Equivalent Circuit Modelling of Bipolar Transistors
Equivalent Circuit Modelling of Bipolar Transistors
Equivalent Circuit Modelling of Bipolar Transistors
by
Monica Dowson
March 1982
ACKNOWLEDGMENTS
with and the provision of data by Mr. P. J. Rankin and Mr. K. W. Moulding
who provided encouragement, food and drink at all the appropriate times.
EQUIVALENT CIRCUIT MODELLING OF BIPOLAR TRANSISTORS
Monica Dowson
ABSTRACT
Page
CHAPTER 1 INTRODUCTION 1
1.1 Background 2
1.3 Achievements 4
3.1 y Parameters 25
3.2 s Parameters 39
S. 4 Improvement of Models 63
REFERENCES 194
I
CHAPTER 1
INTRODUCTION
sought which simulate the behaviour of the device. The work undertaken
by the author and described here, has been concerned entirely with
models are popular for several reasons: circuit designers tend to feel more
construction to be mapped. onto the model; and there are many circuit
analysis packages now available that use equivalent circuit models of tran-
transistors from Philips Research Laboratories for which they were inter-
also provided a definite direction for the research. The requirement was
for small-signal (implying linear) models for transistors designed to
operate at frequencies in ranges between 0.5 MHz and 2 GHz. MacleanI has
Although the techniques described here were originally devised for the
here, which was a bias dependent model, the author obtained models for the
used one optimisation technique to model both bipolar and field effect
transistors.
1.1. Background
ted theses more concerned with the optimisation processes. This expertise
however, only used for a short time until the analysis routine written by
used was not considered suitable for the problems being studied. by the author.
technique involved the use of a two-part algorithm and was designed to solve
problems to be solved by the author one part, the Gauss Newton section,
with adaptations to suit the problem type, was an ideal technique. Details
of the optimisation problem and the techniques used are given in Chapter 4.
3
The first set of data13 was for a lateral p-n-p transistor. This was
with the measured characteristics of the devices. Thus, the aim of this
to optimise the topology of the models together with the element values
This new technique was then tested on the second set of data provided
by Philips Research Laboratories 14 This data was for integrated
- another
circuit transistor, this time a vertical n-p-n transistor, with data for
each of the common emitter, common base and common collector configurations.
Finally, a set of data'5'16 was provided for two similar n-p-n tran-
4
two voltages and at six different currents for the first type and at two
voltages and at seven different currents for the second type. Slatter17
had published a model for a limited bias range based on physical aspects
of these devices. The author's aim was to improve on this model by ob-
taining more accurate bias dependent models covering a wider bias range
using as few bias dependent elements as possible consistent with the use
in detail in Chapter 7.
Achievements
As will be seen later the author has produced a method by which equi-
the topology of the model can be modified until the required accuracy or
complexity is obtained.
Using this technique a number of models have been produced using data
model for a lateral p-n-p transistor which operated in the frequency range
0.5 MHz to 10 MHz. The maximum absolute errors in the parameters
s
of the final model were 0.1 dB in the modulus and 1.2* in the phase with
r. m. s. (root mean square) errors of 0.04 dB and 0.4*. This final model
only.
The next device modelled was a vertical n-p-n transistor model which
5
operated in the frequency range 0.1 GHz to 1 GHz. Separate models were
produced for each of the common emitter, common collector and common base
became a task of tremendous volume and was only partially successful giving
Finally, a bias dependent model was produced for two similar bipolar
model was produced for both transistors and was valid for the linear oper-
the bias dependent elements were developed and good agreement with the
Thus, this modelling technique has been extensively tested and although
these tests have been with s parameter data for bipolar transistors, the
The majority of the computing work has been done on the CDC Cyber 73
easily transportable between the two. The main feature of these machines
is the word length which is 60 bits. This means that high precision can
6
'powerful' than the Leicester Cyber 73, thus for larger jobs where the
Chapter 6 required the CDC 7600, whilst the majority of the computing for
the bias dependent model in Chapter 7 was done on the Cyber 73.
with the GHOST library of graphical subroutines. The model circuit diagrams
circuit layout program written by the author and which used the GINO
library. Flow charts were also produced using the PDP-11/44, using
CHAPTER2
behaviour have been sought. Both circuit and device designers need
models that will simulate the performance of the transistor over a wide
based on the physical structure of the device in order that the effects of
On the other hand, circuit designers are more concerned with the accuracy
of the model and are less interested whether the elements in the model have
with computer aided design techniques to assist in the design and analysis
and its applications. Then some of the popular models are described
beginning with the classical hybrid pi model and going on to more compli-
cated models such as the Ebers Moll model and its variants.
8
may be either an n-p-n or p-n-p type. Block diagrams of both types are
reverse biased.
by using a single battery together with the voltage divider network con-
EC 71 Ec
IpIn pI npn
L 0
emitter IB collector emitter 'o"
coltector
B
base base
EC EC
B
p-n-p type n-p-n type
Figure 2.1. The Structure and Circuit Symbols for Bi polar Transistors
r T_
L
Ve Ib vc
Figure 2.2. Typical Bias Arran ent for an N-P-N Transistor Biased in
the Active Rer,,ion
10
V it
Vout
Vin
0- %%ý wo
The 0- u Ttme
Ic
F
out
Vid
F] ýý, f Time
Vout4
1ime
In the saturated and cutoff modes the bipolar transistor can be used
2.4. The two parts of the switch are the emitter and collector terminals
current ificrease to the base current which causes this increase, can be
the transistor in the inverse mode would be identical with the active mode
but with the collector and emitter interchanged. However, in practice
the emitter is more heavily doped and the collector cross-sectional area
is greater than that of the emitter. Thus the p-n-p transistor should
be more accurately described as a p+-n-p type.
A more accurate diagram of the construction of a planar n-p-n trans-
istor than that given in Figure 2.1 is now shown in Figure 2.5 . This diagram
will aid the identification of some of the physical elements in the models
described in the following sections of this chapter and is also relevant
to the modelling of the integrated circuit transistors described in later
chapters.
A full description of the theory and manufacturing processes involved
n+ emittercontact
p base contacts
n+ coltectorcontact
P+substrate contact
p+isolation channel
AP (base) I/
Ln/epitaxiat
p layer
collector)
ýýL4 n+
buried[aye
r
substrate
isolate the device from the rest of the integrated circuit. The p-type
base and n-type emitter are formed by a diffusion process and this is
paring the model with Figure 2.5 which showed the construction of a tran-
Wc
b C
rce
e
Figure 2.6 Common Emitter Ilybrid Pi Transistor Model
14
Of particular interest is r bbt" which Chua and Lin24 term the base
Searle27 describe how this element is inserted to allow for the transverse
voltage drops in the base region caused by the drift of base region
majority carriers flowing into the active region between the emitter and
Searle27.
measurements to determine the values of the elements in the model for any
particular case.
is also generally agreed that interlead and case capacitances are also
bc
Cb Cce
higher than 300 MHz, especially if some emitter lead inductance is also
included.
measurements which must be made to determine the element values for this
model.
The Ebers Moll model is probably the most popular non-linear model
for bipolar transistors. This model was first published by Ebers and
M01127 in 1954 and there have been numerous variations on this original
model published since then. This model is included here because, even
though the work in this thesis was concerned mainly with small-signal
linear models, the final problem involved a transistor which entered the
non-linear region. Also, the non-linear modelling problem and the Ebers
Moll model are too important to be ignored. Here the notation used by
the models.
A diagram of the basic Ebers Moll model, is given in Figure 2.8. The
in all modes of its operat -ion. The model is described by its reference
q Vbe
(2.1)
kT
es
Boltzmann's constant
A6soluke. +etr,?
er&ture
base junction
18
(2.2)
kT
IR=I e-
cs -11
Coupling between the junctions of the transistor is provided by the
base region and is modelled by the two current dependent current sources.
Thus
Ic=FIF-I (2.3)
aF (2.4)
F1aF
ol
R
R 1- aR (2.6)
GetreU20 shows how the linearised version of this model for operation
in the forward active region reduces to the d. c. portion of the small-
next stage in the development of the model is to take into account the
Figure 2.9, the basic EMILmodel has been transformed. Expressions for the
currents I,,, 12 and 13 are derived by Getreu. Around this 011 model a
number of elements have been added. The three resistors r bb" rc-
cl,
storage effects are modelled by (i) two junction capacitors CJc and
,
C which model the incremental fixed charges stored in the transistor's
Je '
space-charge layers for incremental changes in the associated junction
-I-
sub
r-- -
rcc Tc
S
III
J12
CDC Cj
r bb' ý
T C
Jljý I EM1 model
b 3T, i (transformed)
1 113
CDe-T
T, CJe
el i
ree/
which he calls the EM3 model. This model includes a further two diodes
and a junction capacitor and models the base-width modulation and varia-
tion of current gain S with current and voltage, and some temperature
effects.
eristics of the EM3 model but is still basically equivalent to the 013
model.
adhering closely to the physics of the device, introduces the new network
Ebers Moll, Linvill and Beaufoy Sparkes models are all similar in their
transistors.
21
For each of the models described in this chapter, at least one set of
element values for any particular model is to optimise the values of the
elements to give the best fit to a set of measured data. Probably the
suggests that time and effort spent on modelling even a complex model has
definite advantages, in particular that it saves the wasted time and
Thus, the author's aim was to present here a method by which models
complexity and accuracy required. It was also hoped that by so doing, the
CHAPTER 3
there are prescribed measurements that may be made to enable the calculation
tions that can be measured for the transistor can also be calculated for the
parameters. These parameters are described later in this chapter and are
related to each other by a set of equations also given later. The y para-
ted by Calahan4l and this approach has been favoured for many years.
SavageB states that the advantage's of this method include good convergence
of the network elements with the consequent easy and accurate formulation of
the derivatives. Also, using coefficient matching the minimum size of the
achieved.
stage.
This is not without precedent: Bassett37, one of the first to use optimisa-
3.1. Y Parameters
small-signal voltages and currents, which are shown in Figure 3.1, and are
y1 vi when V0
2
yIl
12 v2 when V, 0
(3.3)
! Z-
Y21 when V
vi 20
!
Y22 = VZ when V
2 .1=0
121-
v TWO-POTv =T
BLACKBOX 2
I
2
V, and V2are the port smaLL
signal voltages
11and12 are the port small signal currents
and 2 are the input and output nodes respectively, and denote, say, node
Current Law to each node in turn we can obtain the equations which in matrix
form give
I= (Y (3.4)
ind3V-
For a two-port network, since the internal nodes are inaccessible, there
cannot be any current injected into these internal nodes from external
negative of the sum of all the admittances connected directly between nodes
i and j, where in general terms the admittance Y between any two nodes
is given by
Y= PC +G+1 (3.5)
pL
the terms in the indefinite admittance matrix due to the controlled source
gM to element Y of Y
ik ind
gM to element Y of Y
it ind
(3.6)
gM to element Y of Y
jk ind
+gM to element Y of Y
jq. ind
reference node and is usually the earth node. At any particular value of
3.1.2.
28
Admittance Matrix
a total of n+I nodes, in which nodes 1 and 2 are the input and output
respectively, plus the earth node which is designated the reference node,
n
yijvj = Ii i=1, n (3.7)
written
A22
Yll 2'' 61122
Y12 A21
61122
(3.8)
A12
y
21 61122
All
Y22
1ý1122
They then describe how, for a purely resistive network, the y parameters
can-be obtained using Gaussian elimination and state that since 61,22 is
be seen that all the determinants involved are polynomials in the complex
and RLC networks. In their method any inverse powers of p are removed
p, ,i=O, m giving
pp2pm a A(p
000 0 0
2m
P, P, P, a A(pl)
(3.9)
pp2pm a A(p
mmm m m
The (m +1) x (m +1) matrix on the left hand side of equation (3.9) is
the Vandermonde matrix and this Cutteridge and di Mambro invert using an
coefficients.
12,, 45
Cutteridge and di Mambro also describe how the partial derivatives
efficiently computed.
current sources with time constants was received by the author. Time
introduce te PT into
constants, T, exponential terms the nodal admittance
xx l+T X2 X3
-5-1-+ 31
PT
thus replacing e by 1+ PT .
The author used this version of di Hambro's program J-n the early stages
there were some errors present, one of these occurring when inductors
were included. However, even on correcting this fault, the method was
the terms involving time constants was considered too inaccurate. Also,
it was felt that a more direct method using pivotal condensation, as des-
into the nodal admittance matrix together with the value of the complex
by Aitken43.
1,
nj in ,
Y
ij y j 1, n-1
nn
The flow chart in Figure 3.2 gives the computational procedure for the
this technique.
yG , I", ij PT (3.12)
ij +1+9me
ýLij -
PT
where 9e was the term due to a current source, which was included as
m
appropriate.
j2Trf (3.13)
Spence 42, the nodal admittance matrix should always have non-Azero elements
33
become very large compared with other elements in the matrix and that this
this it was necessary to scale the values by multiplying the frequency values
9
by 10- and the reactive by 109.
component and time constant values
Bearinj in mind the above points the author decided first of all, that
that the computers in use had a 60 bit word length this, it was felt, should
method, partly for reasons of consistency but also to avoid the presence of
Checks were then made using the circuit shown in Figure 3.3 at frequen-
cies of 0.5 MHz, 10 ?fflz and 100 MHz and with time constantS T on the current
cl
Ar C-'
1
2
0
Node 1- Input
Node 2- Output
Node 0- Ground
g, g (6-3) = 6.84E-01 S
T0 -0.01 ns , +0.01 ns
S.
V across nodes 3
p= j2rrf
00 0
-110
JwLj jwL,
jWT
0- (G2 +jwc +g w-r)0
0+G 0 3WL +jwc, +gle 2 1ej
3
(3.14)
100 jwc 0
-..JWL jwL -+G
--! +jwC -G
1111
00 jw (C +C? )+GlG2 0
-G 2-jWC2 -G, -jwC, 1
Using the notation of Section 3.1.2 and cancelling terms where appropriate
we get
All 0
CiR CI
A12
w l -LlL2L
+GI+jGR+
3
A21 0
(GIC2 + G-,Cj) CJC? GG
ý122 - j (3.15)
w,' LlL2L3 w LlL2L3
jWT
where R= Real part of g1e
I= Imaginary JWT
part of g1
w= 21rf
and f= frequency.
36
Agg
Yll
1ý1122
A21
Y12 0
61122
(3.16)
A12
y 61122
21
Ali
A1122
Table 3.1. Comparisons were then made between the values obtained using
significant figures that corresponded with those in Table 3.1 are given
compared were:
di Mambrols method
as appropriate
t-1 e te)
00 r- 00 ý4 r*I P-4 «e C 0
M
r- (M
CD \O r-1 m
1--4 (N t'n Ln
CD le
rlý
"
Glý e -1 1-4 00 00
r- "0 le o 00 (2) CD fli
m %0 Lr) r-i tn CD 00
t4) cli r--1 0 m ý4 %0
lqt cq r- r, 1 ci t- e C) CD
t- CO 00
le 1-4 Ln CD r_
M Ln M Ln
-4 Ln --1 cý
"l 00 CD -zr Co CD M oo
1-1 CD \o P-4 CD 'Z --. i CD LA
L', -4
4. -1 r-ý Lt-i -4 Ln
r-i r-- Ln
ýi 0 ý Z (ý "ý c;
';
\o
-e %0 00 m
C% CD m CD c71 %0
P-4
%0 rq ei LA -4 r41 (N tn
4 00 (111 cý M Ln 00 -t cn
tn Itt t- ýo tn r, 4 r-- CD C%
" %0 Cý r- -4 tn CD LI- CD
nt "q
er) 00
t41 00 t41 r, 4 00 00 t41 00 0
P-4 00 r4 ýI r, 1 t-' f.-4 qt it,-
\C tn rq e (N e %0 Lr) r_
Ln m CD ul Cý m Ltli (M CD -, -4
\C m r- 1,0 C) \JD \O m r_ 4-4
c; lý
u
., 4
4.4
. r4
0
CD
m CD CD Ln r- r-
r- rq r- tn M P-4
let 00 00 Cä tn "4 Ul %0 CD
rq - Izt cý Lt') tl rq C% 1-
CD P-4
le (M M-4 Co \o gt N
Vý CD Ln e 00 t-
e Co CD e cý e P-4
rq 00 (71 rq -4 -4 r, 1 Ln r-
00 00 tn 00 NI CD Co Izt bi
cq Izt le r-4 00 c) (N cý 00
t4ý e en tn e CD tn " Ln
le Ln P-4 ýr -e ul 1-t le r-
Ln Ilt ul e-i
CD 0
-i t-ý rn r, In In 44
r-, rn r. ' r-,
0
+ + + + + + + + +
Ln Lri 00
rq IKT, M et 00
00 tn CD vj P-4
Ln P--1
\JD tn Lt)
r- 10 nt (ýq P-4 (N rq cz
le Lti le
Ln 0) Ln te gt
(14 %M cý 00 rý
r- vj "-1 %0 (-4 tn e
-I Lr) bl CD t4ý CD ei t-- 00
CD -4 00 CD \Z Ln (Z e 0
rn \Z (N tn 00 \JD tn tn c
00 \o le 00 Ln 00 00 r-- CD
r-4 Co F-4 CN 00 rý (N 00 Lf)
cý oh cl- 0
.0
Ln (Z CD LA C CD Ln CD (Z
; ; ; ; 7
zi c c c c c
(Z C
38
P9 13 14 13 14 14 14 14 14 14 14 14 14
M 14 11 14 14 14 12 14 14 14 14 13 14
0
M9 14 11 14 14 14 12 14 14 14 14 14 14
m9s 13 9 13 14 13 11 13 14 14 13 13 14
P9 13 10 9 10 8 7 7 7 6 4 5 5
0.01
M9 14 10 14 14 14 11 14 14 14 13 14 14
P9 13 10 9 10 8 7 7 7 5 4 5 5
0.01
M9 14 11 14 14 14 11 14 14 ý14 13 14 14
1
The results using method P are not shown in Table 3.2 because the
errors were so large that none of the values corresponded even to one signi-
As can be seen from Table 3.2, with T=0 method P9 was the most
had a major effect on method P, there was only a very slight improvement
when they were applied to method M. However, the differences between the
single and double length arithmetic versions of methods M9 and M9S were
more marked although M9S might still be considered accurate enough for
most purposes.
Taking the two most accurate methods, P9 and M9 for further compari-
,
sons with time constants T of -0.01 ns and +0.01 ns , it can be seen
from Table 3.2 that at 0.5 MHz method P9 has lost some accuracy and that,
is slightly more accurate when no time constants are applied to any current
applied.
In all cases the computed values for y 12 and y 22 had absolute errors
10- 20.
of less than
For the given analyses, method M9 took 90% of the computation time
savings in calculation time and storage space make it the best choice.
3.2. S Parameters
because ok the difficulty in obtaining good short circuits and because short
that they describe the inputs and outputs of a two-port black box, but have
the advantage that they are easily measured up to the GHz frequency range.
incident and reflected parameters (a,, bl) and (a., b. ) which are shown
1
[a -
vrz, + rz-
0
I,
0
2 rz 0
0 (3.17) '
12
+YZ0 12
/Z--
0
12- r-
b2 'ý '(Z 12
rz- 0
,
0
b, = s1la, + Sl2a2
(3.18)
s a1S a2'
21 22
12
2
vi
TWO-PORT CL2
V2 z0
bl b2
BLACKBOX
Fig. 3.4 Incident and Reflected Parameters Used to Define the s Parameters
41
circuiting the input or output port and taking the appropriate admittance
b,
s when a2ý0
b
s -L when a=0
12 a21
(3.19)
s221 when a0
a12
s222 when a, 0
a2
and incident voltages at the input and output ports when one port is matched.
zero may be achieved by using SOQ transmission lines to connect the device
under test and then terminating the transmission line in its characteristic
impedance, Z0 in Figure 3.4. A measurement jig is described by Hewlett
Packard 49, incorporating the 500 transmission lines and SOQ loads in
such a way that the device under test is easily reversed thereby enabling the
with a Hewlett Packard 8405A vector voltmeter which is used to measure the
this chapter and then applying the conversion formulae given in reference
12 dy
(3.20)
2 y2l
21 dy
(1 + Yll)(1
-d Y22) + Y12Y21
22 =
y
2 sig
Y12- ds
(3.21)
2 IZ1
Y21
dt,
(I + S22) + Si2s2i
Y22 ý-
ds
The y parameters given in equations (3.20) and (3.21) are all norma-
ij yii
z01,2;
j=1,2 (3.22)
43
CHAPTER 4
NUMERICAL OPTIMISATION
most basic, optimisation can be described as a means by which the best value
application but usually some error function is defined* which indicates the
or more variables and the value of that parameter currently available. The
most problems there are a number of criteria, therefore the sum of squares
function
F(x) = (4.1)
The sum of squares function is also useful in the solution of nonlinear sim-
then it can be seen that a zero minimum of the sum of squares function would
There are numerous other formulations that are suitable for an over-
all error function. Another popular one is the minimax function which
where f1 (x), i=l, m are individual error functions. The general term
equation (4.1) is when F(x) =0, although a zero minimum may not necess-
then a global minimum has been found since F(x) cannot be negative for
F (x ) F (x) (4.4)
ý<
not least because of the uncertainty of knowing whether the lowest minimum
found so far is the lowest minimum overall. Dixon, Gomulka and Szeg650
with regard to solving global optimisation problems is very poor. The most
popular methods are random methods, including multistart methods where local
ing points, and those based on Branin's trajectory method, reviewed by Dixon
values may occur at any values between and +- In many real optimi-
-to .
sat ion problems, however, there are restrictions on the acceptable solution
45
xm<M! (4.7)
where m= 2(n - 1) .
Thereafter each further term improves the accuracy of the approximation but
it could be argued that only an infinite number of terms will give the true
value.
to get a true match. Therefore, in this context a good model should contain
46
only those elements that h&ve a, majewafficttogether with those that make a
sation methods that have been used in similar applications to those of the
author and secondly with the optimisation method actually used by the author.
categories.
searches are more systematic. A simple pattern search was used by KnudsenS5
of this method are its simplicity and its suitability for small computers.
narrow, curved valleys in the function. The simplex method of Nelder and
ted or reflected about a vertex, the process being repeated until a suitable
trical circuits, although they noted that their method was limited due to
51#52
the large computing time required. The global search algorithm of Price
best points currently available. This algorithm was used by the author6O
k+l k9k
x = xk (4.8)
however, a variation on the method was used by Agnew" who, for the optimisa-
tion of filter circuits, used only active residuals at any particular stage
which was used at the initial stage in a two-part algorithm for the solution
denote the iteration number and T to denote the transpose of a vector, the
k+ 1k k+l
x_ Xky
(4.9)
k+l kkk
y9y
and ak is given by
k
a0 for k=O (n+l), 2(n+l)
,
(ak-1) otherwise
ýk
by both Krzeczkowski6 and Savage8 for the initial stages in the synthesis
F (X + 6) =FW+6Tg+... (4.11)
Ignoring higher order terms, differentiating and equating the first deriva-
0g+ H6 (4.12)
along the direction given by 6, thus the Newton Raphson iteration becomes
k+l k) -1
=xk-Xk (H g (4.13)
The Newton Raphson algorithm thus requires the evaluation of the Hessian
ithms which attempt to approximate equation C4.13) using the first deriva-
tives only. Among these algorithms are those of Davidon 64 and the later,
k+l
s 6khk (4.14)
k k+l k
where x-x
k+l k
hk
A variable metric algorithm was used by Lanca and Nichols" for the design
Powell was used by Orlik-35 to obtain some of the element values in Gummel
i-i
where fi Cx) ,i=l, m, are the component functions dependent on the vector of
tion and its derivatives, there are also available the values of the compon-
then taking the Taylor series expansion of each component function gives
n afi(x)
(X + 6) =f
fi-i-j11 (x) + 6. + i'm (4.17)
axj 3
f (X + 6) =f (x) + j6
16
.
Newton iteration
k+1 k kT kkTk
xx_ Xk (P )i (i f (x (4.22)
k) k)=0
Dimmer" shows that provided a point at which (J f(x has not
k)ik
been reached and (J is non-singular, then there will be a value of
X>0 which will reduce the objective function. Dimmer also emphasises
the differences between the Gauss Newton and Newton Raphson methods and
notes a superiority of performance by the Gauss Newton method over the Newton
networks.
Most other methods designed for minimising sum of squares functions
are based on the Gauss Newton method and are concerned either with avoiding
method or with the use of approximations for the Jacobian matrix thus
It can be seen from section 4.1 that just as no single model of the
algorithm has become the standard even for applications of a similar nature.
It was mentioned in section 4.1.1 that the author used the global
search of Price to obtain the starting values for one of the trial problems;
however it was the author's intention to, initially, use the two-part algor-
62A 10 for
ithm of Cutteridge and Henderson the major part of the model opti-
misation. At this early stage the reason for this choice was largely
52
The author soon discovered that the gradient descent section was not in fact
needed in her particular application since the Gauss Newton section proved
this second section from the modelling point of view also became apparent.
Therefore, this Gauss Newton algorithm was the major optimisation method used
Should this not be achieved, the preliminary search continued until the pre-
slowness of convergence onto the minimum through taking new points for eval-
uation too close to the current three points used in the quadratic interpola-
tion.
The first of these, which was not really relevant in the author's application,
concerned conditions when it was not advisable to attempt using the algorithm.
The second concerned successful termination which was based on the absolute
category described criteria for failure due to the JTJ matrix being or
53
almost being singular. The fourth and final category predicted ultimate
CHAPTER S
Philips Research Laboratories (PRL) who provided data for the author
specified certain conditions for their models. One of these was that the
s parameters be fitted directly. The reasons given for this request were
that (i) these were the quantities actually measured, and (ii) operation on
formation formulae would alter the weightings of the fit. PRL also provided
models which they thought would provide a suitable starting point for a model
optimisation exercise.
Thus, this chapter first describes how the component parts of the
three previous chapters were combined together to form a single method des-
igned, initially, to calculate the optimum element values for any particular
to the conditions set by PRL. Details are then given of the possible ways
in which models may be improved and of the methods chosen by the author.
At each stage examples are given of the results obtained for the
first model developed for PRL, that of a lateral p-n-p transistor. The
It should be noted that although the major criteria for the final modelling
algorithm were developed during this first exercise, the process of refining
and confirming the techniques continued throughout the period of research and
The first set of data provided by PRL was for a lateral p-n-p
transistor connected in common emitter configuration and consisted of the
S S12 S S
Zo -, 5006 11 21 22
It can be seen that the measurements for each of the s parameters are in
terms of modulus in dB and phase in degrees and are given at four frequen-
cies in the range 0.5 DIHz to 10 MHz except for s for which a measurement
12
at 0.5 MHz was not available. The accuracies of these measurements were
given as "about t O. S dBII for the modulus and "phase angles near 0" are
they suggested would be a good starting point for the derivation of a model.
This model, which is shown in Figure 5.1, consisted of four nodes (including
the node 0) and five elements of which two elements were voltage con-
take specific values with zero time constants. These values are given in
the diagram. It was felt that this model was too basic and so a further
three elements were added by the author. The global search technique of
51,52
Price was then used to minimise the single overall error function des-
cribed in Section 5.2.1. in order to obtain suitable starting values for the
unknown element values. This model, which is shown in Figure 5.2, was then
used as the init-ial trial model for the lateral p-n-p transistor.
S6
1 2 No. oF nodes =i
c No. oF elements
(2-3) = 3.53E-03 S
(T :- 0)
(V across nodes P
(0-3) = 2.07E-05 S
(T = 0)
(V across nodes 31)
Node I 9ase
Node 2 Collector
Node 0 Substrate and Emitter
2 No. oF nodes =i
1 No. oF elements
C.
b
g (2-3) = 3ME-03 S
(T = 0)
(V across nodes 3 1)
(0-3) = 2.0? E-05 S
(T = 0)
(U across nodes 31)
G (1-3) G. 12P--10 S
C (1-3) B. GSE-02 nF
G (2-3) S. 22E-OS S
C (2-3) S. OGP--04nF
C (0-3) G. 77E-03 S
C (0-3) 3.71E-06 nF
Node I- Base
Node 2- rollector
Node 0- Substrote and Ernitter
trial model as shown in Figure 5.2, the first requirement by the optimisa-
tion of how good the model was. Since the author had chosen the Gauss
the objective function were also required. The element values of the model
were obviously the variables, but it was necessary to decide whether any con-
obtained when a local minimum had been found by the optimisation algorithm.
Therefore, the exit criteria of the Gauss Newton algorithm are also discussed.
discussed in Chapter 3. The method first used by the author, and used for
the examples given in this chapter, was to calculate the y parameters using
the method of dillambro described in Section 3.1.2. and to then use the con-
The calculation of the yý parameters was later done by the method of pivo-
tal condensation as described in Section 3.1.3, and all the examples in the
the given frequency values for each of the s parameters. There was a
real and imaginary parts of the s parameters or whether the modulus and
former was that by weighting equally the errors in the real and imaginary
58
for errors in the modulus in. decibels and in the phase in degrees could be
errors in terms of decibels and degrees and that by allowing the weightings
The weightings chosen by the author, based on discussions with PRL, were such
phase. It was felt that absolute errors were more appropriate than rela-
tive errors, particularly in the case of the phase errors where the paradox
large relative error while a similar absolute error in a phase angle close
problem occurs for the modulus when it is stated in decibels. Since the
there is negligible gain and many engineers would feel that an error of ldB
the weighted absolute errors would be suitable for both types of individual
error functions.
sum of the squares of the weighted individual errors in the modulus and phase
the initial problem was one of minimising a sum of squares function made up
When analytic first derivatives are not available as was the case
here, then numerical estimates of the Jacobian matrix can be used. The
afi (ý) / axia displacement vector dx is defined such that dxk = d. x k=j
y
and 6xk0O, k ýj, where d is some small value, then
some results were obtained which proved that the value of Sx recommended
I
by Henderson" gave sufficient accuracy. This value was given by
7
2 6x 10- (1 + Ix 1) (5.2)
3= i ,
which was based on the reasoning of Gill and Murray70, who suggest a fixed
small value governed by the word length of the computer in use, assuming
Ix to be of the order unity, but which also allowed for the possibility
iI of
Figure 5.2, this provides a further reason for the use of the factors
can be reduced if forward differences are used since the current function
reduces the accuracy of the estimates of the Jacobian matrix, Brown and
60
minimisation.
take physically feasible values, most engineers find models more acceptable
one of the initial models provided by PRL, who were most insistent that
model elements should have some physical significance, which included a nega-
then A, negative value of g. would merely indicate that the current was
cating a time delay. However it was found that there were many instances
where a small positive value was the optimum value for a particular stage in
the development of a model. The author decided to allow the time constant
to take both negative and positive values since it could be that a positive
value at any particular stage merely compensated for elements which were not
yet present in the model. Certainly there were cases where a positive time
gressed. So that the same constraints could be applied to all the variables,
in this situation, the logarithmic transformation has been used with some
the one aimed for was the successful termination characterised by all the
this occurred then one could say that the elements were all necessary as
part of the present model and that the optimum element values-had been ob-
that the model was not suitable and that it should be modified. It would
then become necessary to study the model, the element values and the reas-
in Figure 5.2, the algorithm failed to converge. The reason for failure
was that there were too many successive increases in the rate of increase of
the maximum modulus correction. This failure mode, which we will refer to
62
as failure mode 1, was in fact the most common mode encountered in all the
ations before the optimisation algorithm terminated and in this example the
10 11
6.12 x 10- S to 9.91 x 10- S while the correction term in the logarith-
mic domain had changed from -8-01 X 104 to -4.95 x 105 . Meanwhile, all
the other element values were virtually unchanged. Taking this progression
model and restarted the optimisation algorithm. Again the algorithm failed
in failure mode 1. This time the element causing the failure was C
3-0
which was increasing in value. It was decided that this element should
current generators, both of which had been allowed to vary, caused a further
generators in the model it was felt that removing one of them would not make
the model invalid. However, there was still the question of whether the
generator should be connected with the current flowing in the opposite direc-
reversed, its value still tended to zero, therefore it too was removed from
was obtained.
The new model is shown in Figure 5.3. It can be seen that the
optimised values of the remaining elements are quite similar to their original
values, the largest change of about 20% being in C At this stage the
2-3*
overall error function had been reduced from 384.2 to 114.9. The individual
errors are shown later in Table 5.2 and it can be seen that the maximum absol-
errors have been reduced from 0.55 dB to 0.43 dB in the modulus and from
ute
I 1. ISE+02
b No. oF nodes =4
No. oF elements
g (2-3) = 3.94E-03 S
(T = 0)
(V across nodes 1)
C (1-3) = 1. OSE-01 nF
G (2-3) = i. 67E-OS S
C (2-3) = 3.90E-Oi nF
G (0-3) = 6.63E-03 S
Node I- Base
Node 2- Collector
Node 0- Su6strake and Emitter
better than one with more elements could be asked. The author's explana-
tion of this would be that firstly, a model containing more elements would
only be better than one with fewer elements if the extra elements were of
the right types and values and in the correct positions, and secondly,
the original element values in this example had been optimised and two of
the elements had optimum values of zero, while a third had an optimum value
of infinity.
functionp it then became necessary to consider how the model could be '
improved. It would not be expected that the removal of any elements from
action. There are many strategies that could be employed for the addition
we should avoid adding extra nodes, Therefore we should first of all con-
tions might be that although we could attempt to replace any current genera-
tor that has been removed, we should not insert any additional generators,
and that we should not insert inductors other than at the terminals since
the most likely physical reason for the presence of inductance would be
lead inductance.
those elements for which some physical explanation for their presence could
be given. Certainly, this is a popular strategy and one which Nabawi and
model at an arbitrary value based on the values of elements of the same type
algorithm and checking the result. If the algorithm converged and the
overall error function was reduced, then the new element was left in the
model and another new element was tried. If the algorithm failed to con-
verge it was generally because failure mode 1 occurred. Usually, this was
caused by the corrections to the new element value. If this was the case
or if one of the other failure modes occurred it was assumed that the new
element would not make a suitable addition. There was one case, which can
65
be seen in the example in the next section, where failure mode 1 was caused
by one of the existing elements in the model, In this case the new elem-
S. S. Example of Results
After failing to replace singly each of the elements removed from the
in which they were added is shown in Figure 5.4. It can be seen that the
tion. These values too, were reasonably close to the optimum values cal-
to converge onto the local minima. The value. of the overall error func-
tion F is given at each stage and it can be seen that it was reduced from
pointed out that the function being minimised at this stage included the
function.
external nodes. The two successful attempts are shown in Figure 5.5. The
first of these was originally inserted at a value of 4nH and was optimised
66
2F=i. iSE+02
C. No. oF nodes
No. nF elements
(2-3) = 3.91E-03 5
(T = 2)
(V across nodes 13)
C (1-3) i. OSE-01 np
C (2-3) 4.67r--QS S
C (2-3) 2.90e-gi nF
G (0-3) G.G'jE--03 F
(S
(I Q
F
b No. nF nodes ai
No. oF elements =6
o (2-3) = 6.35S-03 5
(T = 01
(V a,:ro.; s nodes 13)
(1-3) =I. 99p-ol np
G (2-3) 4..jgE-GSS
C (2-3) nP
r1 (0-3) 4,91E-03 S
C (1-0) 2. ýSF-92 nF
f5
F-S. SGE+gl
No. nP nodes m4
No. oF etements z7
9 (2-3) =i . 36F-02 S
(U across nodes 1)
C (1-3) 21.47E Ol nP
G (2-3) = 4.29F--OS S
c (2-3) - 4.02E-04 nF
G (2-3) - 3.23E-03 S
C (1-0) a 3.20E-02 mp
G (1-9) = 2.1! E-94 S
F=S. GSE+0I
No. nF nodes =I
No. nF elements a8
Made i- Base
liode 2- Collectm-
Node a- Substrate *n4 Emitter-
Figure 5.4 Addition of Elements Between Nodes 1 and 0 and Nodes 2 and 0
b7
2 F=i. 95E+gl
C
No. oF modes =S
No. oF elements
9 (2-3) I. IGE-02 S
(T=
(V across nodes 13)
C (1-3) = 1.53F-01 np
G (2-3) = 1.3sr--OS S
C (2-3) = 4-19r--04 nF
G (4-3) 3.14E-.-03 S
C (1-4) 3.42E-02 nF
G (1-4) 2.27E-04 S
C (2-4) i-SOE-03 nF
L (4-0) 1.27E+01 nH
0
s
I 2 F :-2.2SE+01
b c No. oF nodes =6
No. aF elements
9 (2-3) = 8,79E-03 S
(T = 0)
(V across nodes sl)
C (S-3) 1.31E-01 nF
C (2-3) I. ISP-OS S
C (2-3) 4.04E-Oi nP
G (4-3) 3.33C-03 S
C (5-4) 2. G?E-02 nF
G (S-4) 1.2se-04 S
C (2-1) 2-IOE-03 nF
L (4-0) G. 4SE+Oj nH
L (5-1) 3. S2E+02 nH
0 So
factors of up to about two, but when the second inductor was added the
The overall error function had now been reduced to 22.45. At this point,
After correction of the error the second model in Figure 5.5 was rý-optimised
and in addition to changes in the element values, the value of the overall
had been reduced to 9.70. This model is shown in Figure S. 6 and it might
have been considered that sufficient improvements had been made. However,
one last attempt was made to replace the current generator that had been
removed from the initial model. This last attempt was successful although
the current was flowing in the opposite direction to that in the initial model.
The values of many of the elements also changed quite considerably as can be
seen from Figure 5.6. However, convergence of the Gauss Newton algorithm was
achieved without difficulty and the overall error function was reduced to 6.13.
The individual errors of this model are shown in Table 5.2 and most
1 S
b No. oF nodes =6
No. oF elemenLs = 10
9 (2-3) - 7.9SE-03 S
(T = 0)
(V across nodes sP
C (S-3) = I. IBE--Gl np
0 (2-3) = i. 17F-05 S
C (2-3) = 4.20E-gi nF
G (4-3) = 3.48G-03 S
G (5-4) = 1.29E-04 S
C (2-4) = 2. SGE--03 nP
L (1-0) = 8. SSE+01 nH
L (S-1) = S. 99e+02 nH
C (1-4) = 2. SOE--92 nF
5
1 G. 13E+00
No. oF nodes =6
No. oF elements
9 (2-3) -- I n; --04 S
(T = 0)
(U across nodes '3)
C (S-3) = i. 9GE-03 mp
G (2-3) = 4.43E-OS S
C (2-3) = 8.19E-gi nF
G (4-3) = 2.? iE-03 S
G (5-1) I. ISE-0i S
C (2-4) 2. Gie-03 np
L (4-0) I. SSE+02 mH
L (S-1) 1.9"E+03 nH
C (1-i) 3.23e-02 rip
0 9 (3-4) 2,72ý-03 S
(T =V
(V across noces 13)
Node I- Base
Node 2- Collector
Node 0- Substrate and Emitter
Having developed a model in a trial and error manner it was felt that
was suitable or not, but it was extremely time consuming to do one batch com-
puter run for each possible stage in the development of a model. Unfortun-
ately, the program was too large and slow to permit on-line execution.
model improvements, rejecting those that were unsuitable and retaining the
possible. However, it was felt that each intermediate stage should be stored
so that it was possible for the user to choose a model from any particular
stage, and perhaps modify it, before manually restarting the modelling
process. This way, the model development would not be inhibited by the user's
prejudices, although the user could still exercise some control over the
71
model would be suitable or not and techniques for the addition of elements
and nodes to a model were also developed. Some details of the finaI mod-
stages. The first stage consists of taking the initial model and attemp-
ing elements from the model until a minimum is obtained. The second stage
involves the addition of new elements within the existing nu! nherpf, nodes. since
we assume that we want as few nodes as possible. The third stage is for
the addition of a new node. Stages two and three can be repeated altern-
The criteria for failure are more stringent than those described
in Chapter 4, in particular, the two most common reasons for failure are
iterations and the mode we could now call failure mode 2, is deemed fatal if
obtaining convergence.
Whichever failure mode occurs the next step is to examine the vector
elements it might be best to remove from the model. Throughout all the
earlier experimentation it was found that removal of the elements having the
72
remove from the model the element giving the maximum modulus correction and
to then restart the Gauss Newton algorithm, generally continuing from the
removal of a node and involves the removal of any other elements which are
inductor would. normally occur only at one of the external nodes, elements
provided that at least one other element is in parallel with it. If there
short-circuited.
73
contain only one current generator. It was felt that if there was a sit-
uation where removal of a current generator was the only remaining option,
the basic scheme is to remove only the element causing the maximum modulus
correction and to then restart the Gauss Newton algorithm, it was found that
when there were a number of elements causing large modulus corrections, the
removal of several elements at one time could save on the number of failure/
removal cycles. The criterion devised for multiple element removal was
16ij j+2
that all those elements whose modulus correction > 10 , where j is
the average of the positive powers of 10 of all the corrections 6 (in the
satisfy this criterion then only the element having the maximum modulus
correction is removed.
remove the element having the second highest modulus correction. If this
is successful then the remaining elements are given their values at the
abandoned. If this should happen then the best course of action is for
the user to first check the initial model for errors and to then try a
different model, perhaps based on the initial model but containing some
additional elements.
74
lGauss Newton I
Yes
ocaL minimum SuccessPuL exit
i
No
> CLIM
Yes
N-o
Y At Least one
cessFuL removaL
No
--
essPu Lýýo FaiLure exit
Yes
function, F the second stage may now be entered. The aim of this stage
min'
is to add extra elements into the model without increasing the number of
nodes. The author decided that only elements of capacitance and resistance
decided that inductors should only be added with a new node at Stage 3.
The author also felt that users were likely to have strong ideas about the
model, and their positions. This factor, coupled with the problems of
The user always has the ability to modify a model at any particular stage
had developed expressions for the optimum values for the placement of vir-
one based on the Gauss Newton corrections of virtual elements. The author
are systeriatically made to insert the new element between each pair of
nodes where an element of that type does not already exist. If the addition
Figure 5.9 Scheme for Stage 2
78
is successful, the criteria for which will be given later, then the element
remains in place, otherwise the model retains its former topology. These
istor has been added in one complete cycle. The reason for these repea--
ted attempts. is that it was found that even if a particular element has
when the model topology has changed slightly. Indeed, in Section 5.7 -
there is an example where an element was deleted from a model at one stage
possible placements and to only add the best of these, followed by further
cycles of evaluation and addition of the best until no more elements could
be added. However, the author felt that this latter option would be
more laborious and that the same final topology would be obtained in most
cases.
Obviously, with so many possible placements being attempted, the
inserted into an existing model, the moduli of the Gauss Newton corrections
to be applied to the existing elements were usually much smaller than the
was capatilized upon by the scheme devised for element addition summarised
in Figure 5.10.
Thus, when attempting to add a new element, first the Gauss Newton
corrections are calculated. If the new element does not give rise to
the maximum modulus correction, then the Gauss Newton algorithm proceeds
L42
Does new eeLement have
0-
max
m correction
xmmod col
No
No
Yy
ess
CCC Correction
Orr
Ensure
sm mod C4 10
AppLyjC C to new eLement
RecaLcuLate
Gauss Newton corrections
YY
ee
ss
s it aR
ILaR
tending 3 to zero
No
Yes
Yes
Remove
appropriate eLement
Continue
Gauss Newton
No
iLed du
ýFaiLed duee to
tc0 2 eLements
: onverged .,:,. ->-< aLready removed
-wL new eLement
eLem
ment
No N
No and F> FMIN es
Yes Yes
Y s
Is iaR
it aL
tending to zerro
z
Yes
No
< O. SGxFMIN FaiLure
exit
I-I
ýSuccessPuL
exit
is no longer the one generating the maximum modulus correction then the
0
Gauss Newton algorithm is continued normally. Similarly, if the mod-
correction is of the opposite sign, then again the Gauss Newton algor-
same sign and its absolute value is greater than before, then it is
then the removal of the node that this suggests is permitted and the Gauss
assumed that the new element is unsuitable for addition into the model.
By applying the Gauss Newton correction to the new element in this way,
quickly, and to aid more rapid convergence of the Gauss Newton algorithm
It has been found that the Gauss Newton algorithm is very reliable
even when the starting values of new elements are quite inaccurate. Thus,
with this additional aid to rapid convergence, the values at which new
elements of capacitance and resistance are added, were chosen quite arbi-
trarily. It was felt that the new elements should be of such values that
they would not have too great an effect on the existing model. Therefore,
the initial value chosen for new elements of capacitance was lpF and of
the time of the addition, but it was felt that this would not offer any
improvement.
Gauss Newton algorithm onto a local minimum has been obtained. These
cases are;
If the Gauss Newton algorithm converges at a new local minimum, the new
than Fmin' the value for the currently accepted model. This is to ensure
that the model contains only those elements that make a significant contri-
If the model is still not accurate enough then the next step is to attempt
element.
Savage found that the most effective technique was the last of these, where
negative value then a T-network, based on that element type and its corres-
the author felt that each of these methods which, apart from (iii), caused
the generation of at least two extra elements, ignored the most likely
rather than disturb the internal structure of the model, new nodes should be
developed at the external nodes. This way, the addition of a new node
can be achieved by the addition of only one new element as in the examples
again involves the addition of only one new element, is to insert a new
5.11, is based on the addition of only one element. Since we assume that
the evaluation of each possible placement in order to select the best new
node and element. The reSt-Lictions on the placements of new elements can
STAGE 3
Inductor at externaL
node I Yes
No
NEXT I
JELement I oP type I
R 6 LI IC g
Attempt to add C
at the
Attempt to add L non-externaL node
iP at externaL node
Attempt
to add R
Attempt to add R
lAttempt to add CI at the
non-externaL node
NEXT I
ýSuccessFuL exit
tor connected.
in series.
6. A voltage controlled current source may not have any element added
in series.
model.
2. The new elements are only added on one side of any particular ele-
Attempt to add
eLement B in series
with eLement A
No
FNýo
onverged
Yes
Remove appropriate IF
eLement
based on the Gauss Newton algorithm. The initial values for new elements
one order of magnitude smaller than at Stage 2. The initial value for
this stage and the only early prediction of failure is if the absolute
to the new node, calculated by Gauss Newton, exceeds 1000. Otherwise, the
converge then, provided the failure to converge was not due to either of the
elements connected to the new node, the appropriate element is removed and
finds a new local minimum that is lower than F the best minimum obtained
min ,
before entry to Stage 3, then the addition is said to be successful.
user and the best of these is then used as the new model. Although the
user may decide at which stage the model should be accepted, the author's
that the maximum benefit be gained from any particular number of nodes.
new node. However the author has found that the addition of a new node can
"
affect the whole model and that, therefore, it is better if all permitted
section, after the models described in Chapters 6 and 7 had been developed,
the first modelling exercise, to model a lateral p-n-p transistor, was re-
peated but this time using the final version of the modelling algorithm
instead of the trial and error method used to produce the models in Figures
From the initial model shown in Figure 5.2, the algorithm proceeded
to remove elements G and C3_,, and then stopped with an indication that
1-3
gO-3 should also be removed. On removing this element manually and re-
starting the algorithm, it then went on to obtain the model shown in Figure
5.3. Continuing from there into Stage 2, the algorithm added the elements
Cl-01 C2-0 and R1_0. This model was thus the same as the last model in
Figure 5.4 even though the element addition was in a different order to that
the element values and the overall error function value were slightly differ-
ent and the new value of the overall error function, F, was 50.97 compared
with 56.50. The algorithm then entered Stage 3 and the new element and
,
node added was L 4-0* Thus the algorithm had produced the same model as the
first model shown in Figure S. S. The values of the elements were, here,
quite different from those given in Figure S. S due to the errors in the
analysis routine and the value of F, which had not been recalculated in the
P=3.84E+02
No. oF nodes =4
No. oF etements =8
9 (2-3) = -; IE-03 3
0),
,T=
(V across nades 13)
(0-3) = 2.07E--95 S
(T = 9)
(V across nodes ',)
G (1-3) = S. 12E-ig S
C (1-3) - 9. GSE-02nF
G '2-3) = S. 22E-OSS
S C (2-2) 2 S. OGE-Oinp
(I
e G (9-3) = 6.77E-03 S
C (0-3) = 3.71E-06 np
Delete G( 1-3) C(2-9) 9(0-3)
Add C(1-0) C(2-9) R(1-0)
Add new node with L(4-0)
F=1.3SE+91
No. oF nodes =S
No. oF elements =9
C( 1-3) =I . 2GE--9l nF
6 (2-3) = i. 64E-03 S
C (2-3) x 3.62E-01 np
9 (2-3) = 9.23E-03 S
(T = 0)
(V across nodes 30
G (4-3) = 3.27E-03 S
C (1-4) w 2.97E-02 nF
C (2-4) !. 39E--9i nF
R (1-4) 7.76E+93 ohm
L (0-4) i. 28E+02 nH
F=7.3GE-+Q9
No. oF nodes -6
Mo. oF elements = 11
C (1-3) I. ISE-91 nF
6 (2-S) 8.28E-96 S
9 (2-3) 9.14E-03 S
(T = 9)
(V across nodes 1)
G (4-3) = 3.42E-03 S
C (1-4) = 2.8GE-02 nF
C (2-1) = I. G?E-94 nF
R (1-1) = 9.29E+03 ohm
L (0-4) = i. iSE+02 nH
C (1-2) = 3.94E--13i np
C (2-0) = 2.21F-93 nF
C (S-3) = 3.20G-03 nF
e
Node i- ease
Node 2- Collector
Made 9-S. i:,atrQte, "a Emmet
that the change in the value of F was actually from 13.50 to 12.27.
In the final modelling algorithm, after one successful node addition, Stage 2
gave a very small reduction in the value of F and the new node selected by
series with G2-3 , and gave a value of F=7.36. This sequence is shown in
Figure 5.13 together with the additions made in the final entry to Stage 2.
development.
V, I S.??E+00
No. oF nodes =6
No. oF elements = 12
C (1-3) = 1.19e-01 nF
G (2-S) = 4.63E-05 S
9 (2-3) = B-47E-03 S
(T = 0)
(U across nodes 13)
G (4-3) 2.98E--03 S
C (1-4) 2.99E-02 nF
C (2-4) 1.18E-04 nF
R (1-4) B. iOE+03 ohm
L (0-1) 4. IOE+02 nH
C (2-0) 2. iGF:-03 nF
C (S-3) 2. S7E-01 nF
'e C (2-3) 3. S3C--Oi nF
R (4-S) 2.44E+03 ohm
Node I- Base
Node 2- Coliector
Node 0- Substrate and Emitter
The values of the s parameters. and the errors are given in Table 5.3.
Although the value of F is smaller for this model than for the
final intuitive model in Figure 5.6, these values being 5.77 and 6.13 res-
pectively, the maximum absolute errors for this model are 0.08 dB and 1.24*
compared with 0.13 dB and 1.03*. Thus the errors are again comparable with
In the author's opinion both of these models are equally valid and she
would suggest that the original model proposed by PRL contained two voltage
However, Callahan's model was for the lateral p-n-p transistor operating in
91 1
the saturated region and therefore, probably not applicable to the set of
data provided by PRL. This could explain why the second current generator
current generator to a model needing only one, a model of much greater com-
that although the model in Figure 5.14 has one element more than that in
The time taken to produce this final model was approximately 50 min-
utes of c. p. u. time on the Leicester University CDC Cyber 73. However, the
point should be made that the author's intention was to produce an effective
method for the development of models and that the computing time was not con-
for means of reducing the computing time, but with the present trend of fas-
ter computers constantly being developed, in the author's opinion the main
CHAPTER6
The second set of data provided by PRL was for a vertical n-p-n
an integrated circuit and the term 'vertical' refers to the way in which
original form.
ments at 12 frequencies between 0.1 GHz and 1.0 GHz for each of the common
for example, the original phase. readings for s1l in the common base con-
0.9 GHz are 355.3*, -11.70, -17.6* and 336.1* on consecutive readings.
Therefore, although the data was read in to the computer program exactly
between -1800 and t1800. As the calculated phase angles of the models
were produced in the same range, this made the calculation of the errors
easier although the algorithm still had to check for cases where, for
example, the measured phase angle was -1780 and the calculated phase angle
was +176*. In this case the correct absolute error is 6* and not 3540.,
the graphs that are shown later, the phase angles are made continuous over
the frequency range, therefore the absolute values of the phase angles
PRL also provided a general model which they had developed. This
93
E IF E IESa
=. i-_ =. I= C== U- U- LL U- U- ILL LL U- LL LL U- LL
CD oooooccccccccccc(n--%. -%CCC
cu %nw
M V- In CV)It* 0.1 In
Cki m V- CD CD CD CD m CU Ln LO CD MMU)
CD cD CD CD CD CD CD CD CD CD CD CD CD CD CD C CS) CD CD CD tA V
It
cr) ....... 111i! III! I aj !I1 4)
1
m 11.1LU Uj
(1) - Cl)
LU LLILU
M C11 CS)
LLJ
CD LLI
lq-
LU
M
Lli
LID CD
LU LULULU Uj LU ULJCU a
V- CD CD Cl) CD CD CD 0
LL;LLJU.;
CU CD M
- V) V- Cr) CD Cý, - Cý, CD CS) M Cl) LO CD CýJ LO CD IC CD CD CD
................ 4J
a$ S- CU (7) V- CD V- lqr - -, cfý 10; -4 .4 V) - V) MM V1 - V) V) Cd
"o Qj 10 Ln ýo
C) It It it It It H It If it 11 If It It If It It It 0 11 It It c
ý C
C: 0) s- i
0-% -% 0-% 0-% e% 01% ý% 0-ý ý% 0% -% -% 0-% ý% .%-. % e% IQ 0% -% -%
Ic
Ir
cle) N
0 :2
S. )
-
qj uM gi
-i
ci E0
cm Ui Li cm
CM m CD
r,
-4
-0 14ý
s-I
94
model and its s parameters give good agreement with the measurements
frequencies the errors become very large as can be seen in the graphs
shown later. In each of these graphs, the black graph refers to this
ments only in models produced for them, this model contains a negative-
in each case was to improve the frequency range of the model without too
Each of the parameters at every frequency was weighted equally, with a ldB
error in the modulus having the same weighting as a 10* error in the phase.
The first individual models attempted were for the common emitter
was taken as the PRL general model with the appropriate terminal connected
good result very quickly and with relatively few changes in the topology
of the model. In the common emitter case, one node and several elements
were deleted in Stage l.. The author therefore decided to try using
the common base case was also attempted using the final common collector
two models were produced. The first of these used the PRL model as the
initial model and this attempt was terminated when the model had reached
second general model that was produced used as the initial model, a com-
bination of those elements of the first individual models that were common
to two or more of those models. The model developed from this contained
one node. more than the first model, but had fewer elements, and its over-
all error function value and maximum individual errors were smaller.
graphs showing the accuracy of the results are given in the following sec-
errors, the vertical scales for each graph of modulus and of phase cover
the same absolute ranges and the ranges of the modulus and phase are in
PRL model. The major feature of each of the new models is that there
probably be expected since the PRL model was based on the primary theore-
tical aspects of the device being modelled. The new models improve on
derived
The initial model for the common collector configuration
from the PRL general model in Figure 6.1 is the first model shown in
was given a value of 0.1 pF instead of -0.8 pF. This change in value
2.4 dB and 240. The maximumabsolute errors were 3.6 dB and 85*, these
After Stage 1, the model had been reduced from 7 nodes and 15
elements to 7 nodes and 12 elements. The new value of F was 1.12E + 03,
Figure 6.2.
and the details of these are given in Figure 6.2. The model obtained in
Stage 2 was sufficiently accurate that Stage 3 was not entered. The
final model thus obtained, the last model shown in Figure 6.2, consisted of
7 nodes and 15 elements and the value of F had been reduced in Stage 2 by
model were 0.13dB and 1.3" and the maximum absolute errors were 0.32 dB
and 1.74*. This model is extremely accurate and it could be that many
people would find the simpler model obtained at the end of Stage 1 quite
acceptable.
The graphs in Figures 6.3a and 6.3b show the s parameters of the
final model in blue. As has already been mentioned, the scales of the
Y axes of all the graphs in this chapter have been standardised to aid
comparison of the errors. Thus the modulus scales cover a total range of
30 dH and the phase scales cover a total range of 3000. The black points
97
r- a S.SiE+94
No. (nF nodes w7
No. oF elements = IS
ý R (3-6) a 1,12E+92 ahm
R (4-6) 2 Z. SDE+93 ohm
R (4-9) = 9.43E+04 ohm
R (4-5) - 4.99E+99 ohm
L (1-3) = 4.79E+Iag nH
L (S-2) = 4. igE-+OgM
C (1-9) S. OOE-OSnF
C U-9) I. GOE-04 np
C (3-4) i. 74E-93 nP
C (4-6) 1,99F-02 nF
9C- C (6-9) 9.20E-OS nF
C (4-0) IME-95 nF
C (2-9) S. OGE-93 Mp
C (1-2) S. OOE-OSnF
9 (0-4) 3.90E-02 S
(T u-9.29E-93 ns
Delete R(4-G) i R(4-9) (V across nodes
Replace C(1-0) bg R(1-9)
Delete C(2-9)
r- a 1.12E+03
No. aF nodes -7
No. oF elements = 12
F=I. GIE+02
No. oF nodes =7
No. oF elements = IS
R (3-G) - 2. QIE+02ohm
e1 R ýi-S) 2 2.22E+91ohm
L ý1-3) a 1.986+91nH
L (S-V = 7.70E+99nH
C 0-9) = 1.41E-03np
C (3-4) = 2.12E-gi nF
C (i-G) = 1.9iE-92 nF
C (6-9) 1.8CE-04np
9 (0-1) 5.71E-92 S
(T = ?. DOE-92ns
0C. (V across nodes
C (1-4) = 2. S6E-03nF
C (2-0) - 8.25E-04 nF
R (1-6) = 1.41E+92ohm
R (3-9) = 1.48r:+24 ohm
C (S-G) = 9.9SE-04nF
C (5-9) a 1.93E-94nF
tiode I- F-miLter
Mode 2- Base
Node 0- C0114ct*r
MEASUREMENTS
PRL MODEL
,EWML
5 50
-100
-150
-20 -200
10 100
5 50
-10 -100
-15 -150 -
MEASUREMENTS
PRL MODEL
00
-REG GHZ
0
REQ GHZ
3
-50
-1310
-20
0 EQ GHZ
-50 -
-15
-100 ý
-20
in Figures 6.3a and 6.3b are the calculated s parameters for tile PRL
negative). The red points are the measurements and thus show the values
aimed for. The graphs show quite clearly the improvement obtained over
model derived from the PRL general model and the second used an initial
model based on the final model produced by the author for the common
collector configuration.
First Attempt
The initial model derived from the PRL general model is the first
model shown in Figure 6.4. As in the common collector case, the negative-
The r. m. s. errors given by this model were 3.0 dB and 30.1-* while the maximum
absolute errors were 10.3 dB and 134.40. As will be seen later, the
After Stage 1, the model had been reduced to 6 nodes and 10 ele-
ments giving a value of F=1.45E + 04. The model giving this first
local minimum is the second model in Figure 6.4. The maximum errors in
101
40.
,)F nodes
4c. iF elements i7
12 R (3-1) i. i 7E-+02 atim
R (ý-S) 2. skýK YAm
(2-5) 9.43E, 04 ohm
(S-ý) 1.39F-92 ohm
(6-9) 6.2CE-90 &IM
ýi-3) 4.; 2F-09 nH
7-9) = !.. OEý02 n.4
*
C ý3-5) i.; 4E-03 np
f (4-5) inF.
C (4-2) P. 20E--Ds nF
C U-S) i. O@E-OSnF
C Q-6) i. 24E-03 np
C ', 3--2) 1. ý30&-94 np
(, -2) S. 29E-OS mF
(1-0) i. 23F-03 np
0 C (2-0) S. SOE-94 nF
e 9 (2-5) 3. NE-02 3
(T =-L. 39F-02 ns
DejeLe '-ý1-3) (deteLing node) i Combine M-2) wiLh C(3--2) (U across nodes
Oelet. e R(2-5) R(S-2) C(2-S) , C(I-9) y r-(2-0)
? i
P. j, 4SE+qj
No. iF nodes
No. oF elements
e
Mode i- Base
Mode 2- CalleCLDI
Node 0- Emittf-r
A,
I2 HA [jrL
, W
ýýE MnDý-L
2COO
1130
2. ri
Cýr
-Z ')
- ;
i1Z
'j
4,
- rr
" ýIlz
10
IC'..
the models for this configuration were on sj, and the graphs in Figure
6.5 show S12 for the second model in Figure 6.4 in blue, together with
the measured values in red and the values for the PRL general model
capacitor) in black. It can be seen that while the PRL model gives
good agreement with the measurements at low frequencies the model is very
poor at high frequencies, but the reduced model gives similar errors over
configuration was abandoned. At the same time as this first attempt was
being made, the second attempt, using an initial model based on the final
model developed for the commoncollector case, was also being made.
Thus, it was already known that the second attempt had produced, after
Stage 1, a model with more nodes and elements which gave a considerably
attempt.
6.6 and it can be seen that this model is identical to the final model
produced for the common collector configuration but with nodes 2 and 0
C and R7-0 in Figure 6.1, that were present in the PRL general model,
3-7
However, in Stage 1 only one element was deleted together with one
No. oF modes w7
No. oF elements = IS
r -, 1.77E#93
No. oF modesjr 7
No. oF elements a 14
r -r 3.94E+22
tic. oF modesw9
No. oF elements x IS
Node I- Base
Node 2- C-illector
Made 0- Elnitter
1.45E + 03.
since PRL had stated that they had wanted an inductor at the collector
terminal of their model but had been unable to insert one successfully.
This new addition gave the last model shown in Figure 6.6 which had a value
of F=9.84E + 02.
the final model shown in Figure 6.7. This final model consisted of 8 nodes
and 16 elements and gave a value of*F = 8.81E + 02. The r. m. s. errors were
0.30 dB and 3.0* and the maximum absolute errors were 1.56 dB and 4.1*.
r=9.916+02
No. oF nodes -, 8
No. oF elements a iG
Node I- Base
Node 2- Collector
Node 9- EmMar
* MI-ASUREMENTS
*PRLM ODFL
* NEW MODIL
Mr-in
Li j0 C
iiI
5 50
J,]riz
G -RES GHZ
r
-5 0
-100
2,0
-
C.
5 DI-
MrLD. -J PHASE'Df-ýG 521
5 2 00
r i 1-110
n ,-o
0
E!ý Gri Z
-15
r
-50
-20
MEA5UREMEN15
PRL M0Fj L
t'
INL-ý MODEL
20-5
20
i 50
-Pli i fl%
r,
ju
-30
Griz
--40
rý LL5G tl Z
IF
b9
10,
10
10
0 F-rlIE5 Griz
0 I GHZ
1r1lo
-10
1 r5c
-13
where the value of the modulus was -17.52 dB. Thus, although this
great significance. The next highest absolute errors in the moduli were
graphs in Figure 6.8a and 6.8b. It can be seen that they are a consid-
erable improvement over the values given by the PRL general model.
The computing time taken to generate the final model in the second
The model for the commonbase configuration based on the PRL general
Changing the value of this capacitor from pF to 0.1 pF, changed the.
-0.8
The attempt was terminated after indications that either the voltage con-
trolled current generator should be removed or the nodes across which the
based on the author's final model for the common collector configuration
was then used. This is the first model shown in Figure 6.9 and consisted
initial model used in the second attempt for the commonemitter configura-
tion, this initial model lacked the substrate elements that were present
in the PRL general model.
EMU
No. oF nodes -7
No. oF elements = iS
R (3-6) u 2. OIE+02 ohm
e1 c R (i-S) x 2.22E+01 ohm
L (0-3) 2 1.99E+91 nH
L (5-1) a 7.70G+29 MH
C (3-2) a 1.41E-03 nF
C (3-4) 2 2.12E-gi nF
C (1-6) 321.9iE-02 nF
,C (6-2) 1.8GE-8i nF
9 (2-1) 5.71E-02 S
(T = 7. DQE-02 ns )
W across nodes '4)
C (0-i) 2. SGE-03 np
C (1-2) p
9.2SE-04 1-,
R (0-6) 1.41E+02 ohm
R (3-2) 1.49e+04 ohm
1) C (S-G) z 8.95E-94 nP
C (S-2) J: 1.83E-94 nF
Made 1- EmlLter
Mode 2- Collector
Mode a- Base
In Stage 2, duting the addition and deletion of elements, one node was
also deleted. The model produced after deletion of this node is the
third model in Figure 6.9. This model gave a value of F=1.88E + 03.
into Stage 3 resulted in the addition of a new node which created the
The value of F was also reduced to 3.63E + 02. This final model,
F=3. G3E+02
No. oF nodes a7
No. oF elements = IG
L (9-3) 1.99S+91nH
c.
e' L (i-I ) 3.03F+91nH
C (3-2? G.77E-gi rf-
c (S-2) 1.99E-04 nF
9 (2-1) 2. G2E--02S
(T -1.98F-01 ns
(U across nodes
6 R (9-S) a 3.40E+01 ohm
R (3-2) a !. i7E-+9I ohm
C (4-S) - IME-02 nr-
C (1-3) = 3.91E-gi np
C (G-0) a 1.42E-03 nF
C (S-9) = Lile-92 nF
R (1-4) 9. nE+gl aN,a
R (3-S) 9.99F:492 nhm
0b C (1-2) S. 97E--GS nP
C (4-3) i. 2Sý-03 nF-
R (G-2) 1.9sr-+91 ohth
Node I- GniLter
Node 2- Collector
Node 9- Sme
VEýSupEwllýý -Is
PRL MOD, ý-L.
F'ý ML
'N
morl, 5 EI nFilýC*)Eri lu
ESGHZ
15Cl
If'
-
- rHZ
-20
_2r5 -1%J
r0k,
'15 '121 PHASE DEG 521
5 0 FQEIIJGriz
0s
0 HZ
-i,,,
-15
-20
1OL
0
2ri
25
ýjrjz
35
-31i
-40
-15C
-45 ýýHZ
10
nki
nj !-jb, 'J22 rj,,. 322
-G
10
FRE5 CjHZ
Dr"
0 Griz
IC'
-I
-13
-200
As can be seen from the graphs in Figures 6.11a and 6.11b, this
expected that the analysis of the errors at any particular stage would take
general model was the PRL general model. In the second attempt, the
author produced an initial model based on the final models generated for
each of the common collector, common emitter and common base configurations.
The PRL general model shown in Figure 6.1 gave an overall error
Fa3. IIE+05
No. oF nodes -9
No. oF elements a 22
2 jE C (1-3)
(1-0)
SME-as nF
1.23E-03 np
I0 (3-0)
(2-G)
S. SQF-94 np
3.92E-02 S
(T w-L. DOE-92 ns )
(U across nodes 5g)
C (2-9) x 1.22E-93 nF
C (1-2) a S. WE-QS nF
C (2-3) x SME-gs nF
No. oF nodes w7
No. oF elements a IS
R (1-4) I. O?E+92 ohm
___ ___ C. G (4-S) 1.38E-93 S
R (6-0) i. 9SE+QIohm
L (5-2) x 1.48E+21nH
C (1-5) 2.! IF--03 nF
C (4-5) 7.8ae-93 MF
C (1-3) 7. ise-gs np
R (3-S) 1.47E+03 ohm
C (2-6) 2.43E-03 nF.
C (1-3) 1.62e-04 nF
C (1-0) - 1.83E-93 nF
R (3-0) 0 3.63E-+03ohm
9 (3-S) a 2.97E-92 S
e2 (T 2 1.7SE-92 ns )
(U across nodes 34)
C (2-0) wI G7E--93nF
C (1-2) u SAIE-94 nF-
0
x G.Vt+24
No. OF modes a8
No. OF elements a 21
3 R (7-4) x 9.3SC-+03ohm
R (G-2) a 1.31E+91 ohm
L (S-2) 2.32E-+9l nH
C (I-S) 2.3SE-03 nF
C (4-S) 3.32E-03 nF
C (4-3) S. ISEAS MF
C (3-G) 2.2SE--93nP
R (3-9) a 8.47E-+23ohm
9 ý3-9) & 2. SK-02 S
(T w-4. S.)G-92 ns )
(V Qcross nodes S)
C (I-G) 2.33E-93 Mp
C (2-G) i . 48ri-.23 nr-
R (1-2) 2.91E+02 ohm
R (1-3) !. D! E+OS ohm
R (2-S) G. 20E+92 ohn
C (1-4) ?. S7E--94 np
L U-1) I. I7E+02 nH
0 C (3-7) I. SGE-04 nF:
R (S-7) 1.71E+97 ohm
C (S-9) 2.92E-OG nF
Made i- Base R (7-9) = 2.99E+93 ohm
Node 2- Emitter R (1-4) 2 1.9? E+92 ohm
hoe 3- Collector
Node 0- Ground
N
32.8". As can be seen from the graphs shown earlier, the model gave a
good fit at low frequencies but there were gross errors at higher
frequencies.
on using this model as the initial model for the general case,
Stage I caused the reduction of the model from 9 nodes and 20 elements
to 7 nodes and 15 elements. These are the first two models shown in
Figure 6.12. The two node deletions were attained by (i) the deletion
of the base terminal inductor and (ii) the deletion of the resistor
Stage 1 was 1.67E + 05. The individual errors now became more evenly
distributed although they were still smallest for the common collector
configuration.
After one entry each to Stages 2 and 3, during the latter of which
the base terminal inductor was re-inserted, and one further entry to
Stage 2, the final model shown in Figure 6.12 had been produced. This
model consisted of 8 nodes and 21 elements and gave a value of F =6.07E +-04.
The r. m. s. errors of this model were 1.45 dB and 14.50, but the maximum
absolute errors were 1.5 dB and 380 in the common collector configuration,
6.3 dB and 530 in the commonemitter configuration and 8.0 dB and 59* in
CDC 7600 to develop. Because of the size of the model each further
r- = 9.78F+GS
No. oF nodes -9
No. oF elements i3
I L (1-4) a 3.20E421 nw
lo q (4-S) x 2.30E+92 Ohm
R (I-S) a 8.29E+91 ohm
C (1-3) x 7.90E-94 nF
R (4-3) it I. OQE+giohm
C (S-6) x 2.30E-92 nr-
9 (3-G) 2 SME-92 S
(T a-1.00E-92 ns )
(V across nodes 11)
L (7-2) a S.02EQ9 01
C (1-6) N 3.20E-03 nF
C (S-3) - 2.20E-94 rF
R (6-7) a 3. ME-+Ql ohm
C (3-7) a S. 09E-04 nr-
C (3-2) a L.QOE-OS Mr.
r- - 3.0SE+0S
No. OF nodes -7
No. OF elements a9
L (1-4) 2 S.SPE+99nH
b R (4-S) a 9.96E+91 ohm
C (4-3) - 2. S7E-94 nF
C (S-6) a i. ISE-92 rf
9 (3-6) a 2.37E-02 S
(T. -SAGE-02 ns
(U across nodes
L (G-2) x S. 9GE+99nH
C (1-6) a 2.25E-93 nP
C (3-6) u 4.37F-OGnp
R (3-2) u S.?SF+03ohm
r- a 4.78F+94
No. OF modes-9
No. OF elements = 19
L( 1-4) I. M+01 rAi
b1 C (4-3) G. GIF-OS nP
U. -G) 2.026-02 S
(T --L. GOE-91ns
(U across nodes
L (6-2) 3.23E-+QlnH
C (1-6) 3.97E-93 np
C (1-9) i. 3SE-03 MF
C (2-1) 1.21E-03 nF
C (2-9) i. i2F-03 np
9 ý4-5) x G.Sir=+33ohm
R (2-4) - 2.; ýIZE+02ohm
C U-3) 2.23E..93 pip
(S-? ) 1.49R+gi ohm
(7-9) 3. ýDE401 ohm
R (1-7) 1.3ýE+93 ohm
C (6-7) 9.22E-24 np
R ý2-3) 2.33E-+24 ohm
0 L (9-9) 2.17E+90 nH
C (1-3) i. ý2E-gi Mp
Mode I ease
Mode 2 Emitter
Mode 3 Collector
Mode 9 Ground
For the second attempt, the first model shown in Figure 6.13 was
used-as the initial model. This model was derived by coinbining those
elements and gave a value of F=9.78E + 05. After Stage 1 the model
This simple model, which is the second model in Figure 6.13, gave r. m. s.
errors of 3.3 dB and 32.5* and maximum errors of 11.0 dB and 1520.
shown in Figure 6.13 had been obtained. There are two points about
this model that provoke comment. The first point is that node 5, at
additions and deletions and until a late stage other elements had been
each of the emitter, collector and base terminals, but in common emitter
formed in the region of the ground node. However, the addition of this
obtained in the first attempt. The r. m. s. errors of this model are 1.3 dB
and 12.9* and the maximum absolute errors are 4.8 dB and 28.1* in the
MEASUREMENTý
PRL MODý-L
NLW MODF-L
50
-100
-150
-20 -200
10 IN
5 50
-10 -100
-15 -150
* MEASUREMENTS
w PRL MODEL
* NEL-W
MODEL
150
5
1100
0 ý-REQGHZ
0
50
0 EQ GHZ
-10 10-1
-50 -
-15
'10
-1 Ili
-20
0 FREQ GHZ
100
)0
50
0 GHZ
-50
-20 -100
-25
PRL MODEL
0 FREU GHZ
FREQ GH7
1;
)0
-10
-15
-20
-200
5 200
-10 50
-15 0 GHZ
-20 -50
* MEASUREMENTS
PRL MODEL
L
-15
200
-20
150
-25
100
-30
50
-35
FREQ GHZ
D0
-40
",--REGGHZ -130
1 10 0
10'
Figure 6. lSb s and S 22 for General Model in Common Emitter Configurat ion
12
_
121
MEASUREMENTS
P.RL M'uDf-L.
NEWMODEL
li jI FIHýCJF-
MOD
FF-fý,
EQ GrIz 1K
r oro
-ý'EQGHZ
rý0
.0
25
rý 0 'IHZ
F L-ý Gmz
50
.
1fil0
150
-10
20 ri
? 50
MEASUREMENTS
PR' '--L
-0
ý41EW mn DL
,0 lý
50
C) -REQ GHZ
30 10
-50
-35
-100
40
-150
45 IKI
Griz
--
7S9
10-,'
10
50
5
0 FREQ GHZ
ýE:Q GHZ )0
0
-50
-5
-100
-10
-150
-15
-200
can be seen that not only are the maximum errors lower than those for
the final model obtained in-the first attempt but the distribution of the
errors is also different. This demonstrates that the model chosen for use
as the initial Model can have a significant effect on the results and that
The S parameters of this general model are shown for each config-
that the new general model is not as accurate as the individual models that
were developed. The new general model has removed the gross errors that
were present in the ýRL model at high frequencies but has lost some of the
The time taken to develop this general model was 8300 c. p. u. sec-
should be attempted.
With the benefit of hindsight, the author now feels that the
approach used to produce the two general models was too ambitious.
Knowing that both the size of the problem and the computing time
required would be large, the first priority should have been to ensure
that some accuracy was lost. Therefore, in the analysis routine single
precision arithmetic could have been used and, as shown in Section 3.1.5,
this would have achieved a 50% reduction in the computing time required
although this, too, involves a loss of accuracy, it would have halved the
124
Further savings could also have been made by using less of the
have been a better approach using the initial model used in the second
attempt. However, in the first attempt, since the PRL model gave good
CHAPTER 7
reference numbers lE2M and 3E2M and were isolated n-p-n transistors made
The JE2M had one emitter stripe and the 3E2M had three emitter stripes.
The data provided consisted of the s parameter measurements for
the two devices at 16 frequencies in the range 0.1 GHz to 2GHz with
currents, Ie, in the ranges 0.5 mA to lOmA for the lE2M and 0.5 mA to
27 mA for the 3E2M. The graphs in Figure 7.1 showing fT and the modulus
I fT
(GHz)
3.0-
1TM
2-0-
: ýýI 1M
E2
.01
Mod
S, I
a?lGHZ
(dB)
le(mA)
Figure 7.1 Transistor Characteristics for Vce ý0*5V
126
ranges of. the transistors up to IGHz were 0.5 mA to ImA for the lE2M
currents such that fT was still less than its peak value. The model
was based on the physical structure of the device and contained a number
general a good fit to the measured s parameters had been obtained for
Slatter16 states that the model for the 3E2Mfits the measured s para-
meters more accurately than that for the lE2M and that s1 for the lE2M
shows larger errors than expected. He also states that the higher fre-
quency limits for the models might be 1GHz for the lE2M and 1.6 GlIz for
the 3E2M.
The author started by producing a model for the lE2M using the
the individual errors in the modulus and phase as were used when modelling
initial model used is shown in Figure 7.2. This model, which consisted
of 8 nodes and 12 elements, and the initial element values shown, were
based on the PRL bias dependent model. A new model was produced for
the IMI over the frequency range 0.1 GHz to 2GHz for V = 0.5 V and
ce
Ie `2 4 mA. This model consisted of 6 nodes and 14 elements and gave an
the limit of the linear region of operation of the lE2M and when it was
used as the initial model for the same transistor with V =3V and
ce
Ie = 0.5 mA, further changes were made resulting in the less complicated
and the maximum absolute individual errors were 0.7 dB and 3.8". The
graphs in Figure 7.4 show the s parameters of this model compared with
the measured s parameters. The scales of the graphs cover the same
CDC 7600. This time includes the two stages in the development of the
model, from the initial model in Figure 7.2, through the first model
with Vce 'ý 0.5 V and Ie=0.5 mA, it was found that excellent results
found that this model, with the element values optimised, also gave good
128
No. oF nodes =8
No. oF elements = 12
bI c. R (1-3) = I. OOE+01ohm
R (3-4) = 2. OOE+01ohm
C (4-S) = 3ME-05 nF
R (5-2) = I. OOE+01ohm
C (3-6) = IME-03 nF
R (1-6) = S. OOE+01ohm
C (4-6) = IME-0i nF
g (S-G) = I. OOE-02S
(T =-2 SOE-03 ns )
.
(V across nodes 4)
R (6-0) = I. OOE+00ohm
C (2-7) = 1,00E-03 nF
R (7-0) 2. OOE+02ohm
C (1-2) I. OOE-OSnF
Mode I- Base
Node 2- Collector
Node 0- Emitter
P= V30+02
No. oF nodes =6
No. oF etements
bI
C (1-2) = 1.33E-Oi nF
9 (2-5) = 2.5SE-02 S
(T =-2. GGE-02ns
(V across nodes s
C (i-S) 8. ISE--Oinp
C (3-5) I. We-03 nF
R (3-1) i. giE+01 ohm
R (1-2) 1.9! E+02 ohm
R (1-5) I. Sep+03ohm
R (O-S) 1. 90"E+01oh-n
C (2-3) 9.07E--OSnP
0
Node I- Base
Node 2- Collector
Node 0- Emitter
Figure 7.3 Final Model for 1E211 for V ý3V and I. = 0.5 mA
ce
129
* MEA5UREMENTS
,k NEW MODEL
DF 1ýiI PH L-'
: - -,
ýEI-,
I, -- -jFlZ F-E riZ
i 5c
15
25ý
1I"
IL, -, , r,
ý- F\ t ýJ 'i Fl Z
0 I '\'
It
ý rýll-ý li'OZ
MEASUREMENTS
NEWMODEL
ti -12
2CIG
-25
r
--3,
IE IJ-1
'I--1ý Iýriz
39 10
-
-40 EQ 1ý z
10
I (\(
1C, I JLI
I 'r
-15C
results for the MM with Ie=0.. 5 mA and with both of the values of V
ce
Following this, each of the four new models was used as the
initial model for the same transistor and the samevalue of Vce with the
next higher value of Ie* Excellent results were again obtained without
further changes in the topology and in every case the optimisation of the
model element values took less than 200 c. p. u. seconds on the Leicester
University CDCCyber 73. Therefore the process was repeated for still
function for each of the models having the same topology as that shown in
Figure 7.3 are given in Table 7.1.
Objective Function, F
mA lE2M 3E2?4
e
V = 0. sV V = 3V V = O. SV V = 3V
ce ce ce ce
Table 7.1. Values of the Objective Function for Optimised Models of the
lE2M and 3EN Transistors
Of Ie up to 4mA for the lE2M and up to 8mA for the 3E2M. Although the
value of F=3.80E + 03 for the 1EN with V = 0.5 V and Ie=4.0 mA,
ce
considered acceptable since the maximum absolute errors of 1.7 dB and 15.2*
occurred at 2GHz, the highest frequency, where the moduli of the s para-
this model for the two types of transistor agree with the suggested small-
accurate model for the small-signal operation of the two similar bipolar
transistors covering the frequency range 0.1 GHz to 2.0 GHz had been
developed.
Although the modelling process had gone through a number of diff-
erent stages, this new model was quite similar to the PRL bias dependent
model. - Looking at the elements in the final model in Figure 7.3 those
which have obvious equivalents in the PRL model in Figure 7.2 are C
1-2"
C and g2-5' Those having a slightly weaker resemblance to elements
3-2
in the PRL model are R, C4-, and R The remaining elements: R
-4. 5-6' 1-32
R and C seem to form a 11 arrangement where in the PRL model there
1-5 3-5 ,
is described by Weinberg77 and the author suggests that there is the possi-
the T arrangement in the PRL model, with some of the extra elements that
resistor, equal to Rl_, in the PRL model, in series with a capacitor, and
Having developed the new model, the next step was to examine the
variation in the values of the elements in the model with the bias condi-
tions. The results of this exercise are given in the next section.
At values of Ie greater than 4mA for the lE2M transistor and
133
8mA for the 3E2M, topology changes were indicated. Details of some
the element values against the emitter current. These graphs are shown
in Figures 7.5a and 7.5b. In each of these graphs the points-are joined
by straight lines. The red lines refer to the lE2M transistor and the
blue lines to the 3E2M. The solid-colour lines are for V = 0.5 V and
ce
the dotted-colour lines are for V = 3v . The black dotted lines are
ce
the functions that were fitted to these points and will be described
later.
These graphs were extremely encouraging as most of the elements
showed a distinct pattern of variation with the emitter current and the
collector-emitter voltage.
Although the model was different from the PRL bias dependent model,
some of the elements in the two models could be compared. The bias
curve fitting routine from the NAG library, E02ADF78, to examine the fits
the model shown in Figure 7.2 for the comparisons with the PRL bias depen-
dent model.
C1-2 - This element had a direct equivalent in the PRL model which
they suggested was not current dependent, with values of 0.33/VV- pF for
ce
the lE2M and 0.621VV pF for the 3E2DI. In the author's model the
ce
134
C' (1
-2)
N' GM "3
ý
30,39
---:
311)
ýý61 0, -ýý X
[iJrN M
NU-C c (4-51 N,
/
/ /
0.
/
-- -- -----------
Li I ',
_) c_)
C kj Q
t?,, _-1, j -ý !JA
C (3 -T
- lE2M Oý.5V
1E2M 3, OV
3lý2M OAV
3
ýUM jý 3v
=
1.0'1 -w=.
ý-L1
iNk'7f C,1,-)
CURQt:NT, NA
20
-- --------
40 40
3 ri
2O
.3r
"r
0-
2'7e
MIA !IA
22'3
IA
14
12
140
203
!JA
(j ý r, ý- KI -, !I
ý
C (2--3) N'
-1 E2M 0.5V
1E2M 3. OV
- ýE2M 0.5V
7 ?m 11 ýv
,
----------
5' 6
However, the value of C1-2 was definitely varying with the value of Ie
the value of C1-2 that a 5% change in the optimum value Of CI-2 made a
functions chosen for C1-2 were functions of both Ie and Vce as shown
in Table 7.2. ThL- black dotted lines on the graphs of C1-2 in Figure
7.5a show the function evaluated at the appropriate values of V and Ie.
ce
g2-5 - PRL suggested that the value of the current generator should be
proportional to Ie A further
relationship stated by PRL was that
1
where hfe re was element R4-6 in Figure 7.2 and h fe 0 so
r.
e
It is possible that R4-6 in Figure 7.2 could be equivalent to element
quite right. The average values were 44.4 for lE2M and 46.3 for 3E2M.
However, taking these average values with the optimised values of Rl_,,,
the values calculated for g2-, were not sufficiently accurate. Thus the
relationship chosen for 92-5 was a piecewise linear function dependent only
on I The sudden change in the value of this element for the 3EN at
e*
Ie=8 mA with V = 0.5 V was taken as a sign that the device was becoming
ce
non-linear. This effect can be seen in several of the model elements
and in each case it occurs at the highest value of Ie for which an acceptable
137
value of F had been obtained with V "" O'S V- No attempt was made to
ce
model this effect and the fits of the piecewise linear functions chosen
can be seen as the black dotted lines on the graphs of g2-5 in Figure 7.5a.
-r - The time constant applied to the current generator in the PRL model was
given the independent values of -25 ps for the lE2M and -26 ps for the
model vary with the emitter current and that the values of T for the
in mind the PRL values for T, it was decided that constant values
should be chosen based on the optimum values at low emitter currents since
these were the values when the devices were most certainly in the linear
region of operation. The values chosen were therefore -28 ps for the
JUDI, the emphasis being on the values for the higher voltage, and -32 ps
C4-5 - The function given for the equivalent element in the PRL model
was A/F where A took a constant value of 1.12 pF. Vi for the lE2M and
1.16 PF. V1 for the 3E2M and F was a function of the collector-substrate
current while at lower values the element values were almost constant.
curves were obtained for both transistors on plotting the optimised values
element.
of the transistors and the combined constant values of these two resistors
in the PRL model were 66 Q for the lE2M and 23 Q for the 3MI. In the
the sensitivity of the model to the value of R3-1 meant that the effect of
changing current was significant. Therefore, it was decided that both the current
R - The equivalent element in the PRL model was given constant values
4-2
of 209LI for the lE2M and 185sl for the 3EN. It was felt that constant
values were the most appropriate and the values chosen by the author were
1890 for the lE2M and 14S Q for the MM. It can be seen from the
graphs in Figuie 7.5b that these values were based on the optimised values
R1_5 - This element has already been mentioned during the discussion of
the two values of V for each transistor. The functions giving the best
ce
fit are given in Table 7.2 and are plotted in Figure 7.5b.
arly for the lE2M for which the value of R0-5 at Ie00.5 mA was as high
element values for V = 0.5 V first and then to modify this function as
ce
necessary for the higher value of V The final complicated function
ce'
given in Table 7.2 is not entirely satisfactory and the fits obtained
using the function are shown in Figure 7.5b. It can be seen that the
values with V `2 0.5 V are considerably more accurate than those with
ce
V=3V. Fortunately, the model was not highly sensitive to these
ce
low values of C2-3'
140
2 A= - 1.12 A= - 0.47
9 (A + B. I ). 10- 1s1 mA B 1s1 mA
2-5 e B= 7 21 e = 4 . 64 e
.
S A= 3.4 A= 1.5
1>1 mA B= 3.0 >1 mA
B= 3.0 e e
T ns A A = 0.028 A=-0.032
C4-5 I- 3
A+ 10- A = 0.54 A=0.34
i+ V17-
nF ce
[A A = 0.024 A=0.024
C B
3-5 + I 92-5
nF e B = 0.018 B=0.055
R A = 62.9 A= 22.9
3-1 B
A V B = 43.1 B= 17.2
+I+C. ce C = 1.7 C=1.2
e
R
4-2 A A = 189 = 145
0
R A = 8.385 A=7.371
1-5 105
A+ B. I . .
Q e B = 108.32 B= 76.19
R A = 1.0 A=0.1
6-5 A+B
C
SI B = 8.0 B=0.8
C 5 A = 3.27 A=8.2
2-3 +BIeV lo-
ce B = 5.95 B=9.7
nF
Table 7.2. Bias Dependent Relationships for the Model in Figure 6.3
141
model elements was discussed in Section 7.2 and functions were developed
ent emitter currents on the model element values could generally be seen,
It was known that some of the functions would not give the
a check on the effectiveness of the new bias dependent model, graphs were
As expected they show that the model is accurate for values of modulus S21
up to its peak value. This appears to be slightly better than the PRL
model which, in the paper by Slatter17 is shown to lose accuracy for the
,
3E211 before the peak value of modulus s is reached. Graphs of the
21
S parameters calculated using the author's bias dependent model, together
o MEASUREDVALUES
MOD'321 lE2M 0.5V ,, MODELVALUES
0 CURRENTMA
ýD 246 10 Je
-10
-15 -0
MODS21 lE2M 3. OV
10
5
n CURRENT
MA
To
-5
-10
-15
lo
5
0- ýQ CURRENT
MA
0m5--- 20 25
-5 ----
-15 - -0
10 -
5
0 CURRENT
MA
5 10 Is 20 25 le
-5
-10
L
-15
Figure 7.6 Modulus S21 at 1GHz for the Bias Dependent Model
14
MEASUREMENT5
NEWMODEL
MIU DHPE: iE iI I
C:
5
'1
_rd HZ
"I
-
I "
20
1,r3 IC
1c
1 (r
rz
HZ
5C r
_________________________
-; -' 'ilL
-\
10' 10
MEASUREMENTS
NEWMODEL
r%r
ýQ
715
40,
Gilz U - -
!-Lý'L
ý ill-iz
-'
PC 1o"
If'
L
-J
0 0 ýAZ
4r rlc
-
15
-
A5UREMENTS
ME:
NEWMODEL
rulDLAI,
r
ic
-
0 HZ
-I 1ý11
C
-I 5c
'Az
10,
2C
15c
113
C -J
50
-
10,
MEASUREMENTS
NEW MODFL
1DC
czc,
,
-09 1ý
"z
.
--
10 1O
MC.C. ID-11 ?
"--'2 PHVjH' Uý-G "'ý22
i rý
5
rr CjrZ
0 ltdz
ri
- 11.1,
-15
up to 1GHZ.
The conclusions drawn from this were that accurate bias dependent
models could be developed using the author's modelling algorithm and that
dence of the model elements on the bias conditions. Although some theo-
values at different bias conditions could indicate the possible forms the
function could take. Then the use of a curve fitting technique with a
After the bias dependent, model for the linear region of operation
had been successfully developed, attempts were made to develop models fot
the transistors at the highest values of emitter current for which data
was provided.
For each transistor, using the s parameter data at the highest
values of Ie given and with both values ofV given, the initial
ce
model used was the appropriate optimised model for Ie=4 mA or Iea8 mA
from the earlier work. Applying the modelling algorithm to these four
these four models for each of the four cases, one of the models proved to
be the most effective with V= '2 0.5 V and I 10 mA and 28 m.A for the
ce e
lE2M and 3E2M respectively, while another model was the most effective for
function in each case was less than 4.3-E + 03 and the maximum absolute
individual errors were 1.5 dB and 12*. The models were again effective
148
No. oF nodes =2
No. oF elemenLs
Node I- Base
Node 2- Collector
Node 0- ErMter
0
IE2M O.SV lOmFl 3E2M O.SV 27mA
bi c
lio. oF modes m6
No. oF eletents a 12
Node I- Base
Node 2- Collectir
Node 0- ErAitter
0
IE211 3.9V lOmA 3E2M 3.2V 2?mA
9 (2-9) =I SOE--g
IS (2-0) = ý. S4e-gl S
ýT =-i. 3SE-22 ns (T ::-3.2iE--02 ns
(v across nodes 2 (V acrass nodes 3) 9)
(1-0) = i. 90E-03 nF (i-0) = 3.16E-03 np
C (3-0) = !. SýE-02 nF C (3--0) = 3.97ý-02 nF
q O-S) = S. 93E+91 ohm q (3---) = 2. Z8F+31 ohm
R (1-2) = 2.91E+02 ohm R (1-2) = 2JOE+02 ohm
9 (5-0) = j. S@E+02ohm p (S-0) - G.S7E+el ohm
C (5-3ý = 9.42E-04 nF C=2. SSS-03 nP
1.98E+Oi ohm = !. ý7E-101ohm
(2-0) 1.39E-01 nP (2-0) i. s3e-oi n"-
L (I-S) 1.9ýE+00 nH (1-5) MH-01 nH
C (1-2) 2,72F-V nP C (1-2) S. iSE-Oi nF-
C (i-3) i. 02E-04 nP C (1-3) 9. ýH-04 nF
MEASUREMENTS
NEV MODEL
IU
MOD DE lE2NI 0.5V I OmA ý?, IF 2M 3%, 1OmA
10
0 `CýE SHZ
0 ý rl, 1, -17
1- ýI
r
MOD DE, 3E2M 0.5V 28mA Iuj
__J 31,2M 3V 28mA
:
2:
' r\"--r-! li HZ
15
10
r5
C
20
-
Figure 7.10 Modulus S Against Frequency for the High Current Models
21
iso
CHAPTER 8
author felt that there was insufficient guidance on the choice of alter-
the author's modelling algorithm which was developed for those cases
accurate, but attempts to produce a single general model for all three
used to develop a suitable model and then to optimise the element values
author.
small, fast computer program. The author's view is that computers are
should be written efficiently, they should not take short-cuts that could
the development of the general model for the vertical n-p-n transistor
models for some large problems. However, these sacrifices might not
of the models,
153
This could offset the loss of accuracy due to (ii) above, and would not
each successful addition should remain in the model. One effect of this
was that sometimes elements were added and then removed again during the
addition of-a different element. Cases have been seen where a parti-
cular element is added and removed several times. There are two alter-
select only the best of these. Then a suitable technique for further
evaluate all the possible additions, taking only the best again. The
former suggestion for further additions might seem to be the most effic-
ient method, but changing a model by the addition of only one element
each possible element addition and then to insert all the successful
For large problems such as the general model of the vertical n-p-n
transistor, even these suggestions might not reduce the required computing
154
choosing the constant values in the rather arbitrary manner used by the
choose which elements should take constant values and to ascertain their
optimum values. This would, like the general model of the vertical
data.
the models are then re-optimised with the constant element values fixed,
model under certain bias conditions by the use of diodes in much the same
these cases, the author feels that an initial model based on the modified
modelling of most bipolar transistors. The author has also shown that
bipolar transistors using simple functions for the bias dependent elements
It was hoped that this regearch might indicate how the standard
described here, the author is of the opinion that a much larger selection
APPENDIX 1
(L) CN CVI LM
41 10 1 1 1 1
Z
0
I
. 10 0 0 0
V
ca
$4
4) 10 C) rý 0
%. cel
41 W '0 r-4 P-4 r4
Aj
19
cn LM KM Co 0
PQ cz
"4
cz
P-4
4
r-i P.4
10
r4
Cd
34
41
93 4)
m 10
41
c4
0 c4
41 10 %ID LM tn
(Li
cn
c4
Cd W 0
cý f-4
. 92
43 M 1 1 1 1 Cli
rn
cr
C)
z
158
cl r-
CIJ rij 0
) ..
i-,
.1.,:, 'Cl
Cj
rx..
r F...
r, Ln-.--j.-T. c-i
rr.. ---v
I..) r,
C.
")
- 4. ;Z7-
........... . c
0
.001.7") -4 ...p f.'IJ CIJ 1=1
z.co:r, cr.r,.
,Z. C. i: ý
41
tw
C.j
ce
N 41
m
CT. Q
C) -4,0 C."
r,
. YJ k 4. d
0 "r-I 0
z 0 E
U, U
CD 0
0 Ln *a) 9:.
C, co C, 1=1 1 0 P-4
cl) .:: 1=11=11=1.: ý
C) Z ;Z
N
:=I tz
Cl Ci
Z IM
14:ýI..'- Z. i.'..-z' .:, Z,..Z, IZ,, Zl
159
S PA;;, V--.E7ERS GF VERTICAL UP14 TRAi. -SISTOR Vi CU'-"! O-*4 CULLEClUR CO-SIGJRATION
1ý01.
CJSEF 117,r-
I-Efill LEFT F',-'ti ! ':', CiUl7E! FPIPS
IE FF,F:F.I-,EIEF'C'
I-, C101-1
IICICI.
6100.r-jc:cl -1.67 -SUL
7 C!C. C,QC-1 13IE-7 S3C
It'
7
It,. rl-, 2.
SOO. OCIC, 4 -
C. 12
.
74 3.1
1C-:
CIO C,EIC, cil 2 e.:1-o -IE. 25 E.S. -, Fl. E. ",
1
ý4 -9 q, 9 t4; *.:I
13 2 ,:., -I . E,
12 -1
. 'Ile
1.7 C.
-SOTH 1101.1EI-TEP
11SM LEFD P-11 ISURTED FRIPS
1,
)=t,lcc I=Ic S FPFf-.t-IEIEFS lllý'ISYIIEG
C
I' EMITIEFI, EhITTEP l,:IIITH 2 MICF.C46)
L,C,LTS= CL.,
F,F:Etil' I-,.
Fi=
FREQ IDA I I*IQ
s2 i
00.001 - 12. IS " C, IS'C,
. S4.5 I C, C-;ý4 I fin 5.0
. cl
Cl. C.QCi - 17 . .0 -E5.46 7,7 1ro- ý8. 15C 1 . -at.
-2
76.5 .
1C.; ý5 151. S
ES7 1 ;-. 2 56 14 1. cl
00. C'CCI 7 C,
I! CO. Octri I Cl. E. IE 5 57.5 E
CO. 11(qu -, C!.; E. - 17 54
. . -, ýC17'.I ":2
C0. CJ Z:,47. .
.6.1. 5.1:
llý 9 "'Ci. 2 E E.
clIt . C-1 -:
",:, I. IE 1%-- 4? ý- 0...
- I C-)1:1 ::. *.:I ..
-, . 10pl: 7
15-C,Q. C,E,C, SCI .a -
I OF-10.0 CIO -7 CIO
',
-5. '74 -4. SE; ';:17. -
I r'-C,C,. C.,
C,cl - 'C
el. 71 -1r. ý- -C.?. 9 5 C%-.
cl
.I 4 f. . -, I . c. .
"ý- #I. E.
.
71 E7 I 0-11
a-ItF. I--. .
E2.6 - 2-. 11 -52. ID
TabLe I Sht. I o; IG
161
*A.
711ý1
BEFIN LEFT F'.,*tl HiCILAIE!, PH F.C-
I , 7-il 9DEC
ý- F-FF.f: t-:El E F'I-- Ur
119.7?
6E U. LIC-10 -SI
1.1
. 13 _26ý. 9
.
2 -f "5 ;?
7 OCf.000 -$. Ze .
C cli -EEC. 11: 1
1 57 4.69 7.2
SOO.CIE-1f) -1 08 ZE 2. C. -IS. 12 4.11 W"
62.1
I C.ecl. CIC-C, 256.9 17. 4 ýo 51
- e5 . 31. 10 79.2 4
12 0 (1. (1E'Cf I
C-50.2 - 16. 'ý9 46.4 I. I.B.5 7%;,.C!
140cf. OCICI -12. SE, 2:45.7 47. E. -7.64 -44.1
-16.18 7ý 6 8. z, 7., c:.C 3145-.0
clcici - 1,ý IE.I 24E. v . 41
. ý -15.71 -312.7 -. _296.4 40 312.3
14.06 I"%- 4'".
9CK4. CICý-4
- C-,:.7 ri
228.2
-
14.e-0 ý C, S-3 59.1
-14.48 -14. E7 45., + 1-1 5S.
.4 -el. -73 -50.3
,201h MOUMBEP
ISAM LEAD P,,ti I-, n.LFilEI, PAIP,",
1.1=1)cE: I=IE: S F,FF.F.t-IETEF.,,:IN
1,19DEG
1EM(l EPIMEP; VAITER WIDTH 2 MIUCINS)
S CUPPENT NA= 4
FREG sli S12 sel
1 EiEi C,Cir_I C4 '2 2
. -62.4 -2-7.95 46.4 S9 1F.O. 5
CE. _:3.. jZ -20 9
-75. C; -;+7. S 10.1, C, ice. 2
-4 4-ki
. _-Z
. g
.
Sht. 2 o; IG
162
Sht. 3 o4 1rc
163
1101.1EI-II-EP
1.S7
'FEAN LEP:, F-1; ISCUIEL PAIPS
I=lr S PFFE-METEPSIN DI! I
YIIEG
IE'r-l'ttl Ef-II11EFI; EMI7TEF. WIIIIH 2 MICRUG)
VÜLTE- 51 CUFF.EliT MR= 10
FFEC' .
5111
C, C, c) ,'ý21 ""22
1C--, . -10.79 -144.2
1to. 1. -;icici -! 1. C-11 ITDc .c Ll ý4
n. cl
-11.01 ISC, 10
EIci k.1ür;1 -11. Ell
1 1 S4. 37.9
15: . -2.2E. 61.2 11
4c.le. -11.16 7 4. S.O 50. IS
-11.12 -11.78 %306.1
-C-. 87, 4E. 2 12.07 2? S..,
-11.06 3'19. C, 5, Cj 12 r'.7 an. I
-
-IC.1.93 -10.1 :3. E9.9
C- I S6.6 -11.62 U. _7
i,e5.1 -19.2E 41. c- - 12'. 64 2-4.1
1Ociel.WO Efo 27C.. 4
-11?. 21 3.91 24.7 1%S 19
-ý42
12190.ccici Is ":,C-1 2w. 3 12. i e.
-Ic, I
1-4EI0. C1C1C1 .
6.09
CiEICI 187'. e. -12.64
16.67 6 12.7?
-I Ci.(q.-. 42.7 6.25 a -
74 16.4 4 42.8 -16.13 41.2 0* -117.7
Sht. .4 of IG
TEXT CUT
OFF IN
ORIGINAL
164
ýVlf 1 0. U.
I: EF,I-l LEFI! P.-It
it: I, EC
2 CITH hCASPSEP
1EN-1 LEAD P.11 ISOLF17H, PAIRS
I.-Qrx I=IrS FFRENVERIS M Dl: yl; EC
"E21"? EN
AT1EF; EMITTEP WIDTH 2 MICF.C111S)
MOO Sf CLIPPENT MA= I
FFE0 C11 S162-1 S21
C1
OC' 77 -22.4 IC
-ac". 05 76.5 11 C,21EI. C.
C1 . 1 -. act
2CC1. C10c,
-1.031:a -31.? -21.03,1 C. 72 1A. I
-1. 0-13 C.
ý4.
-1
a7 -2 "4
500. C1 CH -2. C17 iF" t. . e- -Iz.. F7 ec.l. 1 _711110 11 !. E.
c,c1c:. 29 . -,:: - 7-1 1.1C. 11
22f-,E7 14. Z
44 .
VC.. c
700.0 0_1 -3.71 C,17, -1.;. UE. C, 45 -z .
ý-CICI. CICIC, -4. IL 249Z, _1* -, e-,4 4, CIO. co :02.
C,0. (10 C1 26 2_iE. C, 67 -C ?
*C.
-1 25.4 2
C,C C' ro C-0 61 co IS ZS 5. E.
6. E12.;,. 4 E.
. Z
cl 0. C;l C.,Q 72 222. c? I,: ý-S* 1:2. E5
.5I. C.
C-0 M I.:,;
c 1 -7.70
C.
E.
i C. e
Sht. 5 0-F1(
165
-; L1
ýA` -7
Els,.
54
ZOO. C10-1 . 67
C-4'
1
11
"Cit.D. Lic-0
T. 1
ZEE
(f
-17.42
E;7
' :a I
1-4
Cit. 2 S7.7
60cl. C'(10 ý4. Cie
- -c. 61.7 &':
33
C.. .
7 CO. C,0.1 I;?. (10 -26?. 110 c "C? C'.69.
*:-,. I Iz. Cl Z, 2-1-Bl. C.
MCI. CIC-Cl C7 CI-Irl-,
E. E2 871.v '. (-,'I C".5o.
222.4 E7. --C.
? Ocl. 00C, 65 ý
-'S .t0 CU2. C,
1COO. CIVEJ 21C... 4 -lE . 75 27.1 10 7 ?. 9.11 -IS 1. C,
I MI. C:C.r-! 211.7 It. 7*,% E5. 77 `72. C.
12 _E116t
14 C,Q. C,(10 20 2 E-7 6f
E. ". .E '67 CVC. (
E-CO.CICICI 1 -29ý4.1
-9.7.0 6.7
20cl. 1 15. CIS 28, CIE
E;Co. C, .1 . 73 56. G -? . 9--
2'
E . ý-'
ES 1S5. So 14.87'
-4. rl_,
c",.I -. 09 51 .S -I (I. CIS, -ý-? . E.
*Cfll-i 11CII.
CE.. Sti-EP IS7!
TEAM LEAD P-41 I"-OLF!TE! i PPIPO
ý'=Vrx lhý'N: ýDEG
I=IE S F"FSTIZIERS
Sht. G a; fG
166
Ir ._II _. _i
Z.7,r,
27 f' E.
ell
ý4 .
-" . 4::
Zt. 2.7
.-. I-- 7-1
-.. -. ., 0 -. , 9. %-
I rjclo. 0 Ou _c _J.
IC-16.rl
_', _p
-4.S3 *-13.77
Z.C0r: QCl -..
1.4 ict'. 7 -1.03 te. "i -1 ýV 11:. Q,
-. . g- %0,
COO ,
-.
11 Ir$11.e -I ...
-2. ict E-e. 6
I C.00. (I Cis) -5. C-48 184.2 -17.37 - S.CIO.5 07 -4 te
i COC-1. -4.91 I El. 71-j Ci1 1.
-16.1- r S:
C11
NO. I
I I'D 28
.
1Sc,0. C,C,CI S 17.
S1?9.911.:
-. 1.1.:-: Oto
1 49
-4. 5111-71. -16.78 ' S-4.8 12. CEO
.1
Sht. 7 o; iG
167
ECITH I-ILAIEWT-EP197,c,
BEAN LEFill P/*Ii PEGLAIED PAIFS
1=lr- S FPFRETEF-C III YI,EC
-'t'!
SEEP ',';.'-Et: ITTER; ENITTER WIDTH 2
QOLTS= .5 CUFFEW 1.: Fi= 27
FFECi sle C'21 I;
1 c-Ci.c 111Q-1 V5 -29.3 1 '.
" -. -"
(". .--
.; '(
.-
sht. 8 oý f C-D
168
2cilFi rio(Erir-Ep
LF-All P/fi ISOLPIED PFIIPS
V=tIrr I=I rS F'FiPf4-IETEF.
S IN DI-9DEG
Sht. IG
.
169
- Sh-t. 10 o; iG
170
ca-Clti
! 'EFit-I LEAD P, 11 I SCILAIED PAIPS
Q=VCE I=I E: S FKPI-VEPI-ý S 1h lil-pliEG
sht. I ia
171
ca-0-THNOI.ENBEP IIS-75
HAM LEfir, P/h I,'-.CUiIE: PPIPS
1)=I, lr-E: I=IF- ',-' FF.P.PPIVERSIN rjivr: EG
5ht. fZ o; IG
172
Z*
iffit-i LEfill FVI'. I'SCUIElf FATFS
Q=Qct I=Irr. 'I-* PF.F.F.t-EcIEFS M M: 91EC
5ht. Q3 c-F fG
,
173
01
&. F: I
H. F:h LEAD F',,h J'C(-U-.-iE1' FAIPS
FF-PENEIEF-S. ill 111-ji, EG
E BEP 1? 75
SEAM LEFID Peff ISOLMED PAIPS
Ll=',Icc I=IE S F'FiF:
f-.tIETEF.S IN Dr-illEG
Sht. 14 o; IG
174
Sht. Is0; 1G
175
I-OkIENTER 1$75
c*-C-17ý,
MEN UK' P.-*tl ISOLME! FAIPS
S FPFF.t-!EIEF",-. IN : 11'tI;EC
Ef-ITTEF;
12121,1Q. El-,
1177E.P wirilH 2 ttiuota)
CLTF.EhT fiFi= 21
FPEC! Si I S21 SC112
C.O. I-Xict -104.7 - 37,
C,:
9 2:4.44 11 -2.11? _C"erZ.
Oct.ciclu I 11 C.). 9
(I(-C, S"', 2 1 64 10-1.4
4-c-cl.
204.0
.4
. -, C. r.,-. ,4
4-
C;c1c, -5.74
17.11'
S44.7
r 00. (10.1 ECK,. 0-: I e". ; Cl 8-9.?
,, 19.2
600. c1clu 47 IS,7.0 -, r -ES2. -Ial 179 - 273.
i cl.
c, (10-1 12.46 -2777.3
EOct. ciCIO -In. IC16 -2E. 71 69.4 11 cl 1
- IS,
0.8
SC,0. C,CIE, -5.13 18,9.5 - 25 . 25 4-2.S' 1 C-6r--0 77.1
12.8.2 -
oc,C'.C,CIO -5.27 -24.74 -72.1 9- 08 76.1
MOO. CICIC, -r , 1. Elt, 7*. E.4 71 'ri
.
1; cl0. ED CIEl -5. EICI 126.1 -22.22 73.4 66 3
1i CIO.Q00 1S'5- C'- - al, 3,3 -287.1 5. Ci-'-- -295.2 .
-5.13 .
1'1:110. C,CIO 183.4 27 72.7 el i. i
-4.78 -F-Cf. -4.24 6(1.9
19199.9 95. -4.7 1 -19., 0 72.2 3.25 501.1
0 -53.7
IG
176
APPENDIX 2
International Conference on
ComputerAided Design
and Pilanulacture.
of ElectLrnnic
Components,Circuits and Systems
3-6July 1979
Organisedbythe
Electronics Division of the
Institution of Electrical Engineers
Venue
Universityof Sussex
178
University of Leicester, U. K.
S1 S12 S21 S.
.2
Frequ 0 ncy modulus Phase Modu I us Pha s 0 %todulus Ph*% : Modulus Phu::
MHI dB degrees dB degri ýcs (III dog re ' 45 egr ,
S S12 S21 3
22
Fr*qu e ncy modulus Phase Modulus Phas: modblus Pha : : Modulus Phase
MH dB degrees da deg re s dB degr , dB degrees
01 a 4.43 x 10- 5
G3s 2. X x 10-Is 223 2.72 x 10-'S C2z6.19x 10 gF LZz 0.16gH
C,,. 1.1S A 110"S 0- 2 C4 xV uF
Uj-;
C4,-3 3.7 3xV'gF
Algorithms supplement
riots bYthe EdRor The pair of equations A-0 and A-0 have three solutions, the
AlSonthm No. 107 is publisW i out AjFCCI t with the arproximate locations at (0,0), (1.5,3.1) and (1.5, -2.5) having
Institute of Mathemtscs &M IU Appliations. and is snociated been obtained by a global search procedure. The WS procedure
with the authon' article in the JIMA. Volum 24. Number 1. was used to obtain a refinement of the solution for which the initial
Aug= 1979. approximation is xt - 1.5. xt - 3.5. A zone size z-I was chosen
so as to exclude the neighbourhoods of the other two solution
AlPrithm 107 points. The precision was specified by ACC - 10-4. The WS
A WEJG117ZD S1.04PLEXPROCEDURE FOR THE SOLLMON procedure achieves the exact solution at (Z4) in six iterations. the
OFS114ULTANEOUSNONLLVEAX EQUA71ONS same number as is required by the NR procedure from the same
W. I- Prke and. W. Dowsm starting point and with the sameprecision.
10 CONTIMA 60 CONIII#UE
DO 40 Jal. b 70 CONTME
jpl. j*l U(NPI. k? 2)su(bpl, l; p2)/U(Npl, Npl)
40 20 K. I, N L-kpl
ti. Np,r
DO*90 Ktal, N
AW IS RMON hUht: 01 UFMMICR hIlk Mhhl AKG4hihT
TO
90API-Kk
NANKING (A-IrChOU %4&hLF,. Wtu VAL6LS IN hAhQ& 0.0 1.0
1.0.0
KFI. K*l
V(Jpl, K)21(K) DO 90 JoKpl. tpt
20 COI-TlhtV fiRB*U(K. J)*U(J, NFf)
CALL fbhCI(X, F,; -) so CONTME
DO 30 Ksl, V VK, I.P2)s(U(L, NP2)-l0/Q(I. K)
KfhýK+N
VCJPI, LFN)SI(k) FIbD k AND L
30 CONTINTL
40 CONTINUE It (U(e. K?2). LT. U(L, NP2)) LsK
IDII If (U(K, KP2). GT. U(h, N?2)) K-K
go CON11hum
#;L*;? t;Tb, WEIGhIED MUM, I RETURN
END
50 CALL 3UBhOUTINE fUhCT(X. F. N)
00 60 ksl, N MIkSION X(N). F(h)
l(K)SO-0 l0)s2-Q*l(l)'*3*X(2)-X(2)*113
Cri 60 JOI, Npl F(2)-6.0*X0)-1(2)0*2+1(2)
X(K)@l(K)+b(jdlp2)bv(j, K) MORN
60 COMILUE ED
k6VALUAIL &I X AND ILbl tch COt.VLhGbJ.CL
00 100till'k dm
an b* 6xsý
IGO LONlIbUL
is carried out efficiently by meansor Gaussianelimination. In the
-COTO 50
following we assumethat pivoting is not necessary. The idea behind
LND the algorithm is not new ýSprague,1960;Leavenworth. 1960).but
SUEbOt,Tlbl&
wehavesuppliedthe procedurefor easyreference.
#A*.PUTrb 11LIGMS AhL FIND MUST POSITIVi 6EIGNI, k The operation count is: 2n -I divisions, 3n -3 multiplications
AWLoLEAST P03111VE VALIGH1, L and 3n -3 additions. The only new feature in our procedureis a
i; ImLNSIC#; V(NPl, h2), U(hPl, kP2)
slightly more efficient use or simply subscripted variables. The
number of such referencesis lOn -5 compared to 13n -9 in
IkITIALISik U ARNAI Leavenworth(1960)and Sprague(1960).The solution is returned
in array D and array B is destroyed.
00 10 tal, h?2 The reasonfor this very detailedoperation count is that on most
u0,10.1.0
10 cohTINUE computersdivision is slowerthan multiplication. and floating point
DO 20 K-2, týPl addition is not so much faster that it makesthe time ;ns:gnif cant.
u(g. 6P2). 0 a Subscriptedvariablesare countednot only becauseof the subscnpt
20 CONT111. Ii2
DO 10 J. I. t. handlingbut alsobecausetheyimply memoryreferences. In cor.trast,
jvl. j*l the simple variables can (and -should, if possible) stay in Ust
jph. J. h
to 0 Kal, Npl
registers, of which most modern computers have sufficiently
many.
30 If severalsystemswith the samecoefficientmatrix are to besolved
subsequently,we can store the intermediateresults from the LU-
Ck)h*lijlk. ý.ElGhTS decomr)ositionin array A for later use when the right hand sides
become available. The operat-on count for subsequcntsystems
IEE
DIVISION
ELECTRONICS
COLLOQUIUMON
FORCIRCUIT
PARAMETERS
'MODEL
IS
ANALYS
ORGANISED BY
PROFESSIONALGROUPE10
(CIRCUIT THEORYAND DESIGN)
AND PROFESSIONALGROUP E3
(MICROELECTRONICSAND SEMICONDUCTOR
DEVICES)
FRIDAY 27 MARCH1981
I
DIGEST No: 1981/25
ON MDEL MODIFICATION
CONSIDERATIONS
Monica Dowson
1. Introduction
For any given device a model based on the physical aspects of the
device is normally available where perhaps some of the elements are
given fixed values but most of the element values are unhnown. These
element values must therefore be optimised to provide a fit to the
required characteristics.
.:).. I.
iss
similar argument may be used for corrections that are large and positive.
Fig. 1 shows a suggested low frequency physical model for a vertical npn
transistor operating at frequencies of 0.1 to 1.0 CH., After applica-
.
tion of the technique described here the model given in Fig. 2 gave coývergence
in Guass Newton over the entire range of frequencies with an attendant
improvement in the high frequency response of S n particular, as shown
12
in Fig. 3.
Where a model is required that gives close agreement with the measured
characteristics, the requirement of adding only physically based elements can
be an unnecessary restriction particularly if the view is taken that any model
is only an approximation to an infinite connection of elements. In order to
be certain that each element added is the best one at that point in the model
growth all possible placements must be attempted. This can be most time
consuming and some restrictions are usually necessary. Thus, to add elements
without creating a new node one can either attempt to add elements in parallel
with elements already present in the model Can approach most physicists would
not find too unorthodox) or attempt to add elements between each combination
of nodes. To add a new node, the simplest way is to attempt to add one
element in series with an existing element although, of course, this does not
cover all possibilities.
Results
Fig. 4 shows an example of a model grown from the reduced model in Fig. 2.
The frequency response of modulus S. for this model is shown in Fij. 5 and
1
in Fig. 6. - As can be seen this model now gives
of modulus and phase S,,
good agreement with the 'required characteristics.
References
2. D-
186
-. S 3 i
". -
a 'Z cm
1A
6t
IS
I 'o
\ 06
S
S
.5
00
CY I'll
LM
PM
2 ,a 0
2
C2 'o -V cy
£
0
I, 010!..
.4S. a
-
-
OT
-ii
I I
I
2.31
187
9 1
46 r. a
ý0 «so
-0
§ r4 Cm ml mi ýr 0800
. ýQ ýr tý4 CD
2
ýo ýr r14
1"
S
I
IU
h
Lcr-
A. J"
- *J
*
Ix
2-. Lm
188
NC0MFECE
EUR0PE ko'-A..
on ELEC-7FO'MICDESIGN
AUTO'wý"IQATIOM,
1-4
September 1981
Organisedbythe
Venue
University of Sussex,
Brighton, United Kingdom
189
24
EQUIVALENT CIRCUIT MCDELLING or A SMALL SIGNAL HIGH FREQUENCY BIPOLAR TRANSISTOR AT DIFFERENT
BIAS CCNDITIONS
Monica Dowson
University of Leicester, UK
The transistors being modelled were bipolar Thus. taking the model as the basis of a bias
transistors designed to be used in beam lead dependent model, the optimisation algorithm
integrated circuits. S-parameter data for described above was then used to obtain suit-
the transistors was available in the fre- able element values at different values of
quency range 100 MHz to 2 GHz for several emitter current.
different emitter currents.
BIAS DEPENDENT MODEL
Using optimisation techniques, the aim was to
produce accurate models simulating the The model shown in Fig. 3 gave good results for
measured S parameters of the transistors. both type A and type B transistors over
with as few bias dependent elements as the entire range of frequencies at emitter
possible. currents within the normal operating regions
described above. However, at currents in
A brief description is given here of the opti- excess of the maxima stated further changes in
Mi33tiOn techniques used and the results ob- the topology of the model were required.
tained are illustrated by diagrams of the
model together with the S parameters of the The original hypothesis was that there would
model compared with those measured for the be a small number of linearly varying elements
transistors. * in the model that would describe the bias
variation. Within the normal operating.
DESCRIPTION OF THE TRANSISTORS regions this was found to be a valid assump.
tion for this particular model. For example.
The transistors being modelled were isolated as is suggested by the graphs in Fij. 4, R
n-p-n transistors made in silicon with im- is inversely proportional to the emitter
planted arsenic emitters and diffused boron current and C, can be approximated by a
bases (Slatter (1)). The first type (A) had function of the form
one emitter stripe and the second type (a)
b
had three emitter stripes. The data for
these transistors was supplied by Philips
Research Laboratories (Slatter (2)) and con-
S parameters where a and b are constants.
sisted of the of the transis-
tors in common emitter configuration at 12
frequencies 100 MHz to ; GHz Taking the model beyond the normal operating
in the range
at emitter currents of O. S MA to 10 MA for regions. evidence of the type of topology
O. S mA to 27 mA for type B. changes required is given by C, which in
type A and of
Fig. 4 shows a rapid increase ,
in'valut as the
From the graphs in Fig. 1 it can be deduced emitter current rites and which the optimisa-
A up tion algorithm deletes open circuit at higher
that the normal operating range of type
to I CHz was O. S mA to 2 mA and of type values of current. and by R2 which is later
B was I mA to 8 mA deleted short circuit.
.
RESULTS
MODELLING TECHNTQUE
Examples of the S parameters calculated for
The modelling process was started by working
the model compared with the measured parame-
on type A only. Using an initial model as A and type B transistors
ters of type are
shown in Fig. 2. an optimised model was ob- in Fils. 5 and 6. As can be seen these
for the whole of the given range of given
tained frequency
frequencies of emitter give good agreement over the entire
at a single value 100 Milz 2 Gil: is a
range to and there
current within the normal operating region of in cases
This shown in smoothing effect where there appear
the device. optimised model,
by using to be measurement errors.
Fig. 3, was obtained successive appli-
cations of Gauss Newton to minimise the
CONCLUSIONS
weighted errors between both the calculated
modLius and phase of each of the S par&-
The bias dependent model obtained shows sever-
meters of the model and the giYen measurements
to the one described by
The correction al similarities
of the device. terms calcu- Slatter (1) which by a different
lated by Gauss Newton were used to indicate was produced
were required, approach to the problem.
where topology changes conver.
gence in Gauss Newton being required at each
The model described here simulates the trans.
stage in the development of the model.
istors being modelled over a wide range of
190
25
ACKNOWLEDGEMENT
REFERENCES
FT(MHz)
3000
2000 Hill.
1000
niiii ---Fi
1IV. O-sv
2-- -
-7-
Mod 0-
S21
dB at
1 GHz -2,
\yp eA Type b
-T,
ilu 10.0 1000.0
Lig 1 T-cL.
nsis2r chcrc: "Mr-,:tics --t, 1GH le(mA)
191
26
Cl
I 2
Ri RI
1 2
C4
Rin
C,pF i-
2400 .-1-
.00
2000
1600 0-7-
1200 0.5
800
0-3
4w
4-
02468 le mA 02468 lemA
R20 C2nF
so 0.04.
0-03-
40
30 0,02-
20. 0.01
10 0-00
0 lemA lemA
-Type A
Type 8
Erg,. ýcmoesof bias de eriddent
L E? etementvalues
_p_
192
27
lu , lu
Freq GHz Freq GHz
-1
-I
28
10 , lu lu, lu FreqGHz
Freq GHz
160
140
120
-100
80
REFERENCES
Leicester.
S. Wright, D. J. (1974), "Studies in Optimisation as an Aid to Circuit
of Leicester.
Redhill.
Mifflin Company.
6 November 1977.
pp. 78-88.
pp. 29-54.
51. Price, W.L. (1977), "A Controlled Random Search Procedure for Global
52. Price, W.L. (1978), "A Controlled Random Search Procedure for Global
56. Nelder, J. A. and Mead, R. (1965), "A Simplex Method for Function
57. Price, VI.L. (1979), "A Weighted Simplex Procedure for the Solution of
.
Simultaneous Non-Linear Equations", J. Inst. Maths. Applics. Vol. 24,
pp.
199
pp. 182-184.
63. Fletcher, R. and Reeves, C. M. (1964), "Function Minimisation by
67. Levenberg, K. (1944), "A Method for the Solution of Certain Non-Linear
75. Callahan, M. J. jr. (1972), I'Models for the Lateral P-N-P Transistor
pp. 122-123.
78. Numerical Algorithms Group (1976), IIE02ADFII, NAG Fortran Library Mark S.