Vous êtes sur la page 1sur 5

THE MONTE CARLO METHOD FOR SEMICONDUCTOR DEVICE SIMULATION

R. W. Kelsall

Introduction

The development of sophisticated techniques for semiconductor device fabrication, such as molecular beam
epitaxy (MBE) and electron beam lithography, has enabled the realisation of complex multi-layer device
structures, with feature sizes previously considered impossible. Examples of ‘new’ devices include the High
Electron Mobility Transistor (HEMT) and the Heterojunction Bipolar Transistor (HBT), but even the lowly
MOSFET has been refined such that operation at gate lengths of less than 0.1pm has been demonstrated [ 1,2].
These high specification devices will find ready application both in high frequency satellite and mobile
communication systems, and for high speed, ultra high integration density (ULSI)logic.
As the device structures becomes more complex, and as the design possibilities become greater and more
varied, empirical routes to new device designs become impractically expensive. In this scenario, device
simulation represents an extremely cost effective tool in the design process; a few hours CPU time can save
many weeks of clean-room time.

Simulation methods

In the semiconductor industry, simulation work generally falls into three categories; process, device and circuit
simulation. Device simulation is understood to be the modelling of a individual active semiconductor device,
such as a diode or transistor. The simplest approach to device modelling is the construction of an equivalent
circuit. Equivalent circuit models have been widely used; there advantages are their simplicity: and the
consequent ease with which they can be incorporated in a complete circuit description. However, such models
generally represent a v e q crude approximation to the actual device performance under various conditions, and
give no physical insight into the operation of the device.
The most widely used physical device model is the so-called drift-diffusion simulation [3]. On a microscopic
level, the behaviour of a population of electrons inside a semiconductor device when subject to excitation -
such as that caused by an electric field - can be represented by the Boltzmann transport equation. Solution of
the Boltzmann equation gives the electron energy distribution function. In the drift-diffusion approach, a first
order approximation to the distribution function is obtained. The method has given very useful results for a
wide range of semiconductor devices, but becomes inaccurate for the sub-micron feature sizes commonly
encountered today. The main disadvantage of the method is that it is based on a near-equilibrium
approximation; electron heating effects are not included. This restriction can be circumvented by using a
hydrodynamic or energy transport simulation 141. The method provides a second order solution of the
Boltzmann equation; the second moment equation gives an expression for energy transport within the device.
Initially, problems were encountered with reliable closure of the system of equations, but of late, good results
have been obtained by this method.
Whilst the energy transport simulation exiends the possible range of device feature sizes which can be
considered, it still represents an approximation to the true electron energy distribution. The Monte Carlo
approach to semiconductor device simulation yields essentially an exact solution to the Boltzmann equation.
The method is therefore ex%remely well suited to modelling ultra-small devices: in which transport occurs
under conditions far from equilibrium.

0 1995 The institution of Electrical Engineers.


Printed and published by the IEE, Savoy Place, London WCPR OBL, UK.

R W Kelsall is with the Microwave and Terahertz Technology Group. Department of Electronic and Electrical
Engineering. University of Leeds.

3/1
The Monte Carlo method for semiconductor device simulation

The term Monte Carlo is used to describe any simulation method which is based on random numbers. The
method is therefore quite general, and can be applied to a wide range of problems. The design of a simulation
involves the identification of random elements in the system, and the derivation of probabili5 distributions to
describe this random behaviour. For semiconductor device modelling, we consider the simple case of an
electron travelling through a semiconductor sample, under the influence of an electric field. The electron is
accelerated by the field, but its motion is punctuated by scattering 'events', which result in changes in the
electron's speed and direction. Three elements of randomness may be identified in such a motion:

i) The time for which the electron travels until a scattering event occurs.
ii) The nature of the scattering event (phonon, impurity, or electronelectron scattering, etc)
iii) The scattering angle; i.e., the direction of motion after scattering.

Probability distributions are calculated for each case, and are used in the simulation to convert uniformly
distributed random numbers into these three stochastic variables. The process is repeated, typically between
lo4 and 10' times, to simulate a typical electron path through the semiconductor. To characterise steady state
transport in a homogeneous sample, it is merely necessary to simulate the path of one electron for a sufficiently
long time. The system is ergodic; this single path is representative of the average of all actual electron paths
in the sample. The electron drift velocity in the semiconductor can therefore be obtained from a simple time
average of the electron velocity during the one simulated trajectory. Such one-particle simulations have
proved exTremely useful for characterising steady state transport in a range of materials, and are often used to
provide parameters - such as mobility and saturation velocity - for use in approximate device simulations.
To stud\- transient phenomena and/or inhomogenous semiconductor structures it is necessan; to simulate the
trajectories of a representative batch, or ensemble, of particles. This is the approach generally used in Monte
Carlo device modelling. In addition, it is usually necessary to ensure that the simulation reproduces the correct
electric fields inside the device. In the absence of magnetic effects, this involves solving only the Poisson
equation. typically in 1 or 2 dimensions. However, to ensure that the field solution and the transport model are
self-consistent, the Poisson equation must be re-solved at regular and frequent intervals throughout the
simulation. At each time-step, the location of all particles in the ensemble gives the charge density
distribution, which serves as the source term for the Poisson equation; the resultant local electric fields are
then used to accelerate the particles throughout the next time-step. Device doping can be included simply as a
fixed charge, and realistic doping profiles can be incorporated if required. The device contacts are included \.ia
boundan conditions for the Poisson equation.

Applicat'ions

A prototype Monte Carlo device simulation was developed nearly 20 years ago at the UniversiG of Reading
151. This program gave a 2-dimensional (2D) simulation of a G a A s planar MESFET, including a self-
consistent solution of the 2D Poisson equation. The basic structure of this code has been widely used since.
for simulations of MESFETs, MOSFETs [6] and HEMTs [7] along with a number of more esoteric devices.
The Monte Carlo method has also been applied to HBTs; for this device, useful results can be obtained using a
l-dimensional(1D) model [SI, although 2D HBT simulators have also been developed.
The 2D self-consistent Monte Carlo simulator can produce a wide range of results, for both macroscopic
parameters - IN curves, transconductance, and equivalent circuit capacitances (e.g. figure l)?and microscopic
quantities -the intemal potential, electric field; current density and electron density distributions (e.g. figures 2
illid 3). The electron energy distribution in any region of the device can be probed, as can the population of the
various energy bands in the semiconductor(s).

Improvements

Desirable improvements to the standard Monte Carlo simulation algorithm may be considered in three
categories:

3/2
i) improvements in the fundamental physics of the model
ii) making the modelled device closer to reality
iii) run-time optimisation

i) The standard Monte Carlo simulation is based on semi-classical physics. For modelling transport in
narrow layered structures andor ultra-small gate length devices, quantum mechanical phenomena should be
considered. Monte Carlo simulations have already been developed to study transport in semiconductor
quantum wells, using a solution of the 1D time independent Schrodinger equation [9,10]. The problem is
complicated by the fact that the scattering rates for the electrons or holes must be calculated using the
appropriate quantum mechanical wavefunctions. Furthermore, the energy band spectrum of the quantum well
system has much more fine structure than that in the bulk material. For devices, such as the HEMT, which
incorporate narrow semiconductor layers, modelling any quantum mechanical phenomena is even more
difficult. Any solution of the Schrodinger equation must properly be self-consistent with both the electric field
-
solution and the transport solution representing an apparently intractable problem. An alternative approach is
to use a so-called quantum Monte Carlo scheme which simulates, not semi-classical but ‘quantum’ trajectories,
yielding the quantum mechanical density matrix evewhere in the device [ 113. This approach has not been
widely used for practical devices.
Even in bulk semiconductors,where size quantisation is not an issue, the electron and hole energy band spectra
must be adequately described. The usual parabolic approximation to the low-lying energy bands leads to
substantial inaccuracies even at moderate electric fields; in most simulations departure from this
approsimation is modelled by inclusion of a nonparabolicit) parameter for each band. For the valence bands,
and for the conduction bands in silicon, anisotropy of the bands must also be included. Nevertheless, for ultra-
short gate devices the intemal electric fields can be of the order of 107V/m, giving rise to electron energies of
several eV. Accurate simulation of transport under such conditions requires a detailed description of the band
spectra at high energies, including an array of higher lying conduction bands normally ignored in transport
studies. Such simulations have been developed, most famously by Fischetti and Law [6]; however, such a
comprehensive model of the energy bands exacts a high price in terms of simulation run-time.
i) Most models, for simplicity, assume a rectangular device domain in 2-dimensions. However, high
technology devices, such as short gate HEWSand MESFETs, include a recessed gate configuration: the
operation of these devices can be strongly dependent on the shape of the recess. Simulation of non-rectangular
device geometries using the Monte Carlo approach is no problem in principle, but the added complexiv of the
algorithm inevitabl), leads to an increase in run-time.

The performance of some high speed semiconductor devices can be limited, not by transport in the active
region, but by the detailed electrical properties of the contacts and surface. Accurate simulation of ohmic
contacts can be particularly troublesome; in Monte Carlo simulations carrier injection at an ohmic contact is
modelled b\* injecting sufficient carriers to maintain charge neutrality immediately undemeath the contact:
however, this is obviousl). an idealised picture. In practice, so-called ohmic contacts can still present a
potential bamer to incoming electrons, with injection achieved by a combination of tunneling and thennionic
emission. In 111-V compound semiconductor FETs, the contact metal often diffuses below the surface of the
semiconductor wafer. leading to a complex, undetermined contact metallurgy. At the semiconductor surface,
localised states occur, due to defects, but, even in their absence, to the fact that the semiconductor crystal
terminates. These states can trap electrons or holes, giving rise to a surface potential which, in general, varies
across the esposed surface area. Some devices may have been subject to surface passivation processing, which
further affects the electrical properties of the surface.
iii) The main disadvantage of the Monte Carlo method has traditionally been the long run-times required. A
range of commonly used techniques esist for optimising different parts of the algorithm. These efforts have
been substantially assisted in recent years by the enormous increase in available desktop computins po\ver. It
is non- possible to run 2D self-consistent FET simulations at a rate of typicall>;15 mins per point on a high
specification desktop workstation [ 121. Although this performance allows a substantial daily throughput, what
semiconductor design engineers really want are essentially instant results - run times of no ore than a few

3/3
seconds per point. One potential route to this goal involves the use of parallel hanvare and software. An
ensemble Monte Carlo simulation is inherently a massively parallel problem: requiring the same algorithm to
be applied to an ensemble of typically 10,000 particles. However, this parallel structure is broken b> the
repeated solutions of the Poisson equation, which require synchronisation of: and data collection from. the
entire ensemble. This limitation does not preclude the use of parallel computing techniques, but it does
substantially reduce the anticipated speed improvement. A further objection is that parallel computing is not
yet a fully mature tool: the implementation of parallel algorithms requires substantial source code
modification and dependency on some non-standard, proprietq compiler. Code portabiliv is therefore
greatly impaired. However. parallel systems are becoming more widely accepted, with many new workstations
-
incorporating a multiprocessor configuration; it is likely that further developments in parallel computing for
example, a full implementation of the parallel capability specified in the FORTRAN 90 standard - will make
the parallel implementation of Monte Carlo simulations less troublesome and more efficient.

Conclusions

There can be little doubt that the Monte Carlo method for semiconductor device simulation has enormous
power as a research tool. It represents a detailed physical model of the semiconductor material(s). and
provides a high degree of insight into the microscopic transport processes. However. if the authority ascribed
to Monte Carlo models of devices at 1pm feature size is to be maintained for devices below 0.1 pm, modelling
of the fundamental physics must be further improved - see (i) above. If the Monte Carlo method is to be
successfkl as a semiconductor device design tool, the device model must be made more realistic - as described
in (ii) above. Success in the industrial sector depends on (ii), but also on achieving fast run-times (iii).
Improvements in (i) will have a large cost in run time; improvements in (ii), a lesser but still significant cost.
Consequently. it is in (iii) - optimisation - where the scope and need for ingenuiv is now greatest.

References

1. M. Ixase, T. Mizuno, M. Takahashi, H. Niiyama, M. Fukumoto, K. Ishida, S. Inaba, Y. Takigami, A.


Sanda. A. Toriumi and M. Yoshimi Electron Device Letters EDL-14 51 (1993)
2. G. G. Shahidi, J. Wamock, S . Fischer, P. A. McFarland, A. Acovic, S. Subbanna, E. Ganin, E. Crabbe'. J.
Comfort, J. Y. C. Sun. T. H. Ning and B. Davari Electron Device Letters EDL-14 466 (1993)
3. C M Snowden Seniiconductor Device Modelling Springer-Verlag, London (1989)
4. K Blotekjaer IEEE Trans Electron Devices ED-17 38 (1970)
5 . R A Waniner Solid State and Electron Devices 1 105 (1977)
6. M V Fischetti and S E Laus Phys Rev. B 38 9721(1988)
7. D H Park and K F Brennan IEEE Trans EZectron Devices ED-37 618 (1990)
8. D T Hughes. R A Abram and R W Kelsall I .Trans EZectron Devices ED42 20 1 (1995)
9. K Yoko?-ama and I(.Hess Phys Rev. B 33 5595(1986)
10. R W Kelsall. A C G Wood and R A Abram Semicond. Sci. Technol. 6 841 (1991)
1 1. R. Bmnetti, C Jacoboni and F Rossi Phys Rev. B 39 104781 (1989)
12. -4pprosimate run time for a 2D self-consistent HEMT simulation on a 20MFLOP workstation

Further Reading

C Jacoboni and L Reggiani Rev. Mod. Phys. 53 685 (1983)


C Jacoboni and P Lugli The Monte Carlo Method for Semiconductor Device Simulation Springer-Verlag.
Vienna (1 989)
K Hess (ed) Monte Carlo Device Sirnulation: Full Band and Beyond Kluwer Academic (199 1)
R M: Hockne5-and J W Eastwood Computer.Simulation Using Particles IOP Publishing, Bristol (1988)
C Moglestue Monte Carlo Siinzilation of Semiconductor Devices Chapman and Hall, London (1993)

3/4
Figure 1

Simulated drain current vs. gate


voltage for a delta-doped (buried
channel) MOSFET with a O.1pm
self-aligned gate. Results are shown
for devices with buried channels
located d=10081 and 200A below the
gate, and for a conventional
MOSFET. The drain voltage is 1V in
all cases.
-1.0 4.5 0.0 0.5 1.0

Figure 2

Simulated conduction band-edge


energy profile for a delta-doped
MOSFET as described above
(d=200A). The gate and drain
voltages are -0.25V and 1V
respectively. The delta-doping creates
a potential well which acts as a
conduction path for electrons.

0.1 micron delta-FET


Vg = 0 2 5 V . Vd = 1.OV
JX Figure 3

Simulated x-component of the current


density under the gate of the delta-
doped MOSFET. Details as in figure
2. Under these bias conditions, the
device operates in buried channel
mode.

0370. ow -I

3/5

Vous aimerez peut-être aussi