Table of Contents - University of Texas at Austin · A.1 Installation ... 2 2 0 4 χ ω χ ε πρ...
Transcript of Table of Contents - University of Texas at Austin · A.1 Installation ... 2 2 0 4 χ ω χ ε πρ...
Copyright
by
Di Li
2002
The Dissertation Committee for Di Li Certifies that this is the approved version of the following dissertation:
Three-dimensional Monte Carlo Simulation of Ion Implantation
Committee:
Sanjay K. Banerjee, Supervisor
Al F. Tasch, Co-Supervisor
Jack C. Lee
Leonard F. Register
Graham F. Carey
Three-dimensional Monte Carlo Simulation of Ion Implantation
by
Di Li, B. S., M. S.
Dissertation
Presented to the Faculty of the Graduate School of
The University of Texas at Austin
in Partial Fulfillment
of the Requirements
for the Degree of
Doctor of Philosophy
The University of Texas at Austin
December 2002
To my wife, my parents and grandparents
Acknowledgement
First of all, I would like to thank my supervisors and my role models, Drs.
Sanjay K. Banerjee and Al F. Tasch for their invaluable guidance and support
throughout my four-year Ph.D. study. Many thanks to Drs. Jack C. Lee, Leonard
F. Register and Graham F. Carey for taking their time to serve on my committee,
and for their comments and help.
Many thanks also to my fellow workers, and former graduate students,
Geng Wang, Yang Chen, Borna Obradovic, Ganesh Balamurugan, Li Lin and
Steve Morris for the inspiring discussions.
Semiconductor Research Corporation (SRC) has contributed to the support
of this work. I am also grateful to Shyh-Horng Yang, Chuck Machala and Texas
Instruments Inc. for supporting my summer internship in 2001.
Finally, I would like to thank my wife, Ding Yuan, my brother, Mai Li,
my parents and my grandparents for their love, understanding, encouragement and
expectation all these years.
v
Three-dimensional Monte Carlo Simulation of Ion Implantation
Publication No.: ___________________
Di Li, Ph.D.
The University of Texas at Austin, 2002
Supervisors: Sanjay K. Banerjee, Al F. Tasch
A physically-based 3-dimensional Monte-Carlo simulator has been
developed within UT-MARLOWE, which is capable of simulating ion
implantation into multi-material systems and arbitrary topography. Introducing
the third dimension can result in a severe CPU time penalty. In order to minimize
this penalty, a three-dimensional trajectory replication algorithm has been
developed, implemented and verified. More than two orders of magnitude savings
of CPU time has been observed. An unbalanced Octree structure was used to
decompose three-dimensional structures. It effectively simplifies the structure,
and offers a good balance between modeling accuracy and computational
efficiency. This simulator was also extensively verified by comparing the
integrated one-dimensional simulation results with Secondary Ion Mass
Spectroscopy (SIMS).
vi
Table of Contents
Acknowledgement ...........................................................................................................v
Table of Contents...........................................................................................................vii
Chapter 1. Introduction ....................................................................................................1
Chapter 2. Ion Stopping Models and Damage Models ....................................................3
Chapter 3. Low Energy Consecutive Implants into SiO2-Capped Si...............................7
3.1 Introduction ........................................................................................................7
3.2 Model Description and Calibration ....................................................................8
3.3 Results and Analysis...........................................................................................9
3.4 Conclusion ........................................................................................................16
Chapter 4. Three-dimensional Simulations....................................................................21
4.1 Introduction ......................................................................................................22
4.2 New 3-D Trajectory Replication Scheme.........................................................24
4.3 Structure Decomposition ..................................................................................34
4.4 Model Verification ...........................................................................................35
4.5 Example application .........................................................................................49
4.6 Conclusion ........................................................................................................49
Chapter 5. Analytical 1-D and 2-D Ion Implantation Models Based on Legendre
Polynomials....................................................................................................................52
5.1 Introduction ......................................................................................................53
vii
5.2 One-dimensional Modeling ..............................................................................54
5.3 Two-dimensional Modeling .............................................................................66
5.4 Two-D Look-up Table Generation and Simulation Results .............................71
5.5 Conclusion ........................................................................................................80
Chapter 6. Conclusions and Future Work......................................................................81
6.1 Conclusions ......................................................................................................81
6.2 Future tasks.......................................................................................................82
Appendix A: A Brief Manual of TOMCAT ..................................................................83
A.1 Installation ........................................................................................................83
A.2 Quick Tutorial...................................................................................................86
Appendix B: Block Diagram and Flow Chart................................................................95
Bibliography ..................................................................................................................99
Vita...............................................................................................................................105
viii
Chapter 1. Introduction
Ion implantation will remain the dominant doping method to introduce
dopant atoms into silicon for its high throughput and precise control over
implanted dose and areas as comparing with emerging alternative methods, such
as Gas Immersion Laser Doping (GILD) [1,2]. Ion implantation is widely used
with energy ranging from sub-keV to several MeV for the purpose of source/drain
extension formation, source/drain formation, Vt adjust implant, halo implant, and
well formation. Among different profiling techniques for as-implanted profiles,
Secondary Ion Mass Spectroscopy (SIMS) [3] provides an efficient and economic
way to probe 1-D dopant profiles. In 2-D, at the present time, accurate and
efficient profiling techniques are still under development [4,5]. Three-dimensional
experimental profiling techniques are as yet impractical. On the other hand, rapid
advances in Ultra-Large Scale Integrated (ULSI) technology introduced radical
and highly non-uniform structures such as raised source/drain silicon-on-insulator
(SOI) devices [6], and double gate devices [7,8]. Therefore, accurate 2- and 3-
dimensional dopant and damage profiles are of crucial importance. In particular,
to reduce technology development cycle time and cost, the ability to predict the
doping profiles accurately and efficiently of implants into 2- and 3-D complex
structures is highly desirable.
1
Currently, three different types of implant simulators exist. They are
analytical (compact) model based on complex functions or polynomials [9,10,11],
Molecular Dynamics (MD) model [12] and Monte Carlo model [13-17].
Analytical model utilizes readily available profile-implant condition relation
(obtained from experiments or other simulation method) and interpolation
techniques (linearly, logrithmatically or so) to cover major user-specified implant
conditions. This type of implant simulators is used in almost all major
commercially available TCAD tools and is very efficient, yet not very predictive.
On the other hand, molecular dynamics (MD) models are fairly accurate and
predictive by explicitly calculating the rigorous interaction of all the particles,
including all the dopant and the silicon atoms, in the volume of interest. The
complete treatment makes MD models very appealing in terms of the first
principles calculations of displacement energy, and migration energy of dopant
atoms to simulate dopant diffusion. Yet the amount of calculations needed to treat
hyper-thermal processes such as implantation almost forbid its usage in
implantation simulation. Therefore, various techniques have been developed and
approximation has been made to make the method feasible for realistic implant
simulations. Finally, Monte Carlo method, based on physically-verified stopping
models and damage model, is both predictive and computationally-efficient and is
going to be studied extensively in later chapters.
2
Chapter 2. Ion Stopping Models and Damage Models
This section reviews the physics of ion implantation and concentrates on
the physical models used in the three-dimensional Monte Carlo simulator
developed in this work, as well as UT-MARLOWE [18]. Verification of these
physical models is presented in later chapters.
2.1. Stopping Model
During the process of ion implantation, energetic ions travel in the solid
and lose their energy through a series of collisions with the atoms in solid. The
target atoms can gain enough energy to be dislodged from their original sites,
travel through the solid and collide with other atoms, resulting in cascade
collisions. The displaced atoms (interstitials) and the vacant sites (vacancies) will
in turn affect the implanted ions later on.
The interactions between ions and a solid are simulated under Binary
Collision Approximation (BCA) approach and therefore, decoupled into three
parts, nuclear stopping, non-local electronic stopping and local electronic
stopping, described in the following.
The nuclear stopping power is calculated under the binary collision
approximation using the ZBL [19] universal potential
( )r
eZZrV u
221⋅Φ= (2.1)
3
where Φu is a pair specific screening function whose value is available
analytically. The interaction is short range and the stopping power is inversely
proportional to the impact parameter. Nuclear stopping causes target atoms to
escape from their sites, become interstitial atoms, and leave behind vacant sites.
Non-local electronic stopping is also mediated through the electric field
and can be described by the following equations.
( )( )ωε
ωωπε ,
10
20
2
1
kd
kdk
vqZi
dxdE kv
kv∫∫+
−
∞
= (2.2)
Taking into account exchange and correlation [Ma, 1992], one obtains the
following simplified formula for proton stopping:
≥
+−
−
+
<
=
=
=
FFF
gp
F
F
p
vvvv
vv
Emv
vvvvk
L
me
Lmve
dxdE
5.1,314
3532ln
5.1,)(
4
4
422
222
2
3
0
22
220
4
χω
χ
επρ
ω
ερπ
η
(2.3)
where m is the electron rest mass, )(1 02 akFπχ = , a0 is the Bohr radius, kF is
the Fermi wave number, and vF is the Fermi velocity. Here Eg is the band gap of
silicon at the gamma point and is equal to 4.8 eV. The stopping power for a
heavier particle can be obtained by multiplying the proton stopping by the square
4
of the ion charge [20,21], which is calculated by taking into account the ionization
of the moving particle and charge screening effects [19].
Local electronic stopping model is based on electron exchange theory
[22]. It assumes that energy transfer from the ion to the target occurs by
exchanging electrons between these two atoms, and the electrons transferred to
the ion cause it to slow down. The amount of energy loss depends on the
velocities of the atoms and on how close they approach [18].
( )( ) ( )
≥⋅
<⋅
+⋅⋅
+⋅=∆
critB
crit
critB
ZZRZZE
ννννν
νννν
23
121min
35
21
)25.5exp(5 (2.4)
where vcrit denotes the critical velocity beyond which the electrons of the two
atoms will not have sufficient time for free interaction, so the energy transfer will
diminish.
In order to obtain a smoother transition across the critical velocity, an
analytical transfer function is used [18] instead,
( )
−+
−=
22
2exp1exp2critcrit
F νν
ννν (2.5)
2.2. Damage Model
To describe the damage accumulation during the implant process, two
models are currently available. A rigorous model which follows every cascade is
the Kinetic Accumulative Damage Model (KADM) [17], while a simpler one is
5
the modified Kinchin-Pease model [23,24]. In terms of computational and
memory efficiency, a modified Kinchin-Pease formula is more attractive, which
describes the number of defects as a function of deposited energy,
dEEn
2κ
= (2.6)
where =0.8 is a constant, and Eκ d is the displacement threshold energy,
commonly accepted value being around 15 eV for crystalline silicon. Not all of
the defects calculated from Eq. (2.6) will survive, since some of the defects will
recombine with other nearby defects generated in this or a previous cascade. The
net increase in point defects after recombination [24, 25] is given by
)1(αN
Nnfn rec −=∆ (2.7)
where n is given by Eq. (2.6), N is the local defect density, is the critical
defect density for amorphization, and f
αN
rec is a species-dependent parameter
denoting the fraction of defects surviving defect interaction during and between
cascades. The factor ( has been introduced in order to account for the
sub-linear damage growth with dose observed in the experiments. That is, for the
same deposited energy, the damage increase in a more seriously damaged region
should be reduced because of greater chances for the generated interstitials and
vacancies to recombine with each other.
)αNN−1
6
Chapter 3. Low Energy Consecutive Implants into SiO2-Capped Si
The low energy as-implanted profile is very sensitive to the cap oxide
layer thickness and Pre-Amorphization Implant (PAI) conditions, in addition to
implant energy, angle and dose. In this work, theoretical and experimental studies
have been carried out quantitatively to investigate these dependencies. It is found
that a 2nm difference in cap oxide layer thickness could result in a 10nm
difference in junction depth even for a heavy species implant such as Sb. With
and without PAI could result in more than a 35nm difference in junction depth for
low energy B implant. Using Ziegler-Biersack-Littmark (ZBL) pair-specific inter-
atomic potentials in the Monte Carlo ion implantation simulator, UT-
MARLOWE, consecutive implants of PAI and P-type Lightly Doped Drain
(PLDD) were simulated and above effects were accurately captured. The
comparison with Secondary Ion Mass Spectroscopy (SIMS) data shows that
excellent agreement is obtained for different cap oxide layer thicknesses and PAI
conditions between the simulated predictions and SIMS data. The comparison
with Rutherford Backscattering Spectroscopy (RBS) data shows that the damage
profiles are also correctly modeled.
3.1 Introduction
As integrated circuits continue to scale beyond the 100 nm regime, ultra-
shallow junctions are crucial in order to realize desired device performance
7
according to the International Technology Roadmap for Semiconductors [26]. To
engineer P-type Lightly Doped Drain (PLDD) implant profiles, Pre-
Amorphization Implant (PAI) with species such as Sb, provide additional
flexibility and is widely used. However, for low energy implants such as PLDD,
the as-implanted profiles are very sensitive to the cap oxide layer thickness and
PAI conditions. Therefore, developing predictable Monte Carlo models, which
can account for these dependencies is desirable for cost efficiency and reducing
development cycle time.
In this study, experiments were designed to study these dependencies
quantitatively. Based on ZBL pair-specific inter-atomic potential [27] and
Kinchin-Pease [23] damage model, an Sb implant model was developed and
incorporated into UT-MARLOWE [18]. Then, the PAI and PLDD steps were
simulated consecutively. The comparisons with experimental data shows that this
model is capable of accurately predicting both the impurity and the damage
profiles for different PAI and PLDD conditions for different oxide layer
thicknesses.
3.2 Model Description and Calibration
The Monte Carlo model is based on the binary collision approximation
(BCA) as described in Chapter 2. In other words, the stopping power of the fast
moving ion in solids is decoupled into three parts: nuclear stopping, local
8
electronic stopping and non-local electronic stopping. Typically, for low energy
and heavy species implants, nuclear stopping dominates among these three kinds
of stopping powers and we need to pay special attention to it. Therefore, instead
of ZBL universal potential [19], ZBL pair-specific inter-atomic potential [27] was
used for improved accuracy in simulating nuclear stopping power of the fast
moving ion. The electronic model mainly follows the ones developed within UT-
MARLOWE. In particular, local electronic stopping follows Firsov’s electron
exchange model [18], while non-local electronic stopping calculates the ion’s
movement through the potential established by polarization of the surrounding
electron sea [18]. To simulate the de-channeling effect caused by the damage
accumulation during the implantation process, the modified Kinchin-Pease [24]
damage model was used and calibrated (specifically, on the recombination factor)
for its simplicity, accuracy and computational efficiency.
3.3 Results and Analysis
A set of experiments was designed in order to understand the dependencies of the
as-implanted Sb profiles on energy, dose and cap oxide layer thickness, as well as
the dependence of as-implanted B and BF2 profiles on PAI. Chemical Vapor
Deposition (CVD) of oxides was used to cap (100) single crystal silicon with
oxides of different thicknesses for accurate control purpose. For the 0 nm cap
oxide wafers, a HF dip was performed to remove any residual oxide, and then all
9
the wafers were kept in nitrogen purge boxes prior to the implants for control
purposes. After the implantation and prior to the Rutherford Backscattering
Spectroscopy (RBS) measurement, the oxide layer was removed to improve RBS
accuracy. The dopant profile dependence on cap oxide layer thickness is shown in
Figs. 3.1-3. A 2nm difference in cap oxide layer thickness (between bare silicon
and 2nm oxide cap) could result in more than 10nm difference in the junction
depth (at the concentration level of 1×1018cm-3) for as-implanted dopant profiles,
even for heavy species such as Sb. As the cap oxide layer thickness increases, the
dopant profile becomes less sensitive to the cap layer differences. This sensitivity
is well captured by the new model when comparing the predicted results with
experimental SIMS data in Figs. 3.1-3. Using the damage profiles predicted by
the Sb implant, a second implant is simulated and the simulation results are shown
in Figs. 3.4-5 for B and BF2 respectively. Under the same B implant conditions,
with and without PAI implant could result in more than 35nm difference in the
junction depth (at a concentration level of 1×1018cm-3) for as-implanted dopant
10
Depth (Å)
0 100 200 300 400 500 600
Antim
ony
Con
cent
ratio
n (c
m-3
)
1e+17
1e+18
1e+19
1e+20
0 A oxide, simulation20 oxide, simulation50 angstrom oxide, simulation0 angstrom oxide, SIMS20 angstrom oxide, SIMS50 angstrom oxide, SIMS
Fig. 3.1 Comparison between model prediction and SIMS.
Implant condition: Sb, 10keV, 1014cm-2, on-axis.
11
Depth (Å)
0 100 200 300 400 500
Antim
ony
Con
cent
ratio
n (c
m-3
)
1e+17
1e+18
1e+19
1e+200 angstrom oxide, simulation20 angstrom oxide, simulation50 angstrom oxide, simulation0 angstrom oxide, SIMS20 angstrom oxide, SIMS50 angstrom oxide, SIMS
Fig. 3.2 Comparison between model prediction and SIMS.
Implant condition: Sb, 10keV, 6×1013cm-2, on-axis.
12
Depth (Å)
0 100 200 300 400 500 600
Antim
ony
Con
cent
ratio
n (c
m-3
)
1e+17
1e+18
1e+19
1e+200 angstrom oxide, simulation20 angstrom oxide, simulation50 angstrom oxide, simulation0 angstrom oxide, SIMS20 angstrom oxide, SIMS50 angstrom oxide, SIMS
Fig. 3.3 Comparison between model prediction and SIMS.
Implant condition: Sb, 15keV, 3×1013cm-2, on-axis.
13
Depth (Å)
0 200 400 600 800 1000
Boro
n C
once
ntra
tion
(cm
-3)
1e+16
1e+17
1e+18
1e+19
1e+20
no PAI, simulationSb 10keV 1e14, simulationSb 10keV 6e13, simulationSb 15keV 3e13, simulationSb 25keV 9e13, simulationSb 25kev 9e13, SIMSno PAI, SIMSSb 10keV 6e13, SIMSSb 10keV 1e14, SIMSSb 15keV 3e13, SIMS
Fig. 3.4 Comparison between model prediction and SIMS.
Implant condition: B, 3keV, 2×1014cm-2, on-axis, through 2nm
oxide.
14
Depth (Å)
0 100 200 300 400
Boro
n C
once
ntra
tion
(cm-3
)
1e+16
1e+17
1e+18
1e+19
1e+20
Sb 10k 1e14, simulationSb 10k 6e13, simulationSb 15k 3e13, simulationSb 25k 9e13, simulationSb 10k 1e14, SIMSSb 10k 6e13, SIMSSb 15k 3e13, SIMSSb 25k 9e13, SIMS
Fig. 3.5 Comparison between model prediction and SIMS.
Implant condition: BF2, 5keV, 1014cm-2, on-axis, through 2nm
oxide.
15
profiles. This dependence is also well predicted by UT-MARLOWE. After
performing the PAI, the as-implanted dopant profiles are no longer sensitive to
the cap oxide layer thickness even for light species implant such as B, as shown in
Figs. 3.6. This change of sensitivity is predicted very accurately by UT-
MARLOWE as well. Throughout these simulations, the defect profiles of PAI
implant play an important role. Moreover, they are also crucial in modeling the
Transient Enhanced Diffusion (TED) in annealing processes after the implants for
providing as-implanted interstitial and vacancy distributions. The comparison of
the defect profiles between simulation and RBS for different consecutive implant
conditions is shown in Fig. 3.7 and fairly good agreement is observed. Finally, the
Sb implant model based on ZBL pair specific potential is valid for a fairly wide
energy range (up to 100keV), dose range (5×1012cm-2-2×1014cm-2) and for both on
and off-axis implants, as shown in Figs. 3.8-9.
3.4 Conclusion
The Sb implant model has been developed and incorporated into UT-
MARLOWE based on ZBL pair-specific inter-atomic potential. Consecutive
implants of Sb and B or BF2 into SiO2 capped single-crystal Si are simulated and
the impact of PAI on PLDD and the impact of cap oxide layer thickness on PAI
and PLDD are correctly modeled.
16
Depth (Å)
0 200 400 600 800
Boro
n C
once
ntra
tion
(cm
-3)
1e+17
1e+18
1e+19
1e+20
1e+21
0 angstrom, simulation20 angstrom, simulation50 angstrom, simulation0 angstrom, SIMS20 angstrom, SIMS50 angstrom, SIMS
Fig. 3.6 Comparison between model prediction and SIMS.
Implant condition: B, 3keV, 2×1014cm-2, on-axis 10keV 1e14
17
Depth (Å)
0 50 100 150 200 250 300
Amor
phiz
atio
n Fr
actio
n (%
)
0
20
40
60
80
100Sb 15k 3e13, RBSSb 15k 3e13, BF2 5k 1e14, RBSSb 15k 3e13, BF2 5k 2e15, RBSSb 15k 3e13, simulationSb 15k 3e13, BF2 5k 1e14, simulationSb 15k 3e13, BF2 5k 2e15, simulation
Fig. 3.7 Comparison between model prediction and RBS.
Implant condition: on-axis, through 2nm oxide.
18
Depth (Å)
0 200 400 600 800 1000 1200 1400
Antim
ony
Con
cent
ratio
n (c
m-3
)
1e+17
1e+18
1e+19
1e+20
2e14, simulation3e13, simulation5e12, simulation2e14, SIMS3e13, SIMS5e12, SIMS
Fig. 3.8 Comparison between model prediction and SIMS.
Implant condition: Sb, 50keV, on-axis, through 1.6nm oxide.
19
Depth (Å)
0 500 1000 1500 2000
Antim
ony
Con
cent
ratio
n (c
m-3
)
1e+16
1e+17
1e+18
1e+19
1e+20
2e14, SIMS3e13, SIMS5e12, SIMS2e14, simulation3e13, simulation5e12, simulation
Fig. 3.9 Comparison between model prediction and SIMS.
Implant condition: Sb, 100keV, off-axis, through 1.6nm oxide.
20
Chapter 4. Three-dimensional Simulations
A physically-based 3-dimensional Monte-Carlo simulator has been
developed within UT-MARLOWE, which is capable of simulating ion
implantation into multi-material systems and arbitrary topography. Introducing
the third dimension can result in a severe CPU time penalty. In order to minimize
this penalty, a three-dimensional trajectory replication algorithm has been
developed, implemented and verified. More than two orders of magnitude savings
of CPU time have been observed. An unbalanced Octree structure was used to
decompose three-dimensional structures. It effectively simplifies the structure,
offers a good balance between modeling accuracy and computational efficiency,
and allows arbitrary precision of mapping the Octree onto desired structure. Using
the well-established and validated physical models in UT-MARLOWE 5.0, this
simulator has been extensively verified by comparing the integrated one-
dimensional simulation results with SIMS (Secondary Ion Mass Spectroscopy).
Two options, the typical case and the worst scenario, have been selected to
simulate ion implantation into poly-silicon under various scenarios using this
simulator: implantation into a random, amorphous network, and implantation into
the worst-case channeling condition, into (110) orientated wafers.
21
4.1 Introduction
With the continued rapid scaling of modern semiconductor devices,
typically ion implantation is performed into topographically complex structures.
Flat topography implant simulations are no longer sufficient. Users need to have
the flexibility to simulate implants into three-dimensional, realistic structures
[28], which could be either outputs from process simulators, or user-defined
arbitrary structures.
This leads to severe challenges in two aspects, mainly in terms of
computational efficiency. First, as the ions are traveling into the solid, the
simulator needs to determine the material type, whether it is crystalline silicon,
silicon dioxide, or another material, at each specific spatial location along the
ion's trajectory, and choose the corresponding propagator dynamically. Secondly,
and more importantly, the calculation of particle propagation in solids is very time
consuming. This is particularly true for a deterministic Monte Carlo propagator,
required to ensure correct modeling of ion channeling in a crystalline lattice.
Especially, as the simulation volume becomes large for three-dimensional
simulations, in order to obtain the same level of statistical significance for the
overall structure, the number of incident ions at each specific grid cell must still
be kept approximately at the same level as for 1-D and 2-D simulations.
Consequently, the required number of ions should be proportional to the number
22
of grid cells present in the simulation volume. Two or three-dimensional
simulation of ion implantation into a flat or slowly varying topography
intrinsically is a one-dimensional simulation. A small number of grid cells (size is
on the order of a few lattice constants) in the lateral direction is sufficient to
obtain the point source results under rotational periodic boundary conditions. The
results then can be translated and superposed to cover the entire window to give
the final 2/3-D results. This is valid when one or two lateral directions are nearly
invariant. Although a relatively large number of grid cells are needed in the depth
direction (largely energy and species dependent), the total number of grid cells
could be kept reasonably small. The above routine is no longer valid when
performing two or three-dimensional ion implantation simulations into sharply
varying structures. For such cases, a large number of grid cells in both the depth
and the lateral directions have to be used. For example, to simulate a 130 nm gate
length MOSFET, the length on both lateral directions is on the order of 250 lattice
constants. This means, that to maintain the same level of statistical significance,
for each additional dimension, the required number of ions for Monte-Carlo
simulation should increase by about two orders of magnitude.
Hossinger et al has demonstrated three-dimensional capability in Monte-
Carlo ion implantation simulator, MCIMPL [29], and parallelization was
proposed to reduce the simulation time when a cluster of workstations is readily
23
available [30]. Furthermore, an interesting computational efficient combined
approach was proposed by utilizing the physically-generated 3-D point response
distribution functions [31]. Especially, in two-dimensional simulations, the lateral
trajectory replication technique has proven quite successful in enhancing the
computational efficiency [32]. Yet more aggressive, efficient CPU time reduction
techniques are needed for three-dimensional Monte-Carlo simulations of ion
implantation into highly non-uniform structures.
Although not adequate to be applied to three-dimensional simulation
directly, the lateral trajectory replication scheme provides good insight into the
problem. The key is to cleverly reuse the physically-computed trajectories for
similar topography. Interestingly enough, reusing the physically-generated
trajectories is the main idea behind most of the reported two and three-
dimensional analytical simulators as well [9,31,33-35]. While the success of this
approach when applied to simple surface topographies is clear, the validity of the
analytical approximations when applied to sharply varying structures remains
unclear. Thus, a computationally-efficient Monte-Carlo simulator with a modified
replication scheme would appear to be the ideal solution for such structures.
4.2 New 3-D Trajectory Replication Scheme
Typically, for realistic structures slowly varying or flat topography often
exists over quite a significant percentage of the total simulation volume. In such
24
regions, the lateral variation of both dopant and defect profiles is small and
negligible. Therefore, in such regions, the physical calculation of a representative
small portion would be sufficient after the statistical criteria in terms of material
and structural homogeneity have been satisfied. In other words, each physically-
computed individual trajectory of that small portion might be translated efficiently
laterally given that the topography is similar, and this needs to be done
dynamically. While replication proceeds as often as possible in the slowly varying
portion of the structure, the ions' behavior is physically and individually
computed in the rapidly varying parts. The new replication scheme is based
precisely on this idea, based on the observation that, typically, similar or identical
topography exists in nearby segments.
Laterally, the new replication scheme is shown in Fig. 4.1, which is
similar to the lateral trajectory replication scheme discussed in 2-D simulations
[29]. The difference is that here it tests the whole neighboring area for replication,
in an effort to achieve the maximum savings on CPU time. While in 2-D, the
starting point and replication directions are randomly decided and the replication
number is set to around 10 in order to maintain some variation of the initial
conditions, in 3-D, replication starts by physically calculating all the required
scattering events and the stopping mechanisms experienced by one particular ion
25
Fig. 4.1 Side view of the replication algorithm proceeding from left to right.
The light trajectories are physically computed, the gray ones show
where replication error occurs, and the black ones are replicated
trajectories.
26
until it stops. After the ion's trajectory is obtained, its relative displacements
between consecutive scattering events are recorded, as well as the states of the ion
before every scattering, the material type and damage situation along the ion's
trajectory. The algorithm then introduces an offset and replicates the model
trajectory at the new incident position. As the line segments are being copied
along the trajectory, replication error is examined according to two factors,
whether the material in the model trajectory matches that in the target trajectory
and whether the damage level is comparable between them. The above scheme
can be somewhat relaxed to achieve some tradeoff between computational
efficiency and accuracy, in an effort to achieve better CPU time savings and
accommodate the replication scheme to the slow varying topography commonly
encountered in realistic structures. Specifically, the total material mismatch
distance is calculated, and the damage levels are categorized logarithmically into
five different levels. The criteria are set that the total mismatch distance should
not exceed two lattice constants and the damage level should belong to the same
category.
In 3-D, the goal is to take maximum advantage of similar topography in
two lateral directions through replication. The new replication scheme proceeds as
shown in the top view in Fig. 4.2. The implant window is divided into small
rectangles or squares. The first ion is set to implant into the left bottom corner
27
Y (Depth direction)
Z (lateral direction)
X (lateral direction)
Fig. 4.2 Top view of the simulation domain illustrating the replication
algorithm. The hollow circles are physically computed, the gray ones
show that replication errors are encountered and trajectories are
therefore physically computed, and the black ones are replicated. The
gray circles could also be represented by the hollow ones since these
trajectories need to be physically calculated as well. Replication starts
from the left bottom corner and proceeds up and to the right of the
original trajectory.
28
segment and the trajectory is physically-computed and recorded. The initial
relative position within the segment is determined randomly and used for every
rectangle until the whole window is covered. Replication first goes up and
continues moving up until a replication error is detected. Replication stops when
an error occurs in the z direction, and the place (Zstop, the rectangle number) is
recorded. Then, replication proceeds to the right of the original (physically-
computed) trajectory. When an error occurs in the x direction at the new position,
the trajectory is recomputed and the model trajectory is replaced with a newly
calculated one. Otherwise, the model trajectory is simply copied to the new
position along the x direction. After finishing the operation either way, replication
goes up and the above procedure is repeated until replication reaches the rectangle
at the right end and an error is detected in the z direction. Then, replication starts
from the left-most column with minimum Zstop, and continues until Zstop of every
column equals to Zmax (total row number). Thus, the simulator is able to
intelligently reuse physically-calculated trajectories in two directions. Since the
CPU time required to replicate is much less than the time required for the actual
physical computation, large savings are expected with the application of the new
replication scheme without sacrificing accuracy, as shown in Figs. 4.3-4.
29
Implant
Fig. 4.3 (a) Three-dimensional simple trench topography viewing from
different angle for replication scheme demonstration. The length of
the whole volume is 800 lattice constants, and the depth is 700
lattice constants.
30
Fig. 4.3 (b) Simulating a boron 5keV implant into the structure shown in Fig.
4.3(a), using a tilt angle of 7°, a rotation angle of 30°, and a dose of
1×1014cm-2. The CPU time is 72 minutes using 5,000,000 ions. The
iso-concentration contours range from 1×1015cm-3 to 1×1017cm-3
viewing from different angle. The estimated CPU time without
replication is over 20 hours.
31
implant
STI
S/D Gate
S/D
Gate
S/D Gate
Fig. 4.4 (a) Source/Drain structure with shallow trench isolation. The side
length is 600 lattice constants.
32
Fig. 4.4 (b) Simulating an arsenic 15 keV implant into the structure show in
Fig. 4.4 (a) using a tilt angle of 0°, a rotation angle of 0°, and a
dose of 1×1013cm-2. The CPU time is 88 minutes, using 5,000,000
simulated ions. The iso-concentration contours range from
1×1016cm-3 to 1×1019cm-3. The estimated CPU time without
replication is over 20 hours.
33
4.3 Structure Decomposition
As the ions propagate through the structure, the determination of the
material type is crucial and could be very CPU time consuming. This is
particularly true for highly non-uniform structures. Since an unbalanced Quadtree
proved to be quite successful in two-dimensional simulations [29, 36], an
unbalanced Octree becomes a natural choice for 3-dimensional simulations and is
adopted in this new simulator. Similar to the corresponding Quadtree
decomposition, Octree also recursively subdivides space until the desired
resolution is reached. Primarily, the Octree structure offers three major
advantages over structured and unstructured meshes. First, it breaks down the
solid into a set of regular cubic volumes of different sizes and simplifies the
structure, i.e., determines the intersections of the ion's trajectory and material
boundary and tests whether the ion is inside a particular cell, in an elegant and
efficient manner. Secondly, the resulting grid is coarse where the material is
uniform, and fine near the material interfaces. This results in great saving in
storage and CPU time. Finally, the tree data structure makes the tracking more
efficient (proportionally to logN, instead of N), and reduces the zoning time
accordingly. Generally, the Octree offers a good balance between accuracy,
memory efficiency and speed.
34
4.4 Model Verification
The physical models used by the new simulator (including nuclear
scattering, local and nonlocal electronic stopping models and modified Kinchin-
Pease models) are based on UT-MARLOWE [18, 37]. Specifically, BF2 implant
was treated as a molecular cluster. In other words, the trajectories of both boron
and fluorine were physically calculated and the damage generated by both species
was simulated [37]. The models have been extensively verified in 1-D within UT-
MARLOWE. However, verification is still necessary in this 3-D simulator, as the
models are implemented in a quite different way. Results are shown in Figs. 4.5-9
and good agreement with experimental SIMS data are obtained.
In this simulator, two models have been provided to simulate ion implantation
into poly-silicon since the grain size and the grain orientation of poly-silicon are
largely dependent on the process flow and hard to analyze. The first model, as the
typical case, uses amorphous silicon to represent poly-silicon and applies a
corresponding stochastic Monte Carlo approach. The second one, which is the
worst- case channeling scenario, uses a deterministic Monte Carlo model and
simulates implant into a (110) wafer. Implantation into a (100) wafer with both tilt
angle and rotation angle set to 45 degrees is equivalent to implantation into (110)
wafer with both tilt angle and rotation angle set to 0 degrees, because the silicon
matrix is identical. Therefore, as shown in Fig. 4.10, after obtaining a point source
35
Depth [µm]
0.0 0.1 0.2 0.3 0.4 0.5
Boro
n C
once
ntra
tion
[cm
-3]
1e+16
1e+17
1e+18
1e+19
1e+20
1e+21
SIMS Simulation
Fig. 4.5 Profiles of 35 keV boron on-axis implants into a (100) Si wafer
with doses of 1×1013cm-2, 3×1013cm-2, 1×1014cm-2, 5×1014cm-2,
2×1015cm-2, and 8×1015cm-2. The gray lines are the simulation
results, while the black lines are SIMS data.
36
Depth [µm]0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40
As C
once
ntra
tion[
cm-3
]
1e+16
1e+17
1e+18
1e+19
1e+20
1e+21
1e+22
SIMS Simulation
Fig. 4.6 Profiles of 50 keV As on-axis implants into a (100) Si wafer with
doses of 1×1013cm-2, 3×1013cm-2, 1×1014cm-2, 5×10-14cm-2,8×10-
15cm-2. The gray lines are the simulation results, while the black
lines are SIMS data.
37
Depth [µm]
0.00 0.02 0.04 0.06 0.08 0.10
Boro
n C
once
ntra
tion
[cm
-3]
1e+16
1e+17
1e+18
1e+19
1e+20
1e+21 SIMS Simulation
Fig. 4.7 Profiles of 15 keV BF2 on-axis implants into a (100) Si wafer with
doses of 1×1013cm-2, 3×1013cm-2, 1×1014cm-2, 5×10-14cm-2, 2×10-
15cm-2, and 8×10-15cm-2. The gray lines are the simulation results,
while the black lines are SIMS data.
38
Depth [µm]
0.000 0.025 0.050 0.075 0.100 0.125 0.150
Boro
n C
once
ntra
tion
[cm
-3]
1e+17
1e+18
1e+19
1e+20
1e+21 SIMS Simulation
Fig. 4.8 Profiles of 35 keV BF2 on-axis implants into a (100) Si wafer with
doses of 1×1013cm-2, 3×1013cm-2, 1×1014cm-2, 5×10-14cm-2, 2×10-
15cm-2, and 8×10-15cm-2. The gray lines are the simulation results,
while the black lines are SIMS data.
39
Depth [µm]
0.00 0.05 0.10 0.15 0.20 0.25 0.30
Phos
phor
us C
once
ntra
tion
[cm
-3]
1e+17
1e+18
1e+19
1e+20
1e+21
SIMS Simulation
Fig. 4.9 Profiles of 30 keV phosphorus on-axis implants into a (100) Si
wafer with doses of 1×1013cm-2, 1×1014cm-2, 5×10-14cm-2, and
2×10-15cm-2. The gray lines are the simulation results, while the
black lines are SIMS data.
40
Integration Direction
Rotation 45°
Fig. 4.10 Figure showing that by rotating the simulation results (tilt
0°/rotation 0°) in a (110) Si wafer by 45° and integrating along the
new depth dimension, the result is equivalent to implantation into a
(100) Si wafer (tilt 45°/rotation 45°).
41
profile in a (110) wafer, we can rotate it by 45 degree and get the 1-dimensional
profile by integrating it according to the new depth axis. Thus, instead of using
actual (110) wafer pieces [38], we can compare the converted 1-D profile with
SIMS data for the same implant condition into a (100) wafer with both tilt and
rotation set to 45 degrees. As shown in Figs. 4.11-12, for B and As, the SIMS data
and the converted simulation results match very well for several different
energies.
In addition, As channeling through a three-layered structure, polysilicon,
gate oxide and crystalline silicon was studied as shown in Fig. 4-13. Major
discrepancies were observed when comparing to simulation results for the same
implant conditions in [38]. At first, in [38], the dopant concentration in
polysilicon falls to or below 1×1018cm-3 at 100nm depth for all energies below
30keV. Our study shows that the concentrations remain above 2×1019cm-3 at a
depth of 100nm except for the 1keV case. Secondly, the peak concentration in
crystalline silicon is well below 1×1019cm-3 in [38], and it reaches 1020cm-3 for
energies above 3 keV according to our simulation. Finally, and most importantly,
in our result, the channeling of 15keV and 30keV implant continues in crystalline
silicon and goes well beyond 200nm and even to 400nm for 30 keV implants.
According to [38], the profile stops in crystalline silicon rather sharply (before
42
Depth [µm]
0.0 0.1 0.2 0.3 0.4 0.5
Boro
n C
once
ntra
tion
[cm
-3]
1e+15
1e+16
1e+17
1e+18
1e+19
Gray line: Simulation result converted to an implant into a (100) wafer, 45/45 (tilt/rotation). Black line: SIMS data of implant into a (100) wafer, 45/45 (tilt/rotation).
Simulation result of implant into (110) wafer, 0/0.
Fig. 4.11 (a) Profiles of a 15keV B on-axis implant into a (110) Si wafer with a
dose of 1×1013cm-2 to simulate the worst case channeling scenario
of an implant into poly-silicon.
43
Depth [µm]
0.0 0.2 0.4 0.6 0.8
Bor
on C
once
ntra
tion
[cm-3]
1e+15
1e+16
1e+17
1e+18
Simulation result of implant (0/0) into (110) wafer
Gray line: Simulation result converted to an implant into a (100) wafer, 45/45 (tilt/rotation). Black line: SIMS data of implant into a (100) wafer, 45/45 (tilt/rotation).
Fig. 4.11 (b) Profiles of a 35keV B on-axis implant into a (110) Si wafer with a
dose of 1×1013cm-2 to simulate the worst case channeling scenario
of an implant into poly-silicon.
44
Depth (µm)
0.0 0.1 0.2 0.3 0.4 0.5 0.6
Ars
enic
Con
cent
ratio
n (c
m-3)
1e+15
1e+16
1e+17
1e+18Gray line: Simulation result converted to an implant into a (100) wafer, 45/45 (tilt/rotation). Black line: SIMS data of implant into a (100) wafer, 45/45 (tilt/rotation).
Fig. 4.12 (a) Profiles of a 15keV As on-axis implant into a (110) Si wafer with a
dose of 2×1012cm-2 to simulate the worst case channeling scenario
of an implant into poly-silicon.
45
Depth (µm)
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Ars
enic
Con
cent
ratio
n (c
m-3 )
1e+15
1e+16
1e+17
Gray line: Simulation result converted to an implant into a (100) wafer, 45/45 (tilt/rotation). Black line: SIMS data of implant into a (100) wafer, 45/45 (tilt/rotation).
Fig. 4.12 (b) Profiles of a 50keV As on-axis implant into a (110) Si wafer with a
dose of 2×1012cm-2 to simulate the worst case channeling scenario
of an implant into poly-silicon.
46
Depth (µm)
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Arse
nic
Con
cent
ratio
n (c
m-3
)
1e+15
1e+16
1e+17
Gray line: Simulation result converted to an implant into a (100) wafer, 45/45 (tilt/rotation). Black line: SIMS data of implant into a (100) wafer, 45/45 (tilt/rotation).
Fig. 4.12 (c) Profiles of a 100keV As on-axis implant into a (110) Si wafer with
a dose of 2×1012cm-2 to simulate the worst case channeling
scenario of an implant into poly-silicon.
47
Depth (µm)
50 100 150 200 250 300 350 400
Arse
nic
Con
cent
ratio
n (c
m-3
)
1e+17
1e+18
1e+19
1e+20
1e+21
1 keV3 keV8 keV15 keV30 keV
Fig. 4.13 Profiles of As on-axis implant into a three-layered structure
(100nm polysilicon, 2.7nm gate oxide and 700nm crystalline
silicon) at 1, 3, 8, 15, and 30 keV with a dose of 1×1015cm-2.
48
150nm). These discrepancies are possibly due to different As stopping models and
calibrations in the two simulators.
4.5 Example application
With unbalanced Octree decomposition scheme, in principle, implantation
into any arbitrary structures can be simulated. Recent studies show that double-
gate device structures such as FinFET [7] have superior scaling potentials over
traditional bulk CMOS or SOI CMOS, and are chosen here (shown in Fig. 4.14-
15). The highly non-planar and non-uniform nature typically makes the
application of analytical model insufficient.
4.6 Conclusion
A new computational efficient three-dimensional Monte Carlo simulator
has been developed within UT-MARLOWE. An unbalanced Octree algorithm
was used for spatial decomposition. A new trajectory replication scheme was
developed and implemented in order to enhance computational efficiency. More
than two orders of magnitude savings on CPU time have been observed.
49
Source/Drain
Gate Electrode
Gate Oxide
Channel Width
Structure decomposition by Octree scheme
Fig. 4.14 FinFET structure decomposed by unbalanced Octree.
50
Source/Drain
Gate Electrode
Fig. 4.15 Implant into FinFET structure and the dopant distribution.
51
Chapter 5. Analytical 1-D and 2-D Ion Implantation Models Based on
Legendre Polynomials
Computationally-efficient ion implantation modeling has become the
essential tool for efficient and accurate CMOS design as aggressive scaling of
devices continues. Specifically, computationally-efficient two-dimensional
analytical models are often more attractive than physically-based Monte Carlo
simulations since the latter are expensive in terms of computational time. Here we
present new computational-efficient analytical models to simulate 1-D and 2-D
impurity and damage profiles. Legendre polynomials are used as basis functions
in view of their orthogonality and good interpolation property. Conventional
superposition approaches for 2-D implant modeling are explained and the
shortcomings are analyzed. A dose splitting approach is incorporated in the new
2-D model to account for the nonlinear de-channeling effect as implantation-
induced damage accumulates. Good agreement with a physically-based and
experimentally verified Monte Carlo simulator (UT-MARLOWE with TOMCAT)
has been obtained for both impurity and damage profiles with a 50× reduction of
computational time for medium energy implants.
52
5.1 Introduction
As silicon MOS devices scale to the deep sub-micron regime, ion
implantation will continue to be an important method to selectively introduce
impurity atoms into the silicon substrate. Accurate and computationally-efficient
ion implantation models with detailed dependence on various implant conditions
are desirable for CMOS design and technology development. This becomes
especially important with increased process complexity and fabrication cost.
Particularly, a detailed knowledge of both impurity and damage profiles is
required to correctly model dopant diffusion during annealing and thermal
treatment steps after the implantation.
At present time, two major ion implantation modeling approaches are
widely used, each with their advantages and disadvantages. The first one uses the
Monte Carlo method [15,32,39,40] to simulate the transport of ions in silicon and
other material of interest, such as SiO2, Si3N4, and silicide. The Monte Carlo
approach can provide detailed understanding of the implant process, in particular,
ion-solid interactions and the damage generation. The major shortcoming is its
rather large computational burden due to the necessity of computing every
collision to simulate the trajectory of each individual ion. Especially, in an effort
to reduce the statistical fluctuations, which are intrinsic to the Monte Carlo
method, a large number of ions are required. Things get even worse for high-
53
energy implant simulations, and it typically takes hours and even days for MeV
implant simulations. On the other hand, the second approach, analytical modeling
[41,42,43], is computational efficient. Basically, it uses parameterized functions
to model the as-implanted profiles. This way, it replaces the calculations of every
collision by a pre-calculated look-up table of model parameters and evaluating a
few function values at different grid points. Typically it saves CPU time by a few
orders of magnitude over the Monte Carlo method, and it is the preferred
approach of most process simulators. The apparent drawback of the analytical
models is the limited range that current look-up tables cover. Within the
experimentally validated and calibrated range, however, analytical modeling is a
powerful tool for technology development.
5.2 One-dimensional Modeling
Several analytical modeling techniques have been proposed during the last few
decades to simulate the as-implanted one-dimensional impurity profiles. Among
those, the dual-Pearson model [41] is quite successful and very widely used. With
two Pearson functions to separately model the random scattering and channeling
components, it can accurately depict a wide range of as-implanted profiles.
However, the parameter extraction procedure in dual-Pearson model is
complicated and a good initial guess is required for the convergence of the
algorithm. Although software has been written to automatically generate a
54
reasonable initial guess, the parameter set may not be the unique one that provides
the best fit for the particular impurity profile.
Generally, the requirements for a good analytical implant model are the
following:
1. Accurate description of the as-implanted impurity and damage profiles.
2. Good interpolation properties.
3. Easy parameter extraction procedures.
When considering all these requirements, Legendre polynomials are a
good choice. They have a simple functional form, and can model both the
impurity and damage profiles very well over a wide range of parameters [9].
Specifically, the 1-D impurity and damage profiles are modeled as a linear
combination of 14 Legendre polynomials,
∑=
⋅⋅=13
0
15 )))('(exp(10)(i
ii xxLaxC , , (5.1) 21' bxbx +=
where, C(x) is the impurity concentration at depth x, Li(x') is the Legendre
polynomial of degree i, b1, b2 are constants that map x from the interval [xmin,
xmax] to [-1,1], and xmin and xmax are the depths between which the impurity and
damage concentration falls within the range of interest.
Legendre polynomials satisfies orthogonality with a unity weight function in the
interval [-1,1], i.e.,
55
(∫−
=1
1
,)()( jidxxLxL ji δ ) (5.2)
Thus, the 16 parameters describing the profiles can be determined
uniquely as follows:
∫−
⋅⋅=1
1
')'()'( dxxLxCai (5.3)
maxmin1
2xx
b+
= (5.4)
maxmin
maxmin2 xx
xxb
+−
= (5.5)
Although a relatively large number of coefficients are required, the
simplicity of the parameter extraction procedure and its good interpolation
property justifies the extra memory requirement.
UT-MARLOWE, a Monte Carlo simulator, was used to generate the as-
implanted impurity and damage profiles. This code has been validated over a
wide range of implant conditions and species. The output from UT-MARLOWE
was first averaged to get a smoother profile. The resulting profile is then fitted
using equations (3)-(5) and the derived parameters are stored in the look-up table.
The comparison between UT-MARLOWE result and fitted result are shown in
Figs. 5.1-9 for various implant conditions. Excellent agreement has been
56
Fig. 5.1 Comparison of the Legendre polynomial analytic model predictions
with boron impurity profiles obtained from UT-MARLOWE
simulations (histogram) for boron off-axis implant at various
energies and doses through 15 Å native oxide.
57
Fig. 5.2 Comparison of the Legendre polynomial analytic model predictions
with phosphorus impurity profiles obtained from UT-MARLOWE
simulations (histogram) for phosphorus off-axis implant at various
energies and doses through 15 Å native oxide.
58
Fig. 5.3 Illustration of the excellent interpolation property of the Legendre
coefficients. The histogram plots are UT-MARLOWE simulation
results and the solid lines are analytic fits. The profiles with the
following doses are actual fits to simulation results: 1X1013 cm-2,
5X1013 cm-2, 5X1014cm-2. The rest are results of interpolation
between model parameters at the above-mentioned doses. The BF2
implants are on-axis and through 15 Å native oxide.
59
Fig. 5.4 Illustration of the excellent interpolation property of the Legendre
coefficients. The histogram plots are UT-MARLOWE simulation
results and the solid lines are analytic fits. The profiles with 100 and
180 keV implant energies are actual fits to simulation results. The
profile with 140 keV implant energy is the result of interpolation
between model parameters at the above-mentioned energies. The
arsenic implants are on-axis and through 275 Å silicon nitride. The
dose is 1×1013cm-2.
60
Fig. 5.5 Illustration of the excellent interpolation property of the Legendre
coefficients. The histogram plots are UT-MARLOWE simulation
results and the solid lines are analytic fits. The profiles with the
following doses are actual fits to simulation results: 1X1013 cm-2,
1X1014 cm-2, 5X1014cm-2. The rest are results of interpolation
between model parameters at the above-mentioned doses. The 35
keV arsenic implants are on-axis and through 150 Å silicon nitride.
61
Fig. 5.6 Illustration of the excellent interpolation property of the Legendre
coefficients. The histogram plots are UT-MARLOWE simulation
results and the solid lines are analytic fits. All of the profiles are
interpolations between model parameters between 50 and 150 Å
silicon nitride, 150 and 275 Å silicon nitride, and 275 and 400 Å
silicon nitride. The 35 keV arsenic implants are on-axis and the dose
is 1X1013 cm-2.
62
Fig. 5.7 Illustration of the excellent interpolation property of the Legendre
coefficients. The histogram plots are UT-MARLOWE simulation
results and the solid lines are analytic fits. The profiles with the
following energies are actual fits to simulation results: 20 keV, 40
keV, and 60 keV. The rest are results of interpolation between model
parameters at the above-mentioned energies. The phosphorus
implants are on-axis and through 150 Å titanium silicide.
63
Fig. 5.8 Illustration of the excellent interpolation property of the Legendre
coefficients. The histogram plots are UT-MARLOWE simulation
results and the solid lines are analytic fits. The profiles with the
following doses are actual fits to simulation results: 1X1013 cm-2,
1X1014 cm-2, 5X1014cm-2. The rest are results of interpolation
between model parameters at the above-mentioned energies. The 20
keV phosphorus implants are on-axis and through 150 Å titanium
silicide.
64
Fig. 5.9 Illustration of the excellent interpolation property of the Legendre
coefficients. The histogram plots are UT-MARLOWE simulation
results and the solid lines are analytic fits. All of the profiles are
interpolations between model parameters between 50 and 150 Å
titanium silicide, 150 and 275 Å titanium silicide, and 275 and 400 Å
titanium silicide. The 20 keV phosphorus implants are on-axis and
the dose is 1X1013 cm-2.
65
achieved when fourteen Legendre polynomials are used to model the impurity and
damage profiles.
The quality of interpolation is another crucial property for an analytical
model in order to correctly predict the impurity profiles for arbitrary implant
conditions within a specified range. This is true for all analytical models mainly
because only a finite number of parameter sets are available for in any range. For
Legendre polynomials, it is found that square root interpolation is sufficient to
obtain good results when interpolating over energy, tilt angle, rotation angle, the
screen oxide, nitride or silicide thickness. To interpolate between two doses, a
logrithmatic interpolation is used and the resulting profile is re-scaled in order to
ensure the correct dose. When good agreement is reached, a finer sampling of
parameter sets may not be necessary; otherwise, an additional parameter set
should be added to the lookup table. Typical interpolation results are shown in
Figs. 5.3, 5.6 and 5.9.
5.3 Two-dimensional Modeling
With the shrinking of device dimensions, two-dimensional effects, such as
the implant profiles around a mask edge (considered impenetrable), become more
and more significant for CMOS device design. To correctly account for these
effects, accurate and computationally efficient predictions of two-dimensional
66
impurity and damage profiles are desirable in an effort to minimize the cost and
cycle time for the development of process technology.
Traditionally, computationally-efficient analytical modeling of two-
dimensional ion implantation has been performed primarily by a superposition
approach [42,43]. Specifically, the point-source profile (the impurity profile over
a very small window) is generated first by a Monte Carlo method. Since the
implant window is small, i.e. 5a long (a is the silicon crystal lattice constant,
=0.543 nm), a relatively small number of ions is sufficient for a statistically
satisfactory result. The point-source profile is then translated to cover the entire
implant window and superposed. In other words, the implant window is broken
into segments (~5a long implant window), the implant into each segment is
simulated using the point-source profile obtained by Monte Carlo method and all
the point-source profiles are superposed. This has proved quite successful when
the implant-induced damage is negligible (low dose and light species). However,
as the damage-induced dechanneling effects become more significant (high dose
and heavy species), the profile predicted by the superposition approach shows
excessive channeling compared with Monte Carlo predictions (Fig. 5.10). Besides
its failure to take into account the nonlinear damage effect within the segment, the
interference from the lateral straggle of neighboring segments also has a
remarkable impact on the channeling (Fig. 5.11).
67
Fig. 5.10 Comparison of 2-D impurity iso-concentration contours generated
using UT-MARLOWE with TOMCAT, with the conventional
superposition approach. Implant condition: species - arsenic , energy
- 50 keV, dose - 5×1014cm-2, tilt angle - 7°, rotation angle - 30°,
implant window size – 50 nm.
68
Fig. 5.11 Typical point-source damage profile, showing how the lateral
straggle extends far beyond the width of the implant window, which
is 5a (a=0.543 nm, silicon lattice constant). Implant conditions:
species - BF2, energy - 30keV, dose - 2×1012 cm-2, tilt angle - 2°,
rotation angle - 0°.
69
In order to overcome these shortcomings, we propose a modified
superposition approach [9], based on UT-MARLOWE with TOMCAT
(TOpography based Monte CArlo Transport) [32]. The basic idea is to split the
nonlinear problem (dechanneling effect caused by the damage build-up of a high
dose implantation) into several near linear problems (relatively minor damage
changes within small dose intervals), while at the same time taking into account the
interference from neighboring segments. Specifically, the given implant dose D0 is
divided into a number of steps Dstep. Dstep is small enough so that within one
interval the linearity, together with the superposition approach, is valid (i.e.
2×1012/cm2 is used for As, BF2 and P). To simulate high dose implants, instead of
using only one point source profile throughout the process, impurity and damage
profiles are selected from a set of profiles in the look-up table (see next section),
and sorted according to the damage present near the particular segment.
Empirically, the damage near a segment can be quantitatively calculated as follows:
∫ ∫+
−
⋅⋅=max
0
),(),(d x
xii
i
i
dxzxCxGdzDσ
σ
σ , (5.6)
where C(x,z) = damage concentration (i.e. number of interstitials/vacancies per
unit volume as a function of lateral and vertical position); G(xi, ) = Gaussian
weighting function with mean x
σ
i and standard deviation σ; where σ is an
empirical number representing the lateral spread of damage, and is a function of
70
the implant energy. The above Gaussian function is centered on the segment of
interest, which corresponds to maximum weighting, and dies out gradually as the
lateral distance increases. In order to characterize the asymmetry of the damage
profile, the damage distribution (Ri) is introduced to describe the ratio of the
integrated damage to the right and left of the segment. This parameter is crucial
for tilted implants, as the damage information may not be adequate to distinguish
the right and left corner of the implant window.
Based on the damage estimation Di and damage distribution Ri,
corresponding impurity and damage profiles are chosen from the look-up table
and superposed. In this way, the impurity superposition can be monitored
according to the damage accumulation through the implantation process. This
modified superposition approach provides the damage profiles as well, which is
essential for multiple implants and has a significant impact on the diffusion of
impurities during subsequent thermal treatment, especially under minimal thermal
budget conditions.
5.4 Two-D Look-up Table Generation and Simulation Results
Legendre polynomials are used to model the point-source profiles as
mentioned above because of their good interpolation property and orthogonality.
Similar to 1-D simulation, a simple integration can be used to extract the
71
Legendre parameters, which makes the initial guess and recursive looping (e.g.
Levenberg-Marquardt algorithm) unnecessary.
Primarily, the point source profiles are modeled as follows:
)))('(.)(exp(10),(13
0
15 xxLzazxC ii
i∑=
⋅= (5.7)
∑=
=9
0))('(.)(
jjiji zzLbza (5.8)
where x and z are the lateral and depth coordinates, respectively, Lj is the jth order
Legendre polynomial, and ai and bij are Legendre parameters to be extracted from
point-source profiles. As shown in equation (5.7-8), 10 Legendre polynomials are
used to model the point-source response profile in the depth direction, while the
number for lateral direction is 14.
To generate the look-up table, first an implantation into a small window
with Dstep is simulated using the Monte Carlo approach. The generated impurity
and damage profiles are extracted to obtain corresponding Legendre parameters
and then superposed over the entire implant window. Based on the superposed
damage information, the second implantation into the small window is simulated
using the Monte Carlo method again, and the impurity and damage profiles are
extracted and superposed similarly. The above processes are repeated until a fairly
high dose level is reached. It is not necessary to store the parameters for every
implant step. As the dose increases, the damage caused by a single implant step
72
has less effect on the implants thereafter. This can be approximated quantitatively
by the average ion penetration depth of each step. As shown in Fig. 5.12 (with a
small number of ions), the average penetration depth of the ions decrease with
implant dose, and ultimately saturates at a value which corresponds to the
amorphization of silicon and the shutdown of the crystalline channels. The
variation of the penetration depth is quantified to determine the levels where the
accumulated damage effect is significant. Only at these dose levels are Monte
Carlo simulations performed with a large number of ions to achieve statistically
satisfactory results. In this way, the sizes of the look-up table are kept reasonable.
Then the parameters are extracted and put into the look-up tables with the
corresponding index of damage estimation and damage distribution.
Since there is no direct reliable experimental method to measure the 2-D as-
implanted impurity and damage profiles yet, a validated Monte Carlo simulator
(UT-MARLOWE with TOMCAT) is used to verify this new analytical approach
for 2-D ion implant simulation. Good agreement has been obtained between
impurity and interstitial profiles generated using the new nonlinear superposition
approach and those obtained using UT-MARLOWE with TOMCAT for As (Fig.
5.13), P (Fig. 5.14) and BF2 (Fig. 5.15). It is evident that Legendre polynomials
are good basis functions for modeling 2-D impurity and damage profiles with
significant savings in CPU time, without compromising accuracy.
73
Fig. 5.12 The dependence of the average ion penetration depth on as-
implanted dose. The curve shows the impact of the as-implanted
damage on subsequent implantation and the tendency to approach
amorphization as the dose increases. Implant conditions: species -
BF2, energy - 50keV, tilt angle - 0°, rotation angle - 0°, implant
window size - 5a (a=0.543 nm, silicon lattice constant).
74
Fig. 5.13 Comparison of 2-D arsenic impurity iso-concentration contours
generated using UT-MARLOWE with TOMCAT, with those
obtained analytically by the new superposition approach. A dose
step of 2×1012cm-2 was used. Implant conditions: species - arsenic,
energy - 50keV, dose - 5×1014cm-2, tilt angle - 7°, rotation angle -
30°, implant window size – 50 nm.
CPU time: 7 hrs 40 mins (TOMCAT);
6 mins (modified superposition)
75
Fig. 5.14 (a) Comparison of 2-D phosphorus impurity profile generated using
UT-MARLOWE with TOMCAT, with those obtained analytically
using the modified superposition approach. A dose step of 2×1012
cm-2 was used. Iso-concentration contours between 1×1017 cm-3
and 1×1019 cm-3 for the impurity profile is shown.
Implant conditions: species - phosphorus, energy - 30keV, dose - 1×1014cm-2, tilt
angle - 7°, rotation angle - 0°, implant window size – 50 nm.
CPU time for Monte-Carlo simulation (400k ions): 5 hrs 13 mins
CPU time for analytical model: 8 mins
76
Fig. 5.14 (b) Comparison of 2-D phosphorus interstitial profiles generated using
UT-MARLOWE with TOMCAT, with those obtained analytically
using the modified superposition approach. Iso-concentration
contours between 3×1018 cm-3 and 1×1021 cm-3 for the interstitial
profiles is shown.
Implant conditions: species - phosphorus, energy - 30keV, dose - 1×1014cm-2, tilt
angle - 7°, rotation angle - 0°, implant window size – 50 nm.
CPU time for Monte-Carlo simulation (400k ions): 5 hrs 13 mins
CPU time for analytical model: 8 mins
77
Fig. 5.15 (a) Comparison of 2-D boron impurity profile generated using UT-
MARLOWE with TOMCAT, with those obtained analytically
using the modified superposition approach. Iso-concentration
contours between 1×1016 cm-3 and 1×1019 cm-3 for the impurity
profile is shown.
Implant conditions: species - BF2, energy - 30keV, dose - 1×1014cm-2, tilt angle -
2°, rotation angle - 0°, implant window size – 50 nm.
CPU time for Monte-Carlo simulation (400k ions): 8 hrs 9 mins
CPU time for analytical model: 4 mins
78
Fig. 5.15 (b) Comparison of 2-D boron interstitial profiles generated using UT-
MARLOWE with TOMCAT, with those obtained analytically
using the modified superposition approach. Iso-concentration
contours between 1×1016 cm-3 and 3×1020 cm-3 for the interstitial
profiles is shown.
Implant conditions: species - BF2, energy - 30keV, dose - 1×1014cm-2, tilt angle -
2°, rotation angle - 0°, implant window size – 50 nm.
CPU time for Monte-Carlo simulation (400k ions): 8 hrs 9 mins
CPU time for analytical model: 4 mins
79
5.5 Conclusion
A new analytical model has been developed based on Legendre
polynomials to simulate ion implantation in both one and two dimensions. It
greatly simplifies the parameter extraction procedure and can model both the
impurity and the damage profiles. In two-dimensional simulations, a dose-
splitting approach has been developed to account for the nonlinear effect caused
by damage accumulation. It has been shown that the results obtained from the
new model agree very well with experimentally verified Monte Carlo simulation
results with major savings of CPU time.
80
Chapter 6. Conclusions and Future Work
6.1 Conclusions
In this study, physical stopping models and the damage models were first
reviewed. In order to correctly model ultra-shallow junction formation, a heavy
species, antimony, implant model was developed and incorporated into UT-
MARLOWE. ZBL pair-specific inter-atomic potential was used, which is more
accurate than the widely cited ZBL universal inter-atomic potential. Furthermore,
consecutive implants of antimony and B or BF2 into SiO2 capped single-crystal Si
were simulated and the impact of PAI on PLDD and the impact of cap oxide layer
thickness on PAI and PLDD were correctly modeled.
A new computationally-efficient three-dimensional Monte Carlo simulator
was developed within UT-MARLOWE. An unbalanced Octree algorithm was
used for spatial decomposition for its accuracy and efficiency. A new three-
dimensional trajectory replication scheme was developed and implemented in
order to further enhance computational efficiency. More than two orders of
magnitude savings on CPU time have been observed. Examples of simulations of
implant into highly non-uniform, non-planar, novel device structures were
presented.
81
Finally, a new analytical model was been developed based on Legendre
polynomials to simulate ion implantation in both one and two dimensions. It
greatly simplifies the parameter extraction procedure and can model both the
impurity and the damage profiles. In two-dimensional simulations, a dose-
splitting approach has been developed to account for the nonlinear effect caused
by damage accumulation. It has been shown that the results obtained from the
new model agree very well with experimentally verified Monte Carlo simulation
results with major savings of CPU time.
6.2 Future tasks
First of all, improved 2-D and 3-D Kinetic Accumulative Damage Model
(KADM) is desirable. This is especially helpful in studying defect generation and
evolution, and it will lead to great insights of diffusion mechanisms on an
atomistic scale.
Secondly, the current simulator does not yet have the capability to cover
all the species and materials used in manufacturing. Furthermore, new species and
new materials are constantly introduced in research and development of
semiconductor industry. To develop accurate models for new species and new
materials, careful experimental design and model calibration are required. In
particular, realistic model of implant into poly-silicon, with detailed consideration
of poly-grain size as well as crystal orientation is desirable.
82
Appendix A: A Brief Manual of TOMCAT
A.1 Installation
The distribution of TOMCAT actually consists of three programs, tomcat,
Sherwood and utview. Sherwood is a 2-D graphical visualization utility, while
utview is a 3-D graphic visualization tool. Both of them allow users to see the
topography in use, as well as monitor the progress of the simulation. The use of
Sherwood or utview is strongly recommended, particularly when first constructing
topography. The following steps should allow you to painlessly install TOMCAT
on your computer. You will need to download two tar files: the tomcat source code,
and the table file.
A.1.1. Create a directory within which you will compile TOMCAT. Copy the
tomcat_release99.tar.gz file (the source code file) into it, and unpack with
gzip -d tomcat_release99.tar.gz
tar -xvf tomcat_release99.tar
The resulting directory tomcat_release99 should have the following
subdirectories:
tomcat xsherwood utview docs
examples test_suite
A.1.2. Go into the tomcat directory. This directory contains the source code for
TOMCAT. Take a look at the available Makefiles in the directory. You
83
should have Makefile.aix, Makefile.sol, Makefile.linux, Makefile.SGI, and
Makefile.alpha, for AIX, Solaris, Linux, IRIX, and DEC Alpha operating
systems, respectively. Compile tomcat by doing the following:
make -f Makefile.xxx clean
make -f Makefile.xxx
Makefile.xxx is the Makefile that corresponds to your system, i.e.
Makefile.sol if you have a Solaris machine.
The compilation should proceed without errors (although some harmless
warnings are possible on some systems). When the compilation has finished,
copy the executable (tomcat) to a convenient directory of your choice (such
as /usr/local/bin), and make sure the PATH variable includes the directory
the TOMCAT executable is in.
A.1.3. Go into the xsherwood directory. Choose the Makefile as in the previous step.
In order to compile xsherwood on your system, you'll need Motif libraries.
Contact your system administrator if compilation fails due to the lack of
Motif libraries. As before, compile using:
make -f Makefile.xxx clean
make -f Makefile.xxx
This creates a dynamically linked executable called xhserwood.
Copy this to a convenient directory (i.e.: /usr/local/bin).
84
A.1.4. Go into the utview directory. Choose the Makefile as in the previous step. In
order to compile utview on your system, you'll need OPENGL or MESA
(freeware) libraries. Contact your system administrator if compilation fails
due to the lack of MESA libraries (downloadable from
http://www.mesa3d.org/). As before, compile using:
make -f Makefile.xxx clean
make -f Makefile.xxx
This creates a dynamically linked executable called utview.
Copy this to a convenient directory (i.e.: /usr/local/bin).
A.1.5. Unzip and untar the table file as in the steps above. A directory called tables
should be created and it should contain the binary scattering tables and other
information used by TOMCAT. Move the files to a convenient directory.
Each TOMCAT user needs to set two environment variables,
TOMCAT_TABLES and TOMCAT_OUT. The first is the directory in
which TOMCAT will look for tables, and the second is the directory into
which output will be sent. If these are not specified, TOMCAT will use the
directory in which the executable is located. Setting the environment
variables is done differently in different shells; check with your system
administrator if you are not sure. In shells derived from sh, you can set the
variables using
85
export TOMCAT_TABLES=/usr/local/tables
where the tables are expected to be at /usr/local/tables (you can, of course,
use a different directory).
A.1.6. Make sure that the PATH variable contains the directories in which tomcat,
xsherwood and utview are located. The installation is now complete. For
instant gratification of 2-D simulation, copy the files as15.in and sd.top from
the examples/first_example directory into your work directory, and run
TOMCAT by typing:
tomcat as15.in sd.top
This should start a TOMCAT simulation, and display the progress
graphically using xsherwood.
A.2 Quick Tutorial
The input to TOMCAT consists of a description of the ion source, and the
structure to be implanted into for 2-D/3-D simulations. The output is a collection
of files that describe the doping and defect concentrations. TOMCAT is run from
the command line using the following syntax:
tomcat [sourcetype] sourcefile [topographyfile]
where sourcetype is an integer that identifies the type of simulation to be run. 0
denotes ion implantation, 1 is for electron transport (the models for the latter are
not yet validated and its use will not be further documented here) and 3 is for 3-D
86
ion implantation simulation. The sourcetype parameter is optional for 1-D and 2-
D simulation, and the default is 0 (ion implantation). The sourcetype is required
for 3-D simulation. Sourcefile is the name of the input file that describes the
source conditions. If you have used UT-MARLOWE before, this will be quite
familiar. The topography argument is a file that specifies the structure to be
implanted into, which is required for 2-D and 3-D simulation.
A.2.1 Description of the Source File
The sourcefile describes the key implant parameters, such as energy, tilt
and rotation angles, and dose.
Input Description for 1-D simulation
The following is an example of a source input file for 1-D simulation:
species=arsenic
energy=15000
dose=1e14
tilt=7
rot=30
numions=5000
Input Description for 2-D simulation
The following is an example of a source input file for 2-D simulation:
species=arsenic
energy=15000
dose=1e14
87
tilt=7
rot=30
numions=300000
leftwind=1
rightwind=599
elevation=550
The first keyword sets the implant species to arsenic. Next, the energy is
set to 15 keV, the dose is 1e14 cm-2, and the tilt and rotation angles are set to 7
and 30 degrees, respectively. The number of simulated ions is set using the
numions command, 300,000 being a reasonable number for 2-D simulation into a
small structure.
The lateral extent of the source is controlled using the leftwind and
rightwind parameters. These correspond to the lateral (x) coordinate of the left
and right edges of the source. The vertical position of the source is controlled with
the elevation command. All coordinates are in silicon lattice constants, unless
otherwise specified (1 lattice constant = 0.543 nm).
The coordinate system used is described in A.2.2.
Input Description for 3-D simulation
The following is an example of a source input file for 3-D simulation:
species=arsenic
energy=15000
dose=1e14
88
tilt=7
rot=30
numions=300000
xleftwind=1
xrightwind=599
yleftwind=1
yrightwind=599
elevation=550
As you can see, it is very similar to the input description for 2-D
simulation except the implant window description. In 3-D simulation, four
coordinates need to be defined xleftwind, xrightwind, yleftwind and yrightwind.
For the definition of other command, please reference to last section, “Input
Description for 2-D simulation”.
A.2.2 Description of the Topography File
The topography file is not needed for 1-D simulation. For 2-D and 3-D
simulation, it describes the structure to be implanted into. In addition, commands
that control the output of TOMCAT are issued here as well. The structure is
specified as a list of geometric objects. For 2-D simulation, currently supported
objects are rectangles (2Dppiped), discs (2Dcylinder), and triangles (2Dtri).
Alternatively, the structure can be imported from a process simulator, such as
SUPREM or FLOOPS. For 3-D simulation, supported objects include box
89
(3Dppiped), triangle cylinder (3Dtri), tetrahedron (tetra), prism (prism), circular
cylinder (3Dcylinder), and Sphere (3Dsphere).
The syntax for commands in the topography file is:
Command (Argument list)
For example, the command 2Dppiped(0,400,600,0,3) creates a rectangle
with the upper left corner at coordinates (0,400), and bottom right coordinates
(600,0), composed of material type 3 (crystalline silicon). Each topography file
must include a 2DBoundVol command, as in 2DBoundVol(0,600,600,0,9). This
sets the simulation volume, with the first two arguments representing the
coordinates of the upper left corner of the simulation volume, and the latter two
representing the coordinates of the bottom right corner. The simulation area (in 2-
D) or volume (in 3-D) must be set up as a square (for 2-D) or a cube (for 3-D).
The final number in the argument list defines the resolution of the grid that will be
generated to represent the topography. The smallest spacing in the grid can be
obtained from the following formula:
min_spacing = size_of_boundary * 2-(n+1)
The size_of_boundary is the side of the square representing the
2DBoundBol, while n is the last argument to 2DBoundVol. This results from the
fact that the grid is obtained by a recursive subdivision of the geometry at most n
times (less in areas of uniform topography). Typical values are 8-10, though care
90
should be taken that the minimum spacing chosen resolves the features of interest.
All dimensions are in Si lattice constants (0.543 nm). See also the ResWindow
command in the next section for more efficient resolution specification.
A.2.3 Simple example of 2-D simulation
The following is an example of a topography structure, which implements
an oxide step on a silicon surface. Note that the pound sign # is conveniently used
to attach comments after any command.
2DBoundVol(0,600,600,0,8) #coordinates_should_be_positive
2Dcylinder(100,400,50,2) #oxide_disc
2Dppiped(0,400,600,0,3) #block_of_crystalline_silicon
2Dppiped(0,450,100,400,2) #oxide_block
2Dppiped(0,403,600,400,2) #native_oxide
TopPlot(TRUE) #this_starts_xsherwood
BinOut(TRUE) #also_needed_for_xsherwood
Scale(403,-1,5.43e-4)
The following source input file can be used to test it.
species=boron
energy=5000
dose=1e13
tilt=7
rot=30
numions=100000
leftwind=1
91
rightwind=599
elevation=550
Note that the elevation has been set above the highest point in the
topography. This is necessary in order to make sure that shadowing by
topography is handled correctly. By default, the elevation parameter is set to the
top of the structure (including vaccum, dicatated by the bounding volume).
Assuming the topography file is called step.top, and the source file is called
test.in, run the simulation by specifying:
tomcat test.in step.top
After a few messages from TOMCAT, xsherwood pops up a graphic
representation of the topography. You can use the Zoom and Pan buttons on the
xsherwood toolbar to inspect it as you wish. The dopant density is updated every
twenty seconds, although output from TOMCAT might be less frequent than that.
TOMCAT outputs a new binary dopant file (assuming that BinOut(TRUE) has
been specified in the topography file) every 15000 ions. You can ask xsherwood
to reload an update dopant file by clicking on reload. Clicking on “Quit” does not
end the simulation, and it merely exits from xsherwood.
A.2.4 Simple example of 3-D simulation
The following is an example of a topography structure, which implements
an oxide step on a silicon surface and shallow trench isolation. Note that the
pound sign # is conveniently used to attach comments after any command.
92
3DBoundVol(0,600,0,600,0,600,8)#coordinates_should_be_positive
3Dcylinder(180,400,100,0,600,2) #oxide_disc
3Dppiped(0,400,100,600,0,600,3) #block_of_crystalline_silicon
3Dppiped(0,430,0,200,401,600,2) #oxide_block
3Dppiped(0,490,0,200,420,600,4) #poly_silicon_block
3Dppiped(0,500,0,200,490,600,2) #oxide_block
3Dppiped(500,400,0,600,0,600,2) #oxide_block
3Dppiped(0,400,0,600,0,100,2) #oxide_block
TopPlot(TRUE) #this_starts_utview
Scale(403,-1,5.43e-4)
The following source input file can be used to test it.
species=arsenic
energy=15000
dose=1e13
tilt=7
rot=30
numions=50000000
xleftwind=1
xrightwind=599
yleftwind=1
yrightwind=599
elevation=550
Note that the elevation has been set above the highest point in the
topography. This is necessary in order to make sure that shadowing by
93
topography is handled correctly. By default, the elevation parameter is set to the
top of the structure (including vacuum, controlled by the bounding volume).
Assuming the topography file is called step.top, and the source file is called
test.in, run the simulation by specifying:
tomcat 3 test.in step.top
After a few messages from TOMCAT, utview pops up a graphic
representation of the topography. The key command for utview is listed in section
6 and you can inspect the topography as you wish. TOMCAT outputs a new binary
dopant file (assuming that BinOut(TRUE) has been specified in the topography
file) every 15000 ions. You can ask utview to reload an update dopant file by
pressing any key. Clicking on “ESC” to exit utview and it will not end the
simulation.
94
Appendix B: Block Diagram and Flow Chart
C++ is the language used to implementation the model described in earlier
chapters for its modularity and polymorphism. The basic structure consists of four
major parts as shown in Fig. B.1: the input, the source, the zoner and the output.
The input part contains the input source file and the input topography file,
which specifies the implant condition and the implanting topography,
respectively. The source part initializes the particles according to the implant
condition. The zoner part includes two major objects, topography and propagator.
The topography performs the actual Quadtree/Octree decomposition, and contains
a collection of decomposed grid cells. Base class, propagator, has many derived
classes, each a self-contained Monte Carlo program, to propagate particles
through different types of materials, crystal silicon, amorphous silicon, silicon
dioxide, silicon nitride, etc. Each decomposed grid cell is assigned a propagator
pointer, which points to the propagator of the particular material type in the grid
cell.
The generic simulation flow chart is illustrated in Fig. B.2. To start the
simulation, at first, the source initiates the particle according to the implant
condition, and then pass it by reference to the zoner. The zoner assigns the
particle or particles to the grid cell. Normally, implanted ion starts from the grid
cells whose material type is vacuum. The particle then will propagate in the grid
95
Output
Graphic tools
Propagator
Source
Implant conditions
Particle initiation
Input/ User interface
Zoner
Topography
Input topography file
Input/ User interface
Fig. B.1 Schematics of the block diagram of TOMCAT
96
No
No
Yes
Are the particles leaving their grid cells?
Is the particle exit condition
satisfied?
Yes
Is the simulation exiting condition met?
Yes
No
Update the simulated structure, and remove the exiting particles.
Finish simulation and simulation results are stored!
Particles are assigned to the corresponding grid cell and propagator
Particles passed by reference to the zoner
The Source initiates the particle(s) Fig. B.2 Simulation flow chart of TOMCAT
97
cell according to its propagator. Ions will propagate until their energy is lower
than 10 eV. If the particles leave their grid cells, the zoner will assign them to new
grid cells by Octree zoning algorithm. The structure will be updated until all
particles in the ensemble complete their propagation. Above process will be
repeated, assigning the particles that have left their grid cells to new propagators,
and requesting new particles from the source. Simulation continues until the
simulation dose is reached.
98
Bibliography
1. K. Weiner, P. Carey, A. McCarthy, and T. Sigmon, “Low-temperature
Fabrication of p-n Diodes with 300 Å Junction Depth”, IEEE Electron
Device Letters, vol. 13, no. 7, p. 369, 1992.
2. B. Yu, Y. Wang, H. Wang, Q. Xiang, C. Riccobene, S. Talwar, M. Lin,
“70nm MOSFET with Ultra-Shallow, Abrupt, and Super-Doped S/D
Extension Implemented by Laser Thermal Process (LTP)”, IEDM, 1999.
3. R. Wilson, F. Stevie, and C. Magee, “Secondary Ion Mass Spectrometry”,
John Wiley & Sons, Inc., New York, 1989.
4. G. Cooke, M. Dowsett, C. Hill, P. Pearson, A. Walton, J. Vac. Sci. Tech.
B12, 243 (1994).
5. R. Kleiman, M. O’Malley, F. Baumann, J. Garno, and G. Timp, “Junction
Delineation of 0.15pm MOS Devices Using Scanning Capacitance
Microscopy”, IEDM, 1997.
6. R. Chau, J. Kavalieros, B. Doyle, A. Murthy, N. Paulsen, D. Lionberger,
“A 50nm Depleted-Substrate CMOS Transistor (DST)”, IEDM, 2001.
7. Y. Choi, N. Lindert, P. Xuan, S. Tang, D. Ha, E. Anderson, T. King, J.
Bokor, C. Hu, “Sub-20nm CMOS FinFET Technologies”, IEDM, 2001.
8. K. Guarini, P. Solomon, Y. Zhang, K. Chan, E. Jones, G. Cohen, A.
Krasnoperova, M. Ronay, O. Dokumaci, J. Bucchignano, C. Cabral Jr., C.
99
Lavoie, V. Ku, D. Boyd, K. Petrarca, I. Babich, J. Treichler, P. Kozlowski,
J. Newbury, C. D’Emic, R. Sicina, and H. Wong, “Triple-Self-Aligned,
Planar Double-Gate MOSFETs: Devices and Circuits”, IEDM, 2001.
9. G. Balamurugan, B. Obradovic, G. Wang, Y. Chen, “New Analytic
Models and Efficient Parameter Extraction for Computationally Efficient
1-D and 2-D Ion Implantation Modeling”, IEDM, 1998.
10. A. Tasch, H. Shih, C. Park, J. Alvis, and S. Novak, J. Electrochem. Soc.,
136(3), 810-814, 1989.
11. K. Levenberg, Q. Appl. Math. 2, 164, 1994.
12. K. Beardmore, Proceedings of the 11th International Conference on Ion
Implantation Technology, 535, 1997.
13. J. P. Biersack, L. G. Haggmark, Nuclear Instruments and Methods 174
(1980), 257-269, North Holland Publishing Company.
14. K. Klein, C. Park, and A. F. Tasch, IEDM Tech. Dig., 745-748, 1990
15. G. Hobler, H. Potzl, "Simulation of Two-Dimensional Implantation
Profiles with a large concentration range in crystalline silicon using an
advanced Monte Carlo Method", IEDM 1991 Technical Digest, p. 67.
16. M. Posselt, Nuclear Instruments and Methods in Physics Research B80/81
(1993) 28-32 North Holland.
100
17. S. Tian, S. J. Morris, M. F. Morris, B. Obradovic, and A. F. Tasch, IEDM
Tech. Dig., 1996.
18. S. J. Morris, B. Obradovic, S. H. Yang, and A. F. Tasch, IEDM Tech. Dig.
1996.
19. J. F. Ziegler, “Handbook of Ion Implantation Science and Technology”,
Ion Implantation Co., 1996.
20. W. Brandt and M. Kitagawa, Phys. Rev. B, vol. 25, No. 9, 5631-37, 1982.
21. R. Mathar, and M. Posselt, Phys. Rev. B, vol. 51, No. 1, 107-116, 1995.
22. O. B. Firsov, Soviet Physics JETP, vol. 36(9), No. 5, 1517-23, 1959.
23. G. H. Kinchin, and R. S. Pease, “Rep. Progr. Phys., vol. 18, 1-51, 1955.
24. G. Wang, “Computationally Efficient Models for Monte Carlo Ion
Implantation Simulation in Silicon”, M. S. thesis, The University of Texas
at Austin, 1997.
25. G. Hobler, A. Simionescu, L. Palmestshofer, C. Tian, and G. Stingeder, J.
Appl. Phys., vol. 77, 3697-3703, 1995.
26. International Technology Roadmap for Semiconductors, (Semiconductor
Industry Association, San Jose, CA, 1999).
27. J. F. Ziegler, Private communication.
28. M. D. Giles, and J. F. Gibbons, IEEE Trans. Computer Aided Design,
4(4), 362-368, 1985.
101
29. A. Hossinger, and S. Selberherr, Proceedings of the 5th International
Symposium on Process Physics and Modeling in Semiconductor
technology, 17.
30. A. Hossinger, E. Langer, and S. Selberherr, IEEE Transactions on
Computer-Aided Design of Integrated Circuits and Systems, 19(5), 560,
2000.
31. A. Burenkov, K. Tietzel, A. Hossinger, J. Lorenz, H. Ryssel, and S.
Selberherr, IEICE Trans. Electron., E83(8), 1259, 2000.
32. B. J. Obradovic, G. Balamurugan, G. Wang, Y. Chen, and A. F. Tasch,
“Monte Carlo simulation of ion implantation into topographically complex
structures”, IEDM 1998. Technical Digest, pp. 513-16.
33. S. Morris, S. Yang, D. Lim, and A. F. Tasch, MRS proceeding 1993.
34. G. Hobler, H. Huong, J. Bevk, A. Agarwal, H. Gossman, D. Jacobson, M.
Foad, A. Murrell, Y. Erokhin, IEDM 1997 Technical Digest, 489.
35. D. Li, G. Balamurugan, B. Obradovic, G. Wang, Y. Chen, and A. F.
Tasch, Proceedings of the 5th International Symposium on Process Physics
and Modeling in Semiconductor technology, 26.
36. G. F. Carey, W. B. Richardson, C. S. Reed, B. Mulvaney, "Computer
Simulation of Semiconductor Processes and Devices-- Mathematical and
Numerical Aspects", John Wiley and Sons, 1996.
102
37. B. Obradovic, G. Wang, Y. Chen, D. Li, C. Snell, and A. F. Tasch, UT-
MARLOWE Manual, Version 5.0, The University of Texas at Austin,
April, 1999.
38. G. Hobler, G. Bevk, and J. Agarwal, IEEE Electron Device Letters, 1999,
20(7), 357.
39. A. F. Tasch, S.-H. Yang and S. J. Morris, "Modeling of ion implantation
in single-crystal silicon", Nuclear Instruments and Methods in Physics
Research B, 102, pp. 173-9(1995).
40. G. Wang, S. Tian, M. J. Morris, S. J. Morris, B. J. Obradovic, G.
Balamurugan and A. F. Tasch, "A computationally efficient ion
implantation model: modified Kinchin-Pease model", Proceedings of
SPIE-The International Society for Optical Engineering (USA), 3212, pp.
324-33(1997).
41. C. Park, K. M. Klein and A. F. Tasch, "A comprehensive and
computationally efficient modeling strategy for simulation of boron ion
implantation into single-crystal silicon with explicit dose and implant
angle dependence", COMPEL, Int. J Comput. Math. Electr. Electron. Eng.
(UK), Vol. 10, No. 4, pp. 331-40.
42. K. M. Klein, C. Park, S.-H. Yang and A. F. Tasch, "Accurate and efficient
two-dimensional modeling of boron implantation into single-crystal
103
silicon", International Electron Devices Meeting 1991. Technical Digest,
pp. 697-700.
43. S. J. Morris, D. Lim, S.-H. Yang, S. Tian, K. Parab and A. F. Tasch, "A
two-dimensional accurate and computationally efficient model for boron
implantation through oxide layers into silicon", Electrochemical Society
Proceedings. 96(4), pp. 468-80.
104
Vita
Di Li was born in Changchun, Jilin Province, People’s Republic of China
on April 6, 1973, the son of Ying Wang and Jingrui Li. After completing his work
at The High School attached to The Northeast Normal University of P.R. China in
1991, he entered Beijing University in Beijing, P.R. China. He received the
degree of Bachelor of Science majoring in Physics from Beijing University in
1996. In August 1996, he entered Rice University, Houston Texas and received
the degree of Master of Science majoring in Applying Physics in June 1998. In
July 1998, he entered the Graduate School of The University of Texas majoring in
Electrical Engineering.
Permanent Address: 3461 Lake Austin Blvd., Apt. C
Austin Texas 78703
This dissertation was typed by the author.
105