PARAMETER ESTIMATION USING MICRODIELECTROMETRY

WITH APPLICATION TO TRANSFORMER MONITORING

V.l '\.
by

MARK CAMERON ZARETSKY

S.B., Massachusetts Institute of Technology

( 1980)

S.M., Massachusetts Institute of Technology

(1982)

SUBMITTED TO THE DEPARTMENT OF

ELECTRICAL ENGINEERING AND COMPUTER SCIENCE
IN PARTIAL FULFILLMENT OF THE REQUIREMENTS
FOR THE DEGREE OF

DOCTOR OF SCIENCE
at. the
MASSACHUSETTS INSTITUTE OF TECHNOLOGY
November, 1987

@ Massac.husetts Institute of Technology, 1987

J . J

Signature of Author
'pepartmeft of Electrical Engineering
and Computer Science

Certified by
::.:~ra~)I·of'~~I", EE(:S
."./ l ' SlIIH'I'\' I :;;o r
..-- ," .'

Accepted by
ettainnall, Depart.ment.al COllllllit.t.('{'
on Graduate Student.s
Archives
~!ASS.~HUScITS 1!f!!Tmrn:
OF rrrlf/lillOGY

MAR 221988
IJIWJIa
\!.,L J
 PARAMETER ESTIMATION USING MICRODIELECTROMETRY WITH
APPLICATION TO TRANSFORMER MONITORING
by
MARK CAMERON ZARETSKY

Submitted to the Department of Electrical Engineering and Computer Science

on November 13, 1987 in partial fulfillment of the requirement for the Degree of
Doctor of Science

ABSTRACT

Two related objectives of this work are to develop the capability of inferring fun-
damental continuum parameters of semi-insulating heterogeneous media using an in-
terdigital electrode microdielectrometer and to utilize this capability to provide new
sensors useful for transformer monitoring. Using a modal approach, a continuulll
model is derived for describing the electrostatic field distribution associated with an
interdigital electrode structure. In this "imposed '''' - k" technique, the medium is
excited at the-temporal (angular) frequency w by means of an interdigital structure
having a spatial periodicity length ,\ = 2w / le and hence a dominant wavenumber k.
Given one fourier component of the normal displacement field in response to an applied
potential, expressed as the surface capacitance density C(w, k), of any linear system
having property gradients perpendicular to the plane of the electrodes, the model
predicts the resulting complex gain taking into account the properties and geometry
of the interdigital structure. Root searching techniques incorporating the continuum
model, using either a secant or Newton's method, are then developed for solving in-
verse problems and hence estimating continuum parameters from experimental data.
Specifically illustrated are absolute measurements of the thickness of deposited poly-
mer films (0-20 I'm), the dispersive complex bulk permittivities of plasma deposited
bromobenzeup., parylene, paper insulation and transformer oil, and the dispersive C0111-
plex surface permittivities of a tin oxide thin film (350 A) and moisture adsorbed ou.
the surface of Si0 2 • These measurements are made using electrode structures of 50 I'm
(microchip) and 1 111111 (macrochip) spatial wavelength. In addition, an algorithm has
been developed for estimating spatial distributions in heterogeneous media utilizing
data obtained by varying the spatial wavelength, rather than temporal frequency, im-
posed by the interdigital electrode structure.
These capabilities are t,hen applied towards developing Sf'nsors for hot.h oil complex
bulk permittivity and oil moisture content. Passivation of t.he mirrochip and macl'Ochip
electrodes with a 5-7 I,m layer of parylene is necessary for pt'event.ion of t.he adsorpt.ion
of moisture from the oil. Consequently, a prel'equisit.e for accurat.e measurement, of t.he
oil complex bulk permittivity is the abilit.y to take into account the finit.e thickness
parylene layer and its dispersive complex bulk permittivit.y. This capability is provided
by the previously mentioned continuum model.
Sensitization of the microchip to moist.ure cont.ent. in the oil is accomplished by
plasma depOSIting a thin (2-7 I'm) film of bromobenzene. This film undergoes changes

2
in its dielectric properties that reflect the adsorption of moisture. The observed fre-
quency response can be normalized t.o external parameters such as moisture content
and temperature. TIlls allows the coating sensitivity to be characterized by one pa-
rameter, chosen to be the frequency at which the phase peak occurs, fp. It is shown
that the coating behaves as a semipermeable membrane, serving as a barrier to the
transformer oil but absorbing moisture proportional to t.he absolute moisture con-
tent in the oil. Sensitivity to moisture contents of less than 5 to roughly 50 pplll is
demonstrated. The transient response of the coated microchips is on the order of a
minute. Preliminary on-line monitoring results provide a real time demonstrat.ion of
the change in moisture content with load cycling of a 50 kVA test transformer. Mois-
ture permeation and electrical conduction within the coating are examined in terms of
the heterogeneous structure of the coating. The aging of the films on a monr.hly time
scale, resulting in decreased moisture sensitivity, is characterized and discussed.
A preliminary effort is made to develop a hydrogen gas sensor using a t.hin film
(500 A) palladium resistor. Changes in bulk resistance in response to both 100 and
1000 pplll H2 in nitrogen at temperatures of 100° and 225°F are observed.

Thesis Supervisor: J ames R. Melcher

Title: Stratton Professor of Electrical Engineering and Computer Science

3
ACKNOWLEDGEMENTS

For many years now, I have had the pleasure of both working for and learning
with Professor Melcher. It has been a long and slippery path, beginning with corn oil,
passing through ferrofiuids, diesel soot (with motor oil), and ending in transformer oil.
The experience has been very stimulating and rewarding (not to mention dirty). His
impres::;ion on both my academic and personal pursuits will be a positive and lasting
one.
This work was supported by a consort.ium of several electric utilities. I would like to
thank them for the OpPOl·tunity to perform scientific research and apply the knowledge
gained to the practical application of transformer monitoring. Thanks to Phil Gattens
for providing the moisture content data presented in Section 6.2.2.
The constructive criticisms and support of the thesis readers, Professors Stephen
Senturia and Alan Grodzinsky, and Doctor Chathan Cooke, was greatly appreciated.
Each reader provided their own unique insights, resulting in a much improved docu-
ment.
Bridge measurements of transformer oil samples, as well as insights into dielectric
behaviour, were provided by Bill Westphal. Avi Benatar was very helpful in learning to
operate the plasma deposition apparatus, obtained on loan from Professor Gut.owski.
Huan Lee and Dave Day, both of Micromet Instruments, were ve-;y generous with their
time - helping in the installation of the microdielectrometer and answering questions
about its operation. The thin film palladium resistors were fabricated by Professor
Mark Wrighton's chemistry lab and Tracy Jones was instrumental in this effort.
Eliot Frank has consistently been a life, I mean computer, saver. He figured Oll t
how to talk to the microdielectrometer in Xenix and his software provided most of t.he
plot.s in this thesis. The data taken in the test facility would not have been possible
without the herculean efforts of Wayne Hagman and Dan Flagg. Testing of t.he t.hin
film palladium resist.ors was perfurmed by Paul Warren, anot.her person I have known
and enjoyed working with for many years. Many friends in the lab have been helpful
over the years, both on an intellectual and social level: Lama Mouayad, Philip Von
Guggenberg, Kevin DelhI, Paul Grimshaw, Patrick Li, Silvano Brewster, Ric Carreras,
Steve Bart and Karen Kirmse. Thanks for bearing with me, folks. A special mention
goes to ABC basketball and Out-of-Towners softball.
My parents, Carole and Herb, who have made it too easy for me, have provided
the role models from which I draw my strength upon. This accomplishment is as much
theirs as mine. Finally, to my girlfriend J udy, I gratefully acknowledge your patience,
understanding, support and love.

4
Contents

Abstract 2

Acknowledgement 4

Contents 5

List Of Figures 10

List Of Tables 21

1 Introd uction 23
1.1 Background - Engineering and Science . 23
1.2 Scientific Motivations .... 23
1.2.1 Illlposed w-k Sensing. 24
1.2.2 Intimate Sensing . . . 27
1.3 Engineering Motivations - On-Line Transformer Monitoring. 28
1.3.1 Background . . . . . . . 29

1.3.2 State of the Art Review 32

1.4 Microdielectrometry . . . . . . 44
1.4.1 Description and Background -l-l
1.4.2 Micro and Macro Sensors -!()

1.4.3 Preliminary "Vork -l7

1.5 Goals 47
1.5.1 Analytical and Numerical Tools . . 47

5
 1.5.2 Experimental Verification of Model . . . . . . . . . . . 48
1.5.3 Electrochemical Sensors For Transformer IVlonitoring . 48
1.6 Overview . . . 49

2 Continuum Model 52

2.1 Outline of Approach 52

2.2 Electrode Structure Boundary Conditions 55
2.3 Bulk Relations . . . . . . . . . . 57
2.4 Potentials at Collocation Points . 58
2.5 Equivalent Circuit Parameters 65
2.6 Determination of Gain . . . . . 66
2.7 Collocation Points, Fourier Modes and Connrgence 67

3 Typical Responses Using Continuum Model 69

3.1 Predicted Responses . 69

3.1.1 Unifo:m Media 70
3.1.2 Variable Thickness Layer 74
3.1.3 Singular Properties at Substrate-rvledium Int.erface . 76
3.1.4 Layer With Surface Conductivity . . . . . . . . 79
3.1.5 Sensitivity to Electrode Structure Parameters. 80
3.2 Working Theorems .. 82
3.2.1 Master Curves 82
3.2.2 Spatial Discontinuities 86
3.3 Comparison With Analytical Solutions 91

4 Parameter Estimation TEc'chniques 103

4.1 General Algorithms. . . . . . . . . . la·!
4.1.1 Single Parameter - Serallt. Met.hod . 104
4.1.2 Multiparameter - Minimizat.ion of Quadratic Error Function . 107
4.2 Single Parameter Estimation . 108
4.2.1 Layer Thickness ... . 109

6
 4.2.2 Complex Permittivity (Bulk or Surface) . . . . . . . . . . . 111
4.3 Mll1t.iparamet,er Est.imat.ion Using Variahle Spatial 'Vav~h'ngfhR . . 118
4.3.1 Motivation .118
4.3.2 Apparatus. . . . 122
4.3.3 Estimation Algorithm .123
4.3.4 Smooth Distribution of Complex Permittivity. · 129
4.3.5 Step Discontinllities .. . . . . . . . . . . . . . .141
4.3.6 Layer Thickness and Complex Bulk Permittivity .. 141
4.3.7 Complex Surface and Bulk Permittivity . . . . . · 145
4.4 Two Parameter Estimation Using Variable Dielectric Media . .153

5 Parameter Estimation Examples & Oil Complex Permittivity Mea-

surements 156
5.1 Calibration · 1.57
5.1.1 Microchip · 1.57
5.1.2 Macrochip. . .158
5.2 Thickness Gauge . .158
5.2.1 High Impedance Measurement · 159
5.2.2 Virtual Ground Measurement .165
5.3 Layer Thickness and Permittivity . · 167
5.4 Complex Bulk Permittivity of Oil . .168
5.4.1 Uncoated Microchip . .170
5.4.2 Passivated Electrodes .178
5.5 Complex Surface Conduction .190
5.6 Estimation From t.he Frf"ql1f'n('y R~sponse · H14

6 Moisture Sensing - Concepts and Design Strategy 202

6.1 Equilibrium Concepts . . . . . . . . . . . . . . . . . .202
6.2 Moisture Distribution In A Transformer (Status of Oil and Paper Insu-
lation - An Example . 208
6.2.1 Model.... . 209

7
 6.2.2 Experimental Evidence .210
6.3 Moisture Sensing in Oil - A Design Strategy .211

7 Moisture Sensor - Experimental Results 216

7.1 Testing in Oil Samples . . . . . .217
7.1.1 Experimental Procedure .217
7.1.2 Master Curves For Coated Chips .219
7.1.3 Coating Sensitivity to Moisture in Oil .226
7.1.4 AC Electric Field Strength Dependence .233
7.1.5 Temperature Dependence .238
7.1.6 Tests with Aged Oil .240
7.1.7 Transient Response. .242
7.1.8 DC Offset . . . . .246
7.2 Testing in Pilot Facility .251
7.2.1 Experimental Results .251
7.2.2 Discussion....... .258

8 Moisture Sensor - Interpretation of Results 263

8.1 Film Morphology - Effect On Moisture Permeation and Electrical Prop-
erties .264
8.2 Dielectric Response of Plasma Deposited Bromobenzene Films . 270
8.2.1 Attempts to Model the Master Curves .270
8.2.2 Low Frequency Dispersion. . . . . . . . .283
8.2.3 Parameter Estimation of the Master Curve .284
8.3 Film Aging and Electrical Conduction rVlechanis1l1 .293

9 Hydrogen Gas Sensors - Preliminary Research 300

9.1 Motivation .300
9.2 Tin Oxide Coated Microchip .302
9.2.1 Background . . . . . . .302
9.2.2 Experimental Results .304

8
 9.3 Thin Film Palladium Resistor. .309
9.3.1 Background....... . .309
9.3.2 Experimental Results .309

10 Conclusions and Future Work 316

10.1 Continuum Model .. .316
10.2 Parameter Estimation .317
10.3 Complex Bulk Permittivity Of Oil .318
10.4 Moisture Sensors .319
10.5 Gas Sensors . . . .322
10.6 Electrode Terminations .. 324
10.7 Double Layer Interactions .324
10.8 Applicatiolls to Other Physical Domains .325

A Rapidly Convergent Series 326

B Surface Capacitance Density of Multiple Layers 328

C Plasma Deposition 332

C.1 Descript.ion of Polymerization Process .332
C.2 Experimental Apparatus . . ~37

C.3 Operational Procedure . . . .341

C.4 Experimental Observations For Coating Microchips. .344

D Experiments With Doped Transformer Oil 348

E Data Acquisition and Processing Routines 354

F Continuum Model and Parameter Estimation Routines 427

Bibliography 584

9
List of Figures

1.1 Representation of Interdigital Electrode Structure and Associated Cir-

cuitry Interacting with a Heterogeneous Media. . . . . . . . . . . . .. 26
1.2 Apparatus for Measurement of Relative Saturation of Oil (from [12]).. . 34
1.3 Electrochemical Fuel Cell for Detecting Hydrogen Gas (from [31]). 36
1.4 Diffusion Cell for Detecting Hydrogen Gas (from [35]). . . . . . . . 36
1.5 Effect of Reducing Gas 011 Work Function (V barrier) at the Grain Bound-
ary of Pd Doped Sn02_z Having Surface AdsOl bed Oxygen. . . . . . .. 38
1.6 Microfabricated Gas Sensing Diodes, Capacitor and Transistor Utilizing
Pd-Oxicle Interface. . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
1.7 Hitachi On-Line Gas Sampling and Monitoring Syst.em (from [62]). . 43

2.1 Schematic Representation of Continuum Model - Required Inputs, So-

lution for Collocation Potentials, Two Port Network Equivalent, and
Complex Gain Output. 54
2.2 Coordinate System Used for Electrode Structure of Fig. 1.1. . 56
2.3 Voltage Distribution Along Discretized Interelectrocle Surface for k=4
Collocation Points. Unknown Volt ages Vj are Introduced at. the Points
Yj. Conservation of Charge is Maillt.aine'cl for Line Se'gme'nf.s Oe'finf'cl hy
the yj. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 60
2.4 Distribution of Tangential Electric Field, E:(y), Along Electrode-Medium
Interface. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 62

3.1 Electrode Structure Below a) Uniform Medium or b) Variable Thickness

Homogeneous Layer and Uniform Medium. . . . . . . . . . . . . . . .. i1

10
3.2 Predicted Responses for Microchip in Uniform Medium, Vary Conduc-
tivity 0'1 (€1 = 2:dO- 11 F Im). . . . . . . . . . . . . . . . . . . . . . . .. 72
3.3 Predicted Responses for Microchip in Uniform Medium, Vary Permit-
tivity €1 (0'1 = 10- 9 S/m). . . . . . . . . . . . . . . . . . . . . . . . . .. 73
3.4 Predicted Responses for Microchip with Variable Thickness Layer in Air
(f2 = 2xlO- 11 F Im, 0'2 = 2.4xlO- 11 S/m). . . . . . . . . . . . . . . . .. 75
3.5 Predicted Response for Microchip in Uniform Medium with Surface
Conductivity on Illterelectrode Surface (€1 = 2xlO- ll F Im, 0'1 = 10- 11 Slm,
0'/10 = 10- 14 S). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 77
3.6 Predicted Responses for Microchip in Uniform Medium, Vary Surface
Conductivity on Interelectrode Surface (f1 = 2x10- 11 F fm, 0'1 = 10- 11 S/m,

0'.0)' • • • • • • • • • • • • • . • • • • • • • • • • • • • • • • • • • • • • •• 78
3.7
face Conductivity in Air (d 2 = 5 ItIll, €2 = 2x10- 11 F Im, 0'2 = 10- 9 S/m,

0'/12) •• • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • 81
3.8 Electric Fidd Distribut.ion as a Function of Oxide Layer Thickness. . 83
3.9 Predicted Responses of Fig. 3.2 Normalized to Frequency of Phase Peak,
Jp. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 85
3.10 a) Multilayered Structure with Spatial Distribution in Complex Bulk
Permittivity and b) Predicted Response using Microchip. . . . . . . .. 87
3.11 Predicted Responses for Microchip with an Insulating Layer of Vari-
able Thickness (d 2 , f2 = 2x10- 11 F/m, 0'2 = 0 S/m) having a Surface
Conductivity (0'/12 = 10- 15 S) in Air. . . . . . . . . . . . . . . . . . . .. 89
3.12 Predicted Responses for Microchip with an Insulating Layer of Variable
Permitt.ivity (d 2 = 1/1.111, f2' 0'2 = 0 S/m) having a Surface COllductivit.y

(0'/12 = 10- 15 S) in Air . . . . . . . . . . . . . ............ 90

3.13 a) Case 1 - Small Illterelectrode Spacing Limit, b) Case 2 - Parallel
Plate Capacitor Limit, c) Case 3 - Microst.l'ip Solution, and d) Case 4
- Long-Wavelength Limit. . . . . . . . . . . . . . . . . . . . . . . . . .. 92

11
3.14 Distributed Transmission Line Approximation to Thjn Film Coating
Microchip in Air (from [71]). . . . . . . . . . . . . . . . . . . . . . . . . 97
3.15 Predicted Responses using Old Microchip (h = 1 /-Lm) having a Surface
Conductivity (0"0 = 10- 15 S) and Variable Surface Permittivity (f,o)' .. 99
3.16 Comparison of Predicted Responses using Transmission Line Model and
Continuum Model for a Thin Film Coating Old Microchip. . . . . . . . 100
3.17 Predicted Responses using Present Microchip having a Suriace Conduc-
tivity (0'''0 = 10- 15 S) and Variable Surface Permittivity (f"o)' . . . . . . 101

4.1 Comparison of a) Secant Method and b) Newton's Method for Root

Searching (from [80, p. 249 and 255]). . . . . . . . . . . . . . . . . . . . 106
4.2 Predicted High Frequency (10 kHz) Gain for Microchip with Variable
Thickness Layer (d 2 ), Vary Layer Permittivity (f2), in Air . . . . . . . . . 110
4.3 Parameter Space for Complex Bulk Permittivity Estimation of a .1) /-Lm
Layer in a Uniform Medium (f~ = 2:dO- ll F /m and f~ = 10- 12 F /m). . 113
4.4 Parameter Space for Complex Bulk Permittivity Estimation of a Ulli-
form Medium. 114
4.5 Parameter Space for Complex Bulk Permittivity Estimation of a 5 I£m
Layer in a Uniform Medium (f~ = 2xlO- ll F /lU and f~ = 10- 8 F /m) ... 115
4.6 Parameter Space for Complex Bulk Permittivity Estimation of a 2 /-Lm
Layer in Air (f~ = fo F /m and f.~ = 0) . . . . . . . . . . . . . . . . . . . . 116
4.7 Complex Permittivity Measurement of Heterogeneous Medium Using
Parallel Plates with Guard Ring . . . . . . . . . . . . . . . . . . . . . . . 119
4.8 Apparatus to Provide Variable Spatial Wavelength in Potential Using
a Set of Switched Electrodes. . . . . . . . . . . . . . . . . . . . . . . . . 124
4.9 Typical Spatial Dist.ribut.ions in Complex Bulk Pel'luit.t.ivity: a) Smoot.h-
ly Varying, b) with Step Discont.inuity, and c) wit.h Complex Surface
Permittivity. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
4.10 Representatklls that can be Used in Variable Spatial \.yavelengt.h Esti-
mation Routill.e: a) Stail'-Step and b} Exponential. . . . . . . . . . . . . 125

12
4.11 Schematic View of Variable Spatial Wavelength Estimation Algorithm
Using Stair-Step Representation . . . . . . . . . . . . . . . . . . . . . . . 127
4.12 Estimated Spatial Distribution Using 6 Spatial Waveleilgths (Case 1). . 131
4.13 Estimated Spatial Distribution Using 6 Spatial Wavelengt.hs (Case 2, d 2
= 50 p.m) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
4.14 Estimated Spatial Distribution Using 12 Spatial Wavelengths (Case 2,
d2 = 50 p.m). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
4.15 Estimated Spatial Distribution Using 6 Spatial Wavelengt.hs (Case 2, d2
= 500 P.lll). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
4.16 Estimated Spatial Distribution Using 12 Spatial Wavelel1gt,hs (Case 2,
d 2 = 500 p.m). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
4.17 Estimated Spatial Distribut.ion Using 6 Spatial Wavelengt.hs (Uniform
Medium, f = 0.1 Hz). . . . . . . , . . . . . . . . . . . . . . . . . . . . . 139
4.18 Estimated Spatial Distribution Using 6 Spatial Wavelengths (Uniform
Medium, f = 1 Hz). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
4.19 Estimat,ed Step Discontinuity Using 6 Spatial Wavelengths (Case 3, d 2
= 50 JLIU) • • • . . • • • • • • • • • • . . . • • . . . • . • . • . . • • . • • . 142
4.20 Estimated Step Discontinuity Using 6 Spatial Wavelengths (Case 3, d2
= 70 JLlll) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
4.21 Estimated Layer Thickness and Upper Medium Complex Bulk Per-
mittivity Usillg 400 and 1000 p.m Spatial Wavelength Data (Case 3,
d2 = 50 JLlll). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
4.22 Estimated Layer Thickness and Upper Medium Complex Bulk Per-
mittivity Using 400 and 1000 ,.,.111 Spatial Wavelengt.h Dat,a (Case 3,
d2 = 70 p.m). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1·4;
4.23 Estimated Layer Th:cklless and Upper ~IerlillJll Complex Bulk Per-
mittivity Using 400 and 1000 I'lll Spatial Wavelellgt.h Dat.a (Case 4.
d2 = 50 P.lll). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148

13
4.24 Estimated Layer Thickness and Upper Medium Complex Bulk Per-
mittivity Using 400 and 1000 lUll Spatial Wavelengt.h Dnt.a (Case 4,
d2 = 70 J.tnl). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
4.25 Estimated Layer Surface Conductivity and Upper Medium Complex
Bulk Permittivity Using 400 and 1000 fLm Spatial Wavelengt.h Data
(Case 5, 0',2 = 10- 16 S). . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
4.26 Estimated Layer Surface Conductivity and Upper Medium Complex
Bulk Permittivity Using 600 and 1000 fLm Spatial 'Navelellgth Data.
(Case 5, 0'.2 = 10- 16 S). . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
4.27 Estimated Layer Surface Conductivity and Upper Medium Complex
Bulk Permittivity Using 400 and 1000 fLlll Spatial Wavelength Data
(Case 5, 0'.2 = 10- 17 S). . . . . . . . . . . . . . . . . . . . . . . . . . . . 154

5.1 Predicted High Frequency Gains for Microchip with Variable Thickness
Layer with Perfect Conductor Above Layer . . . . . . . . . . . . . . . . . 161
5.2 Measured and Predicted Frequency Responses of Microchip wit.h Oil
Layer of Variable Thickness in Air . . . . . . . . . . . . . . . . . . . . . . 162
5.3 Measured High Frequency Gain of Microchip During Deposition of Plas-
ma Deposited Bromobenzene and Estimated Layer Thickness . . . . . . . 164
5.4 Predicted Normalized Conductance as a Function of Normalized Layer
Thickness and Relative Conductivity of Upper Medium for Microchip
with Virtually Grounded Electrode. . . . . . . . . . . . . . . . . . . . . 166
5.5 Measured and Predicted Frequency Response for Uncoated Microchip
in Unstirred, Unused Oil with 70 ppm Water . . . . . . . . . . . . . . . . 171
5.6 Measured and Predicted Frequency Response of Uncoat.ed lHicrochip in
Unstirred, Unused Oil with 53 ppm \Vater. . . . . . . . . . . . . . . . . 172
5.i Measured and Predicted Fl'e(luent"Y Hesponse of lIumatetl Minodlip in
Unstirred, Userl Oil wit.h 50 ppm Wat.er . . . . . . . . . . . . . . . . . . . 1 j 4.
5.8 Measured Frequency Response of Uncoated Microchip in Unst.irred, Un-
used Oil at 14 ppm Water. . . . . . . . . . . . . . . . . . . . . . . . . . 1i5

14
5.9 Measured Frequency Response of Uncoated Microchip In Uustirred,
Evacuated, Unused Oil at roughly 10 ppm 'Vat.er. . . . . . . . 176
5.10 Measured Frequency Response of Uncoated Microchip in Unstirred, Un-
used Oil at 70 ppm Water. . . . . . . . . . . . . . . . . . . . . . . . . . 177
5.11 Effect of Attenuating Applied Potential on Measured Frequency Re-
sponse of Uncoated Microchip in Unused Oil. . . . . . . . . . . . . .. 179
5.12 Description of Parylene Coating Process a.ld Molecular Structure (from
[85]). . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
5.13 Measured Frequency Response of 6.7 J.Lm Thick Parylene Coated l\ii-
crochip in Air. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
5.14 Measured and Predicte(! Frequency Response of a 6.7 J.Lm Thick Coated
Microchip in Unstirred, Oxidized Oil at. 25°C . . . . . . . . . . . . . . . . 184
5.15 Estimated Dispersion in Complex Bulk Permittivit.y of ParyJene From
Measured Response of Fig. 5.13. . . . . . . . . . . . . . . . . . . . . . . 185
5.16 Estimat.ed Dispersion in Complex Bulk Permittivit.y of Oil From Mea-
sured Response of Fig. 5.14 and using Est.imated Dispersion of Parylene
(Fig. 5.15). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
5.17 Estimated Dispersion in Complex Bulk Permittivity of Oil From Mea-
sured Response of Fig. 5.13 using Micromet Values. 188
5.18 Measured Frequency Response for a Parylene Coated Microchip in Stirr-
ed, Unused Oil at 8 ppm Water. . . . . . . . . . . . . . . . . . . . . . . 189
5.19 Measured and Predicted Frequency Response for a 100 A Thick Palla-
dium Coated Microchip in Air. . . . . . . . . . . . . 192
5.20 Measured and Predicted Frequency Response for a \ Thick Tin
Oxide Coated Microchip in Air at 3~,%RII. . ....
5.21 Estimat.ed Dispersion ill Complex Surface Permit.tivity using l\Ieas\Il'ed
Response of Fig. 5.20. . . . . . . . . . . . . . . . . . . . . . . . . . .. ] !Jr,
5.22 Parameter Space for Complex Surface Permittivity Est.imat.ion of a. Mi-
crochip in Air. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196

15
5.23 Measured and Predicted Frequency Response of PP-Blomohenzene Coat-
ed Microchip in llnstirrE"d, Oxi.lizr>d (JiI at. fl9°C . . . . . . . . . . . . . . 107
5.24 Measured Frequeh('y Response of Parylene Coated Macrochip with Pa-
per in Unstirrt!d, l.Tnuse~1 Oil at 28 ppm Water (from [23, p. 85]) . . . . . 200
5.25 Estimated Dispersion in Paper using Measured Response of Fig. 5.24. . 201

6.1 Moisture a) Sorption anti b) Allsorptiun Curves for Wood Pulp and c)
Moisture Equilibrium Curves for Paper-Oil System (from [14]) . . . . . . 205
6.2 Moisture Content Versus Inverse Absolute Temperaturf: for Pilot Facil-
ity and Harrison Power Station . . . . . . . . . . . . . . . . . . . . . . . . 212

7.1 Apparatus for Conducting Moisture Sensitivity Measurements with Coat-

ed Microchip Moisture Sensors . . . . . . . . . . . . . . . . . . . . . . . . 218
7.2 Measured Responses of Fresh Moisture Sensor Z7, 6.8 J.Llll PP--Bromo-
benzene Coated Microchip, in Stirred, Unused Oil. Oil Moist.ure Con-
tent Varied From 12 to 8 ppm . . . . . . . . . . . . . . . . . . . . . . . . 220
7.3 Master Curve for Fresh Moisture Sensor Z7 Produced by Normalizing
Responses of Fig. 7.2 to fp. . . . . . . . . • • . . . . . . . . . . . . . • . 221
7.4 Measure(J Responses of Fresh Moisture Sells'.)r Z3, 0.5 IHll PP-Bromo-
benzene Coated Microchip, in Stirred, Unused Oil. Oil Moisture Con-
tent Varied From 24 to 53 ppm . . . . . . . . . . . . . . . . . . . . . . . . 222
7.5 Master Curve for Fresh Moisture Sensor Z3 Produced by Normalizing
Responses of Fig. 7.4 to fp. • . . • . . . . . . . . • . . . . • . . . . . . . 223
7.6 Measured Responses of Fresh Moistu~e Sensor Z20, 6.0 ftIll PP-Bro-
mobenzene Coated Microchip, in Stirred, Ullused Oil. Oil Moisture
Cont.ent. Varied From 2 to 7.6 ppm . . . . . . . . . . . . . . . . . . . 22·1
7.7 Master Curve for Fresh Moist.ure Sensor Z20 Produced by Normalizing
R esponses 0 f F·Ig. ...I. 6 t.0 f p' • • • • • • • • • • • • • • • • • • • • • • •• ·n~.
__ "

7.8 Compiled Moisture Sensitivity Dat.a for Freshly Coat.ed Moist.ure Sen-
sors Z3,4,7,8,17,20,23 in Stirred, Unused Oil. '" . . . . . . . . . . . 228

16
 7.9 Moisture SeIlsitivity of Fresh and Aged Coated Microchip in Stirred,
Unused Oil for a) Z3 and b) Z23 . . . . . . . . . . . . . . . . . . . . . . . 229
7.10 Change in Moisture Sensitivity with Aging of Moisture Sensor Z7 in
Stirred, Unused Oil. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
7.11 Change ill Moisture Sensitivity with Aging of Moisture Sensor Z20 in
Stirred, Unused Oil. . . . . . . . . . . . . . . . . . . . . . . . . .232
7.12 Change in Master Curve with Aging of Moisture Sensor Z3. .234
7.13 Change in Master Curve with Aging of Moisture Sensor Z7. .235
7.14 Change in Master Curve with Aging of Moisture Sensor Z20. .236
7.15 Independence of Master Curve on Applied Electric Field Strength for
Moisture Sensor Z7 in Stirred, Unused Oil at '" 4 ppm Water. . . . . . 237
7.16 Independence of Master Curve on Tt" .nperature for Moisture Sensor Z7
in Stirred, Unused Oil. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
7.17 Temperature Dependence of Moisture Sensitivity of Aged Moist.ure Sen-
sor Z7 in St.irred, Unused Oil. . . . . . . . . . . . . . . . . . . . . . . . . 241
7.18 Moisture Sensitivity of Aged Moisture Sensors ill Stirred, Unused and
Oxidized Oil. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
7.19 Me~sured Response of Aged Moist.ure Sensor Z7 ill Stirred, Oxidized
Oil at 8.5 ppm Water . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244
7.20 Transient Response at 1 kHz of Fresh Moisture Sensor Z7 in Unused Oil
Going From 6 to 26 ppm Water. . . . . . . . . . . . . . . . . . .' ..... 245
7.21 Transient Response at 10 kHz of Fresh Moisture Sensor Z7 in Unused
Oil Going From 26 to 11 ppm Water. . . . . . . . . . . . . . . . . . . . 247
7.22 Measured Behaviour of DC Offset Voltage During a Frequency Scan
for :Moisture Sensor Z20 in St.irred, l.Tllusecl Oil a) at. 4.-1 ppll1 \Vat-el':
(e) - Attenuated Voltage to 8.8%, (X) - Full Volt.age, 100% and h) at.
6.9 ppnl Water . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2..J:8
7.23 Effect of DC Offset on Measured Response of Aged Moist.ure Sensol' Z7
in Stirred, Unused Oil at 4.7 pplll \oVateI' and 47°0 . . . . . . . . . . . . . 250

17

•
7.24 Moisture Sensit!vity of Aging Moisture Sensor Z17 in Oil Circulation
Loop of 50 k VA Test Transformer. . . . . . . . . . . . . . . . . . . . . . 253
7.25 Measured Master Curves of Moisture Sensor Z17 Prior to Installation
(0) and 14 Weeks After Installation (X) in Oil Circulation Loop of
50 k VA Test Transformer. . . . . . . . . . . . . . . . . . . . . . . . . . . 254
7.26 Moisture Sensitivity of Aging Moisture Sensor Z20 ill Oil Circulation
Loop of 50 k VA Test Transformer. . . . . . . . . . . . . . . . . . . . . . 256
7.27 On-Line Monitoring of Gain, Phase, Offset and Temperature of Mois-
ture Sensor Z17 at 400 Hz ill Oil Circulation Loop of 50 kVA Test
Transformer. Current Loading Cycled Between 0 and 210 A. . . . . . . 257
7.28 fp, Measured and Predicted Moisture Content, and Residual Using Data
of Fig. 7.27 and Equilibrium Moisture Model. . . . . . . . . . . . . . . . 259
7.29 On-Line Monitoring of Gain, Phase, Offset and Temperature of Mois-
ture Sensor Z17 at 400 Hz in Oil Circulation Loop of 50 kVA Test
Transformer. Current Loading Increased From 150 to 200 A. . . . . . . 260
7.30 fp, Measured and Predicted Moisture Content, and Residual Using Data
of F;g. 7.27 and Equilibrium Moisture Model. . . . . . . . . . . . . . . . 261

8.1 SEM Micrograph of Witness Strip Coated Along;.;ide Moisture Sensor

Z22 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
8.2 SEM Micrograph of Witness Strip Coated Alongside Moisture Sensor Z3.267
8.3 SEM Micrograph of "\Vitness Strip Coated Alongside Moisture Sensor Z7.268
8.4 Measured Master Curve for Moisture Sensor Z7 Compared to Predicted
Response using Bulk Ohmic Conduction Model. . . . . . . . . . . . . . 273
8.5 Measured Master Curve for Moist.ure Sensor Z24 Compared t.o Predicted
Response using Bulk Ohmic Conduct.ion I\Iodel. . . . . . . . . . . . . . 27-l
8.6 Measured Master Curve for Moisture Sensor Z'j Compared t.o Predict.ed
Response using Bulk and Surface Ohmic Conduction Model. . . . . . . 2if3
8.7 Measured Master Curve for Moisture Sensor Z24 Compared to Predicted
Response using Bulk and Surface Ohmic Conduct.ion Model. . . . . . . 277

18
8.8 Measured Master Curve for Moisture Sensor Z7 Compared to Predicted
Response using Linear Distribution in Bulk Ohmit' Conduction . . . . . . 2;9
8.9 Measured Master Curve for Moisture Sensor Z24 Compared to Predicted
Response using Linear Distribution in Bulk Ohmic Conduction . . . . . . 280
8.10 Measured Master Curve for Moisture Sensor Z7 Compared to Predicted
Response using Dispersive Complex Bulk Permittivity Arising From
Lorentz Sphere Model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
8.11 Experimental Data Showing Normalized, Power Law Dispersion of a)
Polydian Carbonate with Temperature Dependence and b) Sand with
RH dependence (from [104, p. 177 and 2;7]). . . . . . . . . . . . . . . . 285
8.12 Measured Response of Aged Moisture Sensor Z24 in Stirred, Unused Oil
at 48 ppm Water. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28;
8.13 Estimated Dispel'sion in Bulk Permittivity and Conductivity of Z24
PP-Bromohenzene Coating using Response of Fig. 8.12 . . . . . . . . . . 288
8.14 Estimated Dispersion in Complex Bulk Permittivity of Z24 PP-BrOlllO-
benzene Coating from Fig. 8.13. . . . . . . . . . . . . . . . . . . . . . . 289
8.15 Estimated Dispersion in Complex Bulk Permittivity of Aged Z; PP-
Bromobenzene Coating. . . . . . . . . . . . . . . . . . . . . . . . . . . . 292
8.16 Estimated Dispersion in Complex Bulk Permittivit.y of Aged Z20 PP-
B:nlllohenzene Coating. . . . . . . . . . . . . . . . . . . . . . . . . . . . 294
8.17 Estimated Dispersion in Complex Bulk Permittivity of Aged Z20 PP-
Bromobenzene Coating using Corrected Response. . . . . . . . . . . . . 295
8.18 Time Dependence of Change in Parameters Describing Moisture Sensi-
tivity for Z; (X) and Z20 (.) with Aging. . . . . . . . . . . . . . . . . . 296
8.19 Correlation of Paramet.ers Desnihillg i\luislnre Sellsiti\'ity for Zi (X)
and Z20 (.) with Aging. . . . . . . . . . . . . . . . . . . . . . . .. . 297

9.1 Schematic View of Tin Oxide Thin Film Deposit.ed on a. Microchip . 303
9.2 Master Curve for Tin Oxide Coat.ed Microchip, in Variable RH Envi-
rOn111ent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305

19
9.3 Relative Humidity Dependence of Estimated Surface Conductivity of
Tin Oxide Coa.ted Microchip . . . . . . . . . . . . . . . . . . . . . . . . . 307
9.4 Measured Response at 10 kHz of Pd/Tin Oxide Coated Microchip in
Gas Flow (Variable Composition). . . . . . .308
9.5 Layout of Thin Film Palladium Resistors. .311
9.6 Response of Thin Film Palladium Resistor to a) 1000 ppm and b)
100 ppm H2 ill Nitrogen at 225°F. . . . . . . . . . . . . . : . . . . . . . 314
9.7 Response of Thin Film Palladium Resistor to a) 1000 ppm and b)
100 ppm H2 in Nitrogen at 225°F. . . . . . . . . . . . . . . . . . . , .. 315

B.1 Medium Above Electrodes Represented by a Multilayered Structure of

P homogeneous layers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 331

C.1 Overall Mechanism of Glow Discharge Polymerization (from [83, p. 180]).334

C.2 Model St.ructure of a Plasma Polymer Film of Ethylene (from [83, p. 154]).334
C.3 Powder and Film Regions at 100 W using Ethylene MOllomer (from [83,
p. 244]). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 336
CA Electrical Configuration of 13.56 MHz Capadtively Coupled RF Plasma
Deposition Apparatus. . . . . . . . . . . . . . . . , . . . . . . .339
C.5 Mechanical Configuration of Plasma Deposition Apparatus. .340
C.6 Response at 1 kHz of Z13, a '" 1 Jlm Thick Bromobenzelle Coated
Microchip, as Room Air was Introduced int.o the Evacuation Chamber
Immediately After Deposition. . . . . . . . . . . . . . . . . . . . . . . . 345
C.7 Response at 10 kHz of Z22, a 3.5 Jllll Thick Bromobellzene Coated
Microchip, as Room Air was Introduced into the Evacuated Chamber .. 346

D.1 Measured (e) and Predicted (-) Response of an lTllcoated l\licl'orhip in

Oil Highly Doped wit.h ASA. . . . . . . . . . . . . . . . . . . . . . . . . 350
D.2 Measured Response of an Un coated l\Iicrochip in Oil Slight.ly Doped
with ASA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353

20
List of Tables

1.1 Reprinted Summary of ASTM Screening Specifications and Functional

Property Requirements (from [6]). .. . . . . . . . . . . . . . . . . . .. 30

2.1 Predicted Gain Output vs. k Collocation Points and N Summation Terms. 68

3.1 Case 1 - Closely Spaced Electrodes, No Ground Plane .. 93

3.2 Case 2 - Parallel Plate Capacitor. 94
3.3 Case 3 - Microstrip Calculation. . 95

4.1 Spatial Wavelength and Normalized Load Capacitance .. .129

4.2 Simulated Data for Case 1. .130
4.3 Simulated Data for Case 2. . .133
4.4 Simulated Data for Uniform Medium. .138
4.5 Simulated Data for Case 3. .144
4.6 Simulated Data for Case 4. .145
4.7 Simulated Data for Case 5. .153

5.1 Two Parameter Estimation Using Coated Microchips. .168

5.2 Oil Complex Bulk Permittivity Measurement With Microdlip. . 173
5.3 Bulk Permitt.ivity Measurement Using An Uncoat.ed f\Iicl'ochip And A
Bridge. ............................... 180
5.4 Bulk Conduct.ivity Measurements \i\'ith Passivat.ed Elect.rodes. . 1no

6.1 Plasma Deposition Process Paramet.el's for Brolllobenzene Coated rvIi-

crochips . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214

21
7.1 Best Fit Parameters for Freshly Coated Chips. .227

8.1 Dispersion Parameters For Moisture Sensors .. .290

E.1 Directory of Sample Input Files and Program Listings .356

F.1 Directory of Sample Input Files and Program Listings .459

F.2 Definition of Symbols used in Input Files .460

22
Chapter 1

Introduction

1.1 Background - Engineering and Science

As with many research endeavors, this work reflects the interplay of two driving forces
- scientific int~rest.s and engineering needs. The scientific interest is understanding
observed phenomena at a fundamental (molecular) level and correlating these obser~

vations with self consistent models. The engineering need is developing eledrochemical
sensors to monitor various chemical constituents within transformer oil. The trick is
to recognize that one interest does not preclude the other and, in fact, to be success~

ful a project must satisfy both interests. In the following sections the scientific and
engineering aspects of this work are discussed along with the goals of this thesis.

1.2 Scientific Motivations

The work presented here applies th.e disciplines of system identification and analytical
and numerical field theory to t.he problem of interpret.ing informat.ion obt.ained using
interdigital electrode structures. Using the fringing fields of a planar eled.rode st.ruc~

ture, the response to an applied potential at. the elect.rode-medium int.erface can he
measured noninvasively. Estimates of the dielectric properties of t.he material can be
obtained from these measurements via parameter estimation routines incorporating a
coutinuum model of the electric field distribution within the material. An increased

23
understanding of the physical structure of a material can be obtained with such esti-
mations.
A specific application of the above approach is to measure the dielectric proper-
ties of thin semi-insulating films. Using the insight gained from these measurements
regarding the physical structure of the films, a strategy can be evolved for develop-
ing electrochemical sensors. Of particular interest here is the development of both a
moisture and a hydrogen gas sensor for transformer oil.
The instrument used to make t!lese dielectric measurements is a microdielectrom-
eter - a device utilizing a set of microfabricated interdigital electrodes for electrost.at-
ically probing a material. A continuum model for the electric field distribution of this
interdigitated electrode stru.cture has been formulated. This model has an emphasis
on generality with respect to the types of media with which the electrode structure
may interact. By embedding this continuum model in simple single or multi param-
eter estimation routines, it is possible to hypothesize and estimate parameters, from
experimental data, describing heterogeneities 0; interactions within a medium under
investigation. Prior to this work the only models for interpreting experimental data
estima.ted the dielectric properties assuming one of two limits: 1) a uniform medium or
2) a thin film « IJ.Lm ) surrounded by a uniform dielectric. Thus, the work presented
here greatly expands the applicabiHty of the microdielectrometer.
The following two sections discuss the advantages of using planar, illterdigital elec-
trode structures and the implications of probing a medium using electroquasistatic
fields.

1.2.1 Imposed w-k Sensing

Classical clielectrometry obtains information concerning t.he physical properties and

state of a medium from the electrical frequency response. The mo.clal approach t.o
dielectrometry, presented in Chapter 2, not only extracts its informat.ion from t.he
cont.rol of the temporal frequency, but also from the imposition of a spatial periodicit.y
as well. In this "imposed w-k" t.echnique, the medium is excited at the temporal
(angular) frequency w by means of an interdigital electrode structure having a spatial

24
 periodicity length ,\ ;:: 211'/ k and hence a dominant wavenumber k.
In Fig. 1.1, the medium is above and the structure used to make the measuremc:-ut.s
is below. Because the electrodes introduce higher spatial harmonics, field quantities,
such as the potential, are expressed as

~ jwt
cp(x,y,t) = ReCP(x,y)e (1.1 )

where
00

~(x,y) =L ~n(X)CoS(knY)ikn = 2n1f'/'\

n=O

representing a superposition of standing waves of dominant wavenulllber k = kl in

the plane of the electrodes. The choice of the cosinusoidal representation exploits the
symmetry of both the electrode structure and the situation above the electrodes (for
example, there is no motion).

The advantages of the w-k approach are many. First, there is the coupling into
the medium from a single surface, allowing measurements to be more readily made
nonillvasive, if desired. For example, dielectric measurements of thin films can be
performed without having to vapor deposit an upper metal electrode for use with a
capacitance bridge, and in many cases, can be taken at the site of the coating process
without changing the ambient environment.

Secondly, because the fields generated by the electrodes In the medium (above
in Fig. 1.1) are quasistatic, they tend to decay into the material exponentially with a
characteristic length that is at most a fraction of the spatial wavelength. Consequently,
the electrodes are sensitive to dielectric materials located within a region roughly ,\/3
or less from the electrode plane. Thus, the spatial sensitivity of t.he dc:-vicc:- CRn be
tailored to individual needs.

A third advantage comes with modern llucrofabrication techniques t.hat. 110t only
make micrometer scale electrodes possible, but make the integration with the signal
processing electronics feasible as well. Electrodes with a very short spatial wavelength,
'\, can be deposited. This refines the spatial resolution. For example, given a spatial

25
 x

~.
E*(X} y

Si02
P - Si ------• •

I ----------
I
I eFT
+ +

1\
VD

I
I I
I
IL _ _ _ _ _ _ _ _-:-_ _ _
-=- _ J

Figure 1.1: Representation of Interdigital Electrode Structure and Associated Circuitry

Interacting with a Heterogelleous Media.

26
wavelength of 50 micrometers, thin films of 20 micrometers or less (perhaps coated on
the electrodes) can be dist.inguished and their dielectric dispersions examined.
Having short wavelengths also improves the sensitivity to measurement.s of surface
conductivity, u.. Sensing conduction on the surface of an insula.tor without making
electrical contact with the surface requires capacitive coupling. Thus, the frequency
must be high enough to make the capacitive reactance on the order of the resistance
of the surface. Roughly, this requires that to measure a surface conductivit.y u" at
the frequency f = w /2rr, the ratio (u. / f fA) must be of the order of u~ity. There
is a practical lower limit on the frequency. Therefore the smaller the wavelength A,
the smaller the surface conductivity that can be measured. Having the electronics
integrated with the electrode structure makes it possible to use fr~quencies as low as
0.005 Hz using the "chip" described in the next section.
Although the emphasis here will be on dielectrometry, the imposed w-k approach
can be used in systems where the electrical response reflects electromechanical and
electrochemical phenomena. This is illustrated by earlier studies where interdigital
electrodes were used to !:>tudy the electromechanics of double layers [1], electrohy-
drodynamic surface waves and instabilitie~ [2] and sensing of electrification effects in
insulating fluids [3,41.

1.2.2 Intimate Sensing

By contrast with remote sensing, where electromagnetic fields described by the wave
equation are often used to resolve properties and states of materials at great distances,
the imposed w-k approach exploits fields that are quasistatic from the electromagnetic
point of view. In the electroquasistatic embodiment described here, the dynamics re-
sult from charge migration and diffusion, and polarizat.ion relaxation. In the magllet.o-
quasistatic analogue, the dynamics would involve current diffusion and magnet.izat.ion
hysteresis. In either of these latter cases, the fields are quasistatic and tend t.o be COll-

fined to the immediate neighborhood of the interdigital electrode structure. Because

measurements reflect the properties and state within a short distance, t.he applica-
tion of such structures to parameter estimation might be dubbed "intimate sensing".

27
Strategies exploiting the wavelength controlled resolution inherent to having fields that
are confined to the proximity of the electrode can utilize the model developed here.
With the combination of microfabrication techn010gy and the "imposed w-k" tech-
nique, a determination of electromechanical or electrochemical structures and dynam-
ics on a m.icroscale becomes feasible. Fluid boundary layers, electrochemical double
layers and thin film structures all fall into this microscale realm. Up to now, the
behavio~r of these phenomena as a function of temporal frequency has been used to

identify various structures. To the extent that this requires the. assump~ion of fre-
quency independent parameters to effectively extract information, this approach t·)
parameter estimation gives little opportunity to see the resulting model as uniqu~.

With the usage of the approach presented here, a continuum model coupled with a
parameter estimation routine, it becomes feasible to extract information at a particular
temporal frequency by varying the spatial frequency of the a.pplied pot.ential. As
mentioned before, the field decay normal to electrodes is directly proportional to the
spatial wavelength of the electrodes. Intuitively, it makes more sense to determine the
structure of heterogeneous media by probing spatially, rather than temporally. This
method does not have the restriction regarding the frequency dependence of parameters
such as the complex permittivity. Practically, this effort amounts to having multiple
electrode structures, each having a different spatial wavelength~ monitoring the same
medium, or perhaps one array of electrodes with switchable terminal connections to
produce the various spatial wavelengths.

1.3 Engineering Motivations - On-Line Transformer

Monitoring
Large scale power transformers (1 MVA and up) are devices that. represent. a large
capital investment, provide an essential service requiring minimal interruption, are
subject to catastrophic failures and are costly to repair. The potential benefits from
improving the decision making process with regards to maintenance and servicing make
it highly desirable to mOllitor these transformers. The following section reviews the

28
field experience gained from monitoring changes in various properties of transformer
oil using off-line techniques.

1.3.1 Background

It is not the intention to extensively review all quantities that might be of interest.
The focus here is the monitoring of the oil circulating within the transformer. Just as a
blood analysis reveals significant information about the condition of the body, it is felt
that an oil analysis will be important in determining the state of a transformer. There
are two major functions served by the oil, heat transfer and insulation. 'T. ....ditionally,
mineral oils consisting mainly of paraffins and cyclo-paraffins have been employed.
Other insulating fluids are continually being examined due to the vagaries of world oil
supply and in the search for less flammable or otherwise more technically advantageous
fluids. However, at present, economics greatly favor using refined petroleum oil. An
excellent review of insulating fluids, their properties and uses has been done by Wilson
[5] .
There are a wide variety of ASTM tests used to characterize the electrical, physical
and chemical properties of transformer oil [6]. The oil must meet certain specifications
both prior to going into a transformer and while in service. Quantities of direct interest
are: dielectric breakdown strength, power factor, dielectric loss, moisture content,
dissolved gas content, acidity, interfacial tension, copper and iron content, sludge and
particulate count. Reprinted in Table 1.1 is a summary of screening specificat.ions for
refined mineral oil without additives and property requirements for new oil to be used
in electrical apparatus. As a result of many years of accumulated experience, both in
the field and laboratory, changes in the above quantit.ies over time (with t.he exclusion
of the first three) can be associated with particular degradation processes occurring
within the transformer [7]. 1

All important example is dissolved gas analysis (DGA). Though there are sev-
eral schemes for calculating and interpreting gas ratios [8,9] it is generally agreed
1 Note - many of the quantities discussed are highly temperature dependent, thus monitoring or
estimating the oil temperature is assumed.

29
 Table 1.1: Reprinted Summary of ASTM Screening Specificat.iolls and Functional
Property Requirements (from [6]).

P'Operl~' Specllied V,11Ie ...ST~ .... ctftod·

I'hJletli:
<.:.,10'. mu I 0"00
flasll "",n,. mln I~ C 1295 fl' 092
1",.,flclll '.11110". min. d~ncs'cm A() 0971
POll' poln,. mu -oIOC I-A() f)' 097
SpeCIfic I':aw"y "' In6115 56 C tCloO/CIoO fl 0.'4 10 O~I 012911
ViSCOIII •• mu. SUS.
~, 37 1a C 1100 FI 65 o la
.,0CIJ2Fl no 00145 .nd 0 :161
CAwr'"elll:
: :WIII'CS nil
C.,mIHncd CO"OIl'. IlIlfu, nonconot."c oms
Ino.... "fC cblondes ~nd IUlr.,CI nil 0811
!'Iculf.II""_ numlle,. mu. m, KOHl, oil 0.05 09701.,,0_
Sludp: by o • ..sa"o" IClI.· mu. percen, 0.06 01313
Sludp ~I·I.' o • ..sal/on ,,"blhl,:
1~ h. mu. pctCH, 0.3 02...0
IM h. mu. percen, :1.7 0:440
Ncu"alIUlIOfl """'_
;~ h. m.... m, KOHl, all
.fIe,
o.lll.,IOfI:
!." 0:...0
11>01 h. ma •• mII\.OH', ,"l :.6 O!...o
W 11.'. ma•• "'"" 35 0153)
£1«",,111.
OlCleClnc lIrcalLdown .011.... ml". k V lO.O o In
Powe, laao, " 100 CUI! FI. OIl HI. mu o.oos to.5 pctCHlI 0924

~t 03487
rULE I F____ ~ R....... aa

Type I Type 11
I'ltmclll:
Aluliae poaat. 'C (6l_7IA) (63-71'" 0611
Colo,.mu O.l O.l o ISOO
FI.s" polnl. man. 'C 145 I.S 0'12
InleriaaailenllOn 11 !5'C. man. dyneSlall 40 .aD 0971
Pou, poan'. mu. 'C -0ICl" -4Cl" o 97
Spccalic &'I\O"Y. IS'C'U'C mu 0.91 0.9\ o 1298
Viscosllv. mu. cSIISUS)II:
1000C
oIll"C
).01361'"
\2.11lb6l'
] 0 (]6),"
12.01661'"
0 ..... S o. 0 88

WC ;/1 UIBOI '6 013S01

Vi,ual CI3mlnluon .:Iear ~nd bn,hl dear .nd bn,hl o 1~24
Elletn,,,'.
DIC/ICfnc b'IIJI"iow" ."""JI ", 00 H:
O,,~ elearodn. mIn. 10. V )0 )0 0 877
voe .Iea,odn. mIn. kV .IIhe'.1 0 t)4().,II. 11 .O!· ~8D '::8° 0 1816
mml,ap • ." >Ill \I~O·'n. I !.lJ4.mm' s.p 56" 16"
D'elrclnc ,,,,,,,,clo.n vnl,,,(r. ''"Pill)' (o"d,noru 0 3300
:'·C. mIn. Io.V. '1.edle ..e.:3l ..e 10 ",here tol~'" 1"~·~
~rnundcd. l'ln I'::~ ~·mm f ~ilP
Pnwrr 'acto' .JI"a It: ,"4'.1. ' ; 0'124
:$'C \J U~ " U~
10ll"C .) )0 It )0
CIt~",lclll.
OIl44'tIOII IIlIb.I'I\' h1rreJ • .,ucJft ,tSU o :"",0
-: h:
... II ..dSe. mu. b)I IDUI 0.15 01'
TOIaIlQd nUIII_. /DU, ml KOH/I 0.' O.lA
1Mb:
... I..dlc. /DU, by IDIW ()J ().~
TOlai a"ad aUIII~'. /DU, aI! KOHII H 04'
O ..dauoa SI.bIIlIY I fDlAWlI bomb IClII. IIWL _UICl 19~ 0 111:::
0114.l ..oa IllJulHla. CaDlcDI. alU.... by ma.u 008 O.l 0 1411 or 0 :6flI'
C"rTOIl'•• ullu. . .:IUD.;.c,riu....c 0 tl"!'5
W ..... m~ •. ppm ~~ )~ 0 IJt~.,. 0 :~n
'euu:auallon nwnPtf, IQla, ~ad numf'Ct mu m, 11113 .,0] o '17.&
";OHIC

30

•
that mechanisms such as arcing, partial discharge and thermal degradation ca.n be
distinguished using DGA. For example, large quantities of hydrogen indicat.e part.ial
discharges, whereas acetylene indicates arcing is occurring [10,11]. Carbon monoxide
and dioxide indicate overheating of the cellulose insulation.
Moisture content is used as an indication of the condition of the pape:- insulation,
existence of free water and leaks to the atmosphere [12]. Typically, paper is dried to
roughly 0.5-1 % moisture by weight prior to filling a transformer with oil. The solubility
of moisture in the paper, relative to the oil, will vary from 3000 times more at 20°0 to
400 times at 60°0 [13]. Thus, during the regular load cycling of the t.ransformer, the
accompanying temperature variations will drive moisture into and out of the paper.
At the higher temperatures the oil can absorb more moisture. This excess moisture
should go back into the paper upon cooling. However, if the oil cools faster than the
time required for equilibrium to occur, then the oil can become supersaturated and
free water will form. Free water can lead to breakdown if it appears in high field stress
locations either in the oil or absorbed into the paper. Thus it is important to monitor
the relative saturation of the oil.
Assuming an equilibrium exists, the moisture content in oil can be directly related
to the moisture content in the paper [14] (this is a point of contention and will be
discussed further in Section 6.1). Paper with a high moisture content (> 2%) will
suffer a significant loss in electrical breakdown strength [15]. Degradation of the oil can
result ill higher moisture saturation levels in the oil, increasing the moisture content
at a given temperature [16]. Thermal degradation of the cellulose insulation will
result in the formation of carbon and water. Moisture content can also affect static
electrification, by altering the charging tendency of the oil [17].
Oxidation of the oil is indicated by changes in oil resistivit.y, acidity, sludge for-
mation or interfacial tension [181-[21]. Analysis of t.he part.iculat.e mat.t.er in oil with
respect to composition, number, size and shape is also important. [221. Copper and
iron particulate are good catalysts for oxidation of t.he oil while excessive amount.s of
cellulose fiber might indicate excessive wear of the insulation. A review of dectricea.l
loss mechanisms in transformer oil, and how oxidat.ion effects its dielectric properties

31
 can be found in [23, Section 1.2].
Oil sampling and analysis, under normal operating circumst.ances, is current.1y per-
formed on a semi-annual basis. There are documeIlted cases of the dissolved gas
analysis alerting operators to internal problems (usually in conjunction with other
tests) [24]-[27] and even preventing failure [28]. The utility of this analysis has been
extended to other types of power equipment such as bushings [291. It is clear that
monitoring dissolved gas content is important and it follows that similar benefits may
be gained through monitoring of quantities such as moisture content and acidity.
However, the above tests all involve removing a sample of oil from the transformer
and introducing the oil into another apparatus. For example, moisture analysis involves
titrating the oil with a Karl Fischer reagent and measuring the resultant current. The
method is based on the reduction of iodine by sulfur dioxide in the presence of water.
Dissolved gas analysis requires stripping the gas from an oil sample and passing t.he
gas through a gas chromatograph. It would be very costly to frequently obtain oil
samples from transformers in the field and analyze them in a lab. Typically, sampling
is performed annually, or semi-annually, unless there is suspicion of a malfunction. In
that case, more frequent monitoring of the oil is performed. This mode of sampling is
haphazard, it will catch some transformers before they fail - ones slowly deteriorating
or rapidly failing at the time of sampling - and completely miss others. There is at
least one documented case of the time to failure being on the order of minutes [30].
In addition, the measurements may be questionable due to factors such as poor or
variable sampling technique or sample transport. Thus, there is an established need
for on-line monitoring of various physico-chemical properties of transformer oil.

1.3.2 State of the Art Review

This section reviews existing sensor and analyzer technology as ei ther directly ap-
plied to transformer oil monitoring or more broadly applied to selective gas detection.
Regarding the latter category, the focus is narrowed to solid state devices for gas de-
tection, avoiding liquid electrolyte detectors. The main reason for this focus is the
emphasis (perhaps unnecessarily) on developing detectors immersible in the oil. The

32

•
sealing problems involved with liquid electrolyte detectors made investigation of this
area unattractive. Of course, operation of detectors in the vapor phase would great.ly
alter this viewpoint. Another factor skewing the focus of this search is the desire t.o
find techniques adaptable to the microchip structure.
Functionally, there are at least two distinct processes involved in the analysis of
either moisture or dissolved gases in oil- extraction and detection. With gas detection
a third process, gas separation, may be necessary. Tlus is dependent on the selectivit.y
of the gas sensor used. All of the sensors discussed below operate in the vapor phase.
Therefore, gas extraction is performed using a reinforced diaphragm, separat.ing the
oil from a sampling chamber. This diaphragm is designed to withst.and t.he int.ernal oil
pressure and allow the pertinent gases to diffuse t.hrough. When equilibrium has been
reached across the membrane, the gas detection process takes place. The time until
equilibrium is reached will depend greatly on the membrane permeability to each gas.

Moist ure Sensors

Only one type of device for on-line monitoring of moisture in oil is available at
present [121. This device allows a small volume « 1 cc) of initially dry air to come

into equilibrium with the oil across a polymer membrane. Thus, the device essentially
is a dew point hygrometer - the collected vapor is cooled through contact with a nurror
surface whose temperature is monitored. An optical system detects the occurrence of
condensation, thus measuring the dew point (Fig. 1.2). This dew point also indicates
the relative humidity of the air. At equilibrium the relative humidity of the air equals
the relative saturation of the oil (see Section 6.1). In this manner, a measurement of the
relative sat.urat,ion of the oil is obt.ained. As discussed earlier, t.he rt"lative SR 11Iral ion

of the oil can indicate the possibility of free wat.er format.ion. In addition, it. ca.n also
be used to help determine the moisture content of the paper insulation as sh')wn in
Chapter 6. No information on field testing of such devices has been report.ed as of t.his
writing.

Gas Detectors

33
 .....
,~

Cl

....
c
';;

..
Cl

Q.
e
<

..
0
.....
VI
~
C1.
'"c;; ea
-'
.:I
"-
~ "-
-' Cl
..0
~ e
0 a
if ...
..::;;

c
VI
...a
...
VI ..
C Cl Go
c..?; c
""0 /

/ > t

I~
VI
Cl
..::l
=
Cl
D
.
c
e
"-
Cl
c.
e rE¥. , I

..
..e ..e ..e.. ...
Cl
,le c 0
a 'E. e
...
~
..0
Cl
::t:
c Cl

:n
"-
Cl
::
0 Vi

Figure 1.2: Apparatus for Measurement of Relative Saturat.ion of Oil (from [12]).

34
FID and TeD - Flame ionization and thermal conductivity detectors are standard,
nonselective gas detectors employed in gas chromatographs. In flame ionizat.ion the
gas is consumed in the flame, generating a signal proportional to the amount of gas
present. Thermal conductivity detectors measure the change in thermal conductivity
of the carrier gas due to the presence of another gaseous species. Due to the lack of
selectivity, both detectors require gas separation, usually accomplished by passing the
gas mixture through one or more gas chromatographic columns.

Electrochemical Detectors - An electrochemical fuel cell for detecting hydrogen gas

is commercially available [31]. The fuel cell is composed of platinum electrodes and an
aqueous solution of H 2S0 4 (Fig. 1.3). In operation, hydrogen is oxidized at one elec-
trode with a corresponding reduction of oxygen at the other electrode. The re-suIting
current flow is monitored and related to the ambient hydrogen concentrat.ion. Indus-
trial tests of this apparatus on current transformers [32] and HVDe valve damping
resistors [33] are ongoing. In addition, this detector is incorporated into a transformer
monitoring system being developed at Westillghouse [34]. As of this writing, thi:s de-
tector is not used with a gas separation section although it is not completely selective.
The problem is the lack of selectivity in distinguishing between several combustible
gases. The detector will also respond, to some extent, to carbon monoxide, acetylene
and ethylene but expresses the result as an equivalent concentration of hydrogen.

Another type of electrochemical device for detecting hydrogen gas is a diffusion cell
[35]. A proton conducting polymer is sandwiched between two noble metal electrodes
(Fig. 1.4). The potential developed across the polymer film is related to the partial
pressure of hydrogen on either side of the det.ector. For the specific device shown
in Fig. 1.4, the Pd electrode has been saturated with hydrogen forming a palladium
hydride reference electrode. Thus, the voltage can be directly related to t.he ambient.
hydrol~en concentration. A major problem with this device is the stabilit.y of t.he
reference electrode. It is very difficult to prevent hydrogen diffusion from t.he reference
over an extended period of time. This renders the device impractical at the present
time for on-line hydrogen gas monitoring.

35
 Air

oxidation reduction

Pt electrodes
/
Figure 1.3: Electrochemical Fuel Cell for Detecting Hydrogen Gas (from [31/).

H2
I I
Pt v
PVA/H3P04
I
I PdHx 1

Substrate

RT P'(H2) Nernstian
V= 2F tj In P""(PdHx) Response
Figure 1.4: Diffusion Cell for Detecting Hydrogen Gas (from [3.51).

36
Solid State Detectors - This category can be divided between structured and un-
structured devices.

Un&tructured devices are doped metal oxides and come in many varieties: 1) pellets

or thick films of silltered tin oxide doped with a noble metal such as palladium or
rhodium and 2) thin films of sputtered tin oxide. The most popular commercially
available device is the Taguchi Gas Sensor (TGS) which is of the first variety. The TGS,
and other devices of this kind, is prepared from a paste of sintered tin oxide powder
doped with palladium [36). A wealth of empirical information on preparation of and
experimentation with these devices call be found in [37). An attempt at quantitatively
charaderizillg the operation and response of these devices is presellted in work by
Clifford and Tuma [38),[39). Work has been done using sputtered thin films of tin
oxide « 5000 A) for sensing hydrogen and hydrogen sulfide [40], ethanol [41), nitrogen
oxide [42] and carbon monoxide [43).
In operation, the DC conductivity of tin oxide gas detector is altered by the presence
of reducing gases such as hydrogen, carbon monoxide or hydrocarbons. The physical
mechanism most commollly used to explain this result is the reducing gas removes
oxygen adsorbed on the grain surfaces (Fig. 1.5). This removal lowers the barrier
potential existing across grain boundaries, freeing up electrons for conduction. As a
result, the conductivity of the device is increased. The presence of the noble metal
dopant serves as a catalyst to enhance the surface reactions.

For proper operation there are several provisos. Oxygen must be present to ensure
reversibility. The devices are always operated at elevated temperatures (> 150°C) to
eliminate problems with moisture adsorption and to improve the adsorption/desorption
kinetics for the gases of interest. Any reducing gas present will affect the measured
conductivity to some extent. Therefore, these det.ectors are inherent.1y nonselect.ive and
are usually used in situations where only one gaseous component. present. is required
to be monitored in a relatively inert background. The selectivity can be improved by
altering the doping concentration, materials and the preparat,jon process but, it. is a
highly empirical approach. An alternative is to use the operating device temperature
to improve selectivity. The temperature dependence of the adsorption/desorption ki-

37
 Pd

1"
air
w/reducing
------~~~~~~~--~----
gas
Grain Boundary

Figure 1.5: Effect of Reducing Gas on Work Function (V barrier) at the Grain Boundary
of Pd Doped Sn02_: Having Surface Adsorbed Oxygen.

38
netics will be different for each gaseous species. By ramping the temperature of the
device, each gas will be driven off at a different temperat.ure, providing a gas sep-
aration analogous to that obtained in gas chromatography. The individual effect of
each gas upon the conductivity is then monitored, providing (ideally) individual gas
concentrations.

Structured devices involve microfabricated structures such as diodes, capacitors and

transistors as shown in Fig. 1.6. In these devices the gas sensitivity is due t.o inter-
facial phenomena occurring at either the Pd-oxide or the oxide-doped semiconductor
interface and manifests itself in changes in i-V, C-V or gate threshold voltage char-
acteristics. Most commonly, it is the sensitivity to hydrogen that is studied [44], [45]
although carbon monoxide [46), {47} and hydrocarbon gas sensitivity {48}, [49} have
also been reported.

The net effect of the pertinent gas is attributed to an alteration of the electrochem-
ical work function (barrier potential) existing across the appropriate interface. On a
device level, this translates into a change in the voltage drop across the device that
shows up directly in the C-V and threshold voltage measurements [48], [50}. For t.he
diodes, the diode current. is an exponential function of this voltage, thus altering the
entire i-V curve shape {51}.
The oxide layer thickness plays a very important role in determining the location
of the interface at which the gas sensitivity occurs. Lundstrom and Svensson [52}
maintain that the change in work function, for devices having oxide layer thicknesses
greater than roughly 100 A, is due to the formation of atomic hydrogen dipoles at the
Pd-oxide interface. For thin oxide layers, Keramati and Zemel propose that the hy-
drogen diffuses through the Pd and the oxide layer and cause changes in t.he interfacial
trap density at the oxide layer-doped silicon interface [44], [53]. The dominant. physics
is very dependent on the structure under examination with regards to materials, purit.y
and perhaps most relevant here, oxide layer thickness. The mechanism of illtedacial
trapping is important for thin « 100 A) oxide layers [54].
Alteration of the Pd layer thickness is also very important. For Pd layers thicker
than 100A, forming a continuous film, only small molecules such as hydrogen can dif-

39
 i
J
Schottky
Pd
V - - - - - - " " , j C - - -.... V
Diode
oxide
I

-. ..
I
Pd Tunnel
V +- oxide V
n-Si Diode
I

.... C
" I
MOS
Pd
Y oxide
., , (upuci tor
n-Si
I V

VT
f
Pd i
I oxide I
....
- MOSFET
n+ J
p-Si
I n+

/
--- ~ PH
2

Figure 1.6: Microfabricated Gas Sensing Diodes, Capacitor and Transist.or Ut.ilizing
Pd-Oxide Interface.

40
fuse through to reach the appropdate interface. This enhances the device selectivity
to hydrogen or gases composed of hydrogen. Sensitizing the device to larger molecu-
lar gases such as carbon monoxide has been demonstrated by using either ultrathin,
discontinuous Pd films [47] or masking the Pd deposition process so as to produce a
porous film with holes (channels) to the oxide layer interface [46].
Typically these devices are also operated at elevated temperatures (> 150°0) both
to enhance adsorption/desorption kinetics and to eliminate the effects of moisture.
The presence of oxygen is again of great importance. With the strong emphasis 011

surface reactions at either the Pd or oxide interface, any other species that either takes
up adsorption sites or reacts with surface adsorbed molecules can significantly alter
device performance. For example, in an inert atmosphere such as argon, introduction
of hydrogen produces a large, rapid response but the subsequent removal of hydrogen
yields a slower change with an irreversihle component. In the presence of oxygen, the
devices are reversible but the gas &ensitivity is reduced due to the reaction of hydrogen
and oxygen at the Pd surface. These class of devices have not been made commercially
available yet. There are problems with long term st.ability due to sulCur poisoning [45]
and hydrogen drift into the oxide layer [55].

Gas Analyzel"s

As pointed out several times above, most of the gas detectors are not very selective.
Thus, the additional step of gas separation must be performed prior to detection. A
logical approach, taken by a group in Japan, has been to automate the standard gas
extraction/chromatography process [56]. An oil sample is extracted and degassed by
pulling a vacuum. The gas mixture is then sent through chromatographic separation
columns and then to t,he FID and TCD detect.nrs. Accompanying t.his dissnl"f'd gas
analysis unit was a total combustible gas (TOG) unit. This unit used t.he same gas
extraction apparatus but only measured total combustible gas content. using a cat-
alytic combustion (TGS) type gas sensor. Val'ious monitoring strategies using t.he t.wo
units were proposed and some field testing performed. Though the units performed
satisfactorily in the field (two samples per day) they are rather large, contain much

41
automated valving, require maintenance of the gas separation columns and periodic
replenishing of the carrier gas cylinders, take one or two hours per sample and must
store the oil samples removed from the transformer. If it is required to monitor on
a timescale of minutes or even perform several samples per day then the units will
consume their resources rapidly and remove a significant quantity of oil.
Most of the effort has been towards developing new techniques for gas extraction
and detection. The common alternative in gas extraction is to expose a polymer
membrane, usually Tefion or a similar fiuoro-polymer, to the circulating oil. The
gases diffuse out of the oil, through the membrane, are collected in a chamber and
then, if DGA is required, separated using a cOllventional gas chromatography column.
Correlation between gas-in-oil concentrations and detector responses require sampling
after time is allowed for the partial pressure of a gas to equilibrate on both sides of the
memIJrane. This equilibration time is a function of the solubility in oil and membrane
permeability of the particular gas to be detected. Further, both of these quantities
are temperature dependent. Equilibration times typically can range from 10 hours for
hydrogen to 50 hours for ethane (depending on the ;).ctual physical dimensions of the
membrane and collection chamber) [571. This criteria, then, imposes an upper limit
011 sampling frequency that mayor may not be acceptable.
The major point of divergence occurs at the gas detectioll end. A group in the
UK has used a fiame ionization detector, common in gas chromatography, combined
with a silicon rubber membrane used for gas extraction [30]. Samples are taken every
40 minutes - no mention was made of equilibration times. Along similar lines, a
group at Westinghouse has used a thermal conductivity detector, also common in gas
chromatography [58,59]. This device samples once every 24 hours. Based on the fielcl
experience of [561 the thermal conduct.ivity det.ect.or is more st.able t.han the flame
ionization det.ect.or. Regardless, bot.h techniques have many of t.he :same drawbacks
mentioned above with the exception of those associated with oil removal.
Another type of detector is a combustion semiconductor gas sensor similar t.o t.hose
used for monitoring emissions from automobilt=s or industrial processes for pollution
control, fire safety, or health reasons. Though not well described in papers on gas

42
 1le45lri-( l.Iba I
ConnActa VaMi3
Valve 2 Gas separation ccium
/
Micro compute
Indic.:lt4r

r
I/elt bcrded
FFA rrerb"ane
'----v--' '-----v---'
Gas separation part Gas detector part Diagnosis part

Figure 1.7: Hitachi On-Line Gas Sampling and Monitoring System (from [621).

monitoring systems for transformers developed at the Hitachi Labs [60,611 the detector
used is believed to be similar in operation to the TGS devices previously discussed.
In the Hitac.hi system, gas separation is performed using chromatographic colulllns
and a carrier gas of filtered air [621 (see Fig. 1.7). This necessitates much valving, an
air pump, and accompanying air filters to clean and dry the air. Demonstration of
the system on field transformers compared well to oil samples analyzed off-line over a
period of a year.

An alternative approach to gas analysis is Fourier Transform Infrared (FTIR) spec-

troscopy. In this optical approach to gas sensing, information about a gaseous mixture
is obtained in the infrared portion of the spectrum where rotational-vibrational modes
are excited. This technique can be used to resolve CO, CO 2 , and hydrocarbon gases
with an accuracy on the order of parts per million without the need for physically
separating the individual gases from the mixture. It will not detect symmetric gas
molecules such as H 2 , N 2 or O 2 , Additionally, the measurement can be performed in
less than a second allowing very rapid lllonitol'ing. Gas ext.raction from t.he oil would
then become the rate limiting step in performing gas analysis. There has been ex-
perience in monitoring air quality in process plants, sampling from several different.
locations, using an FTIR spectrometer [63].

43
1.4 Microdielectrometry

1.4.1 Description and Background

Microdielectrometry, as developed by Senturia and co-workers [64,65], is a commer-

cially available technique for measuring complex permittivity utilizing microfabricatioll
technology to incorporate both the sensing electrodes and associated circuitry on the
same microchip (Fig. 1.1) [66]. A set of interdigital, planar electrodes are deposited
on an insulating oxide layer. In addition to the interdigital electrodes, two FET tran-
sistors are also fabricated on the microchip. The electrical excitation and response
is obtained using a microdielectrometer, specifically a Eumetric SYSTEM II Microdi-
electrometer [67]. One set of electrodes is driven by a variable frequency (.005 - 10,000
Hz) AC voltage. Digital processing is used for both the synthesis of the driven voltage
and the measurement of the response. There are 58 discrete frequencies availabie for
the driven voltage. The response is the magnitude and phase of the complex ratio of
the potential developed at the second electrode to the driven voltage. This electrode is
allowed to float by connecting it to the ga.te of one of the transistors. The other tran-
sistor serves as a reference. Using the unity gain feedback circuit shown ill Fig. 1.1, in
conjunction with the transistors, the floating gate potential is reproduced at the gate
of the reference FET. Thus, a very high impedance measurement of the floating gate
voltage can be obtained, even at very low frequencies. Having two transistors on the
same chip automatically compensates for variations in transistor characteristics .:lue to
changes in the local environment such as temperature or pressure. The output of the
device, defined as the complex gain C, is the complex ratio of the floating to driven
gate voltage, (VF/VD)' and is obtained using a correlation analysis of the two volt.-
ages. This analysis also provides the va.lue of any DC component. of t.he signal, referred
to as the DC offset. DC offset can arise from: 1) a mismat.ch in characterist.ics of the
two FETs and 2) static charge buildup either on the electrodes or in t.he region sensed
by the electrode array. An in-de.pth analysis of the syst.em design and performance
of low frequency dielectric measurement systems, with the microdieled.rometer as a
specific example, can be found in a dissertation by CoIn [68].

44
 Note that the upper limit on frequency for the microdielectrometer of 10 KHz is
not a fundamental one regarding the electroquasistatic field approximation. There
is no reason why the range could not be extended upwards into the megahertz, and
possibly even gigahertz region, as long as the properties of the insulating oxide layer
are appropriately taken into consideration. The electroquasistatic approximation will
be valid for frequencies f < (cl A) where c = Jl/foJLo, the speed of light in free space.
For the microchip, the critical dimension is the length of the electrodes, on the order of
2 mm. The electrodes must be regarded ,as equipotentials along this lengt.h, implying
a minimulll electric field wavelength of 10-100 times greater. This yields a cutoff
frequency on the order of 109 Hz. However, this cutoff can be lowered considerably if
the inductance of the particular electrode strudure used resonates with the electrode
capacitance at a lower frequency.
Two specialized models have been developed for interpreting data obtained with
the microdielectrometer. One model is for the application of the microdielectrometer
to the monitoring of epoxy resin cures [69,70]. A microchip is embedded in a sample of
epoxy and the change in its dielectric properties with curing is monitored. This model,
incorporateu in the form of a lookup table supplied with the microdielectrometer
system [67], makes it possible to relate the measured complex gain to the complex
permittivity of a uniform infinite half space. A relat.ionship between the dielectric
properties and the curing process was established. For this case, a finite difference
simulation of Laplace's equation was used to determine the electric field distribution,
treating the region above the electrode structure as a semi-infinite, isotropic medium
[70, Chapter 3]. The second model is for the application of developing a humidity
sensor. In tIus case a thin film « 1 JLm) was deposited onto the electrode structure
and the change in its sheet resistance as a fUllction of the l'elative humidity in air
was monitored [71,72]. For modeling purposes, t.he t.hin film's influence was t.aken
to be predominantly in the interelectrode region. Thus, the thin film-oxide layer was
represented as an R-C transmission line shunted by lumped capacit.ors [ill. This made
it possible to represent the gate voltage as the transmission line response to a driving
voltage. Comparisons with both of these models will be made in Sect.ion 3.3.

45
1.4.2 Micro and Macro Sensors

Presently, the interdigitated electrode structures are available in two different sizes.
The microchip has an aluminum electrode structure of spatial wavelength (..\) of 50
I'm deposited on a thin Si0 2 layer roughly 1.5 I'm thick and having a permittivity
of 3.9fo. A ceramic macrochip, developed for high temperature applications, is also
available [74]. This ceramic macrochip, produced using thick film technology, has
PD-silver alloy electrodes of 1 mm spatial wavelength deposited on an Ai 2 0 3 layer
of permittivity 9.8fo and 250 I'm thick. There are no on-board electronics with the
macrochip. The longer spatial wavelength allows dielectric measurements to be made
on a larger length scale (up to 330 I'm). A detailed discussion of the macrochip,
including its design, operation and experimental results, can be found in a master's
thesis by Li [231. References to Li's work will be frequently made throughout this thesis
and is viewed as a companion document to this one. An important contrast bet.ween
the micro and macrochip is the decreased sensitivity of the macrochip to changes in the
dielectric medium above the electrodes. This is due to both the higher permittivity
and smaller ratio of spatial wavelength to layer thickness of the oxide layer for the
ceramic sensor. These parameters will be discussed further in Section 3.1.5. Also,
fringing fields are a greater problem with the macrochip due to the smaller aspect
ratio of the electrode fingers.
For an un coated microchip, the gain varies from roughly -40 dB in air to almost
o dB when the electrodes are shorted together. Due to a slight mismatch in transistor
characteristics the latter measurement is usually 0 ± 0.2 dB. The phase varies between
00 and roughly -90 0 , except for certain situations described in Sections 3.1.2 and 3.1.3
where the phase can go more negative. Typically, DC offset.s are negative and on t.he
order of a couple of hundred millivolts. There are various inst.ances under which t.he
DC offset will change significantly and these are described in Sect.iolls 5,4 and 7.1.8 for
un coated and coated microchips in oil, respectively. Generally, large changes in DC
offset occurred when low moisture content oil was flowing by a microchip. Increased
oil flow rates and temperature exacerbated this problem. Uncoated microchips were
much more susceptible to this problem than coated ones. For DC offset information

46
conc'erning the macrochip, consult Li's thesis [23, Section 2.1.2].

1.4.3 Preliminary Work

The early work on developing a continuum model representation of the microchip and
in gaining practical experience of using an uncoated microchip in transformer oil was
performed by Mouayad [73].
Experimentally, it was observed that the response of the microchip was sensitive to
adsorption of moisture at the surface of the silicon dioxide insulating layer, particularly
if this surface was dirty. Adsorption of moisture at the surface created a surface con-
ductance between the electrodes, thus altering the response. This effect occurred both
in air and in moderately wet oil (20-30 ppm). Therefore an uncoated microchip would
not be well suited as a sensor unless the surface adsorption and resulting conductance
was well controlled. This will be discussed further in Section 5.4.
Evidence for polarization at the microchip-liquid interface was observed in experi-
ments where the conductivity of the oil was varied through the addition of an antistatic
agent. Transformer oil is highly insulating, typically having a conductivity of 10- 13 to
10- 12 S/m. Double layer thicknesses resulting from charge separation at solid/liquid
interfaces vary with 1/ JU and can be significant in size with respect to the spatial
wavelength of the microchip. Stirring of the oil altered the measured response. It was
believed that the structure of the diffuse double layer was affected by the fluid motion.
These results will be presented in Appendix D.

1.5 Goals

1.5.1 Analytical and Numerical Tools

The first task is to develop a generalized cont.inuum model of t.he elect.rost.at.ic field
distribution about an interdigital elect.rode structure. This will allow simulation of t.he
wide variety of physical situations in which the microchip might. be used. In practice,
not all the parameters required to completely specify the model are known. A key

47
parameter, such as the complex permittivity of a medium under investigation, will be
unknown. By incorporating the continuum model into various paramet.f'r estimation
routines, interpretation of experimental data will be possible. Both single and multi-
parameter estimC\.tion routines will be developed for estimating quantities such as layer
thickness, complex bulk and surface permittivities and spatial distributions of these
properties.

1.5.2 Experimental Verificatioll of Model

Experimental verification of the continuum model and parameter estimation routines

must be performed to gain confidence in the ability to make absolute measurements.
Measurements of thickness and complex bulk permittivity, as obt.ained with the inter-
digital electrode structures, and the continuum model, will be compared to indepen-
dent measurements using standardized techniques. The development. of the continuum
model allows the microchip to be used for several new applications. One fairly straight-
forward development is to use the microchip as a thickness gauge for thin films formed
by deposition, evaporation or sedimentation of material.

1.5.3 Electrochemical Sensors For Transformer Monitoring

The primary task is to adapt the microchip and macrochip structures for transformer
oil monitoring. The motivation is the potential advantage of monitoring the oil almost
instantaneously with something as simple and cheap as a microchip. The simplest
measurement, that of the complex bulk permittivity of the oil, is not a straight.for-
ward measurement because of problems encountered with moisture adsorption by the
electrode struct.ures. Passivat.ion of t.he microchip a.nd lTIIH'roC'hip is a C'ritical ~tpp in
making this simple measurement.. This passivat.ion procedure, coat.ing t.he elect.l'Odes
with a moisture impermeable membrane, makes necessary the continuumlllodel t.o ex-
tract the correct complex bulk permittivity from t.he experiment.al data. While t.here
has been no conclusive links made between specific degradation processes and dielec-
tric data, the frequency information available with microdielectrometry may shed new

48
light on the situation.
Next, it is desirable to be able t.o monitor the moisture and dissolved gas mntt·nt. of
the oil. The key is to sensit.ize the microchip to a particular contaminant. This will be
achieved by coating the microchip with a thin film that preferentially au- or absorbs the
contaminant of interest and exhibits a change in its dielectric properties proportional
to the contaminant concentration. Again, the continuum model is required to correctly
extract the dielectric properties when using a coated microchip. In addition, parameter
estimation of the dielectric properties over the entire frequency range can help in
forming a physical model of the coating structure and in understanding the basis for
the sensitivity to the specific contaminant.
Efforts will be concentrated on developing a moisture sensor and preliminary de-
,
velopment of a hydrogen gas sensor, as there is previous evidence in the literature
of materials sensit.ive to these two constituents. Passivation of the microchip surface
is an essential step in developing gas sensors and is also important for developing a
moisture sensor.

1.6 Overview
The first part of the thesis concerns itself with developing and verifying the continuum
model, and parameter estimation routines, necessary for applying the microdielectrom-
et er to transformer monitoring. A derivation of the continuum model is presented in
Chapter 2. In Section 3.1, a rapport is developed for the typical responses of the
llucrodielectrometer when used to measure homogeneous media, thin films, surface
conductivities and/or complex bulk permittivities. Some general working theorems for
interpret.ing gain-phase responses are elaborated in Sect-ion 3.2. A In'j,,f comparison
with other models, for various physical limits, is presented in Section 3.3, Develop-
ment of the parameter estimation routines is covered in Chapt.er 4. Section 4.2 deals
with single parameter estimates of thickness 01' complex permittivity. Multiparamet.er
estimation of spatial distributions of dielectric properties, as obtained by varying t.he
spatial, rather than the temporal frequency, is examined in Section 4.3. Two param-

49
eter estimation of layer thickness aud permittivity by varying the dielectric medium
above the layer is discussed iu Section 4.4. Experimental verification of the cont.inuum
model and parameter estimation routines is presented in Chapter 5. Calibration of the
microchip and macrochip is described in Sectiou 5.1. The specific applicatiou of usiug
the interdigital electrode structure as a thickness gauge, wheu configured in both the
low and high impedance modes, is discussed in Sections 5.2.1 and 5.2.2. The effort
towards developing a sensor to monitor the complex bulk permittivity of the oil are
outlined in Section 5.4, including results obtained using sensors passivated with pary-
leue. Estimatiou of complex surface conductivities from experimeutal observations is
performed in Section 5.5.
The specific application of adapting microdielectrometry to t.he monitoring of mois-
ture in transformer oil is the subject of Chapters 6, 7 aud 8. Chapter 6 outlines the
potential increase in understanding of the condition of the paper insulation and oil
degradation via on-line moisture monitoring. A design strategy for sensitizing the
microchip to the moisture content in oil is given in Section 6.3. This sensitization
is accomplished through the plasma deposition of a thin film of bromobenzene onto
the microchip electrodes. Experimental development and testing of these microchip
moisture sensors is presented in Chapter 7. Section 7.1 gives the experiment.al results,
obtained ill controlled oil samples, used to characterize the sensitivity of the moisture
sensors. Results from testing of the moisture sensor in the oil circulatiollloop of an in-
strumented pole type (50KVA) transformer are given in Section 7.2. An investigation
of the electrical and physical properties of the plasma deposited btomobenzene films
is conducted in Chapter 8. The biphasic structure of the film, particles embedded in
a rigid film, is examined in Section 8.1. This structure is examined with respect to
the moisture permeabilit.y of the films and is related to evidence ill th~ lit~ratllJ'(> on
similar film structures. In Section 8.2, the dispersive nat.ure of t.he dielectric propert.ies
of plasma deposited bromobenzene is revealed and shown t.o follow a power law disper-
sion, common to many amorphous st.ructures. Experience from t.he lit.erat.ure on film
aging, and the measures that cau be takeu to increase film stability, are summarized
in Sectiou 8.3.

50
 Chapter 9 summarIzes preliminary attempts developing a. hydrogen gas sensor.
After mot.ivating this effort. in Section 9.1, inconclusive resu1t.s using mkrodlips coat.ed
with tin oxide and palladium are given in Sect.ion 9.2. Encouraging results using thin
film palladium resistors are shown in Section 9.3.
Finally, conclusions and an outline of possihle fut.ure directions are contained in
Chapter 10.
In Appendices A and B, mathematical derivations are presented in greater detail.
Appendix C documents the experience with the plasma deposition apparatus. Experi-
mental results examining double layer effects on the microchip response in doped oil are
presented in Appendix D. Data acquisition software used to take frequency and time
scans with the microdielectrometer is documented in Appendix E. All the parameter
estimation software, along with the continuum model, is presented in Appendix F.

51
Chapter 2

ContinuUlll Model

2.1 Outline of Approach

The objective here is to develop a general and adaptible model for determining the
output of the interdigital electrode dielectrometer. In Fig. 2.1 is shown a schematic
view summarizing the structure of the continuum model developed in the next section.
One set of inputs to the model, the lower box on the left in the diagram, describes the
electrode array structure. Parameters such as interelectrode spacing (a) and the spatial
wavelength (,\) are specified here. In addition, for the specific electrode structure
considered here, interdigital electrodes deposited on an oxide layer with ground plane
underneath, the electric field distribution below the electrodes is completely specified
by the insulating oxide layer thickness (h) and permittivity (Eoz). The other input,
the upper box on the left, describes the half space of media above the electrodes. This
description is in the form of a complex surface capacitance density C:, representing
the response of the half space to one fourier component of a potential applied at the
electrodes. As shown in the upper part of the figure. this response is t.he complex
ratio of the normal displacement field (D:) measured at the interface t.o the interfacial
potential (~~). The ~ signifies a complex quantity, the superscript indicates that. the
function is t.o be evaluated just above the interface and the subscript indicates t.he
nth Courier component. The quantity C: will be dependent on the particular medium
under investigation. All the heterogeneity and structure of the medium is encapsulated

52
in C:. Once these inputs have been defined the fields are completely specified.
The 'V/ box represents the step of solving the electrostatic field problem. Field
quantities in the oxide layer and in the medium are represented in terms of their fourier
components. Boundary conditions at the electrode interface are used to match up the
field solutions for the regions above and below the electrodes. A mixed boundary
value problem occurs at this interface, as the potential is constrained along that part
occupied by the electrodes while conservation of charge and Gauss' law combine to
put a constraint on the potential and on the jump in its normal gradient along the
interelectrode part of the surface. The solution involves discretizing the interelectrode
surface by introducing a grid of unknown voltages Vj at k collocation points. The
potential is assumed to vary linearly between these collocation points. A system of
equations is generated by maintaining a discretized conservation of charge along surface
segments bracketing these collocation points. These equations are written in the form
A. V =X where the elements of A are the coefficients of the unknown voltages Vj and
the elements of X represep.t the known voltage, VD' Inverting the matrix A yields
the voltage distribution along the interelectrode surface, and thus, the electric field
distribution.
In the final box, the evaluation of the Yi/s is tantamount to a determination of
the response with any terminal configuration. Due to the symmetry of the electrode
structure, Y u -:- Y 22 • All of the admittances representing the electrode structure are
determined by finding the electrical terminal currents, iD and iF' wit.h the floating gate
electrode grounded. These currents are obtained by integrating the current density
over the surfaces of the electrodes. Using these admittances, the response of the
microdielectrometer with an arbitrary load is the complex gain, the ratio of the floating
to driven gate voltage

(2.1 )

where Y i is the load admittance of the floating gate FET.

III the medium above the electrodes and in the oxide substrate below, the potential

53
 "a
<%>n
specification
of I\a
medium en
1\
Cn

YI2 -
+ +
A A r---- ,..- ..... A 1\
V· ~0 V. G
J YII y~
'ril
L-. r-- '--r--

-- -
electrode
array
specification
0, ~,h, EOI

L r+.. f ..

Figure 2.1: Schematic Representation of Continuum Model- Required Input.s, Sol ut.ion
for Collocation Potentials~ Two Port Network Equivalent, and Complex Gain Output.

54
is represented by (1.1) and the fields are taken as electroquasistatic, the electric field
intensity is irrotational and hence represented by the potential ~.

(2.2)

For completeness, the oxide layer will be given a complex permittivity, f b.

2.2 Electrode Structure Boundary Conditions

The electrode structure of Fig. 1.1 is modeled as shown in Fig. 2.2. The electrodes are
treated as being one dimensional, one set driven sinusoidally with a peak voltage of VD
and frequency w, the other set grounded. It is possible to include the finite thickness
of the electrodes using the techniques outlined here but the effort is not judged to be
worthwhile at present. For generality, an interelectrode spacing of arbitrary width a
is allowed. Thus, in a half period, the endpoint of one electrode is at y = yo and the
beginning of the other electrode is at y = 'Yo + a =YIc+1'
At the electrodes, the potential is constrained, whereas along the surface between
the electrodes there is only continuity of the potential

cI»b(y,t) = Re {Vne iwt } for O~y~Yo (a)

~b(y, t) = 0 for YIc+1 ~ Y ~ ~ (b) (2.3)
~b(y,t) = ~CI(-y,t) for O~y~~ (c)
On the interelectrode surfaces, Gauss' Law relates the discontinuity in the normal
displacement field to the amount of fr~e surface charge density, O',u, (ii is the unit
normal to the interface)

(2.4 )

where ~ ~ implies the difference ill the function within the brackets evaluated just above
and just below the interface and Vl:. is the surface divergence. The surface electric
displacement vector, 15" represents the possibility of polarization phenomena concen-
trated within the interface. Conservation of charge accounts for the accumulation

55
 _--'______ ~------L-- _____ _

"VD
----'--------------'------------ -
,
r~---------------------------------------------------A_---------------------------------------~,
x

A
2

Figure 2.2: Coordinate System Used for Electrode St.ructure of Fig. 1.1.

56
of O'.u via the discontinuity in normal conduction current 0), including conductivity
within tht: interface represented by the surface current density I.,

(2.5)

Substituting (2.4) into (2.5), using the complex notation introduced by (1.1) and
assuming sinusoidal steady state (:t --? jw) yields the following boundary condition

~ ~

ii. ~e·E~ + VI). (e:E,,) = 0 (2.6)

where complex bulk and surface perlllittivities have been introduced, e· = e' - j e"
and e: = e~ - j e~, and the conduction and displacement volume and surface current
densities have been combined (E" is the component of E in the interfacial plane),

~ ~ ~ . ~ .
jwe·E = J + jwD jwe:E. = I. + jwD,j (2.7)

For an ohmic conductor e' and e~ are constant, and e" = (0' /w) and e~ = (O',,/w), where
0' and 0', are constant.
For uniform surface properties (2.6) yields the second interelectrode region jump
condition,

(2.8)

2.3 Bulk Relations

As discussed earlier, the field distribution in the half space of material either above
or below the electrodes can he represent.eel hy the appropriat.e surfacf." capadt.ance
density. This quantity is the ratio of the the complex fourier amplitudes of normal
electric displacement to potential, both evaluated on the (a.) or (b) side of the int.erface.

(2.9)

57
Here, the subscript n refers to the order of the fourier modes. Given the distribution
of a complex permittivity e:( x) ahove the electrodes, with the coordinate iV perpendic-
ular to the interface (including surface singularities in that distribut.ion), the complex
surface capacitance density C: is found by determining the Fourier amplitude of the
normal flux density at the surface (a) caused by the nth term of the imposed potential
having the form
00

~(x,y) =L ~n(x)cos(knY) ; k n = 2nrr/)..

n=O

Examples of surface capacitance densities for various types of homogeneous and

heterogeneous media are given in Section 3.1.

2.4 Potentials at Collocation Points

The Fourier amplitudes of (1.1) are

~
cp~ = iJoI>' cpa(y) cos(kny) dy for n = 1,2 ...
(2.10)

for n = 0

The potential, ~a, is known only at the electrodes. On the surface, in the interelec-
trode region, the potelltial assumes a distribution such that it is cOlltinuous, (2.3c),
and cOllservation of charge, (2.5), is satisfied. The strategy taken here is to represent
the interface potential by piece-wise contilluouS linear segments having the collocation
potential Vj at each collocation end-point yj. Illustrated in Fig. 2.3 are four colloca-
tion points, illtroduced in the interelectrode region between Yo and Yk+l' An illit.ial
effort divided the region into equal length line segments [;3, p. 33]. A (.'osinusoidal dis-
tribution has been utilized here to give finer resolution in the regions of higher elect.ric
stress (adjacent to the electrodes). Thus, t.he collocation points are at.

Yi = (yo + ~) - (~) Cu~ (k ~ 1) for 0::; j ::; k + 1 (2.11 )

and the potelltial distribution is

58
 for 0:5 Y :5 Yo

v.(y - Yj-l)- V . _1 (y - Yj)

-~- . ___ ._1...: for Yj-l:5 Y :5 Yi
(Yj - Yi-l) (2.12)
j = 1,2, ... ,k + 1

o for Yk+1:5 Y :5 ~

Substitutillg for ~a(y} iuto (2.10) yields the Fourier series coefficients

~~ _ (A VD}2 ( 1 ) [cos(knyo} - cos(knYdJ

mr Yt - Yo

le •
" AV
+ Vi - 1) (k )
i [( 1 cos(knYHd
+~(mr)2 Yi+1 - Yj Yj-l cos 'nYj - Yj+1 - Yj

(2.13)
_ cos(knYj-dJ
Yi - Yj-l

The problem has been transformed to one of solving for k unknowns, the V;'s, with
the contilluity of potential satisfied by (2.3c). The second continuity condition at the
interface, (2.8), remaills to be satisfied. By dividing the interelectrode surface into
k intervals, centered on the collocation points, and satisfyiug conservation of charge,
(2.8), within every interval, k equations for the k Vi's are obtaiued. The intervals are
demarcated by points yj located at t.he edges of the electrodes and at. t.he midpoint.s
of the line segments formed by consecutive collocat.ioll point,s, Yj and YHl (Fig. 2.3).

Yo for j = 1
Y~J = HYj + Yj-l) for 2:5j:5k (~.14'

y/c+1 for j=k+l

59
 "V

k =4
- - - _....

- -----,

"V3 --- I - I -
I I
1\ I I
I
V4 --- I

0 YO'1I '12 Y3 '14 '15 Y

1* I
Y, Y2* '13* Y4* '15·
~ ~I
a

Figure 2.3: Voltage Distribution Along Discret.ized Illterelectrode Surfa.ce for k=4 Col-
location Points. Unknown Volt ages Vj are Introduced at t.he Points Yj. Conservation
of Charge is Maintained for Lille Seglllellts Defilled by the yj.

60
With this choice of y;, the ambiguity in the value of the tangential electric field at the
collocation points Yj is avoided. URing (2.2) and (2.12) yieltls

E(I(y~) = _
Y )
(Vj -Vj-l)
y') - y')-1
(2.,15)

Note that the tangential field is antisymmetric about the x-axis. As shown in Fig. 2.4,
E:(y) is a piecewise discontinuous function that is not well defined at the collocation
points yj. The intervals over which charge is conserved (between broken lines in
Fig. 2.4) extend half a segment in either direction about a collocation point Yj except
for the first and last intervals, which respectively have end points at the electrode
edges. In addition, the values of the tangential field used for the endpoints Y; and
Yk+l are not averaged though they are at a discontinuity. It is more realistic to assume
that these values, for their respective segments, extend to the electrode as shown in
Fig. 2.4.

Integrating (2.8) over each interval, from yj to Yj+u and setting the result to zero
yields k equ8tions for k unknowns,

y;Y;+I
/.
[ • (I
ft • b
€(lE:z:(Y) - fbEAy)
ft

+ f.fO• 8E:(Y)]
8y dy = 0 for j = 1,2 ... k (2.16)

The third term, reflecting contributions due to surface effects, integrates to

(2.17)

and using (2.15) yields,

€.
.f0
[( 1
Yj+l - Yj
+
Yj - Yj-l
1 ) V. _ Vj+1
) Yj+t - Yj
_ Vj - 1
Yi - Yj-l
1 (2.18)

Integrating the remaining two terms of (2.16), using the cosinusoidal dependence of
the normal electric field upon x implied by (1.1) and (2.2), yields t.he set of equat.ions

61
 c<W~ 4-----------~ ...------------

Figure 2.4: Distribution of Tangential Electric Field~ E:(y)~ Along Electrode-MediulU

Interface.

62
 f.k~ (€:E: - €bE~) [sin(knY;+l) - sin(knYJ)] + [€b E; - €:E:] (Y;+l - yj)
n=1

€ • [(
+'0 1
Yj+l - Yj +- -1--
Yj - Yj-l
~J -
) V, VLJ..l
- L - J ____
Yj+l - Yj
- V'_l
--.:....L::.__
Yj - Yj-l
1
-
-
0 for .-
J -
1 2
, ...
k
(2.19)
Again substituting for the electric fields, this time using (2.9) (remembering continuity
of the potential across an interface) into (2.19) yields

+ €.o• [( Yj+l 1- Yj + Yj . - 1Yj-l ) V,J _ Yj+l

Vj +- 1Yj _
Yj -
Vj -Yj-l
1 l} --
0£ .- 1 2 k
or J - , ...
(2.20)
The set of equations described by (2.20) can be expressed in matrix form, A.V=X,
with the elements of A representing the coefficients of the unknown normalized volt ages
Vj = (V j/VD), and the elements of X representing the coefficient of the known drive.
In order to generate the matrix elements, (2.13) is substituted into (2.20) and terms
are rearranged such that all terms proportional to the drive are moved to the right
hand side. Any expression involving C; also goes on the right hand side as it can only
arise from an externally applied field and thus, is also a driven signal. This yields

(2.21 )

where

63
 • [(_1._
Yj+l - Yj
+ .__ J_) cos(k y.) _ ~~_s(k..~Yj-j-Jl
Yj - Yj-l Yi+l - !lj
n J

(a)
(2.22)

(b)

A3(i,j) _ E·
'0
[(_1_
Yi+l - Yi
+ -~-) o(i _}') _ §{(- i.±Jl
Yi - Yi-l Yi+l - Yi
(c)

_O(i-j-l)]
Yi - Yi-l
and

(2.23)

where

•( 1 ) [cos(knyo) - cos(knYdl (a)

Yl - Yo
(2.24)

X 2 ( i) - - [C~ - C! (Yo t Yt ) ] (Yt+ 1 - yt) (b)

( ')
X31 • 6(i - 1)
=E---' (c)
'0 Yt - Yo
where Yj and yj have been previously defined and 6( i) is defined as unity if t.he argu-
ment is zero and zero if it is nonzero. Using this representation, the set, of k equat.ions
in k unknowns can be numerically inverted to determine the V/so A mort. rapidly
convergent series for (2.22a) and (2.24a) is derived. ill Appendix A.

64
2.5 Equivalent Circuit Parameters
The pi network of Fig. 2.1 is helpful for formalizing the transition from the field
analysis just undertaken to the general situation where the detection electrodes are
not grounded but rather, are free to assume a potential consistent with the attached
circuit. With the driven gate at potential VD and the floating gate grounded, Y u
and Y 12 are directly related to the electrode currents, and hence calculable from the
electric fields just determined with the floating gate grounded.

IF - -Y 12 VD
(2.25 )
ID (Yu + Y 12 ) VD
The electrode currents are found by integrating (2.5), conservation of charge, through-
out a volume enclosing each electrode. Assuming a total electrode meander length of
k1e/' these currents are

(2.26)

Solving (2.25) for Y u and Y 12 in terms of iD and iF and incorporating (1.1), (2.2),
(2.9) and (2.15) yields

65
 (2.27)

- cfJ~ O [ Co
~ a - Co
~ b] (..\2 - YIe+1
)}

where Vj =(Vj/VD)' Once (2.27) is programmed, it is only the surface capacitance

densities, C:, that are altered when the medium above is changed.

2.6 Determination of Gain

At this point it is instructive to recap the major steps for calculat.ing the Yi/s and
hence, the gain. Solution of a system of equations of the form A. V=X with the matrix
elements defined by (2.21)-(2.24) determines the Vi's and the electric field distribution.
With these V/s, the fourier coefficients for the potential at the interface, cf?~ and cf?o,
using (2.13), and hence, the circuit parameters Y 11 and Y l2, using (2.27), are evaluated.
FiIlally, given a load admittance Y" the gain is then calculated according to (2.1). The
software routine for computing the gain-phase response is called mz1.for and may be
found in Appendix F.

66
2.7 Collocation Points, Fourier Modes and Conver-
gence
Several tradeoffs must be considered for deciding the optimum number of collocation
points and fourier modes. Increasing the number of collocation points improves the
representation of the actual potential distribution - at the cost of greatly increasing
the computing time required to calculate each element of the matrix A and to COlll-

pute the matrix inverse. Increasing the number of fourier modes used to compute the
voltage distribution increases the accuracy - of representing this distribution with a
piecewise linear function. Beyond a certain number of modes the piecewise represen-
tation may differ significantly from the actual one. However, as previously mentioned
and elaborated in Appendix A, a more rapidly convergent series is used to SUlll the
fourier modes for the elements of A matrix. This series effectively sums up roughly
1,000 modes. In this case, it is quicker to comput.e either a small or a large number of
modes than t.o compute an intermediate number. Using the large number of modes,
the number of fourier modes summed in (2.27) when computing Y 11 and Y 12 can be
altered. Increasing the number of terms used here has a much smaller, though still
significant, effect on the computing time.
Determination of convergence requires a specification against which various results
can be compared. For the case of the microdielectrometer, where comparisons are
made with an actual output, the accuracy of the gain is given as 0.1 dB [67]. This
accuracy will be used as a tolerance for measuring convergence. Modeling an uncoated
microchip in air, the high frequency coupling capacitances per unit length (normalized
to the oxide layer permittivity) and resulting gains are tabulated for a matrix of values
for the number of collocation points and fourier modes (Table 2.1). The normalized
load capacitance used was 1.46. It was observed that using 25 collocat.ion point.s and
100 summation t.erms provides sufficient convergence when compared to t.l~e maximulll
values of 35 and 1000, respectively (as well as taking a reasonable amount of computing
time). For the rest of this thesis, unless otherwise noted, 25 collocation points and 100
summation terms will be the standard numbers used.

67
Table 2.1: Predicted Gain Output vs. k Collocation Poiuts and N Summat.ion Terms.

Model Par.ameters Coupling Capacitances Gain (dB)

k N I CU/fozM I el C12/fozlvlel I Gp
10 50 8.966 .2055 -34.275
100 9.054 .2090 -34.203
150 9.027 .2107 -34.114
1000 9.023 .2096 -34.153
~--------------~---------+----------~----------
25 50 8.822 .2042 -34.210
100 8.939 .2095 -34.089
150 9.007 .2112 -34.074
1000 9.110 .2150 -34.006
f-----------+-----+-------t------.-.-
35 50 I 8.809 .2044 -34.194
100 8.914 .2090 -34.089
150 8.970 .2107 -34.064 I
1000 9.132 .2I.59 -33.989 I
An additional note on the implementation of the model - there is a symmetry in
the matrix A1(i,j) of (2.22) that is exploited to reduce the number of computations.
It turns out that for k collocation points, producing a kxk matrix, At{i,j) = Al(k -
i + l,k - j + 1), thus halving the ~umber of summations that need to be computed.

68
Chapter 3

Typical Responses Using

Continuum Model

As a first step toward developing a scheme for identifying continuum parameters,

Section 3.1 is devoted to examining predicted frequency responses for a variety of
physical situations that would be commonly encountered. Insights formed from such
an examination, combined with confidence gained from experimental evidence, result in
the observations of Section 3.1.5. In addition, some working theorems are developed
in Section 3.2 that lllay be useful in interpreting experimental gain-phase data of
heterogeneous media. These theorems are exploited in Chapter 7 to interpret responses
obtained using the moisture sensors. Comparisons of results using the continuum
model with several limiting cases having analytical solutions are made ill Section 3.3.
The good agreement between models enhances the validity of the continuum model
developed here.

3.1 Predicted Responses

AG Aa
For all of the following examples Eo = 0 (n = 0 mode) thus Co = O. Using t.he form
expressed in (1.1), solution of Laplace's equation for t.he potential in the insulat.ing
oxide layer below the electrodes gives

69
 ftb C
ft b
Cn = -fczkn coth(knh) j '0 = - h-
1:0:1:
(3.1)

where the layer has a thickness h and is bounded from below by a "perfectly conduct-
iug" substrate.

3.1.1 Uniform Media

This situation, a semi-infinite half space of uniform complex permittivity, is pktured

in Fig. 3.la. The potential satisfies Laplace's equation and decays to zero as :r: goes to
infinity. Thus,

ft4

C: = f~En = Eikn
(»4
n
(3.2)

(These expressionf. illustrate how transfer relations naturally fit into the spectral field
description [75, p. 2.33]).

Typical responses, shown in Fig. 3.2, have the features of a linear system with one
time constant - the charge relaxation time, Te = (Et /O't). At high frequencies, WTe ~ l,
the effect of dissipation currents is small. The coupling between the two electrodes
is purely capacitive, explaining the low gain at high frequencies. As the frequency
is lowered, WTe approaches unity and conduction currents become appreciable. Thus
the gain increases and a nonzero phase occurs. Specifically, there is a peak in phase
and the slope of the gain is -20 dB/decade, both indicative of a uniform medium
(which, for constant permittivity and conductivity, is characterized by a single time).
At very low frequencies, T ~ Te , the medium becom~s an extremely good conductor
and consequently the electrodes are shorted together (0 dB gain and zero phase). It is
evident from the equations that the frequency enters only as a rat.io, (WEt/Ut) == v)Te •

Thus, the frequency scales with the bulk conductivity - altering t.he conductivity
merely shifts the frequency response without changing the curve shapes (Fig. 3.2).
Changing the bulk permittivity alters both the breakpoint. (due t.o a change in Te and
the high frequency gain (due to a change ill the capacitive coupling), as shown in
Fig. 3.3.

70
 (0 )

)
< EOX·(T =0
( b)

Figure 3.1: Electrode Structure Below a) Uniform Medium or b) Variable Thickness

Homogeneous Layer and Uniform Medium.

71
 o~--~----~--~----~--~~--~--~

-10 0"1 =I. 3 x 10-6S/ m

-
£D

-
-c -20 I x 10- 9
.-cc:
(!) -30

El = 2 x 10-11 F/m
-40

- 20

QJ
(I) - 40
c
.t::
a.
- 60

- 80

- 3 -I 3
log f

Figure 3.2: Predicted Responses for Microchip in Uniform Medium, Vary Conductivity
0'1 (€1 = 2:r10- 11 F fm).

72
 O~---r----~--~----~---.----.----,

-
CD
I x 10- 10

-.-
"'C

c:
c
(!) 2 x 10- 11
eT1 = I x 10-9S/m
-40

- 20

-10
Cl) E,=2xI0 F/m
(/) -40
c I x 10- 10
~
a..
-60

-80

-3 -I 3
log f

Figure 3.3: Predicted Responses for Microchip in Uniform Medium (0"1 = 10- 9 S/m),
Vary Permittivity €t.

73
 One of the first applications of the microdielectrometer was to monitor the curing
of epoxy resins. A finite difference simulation, employing grid elt'ments and relaxation
methods, to determine the electric field distribution was developed by Lee [70, Chap-
ter 3). Under the assumptions of dealing only with linear, homogeneous systems (no
spatial variation in fl or 0'1) and setting aside any surface conduction phenomena,
the simulation calculates the gain and phase at a specified frequency for a material
with a given (;., and fll. A look-up table, compiled from many simulations, provided
a unique conversion from gain, phase and frequency information to (;.' and f". Dielec-
t.ric relaxation information can then be extracted from plots of fit vs. f' (Cole-Cole
plots). When the phase is plotted versus gain for the curves shown in Figs. 3.2 and
3.3 (eliminating frequency as the independent parameter), the graph reproduces the
look up table generated by this finite difference simulation.

3.1.2 Variable Thickness Layer

A layer of thickness d2 and characterized by a uniform complex bulk permitt.ivity is

placed immediately above the electrodes (Fig. 3.1b). Solution of Laplace's equation in
a piece-wise fashion, perhaps using transfer relations (see Appendix B or [i5, p. 2.331),
glves

(3.3)

As background for inferrin~ the layer thickness from a measurement of the gain,
the effect on the frequency response of varying d2 while holding the complex bulk
permittivities of each region constant is illustrated by Fig. 3.4. In t.his c,ase, the half
space is taken to be more insulating than t.he layer and t.he surface properties of t.he
layer-half space interface are taken to be zero.

For a thick layer, d2 > (A/21r), the response is that. of a uniform half-space, de-
termined solely by the propert.ies of the layer. The electrodes do not. "see" beyond
the thick layer for purely geometrical reasons. As t.he layer thickness is decreased t.he
response is still a bulk one, but now it also reflects the fact that t.he fields have begun

74
 O----~----~----r---~----,-----~---,

-10

-m - 20 d2= 50JA-m

-
"C

.-cc: - 30
5

(!)

-40

- 50

I--+- d2 = 50fLm
-40 /-1---+--- 5
r---H'----+--I
Q)
(I)

.,g -80 -11

a.. E2 =2x10 Fim
0"'2 =2.4 x 10-11 S/m
-120
El = Eo F/m

- 3 -I
log f

Figure 3.4: Predicted Responses for Microchip with Variable Thickness Layer in Air
(E2 = 2:c10- 11 F Im, (T2 = 2.4:&10- 11 S/m).

75
to penetrate into the semi-infinite region, where the permittivity is less (accounting for
the decrease in high frequency gain) and the conductivity is zero. At d 2 '" .1(A/211"),
the layer is so thin that it might well be described by a surface conductivity on the
interelectrode interface. This case is discussed in the next section. Further decreases
in the layer thickness reduces the effective interelectrode surface conductivity. This
results in a shift of the interelectrode surface type response to lower frequencies.
Generalization of the en to include multiple layers is presented in Appendix B.

3.1.3 Singular Properties at Substrate-Medium Interface

This is again the case of a uniform half-space (Fig. 3.1a) with e: given by (3.2),
but with a surface conductivity at the Si0 2 -mediulll interface between the electrodes
f: o = -j(u80 /w}. A typical response is shown in Fig. 3.5 - clearly distinguishable
from a bulk response. There is an additional characteristic time, T8t! = (fd kn ( 80 ),

representing surface charge diffusion along the interface. Again the interplay between
surface and bulk conduction and displacement currents accounts for the transition
from low to high gain and a peak in phase. However, the slope of the gain is greater
than -20 dB/decade and the gain has an overshoot. The phase curve is asymmetric
and the phase peaks at a larger angle than previously seen. All these are characteristic
of the surface charge diffusion from driven to floating electrodes made possible by the
combination of the distributed surface conduction and the shunt capacitance between
the interface and the highly conducting substrate.

As the surface conductivity decreases (T.e increases), the curves pass through a
regime where surface and bulk conduction are equally important to a point beyond
which a purely bulk response is observed (Fig. 3.6).

The overshoot phenomena, a prediction of gains below t.he purely capacit.ive high
frequency gain, is further evidence of the surface charge diffusion process. This can be
explained by separating the coupling between the electrodes into two component.s, one
through the bulk medium and one through a distributed transmission line composed
of the interelectrode surface and the insulating oxide layer. The coupling through the

76
 0
-11
E'I =2 x 10 F/m
-10
= 1 x 10-11 S/m
-
(J.
I
£D DSo I x 10~4S
-
"'C

.-cc:
t!) -30

-40

-40
Q)
en
C
..c:
a..
-80

-120~--~ __~____~__~__~____~__~
-3 -I 3
log f

Figure 3.5: Predicted Response for Microchip in Uniform Medium wit.h Surface Con-
ductivity all Illterelectrode Surface (El = 2x10- 11 F jm, 0"1 = 10- 11 Sjm, 0".0 = 10- 14 S).

77
 o~----~--~----~--~----,---~-----,

-18
U sa =10 S
~~_IO-16
-CD __-\--10- 15

-"'C

c
\000---10- 13

o
<.!) - 30
_11 ~~:f,III==---4-.,e:=::::::..--~
El =2 x 10 F/m
-40 -11
0'"1 =I x 10 S/m

-40
Cl)
(/)
c
~
a.
-80
10-18
-16
10 10-15
-120~--~--~ ____ ~ __ ~ ____ ~ ________
-3 -I 3
log f

Figure 3.6: Predicted Responses for Microchip in Uniform Medium, Vary Surface
Conductivity on Interelectrode Surface (El = 2:z:10- 11 F Im, 0'1 = 10- 11 S/m, 0"0)'

78
bulk makes a smooth transition from purely conductive to purely capacitive just as
in the bulk response and therefore involves at most a 90 0 phase shift. However, t.he
part of the signal resulting from transmission along the distributed RC transmission
line comprised of the interface and the insulating oxide layer suffers larger phase shifts
and hence a contribution that tends to cancel that due to the "direct" coupling. The
result is a frequency range over which the gain is smaller than that at high frequency
and the phase shift larger than 90 0. The diminution of surface conductance coupled
with an existing large ~ulk capacitance "snaps" the response from a surface conduction
coupling to a purely capacitive coupling and accounts for the rapid changes exhibited
in the frequency response.
An important parameter here is the silicon dioxide layer thickness, h. The thinner
the oxide layer the greater the coupling between the electrodes and the substrate. This
results in a reduced capacitive coupling between the electrodes through the oxide layer
and necessitates the incorporation of a surface permittivity for the model to predict
phases within the limits of 00 and -180 0 (see Fig. 3.15). Details on the effect of
altering h are in Section 3.1.5 and on the requirement of a surface permittivity are in
Section 3.3, case 4.

3.1.4 Layer With Surface Conductivity

Polymer films are often coated with conducting surface agents. With the objective
of measuring the surface conductivity of these films without electrical contact, the
frequency response of a layer (Fig. 3.1b) with surface conductivity att.ributed t.o the
layer-medium interface is of interest. The surface capacitance density is given by (3.3)
with €:2 having only an imaginary component.
III Fig. 3.7 is shown the frequency responses for an electrode st.ructure wit.h a layer
thickness d 2 = .L\ in air. By increasing t.he layer's surface conductivit.y, as shown,
two relaxation phenomena are observed. The higher frequency one result.s from t.he
spectrum of times associated with diffusion of surface charge along the upper surface
of the layer. The lower frequency one is from relaxation determined by the bulk
properties of the layer. If (T~2 <t:: (T2'x, so that bulk conduction in the layer dominates

79
that of the surface, theu relaxation through the bulk of the layer will come iuto play
at a higher frequency and the relaxation due to the surfa.ce condudivit.y wi1l not. be so
easily observed. Altering the layer thickness, with an observable surface conductivity,
will affect the capacitance of the layer and thus, the mid-frequency gain plateau and
the higher frequency phase peak.

3.1.5 Sensitivity to Electrode Structure Parameters

There are four parameters describing the electrode structure, each with its individual
effect on the gain-phase response. The spatial wavelength, A has been discussed pre-
viously with the application of intimate sensing. Thus, this value should be consistent
with the characteristic length of the medium under measurement.
Two other parameters are the insulating layers' thickness, h, and permittivit.y,
f oz • They are dependent on the material used (for foz) and the fabrication method
(for h). As mentioned earlier, varying h can significantly affect the response to a
surface conductivity at the electrode boundary. This effect can be extended further
to includf! the response to thin (d 2 '" O.1{Aj21T')) layers having either bulk or surface
conductivities. An example is the response of the macrochip to surface conductivities.
With its current, relatively thick oxide layer with high permittivity, there is no surface-
like response with overshoot gains and asymmetric, large phase peaks when examining
such thin layers [23, Section 2.31. As shown by Li, this can be altered so as to optimize
sensitivity to these type of phenomena by reducing hand f oz •

This lack of sensitivity with thick, high permittivity insulating layers actually ex-
tends to measurements of not only surface but bulk properties of thick and thin layers.
Referring to (2.1), the gain changes wit.h the material above due mainly to changes in
Y 12 • The admittance Y u is basically capacitive and much larger than t.he capacitive
part of Y l2 • As shown by Li [23, Section 2.2.21, Y 12 can be thought of as representing
two parallel couplings between the floating and driven gates, one above the interface
and through the medium, and the other below the interface and through the insulating
layer. A contrast in this coupling above and below is shown in Fig. 3.8 for large and
small values of h with the permittivity the same above and below the electrodes. With

80
 - 10

-aJ
~ -20
E2 =2
-9
0'"2= I x 10 S/m
-11
x 10 F/m

£:
d2 = 5JLm
(5 - 30 El =EO F/m
0-,-0
I -
-40

-40
OJ
(J)
0 10- 12
.c.
a... 10- 14
10- 16
- 80

-3 -I 3
log f

Figure 3.;: Predicted Responses for Microchip with a Layer Having a Variable Surface
Conductivity ill Air (d 2 = 5 p,m, €2 = 2xlO- 11 F /m, (12 = 10- 9 S/lll, (1,,2) •

81
 h large, the couplings above and below are equal. Changing the permittivity of the
medium above will change the net coupling (the sum of the two couplings) by a much
smaller percentage. If the permittivity below is much larger than above, this percent-
age change shrinks even further. With h small, the coupling below is much smaller
than above. Most of the field below couples to the ground plane, thus the contribution
to Y12 from below is greatly reduced. Now changes in the medium above a~e more
fully reflected in changes in Y 12 and the sensitivity is greatly enhanced. Thus, for
maximum sensitivity, the oxide layer thickness and permittivity should be minimized
(subject to fabrication constraints).

The fourth parameter is the interelectrode spacing, a. Using physical iniuition, if

a is decreased then the electrode-ground plane struct ure approaches that of a parallel
plate capacitor. It is as if a is an "electric" shutter controlling the amount of electric
field penetrating into the medium. Sensitivity to changes above will be reduced with
small values of a. However, increasing a, while improving sensitivity, decreases the net
coupling between the electrodes. This reduces the floating gate voltage and requires
the electronics to accurately measure very small volt ages with correspondingly small
sigllal-to-noise ratios. Thus, it appears that the (A/4) width for a is a valid middle
ground.

3.2 Working Theorems

The following paragraphs represent a set of guidelines established while interpreting the
gain-phase responses of a variety of media examined with the micro and macrochips.
It is nut meant to be a complete list, rather, it is just a set commonly encountered in
the work performed here and in Li's [23] thesis.

3.2.1 Master Curves

Definition 3.1 A master curve is the gain-phase response of a medium that displays
shape invariance with respect to variations in an external, physical parameter.

82
 J.. _ I
- -
A- .la...
'10 -v.0.
\
--

<'l}

,/
/'
,- -~- - "-
"-
/

I
,/
./
'-)--

(b)

Figure 3.8: Electric Fit-id Distribution as a Function of Oxide Layer Thickness.

83
Master curves are obtained by normalizing various frequency responses to the appropri-
ate parameter. For example, a master curve can be drawn for t.he set of curves shown in
Fig. 3.2 by normalizing the frequency to the phase peak frequency, JP (Fig. 3.9). Once
a master curve has been produced and the sensitivity of fp to changes in the physical
parameter of interest established, interpretation of experimental data is straightfor-
ward.
There are two types of media frequently encountered in this work that produce
master curves - ohmic conductors and dispersive media. The special c~se of a semi-
infinite ohmic conductor was just given as an example. This case will be shown to be
a subset of dispersive media having a spatial distribution of complex permittivity.

Theorem 3.1 If the comple:r: permittivity of a medium having a ~patial di~tribution

of di~per~ive dielectric propertie~ can be e:r:preued a~

E"' = E'(~,:c) -jf"(~,:c) (3.4)

Wp wp
where the dependence on e:r:ternal parameter~ can be lumped into

Wp = f(T,RH,c ... ) (3.5)

then a master curve can be produced.

Note that the real and imagillary parts can have different spatial depelldencies. The
frequellcy dependence is scaled by W p , a parameter contaillillg the dependence on exter-
nal variables such as temperature, humidity, or concentration. If it is possible to lump
such dependencies into Wp then the relevant para"11eter is the normalized frequency
(w/w p ). Thus, the responses will be shape invariant. and jl1~t shift in freqllf"llcy, pro-
ducing master curves. Examples of such dispersive phenomena are encount.ered in t.he
moisture sensor experiments of Chapter 7.
For the special case of a nondispersive medium, f' is independent of wand Elt varies
inversely with (w/wp). This yields an ohmic conductor with a spatial dist.ribution o.f
properties that remains constant with variations in external parameters. Again, the

84
 o .,
-5
, ~x
-10 X
X
-15
X
$: -20 X
"" X
IU X
tJ) -25
X
-30 \. X
.

-35 X~~~

-40

-45
-3 -2 -1 0 1 2 3 4

0 ..
~
IV
. ,:;)!.
-10 X'
.><
X X
-20 X X
",....
-30 X
X
X
Gl
Co?
-40
X X
/U
~
P-4
-so X X

-60 X
X
X X
-70 X
X
-80 X
X
-90
-3 -2 -1 0 1 2 3 4
log ({'/Fp)

Figure 3.9: Predicted Responses of Fig. 3.2 Normalized to Frequency of Phase Peak,

85
responses can be normalized into one master curve. Examples of responses produced by
such spatial distributions may be found ill Li's work [23, Section 5.3 and Appendix AI.

3.2.2 Spatial Discontinuities

Very often, the medium uncler examination is a multilayered structure that can be
modeled as a piecewise homogeneous structure (Fig. 3.10a). Further, the relaxation
time constants associated with each layer may be quite different from each other as
the layers vary from insulating to semi-insulating to semi-conducting as one moves
away from the electrodes. Assuming the layers are within the sensing region of the
particular spatial wavelength sellsor used, the gain-phase response associated with
such a structure will then exhibit the relaxation phenomena of each layer separately
as shown ill Fig. 3.10b. Each plateau in gain, accompanied by roughly zero phase,
corresponds to the purely capacitive coupling between the electrodes through all t.he
layers of the medium that remain insulating and, this is important, are below the
first layer from the electrode plane that has become essentially perfectly conduding.
This perfectly conducting layer forms an equipotential above which the fields do not
penetrate. Note that if this perfectly conducting layer is the lowest layer, adjacent to
the electrodes, it becomes impossible to see anything beyond this layer using one spatial
wavelength and temporal information. It will be shown in Section 4.3 however, that
distributions of conductivity that decrease away from the electrodes can be discerned
using variable spatial wavelength iuformation.

Thus, the plateaus in gain are determined by the permittivity and thickness of t.he
layers below the first perfectly conducting layer. If all the permittivities are known
and all but one thickness is known, then t.hat thickness can be t"st.imat.ed from t.he
experimental gain (see Section 4.2.1). For example, using t.his t.edllli(IUe one can
estimate the air or oil gap thickness for a material not in int.imate contact. with t.he
macrochip [23, [po 56 and p. 75].
Not only can layer thicknesses be estimated, but dielectric propert.ies can be es-
timated and associated with the appropriate layer. The gain transit.ions bet.ween
plateaus, and accompanying phase peaks, are the result of a layer going from capac-

86
 "- (o.,<r- 0

(t;t )

°1
-5 f-
-10

-15

"'
-20
-25
• t :;
O't
10- 11 F 101
= 10-; 5/01
'""----- --- ...... ,

"''''--JI
~
42 = 4 I'm
-30 ~J '" 7zlO" 11 F/m
0'2 = 10- S/m 10

-35 4Ia = 2 I'm

ts =3zl0- F/m 11

-40
I
O'J = lO-U S/m I
-45 1
j
-3 -2 -1 0 1 2 3 4

-10 L
Or-------------------------------------~
~'-.
~ ,..-----1
//
---.-', /'
I' ,,_/

-20
-30
G.l
(11
-40

"'
..c:
~
-50

-60

~:t
-3
I
-2 -1 0 1
log f'req
I
2 3 4

Figure 3.10: a) Multilayered Structure with Spatial Distribution III Complex Bulk
Permittivity and b) Predicted Response using Microchip.

87
itive to conductive as ihe frequency is decreased. Again the requirement is that the
relaxation times are distinct. enough so that. these transit.ions anti peaks do not. signifi-
cantly overlap one another. A good example of this is the estimation of the dispersive
property of paper insulation within a multilayered structure [23, Section 5.31.

Theorem 3.2 Given a multilayered structure of piece wise homogeneous dielectric prop-
erties, each layer, of known thickness, having an associated relazation time. distinct
enough from the others such that the gain has plateaus and the phase goes to zero at
some time during those plateaus, and further, the relazation times decrease with dis-
tance from the electrodes, then the dielectric properties and thickness of each layer can
be uniquely determined from the frequency response.

In addition to bulk dielectric properties, this approach can be extended to include

surface properties. The proviso of distinct relaxation times must be satisfied by the
surface properties as well, as demonstrated in Fig. 3.7. For (J'.2 < 10- 16 S the response
is dominated by the bulk relaxation within the layer. Only at (J'.2 2:: 1O- 12 S does
the surface coupled relaxation significantly stand out in the response. There is a
question of uniqueness in ascertaining whether the response is truly due to a surface
or bulk conductivity. When the surface conductivity is upon eithet· a relatively thick
layer, at some distance from the electrodes, or a relatively high permittivity layer, the
response will lose the overshoot in gain and large, asymmetric phase, characteristic
of the surface charge diffusion problem, and begin to resemble the response due to a
bulk conductor (Figs. 3.11 and 3.12, respectively). It must be known a priori whether
the conduction is bulk or surface in origin to perform the correct estimation using
the temporal frequency response (see Li for estimation of surface conductivity of pink
polyethylene [23, p. 56]).

It has been demonstrated that under certain conditions information on t.he spat.ial
dependence of dielectric properties can be est.imated from t.he frequency response.
A more natural approach, as alluded to in the section on intimate sensing (Sect.ioll
1.2.2) is to estimate spatial distributions using variable spatial wavelengt.h data. This
approach will be examined in Section 4.3.

88
 0

d2 =.5 #,m (I = to F/DI

-10 0'1 =
0
1 "2 = 2z10- 11 F/m
0'2 = 0

-20
0',2 =to- 16 S

....s: 5

'"tn -30 10

-40
•

-50~-----L------~----~------~----~------~----~
-3 -2 -1 o 1 2 3 4

-20

-40
CD
tit
-60
'"
.a::
$:Lt

-90

-100

-120~----~----~----~----~----~~----~--~
-3 -2 -1 o 1 2 3 4
log Preq

Figure 3.11: Predicted Responses for Microchip with an Insulating Layer of Vari-
able Thickness (d 2 , f2 = 2:dO- ll F /111, 0'2 = 0 S/111) having a Surface Conductivity
(0'62 = 10- 15 S) in Air.

89
 0

-5
-10

-15

-20
.....u
$:

tn -25
-30
E) = Eo F/m
0') =0
-35
d2 = II'm
0'2 = 0 lz10- 10
0'.2 = 10-)5 S
-40
-45
-3 -2 -1 o 1 2 3 4

°r-~~--------------~~~------------l

-2[)

-40

-80

-100~----~----~--~~----~----~----~----~
-3 -2 -1 o 1 2 3 4
log f'req

Figure 3.12: Predicted Responses for Microchip wit.h an Insulating Layer of Va.ri-
able Permittivity (d 2 = 1 p.m, f2, U2 = 0 S/m) having a Surface C()!1~"" vity
(U.2 = 10- 15 S) in Air.

90
3.3 Comparison With Analytical Solutions
Analytical solutions for the coupling admittances between the electrodes can be found
for various limits of (al >..) and (hi>"). For simplicity, only pure dielectrics will be used
for all the limits considered. Thus, only the normalized coupling capacitances per unit
length Oll/(E.l\tIt ) and Ol2/(El',[t) will be of interest. The four limiting cases are shown
in Fig. 3.13.

Case 1 - In the limit of (ai>") ~ 1 and (hi>") ~ 1 an analytical approximation

to the electric field distribution exists. The problem is drawn in Fig. 3.13a for one
wavelength. The electric field in polar coordinates has the simple solution,

(3.6)

Integrating the surface charge 011 the grounded electrode, obtained using (2.4), and
dividing by the voltage Vo yields the capacitance per unit electrode length

(3.7)

Performing the integration yields

Cl !YIel = - Ea : Eb In (~ - 1) (3.8)

With a long spatial wavelength, the coupling between the electrodes is concentrated
in the adjacent halves of each electrode (Iyl < (>"14». For a periodic structure, the
total capacitance per electrode unit length is just lllultiplied by 2. Thus, the effective
0l2/A1el for this analytical limit is

C12 I AIel = - 2 fa + fb In ( ~,\ -

1r 1
)
(3.9)

and can be compared to the value obtained from t.he continuum model. A tabulation
of coupling capacitances for various values of a, k and N (number of collocation points
and summation terms), with (hi)..) '" 10 7 and fa = Eb, is given in Table 3.1. It is seen
that the two models are in agreement to better than 4% for (ai>") = .01 and .001. 1;:he

91
Figure 3.13: a) Case 1 - Small Interelectl'Ode Spacing Limit, b) Case 2 - Parallel Plate
Capacitor Limit, c) Case 3 - Microstrip Solution, and cl) Case 4 - Long-Wavelength
Limit.

92
 Table 3.1: Case 1 - Closely Spaced Electrodes, No Ground Plane.

Model Parameters Numerical Analytical Error

a/A k N C 12 / fbA-fel C 12 / fbJ\;lel %
10- 1 25 100 3.11 11.1
10- 1 25 1000 3.23 2.80 -15.4
10- 1 45 1000 3.20 -14.3
10- 2 25 100 5.74 1.7
10- 2 25 1000 6.05 5.85 -3.4
10- 2 45 1000 6.05 -3.4

10- 3 25 100 7.43 15.5

10- 3 25 1000 8.66 8.79 1.5
10- 3 45 1000 8.64 1.7
10- 4 25 100 7.51 36.0
10- 4 25 1000 10.39 11.73 11.42
10- 4 45 1000 10.20 13.0

upper bound on a results from the requiremPllt that (3.6) be valid. The lower bound
on a limits the intensity of the fringing field at the electrode edges, thus validating the
analytic approximation and allowing a reasonable number of collocation points to ap-
proximate the field solution. For (a/A) =.1 the analytical solution underestimates the
electrode coupling. For (a/ A) = .0001, although there is convergence of C 12 / €bJ\;lel with
N, 45 collocation points are not enough to effectively map the potential distribution.
This is due to the intense field concentration at the electrode edges.

Case 2 - In the limit of (a/ A) «: 1 and (h/ ,\) «: 1 t.he problem approaches t.hat. of a
parallel plate capacitor (Fig. 3.13b). Within t.he insulating layer t.he fields are uniform
except for the intense fringing at the electrode gaps. The field is almost tot.ally excluded
from the region above the electrodes. Therefore, the coupling capacit.ance betwc-en t.he
electrode and grouml vl&.ue should be approximated by

93
 Table 3.2: Case 2 - Parallel Plate Capacitor.

Model Parameters Numerical Analytical Error

a/A h/A Cll/fbl\Id Cll / fbl'vld %
.01 10- 2 49.48 49.00 -.98
.01 10- 1 4.97 4.90 -1.4
.1 10- 2 -41.57 40.00 -3.9
.1 10- 1 4.68 4.00 -17.0

(3.10)

A comparison between analytical and numerical results gives very good agreement
(Table 3.2). For this case, k=25 and N=100, the standard number of collocation
points and summation terms.

Case 3 - This case utilizes a solution to the fields problem as derived by Ellgan [761.
The particular application was for surface acoustic wave devices using interdigital
electrodes. In this case, (h/A) ~ 1 and {a/A} is arbitrary (Fig. 3.13c). For a uniform,

isotropic medium above the electrodes, the solution to the mixed boundary value
problem is a Legendre function of the first kind. The resulting coupling capacitance is

(3.11 )

with

~ = cos( 7ra/ A) (3.12)

and 1((0 is a complete elliptic integral of t.he first. kind. Values of /\(0 are tabulat.ed
as a function of sin- 1 ~ which reduce, in this case, t.o

SiIl-1~ 7r{.5+a/A) (3.13)

sin- 1 Jl - e 2 7ra/A ( 3.14)

94
 Table 3.3: Case 3 - Microstrip Calculation.

Analytical Solution I
a/>. sin- 1 e sin- 1 V1 - e2 K(e) 1(( V 1 - I
e2 ) C12 /fbNlet J

.25 45° 45° 1.8541 1.8541 2.000

.1 72° 18° 2.5998 1.6105 3.229

Numerical Solution Error

a/>. N I C 12 / fb/Met %
.25 100 1.93 3.5
.25 1000 1.995 .25
.1 100 3.11 3.7
.1 1000 3.226 .09

Again, with (h./ >.) '" 1 and fa = fb, very good agreement is obtained (Table 3.3).

Case 4 - This case is the long wave limit treated by Mouayad [73, Section 3.4] and
shown to have an analytical solution corresponding to the transmission line model of
Garverick [71]. Here (a/h) ~ 1, (h./>.) «: 1, fb = fo:z: and If:/f:ol «: 1 (Fig. 3.13d).
The first condition allows the approximate normal electric field distribution along the
interface

(3.15)

The last condition implies the normal current into the boundary between the electrodes
is negligible compared to the surface currents (see (2.8)), yielding a total differential
equation for the potential at the iut.erface

(3.16)

\Vith the boundary conditions of (2.3) the solution becomes

(3.17)

95
with

1
(3.18)
~ _ jl1',qh
f oz Wfoz

In support of the application of the device as a humidity sensor, a distributed

parameter model was developed and implemented by Garverick [71J. In this situa-
tion, a thin film «1 pm) was coated onto the electrode structure of the OFT and
measurements were taken in a temperature and RH controlled environment. Garv-
erick employed a quasi-one dimensional transmission line model. Instead of deter-
mining an electric field distribution first and then identifying circuit parameters, each
electrode-to-electrode interaction (including the grounded substrate) was immediately
represented by a lumped or distributed parameter (Fig. 3.14). In this case

1
a2 = -;::::-----=,.-- (3.19)
~-j~
VT WVT

It is interesting that the quasi-one dimensional lumped parameter approach to an

inherently two dimensional problem works so well in predicting the measured response.
In representing the driven-to-floating electrode impedance through the air as a simple
lumped capacitance of constant value, the actual complex field distribution is being
disregarded.
A point not well understood is the requirement of a sheet susceptance, CA. In com-
paring the expressions for a 2 , this susceptance corresponds to a surface permittivity,
f,o. Using the chip parameters given by Garvel'ick [711 (A = 50 I'm, h = 1 I'm, load
capacitance of 4.3 pF, f oz = 3.9fo F fm, and total illterelcctrode meander length of
10.15 mm) it is seen (Fig. 3.15) t.hat. a s11I'fnce pPt'l11iHivit.y is rp(l1ait'prl hy the cnntin-
uum model to prevent the phase from going positive (alt.hough t.he st.andard responses
reported by Gaverick did not have positive phase, it has been observed [IIJ). The
proposed physical explanat.ion for this surface permittivity is the format.ion of a space
charge layer throughout the thin film caused by the blocking of current flow at the
aluminum electrodes. Typical sheet susceptances, when divided by the t.hickness of

96
 cx

" --
_ _....._ • • • • • • . . • . . 1. 1-..,.- ~. . ... . .. .
I \
\ J
---- ...... . . . . . \. --- .....- . . " . . . . . .
I
,.

.... --'

'--~

Transmission Line

1
C dx
a

cl~-L)
h~X+dx)
Vex) G dx
a r C /dx V(x+dx)
t

< dx >

Figure 3.14: Dist,ributecl 'I'ransmissioll Line Approximat.ion to Thin Film Coat.ing

Microchip ill Air (from [71]).

97
the thin film, yield enormous bulk permittivities that are not physically justified.

A comparison between the continuum model and the transmission line model
yield~d similar frequency responses, when modeling the old version of the microchip,
but values for the surface conductivity and permittivity that differed by ",,30% (see
Fig. 3.16). The transmission line response was generated using the chip parameters
given in [71] producing a high frequency gain in air of -49 dB, a surface conductivity
(l/Rs) of 1.56x10- 14 S and a surface permittivity (Cs) of 1.6:1:10- 16 F (corresponding

to CT/C A = 30). However, with a normalized load capacitance of 24.56, the contin-
uum model response was generated with a surface conductivity of 2.02x1O- 14 S and a
surface permit.tivity of 1.35:z:1O- 16 F. The difference in parameters is significant as the
response is very sensitive to the values of the surface parameters.

Note that the electrode structure parameters play an important role here. The chips
used in these thin film studies [71,721 have a thinner oxide layer than t.hose currently
manufactured. Taken with the fact that the measurements were made in air (a low
dielectric) the RC transmission line capacitive surface coupling provided by the thin
film was essential for accurate interpretation of the data. If this situat.ion is altered by
having a thicker oxide layer, thicker moisture sensitive coating, or taking measurements
in a higher permittivity medium, the resulting increase in the illterelectrode capacitive

• •
coupling will greatly reduce the need for a surface permittivity. The effect of the
thicker oxide layer of the present chip is shown in Fig. 3.17. No positive phase is
observed here.
In summary, it is seen that the continuum model is applicable over a wide range
of parameter values, encompassing many specialized models. A good contrast in im-
plementation is provided by the finite difference approach. The simplicity of the con-
tinuum approach lies in the ease with which various mult.ilayered st.ructures ca.n be
represented by altering a subrout.ine t.hat. comput.es t.he surface capacit.ance densit.y.
Changing the finite difference model requires, at a minimum, restructuring t.h(" mesh
pattern, not as easily automated. On the other hand, variations in t.he t.angent.ial
(y) direction are easier to model using the finite difference rout.ine, i.e. incorporat.ing
the thickness of the electrodes. The tradeoffs should be examined for the particular

98
 o~----~----------------------------------------~

-11

-22 E.., = 10- 15 F

....
-33

-44 El = Eo F/m
0'1 =0
0'.0 = 10- 15 S

-55~ ____ ______ ____ ______L -_ _ _ _

~ ~ ~ ~ _ _ _ _ _ _L -_ _ _ _ ~

-3 -2 -1 o 1 2 3 4

Or-----------------~~~~----------------~

-·60

-120

-240

-300

-360~----~------~----~------~----~------L-----~
-3 -2 -1 o 1 2 3 4
log Freq

Figure 3.15: Predicted Responses using Old Microchip (It = 1 ~tll1) having a Surface
Couductivity (0'.0 = 10- 15 S) and Variable Surface Permittivity (€"o).

99
 trans3.3c
°r-~"~"~~------------------------I

-10 \
¥
~
-20 ¥
:t:
.....= ~
IG
tD
-30
:.«
:.:
:a::
-40 :t

:t:
-50
-3 -2 -1 0 1 2 3 4

-30
¥\ ~
/
~

I)
-60
\ :.« ~

(,4)
-90
:.:
IG
..c= ~ ~
~

-120 \\;1
-150

-1BO~----~----~----~------~----~----~----~
-3 -2 -1 o 1 2 3 4
log Ereq

Figure 3.16: Comparison of Predicted Responses using Transmission Line Model (.)
and Continuum Model (X) for d. Thin Film Coating Old Microchip.

100
 Or-----__~--------------------------------__.

°r---------------~~~~~------------I

-30

-60
ID
U)
~ -90
.s::
~

-120

-150

-180~----~----~----~~----~----~----~----~
-3 -2 -1 o 1 2 3 4
log Freq

Figure 3.17: Predicted Responses using Present Microchip having a Surface Conduc"
tivity (u,o = 10- 15 S) and Variable Surface Permittivity (€,o).

101
application desired.

102
Chapter 4

Parallleter Est illlat iOIll Techniques

In the previous chapter, the problem was posed as one of determining a frequency
response given all the pertinent complex permi~tivities, layer thicknesses and other
relevant parameters describing the media. In practice, the situation is usually re-
versed. The unknown quantity is a material property such as complex permittivity
(bulk or surface) distribution if inhomogeneous) or layer thickness. What is known
is one or more experimentally measured responses, perhaps at different temporal or
spatial frequencies. A parameter estimation scheme is required in which all a priori
knowledge of the physical situation is utilized, i.e. number of layers, values of complex
permittivities for each layer, layer thicknesses, and, based upon one or more measure-
ments with respect to changes in a known parameter, the quantities of interest are
deduced using a root searching routine. Noise or other st.ochastic processes affect-
ing the experimental data call for more sophisticated techniques, not considered here.
Thus, here the estimation routines may be viewed either as root searching or function
minhrJzation routines [78, Section 7.2].
General algorithms for single and multi- parameter est,imation, as t.ypically seen
in optimization and numerical analysis references [78, Section 10.21[79, Chapt.ers 4
and 5], are presented in Section 4.1. The mot.ivat.ion for using CL secant. met.hod for
single paranieter estimation, and a derivation of the Newton search rout.ine using a
quadratic norm for multiparameter estimation, is given. A discussion of single param-
eter estimation routines for estimating layer thickness and complex bulk and surface

103
permittivities is presented in Section 4.2. The ability to uniquely ascertain spatial
distributions in complex bulk permittivity using variable spat.ial wavelength dat.a is
described in Section 4.3. Several case studies are given, demonstrating the utility
of extending the single parameter estimation routine to this multiparamet.er estima-
tion problem. Finally, Section 4.4 describes a two parameter estimation routine that
requires the use of a the Newton's search method.
For application to microdielectrometry, acceptable tolerances for convergence are
determined by the experimental error in the response. The present version of the
device, when using a microchip sensor, has an accuracy of 0.1 dB in gain and .1° in
phase [67]. In the routines, the gain tolerance is tightened to 0.05 dB. This translates
to a tolerance of 0.5% in gain at -40 dB and 0.1 % in phase at -90°.
These routines were developed on an IBM PC/ AT with an 8 Mhz clock and an
80286/287 microprocessor/mat.h co-processor. The emphasis was on developing search
routines that converged and were somewhat robust, not on generating the optimal
search method. Most of the validation of these search algorithms was obtained using
experimental dat.a on semi-insulating materials. Consequently, there are areas of pa-
rameter space where the convergence and robustness of these routines have not been
examined.

4.1 General Algorithms

4.1.1 Single Parameter - Secant Method

The search is for the root to the equation

e(B) =0 ( 4.1 )

where e( 0) is a set of error functions (usually t.he difference between one or more
measured and predicted gains) and e is a list of parameters t.o be est.imat.ed (such
as complex permittivities). Of course: t.he particular physical phenomenon associated
with the parameters to be estimated must make a significant cont.ribution to the gain,

104
and hence, the error.
There are many different met.hods used to conduct searches for root.s. Some require
only evaluation of a function, others, such as Newton methods, also use derivative
information. In one dime:~sion, the Newton-Raphson update to ~, guess is

(4.2)

Along with the extra information provided by the derivative is an extra cost. If an
analytic derivative is not available then it must be approximated numerically, greatly
increasing the number of function evaluations. For a one dimensional search, if the
method without derivatives converges in less than twice the number of iterations of a
search using derivatives, then the advantage of the derivative is lost. This tradeoff be-
comes more important as the number of parameters increases and as the computation
time for a function evaluation increases.
Given the complex nature of the continuum model, there is no simple analytic
d~rivative with respect to parameters such as layer thickness or complex permittivit.y.
The tradeoH' described above applies here, especially given the time required for a
funct.ion evaluation. A secant method of searching was employed for most of the
parameter estimations because of this tradeoH' [79, Section 5.1.1]. In this method,
guesses, 0i, are updated by AOi using the secant formed by the two most recent guesses,

(4.3)

ill the neighborhood of the root, the ~ecant approaches t.he tangent., t.hus t.he secant.
method is also known as a quasi-Newt.on method. The danger I.hat t.he root. does
not necessarily remain bracketed by t.he two guesses is decreased by providing bounds
on the minimum and maximum values of the estimated parameter La~ed upon phys-
ical constraints. The difference between the secant and Newton-Raphson methods is
illustrated graphically in Fig. 4.1.

105
 Secant method. Extrapolation or interpolation lines (da.~hed) are drawn
through tht' two mObt recently ,!valuated points, whether or not they bracket the function.
The points are numbered in the order that they are used.

(a)

Newton's ml'thod extrapolates the local derivative to find the next I'1Itimatl'
of the root. In this example it works well and converge!! quadratically.

(h)

Figure 4.1: Comparison of a) Secant Method and b) Newton's Method for R.oot. Search-
ing (from [80, p. 249 and 255]).

106
4.1.2 M ultiparameter - Minimization of Quadratic Erll"Or Func-
tion

The general approach for multiparameter estimation using derivative information is

outlined below (adapted from [78, p. 2821. This formalism will be used for the multipa-
rameter estimation of layer thickness and complex bulk permittivity and complex bulk
permittivity and surface conductivity. For these estimation problems the objective is
to minimize a quadratic norm defined as

V(O) = ~eT(O)We(O) ( 4.4)

where V is a nonnegative scalar, 0 is a list of one or more parameters to be estimated,

e( 0) is a column vector of errors and W is a symmetric, positive semidefinite weighting
matrix. Although these quantities are usually treated as real functions of real variables,
there appears to be no reason why the concept cannot be extended to include complex
functions of complex variables, as is done in Sections 4.3.6 and 4.3.7. When 0 = eo,
the correct values, the quadratic norm will be at a minimum and its gradient, V'( eo)
will be zero. Expanding in a Taylor series in the neighborhood of eo yields

V'(e o ) = V'(O) + V"(O)LlO = 0 (4.5 )

or

Lle = -[V"(e)]-lV'(e) (4.6)

This form of the increment to update guesses is a common one for many search routines
utilizing the derivative of a function. For methods ot.her than the Newton sear('h,
alterations can made to V"( e) to improve ('onvergen('e and robust.ness. lYsing t.he
quadratic norm defined above (4.4)

V'( 0) - e,T (e)We( e)

( 4.7)
V"(e) e,T(e)We'(e)
where e'( e) is the jacobian of e( 0). The form of the jacobian is

107
 • .J

e'( 9) - • (4.8)

t 33:
where there are n experimental data points, m parameters to be estimated and n ~ m.
The increment to the old guess becomes

(4.9)

which, if e'(9) is square and both e'(9) and Ware nonsingular, reduces to

( 4.10)

It is easily verified that (4.10) reduces to (4.2) for one dimensional searches.

4.2 Single Parameter Estimation

For these estimations, the error is defined as the difference between the measured and
the predicted complex gain,

(4.11)

where 9 is the complex parameter to be estimated. Here, gm and gp are formed by

evaluating the complex logarithm of the complex voltage ratio,

Q == log G - log ~F +j log e L(VF/Vn) (4.12)

VD
where the real part is the gain ill decibels divided by 20 and the imaginary part. is t.he
phase in radians (closely related to the output of the microdielectrometer). If C;, as
defined in (2.1), is a complex analytic function of the complex variable 9, thp.n Qp and
hence e are also. Thus, for the class of constitut.ive laws used here, it is meaningful to

108
directly apply one-dimensional root finding techniques to the estimation of a complex
parameter.
Note that although one paramet.er is being estimated, this does not constrain the
medium to be uniform. Using the surface capacitance density, C:, heterogeneous media
can be represented as long as all but one of the parameters necessary to charact.erize
the medium's dielectric properties is known.

4.2.1 Layer Thic~ness

This is the case of a layer and a surrounding uniform medium, as modeled in Sec-
tion 3.1.2. At high frequencies, when the electrodes are purely capacitively coupled
through the material, the phase will be zero and the gain will be a real function of the
layer thickness and the permittivities of the layer and the surrounding medium. Given
any three of these four real quantities, it is possible to estimate the fourth. Here the
layer thickness (d 2 ) is unknown thus () = d2 and the search is for a purely real root.
For the case of air as the surrounding medium, the curves of gain versus layer
thickness for various layer permittivities are shown in Fig. 4.2. Finding the root of these
functions is relatively straightforward. For robustness, t.he routine first checks whether
the experimental data makes sense by calculating the gain of a uniform medium having
t.he lesser of the two permittivities. This gain should be less t.han the experimental
gain if a layer thickness is to be estimated. After passing this check the routine uses
the secant method to find the root. The search is conducted using the finest grid of
25 points (k) and 100 summation terms {n}. Due to the well behaved nature of the
curves, the thickness is uniquely estimated with two place accuracy in the associated
gain within 3 to 5 iterations. For a 50 micrometer wavelength electrode struct.ure this
implies a hundredth of a micrometer accuracy in t.he estimate for thicknesses in ~.he

range of 1 to roughly 15 micrometers. The lower bound may be set by feat.ures of

the electrode that are not modeled such as electrode thickness or by the experiment.al
error in the measurement. The upper bound reflects t.he decrease in sensitivit.y due t.o
the exponential decay of the electrostatic fields.

The software for this thickness estimation l"Outine, thest.for, may be found in Ap-

109
 - 10

-
CD
~ -20
I x 10- 10

.-cc:
<.!) - 30 3x I -11
2 x 10- 11

-40~~________~__________~____~
EO

o 4 8 12 16 20
Thickness d 2 fLm

Figure 4.2: Predicted High Frequency (10 kHz) Gain for Microchip with Variable
Thickness Layer (d 2 ), Vary Layer Permitt.ivit.y (1:2), in Air.

110
pendix F. This routine was used to interpret experimental results regarding deposition
rate for the rlasma deposition process and evaporation of t.hin oil films (Sect.ion 5.2.1),
and sedimentation of monodisperse particl~s [23, Chapter 6].

4.2.2 Complex Permittivity (Bulk or Surface)

As long as gp( 9) is an analytic function of a complex bulk or surface permittivity,

the search is only slightly more complicated. A complex bulk or surface permitt.ivity
is estimated using a complex data point (gain and phase) and specifying all other
relevant parameters such as the complex bulk permittivities of the other regions, the
complex sutface permittivities at the interfaces and the layer thicknesses. Now, in
(4.3),8= f' - jf" or f~ - jf~ and e(9), in (4.11), is complex.
The economy of using the secant, instead of the Newton method, is demonstrated
here. The Newton method would require a fUllction evaluation in the neighborhood
of the most recent guess so as to compute a local difference approximation t.o the first
derivative. When guesses are far from the root this second function evaluation will
usually have no chance of producing the root. However, with the secant method, each
evaluation has a reasonable chance of producing the root.
The search routine can be started either with a user input guess or wit.h its own
"guess" as to the complex permittivity. For bulk properties this "guess" corresponds
to that of an insulating, nonpolar medium. The first update only is calculated using
the local derivative. For the secant method the order of convergence is the 'golden
ratio' of 1.618 ... [80, p. 248]. The key here is to get in the neighborhood of the root
where the convergence is rapid. A substantial reduction in the time to convergence
is obtained by letting the routine start with its own "guess" and using a technique
of varying the number of collocation points used for the function evaluation. This
technique performs sequential searches beginning with a coarse discret.i7.at.ioll of 2 grid
points and ending with 25 grid points. After a root is found at 2 grid point.s, that root.
is used to begin a new search using 10 grid points. Again, once convergence has been
reached, this new root is used to begin a search at 25 grid points. Generally anywhere
from 20-40 iterations are required for convergence with 2 grid points resulting in only

111
1-10 iterations at 2.5 grid points. Without this technique, convergence can easily take
2.0 or more iterat.ions when using only 2.5 grid point.s. It takes roughly 12. times longer
to compute a gain-phase response using 2.5 collocation points versus 2. collocation
points. Thus, a substantial savings in computing time ca!l be realized.
A two-dimensional view of the space over which the routine searches when es-
timating the complex bulk permittivity of a layer 5 micrometer thick surrounded
by a medium with a bulk permitt.ivitY of E~ = 2.x10- 11 F /m and a loss factor of
E~ = lx10- 8 F /m is shown in Fig. 4.3. This space is constructed by computing the
gain-phase response for a matrix of complex bulk permittivities of the layer at a
constant frequer.cy. The contours are lines of constant E~ and E~ and are orthogonal
because 9 is an analytic function of E·. Reiterating, this orthogonality justifies the
computation of the complex slope used by the one dimensional secant search routine.
As required by (4.12.), the gain is in decibels divided by 20 and the phase in radians
multiplied by log e. This figure is analogous to the look up table given by Senturia, et
al. [69] for the case of a uniform medium. With the gain defined by (4.12.) the contours
are orthogonal for the uniform medi um case too (Fig. 4.4).

In Fig. 4.3, the values of E~ that make the layer more lossy than the upper medium
are concentrated near the origin. The dot in Fig. 4.3 is where the layer and surrounding
medium have equal complex permittivities. By contrast, in Fig. 4.5 the upper medium
is sufficiently insulating that the point of equal ~omplex permittivities (the dot) is near
the zero phase axis. Again, with a 5 micrometer thick layer, the surrounding medlUll1
now has a bulk permittivity of E~ = 2.xl0- 11 F /m and a loss factor of t:~ = 1x10- 12 F /m.
The kink that appears in the left side of the plot is a result of charge diffusion along
the layer, as discussed earlier in regards to the thin film I, '''lidit.y sensor. This kink
is more evident when the layer is in a perfectly insulating medium such as air and/or
the layer is thinner, say 2./tlll (Fig. 4.6).

Given the well behaved nature of t.he spaces represented by Figs. 4.3 and 4.5, it. is
expected that estimation of unique complex permit.tivities is a st.raight.fol·ward process.
The sensitivity of the estimation routine can be observed from the t.wo figures. Poor
estimates occur in regions of high contour density. For example, gain-phase data

112.
 (saaJoap ) aso4d
o ()
000
C\I <D CD
o C\.I

~~~'----~---T----r'--~r---~I ()

- ()

.......
£D
()"'C
C\I-
c:

! ()
r0

Figure 4.3: Parameter Space for Complex Bulk Permittivity Estimation of a 5 pm

Layer in a Uniform Medium (fi = 2x10- 11 F Im and f~ = 10- 12 F 1111) with t~ and f~
in units of fo F Im.

113
 phase (degl'ees)

~
~ fBI ~ ~
o I
~=-~----'-----r----'C
I I

c
~
I

iI

L - - -_ _- - - - ' ~
I

Figure 4.4: Parameter Space for Complex Bulk Permittivity Estimation of a Uniform
Medium with fi and f~ ill units of fo F Im.

114
 (saaJoap) aSD4d
o oq- o o oo U)
o
C\.I
C\.I CX>
I

o
-I

-
al

-- 0
C\.I
I
-
"0

.-c0:
(!)

o
/'0
I

:::N
\AI

~ __ __ ____ __ ____ __
~ ~ ~ ~ ~ ~o
q-
I

Figure 4.5: Parameter Space for Complex Bulk Permittivit.y Estimation of a 5 pm

Layer in a Uniform Medium (fi = 2xlO- ll F /m and f~ = 10- 8 F /m) with f~ and f~ in
units of Eo F /111.

115
 phase (degrees)

~I ~ ~I
o I

o
..-f
I

~
I

Figure 4.6: Parameter Space for Complex Bulk Permittivity Estimation of a 2 /1.111
Layer in Air (f~ = fa F /m and €~ = 0) with €~ and €~ in units of fa F /m.

116
adjacent to the gain axis yield poor loss factor estimates while data with gains near zero
yield poor permittivity estimates. The latt.f"r is a consequence of t.hf' high impp.dnnce
mode of operation which utilizes the measurement of a floating voltage.
For the special case of ohmic media, f" = «(J' / IN). Thus, contours of constant f.'
are also the gain-phase trajectories obtained if the temporal excitation frequency was
varied. This may also be seen by eliminating the independent parameter of frequency
from the plots in Fig. 3.2.
For the gain-phase spaces of Figs. 4.3-4.6, valid data must lie in the region having a
lower bound given by the f.i = fo curve. Upper and lower bounds can be placed on the
parameter estimates to prevent the search from wandering too far oft'. These bounds
are physically motivated - no bulk permittivities less than f.o and no conductivities
less than zero. Upper bounds are chosen based on reasonable guesses of the order
of magnitude of the dielectric properties of the materials to be measured. Another
common technique for preventing endless searches is to limit the maximum number of
iterations a routine can search. Based on experience in using the routines documented
here an upper limit of 50 iterations for a given grid point estimation is established.
In addition to estimating bulk parameters, this routine works just as well for esti-
mating complex surface parameters.
The software for this complex permittivity estimation routine, parest.for, may be
found ill Appendix F. This routine was used to estimate the complex bulk permittivities
of oil (Section 5.4 and [23, Chapter 4]), the complex surface permittivities of adsorbed
layers (Section 5.4.1 and 5.5) and the dispersive dielectric properties of the plasma
deposited bromobenzene (Section 8.2.3).

117
4.3 Multiparameter Estimation Using Variable Spa-
tial Wavelengths

4.3.1 Motivation

Using the temporal frequency response, it is 110t possible to uniquely determine the
spatial distribution in dielectric properties of a heterogeneous medium. However, this
spatial distribution is u~quely observable through extraction of information using
variable spatial wavelength electrodes [82].
The first claim is substantiated with the following example. Let a sample having a
spatial distribution in complex permittivity, i(:z:) be placed between a pair of parallel
plate electrodes as in Fig. 4.7. An electric field is imposed on the sample via a voltage
source and the charge on the electrodes is sampled with an electrometer. With the
charge on the electrodes defined as q such that the current i = (dq/dt), the response
to the applied voltage having the frequency w is defined as

C(w) _ ~ ( 4.13)
v
In the bulk of the material the electroquasistatic fields are described by Gauss' Law

( 4.14)

with the requirement that the electric field be irrotational and therefore related to the
complex potential v by

E- -V~ (4.15)

The purpose of the guard electrode is to elilllinat.e any fringing fields, producing
an electric field having only a normal (x) component and uniform within t.he bulk.
However, due to the spatial variation in complex permittivity with the normal COlll-

ponent, the electric fields will be 110nuniform in z. The effect of this nonuniform.it.y ill

118
 'I(

guard
ring
area A /
f
yew) h

..~~~~~~~~~~!

Figure 4.7: Complex Permittivity Measurement of Heterogeneous Medium Using Par-

allel Plates with Guard Ring.

the measured temporal frequency response can be seen by writing (4.14) and (4.15)
as, respectively

o ( 4.16)

dei»
(4.17)
dx
Integration of (4.16) yields

(4.18)

where u.u is the magnitude of t11e free surfflce charge density nn f'it IH"r elt"ctrode.
Integration of (4.17) from the lower electrode, x = 0, where the potent.ial is zero,
to the upper electrode, x = h, where the potential is the applif'd voltage i" and
substituting for Ez using (4.18) yields

i1 _ u rh <lx
.u 10 f(X) ( 4.19)

119
Using q = AiT,u, where A is the surface area of the cent er electrode, and substituting
(4.19) into (4.13) yields

C(w) A rh dx (4.20)
= 10 i(x)
Measurement of C(w) at a single temporal frequency therefore gives, at best, a spatial
average of the complex permittivity distribution. Thus, there can be no unique infer-
ence of the spatial distribution from the temporal frequency response alone. Using an
interdigital electrode structure, one that utilizes the fringing fields between electrodes,
th~re are certain conditions under which a definitive spatial distribution can be esti-
mated. This has been discussed in Section 3.2.2. However, in general, these conditions
will not be satisfied and an alternative approach 111ust be taken.
It is helpful to re-evaluate C(w) when probing with the fringing field of an interdig-
ital electrode structure. Now there will be fields above and below the electrodes. The
Ab
fields below are readily calculable because C is known. Thus, we are concerned with
the calculation of Ca, defined in (2.9). Using the field representat.ion of Chapter 1, in
a more general form,

jwt
4>(x,y,t) = Re4>(x,y)e
A

(4.21 )

where
00

~(x, y) = L ~n( x )e-jle"Y j k n = 2n1r / >..

n=-oo

the electric field has both an x and y dependence, thus (4.16) is repla.ced by

8(E:E z ) AaEy
ax + f -ay- - =0 (4.22)

From (4.15) we get

a~ 00

EY -
ay J L
n=-oo
kn4tn(x)e-jknY (4.23 )

Substituting this result into (4.22) and integrating from x = 0, where iE z = faE:, to
x = 00, where fEz = 0, yields

120
 ( 4.24)

and using the definition of the surface capacitance density, (2.9)

(4.25 )

In general, the potential ~ n (:c) will decay exponentially with :c, and the higher spatial
harmonics (larger n), decaying even more rapidly. The spatial distribution in complex
.
permittivity f(:C) can be viewed as a weighting function for the potential distribution.
Thus, unlike (4.20), the effect of the specific distribution in complex permittivit.y will,
in general be unique.
There are several advantages to probing a medium using the variable spatial wave-
length approach. The first is the ability to estimate a unique spatial distribution from
the measured responses. As discussed in Section 1.2.1, the electroquasistatic fields gen-
erated by interdigital electrode structures decay exponentially in the direct.ion normal
to the electrode plane. Furthermore, the characteristic decay length is proportional to
the spatial wavelength, roughly (A/3). The response to a short wavelength will reflect
the dielectric properties of the material in the immediate vicinity of the electrodes.
The longer wavelength electrodes will sample both the material adjacent to the elec-
trodes and material further out. As the wavelength is further increased, the dielectric
properties at an increasing range from the sensing surface will influence the measure-
ment. Unlike the temporal frequency response, the measurements here uniquely reflect
the spatial distribution of dielectric properties.
Another advantage is the potential for simplification of the estimation problem.
One ("ould take t.he n measurement.s at. n dirrerent. 8pRtial wRv("J,'ngt hR nn,] f'Rtill1fdf:' m
parameters of a function describing the spatial distribution all at once (11 2 m). This
would entail computing and inverting an nxm. jacobiall (see (4.8)) for updat.ing each
guess. This estimation process is likely to involve a large amount of computer t.ime
for its implementation. However, by exploiting the exponentially decaying nature of
the applied fields, the estimation procedure can be reduced t.o a sequence of single

121
parameter (or two parameter) estimation. This will be shown in Section 4.3.3.
The third advantage of this approach lies in the fact t.hat. variable spatial wave-
length measurements can be taken at the same temporal frequency. Thus the spatial
distribution can be deduced at each frequency without any restrictions or assumptions
about the frequency dependellce of the dielectric properties. This additional freedom
allows an unconstrained evaluation of the possible physical mechanisms underlying the
dielectric properties (such as molecular or ionic dynamics).

4.3.2 Apparatus

One stra;ghtforward implementatioll of this approach is to fabricate several interdigital

electrode structures, each having a different spatial wavelength. If the response must
be measured with all the electrodes contacting the material simultaneously, then there
is a requirement that the spatial distribution within the material be the same at each
separate location probed by the individual electrode structures. This constraint. can
be relaxed somewhat if the measurements can be made consecutively, thereby allowing
each array to make contact with the material at roughly the same location.
It would be more convenient if just one set of electrodes could be used and the
spatial wavelellgth of this array could be varied through changes in the electrical ter-
minations of the electrodes. Such an array is shown in Fig. 4.8. The electrodes may be
connected to either a driven voltage input or a floating gate output or may be totally
disconnected. The shortest spatial wavelength can be imposed by connectillg every
other electrode together, one set drivell and one floating. The next longer wavelength
may be obtained by connecting pairs of electrodes together, again half the pairs driven
and half floating. Beyond this, some of the electrodes between driven and floating elec-
trodes must remain disconnected. They will assume a potential somewhere bet.ween
the driven and floating voltage, as dictated by the constraint that they carry no net.
current. These electrodes are used to prevent the higher spatial harlllonics (arising
from the coupling between the edges of adjbCent electrodes) fl:Olll dominating t.he re-
sponse. An array of switched electrodes base~ on this concept has been constructed
for the purpose of measuring the thickness of a layer composed of ice and seawater

122
[81]. In this case, the measurement is of the impedance between the driven and float-
ing electrodes, obtained by connecting the floating electrode to a viri.ttll.l ground and
'measuring the current flow.

4.3.3 Estimation Algorithm

Typical distributions of complex permittivity are shown in Fig. 4.9 as a function of
the normal coordinate (:z:) to the surface of the electrode array. In Fig. 4.9a the
distribution is smooth. Those shown in Figs. 4.9b and c showcase the situation where
there is a discrete discontinuity in the medium. The layer of thickness d is typical of
those where there is a step discontinuity at the interface between layers. In Fig. 4.9c
there is the addition of a complex surface permittivity associated with the interface.
An ex~ple of tItis is a layer of material having a thickness d with adsorbed material
on its interface giving rise to conduction in the plane of the interface (the dimensions
of which are smaller than can be resolved by the measurements).

Representations of the complex permittivity distribution that can be used are

shown in Fig. 4.10. These are general enough to represent any of the cases shown
in Fig. 4.9. In Fig. 4.10a the distribution is represented by layers, each having a uni-
form complex permittivity. In Fig. 4.10b each layer has an exponential distribution.
Included in each representation is the possibility of having complex surface permittiv-
ities at one or more of the interfaces. These are the impulse functions in Fig. 4.10.
Without these complex surface permittivities, the representations of Figs. 4.10a and
b approximate the distribution by stair-step and piece-wise continuous functions, re-
spectively. Other functions can be used to represent the spatial distribution of proper-
ties, resulting in est.imations of different. paramet.ers relevant t.o f.he individuAl function
chosen.
The following example demoIlstrates the simplification in the estimat.ion problem
available with this approach. Suppose that the spatial distribution is smooth and the
representation chosen is a stair-step one. Using the response at the shortest wave-
length and a uniform medium model, an estimate is made of the complex permittivity

123
 u..

u..
~t& 4---_
_---+-O X +-+--~~-I-- ..... -0

u..
<>
Figure 4.8: Apparatus to Provide Variable Spatial Wavelength in Potential Using a
Set of Switched Electrodes.

124
x x x

x=d ------. x=d

e(x) e(x) E(X)

(a) (b) (c)

Figure 4.9: Typical Spatial Distributions in Complex Bulk Permittivity: a) Smoothly

Varying, b) with Step Discontinuity, and c) with Complex Surface Permittivit.y.

:
:
.. :
:
..
···r····L
• •

....•......
: :.
..•• +• .••.••
•••

••
•
•••
•

. ...•••••.. +• ..

·•••• '. •
••
••
••
•
•
•• •
••

:
•
•
•• •
:•
••
. ·•: :..
:.
•• " .~

····~······~·--·······L ····~······~·-~·~:····L
···· '. ..
• • • ••
••• ••• ""
•• •• ""
•••• A •••••• .:. •••••••••• • ••• ~ •••••• .!, •• ~ •••••• .!"
e(x) EJ - 1 EJ EN•. , eN E(x)
(b)
Figure 4.10: Representations that can be Used in Variable Spatial Wavelength Est.i·
mation Routine: a) Stair-Step and b) Exponential.

125
of the material adjacent to the electrodes. This is represented in the upper left portion
of Fig. 4.11. Then, using the response at the next longer wavelength and a two region
model (layer plus uniform medium), an estimate is made of the complex permittivity
of the material a little further out (middle left in Fig. 4.11). Typically, a measurement
at spatial wavelength Ai will reflect the Ulaterial properties out to about (Ai/4), de-
terminiug the thickness of the previously estimated layer. Following this procedure,
successive estimations are performed using previously estimated parameters and layer
thicknesses determined by the spatial wavelengths associat~d with previous data. Once
the first pass has been made through the data, working from short to long wavelengths
(lower left in Fig. 4.11), the parameters can be refined by passing through the data
again. In these subsequent iterations, one method of refinement would be to set the
parameters of layers other than the one being estimated to their previously estimated
values (upper right in Fig. 4.11). Other refinement schemes, involving smoothing the
initial parallleter estimates, may be more appropriate for certain cases, such as when
the distribution is known to be smoothly varying.

At each spatial wavelength, the sensing depth into the material is a fraction of
the wavelength, etAi. With the stair-step or exponential representations of Fig. 4.10,
the thickness of each layer di is determined by two spatial wavelengths, Ai and Ai-I'
Proceeding from the shortest (i == 1) to the longest spatial wavelength (i == N), the
accompanying layer thickness is

00 for i = 1
di == et( Ai - Ai-I) for·i == 2,3, ... ,N - 1 (4.26)
for i == N
For the cases considered here et == .25, allowing some overlap bet.wepn spa1inl wave-
lengths. In general, the oxide lal er thickness and permit.tivity will affect., to some
degree, the penetration depth of the fields above the electrodes. However, when using
fOil! == 3.9fo and varying AI h from 2 to 100, this effect is negligible.
If et was much larger, the spatial wavelengths would be completely uncoupled.
Parameter estimates due to a particular wavelength would have no effect on the gain-

126
 ~~
~•
• :§~fffkTH
GAIN AT SHORTEST
A:":flh-RNITT/YITr
T///§F,Rffs1 NATERIAL lAYER

!
NEASUffsf:':/ISrtNATI~::/J:SER.
LONGER IIAVELENGTH

~riST TY fOR°Ef'ffcf,o NATERIAL

IIITTlY!lAYER.
LAYER

1Eflfi;f/fJ~1~f
I TES Of THE fIRST AND
WI!/f1/JRf!ttER
ESTlIIATlONS,!~t~G ESTlIIATED
THAN THEON~

, hIe SpatIa
F 'gure 4 . 11'. Schematic V'tew . of Varta ' 1 Waveleng_th Estimat.ton
. Algorithm
Using
1 Stair-Step Repre selltatlO11.
phase response as measured by any of the shorter spatial wavelengths. No refinement
of the estimates could be accomplished by subsequent passes. If this lengt.h was much
smaller, the coupling between the spatial wavelengths would be too intimate. Any
slight change in a parameter would ripple through much of the estimation, greatly
increasing the time to convergence. The value chosen, (c:t = .25), represents a compro-
mise given the above considerations and also is physically based, as discussed in the
intimate sensing section (Section 1.2.2).
The following three sections describe implementations of the estimation algorithm
for examples involving smoothly varying distributions in complex permittivity and step
discontinuities. Conduction in the materials is assumed to be ohmic, thus the complex
permittivity is specified in terms of the permittivity and conductivity. Experimental
data at a constant frequency is simulated using the contilluum model. For these sim-
ulations, a layer in which there is a smoothly varying distribution is approximated
by breaking it into 90 discrete sublayers having uniform properties. The gain-phase
response is computed for several different spatial wavelengths at the same temporal
frequellcy, here it is 0.1 Hz. A stair-step representation (Fig 4.10a) is used for all the
examples. The sequence of single parameter estimations is performed using the esti-
mation routine described in Section 4.2.2. This includes the technique of using variable
grid points to rapidly find the neighborhood of the root. Convergence is defined using
the same tolerances on gain (.05 dB) and phase (.1 0 ) as for previous estimations using
the microchip, unless otherwise denoted. The software for this estimation routine is
called betest.for and may be found in Appendix F.
For these case studies, an apparatus such as the one shown in Fig. 4.8 is envisioned.
Therefore the oxide layer thickness is the same for all spatial wavelengths and is chosen
to be 10 fLm, having a permittivity of 3.45:el0- 11 F /m. The wavelengt hs nsed, find cor-
responding normalized load capacitallces, are compiled in Table 4.1. The normalized
load capacitance is adjusted so as to produce the same dynamic range for all spat.ial
wavelengths when operated in air. This capacitance is a strong funct.ion of the oxide
layer thickness. Here, with the stated thickness of 10 /-1.111, the dynamic range chosen
was 0 to -60 dB. Although this may be an unreasonable dynamic range in practice, it

128
 Table 4.1: Spatial Wavelength and Normalized Load Capacitance.

h = 10 J.Lm, Eo;r; = 3.45:z:l0- 11 F fm, a = .25

,\ (Cd Eo;r;.Nle/)
(p.m)
20 768.78
50 370.87
100 165.84
200 90.50
300 72.34
400 63.11
500 56.73
600 51.48
700 47.39
800 43.45
900 39.86
1000 36.62

was felt to be important that each spatial wavelength operate wit.h the same dynamic
range. Any smaller range would have required negative load capacitances at the longer
spatial wavelengths.

4.3.4 Smooth Distribution of Complex Permittivity

Case 1 - The first case considered is tI. 111f'"f1inlTI WhORf" mmplf'x pf"rmitfh'ity "Hies
exponentially away from the electrodes. The permittivit.y decreases exponentially from
3xl0- 11 F /m at the electrodes to 2x10- 11 F /m at a distance of 50 fl.1U, at. which point
the permittivity remains constant. However, the conductivity increases exponent.ially
from 10- 12 S/m at the electrodes to 10- 10 S/111 at a distance of 100 J.Lm , also remaining
constant beyond this distance. This distribution is represented by the thick line in

129
 Table 4.2: Simulated Data for Case I.

A gam phase
(I'm) (dB) (deg.)

20 -55.18 -2.13
100 -49.24 -10.20
200 -45.68 -25.20
400 -41.10 -39.04
800 -37.13 -51.96
1000 -36.18 -56.4 7

Fig. 4.12. The simulated data, tabulated in Table 4.2, is generated using six different
spatial wavelengths using the oxide layer thickness and normalized load capacitances
mentioned above and a temporal frequency of 0.1 Hz. Note that the spatial wavelengt.hs
vary from 20 to 1000 I'm, effectively probing out to a distance of roughly 250 I'm.

Also shown in Fig. 4.12 are the resulting estimates of 10 "marches", or passes,
through the data. Using the stair-step representation, at each spatial wavelength
there is an associated layer thickness having its own estimated complex bulk permit-
tivity. The estimated distribution after the first pass through the data is denoted by
the triangles, and the final estimate, after the tenth pass, by the crosses. The inter-
vening iterations are denoted numerically, where possible, to indicate convergence or
divergence. Convergence to a reasonable stair-step approximation of the exponential
distribution is observed. The nomenclature established here regarding the thick line
and symbols is used for all the case studies in this section.

Case 2 - The second case considered also utilizes an exponellt.ial dist.ribut.ion in com-
plex permittivity over a distance of 50 l.t.lll. The permittivity varies t.he same as before
but the conductivity varies oppositely, decreasing expollellt.ially from 10- 10 S/m at. I.he
electrodes to 10- 12 S/m at a distance of 50 Itm and constant thereafter. The t.hick
line in Fig. 4.13 again represents the actual distribution and the simulat.ed data (at
0.1 Hz) for this case can also be fonnd in Table 4.3.

130
 viZ - .25. exp. dist.
3.5 _.
A

,-&
fat
3
~-.
~
A
A
&
_1
J

....
.... 2.5 r- \ ~2.
I
Cl
,
-
I

....W ~
-
_7
,- '" ....
s::
0
.....
....
2 - .

- ~
.a.
:a..

"
~
I)
1.5 - ," A

1 I
h' I
o 50 100 150 200 250

1000,--------------------------------------------,

r-,-
S
Cf.)

100
....
N
I

....
W

-Id
&
....
CD
10
"
CD
.....0

1~--------~--------~--------~------~~------~
o 100
50 150 200 250
distance from electrodes (Microns)

Figure 4.12: Estimated Spatial Distribution Using 6 Spatial Wavelengths (Case 1).
Simulated Data in Table 4.2. (6) - first estimation, (X) - final estimation.

131
 Using the same 6 spatial wavelengths as in the previous example, the estimation
pro(,E"clure ra.n int.o prohl('ms wit.h st.nhilit.y. During the first. pa.ss (h·in.nglp.f! in Fig. 4.13),
there was a problem estimating the conductivity for the fourth layer (extending 50-
100 I'm). The result, indicated as 1:c10- 14 Slm, is actually a default value used for
plotting the data, as the estimate actually wanted to go to negative conductivities and
was set to zero. After the final pass (crosses), this problem shows up on the ot.her side
of the discontinuity. Furthermore, it is observed that the permittivity estimates are
diverging from the actual values and from the estimates obtained on the first pass. This.
same result is obtained if the same spatial distribution is probed using all 12 spatial
wavelengths for finer resolution (Fig. 4.14). Here, the first pass produces a reasonable
permittivity profile (triangles) but agaill has a problem with the conductivity following
the discontinuity. After the tenth pass, the permittivity and conductivity estimates
are diverging, and the conductivity estimates at 800 and 900 J.Lm have been set to zero.

It appears that the iteratioll process is unstable. Even the result of Fig. 4.12
appears to begin diverging slightly at the end of the 10 passes, it's just that the first
pass produced such a poor estimated profile that the following iterations had to begin
converging first., before diverging. Altering the tolerance limits for convergence, or the
parameter determining the layer thickness (er), did not improve convergence. Moving
the discontinuity out to 500 J.Lm while retaining the exponential distribution, did not
eliminate the divergence problem. The simulated data for this estimation problem
is also tabulated in Table 4.3. The results llsing 6 and 12 spatial wavelengths are
shown in Figs. 4.15 and 4.16 respectively. In both cases, the best estimate is provided
by the first pass through (trian.gles) although the only significant di vergence using 6
wavelengths was for the last layer (> 200 J.Lm). Probing the medium with 4 or 8 spatial
wavelengths did not improve the situation.

Finally, the case of a uniforl11mediulll havillg no spatial dist.l'ihut.ioll was examined.

The permittivity of the medium was 2:c10- 11 F Im and the conductivity was 10- \0 S /m.
The simulated data, generated at both 0.1 and 1 Hz, is tabulated in Table 4.4. Again
the first pass is close to the actual distribution with subsequent iterations resulting ill a
diverging solutioll (Figs. 4.17 alld 4.18). This divergence occurs for the search at both

132
 Table 4.3: Simulated Data for Case 2.

d2 = 50 I"m d2 = 500 I'm

A gain phase gai0J;ase
(I'm) (dB) (deg.) (dB) (deg:L
-
20 -45.30 -71.25 -44.42 -73.03
50 -41.04 -75.54 -38.59 -79.17
-
100 -37.53 -80.66 -32.51 -85.49
200 -39.77 -87.56 -29.46 -94.46
--
300 -45.46 -81.38 -29.99 -100.93
400 -49.65 -57.92 -31.41 -105.07
500 -50.64 -36.99 -33.13 -107.15
600 -50.69 -27.00 -34.94 -107.38
700 -50.83 -22.23 -36.81 -105.98
800 -51.02 -19.12 -38.59 -103.16
900 -51.21 -17.05 -40.27 -99.10
1---
1000 -51.40 -15.41 -41.81 -94.04

133
 yl1 - .25. expo dist.

~-
-"
~
&
2.7
.

........
I
2.4 i-- ., ~

"

-=
....
~

2.1 -
3-

...
.a

-.....,
0

."
a.. 1.8 I--
.&

i ,..
3
...

f
~

1.5 I I I t
o 50 100 150 zoo 250

100
~

-&-" 10 e-\
~-

~
Cl:)

....
-
N
I

\ -
-....
~

IU
&
1 i--

""
~

~
....

"
-

....tD
(IJ
0.1 i--
tD

-0

... I I
0.01 ~
".
o 50 100 150 200 250
distance i'rom electrodes (Microns)
Figure 4.13: Estimated Spatial Distribution Using 6 Spatial Wavelengths (Case 2).
Simulated Data in Table 4.3, d2 = 50 p.m. (6) - first estimation, (X) - final estimation.

134
 vIS - .25" expo dist." 12 lambda

..-.-. S
&

"
fa.
~
,
~
I
~

-
~

.-...
0

tn
~
ID
1.8 ,
1.5~------~--------~------~--------~------~
o 50 100 150 200 250

100

. ..-.-.

"
~

N
~
&

I
10
~~ . ~
.. a..
~

-
~
1 r- \ v

t..
~

.A

~
~

A 01
~
....
2:
....rn
t:n
"z.
-""'-
-
A
:L

0.1 f--
~
t:n

- 0
1 "
0.01 I a. 1 ~ ~

o 50 100 150 ""200," 250

di stance ~rom el ectrodes (Microns)

Figure 4.14: Estimated Spatial Distribution Using 12 Spatial Wavelengths (Case 2).
Simulated Data in Table 4.3, d2 = 50 I'm. (6) - first estimation, (X) - final estimatioll.

135
 vl11 - .25. exp. dist .• 6 laMbda

a.1~-----------------------------------------------,

-& 2.9
"-
s;..

......
... I

-
u:I
2.7

--
s:
0

OS
s:a..
GJ
2.5

~~

2.3
0 50 100 150 200 250

100~~~--------------------------------~

r-E
"-
~

...
N

...
I

-u:I

Id
&
10

....=
OS

=
-0

1
0 50 100 150 200 250
distance ~rom el ectrodes (..u crons)
Figure 4.15: Estimated Spatial Distribution Using 6 Spatial Wavelengths (Case 2).
Simulated Data in Table 4.3, d 2 = 500 p.m. (6) - first estimation, (X) - final est.ima-
tion.

136
 vl6 - .25" exp. dist." 12 lambda.
5~-----------------------------------------------,

- &
4

"~

....
....
I
3

....
~

-= 2
.,
-c
."
en
s:a..
Q)
1

o~------~--------~--------~--------~------~
o 50 100 150 200 250

100r-~~;-----~~--------------1

. -E

"
~

10
....
N
I

....
~

-fU
E

-~

en
~
1

-c

0.1~------~--------~------~~------~------~
o 50 100 150 250
di stance E'rom el ectrodes (mi crons)

Figure 4.16: Estimated Spatial Distribution Using 12 Spatial Wavelengths (Case 2).
Simulated Data in Table 4.3, d2 = 500 p.m. (6) - first estimation, (X) - final estima-
tion.

137
 Table 44: Simulated Data for Uniform Medium.

f = 0.1 Hz f = 1 Hz
A galll phase galll
~~e
(ILm) (dB) (deg. ) (dB) (deg. )
- . --=
20 -44.54 -76.42 -56.46 -22.16
100 -32.11 -89.50 -50.74 -42.31
200 -28.28 -98.35 -49.42 -46.99
400 -28.33 -108.03 -49.73 -45.80
800 -30.73 -109.78 -50.12 -43.19
1000 -31.62 -107.48 -50.24 -42.47

frequencies, the difference lying in the estimated par;tmeter that has the most signifi-
cant divergence. At 0.1 Hz the permittivity estimate has the larger deviation whereas
a 1 Hz it is the conductivity estimate that has the larger deviation. This behaviour
indicates the importance of knowing where in the gain-phase mapping (see Figs. 4.3-
4.6) the estimation is being performed. At higher frequencies, the estimation routine
is less sensitive to ..:onductivity (closer to gain axis) whereas at lower frequencies, the
routine is less sensitive to permittivity (doser to origin).

One problem may be that on the subsequent mar -hes through data, in the effort to
refine the estimates, the errors from the longer spatial wavelengt.h estimations may be
unduly influencing the estimates using the shorter spatial wavelengths. For example, if
the estimate £2' resulting from A2, perturbs in one direction the predicted response, as
seen by the shorter wavelength A1, on the following iteration, then the refined estimate
er will t,ry to compensate by perturhing in the other direction. VII fnrt11 nateiy, this
can be seen as a type of positive feedback situation, resulting in an unst.able process
producing divergiilg complex permittivity profiles. This might be t.he cause of t.he
divergence in the examples using a uniform mediulll (Figs. 4.17 and 4.18). If t.he
refinement process is unstable, any perturbation errors resulting from 1.he first pass
may cause growing, rather than decaying, errors ill the subsequent iterations.

138
 vl21 - .25. uni~orM. ~=.lHz
2.0

~
.. ,
-"'
Iil
fa..
If
I
7

2.3 f-- 3
'Pt
'Pt 2-
,,~

I
r..l -....
-
'Pt

s::
., UIo
.~

-
...
A

~ d

-0
.~

tIJ
~
1.8 I-
2. -
'I
.:J

8
G)
~
"
" "
.. ~q

1.3 I I I I
o 50 100 150 200 250

1.04

~,
15 1 . 02 I- Od"
" ......
"'
CoO
C
.
.a. '2..2.I#~ '" (,1.j
'Pt
~ 1 .... 2...Jr

-
~
- I

-.
'Pt A

lIS
Iil
tJ)
" 7
5~
·~O. 98 -
" ~

"
0.96 I 1 1 I
o 50 100 200 150 2SO
di stance f'rom electrodes (nd crons)

Figure 4.17: Estimated Spatial Distr:ibution Using 6 Spatial Wavelengths (Uniform

Medium). Simulated Data in Table 4.4, f = 0.1 Hz. (6) - first estimation, (X) - final
estimation.

139
 vl22 - .25. unif'orM. 1'= 1Hz
2.1

-~2.05
~
I--

~
'fJ. " ,
~
:3
'Li' ,
"

..
i
....
..
- ~ v
J.,
1-

'/,~~7~
.to

" .. Cl
'If

-....
s::
o
(1)1.95 r-
~
Q)
,..
:3
¥
~ et· :L

3.1/
~
7'fi o.1

S ",,7.8
"

1.9 I I I I
o 50 100 150 200 2SO

1.125
....
'.~.'f

-"
E
Cl.) 1. 05 I--
3
.:L
s
'1
7

• ."fi"
-t
~

o
....
I
6
..,.S
.
Jt

- ..
~
.... W~
~
- 2-
.to

-
I-- ;z. 3
....
tI)
la
. "I
-..l

1"Cf

0.9 I I
"I I
o 50 200 100 150 2SO
distance f'rom electrodes (microns)
Figure 4.18: Estimated Spatial Distribution Using 6 Spatial Wavelengths (Uniform
Medium). Simulated Data in Table 4.4, f = 1 Hz. (6) - first est.imation, (X) - final

estimation.

140
 Another problem may lie in the usage of the stair-step representation for approxi-
mating a smooth distribution. This amonnt.s to represent.ing distrihutf.'d dlargf"S wit.hin
the bulk usiug impulse charge distributions at the interfaces. The discontinuities in-
troduced by such a representation may be placing an unfair demand on the estimation
routine.

4.3.5 Step Discontinuities

Case 3 - In many instances, the distribution may actually consist of two different,
homogeneous materials, forming a step discontinuity. A stair-step representation is
particularly appropriate for the first example presented here. The permittivi'ty of the
two materials is the same, 2z10- 11 F Im, but there is a step change in conductivity at a
distance of 50 p.m from the electrodes, decreasing from 10- 10 to 10- 12 Slm away from
the electrodes. The simulated data used is ill Table 4.5. It is seen in Fig. 4.19 that
the estimation routine converges almost exactly to the discontinuity. This convergence
is very dependent on the location of the discontinuity relative to the stair-step layer
thicknesses, as established by the spatial wavelengths used. For example, with the
discontinuity moved to 70 p.m, and using the same spatial wavelengths, the estimation
does a very poor job of finding the distribution (Fig. 4.20). It cannot compensate
for the fact that the discontinuity occurs ill the middle of one of the layers. A more
natural approach to this problem, when it is known that a discontinuity exists, is to
use variable wavelength data to estimate both the layer thickness and complex bulk
permittivity, as shown in the following section.

4.3.6 Layer Thickness and Complex Bulk Permittivity

Case 4 - The previous example of a st.ep discontinuity is t.reated in a different. manner

here. It is assumed at the outset that t.he complex bulk permittivity of t.he layer
adjacent to the electrodes has been estimated using a very short spatial wavelength
measurement, one that probes only this layer. Then, a two parameter estimation is
performed to obtain the layer thickness and complex bulk permittivity of the mat.erial

141
 v13 - .25. step discon.

3 - ~

-E 2.7 ~
3

"....
fat a.
~
'I
....
I 2.4 I--
If S
....
r.l

-~ 2.1 -
,
":C G
~

--"
Q '1 11

~ ..,
~ ~
,
·1I·..r lA
~ .:J
1.B I--
"a
Cl)

-.
2..

.. 2.
1.5 I
o 50 100 150 200 250

100 ..,
..,~
~ .

r-E A

"
~ 10 I-- .J:
.,s
.3
....
N
•
I
."./'
....
r.l

-..,
1 -
" ~

~
~

-"
E
."

."
0.1 -
;
&
If

-Q

0.01
o
I
50 100
distance f'rom electrodes (microns)
-l
150
#~'l.c
.
200 -.2.iJ
2SO

Figure 4.19: Estimated Step Discontinuity Using 6 Spatial Wavelengths (Case 3).
Simulated Data in Table 4.5, d2 = 50 p.m. (~) - first estimation, (X) - final estimation.

142
 vl4 - .25. step di scan.
2.9

,......
-"
~
&
2.6 f- oS

3
"
If

..t ... ~

..t 2.3 f- ,,~

I ~
W ~ :t..

-
..t

s: 2 ...- -
...
-
-.....,
0

en
~ 1.7 I--
" -...
A

-.~;3
1.4 I
o 50 100 150 200 250

100 ",,--

f""'"-"'-

-"
E
v.. 10 f-
I-a.,
I-'

N
..t
I
-...

-
W
..t

~
1 I--

-"
lE
m

t»
0.1 ---
-0

I
0.01. .I ......
o 50 100 1~ 200 ~
250
distance f'ran electrodes (Microns)
Figure 4.20: Estimated Step Discontinuity Using 6 Spatial Wavelengths (Case 3).
Simulated Data in Table 4.5, d2 = 70 p,m. (b.) - first estimation, (X) - final estimation.

143
 Table 4.5: Simulated Data for Case 3.

d2 = 50 J.Lm d2 = 70 J.Lm
A gain phase gam phase
(J.Lm) (dB) ( deg.) (dB) (deg. )
-

20 -44.54 -76.42 -44.54 -76.42

100 -32.14 -89.48 -32.11 -89.50
1
200 -29.06 -98.79 -28.50 -98.48
400 -33.32 -116.12 -30.87 -112.24
800 -52.94 -121.35 -43.93 -128.72
1000 -56.65 -32.27 -53.70 -111.09

beyond this layer using the data from two or more longer spatial wavelengths, t.hose
capable of probing beyond the first le.yer. This two parameter estimation is obtained
using the formalism described in Section 4.1.2. The errors, and thus t.he elements of
the jacobian, are complex because the form of the gain used is (2.1). The complex
jacobian is formed by taking numer5cal differences to approximate the local derivatives.
Here, the optional initial guess that is programmed into the routine begins with a
layer thickness of (d 2 = At/8), where Al is the shortest spatial wavelength used for this
estimation, and an insulating, nonpolar material above this layer (El = 9.07:dO- I2 F /m
and 0"1 = 3x10- 12 S/m). As is t.ypical of Newton search routines, these programs are
not very robust when the guesses are very far from the root. This is the reason for
beginning the routine with a programmed guess for the layer thickness in the middle
of the sensitivity range of the shortest wavelength. The update to the thickness (Ad 2 )
ut.ilizes only the real part of t.he complex update. The soft,ware rotlt,ine is disconl.for
and may be found in Appendix F.
The first set of examples reexamines t.he discontilluit.ies presented in Case 3. The
complex bulk permittivity of the lower layer is known from a short. spat.ial wavelengt.h
estimation, say a 50 J.LID measurement (€2 = 2x10- 11 F /m and 0"2 = 10- 10 S/m). Using
only the data at 400 and 1000 fun from Table 4 ..5, the two parameter estimation of t.he

144
 Table 4.6: Simulated Data for Case 4.

d2 = 50 I'm d2 = 70 I'm
>. gain phase gam phase
(I'm) (dB) (deg.) (dB) ( deg.)
=--=
400 -50.94 -15.79 -51.41 -9.43
1000 -50.26 -32.11 -50.48 -27.67

lower layer thickness and the upper layer complex bulk permittivity was performed.
Convergence to the correct profile was obtained in 6 iterations for the 50 I'm discon-
tinuity (Fig. 4.21) and in 4 iterations for the 70 I'm discontinuity (Fig. 4.22). Again
the triangles indicate the initial estimate and the crosses, the final estimate. The layer
thickness estimates are indicated 'at each iteration, as well as the permittivity and
conductivity estimates.

A second set of examples involves estimating the same discontinuities with the
layer properties reversed, namely the lower conductivity layer is now adjacent to the
electrodes. The simulated data, again using 400 and 1000 I'm spatial wavelength
electrodes but now at 1 Hz, is given in Table 4.6. Convergence to the correct profile
was obtained in 6 iterations for the 50 I'm discontinuity (Fig. 4.23) and in 5 iterations
for the 70 I'm discontinuity (Fig. 4.24). As with all parameter estimations, the accuracy
of the parameter estimation will depend on the specific region of the gain-phase plot
(of the type shown in Fig. 4.3) in which a search is being conducted. This, in turn,
will depend on the temporal frequency at which the data was obtained. For example,
repeating this estimation problem using data at 0.01 Hz results in a poor est.imate of
the conductivity of t.he upper layer.

4.3.7 Complex Surface and Bulk Permittivity

Case 5 - Unlike the previous case, the layer thickness is known here. Instead, t.here
is an unknown surface conductivity on the surface of this layer in addition to the un-
known bulk conductivity beyond the layer. Again, it is assumed that the complex bulk

145
 dd100 - e2=2E-11. s2= 1E-10. ~ =.1

3
3

,-E
fa..
3

2 ..... ....
....
....
,...
IIf
,~

oS
I
~ 'f

-
....
~
~~
~

--
s:
0

(4
s:lo.
IU
1 I-
-
a

o I I I I
o 20 40 60 BD 100

1(0)

-,-e 100
~
,

t
....
N
3
I

....
~
10 r- " 1I Z.

-IU
e 1 I-
:I
It.

" ~

-
~

en
(4

en 0.1 r-
-0

0.01 I 1 I 'f I
o 20 40 60 BD 100
di stance ~rom el ectrodes (rni crons)
Figure 4.21: Estimated Layer Thickness aud Upper Medium Complex Bulk Permit.-
tivity Using 400 and 1000 fUll Spatial Wavelength Data from Table 4.5 (Case 3,
d2 = 50 p.m).

146
 dd101 - e2=2E-11. s2= 1E-10. F=.1

-S

"
fa.
'9"'4
2 "
3
'9"'4 ~ 2-
I
fa1

-=
'9"'4 .~

-"
0

At
1 f-
-
A

o I 1 I I
o 20 40 60 80 100

1000

.-S

"
C4
N 100 A .. ~

'9"'4
I
fa1

-'"
'9"'4

S
~
3'1
....
tD
10 - .4!' <:
"
tD
...,
-0 7.

:5
1 I I I I
,..
o 20 40 60 80 100
di stance From el ectrodes (Mi CX"01a8)
Figure 4.22: Estimate-:l Layer Thickness and Upper Medium Complex Bulk Permit.-
tivity Using 400 and 1000 I'm Spatial Wavelength Data from Table 4.5 (Case 3,
d2 = 70 I'm).

147
 dd103 - e2=2E-11. s2= 1E-12. ~ =1

- &
3L....
3
"
~
2 ., ,~

~
~ . at ~
~
I
W
~

-s::
.~

.-...
0 1 f--
A

"
Pot
I)

o I I I l
o 20 40 60 80 100

1£DJ

-"
-&
c;.:,
N
~
100 I""'- 3 --'1
-""-'I' ...
I

-
W
~
2-

~ s
&
....
t:D 10 f--
3
If ,
"
t:D
.J.

- 0

I I
Ij •
A

I
1 ...
o 20 40 60 BO '100
di stance ~rom el ectrodes (Mi CI"Ons)
Figure 4.23: Estimated Layer Thickness and Upper Medium Complex Bulk Permit.-
tivity Using 400 and 1000 I,m Spatial vVavelength Data from Table 4.6 (Case 4,
d2 = 50 I'm).

148
 dd1.04 - eZ=2E-1.1.,. 82= 1.E-1.2,. f'=1.

-"
~
15
31
3

......
2 "-
~I "'1
I ~
...
-
~
~~

s::
-....
0

tit
~
1 r--
-
A

Gl

o I I I I
o 20 40 60 eo 100

1000

-
.- 15

"
c:.-,
If
...
N 100 I -
~

...
I

-
~
1-

...s

-15
0)

U)
10 -
;Z
~ I~ 3

-tn
0

I I I 1
,j~
A

1. ~
o 20 40 60 BD 1.00
distance f'roM electrodes (Microns)
Figure 4.24: Estimated Layer Thickness and Upper Medium Complex Bulk Permit-
tivity Using 400 and 1000 fUll Spatial Wavelength Data from Table 4.6 (Case 4,
d 2 = 70 pm).

149
permittivity of the layer adjacent to the electrodes has been estimated using a very
short spatial wavelength measurement. Now, a two parameter estimat.ion is performed
to obtain the surface conductivity and complex bulk permittivity of the material be-
yond this layer using the data from two or more longer spatial wavelengths. This t.wo
parameter estimation is again obtained using the formalism described in Section 4.1.2.
A complex jacobian is formed by taking numerical differences to approximate the local
derivatives. Here, the optional initial guess that is programmed into the routine begins
with a complex surface permittivity that is purely imaginary, with 0'12 = lWfo At/4) S,
where Al is the shortest spatial wavelength used for this estimation and W is the an-
gular temporal frequency at which the response was measured, and a semi-insulating,
nonpolar material above this layer (fl = 9.07:z:10- 12 F Im and 0'1 = Wfo S/m). Again,
to improve robustness, the routine begins with a programmed guess for the surface
conductivity in a regime where ((O'a24/wfoA) '" I}. It is only in the regime where the
surface conductivity contributes to the loss that the parameter may be estimated. The
routine allows suppression of the real part of the complex surface permittivity if only
a surface conductivity is to be estimated. The software routine is discon2.for and
may be found in Appendix F.
The first example involves a layer 50 I'm thick, having a bulk permittivity of
2:z:10- 11 F Im, a bulk conductivity of 10- 10 S/m. and a surface conductivity of 10- 16 S,
below a medium of unknown complex bulk permittivity. The correct parameter values,
along with the estimated values at each iteration, are shown in Fig. 4.25. In the plot
of surface conductivities, the square denotes the actual value while the triangle and
cross indicate the first and last iteration, as before. This search was conducted using
simulated data obtained from 400 and 1000 I'm electrode structures at 0.01 Hz. Con-
vergence occurs in 4 iterations. This search is repeated using simulated tlnt.n obtained
from 600 and 1000 I'm electrode structures at 0.01 Hz (Fig. 4.26). Convergence again
occurs in 4 iterations but the accuracy of the parameter estimates is not. as good. The
closer spatial wavelength pair, here 600 and 1000 PIll, are less sensitive to the exact.
parameter values and consequently, are les:; able to discern the exact parameter values.
The simulated data used is tabulated in Table 4.7.

150
 dd2fXl - F =.01. 400 a 1000 um

- E
J

"
~
2 ""
"
u

....
~

....'-1
-
'1.

-- -
0
1 ~
...
r
m CQ
Po.
GJ
....
(D

I
'-1
o I I I I ....
o 50 100 150 200 2Slr'
.a
E
C)

11XlO

r- E
"
~ 100 \0

"
....
N
I
....
'-1

-.a
E
10 ~

C)
.~

m ~
1 f-
tn ~ 0

-0

0.1 I I I I O. 5 I I I
~_I.-_I.-_I.-_I.-----J

o 50 100 150 200 250 0 50 100 150 200 250

di stance (Mi. crons) distance (microns)
Figure 4.25: Estimated Layer Surface Conductivity and Upper Mediulll Complex Bulk
Permittivity Using 400 and 1000 p.m Spatial \Vavelellgth Data from Table 4.7 (Case
5, U.2 = 10- 16 S).

151
 dA2(J1 - ~ =. ()jL # 60J a 1()(x) um

1.2,.------------,

-lE
"- ..v
fir".
..-4
..-4
2 --
T'
4
I 1.1 -
~

-
..-4

$:
~

-.....,
0

M
~
1 -
- A.

-
~

(D
..-4 1 r- J1
I

o ,- t I I
~
..-4
o 50 100 150 200 250'"
IU
IS
....
~

1000

,..- lE
"- ...
CI2 lOO
N
"
..-4
I

-
~
..-4

IU
S
10 ~

0.8 r-

....
~

M '2.,3
1 I-

- '}
~
0 -
0.1 I I II 0.7~~--~--~--~--~
j I I
o so 100 150 200 2SO 0 50 100 lSO 200 250
di stance (m:i crons) di stance (m:i crons)
Figure 4.26: Estimated Layer Surface Conductivity and Upper Medium Complex Bulk
Permittivity Using 600 and 1000 p.m Spatial Wavelength Data from Table 4.7 (Case
5, 0'.2 = 10- 16 S).

152
 Table 4.7: Simulated Data for Case 5.

0',2 = 10- 16 S 0',2 = 10-17 S

.A
(I'm) (dB)
gain 1 phase
(deg.)
gain
(dB)
phase
(deg.)
- -
400 -8.96 -74.18 -50.69 -34.12
600 -10.91 -83.67 - -
1000 -15.02 -104.40 -51.95 -13.03

A second example, having a layer closer in conductivity to the upper medium

(E2 = 2:1:10- 11 F Im, 0'2 = 10- 12 S/m), is shown in Fig 4.27. Here, the search used
simulated data from 400 and 1000 I'm electrode structures again at 0.01 Hz. Conver-
gence occurred in 4 iterations and yielded very good estimates of the actual parameter
values.

4.4 Two Parameter Estimation Using Variable Di-

electric Media
Very often, both the layer thickness and permittivity are unknown, requiring two
data points for the estimation problem. One parameter that can be easily varied, in
many situations, to obtain the two data points is the permittivity of the surrounding
medium. As with the layer thickness estimation of Section 4.2.1, the search is in the
high frequency, purely capacitively coupled regime. Thus, the gain used in (4.11) is
purely real.
Unlike the complex permittivity estimation, this two dimensional search cannot
be converted to a one dimensional search of a complex parameter. Therefore, t.he
minimization technique present.ed in Section 4.1.2 is utilized here. The guesses are
updated using a 2:1:2 jacobian matrix (4.8). Here, with (Jl = d2 (layer thickness),
(J2 = E2 (layer permiitivity) and Gp the real part of the complex gain, the jacobian is

153
 dd202 - 1'=.01. 400 a 101XJ um
G~-------------.

-
"-
le
:2
~

~
~
2 --
Y'

~
,.~ ~

51--
I
W

-
~

s::
- 0
....1'4
~
1 f-
-A

-Cl.) 41--
CD
E'o-
~
I
W
o I I I I .-4
o 50 1.00 1.50 200 25()'"
Id
le 3 I--

100
....
tD
Cl)

CD

-,-"-
CJ
Id 2.
~
le ~
Cl.)
.., =
Cl) 2 I--

('I') 1.0 ....,..

.-4
I •
W

-
~

~
:I-
3
le 3J '1 1 I~,
....1'4
tD
1 ~ ,.
t-

-
tD
0

0.1 I I I I o 1
L....--1-._L-.- _ 1L -.-_L-.-1--I
o 50 100 150 200 2SO o so 100 150 2lXl 2SO
di stance (DU crons) di stance (Mi. crons)
Figure 4.27: Estimated Layer Surface Conductivity and Upper Medium Complex Bulk
Permittivity Using 400 and 1000 p.m Spatial Wavelength Data from Table 4.7 (Case
5, (T.2 = 10- 17 S).

154
 . -
e'( 6) - "
. . ( 4.27)

812 8;P2
8
2 E2
For each guess, 4 additional function evaluations are required to obtain the partial
derivatives. The time required for a search was reduced somewhat by again using
the variable collocation point technique, searching first with 2, then 10 and finally
25 collocation points. Convergence usually takes anywhere from 5-20 iterations. No
check on the validity of the data is made in tbis routine. Bor robustness, guesses are
again bounded using physically motivated limits. There is also a maximum limit on
the number of iterations.
The software for this two parameter estimation routine, th4est.for, may be found
ill Appendix F. This routine was used to estimate the thickness and permittivity of
the plasma deposited bromobenzene layers (Table 5.1).

In summary, all of the estimation routines described here demonstrated reason-

able convergence, with the exception of the variable spatial wavelength estimations of
Section 4.3. In the following chapter, examples will be presented demonstrating the
utility of the layer thickness and complex bulk and surface estimation routines.

155
Chapter 5

Parallleter Estilllation Ex·arnples &;

Oil COlllplex Perlllittivity
MeasureIhents

There are several topics to be covered in this chapter. First, it is important to obtain
experimental verification of the absolute measurement capability of the interdigital
electrode structures when used in conjunction with the cOlltilluum model, and the
parameter estimation techniques previously described. Second, the test cases chosen
for the verification were motivated by possible applications to transformer monitoring
and involve measurements of the dielectric properties oi oil, paper insulation and
moisture sensitive thin film coatings. Third, some of the experiellce, when using the
microchip and macro chip for various monitoring applications, is related. This includes
situations when problems occur, such as the complex bulk permittivity measurement
of insulating fluids with a microchip.
After outlining the calibration techniques for the microchip and lllacrochip in Sec-
tion 5.1, several cases are cited in which parameters are both estimated from experi-
mental data and measured independently. These cases consist of bot.h the layer t.hick-
ness of the thin film coatings (Sections 5.2 and 5.3) and the bulk complex permitt.ivity
of oil (Section 5.4). Additional cases are presented where the experimental response
can be predicted using the continuum model, though no independent measurement

156
was performed. These cases include monitoring the sedimentation of small particles
(Section 5.2.1) and the surface adsorption of moisture (Section 5.4.1), when it is diffi-
cult to envision simple, alternative techniques for these situations. The sedimentation
example arose from the idea of monitoring the particulate concentration in transform-
ers while the surface adsorption example was the result of a problem encountered when
using an uncoated microchip in wet oil.
For some cases, where the medium is not ohmic but dispersive in complex per-
mittivity, the dispersion can be estimated, allowing either a better understanding of
the loss mechanisms or an interpretation of the response when using the medium in
other physical situations. Examples of the former include the study of the surface
conduction between electrodes due to moisture adsorption or the behaviour of submi-
cron films deposited on the microchip (Section 5.5). The submicron films were vacuum
deposited palladium and/or tin oxide and were motivated by the idea of adapting the
microchip to a hydrogen gas sensor. An example of the latter is using a passivated
electrode structure to estimate the complex bulk permittivity of oil (Section 5.4.2).
Examples of both instances include measurements of the complex bulk permittivity of
paper insulation (Section 5.6) and moisture sensitive thin film coatings (Section 8.2.3).

5.1 Calibration

5.1.1 Microchip

For the case of the microchip, there is significant variability in the oxide layer thickness
due to the fabrication process used. Therefore, before any parameter estimation can
be made, the response of an ullcoated microchip must be calibrated by est.imation of
the oxide layer thickness. This is usually performed using the high frequency gain of
the bare microchip in air. Given this gain, and the load admittance (Y i ), t.he oxide
layer thickness can be estimated using a secant search similar to the one outlined in
Section 4.2.1 for layer thickness. The error tolerance for convergence is 0.03 dB. This
routine is called best.for and may be found in Appendix F.

157
 Calibration of the microchips, as performed by Micromet, is slightly different. The
expected range of high frequency gain in air, given the variabilit.y in oxide layer thick-
ness, is from -39.9 to -43.0 dB. This range is divided into 9 intervals of 0.3 dB width
and 1 interval of 0.5 dB width, with a chip identification number assigned to each
interval. This ID number, when input into the microdielectrometer, indicates the par-
ticular look up table to be used for mapping the gain-phase space into the complex
bulk permittivity space (using a uniform medium model). The 0.3 dB width intervals
correspond to a variation of 0.02fo F Im in permittivit.y.

5.1.2 Macrochip

For the case of the macrochip, as mentioned in Section 1.4.2, there was a problem as-
sociated with fringing fields and ascribed to the smaller aspect ratio of the electrodes
presently available [74]. The deduction was that fringing fields resulted in a measur-
able reduction in sensitivity to the medium above the electrodes than predicted by the
continuum model. A compensation technique for calibrating the macrochip was devel-
oped by Li [23, p. 40]. In this technique, drawing on the discussion of Sect.ion 3.1.5, the
coupling admittance Y12 is separated into two parallel admittances representing the
coupling through the medium above and below the electrodes. The effect of fringing
is modeled as contributing a purely capacitive component that is independent of the
medium above. Calibration is performed using a high frequency measurement in two
different dielectric media and solving for the load capacitance (Cl) and frillging field
capacitance (CII:) using the experimental gains and satisfying the relation between the
gain and the admittallce parameters, (2.1). With a redesign of the macrochip elec-
trodes, specifically an increase in the electrode aspect ratio, it is expected t.hat. t.he
fringing field component will be negligible.

5.2 Thickness Gauge

A simple application of these interdigital electrode structures, in conjunction with the

parameter estimation routines discussed in Chapter 4, is as a measure of the thickness

158
of dielectric or conducting layers. If measurements are taken at a frequency such that
the coupling between the electrodes is either purely capacitive (or cOlUluctive), the
response will vary only with the permittivity (or conductivity) and thickness of the
medium above the electrodes. This allows measurements of layer thicknesses that
would be difficult to obtain, particularly in a dynamic mode, where the layer thickness
evolves with time. With the availability of microfabricated structures, this technique
projects the thickness measurement into the micrometer and submicrometer range.
The microdielectrometer provides the ability for performing a high impedance mea~
surement (capacitive limit). Several examples of this capability are given in the next
section. An examination of the virtual ground measurement (conductive limit) is also
presented although no experimental measurements have been made for this case.

5.2.1 High Impedance Measurement

Gain to Thickness Mapping

When configured in the high impedance mode using a floating gate electrode, the
device output at relatively high frequencies will reflect the purely capacitive coupling
between the electrodes. There will be an attenuated response, in phase with the
drive, varying only with the permittivity and thickness of the medium above the elec~

trodes. This output is inherently dependent on the parallel capacitances of the media
above and below the electrodes. Thus, the output gain is specific to the particular
oxide layer permittivity (Eo:c), thickness (h) and load admittance (Y i ) describing the
microchi p. The ratios of interest that completely describe the problem are (Et! Eo:!:),
(E2/Eo:c), (d 2 /).), and (hi).). Given the variety of parameter combinations, no single
two~dimensional plot can be made to represent all t.he possibilities.
One limit, shown in Fig. 4.2, is the high frequency gain as a funct.ion of thidmess
for a dielectric layer, with air above, on t.op of a commercially available microchip
[66]. Again, knowing the permittivity of both the layer and the semi-infillt.e region
above, and given an experimentally measured high frequency gain, a unique layer
thickness can be estimated. Observe that the sensitivity to layer thickness is practically

159
nonexistent beyond roughly A/3. For a practical implementation, the mapping must
take into account the specific oxide leyr thickness of t.he microchip, a value t.hat can
vary by roughly 20%.
Another interesting limit is that of a dielectric layer with an equipotential at its
upper surface. This equipotential may be produced by either a highly conducting
medium above the layer, or a high surface conductivity on the layer. The ratio of
(eI/ern:) is unimportant here. For this case, a two dimensional plot can describe all the
possible combinations of layer thickness and permittivity when using a microchip. The
high frequency gain as a function of thickness for various ratios of (e2/ ern:), again using
a microchip, is shown in Fig. 5.1. Here, the region of high sensitivity (large slope)
occurs at higher signal levels (gain closer to unity). Thus, this mode of thickness
measurement may provide better accuracy in practice.

Experimental Results
The fh'st demonstration of applying the microchip to thickness monitoring is a
qualitative one, involving a mass transfer process using evaporation. A small volume
of transformer oil ("" 1 #-tL) was placed on the microchip electrodes and the initial
frequency response was measured (boxes of Fig. 5.2). The microchip, with the oil
sample, was then placed in a chamber and evacuated to remove the oil via evaporation.
After a time, the chamber was repressurized to atmospheric and a second response
was taken (dots). The third response (crosses) was taken after further evacuation,
resulting in complete removal of the oil. The transition of the response from bulk-like
to surface-like clearly demonstrates the effect of thinning down the oil layer.

Jumping ahead slightly, the bulk properties of the oil layer prior to evacuation were
estimated from the first response, assuming the oil layer was thick enough so that the
field never penetrates the air above. The predicted response using t.hese est.imat.ed
parameters (f = 2:1:10- 11 F/m, (J' = 2.87:1:10- 11 8/m), yielded a good fit. t.o t.he dat.a
(boxes and solid line). The estimated value for the oil permittivit.y was t.hen used
to estimate the oil layer thickness associated with the second response (after some
thinning of the layer had occurred). Using this estimated oil layer thickness of 1.09
#-tm, the bulk properties of the oil layer were then re-estimated (e = 2.35:1:10- 11 F/m,

160
 o~----------------------------------------------~

-10
5

-..,-"' -20 2

....IU
~ 1

~
-30
.5

-40
.2

4 B 12 16 20

Layer Thickness. d2 (um)

Figure 5.1: Predicted High Frequency (10 kHz) Gains for Microchip with Variable
Thickness Layer (d 2 ), Vary Layer Permittivity (E2), with Perfect Conductor Above
Layer.

161
 o~------------------------------------------~
(0) - Thick Oil Layer; H - tN = 2.00zl0- 1I F Im
tiN = 2.81zl0- 11 S/m

-10 (e) - Thin Oil Layer and Air; (-) - dN = 1.09 pm

t ..., = 2.35%10- 11 F Im
tiN = 4.00%10- 11 S/m
(X) - No Oil, Completely Evapora~

-20

-
$:
~
en
-30

-40

-50~-----L------~----~------~----~------~----~
-3 -2 -1 o 1 2 3 4

.' -28

-56

-112

-14O~----~~~~----~~----~----~----~----~
-3
log freq

Figure 5.2: Measured and Predicted (-) Frequency Responses of Microchip with Oil
Layer of Variable Thickness in Air.

162
u = 4.90:z:10- 11 S/m). Again, the predicted response using these estimated parameters
yielded a good fit to the data (dots anI I solid line). The loop, howt'ver, is 11ot. a
closed one for several reasons. Neither the oil layer thickness nor the complex bulk
permittivity were measured independently. Furthermore, the oil permittivity used to
estimate the oil layer thickness did not match the re-estimated oil layer permittivity. It
is possible that there is a significant change in both the permittivity and conductivity
of the thinned down oil layer, possibly due to the removal of lower molecular weight
material or changes in ion concentration. A more consistent procedure for estimating
the oil layer thickness and complex bulk permittivity (assuming an ohmic model is
valid for this thin oil layer ) would be to simultaneously estimate bOUl quantities from
two or more frequency data points.
The opposite process, that of mass transfer by deposition, provides a quantit.ative
example of a thickness gauge. Using the plasma deposition apparatus described in
Appendix C, and used to coat microchips with moisture sensitive thin films, an ex-
periment was conduded whereby a microchip was coated with a very thick film ("-' 33
J.Lm). By taking a measurement at several times during the deposition process, with
the microchip in the evacuated deposition chamber but with the RF field turned off, a
plot of high frequency gain versus time was obtained (Fig. 5.3). After a long deposition
time, the gain saturates due to the thickness of the coating, as also seen in Fig. 4.2. At
this thickness, the permittivity can be estimated using a semi-infinite layer thickness.
Assuming the permittivity is constant throughout the deposition process, the thick-
ness at each measurement can then be estimated, yielding the linear deposition rate of
165 A/min. Given the particular configuration used for deposition, this rate compares
reasonably well with those reported elsewhere [83, Section 8.3.3.2 and p. 324].

A similar problem is one of monit.oring the sedimentation of small particles in a

fluid. Specifically, Li [23, Chapter 6] examined the sedimentation of 1l1011odisperse,
semi-insulating glass beads in transformer oil onto a passivated macrochip sensor. A
high frequency (10 kHz) gain versus time plot displayed the same trend as shown in
Fig. 5.3 - beginning linearly and ending at a saturated value. Correlations were made
between the observed settling velocity and rate of rise of the sedimented layer with

163
 -28~------------------------------------------------~

-31.
x

--=
~
-34

- $:

iG
tn
-37

-43L-------------~------------~~----------~
o 500 1000 1500

15

-
.
•
12

-
-"
E
::s
9
slope ,... 165 A/min
"s::
Gl

~ 6
u
....
..c::
~

o~------------~--------------~------------~
o 500 :l0lXJ 1500
time (rni n)

Figure 5.3: Measured High Frequency Gain of Microchip During Deposition of Plasma
Deposited Bromobenzene and Estimated Layer Thickness using (;2 = 3.0xlO- 11 F /m.

164
estimates of the effective layer permittivity and layer thickness. As far as the feasibility
of a particulate monitor was concerned, it was determined to be impra.ctical due to
the polydispersity and wide range of dielectric properties of the particles present in
transformer oil.

5.2.2 Virtual Ground Measurement

Conductance to Thickness Mapping

In this limit, the "floating" electrode would be terminated with a virtual ground,
providing a direct measurement of the admittance Y12 (see Fig. 2.1). At a temporal
frequency low enough such that WTe ~ 1, where Te is the charge relaxation time within
the medium of interest, the coupling between the electrodes would be purely conduc-
tive. The measured admittance would be purely conductive, G 12 , and would depend
only on the conductivities of the layer and medium above, and the layer thickness.
Thus, given the two conductivities and a measured conductance, the layer thickness
could be estimated, using the same estimation routine. Furthermore, assuming the ox-
ide layer remains insulating, the only ratios of importance now are (O't /0'2) and (d 2/ >.).
A two-dimensional plot can describe all the possible combinations of materials above
the electrodes. This assumes the conductivities are ohmic or that the dispersive nature
is well characterized in the frequency regime of interest.
A plot of the normalized conductance, (G12/0'2Med as a function of normalized
thickness, (d 2/>.), for various ratios of (0'1/0'2), is shown in Fig. 5.4. As expected, the
measured conductance will be larger for greater ratios of (O't /0'2) and approaches the
conductance of the lower layer as the layer thickness increases. This limiting normal-
ized conducta.nce is 0.967, not. nnity. In CL ('r1Hl". Rf"tlSf", t.llf" normn1izf"d ('onrll1rt.nll('e
seen by the electrodes in this limit represents t.he rat.io of t.he electrode width (A/4)
to the effective "length" of the conductor. The iact that this rat.io is less t.han unit.y
implies that the path length of the conduction, in some average sense, is slight.ly longer
than (>./4). This is not surprising as the fields are distributed throughout the medium
and not just concentrated at the electrode edges.

165
 100

((J;,/(fj.) =

--•
~
10

.,Q

.....tD
-"'"
N
1
.r

-
.-4
<.!'

tD

-
" 0
0.1

0.01~----------------------~--------------------~
0.01 0.1 1
log [d2/l aNbda1

Figure 5.4: Predicted Normalized Conductance (0 12 /0'&;\) as a Function of Normal-

ized Layer Thickness (d 2/;\) and Relative Conductivity of Upper Merliu1l1 (O'o./U&) for
Mi.:rochip with Virtually Grounded Electrode.

166
 A discussion of the practical issues involved in implementing a virtual ground de-
sign may be found in CoIn's thesis [68, Section 2.3]. It was C"Ondncled f.Imt for t.he
application of monitoring epoxy resin cures using relatively low frequencies, this ap-
proach was not the most practical one. However, the analysis should be reconsidered
for applicati.ons in other areas and in other frequency regimes. Aside from the elec-
tronic considerations, some emphasis should be giv~n to t he potential problems that
could arise witb tbe occurrence of significant conduction currents. Tbe eledrocbem-
istry occurring at the electrode-medium interface becomes ~ven more important than
in tbe higb impedance configuration of the microdieledrometer, witb tbe possibility
of significant redox reactions and electrode polarization.

5.3 Layer Thickness and Permittivity

It is not always possible to either know the bulk permittivity of the deposited material
beforehand, or deposit a layer thick enough so as to estimate the bulk permittivity
independently of tbe layer tbickness. The moisture sensors introduced in Chapter 6
are a good example of this problem. As discussed in Section 4.4, a multipara&neter es-
timation of both the layer thickness and permittivity can be performed by significantly
varying the permittivity of the medium above the deposited layer. In this particular
case, coated chips were operated both in air and in transformer oil. This provided
enough information so as to uniquely estimate both layer properties. A compilation of
such estimates for a large set of plasma polymerized bromobenzene coated microchips
is presented in Table 5.1. Tbese coated microchips were the moisture sensors studied
in Chapters 7 and 8. The high frequency gain of an uncoated microchip in air is used
to compute the actual oxide layer thickness, thus calibrat.ing the microchip. The t.wo
high frequency gains using a coated microchip, ill air and in oil, provide t.he dat.a nec-
essary for the two parameter estimation. Witness strips composed of silicon were also
coated. Tbese strips were partially masked off, producing a step change in height.. The
coating thickness was measured using a Dektak profilometer, a device that measures
step changes in height using a stylus that tracks across the sample. The correlation

167
 Table 5.1: Two Parameter Estimation Using Coated Microchips.

ID HFG HFG I HFG Thickness Pernlittivity

# Uncoated Coated
(in Air) (in Air) (in Oil) IDeas. est. est.
(dB) (dB) (clB) (I'm) (I'm) (:z:10- 11 F Im)
- --
Zl -40.30 -39.60 - 32.30 0.7 0.6 3.01
Z3 -41.86 -40.55 -33.80 0.5 1.6 2.44
Z7 -40.87 -33.60 -31.39 7.4 6.8 2.60
Z8 -40.85 -33.74 -31.41 7.4 6.6 2.58
Z17 -42.56 -40.70 -34.30 2.0 2.1 2.63
Z20 -41.00 -34.67 -31.89 - 6.0 2.50
Z22 -41.01 -37.20 -32.30 '" 3.4 3.5 2.55
Z24 -42.93 -40.52 -34.30 2.4 2.2 2.95
Z25 -42.83 -41.72 -34.75 ,..." 1.1 1.29 2.69
C3 -42.63 -35.36 -32.95 6.4 6.7 2.7

between estimated and measured thicknesses is good to within 10%. Included in Ta-
ble 5.1 is a result obtained with a microchip passivated with a parylene coating (C3).
For the parylene sample, a witness strip was produced using a piece of aluminum. The
film was peeled off the aluminum and measured with a micrometer.

5.4 Complex Bulk Permittivity of Oil

One of the early goals of this work was to use the microchip as a monitor of the complex
bulk permittivity of the transformer oil. As the t.ransformer and oil aged, degrada-
tion byproducts might introduce detectable changes in t.he complex bulk permitt.ivit.y,
changes that would permit early notification of the degradat.ion process. As a nat.ural
extension of the epoxy resin application, an uncoated microchip was simply placed
in the oil. Besides serving as a complex permittivity monitor, this experiment also
provided a direct comparison between the continuum model, using parameter values

168
obtained with the estimation routine discussed in Section 4.2.2, the finite difference
estimate, supplied by the microdidedrometer, and a st.an(lnrfl hridge meaIHlrf"ntf"nt..
One advantage of using the microdielectrometer is its capability to measure the
dispersion at low frequencies, if any, in the complex bulk permittivity of the oil. A
limitation on standard capacitance bridges, such as a Schering bridge, is the mini-
mum measurable dissipation factor, or loss tangent (f." / f.') and is determined by the
dissipation of the capacitors within the bridge. As an example, to measure the loss
tangent of oil having a bulk conductivity of 10- 13 Slm at 0.01 Hz implies a capability
of measuring a loss tangent of 0.08 at this low frequency, not a simple nor inexpensive
requirement. The microdielectrometer, unlike the capacitance bridge, does not try to
match the impedance but performs an absolute measurement of the coupling due to
the effect of the impedance. Without the problem of having to provide low dissipation
capacitors, low frequency measurements of semi-insulating material are now available.
The interest in low frequency measurements lies in the possibility of extracting
additional information about the state of the transformer oil from any observed dis-
persion in the frequency response. Unlike the epoxy resins, dipolar dispersions that
might indicate changes in the molecular structure of the oil cannot be observed at these
low frequencies. Instead, the observable loss mechanisms at these frequencies are ionic
conduction and space charge polarization. The focus of this section is losses due to
ionic conduction. The uncoated microchip, haviug a spatial sensitivity of roughly 15-
20 p.m, is very sensitive to interfacial effects such as moisture adsorption and electrode
polarization. These properties, in turn, are dependent on external parameters such as
the moisture content and temperature of the oil. Very often, the frequency response did
not rellect the condition of the bulk oil but the conditions existing at the microchip-
oil interface. An attempt at. improving reproducihilit.y was made hy passivating t.he
microchip with a thin layer of parylene, an insulat.or known for Hs electrical st.abilit.y
and inertness, and used in biological applications [841. The parylene coating displayed
a dispersive complex bulk permittivity and increased the complexit.y of det.ermining
the oil properties. However, with the use of the continuum model, the dispersion could
be estimated and taken into account when estimating the complex bulk permittivity

169
of the oil. This technique worked fine except under conditions of dry, circulating oil.
Under these conditions, prohlems would arise due to the same illt.("rfnrini efft"rt.s, now
removed to the coating-oil interface. Though these effects represented an opportunity
for using the microchip to study electrokinetic processes, they stood in the way of the
goal of measuring the complex bulk permittivity of the oil. By going to a parylene
coated macrochip the deleterious interfacial effects were overcome. The longer spatial
wavelength of the macrochip prevented the conditions at the interface from domi~

nating the response. However, due to the less than optimum design of the present
macrochip (see Section 1.4.2), it is very insensitivity compared to the microchip (gain
accuracy of only 0.5 dB). This results in very poor accuracy in the estimates of the
bulk permittivity. Thus, at present, the combination of an uncoat.ed microchip (for
bulk permittivity) and passivated macrochip (for loss) must be used to provide the
complex bulk permittivitJ' of transformer oil. There is no reason why a redesigned
macrochip, when passivated, could not provide the complex bulk permittivity of t.he
oil over the entire frequency range. The experiments upon which these conclusions
were based are presented in the following two sections.

5.4.1 U ncoated Microchip

There were instances where an uncoated microchip appeared to be responding to the

bulk properties of the transformer oil. The response of a microchip in unused oil, at
room temperature and having a 70 ppm moisture content, is shown in Fig. 5.5 (dots),
along with a predicted bulk r~sponse using a complex permittivity estimated from the
experimental data at 0.5 Hz (solid line). A similar match is shown in Fig. 5.6 for a
different microchip in unused oil, also at room temperature, wit.h a moisture content
of 53 ppm. These results are tabulated in Table 5.2 and compared t.o t.he complex
bulk permittivit.y as estimated by the finit.e difference simulation model used by t.he
microdielectrometer. Some of the small discrepancy between the values may lie ill t.he
coarser calibration method employed by the microdielectrometer (see Section 5.1).

However, there were more instances in which interfacial phenomena dominated

170
 vd4.pvd4
0

-5

-10

-15

-
s: -2J)
IU
tI) -25

-30
-35

-40

-45
-3 -2 -1 0 1 2 3 4

-10

-20
-30
ID -40
"
IU
..c: -50
~

-60

-70
-BO
-90
-3 -2 -1 0 1 2 3 4
log ~req
Figure 5.5: Measured (.) and Predicted (-) Frequency Response for UIlcoated Mi-
crochip in Unstirred, Unused Oil with iO ppm Water (use fl = 2.03:dO- ll F/m and
0'1 = 1.46z10- 11 S/m).

171
 ~2.pzOOf'2
0

-5
-10
-15
-20
....
$::

IG
m -25

-30
-35

-40

-45
-3 -2 -1 0 1 2 3 4

-10
-20
-30
GJ -40
rn
IG
~ -50
~

-60
-70
-80

-90
-3 -2 -1 0 1 2 3 4
log ~req
Figure 5.6: Measured (.) and Predicted (-) Frequency Response of Uncoated Mi-
crochip in Unstirred, Unused Oil with 53 ppm Water (use El = 1.96x10- 11 F/m and
0'1 = 2.77:z:10- 12 S/m).

172
 Table 5.2: Oil Complex Bulk Permittivity Measurement With Microchip.

Note - E' and fll relative to fo F /m

- -------
ID # Oil HFG Response in Oil Micromet Cont. Model
Cond. (in Air) freq. gain phase f' fll f' fll

88991.23 unused, -41.03 0.5 -32.94 -14.22 2.25 0.5083 2.29 0.5249
70 ppm
88025.24 unused, -42.90 0.1 -35.25 -13.98 2.12 0.4674 2.21 0.4979
53 ppm
--'--- - . - '-- --.- -. -..-

the response. One such instance, shown in Fig. 5.7 (dots), displays a response for
an uncoated microchip in oil from a failed transformer, having a moisture content
of 50 ppm. This response was modeled using by giving the interface between the
electrodes a surface conductivity 0"0 = 2.52:dO- 14 S and a surface permittivity f.o = 0
(solid line). Adsorption of moisture and other ionizable species along the silicon dioxide
interface could account for such a surface conductivity.

Examples of other cases where the frequency response differed significantly from
a purely bulk response are given in Figs. 5.8-5.10. Fig. 5.8 shows the response in
unstirred, unused oil with a moisture content of 14 ppm. This response cannot be
accounted for using a combination of surface and bulk conductivities, as can be seen
by examining Fig. 3.6. It is believed that charge separation effects, such as double
layer formation, are dominating the response in this frequency range, producing a
relaxation higher in frequency than would be expected based on the bulk properties
of the oil.
Operating a microchip while evacuating and stirring the oil gave rise t.o dra.st.ic
and irreversible changes in the DC offset (defined in Section 104.1). The response
shown in Fig. 5.9 was taken while the oil remained under evacuation (30 mT), ha.ving
a moisture content of roughly 10 ppm, but now was unstirred. Large, negative DC
offsets of roughly -2 V were observed in the response. It is difficult t.o guess what
is occurring in Fig. 5.9 - the response looks somewhat like the surface conductivity

173
 kni'5. pknf'Sa
0

-5
-10
-15
-20
....
~

IG
tn -25
-30
-35

-40
-45
-3 -2 -1 0 1 2 3 4

r
-23

-46

-92

-115~----~----~----~----~~----~----~----~
-3 -2 -1 o 1 2 3 4
log ~req
Figure 5.7: Measured (e) and Predicted (-) Frequency Response of Uncoated Mi-
crochip ill Um;tirred, Used Oil with 50 ppm Water (use El = 2.0:z:10- 11 F/m and
u,o = 2.52:z:10- 14 S).

174
 0

-9

-
s:: -18
~
tn -Z/'
-A.
.~"

....
-36
r.......
-- ...-' -- ..-.-

-45
-3 -2 -1 0 1 2 3 4

o ~,...-
.--

-30
,I
•
....

.'
_..tI

.- -90
-3 -2 -1 0 1 2 3 4

-0.16

-0.18
• --"'' ' ....... '.
~
11)
"-0.2
•
• • ..&
--- ,..-
......
--
....

•..•
CM
CM
o
-0.22
•••
-0.24
-3 -2 0 1-1 2 3 4
log Ereq
Figure 5.8: Measured Frequency Response of U neoat.ed Microchip in Unstirred, Ullused
Oil at 14 pplll Water.

175
 0
•
-9 '4 "
....
~

~
~
-18

-Z7
'.•
....

...
-36 -'
-45
-3 -2 -1 0 1 2 3 4

-30
." .••"
•.r-

'\4
•
• •
"....
t
.~

-90
-3 -2 -1 0 1 2 3 4

-1.64

.•.'
-1.69
..., .......
".
Gl
Cf.l.74
c.,.
c.,.
• .-..
••
......
o
-1.79 .
...-.
•••
.~
W~ - ......•
-1.84
-3 -2 -1 0 1 2 3 4
log Ereq
Figur~ 5.9: Measured Frequency Response of Uncoated Microchip in Unstirred, Evac-
uated, Unused Oil at roughly 10 ppm Water.

176
 wd7
0
~
'..
-9
,
".
s: -18
......,
tn -Z?
-36
,
'-....

-45
-3 -2 -1 0 1 2 3 4

o _rw .

-30 t
•
.. .'
• ,••
't. ,.
''IN
,
-9{)
-3 -2 -1 0 1 2 3 4

-1.9

-2 -.
.J~ J ....
"'I ......
-....
.
A.

- I .......
'lA
....
••
-2.3 •
-2.4
-3 -2 -1 0 1 2 3 4
log £req
Figure 5.10: Measured Frequency Response of Uncoated Microchip in Unst.irred, Un-
used Oil at 70 ppm Water.

177
response of Fig. 5.7 with the rapidly rising gain and asymmetric phase. However, the
gain levels out below unity gain, indicating a considernhlt' low frf'ql1t>nry rnpn.cHanre.
The occurrence of both a surface conductivity and a low frequency capacitance is
inexplicable as it would be expected that the surface conductivity would shunt the
electrodes together at low frequencies.
After wetting the oil to 70 ppm, the response became that shown in Fig. 5.10. Here,
the effect of the large DC offset is seen in the upward shift of the entire gain curve -
resulting in low frequency gains greater than unity. The phase curve is significantly
broader than those shown in Figs. 5.5 and 5.6. Again, the deviation from a purely
bulk response is attributed to interfacial phenomena.
In another experiment, slight changes in the frequency response were observed
when attenuating the driven signal by as much as 20 dB (Fig. 5.11). The gain curves of
Fig. 5.11a are normalized to give the same gain at 1 Hz in Fig. 5.11b. It is believed that
the change in gain and phase (Fig. 5.11c) with field strength indicate the possibility
of field dependent conduction in the interfacial region. However, this behaviour may
also be the result of the change in offset (Fig. 5.11d) at each attenuated voltage level.

From the above observations it was felt that, except in the case of large DC offsets,
the high frequency gain of the uncoated microchip could give a reliable estimate of the
bulk permittivity of the oil. Comparisons between estimations from the microdielec-
trometer, the continuum model and a bridge measurement for various oil samples, all
at relatively high frequencies, are summarized in Table 5.3. The correlation between
all the measurements is within the accuracy of the estimation procedures.

5.4.2 Passivated Electrodes

A microchip was coated by Paratronix, Inc. [851 with a thin film (6.4 I1m) of parylent>
to remove the problem of moisture adsorption on the elect.rode interface. Parylene
is a trademark encompassing a family of linear polymers produced by pyrolyzing Di-
para-xylylene vapors and vacuum depositing the resulting monOlller (Fig. 5.12). The
actual film deposited was Parylene-C, consisting of a chloride atom on each benzene

178
 wnl.wa2-5
-15....-----------, 0....----------,

-25 -9

-35
.- Id
."
-45

-55

-65 -45
-3 -2-1 0 1 2 3 4 -3 -2-1 0 1 2 3 4
~q)
log Ereq
(b)
0 -l.B

i\
X
X
-30 -1.9 X
~
I) ~X
..,
0
.,c:;
I)
0
c... ~
s:a.. c... 0-
0 l<
-60 -2

-90 -2.1~~~-~~~-~~
-3 -2 -1 0 1 2 3 4 -3 -2 -1 0 1 2 3 4
(c.)
log Ereq
(d)
Figure 5.11: Effect of Attenuating Applied Potent.ial on .Measured Frequency Response
of Uncoated Microchip ill Unused Oil (X - full value, 6. - 55%, 0 - 24.5%, + - 14.5%,
and 0 - 7.2%).

179
Table 5.3: Bulk Permittivity Measurement Using An Un coated Microchip And A
Bridge.

Notes - ID # = 88025.39, HFG in air = -42.60 dB

measurement with microdielectrometer at 1 kHz, with bridge at 100 Hz

e' relative to eo F Im

Oil Condition Response in Oil Micromet Cont. Model Bridge

gain e' e' e'
-
unused -35.35 2.16 2.16 2.17
11
-35.18 2.20 2.19 2.17
oxidized -35.05 2.23 2.23 2.25
test facili ty -35.20 2.20 2.19 2.17

ring. There are some similarities in the growth mechanisms of Parylene films and films
produced by plasma polymerization [83, Section 5.4.2}.

The response of a parylene coated microchip in room air is shown in Fig. 5.13.
The film is very insulating but does exhibit some relaxation at the lower frequencies.
The extended decrease in gain and slight phase with increasing frequency indicates a
non-ohmic relaxation. For estimation of the complex bulk permittivity of the oil, the
effect of the dispersion in the parylene must be taken into account. This will be done
by first estimating the complex bulk permittivity of the parylene, from the response of
Fig. 5.13, at the specific frequency of interest. Then, assuming this dispersion remains
the same after immersion of the passivated microchip into the oil, the complex bulk
permittivity of the oil will be estimated from data taken at a frequency at. which the
complex bulk permittivity of the parylene is known (via the previous est.imat.ion).

For example, after placing the parylene coat.ed microchip in oxidized oil, t.he fl'e-
quency response of Fig. 5.14 was measured (dots). First, an estimate was made of t.he
complex bulk permittivity of the parylene bulk properties at 0.08 Hz from the response
in Fig. 5.13 (-34.23 dB and -1.73°). The parylene layer thickness was obtained from

180
 THE PAAYLENE PROCESS
TYPES OF
PARYLENE

PARYLENEN

PARYLENEC

{<H.-{5-... }
Cl
PARYLENED
MECH""'t:.foL
V~'''UM PUMP

Figure 5.12: Description of Parylene Coating Process and Molecular Structure (fro111
[85]).

181
 c3f'1
0

-5

-10

-15

-s:: -20 r - -
la
m -25

-30

-35 •• ....... • . . . A . . A, AA .........., .................. .... ..-.,

~ .....-...lJ1
-40

-45
-3 -2 -1 o 1 2 3 4

-10 .. '
.............. ........ ........
' ........ .......
, ' ,

"
-20
-30
u -40
"
~
..c:: -50
~

-60
-70

-00

-90
-3 -2 -1 1 o 2 3 4
log f'req
Figure 5.13: Measurec Frequency Responke of 6.7 /-Lffi Thick Parylene Coa.t.ed rvli-
crochip in Air.

182
the two parameter estimation described in Section 5.3. Using this estimated complex
bulk permittivity (3.11:c10- 11 F Im and 5.12:c10- 13 S/m), an estimat.e was made of
the complex bulk permittivity of the oxidized oil from the response of Fig. 5.14 at
0.08 Hz (-28.89 dB and -22.13°). The predicted frequency response (solid line in
Fig. 5.14) was obtained using the estimated values for the oil (e = 2.00:c10- 11 F Im,
(T = 2.13:c10- 11 S/m) and the estimated layer thickness but treating the parylene layer
as being a pure insulator of 2.7:c10- 11 F/m. The dispersion in the parylene was ne-
glected in Fig. 5.14 to reduce the amount of computation. The neglect of the dispersion
in the parylene accounts for much of the difference between the predicted and mea-
sured responses in Fig 5.14. A standard bridge measurement of the oil at 20 Hz yielded
values of 1.94:c10- 11 F Im and 1.56:c10- 11 S/m. Some of the difference between the
two methods of measurement, microdielectrometer versus bridge, may be accounted
for by the different frequencies at which the measurement was made, 0.08 Hz versus
20 Hz. Without verifying the lack of dispersion in the oil, a comparison based on an
ohmic model is invalid.
An attempt at examining the dispersion of the oil was made by repeating the
process described above at one frequency per decade. The results show the slight
dispersion in the parylene (Fig 5.15) and the apparent lack of dispersion in the oil
(Fig. 5.16). The solid line drawn in Fig. 5.16 represents the frequency dependence in e"
of an ohmic conductor. The deviation from an ohmic behaviour at the high frequencies
represents the sensitivity limit of the microdielectrometer, roughly a minimum loss
tangent of 0.005. The deviation in the oil complex bulk permittivity estimates below
0.1 Hz is puzzling. There is still significant phase information the response at these
frequencies, providing enough sensitivity in the estimates. Perhaps the assumption
that the dispersion in the parylene remains the same in air and in oil is incorrect. at
these frequencies. Or, perhaps charge separation effects occurring at t.he coating-oil
interface are the cause of the discrepan:cy (though one might expect the polarization,
and hence, the permittivity, to increase in this case).

The inadequacy of the uniform medium model (incorporated into the microdieiec-
trometer) for a coated microchip, is demonstrated in a plot of the estimated values

183
 c3f'32.. pc3f'32
0

-5

-10 El = 2.00.dO- ll F /m
al = 2.13:&10- 11 S/m
-15 ~ = 6.7 p.m
E2 = 2.70:&10- 11 F /m

.-~

.u
~
-20
-25 ........•
a2 =0

-30
-35

-40

-45
-3 ...;,2 -1 0 1 2 3 4

-10

-20
-30
CD -40
Ul
.u
,oS:;
~
-SO
-so
-70

-80

-90
--3 -2 -1 0 1 2 3 4
log f'req

Figure 5.14: Measured (.) and Predicted (-) Frequency Response of a 6.7 /l.m Thick
Coated Microchip in Unstirred, Oxidized Oil at 25°C.

184
 S~------~--------~------~--------,--------.

--a
~
...... 3
I
0
P"'4
..-
-..;'
2

1~-------+--------+--------4--------4-------~

OL-------~------~~------~--~--~------~
0.01 0.1 1 10 100 1000

1~------~--------~-------~--------~------~

.....
b 0.1 ~------~~------_4--------~--------_+--------~
P"'4

0.01L-------~------~--------~------~-------
0.01. 0.1 1 10 100 1000
log Ereq

Figure 5.15: Estimated Dispersion in Complex Bulk Permittivity of Parylene From

Measured Response of Fig. 5.13.

185
 5r--------r--------~------~--------,_------_.

4
-s
.......
......~ 3
I

-.,.."
....
0

1~--~.~--~--------~--------4---------+-------~

o~------~--------~--------~--------~--------~
0.01 0.1 1 10 100 1000

1000

100

a.......
~ 10 ~
...... ~
-
I
....

'"
0

1
-:
~
- 0.1
110

~
0

~
0.01 ~
~
0.001
0.01 0.1 1 10 100 lfI10
log ~req

Figure 5.16: Estimated Dispersion in Complex Bulk Permittivity of Oil From Measured
Response of Fig. 5.14 and using Estimated Dispersion of Parylene (Fig. 5.15).

186
from the microdielectrometer (Fig. 5.17). The estimated bulk permittivity using this
model is 2.3:1:10- 11 F Im, representing some average of the parylelle layer and the oil.
The straight line in Fig 5.17 represents the -1 slope that an ohmic material would
display in this log-log plot. At 1 Hz, the estimated conductivity is 9.17:1:10- 12 Slm,
again representing some average value.

Problems with the domination of interfacial effects, at the coating-oil interface,

can still arise at certain oil conditions. Typically, distorted frequency responses arise
with high oil flows and decreased moisture content. An example of this is shown
in Fig. 5.18 for a parylene coated microchip in stirred unused oil having a moisture
content of 8 ppm. The occurrence of a large, negative DC offset is a good indicator of
surface charging pro·blems at the interface.

Although much of the problem can be avoided by performing measurements in

unstirred oil, the best solution is to use a passivated macrochip. The advantage of the
longer spatial wavelength is that more of the oil is sampled, reducing the significance of
interfacial effects. These macro chips must also be encased in parylene as the alumina
that forms the insulating oxide layer will absorb moisture and develop a measurable
conductivity. The effect of the parylene coating on the response will also be small
due to the longer spatial wa.velength. These passivated macrochips then provide rea-
sonable estimates of the oil conductivity, as shown by Li [23, Section 4.2]. Results
of several oil measurements using passivated microchips and macrochips, and using a
standard bridge, are given in Table 5.4. Generally, the conductivities were estimated
using microchip and macrochip measurements in the range of 0.006-.2 Hz whereas
the bridge measurements were made at 20-100 Hz. The estimated bulk conductivities
using the passivated macro chip were obtained without including the dispersive nature
of the parylene, perhaps accounting for some of the discrepancies. Also, the poor esti-
mates of the bulk permittivity of the oil using the macrochip, due t.o its insensit.ivit.y,
are not presented. For the passivated microchip, oil conductivity estimates obt.ained
wbile accounting for the dispersion in the parylene, are presented. The ('diuiat.ed oil
bulk permittivities, even using high frequency data, were 10-50% lar~e\ tI m valut"s
obtained by eitber an uncoated microchip or the bridge (tbe values given Jll Table 5.3).

187
 6

I.

-& 4
'. •
"-
"'"
• I
.
•
2
•'" .... .'" .... ....... , , , . , I •• I

o
-2 -1 o 1 2 3

10
I
I

-~
"'"
..-4
..-4
1 ... ...... '\
~

I
~

-
..-4

N
~a

.,'. ..
CD 0.1
m
~.'. •
-0 • ," I.
I
I , I

0.01
-2 -1 o 1 2 3
log Ereq

Figure 5.17: Estimated Dispersion in Complex Bulk Permittivity of Oil From Measured
Response of Fig. 5.13 using Microlllet Values.

188
 c3f"6
0

-9

s: -18
-"
Id
-.
tft -Z?

-36
-
-.,
-~

-45
-3 -2 -1 o 1 2 3

o
".
,..-
~'"'W'II
~ ........
,..... .......
-30

-90
-3 -2 -1 o 1 2 3 4

-1

~
la)
tn -1.1 ". '.
"
CM

-
CM
o 1-
-- ,A._
.......
.... ,~

--
-1.2
-3 -2 -1 o 1 2 3 4
log f"req
Figure 5.18: Measured Frequency Response for a Parylene Coated Microchip in Stirr-
ed, Unused Oil at 8 ppm Water.

189
 Table 5.4: Bulk Conductivity Measurements With Passivated Electrodes.

Oil Condition Conductivity (S/m)

Estimated Measured
Microchip Macrochip Bridge

unused 4.52:z:10- 13 8:z:10- 13 2.3:z:10- 13

" 5.43:z:10- 13 4:z:10- 13 5.7:z:10- 13

oxidized '5.42:z:10- 12 5:z:10- 12 4:z:10- 12
test facility 1.94:z:10- 11 2:z:10- 11 2.55:z:1O- 11

The error here is due to DC offsets, on the order of -.7 V, that significantly affect the
gain curve, and therefore, the estimated bulk permittivity, while not seriously changing
the phase curve.

5.5 Complex Surface Conduction

Surface conduction, or conduction in regions so thin (relative to the spatial wavelength
used for probing) that they can be modeled as surface conduction, has been observed in
several instances. One example, presented in Fig. 5.7, represented surface conduction
between elt!ctrodes along the silicon dioxide-oil interface arising from moisture adsorp-
tion and ensuing ionization of either the moisture or other adsorbed species. Using an
ohmic conduction model, and representing the oil as a pure dielectric of permittivity
2:z:10- 11 F /m, a surface conductivity of 2.52:z:1O- 14 S was estimated from the response.
The predicted response using this ohmic model matches the measured response fairly
well (solid line in Fig. 5.7).
Another example of surface conduction between the elect.rodes was observed when
a thin film «100 A) of palladium, probably discontinuous, was vapOl' deposit.ed ont.o
the electrodes. In vacuum, the frequency response was that shown in Fig 5.19 (dots).
Modeling the thin film as an ohmic material, a complex surface permit.t.ivit.y was
estimated from the data at 1 Hz. The predicted response (solid line) using t.his complex
surface permittivity (e. = 2.15:z:1O- 11 F and (J'. = 1.12:z:10- 16 S) matches the phase

190
peak fai riy well but provides a poor match to the high frequency gain. The surface
permittivity is contributing an unwanted capacitive component at. high fl·equel1des.
Using only the surface conductivity and no surface permittivity, the predicted response
(crosses) matches the measured high frequency gain but overshoots some on the phase,
actually going positive at 1.5 Hz.

A second example of surface conduction due to a thin film coating, 350 A of vapor
deposited 8n02, is shown in Fig. 5.20. The frequency response was measured while in
room air at 35% RH .. Again a complex surface permittivity was estimated, this time
from data at 2.5 Hz. The predicted response (solid line) using this complex surface
permittivity (f. = 4.06:dO- 17 F and (1'. = 3.42~10-14 S) again matches the phase peak
but no' the high frequency gain. Using only the surface conductivity, the predicted
response (crosses) not only misses on the phase but also does not match the high
frequency gain. This indicates that either a surface permittivity is required at high
frequencies or that the microchip is not correctly calibrated.

Assuming the microchip is correctly calibrated, the complex surface permittivity

was estimated at several different frequencies (Fig. 5.21). It is useful to refer to a
plot of the gain-phase space being searched in this case (Fig. 5.22). For the low
surface permittivities of this example, there is a small window in the response within
which accurate estimates can be made. Estimating any quantity at the OY 'rshoot,
represented by the extreme kink in Fig 5.22 (in the neighborhood of -42 dB and
-15°), is very unreliable. In fact, no estimates were able to be made at 80 and 100 Hz
due to lack of convergence. Above 100 Hz, corresponding to phase between 0 and
-15°, there is no sensitivity to surface permittivities less than 10- 17 F. At the low
frequencies, the dispersion below 0.5 Hz is artifactual as the estimated permit.tivity
values reflect the fact that the gain does not actually go to unity at low frequencies.
Due to possible mismatches in the on-chip transistors, the gain levels out at. roughly -
.2 dB. This mimics the effect of a low frequency capacitance or blocking layer, hence t.he
estimated dispersion. This leaves a window from 0.5--100 Hz where reliable estimat.es
of the complex surface permittivity may be made. In the estimated dispersion shown
in Fig. 5.21, the log-log slope of f~ is -.95, indicating that an ohmic model for the

191
 pd3f'2,. ppd3f'2
Or-~~~~---------------------------------'

-11

-22

-•
s:
tft
-33

-44

-55~----~----~-----~----~----~-----~---~
-3 -2 -1 0 1 2 3 4

-30

-60
Il

"Id
..£:
p.
-90

-120

-150

-1BO~----~----~----~----~------~----~----~
-3 -2 -1 o 1 2 3 4
log ~req

Figure 5.19: Measured (.) and Predicted Frequency Response for a 100 AThick Pal-
ladium Coated Microchip in Air (e,o = 2.15:dO- 1i F and (7.0 = 1.12:c10- 15 S). For (-)
use Complex Surface Permittivity, for (X) use only Surface Conductivity.

192
 sno2l'4. p811D2f'4
°r-~~"~~~~'-------------------I

-11

-22
....
s:
/G
m
-33

-44

-55~----~-----L----~------~----~----~----~
-3 -2 -1 o 1 2 3 4

.- -30

-180~----~----~----~----~----~~----~--~
-3 -2 -1 o 1 2 3 4
log ~req
Figure 5.20: Measured (.) and Predicted FrequeI1("y Response for a 350 A
Thick Tin Oxide Coated Microchip ill Air at 35%RH. (fso = 4.06:elO- 1i F and
u.o = 3.42:el0- 14 S). For (-) use Complex Surface Permittivity, for (X) use only Surface
Conductivity.
193
surface conduction might be valid. However, there is also a dispersion in the surface
permittivity (log-log slope of -.52). This surface permittivity ml\y itu1irftt.p. pnlarizat.ion
occurring at blockiIlg electrodes (similar to the effect seen with the thin film humidity
sensors discussed in Section 3.3, case 4).

5.6 Estimation From the Frequency Response

An example is now given where a spatial distribution can be estimated from the
frequency response, using the techniques demonstrated above. A plasma deposited
bromobenzene coated microchip, Z24, was monitored in unstirred, oxidized oil that
was heated to 99° C. The measured frequency response is shown in Fig. 5.23 (dots).
The bromobenzene coating is thin enough, 2.4 p.m, to allow the electric fields to sample
the oil as well as the coatiIlg. There are clearly two relaxations occurring, one in the
coating and one in the bulk of the oil. Using Theorem 3.2, each relaxation can be
matched with the correct medium. Notice that between the relaxations there is a
plateau in gain and the phase goes back towards zero, similar to the example given in
Section 3.2.2 and shown in Fig. 3.10. If the higher frequency relaxation was occurring
in the coating, then it would be expected that the gain would go to unity (0 dB)
at frequencies below tIus relaxation as the electrodes would be intimately coupled.
However, there is a plateau in the gain at roughly -18 dB, indicating that the coating
is still relatively insulating at these frequencies. Thus, it is the oil that is relaxing
at the higher frequencies and the coating that relaxes at the lower frequencies. The
oil is highly conducting oil for two reasons: 1) the oil has been significantly degraded
by oxidation (heating i.u room air at 90°C for a week) and 2) the oil conductivity
varies inversely with viscosity, thus incre:asing at. highp.r t.empp.ratul'p.s. Thp. l'p.ason for
the insulating bromobenzene coating is that a. sample of heat.p.d oil, exposed t.o t.he
atmosphere, will be dried out and have a very low moisture cont.ent..

The coating permittivity and thickness have bep.n previously p.stimat.ed III Sec-
tion 5.3. The objective here is to obtain the complex bulk permittivities of the oil and
the coating. For simplicity, it will be assumed that both the oil and the bromobenzene

194
 100000

.
)1-' v
10000
• ~
,~
1000 I'

X ~

a-....
~
~
100

10
)(

~
'"'"~
~
~
"~
I
Q
P"1
"-" I,
1 X :.:: k)(
'"'"
~ .
o
" • "

0.1
,
0.01
0.001. 0.01 0.1 1 10 100 1000 10000

Best Fit Lines

log(E~o) = -.52Iogf - 15.481
log( E~o) = -.95 log f - 14.315

Figure 5.21: Estimated Dispersion in Complex Surface Permittivity using Measured

Response of Fig. 5.20. (X) - E~o, (.) - f.~o'

195
 phase (degrees)

o
N
0
u)
8 i
.-f .-f
e
.-f
~=I-,~I~~~I-,~I~~IC

C
.-f
I

~
I

&1I

~
I

Figure 5.22: Parameter Space for Complex Surface Permit.tivity Estimation of a Mi-
crochip in Air with f~o in units of 10- 18 F and f~o in units of 10- 17 F.

196
 z24f'52. pz24f'5Z
0

-5
El = l.99:dO- ll F /m
-10 0'1 = 7.56z10- 10 S/m
d2 = 2.2 pm
-15 E2 = 2.91.%10- 11 F /m
0'2 = 9.14z10- 1S S/m
-20
....,.,s::
tJ) -25
-30
•
-35

-40

-45
-3 -2 -1 0 1 2 3 4

-10

-20
-30
Gl -40
,.,fit
..c;
~
-so
-60

-70
-BO
-90
-3 -2 -1 0 1 2 3 4
log ~req

Figure 5.23: Measured (.) and Predicted (-) Frequency Response of 2.2 Inll Thick
PP-Brolllobenzene Coated Microchip in Unstirred, Oxidized Oil at 99°C.

197
coating are ohmic conductors. It will he shown in Section 8.2.3 that the bromobenzene
coating is actually dispersive. According to Theorem 3.2, bot.h of t.hese complex hulk
permittivities can be estimated from the frequency response. At frequencies higher
than 40 Hz, the contribution of the coating to the response is purely capacitive and
depends only on the previously estimated coating thickness and permittivity. Using
these estimated values, the complex bulk permittivity of the oil can he estimated from
a high frequency datapoint, here the response at 50 Hz was used. Now, given the
estimated complex bulk perm.ittivity of the oil and the estimated coating thickness, an
estimate can be made of the complex bulk permittivity of the coating from a low fre-
quency datapoint, here the response at 0.006 Hz was used. The estimated parameters
predict a frequency response that matches the measured response quite well (solid line
in Fig. 5.23). Notice that the estimated bulk permittivity for the coating, obtained
from the low frequency data in Fig. 5.23, matches the value obtained from the two
parameter estimation problem treated in Section 5.3.
Another example of estimating spatial distributions from the frequency response is
a measurement made of paper insulation using a passivated macrochip. This example
illustrates the estimation of both a spatial distribution and a dispersion in complex
bulk permittivity. The experiment, as performed and analyzed by Li [23, Section 5.2.2],
involved spring loading a passivated macrochip against an 84 fLm thick piece of paper
insulation and immersing the entire setup in transformer oil. The moisture content
of the oil was varied, significantly altering the paper moisture content due to the
small volume ratio of paper to oil. A typical response is shown in Fig. 5.24. As
the moisture content of the oil was decreased, the relaxation at higher frequencies
would shift down in frequency while retaining the same shape. This relaxation was
attributed to t.he paper insulation. The lower frequency relaxation I'f."lllainell const.ant.
during these changes and, in view of the insulat.ing nat.ure of t.he parylene, could not.
be attributed to the parylene. In addition, the plateau in gain in the neighbol'llOod of
1 Hz was lower than would be expected if attributed entirely to t.he capacit.ance of t.he
parylene coating. Thus: it was deduced that the paper was not in intimate cont.act.
with the macrochip and that a layer of oil existed between the parylene coating and

198
the paper. Given the permittivity of the oil, available from other measurements, and
the properties of t.he parylene, the thickness of the oil layer was f'st.imat.ed. From
low frequency measurements, where the paper was treated as a perfect conductor, the
complex bulk permittivity of the oil was estimated. At high frequencies, where the oil
was a perfect insulator, the complex bulk permittivity of the paper was estimated.

It was observed that the relaxation in the paper was not ohmic but, in fact, was
dispersive. The estimated dispersion, using the known parameters for describing the oil
and the parylene, is shown in Fig. 5.25. The real part of the complex permitt.brity (dots)
has the static, nondispersive dielectric constant (1.6:c1O- 11 F IIll) subtracted out and
the imaginary part (crosses) has the nondispersive, dc conductivity (2.3:c10- 11 S/m)
subtracted out. This dispersion is best represented by a power law dependence of the
general form

E• = B('JW )n-l (5.1)

This type of of dispersion is discussed in greater detail in Section 8.2. Pertinent

here is the fact that this dispersion was estimated from t.he frequency response of a
multilayered structure.

The layer thickness and complex bulk and surface permittivity estimation routines
of Sections 4.2 and 4.4 have been methodically applied to various experimental ob-
servations in this chapter. A rough mapping of estimation routine to experiment is
given: 1) layer thickness - Sections 4.2.1 and 5.2.1, 2) complex bulk permittivity -
Sections 4.2.2 and 5.4, 3) complex surface permittivity - Sections 4.2.2 and 5.5, and
4) layer thickness and permittivity - Sections 4.4 and 5.3. In addition, experiments
requiring the use of a combillat.ion of two or more of the est.imation rout.ines were given
in Section 5.6. A significant piece of informat.ion not. given with t.he est.imat.ed param-
eters is the accuracy of the estimates. Although some feeling for t.he sensitivit.y of t.he
estimation routines is given in the appropriate secHons in Chapter 4, no variances were
actually computed.

199
 Po r-/lene chip .f.... ~••• ' ~"<lper in oil
,., .• pcam:e",. ""c,,n ail;pp~::.l
.~I

-5

-1':

-1!· I
i I, I I, I I I
i ,!
,.... I I
iI I
..,
ol:
....,
-20

.11 !, ~
I;:
I
I
" -.! I
no
•
I
I
-!.C·
. . __ Cl
--. .... II-"~III
-35
<r,

-. 11-.....

-.LO

-45
I
-1

IDIjI fn:q
a 1st .. 2nd

Pary!ene chip A w/ paper In oil

17.2 ppm;c:t...,,,rcn cil:ppc:/.a.
0
Cl
;!l ~~Cb..
.~rj!JllfJ
. . . . . . .111
....- i
-1" rr¥ ''''I
CjlQQlol-
g'
-20
~':I
0
-~o

++.!
,.u
0
-40

-5. -se

-eo
-70

-eo
-90
-3 -1

' ' ' ' frwq

D 1st .. 2nd

Figure 5.24: Measured Frequency Response of Parylene Coated Macrochip wit.h Paper
ill Unstirred, Unused Oil at 28 ppm Water (from [23, p. 85]).

200
 Paper (ppo23): extrapolate el and e2

einf=1.6E-ll, sig~ (ohNdc)= 2.3E-l1

10000 i
I I
,
I Ii i
I
I
I
II I

& 1000 ,i I
"ra.
....
I
I I
I
I

I
....
I

....
W

- 1001' I /1

• t • • •• • • • • 1
l°I~----+I----~i-----rI----~T-.r-~I~--~--~
,
I
c..,

.-= , I
I
'X
v... X
I
+*f I
.
I
cu
I
....
·cu
1 I ....'
~ .,
:
;, I
I
X
• X I
I

;l.-
X ' I"X
;-~
I
0.1 1 '/
i
0.001 0.01 0.1 1 10 100 1000 10000
log w

Figure 5.25: Estimated Dispersion in Paper using Measured Response of Fig ..5.24. (.)
- e', (X) - f" (from [23, p. 88)).

201
Chapter 6

Moisture Sensing -. Concepts and

Design Strategy

At the outset, it is important to get a good understanding of the various terms used
in quantifying the amount of moisture in the gas, liquid and solid phases present
in a transformer. Section 6.1 provides definitions for these terms and describes the
many different equilibrium situations encountered. These concepts will be applied
towards understanding how a moisture sensor would respond to moisture in the oil
and providing a framework for relating on-line moisture measurements to parameters
of direct interest in a transformer, such as water solubility in oil and paper moisture
content. An example of using this framework to interpret actual data, obtaiued from
both a test facility and an operating power transformer, is presented in Section 6.2.
After demonstrating the usefulness of moisture monitoring, the design strategy for
developing a moisture sensor is presented in Section 6.3. In this context, the rationale
for coating the microchips using a plasma deposition process is also presented.

6.1 Equilibrium Concepts

There are two existing, commercially available met.hods for measuring moisture in oil.
As described in Section 1.3.1, the Karl Fischer method measures absolute m.oistu1'f
content whereas the dew point hygrometer measures relative saturation. A pertinent

202
question arises: what information about the state of the transformer can be deduced
from the relative saturation and t.he absolute moisture content?
Answering this question will require a reexamination of the various equilibria in a
system containing gas, oil and paper. There are gas-liquid (air-oil), gas-solid (air-
paper) and liquid-solid (oil-paper) interactions that must be understood, at least in
the case of equilibrium. Before that, some commonly used terms are defined. Relative
humidity (RH), of a gas, is the ratio of the partial pressure of the water vapor to the
vapor pressure of pure water at the same temperature [86, p. 279]. The solubility limit
is the maximum amount of moisture held by the gas at 100% RH and a given tem-
perature, corresponding to complete saturation. The relative saturation, of a liquid, is
defined as the mass of moisture dissolved in a liquid divided by the maximum amount
of moisture that could be dissolved at a given temperature. The solubility limit in
a liquid is again the maximum amount of moisture held by the liquid, now at 100%
relative saturation. Beyond this point, additional water will precipitate out, forming
free water. This is true for both the liquid and gas phase. Moisture content., or ab-
solute moisture content, when referring to a liquid, is the amount of water contained
in terms of parts per million (ppm) on a mass basis. This equals the product of the
relative saturation and the solubility limit.
The simplest situation is the gas-liquid equilibrium. Raoult's Law st.ates that,
at a given temperature, the vapor pressure of the solvent over a solution equals the
vapor pressure of the pure solvent lllultiplied by the mole fraction of the solvent in the
solution. At equilibrium, using Raoult's Law for an aqueous solution, the RH is equal
to the mole fraction. [86, p. 279]. It has been shown that the solubility of moisture in
oil is linearly proportional to the relative humidity [87]. Thus, in equilibrium, the RH
0/ the gas and the relative Jaturation 0/ the 0;/ mltsf be eqltal. If t.he volume of oil is
small compared to that of the gas (at st.andard temperat ure and pressure) t.hen t.he RH
of the gas will drive the oil to the appropriat.e relative saturat.ion. By small volume,
it is meant that the amount of moisture exchanged between the gas and the oil does
not significantly affect the RH of the gas. If the volume of oil is large compared t.o
that of the gas (again at STP), then the relative saturation of the oil will drive the gas

203
to the appropriate RH. Now it is the relative saturation of the oil that should remain
constant with the exchange of moisture. Thus, the requireOlf"nt. oC a volume ra.t.io of
roughly 1000:1 for the dew point hygrometer.
For the gas-solid equilibrium the picture is not so simple. The paper insulation
contains moisture in several different ways. There is the 'water of constitution', an
integral part of the cellulose fiber structure. In addition, there is: 1) chemically
absorbed water, 2) physically absorbed water and 3) capillary held water [88, p. 309].
Given this variety, all methods used to measure the moisture content of paper are
calibrated against the definitive measurement, which is a gravimetric one. Paper is
conditioned at a specific RH and temperature, weighed, dried in an oven at 105°C and
reweighed. The resultant percent change in weight is defined as 'percent water, dry
basis' or 'percent moisture content', depending on whether the change is taken from
the dry or wet paper weight. Early data on sorption for wood pulp obtained by leffries
[89] was plotted by Oomen [14] (see Fig. 6.1a). Notice the hysteresis in the curves due
to the different ways in which moisture is held in the paper. In general, the adsorpt.ion
curves have a sigmoidal shape. As the gas temperature is raised, at constant RH, the
paper moisture content decreases (Fig. 6.lb). At higher temperatures, the gas has a
greater affinity for moisture. Note that when conditioning the paper, the volume of
gas is large enough so that the amount of moisture exchanged with the paper does
not significantly alter the state of the gas. Thus, in equilibrium, the absolute moisture
content in paper is directly related to the gas RH and temperature.

Given that the RH of the gas is in equilibrium with, in one case, the relative
saturation of the oil, and in the other case, the moisture content of the paper, it is
reasonable to draw the conclusion, barring unforeseen oil-paper interactions, that the
relative saturation of the oil is related to the moisture cont.ent. of t.he paper. Thus,
the x-axis oC Fig. 6.lh can he replaced with the relative saturat.ion of t.he oil. Vsing
experimentally measured solubility limits oC a naphthenic oil, at. various t.emperat.ures,
and the Cact that the moisture content in oil is linearly proportional t.o the relative
saturation, Oomen Curther transfOl'ms the x-axis Crom relative ~aLllraLioll lo moisture
content (Fig 6.lc) [14]. An attempt to experimentally reproduce these curves would

204
 !I
i to
ca
0

• •
0

(q) I

e••
.
•
,
!/
: I

F1gur. 2. SorpUoD Curves for Wood

I Pulp fro_ J.ffr1.s' Data, Tabl. 1

I
Solub1lity Lt_it, PPM
20 )0 ,0 80 120

100
.... ~.oa.

Figure 4. Adsorpt10D Curves for Wood

Pulp ....d OD Figure 3.
Figure 6. Ho1sture !qu1l1br1u. Curves
for Paper-01l Syste_, Low Ho1sture Region
Cb)
(G)
(4'igure 6.1: Moisture a) Sorption and b) Adsorption Curves for Wood Pulp and c)

Moisture Equilibrium Curves for Paper-Oil System (from [14]).

205
be very difficult to do using the gravimetric method, though other techniques for
observing changes in paper moisture content, such as electrical meRflurt'mputs, may be
more feasible. The conclu3ion drawn by Oomen i" that there i" a direct relationship
between the moiJture content of the paper and the combination of the temperature and
the abJolute moiJture content of the oil. Again, this assumes that the solubility limit
is a known function of temperature.
There is a problem with this conclusion, however. As shown by Stannett [16],
there can be large changes in solubility with degradation. What happens when the
solubility limit of the oil changes due to degradation? Now solubility is a function of
both temperature and aging. It is claimed here that Fig. 6.1c is still valid, except for
the temperature markings. In other words, given the moisture content and solubility
limit of the oil (at the same, but unknown, temperature), the paper moisture content
can be determined. For example, oil sample 1, containing 5 ppm moisture and having
a solubility limit of 50 ppm, will be in equilibrium with paper having a moisture
content of 3%. Oil sample 2, containing 10 ppm moisture and having a solubility
limit of 80 ppm, will be in equilibrium with paper also having a moisture content of
3% . The samples could represent oil at different temperatures not having suffered
any degradation. According to Fig. 6.1c sample 1 would be at 20°0 while sample
2 would be at 30°0. However, it could also be that both samples were at the same
temperature but the oil has been significantly degraded so as to increase its solubility
limit at a given temperature. Either way, the curves of Fig. 6.1c represent the paper-oil
equilibrium. ThuJ, the moiJture content of the paper i" determined by the "olubility
limit of the oil and either th~ relative Jaturation or the abJolute moi"ture content of
the oil.
At any instant in time, if t.he relation bt't.wt'en temperature Rnd suluhility is known,
then knowing the temperature is as good as knowing the solubility limit.. This leads
one back to the conclusion of Oomen. Further, if t.his relation was known at. severa.l
instants in time, then the evolution of the relation gives some indicat.ion of t.he aging
of the oil. In R transformer, the paper drives the equilibrium, as it holds much more
moisture than the oil could ever dissolve. From the above conclusion, for a given paper

206
moisture content, the oil with a higher solubility at temperature To, will be at a higher
relative saturation (and higher absolute moist.ure content).
There is a major assumption in drawing the above conclusion. The assumption
is that the influence of temperature on the curves of Fig. 6.1 b is solely through the
different solubility limit associated with each isotherm. Pa.per holds moisture through
a variety of physical and chemical mechanisms. If the contribution to the moisture
content due to physical adsorption is not a function of the solubility in the vapor,
but is a function of temperature, and if this contribution is significant, then both the
temperature and the solubility are relevant parameters. A way to resolve this question
would be to measure the adsorption isotherms, varying temperature and RH, in air
and in another atmosphere that has a significantly different solubility dependence on
temperature. If the assumption is valid, then the second set of isotherms will be
the same family as the first set, with a different temperature associated with each
individual isotherm. If the assumption is invalid, then the temperature influences
the equilibrium in ways other than just through changes in the solubility limit. This
means that Fig. 6.1c is only valid as long as the solubility of the oil does not change. As
soon as degradation occurs, the adsorption isotherms used to construct Fig. 6.1c are
incorrect and no estimate of the paper moisture content can be made, even when using
the temperature, moisture content and solubility of the oil. As a practical matter, it
may be useful to treat the assumption as being reasonable, as the problem of hysteresis
has not been addressed at all and the required accuracy of the paper moisture content
is not very great.
Finally, the answer to the question posed must take into account the relation be-
tween moisture in the oil and problems associated with excess moisture - electrically
weakened paper insulation, degraded oil and formation of free wat.er. The physical
parameters associated with these problems are t.he moisture cont.ent. of t.he paper in-
sulation for the former problem and the solubility limit of t.he oil for t.he lat.t.er t.wo.
Relative saturation can indicate the potential for free water formation. However, it.
does not provide a meaSl1l'e of the solubility limit of the cH. Thus, it cannot. indicate oil
degradation (oxidation), a phenomenon accompanied by increased acidity and interfa-

207
cia! tension. The transformer decay products, such as sludge, are hygroscopic and will
cause an increase in the solubility of the oil at a given temperat.nre [nO]. Tht" previous
discussion has shown that unless the oil solubility remains constant (no degradation),
the relative saturation does not provide enough information for a determination of
the moisture content of the paper. However, a combination of relative saturation and
absolute moisture content will provide such information.

6.2 Moisture Distribution In A Transformer (Sta-

tus of Oil and Paper Insulation - An Example

As concluded in Section 6.1, both the absolute moisture content and the relative sat-
uration of the oil are required to determine the status of both the oil degradation (via
changes in the moisture solubility) and the moisture content of the paper insulation.
Using a moisture sensor of the type described in Chapter 7, an on-line means for con-
tinuously monitoring the absolute moisture content is possible. However, the method
for measuring relative saturation is a slow process due to the need to establish an
equilibrium between the gas and liquid phases.
An alternative approach is to use the information obtained during the regular load
cycling of the transformer. Variations in the load cause changes in the operating
temperature. Of particular importance are the changes in the winding temperature as
it is this temperature that drives the exchange of moisture between the paper and the
oil. If the thermal changes are slow compared to the kinetics of moisture absorption by
the oil, a. quasi-equilibrium analysis call be made of the moisture distribution between
the paper and the oil. It will be shown that the paramt"t.ers descrihing t.he condit.ion
of the paper and the oil can be extracted llsing this analysis. Experiment,al dat.a fr0111
both a test transformer and a 784 MVA operating t.ransformer will t.hen be illt.eq>ret.ed
using this analysis.

208
 6.2.1 Model

A Freundlich isotherm is used to represent the low and intermediate moist.ure ranges
of the adsorption curves of Fig. 6.1b [14]. Normally the moisture content of the paper
is constant as the paper is essentially an infinite reservoir of moisture for the oil.
Therefore, the relationship is written as

RS oil = g(T) (%mcpGper )n(T) (6.1)

where g(T) and n(T) incorporate the temperature dependence of the adsorption curves
of Fig. 6.1b, %mcpGper is the percentage moisture content of the paper, and the RH
has been replaced by the relative saturation. By definition, the relative saturation is
also related to the solubility limit (S)

IllCoil
RS oil = "8(1') (6.2)

Combining (6.1) and (6.2) and solving for the moisture content of the oil yields

mCoil = S(T) g(T) (%mcpGpert(T) (6.3)

I From experimental data, Oomen [141 has characterized the dependence of the solubility
limit as

Iog S = E.~1 -
[(2
-- (6.4)
T
As the oil temperature increases, the oil can hold more moisture. Taking the logarithm
of (6.3) and substituting in (6.4) yields

[(2
log(mcoi,) = [(1 - T + logg(T) + n(T)log(%mcpoper) (6.5)

This equation relates the temperature dependence of the absolut.e moist.ure cont.ent.
to parameters of direct interest.. [(1 and [(2 describe the solubilit.y of t.he oil and
will be a function of oil degradation. The effect of paper moisture cont.ent is also
very evident. Provided the temporal variations are slow enough to insure equilibrium

209
between the moisture in the paper and in the oil, this functional dependence can be
ohtainecl expeorimentally t.hrough t.he load cycling of the traOf~formf"r, f"limilla.t.ing t.he
need for a measurement of the relative saturation. Changes in intercept and/or slope
over time can be interpreted as changes in the fundamental parameters of the paper
insulation and/or the solubilit!' of tile oil.
The temperature required has deliberately been left ambiguous. It is clear that
the temperature of the windings is important as it reflects the heat driving moisture
out of the paper insulation. The ambient oil temperature is also important because
it determines the solubility of the oil. Ultimately, one or more temperatures may be
required for this type of analysis. For the present purpose of illustration, the ambient
oil temperature will be used.

6.2.2 Experimental Evidence

A fully instrumented 50 kVA pole-type distribution transformer was used for the
experiments. Loading was provided by an identical transformer. An oil circulation
system was installed to simulate the oil flow present in large power transformers. Seven
thermocouples were mounted at various locations inside the transformer. A moisture
sensor was mounted in the oil flow inlet, directly exposed to the flow ( ...... 9 gal/min)
but outside of the transformer tank.
Over a period of 7 months, oil samples were extracted from the test facility and
the moisture content was measured using a moisture analyzer. This data is plotted
versus inverse absolute temperature in the upper part of Fig. 6.2 along with a best
fit line. The other set of data, the boxed symbols, were obtained from Allegheny
Power System. They represent the moisture content of oil samples removed from a
;84 MVA power transformer located in the Harrison Power St.at.ion. The sampling
was done over a period of 2.5 years. The boxed plus signs represent. l'esult.s with one
oil in the transformer, the boxed asterisk, stars and crosses are results wit.h a second
oil. The oil change was due to a large measured static charge. The boxed ast.erisk
represents the status immediately after the oil change, and t.he boxed stars represent
later data with this second oil. The boxed crosses represent still later results obtained

210
after the second oil was degassed. The oil temperature used probably represents the oil
sample temperature when removed from the transformer. It is unknown whether t.he
moisture distribution was in steady state within the transformer when the sampling
was performed.

With these provisos in mind, a best fit line was also drawn for all the data for
the Harrison facility. Although there was a lot more scatter in this data, a complete
description of the moisture content versus temperature could be obtained only by
using the complete data set. The slopes of the two lines are very close, -1303 versus
-1261, perhaps reflecting the dominant influence of the temperature dependence of
the moisture solubility in the oil. There is a significant difference in the intercepts,
5.752 versus 4.851. TIns clearly reflects the fact that the paper insulation in the test
facility is much wetter than the operating ;Jwer transformer. This is not surprising
as no effort was made to dry out the paper insulation prior to energization of the
test facility. Using (6.5) an estimate can be made of the difference in paper moisture
content once the paper adsorption parameters have been sufficiently characterized.

6.3 Moisture Sensing in Oil - A Design Strategy

As shown in Chapter 5, moisture can alter the response of an uncoated microchip
either by surface adsorption or double layer (charge separation) effects, both of which
are not reproducible. Furthermore, it is known that the bulk conductivity of the oil
changes only by small amounts (20-30%) with changes solely in moisture content. For
example, used oil from a transformer exhibited a reduction in conductivity of 28%
(from 1.04:dO- ll to 7.41:z:10- 12 S/m) as the moisture content was reduced from 48
to 10 ppm. However, with significant degradation of t.he oil, perhaps resuit.ing in
an increase in ionizable species, the effect of moiRt.ure can be significant.ly increased.
Hence, changes in oil conduciivity are not a good indicat.ion of moisture cont.ent..
The approach taken here is to deposit a thin film of material onto t.he microeled.l'ode
structure a.nd monitor changes in its dielectric properties. There are several crit.eria
that must be satisfied by the deposited material. First and foremost, it must interact

211
 Moisture content vs. teMp.

pilot ~acility a ~ison PS (boxed)

100~------~c-----------------------------------~

-Si
~ 1*1+1
-
~
~

s::
CD
~
s::
0
u 10
CD
~

....tn=
....0
Si

- m
0

1~------------~~------------~~------------~
2.8 3.05 3.3 3.55
1/1' (units ~ 10E-3/Keivin)

I±l- Oil 1
tiJ- First Result with Oil 2
lil- Subsequent Results with Oil 2
fXf- Degassed Oil 2
Best Fit Lines
Pilot: log(mc) = .5.752 - 1~3
Harrison: log(mc) = 4.851 -14fll

Figure 6.2: Moisture Content Versus Inverse Absolute Temperature for Pilot Facilit.y
and Harrison Power Station.

212
with the moisture in the oil in such a way as to translate moisture content variations to
changes in its dielectric properties. Second, t.he sensitivity to moist.ure cont.ent. should
be reproducible and selective, that is, independent of other chemical changes occurring
in the oil. Third, the coating should be able to withstand exposure to hot. oil over
long periods of time and retain its physical integrity and moisture sensitivity. Fourth,
the coating should move the region of interest, i.e. where the conduction process takes
place, away from the oxide layer surface.
There are some additional constraints imposed by the availability of fabrication
resources and the commercially available form of the microchips. Polyimide is used
at some stage of the microfabrication process and the microchips are encapsulated in
a Kapton package, limiting the temperature range of the device to roughly 200°C.
This places a limitation on the types of deposition processes that can be considered.
Coating of the microchips before packaging would necessitat~ additional fabrication
steps, including masking to protect leads. Sputtering processes can be performed
at low temperatures but are expensive and are generally used to deposit. submicron
thickness films.
Plasma deposition is a low (ambient) temperature coating process that can produce
highly crosslillked polymer films of micron or submicron thickness. A description
of the plasma deposition process, the apparatus used and the operating experience
for coating the microchips can be found in Appendix C. A deposition apparatus,
having a capacitively coupled RF plasma design, was available on a full-time basis.
Operating parameters for producing good quality films had been previously established
for several materials, including brolllobenzene, by Benatar [91, p. 68]. Satisfying all
the constraints imposed by the microchip packaging, this process was used to deposit
bromobenzene for the purpose of developing a moist.ure sensor. The relevant., and
measurable, process parameters for the coat.ed chips discussed are given in Table 6.1.
Bromobenzene coated chips are identified by Z# where # denotes t.he cont.ing run
number. Estimated coating thickness and permit.tivity have been reproduced from
S~d·ion 5.3. Deposition rates were calculat.ed using the est.imated thickness.

All important issue that will be resolved in the next chapter is to determine whether

213
Table 6.1: Plasma Deposition Process Parameters for Bromobenzene Coated Mi-
crocllips

Flow Rate (F) '" 4 seem, Molecular Weight (M) = 157

W /FM '" 5:z:107 J /kg @ 25 Watts
ID Date RF Dep. Dep. Thick. Perm.
# Power Time Rate est. est.
(watts) (min) (A/min) (JLm) (:z:l0-1l F/m)
~.

Z3 3/5/86 150 38 421 1.6 2.44

Z4 3/7-11/86 50 1293 185 .
- 2.90
Z5 3/19/86 50 60 320 1.9 2.68
Z7 4/17/86 25 167 407 6.8 2.60
--
Z8 4/17/86 25 167 395 6.6 2.58
Z17 8/14/86 50 62 339 2.1 2,63
Z20 9/2/86 25 156 385 6.0 2.50
Z22 9/18/86 25 96 365 3.5 2.55
Z23 12/17/86 25 100 - - -
Z24 12/17/86 25 100 310 2.2 2.95

214
the coated chips respond to absolute moisture content, relative saturation or some
combination of the two. For the answer to tltis question, it. is necessary t.o continue
the discussion of liquid-solid equilibria. If the coating behaves as the paper does then,
based on the conclusions regarding the paper-oil system, it would be expected that
the moisture content of the coating would likewise depend on the solubility limit and
either the relative saturation or the absolute moisture content of the oil. Jumping
ahead slightly, the electrical response of the coated chip is related to the absolute
moisture content within the coating. Thus, the conclusion would be that the moisture
sensor does not measure either purely moisture content or relative saturation of the oil.
Instead, a set of curves similar to those of Fig. 6.1c would be needed to characterize
the response of the moisture sensor. However, the plasma deposited bromobenzene
coating may hold moisture in a very different manner than the paper. If it acts as a
semipermeable membrane, allowing moisture to enter the coating while blocking any
oil penetration, then the moisture content within the coating will be a function of the
osmotic pressure. Thus, the moisture content within the coating will depend on the
concentration, or absolute moisture content, in the oil. Measurements of t.he coat.ed
chip sensitivity with oil at different temperatures (Section 7.1.5), and degraded oil
having a different solubility limit (Section 7.1.6), indicate that t.he moisture sensors
do indeed respond to absolute moisture content in the oil.

215
 I
Chapter 7

Moisture Sensor - Experilllental

Results

The bulk of the experimental results using plasma polymerized bromobenzelle coated
chips is presented in this chapter. Section 7.1 focuses on results obtained in controlled
laboratory oil samples. First, it is demonstrated that the coated chip response to mois-
ture in the oil can be normalized into a master curve, resulting in a simple calibration
technique. Next, the moisture sensitivity of both fresh and aged coatings is examined.
The coated chip frequency response is shown to be linear with changes in the electric
field strength and ambient temperature. Experiments at variable oil temperature and
oil degradation validate the concept of all absolute moisture content sensor. The rapid
transient response to changes in moisture content is also presented. The problem with
large excursions in the DC offset of the device, associated with high flow rates of dry
oil across the film, is discussed. Finally, preliminary test results using the sensors
on-line, in the oil circulation loop of an instrumented test transformer, are presented
in Section 7.2.

216
7.1 Testing in Oil Samples

7.1.1 Experimental Procedure

Experiments were conducted in pyrex reaction kettles containing transformer oil and
sealed from the atmosphere (Fig. 7.1). A magnetic stirrer was used to maintain uni-
fOrInity within the kettle. The coated microchips were immersed in the fluid wHh
electrical leads extending through rubber stoppers. Increases in moisture content were
obtained by bubbling llitrogen through a beaker of tap water and then through the oil.
At room temperature, the solubility limit is roughly 40-50 ppm for unused oil 'out of
the barrel'. Drying of the oil was performed by either pulling a vacuum or bubbling
dry nitrogen through the oil. Direct measurements of the moisture content were ob-
tained by injecting oil samples into a Karl Fischer type moisture analyzer (Mitsubishi
Moisturemeter, Model CA-03). The moisture contents reported represent the average
of 4 to 6 samples from a particularly conditioned oil. Each sample volume of oil had at
least 10 p.g of water, where possible, to minimize errors due to the limits of precision
of the device (± 3 p.g). The lowest measurable limit of moisture in oil is 2-3 ppm.

A problem encountered when performing experiments at elevated temperatures was

one of maintaining significant moisture in the heated oil. As the oil-air system tried
to reach equilibrium, the heating of the oil tended to drive the moisture into the air,
resulting in very dry, hot oil. Attempts to increase the moisture content of the hot oil
by bubbling in wet nitrogen were not very successful. A variation on this approach was
to bubble nitrogen through boiling water and then through the oil. This worked too
well in that free water formed in the oil due to condensation at the lower temperature.
The solution was to place a piece of wet paper insulation in t.he hot. oil. In t.his case,
the mass transfer of moisture t.o the oil was greatly enhanced t.lll'ough t.he vehicle of
the paper insulation.

217
 TO
VACUUM
PUMP TO NITROGEN
T ,......-.- / (WET OR DRV)

GASKET - - t

REACTION
KETTLE -----.

HOT PLATE STIRRRER

Figure 7.1: Apparatus for Conducting Moisture Sensitivity Measurements with Coated
Microchip Moisture Sensors.

218
7.1.2 Master Curves For Coated Chips

The frequency response for a 6.8p.m thick coated chip, Z7, as the moisture content of
the oil was varied between 8 and 12 ppm, is shown in Fig. 7.2. A key feature of this
behaviour is the shape invariance of the frequency response. The coated chip response,
with moisture content as the external variable, fits Definition 3.1 and thus, a master
curve can be constructed. Using JP as the normalization parameter, the frequency at
which the phase peak occurs, a master curve for Z7 is shown in Fig. 7.3. A cubic spline
representation (supplied with the plotting routine) of' the discrete frequency data is
shown for clarity.

Two more examples of coated chip respOllses converted to master curves are for
a 0.5 p.m thkk coating, Z3 (Figs. 7.4 and 7.5) and for a 6.0 p.m thick coating, Z20
(Figs. 7.6 and 7.7). There are large differences in the range of moisture contents to
which these chips are sensitive - Fig 7.4 represents a change from 24 to 53 ppm whereas
Fig. 7.6 represents a change from 2 to 7.6 ppm. A discussion of these sensitivities is
given in the following section. Note also that the phase curves are broader, with
shallower peaks, than the curves of either Fig. 3.2 or Fig. 3.4 - particularly in the case
of Z7 and Z20, the thicker coated chips. Predicting the shape of these responses is a
difficult task and is left for Section 8.2.
The beauty of the master curves lies in their simplicity. Through application of
Theorem 3.1, the dependence on moisture content can be reduced to a change in one
parameter, Jp. Using a spline function to represent the master curve, any future gain-
phase data can easily be interpolated to estima.te JP and thus obtain the moisture
content of the oil. From now on, evaluation of the moisture sensor's performance
will be in terms of Jp. This technique is used ext.ensive-Iy for t.he on-line t.e-st. r~sults

presented in Section 7.2.

Furthermore, Theorem 3.1 lends credence t.o the belief t.hat the same physical
mechanism is governing the response of the coat.ing for all the moisture cout.ent. changes.
Tracking moisture content using JP is analogous t.o t.he usage of t.hin film humidit.y
sensors in air, where changes in RH are monitored through changes in resistivity of

219
 0

-5
z7f'12.13.15

~~
A
~·ti
e.
:, ~k
-10
AA
~
••-.. X
II -~

-15
A
--.;;ro •-.0', ~

>xX
-20
A
~t- •
....s::
.
'I'
IU
tn -25
6
II ••
....A
AA..
' ....
-30 ~~ 1\ A AA.A

-35
-40

-45
-3 -2 -1 o 1 2 3 4

-10
'~
),C
........

-A
.. ,
Y' :,
•'. 01"-X A
t£t
l:J.
t?CSp.

AA
-20 A

~
•, ~ )i(
~
,
-30 -, "'Y6 A
,
I) -40
A
'"'
'52
~
..
•
Cl')
IV A )<1.-
,J:; -SO A
s:a.. •
A
't A x·
-60

-70
'\
~~
At.
... ......
-..~~

-80

-90
-3 -2 -1 o 1 2 3 4
log ~req

Figure 7.2: Measured Responses of Fresh Moisture Sensor Z7, 6.8 Il.m PP-Brollloben-
zeue Coated Microchip, in Stirred, Unused Oil. Oil Moisture Content Varied From 12
to 8 ppm.

220
 z7112.13.15
0

-5

-10
-15

-2J.)
....cs
~

tn -25
-30
-35

-40

-45
-4 -3 -2 -1 0 1 2 3

0
-10
-2J.)

-30
I) -40
~
cs
..c:
~
-so
-60

-70

-80

-90
-4 -3 -2 -1 0 1 2 3
log Ereq
Figure 7.3: Master Curve for Fresh Moisture Sensor Z,' Produced by Normalizing
Responses of Fig. 7.2 to fp.

221
 z3f'2,,5,,13

-5
5

0 ...
- .,A....

••
..,. ~

"

~~
-10 ", v
, X
X
•
-s: -15
tG
~ -20
~A
~
•
,..
X
X
x
-25 ~
.a
-3(]
,~ '. ...... ~
/\1>1'!6. "AA X xm.m~ ~MUM'~ ...
-A.
..,.~ I-

-35

-40
-3 -2 -1 o 1 2 3 4

-90~----~------~----~------~----~------~----~
-3 -2 -1 o 1 2 3 4
log ~req
Figure 7.4: Measured Responses of Fresh Moisture Sensor Z3, 0.5 J.Llll PP-Bromoben-
zene Coated Microchip, in Stirred, Unused Oil. Oil Moisture Content Varied From 24
to 53 ppm.

222
 z3f'2.5.13
5
0

-5
-10
-15
....cs
$:

m -20

-25
-30
-35

-40
-4 -3 -2 -1 0 1 2 3

-10
-20
-3(1

IJ -40
"cs
~
~
-50

-60

-70

-80

-90
-4 -3 -2-1 0 1 2 3
log Ereq
Figure 7.5: Master Curve for Fresh Moisture Sensor Z3 Produced by Normalizing
Responses of Fig. 7.4 to Jp.

223
 z2lWS.12.15
0
LhA
....... '" .......... .... l"·' ,.,.. :~"
,..~

'~
-5 A .... .~
-10
A
1:1
- •
A.!\., •• /'::
" 'X
-15 ...
s:: -20
A
j
'.. \
....IU A
A4.
•• ~
m -25
-30
A
41.0. '..... if'

'Q~t:tM Al'vv\.AN\ A. N\M~~ IAM .... a;

-35

-4{)

-45
-3 -2 -1 o 1 2 3 4

0
..... ~
.......
/)\'V\, X~.

'Y.~ ~~---
t\
....
f" ')\{.
A

.'
.. ....
VI. ,AA

'

-10

-20
'. • A
AA
1l ' >Sx v •
-30
A
A
•• X ••
A 1:1
•• ,. X·
-~
A
G) -40
t,')
IU .0. .0..0. •• •\ J~
..s:: -50 •
....... •
A
~ .0.....
.0.1:1
-4.,.. A A
.0. ~ ~~
-60

-70

-80

-90
-3 -2 -1 o 1 2 3 4
log Freq
Figure 7.6: Measured Responses of Fresh Moisture Sensor Z20, 6.0 J-illl PP-Bromo-
benzene Coated Microchip, in Stirred, UIlused Oil. Oil Moisture Content Varied From
2 to 7.6 ppm.

224
 z2lW6.12.15
0

-5

-10

-15

-..,
s: -20
tft -25
-30
-35

-40

-45
-4 -3 -2 -1 0 1 2 3

-10

-20
-30
ID -40
..,Gtt
.,s::; -50
~

-60
-70

-80
-90
-4 -3 -2 -1 0 1 2 3
log f'req
Figure 7.7: Master Curve for Fresh Moisture Sensor Z20 Produced by Normalizing
Responses of Fig. 7.6 to Jp.

225
the thin film.

7.1.3 Coating Sensitivity to Moisture in Oil

Fresh Coatings

Measurements of the frequency response of moisture sensors in oil were t.aken soon
after (within several days) coating the microchips. Fresh coatings refer to the length
of time the coated microchips have been immersed in oil. Extracting fp from the
responses, the relationship between moisture content, as measured by the Mitsubishi
Moisturemeter, and fp, is plotted in Fig. 7.8 for several different sensors. A log-
log scale was chosen to show that all the coated chips exhibited roughly the same
sensitivity though the coating thickness varied from 1.6 p.m to 6.8 p.m. Using a power
law relation between fp and moisture content in the oil (mcoid,

(7.1 )

yields a linear relationship on a logarithmic scale,

(7.2)

Note that high sensitivity to moisture implies a relatively flat curve. This convention
was chosen because the moisture sensor will ultimately provide an on-line moisture
content based on the fp. A least squares linear best fit was used to relate log( mCoil) to
log(fp}. Best fit parameters for individual chips, and an overall average best fit for all
but Z3, are provided in Table 7.1. The lone exception, Z3, a 0.5 I'm coated chip, was
coated at a much higher RF power than all the others. This may have significantly
affected the structure and properties of t.he deposit.ed film (see Sect.ion 8.1). Though
the sensitivity of Z3 is comparable, there is an increased int.ercept, result.ing in a chip
responsive to high moisture contents within the operational frequency range of t.he
microdielectrometer. Differences in the process parameters can be seen in Table 6.1.
In general, it is observed that the freshly coated chips were highly sensitive to low
levels of moisture, with fp varying over many orders of magnitude for a change of

226
 Table 7.1: Best Fit Parameters for Freshly Coated Chips

log( mCDiI) = A + B log( Jp)

I Symbol I ID # I Thickness (p.m) A B

-
0 Z17 2.1 0.497 0.184
Z23 ",2 0.350 0.206
*
0 Z20 6.0 0.529 0.125

+ Z7 6.8 0.565 0.147

0 Z8 6.6 0.487 0.182
6. Z4 32 0.386 0.162
Average parameters of above coated chips - 0.489 0.159
X Z3 1.6 1.659 0.091
-_._---

roughly one order of magnitude in moisture content. There does 110t seem to be any
consistent pattern with regards to coating thickness.

Aged Coatings

After sitting in unstirred, unused oil at room temperature for 11 weeks, there was
a noticeable change in the sensitivity of the coated chips. Fig. 7.9a shows a decrease in
JP of roughly 2 decades for Z3. For this combination, low frequency JP's at high mois-
ture contents, no sensitivity curve can be obtained. Reducing the moisture content
would result in JP's below 0.005 Hz, the lower frequency limit of the microdielectrom-
eter. In addition, the relaxation of the oil starts influencing the response at these low
frequencies. The upper limit on JP is established by the solubility limit of the oil at
room temperature.

A slightly more detailed change is shown in Fig. 7.9b for Z23. Here t.he clerrease
in sensitivity occurred over a 10 week span while t.he coat.ed chip was immersed in
unused oil at 1'00111 temperature. It should be noted that, after the 10 weeks of aging,
the coated chip was removed from t.he initial oil sample and placed int.o a fresh, unused
oil sample for the second set of measurements.
More complete characterizations of the reduction in sensitivity are shown in Figs. 7.10

227
 log Moisture content vs. log fp

f'resh coa.ti ngs

100r---------------------------------------------.

- E
Po.

-
+'
Po.
X

=
IJ
+' B++
=
0
(.)
10
IJ
~
:s
+'
....0
~

E
tD

-0

1~----~------~------~------~----~------~
0.01 0.1 1 10 100 11DO 10000
log fp (Hz)

Figure 7.8: Compiled Moisture Sensitivity Data for Freshly Coated Moist.ure Sensors
Z3,4,7,8,17,20,23 ill Stirred, Unused Oil.

228
 65

- &

-....
~
~ 55
x
s:
x
....s:
Q)

0
45
u
Q)

"
....= 35
...."0
& [i]

25
0.01 0.1 1 10 100
log ~p CHz)
l~)
100
-&
~

-....
~

s:
...s:
Q)

0
u 10
Q)

"
....=
...."0
&

=
-0

1
1 10 100 1 ()(X)
log ~p CHz)
(b)
Figure 7.9: Moisture Sensitivit.y of Fresh and Aged Coated Microchip in St.irred, Ull~

used Oil for a) Z3 and b) Z23.

229
and 7.11 for Z7 and Z20, respectively. In both cases, the fresh coating response is rep-
resented by the boxed star symbol. Not.e that a reduction in sensitivit.y ('orresponds
to an iucrease ill slope due to the choice of the abscissa and ordiuate axes. For Z7
(Fig. 7.10), the first reduction in sensitivity (crosses) is again due to aging for 11
weeks in oil at room temperature, as with Z3. This first aged response cau be divided
into two r~gions, a very high sensitivity for low moisture contents « 6 ppm) and a
lower sensitivity, represented by the liue, for higher moisture contents. Further aging
occurred due to the combiuatioll of time and temperature. The total time elapsed
between the data represented by the crosses and the triangles was 16 weeks. Measure-
ments of the sensitivity were taken at elevated temperatures of 40°-60°0 for 6 weeks
of that time period. For the rest of the time, the chip was at room temperature in the
oil. There was a significant decrease in sensitivity, and change in intercept, due to this
aging. Between the data represented by the triangles and the circles, again 16 weeks
had passed. This time only 1 week was spent at an elevated temperature of 42°C. A
further decrease ill sensitivity was observed. After another 6 weeks at room tempera-
ture, yet another decrease was observed (asterisks). The inset in Fig. 7.10 shows that.
the sensitivity of the aged coating decreases to a linear relationship, approaching a
unity slope.

A similar decrease in sensitivity with aging, due to time and temperature, is shown
in Fig. 7.11 for Z20. In this case, 18 weeks passed between the fresh coating measure-
ments (boxed stars) and the first set of aged data (circles). For 5 weeks of this period
the chip was operated at 40°C. Between the next two sets of data (circle and plus)
the coated chip was operated in the oil circulation loop of an instrumented 50 k VA
test transformer for 12 weeks. Details of this experiment are described in Sect.ion 7.2.
The only information relevant to this sect.ion is that for those 12 weeks t.he coating
was subjected to temperature variations of 100-70 0e and oil flows of up t.o 8 gal/min ..
After removal from the test facility, the coated chip was placed in unused oil for (j

weeks at room temperature. The sensitivity decreased again, as represent.ed by t.he

squares.

A discussion of the aging phenomenon is put off until Section 8.3. However, it is

230
 100 0
0 0.1 0.2 0.5
'p (I.)

t
-lE ~
( A~/",

-
~
~
~
~

III
~
~
0
Cl 10
IU
s.. X
::s xX
~

....0
", X
lE
* X X

-
tn
0

1
0.001. 0.01 0.1 1 10 100 1000 10000
log f'p

Symbol Sampling Period Intercept (A) Slope (B)

5/3/86-5/9/86 .565 .147
*
X 7/30/86-9/26/86 .467 .385
b. 1/15/87-1/21/87 1.719 .721
0 5/12/87-5/18/87 2.398 .940
6/30/87-7/23/87 2.868 1.005
*
Figure 7.10: Change in Moisture Sensitivity with Aging of Moisture Sensor Zi ill
Stirred, Unused Oil. Inset Shows Linear Sensitivity of Aged Coating.

231
 100

-E
~
( Ae l"8

-~

+'
s:::
I)
+'
s:::
Q
U 10
C)
So.
:::::s
+'
....
'0
0

&:

-
t:r)
Q
~

1~----~------~------~------~----~------~
0.01 0.1 1 10 100 1(](J(J 10lXJ0
log fp (Hz)

Symbol Sam~ling Period Intercept (A} Slope (B)

9/9/86-9/21/86 .529 .125
*
0 1/26/87-1/27/87 .538 .333
+ 4/23/87-5/18/87 .815 .603
0 6/26/87-7/23/87 1.108 .667
•

Figure 7.11: Change ill Moisture Sensitivity with Aging of Moisture Sensor Z20 ill
Stirred, Ullused Oil.

232
observed here that this aging does not appear to be correlated with any degradation
in the mechanical integrity of f.he coat.ings, such as crarking or pef"ling. If I his type of
degradation had occurred, a significa.nt change in the capacitance of the coating would
result. This would be reflected in a change in the high frequency gain of the coated
microchips in oil. No such change was observed in the high frequency gain of fresh
and aged coatings.
Due to a combination of time and temperature, the sensitivities of the coatings,
as reflected in fp, changed monotonically over the 6 to 12 months of testing. There
were also small changes in the shape of the master curves, reflecting changes in the
dispersive nature of the coatings. The latter changes for chips Z3, Z7 and Z20 are shown
in Figs. 7.12, 7.13 and 7.14, respectively. In general, the phase seems to broaden, with
reductions ill the peak, with aging. The gain becomes more S-shaped, changing more
gradually with frequency as a result of the aging. Interpretation of these changes in
shape are put off until Section 8.2.3.

7.1.4 AC Electric Field Strength Dependence

A simple resistive divider was placed between the AC voltage source and the set of
driven electrodes 011 the microchip. By varying a potentiometer, an attenuation of
as much as 1000:1 in the driven voltage could be effected. The maximum practical
reduction that could be used was only one order of magnitude due to the reduced
signal to noise ratio. This 10:1 attenuation corresponded to a decrease of 20 dB in
the measured gain. (Gains are always computed relative to a 1 volt peak amplitude
signal in the microdielectrometer). A compariton between responses obtained using
attenuated and full strength signals for Z7 are shown ill Fig. 7.15. These reslllt.s Wf're
typical in that the responses of all coated chips with an att.enuat.ed signalma.t.ched the
respOIlse with full signal strength. Therefore, the mast.er curve was independent. of the
applied AC electric field strengt.h. Linearity in t.he frequency response of t.he coa.t.ed
microchips, with respect to the electric field, was obeyed here.

233
 z3i'2 .. 5 .. .1.3 .. 21. .. 22
5

0
-5

-10

-15
....s::cs
tD -20
-25
-30

-35
-40
-4 -3 -2 -1 0 1 2 3

0
-10

-20
-30
., -40
•"s:a.
.s:; -50

-60
-70
-80

-90
-4 -3 -2 -1 0 1 2 3
log Ereq
Figure 7.12: Change in Master Curve with Aging of Moisture Sensor Z3.

234
 0

-5

-1.0

-1.5

.-,.,
s:: -20
t» -25

-30

-35

-40
-45 I
-4 -3 -2 -1. o 1. 2 3

-1.0

-20
-30
4)
-40
,.,en
,J:;
Po.
-50

-60

-70

-80

-90
-4 -3 -2 -1. o 1. 2 3
log f'req
Figure 7.13: Clu:l.uge in Master Curve with Aging of Moisture Sensor Z7.

235
 0

-5
-10

-15
-20
....c:
IG
tn -25

-30
-35

-40
-45
-4 -3 -2 -1 0 1 2 3

-10

-20
-30
., -40
"
IG
.&:::
~
-so
-60

-70
-eo
-90
-4 -3 -2 -1 0 1 2 3
log Freq
Figure 7.14: Change in Master Curve with Aging of Moisiure Sensor Z20.

236
 z7£S3 .. 54
0

-5 t--

-10
~
\
~

-15 t--

s:: -20
....
\
~
~

tG
= -25 -
-30 ~
,~
-35 t--

-40 ~

-45 I I I I I J
-4 -3 -2 -1 0 1 2 3

-10

-20
-30
Gl -40
en
tG
~ -51]
a..
-60

-70

-80

-90
-4 -3 -2 -1 o 1 2 3
log £'req
Figure 7.15: Independence of Master Curve on Applied Electric Field Strengt.h for
Moisture Sensor Z7 in Stirred, Unused Oil at f'V 4 pplll Water. (X) - full voltage,
100%, (D) - attenuated to 9%

237
7.1.5 Temperature Dependence

The reproduction of the master curve was also observed with variations in the ambient
oil temperature. This invariance is shown in Fig. 7.16 for Z7 as the temperature
was varied from 27° to 60°0. There was actually a very slight deviation that can be
attributed to a slight decrease in ,the oil permittivity of 2% with increased temperat.ure.
Although the oil conductivity was expected to increase with temperature, mainly due
to the decrease ill viscosity, this change resulted in a relaxation at frequencies still
significantly below the relaxation in the coating. Thus, neither the change in oil
permittivity nor conductivity significantly altered the curve shape.

The effect of temperature on the sensitivity of the moisture sensors could be the
result of two effects. One effect would be the amount of moisture absorbed by the
coating, a physicochemical effect. The second would be the change in dielectric prop-
erties due to temperature, an electrical effect. If the coating behaves as the paper
does, then the amount of moisture within the coating would be a function of the sol-
ubility limit of the oil which, in turn, is a function of temperature. If the coating acts
as a membrane, then the amount of moisture absorbed would be independent of the
solubility and thus, temperature would only influence the response as related to the
electrical effects. Using this second model, the membrane, the resulting relationship
between fp, moisture content and temperature, is hypothesized to be

(7.3)

where EA is a thermal activation energy, k is the Boltzmann constant and T is the

absolute temperature. Paralleling (7.2), the relationship is expressed on a logarithmic
scale as

EA
log( mcoi,) = A + B log(fp) + 2.303kT
(7A)

This model predicts that sensitivity measurements, such as those shown in Figs. 7.8-
7.11, at various temperatures, should produce lines with the same slope (B) but dif-
ferent intercepts, due to the temperature contribution. An example of this is shown

238
 z7F54. 64. 77
0

-5

-10

-15

-s::
fG
m -25
-zo

-30

-35

-40

-45
-4 -3 -2 -1 0 1 2 3

-10
x
-zo
-30 \
u. -40
"
CS
.s::; \ I
V
s:L.
-SO
-60

-70
-80

-90
-4 -3 -2 -1 o 1 2 3
log Freer
Figure 7.16: Independence of Master Curve on Temperature for Moist.ure Sensor Z7
in Stirred, Unused Oil. (.) - 27°0, 2 ppm, (X) - 45°0, 21 ppm, and (D) - 58°C,
11.2 ppm.

239
for the aged coated chip Z7 (Fig. 7.17). These lines may be considered parallel within
t.he experimental error of t.he moisture analyzer. Using the difference ill t.he intercept.s
of the two lines (aint), an average slope (B) of 0.94, and

EA = (8.314:cl0- 3 )(aint) eV (7.5)

(96.48)B(1og e )d( ~)
yielded an activation energy of 0.66 eV. This value is close to the low temperature
activation energy of 0.7 eV obtained by Tyczkowski and Kryszewski in the study
of plasma polymerized organosilicon films [92]. It was not possible to obt.ain JP for
Z7 at higher temperatures for a significant range of moisture contents. The increased
conductivity of the heated oil, combined with the low JP's of Z7, resulted in the overlap,
in frequency, of relaxation in the oil and in the coating. The response would no longer
follow t.he master curve. It was difficult to obtain more experimental data as the other
chips, not having aged as much as Z7, would exh!bit ~ significant change in moisture
sensitivity during the week it took to generat,.: the several data points needed at an
elevated temperature.

7.1.6 Tests with Aged Oil

Further proof of the independence of the moisture sensor from variations in solubility
levels was obtained using an aged oil. This oil had been degraded by heating it at 90°0
for 10 days while exposed to the atmosphere. There was a very noticeable darkening
in color, even after 3 days. At room temperature, after bubbling wet nitrogen through
the oil for several hours, the moisture content was 43 ppm as compared to 33 ppm for
unused oil. The room temperature sensitivity of Z7 (Fig. 7.18a) and Z20 (Fig. 7.18b),
in degraded and unused oil, is very similar. There appears t.o he 80111(" hysteresis in
the behaviour of Z20. It was not possible to obt.ain JP using a mast.er curve for t.he
case of the aged Z7 in degraded oil having a llloisiure cont.ent lUuch less than 25 pp111.
Again the increase in conductivity of t.he oil, now due to degradation, combined with
the low JP's of Z7 to result in the overlap, in frequency, of relaxation in ~he oil and in
the coating. The response is no longer the master curve, as can be seen in Fig. 7.19

240
 100~------------------------------------------~

- I:
D.

-.,
~
D.

s::
."
~
s::
0
u
.,
S.
10

....en
:::::I

.'
....0 X
I:

-
D)
0

1~------------~---------------L------------~
O.DID. o.m 0.1 1
log f'p

(X) - 30°C, log(mc) = 2.84 + .991og(fp)

(e) - 48°C, log(mc) = 2.27 + .891og(fp}
EA '" 0.66 eV

Figure 7.17: Temperature Dependence of Moisture Sensitivit.y of Aged Moisture Sensor

Z7 in Stirred, Unused Oil.

241
for Z7 in degraded oil at 8.5 ppm. It would be possible to estimate JP by measuring
the complex bulk permittivity of the oil and including its effect when estimating the
properties of the bromobenzene coating from the frequency response of Fig. 7.19. This
approach is similar to the one taken for measuring the complex bulk permittivity of the
oil with a passivated microchip and including the dispersion in the parylene coating
(Section 5.4.2).

7.1.7 Transient Response

The transient response of Z7 to the addition or removal of moisture to the oil was
monitored. Shown in Fig. 7.20 is the transient response associated with the addition
of moisture through the bubbling in of wet nitrogen. A change in moist.ure content of
6 to 26 ppm was effected while monitoring the chip response at 1 kHz. Initially the
oil was dry, thus the low gain and phase. As moisture was added, the gain rapidly
increased and the phase passed through a peak and then returned to a small value.
This correspondR to tracing out the master curve, beginning with (J / Jp) > 1 and
ending with (J / Jp) < 1.
Aside from the rate at which the moisture content was being changed, there was
the question of how rapidly the change in moisture content reaching the coated chip.
Assuming a turbulent mixing model for the moisture distribution within the reaction
kettle, the rate at which moisture diffuses through the viscous laminar boundary layer
must be examilled. Using a boundary layer thickness on the order of 10 p.m [93], and a
diffusion coefficient of water in oil of IO-9 m 2/ s at room temperature [94], the diffusion
time

e2
'TD =D '" 0.1 sec (7.6)

is much smaller than the relevant time scale of the t.ransient response. Even if t.he
thickness ewas 10 times larger, 'TD would still be much shorter than the time scale of
the response. Thus, diffusion through the lamillar sublayer is estimated not to be the
rate limiting step.

242
 50

-lE
s:a.. 40 - +
+

-~
s:a..

s::: 30 ~ X 0
X +
Q)
~ +
~
0 )«-
u 0
Q)
20 ~

~
~ XX
~ 0
....0
fI)
10 - ~
lE
X
0 I I
0 0.03 0.06 0.09
fp (Hz)
(It)

- lE
s:a..
100

-
,s:a..

~
~
+
t
*
u *
~
~ +
0
u
+ '*"
10
~*+
III
~
~
~

....0
fI)

lE

=
-0

1
0.1' 1
I
10
log fp (Hz)

(h)
Figure 7.18: Moisture Sensitivity of Aged Moisture Sensors in Stirred, Unused and
Oxidized Oil. a) Z7 in Unused (X,o) and Oxidized (+) Oil, a.nd b) Z3 in Unused (*)
and Oxidized (+) Oil.

243
 z7£136
0
-5
-10

-=
IU
-15

-20
.
' • ...
' .. ....
tn ~25

-30
" .....-.-.' ........ ........ ........ ........
-35

-40
-45
-3 -2 -1 o 1 2 3 4

0
•
.......,..-_ - - .............
.. ............
'

.'
A'
-10

-20 ••
••
•
ID
-30

-40 ... •
........
fn
IU
~
~
-so
-60

-70

-80
-90
-3 -2 -1 o 1 2 3 4
log £req
Figure 7.19: Measured Response of Aged Moisture Sensor Z7 in Stirred, Oxidized Oil
at 8.5 pplll Water.

244
 z7t21
0
.. 11111111 •• 1•••••••••••••••••
....
-5
, I'

•
f--

-10 f-- ,
s: -15 -
.....,
m -20 -
I

-25 f-

-30
.............
f- •

-35~------~I--------~I------~I------~
o 10 20 30 40

0
...,
-10
_. ..~
'

• •" I
-20 - •
I) -30 - I

"cs
..t:;
I

~ -40 -
-50 - I

-60 -
•
-70 I I I
0 20 10 30 40
time (Min)
Figure 7.20: Transient Response at 1 kHz of Fresh Moisture Sensor Z7 in Unused Oil
Going From 6 to 26 ppm Water.

245
 Removal of moisture by bubbling in dry nitrogen resulted in a transient response on
a similar time scale as addition of moisture (Fig. 7.21). As before, the pat.h t.raced out
followed the master curve, corresponding to the decrease in JP with moisture content.
The two step response corresponds to the introduction of dry nitrogen twice, the first
time for 4 minutes, the second time for 2 minutes. Note that the speed of the response
involves the rates at which moisture is removed from both the oil and tl~e coating.

The time scale over which changes in the moisture content of the oil are reflected
in the detector's response is on the order of minutes. With the thin coatings involved,
7 p.m or less, this rapid response is to be expected.

7.1.8 DC Offset

Under normal operation, the DC offset displayed the following characteristics during a
frequency scan: 1) a negative value of several hundred millivolts at high frequencies and
2) a trallsition to a less negative value as the frequency varied from high to low (.0. V",
60 - 120 m V), accompallying the gaill and phase transition. These characteristics are
shown in Fig. 7.22a for the freshly coated Z20 in stirred, unused oil, having a moisture
cOlltellt of 4.4 ppm, and at room temperature. Two scallS were taken, the first using an
attenuated voltage (dots) and the secolld at full voltage (cross). The high frequellcy DC
offset of the attenuated scan begins at a much less negative value and, consequently,
there is a much smaller transition. The transition itself may be a characteristic of
the operation of the FETs as it was observed for coated chips operated in air. This
transition was also observed for uncoated chips embedded in 5 minute epoxy samples.
Thus, this behaviour constituted normal operation of a coated microchip.

Dramatic changes in the average DC offset of Z20 were ohservprl a.fter hent,j ng t he oil
to 40°C and cycling the moisture content from dry to wet and back to dry aga.in. The
average DC offset changed upon the second drying out, increasing from roughly -.2 V
to 2.9 V. Upon cooling back to room temperature, the average DC offset decreased
to '" 1.9 V. At a moisture content of 6.9 ppm, the transition was still observed,
again increasing roughly 50 mV as the frequency was decreased (Fig. 7.22b). This

246
 z7t23
0
• • 111111111 ....
-5 • I
I ..

-10 "
,
111111 11111'",

s: -15 ••
-"
m -20

-25

-30

-35~------~--------~------~--------~------~
o 20 40 60 00 100

O~--------------------------------------------,

-lOt-
• • • • 1...
-20t- •
•
I)
(,t
.u
-30 ~

•
...:;
Po. -40 ~ •
-so I--
••
"
-60 I-- Iltlll'III.....
....
-70 I I -f I
0 20 40 60 00 100
time (min)
Figure 7.21: Transient Response at 10 kHz of Fresh Moisture Sensor Z7 in Unused Oil
Going From 26 to 11 ppm Water.

247
 z20f'13.14
-0.1

....
~
~~ ... ~~~
, r-
.... ........ ........ ........
....

-0.15 v

X
X
X
-0.2
X
,

r ~~~. u ... w .. .>..

XVl'JIV
1'. •""YJ<"" I I'
X
X
><
IF-X
-0.25
-3 -2 -1 o 1 2 3 4
(~)
z20f46
.
1.9
'.'
... , •
•
•
1.88
,.a
IU
o
•
""'""'o
1.86 •

. . .. ........
• ••
• . ' ••....
t •
1.84
-3 -2 -1 o 1 2 3 4
log ~req
(h)
Figure 7.22: Measured Behaviour of DC Offset Voltage During a Frequency Scan for
Moisture Sensor Z20 in Stirred, Unused Oil a) at 4.4 ppm Water: (.) - Attenuated
Voltage to 8.8%, (X) - Full Voltage, 100% and b) at 6.9 ppm Water.

248
considerable average DC offset had no effect on the shape of the gain-phase response.
At most, a slight vertical shift in the gain might be expected due to a. possible error
introduced into the correlation analysis by the DC offset. After the change in Z20
described above, its DC offset was very reproducible during further cycling of the oil
temperature and moisture content while in the laboratory. Simultaneous conditioning
of Z7 in the same oil did not result in this shift of DC offset voltage. The DC offset
of Z7 was very reproducible throughout its entire lifetime except for the following
problem.
Occasionally, the DC offset was associated with changes in the shape of the gain-
phase response. This phenomenon occurred when the oil was so dry, and fp so low,
that the relaxation in the oil was beginning to be observed. An example of this is
the response of Z7 in stirred, unused oil, having a moisture content of 4.7 ppm and at
47°C (Fig. 7.23). The DC offset changed rapidly at low frequencies, accompanied by a
deviation from the master curve in the gain-phase response. This deviation prevents
accurate calculation of fp, making the moisture sensor unusable at this operating
condition.
At the time of this example, Z7 was an extremely aged coating, having operated
in oil of various temperatures for over a year. Consequently, the fp's were generally
much lower than when the coating was fresh, following the trend shown in Fig. 7.10.
In addition, increasing the oil temperature shifted the relaxation in the oil to a higher
frequency because of the increased oil conductivity. The combination of these two
effects resulted in the overlap of relaxation in the oil and in the coating, altering the
shape of the gain-phase response. When the oil temperature was lowered and/or the
moisture content increased, the separation between the two relaxations increased, re-
sulting in the master curve response. The moist.ure sensor wa.s t.hus rendered unusable
only until the condition of the oil changed. However, t.here is t.he risk of permanent.
damage if there is too much DC offset. The damage can occur to the FET transistor
connected to the floating electrode due to excessive DC fields across t.he channel.
The observed changes in DC offset were associated with static charge buildup at
the coating-oil interface. In many ways, the conditions leading to this effect resemble

249
 z7f'113
0

-9

....
~

IG
tJ)
-18

-~
'11 ....
, ' ...... - -. ....
-36

-45
-3 -2 -1 0 1 2 3 4

o
,.-- ....-
~
.". ". .,,-
~.

.f
-30
... ..",'

-90
-3 -2 -1 0 1 2 3 4

-0.2
•
"
~-O.4
III ...
t ..

~
CM \~
.......
.- ......... ................
CM
0-0.6 .~ ~....-.
IAao..
~
I ........
~

-0.8
-3 -2 -1 0 1 2 3 4
log f'req
Figure 7.23: Effect of DC Offset on Measured Response of Aged Moisture Sensor Z7
in Stirred, Unused Oil at 4.7 pplll Water and 47°C.

250
the conditions attributed to causing static electrification problems in transformers -
large volume flow rates of hot, dry oil. However, the prohlem (,8,n hE." 8.yoi,lt·rl hE."re hy
keeping the JP of the coating from drifting to such low frequencies that the coating
becomes as insulating af' the oil.

7.2 Testing in Pilot Facility

7.2.1 Experimental Results

The purpose of tIus section is to describe prelinunary tests of the moisture sensor,
and prinutive trend analysis schemes, in an on-line monitoring capacity. The sensor
is mounted in the oil circulation loop of the test transformer (see Section 6.2.2). As
the load on the transformer is altered, the temperature of the transformer changes. In
particular, the change in temperature at the WilldillgS tends to drive the exchange of
moisture between the paper insulation and the oil. Under normal operation, and at
quasi-equilibrium (as defined in Section 6.2.2), this redistribution in moisture displays
a reproducible dependence on temperature, as reflected by Fig. 6.2. This dependence
can be used to predict the moisture content using

(7.7)

The moisture sensor also gives the moisture content of the oil, using a simple
characterization such as (7.2). Taking the difference between moisture content as
measured by the moisture sensor and as predicted uuder normal operation produces
a residual. This residual may indicate that the transformer is deviating from normal
operation. Further analysis, along the lines discnssed in Sp.dion 6.2.1, tlllly indkat.e a
significant change in the moisture content of the paper insulat.ion.
The software implementation of the data acquisition and analysis required ior t.he
on-line monitoring can be found in Appendix E in the program wmod2.c. This pro-
gram, or water module, performed the following t.asks daily. Beginning at midnight,
a frequency scan of the moisture sensor was taken to re-establish the master curve.

251
A spline function was constructed to represent the master curve. After the frequency
scan, the moisture sensor response was sampled, at a fixed frequC"ncy, on a periodic
basis for the rest of the day. Typically, the sampling interval was once every 5 minutes.
Using the spline function and the sampled data, an interpolated value of JP was ob-
tained and two moisture contents were estimated, one using (7.2) and the other using
(7.7). The parameter values used in these equations were calculated from earlier sets
of data and were stored in data files accessible to wmod2.c. Finally, a residual was
computed from these two moisture contents.
The parameter values for (7.2) were calculated using data obtained from calibrating
a moisture sensor (ZI7, 2.1p.m) with the moisture analyzer. This was performed using
oil samples removed from the test facility (Fig. 7.24). Due to aging of the coating
with time and temperature, this calibration changes, as can be seen by the shifts in
Fig. 7.24. Initial moisture sensitivity data (crosses) represents calibration samples
taken over the first 7 weeks of testing. The next set of data (clots) was taken over a
four week period beginning 4 1/2 weeks after the end of the previous sampling period.
The shift is not the same as those observed earlier, in Figs. 7.8-11. The difference lies
in the fact that the earlier changes were recorded under isothermal conditions whereas,
in the test facility, the change in temperature is driving the change in moisture content..
Therefore, there are actually two components causing the change in Jp, as described
by (7.4). Note that although there was a significant change in moisture sensitivity, the
master curve remains the same before placement in the transformer and after operating
for roughly 14 weeks in the oil circulation loop (Fig. 7.25).

Earlier testing of Z20 in the test facility also displayed the same shift in sensitivity
over time (Fig. 7.26). Again, it was difficult to sort out the moisture sensitivity
component from the temperature component.. The significant decrease in moist.ure
sensitivity, after being exercised in the transformer for 12 weeks, is shown in t.he
difference between the circle and plus data in Fig. i.11. The constancy of t.he mast.er
curve before and after operating in the test facility is evidenced by the small changes in
Fig. 7.14. This sensor was removed from the facility due to intermittent malfunctions.
As discussed in Section 7.1.8, Z20 had a large DC offset voltage prior to installation

252
 100~----------------------~--~---------------,

-
~
s::
IJ
~
s::
o
u
.,
N
::s
~

'0
...."
2
m
- o

10~--------------------~----------------------~
100 1 (I)() 1()()(J]
log f'p (Hz)

(X) - log(mc) = 0.19 + 0.54Iog(f,,) (5/1/87-6/19/87)

(e) -log(mc) = 0.64+0.441og(f,,) (7/22/87-8/17/87)

Figure 7.24: Moisture Sensitivity of Aging Moist.ure Sensor Zl i in Oil Cil'l'tllat.ion

Loop of 50 kVA Test Transformer.

253
 z17F22.F~4
0
x~
-5
-10

-15
,
-20
....s::
IU
m -25
-30
-35
-40
'aw"" 11 , '""'1111111'

-45
-3 4

-10

-20
-30
u -40
•IU
.s= -50
~

-GO
-70
-00
-90
-3 -2 -1 o 1 2 3 4
log Freq
Figure 7.25: Measured Master Curves of Moisture Sensor Z17 Prior to Inst.allation
(0) and 14 Weeks After Installation (X) in Oil Circulation Loop of 50 kVA Test
Transfornter.

254
In the test facility. It is believed that the extreme flow rates the moisture sensor
was exposed to (",8 gal/min), in combination wHh t.ilis Jargf" nc nfT~f't., rf"sl1Heti in
damage to the FET transistor connected to the floating electrode. After removal,
further testing in gently circulating oil was possible, producing the rest of the results
shown in Fig. 7.11. The intermittent malfunction was associated with static charge
buildup because the problem could be alleviated by disconnecting and reconnecting
either t.he sensor to the interface box, or the cable connecting the interface box to the
microdielectrometer.
An example of the on-line monitoring using Z17 is shown in Fig. 7.27. The gain,
phase, offset and temperature (at the sensor) are sampled at 400 Hz every 5 minutes
over the span of one day with midnight corresponding to t = o. At this period in time,
the tra.nsformer excitation current was being cycled on and off daily. The amplitude
of the excitation current for this particular scan was 210 A. The sensor temperature
(lower right), reflecting the ambient oil temperature flowing by the sensor, is decreasing
prior to the application of the excitation current because the transformer is still cooling
off from the previous day's cycling. At a little after t = 500 min. the temperature
begius to increase, reflecting the application of the excitation current. In the evening
the current is turned off, resulting in a delayed decrease in temperature. As is evident
from the plot, the transformer is not in thermal equilibrium with this daily cycling.

The gain and phase in Fig. 7.27 (upper left and upper right respectively) appear
to mimic the time variation in the t.emperature. The response is being monitored
at 400 Hz, which is on the low frequency side of the phase peak (J < Jp). As the
temperature decreases, the oil becomes drier, shifting the response to lower frequencies
(Jp decreases). Therefore, both the gain and phase become more negative initially.
With the temperature increase the oil absorbs more moist.ure from (.he paper result.illg
in an increasing Jp, gain and phase. The offset (lower left.), for some reasoll. also
appears to follow the change in temperature.
In the plot of JP (upper left of Fig. 7.28), t.he discret.eness observed is a result of t.he
tolerance used in the interpolation routine. The predict.ed moisture content using JP
in (7.2), with A=.19 and B=.54, is shown in Fig. 7.28 (upper right). These parameter

255
 Moisture content vs. log Ep

220 in pilot Eacility

100.--------------------------------------------,
o
-&

-=
~
a..
a..

ID
~
.,

=
0
(,)

ID
s..
:s
~

....
Gt
0
E
tn

-0

10~------------~--------------~------------~
1 10 100 1000
log fp (Hz)

(X) -log(mc) = 0.92 + O.3291og(f,) (2/11/87-2/14/87)

(Cl) -log(mc) = 1.01 + 0.3531og(f,) (2/15/87-3/16/87)
(e) -log(mc) = 1.18 + 0.3271og(f,) (3/18/87-4/10/87)

Figure 7.26: Moisture Sensitivity of Aging Moisture Sensor Z20 in Oil Circula.tion
Loop of 50 kVA Test Transformer.

25G
 tw0630
-12. a ,...----..,.-----r----, -49 ,..-----r------,----,

-14.3 -52

-c
td
tn
Q)
en
td
..c;
s:a..

-16.3 -55

-lB.3 '----.:...._---'------' -58 L - -_ _- ' - -_ _ _ _ ~ ..___I

o 500 lOOl 1500 0 500 1000 1500

-0. 45 L..-.._~~I---~-----I 4B L--_--.:::II~--~------I

o 1(1]() 1500 0 500 100l 1500
time (..an) time (Min)
Figure 7.27: On-Line Monitoring of Gain, Phase, Offset and Temperature of Moist.ure
Sensor Z17 at 400 Hz in Oil Circulation Loop of 50 k VA Test Transformer. Current
Loading Cycled Between 0 and 210 A.

257
values are taken from the crosses in Fig. 7.24 - prior to further aging of the coating.
The expected moisture content using (7.7), the sensor temperature, and C=5.443 aId
D=-1201, is also shown in Fig. 7.28 (lower left) along with the resulting residual (lower
right). The values for C and D were obtained from the data, shown in Fig. 6.2, that
was available at that time. It is observed that the residuals are very small, indicating
normal changes in the moisture content with temperature.

Another set of data, taken 5 weeks later, is shown in Figs. 7.29 and 7.30. Here,
the transformer had been operating at an excitation current of 150 A long enough to
be in steady state. At a little after t = 750 min. the excitation current was increased
to 200 A. The glitch at t = 500 min. was a result of turning the transformer off so
as to remove oil samples for gas and moisture analysis. Again the measured response,
monitored at 400 Hz, follows the changes in temperature very closely. However the
residuals are very large this time. The parameter values used in (7.2) were different,
A=.74 and 8=.38, reflecting the aging of the moisture sensor (only 4 of the 7 dots
in Fig. 7.24 were known at the time of the experiment resulting in a different best fit
line), but the parameter values used ill (7.7) were the same as before. The change in
residual accompanying the change in excitation current is slightly larger than before,
6 ppm vs. 4 ppm, but there is an overall shift of 8-10 ppm in the average residual.
This shift most likely reflects changes in the sensitivity parameters of the moisture
sensor due to aging. Oil samples from the transformer, measured by the moisture
analyzer, produced moisture content values that correlated with the normal operation
of the transformer, as indicated by the upper set of datapoints in Fig. 6.2.

7.2.2 Discussion

There are many flaws in this implementation. The problem wit.h t.he moist.ure sensor is
the unstable nature of the coatings, casting some doubt 011 the correct. paramet.er val ues
for the moisture sensor beyond, roughly, a one or two week time period. On a mont.hly
time scale, recalibrat.ion of the present moisture sensor would be a necessit.y. Due to
the double valued nature of the phase there is also a problem interpolating JP when the

258
 t.w0630
1560 ,----...,...-------r-----, BD __--......--.--r----,76
• •

••
- 1260
••• ••• - 70 64
f+-
CD

-
N E a
:::c:
•• ••
--
~

~
c...
- ~

Cl
I:
'1:S

( "')

960 60 52

660 L..~IIIIIIMJ---L-_.....J 50 L---=:!~L---L---l40

o 500 1000 1500 0 1000 1500

B1 1.3 70

,
- 71 - I:
0.3 64

-
a~
a.. - ~
~
f+-
CD
a

--
58 '1:S
0;;-0.7
~ ='
...s ( "')
Cl ....UJ
I: 61 .,
s..-1.7 52

51 L--_~. . . .- - - . . I . . . - - - - - J -2. 7 1...-.--....1.-......----'------'46

o 1000 15lXJ 0 500 1()(x) 1500
time (min) time (Min)
Figure 7.28: jp, Measured and Predicted Moisture Content, and Residual Using Data
of Fig. 7.27 and Equilibrium Moisture Model.

259
 twOB04
-15. 3 r---.,..-----r----:~ -54

-16
-55 ~

~
••cs
m
-16.7
Q,)
C4
cs -56
,.s:;
V
,j

s:Lt
-17.4

j
-1B.1
-57
- ~

-18.8 '----------'-------' -58

o 500 1000 1500 0 500 1000 1500

-0.256 ,..----.,..-----r------, 6B r-----~--___,-___==_...,

66 t-----t------1.----i
-0.261~--4_---_+~-~

Q,)

; 64 ~---+---~----f
~ 't ~
Q,) CS
"0.266 t-----+-----.+----I So.
~ Q,)

~ ~ 6Z t-----+---._+----I
Q,)
~

-0.271
60
,
-0.276 '--_ _ -&;;...._ _--L.._ 58 L..-.---.l-_ _ _"--_ _--J
_- - - J

o 500 1000. 1500 0 1000 1500

time (..un) time (..un)
Figure 7.29: On-Line Monitoring of Gain, Phase, Offset and Temperature of Moist.ure
Sensor Z17 at 400 Hz in Oil Circulation Loop of 50 k VA Test Transformer. Current
Loading Increased From 150 to 200 A.

260
 twOB04
•
B40

•••- 69

67
68

66

••
-
N
740
- S 65 64
c+
CD

-
::c
s:a.. -
~
~ (.) (A)
B
e,:s

--
n

,
CM u 63 62
s
640

61 60

540 L -_ _~_ _~_ _----I 59 ' -_ _"'*-_ _--L._ _---l5B

o 1000 1500 0 1000 1500

B5 ,-----r------r---., -6 .-------r------r------,69

-
00
- &
~
-8 66

S
-- ~
c+
CD

-
Q,. B
~

~
u
S
7S
~
IG

....en
~

Cl)
-10 63
--
e,:s

~ -12 60
70

65 L -_ _........_ _- - ' -_ _----I -14 L--._ _-'--_ _---1~~~..J57

o 500 1000 1500 0 500 1.000 1500

time (Min) time (Min)
Figure 7.30: fp, Measured and Predicted Moisture Content, and Residual Using Data
of Fig. 7.27 and Equilibrium Moisture Model.

261
actual frequency is close to Ip. This problem can be avoided by automatically shifting
the monitoring frequency so as to keep the ratio (f / Ip) const.ant.. As t.he moisture
content increases, Ip increases, thus the monitoring frequency I would also increase.
The opposite behaviour would occur as the moisture content decreased.
As to the modeling issues, the simplified models used to compute the moisture
content need to be updated. Characterizatbn of the moisture sensor, presently using
(7.2), should be expanded to include the temperature change recorded by the sensor,
as in (7.4). It remains to be justified whether a quasi-equilibrium model for.describing
the moisture distribution in the transformer, using (7.7), is valid. A more complex
model, including lagged temperatures, may be required to adequately simulate the
dynamics driving the moisture exchange. It also is more reasonable to be using the
winding temperature in (7.7) instead of the oil temperature as measured by the mois-
ture sensor. However, increasing ~he complexity of these models should be governed
by a determination of how accurate the estimates need to be.
Nevertheless, this work represents an important step in taking real-time moisture
data from a transformer and attempting to correlate the observed behaviour wit.h basic
models of "normal" operation.

262
Chapter 8

Moisture Sensor - Interpretation of

Results

The experimental results presented in Chapter 7 established the sensitivity of the

coatings to absolute moisture content in the oil. The issues in this chapter are un-
derstanding the physical mechanism accounting for this sensitivity, interpreting the
dielectric response of the coatings and examining the problem of coating stability.
Much of the discussion in Sections B.1 and B.3 draws upon related work reported in
the literature. Film morphology and its connection with the moisture permeation and
electrical parameters of the film is presented in Section 8.1. Section 8.2 demonstrates
the utility of the continuum model, and associated parameter estimation routines, in
judging the validity of various electrical conduction models as applied to the plasma
deposited bromobenzene material. Following the rejection of several different models
relying on an ohmic conductivity, the dispersive nature of the coatings are revealed.
Examples from the literature of other materials displaying similar dispersive prop-
erties, having responses that can be normalized with respect t.o changes in RH and
temperature, are given. The change in moisture sensitivity with aging of t.he coat.ings
is examined in Section 8.3. Examples are taken from the literature of techniques used
during, and after, the plasma deposition process to improve stability.

263
8.1 Film Morphology - Effect On Moisture Perme-
ation and Electrical Properties
The st.:ucture of plasma deposited thin films is a strong function of the operating
conditions during deposition, as discussed in Appendix C. Coatings can vary from
rigid to oily to powdery films, with combinations pI)ssible as well. SEM photographs
of the witness strips for both thick and thin coatings indicate the existence of a biphasic
structure, with particles embedded in a homogeneous film matrix. Fig. 8.1a shows the
surface of the witness strip coated alongside chip Z22, 3.5 p,m thick coating. The size
scale is indicated by the bar at the lower right, witb the marking of 20 ILm underneath.
A magnified view of a section of tltis coating is presented in Fig. 8.1b, the bar width
at the lower right indicating 4 p,m. Particles both spheroidal and box-like in shape
appear to be both on the surface of, and embedded within a rigid film. The film is
judged to be rigid, not oily, based on the retention of its physical and electrical integrity
after physical handling and immersion in hot oil. Particle dimensions range from 2 p,m
down to submicron. A large particle of about 4 p,m, located in the background and
slightly off center to the lower left, appears to be a grouping of smaller, 2 p,m sized
particles.
Though no samples were made of the cross-section, it is believed that the particles
exist throughout the film. This belief is based on 1) the observation of particles, besides
the ones clearly on the surface, that appear to be embedded within the film and 2)
similar observations in the literature. For example, by pyrolyzing plasma polymerized
hexamethylcyclotrisilazane (HMCTSN) films for various durations, Grebowicz, et a1.
[95] revealed that the particles, in his case, were distributed throughout the entire film
and were not just at the upper surface. The particle diameters in t.hese Hl\ICTSN
films ranged from 1000-2500 A.
The effect of the particle density on moisture sensitivity may be illf~l'l'ed by first.
comparing the SEM photograph of Fig. 8.2a to Fig. 8.1a. In Fig. 8.2a the witness
strip examined had been coated alongside chip Z3, 0.6 lUll thick coating. The scaling
in Fig. 8.2a is given by the bar at the bottom of the picture, its length representing

264
 • •
• 4." ...................
~
"4 ••
....••... ....
•

... ""...
~.~
.
... •

•
...
•
~ ••• :.
... •
~ ~ ~ ........ :
!IO'.. .,'
•.
•
~.: ~•••A~•••••
~ ~
••• ' .. '.. • •• _ • •4 . " ........ .

.. .•
..- .... : ' .. III • ..... • •• " .. " ' . ..
" ... e

'., .....
4 ... ",.." .... .•
... " "'IIII!I?"... .. • • ". • .... .
. •.•. ", .. ,.;."." I. .,.... e "-e_"'.~
A ~,
-.. .,.Ate .,..~.
\~
".".,
"... ".. • .......
. "

• ~ .. .. -.."'- ·'''e: ' . ~.. . e"

.

•
,... ........
o _.~

..... .• ...-.'. ...'..."......•.......,.....•

... .". •• ~..

.. "'..'
.
~~. ~
".'
..

4-
~.
.•
.....
.....
•• '. .,.... . -.a:,' ......... ,
. ~
.
-er.".. .
•

.
•
.~'
-a tt. . . .
.. __ ... '.....

• . . . .-
:
",'

. . . .'
~ •

•
•
•

'''"

,,~.
..... .
." ..... .~~.
e. ~"",
. . ' . ~"_ .. ...& ~
A
1It," • • ~ -~ .... • 00 068 S
- · v

FigI\J'I' K.I: SE\I \li .... Clgraph Cif \\'illll':-':-' SI rip ('Clilk.! l\ICllIg:-.idt· \llIi~II1I1' St'II:-'lIr Z'.!.'.!.,

iI) Se"l" I:, 111111 :!tI/lIlI (ti:,tI\), I,) Sealt· I:, 1111111 11111 (T.!tlO\ I.

26J
100 P.lll. A magnified view of this coating is shown in Fig. 8.2b, the primary object in
this picture being one of the more spheroidal particles of Fig 8.2a. TTf'ft" , tht> Rcaling
represented by the bar at the bottom is 1 p.m. The particle density in Fig. 8.2a is an
order of magnitude less than the density in Fig. 8.la. The main fabrication difference,
besides the coating thickness, is the RF power level during deposition - 150 W for Z3
vs. 25 W for Z22. Experimentally, the moisture sensitivity of Z3 was much lower than
all the other coated chips (Fig. 7.8).

For completeness, SEM photographs of a witness strip coated alongside chip Z7

is also shown in Fig. 8.3. The surface of this coating looks drastically different from
the previous photographs. Large particles (5-20 p.m in width) occupy a large part of
the surface area. These particles appear to consist of two or more smaller particles
agglomerated together (see Fig. 8.3a). This coating was produced by plasma deposit-
ing for a longer period of time than the previously shown coatings (see Table 6.1).
Consequently, this coating was relatively thick, 6-7 p.m. The moisture sensitivity of
Z7, and Z20, was studied extensively in Chapter 7 and found to be just as sensitive as
other chips having thinner coatings, such as Z22 (3.5 p.m).

It is believed that the large particles on the coating surface, Figs. 8.la and 8.3a,
are not relevant to the moisture sensitivity or the electrical conduction process. Pri-
marily, the justification for this conclusion is that the shape of the frequency responses
obtained with thinner coated chips lacking these large surface particles, such as Z22,
are very similar in appearance to the responses of Z7 and Z20. Also, given the isolated
nature of these large particles, their contribution to the net conductivity of the ma-
terial would be highly dependent on the electrical properties of the film between the
particles. This scenario is not a satisfactory one, as will be shown in Section 8.2.
Instead, it is believed that it is the particles embedded within t.he film t.hat. are
.
important. These particles can be seen in the background in Fig. 8.:3a but. are more
clearly seen in the magnified view of Fig. 8.3b. They vary in size from roughly 2 I'lll t.o
submicron. It is thought that these embedded particles increase t,he surface area and
porosity within the film, resulting in greater moisture permeation, and thus, higher
moisture sensitivity. This conclusion is borne out by the lmver moisture sensitivity

266
il I S('id,· ~,'.! I1IIII 11111 ,,111 ,''''1).\ .1. 1" ~,.d,· 11 111111 I "111 1 ~'II.IItIt'.\'

'.!.c 7
l,'iJ.!.lIl"· ~,:~: SI·:\I \linllJ.!.l'ilph lit \\'it II!''' " SII ip ('II;d,·d ,\I""l.!."id" \1"1,1111'" .... '·"""1 I.~

iI) Sl"ill,' 111111111 1111/111 IIIIIHI.\ 1.1" SCilk 11; 111111 11/111111,.111111\ I,

2htl
of Z3, having a coating with low partiCle density. This conclusion also implies that
moisture sensitivity is not dependent on coating thickness, pe,' .. e, hut on t.he voidage
and porosity of the coating. The moisture sensitivity plot of Fig 7.8, coupled with
the observation of similar high particle density coatings for several other moisture
sensors of varying coating thickness, demonstrates this lack of dependence on coating
thickness.
The relationship between particle density and moisture permeation was also studied
by Wertheimer, et al. [96]. Using hexamethyldisiloxane (HMDSO) and hexamethyldis-
ilazane (HMDSN) monomers, the particle density was varied by changing the substrate
temperature during deposition. Increasing the temperature from 25°C to 400°C re-
sulted in the almost complete elimination of particles. This effect was reproduced by
increasing the power level, as observed in this thesis also. In addition, the increase
in either power level or substrate temperature resulted in an increasingly inorganic
film (decr.easing carbon content relative to the silicon, hydrogen and oxygen content.s).
Finally, these changes were accompanied by an order of magnitude decrease in the
moisture permeation coefficient from 5 to .5:dO- 8 standard cm3 -cm/cm2-s . The in-
crease in film density was only roughly 10%. The decrease in moisture permeation was
attributed to the reduced ability of the films to swell as they became more inorganic
and more rigid.
The effect of film morphology on moisture permeation was also studied via mea-
surements of the 1055 tangent and capacitance of thin, plasma polymerized HMDSO
films by Ramu, et al. [97]. Again, increasing the substrate temperature (during de-
position) to 400°C resulted in little, if any particle formation. The dispersion in the
loss tangent was shown for samples prepared at various substrate temperatures as
a function of RH. Measurements were made in the frequency range of 0.01-100 lIz.
For a given sample, the peak loss shifted to higher frequencies wHit increasing RH.
The magnitude of the loss appeared to decrease by a fact.or of three as t.he subst.rat.e
temperature was increased from 25° to 250°C but the sensitivity, or change in JP as a
function of RH, increased with substl'ate temperature. The change in capacitance was
a linear function of RH with a slope that decreased by an order of magnitude with

269
the change in substrate temperature. Again, the picture that emerged was one of re-
gions of lower density t.hat were surrounding the microspheres, allowing p('rment.ion of
moisture. Furthermore, this moisture permeation was decreased by reducing particle
formation at higher substrate temperatures.
An important observation to be made here is that the two studies characterizing
film behaviour were performed in air. No study was ever found describing behaviour
of these films in liquids, much less transformer oil.

8.2 Dielectric Response of Plasma Deposited Bro ..

mobenzene Films

8.2.1 Attempts to Model the Master Curves

The following sections discuss the many attempts that were made to predict the mas-
ter curve response of the coated microchip moisture sensors using standard models of
dielectric phenomena. After distinguishing between interfacial versus bulk 105s mech-
anisms, and dipolar versus charge carrier loss mechanisms, various models utilizing
ohmic conduction (using uniform or heterogeneous medi urns) are considered and re-
jected. Finally, a mathematical fit to the measured response is obtained using a power
law dispersion relation, possibly indicating the physical mechanism of a hopping con-
ductivity.

Interfacial Phenomena

The first possibility considered is that the form of the frequency response was dom-
inated by interfacial phenomena at. the electro(les due to polariz<l.tion, dH\rp;p inject ion
sl1ch as that occurring at a Schottky barrier, or eledrochemical rea.ctions cat.alyzed
by the adsorbed moisture. This interp,zetation was sOlilewhat appealing in view of
the facts that there seemed to be no discernible dependence of moist.ure sensit.ivity
on coating thickness and that many of the observed master curves displayed similar
S-shaped gain functions and broad phase peaks.

270
 A typical model of the coupling between the electrodes through the coating would
have a parallel RC element representing the interfacial impedancf" behveen t.he metal
electrode and the bulk in series with another parallel RC network represent.ing the bulk
coating. For the occurrence of a mast.er cun-e with a significant interfacial impedance
one of two scenarios must take place: 1) the interfacial impedance dominates or 2)

.
both impedances change proportionately with external variables. The first scenario
implies that the bulk impedance is insignificant relative to the int.erfacial impedanc.e.
This is inconsistent with the experimental observation that for J ~ JP the response
is low gain and zero phase - indicat.ing either an insulating bulk or unreasonably
large electrode capacitances. As for the second scenario, the experimental evidence
demonstrated independence in curve shape with respect to AC electric field strength
and temperature variations, casting considerable doubt on the affect of the interfacial
impedance. The behaviour of charge injection and electrochemical reactions t.ends to
exhibit very non-linear dependencies on temperature and field strength. Significant
variations in these parameters would induce large changes in the interfacial impedance
without changing the volume impedance. This would not satisfy Theorem 3.1 and
thus not reproduce the master curve response observed experimentally.
Regarding the possibility of electrode polarization, the consequence of gains, at
low frequency, approaching unity (0 dB) implies little or no polarization. Actually,
the gain at low frequency usually approaches -.2 and not 0 dB. However, this small
deviation call arise from experimental errors due to: 1) the slight resistive drop in
the coating, 2) the slight mismatch in the on-chip FETs (see Section 1.4.2), and 3)
significant DC offset volt ages (see Sections 5.4.1 and 7.1.8). For these cases, given
the small deviations from unity gain, it is judged that the experimental errors are
more significant than electrode polarization effects. Examples of significant. eledl'Ode
polarization when using the microchip sensor can be found in t.he wOl'k of Day. et. a!.
[99] in the monitoring of epoxy resin cures.
Thus, it is believed that the observed response was not due t.o any significant.
contribution by interfacial impedances.

Bulk Phenomena

271
 Dipolar Lou

The forUl of dielectric dispersion due to a siUlple Debye dipole loss model is

E' = f + Eat - Ego

00 1 + (w1f
(8.1)
E" = E • -
-I"-~wr
1 + (wr)
E

where Eoo is the high frequency optical dielectric constant, E,t is the static dielectric
constant and T is the dielectric relaxation time [98, p. 181. At low frequencies E' = E.t

and E" -+ 0, resulting in a purely capacitive coupling. For a coating exhibiting only
dipolar losses, this would imply a low frequency gain significantly different from unity
and a phase of zero. However, the low frequency gain did approach unity (wit.hin the
aforementioned experimental errors), iuconsistent with the dipolar loss wodel.
This low frequency unity gain is consistent with losses due to charged carrier con-
duction. With charge carriers, as w -+ 0, the resistive coating essentially 'shorts' the
electrodes, providing an almost lossless coupling.

Uniform Ohmic Conduction

The simplest approach was to model the coating as a uniform ohmic conductor. A
comparison between the predicted response for a 6.8 P.lll thick ohmic conductor in oil
and a measured master curve for Z7 (fresh) is shown in Fig. 8.4 and for a 2.2 p.m thick
coated chip (Z24) in Fig. 8.5. Note that the exact value of the conductivity used for
the predicted response is irrelevant here as the shape of the curve is independent of
conductivity. The predicted gain rises more steeply as frequency is decreased, and the
phase peak is narrower and much larger than the experimental data. Hypothesizing
that the coating thickness was nonuniform does not appear to help much. It. is expeded
that the respOIlse would then reflect an average thickness, perhaps producing a more
S-shaped gain and a broadened phase. However, judging from Fig. 3.4, it. is not
expected that a variation of even ±1 pm would produce sufficient alterat.ion in t.he
response. Moreover, measurements of the witness strips, having t.he same surface area
as the electrodes, indicated variations in thickness of only ± 0.2 pm at most.

272
 z7~x.12.13.15
0

-5 el = 2%10- 11 F /m
0'1 = 10- 12 S/m
-10 d z = 6.8 JJm
e2 = 2.60%10- 11 F /m
-15 0'2 = 10-1» S/m

-s:: -20
ft:I
tn -25

-30
-35

-40

-45
-3 -2 -1 0 1 2 3 4

-10

-20
-30
u -40
"
ft:I
~
~
-50

-60

-70
-80

-90
-3 -2 -1 0 1 2 3 4
log ~JreCI

Figure 8.4: Measured Master Curve for Moisture Sensor Z7 Compared to Predict.ed
Response using Bulk Ohmic Conduction Model.

273
 z24t'51, pz24t'x
0

-5
El = 2%10- 11 F/m
-10 t11= 10-12 S/m
~ = 2.2 pm
-15 E2 = 2.95%10- 11 F /m
t12 = 10-e S/m

-s: -20
IG
en -25

-30
-35

-40
-45
-3 -2 -1 o 1 2 3 4

°r---~----------------------~~~~~~6~A~

-20

-40
.,
•"
,..c;
At
-60

-80

-100

-12OL-----~----~-----~----~-----L----~----~
-3 -2 -1 o 1 2 3 4
log ~req
Figure 8.5: Measured Master Curve for Moisture Sensor Z24 Compared t.o Predict.ed
Response using Bulk Ohmic Conduction Model.

274
 Surface and Bulk Ohmic Conduction

The next thought was that perhaps the response reflected the influence of a surface
conduction along the oil-coating interface, possibly as a result of a charge separation
process at this interface. This view would still necessitate a bulk conduction occurring
in the coating or else the low frequency gain would be less than unity. The response
produced by such a model does not match the experimental data for Z7 (Fig. 8.6). 'rhis
thought model fails even more so when trying to predict the response for a thin coating
such as Z24 (Fig. 8.7). Additionally, in accordance with Theorem 3.1, both the surface
and bulk conductivities would have to scale in the same manner with moisture content
in order to reproduce a master curve. Given the different conduction mechanisms that
could occur within the bulk versus along the surface, the likelihood of such a scenario
appears to be very remote.

Spatial Distribution in Bulk Ohmic Conductivity

A more complex hypothesis introduced the possibility of heterogeneity in t.he coat-

ing, perhaps as a result of the deposition process, in the form of a spatial distribution
of dielectric properties. The conduction process was still envisaged as being ohmic but
there would be a spatial distribution in ohmic conductivities. Various dist.ributions
were used, both exponential and linear in nature. Success was obtained in modelillg
the response of Z7 using a linear distribution in ohmic conductivity, with an increase of
two orders of magnitude proceeding from the electrode interface to the oil, with a con-
stant permittivity (Fig. 8.8). This response was obtained by discretizing the coating
(and spatial distribution) into 90 individual, homogeneous layers. In view of Theorem
3.1, in order to obtain a master curve with variations in moisture content, the spatial
distribution of the conductivity would have to remain C'Ollst.ant. wHit moisture. ,yltile
this scenario is plausible, there is nothing ill the deposi1.ion proC'{-ss t.o indicat.e why t.he
linear dist.ribution in properties would arise. Furthermore, no dist.ribut.ion in dielectriC'
properties, using ohmic conductivities, could aC'count for the response in t.he t.hinner
coatings. Any ohmic conductivity in the thin coat.ings gave rise 1.0 very ld.rgt: aud
narrow phase peaks as typified by Fig. 8.9. Thus, a spatial distribution in dielectric

275
 z7EI9.psur61.63
0

-5

-10

-15

-s: -20
~
In -25
Et = 2,dO-11 F Im
-30 0'1 = 10-12 S/m
-35
d2 = 6.81lm
E2 = 2.60:z:10- 11 F Im
0'2 = S:z:10- 10 S/m
-40
-45
-4 -3 -2 -1 o 1 2 3

0
-10

-20
-30
u -40
"
ftS
~
s:t..
-50

-60

-70
-00
-90
-4 -3 -2 -1 0 1 2 3
log Ereq
Figure 8.6: Measured Master Curve for Moisture Sensor Z7 Compared to Predicted
Response using Bulk and Surface Ohmic Conduction Model.

276
 0

-5
-10

-15

-20
....s::
iU
en -25
-30
El Im
= 22:10- 11 F
0'1 = 10- 12 S/m
d 2 = 2.2 I'm
-35 E2 = 2.95.dO- l l F Im
0'2 = 52:10- 10 S/m
-40

-45
-4 -3 -2 -1 o 1 2 3

0r-------~~~--------------~~~~

-20

-40
ID

"
CS
.s:;
s:a..
-60

-80

-100

-120~----~----~----~~----~----~----~----~
-4 -3 "-2 -1 o 1 2 3
log Preq
Figure 8.7: Measured Master Curve for Moisture Sensor Z24 Compared to Predicted
Response using Bulk and Surface Ohmic Conduction Model.

277
properties displaying ohmic conduction WCI.'! ruled out.

lvla:r:well- Wagner Polarization

Another possibility, motivated by SEM photographs discussed earlier (Figs. 8.1-

8.3), WCI.'! the existence of a dispersion in the dielectric properties of the coatings due to
the heterogeneity of a biphasic material. This usually goes by the name of Maxwell-
Wa.gner interfa.cial polariza.tion, the charging and discharging of interfaces within a
heterogeneous material. Using a. Lorentz sphere model for the dispersion of small
spheres embedded in the coating matrix, the effective coa.ting permittivity would be

f
•
- f
* [(2Er + e;) -
2q,(E;: - E;)] ( 8.2)
ef f - 1 (2 Ei + ei) + q, (fi - Ei)

f*
1 bulk phase complex permittivity El - j(J't/w
e*2 particle phase complex permittivity E2 - j(J'2/W
particle volume fraction
An a.ttempt was ma.de to use this type of dispersion and fit the response for Z7
(Fig. 8.10). At a volume fraction of 0.4 (a fairly high a.mount) the magnitude of the
main phase peak can be brought down by attributing a. relaxa.tion to the particle phase
occurring at higher frequencies. However, this second relaxation was never observed
and the gain and phase curves are still not well matched in shape. TIllS type of
dispersion cannot be used to explain the experimental responses.

In the work of Ramu, et al. [97] mentioned earlier, a Maxwell-Wagner dispersion

relation was used to explain the observed frequency shifts in the peak loss tangent
with RH. For this case, the assumptions were q, ~ 1, El ~ f2 and (J'1 ~ (J'2. Although
the exact formula given to predict the frequency of maximulU loss tangent

t - q,(J'2 . (8.3 )
m - 21r( El + q, E2 )
does not a.ppear to come from t.he above dispersive relat.ion, t.he form of the dispersion
still requires two loss peaks, one associated with the particle phase and one at a lower
frequency associated with the bulk medium. There should then be some intermediate

278
 z7f'19, Inz7
0

-5

-10
-15

-20
.....s::
~
CO) -25 El = 2:dO- l l F /m
UI = 10- 12 S/m
-30 d z = 6,8 pm
E2 = 2,60:z:10- 11 F /m
-35
( :z: In 10~;o)
-40 uz(:z:) = 10-loe a; 10 S/m

-45
-4 -3 -2 -1 0 1 2 3

-10

-20
-30
CU -40
"~
..s::; -50
~

-60

-70
-80

-90
-4 -3 -2 -1 0 1 2 3
log f'req
Figure 8,8: Measured Master Curve for Moisture Sensor Z7 Compared t.o Predicted
Response using Linear Distribution in Bulk Ohmic Conduction.

279
 z24f'51.1nz24
0

-5
El = 2:1:10- 11 F /m
0'1= 10-12 S/m
~ = 2.2 I'm
-10 E2 = 2.95z10- 11 F /m
lO-8
-15 0'2(Z) = 10- loect In 10- to ) S/m

-20
....s:
/G
en -25

-30
-35

-40
-45
-3 -2 -1 o 1 2 3 4

-15

-30

., -45
"
/G
.s:;
s:I4 --60

-75

-90

-105
-3 -2 -1 o 1 2 3 4
log f'req
Figure 8.9: Measured Master Curve for Moisture Sensor Z24 Compared t.o Predicted
Response using Linear Distribution in Bulk Ohmic Conduction.

280
 z7l"19,cispl0,11
0
.6.4h
-5 ~.!h
A
-10 .ca.
A
A
-15 .ca.
A
A
-20
....s::
IG
I:J.
I:J.
10-7
A
t:D -25 .6.
El = 2:.r:l0- 11 F /m A
-30 0'1 = 10- 12 S/m .6.
d'J = 6.8 JJm
-35 Et = Ep = 2.60:.r:l0- 11 F /m
0'6 = 10- 10 S/m
-40 tP = .4

-45
-3 -2 -1 0 1 2 3 4

-10

-20
-30
ID -40
"
~
Id
At
-51]

-60
-70
-80

-90
-3 -2 -1 0 1 2 3 4
log l"req
Figure 8.10: Measured Master Curve for Moisture Sensor Z7 Compared t.o Predicted
Response using Dispersive Complex Bulk Permittivity Arising From Lorentz Sphere
Model.

281
region during which the capacitance of the bulk material is evident. No such evidence
is given. In fact, no mention is made of the existence, or nonf'xi~t,f'n('f', of llispf.'rsion
in the real part of the permittivity. Without presenting the dispersion of the complex
permittivity it is not possible to ascribe any particular dispersive molecular model to
the conduction process.

Di&per&ive Tran&port

All of the above charge carrier scenarios utilize nondispersive, ohmic conductors.
III general, charge transport and relaxation in amorphous semiconductors and solid
dielectrics are dispersive [100]. Very often, this dispersive transport is observed in
the time domain. A step voltage is applied across a sample and the current dC".cay is
monitored.
One type of decay observed can be represented mathematically as

(8.4)

and is known as the Williams-Watts function. This function is very similar t.o (8.1)
and thus, a physical interpretation of the existence of a distribution of dipole relax-
ation times is given. Another physical model, proposed by Scher and Montroll [101],
uses the formalism of continuous time random walk to describe carrier motion with a
distribution of event times. The event time could be due to either hopping between
the energy levels of localized sites or release from a trap. The resulting behaviour of
the dispersion is (8.4).
Other observed time decays are of the form

i( t) '" r,(t) o < s(t) < 2 (8.5 )

and are usually accounted for mathematically using distributions in charge relaxa.t.ioll
times or charge carrier hopping probabilities. There are several theories t.hat. aUempt.
to provide a physical basis for these distributions.
One such theory is a defect-diffusion model proposed by Shlesillger [102] uses the
concept of {racta! time stochastic processes to describe the relaxation of a mobile defect

282
in a volume containing frozen-in dipoles (traps). When there are many more defects
than traps, the form of the decay is (R.4). When t.he rev~rse sit.uation applif's, more
traps than defects,

0<a<2 (8.6)

In a related model, according to Jonscher [103], low frequency dispersion (LFD) is

associated with very large polarization arising from the migration of charges through
large but limited paths. In particular, LFD is associated with hopping, that is, low
mobility charge carriers, whether ionic or electronic. The physical model involves a
many-body interaction and results in a time domain response of the form of (8.6).
The corresponding frequency domain response is

(8.7)

It will be shown that the observed behaviour of the moisture sensors follows this
dispersive relationship and will produce the invariance in frequency response with re-
spect to moisture content and temperature. This is no~ ~leant to imply any preference
for one physical model over another. The emphasis here is on the correlation of the
observed dispersive behaviour.

8.2.2 Low Frequency Dispersion

The power law dispersion, as described by Jonscher [104, p. 87], has been observed
experimentally for a wide range of mat~rials over many decades in frequency. Some
of Jonscher's work can be found in [103] and [105] but, for a complete clllllpilation
of experimental results, the reader is referred to [10-1]. III it, examples are givPIl of
dispersion relations that scale with temperature and relat.ive humidit.y in t.he manller of
Theorem 3.1, producing what Jonscher calls universal responses but. what. are ('eferred
to here as master curves. Two examples are reproduced in Fig. 8.11. The first.,
Fig. 8.l1a, shows the normalization of lU;.b veaks for the a relaxation in polydian
carbonate as a function of temperature. The second example, Fig. 8.l1b, shows the

283
normalization of the complex permittivity of a sample of saud as a function of relative
humidity.

Furthermore, experimental evidence of low frequency dispersion has been observed

in plasma polymerized films by Hetzler and Kay [1061. Measurements of the dispersion
in plasma deposited films of tetrafiuoroethylene (TFE) mixed with argon as a transport
gas displayed a power law.dependence at 25°C with n '" 1. Film thicknesses varied
from submicron to 3 p,m. Using an optical absorption technique, it was discovered that
the observed extinction was due to Rayleigh scattering from particles up to several
hundred A in diameter. Thus, these films also exhibited a heterogeneity due to the
inclusion of microspheres. However, no connection between the film morphology and
the ambient temperature conduction process was made. Other than attributing the
conduction mechanism to a hopping conductivity, there was no further discussion of
the room temperature results as most of the effort was spent in understanding the
dipole losses exhibited at elevated temperatures of 180°0. No mention was made of
the RH of the atmosphere while the measurements were made.
As discussed earlier in Section 8.1, Ramu, et al. [971 observed a dispersion in
plasma polymerized HMDSO films. Examination of the full behaviour of the complex
permittivity might reveal a power law dependence of the dispersion.

8.2.3 Parameter Estimation of the Master Curve

Using the parameter estimation routines developed in Chapter 4, and the conclusion
of bulk dispersive transport reached in the previous sections, an estimation was made
of the dispersive dielectric properties of the coating from the experimentally measured
responses.
An example of this estimation process was to use a response of chip Z2·1, 2.2 Illll

thick coating, to unused oil having 48 pplll moisture cont,ent (Fig. 8.12) and est.imat.e
the permittivit.y and apparent. conductivity at 3 points per decade as shown in Fig. 8.13.
This estimation process was performed using a complex hulk permittivlt.', " ';;l -

2:z:10- 11 F Im and 0'1 = 0) for the oil above the coating. Essentially, tite ull is so
much more insulating than the coating that its conductivity may be neglected for this

284
 DJ
DJ
,.
0.05

0.02

0.0'

J 4 5
log ((Hz)

~ b)
',0

O.l

-z -,, - D 2
I
J
I

logft/f.)
,-
•

.:.

..
o

..• .. o."'"
••
•, 0 3:1
.
'.64 ••
....
0
.. .......
0

, _
.. .12··...
, .... ..
..... . .....
... . ,.
:~*"" le" :a :t :n
- ~... ,.. o~I.\U..................... !.• I!:J
. ........... ,
6 ' ,",,:.c ,.
'" '~:'".1·'... .. ~ ;--:1

""
of '1 .,
log 4: " '..
" l, ~
Iog ,
4
..e. ..., It ~..... . .~-......... ..
,~ S\
'x -.-- -!'_- --.-]}--!~.!,•
o
'10~
...... ...........-._._......
.. I?~ -•••
a
o 2 6
• 10 14 16
log(f/foJ

(b)

Figure 8.11: Experimental Data Showing Normalized, Power Law Dispersion of a)

Polydian Carbonate with Temperature Dependence and b) Sand with RH dependence
(from [104, p. 177 und 277]).

285
estimation problem. However, it is very important to include the oil permittivity as it
will significantly affect the coating bulk permittivity estimate. Note that t.he coating
permittivity estimate tends to the static dielectric constant (E.e = 2.95z10- 11 F /m) at
the higher frequencies, and that tItis value corresponds to the estimated permittivity
for Z24 in Table 5.1 (obtained using a different set of data). The apparent conductivity
is dispersive, significantly deviating from an ohmic behaviour. In Fig. 8.14 the results
are replotted to display the dispersive nature oC the coating. The estimated static
dielectric constant was subtracted from the permittivity (E' - E,d (dots), and the
conductivity was converted to E" by dividing by w = 27rf (crosses). By drawing a best
fit line through the data on a log-log plot, it appears that the dispersive relationship
between the complex permittivity and frequency follows a power law dependence.

The form of the power law dispersion, adapted from Jonscher [104, p. 87] but
expressed in terms amenable to Theorem 3.1, is

E·(W) t' - jE"

B(jw/wp)n-l (8.8)
- B[sin (n7r /2) - j cos (n7r /2)](w /wpt- 1
with 0 < n < 1 and wp = f(T,RH,c . .. ).
The constraint on n arises from the requirement, for a linear, causal system, that
the Kramers-Kronig transformations be satisfied [104, Section 2.8]. Using Courier
analysis and complex integration, these transformations relate the real and imaginary
parts of the complex permittivity, after subtracting out any dc conductivity or static
dielectric constant. For chip Z24, n '" 0.38 and B '" 9.h:l0- 11 F /m. Using (8.8),
the amount by which log( E' - E,t) is offset from log( E") is log( tan( n7r /2)), and this
correlates well with the difference in the estimated intercepts obtained using the best
fit lines, as noted in Table 8.1.

The above exercise was repeated for the response of chip Z7, 6.8 fUll thick coating,
in unused oil having a moisture content of 14.2 ppm (Fig. 8.15). Again a power law
dispersion was observed with E,t = 2.53xl0- 11 F /m, roughly the same value as given
in Table 5.1. Here, n '" 0.26 and B '" 3.0:cl0- 11 F /m. The lack of estimates of E' at

286
 z2-«51
0

-5
... ....
-10
'.. •
•
"•
'.r.•
-15

c -20
.....a
m -25
-30
...
.....
' ....
........ ........
-35 •• .. H ••

-40

-45
-3 -2 -1 o 1 2 3 4

0
••••
........ .
,...... '"

-10
"• •.-
-20 •
-30
••.... •
••
G

'cs"
..c;
s:a..
-40

-so
-60
'. .......
'.
-
,"I·
•

-70

-BO ,-

-90
-3 -2 -1 o 1 2 3 4
log ~req

Figure 8.12: Measured Response of Aged Moisture Sensor Z24 in Stirred, Unused Oil
at 48 ppm Water.

287
 1.lXlOO

-"
~
S 1()(x)

-'
-'
•
I
•
-"
r&I 100
-' •
~
G)
• •
10 •
-."
0 •
•• •• ••
1
0.001 0.01 0.1 1 10 100 1000 10000

100

-
.
-"
S
Cf.)
10 •
•
•
en •
• •
I

-
r&I
~

•
-"
tn
1
•
•
•

-
tn
0

• •
0.1
•
0.001 0.01 0.1 1 10 1.00 1.000 1.0000
log Ereq

Figure 8.13: Estimated Dispersion in Bulk Permittivity and Conductivity of Z24

PP-Bron1(lbenzene Coating on Z24 using Response of Fig. 8.12 (d 2 = 2.23 J.Lm).

288
 1000~----~-----r----~~----~-----r----~~--~

a....... 100
..ra.
I

---
....
0

10
\,"

-
tIO
0

-...-.
~

--
"101"
tIO
0

0.1

0.01~----~----~------~----~------~----~----~
0.001 0.01 0.1 1 10 100 1000 10000
log f"req

Figure 8.14: Estimated Dispersion in Complex Bulk Permittivity of Z24 PP-Brolllo-

benzene Coating from Fig. 8.13. (.) - log( E~ - E.e), (X) - log( E~).

289
 Table 8.1: Dispersion Parameters For Moisture Sensors

--
Chip Chip n B fd log tan( n2ff ) A(int)
ID COlld. (10- 11 F/m) (10- 11 F /m)

I Z24 I Aged 1 0.381 9.1 2.95 -.167 I -.201 I

Z7 Fresh 0.16 2.3 2.45 -.591 -.417

" Aged I, 0.26 2.7 2.65 -.364 -.265

Adjusted

" Aged II 0.26 3.0 2.53 -.364 -.335

" Aged III

-
0.28 3.4 2.60 -.327 -.309
Z20 Fresh 0.24 3.1 2.60 -.402 -.422
- --
" Aged 0.36 4.4 1.94 -.455 -.249
11
Adjusted, 0.32 6.4 2.64 -.260 -.383
Aged

290
the high frequencies is due to the occurrence of a slight positive phase and dip ill gain,
a parasitic effect associated with the different capacitance seen hy t.he flOill.ing gat.e
FET versus the reference FET [70, Section 5.5].

The dispersive relations characterizing Z7 at several different stages of aging are

given in Table 8.1. Aging of the films altered the master curves somewhat, as shown
in Figs. 7.12-14. This effect can be seen in changes in the power law dependence as
the coating of Z7 ages. The broadening effect with aging shows up in the increase
of nand B, with some apparent change in f,t also. Of course, the drastic change in
moisture sensitivity with aging appears in the parameter fp.
Presented next is the dispersive relation describing chip Z20, 6.0 JLm thick coating,
in unused oil having a moisture content of 17 ppm (Fig. 8.16). The resulting dispersion
relation appears to exhibit a. similar power law dependence with f,t = 1.94:clO- l l F Im
-
except for the low frequencies. Fo'" many of the dispersions, deviations from the power
law dependence occurred at the low frequencies, as seen in this case. The permittivit.y
appears to level off and the loss appears to peak and begin descending. If this result
were taken at face value, it would indicate a capacitance, perhaps due to blocking at an
interface, dominating at these lower frequencies. However, as previously mentioned, it
is known that the gain rarely ever goes to exactly 0 dB (unity gain) at low frequencies
due to experimental errors. Here, the culprit is a large DC offset, on the order of -1
to -2 volts. This problem results in a roughly constant shift in the gain curve. A
correction to the problem can be made by shifting the gain curve such that it does
approach 0 dB, subtracting the difference between the low frequency gain (when it
plateaus) and 0 dB, and re-estimating. Again, this correction assumes there is no
significant electrode polarization and that the gain error is constant with frequency.
This results in significant alterations in the dispersion only at. f.he low frequendes. The
results of such a computation (a gain correction of 1.9 dB) for t.he response of Z20 used
above is shown in Fig. 8.17. The change in nand B give some idea of t.he sensitivit.y
of these parameters. An additional benefit of this correct.iOll is that t.he re-est.imat.ed
f,t of 2.64:clO- l l F Im is much closer to the value given in Table 5.1. It is seen that t.he
behaviour of the estimated dispersion at the low frequencies is dominated by whether

291
 10000

1000

-a
-
......f&. 100 )(
• I

~~
I
....
0

-
"-

~
"- 10 ,""
-~

!~ k
-
.:cs
";
III
1 .
~
kx
-,s'
"-

- 0.1
tIIO
0

0.01
0.001. 0.01 0.1 1 10
~
~
100
.
x

1000
,
'~
x

10000

Figure 8.15: Estimated Dispersion in Complex Bulk Permitt.ivit.y of Aged Z7

PP-Bromobenzene Coating. (e) -log(f~ - e.t}, (X) -log(e~).

292
the nonzero dB gain is real or artifactual.

8.3 Film Aging and Electrical Conduction Mecha-

•
nlsm

As the coatings aged, the parameters tabulated in Figs. 7.10 and 7.11, expressing the
moisture sensitivity for Z7 and Z20, changed. In fact, both the slopes and the inter-
cepts increased mOllotonically with time, as shown in Fig. 8.18. The different paths
taken reflect the different temperature cycJings for each coating. Eliminating time as
the independent variable and plotting slope-intercept pairs, as in Fig. 8.19, yields a
very interesting result. The slope and intercept appear to be relat.ed to each other
throughout the aging process. In fact, there appears to be a distinct trajectory along
which the coatings age although the sample size is rather small (only two coatings).

The issue of aging is intimately linked with the conduction mechanism responsible
for the electrical response of the film. What is carrying the conduction current and
why is it changing with time? The following discussion relates work reported ill the
literature on this subject to the experimental observations of the previous chapter.
Methods used to improve the stability of the coatings are also reported.
The physical picture given by Jonscher [104, Chapter 81 or Shlesinger [1021, and
associated with the dispersion relation of Section 8.2.3, is one of charge carriers (or
mobile defects) with low mobilities traveling through paths limited by frozen dipoles
or traps. These charge carriers could be ionic or electronic. This might be a good
description of the interaction of impurities with free radicals trapped in the film. As
discussed in Appendix C, free radicals play an important pArt in tIll" formntion of films
via plasma deposition and can be present wit.hin t.he film in significant. COl1cent.ratioIls.
The effect of moisture might then be to enhance the mobilit.y of t.he impurit.ies. Ty-
czkowski and Kryszewski have shown that moisture does not. direct.ly cont.ribut.e t.o
the conductivity by comparing results using regular and heavy water [921. The ion-
ized species from these two molecules have sufficiently different mobilit.ies such that

293
 1(0)()

1CXlO

----a • • •
..~100
I
x; 1"'"..!- -t", <
-
o ~ x
~
.-I
)
...-.
=: .. X r-
~ 10
~

~~
1
i(

-.
...-.
00J
1

~
0.1
~)'
-

0.01
0.001 0.01 0.1 1 10 100 1000 10000
log f'req

Figure 8.16: Estimated Dispersion ill Complex Bulk Permitt.ivity of Aged Z20
PP-Bromobenzene Coating. (.) - log( f~ - f,e), (X) - log( f~).

294
 1Ul1lO~-----r------~----~------~----~------'------.

1000
X
.......
a
•
.........
~
......I 100

---
0
.-4

~
10
-.......
~
bO
0

-.
\Cl
1

--..:
- bO
0

0.1

0.01~----~----~------~----~----~------~----~
0.001. 0.01 0.1 1 10 100 1000 10000
log f'req

Figure 8.17: Estimated Dispersion in Complex Bulk Permittivity of Aged Z20

PP-Bromobenzene Coating using Corrected Response. (.) -log( f~ - f6t), (X) - log( f~).

295
 --
1

o~--------------~--------------~--------------~
o 20 40 60

1 ~----------------------------------7'~--~---.

0.8

--
a:a 0.6
I)

-Po..
0
Cl.)
0.4

0.2

o~--------------~--------------~--------------~
o 20 40 60
ti me (weeks)

Figure 8.18: Time Dependence of Change in Paral:'.let.ers Describing Moisture Sensi-

tivity for Z7 (X) and Z20 (.) with Aging.

296
 3~----------------------------------------~
X

--
<I:
2 -
X
+*
~
IJ
0
J".
IJ
•
+*
•
-
s:: 1 ~

•
•X • x

0 I I I I
0 0.2 0.4 0.6 0.0 1
Slope (B)

Figure 8.19: Correlation of Parameters Describing Moisture Sensitivit.y for Z7 (X) and
Z20 (.) with Aging.

297
the effect on the conductivity due to dissociation of the absorbed moisture would be
apparent. There was no measurahle differenre in conductivity so the conclusion was
that the resulting moisture sensitivity was electronic in llature, due to an activation of
initially inactive intrinsic acceptor centers. This conclusion does not take into account
the possibility of absorbed moisture ionizing other impurities within the film or inter-
action with the trapped free radicals. Many examples of enhanced conductivity due to
ionization arising from absorbed moisture in standard polymeric solids were compiled
by Barker [107]. At low moisture concentrations (ppm), before any significant effect
on the macroscopic dielectric constant occurred, partial dissociation of moisture and
enhanced dissociation of impurities resulted in increased DO conductivities.
Aging might then be associated with a decreasing concent.ration of charge carriers
over time. Lack of coating stability has been related to the in.teraction of oxygen wit.h
the trapped free radicals [83, Section 11.2.3]. One effect is the formation of polar groups
arising from this interaction increasing the dipolar losses of the film. For example,
Hetzler and Kay [106], in examining plasma polymerized TFE at 180°0, concluded
that dipole losses dominated, a confusing result given the small amount of dipole
concentration due to CF z bonds. Using FTIR, it was discovered that C-O bonds were
responsible for the dipole losses and that these bonds formed upon exposure of the film
to oxygen in the air. In another type of interaction, DC conduction measurements by
Morita, et al. [108], on fresh plasma deposited styrene films, displayed an irreversible
decrease in conduction currents on exposure to oxygen but a reversible increase with
exposure to humid air. Experiments by Tyczkowski and Kryszewski (92) on plasma
polymerized organosilicon films showed all increase in DC conduction currents with
exposure to oxygen at an elevated temperature of 380°0.
While the particular direction of t.he effed dc-pends on t.he specific 111onome1' used,
be it hydrocarbon or organosilicoll or fiuorocarbon, it. is deal' from t.he above refer-
ences that oxygen is having a significant effect on t.he elect.1'ical propert.ies of t.he film.
According to Yasuda [83, Section 11.2.3] the inst.abilit.y of the film over a long period of
time is attributed to the low mobility of large molecule:; wHhiu tIlt: highly cross-linked
structure. The surface of the film may be in equilibrium with t.he ambient oxygen

298
levels but it will take a significant amount of time for oxygen to penetrate the film and
interact with free radicals trapped deep down ill the bulk of the film. There are several
post treatment methods that reportedly improve the stability of the films. One is to
quench the free radicals by deliberate exposure to oxygen [83, Section 11.2.3]. How-
ever, this may result in undesirably high dielectric losses. The second method is to
thermally anneal the films in vacuum. The effectiveness of this method is mitigated by
the low mobility of the polymer segments. Without an increase in this mobility at ele-
vated temperatures there is the possibility of thermal creation of new free radicals due
to the breaking of chemical bonds. Plasma polymerized organosilicon films, treated for
2 hours at 350°C in a nitrogen atmosphere, exhibited reduced DC conduction currents
and were insensitive to changes in humidity [92]. Tuzov, et al. [109] found that the
capacitance of plasma polymerized HMDSO films, heat treated in vacuum at 350°C
for 1 hour, stabilized to -120% of its initial value after 4 days. The largest. change in
capacitance and least stable film was produced by heating the film in air at 350°C for
1 hour. The exact conditions required for maximizing stability are a function of the
monomer and polymerization conditions as well as the time and tempc;:'dture of the
posttreatment. The third method for improving stability is to increase the wbstrate
temperature during deposition. Ramu, et al. [97] found that the loss tangent of plasma
polymerized HMDSO films stabilized within 2 to 3 days when produced at a substrate
temperature of 400°C.
However, there is a price to pay for any method of stabilizing the film properties -
the resultant film may not have the required sensitivity or properties for the specific
task. This qualification must be taken into account when investigating methods to
reduce aging effects.

299
Chapter 9

Hydrogen Gas Sensors -

Prelinlinary Research

9.1 Motivation
Among the various approaches to gas analysis outlined in Section 1.3.2, the one chosen
for development was Fourier Transform Infrared (FTIR) spectroscopy. The develop-
ment and testing of this approach is beyond the scope of this thesis. However, this
option is not able to detect H2 , thus, there is still a need to develop an H2 gas sen-
sor. Most of the H2 gas detectors discussed in Section 1.3.2, with the exceptions of
the flame ionization detector (FID), thermal conductivity detector (TCD) and electro-
chemical fuel cell, are still in the development stages. The unstructured doped metal
oxides devices, such as the Taguchi Gas Sensor (TGS), are not selective enough to H2
. However, the hydrogen contribution could possibly be determined by subtracting
out the other gaseous components as measured by the FTIR. The st.ruct.ured metal
oxide devices are not available for commercial use, presumably due t.o pOOl' perfor-
mance during long term operation. In any case, it. was felt that. the area of hydrogen
gas sensing was still at such a developmental st.age t.hat. it. was wort.hwhile t.o explore
promising techniques, perhaps exploiting t.he capabilities of the microdielect.romet.er.
As with any gas sensor, the five major areas of attention are: selectivity, sensit.iv-
ity, reversibility, stability and operating conditions. The last area is of great design

300
hilportance - whether to operate the sensor immersed in the oil, or in a gas phase that
is in equilibrium with the dissolved gases ill the oil. In either ('ase t he gas sensor will
be coated with a thin polymer film (membrane) such as parylene. This coating will
improve the selectivity to hydrogen by impeding the diffusion of larger gas molecules
and moisture, and/or protecting the device from the oil. The advantage to operating
in the oil is the anticipated rapid equilibrium between the local gas concentration in
the oil and the sensor via the thin ( . . . . 5 p.m) polymer film. However, a disadvantage is
that the film must remain stable in its mechanical and physicochemical characteristics
for a long period of time (years) ill the hostile environment of hot transformer oil.
Another important constraint is that. the sensor must remain at the temperature of
the oil « 90°C). Typically, the solid state gas sensors are operated at higher temper-
atures (2: 150°C). This lower temperature constraint, although aiding in the hydrogen
selectivity through the favoring of the adsorption-desorption kinetics for hydrogen,
generally results in a reduced sensitivity and slower transient response. Operating the
sensor in a gas phase in equilibrium with the oil does not suffer from this constraint,
nor the previous one. The major disadvantage to operating a solid state sensor in the
gas phase is the requisite increase in complexity of having to establish yet another
equilibrium, between the oil and gas phase. If the sensor is operated behind a mem-
brane separating the gas and oil phase, the membrane must be designed as thin as
possible so as to minimize the diffusion time through the membrane of the gases of
interest.
For the preliminary work presented here, the devices will be evaluated in the gas
phase. At present, parylene is being envisioned as serving as the membrane material,
given its good performance as a moisture barrier and its compatibility with trans-
former oil. Two approaches that were explored are described in t.he following sedions.
The first approach was to sensitize a microchip t.o hydrogen gas by coat.ing it. wit.h a
thin film of tin oxide. This approach relies on t.he known gas sensit.i vHy of t.in oxide
(see Section 1.3.2) but tries t.o adapt the concept. to a high impedance measurement. of
the AC surface conductivity of thin films. No significant gas sensit.ivit.y was observed.
However, these experiments were conducted at room temperature wit.h no attempt

301
at controlling temperature variations. As a consequence, the RH and temperature
variations had a large, and perhaps dominant., effect on the surface cond uel h·it.y. The
second approach involved the use of thin film palladium resistors. Here, the tempera-
ture of the experiments was carefully controlled and varied between 100° and 300°F.
Bulk absorption of hydrogen by the resistors produced measurable changes ill the
resistance.

9.2 Tin Oxide Coated Microchip

9.2.1 Background

This approach combined the gas sensitivity of doped metal oxide thin films (see Sec-
tion 1.3.2) with the detection capabilities of the microdielectrometer. Thin films (::;
2000 A) of tin oxide (Sn02) were vacuum deposited onto an uncoated microchip. Here,
instead of measuring a change in DC conductivity of a macroscopic pellet of sintered
tin oxide powder, an AC, high impedance (low current) measurement of the conduc-
tivity of a thin film was made. The potential advantages of this approach are: 1)
AC operation avoids any DC polarization effects that may adversely alter long term
performance, 2) high impedance implies low currents, yielding negligible temperature
variations arising from the measurement and negligible formation of byproducts due
to electrochemical reactions, and 3) given a microdielectrometer type of device in the
monitoring system, adding another microchip sensor is cheap.
Uncoated microchips were placed inside a bell jar vacuum chamber and a gram of
stannic oxide powder (Mallinckrodt) was placed in a tungsten boat. After the chamber
was evacuated, high currents were passed through the hoat. - heat.ing and vaporizing
the powder. This vaporized powder diffused throughout t.he whole rhamlJer. coat.ill~

all surfaces including the microchip electrodes. The t.hickness was monit.ored using a
quartz crystal mass monitor. Aft.er coating, th~ microchips were immediat.ely exposed
to the atmosphere, no post-treatment was performed.
An exploded schematic view of the deposited, granular thin film of t.in oxide and

302
the microchip is shown in Fig 9.1. Although the surface conductivity between the
elect.rooes was the quantity of interf>st. hf>re, there is no reason why t.he t.hin film could
not be deposited on an insulating coating instead, removing the question of contact
resistance (or polarization) between the electrodes and the thin film. In addition,
it might be preferable to use the macrochip here because of the greater latitude in
operating temperatures with the ceramic device. There will be an upper limit. on
coating thickness due to increasing film conductivity with thickness. At some point,
the film would be too conducting to observe its behaviour with the frequency range of
the microdielectrometer.

Sn02 (w/Pd)

Figure 9.1: Schematic View of Tin Oxide Thin Film Deposited on a Microchip

The gases used were nit.rogen or a factory mixt.ure of 1000 ppll1 or 100 pplll TT 2
11l nitrogen. All the gases were of Matheson PurHy grade (99.999~1%) llwanillg '<
5 ppm total of oxygen, carbon monoxide, carbon dioxide, moist.ure, argon and total
hydrocarbon content.

303
9.2.2 Experimental Results

Tin Oxide

A typical frequency response for the tin oxide coated microchips was shown in
Fig. 5.20. That response was for microchip Sn02, having a 350 Athick coating of till
oxide . .\S the external environment was varied, temperature or RH, the curves would
remain the same shape and just slide up and down in frequency. Again normalizing
the frequency to JP a master curve can be produced (Fig. 9.2). This is reminiscent of
the master curve behaviour of the moisture sensors, only now the response is due to
a complex surface permittivity. As shown in Section 5.5, the response reflects what
appears to be an ohmic surface conductivity with a dispersive surface permittivity,
perhaps due to polarization at the electrodes. This response may be better modeled
by separating the ohmic surface ~onductance from a barrier impedance due to po-
larization. The dispersion in the barrier impedance may then be estimated, perhaps
yielding a power law dependence similar to those observed for the moisture sensors.

A plot of the estimated surface conductivity as a function of RH for microchip

Sn02 is shown in Fig. 9.3. The dependence is linear 011 a log-log scale, producing a
power law dependence of the surface conductivity on RH,

(9.1)

For these results, (To = 3.33z10- 12 Sand f3 = 1.7. At this point it is appropriate to
address the issue of possibly using this tin oxide coated microchip for moisture sensing
in the oil. One problem with this approach is that the tin oxide, deposited by vacuum
vaporization, is not bound very strongly to the material. Either a protective overcoat
needs to be applied or the tin oxide must. be deposited using a t.echnique t.hat produces
better adhesion, such as sputtering. A more crit.ical issue, t.hough, is whet.her t.he t.in
oxide film would be sensitive to the small concentrat.ions of moisture in the oil. It. is felt
that moisture sensitivity in air does not imply sensitivit.y in oil. One of the virtues of
the bromobenzene moisture sensors was the amplification of changes in small moist.ure
concentrations via bulk absorption. It is not clear that the thin tin oxide film will

304
 0

-6.1

-Il.l

-/13

.-= -30.s'I
~
bO
""''1.

-~.,

-'a.7
-nl
-)'1.<i
-4 -3 -2 -1. 0 1. 2 3 4 5

o
-15.8
-3:,7

_'11.5

-/10.'1

-/~

-f~~ ~---~---~~--~--~~---~---~-----~---~--~
-4 5
log(f / Jp)
Figure 9.2: Master Curve for Sn02, a 350 A Tin Oxide Coated Microchip, in Variable
RH Environment.

305
produce the same effect.

Pd/Tin Oxide

The dominant effect of temperature on the response of the tin oxide thin films is
shown in Fig. 9.4. This response was measured at 10 kHz using microchip 5n02, having
in addition to the 350 A thick coating of tin oxide, a 10 A overcoat of palladium. This
palladium overcoat was also vacuum deposited. It was thought that the palladium
migllt serve as a catalyst and enhance the gas sensitivity, as observed in the literature.
The main effect of the palladium appeared to be the dramatic increase in surface
conductivity, necessitating the monitoring at high frequency. The measured phase
at 10 kHz in Fig. 9.4 corresponds to the phase at 0.5 Hz in Fig. 5.20 (before the
overcoating with palladium).

The microchip was placed inside a sealed beaker and exposed to eit.her the dry
nitrogen or the 1000 ppm hydrogen/nitrogen gas flow. Initially, the microchip was
exposed to the dry nitrogen. There was a rapid but small decrease in gain, but no
other significant change was observed. After roughly 4 minutes, the flow was switched
to the 1000 ppm hydrogen/nitrogen gas mixture. Another rapid decrease in gain,
followed by a slower decrease, was accompanied by a decrease in phase, increase in
DC offset and decrease in temperature. Finally, after 25 minutes, the microchip was
removed from the beaker and placed in a dessicator at 2-4% RH. The changes in
the response produced by this action are on the same order as the gas composition
changes. Although the removal of the microchip to the dessicator involved exposure
to the ambient atmosphere, it is felt that the dominant effects were the change in
RH and temperature. Notice that the change in gain, phase and temperature were
the same regardless of gas composition but were opposite in dired.ion upon removal
to the dessicator. A plausible scenario is that. the !lJl' ,. "hip was const.antly heing
both dried and cooled by the gas flow, and absorbed moist. ure and heat.ed up sligh t.ly
upon removal from the gas flow. Once in the dessicator t.he absorbed llloist.l11'e was
slowly desorbing, as indicated by the slow decrease in gain during t.he second half of
the observed time period. In any event, given the lack of temperature and humidity
control, it does not appear that the gas sensitivity of these devices were truly tested.

306
 Sn02 - 350 Angstrom
1000r---------------------------------------____-,

100

.....-
e ll

-
0
.-4

b
0
10

- 1:10
0

O.1~------------~--------------~----
0.001 0.01
__________~
0.1 1
log (RH)

log(O',o) = 1.7Iog(RH) -11.478

Figure 9.3: Relative Humidity Dependence of Estimated Surface Conductivity of

350 ATin Oxide Coated Microchip.

307
 sno2t12
-9 r----~----r--_, -19.1 .----.----"'"T'"-----.,

-20.1

-9.5

s::
•
....IG
~
~22.1
-10

-23.1

-10. 5 L-_~ 1 - - - - ' ' ' - - - - - - - - - ' -24. 1 L - -_ _ -'*""_ _--'-_ _---'
05075 0 50 75
f t
"'- Ha.
·t
D&ssi~.r
f t
1'11. Hl.
-1.86 25.8~--~--~--~

-1. 9 r---~---+-----I CD 25.3 I - - - - + _ -

~
'+' ::s
~ '+'
o IG
Clot ~
Clot CD
o =-
-1. 94 ~--....---+--~~ 24.8 t - - - " . - - - t - - - - t - - - - - - I

-1.98 L - -_ _....L-_ _- - ' -_ _----' 24.3 L - -_ _

o 2S 50 75 0 2S 50 75
time (Min) time (Min)
Figure 9.4: Measured Response at 10 kHz of Pd (10 A)/Till Oxide (350 A) Coat.ed
Microchip in Gas Flow (Variable Composition).

308
9.3 Thin Film Palladium Resistor

9.3.1 Background

The approach explored here was to exploit the uniquely high solubility of hydrogen
in palladium. Basically, once adsorbed onto the surface of palladium, hydrogen is
rapidly absorbed into the bulk, giving rise to small, but measurable, decreases in bulk
conductivity. Due to this unique property, it is expected that devices based on this
principle will be highly selective to hydrogen in the presence of a gas mixture.
This behaviour is detailed by Lewis [110, Chapter 4], including descriptions of
early results obtained using fine palladium wires. Percent changes in the resistance
of the wires were related to the atomic ratio of hydrogen to palladium. The latter
quantity was determined by measurement of gas volumes evolved during electrolysis.
Changes in resistance of up to 70% were obtained at atomic H/Pd ratios of 0.7 at
room temperature. Room and low temperature (77°K) experiments at low pressures,
using evaporated Pd films (80-90 A thick), exhibited much smaller percent changes in
resistance [111]. Here, a pressure of 10- 3 mm Hg of pure hydrogen, corresponding to
roughly 0.1 H/Pd, resulted in a 1.5% increase in resistance. The kinetics of the reaction
H2 r= 2H (dissolved) on thin Pd and Pd-Ag foils (2.5-5 /Lm) was studied by Auer
and Grabke [112). It was shown that the rate determining step is the dissociation
of the H2-molecule on the metal surface. Therefore, the cleanliness of the surface,
and the availability of surface adsorption sites, are very important in determining the
magnitude change in the resistance and the speed of the transient response.

9.3.2 Experimental Results

Fabrication

The thin film palladium resistor was fabricat.ed by t.he chemistry lab of Professor
Mark Wrighton. Four palladium resistors, 500 A thick by 100 lUll wide and roughly
2.4 mm long, were sputtered onto an insulating substrate (Fig. 9.5). The subst.rate

was composed of a silicon wafer base, a middle layer of thermally grown Si0 2, and an

309
upper layer of low pressure chemical vapor deposited (LPCVD) silicon nitride (Si3 N4)'
Two of the resistors were sealed against exposure to the ambient. atmosphere using a
coating of Epoxi-Patch Kit (Hysol Aerospace & Industrial Products Div.), whereas
the other two were exposed over 1/3 of their length (.8 mm). The 'chip' containing
the resistors was then mounted and sealed in a glass tube with a septum at one end.
This allowed testing of the resistor to be performed in an enclosed volume of roughly
0.4 mL.

Experimental Setup and Procedure

A two terminal measurement of the resistance was made using an HP3455A digital
voltmeter, controlled by all IBM PC/XT. It was observed that, due to the small mass
of the resistor, significant heating of the resistor occurred if the voltmeter was operated
cOlltinuously in the resistallce measurement mode and thus, always supplying current.
The heat rise can be calculated assuming all the energy dissipated in the volume (V)
of the resistor goes into heating

[2R
AT = --dt (9.2)
pVcp
Using a current [ = .7 mA, nominal resistance R = 1000, density of palladium p =
12.d03 kg/m 3, the dimensions of the resistor, and a heat capacity cp = .058 cal/gm-DC
(243J/kg in MKS units) yields

AT = 6.73:c10 3 dt (9.3)

With the current applied continuously, a significant rise in temperature can occur
causing a constant increase in resistance, as was observed experimentally. Therefore,
the voltmeter was left in a voltage measurement mode (open circuit) except when
taking the periodic resistance measurcment..
The operatillg temperature of the device was cont.rolled uSlllg an Omega 6100
temperature controller and a thermocouple. The resistor, and a thermocouple, were
placed illside a larger diameter pyrex tube having a heating coil wrapped around it.
Paper was used to fill up the remaining volume within the tube so as to prevent air

310
Figure 9.5: Layout of Thin Film Palladium Resistors.

311
circulation.
The same high purit.y gases mentioned above were flowed C'onfinno1ls1y t.hrough
the enclosed volume via inlet and outlet syringe needles piercing through the sep-
~Ulll. Using the experimental flow rates (roughly 280 mL/min.) and needle lengt.h,
it was calculated that the gas temperature would be in equilibrium with the device
temperature IOllg before it actually reached the device.
Tests were conducted after t.he device had reached thermal equilibrium and had
been exposed to the dry nitrogen flow for roughly one hour. Without this initial
conditioning with the nitrogen, the early results would show a decrease ill resistance
when exposed to either the nitrogen or the H2 /nitrogen mixture. The devices would
be left exposed to room air overnight, thus the conditioning may be drying the surface
of the thin film resistors and/or removing other adsorbed species, such as oxygen,
occupying valuable surface adsorption sites.

Results

At a temperature of 225°F, the device is sensitive to both 1000 ppm (Fig. 9.6a) and
100 ppm (Fig. 9.6b) H2 in nitrogen. The illtroduction of hydrogen illto t.he resistor
results in an increase in the resistallce, as was observed in the references mentioned
above. The transient response has both a rapid ("" 1-5 minutes) and a slow ("'"' 10-
100 minutes) component. The slow component may be due to the slow diffusion of
hydrogen through th~ epoxy barrier, the lateral diffusion of hydrogen throughout the
length of the resistor to regions not exposed to the atmosphere, or slow diffusion of
hydrogen into the substrate. The characteristic diffusion time is

(9.4)

where D is the molecular diffusivity and fJ is t.he charact.erist.ic lengt.h for t.he diffusion
process. For hydrogen ill palladium, D ,..". 10- 9 m 2 /s at. room t.emperat.ure [1131. The
vertical diffusion through the thin resistor (6 = 500 A) is very rapid (TD ,....., 2.5ps) bu t.
can be much longer in the lateral direction (6 ,.",. 1 mm -" TD ,.",. 10 3 s). The responses
appear to be reversible to a large degree. The magnitude change in resist.ance at

312
1000 ppm is 0.45!l versus 0.0611 at 100 ppm. The expected dependence would be as
the square root of the concentrat.ion [112].
Test were also conducted at a reduced temperature of 100°F. The response to
1000 pplll hydrogen (Fig. 9.7a) is comparable to the higher temperature result (Fig. 9.7b).
However, at 100 ppm hydrogen the response (Fig. 9. 7b) is beginning to get lost in the
experimental error. Also, the transient response at the lower temperature and concen-
tration appears to be much slower than the higher temperature response.

313
 ·msT 14 - 1000ppm H2

~~.1~----------------------------------------~

101

100.9

c-:.
~ 100.8
o

100.7

100.6

H~"h
t 1 H~
t
1oo.5~----~~----~------~------~----~~--~
o 60 120 100 240 300

1EST 16 - 100ppm H2
101.B~------------------------------------------~

.-

101.7

CIl
I:
::z:
0

101.6
t
H,..

101.5~----~~----~------~------~------~----~
o 60 120 100 240 300
TInE (HIH)
Figure 9.6: Response of Thin Film Palladium Resistor to a) 1000 ppm and b) 100 ppm
H2 ill Nitrogen at 225°F.

314
 "lEST 25 - 1000ppm H2

92.4~----------------------------------------~

92.2

91.80~----~----~----~~~----~------~----~
30 t 60 120 150 180
fb. N,.
~ 20 - 100ppm H2
91.7~----------------------------------------~

91.68

91.66
Cl.)
E:
:::I:
o
91.64

91.62

91.6~--------~~~~----~--~----~------~--~
o f 60 90 120
H,.. TIME (nItt1- t
Il
Figure 9.7: Response of Thin Film Palladium Resist.or t.o a) 1000 pplll and b) 100 pplll
H2 in Nitrogen at 225°F.

315
Chapter 10

Conclusions and Future Work

10.1 Continuum Model

A modal representation relating the gain-phase response of the microdielectrometer to
the continuum properties of a medium has been formulated so as to retain generality
and flexibility in its application to various linear systems. This approach greatly
extends the usefulness of the device by providing a framework for interpreting data
obtained for heterogeneous media in terms of absolute continuum parameters rather
than "lumped equivalents".
There are several areas that could either be improved or generalized to further
increase the usefulness of the model. One area is computation speed. By taking ad-
vantage of fast fourier transform (FFT) techniques, the time savings for performing the
summations in (2.22a) and (2.24a) could possibly be increased even more than already
obtained using the approach outlined in Appendix A. A further generalization of the
method would be to allow for the capability of representing the entire boundary wit.b
collocation points. This approach would enable implement.ation of a mo!'*" det.a.iled
description of the boundary condition at t.he electrodes, possibly allowing t.he incor-
poration of electrode polarization effects. This capability miJ:!;ht. also allow a met.hod
with which the finite thickness of the electrodes could be simulated as well.

316
 10.2 Parameter Estimation

Development of simple single and multiparallleter estimation routines incorporating

the continuum model provides the capability to infel" continuum properties and geom-
etry from experimental data. The key to parameter estimation, as applied to interdig-
ital electrode structures, is in identifying the types of data that can be mapped into
the continuum properties and geometry. Illustrated here have been one dimensional
schemes for the identification of film thickness, the complex bulk permittivity of fi-
nite thickness films and the complex surface permittivity of thin films. These schemes
have exploited the analyticity of the functional dependence of the gain-phase response
on these parameters. Also developed was the estimation of both the thickness and
permittivity of a film from two measurements of gain. Here, the known variable is
_ the dielectric medium above the layer. This technique was useful for estimating the
thickness and bulk permittivity of thin film coatings.
Several experiments, involving measurements of oil, paper insulation and thin poly-
mer and metal films, demonstrated both the utility of the parameter estimation rou-
tines and the validity of the continuum model.
It was also demonstrated that a unique interpretation of spatial distributions in
dielectric properties in a heterogeneous medium could not be made from the temporal
frequency response unless several assumptions regarding the distribution of relaxation
times, and the dispersive nature of the properties, were made. However, by varying
the fundamental spatial wavelength of the electrodes, a new approach call be taken to
sorting out distributions of properties. Exploiting the 'imposed w-k' approach, and
exponential decay of the electric fields normal to the plane of the electrodes, the mul-
tiparameter estimation problem was reduced t.o a sE-rit's of seqnent.ifll Ringlf' pflfflnlf'ter
estimations that yielded smoot.h distributions of complex bulk permit.t.ivit.ies. In COll-

junction with this scheme, cases were examined where known discont.iuuities ('ould be
estimated using a two parameter, instead of single paramet.er, search.
For several of these multiparameter estimat.ions, iterating to refine the good es-
timates obtained on the first pass through the data appeared to actually produce

317
diverging subsequent estimations. It is believed that at least part of the problem lies
in using the stair-step representation, introducing unrealistic cliscontinuit.ies t.hat con-
tribute to the instability of the iterations. The next step would be to implement the
exponential representation of Fig. 4.10b. With this representation, the spatial dis-
tribution is continuous, with the discontinuity first appearing in the first derivative
of the distribution. Another possible remedy might be to go from an exactly con-
strained to an overspecified problem. That is, the number of experimental datapoints
is greater than the number of parameters to be est~mated. This approach may yield a
convergent search but the cost is the increase in computation due to the necessity of
performing the multiparameter estimation simultaneously rather than as a sequence
of single parameter estimations.
All of the estimation routines described are simple Newton or secant method
searches. The robustness of these routines, particularly the two parameter estima-
tion and the multiparameter estimation using variable spatial wavelength data, could
be improved using 'smarter' search routines such as Levenburg-Marquardt [;9, Sec-
tion 4.2.2]. Also, the confidence level of the estimates has not beell stated 011 a quall-
titative basis.
As demonstrated in Section 5.6, there are occasions where parameters may be
estimated from the frequency response. It would be valuable to perform not only
single, but multiparameter estimations using the frequency response. An example of
the utility of such a routine would be to estimate both the layer thickness and complex
bulk permittivity of the evaporating oil layer mentioned in Section 5.2.1.

10.3 Complex Bulk Permittivity Of Oil

The microchip and macrochip have been successfully adapted to provide measurement.
of the complex bulk permitt.ivity of transformer oil. This effort required t.he combi-
nation of passivation of the electrode structures to block moisture adsorpt.ion, and
development of the continuum model and parameter estimation routines to allow ~st.i­

mation of the dielectric properties of the oil while correctly accounting for t.he dielectric

318
properties of the passivatiug medium. Parylene C was found to provide an excellent
moisture barrier and appears to be compatible with transformer oil.
In the presence of dry flowing oil, interfacial problems would arise when using the
passivated microchip. This can be avoided by placing the sensor in a minimal oil
flow region. Another option is to use the longer spatial wavelength of the macrochip
to reduce the sensitivity to localized effects at the interface. The problem with the
present macrochip, that ofinsensitivity in gain, is a design problem, not a fundamental
limitation. Therefore, it is felt that a redesigned passivated macro chip might prove to
be the best solution.
With this sensor, the opportunity now exists to gam expenence on monit.oring
the complex bulk permittivity of the oil. Therefore, the next step is t.o begin OIl-line
monitoring in the test facility. A database must be built up to determine whether
there is any useful information about the status of the oil to be gained from t.his
measurement. There is very little, if any, experience in the literature on monitoring
this quantity due to the lack of a simple measurement method (and perhaps lack of
interest as well).

10.4 Moisture Sensors

A design strategy for developing electrochemical sensors in general, and moisture sens-
ing in particular, using thin film microdielectrometry, was demonstrated. The concept
of coating an electrode structure with a thin film whose dielectric properties were
sensitive to the chemical constituent of interest in the oil. \,,'as successfully applied
to moisture sensing using a plasma deposited bromobenzene coating on a microchip.
Moreover, this effort represents the first time the moist.ure contf'nt. of t he oil has been
measured in Jitu. The coating behaved as a membrane ill that. t.he moist.ure con-
tent ill the coating was directly related t.o the absolute moist.me cont.ent. of t.he oil.
Further, the sensitivity to moisture in the oil, on the order of parts per milUon, pro-
duced measurable changes in the dielectric response of the coating when using the
microdielectrometer. This dielectric response displayed a shape that was invariant

319
with temperature and moisture content (as long as the relaxation in the coating oc-
curred at frequencies higher than the relaxation in the oil), producing a master ('urve
that allowed for easy calibration. Coating thickness was not seen as a crucial pa-
rameter, a range of 2-7 p.m in thickness appeared to be appropriate when using the
microchip. The film morphology, a biphasic structure of particles embedded in a rigid
film, was concluded to be responsible for the permeation of moisture into the film.
This moisture sensor is by no means a finished product, ready to be commercialized.
The critical problem is the lack of stability of the coatings over time and operating
temperature. More experimental results on coated microchips need to be performed
to characterize the dependence of aging on temperature and time. Controlled exper-
iments should be performed to verify that the coatings degrade only when immersed
in oil and not when exposed to the atmosphere. It would be useful to examine, under
an SEM, the coatings after aging in the oil over a long period of time. Perhaps a
change in the biphasic structure is resulting in the decreased moisture sensitivity. As
discussed in Section 8.3, there are several possible fabrication techniques that may
improve the coating stability. If the plasma deposited brolllobenzene proves inade-
quate, theti. other polymeric materials that have demonstrated better siability should
be tried. For example, plasma deposited HMDSO appears to be a material that has
been plasma polymerized to a high degree of stability and has been investigated by
many researchers. A major constraint has been the maximum temperature to which
the microchip can be subjected, due to the polyimide used on the chip itself and the
KaptC'n used to encase the microchip. This limits the number of deposition processes
that may be investigated. It may be helpful to either coat the microchip before the
polyimide is applied or coat a macrochip. The increased maximum temperature may
open the door to consideration of coating materials via alternat,ive processes. If t.he
macrochip is used, the coating should still be thin, on the order of micromet.ers, so as
to maintain a rapid transient response. In any case, the procedure for developing a
moisture sensor using other materials would be very similar t.o t.he pat.h t.aken ill t.his
wOJ.'k.
An important unknown, perhaps critically relat.ed to the aging issue, is the physi-

320
cal mechanism responsible for electrical conduction ill the bromobenzene. It has been
demonst.rated that the process is dispersive and references were made t.o seve-ral phys-
ical models for interpreting the dispersion, but more physical and chemical analysis of
the coating needs to be done. Many of the analytical techniques used for examining
plasma deposited films, described in Appendix C, remain to be applied to the study
of this material. Given the importance, in the literature, ascribed to the presence of
free radicals in the film, more attention should be given to the free radical concentra-
tion, its dependence on time, temperature and environment, and its relevance to the
electrical conduction process.
Oue possible alternative for moisture sensing would be to use the paper insulation
itself as the moisture sensitive material. Based on the discussion of Section 6.1, using
a small mass of paper in a large volume (mass) of oil, the moisture cont.ent of t.he
oil would determine the equilibrium paper moisture content. As shown by Li [23,
Section 5.31, the dispersion in the paper (Fig. 5.24) can be normalized to the moisture
content of the oil. That is, this dispersion call be normalized into a master curve, as
was done for the plasma deposited bromobenzene. Therefore, a calibration of moisture
content vs. JP can again be made. The most likely location for the moisture sensor
will be in the oil circulation loop, outside the main body of the transformer. In this
region, the oil temperature never gets extremely hot, thus the paper should not suffer
much degradation and remain relatively stable ill its properties.
Once the moisture content in the oil is measured in an on-line fashion, the next
step is to interpret the results. There are two modes of analysis here, one dynamic
and the other static. A primitive attempt at relating the moisture content in the
oil to the moisture content of the paper and the condition of the oil was outlined
in Section 6.2.1. Under normal operating conditions a stat.ic model, such as t.he one
represented by (6.5), is established, complet.e wit.h paramet.er values. In a dynamic
mode, on-line moisture measurements are made every 5 minutes, as demonstrat.ed
in Section 7.2. These measurements are then compared t.o predictions made using a
dynamic model describing the moisture distribution between the paper and t.he oil as
a function of one or more temperatures. This dynamic model mayor may not be t.he

321
same as the static one, depending on how rapidly the moisture redistributes within the
transformer. For the residuals computed in Sect.ion 7.2, t.he llloch·ls were t.aken t.o he
the same. However, other models, perhaps including information on the temperature
and oil flow distribution within the transformer, may be more appropriate. The point
is that the parameters used ill the static model must be related to the parameters
used in the dynamic one. Over a period of time on the order of 24 hours, significant
residuals will indicate, on a basic level, a change in the static model parameters. In
turn, this change represents the degree to which the moisture level deviates from
normal operation and indicates a possible change in the state of the paper or the oil.

10.5 Gas Sensors

With the decision to use FTIR spectroscopy for on-line gas analysis, the only relevant
gas of interest not detected using this apparatus is hydrogen. Two attempts were made
at developing hydrogen gas sensors based on knowledge obtained from the literature.
Doped tin oxide, known to be sensitive to reducing gases such as hydrogen, was de-
posited on a microchip. However, the results were not very conclusive. Several factors
may be at work here: 1) the low temperature limitation of the microchip, 2) lack of
proper doping levels using a noble metal such as palladium, 3) the tin oxide may need
to be sintered to produce the correct granular structure, and 4) better control of the
experimental temperature. Using a macrochip, rather than a microchip, perhaps the
considerations of 1) and 3) may be corrected. Another approach might be to capitalize
on the results using discontinuous ultrathin « 100 A) films of palladium, obtained
by Jelley and Maclay [47]. There, the sensitivity of the impedance of these films to
hydrogen at room temperat.ure, and carbon monoxide at. elevated f.empf'l'ntlll'es, was
demonstrated. Finally, an approach that parallels t.he sint.ered t.in oxide, but. uses
plasma deposited organomet.allic films inst.ead, may prove t.o be very adapf.ible t.o t.he
microchip. Starting with tetramet.hylgermaniulll and tet.ramet.hylt.in lllonomers, Ina-
gaki and Hashimoto plasma deposited these monomcrs onto nn intcrdigital structure
[114]. Heating these structures anywhere from 200-400°0, the DC resistance was mon-

322
itored as the ambient gas concentratioll was varied. Sensitivity to combustible gases
such as propane, hydrogen, methanc:a, Rnd carhon monoxide, was demonstrfttc:ad nt. con-
centrations of 0.1 molar % (1000 ppm). In addition, sensitivity was increased by the
vacuum evaporation of a thin film ('" 100 A) of palladium chloride on top of the plasma
deposited layer. Using the microdielectrometer, the impedance of these films at lower
temperatures may be observable alld provide a hydrogen gas sensor that operates in
a manner very similar to the plasma deposited bromobenzene moisture sensor. This
approach might prove to be better suited to adaptation to the microdielectrometer
due to the upper limit on the measurable surface conductivity of the thill metal oxide
films versus the required grain structure llecessary for gas sensitivity.
Preliminary work using a thin film (500 A) palladium resistor yielded encourag-
ing results. Sensitivity to H2 dowll to a 100 ppm level at temperatures ~ 1000 F was
demollstrated. More data lleeds to be takell to qualltify the sellsitivity to H2 concentra-
tion. Heating the device is important as moisture will corrupt the measurement. It i;s
believed that the sensitivity could be improved further, perhaps by an order of magni-
tude, using a bridge arrangement. This differential measurement, using all 4 palladium
resistors, would also automatically compensate for local temperature variations. An
improved barrier to prevent hydrogen from affecting the t.wo reference resistors will be
needed. Although the device is expected to be selective to hydrogen due to the nature
of the physical mechanism responsible for the change in resistance, this selectivity has
not been verified experimentally. It would also be interesting, though not crucial, to
test the device immersed in the oil.
The goal of developing microfabricated sensors, selectively tuned to the various
gases of interest and immersed in the oil, is a much longer term project. Some idea of
the large effort that is currently being given to developing met.hods of tlerivat.izing and
testing microelectrode structures using various redox active polymers can he oht.ained
from [115]. At present, the emphasis has been on st.udying t.he behaviour of redox poly-
mers when there is a chemical change ill the environment.. With t.he polymer suit.ably
specialized by a catalyst or elect.roactive center, t.he chemical change can be translat.ed .
into a physical change in the electron distribution within the polymer resulting in an

323
alteration of the conductivity or standard potential of the polymer. This resultant
electrical change can be monitored using a two or three terminal configurat.ion. In
general, there is a literature of redox polymers that interact with molecules such as
methane, carbon monoxide, carbon dioxide, and ethylene. This fact, coupled with
the concepts to modify electrodes with polymers consisting of the chemically sensitive
redox subunits, leads to the opportunity to fabricate a variety of chemically responsive
devices.

10.6 Electrode Terminations

It would be desirable to have the capability of operating the microdielectrometer in
either a high or low impedance mode. That is, being able to switch between a mea-
surement of a complex gain using the floating gate potential and a measurement of
the coupling admittance (impedance) using the current flow to a virtual ground. This
would remove the limitation on extracting information at gains near unity and provide
a broader view of the dispersive nature of materials such as the plasma deposited bro-
mobenzene. It would also allow a check on the linearity of the dielectric behaviour by
providing a larger range over which the Kramers-Kronig relationship may be verified.

10.7 Double Layer Interactions

One of the benefits of the m.icrofabricated nature of the microchip, coupled with the
modeling approach taken here, is to be able to study phenomena having a characteris-
tic length on the order of micrometers. An example of this is the diffuse double layer
which, for the insulating oils used in transformt'rs, can ht' anY"'ht'l"t' rrom 1 100 l'lll in
length. This charge separation layer, occurring at. t.he paper insulat.ion-oil interface,
has been cited as the source of charge giving rise t.o static electrificat.ion problems
[116,117]. Using a microchip at low field strengt.hs, it. may be possible t.o pert.urb t.he
double layer and observe the response. Experiments demonst.rating the feasibilit.y of
this approach are reported in Appendix p. Development of the appropriat.e model for

324
the perturbed double layer, using self consistent charge generation and recombination
terms, will provide better insight. into t.he structure of the douhle layer. l.Tnderst.R.nd-
ing the equilibrium at the paper-oil interface is a necessary step along the path to
explaining what occurs in an energized transformer when charge is injected int.o the
oil.

10.8 Applications to Other Physical Domains

Finally, it is not to difficult to imagine applying the concepts developed here to a
wide variety of applications, ranging from measuring a complex susceptibility using
current loops to measuring thermal properties using meandering resistors. In each
case, the combination of microfabrication techniques, the 'imposed w-k' approach, the
spectral solution of the resulting mixed boundary value problem, parameter estimation
techniques, and cC''1lputing power, allow absolute measurements of physical parameters
and observations of physical phenomena not previously available.

325
Appendix A

Rapidly Convergent Series

The individual terms in (2.22a) and (2.24a) that are to be summed over the iudice n
have the general form of

~ F. sin( no:) cos( n~)

~ 2 n (A.l)
n=l n
A more rapidly convergent series was used to reducing the computational time required
for this summation. Expression (A.l) can be expanded to the form of

~ H sin( n"Y )
~ n 2 (A.2)
n=l n
Assuming H n -t H min as n -t 00 the sum

II . ~ sin(1!"Y) (A.3)
mm~ 2
n=l n
will be computed and the difference between Hn and Hmin will be accounted for later
on. From Oberhettinger [118] comes the expression

fsin(~;r) = _ fo:L" III [2sill(B/2)] dB

n=l n Jo
(AA)

- -;r In[2] - fa'" In [Sill( B/2)] dB

Using a series expansion [119]

326
 (A.5)

where Bi are the Bernoulli .numbers, and substituting ill (AA) yields

10r [In[8/2]- ~l -l2m(2,;j)

~ sin{nz) 00
m+1B 8
2 2m ]
.t- n2 -z In[21 - d8
n=1

(A.6)

Due to the factorials present in the denominator in the summation term of (A6), only
5 to 10 terms need to be computed to obtain a high degree of accuracy (comparable
to computing the first 1000 terms of the original series). At this point, a calculation
of the terms for which Hn is significantly different from H min is performed and the
difference is added to the results of (A.6).

327
Appendix B

Surface Capacitance Density of

Multiple Layers

The medium may actually be comprised of a finite number of layers, each layer de-
scribed by a set of parameters, or layers may be used to approximate what is actually
a smoothly inhomogeneous material. With the objective of describing either of these
situations, suppose that the medium is composed of P layers, as shown in Fig. B.l.
The jth layer has a thickness dj and an upper surface designated by j. Surface proper-
ties of the P + 1 surface, the electrode-medium interface, are handled differently from
other continuum properties and are designated by €:o' There is no €:l as this surface
property is associated with the j = 1 surface and will normally be at infinity. Solution
of the field laws within the layer gives the transfer relations [75, p. 2.331,

(B.1)

wIlere D (j+1)
A
n an d DA (1+1)
n • Iy tie
are respective 1 comp 1ex amp l'It.tl(1es 0 f t. Ile 11t. I
1 F '
'ourlel'
components of the dielectric flux density and pot.ent.ial, evaluated just. above t.he lower
interface of the layer and i>~)' and 4>~)' are respectively t.hese quant.ities evaluat.ed just.
below the upper interface. In the case of a layer having a uniform complex permittivity
[75, p. 2.331,

328
 (B.2)
-(j) -(j) fjkn
An = -A21 = sinh(kndj )

At the jth interface, the potential is continuous, but there is a complex surface per-
mittivity f:j and hence a discontinuity in the dielectric flux.

~(j) - ~(j)' (B.3)

(B.4)

The surface capacitance density of the jth interface is defined in terms of the quantities
evaluated just above that interface.

(B.5)

From these equations, it follows that

A (j) • (j)
C(;+1)
A = A- (j)
22
+ A21 A12 (B.6)
n
- ) - AA ()') + k 2 f .,
C
( ')
n 11 n .)

This expression can be used repeatedly, starting from the top layer (j = 1) and working
down to the Pth layer. By definition, the surface capacitance density called for in
evaluating the complex gain is

(B.7)

In the case where the first surface is at iufinit.y, t.he lllut.ual t.el'1US in (B.6) are zero so
-(2) -(I)
that Cn can be evaluated without en . For the finit,e thickness layer bounded by an
infinite half space, P = 2 and (3.3) follows fro111 evaluation of (B.6) with C~:12' = A~~).
By flipping Fig. B.1 about the electrodes 180 0 the multiple layer <\ for the half
space below can be derived. The form of the solution is the same as above with the

329
only difference being several minus signs. Thus, below the electrodes the quantity
required is again (B.7) where P is now t.he number of layers below, and

_ (j) _ (i)
6(i+l)
n
= -A. (j)
22
+ _ (j)
A21,-,-A,-::1:..::.2_ _
• (j) (B.B)
en + All - k~€:j
• .(2) -(I)
where the Aij are as previously defined and On = - A22 •

The ground plane of the microdielectrometer can be accounted for by using a very
large surface conductivity at that. boundary. This will result in t.he expression given
in (3.1).

330
 j -

~ j'th surface

j j' t h I a yer

j =P

Figure B.1: Medium Above Electrodes Represented by a Multilayered Struct.ure of P

homogeneous layers.

331
Appendix C

PlasIDa Deposition

e.l Description of Polymerization Process

Plasma deposition is a coating technique that is attractive because it can produce
highly cross-linked thin films. Possible applications for these thin films have been as
dielectrics, optical coatings [120], semipermeable membranes (for gas separation and
reverse osmosis) [121,122], moisture barriers [123,124] and as an intermediary to im-
prove adhesion [125]. Plasma polymerization falls under the general heading of a glow
discharge deposition process, the specialization will be made clearer a little further
on. It is referred to as a thermally nonequilibrium deposition process because the
ions, electrons and neutrals comprising the plasma are at very different energy levels
(temperatures) than the roughly ambient temperature of the substrate. Formation of
thin films via this process was recognized as early as the mid-1870's, but only recently,
since the 1960's, has the advantages of such films been realized. Consequently, though
the studies of the process are quantitative now, there is a great deal of room for im-
provement and inllovation regarding processing techniques, predicting coat.ing quali t,y
and properties, and reproducing results. The following represent.s a brief int.roduct.ioll
into the area of plasma deposition. There are several good references surveying t.he
field over the years [126]-[129] and a good in-depth analysis provided by Yasuda [83].
In plasma deposition, a chamber containing the substrate to be coated, is evacu-
ated tO,roughly 35 p.m Hg to remove gaseous impurities. A monomer vapor, forming

332
the basic building block for the resulting polymer, is then drawn through the chamber,
accompanied by an increase in pressure. Either inductively or cf.pacit.ivdy, an <,lect.ric
field is produced within the evacuated chamber at a high enough field strength such
that fragmentation and ionization of the monomer occurs. This process causes the for-
mation of free radicals and ionized material. Diffusion of both to the su bstrate results
in the formation of a thin film. This is accompanied by a drop in operating pressure
as some of the vapor phase is being converted to solid phase. The thin film is actually
formed via two different reaction paths as outline by Fig. C.l. One mechanism is radi-
ation induced polymerization - the fragmented monomer diffuses to the substrate and
polymerizes after deposition, while in the solid phase. Another mechanism is glow dis-
charge or plasma polymerization, whereby the fragments nucleate in the vapor phase,
forming a polymer intermediate, and then diffuse to the substrate - undergoing further
crosslinking with the rest of the film matrix. This second mechanism is responsible for
the formation of microspheroids which are then embedded in the film matrix, yielding
a biphasic structure [95]. By varying operating parameters such as monomer flow rate,
pressure, power density and substrate temperature, powder (microspheroid) formation
can be controlled [83, Section 8.4.3]. Scaling specific operating parameters from one
process chamber to another is a difficult task. However, powder formation appears to
be enhanced by operating conditions such that a high density plasma meets a high
concentration of monomer in a localized space. It also has been shown that increasing
the substrate temperature can reduce powder formation [96]. This dependence was
attributed to thermal effects on: 1) chemical reaction rates, 2) desorption of loosely
bound constituents, 3) heat transfer from the substrate to the gas phase and 4) surface
mobility of the impinging species.

Due to these different film formation mechanisms, it is 110tewOl·thy t.hat. t.he end
product is very different from a conventionally polymerized mat.erial. Plasma poly-
merized styrene does not have the same characteristics as polystyrene, although bot.ll
start from the same monomer. An example of the fragmenting and resultant three
dimensional, highly cross-linked structure of a plasma polymerized ethylene film is
shown in Fig. C.2.

333
 r:'-'-'-'-'-'-'
I ISTARTING
-------
N-POLYMER- EFFWENT
MATERIAL. FORMING GAS GAS

I
I
,,
I I
I
.I
Plasma 01' I I ,
Glow DndIafge I
PoIym~iJatiOll
I : I
I I ,
I

\ I I I
I I I

1\
I I
I ,
Platma-lncluc.CI POLYMER-FORMING I I
Ablallan
INTERMEOIATES I ,
L--....-_-r-_---J I I
I ,
, I

I
I I
f I
I
I PLASMA
BOUNOARY

I I

Figure C.1: Overall Mechanism of Glow Discharge Polymerization (from [83, p. 180]).

Figure C.2: Model Structure of a Plasma Polymer Film of Ethylene (from [83, p. 154]).

334
 There are a several process parameters that are relevant. One major factor is the
geometry of the vacuum chamber, either a bell-jar or, more likely nowadays, t.ubular.
Included with this is the manner in which the discharge is produced, the location of
the monomer and inert carrier gas (optional) inlet relative to the discharge and the
location of the substrate within the chamber. Inductive coupling uses coils wrapped
around the outside of the chamber. Capacitive coupling uses parallel plates placed
inside the chamber. Other parameters are the flow rate of the monomer vapor and
an optional inert carrier gas, background pressure before introduction of the monomer
and operating pressure with the monomer flow. Varying the flow rate and pressure will
tend to favor one of the two growth mechanisms described above. ~~ t a given operating
power, films can vary from oily to rigid to powdery, as shown in Fig. C.3. This mapping
will also change with power density and substrate temperature, as indicated above.
The frequency, if AC, and power level of the discharge are also crit.ical. The majority
of the discharges are RF and performed at the FCC regulated frequency of 13.56 MHz.
If the power level is too high, the competing mechanism of film ablation may result
in etching of the substrate instead of film growth. Yasuda introduces the composite
parameter W /FM which combines the power level (W) in watts, mono mer flow rate
(F) in sccm and monomer molecular weight (M) in grams/mole. It is usually expressed
in J /kg by multiplying the former result by 1.34~109. Although this is st.ill a system
dependent parameter (chamber geometry) it shows that there is some relation between
changes in flow rate and power and allows some crosschecking with various monomers.

These parameters affect the deposition rate, quality, uniformity and electrical, me-
chanical and physicochemical properties of the film. A good summary of the empirical
relationship between these parameters and the film produced can be found in Jensen,
et al. [130] for polymerization of ethane, Wr6bel, et a1. for organosilicones [131,132]
and Yasuda [83, Chapter 9] for a wide variety of monomers. An obvious paramet.er
not mentioned previously is the monomer used. Generally, due to the fragment.at.ion of
the monomer and the atomic nature of the plasma polymerization, the resulting films
are categorized more by the family ill which the monomer is part of than the specific
monomer. The various families are hydrocarbons, nitriles, organosilicones, fluorocar-

335
 ,,
\

Unstable
\
,
4 .... dilCharge ,,
\
. ..
.. ... I
.. . ,I Film
i 3 .. . ... ,\
-=!
,.. .,. ..
' ... ,.. -" "
I
~,
..
,' .. . ..
,, ,
Go 2
......
, powder formltion
rlpid Very ..
,.
I
,,
\
\

.. ..... ,
. Powder
Indfilm
" ,
0
10 20 30 40 50 70
Ethyl... Flow RI. (c:mh"/minl

Figure C.3: Powder and Film Regions at 100 W using Ethylene MOllomer (from [83,
p.244]).

bons, silicones, organometallics and oxygen containing compounds. Hydrocarbons can

be further broken down into the various aromatics, chained, saturated, double and
triple bonded compounds. Almost any vapor, whether it is the natural phase of the
compound under STP or is pulled off of the liquid or solid phase, can be deposited. For
example, plasma polymerized films of methane have been formed and have exhibited
good adhesive qualities [125].
Various analytical tools are used to study the structure of these thin films. Elemen-
tal analysis is performed using NMR and mass spectrometry. Structural information
on bonding within the film is obtained using electron spectroscopy for chemical anal-
ysis (ESCA), X-ray photoelectron spectroscopy (XPS), IR and FTIR. spectroscopy. A
technique called electron spin resonance (ESR) is used to determine t.he concent.rat.ion
of free radicals [83, Section 6.4.2]. Free radicals are negatively charged subgroups, hav-
ing an unpaired electron, that form due to the fragmentation of the lllonomer [133].
They serve an important function in the growth of both conventional polymers and
plasma polymers. Many of the free radicals participate in the growth reactions of

336
film formation but there may be a substantial concentration of free radicals trapped
within the film after the deposition process. The quantity of t.rapped free radicals is a
strong function of the chemical structure of the monomer. Once trapped in the film,
these free radicals can affect the properties of the film such as electrical conduction
and permeability. The radicals are usually difficult to remove or eliminate but their
concentration can change slowly over time, contributing to the instability of the film
properties.

C.2 Experimental Apparatus

The electrical configuration of the apparatus is sketched in Fig. C.4. The capacitively
coupled RF system is made up of all RF gellerator, impedance matchillg tuner, voltage
multiplying coil and parallel plate electrodes. The RF generator (ENI Power Systems,
Inc., Model ACG-5) is capable of supplying up to 500W and is preset to operate
at 13.56MHz. An impedallce matching tuner (MFJ Enterprises, Inc., MFJ Versa
Tuner V, 3kW series, model MFJ 989) was used to tune the resonant circuit of the
electrode-coil system to 13.56MHz and thus cancel the imaginary contributions to the
load impedance of this system and the plasma. The unit needed to be tuned only once
and was performed under typical operating conditions. Once tuned (SWR as close to
unity as possible), the load presented to the generator was a pure 500, through proper
design of the voltage multiplying coil and plate electrodes. The voltage multiplying
coil was constructed from 1/4" copper tubing wound into a 13 turn (n) air core coil
with a coil radius (a) of 1" and extending a length (b) of 5.5". The ends of the coil are
connected to the electrodes, the coil is grounded at its midpoint and the RF power is
tapped in approximately 1 1/4 t.urns from f.he grollnd contact.. Thl1s the tnrtls ratio
yidds a voltage amplification of 10.4. The electrodes are 8"x12" (A) aluminum plat.es
spaced (cl) 2.5" apart. Using

337
 C
(C.I)
a2 n 2
L(JLH) = 9a + lOb
gives an inductance of 2.64 pH in parallel with a capacitance of 8.64 pF [134]. The re-
sulting resonant frequency would be 33.3 MHz, emphasizing the need for the impedance
tuner. This capacitance is not quite correct because there is a 8"x14"xl/16" aluminum
plate located midway between the electrodes and grounded. It is 011 this plate that
the substrates to be coated were mounted.
The mechanical configuration of the apparatus is shown in Fig. C.5. The main
vacuum chamber was a 9"x6" Corning pyrex glass reducing tee. The chamber was
sealed with 3/4" thick aluminum plates and teflon gaskets. All the aluminum plates
were grouuded. Vacuum pressure was monitored with an MKS vacuum gauge wit.h
digital readout (222AHS-MKS absolute vacuum gauge, PDR-C-IB MKS power sup-
ply) screwed into the plate that sealed the 6" diameter tee section. It is important to
choose a pressure gauge whose operation would be unaffected by the RF fields. Vacu-
ums down to 35 I'm Hg were obtained using a Welch Duoseal (modeI1397) mechanical
roughing pump connected to the chamber through a cold trap. A butterfly valve at the
bottom of the chamber allowed faster evacuation while the two Whitey valves allowed
coarse and fine control of the pressure during deposition. The monomer liquid was
placed in a flask immersed in an ice bath to prevent condensation of the vapors in the
tubing. Monomer flow rate was controlled by a Whitey valve at the upper inlet to
the chamber. Diffuser boxes at the inlet and outlet were used to provide an evenly
distributed monomer flow between the electrodes and to help prevent sneakage around
the electrodes. The entire device was enclosed by a wire mesh when fhe RP was on.
This was both for safety reasons and to prevent noise on the radio bands.

338
 ELECTRODES

VOLTAGE
A~PLlFVING ----.
COIL

IMPEDANCE
MATCHING
TUNER

RF
GENERATOR

Figure C.4: Electrical Configuration of 13.56 MHz Capacitively Coupled RF Plasma

Deposition Apparatus.

339
MONOMER-----~~

ICE BATH - - -..... - . GAS INLET

VALVE

~__+-__________ CENTERPLATE
FOR SAMPLE
DIFFUSER BOXES

PRESSURE
+- TRANSDUCER

COARSE &
FINE ----~~~-,
L-========= TO
RF POWER SUPPL V

VACUUM
VALVES COLDL.....-._.....J VACUUM
PUMP

BUTTERFLV
VALVE

Figure 0.5: Mechanical Configuration of Plasma Deposition Apparat.us.

340
C.3 Operational Procedure

Generally, only the upper aluminum plate was unsealed for placement and removal of
substrates, and cleaning of the chamber. The bottom and side plates were unsealed
and cleaned very infrequently.

1. If the vacuum chamber, electrodes or aluminum plate were coated ft'om a pre-
vious deposition then dean them with acetone. For an accumulation of coating
residues a stronger solvent, methyl chloride was more effective (but required
safety precautions for handling). A thin buildup did not seem to adversely affect
the process.

2. Using Scotch tape, affix the substrates to be coated onto the aiuminum plate.
Generally, the chips and witness strips were located 4" downstream from the
diffuser box at the inlet, 2.5" below the top of the electrodes, along a horizontal
band 3" wide centered about the middle of the 8" wide platt':.

3. Slide the aluminum plate into place between the electrode~ in the chamber.

4. Check to ensure that the electrodes are isolated frolD ground.

5. Using a high vacuum grease with low vapor pressure, fmch as Cello-Grease™,
liberally grease the upper lip of the chamber, both sides of the gasket and the
aluminum plate and reassemble. Hand tighten the flange bolts.

6. Make sure all the valves are closed, including the hutterfly valve. Fill the cold
trap with liquid nitrogen and start the vacuum pump. After allowing the tubing
leading to the chamber to be evacuated, open the butterfly valve to evacuate the
chamber.

7. Evacuate down to 35-50 pm Hg. Check the flange bolts and ret,ight,en if neces-
sary. It is important to get a good seal to prevent contamination by atmospheric
gases during deposition. Check the leak r<"te by closing the butterfly valve and
monitoring the rise in pressure. It should be less than 2-3 plD Hg per minute.

341
 Due to the large volume of the chamber, this took several hours to evacuate and
t.o form a good seal.

8. When the chamber is sufficiently well sealed, place the flask containing the
monomer in an ice bath. Slowly open the inlet valve, allowing the monomer
to degas, the tubing to evacuate and the pressure to drop back down to its
prior low level. If microchips are being coated, their high frequency gain in air,
used in computing the normalized capacitance, should be recorded now. Unfor-
tunately, the monomer flow rate usually was not monitored. Instead, a rough
estimate was made at this point by closing the butterfly valve and monitoring
the initial pressure rise. At a fixed volume and temperature, the system pressure
is proportional to the number of gas molecules present and thus the change in
pressure is proportional to the· change in the number of molecules. Therefore the
volume flow rate in torr-liter/sec equals the chamber volume times the initial
rate of pressure rise [135, p. 16]. Assuming the monomer flows mainly through
the straight section (radius of 4.5", length of 18") yields a volume of 18.8 L. A
typical pressure rise was 2.7 mT /sec giving a volume flow rate of 0.05 tL/s which
is 4 standard ccm.

9. Open the coarse control valve at the bottom fully, partially open the fine cont.rol
valve and dose the butterfly valve. By adjusting these control valves and the
inlet flow valve, let the chamber pressure equilibrate at roughly 200 I'm Hg.
Operating conditions are a function of the monomer used. The flow rate must
be high and the pressure low to obtain rigid films (as can be seen from Fig. C.3).

10. Switch on the AC power for the RF generator and increase t.he power level. By
flipping the power meter switch the forward and reverse power levels can be
checked. If the tuning is off t.here will be significant. reverse power refled.ed back
from the load. The plasma glow will start slowly as the ionization process begins
to buildup. This is accompanied by a decrease ill a pressure as t.he 1ll01l0mer
vapor starts nucleating, polymerizing and depositing. Usually a drop of 100
I'm Hg from pre-discharge occurred in normal operation. When coating at 25

342
 watts it was necessary to begin at 50 watts to 'get the plasma started' and
then turn the power back down to 25 watts after a minute or t.wo. The color
of the discharge with bromobenzene was quite purplish. This is in contrast to
discharges with air flowing through, which were more pink colored.

11. Leave the power on and coat for as long as desired. During first 1/2 hr. of coating
the power may rise 5 watts as the discharge establishes itself - reset the power
level if necessary. Check to make sure there is enough liquid nitrogen in the
cold trap and ice in the ice bath. When there is too little of either coolant the
pressure will start rising again. If this occurs turn the RF power off and replenish
the appropriate coolants, leaving everything else operational. Then restart t.he
discharge.

12. When finished coating turn off the RF generator. Close both pressure control
valves and the monomer inlet valve. Seal off the monomer flask to prevent any
leaking of vapors and disconnect from the vacuum chamber inlet.

13. If post-treatment is desired, Le. letting the coated samples remain in the monomer
vapor overnight, then leave the chamber sealed overnight. With a good seal, the
chamber should not rise above 4 mm Hg overnight. Close the valve connecting
the pressure gauge to the chamber as there is no need to expose the pressure sen-
sor to possible degradation through film formation. The next day go to step 14.
Due to the nature of a highly ionized and reactive surface upon removal of the
discharge, introduction of oxygen immediately afterwards may lead to oxidation
of the surface, resulting in unwanted surface properties regarding wettability.

14. If no post-treatment is desired then open t.he htttt.erfly valve and l'f"-f"VRCURt.e t.he
chamber for 1/2 hr. to remove the 1ll01l01ller vapor. Close t.he but.t.edly valve,
turn off the vacuum pump and slowly open t.he lllonomer inlet. valve and allow
chamber to reach ambient pressure. Open the chamber up and vent. the chamber
through an exhaust hood via ducting to remove any remaining "apors.

343
C.4 Experimental Observations For Coating Mi-

crochips
An initial concern with the deposition process was the danger of damaging the FET
connected to the floating gate when the plasma was on due to charge accumulating on
the electrodes. In fact, it was thought that the first several attempts were unsuccessful
because of this effect. It turned out that all but one of these coated chips worked fine
eventually, just not immediately after turning off the RF power. The leads of t.he
microchips were always left floating (electrically), without any ill effects. The only
chip actually damaged irreversibly was one coated immediately after the impedance
matching tuner was used for the first time and an unnecessarily high power level of
450 watts was applied. Most chips were not functional immediately after removal
of the discharge due to the large DC offset values associated with trapped charge
within the coating. These charges seemed to remain indefinitely while the chamber
was evacuated. However, upon exposure to air the charge was able to leak off, leaving
the coated chips usable. Two examples of this are shown in Figs C.6 and 0.7 for
chips Z13 ('"'" 1.2/Lm)and Z22 (3.5 /Lm). For Z13, room air was introduced into the
evacuated chamber immediately after the deposition process. However, Z22 remained
under vacuum for 3 days before the introduction of room air into the chamber. As the
offset decreases to a reasonable level the response goes to a low gain and zero phase
- just what is expected in t.he purely capacitively coupled regime. Note that both
polarities of offset were obtained. A glaring exception to this behaviour is the tracking
of the deposition process for Z4 (see Section 5.2.1). These measurements were made
with the RF off but the chamber was still evacuated. It is unclear as to the reason
why the offset was not a problem here.

An interesting phenomenom was observed before the realizat.ion t.hat. the coated
chips were undamaged and usable. In an attempt to reduce the charge reaching the
microchips, coatings were deposited on witness strips shielded with a metal mesh. The
thought was to allow air flow in a zero field region above the substrate. The result,
however, was the deposition of a powder inst.ead of a rigid film. Presumably, based

344
 -41.

-42 ..
" i, j .~r -
- s::
,.,
en
-43 ,It......"..,'.,
','
~
-44 ,
-45
o 10 20 30 40 50

•
ID
~
,., -0. 9
...,,,.'.......:....
,
~
,
.&

.s:::
~

.' . \~ •
-1.8
•
o 10 20 30 40 50

4~----~----~-----r-----.-----'

..
ID
~
c...
3~~--+-----~----~-----+-----4

2~~'-+-----~----~-----+-----4
c...
c
1 ~----~k---~----~-----+-----4

oL-~--~~__~~~
o 10 20 30 40 50
time (Min)
Figure C.6: Response at 1 kHz of Z13, a '" 1 lUll Thick Bromobenzene Coated Mi-
crochip, as Room Air was Introduced into the Evacuation Chamber Immediately After
Deposition.

345
 zZ2t1.
-35. 5 r------.-------r------r------.

-37 . 5 I---~~~

....
IG
~
- 39. 5 1---li----+----I-----+------1

-41. 5 ___--.....L....------I~--.....L....-----'
o 50 1.00 150 200

4)
th
IU
..s:;
-3 ~ ...

-6
,i
~
":..
.!
w,
S2.
wilt
-9
j'
-1.2
o 50 100 150 200

-0.3

~-3.3
4)
th
q..
, ~
~

t; -6. 3

j
-9.3
-
o 50 100
time (..un)
1.50 200

Figure C.7: Response at 10 kHz of Z22, a 3 ..5 plll Thick Bl'omobenzene Coated Mi-
crochip, as Room Air was Introduced into the Evacuat.ed Chamber.

346
on Fig. C.3, this was due to a very low flow rate underneath the mesh. In any event,
powder formation was always seen on the lower 1/3 of the cellter aluminum plate. This
gave an indication of the uneven distribution of deposition throughout the chamber.
Another issue was the adhesion of the coatings. After many months of testing
coated chips in transformer oil with temperature cycling and testing in vacuum ovens
at temperatures up to 115°C there appeared to be no problem with adhesion to the
electrodes. However, the coating did not adhere well to the smooth Kaptoll packaging
that encapsulated the microchips.

347
Appendix D

Experiments With Doped

Transformer Oil

In electrochemistry, the charge separation layer at an interface is generally described

in terms of a lumped parameter impedance (perhaps frequency dependent) and is asso-
ciated with polarization occurring at the electrodes [136, p. 345]. Using the microchip,
an attempt can be made to describe the double layer using more fundamental, phys-
ical parameters. As an example of tIus, a sample of transformer oil was doped with
ASA-3, a Shell antistatic agent [137]. The conductivity of the oil could be altered
over several orders of magnitude by varying the doping concentration. This change,
in turn, greatly affects the diffuse double layer tIuckness (AD) according to

(D.1)

where D is the molecular diffusivity of the ions in the fluid. At a high doping con-
centration, the concentration of ionizable species in the oil greatly increases, resulting
in a large increase in conductivity. This results in a thinning down of t.he double
layer thickness. These type of experiI1lents were first performed by l'vlouayad [73, Sec-
tion 5.3.2] and were repeated here. The frequency response of an uncoat.ed microchip
in a highly doped oil sample is shown in Fig. D.l. The relaxation observed is due to
the highly conducting oil. However, notice that the low frequency gain does not go to
unity, but levels out and then starts proceeding upwards again. This is accompanied

348
by another downturn in phase at the lowest frequencies. Another relaxation is occur-
ring here, having a phase peak below the freqnency range of t.he microdi«·l~ctrom~t,er.

This relaxation is attributed to polarization occurring at the electrodes and, as a first

attempt, is modeled using a thin, insulating layer adjacent to the electrodes, under-
neath the surrounding oil. First, the properties of the oil are estimated with data at
80 Hz, using a uniform medium model. The estimated oil permittivity and conduc-
tivity are 1.94:dO- ll F /m and 1.80:dO- 9 S/rn, respectively. Second, the double layer
thickness and permittivity are estimated using the two parameter estimation technique
described in Section 4.4. Here, the two datapoints are the high frequency gain using
the oil permittivity, and the low frequency gain at 0.015 Hz, where the plateau occurs,
using a very large permittivity to simulate a perfectly conducting surrounding medium.
This gives a double layer thickness of 484 A and permittivity of 1.04:z:10- 11 F /m. Now,
using the estimated bulk properties of the oil and the estimated layer thickness, the
bulk properties of the double layer are estimated using t.he data at 0.005 Hz, where
there is evidence of the second, lower frequency relaxation. The estimated double layer
permittivity and conductivity are 1.04:z:10- 11 F /m and 1.34:z:10- 13 S/m, respectively.
Note that the re-estimated bulk permittivity of the double layer agrees with the pre-
vious estimate obtained from different frequency datapoints. The resulting predicted
frequency response matches the data very well (Fig. D.l).

The effect on the frequency response of stirring the oil was minimal in this case,
indicating a thin double layer lying within the boundary layer created by the oil flow.
The oil flow was on the order of 5 cm/sec. The double layer thickness can be computed
using (D.l) and the Einstein relation relating diffusivity and mobility,

D kT
( 0.2)
b q
where the thermal voltage (kT / q) is 25 m V, and b is t.he ion mobilit.y. In hydrocarbon
liquids, b is taken to be (2:z:10- 11 /1]) m 2 /V -s where 11 is the absolut.e viscosit.y in
MKS units [75, p. 5.4]. For transformer oil at. room temperature, the viscosity was
measured to be 0.02 kg/m-sec. This gives a double layer thickness of 5,190 A. The
estimated double layer thickness is an order of magnitude less than computed from the

349
 zf!16. pzf'16
0
-5

-10
-15

-20
••11-=
~ -25

-30 El = 1.94:1:10-11 F /m
0'1 = 1.80:1:10-8 S/m

-35 ~=480A.
E2 = 1.04:1:10- 11 F /m

-40 0'2 = 1.34:1:10- 13 S/m

-45
-3 -2 -1 0 1 2 3 4

-10

-20
-30
tl -40
"
~
,.&::; -50
s:a..

-60
-70

-BO
-90
-3 -2 -1 0 1 2 3 4
log ~req

Figure 0.1: Measured (e) and Predicted (-) Response of all Uncoated Microchip in
Oil Highly Doped with ASA.

350
standard balancing of migration and diffusion currents. An important consideration
is the electric field strength within the double layer relative t.o t.he appliell AC decidc
field strength. The AC electric field strength is roughly 1 V divided by 12.5 J.Lm,
giving a field strength of 8x10 4 V/m. However, at the low frequencies the gain was
roughly -2.5 dB, implying a voltage difference of 0.25 V between the electrodes and
a field reduction by a factor of 4. The double layer field can be approximated by
dividing the thermal voltage by the layer thickness, giving a field strength of 5x1i)6 V/m
using the estimated thickness and a field strength of 5x10 4 V/m using the comput~d
value. Perhaps the discrepancy in estimated and computed layer t.hicknesses can be
explained by the following scenario. Given the relative field strengths, the applied
AC field is strong enough to disrupt the charges further out in the diffuse region but
only perturbs the equilibrium of the charges closer in. This might account for the
much thinner 'average' layer thickness. Part of the problem also might be in trying
to describe the space charge region ill terms of a complex bulk permittivity inst.ead of
more fundamental parameters such as charge carrier concentrations.
The response of an uncoated chip in oil having a lower concentration of do pant is
shown for both the stirred and unstirred case in Fig. D.2. Unlike the previous exam-
ple, there is a large effect produced by the stirring of the fluid. The stirred response
looks very much like it could be modeled just using the complex bulk permittivity of
the oil (after correcting for the large, positive DC offset that is responsible for the
upward shift of the gain curve). However, the unstirred response displays overlapping
relaxations, making it is impossible to determine the complex bulk permittivities and
double layer thickness here via a sequence of single and two paramet.er estimation
schemes, if indeed such a model is relevant in the first place. Perhaps using a multi-
parameter estimation routine based on frequency measllrement.s ('QuId produce t.hese
parameters. The apparent decrease in t.he bulk conductivit.y of the oil is due 1.0 t.he
decreased dopant concentration. Using t.he Micromet estimat.ed value for f/l at. 0.15 Hz
for the stirred experiment yields an oil conductivit.y of 8.0LrlO- 1:S S/m, significantly
more insulating than in the highly doped case. According to (D.1), the double layer
thickness here is 20 J.l.m, much larger than ill the previous caSf:. Therefore, it is not

351
surprising that the response here is susceptible to the effects of stirring, as observed.

352
 %£23.26
0

-9

-18
....s::cs

-
~ -Z7 ..... ~.
~ !:'lIIIIII.o

-36 la
-45
-3 -2 -1 0 1 2 3 4

o
,. tl ~~
.~
lr
!]'V

-30

-90
-3 -2 -1 0 1 2 3 4

0.5 -
~
.,
~ " 0
~
o .,... ~

..
-0.5

-1
-3 -2 -1 0 1 2 3 4
log Freq·
Figure D.2: Measured Response of all Ullcoated Microchip ill Oil Slightly Doped with
ASA. (0) - Stirred, (*) - Ullstirred).

353
Appendix E

Data Acquisition and Processing

Routines

Data Acquisition

Main Routines

WMOD.C - complete water module gOi~lg from time scan information to moisture
contents and residuals.
TSCAN.C - gets raw data with a time scan at a selected frequency.
FCHIP5.C - performs frequency scan, used to generate raw data for master curve.

Auxiliary Subprogram

NEWFUNCT.C - contains microdielectrometer subroutines, issues commands to

microdielectrometer and gets data back.

A uxiliary Files

MMET.H - specifies device port for microdielectrometer.

HEADER - contains the heading that goes at top of data files.
CHIPIDS - contains the chipid for each chip hooked up.
MCFP.PAR - contains relation between log(fp) and log(mc) for each miosture chip.
MCT.PAR - contains relation between log(mc) and liT for transformer.

354
 Data Stripping

PLTFILE.C - strips raw data from frequency scans produced by FCHIP5.C into
format suitable for XENIX plotting routine.
TPLTACM.C - strips raw data from time scans produced by WMOD.C in .acm files
into format suitable for XENIX plotting routine.
TPLTHST.C - strips processed data from time scans produced by WMOD.C in .hst
files into format suitable for XENIX plotting routine.
MSC.F - computes master curve spline function from raw data.
FPCALC.F - computes moisture content and residual from. latest reading.
SPLINE.F - Computes spline function from freq. response.
SPLINT.F - Performs spline interpolation to find fp.
BRENT.F - Finds function minimum (fp).

355
Table E.1: Directory of Sample Input Files and Program List.ings
Page Number
Program Sample Input File Listing
WMOD.C 357 363
TSCAN.C 358 386
FCHIP5.C 359 392
NEWFUNCT.C - 399
MMET.H 403
HEADER 403
CHIPIDS 403
MCFP.PAR 403
MCT.PAR 403
PLTFILE.C 404
TPLTACM.C 408
TPLTHST.C 412
MSC.F 361 416
FPCALC.F 362 419
SPLINE.F 422
SPLINT.F 423
BRENT.F 424

356
 WMOD.C

This program is a complete water modu~e, executing TSCAN.C alld converting raw
data to fp, moisture content and generating a residual in one routine. Requires NEW-
FUNCT.C, MMET.H, STDIO.H, SGTTY.H, MATH.H, TIME.H and STRING.H to
compile.

Input File (INWMOD)

illput number of channels to be monitored <CR>
loop: input channel # and freq. to monitor <CR>
input sampling interval (sec) <CR>

Sample Input File

2
2200.0
4 1000.0
300

Run in background with command: wmod <inwmod> 0 & <CR>

WMOD.C looks for: 1) file HEADER to get heading for data files, 2) file CHIPIDS
to get the proper chipid, 3) CHX.MC to get spline information for channel x, 4) file
MCFP.PAR for parameters describing relationship between log(mc) and log(fp) for
each channel having a moisture chip, and 5) file MCT.PAR for parameters describillg
relationship between log(me) and liT for the transformer. WMOD.C opens 4 files
per each channel monitored per day. These filenames are formed by cOllcatenating
twmmdd- (obtained from timestamp) with the channel # and either .acm, .rdg, .chr
or .hst. The .aem file holds all the raw data for the scan, the .rdg file contains t.he
raw data for the most recent reading. The .hst file holds some raw data and all the
converted data, the .chr file contains the most recent entry to the .hst file. WMOD.C
runs forever, opening a new file at the beginning of each day.

357
 TSCAN.C

This program monitors several chips, each at a specified frequency, sampling at a regu-
lar interval for a given period of time. Requires NEWFUNCT.C, MMET.H, STDIO.H,
SGTTY.H, TIME.H and STRING.H to compile.

Input File (INTSCAN)

input number of channels to be ~onitored <CR>
loop: input channel # and freq. to monitor <CR>
input generic filename (twmmdd-#) <CR>
input heading followed by ! (for operating conditions) <CR>
input sampling interval (sec) <CR>

Sample Input File

2
2 200.0
4 1000.0
tw0616-1
voltage is on, cycling xformer through loads
begin at 8:55AM - MZ

300

Run in background with command: tscan <intscan> 0 & <CR>

TSCAN.C also looks for a file CHIPIDS to get the proper chipid TSCAN.C opens 2 files
per each channel monitored. These filenames are formed by concatenating t.wmmdd-#
(as input) with the channel # and either .acm or .rdg. The .acm file holds all t.he raw
dab. for the scan, the .rdg file contains only the raw data for the most recent reading.
The .rdg file is in the proper form to be read by FPCALC.F routine which convert
from raw data to fp and moisture content.

358
 FCHIP5.C

This program runs a specified frequency scan for any amount of chips. Requires NEW-
FUNCT.C, MMET.H, STDIO.H, SGTTY.H, TIME.H and STRING.H to compile.

Input File (INFCHIP5)

input number of channels to be scanned < CR>
loop: input each channel # <CR>
input heading followed by ! (for operating conditions) <CR>
input minimum frequency <CR>
input maximum frequency <CR>
input # of measurements at each freq. <CR>
ascending or descending (for now make always ascending) <eR>
if more than 1 reading at each freq. then do many sweeps (0) or one sweep (a) < CR>

Sample Input File

2
1
2
voltage is on, cycling xformer through loads
begin at 8:55AM - MZ

1
10000
1
a

Run in background with command: fchip5 <infchip5> 0 & <CR>

Note - to generate data for master curve spline function computation the scan should
be done only once and done ascending

359
FCHIP5.C looks for a file caIled CHIPIDS to get the chipid #'s for each channel.
FCIIIP5.C opens 2 files per each channel scanned. These filenames are fm'lllf"d by
concatenating fwmmdd-# (from timestamp) with the channel # and either .acm or
.str. The .acm file holds all the raw data for the scan, the .str file contains the
information for MSC.F, a routine that computes the master curve spline function.

360
 MSC.F

TIlls program computes a master curve spline function from a frequency scan. Requires
BRENT.F, SPLINE.F, and SPLINT.F.

Inputs
input channel # <CR>
input generic filename (must be in form fwmmdd)! <CR>

Run with command: msc <CR>

MSC.F looks for the stripped data file as output from FCHIP.C. This file is a concate-
nation of fwmmdd-, the channel number and .stl'. This file contains only the frequency,
gain and phase for one scan, with frequency ascending from low to high.

361
 FPCALC.F

This converts raw data to fp. Requires SPLINT.F.

Inputs
input channel # <CR>
Run using command: fp < CR>

FPCALC.F reads most recent raw data from file TWCHX.RDG where X is the channel
# and this file is the one produced from TSCAN .C. FPCALC.F also reads CHX.MC
which is produced by MSC.F and contains the master curve spline function. FP-
CALC.F opens a .chr and .hst file, the .hst file containing the history of converted
values, the .c1u file containing the most recent entry to the .hst file.

362
,,.........................................,
,.,. .,
wmod.c .,
6/23/87

,. scans several channels, each at .,

,. an individual frequency.'
,. records accumulated data in .,
, • . acm file, present status in .,
, •• rdg file, fp t mc data .,
,. in .chr and .hst files .,
,. requires chx.mc for spline info .• '
,. mcfp.par for moisture into . • ,
,. mcT.par tor xtormer into •• /
,. and header tor tile heading .,
,. at midnight runs fChip6 and .,

,........................................,
/. recalibrates master curve ./

'include <stdio.h>
'include <math.h>
'include <sgtty.h>
'include <time.h>
'include <string.h>
'include "mmet.h"
'include "newtunct.c"

float sign(a,b)
float a,b;
{
i t (b>=O.O)
return(fabs(a»;
else
return(-tabs(a»;
}

int max(m,n)
int m,n;
{
it (m > n) return (m);
else return (n);
}

tloat tmax(m,n)
float m,n;
{
if (m > n) return (m);
else return (n);
}

363
 float fmin(m.n)
float m.n;
{
if (m > n) return (n):
else return (m);
}

1* computes fp from spline function *1

1* of master curve and input gain-phase value *1
1* reads chx.mc *1
float fpcalc(chan,x,g,y)
float g,y,x:
int chan;
{
float p[200] ,f[200].y2[200] ,gmid.pl,p2,tol,fest.fp;
extern float rtbis(),splint():
int N,i.kmid,klo,khi,k;
extern FILE *fopen();
FILE *fpd;
extern int fclose();
extern int fprintf();
char mcfl[20].dum[60],ext.done:

1* open .mc file and get spline information *1

sprintf(ext."chy'd.mc".chan);
strcpy(mcfl.ext.1):
fpd=fopen(mcfl. Ur") ;
fscan:f(fpd, "Y.f" ,igmid);
fscanf(fpd,"y'd",iN):
fscanf(:tpd,"Y.d" ,ikmid):
for (i=1:i<=N:i++) {
f s canf (fpd ,"Y.f Y.f Y.f 11 ,if [i] ,ip [i] ,iy2 [iJ ) ;
1* printf("Y.f Y.t y'f\n",f[il.p[il.y2[i]):
*1 }
fclose(fpd);

1* determine on which sid8 of rhas~ peak we're on *1

1* set bracketing interval from one end of scan *1
1* to 101. over fp *1
if (gmid<g) {
klo=1:
khi=kmid;
f [khi] =1.16;
}
else

364
 {
klo=kmid;
khi=N:
f [klo] = . 86 ;
}

p1=splint(f,p,y2,N.1.0)-y:
if (p1>=O.O) printf("problem - phase is lower than estimated peak\n");
/* printf(lpmin=Y.f\n",p1):
p1=splint(f.p,y2.N,f[klo])-y;
p2=splint(f,p.y2.N.f[khi])-y;
printf("brackets for routine\n");
printf("klo=Y.d freq=Y.9.3f delta phase=Y.7.2f\n".klo.f[klo].p1);
printf(IIkhi=1.d freq=Y.9.3f delta phase=1.7.2f\n".khi.f[khi].p2);
*/
/* use bisection method to find root */

tol=.01;
fest=rtbis(f[klo].f[khi].tol.y.f.p.y2.N);
if (fest>O.O)
fp=:x:/fest;
else
fp=:x:;
printf("fp=1.9.3f Bz\nn.fp):
return(fp);
}

/* bisection root finding algorithm */

/* from Numerical Recipes. p. 247 */
/* converted to C */

float rtbis(:x:1.x2.:x:acc.dc.f.p.y2.N)
float x1.x2.xacc.dc.p[200].f[200].y2[200];
int N;
{
int jmax.j;
float fmid.f1.dx.rtbis1.xmid.dx1;
jmax=40;
fmid=splint(f.p.y2.N.x2)-dc;
f1=splint(f.p.y2.N.x1)-dc;
if (f1*fmid>=O.O) return(O.O):
if (f1<O.O) {
rtbis1=x1;
dx=x2-x1j
}
else
{
rtbis1=x2;

365
 dx=xl-x2;
}
for (j=l:j<=jmax;j++){
dx=dx .... 6:
xmid=rtbisl+dx:
fmid=splint(f.p.y2.N,xmid)-dc:
if (fmid<O.O) rtbisl=xmid;
if «fabs(dx)<xacc) I I (fmid==O.O» {
/ ... printf("iter=Y,d fest=Y,f\n".j.rtbisl):
... / return(rtbisl):
}
}
printf("too many bisections\n ll ) :
}

/... spline interpolation routine */

/ ... from Numerical Recipes. p. 289 */
/... converted to C "'1

float splint(xa,ya.y2a.n.x)
float xa[200].ya[200].y2a[200],x;
int n;
{
float h.a.b.y:
int klo,khi.k;
klo=l:
khi=n;
while (khi-klo>l) {
k=(khi+klo)/2;
i f (xa [k] >x)
khi=k:
else
klo=k:
}
h=xa[khi]-xa[klo];
i f (h==O.O) {
printf(lIbad xa input\n")j
return(O.O);
}
a=(xa[khi]-x)/h;
b=(x-xa[klo] )/h;
y=a"'ya[klo] +b*ya[khi] +«a*a*a-a)*y2a[klo]+(b*b*b-b)*y2a [khi])*(h*h)!6;
return(y):
}

/* counts lines in a file between $'s ... /

countlines(fpa)

366
PILE *:tpa;
{
int n;
char Cj
n=O;
vhile«c=getc(fpa»!='.'){

}
c='i':
vhile«c=getc(fpa»!='.')
if (c=='\n') n++;
return(n-1);
}

1* spline computing function *1

1* from Numerical Recipes, p. 88 *1
1* converted to C *1
spline(x.y.n.yp1.ypn.y2)
int n;
float x[200].y[200].y2[200].yp1.ypn;
{
int nmax,i.k:
float u[200].p.sig.qn.un;
nmax=200;
if (yp1>.99E30) {
y2[1] =0.0;
u[1] =0.0:
}
else
{
y2 [1] =( -0.6):
u[1] =(3. 0/(x[2] -x [1]) )*( (y[2] -Y[1] )/(x[2] -x[1] )-yp1) :
}
for (i=2:i<=(n-l);i++){
sig=(x[i]-x[i-l])/(x[i+1]-x[i-1]):
p=sig*y2[i-1]+2.0;
y2[i]=(sig-1.0)/p;
u[i]=(6.0*«y[i+l]-y[i])/(x[i+1]-x[i])-
(y [i] -y [i-1] )/(x [i] -x[i-1]» I (x [i+l] -x [i-l])-
sig*u[i-1])/p;
}
if (ypn>.99E30) {
qn=O.O;
un=O.O;
}
else

367
{
qn=0.6; .
un=(3.0/(x[n]-x[n-l]»*(ypn-(y[n]-y[n-l])/(x[n]-x[n-l])):
}
y2[n]=(un-qn*u[n-l])/(qn*y2[n-l]+1.0):
for (k=n-l:k>=l;k--) y2[k]=y2[k]*y2[k+1]+u[k];
returnj
}

1* minimization routine *1
1* from Numerical Recipes, p. 263 *1
1* converted to C *1
float brent(ax,bx,cx,tol,xmin,f,p,py2,N')
int N:
float ax,bx,cx,tol,*xmin,f[200],p[200],py2[200]j
{
extern float splint() ,fmin() ,fmax(),sign();
int itmax,iter;
float cgold,zeps,a,b,u,v,v,x,e,d,fx,fv,fu,fw;
float xm,toll,tol2,r,q,pl,etemp;
itmax=1000;
cgold=.3819660;
zeps=1. OE-l0;
a=fmin(ax,cx);
b=fmax(ax,cx):
v=bx:
v=v:
x=v;
e=O.O;
fx=splint(f,p,py2,N,x);
printf("initial freq='l.f\n",x);
printf("p='l.t\n",fx);
fv=fx;
fv=fx;
for (iter=l;iter<=itmax;iter++){
xm=0.6*(a+b);
toll=tol*fabs(x)+zeps;
tol2=2.0*tol1;
1* printf (lIh.bs (x-xm) ='l.t\n" ,fabs(x-xm»;
printf ("tol2=Y.f halfdif=Y.f\n", tol2, .5* (b"a» ;
printf("toler='l.f\n ll ,tol2-0.5*(b-a»;
*1 if (fabs(x-xm)<=(tol2-0.5*(b-a») {
*xmin=x:
return(fx) ;
}
if (fabs(e»toll) {
r=(x-v)*(fx-fv);

368
q=(x-v)*(fx-fw);
pl=(x-v)*q-(x-w)*r;
q=2.0*(q-r);
if (q>O.O) pl=(-1.0)*plj
q=fabs(q);
etemp=e;
e=d;
if «fabs(pl»=fab~(.S*q-etemp»II(pl<=q*(a-x»II(pl>=q*(b-x») goto alpha;
d=pl/q;
u=x+d;
if «(u-a)<to12) I I «b-u)<to12» d=sign(toll,xm-x);
goto beta;
}
alpha:
if (x<=xm)
e=a-x;
else
e=b-x;
d=cgold*e;
beta:
if (fabs(d»=toll)
u=x+dj
else
u=x+sign(toll,d);
fu=splint(f,p,py2,N,u)j
/* printf(liter=1.d\n",iter)j
printf("guess freq=1.f\n".u)j
printf(lpmin=Y.f\n".fu);
*/ if (fu<=fx) {
if (u>=x)
a=x;
else
b=x;
v=w;
fv=fw;
w=x;
fw=fx;
x=u;
fx=fuj
}
else
{
if (u<x)
a=uj
else
b=u;
if «fu<=fw) I I (w==x» {
v=w;

369
fv=fw:
w=u:
fw=fu:
}
if «fu<=fv)I I (v==x) I I (v==w» {
v=u:
fv=fu;
}
}
}
printf("brent exceed maximum iterations\n");
*xmin=x;
return(fx);
}

1* calculate the second derivatives for *1

1* spline interpolation and fp for master curve *1
1* important - data must be input from high to low freq. *1
1* reads data from fwmmdd-x.str *1
1* writes to chx.mc *1

msc(chan)
int chan;
{
extern FILE *fopen();
extern int fclose();
extern int fflush():
extern int fprintf();
extern float brent(),splint();
extern struct tm *tmstr:
extern float id[10]:
FILE *fpmc,*fpstr;
float g[200] ,p[200] ,f[200],py2[200] ,imean,pmin,fp,tol;
float gy2[200],gmid:
int N,i,klo,khi,kmin,k:
char dfile[20],o1[20],ext,c;

1* make .str filename (fwmmdd) *1

strcpy(dfile, Ill') ;
if «*tmstr).tm_mon<10)
sprintf (ext. "fwO'l.d" , (*tmstr) . tm_mon +1) ;
else
sprintf(ext, "fw'l.d", (*tmstr). tm_mon+O:
strcat(dfile,ext,1);
if «*tmstr).tm_mday<10)
sprintf(ext,"O'l.d-'l.d.str".(*tmstr).tm_mday,chan):
else

370
sprintf(ext,"y'd-y'd.str",(*tmstr).tm_mday,chan);
strcat(dfile,ext,1);
,* printt("tname tor msc = 1,s\n",dtile)i ,

,* read in data, requires increasing order in x coordinate *,

fpstr=topen(dfile,"r");
N=countlines(fpstr);
fclose(fpstr);
printf(". of lines in Y.s= y'd\n".dfile.N);
fpstr=fopen(dfile."r");
while«c=getc(fpstr»!='.'){

}
for (i=1;i<=Nii++) {
f s canf (fpstr • "1.f l.f Y.f 11 • if [i] , ig [i] , ip [i] ) ;
,* printf(tly'f Y.f Y.f\ntl.f [i] .g(i] ,p[i]);
*, }
fclose(fpstr) ;

,* calculate second derivatives for spline routine *,

,* use natural spline so set endpt. second derivatives to 0 *,

spline(f,p,N,1E31,1E31,py2);
spline(f.g,N.1E31,lE31.gy2);

,* now find fp, where phase is at m1n1mum *,

,* using Brent's search routine *,

,* first compute bracketing interval *,

kmin=1;
pmin=p(1] ;
tor (k=2;k<=N:k++){
if (p [k] <pminH
pmin=p[k];
kmin=k;
}
}

klo=kmin-1:
khi=kmin+l ;
printf("brackets for routine\n");
printf("klo=1.d freq=1.f\n",klo.t[klo]);
printf(tlkhi=y'd freq=Y.f\ntl,khi,f[khi]);

,* compute fp by searching for phase minimum *,

,* and also determine gain at fp *,

371
fmean=f[kmin];
tol=.OOl:
pmin=brent(f[klo],fmean,f[khi],tol,tfp,f,p,py2,N);
gmid=splint(f,g,gy2,N,fp);
printf("fp=Y.f gmid=Y.f pmin=y'f\n",fp,gmid,pmin);

/* write to tile phase, normalized freq and second derivs. */

/* open chx.mc */

strcpy(ol,"ch"):
sprintf(ext,"y'd.mc",chan):
strcat(ol,ext,l);
fpmc=fopen(ol,"w");
fprintf(fpmc,"1.6.2f\n",gmid);
fprintf(fpmc,"1.d\n",N):
fprintf(fpmc,"Y.d\n",klo+l);
for Ci=l;i<=N;i++) fprintf(fpmc,"Y.i2.6f Y.7.2f y'13.6f\n",f[i]/fp,p[i],py2[i]);
fclose(fpmc);
return;
}

/* performs frequency scan at midnight */

int fchip6C)
{
extern FILE *fopen(),*fpo,*fpi;
extern int fclose(),port;
extern int fprintf();
extern struct tm *tmstr;
extern float id[10];
extern int nscan.q3[10] .nchan.computeC);
extern char bufter[aO] .s5[80];
FILE *fpd. *tpdl :
char c,q2.o1[20].ext.dfile[20];
float tem.fmin.fmax,qf[9].x.scale,f;
int j,k,l.m.n.rep.num.all.i.ql:

1* make new filename (fwmmdd) *1

strcpy(dfile,""):
if «*tmstr).tm_mon<10)
sprintf (ext. "fw01.d" , (*tmstr) . tm_mon+1) ;
else
sprintf (axt, "tw1.d" • (*tmstr) . tm_mon+1) ;
strcat(dfile.ext.l);
if «*tmstr).tm_mday<10)
sprintf(ext,"01.d".(*tmstr).tm_mday);

372
else
"printf(ext, u1.d u , (*tmstr) .tm_mday) j
strcat(dfile,ext,1)j
1* printf(Ufname = l.s\n U,dfile)j *1
for (i~1ji<=nscanji++){
strcpy(01,dfile)j
sprintf(ext."-'l.d.acm u .q3[i])j
strcat(01,ext,1)j
fpd=fopen(01. uv U)j
fprintf(fpd. U\t\t\t\t1.s\n\n u .o1);
1* printf(Ufilename=1.s\n".o1); *1
fclose(fpd);

1* this just opens .str file the first time *1

strcpy(01.dfile);
sprintf (ext, u-1.d. str U, q3 [i] ) ;
strcat(01.ext,1)j
fpd=fopen(01. Uv");
fprintf(fpd,U$\n U):
fclose(fpd)j
}

1* get heading *1
for (i=1;i<=nscanji++){
strcpy(01,dfile):
sprintf(ext. u-'l.d.acm U,q3[i]);
strcat(01,ext.1);
fpd=fopen(01,U a U);
fpd1=fopen(Uheader U,Ur U);
sprintf(buffer,UU):
vhile (*buffer != '!'){
fscanf(fpd1,u'l.su,buffer);
fprintf(fpd,"'l.s u,buffer);
}
fclose(fpd1);
fclose(fpd);
}

1* set frequency range *1

fmin=1.0;
fmax=10000.0:
fmin=fmin-.0001;

1* no. of measurements at each freq. *1

373
m=1:

/. for the present make all scans ascending ./

/. this makes operation of fp routine simpler .,

q1=1:
rep=1:
if (m>1)
{
printf(lIspecify mode (0 or a)\n\n"):
printf("o - do m sweeps of freq. range\n"):
printf("a - take m readings at each freq., sweep once\n"):
scanf(l1.s",tq2):
if (q2=='a')
rep=1:
else
rep=m:
}

,. get chipids written in file .,

,. and get initial temperatures .,

for (i=1:i<=nscan:i++)
{
multi(q3[i]) :
chipid(id[q3[i]]):
chiptemp(ttem):
printf("Channel 1.1d \tTemperature= 1.4.2f Celsius\n",q3[i],tem):
strcpy(01,dfile):
sprintf(ext,"-1.d.acm",q3[i]):
strcat(01,ext,1):
fpd=fopen(01. 11 a" ) ;
fprintf(fpd,"\n\n\t\tChipid # : 1.6.2f\n".id[q3[i]]):
fprintf(fpd,"\t\tTemperature= 1.4.2f Celsius\n".tem);
fprintf(fpd,"\nfreq\t\ttemp\tchan\tgain\tphase\te1\te2\toffset\n$\n"):
fclose(fpd);
}

.,.,
printf("this is a wasted reading\n"):
freq(1000 .0); /* this is
measure(tf,tf); ,. a wasted
freq(O.O);
convert(f.f.f.tf,tf): ,. reading .,
printf(" •••• *••••••••••••••••••••••••••• \n"):

/. subroutine compute() is used ./

/. to take actual readings and .,
/. write them into the correct .,

374
1* this is ascending scan *1
if (ql==l){
qf [0] =6.0:
qf [1] =6.0:
qf [2]=8 .0;
qf[3]=10.0:
qf[4J=16.0;
qf[6J=20.0:
qf [6] =26.0:
qf [7] =30 . 0 ;
qf[8]=40.0:
for(k=l:k<=rep:k++){
scale=.OOOl:
for(i=(-3);i<=3;i++){
scale=scale*10.0;
for(j=O:j<=8:j++){
f=qf [j] *scale:
if «f<=fmax)tt(f>=f~in-.001»{
if (rep==l){
for(num=l;num<=m;num++){
for (i=l;i<=nscan;i++){
multi(q3[i]) ;
compute(i.f.dfile):
}
}
}
else {
for (i=l:i<=nscan:i++){
multi(q3[i]) :
chipid(id[q3[i]]):
compute(i.f.dfile):
}
}
}
}
}
}
}

1* this is for descending scan *1

else {
qf[0]=10.0;
qf [1] =8.0;
qf [2] =6.0:
qf [3] =6 .0:
qf[4]=4.0;

375
qf[S]=3.0:
qf[6]=2.S:
qf [7]=2 .0:
qf[8]=1.S:
for(k=l;k<=rep;k++){
Bcale=10000.0;
for(i=3;i>=(-3);i--){
scale=scale/l0.0;
for(j=0;j<=8;j++){
f=qf [j] *scale;
if «f>=fmin)ii(f<=fmax»{
if (rep==l){
for(num=l;num<~:num++){
for (i=l:i<=nscan:i++){
multi(q3 [i]);
chipid(id[q3[i]]):
compute(i.f.dfile);
}
}
}
else {
for (i=lii<=nscan;i++){
multi(q3 [i]):
chipid(id[q3[i]]):
compute(i.f.dfile);
}
}
}
}
}

}
}
for (i=l;i<=nscan:i++){
strcpy(ol.dfile);
sprintf(ext."-Yod.acm n .q3[i]);
strcat(ol,ext,l);
fpd=fopen(ol, na n);
fprintf(fpd,lI$\n ll ) :
fclose(fpd);
strcpy(ol,dfile);
sprintf(ext, II-Yod. str ll , q3 [i] ) ;
strcat(ol,ext.l);
fpd=fopen(ol, nail);
fprillt:i(fpd,II$\nn):
fclose(:fpd) ;
}
printf("output in files Yos#.acm\nll,dfile);

376
 [88} R.J. McGill, Mea"uremeni and Control in Papermaking, Adam Hilger Ltd., Eng-
land, 1980.
[89] R. Jeffries, "The Sorpt.ion of Water by Cellulose and Eight. Other Textile Poly-
mers", J. of the Textile Inst. Trans., Vol. 51, No.9, pps. 339-374,1960.
[90] L. Szepes, K. Torkos, R. Dobo and A. Szekely, "A New Analytical Method for
the Determination of the Water Conent of Transformer Oils", IEEE Trans. on
Electr. Insul., Vol. EI-17, No.4, pps. 345-349, August, 1982.
[91] A. Benatar, Effect" of Surface Jvladification of Graphite Fibers by Pla"ma Treat-
ment on the Hygral Behaviour of Compo.,ites, SM Thesis, ME~ MIT, Cambridge,
MA, May, 1983.
[92] J. Tyczkowski and M. Kryszewski, "Electrical Conduction in Plasma-Polymer-
ized Organosilicon Films: Influence of Water and Oxygen", J. Appl. Polym. Sci.:
Appl. Polym. Symp. 38, pps. 149 161, 1984.
[93] S.M. Gasworth, J .R. Melcher and M. Zahn, "Flow-Induced Charge Accumula-
tion and Field Generation jll Thin Insulating Tubes", Con£. on Interfacial Phe-
nomena in Practical Insulating Systems, NBS, Gaithersburg, MD, Sept. 19-20,
1983.
[94] A.F. Howe, "Diffusion of Moisture Through Power-Transformer Insulation",
Proc. lEE, Vol. 125, No. 10, pps. 978-986, Oct., 1978.
[95} J. Grebowicz, T. Pakula, A.M. Wrobel and M. Kl'yszewski, "Electron Microscopy
Studies of Plasma-Polym~rized Organosi1icon Thin Films", Thin Solid Films,
Vol. 65, pps. 351-359, 1980.
[96] M.R. Wertheimer, J.E. Klemberg-Sapieha and H.P. Schreiber, "Advances in
Basic and Applied Aspects of Microwave Plasma Polymerization", Thiu Solid
Films, Vol. 115, pps. 109-124, 1984.
[97] T .S. Ramu, M.R. Wertheimer and J .E. Klemberg-Sapieha, "Dielectric Properties
of Plasma-Polymerized Hexamethyldisiloxane Films: 1 Complex Permittivit.y",
IEEE Trans. on Electr. Insul., Vol. EI-21, No.4, pps. 549-556, Aug., 1986.
[98] V.V. Daniel, Dielectric Relazation, Academic Press, NY, 1967.
[99] D.R. Day, T.J. Lewis, ILL. Lee a.nel s.n. Sf"ntltl'ia, "The note of Donndnl'Y Lnyel'
Capacitance at Blocking Electrodes in I.he Int.erpretation of Dielecf.rie ('me Data
in Adhesives", J. Adhesion, Vol. 18, p. 18, 1973.
[100] H. Scher, "Dispersive Transport and Relaxat.ion", Proc. 2nd lIlt.. Con£. 011 ('ond.
and Breakdown in Sol. Diel., IEEE/Elect.. lustll. Soc" pps. 13-21, 1986.
[101] H. Scher and E. Montroll, "Anomalous transit-time dispersion ill amorphous
solids", Phys. Rev. B, Vol. 12, No.6, pps. 2455-2477, Sept., 1985.

590