i
PREDICTIVE STATISTICAL ANALYSIS OF
EMBEDDED MEANDER RESISTORS VIA
MEASUREMENT OF CANONICAL BUILDING
BLOCKS
A Thesis Presented to
The Academic Faculty
by
Lawrence A. Carastro
In Partial Fulfillment of the Requirements for the Degree of
Doctor of Philosophy In Electrical and Computer Engineering
Georgia Institute of Technology April 2002
Copyright ©2002 by Lawrence A. Carastro
ii
PREDICTIVE STATISTICAL ANALYSIS OF
EMBEDDED MEANDER RESISTORS VIA
MEASUREMENT OF CANONICAL BUILDING
BLOCKS
Approved:
____________________________________
Martin A. Brooke ____________________________________ Steveen Kenney
____________________________________
John Dorsey
Date Approved__________________
iii
DEDICATION
To My Wife Sheri
ACKNOWLEDGEMENT
iv
SUMMARY
A further developed passive device modeling methodology that accurately
predicts the high frequency electrical behavior, and the statistical variation of embedded
passive devices on a generic substrate, is presented. This methodology accurately models
new passive devices by utilizing a small number of test structures to deembed the lumped
element circuit values from canonical building blocks.
In essence, a nonlinear optimizer is used to find the optimal building block
models, by fitting extensive high frequency measurements of the test structures behavior.
This process is repeated on numerous fabricated test structures producing a relative
variation in the lumped element circuit values of the building block models. A predicted
range of operation for the actual modeled device is then generated using a MonteCarlo
simulation tool, and the relative variations of the building block models.
Two sets of thirty-two test structures are fabricated by depositing gold on a 96 %
Alumina substrate, and s-parameter measurements taken. The Hspice Optimization tool
uses these measurements to deembed the lumped element circuit values from the building
block models; which are contained within the fabricated test structures. Relative
variations of the deembedded element values are calculated and added to the device
model circuit file, for simulation in the Hspice MonteCarlo tool. By means of
comparison, between the range predicted by the MonteCarlo simulations and the actual
measured data, the statistical variations of the component values were shown to provided
an accurate representation of the overall passive device performance.
v
This method has been experimentally verified for a 9-segment meander resistor,
by demonstrating that the variation in the complete equivalent circuit models, based only
on the circuit building blocks, can be used to predict such variations in the actual
fabricated device. An accurate statistical modeling of the same 9-segment meander
resistor utilizing only one test structure to deembed building block equivalent circuits will
also be successfully proven in this research.
To automate this research, Unix scripts were developed which act as interface
between user and computer by; automatically, running the complete Hspice Optimization
simulations, calculating the relative variations of the deembedded element values,
generating the device model circuit file, running the Hspice MonteCarlo analysis,
extracting and converting all s-parameters to polar coordinates, and saving and localizing
all circuit and optimization files.
1
CHAPTER 1
INTRODUCTION
Passive components are referred to as “glue components” because they “glue”
integrated circuits together to make a system [1]. Currently, the majority of passive
components functioning in electronic systems are discrete, and dominate the area of a
printed wiring board in a typical electronic product. For example, a cellular phone may
consist of only about 20 integral circuits compared to 300-400 passives. Clearly, discrete
passives substantially influence system cost, size, and reliability. To meet the
requirements for the next generation of electronic packaging (i.e. smaller, lighter, faster,
cheaper, and more reliable), alternatives to discrete passives must be realized.
Recent advances in technology have allowed the integration of more functionality
onto a single IC by increasing on-chip transistor counts. This integration has lead to an
increase in passive component board area, and is now a limiting factor for further
reductions in board size. These further reductions would result in much smaller, lighter,
and more reliable systems, and potentially impact every component using electronic
circuits. Therefore, the advantages of board reduction or miniaturization have forced the
development of new technologies to remove passives from the top of the printed circuit
board, and to embed them within a substrate or a package.
Integral Passives defined, as any functional passive element that is formed on a
layer within the interconnecting substrate, is that new developing technology. Integral or
2
embedded passives have the potential for increased reliability, improved electrical
performance, size shrinkage, and cost reduction. This potential has lead to a constant
focus on higher levels of system integration and miniaturization. Processes such as thick
and thin film processing have enabled materials to be deposited at the mil and micron
level for linewidths and spacing within multichip modules. Techniques such as low
temperature cofired ceramic (LTCC) processing can use thick film printing and stacking
of passive devices in a multilayer low-cost substrate. LTCC processing can create up to
30 layers of passive component integration with the possibility of combining it with an
integrated circuit within a standard IC size package. For the accurate design and
fabrication of these compact high performance systems, accurate modeling of embedded
passive components is becoming very important.
However, designing circuits with embedded components is non-trivial due to
electromagnetic interactions that lead to parasitics, and ultimately non-ideal frequency
behavior. Passive devices generally have complex geometries, non-uniform current flow,
and correspondingly complex field patterns; therefore, suffer from parasitic effects that
influence the electrical behavior of the device at different frequencies. For this reason,
accurate models of embedded passive components and the examination of component
value statistical variation are crucial for designing and characterizing high performance
systems.
3
1.1 Thesis Organization
In this dissertation, a statistical analysis procedure is detailed and applied to a 9-
segment meander resistor, using a novel passive device modeling methodology. The first
part of this thesis examines the development and procedures of the modeling method.
The second section details the improvements to the modeling methodology, and
demonstrates the application of the analysis procedure to the statistical variation of a 9-
segment meander resistor using two separate deembedding procedures. The first
procedure uses two test structures to model four building blocks, which are comprised of
twenty-two different deembedded element values. The second procedure uses only one
test structure to model four building blocks, which are also comprised of twenty-two
different deembedded element values.
The statistical modeling of the 9-segment meander resistor with interactions is
examined, and good results are shown for the device fabricated on a 96% Alumina
substrate. A brief chapter-by-chapter outline of the thesis is given below.
Chapter 2 discusses several methods developed for the characterization,
modeling, and design of microwave and millimeter-wave passive structures. An
overview of various lumped element modeling methods and how they apply to statistical
prediction is presented.
Chapter 3 discusses the basis for this research which is the work done by Dr. Ravi
Poddar, Dr. Nan Jokerst and Dr. Martin A. Brooke in 1997 entitled “Accurate, High
Speed Modeling of Integrated Passive Devices” [28]. A test case using this modeling
4
methodology is examined in detail for a 9-segment meander resistor, which demonstrates
that the predictive capability is both accurate and very fast.
Chapter 4 presents a detailed account of the improvements made to the passive
device modeling methodology detailed in chapter 3. The same 9-segment meander
resistor test case will be examined in detail and compared to the original methodology.
Chapter 5 will detail the successful modeling of the same 9-segment meander
resistor; however, only using one test structure to model the four building blocks and the
subsequent 22 deembedded element values. The results will be compared to the original
methodology (detailed in Chapter 3), and the improved methodology (detailed in Chapter
4).
Chapter 6 will present the statistical analysis procedure developed under this
research program. A question concerning whether the deembedded element values are
correlated or independent of one another, and how this applies to the construction of the
modeled device circuit file is discussed. Several statistical analysis techniques used to
evaluate and render analysis results will also be explained in detail.
Chapter 7 will detail the application of the statistical analysis procedure to a 9-
segment meander resistor using the improved modeling methodology detailed in Chapter
4. A comparison between correlated and independent element value variations, which
directly influence the construction of the modeled device circuit file, used in the
MonteCarlo analysis will be shown.
Chapter 8 will detail the application of the statistical analysis procedure to a 9-
segment meander resistor using only one test structure to model the canonical building
blocks as described in Chapter 5. A comparison between the results of Chapter 7
5
(correlated and independent) and the results using only one test structure (independent)
will also be shown.
Chapter 9 will act as a tutorial for the operation of the Unix scripts, which
automate the entire modeling and statistical analysis process.
Chapter 10 will demonstrate the application of the statistical analysis procedure
by comparing MonteCarlo predictions with actual measured devices. Conclusions will
also be extracted regarding this analysis procedure along with recommendations for
further research.
6
CHAPTER 2
BACKGROUND
2.1 Introduction
As presented in Chapter 1, accurate modeling of embedded passive components
and the examination of component value statistical variation are crucial for the design,
and fabrication of high performance systems that utilize these devices. Highly
miniaturized passive components, fabricated on GaAs, and high speed silicon substrates,
are extensively used in the microwave/RF community for use in microwave circuits;
whereas, good frequency dependent models must be obtained for successful design. As a
result of this requirement, much of the work in the area of passive component modeling
originates from the microwave engineering community [4].
A number of numerical methods have been developed for the characterization,
modeling and design of microwave and millimeter-wave passive structures. These
methods can be divided into 3 basic areas. The first is the numerical full-wave method,
which applies Maxwell’s equations to the structure and computes the electric and
magnetic field patterns. This method is the most flexible and general, however
simulation times are very long for complex structures.
Second is the analytical equation based method, which is derived from first
principles or from curve fitting using simulation or measurement. Generally speaking,
7
computer efficiency increases with the amount of analytical effort required, but at the
same time, versatility of the method is reduced [4]. This is due to the limited number of
simple shapes that analytical forms are developed for, under some simplifying
hypotheses. This method is accurate and extremely fast.
Third is the measurement-based method that can be applied to any arbitrary
structure, which takes into account processing effects, and is very accurate at any
frequency range.
On a whole, applying any of these three methods to non-lumped element
simulation and modeling methods is difficult and time-consuming due to the complexity
of the modeled device. However, when lumped element models are used, the electrical
behavior information of the device at different frequencies must be synthesized to
deembed the equivalent circuit element values. To accomplish this goal, the
deembedding process incorporates four different methods: 1) Physical Equations, 2) EM
simulations, 3) Algorithms, and 4) Measurements.
In this chapter, an overview of various lumped element modeling methods and
how they apply to statistical prediction will be presented. Details can be found in the
various references.
8
2.2 Lumped Element Models
Passive device modeling using lumped element models yields circuits or sub-
circuits that are easily and quickly modeled using a Spice type circuit solver. The
generated circuit models simulate very quickly, usually on the order of minutes or
seconds; therefore, providing a major speedup over methods that do not utilize lumped
elements [4]. These highly flexible circuit level models can be incorporated into the
circuit design stage to investigate affects on circuit performance. Recently, computer-
aided design tools such as Hspice have become indispensable in IC design; whereas,
accurate circuit simulation is dependent on both the structural validity of the device
models and the accuracy of the values used as model parameters. Therefore, the
extraction of an optimum set of device model parameter values is crucial in
characterizing the relationship between the model and the measured behavior. Fig. 2.2-1
is a schematic of a generic lumped element equivalent circuit model for a spiral inductor.
This equivalent circuit can also model a capacitor when the circled capacitor is added.
9
Ls Rs
Cs
C
CsiRsi Csi Rsi
Cox Cox
Capacitance and Resistance of Silicon Substrate Modeled by Csi and Rsi
Capacitance and Resistance of Silicon Substrate Modeled by Csi and Rsi
Oxide CapacitanceBetween
Spiral and SiliconModeled by Cox
Oxide CapacitanceBetween
Spiral and SiliconModeled by Cox
Inductance of SpiralModeled by Ls
Inductance of SpiralModeled by Ls
Feed-Through Path ModeledBy Series Capacitance Cs
Feed-Through Path ModeledBy Series Capacitance Cs
Resistance of Metal TraceModeled by Rs
Resistance of Metal TraceModeled by Rs
Added for Capacitor ModelModeled by C
Added for Capacitor ModelModeled by C
Figure #2.2-1: Lumped Element Circuit
The extraction of lumped element circuit values fall into two basic classes, the
first obtains electrical behavior data (s-parameters) by calculation, using an EM
simulation tool, or by measurement from an actual fabricated device. A unique
equivalent circuit model is created, and the element values deembed using an
optimization routine such as Levenberg-Marquardt or polynomial equations in a black-
box configuration. The model is then simulated in a Spice type circuit solver. The
second is a direct calculation of the lumped element values using scalable process and
configuration dependent physical equations, or algorithms derived from first principles
and simplified using assumptions. The lumped element values are inserted into a unique
equivalent circuit model and simulated in Spice type circuit solver.
Individual modeling goals further divide the lumped element extraction process
into three groups. The first group is the scalable models whose lumped element values
are a function of configuration such as length, width, and metal trace separation. The
10
second group is the subset models that partition a device into subsets and model the
individual subset, which are then assembled to create the device model. The third group
is the polynomial equation models that use transfer functions and predictive polynomial
equations to model the device.
2.2.1 Scaleable Lumped Element Models
Scaleable lumped element models can be extracted using four different methods.
The first method uses physical equations and optimization routines to deembed element
values using fixed and free variables, and empirical and semi-empirical methods.
In the paper by Lakshminarayanan, B., Gordon, H.C., and Weller, T.M., the
lumped element values are deembedded using a combination of circuit optimization (free
variables using measured s-parameter data) and process and configuration dependent
equations (fixed variables) [5]. The free variables improve the prediction of substrate
dependencies while the fixed variables ensure scalability for different capacitor values.
The inclusions of the two RLC branches model higher order resonances and are
optimized using the original element values as initial guesses. This model accurately
captures significant changes in frequency response by changing only the height of the
microstrip substrate; therefore, parameter values for intermediate capacitor sizes are
predicted with a high degree of accuracy. The uniform variation of parameters is a
consequence of model’s close ties to physical properties. Fig. 2.2.1-1 shows the cross
section of the ceramic substrate and the lumped element equivalent circuit used to model
the capacitor.
11
Electrodes
Metalization
MicroStrip
Substrate
GroundPlane
t
εc
rε
Cross Section of CMC In Typical Micro strip MountCross Section of CMC In Typical Micro strip Mount
P1 P2ESL ESR
Cs
Cg Cg
C
L1 R1C1
L2 R2C2
Substrate-Dependent Model for Ceramic Multilayer Chip Capacitors
Substrate-Dependent Model for Ceramic Multilayer Chip Capacitors
Figure #2.2.1-1: CMC Substrate and Equivalent Circuit Model
In the paper by Ronkainen, H., Kattelus, H., Tarvainen, E., Ruhisaari, T., and
Andersson, M., [6] a systematic study of modeling, design, and fabrication of integrated
planar inductors on silicon with different inductor geometries and number of turns was
examined. This methodology developed a purely empirical and a semi-empirical lumped
element model of an inductor. Purely empirical models are basically useless in the
physical layout of inductors, and are also not capable of describing device and process
parameters changes; however, they do accurately reproduce measured high frequency
features. The accuracy is improved by increasing the number of sub-partitions using the
distributed model shown on the left in Fig. 2.2.1-2.
The more physical or semi-empirical model, models components of the equivalent
circuit using functions related to inductor layout and material parameters. These
functions or equations provide scalability, and the minimal number of optimized
parameters provides the speed [6]. The probe pads are modeled separately to determine
their parasitics, which enhances accuracy. This semi-empirical model was found to be
the optimum method in terms of speed and accuracy due to the scaleable design equations
12
and minimal optimized parameters. Fig. 2.2.1-2 shows the equivalent circuit for the
distributed (empirical) and semi-empirical inductor model.
CP
RS LS
RIN ROUT
CIN COUT
Equivalent Circuit of Distributed Inductor Model
Equivalent Circuit of Distributed Inductor Model
Cint
R L
Csub1 Csub2
C1 C2
Rsub1 Rsub2
Rsub12
nin nout
Semi-Empirical Sub circuit Model for Planar InductorsSemi-Empirical Sub circuit Model for Planar Inductors
Figure #2.2.1-2: Equiv. Circuit for Distributed and Semi-Empirical Inductor Model
The second method uses a custom program to calculate the coefficients in the
physical equations using an EM solver.
In the paper by Kmayman, N. and Jain, N., [7] an EM solver is used to calculate
coefficients in the design equations. Basically, the designer inserts the size and the EM
solver solves the unknown coefficients, which yield the element values of the lumped
element circuits [7]. The method is completely automated using shell-scripts as interface
between user and computer. The models are separated into different families according
to structures, to improve model accuracy and speed by reducing unknowns in equations.
Newton-Gauss’s singular-value-decomposition algorithm is also used to improve
convergence and simulation time. Fig. 2.2.1-3 shows the design equations and the
equivalent circuit for a MIM Capacitor and a square spiral inductor.
13
C L RsCp1 Cp2
Rp Rp1
ωC. tanδn.=
Rs r 2ψ. f.. 106.=
2 GHz < f < 35 GHz
25 µm< l < 400 µm30 µm< w < 100 µm
h = 200 µm εr=4
Capacitor
Substrate
C α12. l w( ). α2l. w.= pF
L α31w. α4w. α5l. α6= nH
r α1lw. α8w. α9l. α10= Ω/ Hz
Cp1 α112. l w( ). α12l. w. α13= fF
Cp1 Cp2=
Model & Scalable Design Equations for MIM Capacitor
Model & Scalable Design Equations for MIM Capacitor
L RsCp1 Cp2
C
2 GHz < f < 35 GHz
150 µm< l < 1 mm1.5 < turn < 5.5
h = 200 µmεr=4
Capacitor
Substrate s =10 µm
w =30 µm
L α11. ln l( ). α2= nH
r α3l. ln l( ). α4= Ω/ Hz
Cp1 α5w. l total. α6= pF
Cp2 α7w. l total. α8= pF
Rs r 2ψ. f.. 106.=
Model & Scalable Design Equations for Square Spiral Ind
Model & Scalable Design Equations for Square Spiral Ind
Figure #2.2.1-3: Equivalent Circuits and Design Equations for MIM Capacitor and
Square Spiral Inductor
The third method optimizes the Q-factor and Inductance using a geometric
programming method.
In the paper by del Mar Hershenson, M., Mohan, S.S., Boyd, S.P., and Lee, T.H.,
[8] a simple and efficient CAD tool for designing on-chip spiral inductors for use in a
variety of RF circuits is proposed. This methodology uses the physical dimensions of the
inductor as design parameters to compute globally optimal trade-off curves between
competing objectives such as Q-factor and total inductor area. The Design goals are
translated into geometric programs, which permits circuit optimization efficiently and
globally [8]. Fig. 2.2.1-4 shows the square spiral inductor geometry and the equivalent
circuit used to model the inductor.
14
WS
d out
d in
Square Inductor Layout and Geometry
Square Inductor Layout and Geometry
Ls Rs
Cs
CpRp CpRp
1 2
Lumped Element Equivalent Circuit Model
Lumped Element Equivalent Circuit Model
Figure #2.2.1-4: Inductor Geometry and Equivalent Circuit
The fourth method is a neural network model that is trained using EM solver
generated s-parameter data.
In the paper by Wang, S., Wang, F., Devabhaktuni, V.K., and Zhang, Q.J., [9] a
sub neural network has been developed that establishes the mapping between original
model input space (physical/geometrical parameters) and approximate circuit model input
space. The space-mapped resistor model (SMRM) was simulated using Hspice; whereas,
the output prediction closely matched the EM simulated data. The model simulated in
minutes and was much faster than original EM simulation. This methodology proves that
hybrid neural models are computationally efficient and have an accuracy that is
comparable to EM simulation. Fig. 2.2.1-5 shows the neural model structure and the
equivalent circuit.
15
IS 21RS 21IS 11
fL R W R
R 1 R 2 L 1 C 1 C 2
CircuitModel
RS11
Hybrid Neural Model Structure
Hybrid Neural Model Structure
L1R1
C1
C2
R2
C1
R2
Equivalent Circuit Model
Equivalent Circuit Model
Figure #2.2.1-5: Neural Model and Equivalent Circuit
2.2.2 Sub-Set Lumped Element Models
Sub-set lumped element models can be extracted using three different methods.
The first method divides the model into sub-units and models a single sub-unit.
The method described in the paper by Belk, N.R., Frei, M.R., Tsai, M., Becker,
A.J., and Tokuda, K.L., [10] decomposes the inductor into repeated sub-units and solves
for the field dependent three-dimensional charge and current distributions to determine
the electrical characteristics of the inductor from a suitable combination of sub-unit
characteristics [10]. Basically, the designer assembles the sub-units, determines the
coupling, and then accurately extracts the electrical characteristics. Skin effect increases
resistance at higher frequencies and starts to reduce Q-factor as shown in Fig. 2.2.2-1.
Fig. 2.2.2-1 also shows the inductor and sub-unit configuration, and measured vs.
modeled inductance and Q-factor.
16
123
Optimized Three-Turn InductorOptimized Three-Turn Inductor Fit of Algorithms to Measured DataFit of Algorithms to Measured Data
Figure #2.2.2-1: Optimized Three-Turn Inductor
The second method divides the model into unconnected concentric rings and
optimizes the Q-factor and Inductance as a function of area using a design space method.
In the paper by Rejaei, B., Tauritz, J.L., Snoeij, P., and Kayali, S., [11] this
methodology replaces an N-turn spiral by a series connection of N circular rings, and
assumes the rings are disconnected by modeling them as a 2N-port system with N Inputs
and N Outputs [11]. Assumptions such as one-dimensional/planar approximations and
circular symmetry simplify the field equations yielding an equivalent lumped element
model whose components are expressed in terms of electric and magnetic Green’s
functions. The lumped element model is too complicated for practical purposes;
therefore, by restricting m to a small integer only a limited number of resonant sub-
circuits are used. The measured vs. simulated results for the inductance (above) and loss
(below) illustrates that the models tend to overestimate the inductance, this is possibly
due to the 2D nature of calculations. The DC current was measured as a function of the
input voltage yielding resistor values in good agreement with the predicted ones. Fig.
17
2.2.2-2 shows the spiral inductor and concentric rings configuration and some measured
vs. modeled data for the inductance and impedance.
R R
SS
Spiral Inductor and Concentric Ring Model
Spiral Inductor and Concentric Ring Model
Measured (solid line) vs. Simulated (dashed line)
Inductor Results
Measured (solid line) vs. Simulated (dashed line)
Inductor Results
Figure #2.2.2-2: Spiral Inductor Model and Measured Data
The third method divides the device into segments; whereby, the modeled device
is a combination of individually modeled segments. The lumped element model also
includes inductive and magnetic coupling, by adding magnetic coupling to the substrate
and including inductive coupling between metal traces and other devices.
In the paper by Pieters, P., and Beyne, E., [12] a method for the realization of an
optimum quality factor for spiral inductors given an inductance value and frequency
range is developed. This method approximates spiraling coils as concentric rings and
solves the algorithms for each separate ring. The equivalent circuit element values are
the function of algorithms and relate to physical effects of the component. An
optimization routine finds the highest possible Q-factor as function of inductor area and
frequency. This methodology creates an inductor design space that accurately predicts
inductor behavior for any possible geometry. The Design space for the evaluated
18
geometrical parameter range is created by plotting the quality factor in a plane of
component values vs. frequency [12]. This research observed that for small inductances
values the highest quality factor was achieved at high frequencies, and the other way
around for larger inductances. Fig. 2.2.2-3 shows the equivalent circuit model and some
measured vs. modeled data for S11 and S22 magnitude, and also the Q-factor.
L R
C3
C1 C2Q
Imag ZinRe Zin
=
Lumped Element Model for Multi-turn Spiral Inductor
Lumped Element Model for Multi-turn Spiral Inductor
Comparison of Measurements and Calculated ModelsComparison of Measurements and Calculated Models
Figure #2.2.2-3: Lumped Element Equiv. Circuit & Measured vs. Modeled Data
In the paper by Arcioni, P., Castello, R., Perregrini, L., Sacchi, E., and Svelto, F.,
[13] a classical and an improved equivalent circuit model for inductors has been
developed. The classical equivalent circuit was unable to distinguish between losses in
metal strips and losses in substrate. The improved equivalent circuit introduces a
substrate resistance into the lumped element model, which is magnetically coupled with
the inductor and accurately describes the loss mechanism in both the metal and in the
low-resistivity substrate [13]. The separate characterization of the metal and substrate
losses allows the designer to separately evaluate their contribution to the inductor’s Q-
factor. The separate s-parameter measurements of the entire inductor and the pad isolate
the individual properties. There is a noticeably better fit for the improved equivalent
19
circuit as shown in Fig. 2.2.2-4. The extracted parameters were compared to the expected
ones estimated from geometrical and technological data calculated using the Grover
method [14] yielding errors of 30% for capacitance values, (15%) for inductance values,
and (8%) for resistance values. The s-parameter measurements of the actual device were
compared to modeling results yielded good agreement. These measurements take into
account the physical phenomena underlying the coupling between metal and substrate.
The new proposed model is more accurate and highlights losses contributing at a given
frequency. Fig. 2.2.2-4 shows the classical and improved equivalent lumped element
circuit and some measured vs. modeled data for S21 magnitude and the Q-factor.
LYL RS
1:1
RSUB
CF
Improved Equivalent Circuit
Improved Equivalent Circuit
L
YL
RS
CF
Classical Equivalent Circuit
Classical Equivalent Circuit
Measured vs. Classical and Improved Model
Measured vs. Classical and Improved Model
Measured Device vs. Modeled Q-Factor
Measured Device vs. Modeled Q-Factor
Figure #2.2.2-4: Classical and Improved Equivalent Circuit Model and Measured
vs. Modeled Data
In the paper by Koutsoyannopoulos, Y.K., and Papananos, Y., [15] a generic and
process independent model for simulating the performance of arbitrarily shaped and
multi-layer inductors on silicon substrates was developed. This method breaks-up the
modeled device into segments and adds a mutual inductance parameter to the lumped
20
element model. The figures on the right in Fig. 2.2.2-5 show all possible relative position
cases between two segments, this information is used to calculate the mutual inductance
between the traces. The Spiral Inductor Simulation Program (SISP) is generic in terms of
geometry and technology and performs a fast segment-by-segment extraction of the
equivalent SPICE subcircuit within seconds [15]. The performance prediction was
successful and accurate for the inductance and quality-factor and critical resonance
frequency. The relative percent errors were below 2% for (L) and (Q), below 5% for
resonance frequency. Fig. 2.2.2-5 shows the segment configuration, equivalent circuit,
magnetically coupled configurations, and measured vs. modeled data for the inductance
and Q-factor of a square spiral and octagonal inductor.
GsCs GsCsRsub
Nth Segment
SubstrateModeling
L R
Coupled to imageCurrent
Cp CpInsulator
Capacitance
K
R
Coupled to AllSegments Cf2
To AdjacentSegment Nodes
Zin Segment
Cf1
Equivalent Two-Port for One Segment of Spiral Inductor
Equivalent Two-Port for One Segment of Spiral Inductor
ρ
l
m
(a) (b)
ρ
m
ρ q
l
(c)
φ
l y
m(d)
ρ φ
l B
DC
v
A
m
µ
(e)
da
vm
D
BAφ
C
(f)
ρφ
l
D
C
Am
µ1
2m
Magnetically Coupled Segments in Special
Configurations
Magnetically Coupled Segments in Special
Configurations
Well
W S
d
thsio2
hsi
Parallel-Coupled Spiral SegmentsOver Si Substrate
Parallel-Coupled Spiral SegmentsOver Si Substrate
Comparison between Measured and Modeled L and QComparison between Measured and Modeled L and Q
Figure #2.2.2-5: Segments, Equivalent Circuit, Magnetically Coupled Segment
Configurations, and Measured vs. Modeled Data
21
2.2.3 Polynomial Equation Models
Polynomial equation models can be extracted using two different methods. The
first method uses a high permittivity circular insert to model substrate effects and
calculate capacitance.
In the paper by Delaney, K., Barrett, J., Barton, J., and Doyle, R., [17A] a high
permittivity circular insert is used to enhance component performance and to help model
substrate effects. A proportional relationship between capacitance and number of inserts
was verified [16]. Three methods were used to complete the full electrical modeling
analysis; analytical modeling, boundary element modeling, finite element modeling. The
finite element method provided the best results however was very slow. A computer
program was written to apply the analytical model, which became the fastest performance
predictor of the three methods investigated. The analytical method calculates the
capacitance contribution of the insert and the LTCC material, and the MICAP program
calculates the fringing effects at interfaces between the two materials; whereas, defects
due to non-optimized process conditions were predicted. Fig. 2.2.3-1 shows the lumped
element equivalent circuits for a buried capacitor, resistor, and the equations used to
determine the element values.
22
Z R1R2
1 ω2R2
2. C2.1i ωL.
ωR22. C.
1 ω2R2
2. C2.=
C α β log ω( )2. λ log ω( )
3.=
D A Blog ω( )4. CD. log ω( )
12.=
CT KN. k. Nper.
E0ψ.
4 d.D2 λ D
tan θ( ) λ D( ).=
CPLATEε DE0
.
tLW.( ) Nψ. γ tan θ( ). γ tan θ D( ).
t tan θ( ). D2
4.=
Rc
Cg
Rg
Ci
Ri
RL
Ci
Ri
R
Resistor
Glass Matrix
InterfacialBarrier
Conductive Particle
[A]
[B]
Resistor
InterconnectingConductors
Equivalent Circuits for Buried Resistor
Equivalent Circuits for Buried Resistor
RcLc
Cp
Rp
CapacitorInterconnecting
Conductors
Equivalent Circuits for Buried Capacitor
Equivalent Circuits for Buried Capacitor
Lumped Element EquationsLumped Element Equations
Figure #2.2.3-1: Lumped Element Equivalent Circuit for Buried Capacitor and
Resistor, and Lumped Element Equations
The second method uses an interpolation technique and the ratio of two
polynomials to deembed lumped element values.
In the paper by Kwang Lim Choi, and Swaminathan, M., [17] a broadband
transfer function of any structure as a ratio of two polynomials is generated using a
minimum number of either measured or simulated data [17]. All lumped element
equivalent circuits or transfer functions generated may be used in Spice and simulated in
both frequency and time domain. Fig. 2.2.3-2 shows the geometrical representation,
equivalent circuits, material and physical parameters, and polynomial equation of a three
turn spiral inductor.
23
Zin s( )a 0 a 1 s. a 2 s2. a3 s3.
b 0 b 1 s. b 2 s2. b 3 s3.=
1114
106
98
8
Metal width = 8
Metal Dimensions of Inductor in Mils
Metal Dimensions of Inductor in Mils
= 5.6
tan δ= 0.001 α= 3e7 S/m
10
10
= 5.6
tan δ= 0.001
Metal
Ground
0.3
0.270 pF
Port 1
0.793 Ω
0.431 Ω
0.399 pF
6810 Ω
419940 Ω
3.543 nH
9.532 nH
εr
εr
Equivalent Circuits and Material and Physical Parameters of Inductor in MilsEquivalent Circuits and Material and
Physical Parameters of Inductor in Mils
Polynomial EquationPolynomial Equation
Figure #2.2.3-2: Geometrical Representation, Equivalent Circuit, and Polynomial
Equation Model for Three-Turn Spiral Inductor
2.3 Statistical Variation Prediction
Process parameters in contrast to electrical parameters are the optimum basis for
statistical circuit modeling of passive devices [21]. Therefore, to statistically model a
device material properties such as oxide and metal thickness variations and physical
properties such as dielectric constant and resistivity variations are required. In
methodologies that utilize lumped elements models, process variations equate to element
value variations.
For element values deembedded using Measurement the process would be to take
measurements of multiple devices, find a uniform parameter distribution in the lumped
element values, and insert percent device deviations into the MonteCarlo simulation tool.
The MonteCarlo simulation tool generates, a range of device performance.
24
For element values deembedded using Physical Equations metal deposition
variations would need to be obtained and inserted into the modeling equations. A large
number of lumped element models would need to be created and the individual element
values examined to calculate a uniform parameter distribution. The process follows the
method described above to obtain overall model performance.
For element values deembedded using EM Simulation and Algorithms process
variations would be entered into the calculations yielding a large number of lumped
element models. As in the categories above a uniform parameter distribution is generated
and MonteCarlo simulations run to accurately predict a device range of operation.
Ideally, the designer would like to demonstrate that the variation in the equivalent
circuit models, based only on the data used to extract them, can be used to predict
variations in actual fabricated devices; therefore, providing an overall accurate
representation of the device performance.
2.4 Discussion
A unique variety of methods for modeling passive devices using lumped element
models have been presented. All methods preformed their functions in diverse ways with
varying computational requirements and goals. While some methods focused on sub-
dividing the device and modeling just that sub-division, others moved towards scalability
and optimization of design space. Still, others focused on enhancement techniques
applied to existing lumped element models and substrates by adding inductive coupling
and substrate inserts. Finally, there was the neural network approach and the broadband
25
transfer function technique. Each methodology added its own uniqueness to the process
of deembedding lumped element values for insertion into a circuit solver.
To arrive at an optimum modeling methodology means defining the trade offs
between speed and accuracy for the different methods, as shown in Fig. 2.4-1.
Method Speed Accuracy NotesPhysical Equations Excellent Moderate Minimal high frequency information.
Physical Equation and Measurement
Very Good Very Good Optimizing only substrate effects.
EM Simulations Slow Excellent Slow due to 3-D field equations, however very accurate.
Algorithms Excellent Good Approximations and assumptions, loss in accuracy.
Measurement Moderate Excellent Contains actual information from substrate (non-ideal material properties) and device.
Figure #2.4-1: Speed/Accuracy Summary for Modeling Methodologies
A purely Physical Equation based method is fast, however does not model high
frequency behavior well. An Algorithm based method is fast and models high frequency
behavior well due to its close ties to numerical methods; however, has no way of
modeling non-ideal material properties. An EM solver based method equates to high
accuracy however speed is definitely lost in 3-D field equation calculations. A
Measurement based method models high frequency behavior well and contains nonideal
material properties, which would be impossible to model mathematically; however,
convergence problems can be open ended in terms of time and the device must be
fabricated before it can be modeled.
It appears that, to develop the consummate passive device modeling methodology,
a modeling method that employs the accuracy of “measurement” coupled with the speed
of “physical equations” must be used.
26
2.5 Summary
The methods developed for the characterization, modeling, and design of
microwave and millimeter-wave passive structures, and an overview of various lumped
element modeling methods and how they apply to statistical prediction was presented in
this chapter. The next chapter will discuss in detail a novel modeling methodology used
as the basis for this research. A test case will be provided and will demonstrate the
accuracy and speed of this modeling methodology.
27
CHAPTER 3
PASSIVE DEVICE MODELING METHODOLOGY
3.1 Introduction
The point of departure for this research is the work done by Dr. Ravi Poddar, Dr.
Nan Jokerst and Dr. Martin A. Brooke in March of 1997 entitled “Accurate, High Speed
Modeling of Integrated Passive Devices” [22]. The fundamental idea behind the
modeling procedure (proposed in [22]) is that most designed passive structures are
comprised of several key geometrical building blocks, that is, they can be constructed
from several building block cells representing individual parts of the structure. These
building blocks can be defined in a number of ways, but careful selection can result in
relatively few building blocks needed. The theory being, if accurate models for each of
the building blocks along with interaction information can be obtained, then any structure
comprised of those building blocks can be modeled accurately using the individual block
models [23].
The building block equivalent circuits are derived from fabricated test structures
and measurements using optimization and extraction routines. Passive RLC models for
each embedded building block are extracted and take into account effects of process
nonidealities and unknown material properties. The objective here is to predict the
28
electrical behavior of arbitrary geometry passive devices in a standard circuit simulator,
thus enabling a major increase in modeling speed over methods that do not utilize lumped
elements.
To be a practical method it is necessary that the number of test structures needed,
to accurately model all of the building blocks required, be as few as possible. In the test
case detailed below, the computed fundamental building block models were used to
predict the behavior of a 9-segment meander resistor, and will show that the number of
test structures can be much fewer than the number of parameters and building blocks
used. Specifically, four building blocks will be modeled with 18 parameters and
extracted from only two test structures. The test case will demonstrate that the predictive
capability of the resulting models for new devices, fabricated from the building blocks, is
both accurate and very fast.
3.2 Design and Modeling Flowchart
The modeling methodology can be described in a flow diagram, as shown in Fig.
3.2-1. The first step in the modeling process is to identify what devices to model in a
process, and identify building blocks and design rules. The next step is to design and
fabricate test structures to characterize. Next, the test structures are characterized for
building block circuit extraction. The measured data is then used to set up optimizations
and determine initial guesses to extract equivalent circuits of the test structures and
building blocks. Once successful optimizations have been achieved, then the building
blocks with associated models and design rules are combined in a library.
29
Accurate ResultsAccurate Results
Simulate in SPICESimulate in SPICE--type type simulatorsimulator
DRCDRC
Design Desired Passive Design Desired Passive DeviceDevice
Canonical Block/Design Canonical Block/Design Rule LibraryRule Library
Generate Design Rules Generate Design Rules Associated with Each Associated with Each
BlockBlock
Model CanonicalModel CanonicalBlocksBlocks
PassPassPass
FailFailFail
Design &Design & FabFabTest StructuresTest Structures
Figure #3.2-1: Design and Modeling Flowchart
3.3 Detailed Resistor Modeling Procedure
3.3.1 Devices Modeled
Most resistors are designed in straight lines or serpentine structures; however, the
layout of choice for high frequency applications is usually straight lines. It is attractive to
explore serpentine or meander resistors for high frequency applications for several
reasons. First, meander resistors are more efficient in substrate area for the same
resistance value, and, if modeled correctly, could have larger application in the high
frequency arena. Secondly, the meander structure presents a difficult modeling problem
due to higher levels of parasitics, such as coupling effects between the segments of the
structure, potentially affecting the overall system response [24].
30
3.3.2 Identify Building Blocks
In the test case below, meander resistors fabricated with 25 µm linewidths and
spacings were considered. The meander (serpentine) geometry dictated three
fundamental building blocks that required characterization: 1.) A square building block
with connections on opposite sides; 2.) A U-shaped section connecting two parallel
segments of the resistor together; and 3.) A coupled block segment to characterize line-
to-line coupling behavior on a per square basis. Due to testing requirements, a probe pad
was added as a fourth building block. Coupling was only considered with respect to
nearest neighbors. A current flow visualization tool could determine the building block
sizes [25].
Current visualization software was used to predict current flow through a
representative serpentine resistor, shown in Fig. 3.3.2-1. The cutoff points for each of the
primitives were at the areas where the current contours stopped changing rapidly,
indicating constant current flow between the boundaries. Using this approach, the pad
primitive was taken to be the large pad square plus one adjacent line square, the material
square and coupled material square were taken as one unit of material square each, and
the U shaped primitive was represented by 3 squares on each of the horizontal and
vertical axes of the U shape.
31
Figure #3.3.2-1: Current Distribution Plot of Serpentine Resistor Structure
3.3.3 Equivalent Circuits
Each building block is modeled as a Spice compatible RLC circuit. For simple
uncoupled building blocks, such as a piece of straight line, simple RLC models based on
the partial element equivalent circuit (PEEC) are used for equivalent circuits [26]. In the
case of blocks where coupling is examined, e.g. coupled material squares (corresponding
to coupled adjacent lines or interacting material squares), coupled PEEC models joined
by coupling capacitances and mutual inductances are used, as shown in Fig. 3.3.3-1.
32
Building BlocksBuilding Blocks1 3
2 4
1 2
R=0 .09 Oh mL= 1E -11 HC = 1.1e-1 5 FCM =0.4 e- 15 FCC = 1.4e-1 5 FLM = 0 .20
LC
R L R
C C CC
LC
R L R
C C CC
LM
LMCM
1
2
3
4R=0 .08 OhmL= 1 E- 11 HC = 2.7 e-15 FCC = 1. 2e -15 F
L
C
R L R
CC C C
1 2
PEECsPEECs
Figure #3.3.3-1: Uncoupled and Coupled PEEC Circuits With Associated Building
Blocks
3.3.4 Test Structures
Once the building blocks have been defined, the next step is to characterize and
develop models. This is achieved through the use of carefully designed test structures.
The test structure set is designed to include all of the predefined building blocks. The
equivalent circuit of the test structure is made up of the equivalent circuits of each
building block. Once designed and fabricated in the process of interest, high frequency
measurements are taken and used to characterize each of the building blocks [23].
By virtue of device measurements, no assumptions are made regarding material
properties, layered dielectrics, or imperfect substrates, because all of these effects are
represented in the measured data. This makes the modeling procedure entirely process
independent.
33
To model the four stated building blocks, two test structures were fabricated, as
shown in Fig. 3.3.4-1. The first test structure is simply a line with probe pads on the two
ends. It is used to characterize basic uncoupled material parameters, including self-
resistance, inductance, and capacitance.
The second test structure is a 3-segment meander resistor. This structure allows
passive characterization of the U-shaped corner segments as well as line-to-line mutual
inductance and coupling capacitance.
Test Structure 1Test Structure 1
Coupled Square PrimitiveCoupled Square Primitive
U-Shaped BendPrimitive
U-Shaped BendPrimitive
Pad PrimitivePad PrimitiveMaterial Square PrimitiveMaterial Square Primitive
Test Structure 2Test Structure 2
Figure #3.3.4-1: Test Structures and Primitives for Meander Resistor Modeling
It is interesting to note that a structure with first order coupling is actually a 4-port
structure, whereas the test structures are only 2-port devices. Thus only standard,
repeatable 2-port measurements are necessary. A 4-port device is more difficult to
measure in practice than a 2-port, since many different excitation and loading iterations
are required. The method of simply measuring two port structures and extracting all
required multiport information is a significant advantage over attempting to measure
coupling between physically disconnected devices [23].
34
3.3.5 Processing and Measurement
The test structure design was fabricated by depositing gold on a 96 % alumina
substrate, which had a surface roughness of approximately +/- 1.5 µm. A
photomicrograph of the fabricated test structures is shown in Fig. 3.3.5-1.
The test structures were measured using network analysis techniques, a DC curve
tracer, and a high precision multimeter. For the high frequency measurements, a HP
8510C network analyzer was used with a Cascade Microtech probe station and ground-
signal-ground configuration probes. Calibration was accomplished using a calibration
substrate and utilization of the line-reflect-match (LRM) calibration method. Data was
gathered for each of the test structures at over 200 frequency points between 500MHz
and 20GHz, and were stored with the aid of computer data acquisition software and
equipment. DC I-V measurements of the test structures were also made in order to
ascertain component resistances. At DC, parasitic capacitance and inductance have no
effect on the response and the measured resistance value, and, once properly apportioned,
can be used directly in the models of the building blocks.
35
Figure #3.3.5-1: Photograph of Fabricated Structures for Meander Resistor
Modeling
3.3.6 Modeling and Parameter Extraction
Circuit models must be generated for each of the defined building blocks. The
circuit topologies and parameters for the uncoupled and coupled building blocks are
shown in Fig. 3.3.6-1. These circuits represent models for the building blocks only; the
test structures and resistor circuits are comprised of many building block circuits
connected in accordance with the structure geometry.
36
Test Structure 1
Deembed MaterialSquare
Pad
R =0 . 08 O h mL = 1 E- 11 HC = 2.7 e- 1 5 FC C = 1.2e - 15 F
L
C
R L R
C C C C
1 2
R =0 .0 8 O hmL = 1 E- 11 HC = 2.7 e- 15 FC C = 1. 2e - 15 F
L
C
R L R
C C C C
1 2
Test Structure 2
Deembed
1Block
Coupled Pair
PEECsPEECs
R =0 .09 O h mL = 1 E -1 1 HC = 1 .1 e -1 5 FC M =0 .4 e - 1 5 FC C = 1 .4 e-1 5 FL M = 0 .2 0
LC
R L R
C C CC
LC
R L R
C C C C
L M
L MCM
1
2
3
4
R =0 . 08 O h mL = 1 E- 11 HC = 2.7 e- 1 5 FC C = 1.2e - 15 F
L
C
R L R
C C C C
1 2
Building BlocksBuilding Blocks
Figure #3.3.6-1: Building Blocks, Equivalent Circuits and Parameter Values for
Meander Resistor Modeling
The extraction of the circuit model parameters was achieved in several steps. Due
to the highly nonlinear nature of the generated system equations with respect to circuit
parameter values, a procedure of hierarchical optimization with respect to measured s-
parameter and DC resistance data was chosen [27]. All optimizations and simulations
were done using the Hspice circuit simulator on Sun Sparcstation 20 series workstations.
The starting point or initial guess of the circuit parameters was crucial for correct
optimization results. Therefore, an initial optimization was done assuming that each test
structure was comprised of one building block utilized repetitively across the length of
the structure on a per square basis. This method was very effective for obtaining a good
starting point for the optimization of the test structure circuits. Optimizations were
37
performed for frequencies up to 10 GHz, and results for the meander resistor test
structure (shown in Fig. 3.3.4-1) are shown in Fig. 3.3.6-2.
0
0.05
0.1
0.15
0.2
0.25
1.00E+08 1.00E+09 1.00E+10 1.00E+11
Frequency (Hertz)
S11
S11(R) MeasuredS11(R) ModeledS11(I) MeasuredS11(I) Modeled
Real
Imaginary
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
1.00E+08 1.00E+09 1.00E+10 1.00E+11
Frequency (Hertz)
S21
S21(R) Measured S21(R) ModeledS21(I) MeasuredS21(I) Modeled
Real
Imaginary
Figure #3.3.6-2: Measured vs. Modeled Results for the Meander Resistor Test
Structure: (a) S21 Real and Imaginary Response; (b) S11 Real and Imaginary
Response
3.3.7 9-Segment Meander Resistor
The computed fundamental building block models were used to predict the
behavior of a 9-segment meander resistor. An equivalent circuit of the resistor was
constructed by replacing each building block in the structure with its equivalent extracted
circuit, as shown in Fig. 3.3.7-1. Since only first level coupling was taken into account,
38
each material square in each segment of the resistor was coupled to its nearest neighbor
by a pair of mutual inductances and a coupling capacitance. As inferred from the circuit
description, the resulting circuit using 25µm linewidths and 300µm line lengths per
segment 9-segment resistor was a complex, highly interconnected system, consisting of
approximately 700 nodes. The longest path length of the resistor was approximately 0.35
wavelengths long at 10GHz.
Coupled LineBuilding Block
U Building Block
Uncoupled LineBuilding Block
Figure #3.3.7-1: Meander Resistor and Associated Building Blocks
3.4 Results
In spite of the large circuit size, the AC small signal circuit simulation was
completed in less than 2 minutes. The predicted s-parameters, shown in Fig. 3.4-1, were
compared to measured values for the same structure. Both real and imaginary parts of S11
and S21 were accurately predicted up to 5GHz. In comparison, the same structure was
39
designed and simulated in a method of moments solver with a 3GHz meshing frequency.
The structure required 72 minutes to complete, while consuming approximately 50MB of
system memory while utilizing 2 processors in a multiprocessing Sun workstation. Thus,
for this example, a speedup factor of approximately 35 was demonstrated.
-0 .1
0
0 .1
0 .2
0 .3
0 .4
0 .5
0 .6
0 .7
1 .0 0 E +0 8 1 .0 0 E +0 9 1 .0 0 E +1 0 1 .0 0 E +1 1
F requency (H ertz )
S11
S 1 1 (R ) M e a s ure dS 1 1 (R ) P re dic te d (This P a pe r)S 1 1 ( I) M e a s ure dS 1 1 ( I) P re dic te d (This P a pe r)
-0 .8
-0 .6
-0 .4
-0 .2
0
0 .2
0 .4
0 .6
0 .8
1
1 .0 0 E +0 8 1 .0 0 E +0 9 1 .0 0 E +1 0 1 .0 0 E +1 1
F requency (H ertz)
S21
S 2 1 (R ) M e a s ure dS 2 1 (R ) P re dic te d (This P a pe r)S 2 1 ( I) M e a s ure dS 2 1 ( I) P re dic te d (This P a pe r)
R ea l
Im ag ina ry
R ea l
Im ag ina ry
Figure #3.4-1: Measured vs. Predicted Results for 9-Segment Meander Resistor: (a)
S21 Real and Imaginary Response. (b) S11 Real and Imaginary Response
40
3.5 Summary
This methodology shows that with enough frequency points and sufficient
sensitivity, accurate equivalent circuit extraction of individual building blocks from multi
building block test structures is possible. The fact that a non-singular system of
equations can be formulated for a linear network, with each equation generated by a
different frequency point, leads to the idea that simply by sampling the system over
frequency, we can obtain all the information necessary to deembed the various circuit
parameters which comprise that system [28]. Therefore, many more than one unknown
sets of equivalent circuit parameters can be de-embedded from the measurements of one
test structure. This theory will be proven for the same 9-segment meander resistor and
detailed in Chapter 5.
41
CHAPTER 4
PASSIVE DEVICE MODELING METHODOLOGY
IMPROVEMENTS
4.1 Introduction
Ideally, the designer would like to demonstrate that the variation in the equivalent
circuit models, based only on the data used to extract them, can be used to predict
variations in the actual fabricated device. Therefore, the more accurate the measured vs.
modeled results of the test structures, the more accurate the modeled device, and
consequently the tighter and more realistic the predicted range of operation.
The passive device modeling methodology described in Chapter 3 has been very
successful in modeling both 2D and 3D passive device structures. However, the
problems with this method lie in the occasional long convergence time experienced when
using the optimization routine. An initial guess and a range of possible element values is
required by the optimizer; whereas, inaccurate initial guesses coupled with non-inclusive
element value ranges and an excessive number of optimized elements, lead to long
convergence times or no convergence at all.
One solution to this problem is to decrease the number of optimized element
values by using physical equations (Table 4.1-1) to calculate some of the lumped element
42
circuit values. In the paper by Ronkainen, H., Kattelus, H., Tarvainen, E., Ruhisaari, T.,
and Andersson, M., entitled “IC compatible planar inductors on silicon,” a more physical
semiempirical model was developed. Several components in the lumped element
equivalent circuit, shown in Fig. 4.1-1, were modeled by functions related to inductor
layout and material parameters, and the remaining were deembedded through circuit
optimization [6].
Cint
R L
Csub1 Csub2
C1 C2
Rsub1Rsub2
Rsub12
nin nout
Semi-Empirical Sub circuit Model for Planar InductorsSemi-Empirical Sub circuit Model for Planar Inductors
Figure #4.1-1: Semi-Empirical Lumped Element Circuit Model
This equivalent circuit and the method used to arrive at the circuit values was
applied to the modeling method described in Chapter 3. The equations in Table 4.1-1
were also singularly applied to the existing elements of the PEECs, from Chapter 3, in an
attempt to increase measured vs. modeled test structure accuracy.
43
Equation Number
L 1.5 µ 0. N2. D. e
3.7 N 1( ). W S( ).
D. DW
0.1.=
µ 0 is vacuum permeability, N number of turns, D length of side of rectangular inductor, and S and W are spacing and width of inductor traces, respectively
(1)
R ρlA
. 1 k f.( )=
ρ is resistivity of inductor metallisation, l length of inductor, A cross-section of wire, and k is empirical parameter
(2)
C 1 C 2 ε oxA ind A pads
t ox.=
ε r is permitivity of oxide, Aind ( Apads ) total area of inductor (pads), and tox oxide thickness
(3)
Table #4.1-1: Semiempirical Expressions for Inductor Model
The outcome from both attempts yielded less than accurate results. It was
apparent that the circuit topology in Fig. 4.1-1 could not hold the circuit behavior
information necessary to model the test structures, into the GHz. After this attempt
failed, the focus on improving the building block equivalent circuit model was contained
within the existing PEECs’.
After thorough investigation, improvements to the overall existing methodology
detailed in Chapter 3 were observed and implemented. These improvements resulted in a
substantial increase in measured vs. modeled accuracy for the test structures as well as
the modeled device. This chapter will follow the format of chapter 3 in detailing the
improvements made to the original methodology. The 9-segment meander resistor will
44
be modeled again using both versions (original & improved) of the methodology, and
will display the modeled vs. an actual measured device results.
4.2 Design and Modeling Flowchart
The improvements to the novel passive device modeling methodology are
described using the same flow chart with the enhancements localized in the “Model
Canonical Blocks” section, shown in Fig. 4.2-1. The first step, as before, in the modeling
process is to identify what devices to model in a process, and identify building blocks and
design rules. The next step is to design and fabricate test structures to characterize.
Next, the test structures are characterized for building block circuit extraction; this is one
of the enhanced areas. The measured data is then used to determine initial guesses and to
set up optimizations to extract equivalent circuits of the test structures and building
blocks; this is also enhanced. Once successful optimizations have been achieved, then
the building blocks with associated models and design rules are combined in a library.
45
DRCDRC
Accurate ResultsAccurate Results
Simulate in SPICESimulate in SPICE--type simulatortype simulator
Design Desired Design Desired Passive DevicePassive Device
Canonical Block/Design Canonical Block/Design Rule LibraryRule Library
Generate Design Rules Generate Design Rules Associated with Each BlockAssociated with Each Block
Model CanonicalModel CanonicalBlocksBlocks
PassPassPass
FailFailFail
Design &Design &FabFabTest StructuresTest Structures
ImprovementsImprovements
Figure #4.2-1: Design and Modeling Flowchart
4.3 Detailed Resistor Modeling Procedure
4.3.1 Identify Building Blocks
As stated previously, the serpentine (meander) geometry of the resistor dictated
four fundamental building blocks that required characterization: 1.) A square building
block with connections on opposite sides; 2.) A U-shaped section connecting two
parallel segments of the resistor; and 3.) A coupled block segment to characterize line-
to-line coupling behavior on a per square basis; and 4.) A probe pad to connect to the test
equipment. Nearest neighbor coupling was also included between the material squares.
The current visualization software used by Dr. Poddar to predict current flow
through a representative serpentine resistor is shown in Fig. 4.3.1-1. The cutoff points for
46
each of the primitives are at the areas where the current contours stop changing rapidly,
indicating constant current flow between the boundaries. Using this approach, the pad
primitive was taken to be the large pad square plus one adjacent line square, the material
square and coupled material square were taken as one unit of material square each, and
the U-shaped primitive was represented by 3 squares on each of the horizontal and
vertical axes of the U shape. However after close examination, it was observed that the
current contours are continually changing throughout the corner primitive. Therefore
instead of modeling them as one block, each piece needs to be individually deembedded
to accurately reproduce the complex electrical behavior as shown in Fig. 4.3.1-1.
Close-UpClose-Up
Test Structure 2Test Structure 2U-Shaped BendBuilding BlockU-Shaped BendBuilding Block
U-Shaped BendBuilding BlockU-Shaped BendBuilding Block
Figure #4.3.1-1: Current Distribution Plot of Meander Resistor Structure
47
4.3.2 Equivalent Circuits
As stated previously, each building block is modeled as a Spice compatible RLC
circuit. For simple uncoupled building blocks, such as a piece of straight line, simple
RLC models based on the partial element equivalent circuit (PEEC) are used for
equivalent circuits [26]. In the case of blocks where coupling is examined, e.g. coupled
material squares, coupled PEEC models joined by coupling capacitances and mutual
inductances are used. The corner primitive was reevaluated and modeled as three
independent PEECs’; 1.) One for the input/output; 2.) One for the two corners; and 3.)
One between the corners. The test structures, the improved building block equivalent
circuits, and the respective PEECs’ are all shown in Fig. 4.3.2-1.
Test Structure 1Test Structure 1
Coupled Square PrimitiveCoupled Square Primitive
U-Shaped BendPrimitives
U-Shaped BendPrimitives
Pad PrimitivePad Primitive Material Square PrimitiveMaterial Square Primitive
Test Structure 2Test Structure 2
Building Blocks / PEECsBuilding Blocks / PEECs
1 3
2 4
R=0 .09 OhmL= 1E- 11 HC = 1.1e-1 5 FCM=0.4e-15 FCC = 1.4e-1 5 FLM = 0.20
LC
R L R
C C CC
LC
R L R
C C C C
L M
LMC M
1
2
3
4
1 2
R =0 .08 O hmL= 1 E- 11 HC = 2.7 e- 15 FCC = 1. 2e- 15 F
L
C
R L R
C C C C
1 2
1 2
R =0.08 OhmL= 1 E- 11 HC = 2.7 e- 15 FCC = 1.2e- 15 F
L
C
R L R
C C C C
1 2
.
R= 0.4 O hmL= 3 .7 E- 11 HC = 5.3e-1 5 FCC=2 7e- 15 F
L
C
R L R
CC CC
1 2
CC CC
2L
C
R L R1
1
2
L
C
R L R
CC CC
1 2
L
C
R L R
CC CC
1 2
L
C
R L R
CC CC
1 2
Figure #4.3.2-1: Uncoupled and Coupled PEEC Circuits with Associated Building
Blocks
48
4.3.3 Test Structures
The building blocks are defined as in Chapter 3, and include the improvement to
the U-shaped bend primitive. The developed models are characterized using the same
carefully designed test structures, which include all of the predefined building blocks.
The structures are designed, fabricated, and measured to characterize each of the building
blocks.
By virtue of device measurements, no assumptions are made regarding material
properties, layered dielectrics, or imperfect substrates, because all of these effects are
represented in the measured data. This makes the modeling procedure entirely process
independent.
To model the four stated building blocks, the same two test structures were
fabricated, as shown in Fig. 4.3.2-1. The first test structure is a line with probe pads on
the two ends, and is used to characterize the basic uncoupled material parameters,
including self-resistance, inductance, and capacitance. The second test structure is a 3-
segment meander resistor. This structure includes the enhanced characterization of the
entire U-shaped corner segments as well as line-to-line mutual inductance and coupling
capacitance.
4.3.4 Processing and Measurement
The test structure design was fabricated by depositing gold on a 96% Alumina
substrate. A photomicrograph of the fabricated test structures is shown in Fig. 4.3.4-1.
The test structures were measured using network analysis techniques. For the
high frequency measurements, a HP 8510C network analyzer was used with a Cascade
49
Microtech probe station and ground-signal-ground configuration probes. Calibration was
accomplished using a calibration substrate and utilization of the line-reflect-match (LRM)
calibration method. Because measured vs. modeled accuracy is directly related to the
number of evaluated frequency points, data was gathered for each of the test structures at
over 400 frequency points between 45MHz and 20GHz, and were stored with the aid of
computer data acquisition software and equipment.
Test Structure 2Test Structure 2Test Structure 1Test Structure 1
9-Segment Meander Resistor9-Segment Meander Resistor
Figure #4.3.4-1: Photograph of Fabricated Structures for Meander Resistor
Modeling
4.3.5 Modeling and Parameter Extraction
Circuit models must be generated for each of the defined building blocks. The
circuit topologies and parameters for the uncoupled and coupled building blocks, pad
building block, and improved multi-block U-shaped bend are shown in Fig. 4.3.5-1.
These circuits represent models for the building blocks only; the test structures and
50
resistor circuits are comprised of many building block circuits connected in accordance
with the structure geometry.
Test Structure 1
Deembed MaterialSquare
Pad
R = 0 . 08 O h mL = 1 E - 1 1 HC = 2 .7 e- 1 5 FC C = 1 . 2 e - 15 F
L
C
R L R
C C C C
1 2
R =0 .0 8 O h mL = 1 E - 1 1 HC = 2.7 e - 1 5 FC C = 1. 2 e - 15 F
L
C
R L R
C C C C
1 2
Test Structure 2
Deembed
Multi-Block
Coupled Pair
PEECsPEECs
R = 0 . 0 9 O h mL = 1 E -1 1 HC = 1 .1 e -1 5 FC M = 0 .4 e - 1 5 FC C = 1 .4 e-1 5 FL M = 0 .2 0
LC
R L R
C C CC
LC
R L R
C C C C
L M
L MC M
1
2
3
4
Building BlocksBuilding Blocks
.
R = 0 . 4 O h mL = 3 .7 E- 1 1 HC = 5 .3 e -1 5 FC C = 2 7 e - 1 5 F
L
C
R L R
CC CC
1 2
CC CC
2L
C
R L R1
L
C
R L R
CC CC
1 2
L
C
R L R
CC CC
1 2
L
C
R L R
CC CC
1 2
Figure #4.3.5-1: Building Blocks, Equivalent Circuits and Parameter Values for
Meander Resistor Modeling
In an attempt to increase the measured vs. modeled test structure accuracy, several
different aspects of the circuit model extraction process were performed, and are shown
in Fig. 4.3.5-2. Because the number of element values extracted and the speed of
convergence are inversely proportional, using the smallest number of optimized elements
51
is desired. Several different circuit configurations and number of deembedded element
values were applied to test structures one and two, and analyzed using the Hspice
Optimization Tool in an attempt to evaluate speed vs. accuracy trade offs.
Test Structure 1
Test Structure 2
Deembed
Deembed
Different CircuitConfigurations
Different CircuitConfigurations
1. Semi-Empirical
2. No Substrate Resistance
3. Substrate Resistance Included
1. Semi-Empirical
2. No Substrate Resistance/1Block Corner
3. Substrate Resistance Included/1Block Corner
4. No Substrate Resistance/MultiBlock Corners
5. Substrate Resistance Included/MultiBlkCorners
Building Blocks1Block
Multi Block
Coupled Pair
MaterialSquare
Pad
Circuit Model Topologies
Figure #4.3.5-2: Test Structures and Primitives for Meander Resistor Modeling
To simplify accuracy evaluation, percent error between measured vs. modeled
was calculated at each frequency point and is shown in the plots below. After looking
deeper into the Hspice Optimization Tool [31], it was noticed that iterative optimization
runs increase measured vs. modeled accuracy. Therefore, Unix scripts were developed to
automatically run 18 iterative optimizations per measured device, while at the same time
increasing the accuracy settings and decreasing the range of possible element values
within the optimization circuit file. The building block characterization process executed
52
1152 individual optimizations to arrive at 32 unique equivalent circuit models of test
structure one, and 32 unique equivalent circuit models of test structure two.
The entire Optimization/MonteCarlo analysis process together with data
calculations, conversions, storage, and circuit file creation was automated using Unix
shell scripts. The detailed automated process will be showcased in Chapter 9.
4.3.6 Test Structure 1 Results
After inspecting the PEECs’ of test structure one and two used in Chapter 3, it
was found that the substrate resistance was held constant at 10Mega ohms while the
remaining building block element values were deembedded using the Hspice
Optimization Tool. Knowing the importance of substrate dependencies, this appeared to
be an area in need of improvement. In the paper by Arcioni, P., Castello, R., Perregrini,
L., Sacchi, E., and Svelto, F., [13] an improved equivalent circuit for silicon-integrated
inductors was presented. This model introduces a substrate resistance, which is
magnetically coupled with the inductor and accurately describes the loss mechanism in
both the metal and in the low-resistivity substrate. The separate characterization of metal
and substrate losses permits the designer to separately evaluate their contribution to the
inductor’s Q-factor. Therefore, the optimization of the PEEC’s with and without
deembedded substrate resistance was performed on test structure one and two for both
sets or 32 fabricated devices.
The percent error between the actual measured device and the modeled device
was calculated for test structure one, and is shown in Fig. 4.3.6-1 and Fig. 4.3.6-2. The
reflection coefficient (S11), which models the terminating impedance at the input of the
53
device, shows a 0.5% difference between no substrate resistance (no deembedded) and a
deembedded substrate resistance. There seems to be an improvement in both the high
and low frequency for the models that deembedded the substrate resistance.
Test Structure 1 No Substrate Resistance S11 Real Measured vs . Modeled
-0.5
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
1.E+08 1.E+09 1.E+10 1.E+11
Frequency
PercentError
Test Structure 1 Substrate ResistanceS11 Real Measured vs. Modeled
-0.5
0
0.5
1
1.5
22.5
3
3.5
4
4.5
1.E+08 1.E+09 1.E+10 1.E+11
Frequency
Percent Error
No Substrate Resistance Deembedded~3 percent Error
Between Measured vs. Modeled
No Substrate Resistance Deembedded~3 percent Error
Between Measured vs. Modeled
Substrate Resistance Deembedded~2.5 percent Error
Between Measured vs. Modeled
Substrate Resistance Deembedded~2.5 percent Error
Between Measured vs. Modeled
Figure #4.3.6-1: S11 Measured vs. Modeled Percent Difference for 32 Fabricated
Structures and 32 Modeled Structures
The transmission coefficient (S21) that models the actual impedance of the
overall device shows a 0.25% difference between no substrate resistance (no
deembedded) and a deembedded substrate resistance. An improvement can clearly be
seen at low frequency for the models that did not deembed the substrate resistance;
whereas, a high frequency improvement is shown for the models that deembedded the
substrate resistance. The dividing line that separates the low and high frequency points
seems to be centered at approximately 3.5GHz. At this point substrate parasitics that are
a function of physical properties such as dielectric constant and resistivity, and skin effect
which are a function of material properties, begin to influence the circuit behavior.
54
Test Structure 1 Substrate ResistanceS21 Real Measured vs. Modeled
-0.5
0
0.5
1
1.5
2
2.5
3
1.E+08 1.E+09 1.E+10 1.E+11
Frequency
Percent Error
Test Structure 1 No Substrate Resistancei S21 Real Measured vs. Modeled
-0.5
0
0.5
1
1.5
2
2.5
3
1.E+08 1.E+09 1.E+10 1.E+11
Frequency
PercentError
No Substrate Resistance Deembedded~0.5 percent Error
Between Measured vs. Modeled
No Substrate Resistance Deembedded~0.5 percent Error
Between Measured vs. Modeled
Substrate Resistance Deembedded~0.75 percent Error
Between Measured vs. Modeled
Substrate Resistance Deembedded~0.75 percent Error
Between Measured vs. Modeled
Figure #4.3.6-2: S21 Measured vs. Modeled Percent Difference for 32 Fabricated
Structures and 32 Modeled Structures
Due to the focus on low frequency behavior (i.e. less that 5GHz), the evaluation
of test structure one clearly points to “not deembedding” the substrate resistance.
4.3.7 Test Structure 2 Results
As stated previously, close examination of the current visualization plot found
that the current contours were continually changing throughout the corner primitive.
Therefore instead of modeling them as one block, each piece was individually
deembedded to accurately reproduce the complex electrical behavior of the corners.
Optimizations were then performed on test structure two; 1.) With and without the
substrate resistance; and 2.) Deembedding a single block corner vs. deembedding a multi-
block corner, for all 32 fabricated devices.
55
The percent error between the actual measured device and the modeled device
was calculated for test structure two, and is shown in Fig. 4.3.7-1 and Fig. 4.3.7-2. The
reflection coefficient (S11) for “no substrate resistance (no deembedded)/multiBlk
corner” shows a 1.5% improvement over “no substrate resistance (no deembedded)/1Blk
corner,” a 3.5% improvement over “substrate resistance (deembedded)/multiBlk corner,”
and a 7.5% improvement over “substrate resistance (deembedded)/1Blk corner.” The
building block model that “does not deembedded a substrate resistor” and that
“characterizes the additional corner primitives” is the optimum circuit model for S11 of
test structure two.
Test Structure 2 Substrate Resistance MultiBlk Corner S11 Real Measured vs. Modeled
-20
24
6
8
10
12
14
1.E+08 1.E+09 1.E+10 1.E+11
Frequency
Percent Error
Test Structure 2 Substrate Resistance 1Blk Corner S11 Real Measured vs. Modeled
-505
10152025303540
1.E+08 1.E+09 1.E+10 1.E+11
Frequency
Percent Error
Sub Resistance Deembedded MultiBlk Corner~6 percent Error
Between Measured vs. Modeled
Sub Resistance Deembedded MultiBlk Corner~6 percent Error
Between Measured vs. Modeled
Sub Resistance Deembedded 1Blk Corner~10+ percent Error
Between Measured vs. Modeled
Sub Resistance Deembedded 1Blk Corner~10+ percent Error
Between Measured vs. Modeled
Test Structure 2 No Substrate Resistance 1Blk Corner S11 Real Measured vs. Modeled
-2
0
2
4
6
8
10
12
14
1.E+08 1.E+09 1.E+10 1.E+11
Frequency
Percent Error
No Sub Resistance Deembedded MultiBlk Corner~2.5 percent Error
Between Measured vs. Modeled
No Sub Resistance Deembedded MultiBlk Corner~2.5 percent Error
Between Measured vs. Modeled
No Sub Resistance Deembedded 1Blk Corner~4+ percent Error
Between Measured vs. Modeled
No Sub Resistance Deembedded 1Blk Corner~4+ percent Error
Between Measured vs. Modeled
Test Structure 2 No_Multi S11 Real Measured vs. Modeled
-202468
101214
1.E+08 1.E+09 1.E+10 1.E+11
Frequency
Percent Error
Figure #4.3.7-1: S11 Measured vs. Modeled Percent Difference for 32 Fabricated
Structures and 32 Modeled Structures
56
The transmission coefficient (S21), which models the actual impedance of the
overall device, shows an overall percent error of 1.0%. As seen in the S21 plots, the
building block model that “does not deembedded substrate resistance” and that
“characterizes the additional corner primitives” is the optimum circuit model for S21 of
test structure two.
Test Structure 2 No Substrate Resistance 1Blk Corner S21 Real Measured vs. Modeled
-2
0
2
4
6
8
10
12
1.E+08 1.E+09 1.E+10 1.E+11
Frequency
Percent Error
Test Structure 2 Substrate Resistance MultiBlk Corner S21 Real Measured vs. Modeled
-2
0
2
4
6
8
10
12
1.E+08 1.E+09 1.E+10 1.E+11
Frequency
PercentError
Test Structure 2 Substrate Resistance 1Blk Corner S21 Real Measured vs. Modeled
-5
0
5
10
15
20
1.E+08 1.E+09 1.E+10 1.E+11
Frequency
Percent Error
No Sub Resistance Deembedded / MultiBlk Corner~1.0 percent Error
Between Measured vs. Modeled
No Sub Resistance Deembedded / MultiBlk Corner~1.0 percent Error
Between Measured vs. Modeled
Sub Resistance Deembedded / MultiBlk Corner~3+ percent Error
Between Measured vs. Modeled
Sub Resistance Deembedded / MultiBlk Corner~3+ percent Error
Between Measured vs. Modeled
No Sub Resistance Deembedded / 1Blk Corner~1.5+ percent Error
Between Measured vs. Modeled
No Sub Resistance Deembedded / 1Blk Corner~1.5+ percent Error
Between Measured vs. Modeled
Sub Resistance Deembedded / 1Blk Corner~10+ percent Error
Between Measured vs. Modeled
Sub Resistance Deembedded / 1Blk Corner~10+ percent Error
Between Measured vs. Modeled
Test Structure 2 No_Multi S21 Real Measured vs. Modeled
-202468
1012
1.E+08 1.E+09 1.E+10 1.E+11
Frequency
PercentError
Figure #4.3.7-2: S21 Measured vs. Modeled Percent Difference for 32 Fabricated
Structures and 32 Modeled Structures
57
4.3.8 9-Segment Meander Resistor
The original/improved fundamental building block models were used to predict
the behavior of a 9-segment meander resistor using the original (Chapter 3) and improved
modeling methodology. An equivalent circuit of the resistor was constructed by
replacing each building block in the structure with its equivalent extracted circuit, as
shown in Fig. 4.3.8-1. Since only first level coupling was taken into account, each
material square in each segment of the resistor was coupled to its nearest neighbor by a
pair of mutual inductances and a coupling capacitance. As inferred from the circuit
description, the resulting circuit using 30µm linewidths and 510µm line lengths per
segment 9-segment resistor was a complex, highly interconnected system, consisting of
approximately 700 nodes. The longest path length of the resistor was approximately
0.175 wavelengths long at 20GHz.
MaterialSquare
R = 0 .0 8 O hmL = 1 E- 1 1 HC = 2.7 e-15 FC C = 1. 2e -15 F
L
C
R L R
C C C C
1 2
Pad
R = 0 .08 O h mL = 1 E- 11 HC = 2.7 e- 1 5 FC C = 1.2e -1 5 F
L
C
R L R
C C C C
1 2
Coupled Pair
R = 0 . 09 O h mL = 1 E -11 HC = 1 .1 e-1 5 FC M = 0.4 e- 1 5 FC C = 1 .4 e -1 5 FL M = 0 .2 0
LC
R L R
C C CC
LC
R L R
C C C C
LM
L MCM
1
2
3
4
.
R= 0 .4 O h mL= 3 .7 E- 1 1 HC = 5.3 e -1 5 FCC = 2 7 e -1 5 F
L
C
R L R
CC CC
1 2
CC CC
2L
C
R L R1
Multi-BlockL
C
R L R
CC CC
1 2
L
C
R L R
CC CC
1 2
L
C
R L R
CC CC
1 2
Pad
Pad
Figure #4.3.8-1: 9-Segment Meander Resistor and Associated Building Blocks
58
4.4 Results
The AC small signal circuit simulation as before was completed in less than 2
minutes. Comparisons in the form of percent differences from the actual measured
device are shown in Fig. 4.4-1. The “improved” modeling methodology clearly
outperforms the original, yielding an improvement of approximately 1.0% (which is
actually a 50% improvement over the original 2% error) at low frequencies and up to
14.0% at frequencies greater than 3GHz. This is due to the enhanced modeling of the
corner primitives, which more accurately follows the complex high frequency electrical
behavior as the current moves through the corners.
Actual Measured 9 Segment Meander Resistor vs. Original and Improved Characterization MethodActual Measured 9 Segment Meander Resistor vs. Original and Improved Characterization Method
9 Segment Modeled ResistorOriginal Characterization Method vs. Improved Method
-2.00E-01
-1.00E-01
0.00E+00
1.00E-01
2.00E-01
3.00E-01
4.00E-01
5.00E-01
6.00E-01
7.00E-01
8.00E-01
1.E+08 1.E+09 1.E+10 1.E+11
Frequency
S11/S21Real
S11 Measured DeviceS11 Original MethodS11 Improved MethodS21 Measured DeviceS21 Original MethodS21 Improved Method
S11R Percent Difference From Measured Device Original Method vs. Improved Method
S11R Percent Difference From Measured Device Original Method vs. Improved Method
9 Segment Modeled ResistorCalculated Measured vs. Modeled Percent Difference
Original Characterization Method vs. Improved Method
-2.00%
0.00%
2.00%
4.00%
6.00%
8.00%
10.00%
12.00%
14.00%
1.E+08 1.E+09 1.E+10 1.E+11
Frequency
S21 RealPercent
Error
Improved MethodOriginal Method
9 Segment Modeled ResistorCalculated Measured vs. Modeled Percent Difference
Original Characterization Method vs. Improved Method
-2.00%
0.00%
2.00%
4.00%
6.00%
8.00%
10.00%
12.00%
14.00%
1.E+08 1.E+09 1.E+10 1.E+11
Frequency
S11 RealPercent
Error
Improved MethodOriginal Method
S21R Percent Difference From Measured Device Original Method vs. Improved Method
S21R Percent Difference From Measured Device Original Method vs. Improved Method
Figure #4.4-1: 9 Segment Meander Resistor Measured vs. Original and Improved
Characterization Methods
59
4.5 Summary
Improvements to the overall existing methodology, detailed in Chapter 3, have
been presented, yielding a substantial increase in measured vs. modeled accuracy and
tool/user interface. The 9-segment meander resistor was again used to showcase the
improvements made to the existing methodology. The percent improvement results
shown in Fig. 4.5-1 clearly demonstrate a 1.0% improvement at frequencies below 3GHz
and a much greater percent improvement for higher frequencies.
Original Method vs. Improved Method~1 Percent Difference Up To 3GHz
Original Method vs. Improved Method~1 Percent Difference Up To 3GHz
9 Segment Modeled ResistorCalculated Percent Improvement
Original Characterization Method vs. Improved Method
-2.00%
0.00%
2.00%
4.00%
6.00%
8.00%
10.00%
12.00%
14.00%
1.E+08 1.E+09 1.E+10 1.E+11
Frequency
S11/S21 RealPercent
Improvement
S11 Percent ImprovementS21 Percent Improvement
Figure #4.5-1: 9 Segment Meander Resistor Original Characterization Method vs.
Improved Characterization Methods
60
CHAPTER 5
DEEMBEDDING BUILDING BLOCK EQUAVILANT CIRCUIT
ELEMENT VALUES USING ONLY ONE TEST STRUCTURE
5.1 Introduction
Another modification to the existing methodology, detailed in Chapter 3, was the
removal of one of the test structures used to deembed two of the four building blocks.
This chapter will detail the deembedding of all 22 element values from four building
blocks using only one test structure, as shown in Fig. 5.1-1. These improvements
resulted in a substantial increase in measured vs. modeled accuracy for the test structures
as well as the modeled device. This chapter will follow the format of chapter 4 in
detailing the departure from a two-test structure deembedding procedure to a one-test
structure deembedding procedure. The 9-segment meander resistor will be modeled
again using this new version of the methodology. A comparison between the original,
improved and the improved using 1 test structure will be displayed via measured vs.
modeled device results.
61
Test Structure 2Test Structure 2Coupled Square PrimitiveCoupled Square Primitive U-Shaped Bend
PrimitivesU-Shaped Bend
Primitives
Pad PrimitivePad Primitive
Material Square PrimitiveMaterial Square Primitive
Figure #5.1-1: Test Structure and Building Blocks
5.2 Design and Modeling Flowchart
The improvements to the novel passive device modeling methodology are
described using the same flow chart with the enhancements localized in the “Model
Canonical Blocks” section, shown in Fig. 5.2-1. The first step, as before, in the modeling
process is to identify what devices to model in a process, and identify building blocks and
design rules. The next step is to design and fabricate one test structure to characterize.
Next, the test structure is characterized for building block circuit extraction; this is the
enhanced area. The measured data is then used to determine initial guesses and to set up
optimizations to extract equivalent circuits of the test structure and building blocks. Once
62
successful optimizations have been achieved, then the building blocks with associated
models and design rules are combined in a library.
DRCDRC
Accurate ResultsAccurate Results
Simulate in SPICESimulate in SPICE--type simulatortype simulator
Design Desired Design Desired Passive DevicePassive Device
Canonical Block/Design Canonical Block/Design Rule LibraryRule Library
Generate Design Rules Generate Design Rules Associated with Each BlockAssociated with Each Block
Model CanonicalModel CanonicalBlocksBlocks
PassPassPass
FailFailFail
Design &Design & FabFabTest StructuresTest Structures
ImprovementsImprovements
Figure #5.2-1: Design and Modeling Flowchart
5.3 Detailed Resistor Modeling Procedure
5.3.1 Identify Building Blocks
As stated previously, the serpentine (meander) geometry of the resistor dictated
four fundamental building blocks that required characterization: 1.) A square building
block with connections on opposite sides; 2.) A U-shaped section connecting two
parallel segments of the resistor; 3.) A coupled block segment to characterize line-to-line
coupling behavior on a per square basis; and 4.) A probe pad to connect to the test
equipment. Nearest neighbor coupling was also included between the material squares.
63
Using the current visualization approach, the pad primitive was taken to be the
large pad square plus one adjacent line square, the material square and coupled material
square were taken as one unit of material square each, and the U-shaped primitive was
represented by 3 individually deembedded squares on each end, as shown in Fig. 5.3.2-1.
5.3.2 Equivalent Circuits
Each building block is modeled as a Spice compatible RLC circuit. For simple
uncoupled building blocks, such as a piece of straight line, simple RLC models based on
the partial element equivalent circuit (PEEC) are used for equivalent circuits [26]. In the
case of blocks where coupling is examined, e.g. coupled material squares, coupled PEEC
models joined by coupling capacitances and mutual inductances are used. The corner
primitive are modeled as detailed in Chapter 4, with three independent PEECs’; 1.) One
for the input/output; 2.) One for the two corners; and 3.) One between the corners. The
test structures, the improved building block equivalent circuits, and the respective
PEECs’ are all shown in Fig. 5.3.2-1.
64
Building Blocks / PEECsBuilding Blocks / PEECs
1 3
2 4
R = 0 .09 Oh mL = 1 E -1 1 HC = 1 .1 e -1 5 FC M = 0 .4 e -1 5 FCC = 1 .4 e-1 5 FL M = 0 .2 0
LC
R L R
C C CC
LC
R L R
C C C C
L M
L MCM
1
2
3
4
1 2
R =0 .08 Oh mL = 1 E- 11 HC = 2.7 e- 1 5 FC C = 1.2e -15 F
L
C
R L R
C C C C
1 2
1 2
R =0 . 08 O h mL = 1 E- 11 HC = 2.7 e- 1 5 FC C = 1.2e - 15 F
L
C
R L R
C C C C
1 2
.
R= 0 .4 O h mL= 3 .7 E- 1 1 HC = 5.3 e -1 5 FCC = 2 7e - 1 5 F
L
C
R L R
CC CC
1 2
CC CC
2L
C
R L R1
1
2
L
C
R L R
CC CC
1 2
L
C
R L R
CC CC
1 2
L
C
R L R
CC CC
1 2
Test Structure 2Test Structure 2Coupled Square PrimitiveCoupled Square Primitive U-Shaped Bend
PrimitivesU-Shaped Bend
Primitives
Pad PrimitivePad Primitive
Material Square PrimitiveMaterial Square Primitive
Figure #5.3.2-1: Uncoupled and Coupled PEEC Circuits with Associated Building
Blocks
5.3.3 Test Structures
The building blocks are defined as in Chapter 4, with characterization and model
development next. The process of characterization and model development is
accomplished through the use of one carefully designed test structures. In this chapter
only one test structure comprised of all the predefined building blocks will be used. The
equivalent circuit of the test structure is assembled of “building block equivalent circuits”
of which it is comprised. The structure is designed, fabricated, and measured to
characterize each of the building blocks.
65
An important issue when designing the test structure is that it must be sensitive to
all the building blocks. Wide band sampling with sufficient frequency resolution is
crucial in order to capture all the major reactive effects over a band of interest [4]. In
most cases, parasitics are small enough that they only manifest themselves at higher
frequencies and in order to deembed them they must be observed. Therefore, the Hspice
optimizations were run using 400 frequency points at a range of 45MHz to 20GHz to
insure that the output parameters were sensitive to all of the important circuit parameters
at some points over the entire frequency range.
As stated previously, by virtue of device measurements, no assumptions are made
regarding material properties, layered dielectrics, or imperfect substrates, because all of
these effects are represented in the measured data. This makes the modeling procedure
entirely process independent.
To model the four stated building blocks, only “test structure two” will be used, as
shown in Fig. 5.3.2-1. Test structure two is a 3-segment meander resistor, and will be
used to characterize the basic uncoupled material parameters (including self-resistance,
inductance, and capacitance) the entire U-shaped corner segments, as well as line-to-line
mutual inductance and coupling capacitance. High frequency measurements of the test
structures were taken and used to characterize each of the four building blocks.
5.3.4 Processing and Measurement
The test structure design was fabricated by depositing gold on a 96% Alumina
substrate. A photomicrograph of the fabricated test structure is shown in Fig. 5.3.4-1.
66
The test structure was measured using network analysis techniques. For the high
frequency measurements, a HP 8510C network analyzer was used with a Cascade
Microtech probe station and ground-signal-ground configuration probes. Calibration was
accomplished using a calibration substrate and utilization of the line-reflect-match (LRM)
calibration method. Data was gathered for test structure two at over 400 frequency points
between 45MHz and 20GHz, and was stored with the aid of computer data acquisition
software and equipment.
9-Segment Meander Resistor9-Segment Meander Resistor
Test Structure 2Test Structure 2
Figure #5.3.4-1: Photograph of Fabricated Structure for 9-Segment Meander
Resistor Modeling
5.3.5 Modeling and Parameter Extraction
Circuit models must be generated for each of the defined building blocks. The
circuit topologies and parameters for the uncoupled and coupled building blocks, pad
building block, and improved multi-block U-shaped bend are shown in Fig. 5.3.5-1.
These circuits represent models for the building blocks only; test structure two and the 9-
67
segment meander resistor circuits are comprised of many building block circuits
connected in accordance with the structure geometry.
Deembed
Test Structure 2
Test Structure 2
MaterialSquare
R = 0 . 0 8 O h mL = 1 E - 1 1 HC = 2 .7 e - 1 5 FC C = 1 . 2 e - 1 5 F
L
C
R L R
C C C C
1 2
Pad
R = 0 .0 8 O h mL = 1 E - 1 1 HC = 2.7 e - 1 5 FC C = 1. 2 e - 15 F
L
C
R L R
C C C C
1 2
Multi-Block
Coupled Pair
R = 0 . 0 9 O h mL = 1 E -1 1 HC = 1 .1 e - 1 5 FC M = 0 .4 e - 1 5 FC C = 1 .4 e -1 5 FL M = 0 .2 0
LC
R L R
C C CC
LC
R L R
C C C C
L M
L MC M
1
2
3
4
.
R = 0 . 4 O h mL = 3 .7 E- 1 1 HC = 5 .3 e -1 5 FC C = 2 7 e - 1 5 F
L
C
R L R
CC CC
1 2
CC CC
2L
C
R L R1
L
C
R L R
CC CC
1 2
L
C
R L R
CC CC
1 2
L
C
R L R
CC CC
1 2
PEECsPEECsBuildingBlocks
BuildingBlocks
4 Building Block Primitives22 Deembedded Element Values
Figure #5.3.5-1: Building blocks, Equivalent Circuits and Parameter Values for
Meander Resistor Modeling
Using only one test structure for the deembedding process will simplify the
fabrication and extraction procedure, and will increase the measured vs. modeled test
structure accuracy. A comparison was made between the methodology used in Chapter 4
(Improved) and the one used in this Chapter (Improved 1TS). Both analyses used the
same building block circuit configurations as shown in Fig. 5.3.5-2, and were analyzed
using the Hspice Optimization Tool.
68
Test Structure 1 & Test Structure 2Used (Improved)
Test Structure 1 & Test Structure 2Used (Improved)
No Substrate ResistanceDeembedded
Building Blocks Circuit Model Topology
Deembed MaterialSquare
PadMulti Block
Coupled Pair
No Substrate ResistanceDeembedded
Deembed MaterialSquare
PadMulti Block
Coupled Pair
Only Test Structure 2Used (Improved 1TS)Only Test Structure 2Used (Improved 1TS)
Figure #5.3.5-2: Analysis Comparison Using Two Test Structures and One Test
Structure
To simplify accuracy evaluation, percent error between measured vs. modeled
was calculated at each frequency point and is shown in the plots below. Iterative Hspice
optimization runs increased measured vs. modeled accuracy. Unix scripts automatically
ran 18 iterative optimizations per measured device, while at the same time increasing the
accuracy settings and decreasing the range of possible element values within the
optimization circuit file. The building block characterization process executed 576
individual optimizations to arrive at 32 unique equivalent circuit models of test structure
two. The entire Optimization/MonteCarlo analysis process together with data
calculations, conversions, storage, and circuit file creation was automated using Unix
shell scripts. The detailed automated process will be showcased in Chapter 9.
69
5.3.6 Test Structure 2 Results
An optimization analysis was performed on test structure two; 1.) Without
deembedding the substrate resistance; and 2.) Deembedding a multi-block corner, for all
32 fabricated devices.
The percent error between the actual measured device and the modeled device
was calculated for test structure two using the methodology detailed in this chapter, and
compared with the results from Chapter 4 as shown in Fig. 5.3.6-1. The reflection
coefficient (S11), using only one test structure to deembed, shows a 0.5% improvement
over the two-test structure method. The transmission coefficient (S21) also shows an
overall percent error improvement of 0.5%.
70
Two Test Structures Used~3 percent Error
Between S11 Real Measured vs. Modeled
Two Test Structures Used~3 percent Error
Between S11 Real Measured vs. Modeled
One Test Structure Used~2.5 percent Error
Between S11 Real Measured vs. Modeled
One Test Structure Used~2.5 percent Error
Between S11 Real Measured vs. Modeled
Test Structure 2Two Test Structures Used to Deembed Element Values
S11 Real Measured vs. Modeled
-1
1
3
5
7
9
1.E+08 1.E+09 1.E+10 1.E+11
Frequency
S11 Real% Error
Test Structure 2One Test Structures Used to Deembed Element Values
S11 Real Measured vs. Modeled
-1
1
3
5
7
9
1.E+08 1.E+09 1.E+10 1.E+11
Frequency
S11 Real% Error
Test Structure 2One Test Structures Used to Deembed Element Values
S21 Real Measured vs. Modeled
-1
0
1
2
3
4
5
6
7
8
1.E+08 1.E+09 1.E+10 1.E+11
Frequency
S21 Real% Error
Test Structure 2Two Test Structures Used to Deembed Element Values
S21 Real Measured vs. Modeled
-1
0
1
2
3
4
5
6
7
8
1.E+08 1.E+09 1.E+10 1.E+11
Frequency
S21 Real% Error
Two Test Structures Used~1 percent Error
Between S21 Real Measured vs. Modeled
Two Test Structures Used~1 percent Error
Between S21 Real Measured vs. Modeled
One Test Structure Used~0.5 percent Error
Between S21 Real Measured vs. Modeled
One Test Structure Used~0.5 percent Error
Between S21 Real Measured vs. Modeled
Figure #5.3.6-1: Measured vs. Modeled Percent Difference for 32 Fabricated
Structures and 32 Modeled Structures
5.3.7 9-Segment Meander Resistor
The fundamental building block models were assembled and used to predict the
behavior of a 9-segment meander resistor using only one test structure to deembed all the
element values of the building block equivalent circuits. An equivalent circuit of the
resistor was constructed by replacing each building block in the structure with its
equivalent extracted circuit, as shown in Fig. 5.3.7-1. Since only first level coupling was
taken into account, each material square in each segment of the resistor was coupled to its
nearest neighbor by a pair of mutual inductances and a coupling capacitance. As inferred
71
from the circuit description, the resulting circuit using 30µm linewidths and 300µm line
lengths per segment 9-segment resistor was a complex, highly interconnected system,
consisting of approximately 700 nodes. The longest path length of the resistor was
approximately 0.175 wavelengths long at 20GHz.
MaterialSquare
R =0 .0 8 O hmL = 1 E- 11 HC = 2.7 e- 15 FCC = 1. 2 e -15 F
L
C
R L R
C C C C
1 2
Pad
R =0 .0 8 O h mL = 1 E- 11 HC = 2.7 e-1 5 FC C = 1. 2e - 15 F
L
C
R L R
C C C C
1 2
Coupled Pair
R=0 . 09 O h mL = 1 E -11 HC = 1 .1 e-1 5 FCM =0 .4 e -1 5 FCC = 1 .4 e -1 5 FL M = 0 .2 0
LC
R L R
C C CC
LC
R L R
C C C C
L M
L MCM
1
2
3
4
.
R= 0 .4 O h mL= 3 .7 E- 1 1 HC = 5.3 e -1 5 FCC=2 7e - 1 5 F
L
C
R L R
CC CC
1 2
CC CC
2L
C
R L R1
Multi-BlockL
C
R L R
CC CC
1 2
L
C
R L R
CC CC
1 2
L
C
R L R
CC CC
1 2
Pad
Pad
Figure #5.3.7-1: 9-Segment Meander Resistor and Associated Building Blocks
5.4 Results
The AC small signal circuit simulation, as before, was completed in less than 2
minutes. Comparisons in the form of percent differences from the actual measured
device are shown in Fig. 5.4-1.
The improved 1TS modeling methodology clearly outperformed the original and
improved methods, yielding an enhancement of approximately 1.0% at frequencies up to
72
9GHz for S11R. However for S21R, the improved modeling methodology outperformed
the original and improved 1TS at frequencies below 1GHz and equaled the improved 1TS
method at frequencies above 1GHz.
This is mainly due to enhanced matching; i.e. using a dependent “test structure 1
& 2” (contained within test structure 2) in contrast to an independent “test structure 1”
and an independent “test structure 2” located at different sites on the substrate. The
number of frequency points (400) in the measured input file also enabled a successful
extraction of all building block equivalent circuit values.
Measured 9 Segment Meander Resistor vs. Original, Improved, Improved 1 TS, Characterization MethodMeasured 9 Segment Meander Resistor vs. Original, Improved, Improved 1 TS, Characterization Method
S11R % Difference From Measured Device Original vs. Improved vs. Improved 1 TS
S11R % Difference From Measured Device Original vs. Improved vs. Improved 1 TS
S21R % Difference From Measured Device Original vs. Improved vs. Improved 1 TS
S21R % Difference From Measured Device Original vs. Improved vs. Improved 1 TS
9 Segment Modeled ResistorCalculated Measured vs. Modeled Percent Difference
Original Method vs. Improved Method vs. Improved 1 TS
-2.00%
0.00%
2.00%
4.00%
6.00%
8.00%
10.00%
12.00%
14.00%
1.E+08 1.E+09 1.E+10 1.E+11
Frequency
S21 RealPercent
Error
Improved MethodOriginal Method
Improved Method 1 TS
9 Se gm e nt M ode le d Re sistorO riginal M e thod v s. Im prov e d M e thod v s. Im prov e d 1 TS
-2.00E-01
-1.00E-01
0.00E+00
1.00E-01
2.00E-01
3.00E-01
4.00E-01
5.00E-01
6.00E-01
7.00E-01
8.00E-01
1.E+08 1.E+09 1.E+10 1.E+11
F re que ncy
S 11/S 21Re a l
S11 Meas ured Dev ic e
S11 Original Method
S11 Improv ed Method
S11 Improv ed Method 1 TSS21 Meas ured Dev ic e
S21 Original Method
S21 Improv ed Method
S21 Improv ed Method 1 TS
9 Segment Modeled ResistorCalculated Measured vs. Modeled Percent Difference
Original Method vs. Improved Method vs. Improved 1 TS
-2.00%
0.00%
2.00%
4.00%
6.00%
8.00%
10.00%
12.00%
14.00%
1.E+08 1.E+09 1.E+10 1.E+11
Frequency
S11 RealPercent
Error
Improved MethodOriginal Method
Improved Method 1 TS
Figure #5.4-1: 9 Segment Meander Resistor
Measured vs. Original, Improved, and Improved 1TS Characterization Methods
73
5.5 Summary
The successful deembedding of 22 element values from 4 building block
equivalent circuits using only one test structure was presented in this chapter. The 9-
segment meander resistor was again used to showcase the improvements made to the
existing methodologies. The percent improvement comparison plots in Fig. 5.5-1 show
that this methodology out performed the original method for S11R and S21R, and
outperformed the improved method for S11R, into the GHz. However, the improved
method for S21R at frequencies less than 3.5GHz out performed the improved 1TS
method.
Percent Comparisons Between The Original, Improved andImproved 1TS Methodologies
Percent Comparisons Between The Original, Improved andImproved 1TS Methodologies
9 Segment Modeled Resistor Calculated Percent Improvement Original Characterization Method vs. Improved Method
-20.00%
-15.00%
-10.00%
-5.00%
0.00%
5.00%
10.00%
15.00%
20.00%
25.00%
30.00%
1.E+08 1.E+09 1.E+10 1.E+11
Frequency
S11/S21 RealPercent
Improvement
S11 Original MethodS21 Original MethodS11 Improved MethodS21 Improved Method
Figure #5.5-1: 9 Segment Meander Resistor Percent Comparisons
74
CHAPTER 6
PREDICTIVE STATICTICAL ANALYSIS VIA MEASUREMENT OF
CONONICAL BUILDING BLOCKS
6.1 Introduction
As stated previously, a further developed passive device modeling methodology
that accurately predicts the high frequency electrical behavior, and the statistical variation
of embedded passive devices on a generic substrate, will be presented. This research will
prove that the variation in the complete equivalent circuit models, based only on the
circuit building blocks, can be used to predict such variations in actual fabricated devices.
Process parameters are the optimum basis for statistical circuit modeling of
passive devices [29]. To statistically model a device material properties such as oxide
and metal thickness variations, and physical properties such as dielectric constant and
resistivity variations are required. Process variations equate to element value variations
when using the aforementioned modeling methodology. Therefore to statistically model
a device, measurements of multiple test structures need to be taken and the equivalent
circuit values deembedded. This deembedding process automatically accounts for the
effects of processing fluctuations and nonideal material properties.
75
Mean and relative variations in the deembedded circuit values from all the devices
are then calculated and used in the Hspice MonteCarlo simulator as limits to generate
random circuit values for a defined number of simulations. A predicted range of
operation for the actual modeled device is then generated from the MonteCarlo
simulations. Ideally, the designer would like to demonstrate that the variation in the
equivalent circuit models, based only on the data used to extract them, can be used to
predict variations in the actual fabricated device. Therefore, accurate measured vs.
modeled test structure results, that generate accurate device models, and consequently
realistic range of operation predictions, will be the basis of this process.
This chapter will explain in detail the procedure of predicting passive device (i.e.
meander resistor) performance variations and yield, when provided with a statistical
distribution of the building block models and a circuit schematic.
6.2 Modeling Procedure
The procedure begins by characterizing and modeling the passive device (i.e.
meander resistor) as stated in Chapter 3. A brief description of the various steps is
provided below and shown in Fig. 6.2-1.
1. The first step is to determine what structure geometries to consider, in compiling a set
of building blocks and test structures that will be measured and characterized.
76
2. The second step is to create building block geometries. Current visualization
software is useful in determining areas of constant current flow, which provide the
boundaries for the building blocks.
3. The third step is to design test structures comprised of the building blocks, which will
enable their accurate modeling.
4. At this point, test structures are fabricated, tested, and high frequency network
analysis and DC resistance measurements taken.
5. The measured test structure data and a circuit file description are then used as input to
the optimization tool. PEEC element value initial guesses are based on the measured
results for each structure. Once optimization for one structure is complete, the results
are used, as initial guess, for the remaining optimizations.
6. The deembedded circuit values of the building block equivalent circuits are obtained
through non-linear optimization using the Hspice Optimization Tool. The designer
now has an accurate spice model of each building block, which can then be used to
design any value/configuration meander resistor.
77
Accurate ResultsAccurate Results
Simulate in SPICESimulate in SPICE--type type simulatorsimulator
DRCDRC
Design Desired Passive Design Desired Passive DeviceDevice
Canonical Block/Design Canonical Block/Design Rule LibraryRule Library
Generate Design Rules Generate Design Rules Associated with Each Associated with Each
BlockBlock
Model CanonicalModel CanonicalBlocksBlocks
PassPassPass
FailFailFail
Design &Design & FabFabTest StructuresTest Structures
Figure #6.2-1: Design and Modeling Flowchart
6.3 Statistical Variations
The ability to predict the performance of an embedded passive device is
fundamental to those attempting to design integrated circuits, modify fabrication
processes, plan production schedules, or specify product-operating characteristics. In
general, a three-step MonteCarlo technique is used to accomplish this prediction: 1.) A
statistical model is formed to characterize the fabrication process used; 2.) A circuit
design is created using a Spice type simulator and applying nominal device values for the
target process; and 3.) Randomly generated element values of the process model are
simulated in a MonteCarlo tool to produce a representative set of output performance
characteristics. The impact of random process variations can be inferred from these
78
simulations, and electrical yield can then be estimated using the percentage of that sample
which meets the performance requirements [29]. Electrical yield refers to the number of
parts that meet the electrical test specifications; whereas, maximizing yield is important
for the overall process efficiency.
6.3.1 Statistical Model
The statistical model used in this research is formed via the passive device
modeling methodology described in Chapters 3, 4, and 5. Test structures are designed
and fabricated in a process of interest and contain the fundamental building blocks of the
larger modeled device. The building blocks, modeled as Spice compatible RLC circuits,
are characterized using actual measured data and a non-linear optimization method.
Circuit models complete with deembedded element values are defined for each building
block, and accurately reproduce the test structure behavior.
Once a defined number of these test structures have been characterized, mean and
relative variations in the deembedded circuit values of the building blocks are then
calculated and used in the Hspice MonteCarlo simulator as limits to generate random
circuit values for a defined number of simulations. Hspice analyzes the yield by using
statistical techniques and observing the effects of element and model parameter variation
[31]. Fig. 6.3.1-1 shows some of the deembedded element values and the syntax used in
the MonteCarlo simulation tool.
79
*************************************** 30 u Line Specs 1**************************************+ c_sq_1 = UNIF(1.262198980e-13, 9.72E-01) + rsq_1 = UNIF(2.510000000e-06, 0.00E+00) + lsq_1 = UNIF(1.816762077e-11, 4.91E-02)+ csq_1 = UNIF(6.520000000e-26, 9.69E-01)+ r_sub_sq = 10mega*************************************** Corner and Coupled Lines Specs 1**************************************.param+ cou_l_1 = UNIF(2.205573309e-01, 2.97E-01) + c_cou_1 = UNIF(6.714851620e-24, 8.15E-01) + r2_1 = UNIF(6.652141530e-01, 4.43E-01)+ c2_1 = UNIF(1.357575649e-15, 2.34E-01)+ l2_1 = UNIF(4.295089543e-12, 3.34E-01)+ c_co2cr_1 = UNIF(8.071681342e-12, 3.66E-01) + r_sub_corrn = 10mega
c_sq csq lsq rsq c_pad rpadlow 3.56172E-15 6.520000E-26 1.7331E-11 0.00000251 1.27E-12 3.754664high 2.49093E-13 4.098440E-24 1.91221E-11 0.00000251 1.88E-12 5.626608median 1.26327E-13 2.081820E-24 1.82265E-11 0.00000251 1.57E-12 4.690636ab var 1.22765E-13 2.016620E-24 8.95575E-13 0 3.06E-13 0.935972rel var 0.971805619 9.686813E-01 0.049135815 0 0.194744 0.19954
cou_1 c_cou l2 r2 c2 c_co2crlow 0.159291519 8.919877E-25 2.76911E-12 0.398398019 1.08E-15 5.56E-12high 0.293682413 8.726346E-24 5.54968E-12 1.032752749 1.75E-15 1.2E-11median 0.226486966 4.809167E-24 4.1594E-12 0.715575384 1.41E-15 8.77E-12ab var 0.067195447 3.917179E-24 1.39028E-12 0.317177365 3.32E-16 3.21E-12rel var 0.296685713 8.145234E-01 0.334250965 0.443248011 0.234432 0.366037
Statistical Variations for Each Element ValueCalculate and Inserted into Hspice MonteCarlo Simulator
Statistical Variations for Each Element ValueCalculate and Inserted into Hspice MonteCarlo Simulator
Figure #6.3.1-1: Mean and Absolute Deviations for Deembedded Element Values
6.3.2 Circuit Design
An equivalent circuit of the modeled device (device under inspection) is
constructed by replacing each building block in the structure with its equivalent extracted
circuit values. To accurately represent the variations in the complete circuit model, the
correlation between the building block element values needs to be investigated.
The correlation between two or more random variables is a question defined as;
does one or more random variable depend in any way on any of the other random
variables [30]. This question was addressed by examining the “Sputtering” technique
that was used to deposit the gold onto the 96% Alumina substrate. Sputtering deposits a
random nonuniform layer of metal onto a substrate surface. This basically means that
one portion of the deposited metal is in no way a function of any other deposited portion;
therefore, the correlation between segments does not exist.
80
Due to the lack of correlation between the building blocks, each element is
basically independent of any other element in the circuit. Therefore, to create an accurate
circuit model file all the elements in the design need to generate a random value for the
defined number of simulations. This circuit file can become quite large, for example a 9-
segment meander resistor needs to generate 1336 random circuit values for each
simulation. Fig. 6.3.2-1 is a block diagram of a 9-segment meander resistor and shows
the amount of random number generations per simulation, for the PEEC element values.
As a comparison, the statistical analysis procedure was also applied to a circuit model file
using the premise that all the building block equivalent circuits were correlated.
MaterialSquare
R=0.08 OhmL= 1E-11 HC = 2.7e-15 FCC = 1.2e-15 F
LC
R L R
CC CC
1 2
Pad
R=0.08 OhmL=1E-11 HC = 2.7e-15 FCC = 1.2e-15 F
LC
R L R
CC CC
1 2
Coupled Pair
R=0.09 OhmL= 1E-11 HC = 1.1e-15 FCM=0.4e-15 FCC = 1.4e-15 FLM = 0.20
LC
R L R
CC CC
LC
R L R
CC CC
L MLM
CM
1
2
3
4
CC
LC
R1
Multi-Block
LC
R L R
CC CC1 2
LC
R L R
CC CC1 2
.
R=0.4 OhmL= 3.7E-11 HC = 5.3e-15 FCC=2 7e-15 F
CC2
L R
LC
R L R
CC CC1 2
LC
R L R
CC CC1 2
Pad
Pad
TOTAL = 1336 Random Generated Element Values
Pads = (2 BBlocks * 8 Elements) = 16 ElementsMaterial Squares = (8 BBlocks * 8 Elements) = 64 ElementsCoupled Squares = [((4 BBlocks * 9 Coupled) * 3) * 8 Elements] + (9 Coupled * 8 Elements) = 936 Elements
Corners = (5 BBlocks * 8 Corners * 8 Elements) = 320 Elements
Figure #6.3.2-1: Number of Random Number Generation for MonteCarlo
Simulation
81
6.3.3 MonteCarlo Analysis
The MonteCarlo analysis uses a random number generator to create a uniform
parameter distribution by means of the relative variation; whereas, the nominal value is
varied by +/- (nominal_value * relative_variation) [31]. A new random variable is
calculated each time a parameter is used. Fig. 6.3.3-1 shows how the relative variation is
calculated.
Uniform DistributionUniform Distribution
Rel_variation = Abs_variation / Nom_valueRel_variation = Abs_variation / Nom_value
Abs_variationAbs_variation
Nom_valueNom_value
PopulationPopulation
Figure #6.3.3-1: Uniform Distribution
Based on the statistics calculated for each element value in the equivalent circuit
model, a MonteCarlo analysis is performed using Hspice. A MonteCarlo simulation with
a uniform parameter distribution is generated for a defined number of sets of circuit
model parameters. Following the simulations, the output response for the modeled
device is compared with actual measured data to determine if the statistical variations of
the model parameter values reside in the range of the MonteCarlo predicted output. To
accurately compare the measured vs. MonteCarlo results, several statistical methods were
used to find a number that represents both sets of data.
82
6.3.4 Statistical Analysis Methods
The field of statistics is concerned with methods of organizing, summarizing, and
interpreting, data [32]. “Data” is defined as information; whereas, as any collection of
information is a collection of data. However for statisticians, “data” means numerical
information. Basically, statistics amounts to a collection of techniques for dealing with
sets of numbers, organizing them, summarizing them, and discovering their meaning. An
excel workbook with a set of worksheets is used to discern all collected data. Mean,
variance, and standard deviation between MonteCarlo generated and actual measured
data are the basis for the techniques described below. Fig. 6.3.4-1 is a portion of the
excel workbook that filters the MonteCarlo and measured data then applies several
statistical techniques to the results.
Excel Spread Sheet Used To Calculate Statistical InformationExcel Spread Sheet Used To Calculate Statistical InformationFrequency MC Meas Filter MC CFD Meas CFD MC CFDN Meas CFDN MC CFDNT Meas CFDNT T-Test
45000000 0.4181 0.399 0.395158 1 0 0.002000 0.000000 0.002000 0.000000 Actual CN T-Score0.4169 0.404 0.398099 2 0 0.004000 0.000000 0.004000 0.000000 -3.013391 0.545626 0.0174250.4114 0.405 0.401041 10 1 0.020000 0.031250 0.020000 0.0000000.4231 0.407 0.403982 24 1 0.048000 0.031250 0.048000 0.031250 Corralation Cofficient0.4253 0.408 0.406924 57 3 0.114000 0.093750 0.114000 0.125000 CN T-Score0.4149 0.408 0.409865 108 7 0.216000 0.218750 0.216000 0.281250 0.974649 0.9972350.4219 0.409 0.412806 179 9 0.358000 0.281250 0.358000 0.3437500.4186 0.41 0.415748 277 11 0.554000 0.343750 0.554000 0.531250 Mean Error0.4089 0.411 0.418689 370 16 0.740000 0.500000 0.740000 0.656250 Actual CN T-Score0.4261 0.414 0.421630 419 20 0.838000 0.625000 0.838000 0.875000 -0.87% 14.36% 0.48%0.4138 0.415 0.424572 464 22 0.928000 0.687500 0.928000 0.9062500.4076 0.417 0.427513 489 27 0.978000 0.843750 0.978000 1.000000 Standard Deviation Error0.4135 0.417 0.430455 497 28 0.994000 0.875000 0.994000 1.000000 Actual CN T-Score0.4179 0.417 0.433396 500 32 1.000000 1.000000 0.998000 1.000000 -43.72% 8.00% -0.20%0.4164 0.418 0.436337 500 32 1.000000 1.000000 1.000000 1.0000000.4216 0.418 0.439279 1.000000 1.000000 1.000000 1.000000 ANOVA0.4025 0.419 Actual CN0.4121 0.42 Filter Adjustments 9.063427 0.2977080.4152 0.421 0.002941 500 320.4162 0.421 4.584301 500 320.4233 0.423 Divider 130.427 0.424
0.4105 0.425 MC = MonteCarlo0.4228 0.426 Meas = Measured0.4195 0.426 CFD = Cumulative Frequency Distribution0.4169 0.426 CFDN = Cumulative Frequency Distribution Normalized0.4124 0.427 CFDNT = Cumulative Frequency Distribution Normalized T-Scores0.4178 0.428
Figure #6.3.4-1: Excel Spread Sheet
83
6.3.5 Mean and Standard Deviation
The mean is a measure of central tendency, “middleness”, or the arithmetic
average. Adding up all the scores and dividing by the number of scores is the method of
calculating the mean. The formula for the mean value of X is:
MxΣXN
Where Mx is the mean of X and N is the number of scores (or numbers in a population).
Therefore, the mean of variable X (symbolized as Mx ) equals the sum of the X scores
(ΣX) divided by the number of scores (N).
The variance is the most frequently used measure of variability [32]. The formula
for the variance is:
s2 N X2. X( )2
N N 1( )
The standard deviation is derived from the variance and also indicates the relative
variability within a group. The formula for the standard deviation is:
s N X2. X( )2
N N 1( )
To aid in the evaluation of measured vs. modeled data, percent differences in
mean and standard deviation will be used. Fig. 6.3.5-1 displays the Mean Value Percent
Difference results using two different sets of data. The first set compares the raw
MonteCarlo generated (population 500) data vs. the raw Measured (population 32) data.
84
The second set compares the normalized cumulative frequency distribution data
(population 16). Both sets are displaying the magnitude of S11 and S21 using data
generated from a circuit model file constructed using the correlated premise. The results
begin to diverge at 1.84GHz for S11 magnitude and 4.5GHz for S21 magnitude.
Mean Percent Error Comparing Cumulative Frequency Distribution Data For
MonteCarlo Prediction (16 Data Points) vs. Actual Measured Device (16 Data Points)
-5.00%
0.00%
5.00%
10.00%
15.00%
20.00%
25.00%
30.00%
35.00%
40.00%
1.E+08 1.E+09 1.E+10
Frequency
S11 Mag% Error Correlated Blocks
Mean Percent ErrorComparing
MonteCarlo Prediction (500 Data Points) vs. Actual Measured Device (32 Data Points)
-0.10%
0.40%
0.90%
1.40%
1.90%
1.E+08 1.E+09 1.E+10
Frequency
S11 Mag% Error Correlated Blocks
Mean Percent ErrorComparing
MonteCarlo Prediction (500 Data Points) vs. Actual Measured Device (32 Data Points)
-1.00%
-0.50%
0.00%
0.50%
1.00%
1.50%
2.00%
1.E+08 1.E+09 1.E+10
Frequency
S21 Mag% Error Correlated Blocks
Mean Percent Error Comparing Cumulative Frequency Distribution Data For
MonteCarlo Prediction (16 Data Points) vs. Actual Measured Device (16 Data Points)
-1.00%
9.00%
19.00%
29.00%
39.00%
49.00%
1.E+08 1.E+09 1.E+10
Frequency
S21 Mag% Error Correlated Blocks
Mean Percent Difference For S11/S21 MagnitudeUsing Data Generated Via Correlated Circuit File Configuration
Mean Percent Difference For S11/S21 MagnitudeUsing Data Generated Via Correlated Circuit File Configuration
1.84GHz1.84GHz 1.84GHz1.84GHz
4.5GHz4.5GHz 4.5GHz4.5GHz
Figure #6.3.5-1: Mean Value Percent Difference Between MC and Measured Data
For S11/S21 Magnitude
Fig. 6.3.5-2 displays the Standard Deviation Percent Difference results using two
different sets of data. The first set compares the raw MonteCarlo generated (population
500) data vs. the raw Measured (population 32) data. The second set compares the
normalized cumulative frequency distribution data (population 16). Both sets are
displaying the magnitude of S11 and S21 using data generated from a circuit model file
constructed using the correlated premise. The results begin to diverge at 2.5GHz for S11
85
magnitude using the first set of data, and at 1.74GHz for S11 magnitude using the second
set of data. The results begin to diverge at 4.2GHz for S21 magnitude using the first set
of data, and at 7.0GHz for S21 magnitude using the second set of data.
Standard Deviation Percent Error Comparing Cumulative Frequency Distribution Data For
MonteCarlo Prediction (16 Data Points) vs. Actual Measured Device (16 Data Points)
-1.00%
4.00%
9.00%
14.00%
19.00%
24.00%
29.00%
1.E+08 1.E+09 1.E+10
Frequency
S11 Mag% Error Correlated Blocks
Standard Deviation Percent Error Comparing Cumulative Frequency Distribution Data For
MonteCarlo Prediction (16 Data Points) vs. Actual Measured Device (16 Data Points)
-1.00%
4.00%
9.00%
14.00%
19.00%
24.00%
29.00%
1.E+08 1.E+09 1.E+10
Frequency
S21 Mag% Error Correlated Blocks
Standard Deviation Pecent ErrorComparing
MonteCarlo Prediction (500 Data Points) vs. Actual Measured Device (32 Data Points)
0.00%
20.00%
40.00%
60.00%
80.00%
100.00%
120.00%
1.E+08 1.E+09 1.E+10
Frequency
S21 Mag% Error Correlated Blocks
Standard Deviation Pecent ErrorComparing
MonteCarlo Prediction (500 Data Points) vs. Actual Measured Device (32 Data Points)
0.00%
20.00%
40.00%
60.00%
80.00%
100.00%
120.00%
140.00%
160.00%
180.00%
1.E+08 1.E+09 1.E+10
Frequency
S11 Mag% Error Correlated Blocks
Standard Deviation Percent Difference For S11/S21 MagnitudeUsing Data Generated Via Correlated Circuit File ConfigurationStandard Deviation Percent Difference For S11/S21 MagnitudeUsing Data Generated Via Correlated Circuit File Configuration
2.5GHz2.5GHz 1.74GHz1.74GHz
4.2GHz4.2GHz 7.0GHz7.0GHz
Figure #6.3.5-2: Standard Deviation Value Between MC and Measured Data For
S11/S21 Magnitude
6.3.6 Frequency Distribution
The s-parameters generated from the MonteCarlo results and the actual measured
devices are transposed from rows representing a particular frequency to columns
representing MonteCarlo simulations and measured device data at a particular frequency
point. The range of values from high to low is split up into sixteen linearly place filter
bins. How often a value occurs within a range of values, defined by the filter bins, is
86
calculated and normalized. These normalized cumulative frequency distribution plots are
generated at every frequency point for S11/S21 magnitude up to a certain frequency.
These plots illustrate that the measured data from each fabricated device at a particular
frequency fit within the MonteCarlo predicted range. They also show how the standard
deviations between predicted and measured compare. Trend lines were added to each
plot to simplify the evaluation; blue is the MonteCarlo results and red is the Measured
data. Fig. 6.3.6-1 displays a normalized cumulative frequency distribution plot at several
frequency points for S11. Fig. 6.3.6-2 display a normalized cumulative frequency
distribution plot at several frequency points for S21.
87
Correlated Premise @ 600MHzCorrelated Premise @ 600MHz
Normalized Cumulative Frequency Distribution For S11 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.397
8
0.400
7
0.403
6
0.406
5
0.409
4
0.412
2
0.415
1
0.418
0
0.420
9
0.423
7
0.426
6
0.429
5
0.432
4
0.435
2
0.438
1
0.441
0
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
Correlated Premise @800MHzCorrelated Premise @800MHz
Normalized Cumulative Frequency Distribution For S11 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.399
9
0.402
8
0.405
6
0.408
4
0.411
2
0.414
1
0.416
9
0.419
7
0.422
5
0.425
4
0.428
2
0.431
0
0.433
9
0.436
7
0.439
5
0.442
3
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
Correlated Premise @ 950MHzCorrelated Premise @ 950MHz
Normalized Cumulative Frequency Distribution For S11 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.401
9
0.404
7
0.407
4
0.410
2
0.413
0
0.415
8
0.418
6
0.421
3
0.424
1
0.426
9
0.429
7
0.432
5
0.435
2
0.438
0
0.440
8
0.443
6
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
Correlated Premise @ 1.15GHzCorrelated Premise @ 1.15GHz
Normalized Cumulative Frequency Distribution For S11 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.405
0
0.407
7
0.410
4
0.413
1
0.415
8
0.418
5
0.421
2
0.423
9
0.426
6
0.429
3
0.432
0
0.434
7
0.437
4
0.440
1
0.442
8
0.445
6
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
Correlated Premise @ 1.5GHzCorrelated Premise @ 1.5GHz
Normalized Cumulative Frequency Distribution For S11 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.412
7
0.415
2
0.417
7
0.420
3
0.422
8
0.425
3
0.427
8
0.430
4
0.432
9
0.435
4
0.437
9
0.440
5
0.443
0
0.445
5
0.448
0
0.450
6
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
Correlated Premise @ 2.0GHzCorrelated Premise @ 2.0GHz
Normalized Cumulative Frequency Distribution For S11 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.424
9
0.427
2
0.429
4
0.431
7
0.433
9
0.436
2
0.438
4
0.440
7
0.442
9
0.445
2
0.447
4
0.449
7
0.451
9
0.454
2
0.456
4
0.458
7
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
Figure #6.3.6-1: Normalized Cumulative Frequency Distribution Plot
88
Correlated Premise @ 800MHzCorrelated Premise @ 800MHz
Normalized Cumulative Frequency Distribution For S21 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.565
0
0.567
9
0.570
8
0.573
8
0.576
7
0.579
6
0.582
5
0.585
4
0.588
3
0.591
2
0.594
1
0.597
1
0.600
0
0.602
9
0.605
8
0.608
7
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
Correlated Premise @ 1GHzCorrelated Premise @ 1GHz
Normalized Cumulative Frequency Distribution For S21 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.564
5
0.567
4
0.570
3
0.573
2
0.576
1
0.579
0
0.581
9
0.584
8
0.587
7
0.590
6
0.593
4
0.596
3
0.599
2
0.602
1
0.605
0
0.607
9
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
Correlated Premise @ 2.0GHzCorrelated Premise @ 2.0GHz
Normalized Cumulative Frequency Distribution For S21 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.559
9
0.562
6
0.565
4
0.568
1
0.570
9
0.573
6
0.576
4
0.579
2
0.581
9
0.584
7
0.587
4
0.590
2
0.592
9
0.595
7
0.598
4
0.601
2
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
Correlated Premise @ 3.0GHzCorrelated Premise @ 3.0GHz
Normalized Cumulative Frequency Distribution For S21 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.552
9
0.555
4
0.558
0
0.560
5
0.563
1
0.565
6
0.568
1
0.570
7
0.573
2
0.575
8
0.578
3
0.580
9
0.583
4
0.586
0
0.588
5
0.591
1
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
Correlated Premise @ 4.0GHzCorrelated Premise @ 4.0GHz
Normalized Cumulative Frequency Distribution For S21 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.543
8
0.546
0
0.548
3
0.550
6
0.552
9
0.555
2
0.557
5
0.559
8
0.562
1
0.564
3
0.566
6
0.568
9
0.571
2
0.573
5
0.575
8
0.578
1
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
Correlated Premise @ 5.0GHzCorrelated Premise @ 5.0GHz
Normalized Cumulative Frequency Distribution For S21 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.533
1
0.535
7
0.538
3
0.540
9
0.543
5
0.546
1
0.548
7
0.551
3
0.553
9
0.556
5
0.559
1
0.561
7
0.564
3
0.566
9
0.569
5
0.572
1
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
Figure #6.3.6-2: Normalized Cumulative Frequency Distribution Plot
6.3.7 Pearson Product-Moment Correlation Coefficient
Using the normalized distribution information, a measure of linear relationship
between the two sets of data can be obtained by means of the Pearson Product-Moment
Correlation Coefficient (r). A perfect correlation between two variables results in r = 1.0,
89
no correlation at all results in r = 0.0, and any number in-between means that there is
some correlation between the variables however the relationship is not perfect. The
degree of correlation is a function of statistical significance, which is determined using
numerical charts to ascertain the outcome. The formula for the Correlation Coefficient is:
γ N X Y.( ). X Y.
N X2. X( )2 N Y2. Y( )2.
Where N is the number in the population; and X Y, are the elements of the population.
Fig. 6.3.7-1 displays a Correlation Coefficient plot for S11 and S21 magnitude via the
data generated from a circuit model file constructed using the correlated premise. The
correlation between the two sets of data begins to diverge at 2.0GHz for S11 magnitude
and 4.5GHz for S21 magnitude.
Corralation CofficientComparing Cumulative Frequency Distribution Data For
MonteCarlo Prediction (16 Data Points) vs. Actual Measured Device (16 Data Points)
0.94
0.95
0.96
0.97
0.98
0.99
1.00
1.E+08 1.E+09 1.E+10
Frequency
S21 MagCorrelation
Coff.Correlated Blocks
Correlation CofficientComparing Cumulative Frequency Distribution Data For
MonteCarlo Prediction (16 Data Points) vs. Actual Measured Device (16 Data Points)
0.91
0.92
0.93
0.94
0.95
0.96
0.97
0.98
0.99
1.00
1.E+08 1.E+09 1.E+10
Frequency
S11 MagCorrelation
Coff.Correlated Blocks
Correlation Coefficient For S11/S21 MagnitudeUsing Data Generated Via Correlated Circuit File Configuration
Correlation Coefficient For S11/S21 MagnitudeUsing Data Generated Via Correlated Circuit File Configuration
2.0GHz2.0GHz 4.5GHz4.5GHz
Figure #6.3.7-1: Correlation Coefficient Between MC and Measured Data For
S11/S21 Magnitude
90
6.3.8 T-Test
The “t” test is one of the most commonly used statistical tests [32]. Its primary
purpose is to determine whether the means of two groups of numbers differ to a
statistically significant degree. The larger the absolute value of t, the more likely it is to
reflect a significant difference between the two groups under comparison. Charts that
evaluate statistical significance for the t test have also been developed and are included in
the appendix. The formula for the t test is:
sp2 N1 1 s1
2. N2 1 s22.
N1 N2 2
s M1 M2
sp2
N1
sp2
N2
τ obt
M1 M2
s M1 M2
Where τ obt is the value of t obtained through the data; N1 N2, is the number of elements in
each group; s12 s2
2, the estimates of the variances of the two populations; M1 M2, is the
means of the two groups; sp2 s M1 M2, are the values needed in order to arrive at τ obt
2.
Fig. 6.3.8-1 displays the t-test results using two different sets of data. The first set
compares the raw MonteCarlo generated (population 500) data vs. the raw Measured
91
(population 32) data. The second set compares the normalized cumulative frequency
distribution data (population 16). Both sets are displaying the magnitude of S11 and S21
using data generated from a circuit model file constructed using the correlated premise.
The t-test results begin to diverge at 1.84GHz for S11 magnitude and 4.5GHz for S21
magnitude.
T-Test For Independent SamplesComparing Cumulative Frequency Distribution Data For
MonteCarlo Prediction (16 Data Points) vs. Actual Measured Device (16 Data Points)
0.00
0.50
1.00
1.50
2.00
2.50
1.E+08 1.E+09 1.E+10
Frequency
S21 MagT Value Correlated Blocks
T-Test For Independent SamplesComparing
MonteCarlo Prediction (500 Data Points) vs. Actual Measured Device (32 Data Points)
-0.50
1.50
3.50
5.50
7.50
9.50
1.E+08 1.E+09 1.E+10
Frequency
S21 MagT Value Correlated Blocks
T-Test For Independent SamplesComparing
MonteCarlo Prediction (500 Data Points) vs. Actual Measured Device (32 Data Points)
-0.50
1.50
3.50
5.50
7.50
9.50
1.E+08 1.E+09 1.E+10
Frequency
S11 MagT Value Correlated Blocks
T-Test For Independent SamplesComparing Cumulative Frequency Distribution Data For
MonteCarlo Prediction (16 Data Points) vs. Actual Measured Device (16 Data Points)
-0.25
0.25
0.75
1.25
1.75
2.25
1.E+08 1.E+09 1.E+10
Frequency
S11 MagT Value Correlated Blocks
T-Test For S11/S21 MagnitudeUsing Data Generated Via Correlated Circuit File Configuration
T-Test For S11/S21 MagnitudeUsing Data Generated Via Correlated Circuit File Configuration
1.84GHz1.84GHz1.84GHz1.84GHz
4.5GHz4.5GHz
4.5GHz4.5GHz
Figure #6.3.8-1: T-Test Between MC and Measured Data For S11/S21 Magnitude
6.3.9 ANOVA
The analysis of variance (ANOVA) is a technique used to test for statistical
significance of the differences among the means of two or more groups. The test looks at
the amount of variability (the differences) between the means of the groups, compared
with the amount of variability among the individual scores in each group; that is, the
92
variance between groups versus the variance within groups [32]. The ANOVA starts
with the total amount of variability in the data and divides it up (partitioning) into various
categories. Basically, it compares the variability among the group means with the
variability that occurred just by chance or error. To calculate an ANOVA the total sum
of the squares:
sT2 X MT
2
NT 1
SST
dfT
is broken down into two parts; 1.) A sum of squares within groups, SSW, which reflects
the degree of variability within groups, however is not sensitive to overall differences
between the groups; and 2.) A sum of squares between groups, SSB, which reflects
differences between groups, however is not sensitive to variability within groups. The
total sum of squares is the sum of the sum of squares within and the sum of squares
between:
SST SSW SSB
The total degrees of freedom is:
dfT dfW dfB
Adding up all df’s within all the groups yields dfW. The number of groups minus 1 equal
dfB. The mean square within, a measure of the variability within groups is:
MSW
SSW
dfW
The mean square between, a measure of variability between groups is:
MSB
SSB
dfB
93
To find out if the between-group differences are significantly greater than chance,
comparing the two by computing the ratio:
Fobt
MSB
MSW
Fobt is the ratio of a mean square between groups to a mean square within groups. The
“obtained” (obt) value is compared with a critical value (extracted from tables) to
ascertain its validity. Fig. 6.3.9-1 displays the ANOVA results using two different sets of
data. The first set compares the raw MonteCarlo generated (population 500) data vs. the
raw Measured (population 32) data. The second set compares the normalized cumulative
frequency distribution data (population 16). Both sets are displaying the magnitude of
S11 and S21 using data generated from a circuit model file constructed using the
correlated premise. The ANOVA results begin to diverge at 1.74GHz for S11 magnitude
and 4.5GHz for S21 magnitude.
94
ANOVA AnalysisComparing Cumulative Frequency Distribution Data For
MonteCarlo Prediction (16 Data Points) vs. Actual Measured Device (16 Data Points)
-0.25
-0.05
0.15
0.35
0.55
0.75
0.95
1.15
1.35
1.E+08 1.E+09 1.E+10
Frequency
S21 MagF(obt) Correlated Blocks
ANOVA AnalysisComparing
MonteCarlo Prediction (500 Data Points) vs. Actual Measured Device (32 Data Points)
-1.00
4.00
9.00
14.00
19.00
24.00
29.00
1.E+08 1.E+09 1.E+10
Frequency
S21 MagF(obt) Correlated Blocks
ANOVA AnalysisComparing Cumulative Frequency Distribution Data For
MonteCarlo Prediction (16 Data Points) vs. Actual Measured Device (16 Data Points)
-0.02
0.18
0.38
0.58
0.78
0.98
1.E+08 1.E+09 1.E+10
Frequency
S11 MagF(obt) Correlated Blocks
ANOVA AnalysisComparing
MonteCarlo Prediction (500 Data Points) vs. Actual Measured Device (32 Data Points)
-0.25
4.75
9.75
14.75
19.75
24.75
29.75
1.E+08 1.E+09 1.E+10
Frequency
S11 MagF(obt) Correlated Blocks
ANOVA For S11/S21 MagnitudeUsing Data Generated Via Correlated Circuit File Configuration
ANOVA For S11/S21 MagnitudeUsing Data Generated Via Correlated Circuit File Configuration
1.74GHz1.74GHz 1.74GHz1.74GHz
4.5GHz4.5GHz 4.5GHz4.5GHz
Figure #6.3.9-1: ANOVA Between MC and Measured Data For S11/S21 Magnitude
6.4 Summary
This chapter explained in detail the method of predicting passive device
performance variations and yield, when provided with a statistical distribution of the
building block models and a circuit schematic. Correlation between the building block
equivalent circuit values was also examined for use in collating the circuit model file
used in the MonteCarlo simulations. A description of the MonteCarlo simulation process
and several statistical analysis techniques, used to evaluate and render the results, were
also presented. The data used in rendering the example plots was generated from a
circuit model file constructed using the “correlated” premise in contrast to the
95
noncorrelated or “independent” premise. In the following chapters, this predictive
modeling methodology will be applied to a 9-segment meander resistor using the
modeling methodology described in Chapters 4 and 5, and using the independent
premise.
96
CHAPTER 7
PREDICTIVE STATISTICAL ANALYSIS OF 9 SEGMENT
MEANDER RESISTOR
7.1 Introduction
As microelectronic technology continues to progress, there is a constant focus on
higher levels of system integration and miniaturization. For example, in many
applications it’s desirable to package several integrated circuits (ICs) together in
multichip modules (MCMs) to achieve further compactness and higher performance.
Passive components (i.e., capacitors, resistors, and inductors) are an essential
requirement for many MCM applications [35]. A significant advantage of MCM
technology is the ability to embed a large number of passive components directly into the
substrate at low cost; whereas, examination of component value statistical variation is
crucial for designing and characterizing their performance.
Successful design of passive structures requires that accurate models of the
various components exist or can be easily obtained. However, for high frequency designs
passive structures can be electrically long, and due to their full 3-dimensional geometries,
have very complex field patterns. Standard modeling methods for microstrip or stripline
based structures do not apply for these components. Therefore in order to successfully
97
design structures for high frequencies, the behavior of the passive components that
comprise the structure must be characterized accurately at those frequencies. Recently,
computer-aided design tools such as Hspice [31] have become indispensable in IC
design; whereby, accurate circuit simulation using Hspice is dependent on both the
structural validity of the device models and the accuracy of the values used as model
parameters. For that reason, the extraction of an optimum set of device model parameter
values is crucial in characterizing the relationship between the model and the measured
behavior.
In this chapter, the improved building block modeling methodology originally
created by Dr. Poddar [28] is further developed by focusing on the statistical variation of
the deembedded element values of the building block models. It will be demonstrated
that the variation in the complete equivalent circuit models, based only on the circuit
building blocks, can be used to predict such variations in actual fabricated devices. The
9-segment meander resistor fabricated on a 96% Alumina substrate will be the device
under inspection.
Three sets of resistor structures were fabricated on a 96% Alumina substrate: 32
straight metal segments (test structure 1), 32 three-segment resistors (test structure 2), and
32 nine-segment resistors (modeled device). Scattering parameter measurements ranging
in frequencies from 45MHz to 20GHz were taken from all the structures. The measured
s-parameter data from test structure 1 and 2 was used to deembed the element values of
the partial element equivalent circuit (PEEC) for both sets of 32 test structures, using a
non-linear optimization algorithm in Hspice. This process generated a unique equivalent
circuit model for each of the 64 test structures. Mean and relative variation was then
98
calculated for each deembedded component from the 64 unique equivalent circuit
models, and used in a MonteCarlo analysis of the 9-segment modeled device. This
analysis generated 500 s-parameter curves of the modeled device electrical behavior.
The s-parameter measurements from the MonteCarlo analysis and the actual
measured device were compared to determine if the measured data resides in the range of
curves generated by the MonteCarlo analysis. By inspection, it was noted that the
electrical behavior from the fabricated devices was in fact contained within the range
predicted by the MonteCarlo results. Therefore by means of comparison between the
MonteCarlo results and measured data, it was shown that the statistical variations of the
component values provide an accurate representation of the overall modeled device
performance.
7.2 Statistical Variations
A three-step MonteCarlo technique was used to accomplish the statistical
prediction: 1.) A statistical model was formed to characterize the fabrication process
used; 2.) A circuit design was created using a Spice type simulator and applying nominal
device values for the target process; and 3.) Randomly generated element values of the
process model were simulated in a MonteCarlo tool to produce a representative set of
output performance characteristics. The impact of random process variations can be
inferred from these simulations, and electrical yield can then be estimated using the
percentage of that sample which meets the performance requirements [29].
99
7.3 Test Structure Description
Two sets (test structures 1 and 2) of 32 test structures were fabricated by
depositing gold, using a sputtering technique, on a 96% Alumina substrate. The metal
thickness used was 3000 angstroms with 30µ widths and spacings for all connecting
segments. All resistor structures were on one layer, with connections to the devices made
using a ground-signal-ground probe pad pattern.
Test structure 1 consisted of two probe pads connected to a 30µ wide strip of
metal approximately 300µ long. Test structure 2 consisted of two probe pads connected
to five material squares which are connected to nine coupled material squares, and two
corners which are made up of five material squares.
The layout and fabrication mask photos of test structures’ 1 and 2 are shown in
Fig. 7.3-1.
Test StructuresTest Structures
Test Structure 2Test Structure 2Test Structure 1Test Structure 1
LayoutLayout
MaskMask
Figure #7.3-1: Test Structures 1 and 2 Layout and Mask Photos
100
These test structures allowed the modeling of the probe pads, the material squares,
the coupled material squares, and the corner primitives.
7.3.1 Processing and Measurement
The 96% Alumina structure was physically designed using integrated circuit
design tools within the Avanti design environment. Rena Huang PhD. Candidate in the
School of Electrical and Computer Engineering at the Georgia Institute of Technology
fabricated the design in Georgia Tech’s MiRC cleanroom. The size of the completed
coupon was approximately 2.25" x 2.25". Metal lines were drawn 30µ wide. The
embedded structures were interfaced accessed using ground-signal-ground probe pads.
The complete test structure coupon is shown in Fig. 7.3.1-1.
Test Structures & Measured DevicesTest Structures & Measured Devices
Figure #7.3.1-1: Test Structure Coupon
The test structures were measured using network analysis techniques. Since very
low loss metal was used in the manufacturing process, DC resistance measurements were
unreliable and not used. For high frequency measurements, an HP 8510C network
101
analyzer was used in conjunction with a Cascade Microtech probe station and ground-
signal-ground configuration probes. Calibration was accomplished using a supplied
substrate and the application of the line-reflect-match (LRM) calibration method. Data
was gathered for each of the test structures at 400 frequency points between 45MHz and
20GHz and stored by means of data acquisition software.
7.4 Modeling Scheme
7.4.1 9-Segment Meander Resistor Modeling Procedure
The improved modeling methodology detailed in Chapter 4 will be utilized to
model the 9-segment meander resistor. The fundamental idea behind this modeling
procedure is that passive structures are comprised of several key geometrical building
blocks. The building block equivalent circuits are derived from fabricated test structures
and measurements using optimization and extraction routines. Passive RLC models for
each embedded building block are extracted and take into account effects of processing
fluctuations and nonideal material properties. The objective here is to predict the
electrical behavior of arbitrary geometry passive devices in a standard circuit simulator;
therefore, providing a major speedup over methods that do not utilize lumped elements.
The steps involved in the modeling procedure for the 9-segment meander resistor
are described in a brief outline below:
102
1. The 64 test structures and 32 9-segment meander resistors were physically designed
(96% Alumina) and fabricated. Rena Huang PhD. Candidate fabricated the designs in
the MiRC cleanroom at Georgia Tech. High frequency s-parameter measurements of
the devices were taken by on-wafer ground-signal-ground probing.
2. The measured data was used in a circuit optimization input file to deembed the
element values for the 2 sets of 32 structures. Initial guesses used in the Hspice
Optimizer were made based on s-parameter measurements for one of each structure.
Once optimization for one structure was completed, the results were used for the
remaining optimizations.
3. 32 unique equivalent circuit models of the two test structures were obtained.
Fig. 7.4.1-1 shows the building blocks, the equivalent circuits, and some of the
element values generated from the Hspice Optimizer.
Test Structure 1Test Structure 1
Coupled Square PrimitiveCoupled Square Primitive
U-Shaped BendPrimitives
U-Shaped BendPrimitives
Pad PrimitivePad PrimitiveMaterial Square PrimitiveMaterial Square Primitive
Test Structure 2Test Structure 2
Building Blocks / PEECsBuilding Blocks / PEECs
1 3
2 4
R =0 . 09 Oh mL = 1 E -1 1 HC = 1 .1 e -1 5 FC M =0 .4 e -1 5 FCC = 1 .4 e -1 5 FL M = 0 .2 0
LC
R L R
C C CC
LC
R L R
C C C C
L M
L MCM
1
2
3
4
1 2
R = 0 . 08 O h mL = 1 E- 11 HC = 2.7 e- 1 5 FC C = 1. 2e - 15 F
L
C
R L R
C C C C
1 2
1 2
R =0 .0 8 O hmL = 1 E- 11 HC = 2.7 e- 15 FC C = 1. 2e - 15 F
L
C
R L R
C C C C
1 2
.
R= 0 .4 O h mL= 3 .7 E- 1 1 HC = 5.3 e -1 5 FCC= 2 7e - 1 5 F
L
C
R L R
CC CC
1 2
CC CC
2L
C
R L R1
1
2
L
C
R L R
CC CC
1 2
L
C
R L R
CC CC
1 2
L
C
R L R
CC CC
1 2
Figure #7.4.1-1: Building Blocks and Equivalent Circuits
103
7.4.2 Statistical Model
Two sets of 32 test structures (1 and 2) were characterized. Mean and absolute
deviations in the deembedded circuit values of the building blocks were then calculated
and used in the Hspice MonteCarlo simulator as limits to generate random circuit values
for 500 simulations. Fig. 7.4.2-1 shows the deembedded element values from test
structure 1 & 2 and the relative variations used in the MonteCarlo simulation tool.
Equivalent Circuit Element ValuesStatistical Variations
Equivalent Circuit Element ValuesStatistical Variations
TS1 Sum of Sq.s' c_sq csq lsq rsqlow 1.11576 3.36E-15 6.52E-26 1.73E-11 2.51E-06high 29.9662 2.49E-13 6.52E-26 1.91E-11 2.51E-06median 2.745285 1.08E-13 6.52E-26 1.84E-11 2.51E-06ab var 1.629525 1.05E-13 0 1.03E-12 0rel var 0.593572252 0.968897 0 0.056241 0
TS1 c_pad rpad lpad cpadlow 1.2655E-12 3.75288 2.79E-11 2.21E-14high 1.87813E-12 5.625027 3.73E-11 2.59E-14median 1.58577E-12 4.37748 3.05E-11 2.44E-14ab var 3.20265E-13 0.6246 2.68E-12 2.27E-15rel var 0.2019619 0.142685 0.087727 0.092907
TS2 Sum of Sq.s' cou_1 c_cou l2 r2low 0.71189 0.159073 9.2E-25 1.27E-13 0.212918high 440.398 0.324847 2.74E-23 8.36E-12 1.589938median 1.690735 0.218506 6.36E-24 4.07E-12 0.753571ab var 0.978845 0.059433 5.44E-24 3.95E-12 0.540653rel var 0.578946435 0.271997 0.855404 0.968861 0.717455
TS2 c2 c_co2cr l2_2 r2_2 c2_2low 2.72533E-16 1.65E-15 1.51E-12 0.097137 2.8E-16high 2.79231E-15 3.38E-11 9.79E-12 1.115453 3.48E-15median 1.32663E-15 7.69E-12 4.16E-12 0.755395 1.42E-15ab var 1.0541E-15 7.68E-12 2.64E-12 0.658258 1.14E-15rel var 0.794567513 0.999786 0.635701 0.871409 0.803456
TS2 c_co2cr_2 l2_3 r2_3 c2_3 c_co2cr_3low 2.35967E-13 8.79E-13 0.278774 1.37E-18 2.77E-12high 1.14784E-11 1.21E-11 1.109648 4.25E-15 2.6E-11median 7.00171E-12 4.42E-12 0.705214 1.41E-15 7.81E-12ab var 6.76575E-12 3.54E-12 0.426439 1.41E-15 5.03E-12rel var 0.966298662 0.801238 0.604695 0.999028 0.644819
Figure #7.4.2-1: Mean and Absolute Deviations for Deembedded Element Values
104
7.4.3 Circuit Design
An equivalent circuit of the 9-segment meander resistor was constructed by
replacing each building block in the structure with its equivalent extracted circuit values.
Due to the lack of correlation between the building blocks, each element is independent
of any other element in the circuit. Therefore, to create an accurate circuit design all the
elements in the design need to generate a random value for the defined number of
simulations. The 9-segment meander resistor generated 1336 random circuit values for
each simulation. This circuit file was automatically generated using Unix shell scripts
detailed in Chapter 9. Fig. 7.4.3-1 is a block diagram of a 9-segment meander resistor
and shows the amount of random number generations per simulation, for the PEEC
element values.
MaterialSquare
R=0.08 OhmL= 1E-11 HC = 2.7e-15 FCC = 1.2e-15 F
LC
R L R
CC CC1 2
Pad
R=0.08 OhmL=1E-11 HC = 2.7e- 15 FCC = 1.2e-15 F
LC
R L R
CC CC1 2
Coupled Pair
R=0.09 OhmL= 1E-11 HC = 1.1e-15 FCM=0.4e-15 FCC = 1.4e-15 FLM = 0.20
LC
R L R
CC CC
LC
R L R
CC CC
L MLM
CM
1
2
3
4
CC
LC
R1
Multi-Block
LC
R L R
CC CC1 2
LC
R L R
CC CC1 2
.
R=0.4 OhmL= 3.7E-11 HC = 5.3e-15 FCC=2 7e-15 F
CC2
L R
LC
R L R
CC CC1 2
LC
R L R
CC CC1 2
Pad
Pad
Pads = (2 BBlocks * 8 Elements) = 16 ElementsMaterial Squares = (8 BBlocks * 8 Elements) = 64 ElementsCoupled Squares = [((4 BBlocks * 9 Coupled) * 3) * 8 Elements] + (9 Coupled * 8 Elements) = 936 Elements
Corners = (5 BBlocks * 8 Corners * 8 Elements) = 320 Elements
TOTAL = 1336 Random Generated Element Values
Figure #7.4.3-1: Number of Random Number Generation for MonteCarlo
Simulation
105
7.4.4 MonteCarlo Analysis
Based on the relative variation calculated for each element value in the
equivalent circuit models, a MonteCarlo analysis was performed using Hspice. A
MonteCarlo analysis with a uniform parameter distribution simulated 500 sets of circuit
model parameters as shown in Fig. 7.4.4-1.
MonteCarlo Analysis Results Using Independent PremiseMonteCarlo Analysis Results Using Independent Premise
S21 RealS21 Real
S11 RealS11 Real
S21 ImaginaryS21 Imaginary
S11 ImaginaryS11 Imaginary
Figure #7.4.4-1: MonteCarlo Analysis Generated 500 Simulation
S-parameter data was generated for each of the 500 simulations, and extracted
from the output file using Unix scripts (detailed in Chapter 9). Due to a Microsoft Excel
limitation of 255 series plots per graph, all the data could not be rendered together.
Therefore, the MonteCarlo maximum and minimum were found for S11/S21
106
magnitude/phase and used as upper and lower limits (red) with the actual measured data
contained within the predicted range. This plot is shown in Fig. 7.4.4-2.
Following the simulations, the output response for the modeled device was
compared to actual measured data from the 32 fabricated 9-segment structures. It was
established that the electrical behavior from the 32 actual measured devices, was
predicted by the MonteCarlo analysis using the statistical variations of the equivalent
circuit element values deembedded from the test structures, as shown in Fig. 7.4.4-2.
S21 Magnitude MonteCarlo Range (Red) vs. Actual Measured Devices
0.36
0.41
0.46
0.51
0.56
0.61
1.E+08 1.E+09 1.E+10 1.E+11
Frequency
S21Mag
S11 Phase MonteCarlo Range (Red) vs. Actual Measured Devices
-90
-70-50
-30
-1010
30
5070
90
1.E+08 1.E+09 1.E+10 1.E+11
Frequency
S11Phase
S11 Magnitude MonteCarlo Range (Red) vs. Actual Measured Devices
0.380.430.480.530.580.630.680.730.78
1.E+08 1.E+09 1.E+10 1.E+11
Frequency
S11Mag
MonteCarlo Predicted Range vs. Actual Measured 9-Segment Resistors MonteCarlo Predicted Range vs. Actual Measured 9-Segment Resistors
S21 Phase MonteCarlo Range (Red) vs. Actual Measured Devices
-90
-70-50
-30-10
1030
5070
90
1.E+08 1.E+09 1.E+10 1.E+11
Frequency
S21Phase
Figure #7.4.4-2: MonteCarlo Predicted vs. Actual Measured
A close-up of the MonteCarlo vs. measured S11 magnitude plot (Fig. 7.4.4-3)
shows that the measured data stays within the MonteCarlo predicted range up to 2GHz.
By inspection, it appears that the parasitics begin to take effect at approximately 1GHz;
107
therefore, the deembedded model for the 9-segment resistor accurately predict the
measured electrical behavior range past the first resonance for S11 magnitude.
A close-up of the MonteCarlo vs. measured S21 magnitude plot (Fig.7.4.4-3)
shows that the measure data stays within the MonteCarlo predicted range up to 5GHz.
Again by inspection, it appears that the parasitics begin to take effect at 1GHz and at
3GHz; therefore, the deembedded model accurately predicts the measured behavior range
for S21 magnitude past the first and second resonance.
The MonteCarlo vs. measured phase plots were not as impressive as the
magnitude plots, because the measured data fell outside the MonteCarlo predicted range;
however, the trends were identical. Looking at the scale of +/- 90 degrees in Fig. 7.4.4-2,
it appears that the error between the MonteCarlo vs. measured is trivial compared to an
inductive 90 degrees and a capacitive –90 degrees of phase shift. The error actually
seems to be some DC offset.
MonteCarlo Predicted Range vs. Actual Measured 9-Segment ResistorsFor S11/S21 Magnitude
MonteCarlo Predicted Range vs. Actual Measured 9-Segment ResistorsFor S11/S21 Magnitude
S11 Magnitude MonteCarlo Range (Red) vs. Actual Measured Devices
0.38
0.4
0.42
0.44
0.46
0.48
0.5
0.52
1.E+08 1.E+09 1.E+10
Frequency
S11Mag
S21 Magnitude MonteCarlo Range (Red) vs. Actual Measured Devices
0.46
0.48
0.5
0.52
0.540.56
0.58
0.6
0.62
1.E+08 1.E+09 1.E+10
Frequency
S21Mag
2GHz2GHz
5GHz5GHz
Figure #7.4.4-3: MonteCarlo Predicted vs. Actual Measured
108
7.4.5 Statistical Analysis Methods
To accurately compare the MonteCarlo range vs. measured results, several
statistical methods were used to find a number that represented both sets of data at each
frequency point.
7.4.6 Mean and Standard Deviation
The mean is a measure of central tendency, “middleness”, or the arithmetic
average. Adding up all the scores and dividing by the number of scores is the method of
calculating the mean. The variance is the most frequently used measure of variability
[32]. The standard deviation, derived from the variance, also indicates the relative
variability within a group and is calculated by taking the square root of the variance.
To aid in the evaluation of measured vs. modeled data, percent differences in
mean and standard deviation will be used. Fig. 7.4.6-1 displays the Mean Value Percent
Difference results using two different sets of data. The first set compares the raw
MonteCarlo generated (population 500) data vs. the raw Measured (population 32) data.
The second set compares the normalized cumulative frequency distribution data
(population 16). Both sets are displaying the mean percent difference in magnitude of
S11 and S21, and are comparing data generated from a circuit model file constructed
using the correlated premise (Blue Plot), and the independent premise (Yellow Plot).
The results for both types of premise begin to diverge at 1.84GHz for S11
magnitude and 4.5GHz for S21 magnitude, however; the results using the correlated
premise have a lower percent difference at low frequencies (less than 1GHz) and a higher
percent difference beyond 1GHz.
109
Mean Percent Error Comparing Cumulative Frequency Distribution Data For
MonteCarlo Prediction (16 Data Points) vs. Actual Measured Device (16 Data Points)
-1.00%
9.00%
19.00%
29.00%
39.00%
49.00%
1.E+08 1.E+09 1.E+10
Frequency
S21 Mag% Error
Independent Blocks
Correlated Blocks
Mean Percent ErrorComparing
MonteCarlo Prediction (500 Data Points) vs. Actual Measured Device (32 Data Points)
-1.00%
-0.50%
0.00%
0.50%
1.00%
1.50%
2.00%
1.E+08 1.E+09 1.E+10
Frequency
S21 Mag% Error
Independent Blocks
Correlated Blocks
Mean Percent Error Comparing Cumulative Frequency Distribution Data For
MonteCarlo Prediction (16 Data Points) vs. Actual Measured Device (16 Data Points)
-5.00%
0.00%
5.00%
10.00%
15.00%
20.00%
25.00%
30.00%
35.00%
40.00%
1.E+08 1.E+09 1.E+10
Frequency
S11 Mag% Error
Independent Blocks
Correlated Blocks
Mean Percent ErrorComparing
MonteCarlo Prediction (500 Data Points) vs. Actual Measured Device (32 Data Points)
-0.10%
0.40%
0.90%
1.40%
1.90%
1.E+08 1.E+09 1.E+10
Frequency
S11 Mag% Error
Independent Blocks
Correlated Blocks
Mean Percent Difference For S11/S21 MagnitudeUsing Data Generated Via Correlated Circuit Configuration vs. Independent Circuit Configuration
Mean Percent Difference For S11/S21 MagnitudeUsing Data Generated Via Correlated Circuit Configuration vs. Independent Circuit Configuration
1.84GHz1.84GHz 1.84GHz1.84GHz
4.5GHz4.5GHz 4.5GHz4.5GHz
Figure #7.4.6-1: Mean Value Percent Difference Between MonteCarlo and
Measured Data For S11/S21 Magnitude
Fig. 7.4.6-2 displays the Standard Deviation Percent Difference results using two
different sets of data. The first set compares the raw MonteCarlo generated (population
500) data vs. the raw Measured (population 32) data. The second set compares the
normalized cumulative frequency distribution data (population 16). Both sets are
displaying the standard deviation percent difference of the magnitude of S11 and S21,
and are comparing data generated from a circuit model file constructed using the
correlated premise (Blue Plot), and the independent premise (Yellow Plot).
The results for both types of premise begin to diverge at 2.5GHz for S11
magnitude using the first set of data, and at 1.74GHz for S11 magnitude using the second
set of data. The results begin to diverge at 4.2GHz for S21 magnitude using the first set
110
of data, and at 7.0GHz for S21 magnitude using the second set of data. However; the
results using the independent premise have a much lower percent difference at all
frequencies for S11 and S21 magnitude.
Standard Deviation Percent Error Comparing Cumulative Frequency Distribution Data For
MonteCarlo Prediction (16 Data Points) vs. Actual Measured Device (16 Data Points)
-1.00%
4.00%
9.00%
14.00%
19.00%
24.00%
29.00%
1.E+08 1.E+09 1.E+10
Frequency
S21 Mag% Error
Independent Blocks
Correlated Blocks
Standard Deviation Pecent ErrorComparing
MonteCarlo Prediction (500 Data Points) vs. Actual Measured Device (32 Data Points)
0.00%
20.00%
40.00%
60.00%
80.00%
100.00%
120.00%
1.E+08 1.E+09 1.E+10
Frequency
S21 Mag% Error
Independent Blocks
Correlated Blocks
Standard Deviation Percent Error Comparing Cumulative Frequency Distribution Data For
MonteCarlo Prediction (16 Data Points) vs. Actual Measured Device (16 Data Points)
-1.00%
4.00%
9.00%
14.00%
19.00%
24.00%
29.00%
1.E+08 1.E+09 1.E+10
Frequency
S11 Mag% Error
Independent Blocks
Correlated Blocks
Standard Deviation Pecent ErrorComparing
MonteCarlo Prediction (500 Data Points) vs. Actual Measured Device (32 Data Points)
0.00%
20.00%
40.00%
60.00%
80.00%
100.00%
120.00%
140.00%
160.00%
180.00%
1.E+08 1.E+09 1.E+10
Frequency
S11 Mag% Error
Independent Blocks
Correlated Blocks
Standard Deviation Percent Difference For S11/S21 MagnitudeUsing Data Generated Via Correlated Circuit Configuration vs. Independent Circuit Configuration
Standard Deviation Percent Difference For S11/S21 MagnitudeUsing Data Generated Via Correlated Circuit Configuration vs. Independent Circuit Configuration
2.5GHz2.5GHz
1.74GHz1.74GHz
4.2GHz4.2GHz7.0GHz7.0GHz
Figure #7.4.6-2: Standard Deviation Percent Difference Between MonteCarlo and
Measured Data For S11/S21 Magnitude
7.4.7 Frequency Distribution Plots
The s-parameters generated from the MonteCarlo results and the actual measured
devices were transposed from rows representing a particular frequency to columns
representing MonteCarlo simulations and measured device data at a particular frequency
point. The range of values from high to low was split up into sixteen linearly place filter
111
bins. The frequency of occurrence was calculated and normalized to the number of data
points (500 for MonteCarlo, and 32 for actual measured). These normalized cumulative
frequency distribution (NCFD) plots were generated at every frequency point for
S11/S21 magnitude up to 7.5GHz. These plots illustrate that the measured data from
each fabricated device at a particular frequency fits, or doesn’t fit, within the MonteCarlo
predicted range. The trend lines generated within the NCFD plots also show how the
standard deviation between predicted and measured compare.
Fig. 7.4.7-1 and Fig. 7.4.7-2 display NCFD plots at several frequency points for
S11 magnitude, and Fig. 7.4.7-3 and Fig. 7.4.7-4 display NCFD plots at several
frequency points for S21 magnitude. Data generated via the correlated premise vs. the
independent premise is compared in all plots. By inspection, the standard deviation
match between the MonteCarlo predicted vs. measured, using the independent premise,
clearly outperforms the correlated premise at both frequencies.
112
MonteCarlo Prediction vs. Actual Measured Device NCFD Plot S11 Magnitude @ 600MHz Using Independent Premise
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.392
0
0.395
9
0.399
8
0.403
8
0.407
7
0.411
6
0.415
5
0.419
4
0.423
4
0.427
3
0.431
2
0.435
1
0.439
0
0.443
0
0.446
9
0.450
8
Filter Bins
NCFD
MonteCarloMeasuredMonteCarlo TrendMeasured Trend
MonteCarlo Prediction vs. Actual Measured Device NCFD Plot S11 Magnitude @ 800MHz Using Independent Premise
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.394
4
0.398
2
0.402
1
0.405
9
0.409
7
0.413
6
0.417
4
0.421
3
0.425
1
0.428
9
0.432
8
0.436
6
0.440
5
0.444
3
0.448
1
0.452
0
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
MonteCarlo Prediction vs. Actual Measured Device NCFD Plot S11 Magnitude @ 950MHz Using Independent Premise
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.396
6
0.400
4
0.404
1
0.407
9
0.411
7
0.415
4
0.419
2
0.423
0
0.426
7
0.430
5
0.434
3
0.438
0
0.441
8
0.445
6
0.449
3
0.453
1
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
600MHz600MHz
800MHz800MHz
950MHz950MHz
MonteCarlo Prediction vs. Actual Measured Device NCFD PlotS11 Magnitude @ 600MHz Using Correlated Premise
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.397
8
0.400
7
0.403
6
0.406
5
0.409
4
0.412
2
0.415
1
0.418
0
0.420
9
0.423
7
0.426
6
0.429
5
0.432
4
0.435
2
0.438
1
0.441
0
Filter Bins
NCFD
MonteCarloMeasuredMonteCarlo TrendMeasured Trend
MonteCarlo Prediction vs. Actual Measured Device NCFD PlotS11 Magnitude @ 800MHz Using Correlated Premise
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.399
9
0.402
8
0.405
6
0.408
4
0.411
2
0.414
1
0.416
9
0.419
7
0.422
5
0.425
4
0.428
2
0.431
0
0.433
9
0.436
7
0.439
5
0.442
3
Filter Bins
NCFD
MonteCarloMeasuredMonteCarlo TrendMeasured Trend
MonteCarlo Prediction vs. Actual Measured Device NCFD PlotS11 Magnitude @ 950MHz Using Correlated Premise
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.401
9
0.404
7
0.407
4
0.410
2
0.413
0
0.415
8
0.418
6
0.421
3
0.424
1
0.426
9
0.429
7
0.432
5
0.435
2
0.438
0
0.440
8
0.443
6
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
600MHz600MHz
800MHz800MHz
950MHz950MHz
Figure #7.4.7-1: NCFD Plot Comparisons Between Correlated Premise and
Independent Premise For S11
113
MonteCarlo Prediction vs. Actual Measured Device NCFD Plot S11 Magnitude @ 1.15GHz Using Independent Premise
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.400
1
0.403
7
0.407
4
0.411
0
0.414
7
0.418
4
0.422
0
0.425
7
0.429
3
0.433
0
0.436
6
0.440
3
0.443
9
0.447
6
0.451
2
0.454
9
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
MonteCarlo Prediction vs. Actual Measured Device NCFD Plot S11 Magnitude @ 1.5GHz Using Independent Premise
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.408
7
0.412
1
0.415
5
0.418
8
0.422
2
0.425
6
0.429
0
0.432
4
0.435
7
0.439
1
0.442
5
0.445
9
0.449
3
0.452
6
0.456
0
0.459
4
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
MonteCarlo Prediction vs. Actual Measured Device NCFD Plot S11 Magnitude @ 2.0GHz Using Independent Premise
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.421
1
0.424
2
0.427
3
0.430
3
0.433
4
0.436
5
0.439
6
0.442
6
0.445
7
0.448
8
0.451
9
0.455
0
0.458
0
0.461
1
0.464
2
0.467
3
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
1.15GHz1.15GHz
1.5GHz1.5GHz
2.0GHz2.0GHz
MonteCarlo Prediction vs. Actual Measured Device NCFD PlotS11 Magnitude @ 1.15GHz Using Correlated Premise
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.405
0
0.407
7
0.410
4
0.413
1
0.415
8
0.418
5
0.421
2
0.423
9
0.426
6
0.429
3
0.432
0
0.434
7
0.437
4
0.440
1
0.442
8
0.445
6
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
MonteCarlo Prediction vs. Actual Measured Device NCFD PlotS11 Magnitude @ 1.5GHz Using Correlated Premise
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.412
7
0.415
2
0.417
7
0.420
3
0.422
8
0.425
3
0.427
8
0.430
4
0.432
9
0.435
4
0.437
9
0.440
5
0.443
0
0.445
5
0.448
0
0.450
6
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
MonteCarlo Prediction vs. Actual Measured Device NCFD PlotS11 Magnitude @ 2.0GHz Using Correlated Premise
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.424
9
0.427
2
0.429
4
0.431
7
0.433
9
0.436
2
0.438
4
0.440
7
0.442
9
0.445
2
0.447
4
0.449
7
0.451
9
0.454
2
0.456
4
0.458
7
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
1.15GHz1.15GHz
1.5GHz1.5GHz
2.0GHz2.0GHz
Figure #7.4.7-2: NCFD Plot Comparisons Between Correlated Premise and
Independent Premise For S11
114
MonteCarlo Prediction vs. Actual Measured Device NCFD Plot S21 Magnitude @ 1.0GHz Using Independent Premise
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.557
5
0.561
5
0.565
4
0.569
3
0.573
3
0.577
2
0.581
1
0.585
1
0.589
0
0.593
0
0.596
9
0.600
8
0.604
8
0.608
7
0.612
6
0.616
6
Filter Bins
NCFD
MonteCarloMeasured
MonteCarlo TrendMeasured Trend
MonteCarlo Prediction vs. Actual Measured Device NCFD Plot S21 Magnitude @ 800MHz Using Independent Premise
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.558
0
0.561
9
0.565
9
0.569
9
0.573
8
0.577
8
0.581
8
0.585
7
0.589
7
0.593
7
0.597
6
0.601
6
0.605
6
0.609
6
0.613
5
0.617
5
Filter Bins
NCFD
MonteCarloMeasured
MonteCarlo TrendMeasured Trend
MonteCarlo Prediction vs. Actual Measured Device NCFD Plot S21 Magnitude @ 2.0GHz Using Independent Premise
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.553
6
0.557
2
0.560
9
0.564
6
0.568
2
0.571
9
0.575
6
0.579
3
0.582
9
0.586
6
0.590
3
0.593
9
0.597
6
0.601
3
0.604
9
0.608
6
Filter Bins
NCFD
MonteCarlo
MeasuredMonteCarlo TrendMeasured Trend
800MHz800MHz
1.0GHz1.0GHz
2.0GHz2.0GHz
MonteCarlo Prediction vs. Actual Measured Device NCFD PlotS21 Magnitude @ 800MHz Using Correlated Premise
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.565
0
0.567
9
0.570
8
0.573
8
0.576
7
0.579
6
0.582
5
0.585
4
0.588
3
0.591
2
0.594
1
0.597
1
0.600
0
0.602
9
0.605
8
0.608
7
Filter Bins
NCFD
MonteCarloMeasuredMonteCarlo TrendMeasured Trend
MonteCarlo Prediction vs. Actual Measured Device NCFD PlotS21 Magnitude @ 1.0GHz Using Correlated Premise
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.564
5
0.567
4
0.570
3
0.573
2
0.576
1
0.579
0
0.581
9
0.584
8
0.587
7
0.590
6
0.593
4
0.596
3
0.599
2
0.602
1
0.605
0
0.607
9
Filter Bins
NCFD
MonteCarloMeasuredMonteCarlo TrendMeasured Trend
MonteCarlo Prediction vs. Actual Measured Device NCFD PlotS21 Magnitude @ 2.0GHz Using Correlated Premise
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.559
9
0.562
6
0.565
4
0.568
1
0.570
9
0.573
6
0.576
4
0.579
2
0.581
9
0.584
7
0.587
4
0.590
2
0.592
9
0.595
7
0.598
4
0.601
2
Filter Bins
NCFD
MonteCarloMeasuredMonteCarlo TrendMeasured Trend
800MHz800MHz
1.0GHz1.0GHz
2.0GHz2.0GHz
Figure #7.4.7-3: NCFD Plot Comparisons Between Correlated Premise and
Independent Premise For S21
115
MonteCarlo Prediction vs. Actual Measured Device NCFD Plot S21 Magnitude @ 3.0GHz Using Independent Premise
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.547
1
0.550
5
0.553
9
0.557
4
0.560
8
0.564
2
0.567
6
0.571
0
0.574
4
0.577
8
0.581
3
0.584
7
0.588
1
0.591
5
0.594
9
0.598
3
Filter Bins
NCFD
MonteCarloMeasured
MonteCarlo TrendMeasured Trend
MonteCarlo Prediction vs. Actual Measured Device NCFD Plot S21 Magnitude @ 4.0GHz Using Independent Premise
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.538
6
0.541
7
0.544
8
0.547
9
0.551
1
0.554
2
0.557
3
0.560
4
0.563
6
0.566
7
0.569
8
0.572
9
0.576
0
0.579
2
0.582
3
0.585
4
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
MonteCarlo Prediction vs. Actual Measured Device NCFD Plot S21 Magnitude @ 6.0GHz Using Independent Premise
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.517
0
0.519
9
0.522
8
0.525
7
0.528
6
0.531
5
0.534
4
0.537
3
0.540
2
0.543
1
0.546
0
0.548
9
0.551
8
0.554
7
0.557
6
0.560
5
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
3.0GHz3.0GHz
4.0GHz4.0GHz
6.0GHz6.0GHz
MonteCarlo Prediction vs. Actual Measured Device NCFD PlotS21 Magnitude @ 3.0GHz Using Correlated Premise
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.552
9
0.555
4
0.558
0
0.560
5
0.563
1
0.565
6
0.568
1
0.570
7
0.573
2
0.575
8
0.578
3
0.580
9
0.583
4
0.586
0
0.588
5
0.591
1
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
MonteCarlo Prediction vs. Actual Measured Device NCFD PlotS21 Magnitude @ 4.0GHz Using Correlated Premise
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.543
8
0.546
0
0.548
3
0.550
6
0.552
9
0.555
2
0.557
5
0.559
8
0.562
1
0.564
3
0.566
6
0.568
9
0.571
2
0.573
5
0.575
8
0.578
1
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
MonteCarlo Prediction vs. Actual Measured Device NCFD PlotS21 Magnitude @ 6.0GHz Using Correlated Premise
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.521
4
0.524
3
0.527
1
0.530
0
0.532
9
0.535
8
0.538
6
0.541
5
0.544
4
0.547
2
0.550
1
0.553
0
0.555
9
0.558
7
0.561
6
0.564
5
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo TrendMeasured Trend
3.0GHz3.0GHz
4.0GHz4.0GHz
6.0GHz6.0GHz
Figure #7.4.7-4: NCFD Plot Comparisons Between Correlated Premise and
Independent Premise For S21
7.4.8 Pearson Product-Moment Correlation Coefficient
Using the normalized cumulative distribution information, a measure of linear
relationship between the MonteCarlo predicted and actual measured data was calculated
using the correlation coefficient. A perfect correlation between two variables results in r
= 1.0, no correlation at all results in r = 0.0, and any number in-between means that there
is some correlation between the variables however the relationship is not perfect.
116
Fig. 7.4.8-1 displays the Correlation Coefficient results using the normalized
cumulative frequency distribution data (population 16) generated from the magnitude of
S11 and S2. The Correlation Coefficients are compared to data generated via the
correlated premise (Blue Plot) vs. the independent premise (Yellow Plot).
The results for both types of premise begin to diverge at 2.0GHz for S11
magnitude and 4.5GHz for S21 magnitude. The results using the independent premise
have a higher Correlation Coefficient than the results using the correlated premise at all
the frequencies.
Corralation CofficientComparing Cumulative Frequency Distribution Data For
MonteCarlo Prediction (16 Data Points) vs. Actual Measured Device (16 Data Points)
0.91
0.92
0.93
0.94
0.95
0.96
0.97
0.98
0.99
1.00
1.E+08 1.E+09 1.E+10
Frequency
S11 MagCorrelation
Coff.
Independent Blocks
Correlated Blocks
Corralation CofficientComparing Cumulative Frequency Distribution Data For
MonteCarlo Prediction (16 Data Points) vs. Actual Measured Device (16 Data Points)
0.94
0.95
0.96
0.97
0.98
0.99
1.00
1.01
1.E+08 1.E+09 1.E+10
Frequency
S21 MagCorrelation
Coff.
Independent Blocks
Correlated Blocks
Correlation Coefficient For S11/S21 MagnitudeUsing Data Generated Via Correlated Circuit Configuration vs. Independent Circuit Configuration
Correlation Coefficient For S11/S21 MagnitudeUsing Data Generated Via Correlated Circuit Configuration vs. Independent Circuit Configuration
2.0GHz2.0GHz4.5GHz4.5GHz
Figure #7.4.8-1: Pearson Product-Moment Correlation Coefficient for MonteCarlo
Predicted and Actual Measured Data
7.4.9 T-Test
The primary purpose of the “t” test is to determine whether the means of two
groups of numbers differ to a statistically significant degree. The larger the absolute
value of t, the more likely it is to reflect a significant difference between the two groups
under comparison.
117
Fig. 7.4.9-1 displays the t-test results using two different sets of data. The first set
compares the raw MonteCarlo generated (population 500) data vs. the raw Measured
(population 32) data. The second set compares the normalized cumulative frequency
distribution data (population 16). Both sets are displaying the t-test values using the
magnitude of S11 and S21, and are comparing data generated via the correlated premise
(Blue Plot) vs. the independent premise (Yellow Plot).
The results for both types of premise begin to diverge at 2.0GHz for S11
magnitude and 4.5GHz for S21 magnitude, however; the results using the independent
premise have a lower t-test value at higher frequencies (greater than 1.14GHz for S11,
and greater then 2.5GHz for S21) and a slightly lower value below 1.14GHz and 2.5GHz
for S11 and S21 respectively.
T-Test For Independent SamplesComparing Cumulative Frequency Distribution Data For
MonteCarlo Prediction (16 Data Points) vs. Actual Measured Device (16 Data Points)
0.00
0.50
1.00
1.50
2.00
2.50
1.E+08 1.E+09 1.E+10
Frequency
S21 MagT Value
Independent Blocks
Correlated Blocks
T-Test For Independent SamplesComparing
MonteCarlo Prediction (500 Data Points) vs. Actual Measured Device (32 Data Points)
-0.50
1.50
3.50
5.50
7.50
9.50
1.E+08 1.E+09 1.E+10
Frequency
S21 MagT Value
Independent Blocks
Correlated Blocks
T-Test For Independent SamplesComparing Cumulative Frequency Distribution Data For
MonteCarlo Prediction (16 Data Points) vs. Actual Measured Device (16 Data Points)
-0.25
0.25
0.75
1.25
1.75
2.25
1.E+08 1.E+09 1.E+10
Frequency
S11 MagT Value
Independent Blocks
Correlated Blocks
T-Test For Independent SamplesComparing
MonteCarlo Prediction (500 Data Points) vs. Actual Measured Device (32 Data Points)
-0.50
1.50
3.50
5.50
7.50
9.50
1.E+08 1.E+09 1.E+10
Frequency
S11 MagT Value
Independent Blocks
Correlated Blocks
T-Test For S11/S21 MagnitudeUsing Data Generated Via Correlated Circuit Configuration vs. Independent Circuit Configuration
T-Test For S11/S21 MagnitudeUsing Data Generated Via Correlated Circuit Configuration vs. Independent Circuit Configuration
2.0GHz2.0GHz2.0GHz2.0GHz
4.5GHz4.5GHz
4.5GHz4.5GHz
Figure #7.4.9-1: T-Test for MonteCarlo Predicted and Actual Measured Data
118
7.4.10 ANOVA
The analysis of variance (ANOVA) is a technique used to test for statistical
significance of the differences among the means of two or more groups. The test looks at
the amount of variability (the differences) between the means of the groups, compared
with the amount of variability among the individual scores in each group; that is, the
variance between groups versus the variance within groups.[terrified] The ANOVA
starts with the total amount of variability in the data and divides it up (partitioning) into
various categories. Basically, it compares the variability among the group means with
the variability that occurred just by chance or error.
Fig. 7.4.10-1 displays the ANOVA results using two different sets of data. The
first set compares the raw MonteCarlo generated (population 500) data vs. the raw
Measured (population 32) data. The second set compares the normalized cumulative
frequency distribution data (population 16). Both sets are displaying the ANOVA values
using the magnitude of S11 and S21, and are comparing data generated via the correlated
premise (Blue Plot) vs. the independent premise (Yellow Plot).
The results for both types of premise begin to diverge at 1.84GHz for S11
magnitude and 4.5GHz for S21 magnitude, however; the results using the independent
premise have a lower ANOVA value at higher frequencies (greater than 1.0GHz for S11,
and greater then 2.0GHz for S21).
119
ANOVA AnalysisComparing Cumulative Frequency Distribution Data For
MonteCarlo Prediction (16 Data Points) vs. Actual Measured Device (16 Data Points)
-0.25
-0.05
0.15
0.35
0.55
0.75
0.95
1.15
1.35
1.E+08 1.E+09 1.E+10
Frequency
S21 MagF(obt)
Independent Blocks
Correlated Blocks
ANOVA AnalysisComparing
MonteCarlo Prediction (500 Data Points) vs. Actual Measured Device (32 Data Points)
-1.00
4.00
9.00
14.00
19.00
24.00
29.00
1.E+08 1.E+09 1.E+10
Frequency
S21 MagF(obt)
Independent Blocks
Correlated Blocks
ANOVA AnalysisComparing Cumulative Frequency Distribution Data For
MonteCarlo Prediction (16 Data Points) vs. Actual Measured Device (16 Data Points)
-0.02
0.18
0.38
0.58
0.78
0.98
1.E+08 1.E+09 1.E+10
Frequency
S11 MagF(obt)
Independent Blocks
Correlated Blocks
ANOVA AnalysisComparing
MonteCarlo Prediction (500 Data Points) vs. Actual Measured Device (32 Data Points)
-0.25
4.75
9.75
14.75
19.75
24.75
29.75
1.E+08 1.E+09 1.E+10
Frequency
S11 MagF(obt)
Independent Blocks
Correlated Blocks
ANOVA For S11/S21 Magnitude Using Data Generated Via Correlated Circuit Configuration vs. Independent Circuit Configuration
ANOVA For S11/S21 Magnitude Using Data Generated Via Correlated Circuit Configuration vs. Independent Circuit Configuration
1.84GHz1.84GHz1.84GHz1.84GHz
4.8GHz4.8GHz 4.8GHz4.8GHz
Figure #7.4.10-1: ANOVA for MonteCarlo Predicted and Actual Measured Data
7.5 Summary
This chapter explained in detail the application of the predictive statistical
analysis methodology, described in Chapter 6, to a 9-segment meander resistor.
Modeling the 9-segment resistor, successfully proved the method of predicting passive
device performance variations and yield, when provided with a statistical distribution of
the building block models and a circuit schematic.
Using the improved method from Chapter 4 and the independent premise when
collating the circuit model file, results from the 9-segment meander resistor analysis were
120
compared to the analysis of the same 9-segment meander resistor; using the same
modeling methodology; and, using the correlated premise when collating the circuit
model file.
The results of the Mean Percent Difference, the t-test, and the ANOVA showed
that both types of premise produced the same behavior at lower frequencies (1GHz for
S11 and 3GHz for S21), however at higher frequencies the independent premise out
preformed the correlated premise.
The results of the Standard Deviation Percent Difference, and the Correlation
Coefficient showed an overall improvement at all frequencies using the independent
premise. The NCFD plots best exemplified the contrast in the data generated by the two
types of premise, with the independent premise outperforming the correlated premise.
121
CHAPTER 8
PREDICTIVE STATISTICAL ANALYSIS OF 9 SEGMENT
MEANDER RESISTOR MODELED USING ONLY 1 TEST
STRUCTURE
8.1 Introduction
The predictive statistical analysis of the 9-segment meander resistor detailed in
Chapter 7 will now be accomplished using only one test structure to deembed all the
building block equivalent circuit element values. The modeling methodology will follow
the procedure described in Chapter 4 except for the elimination of test structure one. The
methodology presented in this chapter will also demonstrate that the variation in the
complete equivalent circuit models, based only on the circuit building blocks, can be used
to predict such variations in actual fabricated devices. The 9-segment meander resistor
fabricated on a 96% Alumina substrate will be the device under inspection.
Two sets of resistor structures were fabricated on a 96% Alumina substrate: 32
three-segment resistors (test structure 2), and 32 nine-segment resistors (modeled device).
Scattering parameter measurements ranging in frequencies from 45MHz to 20GHz were
taken from all the structures. The measured s-parameter data from test structure 2 was
122
used to deembed the element values of the partial element equivalent circuit (PEEC) for
the 32 test structures, using a non-linear optimization algorithm in Hspice. This process
generated a unique equivalent circuit model for each of the 32 test structure devices.
Mean and relative variation was then calculated for each component from the 32 unique
equivalent circuit models, and used in a MonteCarlo analysis of the 9-segment modeled
device. This analysis generated 500 s-parameter curves of the modeled device electrical
behavior.
The s-parameter measurements from the MonteCarlo analysis and the actual
measured device were compared to determine if the measured data resides in the range of
curves generated by the MonteCarlo analysis. By inspection, it was noted that the
electrical behavior from the fabricated devices was in fact contained within the range
predicted by the MonteCarlo results. Therefore by means of comparison between the
MonteCarlo results and measured data, it was shown that the statistical variations of the
component values provide an accurate representation of the overall modeled device
performance.
8.2 Statistical Variations
The same three-step MonteCarlo technique, as in Chapter 7, was used to
accomplish the statistical prediction: 1.) A statistical model was formed to characterize
the fabrication process used; 2.) A circuit design was created using a Spice type
simulator and applying nominal device values for the target process; and 3.) Randomly
generated element values of the process model were simulated in a MonteCarlo tool to
123
produce a representative set of output performance characteristics. The impact of random
process variations can be inferred from these simulations, and electrical yield can then be
estimated using the percentage of that sample which meets the performance requirements
[29].
8.3 Test Structure Description
One set (test structures 2) of 32 test structures were fabricated by depositing gold
on a 96% Alumina substrate. The metal thickness used was 3000 angstroms with 30µ
widths and spacings for all connecting segments. All resistor structures were on one
layer, with connections to the devices made using a ground-signal-ground probe pad
pattern.
Test structure 2 consisted of two probe pads connected to five material squares
which are connected to nine coupled material squares, and two corners made up of five
material squares.
The layout and fabrication mask photos of test structure 2 and the modeled 9-
segment meander resistor are shown in Fig. 8.3-1.
124
Test Structure 2Test Structure 2
9-SegmentMeander Resistor
9-SegmentMeander Resistor
Figure #8.3-1: Test Structure 2 and 9-Segment Resistor Layout and Mask Photos
These test structure 2 allowed the modeling of probe pads, material square,
coupled material squares, and corner primitives.
8.3.1 Processing and Measurement
The 96% Alumina structure was physically designed using integrated circuit
design tools within the Avanti design environment. Rena Huang PhD. Candidate in the
School of Electrical and Computer Engineering at the Georgia Institute of Technology
fabricated the design in Georgia Tech’s MiRC cleanroom. The size of the completed
coupon was approximately 2.25" x 2.25". Metal lines were drawn 30µ wide. The
embedded structures were interfaced accessed using ground-signal-ground probe pads.
The complete test structure coupon is shown in Figure 8.3.1-1.
125
Figure #8.3.1-1: Test Structure Coupon
All the test structure two devices were measured using network analysis
techniques. Since very low loss metal was used in the manufacturing process, DC
resistance measurements were unreliable and not used. For high frequency
measurements, an HP 8510C network analyzer was used in conjunction with a Cascade
Microtech probe station and ground-signal-ground configuration probes. Calibration was
accomplished using a supplied substrate and the application of the line-reflect-match
(LRM) calibration method. Data was gathered for each of the test structures at over 400
frequency points between 45MHz and 20GHz and stored by means of data acquisition
software.
126
8.4 Modeling Scheme
8.4.1 9-Segment Meander Resistor Modeling Procedure
The improved modeling methodology using only one test structure, detailed in
Chapter 5, will be utilized to model the 9-segment meander resistor. The fundamental
idea behind this modeling procedure is that passive structures are comprised of several
key geometrical building blocks. The building block equivalent circuits are now derived
from only one fabricated test structure and measurements using optimization and
extraction routines. Passive RLC models for each embedded building block are extracted
and take into account effects of processing fluctuations and nonideal material properties.
The objective here is to predict the electrical behavior of arbitrary geometry passive
devices in a standard circuit simulator; therefore, providing a major speedup over
methods that do not utilize lumped elements.
The steps involved in the modeling procedure for the 9-segment meander resistor
are described in a brief outline below:
1. The 32 test structures and 32 9-segment meander resistors were physically designed
(96% Alumina) and fabricated. Rena Huang PhD. Candidate fabricated the designs in
the MiRC cleanroom at Georgia Tech. High frequency s-parameter measurements of
the devices were taken by on-wafer ground-signal-ground probing.
2. The measured data was used in a circuit optimization input file to deembed the
element values for the 32 structures. Initial guesses used in the Hspice Optimizer
127
were made based on s-parameter measurements for one of each structure. Once
optimization for one structure was completed, the results were used for the remaining
optimizations.
3. 32 unique equivalent circuit models of test structure two (the only test structure used)
were obtained.
Fig. 8.4.1-1 shows the building blocks, the equivalent circuits, and some of the
element values generated from the Hspice Optimizer.
Building Blocks / PEECsBuilding Blocks / PEECs
1 3
2 4
R =0 .09 Oh mL = 1 E -1 1 HC = 1.1 e -1 5 FC M =0 .4e - 1 5 FCC = 1 .4 e-1 5 FL M = 0 .2 0
LC
R L R
C C CC
LC
R L R
C C C C
L M
L MCM
1
2
3
4
1 2
R =0. 08 O hmL = 1 E- 11 HC = 2.7 e- 1 5 FC C = 1.2e - 15 F
L
C
R L R
C C C C
1 2
1 2
R =0 .08 O h mL = 1 E- 11 HC = 2.7 e- 1 5 FC C = 1.2e- 15 F
L
C
R L R
C C C C
1 2
.
R= 0 .4 O hmL= 3 .7 E- 11 HC = 5.3 e -1 5 FCC = 2 7e - 1 5 F
L
C
R L R
CC CC
1 2
CC CC
2L
C
R L R1
1
2
L
C
R L R
CC CC
1 2
L
C
R L R
CC CC
1 2
L
C
R L R
CC CC
1 2
Test Structure 2Test Structure 2Coupled Square PrimitiveCoupled Square Primitive U-Shaped Bend
PrimitivesU-Shaped Bend
Primitives
Pad PrimitivePad Primitive
Material Square PrimitiveMaterial Square Primitive
Figure #8.4.1-1: Building Blocks & Equivalent Circuits Using Only Test Structure 2
128
8.4.2 Statistical Model
Thirty-two test structures (test structure 2) were characterized. Mean and relative
variation in the deembedded circuit values of the building blocks was calculated and used
in the Hspice MonteCarlo simulator as limits to generate random circuit values for 500
simulations. Fig. 8.4.2-1 shows the deembedded element values from test structure 2 and
the relative variations used in the MonteCarlo simulation tool.
Test Structure 2Test Structure 2
Statistical Variations of Equivalent Circuit Element ValuesStatistical Variations of Equivalent Circuit Element Values
Simulation Sum of Squares c_sq csq lsq rsq c_pad rpad lpad cpadlow 0.558429 3.96E-14 3.04E-27 1.74E-11 2.11E-06 1.37E-12 3.213313 2.52E-11 2.1E-14high 63.4942 1.65E-13 2.12E-25 1.95E-11 3.87E-06 2.78E-12 5.772112 4.38E-11 2.66E-14median 1.53612 1.02E-13 6.21E-26 1.84E-11 2.47E-06 1.75E-12 4.452542 3.23E-11 2.46E-14ab var 0.977691 6.21E-14 5.91E-26 9.67E-13 3.59E-07 3.83E-13 1.23923 7.09E-12 3.66E-15rel var 0.636467854 0.610554 0.95101 0.052618 0.145617 0.218885 0.27832 0.219803 0.148648
Simulation Sum of Squares cou_1 c_cou l2 r2 c2 c_co2cr l2_2 r2_2 c2_2 c_co2cr_2 l2_3 r2_3 c2_3low 0.558429 0.167882 3.25E-24 3.24E-12 0.283159 8.62E-16 5.13E-12 2.83E-12 0.364671 5.2E-17 4.71E-12 5.43E-13 0.405914 6.73E-16high 63.4942 0.256087 1.19E-23 7.58E-12 1.292548 2.19E-15 1.28E-11 7.61E-12 0.904278 2.55E-15 1.06E-11 5.92E-12 0.965507 2.57E-15median 1.53612 0.211063 6.98E-24 4.52E-12 0.743332 1.36E-15 7.78E-12 4.83E-12 0.72981 1.48E-15 6.77E-12 4.4E-12 0.706911 1.36E-15ab var 0.977691 0.043182 3.73E-24 1.28E-12 0.460173 4.93E-16 2.65E-12 1.99E-12 0.365139 1.43E-15 2.06E-12 3.86E-12 0.300997 6.85E-16rel var 0.636467854 0.204592 0.534402 0.282134 0.619067 0.36382 0.340874 0.412802 0.50032 0.964868 0.30473 0.876576 0.425792 0.504412
Figure #8.4.2-1: Mean and Relative Variations for Deembedded Element Values
From Test Structure Two
8.4.3 Circuit Design
An equivalent circuit of the 9-segment meander resistor was constructed by
replacing each building block in the structure with its equivalent extracted circuit values.
Due to the lack of correlation between the building blocks, each element is independent
of any other element in the circuit. Therefore, to create an accurate circuit representation
all the elements in the design need to generate a random value for the defined number of
MonteCarlo simulations. The 9-segment meander resistor generated 1336 random circuit
values for each simulation. This circuit file was automatically generated using Unix shell
129
scripts detailed in Chapter 9. Fig. 8.4.3-1 is a block diagram of a 9-segment meander
resistor and shows the amount of random number generations per simulation, for the
PEEC element values.
MaterialSquare
R=0.08 OhmL= 1E-11 HC = 2.7e-15 FCC = 1.2e-15 F
LC
R L R
CC C C
1 2
Pad
R=0.08 OhmL=1E-11 HC = 2.7e-15 FCC = 1.2e-15 F
LC
R L R
CC C C
1 2
Coupled Pair
R=0.09 OhmL= 1E-11 HC = 1.1e-15 FCM=0.4e-15 FCC = 1.4e-15 FLM = 0.20
LC
R L R
CC CC
LC
R L R
CC CC
LMLM
CM
1
2
3
4
CC
LC
R1
Multi-Block
LC
R L R
CC CC1 2
LC
R L R
CC CC1 2
.
R=0.4 OhmL= 3.7E-11 HC = 5.3e-15 FCC=2 7e-15 F
CC2
L R
LC
R L R
CC CC1 2
LC
R L R
CC CC1 2
Pad
Pad
Pads = (2 BBlocks * 8 Elements) = 16 ElementsMaterial Squares = (8 BBlocks * 8 Elements) = 64 ElementsCoupled Squares = [((4 BBlocks * 9 Coupled) * 3) * 8 Elements] + (9 Coupled * 8 Elements) = 936 Elements
Corners = (5 BBlocks * 8 Corners * 8 Elements) = 320 Elements
TOTAL = 1336 Random Generated Element Values
Figure #8.4.3-1: Number of Random Number Generation for MonteCarlo
Simulation
8.4.4 MonteCarlo Analysis
Based on the relative variation calculated for each element value in the
equivalent circuit models, a MonteCarlo analysis was performed using Hspice. A
MonteCarlo analysis with a uniform parameter distribution simulated 500 sets of circuit
model parameters, as shown in Fig.8.4.4-1.
130
MonteCarlo Analysis Results Using OnlyOne Test Structure To Deembed
Building Block Equivalent Circuit Values
MonteCarlo Analysis Results Using OnlyOne Test Structure To Deembed
Building Block Equivalent Circuit Values
S21 RealS21 Real
S11 RealS11 Real
S21 ImaginaryS21 Imaginary
S11 ImaginaryS11 Imaginary
Figure #8.4.4-1: MonteCarlo Analysis Generated 500 Simulations
S-parameter data was generated for each of the 500 simulations, and extracted
from the output file using Unix scripts (detailed in Chapter 9). Due to a Microsoft Excel
limitation of 255 series plots per graph, all the data could not be rendered together.
Therefore, the MonteCarlo maximum and minimum were found for S11/S21
magnitude/phase and used as upper and lower limits (red) with the actual measured data
contained within the predicted range. This plot is shown in Fig. 8.4.4-2.
Following the simulations, the output response for the modeled device was
compared to actual measured data from the 32 fabricated 9-segment structures. It was
131
established that the electrical behavior from the 32 actual measured devices, was
predicted by the MonteCarlo analysis using the statistical variations of the equivalent
circuit element values deembedded from the test structure, as shown in Fig. 8.4.4-2.
MonteCarlo Predicted Range vs. Actual Measured 9-Segment Resistors MonteCarlo Predicted Range vs. Actual Measured 9-Segment Resistors
S11 Magnitude MonteCarlo Range (Red) vs. Actual Measured Devices
0.38
0.43
0.48
0.53
0.58
0.63
0.68
0.73
0.78
1.E+08 1.E+09 1.E+10 1.E+11
Frequency
S11Mag
S21 Magnitude MonteCarlo Range (Red) vs. Actual Measured Devices
0.36
0.41
0.46
0.51
0.56
0.61
1.E+08 1.E+09 1.E+10 1.E+11
Frequency
S21Mag
S21 Phase MonteCarlo Range (Red) vs. Actual Measured Devices
-90
-70
-50
-30
-10
10
30
50
70
90
1.E+08 1.E+09 1.E+10 1.E+11
Frequency
S21Phase
S11 Phase MonteCarlo Range (Red) vs. Actual Measured Devices
-90
-70
-50
-30
-10
10
30
50
70
90
1.E+08 1.E+09 1.E+10 1.E+11
Frequency
S11Phase
Figure #8.4.4-2: MonteCarlo Predicted vs. Actual Measured
A close-up of the MonteCarlo vs. measured S11 magnitude plot (Fig. 8.4.4-3)
shows the measured data stays within the MonteCarlo predicted range up to 1GHz. By
inspection, it appears that the parasitics begin to take effect at approximately 1GHz;
therefore, the deembedded model for the 9-segment resistor accurately predict the
measured electrical behavior range up to the first resonance for S11 magnitude.
A close-up of the MonteCarlo vs. measured S21 magnitude plot (Fig. 8.4.4-3)
shows the measure data stays within the MonteCarlo predicted range up to 2.7GHz.
132
Again by inspection, it appears that the parasitics begin to take effect at 1GHz and at
3GHz; therefore, the deembedded model accurately predicts the measured behavior range
for S21 magnitude past the first and up to the second resonance.
The MonteCarlo vs. measured phase plots were not as impressive as the
magnitude plots, because the measured data fell outside the MonteCarlo predicted range;
however, the trends were identical. Looking at the scale of +/- 90 degrees in Fig. 8.4.4-2,
it appears that the error between the MonteCarlo vs. measured is trivial compared to an
inductive 90 degrees and a capacitive –90 degrees of phase shift. The error actually
seems to be some DC offset.
S21 Magnitude MonteCarlo Range (Red) vs. Actual Measured Devices
0.46
0.48
0.5
0.52
0.54
0.56
0.58
0.6
0.62
1.E+08 1.E+09 1.E+10 1.E+11
Frequency
S21Mag
S11 Magnitude MonteCarlo Range (Red) vs. Actual Measured Devices
0.38
0.4
0.42
0.44
0.46
0.48
0.5
0.52
1.E+08 1.E+09 1.E+10 1.E+11
Frequency
S11Mag
MonteCarlo Predicted Range vs. Actual Measured 9-Segment ResistorsFor S11/S21 Magnitude
MonteCarlo Predicted Range vs. Actual Measured 9-Segment ResistorsFor S11/S21 Magnitude
1GHz1GHz
2.7GHz2.7GHz
Figure #8.4.4-3: MonteCarlo Predicted vs. Actual Measured
8.4.5 Statistical Analysis Methods
To accurately compare the MonteCarlo range vs. measured results, several
statistical methods were used to find a number that represented both sets of data at each
frequency point.
133
8.4.6 Mean and Standard Deviation
The mean is a measure of central tendency, “middleness”, or the arithmetic
average. Adding up all the scores and dividing by the number of scores is the method of
calculating the mean. The variance is the most frequently used measure of variability
[32]. The standard deviation, derived from the variance, also indicates the relative
variability within a group and is calculated by taking the square root of the variance.
To aid in the evaluation of measured vs. modeled data, percent differences in
mean and standard deviation will be used. Fig. 8.4.6-1 displays the Mean Value Percent
Difference results using two different sets of data. The first set compares the raw
MonteCarlo generated (population 500) data vs. the raw Measured (population 32) data.
The second set compares the normalized cumulative frequency distribution data
(population 16). Both sets are displaying the mean percent difference in magnitude of
S11 and S21, and are comparing data generated from a circuit model file constructed
using the correlated premise (Blue Plot), the independent premise (Yellow Plot), and the
independent premise using only 1 test structure (independent 1TS, Red Plot).
The results for all three types of premise begin to diverge at 1.84GHz for S11
magnitude and 4.5GHz for S21 magnitude, however; the results using the independent
1TS premise have a lower percent difference at lower frequencies (less than 1GHz) and a
higher percent difference beyond 1GHz for S11 magnitude. The independent 1TS
premise also has a lower percent difference at frequencies lower than 2.5GHz for S21
magnitude, and basically equal the other two types of premise at frequencies greater than
2.5GHz.
134
Mean Percent Error Comparing Cumulative Frequency Distribution Data For
MonteCarlo Prediction (16 Data Points) vs. Actual Measured Device (16 Data Points)
-1.00%
9.00%
19.00%
29.00%
39.00%
49.00%
1.E+08 1.E+09 1.E+10
Frequency
S21 Mag% Error
Independent Blocks
Correlated Blocks
Independent Blocks 1TS
Mean Percent ErrorComparing
MonteCarlo Prediction (500 Data Points) vs. Actual Measured Device (32 Data Points)
-1.00%
-0.50%
0.00%
0.50%
1.00%
1.50%
2.00%
1.E+08 1.E+09 1.E+10
Frequency
S21 Mag% Error
Independent BlocksCorrelated Blocks
Independent Blocks 1TS
Mean Percent ErrorComparing
MonteCarlo Prediction (500 Data Points) vs. Actual Measured Device (32 Data Points)
-0.10%
0.40%
0.90%
1.40%
1.90%
1.E+08 1.E+09 1.E+10
Frequency
S11 Mag% Error
Independent Blocks
Correlated Blocks
Independent Blocks 1TS
Mean Percent Error Comparing Cumulative Frequency Distribution Data For
MonteCarlo Prediction (16 Data Points) vs. Actual Measured Device (16 Data Points)
-5.00%
0.00%
5.00%
10.00%
15.00%
20.00%
25.00%
30.00%
35.00%
40.00%
1.E+08 1.E+09 1.E+10
Frequency
S11 Mag% Error
Independent Blocks
Correlated Blocks
Independent Blocks 1TS
Mean Percent Difference For S11/S21 Magnitude Using Data Generated Via Correlated Circuit Configuration vs. Independent Circuit Configuration
Mean Percent Difference For S11/S21 Magnitude Using Data Generated Via Correlated Circuit Configuration vs. Independent Circuit Configuration
1.84GHz1.84GHz 1.84GHz1.84GHz
4.5GHz4.5GHz 4.5GHz4.5GHz
Figure #8.4.6-1: Mean Value Percent Difference Between MonteCarlo and
Measured Data For S11/S21 Magnitude
Fig. 8.4.6-2 displays the Standard Deviation Percent Difference results using two
different sets of data. The first set compares the raw MonteCarlo generated (population
500) data vs. the raw Measured (population 32) data. The second set compares the
normalized cumulative frequency distribution data (population 16). Both sets are
displaying the standard deviation percent difference of the magnitude of S11 and S21,
and are comparing data generated from a circuit model file constructed using the
correlated premise (Blue Plot), the independent premise (Yellow Plot), and the
independent 1TS premise (Red Plot).
The results for all three types of premise begin to diverge at 2.5GHz for S11
magnitude using the first set of data, and at 1.74GHz for S11 magnitude using the second
135
set of data. The results begin to diverge at 4.2GHz for S21 magnitude using the first set
of data, and at 7.0GHz for S21 magnitude using the second set of data. However; the
results using the independent and independent 1TS types of premise have a much lower
percent difference at all frequencies than the correlated premise for S11 and S21
magnitude.
Standard Deviation Percent Error Comparing Cumulative Frequency Distribution Data For
MonteCarlo Prediction (16 Data Points) vs. Actual Measured Device (16 Data Points)
-1.00%
4.00%
9.00%
14.00%
19.00%
24.00%
29.00%
1.E+08 1.E+09 1.E+10
Frequency
S21 Mag% Error
Independent Blocks
Correlated Blocks
Independent Blocks 1TS
Standard Deviation Pecent ErrorComparing
MonteCarlo Prediction (500 Data Points) vs. Actual Measured Device (32 Data Points)
0.00%
20.00%
40.00%
60.00%
80.00%
100.00%
120.00%
1.E+08 1.E+09 1.E+10
Frequency
S21 Mag% Error
Independent Blocks
Correlated Blocks
Independent Blocks 1TS
Standard Deviation Pecent ErrorComparing
MonteCarlo Prediction (500 Data Points) vs. Actual Measured Device (32 Data Points)
0.00%
20.00%
40.00%
60.00%
80.00%
100.00%
120.00%
140.00%
160.00%
180.00%
1.E+08 1.E+09 1.E+10
Frequency
S11 Mag% Error
Independent Blocks
Correlated Blocks
Independent Blocks 1TS
Standard Deviation Percent Error Comparing Cumulative Frequency Distribution Data For
MonteCarlo Prediction (16 Data Points) vs. Actual Measured Device (16 Data Points)
-1.00%
4.00%
9.00%
14.00%
19.00%
24.00%
29.00%
1.E+08 1.E+09 1.E+10
Frequency
S11 Mag% Error
Independent Blocks
Correlated Blocks
Independent Blocks 1TS
Standard Deviation Percent Difference For S11/S21 MagnitudeUsing Data Generated Via Correlated Circuit Configuration vs. Independent Circuit Configuration
Standard Deviation Percent Difference For S11/S21 MagnitudeUsing Data Generated Via Correlated Circuit Configuration vs. Independent Circuit Configuration
2.5GHz2.5GHz1.74GHz1.74GHz
4.2GHz4.2GHz7.0GHz7.0GHz
Figure #8.4.6-2: Standard Deviation Percent Difference Between MonteCarlo and
Measured Data For S11/S21 Magnitude
8.4.7 Frequency Distribution Plots
The s-parameters generated from the MonteCarlo results and the actual measured
devices were transposed from rows representing a particular frequency to columns
136
representing MonteCarlo simulations and measured device data at a particular frequency
point. The range of values from high to low was split up into sixteen linearly place filter
bins. The frequency of occurrence was calculated and normalized to the number of data
points (500 for MonteCarlo, and 32 for actual measured). These normalized cumulative
frequency distribution (NCFD) plots were generated at every frequency point for
S11/S21 magnitude up to 7.5GHz. These plots illustrate that the measured data from
each fabricated device at a particular frequency fits, or doesn’t fit, within the MonteCarlo
predicted range. The trend lines generated within the NCFD plots also show how the
standard deviation between predicted and measured compare.
NCFD plots at several different frequencies for S11 and S21 magnitude are shown
in Fig. 8.4.7-1 through Fig. 8.4.7-7 and in Fig. 8.4.7-8 through Fig. 8.4.7-14 respectively.
Data generated via the correlated vs. the independent vs. the independent 1TS premise
are compared in these NCFD plots. By inspection, the standard deviation match between
the MonteCarlo predicted vs. measured for the independent premise clearly outperforms
the correlated and the independent 1TS premise at all frequencies.
137
Normalized Cumulative Frequency Distribution For S11 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.392
0
0.395
9
0.399
8
0.403
8
0.407
7
0.411
6
0.415
5
0.419
4
0.423
4
0.427
3
0.431
2
0.435
1
0.439
0
0.443
0
0.446
9
0.450
8
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
Normalized Cumulative Frequency Distribution For S11 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.396
6
0.399
8
0.403
0
0.406
2
0.409
4
0.412
6
0.415
8
0.419
0
0.422
2
0.425
4
0.428
6
0.431
8
0.434
9
0.438
1
0.441
3
0.444
5
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
Correlated PremiseCorrelated Premise Independent 1TS PremiseIndependent 1TS Premise
Independent PremiseIndependent Premise
Data Extracted at:600MHz
Data Extracted at:600MHz
Normalized Cumulative Frequency Distribution For S11 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.397
8
0.400
7
0.403
6
0.406
5
0.409
4
0.412
2
0.415
1
0.418
0
0.420
9
0.423
7
0.426
6
0.429
5
0.432
4
0.435
2
0.438
1
0.441
0
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
Figure #8.4.7-1: NCFD Plot Comparisons Between Correlated, Independent, and
Independent 1TS Premise For S11 at 600MHz
Correlated PremiseCorrelated Premise Independent 1TS PremiseIndependent 1TS Premise
Independent PremiseIndependent Premise
Normalized Cumulative Frequency Distribution For S11 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.394
4
0.398
2
0.402
1
0.405
9
0.409
7
0.413
6
0.417
4
0.421
3
0.425
1
0.428
9
0.432
8
0.436
6
0.440
5
0.444
3
0.448
1
0.452
0
Filter Bins
NCFD
MonteCarloMeasuredMonteCarlo TrendMeasured Trend
Normalized Cumulative Frequency Distribution For S11 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.398
8
0.402
0
0.405
1
0.408
2
0.411
4
0.414
5
0.417
7
0.420
8
0.423
9
0.427
1
0.430
2
0.433
4
0.436
5
0.439
6
0.442
8
0.445
9
Filter Bins
NCFD
MonteCarloMeasuredMonteCarlo TrendMeasured Trend
Data Extracted at:800MHz
Data Extracted at:800MHz
Normalized Cumulative Frequency Distribution For S11 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.399
9
0.402
8
0.405
6
0.408
4
0.411
2
0.414
1
0.416
9
0.419
7
0.422
5
0.425
4
0.428
2
0.431
0
0.433
9
0.436
7
0.439
5
0.442
3
Filter Bins
NCFD
MonteCarloMeasuredMonteCarlo TrendMeasured Trend
Figure #8.4.7-2: NCFD Plot Comparisons Between Correlated, Independent, and
Independent 1TS Premise For S11 at 800MHz
138
Correlated PremiseCorrelated Premise Independent 1TS PremiseIndependent 1TS Premise
Independent PremiseIndependent Premise
Normalized Cumulative Frequency Distribution For S11 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.396
6
0.400
4
0.404
1
0.407
9
0.411
7
0.415
4
0.419
2
0.423
0
0.426
7
0.430
5
0.434
3
0.438
0
0.441
8
0.445
6
0.449
3
0.453
1
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
Normalized Cumulative Frequency Distribution For S11 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.400
9
0.404
0
0.407
0
0.410
1
0.413
2
0.416
3
0.419
4
0.422
5
0.425
6
0.428
7
0.431
8
0.434
8
0.437
9
0.441
0
0.444
1
0.447
2
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
Data Extracted at:950MHz
Data Extracted at:950MHz
Normalized Cumulative Frequency Distribution For S11 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.401
9
0.404
7
0.407
4
0.410
2
0.413
0
0.415
8
0.418
6
0.421
3
0.424
1
0.426
9
0.429
7
0.432
5
0.435
2
0.438
0
0.440
8
0.443
6
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
Figure #8.4.7-3: NCFD Plot Comparisons Between Correlated, Independent, and
Independent 1TS Premise For S11 at 900MHz
Correlated PremiseCorrelated Premise Independent 1TS PremiseIndependent 1TS Premise
Independent PremiseIndependent Premise
Normalized Cumulative Frequency Distribution For S11 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.400
1
0.403
7
0.407
4
0.411
0
0.414
7
0.418
4
0.422
0
0.425
7
0.429
3
0.433
0
0.436
6
0.440
3
0.443
9
0.447
6
0.451
2
0.454
9
Filter Bins
NCFD
MonteCarloMeasured
MonteCarlo TrendMeasured Trend
Normalized Cumulative Frequency Distribution For S11 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.404
1
0.407
1
0.410
1
0.413
1
0.416
1
0.419
1
0.422
1
0.425
2
0.428
2
0.431
2
0.434
2
0.437
2
0.440
2
0.443
2
0.446
3
0.449
3
Filter Bins
NCFD
MonteCarlo
MeasuredMonteCarlo TrendMeasured Trend
Data Extracted at:1.15GHz
Data Extracted at:1.15GHz
Normalized Cumulative Frequency Distribution For S11 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.405
0
0.407
7
0.410
4
0.413
1
0.415
8
0.418
5
0.421
2
0.423
9
0.426
6
0.429
3
0.432
0
0.434
7
0.437
4
0.440
1
0.442
8
0.445
6
Filter Bins
NCFD
MonteCarlo
MeasuredMonteCarlo Trend
Measured Trend
Figure #8.4.7-4: NCFD Plot Comparisons Between Correlated, Independent, and
Independent 1TS Premise For S11 at 1.15GHz
139
Correlated PremiseCorrelated Premise Independent 1TS PremiseIndependent 1TS Premise
Independent PremiseIndependent Premise
Normalized Cumulative Frequency Distribution For S11 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.408
7
0.412
1
0.415
5
0.418
8
0.422
2
0.425
6
0.429
0
0.432
4
0.435
7
0.439
1
0.442
5
0.445
9
0.449
3
0.452
6
0.456
0
0.459
4
Filter Bins
NCFD
MonteCarlo
MeasuredMonteCarlo Trend
Measured Trend
Normalized Cumulative Frequency Distribution For S11 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.412
1
0.414
9
0.417
7
0.420
5
0.423
4
0.426
2
0.429
0
0.431
8
0.434
7
0.437
5
0.440
3
0.443
1
0.446
0
0.448
8
0.451
6
0.454
5
Filter Bins
NCFD
MonteCarloMeasuredMonteCarlo Trend
Measured Trend
Data Extracted at:1.5GHz
Data Extracted at:1.5GHz
Normalized Cumulative Frequency Distribution For S11 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.412
7
0.415
2
0.417
7
0.420
3
0.422
8
0.425
3
0.427
8
0.430
4
0.432
9
0.435
4
0.437
9
0.440
5
0.443
0
0.445
5
0.448
0
0.450
6
Filter Bins
NCFD
MonteCarlo
MeasuredMonteCarlo Trend
Measured Trend
Figure #8.4.7-5: NCFD Plot Comparisons Between Correlated, Independent, and
Independent 1TS Premise For S11 at 1.5GHz
Correlated PremiseCorrelated Premise Independent 1TS PremiseIndependent 1TS Premise
Independent PremiseIndependent Premise
Normalized Cumulative Frequency Distribution For S11 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.421
1
0.424
2
0.427
3
0.430
3
0.433
4
0.436
5
0.439
6
0.442
6
0.445
7
0.448
8
0.451
9
0.455
0
0.458
0
0.461
1
0.464
2
0.467
3
Filter Bins
NCFD
MonteCarloMeasuredMonteCarlo Trend
Measured Trend
Normalized Cumulative Frequency Distribution For S11 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.424
5
0.427
1
0.429
6
0.432
2
0.434
7
0.437
3
0.439
9
0.442
4
0.445
0
0.447
5
0.450
1
0.452
6
0.455
2
0.457
7
0.460
3
0.462
9
Filter Bins
NCFD
MonteCarloMeasuredMonteCarlo Trend
Measured Trend
Data Extracted at:2.0GHz
Data Extracted at:2.0GHz
Normalized Cumulative Frequency Distribution For S11 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.424
9
0.427
2
0.429
4
0.431
7
0.433
9
0.436
2
0.438
4
0.440
7
0.442
9
0.445
2
0.447
4
0.449
7
0.451
9
0.454
2
0.456
4
0.458
7
Filter Bins
NCFD
MonteCarlo
MeasuredMonteCarlo Trend
Measured Trend
Figure #8.4.7-6: NCFD Plot Comparisons Between Correlated, Independent, and
Independent 1TS Premise For S11 at 2.0GHz
140
Correlated PremiseCorrelated Premise Independent 1TS PremiseIndependent 1TS Premise
Independent PremiseIndependent Premise
Normalized Cumulative Frequency Distribution For S11 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.434
4
0.437
2
0.440
0
0.442
7
0.445
5
0.448
3
0.451
0
0.453
8
0.456
5
0.459
3
0.462
1
0.464
8
0.467
6
0.470
4
0.473
1
0.475
9
Filter Bins
NCFD
MonteCarloMeasuredMonteCarlo TrendMeasured Trend
Normalized Cumulative Frequency Distribution For S11 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.437
8
0.440
1
0.442
4
0.444
7
0.446
9
0.449
2
0.451
5
0.453
8
0.456
1
0.458
3
0.460
6
0.462
9
0.465
2
0.467
5
0.469
7
0.472
0
Filter Bins
NCFD
MonteCarloMeasuredMonteCarlo TrendMeasured Trend
Data Extracted at:2.5GHz
Data Extracted at:2.5GHz
Normalized Cumulative Frequency Distribution For S11 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.437
9
0.439
9
0.441
9
0.443
8
0.445
8
0.447
8
0.449
8
0.451
7
0.453
7
0.455
7
0.457
6
0.459
6
0.461
6
0.463
6
0.465
5
0.467
5
Filter Bins
NCFD
MonteCarlo
MeasuredMonteCarlo Trend
Measured Trend
Figure #8.4.7-7: NCFD Plot Comparisons Between Correlated, Independent, and
Independent 1TS Premise For S11 at 2.5GHz
Correlated PremiseCorrelated Premise Independent 1TS PremiseIndependent 1TS Premise
Independent PremiseIndependent Premise
Data Extracted at:800MHz
Data Extracted at:800MHz
Normalized Cumulative Frequency Distribution For S21 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.565
0
0.567
9
0.570
8
0.573
8
0.576
7
0.579
6
0.582
5
0.585
4
0.588
3
0.591
2
0.594
1
0.597
1
0.600
0
0.602
9
0.605
8
0.608
7
Filter Bins
NCFD
MonteCarlo
MeasuredMonteCarlo Trend
Measured Trend
Normalized Cumulative Frequency Distribution For S21 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.558
0
0.561
9
0.565
9
0.569
9
0.573
8
0.577
8
0.581
8
0.585
7
0.589
7
0.593
7
0.597
6
0.601
6
0.605
6
0.609
6
0.613
5
0.617
5
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
Normalized Cumulative Frequency Distribution For S21 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.562
8
0.566
0
0.569
3
0.572
5
0.575
7
0.579
0
0.582
2
0.585
4
0.588
7
0.591
9
0.595
1
0.598
4
0.601
6
0.604
9
0.608
1
0.611
3
Filter Bins
NCFD
MonteCarloMeasured
MonteCarlo TrendMeasured Trend
Figure #8.4.7-8: NCFD Plot Comparisons Between Correlated, Independent, and
Independent 1TS Premise For S21 at 800MHz
141
Correlated PremiseCorrelated Premise Independent 1TS PremiseIndependent 1TS Premise
Independent PremiseIndependent Premise
Data Extracted at:1.0GHz
Data Extracted at:1.0GHz
Normalized Cumulative Frequency Distribution For S21 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.564
5
0.567
4
0.570
3
0.573
2
0.576
1
0.579
0
0.581
9
0.584
8
0.587
7
0.590
6
0.593
4
0.596
3
0.599
2
0.602
1
0.605
0
0.607
9
Filter Bins
NCFD
MonteCarlo
MeasuredMonteCarlo Trend
Measured Trend
Normalized Cumulative Frequency Distribution For S21 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.557
5
0.561
5
0.565
4
0.569
3
0.573
3
0.577
2
0.581
1
0.585
1
0.589
0
0.593
0
0.596
9
0.600
8
0.604
8
0.608
7
0.612
6
0.616
6
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
Normalized Cumulative Frequency Distribution For S21 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.562
3
0.565
5
0.568
7
0.571
9
0.575
1
0.578
3
0.581
6
0.584
8
0.588
0
0.591
2
0.594
4
0.597
6
0.600
8
0.604
1
0.607
3
0.610
5
Filter Bins
NCFD
MonteCarlo
MeasuredMonteCarlo Trend
Measured Trend
Figure #8.4.7-9: NCFD Plot Comparisons Between Correlated, Independent, and
Independent 1TS Premise For S21 at 1.0GHz
Correlated PremiseCorrelated Premise Independent 1TS PremiseIndependent 1TS Premise
Independent PremiseIndependent Premise
Data Extracted at:2.0GHz
Data Extracted at:2.0GHz
Normalized Cumulative Frequency Distribution For S21 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.559
9
0.562
6
0.565
4
0.568
1
0.570
9
0.573
6
0.576
4
0.579
2
0.581
9
0.584
7
0.587
4
0.590
2
0.592
9
0.595
7
0.598
4
0.601
2
Filter Bins
NCFD
MonteCarlo
MeasuredMonteCarlo Trend
Measured Trend
Normalized Cumulative Frequency Distribution For S21 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.553
6
0.557
2
0.560
9
0.564
6
0.568
2
0.571
9
0.575
6
0.579
3
0.582
9
0.586
6
0.590
3
0.593
9
0.597
6
0.601
3
0.604
9
0.608
6
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
Normalized Cumulative Frequency Distribution For S21 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.557
5
0.560
5
0.563
6
0.566
6
0.569
7
0.572
8
0.575
8
0.578
9
0.581
9
0.585
0
0.588
0
0.591
1
0.594
1
0.597
2
0.600
3
0.603
3
Filter Bins
NCFD
MonteCarlo
MeasuredMonteCarlo Trend
Measured Trend
Figure #8.4.7-10: NCFD Plot Comparisons Between Correlated, Independent, and
Independent 1TS Premise For S21 at 2.0GHz
142
Correlated PremiseCorrelated Premise Independent 1TS PremiseIndependent 1TS Premise
Independent PremiseIndependent Premise
Data Extracted at:3.0GHz
Data Extracted at:3.0GHz
Normalized Cumulative Frequency Distribution For S21 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.552
9
0.555
4
0.558
0
0.560
5
0.563
1
0.565
6
0.568
1
0.570
7
0.573
2
0.575
8
0.578
3
0.580
9
0.583
4
0.586
0
0.588
5
0.591
1
Filter Bins
NCFD
MonteCarlo
MeasuredMonteCarlo Trend
Measured Trend
Normalized Cumulative Frequency Distribution For S21 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.547
1
0.550
5
0.553
9
0.557
4
0.560
8
0.564
2
0.567
6
0.571
0
0.574
4
0.577
8
0.581
3
0.584
7
0.588
1
0.591
5
0.594
9
0.598
3
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
Normalized Cumulative Frequency Distribution For S21 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.550
2
0.553
1
0.556
0
0.558
8
0.561
7
0.564
6
0.567
4
0.570
3
0.573
2
0.576
0
0.578
9
0.581
8
0.584
6
0.587
5
0.590
4
0.593
3
Filter Bins
NCFD
MonteCarlo
MeasuredMonteCarlo Trend
Measured Trend
Figure #8.4.7-11: NCFD Plot Comparisons Between Correlated, Independent, and
Independent 1TS Premise For S21 at 3.0GHz
Correlated PremiseCorrelated Premise Independent 1TS PremiseIndependent 1TS Premise
Independent PremiseIndependent Premise
Data Extracted at:4.0GHz
Data Extracted at:4.0GHz
Normalized Cumulative Frequency Distribution For S21 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.543
8
0.546
0
0.548
3
0.550
6
0.552
9
0.555
2
0.557
5
0.559
8
0.562
1
0.564
3
0.566
6
0.568
9
0.571
2
0.573
5
0.575
8
0.578
1
Filter Bins
NCFD
MonteCarlo
MeasuredMonteCarlo Trend
Measured Trend
Normalized Cumulative Frequency Distribution For S21 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.538
6
0.541
7
0.544
8
0.547
9
0.551
1
0.554
2
0.557
3
0.560
4
0.563
6
0.566
7
0.569
8
0.572
9
0.576
0
0.579
2
0.582
3
0.585
4
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo TrendMeasured Trend
Normalized Cumulative Frequency Distribution For S21 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.540
8
0.543
5
0.546
1
0.548
7
0.551
3
0.553
9
0.556
6
0.559
2
0.561
8
0.564
4
0.567
0
0.569
7
0.572
3
0.574
9
0.577
5
0.580
2
Filter Bins
NCFD
MonteCarlo
MeasuredMonteCarlo Trend
Measured Trend
Figure #8.4.7-12: NCFD Plot Comparisons Between Correlated, Independent, and
Independent 1TS Premise For S21 at 4.0GHz
143
Correlated PremiseCorrelated Premise Independent 1TS PremiseIndependent 1TS Premise
Independent PremiseIndependent Premise
Data Extracted at:5.0GHz
Data Extracted at:5.0GHz
Normalized Cumulative Frequency Distribution For S21 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.533
1
0.535
7
0.538
3
0.540
9
0.543
5
0.546
1
0.548
7
0.551
3
0.553
9
0.556
5
0.559
1
0.561
7
0.564
3
0.566
9
0.569
5
0.572
1
Filter Bins
NCFD
MonteCarlo
MeasuredMonteCarlo TrendMeasured Trend
Normalized Cumulative Frequency Distribution For S21 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.528
3
0.531
1
0.533
9
0.536
7
0.539
5
0.542
3
0.545
0
0.547
8
0.550
6
0.553
4
0.556
2
0.559
0
0.561
8
0.564
6
0.567
3
0.570
1
Filter Bins
NCFD
MonteCarlo
Measured
MonteCarlo Trend
Measured Trend
Normalized Cumulative Frequency Distribution For S21 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.529
8
0.532
9
0.535
9
0.539
0
0.542
0
0.545
1
0.548
1
0.551
1
0.554
2
0.557
2
0.560
3
0.563
3
0.566
4
0.569
4
0.572
4
0.575
5
Filter Bins
NCFD
MonteCarlo
MeasuredMonteCarlo Trend
Measured Trend
Figure #8.4.7-13: NCFD Plot Comparisons Between Correlated, Independent, and
Independent 1TS Premise For S21 at 5.0GHz
Correlated PremiseCorrelated Premise Independent 1TS PremiseIndependent 1TS Premise
Independent PremiseIndependent Premise
Data Extracted at:6.0GHz
Data Extracted at:6.0GHz
Normalized Cumulative Frequency Distribution For S21 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.521
4
0.524
3
0.527
1
0.530
0
0.532
9
0.535
8
0.538
6
0.541
5
0.544
4
0.547
2
0.550
1
0.553
0
0.555
9
0.558
7
0.561
6
0.564
5
Filter Bins
NCFD
MonteCarloMeasuredMonteCarlo Trend
Measured Trend
Normalized Cumulative Frequency Distribution For S21 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.517
0
0.519
9
0.522
8
0.525
7
0.528
6
0.531
5
0.534
4
0.537
3
0.540
2
0.543
1
0.546
0
0.548
9
0.551
8
0.554
7
0.557
6
0.560
5
Filter Bins
NCFD
MonteCarlo
MeasuredMonteCarlo Trend
Measured Trend
Normalized Cumulative Frequency Distribution For S21 Magnitude MonteCarlo Prediction vs. Actual Measured Device at 1GHz
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0.517
8
0.521
2
0.524
6
0.528
0
0.531
5
0.534
9
0.538
3
0.541
7
0.545
1
0.548
5
0.552
0
0.555
4
0.558
8
0.562
2
0.565
6
0.569
0
Filter Bins
NCFD
MonteCarlo
MeasuredMonteCarlo TrendMeasured Trend
Figure #8.4.7-14: NCFD Plot Comparisons Between Correlated, Independent, and
Independent 1TS Premise For S21 at 6.0GHz
144
8.4.8 Pearson Product-Moment Correlation Coefficient
Using the normalized cumulative distribution information, a measure of linear
relationship between the MonteCarlo predicted and actual measured data was calculated
using the correlation coefficient. A perfect correlation between two variables results in r
= 1.0, no correlation at all results in r = 0.0, and any number in-between means that there
is some correlation between the variables however the relationship is not perfect.
Fig. 8.4.8-1 displays the Correlation Coefficient results using the normalized
cumulative frequency distribution data (population 16) generated from the magnitude of
S11 and S2. The Correlation Coefficients are compared to data generated via the
correlated premise (Blue Plot) vs. the independent premise (Yellow Plot) vs. the
independent 1TS premise (Red Plot).
The results for all three types of premise begin to diverge at 2.0GHz for S11
magnitude and 4.5GHz for S21 magnitude. The results using the independent and
independent 1TS types of premise have a higher Correlation Coefficient than the results
using the correlated premise at all frequencies below 5GHz.
145
Corralation CofficientComparing Cumulative Frequency Distribution Data For
MonteCarlo Prediction (16 Data Points) vs. Actual Measured Device (16 Data Points)
0.94
0.95
0.96
0.97
0.98
0.99
1.00
1.01
1.E+08 1.E+09 1.E+10
Frequency
S21 MagCorrelation
Coff.
Independent Blocks
Correlated Blocks
Independent Blocks 1TS
Corralation CofficientComparing Cumulative Frequency Distribution Data For
MonteCarlo Prediction (16 Data Points) vs. Actual Measured Device (16 Data Points)
0.91
0.92
0.93
0.94
0.95
0.96
0.97
0.98
0.99
1.00
1.E+08 1.E+09 1.E+10
Frequency
S11 MagCorrelation
Coff.
Independent Blocks
Correlated Blocks
Independent Blocks 1TS
Correlation Coefficient For S11/S21 MagnitudeUsing Data Generated Via Correlated Circuit Configuration vs. Independent Circuit Configuration
Correlation Coefficient For S11/S21 MagnitudeUsing Data Generated Via Correlated Circuit Configuration vs. Independent Circuit Configuration
2.0GHz2.0GHz4.5GHz4.5GHz
Figure #8.4.8-1: Pearson Product-Moment Correlation Coefficient for MonteCarlo
Predicted and Actual Measured Data
8.4.9 T-Test
The primary purpose of the “t” test is to determine whether the means of two
groups of numbers differ to a statistically significant degree. The larger the absolute
value of t, the more likely it is to reflect a significant difference between the two groups
under comparison.
Fig. 8.4.9-1 displays the t-test results using two different sets of data. The first set
compares the raw MonteCarlo generated (population 500) data vs. the raw Measured
(population 32) data. The second set compares the normalized cumulative frequency
distribution data (population 16). Both sets are displaying the t-test values using the
magnitude of S11 and S21, and are comparing data generated via the correlated premise
(Blue Plot) vs. the independent premise (Yellow Plot) vs. the independent 1TS premise
(Red Plot).
146
The results for all three types of premise begin to diverge at 1.8GHz for S11
magnitude and 4.5GHz for S21 magnitude, however; the results using the independent
1TS premise has a lower t-test value at frequencies less than 1.0GHz for S11, and less
than 2.5GHz for S21.
T-Test For Independent SamplesComparing Cumulative Frequency Distribution Data For
MonteCarlo Prediction (16 Data Points) vs. Actual Measured Device (16 Data Points)
-0.50
0.00
0.50
1.00
1.50
2.00
2.50
1.E+08 1.E+09 1.E+10
Frequency
S21 MagT Value
Independent Blocks
Correlated Blocks
Independent Blocks 1TS
T-Test For Independent SamplesComparing
MonteCarlo Prediction (500 Data Points) vs. Actual Measured Device (32 Data Points)
-0.50
1.50
3.50
5.50
7.50
9.50
1.E+08 1.E+09 1.E+10
Frequency
S21 MagT Value
Independent Blocks
Correlated Blocks
Independent Blocks 1TS
T-Test For Independent SamplesComparing Cumulative Frequency Distribution Data For
MonteCarlo Prediction (16 Data Points) vs. Actual Measured Device (16 Data Points)
-0.25
0.25
0.75
1.25
1.75
2.25
1.E+08 1.E+09 1.E+10
Frequency
S11 MagT Value
Independent Blocks
Correlated Blocks
Independent Blocks 1TS
T-Test For Independent SamplesComparing
MonteCarlo Prediction (500 Data Points) vs. Actual Measured Device (32 Data Points)
-0.50
1.50
3.50
5.50
7.50
9.50
1.E+08 1.E+09 1.E+10
Frequency
S11 MagT Value
Independent Blocks
Correlated Blocks
Independent Blocks 1TS
T-Test For S11/S21 Magnitude Using Data Generated Via Correlated Circuit Configuration vs. Independent Circuit Configuration
T-Test For S11/S21 Magnitude Using Data Generated Via Correlated Circuit Configuration vs. Independent Circuit Configuration
1.8GHz1.8GHz1.8GHz1.8GHz
4.5GHz4.5GHz
4.5GHz4.5GHz
Figure #8.4.9-1: T-Test for MonteCarlo Predicted and Actual Measured Data
8.4.10 ANOVA
The analysis of variance (ANOVA) is a technique used to test for statistical
significance of the differences among the means of two or more groups. The test looks at
the amount of variability (the differences) between the means of the groups, compared
with the amount of variability among the individual scores in each group; that is, the
147
variance between groups versus the variance within groups [32]. The ANOVA starts
with the total amount of variability in the data and divides it up (partitioning) into various
categories. Basically, it compares the variability among the group means with the
variability that occurred just by chance or error.
Fig. 8.4.10-1 displays the ANOVA results using two different sets of data. The
first set compares the raw MonteCarlo generated (population 500) data vs. the raw
Measured (population 32) data. The second set compares the normalized cumulative
frequency distribution data (population 16). Both sets are displaying the ANOVA values
using the magnitude of S11 and S21, and are comparing data generated via the correlated
premise (Blue Plot) vs. the independent premise (Yellow Plot) vs. the independent 1TS
premise (Red Plot).
The results for all three types of premise begin to diverge at 1.84GHz for S11
magnitude and 4.5GHz for S21 magnitude, however; the results using the independent
1TS premise have a lower ANOVA value at frequencies less than 1.0GHz for S11, and
less than 2.5GHz for S21.
148
ANOVA AnalysisComparing Cumulative Frequency Distribution Data For
MonteCarlo Prediction (16 Data Points) vs. Actual Measured Device (16 Data Points)
-0.25
-0.05
0.15
0.35
0.55
0.75
0.95
1.15
1.35
1.E+08 1.E+09 1.E+10
Frequency
S21 MagF(obt)
Independent Blocks
Correlated Blocks
Independent Blocks 1TS
ANOVA AnalysisComparing
MonteCarlo Prediction (500 Data Points) vs. Actual Measured Device (32 Data Points)
-1.00
4.00
9.00
14.00
19.00
24.00
29.00
1.E+08 1.E+09 1.E+10
Frequency
S21 MagF(obt)
Independent Blocks
Correlated BlocksIndependent Blocks 1TS
ANOVA AnalysisComparing Cumulative Frequency Distribution Data For
MonteCarlo Prediction (16 Data Points) vs. Actual Measured Device (16 Data Points)
-0.02
0.18
0.38
0.58
0.78
0.98
1.E+08 1.E+09 1.E+10
Frequency
S11 MagF(obt)
Independent Blocks
Correlated Blocks
Independent Blocks 1TS
ANOVA AnalysisComparing
MonteCarlo Prediction (500 Data Points) vs. Actual Measured Device (32 Data Points)
-0.25
4.75
9.75
14.75
19.75
24.75
29.75
1.E+08 1.E+09 1.E+10
Frequency
S11 MagF(obt)
Independent BlocksCorrelated Blocks
Independent Blocks 1TS
ANOVA For S11/S21 MagnitudeUsing Data Generated Via Correlated Circuit Configuration vs. Independent Circuit Configuration
ANOVA For S11/S21 MagnitudeUsing Data Generated Via Correlated Circuit Configuration vs. Independent Circuit Configuration
1.84GHz1.84GHz
1.84GHz1.84GHz
4.8GHz4.8GHz 4.8GHz4.8GHz
Figure #8.4.10-1: ANOVA for MonteCarlo Predicted and Actual Measured Data
8.5 Summary
This chapter explained in detail the application of the predictive statistical
analysis methodology, described in Chapter 6, to a 9-segment meander resistor.
Modeling the 9-segment meander resistor using only one test structure, successfully
proved the method of predicting passive device performance variations and yield, when
provided with a statistical distribution of the building block models and a circuit
schematic.
149
Using the improved method from Chapter 5 and the independent premise when
collating the circuit model file, results from the 9-segment meander resistor analysis
(using only one test structure to deembed and model the device) were compared to the
analysis of the same 9-segment meander resistor; using the modeling methodology in
Chapter 4; and, using both the correlated and independent premise when collating the
circuit model file.
The results generated using the raw data vs. the normalized filtered data for all the
statistical analysis techniques were found to mirror the behavior of one another, which
was expected. The Mean Percent Difference, the t-test, and the ANOVA, all rendered the
independent 1TS premise superior for S11 and S21 magnitude up to 1GHz and 2.5GHz
respectively. The Standard Deviation Percent Difference displayed a similar behavior for
the independent and independent 1TS types of premise for S11 and S21 magnitude up to
the same frequencies respectively. The NCFD plots best exemplified the contrast in the
data generated by all three types of premise, with the independent premise outperforming
the independent 1TS premise, and correlated premise.
150
CHAPTER 9
RESEARCH AUTOMATION USING UNIX SHELL SCRIPTS
9.1 Introduction
This research further develops a passive device modeling methodology that
accurately predicts the high frequency electrical behavior, and the statistical variation of
embedded passive devices on a generic substrate. By utilizing a small number of test
structures, to deembed the lumped element circuit values of canonical building blocks,
accurate statistical models of new passive devices can be created using this methodology.
In summary, a nonlinear optimizer is used to find the optimal building block
models, by fitting extensive high frequency measurements of the test structures behavior.
This process is repeated on numerous fabricated test structures producing a relative
variation in the lumped element circuit values of the building block models. A predicted
range of operation for the actual modeled device is then generated using a MonteCarlo
simulation tool, and the relative variations of the building block models.
To begin this research, two sets of thirty-two test structures are fabricated by
depositing gold on a 96 % alumina substrate, and s-parameter measurements taken. The
Hspice Optimization tool uses these measurements to deembed the lumped element
circuit values of the building block models, which are contained within the fabricated test
151
structures. Relative variations of the deembedded element values are calculated and
added to the modeled device circuit file, for simulation in the Hspice MonteCarlo tool.
By means of comparison, between the range predicted by the Monte Carlo simulations
and the measured data, the statistical variations of the component values were shown to
provided an accurate representation of the overall passive device performance.
The goal of this research is stated above, however investigation into the accuracy
of “measured vs. modeled” needed to be accomplish using the premise; the more accurate
the measured vs. modeled test structure results, the more accurate the modeled device,
and consequently the tighter and more realistic the predicted range of operation. This
investigation encompassed test structure circuit file examination along with statistical
analysis comparisons. The aforementioned generated a huge amount of data that needed
to be synthesized; therefore, Unix scripts were developed to act as interface between user
and computer to automate the entire analysis process. The scope of the automation
includes; 1.) Hspice Optimization simulations; 2.) Extracting and calculating percent
difference in measured vs. modeled s-parameters; 3.) Statistical analysis calculations
such as the relative variations of the deembedded element values; 4.) Generating the
device model circuit file; 5.) Hspice MonteCarlo analysis; 6.) Extracting and converting
s-parameters to polar coordinates; and 7.) Saving and localizing all circuit and
optimization files. A block diagram of the basic components of the automation flow is
shown in Fig. 9.1-1.
152
HspiceHspiceMonteCarloMonteCarlo
AnalysisAnalysis
HspiceHspiceOptimization Optimization
AnalysisAnalysis
Output Data Output Data Synthesis And Synthesis And
StorageStorage
Plotting Plotting FormatFormat
Analysis Analysis InputInput
Statistical Analysis Automation FlowStatistical Analysis Automation Flow
Figure #9.1-1: Main Components of Automation Flow
The automation detailed in this chapter allowed this research to successfully
complete in an elegant timely fashion. The statistical prediction of a passive device using
this methodology was experimentally verified for a 9-segment meander resistor, by
demonstrating that the variation in the complete equivalent circuit models, based only on
the circuit building blocks, can be used to predict such variations in the actual fabricated
device.
This chapter will act as a tutorial by explaining the functionality of each block,
and how to successfully complete an analysis. Messages detailing the progression of the
automated flow have been designed into the analysis scripts, and are visible in the Unix
shell. The Analysis Input shown in Fig. 9.1-2 will be the first block presented.
153
HspiceHspice MonteCarloMonteCarloAnalysisAnalysis
HspiceHspice Optimization Optimization AnalysisAnalysis
Output Data Synthesis Output Data Synthesis And StorageAnd Storage
Plotting FormatPlotting Format
Analysis Analysis InputInput
Input DataUsed in Statistical Automation Flow
Input DataUsed in Statistical Automation Flow
Figure #9.1-2: Analysis Input Block
9.2 Statistical Analysis Flow Input
The first step in using the analysis flow is to place the measured device data in a
predefined location. The s-parameter measurements of test structure one, test structure
two, and the 9-segment meander resistors were placed in the r1_txt, r2_txt, and r9_txt,
directories respectively. The naming convention for the 32 measured device files are as
follows; R1_<number>.txt, R2_<number>.txt, and R9_<number>.txt.
To start the analysis open a Unix shell and type in "sh hspice_runs". The message
“Hspice Optimization Analysis” will appear in the Unix shell and the window shown in
Fig. 9.2-1 will emerge. As stated previously, an investigation into the measured vs.
modeled accuracy was required which included test structure circuit file examination,
154
along with statistical analysis comparisons. The window in Fig. 9.2-1 asks the user to
choose which test structure circuit file configuration to analyze.
Circuit Configuration AnalysisCircuit Configuration Analysis
Figure #9.2-1: Choose Circuit Topology
After clicking on a choice, the user will then press “enter” the message “Pick Test
Structure” will appear in the Unix shell, and the window shown in Fig. 9.2-2 will emerge.
The deembedding procedure dictates which test structure will be analyzed first,
depending on a one (Chapter 5) or two (Chapter 4) test structure analysis.
155
Test Structure AnalysisTest Structure Analysis
Figure #9.2-2: Number of Test Structures Characterized
After clicking on a choice, the user will press “enter” the message “Insert Number
of Measured Test Structures” will appear in the Unix shell, and the window shown in Fig.
9.2-3 will emerge. The user will enter the number of measured test structure (1 and 2)
that have been placed in the r1_txt and r2_txt directories.
Measured Test StructuresMeasured Test Structures
Figure #9.2-3: Number of Measured Devices
The user will then press “enter” the message “Insert Initial Guesses” will appear
in the Unix shell, and the window shown in Fig. 9.2-4 will emerge. The user will insert
on the appropriate line, the initial guess values for each deembedded element value in test
156
structure one and test structure two. These values will be initially used to set up the first
extraction of optimized values from the output file.
Element Value Initial GuessElement Value Initial Guess
Figure #9.2-4: Initial Guess Values
The user will then press “enter” the message “Removing Previous .out File” will
appear in the Unix shell. This message tells the user that the optimization analysis has
begun and any previously generated .out file will be removed to avoid a conflict within
157
Hspice. The next block presented will be the “Hspice Optimization Analysis” block, and
is shown in Fig. 9.2-5.
HspiceHspice MonteCarloMonteCarloAnalysisAnalysis
HspiceHspiceOptimization Optimization
AnalysisAnalysis
Plotting FormatPlotting Format
Analysis InputAnalysis Input
Hspice Optimization AnalysisUsing Automation Flow
Hspice Optimization AnalysisUsing Automation Flow
Output Data Synthesis Output Data Synthesis And StorageAnd Storage
Figure #9.2-5: Hspice Optimization Analysis Block
9.3 Hspice Optimization Analysis
The modeling and statistical analysis of the 9-segment meander resistor
has begun. The building block characterization portion of the analysis will execute 1152
individual optimizations to arrive at 32 unique equivalent circuit models of test structure
one, and 32 unique equivalent circuit models of test structure two. The automated flow
will then calculate the relative variation in each element value, construct and collate the
.cir file, and run a MonteCarlo analysis consisting of 500 simulations. The output data
will be synthesized, converted to polar coordinates, compared, and formatted for
Microsoft Excel rendering.
158
This section will demonstrate an actual analysis while detailing the messages
generated in the Unix shell as the analysis proceeds. The first message “Updating
Resistor File Number To 1” means that the first set of s-parameters from the r1_txt
directory (R1_1.txt) will be used as input for the optimization of test structure one.
The next message “Generating .out File Using Initial Guesses” means that Hspice
is using the initial guess values to deembed the first set of optimized element values using
a very low-resolution setting, a very wide range of possible values, and the “steepest
decent” curve fit algorithm.
At this point, a summary of the Hspice optimization tool and an explanation of the
Gauss-Newton Levernber-Marguardt algorithm, used to perform the optimizations, need
to be conveyed.
9.3.1 Hspice Optimization Analysis Tool
The Hspice Optimization Tool uses nonlinear estimation to fit element parameters
of a circuit file to measured s-parameter data. Nonlinear estimation, or nonlinear
optimization, is the process of fitting a mathematical model to experimental data to
determine unknown parameters in the model. The optimization process is often nonlinear
because the observed data does not vary in direct proportion to the parameters in
question; whereas, parameters are chosen so that the output of the model is the best match
to the observed data. Having found accurate values of the estimated parameters,
designers can then make accurate predictions of electrical behavior from the underlying
model. These results are an integral part of circuit design, process refinement, and many
other aspects of the design process.
159
Nonlinear optimization is an ever-present tool of modern technology, and the
Gauss-Newton Levenberg-Marquardt algorithm is the breakthrough that accounts for
much of its commonplace use [40].
9.3.2 Gauss-Newton Levenberg-Marquardt Algorithm
The Gauss-Newton Levenberg-Marquardt algorithm (GNLM), which combines
the robustness of a steepest descent method with the efficiency of a Gauss-Newton step,
is used to accomplish the nonlinear estimation in the Hspice Optimization Tool [31]. The
GNLM method switches continuously from a gradient method (large lambda) far from
the minimum, to a Gauss-Newton step as the minimum is approached and lambda is
reduced. Basically this is a compromise between steepest descent (very slow
convergence, but converges from anywhere) and the Gauss-Newton method (very fast
convergence, but converges only close to the optimum). To further understand the
GNLM method, a closer look at Gauss-Newton component of the algorithm is necessary.
Consider that pbar is the parameter vector and the components p1 p2 pn.. Tneed to
be estimated. The Gauss-Newton optimization method is specifically designed for
minimizing the objective function, which has the form of “the sum of squares” functions,
such as the one given in equation (1) [40].
1
M
m
Wm2 H
mc pbar Hmobs pbar
2.
=
(1)
or
160
E pbar
1
M
m
fm pbar2
=
(2)
where
fm pbar Wm Hmc pbar H
mobs pbar. (3)
where Hmc pbar is the calculated head corresponding to measurement H
mobs pbar ,and Wm
is the weighting coefficient of measurement m. The first order derivatives of E pbar can
be obtained from (2)
pi
Edd
2
1
M
m
fm pi
fmdd.
=
. i 1 N..( ) (4)
where N is the number of parameters to be estimated. The second derivatives are
d2 Edpi dpj
2
1
M
m
dfm
dpi
dfm
dpj
fm
d2 fm
dpi dpj
.
=
. i 1 N..( ) ; j 1 N..( ) (5)
In equation (5) fm pbar is residual, when pbar is not too far from the optimum
value, it can be assumed that the value of fm pbar is small and the second order terms of
the right-hand side of (5) can be ignored, thus
d2 Edpi dpj
2
1
M
m
dfm
dpi
dfm
dpj=
. (6)
161
It is possible to define a new matrix
A
df1
dp1
df2
dp1
.
.
.
dfM
dp1
df1
dp2
df2
dp2
.
.
.
dfM
dp2
.....
.....
.
.
.
.....
df1
dpN
df2
dpN
.
.
.
dfM
dpN
(7)
which consist of derivatives of functions f1 pbar f2 pbar fM pbar.., with respect to the
variation of each parameter component p1 p2 pN.., . Usually M is much larger than N,
therefore A is not a square matrix. Using matrix A and considering (4), gradient ∆ E can
now be represented in matrix form
∆ E 2 AT. f. (8)
where
f f1 f2 fM.. T (9)
Considering also (6) and matrix A, the Hessian matrix G can be replaced
approximately by
162
G 2 AT. A. (10)
Now (9) and (10) can be substituted by Newton’s equation given in equation (11)
pk 1 pk G 1 gk. (11)
yielding
pk 1 pk AkT Ak
.1
AkT fk
.. (12)
where the subscript k in (12) indicates that Ak and fk are evaluated at p bark. Equation
(12) can also be written in the form
AkT Ak
. ∆ pk AkT fk (13)
which is a linear system of equations where AkT Ak
. is an N by N square matrix and AkT Ak
.
is an N-dimensional vector and ∆ pk pk 1 pk is the unknown vector, which can be solved
by Gaussian elimination or decomposition methods from (13) [40].
The Gauss-Newton method is different from Newton’s method because the
Hessian matrix is calculated using only the first derivatives. In the generation of a
Gaussian-Newton sequence for inverse solutions, several problems may be encountered.
First, it is possible that the search sequence does not converge, i.e. E pk 1 E pk> for
163
some k. Second, matrix AkT Ak is near singular (elements very close to zero), and
solution of (13) cannot be obtained. Third, the displacement vector ∆ pk obtained for (13)
is so large that parameter values are not in the admissible region. In order to avoid the
above-mentioned difficulties, it was necessary to develop the modified Gauss-Newton
algorithm by incorporating the Levenberg-Marquardt method.
The contribution of the Levenberg-Marquardt method was to guarantee that
E pk 1 E pk> . The method then took on the name “Gauss-Newton-Levenberg-
Marquardt” algorithms whose modification included and additional term added to AkT Ak
to avoid the singularity
AkT Ak λ M I ∆ pk
. AkT fk (14)
where λ M is a coefficient and I is the unit matrix. When λ M 0, the method reduced to
Gauss-Newton, and if λ M tends to infinity, ∆ pkturns to so called steepest descent direction
and the size ∆ pktends to zero. Therefore, E pk 1 E pk< can always be expected by
increasing he value of λ M.
The procedure for selecting λ M is as follows:
1. Compute E pk
2. Pick a modest value for λ M, say λ M 0.001
3. Solve the linear system of equations (14) to find ∆ pk and calculate E pk ∆ pk
164
4. If E pk ∆ pk E> pk , increase λ M by a factor of 10, and go back to 3
5. If E pk ∆ pk E pk< , decrease λ M by a factor of 10 update the trial solution by
replacing pk with pk ∆ pk and go to the next time step
9.3.3 Optimization Analysis Flow For Test Structure One
Continuing on with the automated analysis, the next message “Inserting CLOSE,
RELIN, RELOUT, and Resistor File Number: 1 Into .cir File” appearing in the Unix shell
mean the settings that control the accuracy, the GNLM optimization component used, and
the measured data used as input, are being updated in the circuit file. In this statement
the optimization model name is “converge”, and convergence success is defined when the
input/output parameters change is less than 1% between successive iterations
(relin/relout). The “close” parameter corresponds to the “lambda” mentioned above. A
high lambda value initiates only the steepest decent component of the GNLM method.
When the gradient of the results functions are less than “cendif” the more time
consuming derivative methods are used. Due to the inaccuracy of the initial guesses, the
goal of the first optimization is to just complete a successful convergence.
Optimization Parameter Settings .model converge opt relin=1e-2 relout=1e-2 cendif=0.1 close=400 itropt=1500
file= '../r1_txt/R1_1.txt'
Table #9.3.3-1: Optimization Settings
165
The next message “Extracting Optimization Information and Element Values
From .out File” appearing in the Unix shell means that the optimization log information
and the most recent optimized element values are being extracted from the .out file.
Optimization Log .model converge opt relin=1e-2 relout=1e-2 cendif=0.1 close=500 itropt=1500 grad= 1.00u close= 500.00 cut= 2.00 residual sum of squares = 5.52464 norm of the gradient = 100.625 marquardt scaling parameter = 500.000 no. of function evaluations = 10 no. of iterations = 0
Optimized Element Values optimized parameters opt1 %norm-sen %change .param c_sq = 1.079756913e-13 $ 7.213396783e-01 -2.251310361e-02 .param rsq = 2.510000000e-06 $ 0. 0. .param lsq = 1.839982236e-11 $ 4.544907570e+01 -9.654611944e-04 .param csq = 6.520000000e-26 $ 0. 0. .param c_pad = 1.590160258e-12 $ 1.099024270e+01 1.007808764e-02 .param rpad = 4.378772295e+00 $ 1.785141128e+01 -5.200190276e-03 .param lpad = 3.049392072e-11 $ 8.867906848e+00 -1.993604484e-02 .param cpad = 2.439993256e-14 $ 1.612002378e+01 -2.764101106e-04
Table #9.3.3-2: Optimization Log and Values
The next message “Checking For Element Values = Zero” will appear in the Unix
shell. Because some of the optimized capacitance values are extremely small,
occasionally the optimizer will return a zero instead of a very small number. This will
create a conflict within Hspice when the next optimization runs. Therefore, the scripts
check for zero values and if found replace the zero with the last optimized non-zero
value. Due to the progressive resolution adjustments shown in Table 9.3.3-5, optimized
values rarely change by any great amount from one run to another.
166
Optimized Element Value Check (rm element_check_c_sq) awk 'if ( $1 == "0." ) print "(cd runs; cp ../spare_1a temp_1a)" ' runs/temp_1a > element_check_c_sq (chmod 777 element_check_c_sq; sh element_check_c_sq) (cd runs; cp temp_1a ../spare_1a)
Table #9.3.3-3: Element Value Check
The next messages “Inserting New Element Values Into .cir File,” “Calculating
The High and Low Ranges Using: 1000,” and “Inserting The High and Low Ranges Into
The .cir File” will appear in the Unix shell. The newly optimized element values are
extracted from the .out file and inserted into the .cir file along with the calculated high
and low element value range. As the optimizations proceed, for a particular measured s-
parameter set, the relin/relout, the close, and the possible range are adjusted to ultimately
converge to the optimum. The range of possible element values is calculated by dividing
and multiplying the optimized element value for a particular run by 10000, 1000, 100, 50,
25, 10, and 5 as the optimization proceed.
167
Optimized Element Value Insertion .param ****************************************** * 30 u line specs ***************************************** + c_sq = opt1(VX_C_SQ,c_sq_LOW,c_sq_HIGH) + rsq = opt1(VX_RSQ,rsq_LOW,rsq_HIGH) + lsq = opt1(VX_LSQ,lsq_LOW,lsq_HIGH) + csq = opt1(VX_CSQ,csq_LOW,csq_HIGH) + r_sub_sq = 10g ************************************** * pad specs ************************************** + c_pad = opt1(VX_C_PAD,c_pad_LOW,c_pad_HIGH) + rpad = opt1(VX_RPAD,rpad_LOW,rpad_HIGH) + lpad = opt1(VX_LPAD,lpad_LOW,lpad_HIGH) + cpad = opt1(VX_CPAD,cpad_LOW,cpad_HIGH) + r_sub_pad = 10g ************************************** .param ****************************************** * 30 u line specs ***************************************** + c_sq = opt1(1.080e-13,1.08e-17,1.08e-09) + rsq = opt1(2.510e-06,2.51e-10,0.0251) + lsq = opt1(1.840e-11,1.84e-15,1.84e-07) + csq = opt1(6.520e-26,6.52e-30,6.52e-22) + r_sub_sq = 10g ************************************** * pad specs ************************************** + c_pad = opt1(1.590e-12,1.59e-16,1.59e-08) + rpad = opt1(4.379e+00,0.0004379,43790) + lpad = opt1(3.050e-11,3.05e-15,3.05e-07) + cpad = opt1(2.440e-14,2.44e-18,2.44e-10) + r_sub_pad = 10g *****************************
Table #9.3.3-4: Element Value Insertion
The next messages “Removing .temp Files,” “Saving and Appending
Optimization Results To Log File,” and “Running TS1 Hspice Optimization Auto Run
Number: 1” will appear in the Unix shell. This step cleans up the previous optimization
run by deleting all .temp files, and by appending the newly generated optimization
information to the master file. The scripts then kick off another optimization, using the
168
previous optimized element values as initial guesses, and adjusting the resolution and
convergence settings for a more accurate analysis. The script lines in Table 9.3.3-5 show
the close, the relin, and the relout settings as the optimizations proceed.
Optimizations Resolution Settings 1 2 3 4 5 6 7 8 9
10 11 12 13 14 15 16
relin=1e-2 relout=1e-2 cendif=0.1 close=300 itropt=1500 relin=1e-2 relout=1e-2 cendif=0.1 close=200 itropt=1500 relin=1e-2 relout=1e-2 cendif=0.1 close=100 itropt=1500 relin=1e-2 relout=1e-2 cendif=0.1 close=50 itropt=1500 relin=1e-2 relout=1e-2 cendif=0.1 close=25 itropt=1500 relin=1e-2 relout=1e-2 cendif=0.1 close=10 itropt=1500 relin=1e-2 relout=1e-2 cendif=0.1 close=5 itropt=1500 relin=1e-2 relout=1e-2 cendif=0.1 close=1 itropt=1500 relin=1e-2 relout=1e-2 cendif=0.1 close=0.1 itropt=1500 relin=1e-2 relout=1e-2 cendif=0.1 close=0.01 itropt=1500 relin=1e-2 relout=1e-2 cendif=0.1 close=0.001 itropt=1500 relin=1e-3 relout=1e-2 cendif=0.1 close=0.001 itropt=1500 relin=1e-4 relout=1e-2 cendif=0.1 close=0.001 itropt=1500 relin=1e-5 relout=1e-2 cendif=0.1 close=0.001 itropt=1500 relin=1e-6 relout=1e-2 cendif=0.1 close=0.001 itropt=1500 relin=1e-7 relout=1e-2 cendif=0.1 close=0.001 itropt=1500
Table #9.3.3-5: Resolution Settings
All the optimization log information generated from each of the 18 analyses is
extracted from the .out file and appended into one file. Table 9.9.3-6 shows the
optimization log information generated from three optimizations using one set of
measured data.
169
Optimizations Optimization Log Information 1
2
3
.model converge opt relin=1e-5 relout=1e-5 cendif=0.1 close=0.001 itropt=1500 grad= 1.00u close= 1.00m cut= 2.00 residual sum of squares = 2.97267 norm of the gradient = 2.090913E-02 marquardt scaling parameter = 67108.9 no. of function evaluations = 57 no. of iterations = 2 .model converge opt relin=1e-6 relout=1e-6 cendif=0.1 close=0.001 itropt=1500 grad= 1.00u close= 1.00m cut= 2.00 residual sum of squares = 2.97267 norm of the gradient = .817667 marquardt scaling parameter = 268435. no. of function evaluations = 33 no. of iterations = 1 .model converge opt relin=1e-7 relout=1e-7 cendif=0.1 close=0.001 itropt=1500 grad= 1.00u close= 1.00m cut= 2.00 residual sum of squares = 2.97267 norm of the gradient = 8.64253 marquardt scaling parameter = 268435. no. of function evaluations = 33 no. of iterations = 1
Table #9.3.3-6: Optimization Log Information
After all 18 analyses have completed for one set of measured data, the “optimal”
optimized element values are inserted into the test structure one simulation file. The
message “Inserting Optimized Element Values Into Simulation File” appears in the Unix
shell window along with the message “Running Simulation File and Extracting S-
Parameter Values.” Table 9.3.3-7 shows the optimized element values calculated after
running 18 optimization simulations.
170
Optimized Element Values In Simulation .cir File .param **************************** * 30 u line specs **************************** + c_sq = 1.230444586e-13 + rsq = 2.510000000e-06 + lsq = 1.831326317e-11 + csq = 6.520000000e-26 + r_sub_sq = 10g **************************** * pad specs **************************** + c_pad = 1.485754792e-12 + rpad = 4.734107069e+00 + lpad = 3.282431715e-11 + cpad = 2.378349977e-14 + r_sub_pad = 10g ****************************
Table #9.3.3-7: Optimized Element Values
The next messages “Extracting TS1 s-parameters from .out File and Calculating Percent
Difference” will appear in the Unix shell. The automated script flow extracts the s-
parameter data from the simulation .out file and compares the results with the actual
measured data used to deembed the element values. Table 9.3.3-8 shows some of the
Unix scripts used in calculating the percent difference between measured vs. modeled.
Percent Difference Between Measured vs. Modeled (awk 'print (100 * (sqrt((1 - ($1 / $2)) * (1 - ($1 / $2))))) ' compare_s11_real_2 > s11_real_compare_2) (awk 'print (100 * (sqrt((1 - ($1 / $2)) * (1 - ($1 / $2))))) ' compare_s11_imag_2 > s11_imag_compare_2) (awk 'print (100 * (sqrt((1 - ($1 / $2)) * (1 - ($1 / $2))))) ' compare_s21_real_2 > s21_real_compare_2) (awk 'print (100 * (sqrt((1 - ($1 / $2)) * (1 - ($1 / $2))))) ' compare_s21_imag_2 > s21_imag_compare_2) (paste '-d \t' done_temp_freq s11_real_compare_2 s11_imag_compare_2 s21_real_compare_2 s21_imag_compare_2 > TS1_data_compared_`cat r_file_number`)
Table #9.3.3-8: Percent Difference In Measured vs. Modeled
171
The next messages “Saving and Appending Optimized Element Values For Run
Number: 1 in Master File” will appear in the Unix shell. The final optimized element
values from each set of measured s-parameters are extracted from optimization run 18
(optimal) and appended into a master file. The test structure two scripts will access this
file during their optimization runs, and the MonteCarlo scripts will access this file to
calculate relative variations in the deembedded element values. Table 9.3.3-9 displays
part of this file.
Optimization Run Number TS1 Deembedded Element Values 1
2
3
residual sum of squares = 2.97264 1.226167537e-13 6.520000000e-26 1.831703392e-11 2.510000000e-06 1.485697892e-12 4.735672319e+00 3.283106370e-11 2.378566732e-14 residual sum of squares = 16.5210 1.019855779e-13 6.520000000e-26 1.842597617e-11 2.510000000e-06 1.582988094e-12 4.384412533e+00 3.027150692e-11 2.486747526e-14 residual sum of squares = 14.0001 1.219019979e-13 6.520000000e-26 1.801390055e-11 2.510000000e-06 1.659259059e-12 4.087316426e+00 2.795404944e-11 2.486223391e-14
Table #9.3.3-9: Deembedded Element Values
The next few messages advise that the simulation of test structure one using the
optimum deembedded element values had been run and the s-parameters, optimization
log, and the simulation circuit file, are being extracted from the .out file and saved. The
messages “Saving s-parameters For Run Number: 1,” “Saving Optimization Run Log
File (18 optimizations) For Run Number: 1,” “Saving Sim File For Run Number: 1” will
appear in the Unix window.
172
The optimization process will now be repeated until all 32 sets of measured s-
parameters have been processed and element values deembedded. The message “End of
TS1 Element Extraction” will appear when the flow scripts have completed test structure
one.
9.3.4 Optimization Analysis Flow Test Structure Two
The automation flow will now move to test structure two and display the message
“Moving TS1 Deembedded Elements File to TS2 Directory.” The optimization of all the
test structure two sets of measured s-parameters will mirror the process completed on test
structure one. The only difference is at the end of each optimization routine (18 runs) the
deembedded element values from the corresponding test structure one and test structure
two optimizations (optimal) will be inserted into the 9-segment meander resistor
simulation file and analyzed. The s-parameter results will also be extracted and percent
difference calculated for each of the 32 measured vs. modeled sets of data. Table 9.3.4-1
shows the analysis flow by displaying the actual messages illustrated in the Unix shell.
173
Steps Test Structure Two Optimizations 1 2 3 4 5 6 7 8 9
10 11 12 13 14 15 16 17 18 19 20 21 22 23 . . . .
24 25 26 27 28 29 30 31 32 33 34 . . . .
35 36
Beginning TS2 Deembedding Using Resistor File: 1 Removing Previous .out File Inserting Initial Guesses and Optimized TS1 Circuit Values Into TS2 .cir file Inserting Opt. TS1 Circuit Values Into TS2 .sim file, and 9-Segment .cir File Generating 1st. .out File Using Initial Guesses Inserting CLOSE, RELIN, RELOUT, and Resistor File #: 1 Into .cir File Extracting & Appending Opt. Info and Element Values From .out File Checking For Element Values = Zero Calculating The High and Low Ranges Using: 1000 Inserting The High and Low Ranges Into The .cir File Removing .temp Files Saving and Appending Optimization Results To Log File Running TS2 Hspice Optimization Auto Run Number: 1 Inserting CLOSE, RELIN, RELOUT, and Resistor File #: 1 Into .cir File Extracting Optimization Information and Element Values From .out File Checking For Element Values = Zero Inserting New Element Values Into .cir File Calculating The High and Low Ranges Using: 1000 Inserting The High and Low Ranges Into The .cir File Removing .temp Files Saving Optimization Results Running TS2 Hspice Optimization Auto Run Number: 2 ........... Running TS2 Hspice Optimization Auto Run Number: 3 ........... ........... ........... ........... ........... Running TS2 Hspice Optimization Auto Run Number: 17 Removing .log Files Inserting Optimized (Optimal) Element Values Into Simulation File Running Simulation File and Extracting S-Parameter Values Extracting 9Seg & TS2 s-parameters and Calculating Percent Difference Saving and Appending Optimized Element Values For Run Number: 1 Saving s-parameters For Run Number: 1 Saving Sim File For Run Number: 1 Saving Runs Log File For Run Number: 1 Compiling and Localizing All Data ........... Beginning TS2 Deembedding Using Resistor File: 2 ........... ........... ........... ........... ........... Beginning TS2 Deembedding Using Resistor File: 32 End of TS2 Element Extraction
Table #9.3.4-1: Test Structure Two Optimizations
At this point, the analysis has generated 32 deembedded building block equivalent
circuit values for test structure one, and generated 32 deembedded building block
174
equivalent circuit values for test structure two. The next block to be presented will be the
“Hspice MonteCarlo Analysis” block, and is shown in Fig. 9.3.4-1.
HspiceHspiceMonteCarloMonteCarlo
AnalysisAnalysis
HspiceHspice Optimization Optimization AnalysisAnalysis
Output Data Synthesis Output Data Synthesis And StorageAnd Storage
Plotting FormatPlotting Format
Analysis InputAnalysis Input
Hspice MonteCarlo AnalysisUsing Automation Flow
Hspice MonteCarlo AnalysisUsing Automation Flow
Figure #9.3.4-1: Hspice MonteCarlo Analysis Block
9.4 Hspice MonteCarlo Analysis
As soon as the Hspice Optimization Analysis has successfully completed the
message “Start MonteCarlo Circuit Construction” will appear in the Unix shell and the
window shown in Fig. 9.4-1 will emerge. The user will insert the number of pads,
number of material squares, number of coupled material squares, and the number of
corners, then the Unix scripts will generate the entire MonteCarlo circuit file. The 9-
segment meander resistor configuration treating each element value independently (i.e.
no correlation) will generate a 2,827 line, 343KB file. Each element value is set up to
generate a new random number on each simulation.
175
9-Segment Meander ResistorMonteCarlo Circuit File Construction
9-Segment Meander ResistorMonteCarlo Circuit File Construction
Figure #9.4-1: MonteCarlo Circuit File Construction
During the construction of the circuit file, messages will appear in the Unix shell advising
the user of the progress. These messages are shown in Table 9.4-1.
Steps 9-Segment Meander Resistor MonteCarlo Circuit File Construction 1 2 3 4 5 6 . . . 7 8 9
10 11 . . .
12 13 14 15 16 . . .
17 18 19 20 21
Inserting .OPTIONS Command Inserting PAD 1 Inserting PAD 2 Inserting Pad .PARAMS Into .Cir File Inserting Substrate Resistance Value Inserting Material Square 1 . . . Inserting Material Square 8 Inserting Material Square .PARAMS Into .Cir File Inserting Substrate Resistance Value End of Material Square Sub-Circuit Construction Inserting Coupled Material Square 1 . . . Inserting Coupled Material Square 9 Inserting Coupled Material Square .PARAMS Into .Cir File Inserting Substrate Resistance Value End of Coupled Square Sub-Circuit Construction Inserting Corner 1 . . . Inserting Corner 8 Inserting Corner .PARAMS Into .Cir File Inserting Substrate Resistance Value End of Corner Sub-Circuit Construction End of Circuit Construction
Table #9.4-1: Circuit File Construction
176
Once the MonteCarlo circuit file has been collated, the 32 sets of element values
from test structure one and test structure two will be analyzed statistically yielding mean
and relative variation values for each deembedded element value. During this process,
messages will appear in the Unix shell advising the user of the progress. These messages
are shown in Table 9.4-2.
Steps Statistical Analysis of Deembedded Element Values 1 2 3 4 5 6 7 8 9
10 11 12
Combining TS1 and TS2 Deembedded Element Values Into One File Sorting The File (Ascending) Calculating The Median Calculating The Absolute Variation Calculating The Relative Variation Calculating The Mean Calculating The Sum of Squares Calculating The Sum Squared Calculating The Variance Calculating The Standard Deviation Calculating The Standard Deviation/Mean Inserting Relative Variations Into 9-Segment MonteCarlo .cir File
Table #9.4-2: Statistical Analysis of Deembedded Element Values
The 9-segment meander resistor MonteCarlo circuit file is now complete. The
message “Running 9Seg MonteCarlo Analysis” will appear in the Unix shell advising the
user that the MonteCarlo analysis has begun. The MonteCarlo analysis generated 500
simulations of the 9-segment meander resistor circuit file while varying each
deembedded element value on each simulation. The Hspice MonteCarlo analysis was run
using a four-processor HP model 1100 computer. The analysis run time was
approximately 2 hours. At the successful completion of the analysis the message “End of
9Seg MonteCarlo Analysis” will appear in the Unix shell. The last block to be presented
will be the “Output Data Synthesis and Storage” block, and is shown in Fig. 9.4-2.
177
HspiceHspice MonteCarloMonteCarloAnalysisAnalysis
HspiceHspice Optimization Optimization AnalysisAnalysis
Output Data Output Data Synthesis Synthesis
And StorageAnd Storage
Plotting FormatPlotting Format
Analysis InputAnalysis Input
Output Data Synthesis and Storage ForStatistical Analysis Automation Flow
Output Data Synthesis and Storage ForStatistical Analysis Automation Flow
Figure #9.4-2: Output Data Synthesis
9.5 Output Data Synthesis And Storage
After the entire analysis has completed there is still an issue concerning the
management of all the data generated during the statistical analysis process. Files need to
be saved, parameters need to be converted, and results need to be compared. Portions of
the data synthesis were preformed as they were generated. Such as the s-parameters
results from test structure one, test structure two, and the 9-segment resistor. During the
optimization process the messages “Extracting TS1 s-parameters from .out File and
Calculating Percent Difference,” and “Extracting 9Seg & TS2 s-parameters and
Calculating Percent Difference” were seen scrolling by in the Unix shell. However, most
of the data management is accomplished at the end of the analysis. For example the
results from both sets of 32 deembedded element values were automatically inserted into
178
a 9-segment resistor Hspice circuit file and analyzed. The results were extracted from the
.out file converted to polar coordinates, and put in separate files (i.e. magnitude/phase,
magnitude, phase) and saved. Another option inserts the element values from a particular
deembedded analysis run into a 9-segment Hspice circuit file, simulates the design then
compares the results with all 32 measured 9-segment devices. This is done for each set
of deembedded element values. Table 9.5-1 shows the extent of the data generated and
saved. All the data generated and documented below is saved on a CD, using
approximately 400Meg its available space.
Steps Data Synthesis And Storage 1 2 3 4 5 6 7 8 9
10 11 12 13 14 15 16 17 18 19 20
Extracting TS1 s-parameters and Calculating Percent Difference Extracting TS2 s-parameters and Calculating Percent Difference Extracting 9Seg s-parameters and Calculating Percent Difference Saving All Percent Difference Files Extracting TS1 Optim. Information and Element Values From .out File Extracting TS2 Optim. Information and Element Values From .out File Saving All Optimization Runs Log File For Run Number: 1 Saving and Appending Optimized TS1 Element Values For Run Number: 1 Saving and Appending Optimized TS2 Element Values For Run Number: 1 Saving TS1 s-parameters For Run Number: 1 Saving TS2 s-parameters For Run Number: 1 Saving 9-Seg s-parameters For Run Number: 1 Saving TS1 Circuit File For Run Number: 1 Saving TS2 Circuit File For Run Number: 1 Saving 9-Seg Circuit File For Run Number: 1 Extracting s-parameters Data For Each (500) MonteCarlo Run Converting All TS1, TS2, 9-Seg S-parameter Data to Polar Coordinates Separating All s-parameters files Into Real, Imag. & Magnitude, Phase Running 9Seg Sims’ Using Deembedded Circuit Values From All Analysis Compiling and Localizing All Data
Table #9.5-1: Data Synthesis and Storage
179
9.6 Conclusion
The Statistical Analysis Automation Flow completed all aspects of this research
except for rendering the plots. This was accomplished using Microsoft Excel mainly due
to the exceptional graphic quality. After learning the mechanics behind the Gauss-
Newton Levenberg-Marquardt Algorithm, running iterative optimizations while adjusting
the resolution, convergence parameters, and range of possible element values, was a
process that had to be preformed. In the quest of statistical prediction, measured vs.
modeled accuracy yields a more realistic outcome. The generation and application of the
Unix shell scripts aided this research immensely. The picture in Fig. 9.6-1 shows the full
scope of the Statistical Analysis Automation Flow.
Convert MC & Model Resistor S-Parameter To Polar Coordinates
HSPICESIMULATIONTS1 & TS2 &
MODEL RESISTOR
Calculate % ErrorMeasured vs. Modeled For TS1 & TS2
Deembedded ElementValues For TS1
HSPICE
OPTIMIZATION
TS1 & TS2
HSPICEMONTECARLO
ANALYSIS
PLOTPLOT
Model Circuit Description
Extract S-ParametersFrom .out files &
Separate Into SingleReal/Imag, Real, Imag
Master Files
Deembedded ElementValues For TS2
Model Circuit Generation
Statistical AnalysisIn Excel PLOTPLOT
PLOTPLOT
Model ResistorMeasured S-Parameters
Save Optimization Files
Save .cif Files
LegendLegendInput Parameters
Hspice Analysis
.cir Files & Measurements
Output File Information
Deembedded Elements
TS1 Initial GuessTS2 Initial Guess
TS1 & TS2 .cir Files
Relative Variation of Element Values
TS1 & TS2 Measured S-Parameters
Number of TS1 & TS2 Measured Devices
Automation of Predictive Passive Device Modeling Methodology Using Unix Shell ScriptsAutomation of Predictive Passive Device Modeling Methodology Using Unix Shell Scripts
Figure #9.6-1: Statistical Analysis Automation Flow
180
CHAPTER 10
CONCLUSION AND RECOMENDATIONS
10.1 Summary of Research and General Conclusions
A further developed passive device modeling methodology that accurately
predicts the high frequency electrical behavior, and the statistical variation of embedded
passive devices on a generic substrate, has been presented. This methodology accurately
modeled new passive devices by utilizing a small number of test structures to deembed
the lumped element circuit values of canonical building blocks. The fundamental idea
behind the modeling procedure is that most designed passive structures are comprised of
several key geometrical building blocks, that is, they can be constructed from several
building block cells representing individual parts of the structure.
The objective of this research is to predict the electrical behavior and statistical
variation of a 9-segment meander resistor in a standard circuit simulator. The number of
test structures needed to accurately model all of the required building blocks must be as
few as possible for the method to be practical; therefore, deembedding 22 element values
from four canonical building blocks using only one test structure was attempted and
successfully achieved, along with rendering a substantial increase in measured vs.
181
modeled accuracy for the test structures as well as the modeled device. This proved that
with enough frequency points and sufficient sensitivity, accurate equivalent circuit
extraction of individual building blocks from multi building block test structures is
possible.
Other improvements to the overall existing methodology such as enhancing the
corner primitive, not deembedding the substrate resistance, and running consecutive
simultaneous optimizations per measured device, also resulted in a substantial increase in
measured vs. modeled accuracy for the test structures as well as the modeled device.
These improvements were fueled by the theory; the more accurate the measured vs.
modeled results of the test structures, the more accurate the modeled device, and
consequently the tighter and more realistic the predicted range of operation.
In an attempt to ease research effort and to gain wider acceptance, the entire
Optimization/MonteCarlo methodology was completely automated. Unix scripts were
developed which act as interface between user and computer by; automatically, running
the complete Hspice Optimization simulations, calculating the relative variations of the
deembedded element values, generating the device model circuit file, running the Hspice
MonteCarlo analysis, extracting and converting all s-parameters to polar coordinates, and
saving and localizing all circuit and optimization files. This automation will go a long
way towards easing the difficulty in using the Hspice Tool, and making the entire
methodology less cumbersome.
182
10.2 Application
The method of predicting passive device performance variations and yield, when
provided with a statistical distribution of the building block models and a circuit
schematic was successfully proven using the 9-segment meander resistor. Randomly
generated element values of the 9-segment meander resistor were simulated in a
MonteCarlo tool to produce a representative set of output performance characteristics.
The impact of random process variations was inferred from these simulations, and
electrical yield was estimated using the percentage of that sample which met the
performance requirement.
Electrical yield refers to the number of parts that meet the electrical test
specifications; whereas, maximizing yield is important for the overall process efficiency.
Fig. 10.2-1 shows the Gaussian distribution curve representing 500 MonteCarlo produced
simulations in terms of mean and standard deviation. The proposition is that 95.44% (2
standard deviations) of the actual 32 measured devices will be contained within the
MonteCarlo predicted range.
183
Mean and Standard DeviationsMean and Standard Deviations
MonteCarlo MeanMonteCarlo Mean
34.13%34.13%34.13%34.13%
13.59%13.59% 13.59%13.59%
00 11-1-1 22-2-2
95.44%95.44%
Standard DeviationsStandard Deviations
Figure #10.2-1: MonteCarlo Predicted Mean and Standard Deviation
10.2.1 Statistical Analysis Results From Chapter 7
The optimum modeling method for the 9-segment meander resistor was presented
in Chapter 7, and the optimum MonteCarlo analysis was generated using the method in
Chapter 7 coupled with the independent premise used in collating the circuit model file.
The MonteCarlo data generated by this method was investigated at two frequency points,
45MHz(DC) & 1.7GHz for S11, and 45MHz(DC) & 4.0GHz for S21. Using the
MonteCarlo generated data; the standard deviation was calculated, multiplied by two, and
added/subtracted to the mean value at each frequency. This created a high and low range
(horizontal red lines) designating the range of possible values and the predicted yield for
the 9-segment meander resistor. The measured data from the 9-segment resistor, at the
corresponding frequency points was inspected to ascertain what percentage of the 32
measured devices fell within the MonteCarlo (two standard deviation) predicted range,
and is shown in Fig. 10.2.1-1Excel plots.
184
MonteCarlo Predicted Two Standard Deviations OutComparison Of Actual Measured Devices At 4.0GHz
0.53
0.535
0.540.545
0.55
0.555
0.56
0.5650.57
0.575
0.58
0 5 10 15 20 25 30
Number of Measured Devices
S21Mag.
Measured Devices
Predicted High Range
Predicted Low Range
MonteCarlo Predicted Two Standard Deviations OutComparison Of Actual Measured Devices At 45MHz
0.5550.56
0.5650.57
0.5750.58
0.5850.59
0.5950.6
0.6050.61
0 5 10 15 20 25 30
Number of Measured Devices
S21Mag.
Measured Devices
Predicted High Range
Predicted Low Range
MonteCarlo Predicted Two Standard Deviations OutComparison Of Actual Measured Devices At 45MHz
0.390.395
0.40.4050.41
0.4150.42
0.4250.43
0.4350.44
0.445
0 5 10 15 20 25 30
Number of Measured Devices
S11Mag.
Measured DevicesPredicted High Range
Predicted Low Range
MonteCarlo Predicted Two Standard Deviations OutComparison Of Actual Measured Devices At 1.7GHz
0.41
0.415
0.42
0.425
0.43
0.435
0.44
0.445
0.45
0.455
0 5 10 15 20 25 30
Number of Measured Devices
S11Mag.
Measured Devices
Predicted High Range
Predicted Low Range
MonteCarlo Predicted Yield and Element Value Variation96.88% Of Actual Measured Devices (31 out of 32) Fell Within Predicted Range
Using Independent Premise
MonteCarlo Predicted Yield and Element Value Variation96.88% Of Actual Measured Devices (31 out of 32) Fell Within Predicted Range
Using Independent Premise
Not IncludedNot Included
Not IncludedNot Included
Not IncludedNot IncludedNot IncludedNot Included
Figure #10.2.1-1: MonteCarlo Predicted Range vs. Actual Measured Devices Using
Independent Premise
By inspection, 96.88% (1.44% more) of the actual measured devices fell within
the MonteCarlo predicted range, generating a rectangular area of possible values and a
predicted yield. Fig. 10.2.1-2 shows the MonteCarlo (two standard deviation) predicted
range (Red), the actual MonteCarlo predicted range (Blue Dots), and the actual measured
data.
185
9-Segment Meander ResistorPredicted Range of Values and Statistical Yield
0.38
0.39
0.4
0.41
0.42
0.43
0.44
0.45
0.46
1.E+08 1.E+09 1.E+10
Frequency
S11Mag
9-Segment Meander ResistorPredicted Range of Values and Statistical Yield
0.53
0.54
0.55
0.56
0.57
0.58
0.59
0.6
0.61
0.62
1.E+08 1.E+09 1.E+10
Frequency
S21Mag
MonteCarlo Predicted 95.44% of Measured Devices Actual Measured 9-Segment Resistors Totaled 96.88% Using Independent Premise Method
MonteCarlo Predicted 95.44% of Measured Devices Actual Measured 9-Segment Resistors Totaled 96.88% Using Independent Premise Method
1.7GHz1.7GHz
4.0GHz4.0GHz
Figure #10.2.1-2: MonteCarlo Predicted Range vs. Actual Measured Devices Using
Independent Premise
10.2.2 Statistical Analysis Results From Chapter 8
The modeling method for the 9-segment meander resistor presented in Chapter 8
coupled with the independent premise used in collating the circuit model file was use to
generated the MonteCarlo simulations. The MonteCarlo data generated by this method
was investigated at two frequency points, 45MHz(DC) & 1.04GHz for S11, and
45MHz(DC) & 2.0GHz for S21. Using the MonteCarlo generated data; the standard
deviation was calculated, multiplied by two, and added/subtracted to the mean value at
each frequency. This created a high and low range (horizontal red lines) designating the
range of possible values and the predicted yield for the 9-segment meander resistor. The
measured data from the 9-segment resistor, at the corresponding frequency points was
inspected to ascertain what percentage of the 32 measured devices fell within the
MonteCarlo (two standard deviation) predicted range, and is shown in Fig. 10.2.2-1Excel
plots.
186
MonteCarlo Predicted Two Standard Deviations OutComparison Of Actual Measured Devices At 2.04GHz
0.555
0.56
0.565
0.57
0.575
0.58
0.585
0.59
0.595
0.6
0 5 10 15 20 25 30
Number of Measured Devices
S21Mag.
Measured Devices
Predicted High Range
Predicted Low Range
MonteCarlo Predicted Two Standard Deviations OutComparison Of Actual Measured Devices At 45MHz
0.565
0.57
0.575
0.58
0.585
0.59
0.595
0.6
0.605
0 5 10 15 20 25 30
Number of Measured Devices
S21Mag.
Measured Devices
Predicted High Range
Predicted Low Range
MonteCarlo Predicted Two Standard Deviations OutComparison Of Actual Measured Devices At 1.04GHz
0.395
0.4
0.405
0.41
0.415
0.42
0.425
0.43
0.435
0.44
0 5 10 15 20 25 30
Number of Measured Devices
S11Mag.
Measured Devices
Predicted High Range
Predicted Low Range
MonteCarlo Predicted Two Standard Deviations OutComparison Of Actual Measured Devices At 45MHz
0.39
0.395
0.4
0.405
0.41
0.415
0.42
0.425
0.43
0.435
0 5 10 15 20 25 30
Number of Measured Devices
S11Mag.
Measured Devices
Predicted High Range
Predicted Low Range
MonteCarlo Predicted Yield and Element Value Variation96.88% Of Actual Measured Devices (31 out of 32) Fell Within Predicted Range
Using Independent 1TS Premise
MonteCarlo Predicted Yield and Element Value Variation96.88% Of Actual Measured Devices (31 out of 32) Fell Within Predicted Range
Using Independent 1TS Premise
Not IncludedNot Included Not IncludedNot Included
Not IncludedNot Included Not IncludedNot Included
Figure #10.2.2-1: MonteCarlo Predicted Range vs. Actual Measured Devices Using
Independent 1TS Premise
By inspection, 96.88% (1.44% more) of the actual measured devices fell within
the MonteCarlo predicted range at all four frequencies, generating a rectangular area of
possible values and a predicted yield. Fig. 10.2.2-2 shows the MonteCarlo (two standard
deviation) predicted range (Red), the actual MonteCarlo predicted range (Blue Dots), and
the actual measured data.
187
9-Segment Meander ResistorPredicted Range of Values and Statistical Yield
0.38
0.39
0.40
0.41
0.42
0.43
0.44
0.45
1.E+08 1.E+09 1.E+10
Frequency
S11Mag
9-Segment Meander ResistorPredicted Range of Values and Statistical Yield
0.55
0.56
0.57
0.58
0.59
0.6
0.61
0.62
1.E+08 1.E+09 1.E+10
Frequency
S21Mag
MonteCarlo Predicted 95.44% of Measured Devices Actual Measured 9-Segment Resistors Totaled 96.88% Using Independent 1TS Premise Method
MonteCarlo Predicted 95.44% of Measured Devices Actual Measured 9-Segment Resistors Totaled 96.88% Using Independent 1TS Premise Method
1.0GHz1.0GHz2.0GHz2.0GHz
Figure #10.2.2-2: MonteCarlo Predicted Range vs. Actual Measured Devices Using
Independent 1TS Premise
10.2.3 Statistical Analysis Results From Chapter 7 Using Correlated Premise
The modeling method for the 9-segment meander resistor presented in Chapter 7
coupled with the correlated premise used in collating the circuit model file was use to
generated the MonteCarlo simulations. The MonteCarlo data generated by this method
was investigated at two frequency points, 45MHz(DC) & 1.0GHz for S11, and
45MHz(DC) & 2.0GHz for S21. Using the MonteCarlo generated data; the standard
deviation was calculated, multiplied by two, and added/subtracted to the mean value at
each frequency. This created a high and low range (horizontal red lines) designating the
range of possible values and the predicted yield for the 9-segment meander resistor. The
measured data from the 9-segment resistor, at the corresponding frequency points was
inspected to ascertain what percentage of the 32 measured devices fell within the
MonteCarlo (two standard deviation) predicted range, and is shown in Fig. 10.2.3-1Excel
plots.
188
MonteCarlo Predicted Two Standard Deviations OutComparison Of Actual Measured Devices At 2.0GHz
0.56
0.565
0.57
0.575
0.58
0.585
0.59
0.595
0.6
0 5 10 15 20 25 30
Number of Measured Devices
S21Mag.
Measured Devices
Predicted High Range
Predicted Low Range
MonteCarlo Predicted Two Standard Deviations OutComparison Of Actual Measured Devices At 45MHz
0.565
0.57
0.575
0.58
0.585
0.59
0.595
0.6
0.605
0 5 10 15 20 25 30
Number of Measured Devices
S21Mag.
Measured Devices
Predicted High Range
Predicted Low Range
MonteCarlo Predicted Two Standard Deviations OutComparison Of Actual Measured Devices At 1.0GHz
0.395
0.4
0.405
0.41
0.415
0.42
0.425
0.43
0.435
0.44
0 5 10 15 20 25 30
Number of Measured Devices
S11Mag.
Measured Devices
Predicted High Range
Predicted Low Range
MonteCarlo Predicted Two Standard Deviations OutComparison Of Actual Measured Devices At 45MHz
0.39
0.395
0.4
0.405
0.41
0.415
0.42
0.425
0.43
0.435
0 5 10 15 20 25 30
Number of Measured Devices
S11Mag.
Measured Devices
Predicted High Range
Predicted Low Range
MonteCarlo Predicted Yield and Element Value Variation81.25% Of Actual Measured Devices (26 out of 32) Fell Within Predicted Range
Using Correlated Premise
MonteCarlo Predicted Yield and Element Value Variation81.25% Of Actual Measured Devices (26 out of 32) Fell Within Predicted Range
Using Correlated Premise
2 Not Included2 Not Included 2 Not Included2 Not Included
2 Not Included2 Not Included 2 Not Included2 Not Included
4 Not Included4 Not Included 3 Not Included3 Not Included
4 Not Included4 Not Included 3 Not Included3 Not Included
Figure #10.2.3-1: MonteCarlo Predicted Range vs. Actual Measured Devices Using
Correlated Premise
By inspection, 81.25% (14.19% less) of the actual measured devices fell within
the MonteCarlo predicted range at all four frequencies, generating a rectangular area of
possible values and a predicted yield. Fig. 10.2.3-2 shows the MonteCarlo 2 standard
deviation predicted range (Red), the actual MonteCarlo predicted range (Blue Dots), and
the actual measured data.
189
9-Segment Meander ResistorPredicted Range of Values and Statistical Yield
0.38
0.39
0.4
0.41
0.42
0.43
0.44
0.45
1.E+08 1.E+09 1.E+10
Frequency
S11Mag
9-Segment Meander ResistorPredicted Range of Values and Statistical Yield
0.55
0.56
0.57
0.58
0.59
0.6
0.61
0.62
1.E+08 1.E+09 1.E+10
Frequency
S21Mag
MonteCarlo Predicted 95.44% of Measured Devices Actual Measured 9-Segment Resistors Totaled 81.25% Using Correlated Premise Method
MonteCarlo Predicted 95.44% of Measured Devices Actual Measured 9-Segment Resistors Totaled 81.25% Using Correlated Premise Method
1.0GHz1.0GHz2.0GHz2.0GHz
Figure #10.2.3-2: MonteCarlo Predicted Range vs. Actual Measured Devices Using
Correlated Premise
10.3 Discussion
An accurate range of possible element values variations and an accurate yield
prediction was generated for the 9-segment meander resistor using the statistical analysis
procedure, and the improved modeling methodology in Chapter 4. A description of the
MonteCarlo simulation process and several statistical analysis techniques, used to
evaluate and render the results, were also presented. The MonteCarlo data generated
from a circuit model file constructed using the “correlated” premise vs. the noncorrelated
or “independent” premise was investigated, and found that no correlation existed between
any of the building block equivalent circuit elements. This was shown to be true in the
poor yield and element value predictions in Figures 10.2.3-1 and 10.2.3-2, and is mainly
due to the non-uniformity of the “Sputtering” technique used to deposit the gold onto the
Alumina substrate. A successful statistical analysis prediction of the same 9-segment
190
meander resistor was also generated using only one test structure and the independent
premise.
10.4 Recommendations
To further refine the measured vs. modeled accuracy of the test structures, higher
order coupling will need to be modeled as designs move further into the giga hertz. With
the automated Optimizatio/MonteCarlo analysis in place, fabricated structures and
measured data is all that is needed to effectively model any type of passive device.
Because no assumptions are made regarding material properties, layered dielectrics, or
imperfect substrates, when using measurement to model a device; a much greater
understanding of device/substrate effects can be extracted by using the automated flow to
its potential.
10.5 Final Conclusion
This research detailed the method of predicting passive device performance
variations and yield, when provided with a statistical distribution of the building block
models and a circuit schematic. A comparison of the MonteCarlo simulation results to
actual measured data of a 9-segment meander resistor revealed that the behavior
variations in the fabricated structures were a subset of the predicted variations obtained
from the MonteCarlo results. This means that the building block equivalent circuits
191
formerly proven to model the entire resistor structure are also accurate enough to model
the variations, created by process fluctuations and non-ideal properties, between devices.
It has therefore been determined that the statistical variation of the component values
provides and accurate representation of the overall resistor performance. This modeling
methodology’s potential was shown in Chapter 7 where performance and parametric
yield of the 9-segment meander resistor was predicted.
Dr. Ravi Poddar’s statement concerning the basis of his novel modeling
methodology was quoted as:
“If accurate models for each of the building blocks along with interaction
information can be obtained, then any arbitrary structure comprised of those
building blocks can be modeled accurately using the individual block models.”
This research has expanded on that quote to now include:
“The variation in the building block equivalent circuit models, based only on
the data used to extract them, can be used to predict the variations in actual
fabricated devices; therefore, providing an overall accurate representation of
the device performance.”
222
REFERENCES
[1] Rector, J., “Economic and technical viability of integral passives,” Proceedings.
4th International Symposium on Advanced Packaging Materials Processes, p.277-83, March 1998.
[2] Yang Rao; Jianmin Qu; Wong, C.P., “Electrical and mechanical modeling of
embedded capacitors,” 49th Electronic Components and Technology Conference , p.506-9, June 1999.
[3] Sorrentino, R., “Numerical methods for passive components,” 1988 IEEE MTT
International Microwave Symposium Digest, Cat. No.88CH2489-3, p.619-22 vol.2, May 1988.
[4] Poddar, R.; Brooke, M.A., “Accurate high speed empirically based predictive modeling of deeply embedded gridded parallel plate capacitors fabricated in a multilayer LTCC process,” IEEE Transactions on Advanced Packaging, vol.22, no.1 p.26-31, Feb. 1999.
[5] Lakshminarayanan, B.; Gordon, H.C.; Weller, T.M., “A substrate-dependent CAD model for ceramic multilayer capacitors,” IEEE Transactions on Microwave Theory and Techniques, vol.48, no.10p.1687-93, Oct. 2000.
[6] Ronkainen, H.; Kattelus, H.; Tarvainen, E.; Ruhisaari, T.; Andersson, M.;, “IC compatible planar inductors on silicon,” IEE Proceedings-Circuits, Devices and Systems, vol.144, no.1 p.29-35, Feb. 1997.
[7] Kmayman, N.; Jain, N., “Automatic and accurate lumped-model generation of millimeter wave passive components using EM simulator,” 29th European Microwave Conference 99, vol.1, p.170-3, Oct. 1999.
[8] del Mar Hershenson, M.; Mohan, S.S.; Boyd, S.P.; Lee, T.H., “Optimization of inductor circuits via geometric programming,” Proceedings 1999 Design Automation Conference, Cat. No. 99CH36361, p.994-8, June 1999.
223
[9] Wang, S.; Wang, F.; Devabhaktuni, V.K.; Zhang, Q.-J., “A hybrid neural and circuit-based model structure for microwave modeling,” 29th European Microwave Conference 99, vol.1, p.174-7, Oct. 1999
[10] Belk, N.R.; Frei, M.R.; Tsai, M.; Becker, A.J.; Tokuda, K.L., “The simulation and design of integrated inductors,” Proceedings 1999 Design Automation Conference, Cat. No. 99CH36361, p.988-93, June 1999.
[11] Rejaei, B.; Tauritz, J.L.; Snoeij, P. Kayali, S., “A predictive model for Si-based circular spiral inductors,” Proceedings of First Topical Meeting on Silicon Monolithic Integrated Circuits in RF Systems, Cat. No.98EX271, p.148-54, Sept. 1998.
[12] Pieters, P.; Beyne, E., “Spiral inductors integrated in MCM-D using the design space concept,” International Conference on Multichip Modules and High Density Packaging, Cat. No.98EX154, p.478-83, April 1998.
[13] Arcioni, P.; Castello, R.; Perregrini, L.; Sacchi, E.; Svelto, F., “An innovative modelization of loss mechanism in silicon integrated inductors,” IEEE Transactions on Circuits and Systems II, vol.46, no.12 p.1453-60, Dec. 1999.
[14] F.W. Grover, Inductance Calculations. New York: Van Nostrand, 1946
[15] Koutsoyannopoulos, Y.K.; Papananos, Y., “Systematic analysis and modeling of integrated inductors and transformers in RF IC design,” IEEE Transactions on Circuits and Systems II: Analog and Digital Signal Processing, vol. 47, no.8 p.699-713, Aug. 2000.
[16] Delaney, K.; Barrett, J.; Barton, J.; Doyle, R., “Characterization of the electrical performance of buried capacitors and resistors in low temperature co-fired (LTCC) ceramic,” 48th Electronic Components and Technology Conference, Cat. No.98CH 36206, p.900-8, May 1998.
[17] Kwang Lim Choi; Swaminathan, M., “Synthesis of RF circuits for embedded passive components in mixed signal,” 48th Electronic Components and Technology Conference, Cat. No.98CH 36206, p.1052-60, May 1998.
[18] R. Poddar and M. Brooke, “Accurate, High Speed Modeling of Integrated Passive Devices,” IMAPS ATW Passive Component Technology, Braselton, GA, March 1997.
224
[19] Carastro, L.A.; Poddar, R.; Moon, E.; Brooke, M.A.; Jokerst, N.M. “Passive device modeling methodology using nonlinear optimization”, ISCAS'99. Proceedings of the 1999 IEEE International Symposium on Circuits and Systems VLSI (Cat. No.99CH36349) p.53-6 vol.6, 1999
[20] Beattie, M.W.; Pileggi, L.T., “IC analyses including extracted inductance models,” Proceedings 1999 Design Automation Conference, Cat. No. 99CH36361, p.915-20, June 1999.
[21] Vandenbossche, E.; Kopalidis, G.; Tack, M.; Schoenmaker, W. Meyer, K.D.; Biesemans, “Statistical modeling based on extensive TCAD simulations. Proposed methodology for extraction of fast/slow models and statistical models,” Simulation of Semiconductor Processes and Devices, p.85-8, Sept. 1998.
[22] R. Poddar and M. Brooke, "Accurate, High Speed Modeling of Integrated Passive Devices in Multichip Modules," Proc. Electrical Performance of Electronic Packaging, Oct. 1996, pp. 184-186.
[23] R. Poddar and M. Brooke, “Integrated Passive Device Design Based Upon Design and Modeling of Test Structures,” IMAPS Next Generation Package Design Workshop, South Carolina, June 1997.
[24] M. Engles and R. H. Jansen, “Modeling and Design of Novel Passive MMIC Components with Three and More Conductor Layers”, IEEE MTT-S Digest, pp. 1293-6, 1994.
[25] R. Poddar, E. Moon, N. Jokerst, and M. Brooke, “Accurate, Rapid High Frequency Empirically Based Predictive Modeling of Arbirary Geometry Planar Resistive Passive Devices,” submitted for publication in IEEE Trans. on Components, Packaging, and Manufacturing Technology, Part B.
[26] A. E. Ruehli, “Equivalent Circuit Models for Three Dimensional Multiconductor Systems,” IEEE Trans. Microwave Theory Tech., vol. MTT-22, pp. 216-221, Mar. 1974.
[27] Y. Eo and W. R. Eisenstadt, “High-Speed VLSI Interconnect Modeling Based on S-Parameter Measurements,” IEEE Trans. On Components, Hybrids, and Manufacturing Tech., vol. 16, no. 5, pp. 555-562, 1993.
225
[28] R. Poddar and M. Brooke, “Accurate, High Speed Modeling of Integrated Passive Devices in Multichip Modules,” IEEE Topical Meeting on Electrical Performance of Electronic Packaging, pp. 184-6, 1996.
[29] David S. Gibson and Ravi Poddar and Gary S. May, “Statistically Based Parametric Yield Prediction for Integrated Circuits,” IEEE Transactions On Semiconductor Manufacturing, pp. 445-458, vol. 10, no.4, November 1997
[30] George R. Cooper and Clare D. McGillern, “Probabilistic Methods of Signal and System Analysis,” New York, NY., Harcourt Brace Jovanovich College Publishers, 1986.
[31] Meta-Software, “HSPICE User’s Manual,” Campbell, CA., Meta-Software Inc., 1992.
[32] G.Kranzler and J.Moursund, “Statistics for the Terrified,” Upper Saddle River, New Jersey, Prentice Hall, 1999.
[33] Daniel Gilly, “Unix in a Nutshell,” Sebastopol, CA., O’Reilly & Associates, Inc., 1992.
[34] Richard M. Stallman, “GNU Emacs Manual,” Boston, MA., Free Software Foundation, version 20.7, June 2000.
[35] R. Brown and A. Shapiro, "Integrated Passive Components and MCMs: The Future of Microelectronics," Proc. Int'l. Conf. Exhibition Multichip Modules, Apr. 1993, pp. 287-94.
[36] Arnold Robbins, “awk Programming,” USA, O’Reilly & Associates, Inc., May 2001.
[37] Nancy Walsh, “Perl/Tk,” Sebastopol, CA., O’Reilly & Associates, Inc., 1999.
[38] Randal L. Schwartz, “Learning Perl,” Sebastopol, CA., O’Reilly & Associates, Inc., 1993.
[39] Paul W. Abrahams and Bruce R. larson, “Unix for the Impatient,” USA, Addison Wesley Longman, Inc., second, 1993.
226
[40] D. Marquardt, "An Algorithm for Least-Squares Estimation of Nonlinear Parameters," J. Soc. Ind. Appl. Math., vol. 11, pp. 431-441, 1963.