boltzmann equation.
The derivation of continuity and current density equations for these materials requires
the solution of a coupled set of Boltzmann equations. Blotekjaer (1970) has shown
that the electron mobility and diffusion coefficients for compound semiconductors can
be expressed as functions of the average velocity in the lower valley, rather than the
localised electric field. The implications of these departures from the behaviour
described by the basic semiconductor equations will be discussed later.
The terminal current response I for semiconductor models is usually obtained by
integrating the total electric current density J across the surfaces of a suitable plane
through the device

I=
I,Jds

where the total current J includes both the particle currents J,, Jp and displacement
(2.20)

current:
J = J,+ Jp+E ~ E dE/dt.
, (2.21)
The description of the physical model is completed by defining the domain, boundary
and initial conditions over which the device equations previously described apply.
The number of dimensions which must be accounted for in the model is largely
determined by the device geometry, contact and surface properties, and uniformity of
field and carrier distributions. Two terminal devices with simple device structures (for
example, vertical transferred-electron devices (TED), pn junction diodes) are often
represented using one-dimensional models. Most planar devices, including those found
in monolithic structures (integrated circuits), require at least two-dimensional models
to account for the multi-dimensional nature of the electric field and carrier distributions.
In devices where the width is much greater than the length, which is usually the case
with discrete devices, a two-dimensional model is adequate. However, for small
geometry devices where device width is of the same order as the device length, such
as is found in VLSI circuits, a three-dimensional model should be used to account for
fringing fields and non-homogeneous current flow.
The analysis domain chosen to represent the device is an approximation to the real
device and the importance of choosing the correct model geometry cannot be under-
stated. For example, planar devices such as MESFET modelled in two dimensions
should have planar ohmic contacts on the top surface in order to fully account for
local electric-field and space-charge effects. However, the model is often simplified to
reduce computation requirements by imposing ohmic boundary conditions on the side
walls at the source and drain perpendicular to the real contacts.
Two spatial boundary conditions and an initial condition are required to solve the
classical differential semiconductor equations. The surface and contact properties
provide the boundary conditions. The presence of current-free boundary regions as
well as conducting contacts on the surfaces leads to mixed boundary conditions. The
absence of current flow through surfaces may be modelled by assuming that the
potential and carrier gradients normal to the surface are zero, if the surrounding media
are assumed to have zero relative permittivity. This leads to differential (Neumann)
boundary conditions. The effects of surface potentials and surface recombination may
also be included as boundary conditions. Ohmic contacts are usually described by
Dirichlet boundary conditions where potential and carrier concentrations are pre-
defined at the contacts.
 Semiconductor device modelling 23 1

Initial conditions for the Poisson and continuity equation are obtained by setting
the potential to zero (except at contacts) and the carrier density to the doping density.
In successive simulations a considerable time saving can be achieved by using steady-
state distributions from previous solutions as the initial conditions.

2.2. Methods of solution for the semiconductor equations

The semiconductor equations have, for many years, formed the foundation of semicon-
ductor device studies. Closed-form analysis solutions have been obtained for many
device structures, but are severely limited in their range of application and accuracy
because of the multi-dimensional non-linear nature of the problem. A more generalised
method of solution often applied to these equations is to solve them using numerical
techniques. This approach requires considerably more computer time than the closed-
form solutions, but yields more versatile and accurate results. The most common
numerical techniques applied to these sets of partial differential equations are the
finite-diff erence and finite-element methods. The boundary integral method has been
used in several semiconductor simulations (De Mey 1974, 1978, De Visschere and De
Mey 1977, Oh and Dutton 1980, De Mey et a1 1984). This review will concentrate on
finite-diff erence and finite-element methods as they are more commonly used, although
the boundary integral method offers a relatively rapid means of solution because of
the requirement for fewer nodes.

2.2.1. Finite-diference schemes. Finite-diff erence techniques applied to semiconductor

modelling are well established and there is considerable information available on the
stability and convergence properties of these schemes (see, for example, Richtmeyer
and Marton 1957, Reiser 1972, Slotboom 1973, Gerald 1973). Most of the earlier
device simulations such as those of Gummel (1964), Slotboom (1969) and Scharfetter
and Gummel ( 1969) used one-dimensional finite-difference schemes. Even the more
sophisticated simulation packages of the early 1970s (Reiser 1972, Slotboom 1973)
relied almost entirely on this method, largely because of the additional complexity of
finite-element schemes.
Finite-difference methods produce solutions for the physical variables $, n and p
as discretised values at individual nodes throughout a mesh which describes the solution
domain. The continuous derivatives of the semiconductor equations are replaced by
discretised standard finite-diff erence approximations. The discretised variables are
represented by values obtained from the solution of the equations at each mesh point
except where boundary conditions determine the parameter values. The finite-diff er-
ence expansions are derived from truncated Taylor series for f ( x + A x ) and f ( x - A x ) ,
which assume that the function f ( x ) is a continuous, single-valued function of x with
continuous derivatives. First-order approximations are obtained from these Taylor
expansions as forward or backward diflerence formulae, respectively, by truncating
the series after the second term. Second-order approximations may also be obtained
by subtracting or adding the truncated Taylor series to yield first and second diff eren-
tials, respectively, which are also known as central-diff erence approximations and are
of the order of A x more accurate than the first-order expressions. In describing
finite-difference schemes it is convenient to use the notation (for two dimensions)

(2.22)
232 C M Snowden

etc, where the relative positions of the points i, j , etc, are in the Cartesian plane. This
notation is used to describe a mesh superimposed on the region requiring analysis.
Uniform meshes are frequently used because of the ease of calculation. Non-uniform
meshes can be advantageous in regions of high- and low-valued derivatives, with fine
and coarse grid spacing, respectively (see figure 1). This allows greater accuracy for
the higher derivatives and a saving in time and memory in regions of low-valued
derivatives (Slotboom 1973). A major disadvantage of the finite-diff erence technique
when compared with the finite-element method is that the requirement for a finer mesh
in one specific area of the device necessitates the presence of a finer mesh on other
parts of the device (see figure 1). This leads to a surplus of nodes in regions away
from the region requiring refinement which would otherwise have a relatively coarse
mesh. The recently reported finite box method of Franz et a1 (1983) avoids this situation
and allows finite-difference schemes to be applied to more complex non-planar
geometries.

I Ax I
(U) ibi
Figure 1. Finite-difference meshes. ( a ) Uniform mesh, ( b ) non-uniform mesh with finely graded region
(shaded area).

Adaptive meshes, which are ‘self-generating’ with localised mesh refinement and
take account of changes in the potential and carrier distributions, are a desirable feature
for improved accuracy and convergence of the solution. The mesh is chosen so as to
minimise the discretisation error and the number of mesh points required, and is
refined adaptively as the computation progresses so as to limit the difference in
+
electrostatic potential between any two mesh points to be less than a specified limit
(Watanabe and Slamet 1983). Adaptive meshes increase the complexity of the coding
and tend to increase the solution time required but provide a considerable improvement
in accuracy and stability. The novel generalised finite-diff erence scheme recently
developed by Franz er al (1983), which uses finite boxes to perform the discretisation,
generates and adapts the grid automatically. The mesh consists of rectangular cells
which permit lines to terminate in any direction and at any intersection in the mesh.
This method offers a significant saving in computer resources as redundant mesh points
are avoided.
The Poisson equation is usually solved using a five-point difference approximation,
where the potential at a node depends on the potential and charge at the node and on
the potential at the four neighbouring nodes. The finite-difference equation for the
elliptic Poisson equation is thus
 Semiconductor device modelling 233

In matrix form this is represented as

[AI[+] = [BI (2.24)

where matrix [A] is the coefficient matrix, normally of tridiagonal form, and matrix
[B] contains the terms on the right-hand side of equation (2.23).
Finite-diff erence representations of the current continuity equation can be realised
in many forms (Gummel 1964, Scharfetter and Gummel 1969, Reiser 1973, Slotboom
1973). There are two principal approaches to solving the continuity equation-schemes
utilising the explicit current densities at the nodes and schemes based on eliminating
the current densities between equations (2.2)-(2.5) using expanded forms of the
equation. This latter approach works well in devices where localised rapid changes
in carrier concentration have little effect on the bulk properties of the device, such as
in TED, or for materials with small diffusion coefficients like silicon (Dubock 1970,
Snowden et a1 1981). Such schemes usually require less computation than the former
method. However, for devices with high carrier concentrations (> loz3m-3), or for
materials with large diffusion coefficients (gallium arsenide, indium phosphide), the
truncation error in higher-order derivative terms can become significant, particularly
in the vicinity of depletion regions. The more popular approach is to develop a
finite-difference scheme based on the current densities at each node (Gummel 1964,
Keiser 1973, Slotboom 1973). A Crank-Nicolson implicit scheme is usually adopted,
which provides greater accuracy and better convergence properties than is possible
with explicit solution methods (explicit schemes usually require excessively small time
steps to guarantee stability). The general implicit form of the continuity equation (for
electrons) based on current densities is
ktI
dn _.ni,j
-- -ni,j
(2.25)
3t At 2q
where the superscript k refers to the time step. If this equation is treated as being
fully implicit in both velocity ( P E )and carrier concentration, it leads to a complicated
system of coupled, non-linear equations, which is difficult to solve. An advantage of
this approach is that it is stable for all values of time and space step (not true of
explicit schemes which have strong tendencies to instability), although smaller values
of Af, Ax and Ay yield more accurate results. The continuity scheme (equation (2.25))
is usually linearised to allow conventional methods of solving linear systems of
equations to be used (Varga 1962). The linearisation usually implemented requires that
U;;' = U:, (2.26)
which makes the equation implicit in carrier concentration and explicit in velocity.
This leads to an asymmetric coefficient matrix which yields complex eigenvalues in
the Jacobi iteration matrix (Reiser 1971, 1972). As a result ofthis, the successive under-
relaxation ( S U R ) technique has to be used to solve this set of equations rather than
the more familiar successive over-relaxation (SOR) technique. The schemes of
Scharfetter and Gummel (1969) and Slotboom (1973) produce positive definite matrices
and do not lead to coefficient matrices with complex eigenvalues and hence can be
solved using SOR techniques (which often have a faster rate of convergence). Both
approaches have been used extensively in finite-diff erence simulations.
Stability analysis for non-linear systems is a highly complicated process and no
generalised analysis is available. Reiser (1972) has shown that for linear explicit
234 C M Snowden

solutions of the continuity equation

(2.27)

where vUmis the high field saturated constant velocity. Richtmeyer and Marton (1957)
indicate that the approximations obtained using linear stability analysis theory also
provide useful guidelines for non-linear problems. However, the explicit treatment of
drift velocity in the linearised continuity equation gives rise to non-linear instabilities
which manifest themselves as standing waves superimposed on the required solution
when large time steps are used (Reiser 1972). Furthermore, these erroneous perturba-
tions are amplified rather than damped if the step width is excessive and it is necessary
to restrict the time steps to less than 1 ps to minimise this effect. In order to maintain
a physically meaningful solution the time step A t should be of the order of the dielectric
relaxation time and the space steps Ax, Ay should not be significantly larger than a
Debye length, both of which are functions of doping density and dielectric permittivity.
The finite-diff erence implementation is completed by obtaining suitable expressions
for the boundary conditions. Dirichlet boundary conditions are easily implemented
using finite-diff erence equations, with fixed potentials and carrier densities at the
contact nodes. The derivative (Neumann) boundary conditions are generally more
difficult to implement and erroneous solutions often occur in the transition from
between Neumann and Dirichlet boundaries. Suitable boundary condition approxima-
tions are obtained from Newton polynomials and Stirling polynomials (Gerald 1973,
Fenner 1978). The approximation obtained from Stirling polynomials requires extra
‘phantom’ nodes outside the surface of the device and leads to singularities in the
solution near contacts. The Newton polynomial approximation avoids these difficulties
and provides a degree of accuracy comparable with the difference expansions of the
Poisson and continuity equations. The use of this type of boundary condition means
that the coefficient matrix does not satisfy the diagonal dominance conditions necessary
to guarantee convergence in interactive solutions. However, despite this, convergence
is normally possible although the optimum relaxation factor must be determined
empirically.

2.2.2. Methods of solvingjnite-digerence equations. The set of simultaneous equations

formed by the finite-diff erence expressions for the Poisson and continuity equations
can be solved by direct or iterative techniques. A direct method of solving tridiagonal
(and pentadiagonal) sets of simultaneous equations, in which only the non-zero entries
in the coefficient matrix need be stored, can be adapted from the Gaussian elimination
technique. Another commonly used technique for the solution of the Poisson equations
is based on the fast Fourier transform (FFT) (see, for example, Hockney’s POT4 (1970)).
This method requires a moderately lengthy coding compared with iterative methods.
However, for a large number of equations with a sparse coefficient matrix, iterative
methods are normally faster, require less memory and are more accurate.
Three common iterative methods are Gauss-Seidel, successive over-relaxation (SOR)
and the cyclic Chebyshev method. The Gauss-Seidel method requires a diagonally
dominant or positive-definite coefficient matrix to guarantee convergence. Successive
over-relaxation can be used to increase the ratio of convergence of the Gauss-Seidel
method by multiplying the residual by a relaxation factor w (the residual being the
difference between the updated value of the variable and the previously calculated
value: it is zero for the exact solution). The relaxation factor can be calculated for
 Semiconductor device modelling 235

diagonally dominant coefficient matrices and lies in the range 1-2 for SOR (Varga
1962). There is no corresponding analytical method for selecting the relaxation factor
for non-rectangular regions. Under-relaxation (0 < w < 1) is usually used where there
is no diagonal dominance in the coefficient matrix.
In the cyclic Chebyshev method, the value of the relaxation factor w is changed
every half-iteration to improve the rate of convergence. The initial rate of convergence
is far higher for the Chebyshev method than for the straightforward SOR scheme, but
if few iterations are required the extra calculation required for w offsets the time saving
obtained by this method.
An important advantage of iterative methods is that round-off errors in the solution
are limited to those incurred in the final iteration and are normally negligible. This is
because the values used in the last iteration are treated as initial estimates and updated
independently of the rounding error. Direct elimination methods are subject to cumula-
tive rounding errors which can become significant for large coefficient matrices or for
ill-conditioned systems of equations.
Several solution techniques are available for solving large systems of simultaneous
equations such as those generated by the finite-difference method. The most efficient
solution technique is the one-step block SOR method, which is relatively easy to
implement and requires half as much storage as the popular alternating direction
implicit ( A D I ) scheme. LU decomposition methods have also been used successfully
to solve the non-linear current continuity equation.
There are three principal approaches to solving the coupled Poisson and continuity
equations-the simultaneous method, the alternating method (Gummel’s method) and
the sequential solution method. The simultaneous method produces a combined
solution for both the Poisson and continuity equations simultaneously and utilises a
matrix twice as large as the individual matrices, requiring substantial computation
effort. The popular Gummel(l964) algorithm starts by obtaining a converged solution
for the Poisson equation to obtain the node potentials. The carrier densities are then
updated using the new node potentials and the Poisson equation is solved once again.
This procedure is repeated until the potential converges to within the preset limit. In
this way self-consistent solutions are obtained for the potential and carrier distribution.
The sequential solution method proceeds by first solving the Poisson equation and
then solving the current continuity equation using the updated potentials. Several
variations on this technique exist and it is particularly suited to time-dependent
solutions.

2.2.3. Examples ofjinite-diference applications. The majority of early numerical device

simulations were carried out using finite-diff erence techniques and the method still
remains a popular choice because of the simplicity of implementing this technique.
There are many published examples of the approach for one-, two- and three-
dimensional simxlations.
Two terminal vertical devices (as opposed to planar structures) have traditionally
been modelled using one-dimensional simulations. Devices in this category include
p-n junctions (De Mari 1968), Schottky varactor diodes (Tasker 1983) and transferred-
electron devices (TED) (see, for example, McCumber and Chynoweth 1966, Kroemer
1966, Thim 1968, Robrock 1969, Freeman and Hobson 1972, 1973, Grubin et a1 1973,
Lakshminarayana and Partain 1980).
Transferred-electron devices are fabricated from GaAs and InP and are of particular
interest because of their ability to generate spontaneous and coherent microwave
236 C M Snowden

current oscillations when biased correctly. Numerical simulations provide the physicist
with a dramatic picture of the physical phenomena behind the operation of this type
of majority carrier semiconductor device. TED,also known as Gunn diodes, are usually
fabricated from epitaxial layers of GaAs in an n+-n-n+ layer structure (figure 2).
One-dimensional models are used to represent this structure because domains propagate
between the parallel ohmic contacts and because the length of the active n layer is
usually much smaller than the device diameter. The term ‘domain’ describes the
non-uniformity in the electric field and electron distribution which characterises the
operation of TED (see Hobson (1974) for a more complete description). The domain
travels through the device towards the anode. TED can operate in a variety of modes,
determined by the conditions governing the formation and extinction of the domains.

C 16 17 18
l g ND

Figure 2. Transferred-electron device (TED or Gunn diode) and a typical doping profile. The section through
the device is typical of the profile used in one-dimensional models.

Numerical simulations of TED can predict essential performance characteristics

such as microwave output power, stability, efficiency and tuning characteristics by
analysing the large-signal behaviour of the device. Such simulations often proceed by
eliminating the electron density from the current density expression by combining
Poisson’s equation with the current density expression and neglecting generation and
recombination effects (see, for example, Grubin 1978, Lakshminarayana and Partain
1980)

(2.28)
The electric field E is then constrained by the applied voltage V ( t ) :

(2.29)

where
V( t ) = V,, - V,, cos ut (2.30)
with the additional boundary conditions:
E(0,t ) =E(L,t )=0 (2.31)
where L is the length of the active device. The diffusion D ( E ) and velocity v ( E ) are
instantaneous functions of the electric field E in this model and follow the relationships
for GaAs shown in figures 15 and 16. Note that the form of equation (2.27) treats
 Semiconductor device modelling 237

electrons as negatively charged particles (with U = - P E ) . Many authors prefer to treat

electrons as positively charged particles in simulations with only one species of carrier,
which requires the sign of the diffusion term to be changed in equation (2.28).
The exact form of the diffusion term in equations (2.4), (2.5) and (2.28) has been
the subject of considerable debate in the literature. The diffusion coefficient D is shown
outside the derivative term in these equations, which is the usual form. However,
within the context of derivations based on the Boltzmann transport theory, it has been
shown that the coefficient D should, strictly speaking, be included inside the derivative
term d(Dn)/dx (Carroll 1970, Stratton 1972). The derivation of both of these forms
assumes that the distribution function is disturbed only slightly from its equilibrium
condition. In highly non-equilibrium situations a displaced Maxwellian method (Bosch
and Thim 1974) is often used to analyse the transport process and it is difficult to
isolate a diffusion term. This latter problem will be discussed in more detail under
the section dealing with non-equilibrium transport.
Considerable simulation effort has been devoted to GaAs TED over the last 18 years.
It has been established that the nature of the contact properties (doping and mobility
variations and interfacial fields), particularly at the cathode, determine the operational
behaviour of the device. Doping notches near the cathode are frequently incorporated
into the models to approximate the domain nucleating mechanism. Typical results
obtained from time-dependent one-dimensional finite-diff erence numerical simulations
of GaAs TED are shown in figure 3. Models have been developed which take into
account the effects of space- and time-dependent ionised impurities due to carrier
generation-recombination mechanisms (Maksym and Hearn 1978). Indium phosphide
TED have also been modelled to investigate the higher efficiencies of this device
compared with GaAs TED (Colliver et al 1973).

0 i 0 li 0
Length( pm 1
Figure 3. Time evolution of domain electric field along the length of the Gunn device operating in the
dipole domain mode (from Lakshminarayana and Partain 1980). f = 10 GHz, V,, = 10 V.

Finite-difference methods have been used to analyse a wide variety of three terminal
and planar devices, including BJT (Slotboom 1973, Heimeier 1973), JFET (junction
field-effect transistors) (Kennedy and O’Brien 1969), MOSFET (Oka et a1 1980, Liu et
a1 1980, Mock 1973, 1981), MESFET (Reiser 1973, Yamaguchi et at 1976, Snowden
1984a), planar TED (Goto et al 1975, Doades et al 1984) and thyristors (Berz et a1
1984). Of all these devices, the MOSFET (or insulated-gate field-effect transistor (IGFET))
23 8 C M Snowden

is the most-studied device because of its crucial importance in the VLSI industry. The
small geometry of the MOSFET is necessary to obtain the highest packing density and
speed possible in integrated circuits. Models of these small geometry MOSFET must
be at least two-dimensional to account for the multi-dimensional nature of the field
and carrier distributions. In order to minimise the costs of developing new types of
devices and circuits, numerical simulations play an important role in the design stage
of new VLSI circuits.
Gate

Bulk substrate

Figure 4. Two-dimensional model of a silicon n-channel MOSFET.

MOS simulations involve two types of carrier and hence it is necessary to solve the
full set of semiconductor equations (2.1)-(2.5) for the model. Figure 4 shows a typical
two-dimensional model of a silicon n-channel MOSFET with heavily doped n+ contact
regions, the p-type bulk and the semiconductor-oxide interface beneath the gate. In
the model the carrier densities at the source and drain contacts are set equal to the
doping concentrations (nt). The boundary condition for the oxide region assumes
that no space charge exists in this region and hence only the Laplace equation has to
be solved:
v2* = 0. (2.32)

At the interface the following boundary condition is applied from Gauss’ law:

a*
€oxide-
ay I oxide
= cs/c-
a*
ay I semiconductor
(2.33)

Surface states have also been included in some models (Sutherland 1980b), although
their effect on the solution is very small and can be accounted for with the flatband
voltage (Selberherr et al 1983). The gate potential is set to the applied bias voltage
minus the flatband voltage. The potential on the source, drain and bulk contacts is
determined by the applied bias voltage plus the built-in potential, which assumes that
these are ideal ohmic contacts. The current-free vertical surfaces either side of the
gate contact and at each end of the device model are assumed to have zero potential
derivative boundary conditions normal to the surfaces ( E , = 0). Many authors assume
a one-dimensional electric field in the oxide (for example, Toyabe and Asai 1979),
although this approximation may not be valid for small devices (Selberherr et a1 1983).
 Semiconductor device modelling 239

Typical results for a device with a uniformly doped substrate ( lo2’ m-3) are shown
in figure 5. The n+ implants below the contacts are doped at 5 x m-3, with the
p-n junction approximately 0.3 p m below the surface (Selberherr et a1 1980).

Figure 5. MOSFET simulation results showing ( a ) doping concentration, ( b ) potential and ( c ) electron
distributions (from Selberherr et a1 1980).

A detailed description of MOSFET modelling techniques is given in Yu and Tan-

traporn’s (1982) article. Selberherr et a1 (1983) also provide a comprehensive descrip-
tion of Mos-transistor modelling. The numerical modelling of MOSFET has become
sufficiently well established and of enough importance to industry to merit the develop-
ment of a host of simulation software packages. Notable two-dimensional application
packages for MOSFET modelling include CADDET (Toyabe et a1 1978), GEMINI (Green-
field and Dutton 1980) and M I N I M O S (Selberherr et a1 1980). MINIMOS is capable of
analysing coupled devices in monolithic silicon circuits such as NMOS and CMOS
inverters.
Another type of semiconductor device which has been extensively researched using
two-dimensional finite-diff erence numerical simulations is the MESFET.The MESFET
is a planar transistor with ohmic source and drain contacts on either side of a Schottky
gate contact (figure 6). The active layer beneath the contacts is a very thin layer of
highly doped n-type semiconductor (typically 0.15 p m thick doped at 1.5 x mP3
for GaAs devices). A buffer layer is frequently interposed between the active layer
240 C M Snowden

and semi-insulating substrate. The GaAs MESFET forms the foundation of active MMIC
(microwave monolithic integrated circuits) and has replaced junction FET (JFET) and
Si BJT as a microwave transistor above 4 GHz. GaAs MESFET technology is also set
to form the basis of the next generation of gigabit logic rate integrated circuits for
digital applications. Current GaAs MESFET have a very low gate length to active layer
thickness aspect ratio, which means that the electric field in the active channel has a
two-dimensional distribution. In these circumstances Shockley’s (1952) gradual chan-
nel model is wholly inadequate to model these devices and it is necessary to resort to
two-dimensional numerical simulations.

Source Gate Drain

~~

Figure 6. Two-dimensional representation of a planar MESFET.

MESFET are unipolar devices and hence the set of classical semiconductor equations
which are frequently used to describe their operation is reduced to one describing
electron transport only. This simplifies the analysis considerably, as generation and
recombination are usually neglected. The current-free surfaces and the boundaries
inside the semiconductor bulk are modelled with zero-valued derivatives normal to
the boundaries for both the potential and carrier density. The ohmic contacts are
assumed to be ideal and have a pre-specified potential and fixed carrier concentration
imposed as boundary conditions. This requires that the ohmic contacts are neutral
and that they are sufficiently remote from the active region of the device that they have
no influence on the electrical behaviour of the transistor. The Schottky gate is charac-
terised by the difference in work function between the metal contact and the semicon-
ductor material, with a potential fixed at the applied voltage minus the built-in voltage
and carrier concentration often based on diffusion theory:
(2.34)
(2.35)
Other Schottky boundary conditions have been modelled using thermionic emission-
diffusion theory and thermionic emission theory (Tasker 1983, Snowden et a1 1983,
Sangiorgi et a1 1984). The source potential is frequently assumed to be zero, while
the gate and drain have static or time-dependent voltages applied.
 Semiconductor device modelling 24 1

Considerable effort has been devoted in the literature to investigating the many
different aspects of MESFET technology. The effects of different doping profiles, buff er
layers and semi-insulating substrates have been considered (see, for example, Reiser
1973, Snowden 1984a). This work has demonstrated the importance of including buffer
layers and substrates in the models to produce an accurate analogy of device operation.
Numerical MESFET simulations also clearly reveal the presence of a static dipole region
in the active channel under the drain end of the gate in GaAs MESFET (Yamaguchi et
a1 1976). In some circumstances GaAs FET structures are capable of supporting
spontaneous current oscillations attributable to the negative differential mobility of
the material and this has been investigated by several workers (Yamaguchi et a1 1976,
Grubin et a1 1980).
Typical results for a two-dimensional simulation of a 0 . 6 ~ mgate length GaAs
MESFET are shown in figure 7. Wada and Frey (1979a, b) have shown that it is necessary
to account for non-equilibrium transport conditions in devices with gate lengths of the
order of 0.5 Fm or less and this will be discussed later in this review. Wada and Frey
(1979b) also describe how the simulation package CUPID has been used to model Si,
GaAs and InP MESFET.Faricelli et a1 (1982) extended this work to characterise the
transient behaviour of short channel MESFET both in isolation and in circuit models.

(a1 (bl

Source Gate Drain Source Gate Drain

0 04 08 12 16 20 0 04 08 12 16 20
x ipmi xipm)

Drain Drain

,102
-a l .
- 3
m

i E 2
C
1
0

Figure 7. Typical potential ( a ) and electron ( b ) distributions obtained from a two-dimensional simulation
of a GaAs 0.55 Fm gate length MESFET, VDs= 6 V, v,, = 0V. ( a ) Values on the curves are given in volts,
( b ) values o n the curves are given in m-3.

Recent advances in finite-diff erence simulation techniques have extended research

into the areas of three-dimensional modelling and the analysis of non-planar devices.
In very small geometry VLSI MOSFET the channel width of the FET is of the same order
as the depletion layer thickness and three-dimensional effects strongly influence the
operation of the transistor. Also on this scale of integration the devices appear highly
non-planar in construction. Husain and Chamberlain (1982) have described a three-
dimensional finite-diff erence simulation, WATMOS, for modelling small geometry
MOSFET.Using this simulation they have investigated the sensitivity of the electrical
242 C M Snowden

behaviour to variations in physical, geometrical and technological parameters. They

have demonstrated that three-dimensional fringing fields affect the threshold potential
of small geometry devices and that the effective width of the device is non-uniform
throughout the channel region.
The application of two-dimensional finite-diff erence techniques to non-planar
device models is discussed in the published work of Greenfield et a1 (1983) and Franz
et a1 (1983). Non-planar device structures are analysed using a non-uniform finite-
difference mesh. Adler (1979, 1980) describe how variable-density meshes can be
implemented for semiconductor simulations. The discretisation of Poisson’s equation
is achieved by integrating equation (2.1) over the local area surrounding each node
and then approximating the integrals to produce a ‘five-point’ difference approximation
similar to that discussed previously. The discretised Poisson equation is then solved
using a highly efficient one-step block SOR technique. This technique has been applied
to the study of non-planar MOS transistors. The simulation B A M B I , based on finite
boxes (Franz et a1 1983, Franz and Franz 1984), has been recently developed to model
unrestricted device geometries and has been applied to bipolar and MOS devices. In
situations where it is decided to use non-rectangular grids, it is necessary to resort to
using finite-element methods, discussed in the next subsection.

2.2.4. Finite-element schemes. The finite-element method, used frequently in analysing

structures in civil and mechanical engineering, provides a flexible means of investigating
complex semiconductor device geometries. This method is fundamentally quite differ-
ent to the finite-difference technique. The geometric model of the device is sub-divided
into smaller regions-the finite elements (triangles or rectangles). The continuous
functions representing the required solution are assumed to have simple analytic forms
within each element (piecewise-linear, cubic). The continuous semiconductor
equations are then used to define an equivalent integral formulation. As with the
finite-difference method, this method transforms the continuous functions of the
semiconductor equations into discretised forms. The finite-element technique is usually
implemented using the Galerkin method of weighted residuals. The domain requiring
analysis is sub-divided using piecewise approximations to produce a mesh with N
nodes at intersections of the mesh (figure 8). Shape functions Bi(x,y ) are defined to
form the basis of the finite-element formulation, where
e i ( x j ,y j ) = 0 for i + j
i = 1-N. (2.36)
e i ( x j ,y j ) = 1 for i = j
These shape functions, which are usually linear functions on triangular elements of
+,
size h, cover the analysis region R and allow approximations to n and p to be made:
N
(2.37)

(2.38)

The finite-element equations are obtained from equations (2.1)-(2.3) by multiplying

them by Bi and integrating over the region Q. This operation corresponds to setting
 Semiconductor device modelling 243

Figure 8. Finite-element mesh with triangular elements. The shaded region indicates localised mesh
refinement.

the residuals orthogonal to the set of functions rather than making the residual zero
as in finite-diff erence schemes. Hence the Poisson and current continuity equations
(‘just taking the case for electrons here) after applying the divergence theorem become,
respectively,

.
Oi EV$ n d l =
5, E V + . VOi - goi(ND- n + p - N A )d V (2.40)

.
OiJ, n dl =
I, qOi dn/dt + V O i . J , - qOiG d V (2.41)

where V is the volume and ii is the normal to the boundary r of the domain R. By
inserting the approximations nh and p h into these equations, a set of finite-element
equations are obtained. The integrals over r vanish because of the boundary conditions
and definition of the shape functions. Assuming p and n are known from the previous
time step, then the Poisson equation can be expressed as
N
[K$$F - qM, ( N b j - n:+p; - N k j )= 0 1SiGM (2.42)
j=i

where

(2.43)

Here M is the number of nodes not lying on contacts ( M < N ) . Expressing this
arrangement in matrix notation:

[KG]~lh-q[M](Nhg-nh+ph-N~)=0. (2.44)
244 C M Snowden

Similarly the continuity equation (here, for electrons) is expressed in the form
anh
q[jt4],,+[K"]nh-Fh=O (2.45)

where

(2.46)

(2.47)

The time derivatives a n h / at, d p h / a t are usually obtained using finite-diff erence schemes
because of the unwanted coupling between time steps frequently produced by finite-
element methods. As in the case of finite-difference continuity schemes, the equation
is usually linearised across one time step (see, for example, Salsburg et al 1983),
working in terms of [ K " ]and F h at the previous time step. Barnes and Lomax (1977)
have also investigated fully implicit (non-linear) schemes, using Newton-Raphson
techniques to extract [ K " ] and [ F h ]at the new time step, although they found that
the additional task of extracting the Jacobian considerably lengthened the solution.
An advantage of the fully implicit system is that a rapid convergence is obtained
compared with the partially implicit schemes.
The finite-element matrix equations for the Poisson and current continuity equations
are usually solved using direct methods (which normally require large amounts of
storage) or an iterative solution such as successive over-relaxation ( SOR). The device
terminal currents and induced charge can be readily extracted for the contacts by
rearranging the matrix expressions for the current continuity and Poisson equations,
respectively.
The choice of shape function O,(x, y ) and the method of dividing up the analysis
domain 0 strongly influence the precision of solution and the rate of convergence. A
linear shape function for triangular elements is most commonly used (Barnes and
Lomax 1977, Adachi et a1 1979) where the shape function is of the form:
cpZ(X,Y) =(a,+b,x+c,y)/2A (2.48)
where A is the area of the triangle and a,, b,, c, are constants determined from the
constraints on cp, (equations (2.35) and (2.36)) which vary from element to element.
The gradient of cp, is constant within an element which can lead to discontinuities in
the functions at element boundaries. The Hermite bicubic approximation described
by Barnes and Lomax (1977) overcomes this difficulty, although it does not allow for
actual discontinuities in the model (such as contacts).
The finite-element method has been used to analyse a variety of semiconductor
devices including GaAs MESFET, bipolar transistors and MOSFET. However, although
it normally requires fewer nodes than finite-diff erence schemes, a matrix re-ordering
algorithm is usually necessary to deal with the complex matrix structure which occurs,
in order to minimise computer storage requirements and avoid slowing down numerical
processing (Berry 1971). Despite this the finite-element method offers several important
advantages. It is easy to analyse irregular geometries and the computational mesh can
be graded to provide a fine mesh in regions of rapid change of variable. This localised
mesh refinement is generally easier to implement than in finite-difference schemes.
Another advantage of the finite-element method is that it allows high-order approxima-
 Semiconductor device modelling 245

tions to be readily produced. A key property of this'technique is that the conservation

equations (including current continuity) are exactly satisfied, even using relatively
coarse numerical approximations. Barnes and Lomax (1977) have shown that the total
current is conserved on an element-by-element basis and that the result is independent
of discontinuities in the shape functions. The principal disadvantages encountered in
using the finite-element method compared with finite-difference schemes are increased
complexity of the programming and the greater density of the matrix equations. This
latter factor is particularly significant if higher-order approximations and non-uniform
meshes are used. A non-uniform mesh, however, usually reduces the number of matrix
equations required which in turn tends to compensate for the additional complexity
of the matrix.

2.2.5. Examples of5nite-element applications. Finite-element methods are ideally suited

to two- and three-dimensional modelling of non-planar devices because of the adaptable
efficient nature of the mesh. Finite-element computational methods have been applied
to simple one-dimensional structures such as p-n junctions (Please 1982, Bettess et a1
1984), but were not generally applied to semiconductor device modelling until the
early 1970s. Barnes and Lomax (1973) reported one of the first two-dimensional
finite-element simulations applied to semiconductor devices, which was used to model
short gate length GaAs MESFET.Since that time, this approach has become comparable
in popularity to the finite-diff erence method.
The popularity of finite-element simulation techniques has grown substantially over
the last ten years and has been applied to the modelling of a wide variety of semiconduc-
tor devices including MOSFET (Oh et a1 1980, Kotecha and Noble 1980, Chamberlain
and Husain 1981, Hachtel et a1 1981, Buturla et al 1981, Salsburg et a1 1983), vertical
metal-oxide semiconductor (VMOS) FET (Cottrell and Buturla 1978, Salsburg et a1
1983), MESFET (Barnes and Lomax 1973, 1974), BJT (Adachi et a1 1979, Hachtel et a1
1981) and non-planar Schottky diodes (Sangiorgi et a1 1984). IBM have developed
the FIELDAY simulation package, based mainly on finite-element techniques, to model
a variety of semiconductor devices with particular emphasis on VLSI structures (Buturla
et a1 1981, Salsburg et a1 1983). This simulation is currently being used in the
computer-aided design (CAD) of VLSI devices to predict device behaviour and optimise
device designs prior to fabrication. This cost-effective method of evaluating new ideas
for integrated circuits makes this feature of physical modelling techniques very
attractive.
The detailed analysis of a silicon bipolar npn transistor, with generation and
recombination included in the model, provides a rigorous test for this method. The
Shockley-Read-Hall model for generation and recombination is commonly used in
bipolar models:
G = ( p n - n t ) [ T n o ( p + ni)+ T ~n +
~ ni)]-'
( (2.49)
where T~,,,T~~ are the electron and hole lifetimes (100 ns each), respectively. Adachi
et a1 (1979) have modelled an npn planar transistor for integrated injection logic ( I ~ L )
applications (figure 9). Results obtained from this model are shown in figure 10 and
include carrier distributions for high injection levels. Adachi et a1 found that the
higher the injection level, the slower the rate of convergence of the solutions. This
model allows the junction capacitances CBEand CBcto be calculated from the charge
Q in the device, as well as the current gain p, cut-off frequency fT and the static
current-voltage characteristics.
246 C M Snowden
Emitter Bose
3 3

2 2
-E
.1
.1 oolAl
01
-E
1
1

CI L 0 12 1 1 6 7 0
Collector ipml

Figure 9. ( a ) Two-dimensional model of a planar npn transistor. Full lines indicate n-type impurity
concentration. Broken lines indicate p-type concentrations. ( b )Triangular finite-element mesh for the model,
with regions of mesh refinement at junction interfaces (from Adachi et al 1979). Values on the curves are
in ioz5

Emitter Base

I
0 16

I 1 -
0 8 16

Figure 10. Simulation results for an npn planar transistor simulation. ( a ) Equipotential plot, ( b ) electron
distribution, (c) hole distribution for high injection level (V,, = -0.8 V, V, = 0 V) (from Adachi et al 1979).
( a ) Values on the curves in volts, ( b ) ( c ) values on the curves in m-3.
 Semiconductor device modelling 247

The application of finite-element techniques to non-planar devices is well illustrated

by Salsburg et a1 (1983) in their modelling of short and narrow silicon MOFSET. A
simplified two-dimensional diagram of a MOFSET device structure is shown in figure
4. Short- and narrow-channel effects can be modelled independently using this two-
dimensional model but to incorporate both effects in the same device model requires
a three-dimensional approach. Short channels cause a decrease in threshold voltage
because of the potential barrier lowering due to the proximity of the large drain
potential, whilst narrow channels have the opposite effect. Buturla et al (1981), who
have analysed devices as short as 0.7 p m and as narrow as 1.5 pm, claim that variations
in channel length may contribute more than 50% to the total threshold tolerance of
MOSFET incorporated into VLSI chips.
The structure of a three-dimensional mJdel for small-geometry MOSFET is shown
in figure 11. Simulated sub-threshold characteristics were used to define the threshold
for a device 1.5 p m long by 1.5 p m wide. It was found that a three-dimensional model
is essential to accurately predict the source to substrate bias dependence of the threshold
(Buturla et a1 1981). The results of modelling the surface potential for two different
structures are shown in figure 11. In the first structure the source and drain diffusions
extend under the field oxide whereas in the second case they terminate at the edge of
the field oxide region. This example clearly shows the importance of three-dimensional
modelling of VLSI circuit elements in modern semiconductor technology.

Field
region

Channel

- Structure A
---- Structure B

(U)

Drain e

Source \

- Thick oxiae
field region field region
Structure A Structure B

(cl
Figure 11. Three-dimensional MOSFET (IGFET)simulation. ( a )Structure of a short and narrow MOSFET.
( b ) The finite-element mesh used to simulate a 1.5 x 1.5 ,um2 short and narrow MOSFET containing 13 planes
of 164 nodes each. ( c ) Modelled surface potential for structures A and B at a drain to source bias of 5 V
and a source to substrate bias of 1 V. The device length and width is 1.5 p m (from Buturla et al 198 I ) .
248 C M Snowden

Finite-element methods provide a flexible means of investigating semiconductor

devices with non-planar geometries and non-linear field distributions. Present trends
in finite-element modelling techniques include the use of different kinds of elements
(triangles and rectangles) in the same grid, thus allowing arbitrary subdivision and
expansion of grid spacings (Machek and Selberherr 1983). This important development
ensures that only the minimum number of nodes is generated. Machek and Selberherr
( 1983) also introduce the concept of variable and potential-dependent shape functions.
These novel shape functions are used for the exponentials of the quasi-Fermi levels
which describe an exponential distribution of the carrier densities exactly.
The development of optimised finite-element meshes is another important step in
achieving improved simulation techniques. Guerrieri and Rudan ( 1983) have recently
presented a technique for obtaining the optimum finite-element solution of the Poisson
equation using a mesh optimisation technique. Mesh optimisation is important because
functions may exhibit sharp variations in isolated regions of the domain while being
almost constant in other regions. Also, the boundary conditions may become non-
analytical in regions where Dirichlet boundary conditions for conducting boundaries
abruptly change into Neumann boundary conditions along current-free surfaces, which
may lead to singularities in the solution (Strong and Fix 1973). Non-analytical bound-
ary conditions can also cause non-physical effects in the solution. Another important
requirement for optimised meshes is in regions where internal boundaries and discon-
tinuities occur.
Bettess et a1 (1984) have attempted to identify optimum methods of solution for
semiconductor models using the finite-element method. They suggest that, amongst
many possible methods of solution, the most promising aids may prove to be special
shape functions and the use of Petrov-Galerkin weightings (Brooks and Hughes 1982)
rather than the usual Galerkin weightings. These techniques may increase the efficiency
of solution and help resolve difficulties in pn junction modelling which occur in the
narrow interface region (where spurious oscillations in the solution often occur).

2.3. Semiclassical transport equations

The basic classical semiconductor equations, which have for many years formed the
foundation for device modelling, are based on the assumption that the carrier energy
distribution remains close to its equilibrium form. However, if the electric fields, carrier
gradient and current densities in the device become excessively large then non-
equilibrium transport conditions will prevail. In practice these conditions occur in
many semiconductor devices used in VLSI, VHSIC, MMIC and discrete high-frequency
applications. The small device structures encountered in advanced semiconductor
technology generally have regions of very high electric fields which cause substantial
electron ‘heating’, creating regions where the electrons attain very high energies relative
to the equilibrium conditions. Ruch (1972) first demonstrated with the aid of Monte
Carlo methods that strongly non-homogeneous electric fields in sub-micron semicon-
ductor devices caused peak velocities far in excess of those attributed to equilibrium
velocity-field characteristics. His investigations into short-channel GaAs and Si FET
revealed that this velocity overshoot effect dominated the carrier transport process in
GaAs and Si FET with active channel lengths of the order of 1 p m and 0.1 pm,
respectively.
Figure 12 shows the time dependence of the average velocity of electrons injected
cold and drifting in a region of uniform electric field for samples of silicon, gallium
 Semiconductor device modelling 249

arsenide and indium phosphide. In both cases the drift velocity overshoots to a value
significantly higher than the equilibrium value for each applied electric field. This
effect can be attributed to the difference in the momentum and energy relaxation times.
In both cases momentum relaxation time is smaller than the energy relaxation time,
causing the electron distribution to be perturbed first in momentum space. As the
energy relaxation becomes effective the distribution function spreads out, causing the
average drift velocity to decrease. Strong intervalley and acoustic phonon scattering
in silicon causes the drift velocity to reach its equilibrium value in less than 0.1 ps.
The transient effects in gallium arsenide last substantially longer because of weaker
influence of polar optical scattering which is the dominant mechanism. Velocity
overshoot thus improves upon the performance of high-speed and microwave devices
as predicted by classical models which do not include carrier heating.

-x
0
7
o
, 1
002
1 1
001
1 1
006
! I
008
"
01 0
,
2
/ l
4
l l
6
l l
8
l \
10
c

c
'c
I M V m-')

0 05 10 15 20 c 2 4 6 0 10
Distance Ipm) Time i p s )

Figure 12. Time and space dependence of the average velocity of electrons injected cold and drifting in a
region of uniform field for samples of Si, GaAs and InP. ( a ) ( b ) Si, ( c ) ( d ) -, GaAs, -. -, InP. V,
Ruch (1972); 0 , Maloney and Frey (1977).

Although Monte Carlo techniques provide a strong insight into the physics of hot
electron effects, the computation time required to conduct a complete FET simulation
is prohibitively excessive, requiring several hours per bias point (Hockney and Reiser
1972). Furthermore, problems in modelling carrier-carrier scattering and quantum
effects limit the validity of the model. An alternative approach is to use a 'hydro-
dynamic' semiclassical model derived from the Boltzmann equation. This method has
been explored by a number of researchers (see, for example, Carnez et a1 1980, Grubin
1980, Higgins and Pattanayak 1982) and a detailed description of this method is given
by Cook and Frey (1982). McCumber and Chynoweth (1966), Blotekjaer (1970) and
250 C M Snowden

Curtice and Yun (1981) have all developed transport models based on electron
temperature.
A set of semiclassical transport equations suitable for describing the effects of
non-equilibrium transport in small semiconductor devices may be obtained by assuming
that the distribution function is symmetrical in momentum space. This yields the
following three conservation equations (for electrons):
particle conservation
an
(2.50)
at

momentum conservation

(2.51)

energy conservation
1
%+U. V f = q u e E --V (2.52)
at n
where 5 is the electron energy consisting of kinetic and thermal components:
5 = itn * v 2 +5 kT, (2.53)
where m* is the effective mass of the electron and T, is the electron temperature. The
collision terms are given as

(2.54)

(2.55)

(2.56)

where T~ and T , are the momentum and energy relaxation times, respectively, which
are themselves functions of the average electron energy. CO is the equilibrium electron
energy corresponding to the lattice temperature To. The energy conservation equation
(2.52) assumes a displaced Maxwellian distribution, where higher-order moments of
the distribution function disappear. Because of this assumption, heat flow due to the
electron gas (the third moment of the distribution function) is not included in equation
(2.52). Blotekjaer (1970) points out that this term may be significant for non-Maxwellian
distributions.
From a practical point of view most researchers are interested in the characteristics
of devices in the context of timescales associated with circuit time constants (typically
10 ps to i 0 ns). This represents, at the very least, an order of magnitude difference in
timescale between the device and circuit responses and implies that a quasi-steady-state
model is adequate for most purposes. This means that it is possible to neglect the terms
d v / a t , &$/at and set V * n u = O in the momentum and energy equations (2.51) and
(2.52). Also in cases where carrier-cooling effects are insignificant, such as in MESFET,
it is convenient to neglect the kinetic energy term in equation (2.53) since the thermal
 Semiconductor device modelling 25 1

energy :kTe is typically an order of magnitude greater. Finally, since the term U * V u
is usually small compared with the other terms in the momentum conservation equation
(2.5 1) it is usually omitted. These approximations allow the conservation equations
to then be re-expressed as follows:
an
-+V. (nu)=O (2.57)
at

(2.58)

(2.59)

Note that equation (2.58) is equivalent to the form

where

(2.61)

The definition of electron velocity here includes the diffusion component and hence
must be distinguished from the drift velocity defined in equation (2.15). The mobility
and diffusion coefficients shown here are functions of average electron energy rather
than electric field (the concept of mobility and diffusion in non-stationary models is
open to some debate). Equation (2.57) has the same form as the classical current
continuity equation. The simplified form of the energy equation (2.59) yields the
average energy required to solve equations (2.57) and (2.58). The current density
equation which corresponds to the velocity equation (2.60) is
J = gnu. (2.62)
The above analysis assumes a single-electron gas model suitable for silicon devices.
In the case of compound semiconductors such as GaAs, with multi-valley band
structures, it is necessary to assume that the average electron temperature of the electron
gases in the upper and lower valleys are the same (Blotekjaer 1970, Curtice and Yun
1981).
Substitution of the velocity equation (2.60) into the current density expression
(2.62) yields an equation comparable with the controversial form containing the term
V ( D n ) . A greater insight into the meaning of this expression can be obtained by
re-expressing equation (2.58) in terms of electron temperature to yield

(2.63)

A comparison of this equation with equation (2.4) reveals that this expansion has an
additional term for the spatial variation in electron temperature, which expresses the
fact that the most energetic (hot) electrons will diffuse out of the collection. This
feature may be significant in devices where there is a high degree of local electron
heating such as in Schottky barriers and at the edges of planar contacts (Doades et a1
1984). Curtice and Yun (1981) have shown that if the temperature gradients are
negligible, then the term containing the electron temperature. gradient in equation
252 C M Snowden

(2.63) may be omitted, yielding an expression identical to equation (2.60). A con-

sequence of the electron temperature model is that the diffusion coefficient obtained
is isotropic rather than anisotropic as predicted by Monte Carlo techniques and
measurements. This anisotropic behaviour occurs because the energy is not completely
randomised.
Another simplification of the momentum and energy conservation equations (2.5 1)
and (2.52) frequently used in semiconductor device models is to assume that spatial
variations are small. This reduces equations (2.51) and (2.52) to the following com-
monly used forms (Maloney and Frey 1977, Carnez et a1 1980, Shur 1981a, Snowden
1984a):

(2.64)

(2.65)

These equations can only strictly be used to model devices where the spatial variation
in energy and velocity are small, and hence are not well suited to analysing Schottky
barrier or planar devices, where equations (2.58) and (2.59) are better approximations.
In models which assume that the electron energy is a function of electric field
alone, which is a reasonable approximation in devices with relatively uniform and
slowly varying fields, it is often assumed that spatial and temporal variations are
negligible. In these circumstances the momentum and energy conservation equations
reduce to the following extremely simple forms:

(2.66)

6-(0=qre(()E* U. (2.67)
It can be immediately deduced that equation (2.66) is equivalent to the steady-state
form U = PE, which is a velocity equation which neglects diffusion. Substitution of
equation (2.66) into (2.67) confirms that the energy 6 is solely a function of electric
field, since re and r p are functions of the electron energy.
Variations in the effective mass m* with electron energy become significant in GaAs
and InP because of intervalley transfer. One possible method of taking account of
this effect is to generate transport equations for electrons in both the upper and lower
valleys (Blotekjaer 1970). This would lead to a highly complex non-linear model and
as previously mentioned it is generally accepted that a single-electron gas model is
justified. Cook and Frey (1982) describe an equivalent single-valley model which
requires that the fraction of total free carriers in the upper valley and the ratio T ~ m*
/
are functions of the localised average energy; the contribution of kinetic energy to the
electron temperature is negligible. This latter point implies that velocity overshoot is
attributable to the dependence of the momentum relaxation time T~ on energy. Bozler
and Alley (1980) equate the thermal energy to the energy separation between the lower
and upper valley ASuL (0.36 eV for GaAs), the proportion of electrons in the upper
valley Gu(5) and the total energy 5:
w e = 5 - G"(6) A6"L. (2.68)
This relationship can be used to modify the momentum conservation equation (2.58)
 Semiconductor device modelling 253

to obtain a model for multi-valley semiconductors, where the velocity is given by

Considerable difficulty is experienced in solving even the simplified energy equation

(2.59) because of the highly non-linear nature of this boundary value problem. The
problem is reduced in complexity by identifying regions of hot electron transport in
the device and regions where the carriers d o not gain energy from the field (Cook and
Frey 1982). Cook and Frey (1982) consider the case of a MESFET and assume that in
regions where carrier heating is significant both the electric field and drift velocity are
nearly one-dimensional, allowing the energy conservation equation to be approximated
(for silicon) as

(2.70)

By assuming that diffusion effects are small compared to the drift velocity in the regions
of hot electron transport the last term in equation (2.63) can be omitted to yield a
one-dimensional transport equation for Si:
1/2
(2.71)

or for GaAs:

(2.72)

where

The energy transport equations described above allow 5 to be determined from the
electric field and boundary conditions on 6. In this way the determination of 5 and
rp(6)correspond to the calculation of p ( E ) and D ( E ) in classical models. Future
non-stationary models may have to address the problem of intervalley carrier relaxation
which, to date, has not been included in the models.
The relaxation parameters rp and re can be obtained from information available
from homogeneous models. In the case of silicon, with constant effective mass m*,
the steady-state relationships of equations (2.66) and (2.67) for the homogeneous case
can be used:

(2.73)

(2.74)

A considerable amount of theoretical and experimental information is available on

the steady-state velocity-field characteristic ( v - E ) which allows the relaxation times
T ~6() and re(5) to be determined for the homogeneous case. The extraction of these
parameters for GaAs is more tedious because of the strong dependence of effective
254 C M Snowden

mass rn* on energy. Carnez et al (1980) utilise steady-state information on the

electric-field dependence of velocity, energy and effective mass for GaAs, shown in
figure 13, obtained from a three-valley Monte Carlo model. A similar approach can
be taken for electron temperature models which use theoretical field-dependent electron
temperature characteristics and temperature-dependent energy relaxation time to deter-
mine the carrier velocity and relaxation time, respectively (Doades et a1 1984, Snowden
1984a).

“6 , , ,

0 1 2

0 1 2

I ic)

005
’/ - ,

0 1 2
E(MV m-’)
Figure 13. Average steady-state drift velocity, energy and effective mass of electrons in GaAs as a function
of electric field E. The full curves indicate the characteristics for a doping density ND of loz3m-3 and the
broken curves correspond to a doping density ND of 3 x m-3 (from Carnez et al 1980).

Hot electron effects in short channel devices have been the subject of considerable
interest in recent years (Maloney and Frey 1977, Price 1978, Carnez et al 1980, Hess
1981, Higgins and Pattanayak 1982, Teitel and Wilkins 1983, Watanabe and Slamet
1983, Cook 1983). The effect of velocity overshoot is illustrated in figure 14, which
compares the classical and hot electron models for a short channel GaAs MESFET.
Simulations have revealed that velocity overshoot is significant in both Si and GaAs
MESFET where the gate length is less than 0.5 pm. The results show that velocity
overshoot is considerably greater in the GaAs FET than in the Si device, as expected.
The drain saturation current in a 0.25 pm gate length GaAs PET has been shown to be
up to three times higher than that predicted by the classical diffusion model (for
V,,=O) (Cook and Frey 1982). The saturation current of silicon transistors with
similar gate lengths is increased by approximately 20% as a result of velocity overshoot
in the channel. Carnez et al (1980) have also demonstrated the increase in drain to
source current due to non-stationary electron dynamics and have investigated the
 Semiconductor device modelling 255

influence of energy relaxation effects on transconductance g , and gain-bandwidth

product. They conclude that the transconductance and maximum operating frequency
increase substantially for the hot electron model, compared with results obtained using
classical approach.

" I o
mE 0 02 04 06 08 10
-
0
r- W
W
x Gate m
%
-
0
W

0 05 10 15
Distance (pm)
Figure 14. Average velocity and energy as functions of distance along the channel for 0.25 pm gate length
Si ( a ) and GaAs ( b ) MESFET, using both energy transport and field-dependent transport models (from
CookandFrey1982). ( a ) LG=0.25pm, VD=2.0V, VG=OV. ( b ) L G = 0 . 2 5 p m , V,,=l.O, V,=OV.

Semiclassical models have been recently applied to the simulation of heterostructure

devices. Asbeck et a1 (1982), for example, have used a modified form of the one-
dimensional bipolar simulation SEDAN to model GaAs/GaAlAs heterojunction bipolar
transistors (HBT),predicting cut-off frequencies of the order of 100 GHz. Heterojunc-
tion device models have also been based on the formulation of Lundstrom and Schuelke
(1982). An analytical model has been presented by Delagebeaudeuf and Linh (1982).
A model similar to that used to represent heavily doped semiconductors, incorporating
Fermi-Dirac statistics, has been described by Lundstrom and Schuelke (1983). This
latter finite-difference model has been applied to HBT and modulation-doped FET to
investigate the influence of the conduction band spike in HBT and the variation of the
two-dimensional gas with gate bias in the FET.
The concept of ballistic transport in very small semiconductor devices has been a
topic of considerable debate over the past five years. Ballistic transport was originally
considered by Shur and Eastman (1979) in connection with the low-temperature (77 K)
operation of low-power, high-speed devices for logic applications. The term 'ballistic
transport' refers to the collision-free motion of electrons, which occurs over short
timescales. Ballistic transport can occur in devices where the phonon mean free path
at the operating temperature is longer than the critical device dimension, the voltage
is sufficiently high so that the diffusion term can be neglected and the voltage is low
enough to minimise interband intervalley transition. During ballistic transport in a
256 C M Snowden

uniform electric field, the electrons move with a velocity given by

9Et
v=-. (2.75)
m*
The timescale over which ballistic transport can occur is determined by comparing the
average energy acquired by the ballistic electron with the energy at which scattering
mechanisms become significant (Teitel and Wilkins 1983). In practice, ballistic trans-
port occurs over such short intervals that it does not lead, in itself, to large velocities.
Over longer periods collision mechanisms limit the velocity to a steady-state drift value.
Shur (1980, 1981b) and Hess (1981) discuss the significance of ‘near-ballistic’ transport
which describes transport in the presence of only a few collisions. Near-ballistic
transport allows large velocities to be achieved as a result of velocity overshoot.
However, ballistic transport (or near-ballistic) cannot, by itself, lead to overshoot
unless the momentum relaxation rate is larger than the energy relaxation rate ( 7p<<7,)
(and is an increasing function of energy), or the semiconductor has satellite valleys
with heavier effective mass at low enough energies to have a significant electron
population in the steady state.
Hess (1981) has shown that for short ( ~ 0 . 3pm) GaAs devices ‘near-ballistic’
transport occurs only at relatively low voltages ( ~ 0 . 0 3 V)
6 and that the transport is
‘near-diffusion-limited’ at higher voltages. The transport properties may be enhanced
by using materials with a higher optical phonon energy and by the screening electron-
phonon interaction. The electron-phonon interaction is weaker at high electron
densities ( m-3) and close to metal contacts because of remote screening effects.
Special device structures with appropriate electric-field distributions can be used to
enhance velocity overshoot effects. A permeable base transistor described by Bozler
and Alley ( 1980) provides suitable conditions for ballistic transport.

3. Monte Carlo particle modelling of semiconductor devices

The Monte Carlo method is a statistical numerical technique for solving mathematical
and physical problems. It was in use long before its application to semiconductor
transport problem (Meyer 1956). The late 1960s saw significant developments in the
application of the Monte Carlo technique to semiconductor device modelling (Rees
1968, 1969, Fawcett et a1 1969, Fawcett and Rees 1969a, b, Ruch and Fawcett 1970).
Since that time, Monte Carlo simulation techniques have grown in popularity and
sophistication and promise to extend the understanding of quantum transport models
(Barker and Ferry 1979).
Monte Carlo simulation techniques provide a means of solviiig the Boltzmann
transport equation by modelling the dynamics of the individual charge carriers. The
method provides a detailed physical interpretation of the operation of semiconductor
devices with arbitrary field distribution. This approach has provided a wealth of
information on the velocity-field (mobility), diffusion, energy, valley population,
scattering mechanisms and effective mass characteristics of carriers in both polar and
non-polar semiconductors. Monte Carlo simulations have also been applied to a wide
range of semiconductor devices to investigate the transport phenomena and determine
terminal characteristics. Both the classical and semiclassical approaches described
previously take advantage of material characteristics obtained from Monte Carlo
simulations.
 Semiconductor device modelling 257

3.1. The Monte Carlo method

The Monte Carlo method of modelling semiconductors is based on simulating the
transport history of one or more carriers, subject to the action of physical stimuli on
the physical system. This type of model represents a particulate view of the semiconduc-
tor physics and are described in detail in reviews by Alberigi-Quaranta et a1 (1971),
Price (1979) and Jacoboni and Reggiani (1979, 1983). The essence of the Monte Carlo
method lies in the generation of sequences of random numbers with specified distribu-
tion probabilities used to describe the microscopic processes, such as scattering events
which determine the time between successive collisions of individual carriers.
Monte Carlo simulations have been used to model steady-state phenomena by
analysing the motion of a single carrier. This assumes that, if the particle history is
followed for a sufficiently long period of time, it will yield information on the behaviour
of the entire electron gas. Because only one particle is involved it means that a relatively
long history can be accumulated in reasonable amounts of computer time. If non-
stationary transport conditions prevail or the device is highly non-homogeneous, it is
necessary to simulate the history of large numbers of carriers (Lebwohl and Price 197 la).
The following description of the Monte Carlo technique applied to semiconductors
examines the case for electrons only in order to simplify the analysis. The potential
distribution is solved using a conventional Poisson solver, utilising a finite-difference
mesh. Consider the case of a single carrier with suitable initial conditions and a
wavevector ko moving in an electric field E. The equations of motion governing the
electron during free flight are
dk qE
- h = h/2n (3.1)
dt h

U=-
dr
(3.3)
dt
where k is the electron wavevector, q is the charge on the electron (negative), h is
Planck’s constant, r is the position and v is the velocity. The generalised electron
energy for cubic semiconductors defined from the bottom of the conduction band is
taken as (Moglestue 1982)

where k,, k, and k, are the k vectors and m:, m: and mT are the effective masses
along the three Cartesian axes. In GaAs and InP the three effective masses are all
equal ( m * ) . CY is a non-parabolicity factor and for the case of GaAs is zero except for
the conduction band. The velocity, energy and position are determined prior to
scattering. At the end of the free flight, the electron is scattered by a scattering mechanism
chosen according to the relative probability of all possible scattering mechanisms.
Scattering mechanisms which operate on the electron include optical and acoustic
phonon scattering, ionised impurity scattering and neutral ionised scattering. The latter
effect is only significant at very low temperatures and is usually neglected. Since optical
phonon dispersion is reasonably flat it is usual to consider only one phonon energy.
Acoustic phonon scattering is often regarded as elastic because the acoustic phonon
258 C M Snowden

energy is small relative to the electron energy. Impurity scattering is assumed to be

elastic and has a strong dependence on wavevector k (Brooks 1955, Moglestue 1982).
Scattering mechanisms are usually assumed to be an instantaneous process although
Barker and Ferry (1980) have demonstrated that for very small devices (<0.1 pm) the
interaction times must be accounted for.
The total scattering rate hi for one process is described by

hi =
I B i ( k ,k ’ ) dk’ (3.5)

= [ dcp [ d 8 cos 8 [ dkk2Bi(k,k’, cos 8 )

J J J
(3.6)

where k and k’ represent the wavevector before and after scattering, respectively. Bi
(k, k‘) is the differential cross section of the scattering mechanism, 8 is the angle
between k and k‘ and cp is the azimuth angle. The sum of all the scattering rates is
given as
r= hi. (3.7)
i=l

The scattering mechanism a specific electron of energy & undergoes is determined

using a random number Y in the range 0 to r. The scattering angles cp and 8 are chosen
using random numbers with a uniform density in the interval 0 to 27~.
The probability that the particle will fly freely for a time t since the last collision is

g ( t ) d t = T [ k ( t ) ]exp ( -IorT [ k ( t ’ ) ]dt’) d t

where g( t ) is the probability density of scattering at time t. Rees (1968,1969) introduced
a notional ‘self-scattering’ concept to overcome difficulties in solving equation (3.8)
which would otherwise require the solution of an integral equation for each scattering
event (Kurosawa 1966). Using this apprpximation, the probability density of scattering
reduces to
d t ) =To exp (-rot) (3.9)
where
n
lT0= A(k, t ) (3.10)
i=O

and ho is the self-scattering rate. A random number r with a uniform density is used
to determine the time of free flight tf:
1
t f = --ln (r). (3.1 1)
TO
The computational time required to obtain tf can be minimised by ensuring that
self-scattering is reduced to a minimum. The parameter To is chosen so that it is not
less than the sum of all the scattering rates r in the region of k space of interest.
After a free flight of duration t the position r and momentum k are determined from

r = ro+-+-
m”
‘ jar jar'
m”
dt‘ dt” E ( t ” ) (3.12)

k = ko+E jotE (dt’ t’) (3.13)

 Semiconductor device modelling 259

where ro and ko are the position and wavevector of a carrier just associated with a
scattering event. The electric field E is assumed to be constant within each mesh cell,
but may change from cell to cell. If it is assumed that during a field-adjusted time
step the carriers continue to accelerate in the field they experienced before they crossed
the boundary to the next potential cell, then
r=r0+,+--4_m:
hk,t
(3.14)
m 2m"
qEt,
k=ko+---. (3.15)
h
This assumption saves on computation time and requires that the particle continues
to accelerate until the occurrence of the next scattering event or the end of the simulation
time step. The simulation time step A t between successive solutions of the Poisson
equation, and subsequent re-evaluation of the electric field, is often chosen using the
expression
min (Ax, Ay)
At = (3.16)
V
where V is the average velocity of the carriers. This relationship ensures that only a
few particles cross the potential cell boundaries during the time step. Stochastic heating
effects (Hockney 197 1) and numerical stability requirements further restrict the size
of the time step At.
The potential I) and electric-field distribution E are determined using Poisson's
equation (2.1) which is solved for the domain of the problem using one of the techniques
described earlier (direct, iterative or FFT (Hockney's POT4)). The device is divided
into a mesh with node spacings Ax and Ay (for a two-dimensional model). The charge
due to ionised dopants and charge carriers is assigned to each cell using a suitable
technique (e.g. Birdsall and Fuss 1969, Eastwood and Hockney 1974). The electric
field E is then obtained using a finite-diff erence approximation. After calculating the
electric field, it is frozen until the next time step. The time step A t is chosen to be as
long as possible, but short enough to minimise errors due to the discretised nature of
the electric field ( A t usually lies in the range 5-50 fs).
The total number of charge carriers in a semiconductor sample of volume Vo with
a uniform density of ionised impurities No is NoVo-a figure which can be of the order
of lo7 carriers in the active channel of a modern GaAs MESFET. This number of
charge carriers is far too large to be analysed on a computer and it is necessary to
restrict simulations to less than 50 000 carriers. In order to satisfy the Poisson equation
it is necessary to assign a charge qp to each particle where
ncqp= q Ax Ay Az No (3.17)
where Az is the dimension of the cell perpendicular to the xy plane and n, is the
number of modelled particles per cell. These particles of charge qp are referred to as
super particles.
When particles reach the boundary of the device they are reflected or absorbed,
depending on the nature of the boundary conditions. Apart from the requirement for
suitable boundary conditions, the Monte Carlo model requires a set of initial conditions.
Unless the simulation time is very long the initial conditions of the electron matrix
will influence the outcome of the simulation. This is not normally a problem when
260 C M Snowden

considering steady-state situations, but can strongly affect the outcome of a transient
phenomenon or non-stationary transport problem. For these types of study it is essential
to take account of the distribution of initial carrier states.
If a time-averaged value of a quantity a [ k ( t ) ]is required from the simulation it
may be obtained over the period T using the expression

(3.18)

separation into the sum of integrals over all flights of duration ti. Using this expression
it is possible to extract time-averaged values of drift velocity, average carrier energy
and carrier density. The electron distribution function which is the solution to the
Boltzmann transport equation can also be extracted in a similar way (Fawcett et al
1970). If a k space mesh is used in the model, the time spent by each carrier in each
cell of the k space mesh can be evaluated over a long period T. By normalising this
time the required electron distribution function can be obtained. Price (1970) has
described an alternative method for obtaining a time average in steady-state simulations
using a synchronous-ensemble technique. The average value of a quantity a is defined
as the ensemble average at time t over the N particles of the system using the
relationship:
1
(a)=-Zai(ti=t). (3.19)
N i
An example of the steady-state evaluation of average momentum and kinetic energy
is given by Aas and Blotekjaer (1974) where the averages are calculated for approxi-
mately 100 collisions. The calculation is repeated for several such time intervals and
is terminated when the deviation of average momentum and kinetic energy are both
less than one or two per cent, or if 40 000 collisions have been exceeded.
Monte Carlo simulations are particularly well suited to modelling time- and space-
dependent phenomena. This is especially important for small devices where it is
necessary to consider both the transient response to external voltage stimuli and the
effects of non-homogeneous field and material properties. This type of problem requires
the simulation of an ensemble of particles (Lebwohl and Price 1971a) unless the
physical phenomena are periodic in space or time (Zimmerman et a1 1978).
The transient response of a homogeneous electron gas to a step change in applied
electric field is typical of the type of problem studied using Monte Carlo programs.
This situation requires many particles to be independently modelled with specified
distributions and initial conditions. The number of simulated particles must be
sufficiently large so that the ensemble average value of the quantity of interest obtained
from equation (3.19) as a function of time will be representative of the average for
the entire gas. The steady-state average of equation (3.18) cannot be used in connection
with time-dependent phenomena. The transient response obtained from the simulation
depends on the initial conditions of the carriers which should be related to the problem.
A Maxwellian distribution at the lattice temperature and no drift velocity, or a
Maxwellian distribution with an electron temperature higher than that of the lattice,
are suitable initial conditions for this type of model.
Space-dependent phenomena in semiconductor devices represent another important
class of problem which is appropriate to Monte Carlo modelling. Here the carrier
transport properties depend upon the spatial distribution of the carriers in the device.
 Semiconductor device modelling 26 1

This class of problem also requires an ensemble of independently simulated particles

to be used with averages taken over groups of particles at specific positions.
A particular important example of a space- and time-dependent phenomenon is
the case of diffusion. Diffusion effects have been rigorously investigated using Monte
Carlo simulations (Butcher et a1 1967, Fawcett and Rees 1969a, b, Ohmi and Hasuo
1970, Bauhahn et a1 1973, Abe et a1 1974). Low field values of the diffusion coefficient
D can be obtained using the Einstein relationship of equation (2.18) by substituting
information on the ohmic (linear) mobility. However, in high field conditions the
Einstein relationship is no longer valid and it is necessary to consider hot electron
diffusion coefficients. Intervalley diffusion effects present in compound semiconductors
also lead to departures from Einstein diffusion even at relatively low field strengths.
Fick’s law, which is often used to model classical diffusion processes, cannot be used
to describe transient phenomenon.
The second central moment of the carrier distribution function can be used to
determine the diffusion coefficient using a Monte Carlo simulation, where

I d
D, = - - (( z - (z))’) (3.20)
2 dt

and D, is the longitudinal component of the diffusion coefficient D along the direction
of the drift velocity. This expression assumes that the distribution n is a function of
only one coordinate, z, parallel to the direction of drift. A number of particles are
independently simulated with their positions recorded at specific times. After long
simulation times the second central moment follows a linearity predicted by equation
(3.20). D, is then obtained from the slope of the characteristic. The diffusion coefficient
can also be calculated from the variance of the distance z travelled by the particle in
time t:

(3.21)

This method has been used by several workers (Ferry 1980, Fauquembergue et a1
1980), but since the solution obtained from this equation only approaches that obtained
from equation (3.20) as t+co, it is necessary to use very long simulation times to
extract D, from equation (3.21). Other components of the diffusion (transverse 0,)
can be obtained using an expression analogous to equation (3.20):

(3.22)

Typical diffusion results obtained from Monte Carlo simulations are shown in figure
15. The anisotropic characteristics of the diffusion coefficients of GaAs and InP are
those obtained by Fawcett and Rees (1969a, b) and Bauhahn et a1 (1973).
Apart from the restrictions on Fick’s diffusion law due to transient behaviour, it is
also important to appreciate that this classical model cannot be applied if the time
during which the concentration gradient varies is short compared with energy relaxation
times or if the distance at which the gradient varies appreciably is large compared
with the electron mean free path. Jacoboni (1974), Jacoboni et al (1981) and others
have suggested more generalised diffusion theories which include the effects of diffusion
at small distances and short times.
262 C M Snowden

? 001-
0

0 0.8 16 2 4
Electric f i e l d (MV m-ll

Figure 15. Anisotropic diffusion characteristics for GaAs and InP obtained from Monte Carlo simulations
at 300 K. -.-, InP, Bauhahn et al (1973); 0 , GaAs, Dll,Fawcett and Rees (1969a, b); 0, GaAs, D _ ,
Fawcett and Rees (1969a, b); -, GaAs, Bauhahn et a/ (1973).

The drift velocity in semiconductor devices can be readily determined for both
steady-state and transient situations using Monte Carlo simulations. This area has
been the subject of substantial modelling effort (Boardman et al 1968, Fawcett et a1
1970, Ruch and Fawcett 1970, Maloney and Frey 1977, Littlejohn et a1 1977). A
steady-state simulation for GaAs starts off with the electrons in the lower band with
a Maxwellian velocity distribution at the lattice temperature. The steady-state velocity-
field characteristics obtained from various published results are shown in figure 16.
Transient results obtained by Ruch (1972) and Maloney and Frey (1977) who used a
modified constant field technique (Fawcett er a1 1970) are shown in figure 12 for Si,
GaAs and InP.

0 04 08 12
E l e c t r i c f i e l d I M V m-’)

Figure 16. Steady-state velocity-field characteristics for GaAs and InP obtained from Monte Carlo simula-
tions at 300 K. InP: 0 , Maloney and Frey (1977). GaAs: 0, Ruch and Fawcett (1970); A, Littlejohn et al
(1977).
 Semiconductor device modelling 263

An important feature of Monte Carlo simulations is that they allow the analysis
of the noise spectrum of velocity fluctuations in the semiconductor (Nougier 1980).
This can be achieved by Fourier analysing the velocity autocorrection function C , ( t )
(Fauquembergue et a1 1980, Ferry and Barker 1981):
C , ( t ) = ( S v , ( t ’ ) S v j ( t’+ t ) ) . (3.23)
The noise spectrum S,( w ) is obtained using the Wiener-Khintchine theorem to realise
the Fourier transform of c( t ) :

(3.24)

In fact, the noise spectrum can be directly related to the frequency-dependent diffusion
coefficient D(o) using the relationship
S , ( w ) = 2D(w) (3.25)
which follows from the dependence of autocorrelation function C ( t ) on diffusion (Reif
1965). At low frequencies the noise exhibits a white spectrum.
The equivalent white noise temperature associated with a semiconductor device
which exhibits a positive value of the real small-signal impedance can be extracted
using Monte Carlo methods (see Jacoboni and Reggiani 1983). The field-dependent
drift velocity v d ( E )and diffusion coefficient D ( E ) are obtained from the Monte Carlo
model and inserted into the expression for noise temperature T,. For the homogeneous
case the lengthy expression for T, (see Jacoboni and Reggiani 1983) reduces to a form
similar to the Einstein relationship:

(3.26)

where p d is the differential mobility.

Monte Carlo simulations are frequently used to investigate impact ionisation
phenomena in semiconductor devices. Impact ionisation is modelled by introducing
the probability of impact ionisation as an independent scattering mechanism in addition
to phonon and impurity scattering mechanisms. Lebwohl and Price (197 1b) have used
an ensemble Monte Carlo method to model impact ionisation, where the electron-hole
pair generation rate per particle per unit time g , is obtained by counting ionisation
events. The impact ionisation rate a I is then obtained using

aI = 811 vd* (3.27)

Single-particle Monte Carlo simulations have been used by other workers, where the
impact ionisation rate a I is obtained by averaging the distance to impact ionisation
over a large number of ionisations (Curby and Ferry 1973, Schichijo and Hess 1981).
A further area of interest applicable to Monte Carlo simulations is in the modelling
of semiconductors subject to a combination of electric and magnetic fields. In this
situation the equation of motion for a carrier in free flight is modified to

_dk-%+s
-_ U xB (3.28)
dt h ch
where B is the magnetic field, U is the group velocity of the particles and c is the
velocity of light in a vacuum. This type of model requires a full three-dimensional
264 C M Snowden

simulation. Boardman et a1 (197 1) has investigated this type Df problem in detail. The
Hall mobility p H ,drift mobility p and longitudinal mobility p E are all readily obtained
using this type of simulation.

3.2. Application of Monte Carlo techniques to device modelling

Monte Carlo simulations provide a powerful method for modelling the operation of
very small semiconductor devices where the carrier transport mechanism is governed
by non-stationary processes and classical models fail. Although Monte Carlo models
require extensive computer facilities which restrict the present capabilities of this
technique, they have been used to study a variety of semiconductor device problems.
Devices which have been investigated using Monte Carlo simulations include TED
(Lebwohl and Price 1971a, Hockney et a1 1974), MOSFET (Park et a1 1983), MESFET
(Hockney et a1 1974, Warriner 1977, Rees et a1 1977, Moglestue 1979, Pone et al 1982),
bipolar transistors (Baccarani et a1 1977), heterostructures (Glisson et a1 1980), Schottky
diodes (Beard and Rees 1981) and photodetectors (Moglestue 1984).
MESFET are perhaps the most extensively researched device using Monte Carlo
methods. The steady-state carrier distribution for single- and dual-gate GaAs and Si
Source Gate Drain

(0)

Source Gate 1 Gate 2 Drain

if)

Figure 17. Distributions of electrons in single- and dual-gate MESFET obtained from Monte Carlo simulations.
( a ) Single-gate MESFET. Distribution of central-valley electrons. Electrodes and gaps are 1 pm across. The
epilayer is 0.23 w1.m thick and the substrate is 0.385 pm thick. The doping concentration is 10’’ II-~. The
gate potential is -0.1 V and the drain potential 3 V. The time is 9.5 ps after switching on. (6) Distribution
of upper-valley electrons. (c) Distribution of all electrons in the microscopic model. ( d ) Distribution of
all electrons in the diffusive-particle model. ( e ) Distribution of all electrons in a Si MESFET. Drain potential
is 2 V. (f)Dual-gate MESFE?. Distribution of all electrons. (g) Distribution of r electrons. ( h ) Distribution
of upper-valley electrons. ( a ) - ( e ) from Hockney et a1 (1974), (f)-(h) from Moglestue (1982).
 Semiconductor device modelling 265

FET are shown in figure 17, which clearly reveals the depletion regions beneath each
Schottky barrier gate and the non-abrupt nature of the depletion boundaries. Moglestue
(1979) has found that the simulated transconductances for this device agree well with
measured values corresponding to the real device. The carrier distributions in figure
17 clearly show the penetration of carriers into the substrate decreasing the transcon-
ductance g, and increasing the output conductance. Hockney er a1 (1974) and
Moglestue (1979) have also investigated the distribution of electrons in the central and
upper valleys for GaAs MESFET (figure 17). This has shown that the current in the
channel below the gate of P 1 pm gate length FET is entirely due to central valley
electrons, whereas in the gap between gate and drain the majority of carriers are
low-mobility upper-valley electrons. The work of Hockney et al (1974) demonstrated
that the commonly used diffusion approximation is inadequate for modelling the
channel region in GaAs MESFET because the assumption of an instantaneous mobility
relationship leads to the stationary accumulation rcgion appearing under the drain
edge of the gate. The use of a microscopic model, which takes account of the finite
time and distance required for electrons to move into the upper valleys, correctly places
the electron accumulation midway between the gate and drain.
Dynamic Monte Carlo simulations have also been performed on FET and diode
models. However, because of the excessive amounts of computer time required for a
true AC analysis over several cycles, dynamic simulations are usually restricted to
transient cases which make use of Fourier transform techniques to extract small-signal
AC parameters (Moglestue 1982). Despite the limitations in computer power, it is
possible to conduct complete AC analyses for small simple device structures operating
over several cycles at millimetric frequencies (=lo0 GHz). An example of this type
of simulation is given in the Schottky diode model of Beard and Rees (1981).

4. Quantum transport theory

The preceding sections in this review have dealt with semiconductor device models
based almost entirely on the Boltzmann transport equation. Quantum transport theory
has a completely independent origin (Kubo 1957, Kohn and Luttinger 1957, 1958,
Chester 1963, Luttinger 1968) even though it often reduces to a Boltzmann transport
model. It is both conceptually and mathematically more complex than any of the
Boltzmann type of model. It is important to distinguish between quantum transport
theory which completely describes a quantum view of a system and the application of
quantum mechanics to the semiclassical picture, where quantum mechanics are used
to provide information on bulk and interface properties (e.g. Price and Stern 1983,
Capassco 1983).
The fundamental concept of classical Boltzmann transport theory is the assumption
that a unique carrier distribution function dependent on the position and momentum
of carriers can he obtained and used to determine the macroscopic properties of the
device (current flow, etc). Quantum transport theory does not allow the simultaneous
specification of position and momentum and requires that the Heisenberg uncertainty
principle be obeyed. It provides a quantum-mechanical description of the system in
terms of a density matrix and examines the time evolution of the system. The concept
of the carrier distribution function is still meaningful for large devices (with critical
dimensions greater than 1 pm), since the spatial variations in the distribution function
occur over distances much larger than the de Brsglie wavelength associated with the
266 C M Snowden

carriers. In devices where the critical dimension is of the order of 0.25 p m the variations
in the distribution function occur on a timescale comparable with the collision times,
but still short compared with transit times. Hence, the distribution function is still
useful in describing the behaviour of carriers, although carrier transport does not
strictly follow the Boltzmann transport theory. The concept of a classical distribution
function is not valid for very small devices where the active channel is less than 250 A
long since the collision time is of the same order as the transit time. It is useful,
however, to attempt to establish an equivalence to the distribution function.
Wigner (1932) has shown that quantum transport theory has an analogy to the
classical concept of a phase-space distribution function by using a probability density
matrix. Quantum transport theory is used to explain, support and set confidence limits
for the Boltzmann transport theory. It is also used to understand and develop genuine
quantum transport phenomena. There are many features of carrier transport which
are neglected by the Boltzmann theory and the review by Grubin et a1 (1982) cites the
following causes as major failings-non-locality of scattering processes, strong driving
forces, strong scattering, dense systems, small systems and the non-classical influence
of driving fields. Quantum transport theory accounts precisely for these shortcomings
in the Boltzmann approach.

4.I. Quantum transport theory

The generalised quantum transport theory is based on the Liouville-von Neumann
equation for the statistical density matrix p ( t ) , which provides a complete description
of the system (Klinger 1979):

where H is the Hamiltonian and F is the coupling to external driving forces. The
usual boundary conditions for t < 0 is
P =Po(H) (4.4)
where po is a thermal equilibrium solution. The driving force is initiated at t = 0. The
Hamiltonian H F is usually partitioned into free carrier He, free scattering H,, carrier-
scatterer interaction V and driving free components F. Hence equation (4.3) expands
to yield
H F = He+ H,+ V + E (4.5)
Phenomenological effects such as current density J, can be obtained by evaluating
quantum-statistical expectation values determined from the density matrix p ( t ) . For
example
( J , ( t ) )= Tr [J,P(t)l (4.6)
and for a finite trace
 Semiconductor device modelling 267

where { I A ) } is any complete set of states. The states l h ) are chosen to diagonalise the
term ( H e + H,). The choice of { I A ) } determines the characteristics and interpretation
of the subsequent transport theory. Normally { / A ) } is chosen so that

IA) = 1 4 s ) (4.9)
so that, using Dirac's notation, the states {le)} and {Is)} diagonalise He and H,,
respectively.
If it is assumed that the current density J depends only on electronic variables and
is commutative with the component He- which requires homogeneous transport in
the absence of magnetic fields then
(4.10)

f(e ) is a time-dependent generalised electron distribution function over the free-carrier

states {le)} of the form:
(4.1 1)

(4.12)
The Liouville equation can be used in conjunction with f ( e ) to produce a transport
equation which may have a form similar to the Boltzmann equation.
In the case of inhomogeneous transport and for transport in quantising magnetic
fields, J is not diagonal and it is necessary to consider off-diagonal elements of the
electron density matrix f = ( P ) ~ .Although it is possible to express f ( e ) in terms of
f ( e , e') the resulting transport theory does not resemble the Boltzmann equation. A
generalised closed-form equation of motion for f(e ) can only be obtained for the case
of independent carrier transport in a stationary scattering system.
A generalised solution for the Liouville-von Neumann equation is not available,
but as only limited aspects of the full solution are of interest, it is possible to consider
special cases such as single-polaron (electron or hole) transport properties. An alterna-
tive method of interpreting quantum transport theory was proposed by Wigner (1932)
and has been extensively adopted (see, for example, Tatarskii 1983, Balazs and Jennings
1984). Wigner's distribution function has no simple interpretation in the sense of
probability theory but can be used to directly calculate expectation values of observable
quantities (current density, etc) in a manner analogous to classical transport theory.
The carrier density n(r, t ) and current density J ( r , t ) , for example, are given by the
expressions

(4.13)

(4.14)

where fn( k, r, t ) is the Wigner one-electron distribution function given by

fn(k, r, t ) ' l d3y exp(-ik . y ) Tr [p(t)+:(F-&)+Ar+$~)] (4.15)

(4.16)
268 C M Snowden

where u( k ) is the group velocity of the electron momentum state 1 k). ( :

)
I r ) and I / J ~r()
are the second quantised creation and annihilation operators, respectively, for a carrier
of spin U and position r (Hedin and Lundquiest 1969). In the case of independent
carriers fv reduces to the form ( r - $ u ~ ( p ) s ~ r + $ where
v) p is a fractional of the
one-electron Hamiltonian (Barker 1980a, b). The probability density matrix obtained
from Wigner distributions is not necessarily positive-definite and this complicates the
interpretation of these distributions.
The electron density matrix f(t ) is obtained from the full density matrix using the
condition
S(t) =Trs [ P ( t ) l . (4.17)
For stationary phonon and impurity distribution the initial thermal-equilibrium density
matrix may be approximately expressed as
Pt=ozfo(He+ V)Rs(Hs). (4.18)
Here fo is of the Maxwellian equilibrium form and R, describes the equilibrium
distribution of scatterers. By re-expressing the Liouville-von Neumann equation (3.26)
it is possible to obtain an equation of the form (Barker 1980b):

+'a t i[ He,f 1 + i[ F, J'] = i d7 e'( ~)f(t - + M F( t )

7) , (4.19)

The terms on the left-hand side of this equation describe collision-free diffusion and
acceleration of carriers, whilst the right-hand side is proportional to the scattering
interaction and describes collision effects by means of the memory term MF(f). The
memory term M,( t ) results from the interaction between the electric field and scattering
processes and is a function of the initial equilibrium state f , ( H e + V). In the
homogeneous case f is a function of momentum only, and if

He=- PL (4.20)
2m"
then the term [ H e , f ]vanishes. In the context of this model (Barker 1980b), the coupling
to the electric field E, obtained from the solution of Poisson's equation, is
F=-qE.r (4.21)
and in the momentum representation i[F,f] reduces to eEaf/ak. For the case of
inhomogeneous transport the term [ H , , f ] gives rise to a component u(k)af/ar.
Barker (1973) has shown that equation (4.19) can be reduced to the Boltzmann
transport equation for conditions of weak, infrequent scattering, translational invari-
ance of the scattering system and point collision occurring on an asymptotic timescale
t >> rc.
A major deviation of the quantum transport theory from the Boltzmann equation
is due to the intra-collisional field effect (ICFE). This describes the influence of the
electric field within a collision event which causes the total energy-conserving 6 function
to broaden and modifies the threshold energy of the emission of an optical phonon
(Barker 1980b). Ferry (1980) has discussed the retardation effect of the ICFE on
collisional interactions. The collisional relaxation rate is retarded, allowing a faster
rise in the transient velocity response. The retardation in the collisional interaction
also causes a faster rise in the electron temperature which results in a more rapid
settling of the response to the steady-state value. These effects have a significant
 Semiconductor device modelling 269

influence on the operation of sub-micron semiconductor devices, increasing the

frequency response and decreasing the distance over which the velocity transient occurs.
When quantum transport theory is used to investigate transport in semiconductors
at very high frequencies it becomes apparent that the collision kernel is frequency-
dependent and that the collision response is largely controlled by the memory term
M F (t ) (Barker 1980a, b). The Boltzmann transport theory cannot account for these
effects and is therefore not valid at high frequencies, even in homogeneous cases.
The generalised quantum transport theory which describes many-carrier, high-field,
inhomogeneous transport in the completely coupled hot-carrier-hot-phonon impurity
system is still a relatively new area of research. Many-body quantum transport theory
is conventionally analysed in the context of the Heisenberg uncertainty principle. The
Wigner density matrices are evaluated as averages of the appropriate annihilation and
creation operators over the initial equilibrium density matrix p ( H ) . The dynamic
evolution of the full system can be obtained from the time dependence of the field
operators by coupling to the self-consistant local electric potential $ ( r , t ) . In many-
body quantum transport theory collision rates are field-dependent retarded functions
of the Wigner electron distribution function and Wigner phonon distribution. The
many-body theory includes Pauli degeneracy factors and generates coupled hot-carrier
and hot-phonon transport equations. It also introduces explicitly and implicitly field-
dependent, self-consistent dynamic screening of the scattering processes which makes
the electron kinetic equation non-linear in J:

4.2. Application of quantum transport theory to semiconductor modelling

Quantum transport theory is difficult to apply directly to device modelling because of
the inherent complexity of this approach. However, it does provide an important
means of characterising the operation of very small semiconductor devices and inves-
tigating the influence of device environment and joint interaction of devices (synergetic
effects). Ferry and Barker (1980) have postulated that the quantum-mechanical
operation of an array of very small devices may exhibit synergetic effects when the
feature size is less than 0.1 pm. Although this aspect of transport theory is still in a
relatively embryonic state, some progress has been made towards the analysis of
synergetic efforts in VLSI structures (Grubin er al 1982). Quantum transport theory
has been used to analyse the operation of very small-scale devices. Monte Carlo
methods may ease the application of quantum transport theory to semiconductor device
problems (Ferry and Barker 1979). Many-body theory has been used to confirm that
the gallium arsenide non-equilibrium intervalley (r- X ) electron-phonon scattering
rate de-screens in fields exceeding 1 kV m-’ (Barker 1980b).
Semiconductor devices which have been investigated using quantum transport
theory include short-channel MOSFET and p-channel MESFET (Barker 1980a). The
effects of high field transport in light-mass semiconductors have been investigated
using quantum-mechanical models (Barker 1978). Quantum transport theory has been
used to demonstrate that the minimum practical limit for channel lengths in conven-
tional MOSFET is of the order of 0.24 pm (Barker 1980a). A novel p-channel MESFET
structure has been suggested by Ferry and Barker and described by Barker (1980a).
This very small device, shown in figure 18, consists of a thin (200A) highly doped
(6 x m-3) p-type Si active channel with a Schottky barrier gate. Ohmic contacts
are not used for the source and drain because quantum-mechanical effects would
otherwise lead to the channel becoming fully depleted. Current flow through the
270 C M Snowden

contacts is dominated by tunnelling. The predicted rapid switching characteristic and

low speed-power product (=lo-" J) make this device an attractive proposition for
future VLSI circuits.

Pt/Pt,SI
Source Drain /

Figure 18. A model for a very small silicon MESFET device (Barker 1980a).

5. Conclusions

Semiconductor device modelling has developed into a major area of interest to both
device physicists and electronic engineers, providing a valuable tool for device design
and playing a crucial role in modern solid-state technology. A particularly important
aspect of modelling is the ability to accurately predict the performance of new devices
prior to fabrication.
Classical transport models based on approximations obtained from Boltzmann
transport equations will continue to provide a relatively simple means of simulating
the bulk properties of semiconductor devices where the critical dimensions (active
channel, gate length) are greater than 10 pm. Recent studies, however, have shown
that the modelling of transport in smaller devices requires more careful consideration.
In particular, non-stationary transport processes and quantum effects must be taken
into account in sub-micron devices. Although hot-carrier transport effects do not
appear to greatly affect the terminal characteristics of devices with dimensions greater
than 1 pm, traditional device simulations have been shown to be invalid for modelling
microscopic effects. The situation changes substantially for devices where the critical
dimension is in the range 0.1-0.5 pm and it is found that both the terminal current
and transport phenomena differ considerably between classical models and those which
account'for hot-electron and quantum effects. Semiclassical models and Monte Carlo
simulations have been developed to account for transient transport conditions and
quantum-mechanical effects in devices where the active channel is greater than 0.1 pm.
Future Monte Carlo simulations must address the problem of including electron-
electron interactions and ionised impurity scattering in heavily doped devices
(> loz4m-3). The development and understanding of new types of device under 0.1 pm
in size relies on the implementation of quantum transport theory. Monte Carlo
techniques may also be useful in the study of such quantum transport models and may
lead to a better understanding of the range of applicability of the Boltzmann transport
equation.
New device structures such as two-dimensional electron gas field-effect transistors
(TEGFET), heterojunction bipolar transistors and other high-frequency devices fabri-
cated from materials such as InP, AlGaAs and GaInAs present another challenge to
 Semiconductor device modelling 27 1

the device physicist because the limited information available on semiconductor

material properties, the difficulty in modelling the complex structures and the need to
consider quantum effects. Monte Carlo techniques may be useful in studying quasi-one-
dimensional structures. Although in principle heterostructure devices can be treated
in a similar way to existing devices, such as MOS structures, the effects of quantised
energy levels, inter-band scattering and band population in the very small active regions
are not fully understood and may prove significant. Existing information obtained
from quantum-mechanical models indicates that the operation of very small-scale
devices (<O. 1 pm) cannot be accounted for by classical or semiclassical models and
requires a quantum transport treatment.
The recent development of industrial semiconductor simulation software packages
based on physical device models represents an important trend in device modelling.
Recent advances in semiconductor technology, particularly in the VLSI area, require
more realistic models which, coupled with a dramatic increase in computing power,
has led to the adoption of sophisticated simulation programs by semiconductor device
manufacturers. Future developments in semiconductor device modelling are likely to
be centred around improving non-stationary and quantum transport representations.

References

Aas E J and Blotekjaer K 1974 J. Phys. Chem. Solids 35 1053-9

Abe M, Yanagisawa S, Wada 0 and Takanashi H 1974 Appl. Phys. Lett. 25 674-5
Adachi T, Yoshii A and Sudo T 1979 IEEE Trans. Electron Devices ED-26 1026-31
Adler M S 1979 Proc. NASCODE I Con$ 3-30
- 1980 Solid St. Electron. 23 845-53
Alberigi-Quarantra A, Jacoboni C and Ottaviani G 1971 Reo. Nuooo Cim. 1 445
Asbeck P M, Miller D L, Asatourian R and Kirkpatrick C G 1982 IEEE Trans. Electron Device Lett. EDL-3
403-6
Baccarani G, Jacoboni C and Mazzone A M 1977 Solid St. Electron. 20 5
Balazs N L and Jennings B K 1984 Phys. Rep. 104 349
Barker J R 1973 J. Phys. C: Solid St. Phys. 6 2663
- 1978 Solid St. Electron. 21 267
-1980a Physics ofNon-linear Transport in Semiconductors ed K Ferry, J R Barker and C Jacoboni (New
York: Plenum) pp 589-606
-1980b Physics ofNon-linear Transport in Semiconductors ed K Ferry, J R Barker and C Jacoboni (New
York: Plenum) pp 127-51
Barker J R and Ferry D K 1979 Phys. Reo. Lett. 42 1779
- 1980 Solid St. Electron. 23 519-49
Barnes J J and Lomax R J 1973 Solid St. Commun. 13 93-5
- 1974 Electron. Lett. 10 341-3
-1977 IEEE Trans. Electron Deoices ED-26 1026-31
Bauhahn P E, Haddad G L and Masnari N A 1973 Electron. Lett. 9 460
Beard J J and Rees H D 1981 Electron. Lett. 17 81 1
Berry R D 1971 IEEE Trans. Circuit Theory CT-18 40-50
Berz F, Gough P A and Slatte J A G 1984 Proc. Int. Conj on Simulation of Semiconductor Devices and
Processes ed K Board and D R J Owen (Swansea: Pineridge) pp 182-203
Bettess P, Greenough C and Emson C 1984 Roc. Int. Con$ on Simulation of Semiconductor Devices and
Processes ed K Board and D R J Owen (Swansea: Pineridge) p p 247-66
Birdsall C K and Fuss D 1969 J. Comput. Phys. 3 494
Blotekjaer K 1970 IEEE Trans. Electron Devices ED-17 38-47
Boardman A D, Fawcett W and Rees H D 1968 Solid St. Commun. 6 305-7
Boardman A D, Fawcett W and Ruch G 1971 Phys. Stat. Solidi a 4 133
Bosch R and Thim H W 1974 IEEE Trans. Electron Devices ED-21 16-25
Bozler C 0 and Alley G D 1980 IEEE Trans. Electron Devices ED-27 1128-41
272 C M Snowden

Brooks A N and Hughes T J R 1982 Comp. Meth. Appl. Mech. Engng 32 199-259
Brooks H 1955 Adv. Electron. Electron Phys. 7 87-182
Butcher P N, Fawcett W and Qgg N R 1967 Br. J. Appl. Pbys. 18 755
Buturla E M et a1 1981 ZBMJ. Res. Dev. 25 218-31
Capassco F 1983 Surj Sci. 132 527-39
Carnez B, Cappy A, Kaszynski A, Constant E and Salmer G 1980 J. Appl. Phys. 51 784-90
Carroll J E 1970 Hot Electron Microwave Generators (London: Arnold) pp 26-48
Chamberlain S G and Husain A 1981 Proc. IEEE Int. Electron Devices Conf:592-5
Chester G V 1963 Rep. Prog. Phys. 26 41 1
Colliver D J, Gray K W, Jones D J, Rees H D, Gibbons G and White P M 1973 Gallium Arsenide and
Related Compounds. Inst. Phys. Con$ Ser. N o 17 ed C Hilsum (Bristol: Institute of Physics) p 100
Cook R K 1983 ZEEE Trans. Electron Devices ED-30 1103-10
Cook R K and Frey J 1982 In?. J. Comp. Math. Electron. Elec. Engng 1 65-87
Cottrell P E and Buturla E M 1978 IEEE Trans. Electron Devices ED-25 1346
Curby R C and Ferry D K 1973 Phys. Stat. Solidi a 15 319
Curtice W R and Yun Y H 1981 ZEEE Trans. Electron Devices ED-28 954
Dacey G C and Ross I M 1955 Bell Syst. Tech. J. 34 1149
Delagebeaudeuf D and Linh N ?’ 1982 IEEE Trans. Electron Devices ED-29 955-60
De Mari A 1968 Solid St. Electron. 11 33-58
De Mey G 1974 Solid St. Electron. 17 977-9
- 1978 Solid St. Electron. 21 595-6
De Mey G , Loret D and Van Calster 1984 Pruc. In?. Cunf: on Simulation of Semiconductor Devices and
Processes ed K Board and D R J Owen (Swansea: Pineridge) pp 69-81
De Visschere P and De Mey G 1977 Electron. Lett. 13 104-6
Doades G P, Howes M J and Morgan D V 1984 Proc. Znt. Conf: on Simulation ofSemiconductor Devices
and Processes ed K Board and D R J Owen (Swansea: Pineridge) pp 353-67
Dubock P 1970 Electron. Lett. 6 53-5
Eastwood J W and Hockney R W 1974 J. Comput. Phys. 16 342
Faricelli J V, Frey J and Krusius J P 1982 IEEE Trans. Electron Devices ED-29 377-88
Fauquembergue R J, Zimmermann A, Kaszynski A, Constant E and Microondes G 1980 J. Appl. Phys. 51 1065
Fawcett W, Boardman D A and Swain S 1970 J. Phys. Chem Solids 31 1363-90
Fawcett W, Hilsum C and Rees H D 1969 Solid St. Commun. 7 1257
Fawcett W and Rees H D 1969a Phys. Lett. 11 731
- 1969b Pbys. Lett. 29A 578-9
Fenner R T 1978 Computing for Engineers (London: Macmillan)
Ferry D K 1980 Physics of Non-linear Transport in Semiconductors ed D K Ferry, J R Barker and C Jacoboni
(New York: Plenum) pp 577-88
Ferry D K and Barker J R 1979 Phys. Rev. Lett. 42 1779
- 1980 Solid St. Electron. 23 545
-I98 1 J. Appl. Phys. 52 8 18
Franz G A and Franz A F 1984 Proc. In?. Conf:on Simulation of Semiconductor Devices and Processes ed K
Board and D R J Owen (Swansea: Pineridge) pp 204-18
Franz A F, Franz G A, Selberherr S, Ringhofer C and Marowich P 1983 IEEE Trans. Electron Devices
ED-30 1072-82
Freeman K R and Hobson G S 1972 IEEE Trans. Electron Devices ED-19 62-70
- 1973 ZEEE Trans. Electron Devices ED-20 891-905
Gerald C F 1973 Applied Numerical Analysis (New York: Addison Wesley)
Glisson T H, Hauser J R, Littlejohn M A, Hess K, Streetman B G and Schichijo H 1980 J. Appl. Phys. 51 5445
Goto G, Nakamura T and Isobe T 1975 IEEE Trans. Electron Devices ED-22 120-6
Grebene A B and Ghandi S K 1969 Solid St. Electron. 12 573
Greenfield J A and Dutton R W 1980 ZEEE Trans. Electron Devices ED-27 1520-32
Greenfield J A, Price C H and Dutton R W 1983 Process and Device Simulation for MOS-VLSZ Circuits ed
P Antognetti et a1 (Amsterdam: Nijhoff) pp 432-80
Grubin H L 1978 IEEE Trans. Electron Devices ED-25 511-9
- 1980 IEEE Trans. Microwave Theory Technique M’IT-28 442
Grubin H L, Ferry D K and Gleason K R 1980 Solid St. Electron. 23 157-72
Grubin H L, Ferry D K, Jafrate G J and Barker J R 1982 VLSI Electronics Microstructure Science vol 3
(New York: Academic) pp 197-299
 Semiconductor device modelling 27 3

Grubin H L, Shaw M P and Solomon P R 1973 IEEE Trans. Electron Devices ED-20 63-78
Guerrieri R and Rudan M 1983 IEEE Trans. Electron Devices ED-30 1097-103
Gummel H K 1964 IEEE Trans. Electron Devices ED-11 455-65
Hachtel G D, Mack M H and O’Brien R R 1981 IBM J. Res. Dev. 25 246-60
Hauser J R 1967 Solid St. Electron. 10 577
Hedin L and Lundquiest S 1969 Solid St. Phys. 23 1
Heimeier H H 1973 IEEE Trans. Electron Devices ED-20 708-14
Hess K 1980 Physics of Non-Linear Transport in Semiconductors ed K Ferry, J R Barker and C Jacoboni
(New York: Plenum) chap 1, pp 1-42
- 1981 IEEE Trans. Electron Devices ED-28 937-40
Higgins J A and Pattanayak D N 1982 IEEE Trans. Electron Devices ED-29 179-83
Himsorth B 1972 Solid St. Electron. 15 1353
-1973 Solid St. Electron. 16 931
Hobson G S 1974 The G u n n Effect (Oxford: Clarendon)
Hockney R W 1970 Merh. Comput. Phys. 9 135-211
- 1971 J. Comput. Phys. 8 19-44
Hackney R W and Reiser M 1972 IEM Res. Rep. R2482
Hockney R W, Warriner R A and Reiser M 1974 Electron. Lett. 10 484
Husain A and Chamberlain S G 1982 IEEE Trans. Electron Devices ED-29 631-8
Jacoboni C 1974 Phys. Stat. Solidi b 65 61
Jacoboni C and Reggiani L 1979 Adv. Phys. 28 493-553
- 1983 Reo. Mod. Phys. 55 645-705
Jacoboni C, Reggiani L and Brunetti R 1981 Proc. 3rd Int. Conf:on Hot Carriers in Semiconductors (J.
Physique Colloq. 42 C7-123)
Kennedy D P and O’Brien R R 1970 IBM J. Res. Deo. 14 95-1 16
Klinger M I 1979 Problems of Linear Electron (Polaron) Transport Theory in Semiconductors (Oxford:
Pergamon)
Kohn W and Luttinger J M 1957 Phys. Rev. 108 590
-1958 Phys. Rev. 109 1892
Kotecha H and Noble W P 1980 Proc. ZEEE Electron Deoices Meeting (New York: I E E E ) pp 724-7
Kroemer H 1966 ZEEE Trans. Electron Devices ED-13 27
Kubo R 1957 J. Phys. Soc. Japan 12 570
- 1966 Rep. Prog. Phys. 29 225-84
Kurosawa T 1966 Proc. Int. Conf:on Physics of Semiconductors, Kyoto (J. Phys. Soc. Japan 21 Suppl. 424)
Lakshminarayana M R and Partain L D 1980 IEEE Trans. Electron Devices ED-27 546-52
Lebwohl P A and Price P J 1971a Appl. Phys. Lett. 19 530-2
-1971b Solid St. Commun. 9 1221
Lehovec K and Zuleeg R 1970 Solid St. Electron. 13 1415-26
Littlejohn M A, Hauser J R and Glissan T H 1977 J. Appl. Phys. 48 4587-90
Liu S, Hoefflinger B and Pederson D 0 1980 IEEE Trans. Electron Devices ED-27 1550-8
Lundstrom M S and Schuelke R J 1982 Solid St. Electron. 25 683-91
-1983 IEEE Trans. Electron Devices ED-30 1151-9
Luttinger J M 1968 Mathematical Methods in Solid-State and Superfluid Physics (Edinburgh: Oliver and Boyd)
McCumber D E and Chynoweth A G 1966 IEEE Trans. Electron Devices ED-13 4-21
Machek J and Selberherr S 1983 ZEEE Trans. Electron Deoices ED-30 1083-94
Maksym P A and Hearn C J 1978 Solid St. Electron. 21 1531
Maloney T J and Frey J 1977 J. Appl. Phys. 48 781-7
Meyer H A (ed) 1956 Symp. on Monte Carlo Methods (New York: Wiley)
Mock M S 1973 Solid St. Electron. 16 601-9
- 1981 Solid St. Electron. 24 959-66
Moglestue C 1979 IEEE Solid St. Electron Deoices 3 133-6
-1982 Comp. Meth. Appl. Mech. Engng 30 173-208
- 1984 Proc. Int. Conf:on Simulation of Semiconductor Devices and Processes ed K Board and D R J
Owen (Swansea: Pineridge) pp 153-63
Nougier J P 1980 Physics ofNon-Linear Transport in Semiconductors ed K Ferry, J R Barker and C Jacoboni
(New York: Plenum) pp 415-65
Oh S-Y and Dutton R W 1980 IEEE Trans. Electron Devices ED-27 2101-8
Oh S-Y, Ward D E and Dutton R W 1980 IEEE Trans. Electron Deoices ED-27 1571-8
274 C M Snowden
Ohmi T and Hasuo S 1970 Proc. Int. Conf: on Physics and Semiconductors (USAEC) p 60
Oka H, Nishiuchi K, Nakamura T and Ishikawa H 1980 IEEE Trans. Electron Devices ED-27 1514-20
Park Y-J, Tang T-W and Navon D-H 1983 IEEE Trans. Electron Devices ED-30 1 1 10-6
Please C P 1982 I M A J. Appl. Math. 28 301-18
Pone J-F, Castagne R C, Courant J-P and Arnodo C 1982 IEEE Trans. Electron Devices ED-29 1244-55
Price P J 1970 I B M J . Res. Dev. 14 12
- 1978 Solid St. Electron. 21 9-16
- 1979 Semiconductors and Semimetals (New York: Academic) p 249
Price P J and Stern F 1983 Surf: Phys. 132 577-93
Puce1 R A, Haus H A and Statz H 1975 Adv. Electron. Electron Pbys. 38 195
Rees H D 1968 Pbys. Lett. 26A 416-7
- 1969 J. Phys. Cbem. Solids 30 643
Rees H, Sanghera G S and Warriner R A 1977 Electron. Lett. 13 156-8
Reif F 1965 Fundamentals of Statistical and Thermal Physics (New York: McGraw-Hill)
Reiser M 1971 Electron. Lett. 7 353-5
_. 1972 Comp. Meth. Appl. Mech. Engng 1 17-38
- 1973 IEEE Trans. Electron Devices ED-20 35-45
Richmeyer R D and Marton K W 1957 Di’erence Methods for Initial Value Problems (New York: Wiley-
Interscience)
Robrock R B 1969 IEEE Trans. Electron Devices ED-16 647-53
Ruch J G 1972 IEEE Trans. Electron Devices ED-I9 652-9
Ruch J G and Fawcett W 1970 J. Appl. Pbys. 41 3843-9
Salsburg K A, Cottrell P E and Buturla E M 1983 Process and Device Simulation for MOS-VLSI Circuits
(Amsterdam: Nijhoff) pp 582-618
Sangiorgi E, Rafferty C, Pinto M and Dutton R 1984 Proc. Int. Conf: on Simulation of Semiconductor Devices
and Processes ed K Board and D R J Owen (Swansea: Pineridge) pp 164-71
Scharfetter D L and Gummel H K 1969 IEEE Trans. Electron Devices ED-16 64-7
Selberherr S, Schutz A and Potzl H W 1980 IEEE Trans. Electron Devices ED-27 1540-50
-1983 Process and Device Simulation for MOS-VLSI Circuits ed P Antognetti et al (Amsterdam: Nijhoff)
pp 490-581
Shichijo H and Hess K 1981 Pbys. Rev. B 23 4197
Shockley W 1952 Proc. IRE 40 1365-77
Shur M 1980 IEEE Trans. Electron Devices Lett. EDL-1 147
- 1981a IEEE Trans. Electron Devices Lett. ED-28 1120-30
- 1981b Solid St. Electron. 24 11
Shur M and Eastman L F 1978 IEEE Trans. Electron Devices ED-25 606-1 1
-1979 IEEE Trans. Electron Devices ED-26 1677-83
Slotboom J W 1969 Electron. Lett. 5 677-8
- 1973 IEEE Trans. Electron Devices ED-20 669-79
Snowden C M 1984a h o c . Int. Conf: on Simulation of Semiconductor Devices and Processes ed K Board and
D R J Owen (Swansea: Pineridge) pp 406-25
-1984b Proc. 14th Europ. Microwave Con$ (Sevenoaks: Microwave Exhibitions and Publishers Ltd) pp
717-22
Snowden C M, Doades G P Howes M J and Morgan D V 1981 Proc. f f t h Europ. Microwave Conf: pp 633-8
Snowden C M, Howes M J and Morgan D V 1983 IEEE Trans. Electron Devices ED-30 1817-24
Stratton R 1972 IEEE Trans. Electron Devices ED-19 1288-92
Strong G and Fix G J 1973 A n Analysis of the Finite-Element Method (Englewood Cliffs, NJ: Prentice-Hall)
Sutherland A D 1980a IEEE Trans. Electron Devices ED-27 1297-8
- 1980b Solid St. Electron. 23 1085-7
Tasker P J 1983 PbD Thesis University of Leeds, pp 57-77
Tatarskii V I 1983 Sou. Pbys.- Lisp. 26 3 I I
Teitel S L and Wilkins J W 1983 IEEE Trans. Electron Devices ED-30 150-3
?him H W 1968 J. Appl. Pbys. 39 3897-905
Toyabe T and Asai S 1979 IEEE Trans. Electron Devices ED-26 453-61
Toyabe T, Yamaguchi K, Asai S and Mock M 1978 IEEE Trans. Electron Devices ED-25 825-32
Turner J A and Wilson B L H 1968 Gallium Arsenide. Inst. Pbys. Con$ Ser. No 7 ed H Strack (Bristol:
Institute of Physics) p 195
Varga R S 1962 Matrix Iterative Analysis (Englewood Cliffs, NJ: Prentice-Hall)
Wada T and Frey J 1979a IEEE J. Solid St. Circuits SC-14 390-412
 Semiconductor device modelling 275

-1979b IEEE Trans. Electron Devices ED-26 476

Warriner R A 1977 IEEE Solid St. Electron Devices 7 105-10
Watanabe D S and Slamet S 1983 IEEE Trans. Electron Devices ED-30 1042-9
Wigner E P 1932 Phys. Rev. 40 749
Yamaguchi K, Asai S and Kodera H 1976 IEEE Trans. Electron Devices ED-23 1283-90
Yoshii A, Kitazawa H, Tomizawa M, Horiguchi S and Sudo T 1982 IEEE Trans. Electron Devices ED-29 184-9
Yu S P and Tantraporn W 1982 VLSI Electronic Microstructure Science vol 3 (New York: Academip) chap
5, pp 163-95
Zimmermann J and Constant E 1980 Solid St. Electron. 23 915-25
Zimmermann J, Leroy Y and Constant E 1978 J. Appl. Phys. 49 3378-83

View publication stats