0% found this document useful (0 votes)

109 views198 pages

Lithography Process Control

Uploaded by

lcy1823626238

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

109 views198 pages

Lithography Process Control

Uploaded by

lcy1823626238

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

You are on page 1/ 198

lithography Process

Control

Downloaded from SPIE Digital Library on 24 Feb 2010 to 130.60.68.45. Terms of Use: http://spiedl.org/terms
TUTORIAL TEXTS SERIES
• Practical Applications of Infrared Thermal Sensing and Imaging Equipment, Second Edition,
Herbert Kaplan, Vol. TT34
• Fundamentals of Machine Vision, Harley R. Myler, Vol. TT33
• Design and Mounting of Prisms and Small Mirrors in Optical Instruments, Paul R. Yoder, Jr.,
Vol. TT32
• Basic Electro -Opticsfor Electrical Engineers, Glenn D. Boreman, Vol. TT31
• Optical Engineering Fundamentals, Bruce H. Walker, Vol. TT30
• Introduction to Radiometry, William L. Wolfe, Vol. TT29
• Lithography Process Control, Harry J. Levinson, Vol. TT28
• An Introduction to Interpretation of Graphic Images, Sergey Ablameyko, Vol. TT27
• Thermal Infrared Characterization of Ground Targets and Backgrounds, Pieter A. Jacobs,
Vol. TT26
• Introduction to Imaging Spectrometers, William L. Wolfe, Vol. TT25
• Introduction to Infrared System Design, William L. Wolfe, Vol. TT24
• Introduction to Computer-based Imaging Systems, Divyendu Sinha, Edward R. Dougherty,
Vol. TT23
• Optical Communication Receiver Design, Stephen B. Alexander, Vol. TT22
• Mounting Lenses in Optical Instruments, Paul R. Yoder, Jr., Vol. TT21
• Optical Design Fundamentals for Infrared Systems, Max J. Riedl, Vol. TT20
• An Introduction to Real-Time Imaging, Edward R. Dougherty, Phillip A. Laplante, Vol. TT19
• Introduction to Wavefront Sensors, Joseph M. Geary, Vol. TT18
• Integration of Lasers and Fiber Optics into Robotic Systems, Janusz A. Marszalec, Elzbieta A.
Marszalec, Vol. TT17
• An Introduction to Nonlinear Image Processing, Edward R. Dougherty, Jaakko Astola, Vol.
TT16
• Introduction to Optical Testing, Joseph M. Geary, Vol. TT15
• Sensor and Data Fusion Concepts and Applications, Lawrence A. Klein, Vol. 'F F14
• Practical Applications of Infrared Thermal Sensing and Imaging Equipment, Herbert Kaplan,
Vol. TT13
• Image Formation in Low-Voltage Scanning Electron Microscopy, L. Reimer, Vol. TT12
• Diazonaphthoquinone-based Resists, Ralph Dammel, Vol. TT11
• Infrared Window and Dome Materials, Daniel C. Harris, Vol. TT10
• An Introduction to Morphological Image Processing, Edward R. Dougherty, Vol. TT9
• An Introduction to Optics in Computers, Henri H. Arsenault, Yunlong Sheng, Vol. TT8
• Digital Image Compression Techniques, Majid Rabbani, Paul W. Jones, Vol. TT7
• Aberration Theory Made Simple, Virendra N. Mahajan, Vol. TT6
• Single-Frequency Semiconductor Lasers, Jens Buus, Vol. TT5
• An Introduction to Biological and Artificial Neural Networks for Pattern Recognition, Steven
K. Rogers, Matthew Kabrisky, Vol. TT4
• Laser Beam Propagation in the Atmosphere, Hugo Weichel, Vol. TT3
• Infrared Fiber Optics, Paul Klocek, George H. Sigel, Jr., Vol. TT2
• Spectrally Selective Surfaces for Heating and Cooling Applications, C. G. Granqvist,
Vol. TT1

Harry J. Levinson

Tutorial Texts in Optical Engineering

Volume 1128

Donald C. O'Shea, Series Editor

Georgia Institute of Technology

SPIE OPTICAL ENGINEERING PRESS

A Publication of SPIE —The International Society for Optical Engineering
Bellingham, Washington USA

Levinson, Harry J.
Lithography process control / Harry J. Levinson.
p. cm. – (Tutorial texts in optical engineering; v. TT28)
Includes bibliographical references and index.
ISBN 0-8194-3052-8
1. Semiconductors—Etching. 2. Microlithography—Quality control. I. Title. II. Series.
TK7871.85.L464 1999
621.3815'2—dc2l 98-44433
CIP

Published by

SPIE—The International Society for Optical Engineering

P.O. Box 10
Bellingham, Washington 98227-0010
Phone: 360/676-3290
Fax: 360/647-1445
Email: spie @spie.org
WWW: http://www.spie.org/

Copyright © 1999 The Society of Photo-Optical Instrumentation Engineers

All rights reserved. No part of this publication may be reproduced or distributed

in any form or by any means without written permission of the publisher.

Printed in the United States of America.

Downloaded from SPIE Digital Library on 24 Feb 2010 to 130.60.68.45. Terms of Use: http://spiedl.org/terms
INTRODUCTION TO THE SERIES
The Tutorial Texts series was begun in response to requests for copies of SPIE
short course notes by those who were not able to attend a course. By policy the
notes are the property of the instructors and are not available for sale. Since
short course notes are intended only to guide the discussion, supplement the
presentation, and relieve the lecturer of generating complicated graphics on the
spot, they cannot substitute for a text. As one who has evaluated many sets of
course notes for possible use in this series, I have found that material
unsupported by the lecture is not very useful. The notes provide more
frustration than illumination.
What the Tutorial Texts series does is to fill in the gaps, establish the
continuity, and clarify the arguments that can only be glimpsed in the notes.
When topics are evaluated for this series, the paramount concern in
determining whether to proceed with the project is whether it effectively
addresses the basic concepts of the topic. Each manuscript is reviewed at the
initial state when the material is in the form of notes and then later at the final
draft. Always, the text is evaluated to ensure that it presents sufficient theory to
build a basic understanding and then uses this understanding to give the reader
a practical working knowledge of the topic. References are included as an
essential part of each text for the reader requiring more in-depth study.
One advantage of the Tutorial Texts series is our ability to cover new fields
as they are developing. In fields such as sensor fusion, morphological image
processing, and digital compression techniques, the textbooks on these topics
were limited or unavailable. Since 1989 the Tutorial Texts have provided an
introduction to those seeking to understand these and other equally exciting
technologies. We have expanded the series beyond topics covered by the short
course program to encompass contributions from experts in their field who can
write with authority and clarity at an introductory level. The emphasis is
always on the tutorial nature of the text. It is my hope that over the next few
years there will be as many additional titles with the quality and breadth of the
first ten years.

Donald C. O'Shea
Georgia Institute of Technology

Preface / ix

CHAPTER 1 INTRODUCTION TO THE USE OF STATISTICAL PROCESS CONTROL IN

LITHOGRAPHY / 1

1.1. The assumptions underlying statistical process control / 2

1.2. The properties of statistical process control / 3
1.3. Situations in lithography where statistical process control cannot be
applied naïvely / 6
1.4. Non-normal distributions / 10
1.5. Process capability / 20

CHAPTER 2 SAMPLING / 27

2.1 Choosing the proper sample size / 27

2.2 Measurement location considerations / 33
2.3 Correlations / 35
2.4 Measurement frequency / 37
2.5 Systematic sources of variation / 37

CHAPTER 3 SIMPLE AND COMPLEX PROCESSES / 43

3.1 Definitions / 43
3.2 Why test wafers are useful / 44
3.3 How to address complex processes in lithography / 46
3.4 Distinguishing between layer-specific and equipment-specific effects / 50

CHAPTER 4 LINEWIDTH CONTROL / 57

4.1 Cause and effect / 57

4.2 Independent variables / 59
4.2.1 Exposure dose / 63
4.2.2 Resist thickness / 66
4.2.3 Focus / 67
4.2.4 Bake temperatures / 72
4.2.5 Resist development / 74
4.2.6 Humidity / 75
4.2.7 DUV resists — special considerations / 77
4.2.8 Contributions from reticles / 78
4.3 Maximizing the process window / 79

VII

5.1 Overlay models / 87

5.2 Matching / 94
5.3 Contributions from processing and alignment mark optimization / 105
5.4 Addressing the problem of non-normal distributions / 108
5.2 Outliers / 112

CHAPTER 6 YIELD / 1 19

6.1 Yield monitor strategy / 119

6.2 Yield models / 124
6.3 Parameters which affect yield / 126

CHAPTER 7 PROCESS DRIFT AND AUTOMATIC PROCESS CONTROL / 131

7.1 Adjusting for process drift / 131

7.2 The exponentially-weighted moving average / 133
7.3 Automatic process control / 135

CHAPTER 8 METROLOGY / 141

8.1 The need for understanding the measurement process: defect detection /
141
8.2 Linewidth measurement using scanning electron microscopes / 143
8.3 Electrical linewidth measurement / 150
8.4 Measurement error budgets / 152
8.5 Measurement of overlay / 154

CHAPTER 9 CONTROL OF OPERATIONS / 161

9.1 Self-control / 161

9.2 Documentation / 163
9.3 ISO 9000 / 165

GENERAL REFERENCES / 169

INDEX / 189

Viii

In the mid-1980s the semiconductor industry underwent a fundamental change.

With superior yields, reliability, and efficiency, Japanese manufacturers of
integrated circuits surpassed their American competitors in terms of market
share. The American semiconductor industry responded with significant quality
improvement and regained the leadership market position by 1993. Today the
integrated circuit business is highly competitive and global.
During the transition years I found that I was frequently consulted by
lithography engineers who were attempting to apply quality improvement
methods, such as statistical process control (SPC), to which they had been
recently introduced. Through discussions with these engineers it became clear
that there were subtle reasons why SPC could not be applied in a
straightforward way to many situations which occurred in lithography. My
explanations to the engineers evolved into a set of class notes, and now this
Tutorial Text.
Many of the quality problems that we were trying to solve arose in
conventional manufacturing situations, while others involved development pilot
lines. Methods for addressing the problems that occur in the context of process
development are rarely addressed in texts on process control. Another objective
of this Tutorial Text is to present control methodologies applicable to
development pilot lines.
To understand this text there are some prerequisites. A basic foundation in
lithography science is assumed. The SPIE Handbook on Microlithography,
Micromachining and Microfabrication. Volume 1: Microlithography l provides
a suitable introduction. It is also assumed that the reader has had some
introduction to basic statistical concepts and statistical process control. It is my
intention that this text be a self-contained tutorial on lithography process
control for readers familiar with the prerequisite lithography science and basic
statistical process control, although some subjects may involve a higher level of
mathematical sophistication than others. The text covers the subject of
lithography process control at several levels. Discussions of some very basic
elements of statistical process control and lithography science are included,
because, when trying to control a lithography process, a number of subtle
problems arise that are related to fundamental issues. To most readers, the
information presented on the foundations of statistical process control should be
familiar. Nevertheless, it is useful to review the foundations of statistical
process control, in order to clearly identify those circumstances in which the
method may be applied, and where it needs to be applied with particular care.
This inclusion of basic topics also allows the text to serve as an introduction to
process control for the novice lithography engineer and as a reference for
experienced engineers. More advanced topics are also included to varying

1 Handbook of Microlithography, Micromachining, and Microfabrication. Volume 1:

Microlithography, P. Rai-Choudhury, Ed., SPIE Press, Bellingham (1997).

Statistical methods need to be part of every lithographer's toolbox, because

lithographic processes contain intrinsic levels of variation. This variation is a
consequence of the nature of the world. For example, petroleum is typically the
starting material from which photoresists are synthesized, and the composition
of crude oil varies from well to well. Lithographic processes and tools are
affected by environmental parameters such as barometric pressure and relative
humidity, and these factors vary with the weather. Lithography is a
manufacturing science implemented and ultimately exercised by human beings,
each of whom is a unique individual, different from all others. When people are
involved, there is a special element of variation interjected into the process. The
analytical methods used by lithography engineers and managers must be
capable of dealing with variation in equipment, materials, and people.
The objective of any process control methodology is the reduction of
variation, in order to maintain conformance to standards or to meet a higher
standard. Variations in gate lengths can lead to degraded yield or slower parts,
which usually sell for a lower price than faster devices. Higher manufacturing
costs result from variation, in the form of scrap, reduced yield, rework, and low
equipment utilization. Often there are costs associated with attempts to reduce
variation. Since the objective of process control is to maximize profitability, the
most effective methods are those which accomplish control in the most cost-
effective manner. Indeed, Dr. Walter Shewhart, the inventor of statistical
process control, titled his book, Economic Control of Quality of Manufactured
Product,' with the first word of the title identifying the monetary considerations
motivating his methods. Generally, the most economical approaches require
that particular levels of variation be tolerated, and the purpose of statistical
process control is the identification of variation in excess of the norms of a
controlled process.
Statistics is the mathematical science for making inferences about
quantities which are probabilistic (in contrast to deterministic) in nature. Many
statistical methods are therefore applicable only to situations that are random, in
which events are independent of each other. A large fraction of the first three
chapters of this Tutorial Text will involve the examination of commonly
occurring situations in microlithography in which measurements are not
independent. The discussion will lead to methods for applying statistical control
techniques correctly in such situations.
A word is in order about the probabilistic versus deterministic nature of
quantities that are encountered in microlithography. Nearly all lithography
processes occur at levels that are well described by classical physics. For
example, optical patterning involves sufficient numbers of photons for the

imaging to be well characterized by classical electrodynamics and optics.

Processes can be understood to a high level of accuracy, but variations occur
because the choice is made, on an economic basis, to not determine actual
process conditions to the level necessary to make accurate quantitative
predictions. There is a significant difference between variations whose causes
are unknowable versus simply unknown. Microlithography is currently far from
quantum limits, where fluctuations arise from physical law.
While statistical methods are essential for identifying situations in which
there are abnormal levels of variation, control is re-established only after
corrective action. Effective response is achieved by using engineering science
to identify and then address causes of variation. Process control is the
application of engineering science, where statistical techniques are used to
interpret data. This tutorial text will discuss statistics, lithography science, and
the interplay between the two. Statistical methods will be the primary topic of
the first three chapters, while Chapters 4 and 5 will focus on the scientific basis
of image formation and pattern placement, respectively. Chapter 6 will deal
with the subject of yield, with a focus on defect monitors. Process drift,
feedback, active control and their relationships to statistical process control will
be discussed in Chapter 7. All control methodologies rely upon data, and the
integrity of the data is crucial. Metrology is difficult in the world of deep sub-
micron geometries, and Chapter 8 will deal with this critical aspect of process
control. In Chapter 9 we move from the mathematical and physical scientific
aspects of processing to considerations of the most direct human elements, and
the text is concluded with a discussion of the control of operations.

1 .1 THE ASSUMPTIONS UNDERLYING STATISTICAL PROCESS CONTROL

The genius of Walter Shewhart was in identifying a methodology applicable to

all random processes, regardless of their underlying probability distributions.
This universal approach was achieved by exploiting the central-limit theorem,
which is summarized as follows:

If individual and independent measurements are sampled from any random

process in subgroups comprised of n individual measurements, and the
measurements of the jth subgroup are averaged

x3 1 n(1.1)
• =—^x i•
1

where x the ith measurement in the jth subgroup, then:

1) The distribution of subgroup averages (x) will approach a normal
distribution (Gaussian or bell curve) as n gets large.

2)The distribution of subgroup averages will have the same average as the
distribution of individual measurements:
N n N
NEx. N Iyxl,^x,
n,N->°°
(1.2)
j =i =1 j=1

where N is the number of subgroups and .x is the mean of the overall

process.
3)The standard deviation cr of the distribution of averages .x^ is given
by:

Cr
6 X =7' (1.3)

where 6 x is the standard deviation of the individual measurements.

Throughout this text, a bar above a quantity indicates a mean value.
Discussions of the central-limit theorem can be found in most basic texts on
probability theory. 2
According to the central-limit theorem, any random distribution can be
effectively transformed to the normal distribution by subgrouping and
averaging. Studies have shown that this can be accomplished effectively with
subgroups as small as three or four, so long as the primary distribution does not
depart too significantly from normality.' Moreover, the resulting normal
distribution has key parameters, the average and standard deviation, that can be
related to their respective values in the original random distribution. By
working with the normal distributions obtained through subgrouping and
averaging, conclusions about processes can be made without knowing the
functional form of the original random distribution. As will be seen shortly, in
lithography the choice of subgroups must be made carefully.

1 .2 THE PROPERTIES OF STATISTICAL PROCESS CONTROL

Statistical process control is based upon the normal distribution, whose

properties are well known. The functional form of the normal probability
distribution, for a random variable x, is

(x -µi
(1.4)
f(x) 6 2^ e 262

where a is the standard deviation of the process and t is the process mean. This
function is plotted in Fig. 1.1. Some of the well-known properties of this
distribution are:

1) Data fall within ± 36 99.73% of the time. In other words, data fall
outside of ± 36 once every 370 times.
2) Half of the data are larger than the mean, and half are smaller. The
probability of having eight points in a row above the mean is1 _
2 8 256
0.0039, and the probability of having eight points below the mean is
0.0039 (0.39%).
3) Two out of three consecutive points are between 26 and 36 or between
-2a and -36 with a frequency of 0.30%.
4) Four out of five consecutive points are between 16 and 36 or between
-16 and - 36 with a frequency of 0.54% .

0.5

0.4

p0.3
ó 0.2
A.

* 0. 1
0 -4 -3 -2 -1 0 1 2 3 4
Parameter

Figure 1.1. The probability density function for the standard normal distribution
with a mean of zero and a standard deviation of 1.0.

These properties of the normal distribution lead to the "Western Electric Rules"
of statistical process control (Fig. 1.2), so called, because SPC and these rules
originated at the Western Electric Company. 4 The idea behind the Western
Electric Rules is that action should be taken if events occur that are improbable
based upon normal variation. For example, data (x) should exceed ±3 6 X only
once every 370 times as long as the process remains in control. Suppose control
charts are updated once per day. In this situation, one should find that x^ falls
outside of ±3 6 x slightly less than once per year, so long as the process remains
in control. With such a low probability of occurrence, the existence of data
outside of ±3 6 X suggests that the process is no longer in control. Similarly, a
long sequence of values all above or below the mean is highly improbable. For
example, the probability of having eight values in a row either above or below

(I) = '

the mean is 2x 128. For a process in control, false alarms, indicating

out-of-control conditions, occur, on average, only once in 128 values of x^ .

The average number of .z^ values between events indicating potential out-of-
control conditions, is referred to as the "average run length." This is the average
number of times that x^ is determined before a Western Electric Rule is
violated.

1)Any single point falls outside of the +/-3a limits.

2) Eight successive points are above the mean, or eight successive points are below the mean.

3) Two out of three successive points are between 26 and 36 or between -26 and -3a.

4) Four out of five successive points are between 1a and 3a or between -16 and -3a.

Figure 1.2. The Western Electric Rules. Corrective action should be taken if any
these situations occur.

Suppose there is a shift in the process mean. The average run length will
decrease, as shown in Fig. 1.3. For no shift in the mean, the average run length
is 370 when monitoring the process by the "±3o' rule, and 128 for monitoring
eight-in-a-row above or below the mean. When the mean has shifted one sigma,
these values decrease to 44 and 8, for the "±3o' rule and the "eight-in-a-row"
tests, respectively. For the "eight-in-a-row" test, mean shifts are detected after
the minimum number of values, eight, while the "±3d' test is less sensitive.
The process can be continued, on average, for 44 more testing intervals before
an out-of-control condition is indicated, when only the "±3d' test is used. This
lack of sensitivity to small shifts in the process is a concern that will be
discussed further in Section 1.5 on process capability.
Another type of process change is an increase in the standard deviation, 6.
When this occurs, the "eight-in-a-row" test will indicate nothing so long as the
process mean remains unchanged and the process distribution remains
symmetric. On the other hand, the "±3d' test will have a decreasing average
run length with increasing a, as shown in Fig. 1.4.

400
3 sigma
a 300 8 in a
row
200
00
M

100

O il
. ------------------^

0 0.5 1.0 1.5 2.0

Shift in mean (in units of sigma)

Figure 1.3. The average run length as a function of mean shift for two of the
Western Electric Rules.

!DI€

300
a
200
an
E 100

0 0 10 20 30 40 50 60 70 80
% change in sigma

Figure 1.4. The average run length as a function of increase in the process
standard deviation, for the +/-36 rule.

1.3 SITUATIONS IN LITHOGRAPHY WHERE STATISTICAL PROCESS CONTROL

CANNOT BE APPLIED NAIVELY

In order to create a process control chart, one needs to estimate 6 z . This is

usually accomplished in one of two ways.
1) A large number of measurements x. can be taken, and the estimate can be
calculated directly:

s _M( — x 1. x)2
(1.5)
M-1

In this equation, s is the estimate of the standard deviation for the overall
x
process, is the process mean, and M
is the total number of measurements
N n
taken. If subgroups of size are taken, then

M = n x N. (1.6)

Eq. 1.3 can then be used to calculate an estimate for 6 x

2) The calculation of the right side of Eq. (1.5), involving multiplication and
square roots, is difficult without calculators or computers. Statistical process
control originated in the 1920's, when such tools were unavailable.
Consequently, much of the methodology of SPC is based upon estimates of
standard deviations based upon ranges. The range R, which is given by:

R = largest value - smallest value, (1.7)

is easily computed for small subgroups without calculators or computers. A

mean range R can be computed as well:

1 N
(1.8)
N ^_1
where R. is the range for the jth subgroup. For normally distributed
measurements, the mean range of subgroup data provides an estimate for the
population standard deviation according tos :

s =* . (1.9)
2

When R has been determined from a large number of subgroups, values for the
parameter d2 are given in Table 1.1. 6 The estimate for 6 X can then be
calculated from Eq. (1.3).
The two methods described above are the most common approaches for
estimating the standard deviation, which is used for determining process control
limits. In one method, the standard deviation is computed directly from all data
collected over time, while the other involves an estimation of the standard
deviation based upon the ranges of data in the subgroups. Unfortunately,
neither of these methods apply in situations which occur commonly in
lithography operations. Consider a monitor of resist coatings, where a wafer is

coated daily, and thickness is measured at nine locations on the wafer. From the
measurements on this wafer, a daily mean and standard deviation can be
computed. The results from a typical process are summarized in Table 1.2.

Subgroup size d*
a
2 1.128
3 1.693
4 2.059
5 2.326
6 2.534
7 2.704
8 2.847
9 2.970
10 3.078

Table 1.1. Factors for estimating the standard deviation from the mean range of
subgroups.

Average uniformity across individual wafers (6) 9A

Average standard deviation of wafer-to-wafer 6A
means, 10 consecutive wafers
Standard deviation of wafer-to-wafer means over three 17
months.
Overall process standard deviation. 19

Table 1.2. Variations observed in a monitor or resist thickness. Eq. 1.7 was used to
estimate the average uniformity across individual wafers.

Suppose that the average resist thickness on the wafers provides the values
x^
for that are used to generate a control chart. One might expect that the nine
measurements across each wafer comprise the corresponding subgroups. The
uniformity across individual wafers (first row in Table 1.2), in practice, could
be obtained by determining the average range of wafer thickness and using Eq.
1.9, and control limits could be calculated using the resulting value for 6 and
Eq. 1.3. However, this is not the right thing to do.
For this process, the control chart of daily wafer means is charted in Fig.
1.5. Using the first row of data in Table 1.2 and Eq. 1.3 to calculate the control
limits results in the absurdly tight limits about the process mean (8000 A) of
3 xc
±3 x 6 x = ± _ ± 9 A. Here we are using the conventional Shewhart
n
±36 control limits about the mean that form the basis for one of the Western
Electric Rules. The nine measured thicknesses across each wafer clearly do not

constitute an appropriate subgroup. Control limits calculated from all of the

thickness data would have been 8000 ± 3 x 19 A = 8000 ± 57 A about the
process mean, as indicated by the data in Table 1.2. However, as will be
discussed shortly, the appropriate control limits for a control chart of the daily
mean resist thickness should have been determined using the data from the third
row of Table 1.2, resulting in control limits of 8000 ± 3 x 17 A = 8000 ± 51 A.
The two most common methods for estimating the process standard deviation,
calculating directly from Eq. (1.5) or estimating from variations within
subgroups, would have led to either inappropriately loose or excessively tight
limits.
The best way to avoid using an incorrect method for calculating the
standard deviation is to take a critical look at the quantity that comprises the
ordinate ("y-axis") of the control chart. The control limits should be calculated
from the standard deviation of that quantity. In the example just discussed, the
parameter being charted was the mean thickness x^ from the single wafers that
were coated daily. The resulting estimate for the control limits (about the
process mean) is:

N JC . — x 2
±3s = ±3 (1.10)

The variations in the mean resist thicknesses could not be inferred from
variations within wafers or from measurements taken over short periods of time
(see Table 1.2). Calculations of the standard deviation using the entire set of
resist thickness measurements overestimates the variation in wafer-to-wafer
means, because it includes across-wafer variations.

8060

8040 correctly calculated

control limit
8020
incorrectly calculated
8000
control limits
7980

7960

7940
Day

Figure 1.5. Control chart for the mean resist thickness, with incorrectly and
correctly calculated control limits.

It is also important to note that the across-wafer variation and drifts in resist
thickness are caused by different physical phenomena. For example, changes in
barometric pressure and relative humidity will cause changes in the average
resist thickness, but there will be a much smaller effect on across-wafer
uniformity. Control limits based on across-wafer non-uniformity are not
applicable to control charts of average resist thickness, because across-wafer
non-uniformities reflect different physical phenomena than do the average
thicknesses. Subgroups must be chosen carefully if variation within those
subgroups are going to be used to generate control limits.
As seen from the data listed in Table 1.2, the average resist thickness varied
little over the short amount of time required to coat 10 wafers, and the variation
across individual wafers was less than the variation over a three month time
frame. These data illustrate two types of correlations — temporal and spatial —
which break randomness. Wafers coated close in time to each other will tend to
have similar resist thicknesses. Across individual wafers, the resist thickness
will not vary rapidly over short distances. These situations, in which
measurements are related non-randomly to the values of other parameters, will
be discussed in Section 2.3.

1.4 NON-NORMAL DISTRIBUTIONS

In the preceding example, the subgroup size = 1. A single wafer was coated
daily, and it was the mean thicknesses of the resist on the individual wafers that
were subjected to process control. As described in Section 1.1, the methodology
of SPC is based upon the ability to transform all random distributions to ones
that are normally distributed, by means of subgroup averaging and the central-
limit theorem. When subgroup sizes of one are used, care must be taken to
ensure that the process being monitored varies normally, or the conventional
SPC formalism may lead to incorrect conclusions.
There are a number of circumstances that arise in lithography operations in
which it is inconvenient to generate subgroups of size greater than one. The
example of resist coating monitors, discussed in the previous section, was one
example. There are many reasons for restricting subgroups to the minimum
size, which is one:
• High measuring costs. Metrology equipment, for measuring defects, film
thickness, linewidths and overlay, is expensive. Reducing the number of
measurements reduces the expenditures required for measurement
equipment.
• The need to maximize equipment utilization. Processing wafers for the
purpose of monitoring equipment detracts from productive time, in which
processing equipment is used to make saleable product. It is clearly
desirable to minimize the time used for processing material that cannot be
sold.
• Data collection is very time consuming. The amount of time required for
people to process, measure, and analyze equipment monitors tends to

increase with the number of wafers processed. Minimizing the number of

wafers reduces the level of human resources needed to control the process.
For all of these reasons, it is frequently desirable in lithography operations to
have sub-groups of size equal to one. Since conventional SPC formalism can be
conveniently applied to individual measurements in situations where the
processes do vary normally, and may be problematic where the processes do
not, it is useful to have methods to determine whether processes are distributed
normally or not. There are graphical and analytical methods that can test for
normality. The former involves the use of normal probability paper, which is
derived from the graph of the cumulative probability of the normal distribution,
shown in Fig. 1.6. The cumulative probability is obtained by integrating Eq.
1.4:
x 1_ (X µ) 2 -

Cumulative probability =f 6
e 2 e2 dx'
27c
(1.11)

Normal probability paper is created by stretching Fig. 1.6 along the ordinate
nonlinearly, such that the cumulative probability curve turns into a straight line.
The result is shown in Fig. 1.7.
To use normal probability paper, do the following:
1)Rank the data from smallest to largest, i = 1, ..., n.
2) Determine the percentage value, F. , for the ranked data.
F 100(i-0.5)
1=1,...,n. (1.12)
n
3) Plot the data on normal probability paper, where the data values are plotted
on the (linear) abscissa and the percentage value is plotted on the ordinate.
Data that are normally distributed will plot as nearly a straight line. A
convenient source of normal probability paper is the book, Graph Paper from
Your Copier, 7 from which the purchaser is free to make photocopies.
Alternatively, there are a number of software packages that plot the data onto
normal probability scales automatically. A listing of software packages which
have this capability can be found in the annual software directory in the
magazines, Quality Progress (usually in the April issue) or Quality (December
issue).

Example: Consider the following 25 ranked measurements:

3.8, 4.6, 4.6, 4.9, 5.2, 5.3, 5.3, 5.4, 5.6, 5.6, 5.7, 5.8, 5.9, 6.0, 6.0, 6.1, 6.3, 6.3,
6.4, 6.5, 6.6, 6.8, 7.0, 7.4, 7.6.

These data are plotted on normal probability paper and shown in Fig. 1.8. The
straight line in Fig. 1.8 is for the normal distribution with the same mean and
standard deviation as these data.

Standard Normal Cumulative Probability Distribution

A, 0.8
ó 0.6
a
0.4
E
c) 0.2

0 -3 -2 -1 0 1 2 3

Variable (in units of 6)

Figure 1.6. The cumulative normal distribution.

Normal Probability Plot

99.99
99.90
99.00
g 90.00

50.00

10.00

1.00
0.10
0.01

Data values

Figure 1.7. Normal probability paper.

Normal Probability Plot

99.99
99.90
99.00

90.00

1a.
50.00

10.00

1.00
0.10
0.01
4.0 5.0 6.0 7.0
Data values

Figure 1.8. Normal probability plot for the data discussed in the text.

Normal probability paper was a tool from an era in which computations

were difficult and laborious. The primary deficiency of this method is the
degree of subjectivity in deciding whether data fit a straight line well or not.
There are numerous analytical and objective methods for testing normality, the
most common of which is the x 2 (chi-squared) test,$ which is based upon
evaluating the statistic:

) 2 (O — E )
(O Ei1 2 (O — E )2
x z =i .,a z+...+z kJ
(1.13)
ElE2 Ek

where Oi is the observed frequency (number of occurences) for event or

category i, and E. is the expected frequency based upon the assumption of
normality. For example, an event or category might consist of having values
within a particular interval. Even for a normal distribution, x2 usually # 0,
because of random variation. A test is needed to determine when the value of x2
is larger than might be expected on the basis of typical random variation. When
this occurs, the assumption of normality can be questioned.
The x2 test involves several steps:
1) From the data, calculate the estimated mean (x) and standard deviation (s),
which will be used to determine the expected frequencies from a normal
distribution where µ = x and a = s.
2) With the parameters estimated in the prior step, it is possible to compute the
expected frequency of data that occur between any two values. The range of the
data should be completely divided into categories such that the expected

frequency is at least five in each category. The requirement that there are at
least five in each category ensures robustness of the test in situations where the
actual distribution deviates significantly from normality.
3) With the range of data divided into categories, the actual occurrences O i
within each category are computed.
4) Compute x2 using Eq. 1.13.
5) Select a desired Type I error level, a. This is the probability that the
hypothesis of normality will be rejected even though the distribution truly is
normal. Typical values for a are 0.05 to 0.10. The larger a is, the more stringent
the testing.
6)Look up xl a,v in a table of the chi-square distribution or calculate it. Tables
for the chi-square distribution can be found in most texts on statistics. The
degree of freedom is v = k - 3, where k is the number of categories.
7) If x2 < xi a,v , then the distribution can be assumed to be normally
distributed to the Type I error level of a.
The data in example above can be used to exemplify the use of the X 2 test.
The mean and standard deviation of the data set are:
x = 5.868 (1.14)
s=0.892 (1.15)
With these parameters, one can calculate the expected frequencies. It is
necessary to create categories in which the expected frequency is at least 5.
Additionally, it should be noted that the best test has the most comparisons
between expected and actual frequency, i.e., when there are the most categories.
For the data in the above example, one has the largest number of categories,
with at least five events in each category, when there are five categories with an
expected frequency of five in each. The probability of being in any particular
category is 0.20. The boundary values that lead to this division of the data are
listed in Table 1.3. These were obtained as follows. The first category consists
of the smallest values. The lower boundary for such a category is — and the
upper boundary u is given by the normal probability distribution with mean and
standard deviation given in Eqs. 1.14 and 1.15:
u 1(x,_x)z
e 2 S2 dx'= 5 = 0.2. (1.16)
s 2n 25

The boundary u forms the lower boundary for the next category, and its upper
boundary is calculated similarly.
Continuing with the rest of the data, one obtains the results shown in Table
1.3, from which x2 is calculated to be 0.80. Letting a = 0.05, we find that

Xl a,v =2C0.95,2= 5.99. (1.17)

In this case,
x2= 0.80<5.99 = xi av, (1.18)

and we can conclude that the data are consistent with the assumption of a
normal distribution.

Boundary value J
Original data Expectedfrequency Actualfrequency
to 5.12
--00 3.8, 4.6, 4.6, 4.9 5 4
5.12to5.65 5.2,5.3,5.3,5.4,5.6,5.6 5 6
5.65to6.10 5.7,5.8,5.9,6.0,6.0 5 5
6.10to6.62 6.1,6.3,6.3,6.4,6.5,6.6 5 6
6.62 to +oo 6.8, 7.0, 7.4, 7.6 5 4

Table 1.3. Summary of the parameters used for applying the x2 test to the data
from the above example.

The xz test indicates whether measurements from a process are distributed

normally or not. There are other tests for non-normality that involve measures
of the shape of the distribution, 9 and which provide additional information that
is often useful to lithographers. Two characteristics, symmetry and peakedness,
are of particular interest, and each of these can be defined rigorously, i.e.,
mathematically.
First, a preliminary term needs to be defined. Suppose there are N
measurements from a process. Then, the rth moment about the mean, m,

N
1(Xi —,x )r
m r = `-1 N, (1.19)

where r = 1, 2, 3, ... Note that the standard deviation is the square root of the
second moment about the mean:

s = m2 . (1.20)

Skewness is the term that mathematicians use to refer to the symmetry of a

distribution. The coefficient of skewness is defined as

a 3 = m3 `^ . (1.21)
( m2/

Note that this is a dimensionless quantity. Symmetric distributions, such as the

normal distribution, have skewness = a 3 = 0. A distribution with positive
skewness is shown in Fig. 1.9.

Skewness =0 Positive skewness

mean 1• mean
r

v
i>
.n
0

Parameter value Parameter value

Figure 1.9. Examples of skewness.

Another shape parameter is the degree of peakedness, called kurtosis,

defined by:

a4 = m2 (1.22)
m2

Kurtosis is also dimensionless. For a normal distribution, a 4 = 3. Distributions

for which a 4 = 3 are also know as mesokurtic. For distributions that have flatter
peaks than the normal distribution's bell curve (see Fig. 1.10):

a4 < 3 : platykurtic (1.23)

Distributions may also have sharper peaks than a bell curve:

a4 > 3 : leptokurtic (1.24)

Platykurtic distributions arise frequently in the measurement of overlay, which

will be discussed in Chapter 5.
The shape of the distribution can often reveal some of the nature of
underlying sources of variation. Consider a situation in which defocus is a
significant source of linewidth variation (Fig. 1.11). In many situations the
linewidth will vary in only one direction as the focus is changed. For the data
shown in Fig. 1.11, the linewidth always decreases as focus is moved away
from best focus. (This behavior — decreasing linewidths because of defocus —

0.8

0.7 leptokurtic
0.6 normal = :Z
mesokurtic
0.4
platykurtic
0.3

á 0.2

0.1

0
-4 -2 0 2 4
x (in units of sigma)

Figure 1.10. Examples of distributions with different types of kurtosis.

target

Defocus

Figure 1.11. As a consequence of defocus, the linewidth always becomes smaller

than the target value, leading to a skewed distribution.

is typical for isolated chrome lines printed into positive photoresist. Behavior
can vary with pitch and the tone of the photoresist. 10 In this text, a line refers to
a long and narrow piece of photoresist, while a space refers to a long and
narrow trench in a sheet of resist.) In such a situation, the effect of defocus will
be to shift the linewidths in only one direction, thus creating a skewed
distribution. For example, suppose the focus varies normally, with a mean of
zero and a standard deviation of 6, and if the linewidth w varies with defocus z
according to

w = az 2 , ( 1.25)

then the linewidth varies as a consequence of the variation in focus3:

probability density = 1e22if w >0 (1.26)

6 27taw

=0 if w<_0 . (1.27)
This probability density function is shown in Fig. 1.12.
3

2.5

r.
a^ 2
b
1.5

1 1
0.5

0' h - -

-0.4 0.0 0.4 0.8 1.2 1.6 2

Variable
Figure 1.12. Probability density of Eq.1.26. The ordinate is in units of a 2Ia.
Knowing that some sources of variation produce skewed distributions, one
can use skewness as a diagnostic tool. Significant skew in the distribution of
linewidths, in a particular direction, can be the result of poor focus control. The
appearance of skew in linewidth distributions might be a clue that there is a
problem with defocus, rather than some other source of increased linewidth
variation. Given the number of possible sources of increased process variation,
information that can reduce the number of potential factors is invaluable for
expeditiously returning a process to a controlled state.
Finite samples from a normally distributed process will produce non-zero
values for skewness and values for kurtosis which do not equal three, because
of random variation. Levels of statistical significance must be established in
order to distinguish between real deviations from normality and natural
statistical variation, i.e., by how much can we expect skewness to deviate from
zero, or kurtosis to deviate from three even when the process is normally
distributed, but the sample size is small. The expected ranges of skewness and
kurtosis for a normally distributed process are given in Tables 1.4 and 1.5.
Values for skewness within the confidence intervals shown in Table 1.4, and
values of kurtosis within the confidence intervals shown in Table 1.5, can be
expected from normal statistical variations.

Example. The skewness for the distribution of Eqs. 1.26 and 1.27 can be
calculated. Since this is a continuous distribution, integrals are required to

calculate the moments. If p(w) is the probability distribution for the linewidth
w, then the rth moment about the mean t is given by:

m r =$ p(w)(w µydw — (1.28)

Evaluating the integrals, one finds that µ = au g , m 2 = 2a 24 , m3 = 8a 36 ,

3
and consequently, a 3 = = 2.82. Looking at Table 1.4, one would expect to
be able to distinguish this distribution from a normal distribution with a
relatively small sample size.

Sample size 95% confidence 99% confidence

interval j interval
25 ± 0.714 ± 1.073
30 ± 0.664 ± 0.985
35 ± 0.624 ± 0.932
40 ± 0.587 ± 0.869
45 ± 0.558 ± 0.825
50 ± 0.533 ± 0.787
60 ± 0.492 ± 0.723
70 ± 0.459 ±0. 673
80 ± 0.432 ± 0.631
90 ± 0.409 ± 0.596
100 ± 0.389 ± 0.567

Table 1.4. The 95 % and 99 % confidence intervals for skewness, assuming a

normal universe. 6

Sample size 95% confidence interval 99% confidence interval

Lower Upper Lower Upper
interval interval interval interval
boundary boundary boundary boundary
25 1.91 4.16 1.72 5.30
30 1.98 4.11 1.79 5.21
40 2.07 4.06 1.89 5.04
50 2.15 3.99 1.95 4.88
75 2.27 3.87 2.08 4.59
100 2.35 3.77 2.18 4.38

Table 1.5. The 95 % and 99 % confidence intervals for kurtosis, assuming a normal
universe.6

1.5 PROCESS CAPABILITY

Statistical process control provides a method for identifying changes in

processes, by comparing current processes to prior levels of performance. The
requirements of a process, in terms of what is needed to produce good products,
do not enter into the SPC methodology. A process may be very stable, but
unable to perform within specifications. A process is said to be capable if it is
able to meet the requirements of the technology (Fig. 1.13). Capability indices
are metrics for gauging process performance versus requirements. The most
fundamental capability index is C, which measures the width of the process
distribution, relative to the specification limits"

C USL — LSL
= (1.29)
p66

where USL is the upper specification limit and LSL is the lower specification
limit. Note that Cp is a dimensionless number. Large values for Cp indicate a
tight distribution, relative to requirements. Hence, bigger is better.
Lower Upper:
Ü
iü spec limit spec limit1

I^ ^I

I I Capable
I I

9 I I
I I
w
I I

Parameter value

c I
I
I
I
I
I
á
I i
I I
I I

I I
Not capable
I I

Parameter value

Figure 1.13. Histograms of data from a capable process and from a process that is
not capable.

The two processes in Fig. 1.14 have the same C, even though the process
centered in the middle of the specification limits is clearly more desirable. In

order to incorporate process centering into the measure of process capability,

additional indices have been defined. The most commonly used metric is
d pk o.

Cpk =( 1—k)C p , (1.30)

where

k= 21T -

USL — LSL
µl , (1.31)

T is the process target, and µ is the process mean. When the process mean
equals the process target, C p = C pk . As the process mean moves away from the
target value, C pk decreases.

(a) (b)

LSL Target USL LSL Target USL

Figure 1.14. Two processes with the same Cp = 2.0.

There are several disadvantages to C pk . First, the quantity is not an

analytical function in its parameters because of the absolute value in Eq. 1.31
(the derivatives aaT and áµ do not exist when T = p). Second, it does not
weigh shifts in the process mean heavily, compared to increases in the width of
the distribution. To address these two deficiencies, another capability index has
been defined:12,13

Cpm = Cp(1.32)
1+ (T —µl
a2

For the two processes in Fig 1.14, we have the process capability values shown
in Table 1.6. Note that the metric Cpm is smaller than CPk for the same
process, when there is a mean shift, since the Cp Z index emphasizes shifts of
„

the process mean.

Process A Process B
Cp 2.0 2.0
Cpk 2.0 0.85
Cpm 2.0 0.56

Table 1.6. Process capability indices for the processes shown in Fig. 1.13.

Process capability is a very powerful management and engineering tool.

With a single number, one is able to assess the extent to which a process can
meet the needs of the product. During development, processes tend to have low
values for Cpm , quite often < 1.0. As the processes mature, Cpm increases, as a
consequence of process learning. However, in the ever-changing
microelectronics industry, processes are usually rendered obsolete before
manufacturing engineers have sufficient opportunity to bring them to full
maturity. Process capability can be made large by having wide specification
limits, but in the semiconductor electronics business this often results in
uncompetitive products. Values for C,,, in excess of three, often experienced
in other industries, 14 are almost never achieved in the world of
microlithography. Typical values for C pm in microlithography manufacturing
are given in the following table:

2.0 < C pm World class

1.3<C,<2.0 OK
1.0 < Cpm < 1.3 Significant improvement is needed
C pm < 1.0 In serious trouble

Table 1.7. Interpretation of process capability.

Process capability forms the basis for the "Six-sigma" methodology

advocated by Motorola. 15 Suppose one has a normally distributed process with
C p = 2.0. When the process is centered (C p = C pm ), only two parts per billion
will be out of specifications. Suppose the process mean shifts by la. In this
situation, only 0.3 parts per million will fall out of specifications, which still
represents good performance. Since SPC methodology cannot identify process
shifts immediately after they occur, a process where the upper and lower
specification limits are 66 from the target, prior to a shift in the process, will
continue to produce quality products between the time a shift occurs and the
time of detection and corrective action. Such "Six-sigma" processes ensure
product quality. On the other hand, suppose one has a process where
C p =1.33, which is a "4d' process. As long as the process mean remains on
target, only 63 parts per million will fall out of specifications, which is
adequate for many microelectronic devices. However, with a shift in the mean
of 16, 1350 parts per million will fall out of specifications, and this may no
longer meet product requirements. Note that the relationship between the
standard deviation and the fraction of non-conforming measurements given in
this section holds only for normally distributed processes. For non-normal
processes, the assessment summarized in Table 1.7 would probably require
adjustment, since there would be different relationships between the values of
the process capability indices and the fraction of products which are non-
conforming.
Process capability indices must be interpreted carefully. Suppose that a
device will function only if all linewidths are within specifications. If the
process for producing this device is centered to the target linewidth, normally
distributed, and C p =1.0 , there will be 0.3 parts per million out of
specifications. If the linewidths that are outside of specifications are distributed
randomly across the wafers, the process would be virtually incapable of
yielding 256M or 1G DRAMs. Fortunately, linewidths do not vary randomly
across wafers, so processes with C p = 1.0 are capable of producing such
products. The non-random distribution of linewidths and overlay will be
discussed in subsequent chapters.
Process capability indices are based upon the process standard deviation
and mean and will have uncertainties associated with knowing each of these
parameters only within the range of a confidence interval. 16 Interpretations of
process capability should be made with this in mind. With limited data there are
opportunities for unwarranted concern or optimism.

' W. A. Shewhart, Economic Control of Quality of Manufactured Product,

D. van Nostrand Co., New York, (1931).

2H. J. Larson, Introduction to Probability Theory and Statistical Inference,

John Wiley & Sons (1969).

3D. C. Montgomery, Introduction to Statistical Quality Control, John Wiley &

Sons, New York (1996).

4Western Electric's Statistical Quality Control Handbook, Delmar Printing

Company, Charlotte (1985).

5A. J. Duncan, Quality Control and Industrial Statistics, 5th Edition, Irwin,
Homewood, Illinois (1986).

6M. G. Natrella, Experimental Statistics, U.S. Government Printing Office,

Washington, D. C. (1966).

J. S. Craver, Graph Paper from Your Copier, HPBooks, Tucson, AZ (1980).

8S. S. Shapiro, How to Test Normality and Other Distributional Assumptions,

American Society for Quality Control, Milwaukee (1986).

9P. P. Ramsey and P. H. Ramsey, "Simple Tests of Normality in Small

Samples," Journal of Quality Technology, pp. 299 — 313 (1990).

10 M. Fujimoto, T. Hashimoto, T. Uchiyama, S. Matsuura, and K. Kasama,

"Comparison Between Optical Proximity Effect of Positive and Negative Tone
Patterns in KrF Lithography," SPIE Vol. 3051, pp. 739 — 750 (1997)

V. E. Kane, "Process Capability Indices," Journal of Quality Technology,

11
Vol. 18, pp. 41— 52 (1986).
12
L. K. Chan, S. W. Cheng, and F. A. Spiring, "A New Measure of Process
Capability: Cpm Journal of Quality Technology, Vol. 20, pp. 162 — 175
,"

(1988).

13R. A. Boyles, "The Taguchi Capability Index," Journal of Quality

Technology, Vol. 23, pp. 17 — 26 (1991).

'4 D. J. Wheeler, A Japanese Control Chart, SPC Press, Knoxville (1986).

15
F. R. McFadden, "Six-Sigma Quality Programs," Quality Progress, pp.
37 — 42 (June, 1993).

16
Y-M. Chou, D. B. Owen, and S. A. Borrego, "Lower Confidence Limits on
Process Capability Indices," Journal of Quality Technology, Vol. 22(3), pp.
223 — 229 (1990).

2.1 CHOOSING THE PROPER SAMPLE SIZE

An engineer responsible for resist processing had to address the following

problem. His resist supplier provided materials with batch-to-batch photospeed
controlled to ± 3%. For his process, this corresponded to linewidth changes of
± 8 nm about the target value of 250 nm. The engineer wanted to determine the
photospeed of a new batch of resist. This information would be used to adjust
exposure doses in order to continue to pattern production wafers with
linewidths at their target values. The engineer wanted to expose the minimum
number of wafers in order to guarantee that the linewidths would change, on
average, by no more than ± 1.5 nm after the facility changed to the new batch
of resist. Because the process had been running production volumes for several
months, the engineer knew that the process with the old batch of resist was well
centered on the target, that is, the process mean

µold = 250 nm. (2.1)

The engineer also knew the (short term) wafer-to-wafer variation very well:

a=2.0nm. (2.2)

Keeping the exposure dose fixed, it was reasonable to assume that ß would be
unchanged with the new batch of resist, only that µ would change.
Because of the wafer-to-wafer variation, estimates of the new process mean
( µnew) could be expected to vary, introducing uncertainty into the
determination of new . Such uncertainty is a function of sample size. Suppose
that n wafers were used to estimate new . Because of the random wafer-to-
wafer variation, the resulting sample mean (x) would not equal µnew exactly.
Processing and measuring groups of n wafers repeatedly would produce values
for x which would vary about new with a standard deviation of

6X =7 . (2.3)
sJfl

This result is true regardless of the nature of the parent distribution; it does not
need to be normal. From the central-limit theorem, it is understood that

repeated measurements of the sample mean x produce a distribution that is

approximately normal, regardless of the parent distribution.
Suppose the resist engineer wanted to know the new process mean to
within ± 1.5 nm to a 95% level of confidence or better. Then
1.96 6 x <_ 1.5 nm , (2.4)

because 95% of all data fall with ± 1.96 a of the mean for a normal distribution.
That is,
1.96 = z a(2.5)
1 2

where a = 0.05 and z is the value of x µ such that a is the area of the
1 a2 a 2
upper tail of the standard normal distribution. (See Eq. 1.4 and Fig. 2.1.)
Using Eqs. 2.3 and 2.4 and the known value for 6,

(2.6)
or
n>_6.8. (2.7)
Thus, the engineer needed at least seven wafers to determine the dose for the
new batch of photoresist to the desired level of confidence, so long as the
process was stable, and the old process was expected to produce mean
linewidths of 250 nm.

Figure 2.1. The standard normal distribution (mean = 0 and standard deviation
= 1), where al2 is the area under the upper tail.

If there was significant long-term drift to the process, then the engineer
would need to expose wafers with both the old and new resist at nearly the
same time, and compare means. By collecting data from the old and new resists
together, equipment drifts would factor out of the comparison. Since the
engineer was interested in knowing the difference between the two batches of
photoresist, he would determine

xold —xnew' (2.8)

where xold is the mean linewidth measured on wafers processed with the old
photoresist, and new is the mean measured linewidth using the new batch of
photoresist. Again, because of statistical variations, these measured quantities
(x01d' anew) will not exactly equal the actual process means (µold'knew ). The
100x(1 - a) % level of confidence for gold — new is given by'

2 2 z 2
b old ó new bold ó new
(xold —a new )—Z a + old — µnew old —x new )+Z a + '
1 2 h old a new 12 H old anew
(2.9)

where c old and a new are the standard deviations of wafers processed with the
old and new resists, and Hold and anew are the number of wafers processed
using the old and new resists, respectively. From Eq. 2.9 one can calculate the
number of wafers required for xold — knew to provide a sufficiently accurate
estimation of told — µnew'
Situations exist in which the standard deviation is unknown. In such
situations, the t distribution is used, instead of the normal distribution. If the
standard deviations for the two processes are approximately the same, then the
100x(1 - a) % level of confidence for µold — new is given by 7

1
aSp
(xold —knew) t a S p 1+
2,v 1 µold µnew< (xold —knew)+t
Fn+
1
Hold anew 1 2'vanew
(2.10)

where
l 2 2
S 2_ (\H old — l l sold +(anew
1 Snew
—
(2.11)
n old + anew —2

and the number of degrees of freedom v is given by:

v =H old +anew -2, (2.12)

and s old and Snew are the sample standard deviations for the old and new
processes, respectively. The factor t ais the analog of z a for the
1--

t-distribution with v degrees for freedom, where is the area of the upper tail.
The preceding discussion focused on the number of measurements
required to determine the mean to a sufficient level of confidence. For the
sample standard deviation there are no relationships analogous to Eq. 2.3,
which are valid regardless of how the data are distributed. The variations in the
standard deviation are known for the normal distribution. If x is normally
distributed with variance 6 2 and

n
1 x —) ß
2

s 2 = ` -1(2.13)
n-1

—1)s
is the sample variance based upon a random sample of size n, then
62
has a x 2 distribution with v = n-1 degrees of freedom. The x 2 probability
has
distribution for the random variable x is given by:

- -i - X

f (x) = x2 e 2 . (2.14)
22

intervals at the 100x(1 - a)% level are found from 4

2 (n -1)s2
1 — a = Prob. xl_a12 — 2— x a2 /2 1
(2.15)

where xß is the point on the x 2 distribution illustrated in Fig. 2.2. From Eq.
2.15, confidence intervals for 6 can be constructed:

BLS< 6 <BUS, (2.16)

where

BL=F n
-(2.17)

and

Bu= 2 -1(2.18)
x1_ 2

0.12
0.1

0.08

0.06

0.04

0.02

x2
R
Figure 2.2. The chi -square distribution.

Values for xß can be obtained from tables or computed directly to obtain the
desired confidence interval.
Suppose we sample a process for which neither the mean nor the standard
deviation are known and find

x = 250 nm (2.19)

and

3s = 25 nm. (2.20)

The 95% confidence intervals for the mean and standard deviation are given in
Table 2.1. Because the x 2 distribution is asymmetric, and it is a distribution for
the variance, the confidence interval for the standard deviation is not
symmetric. Note that the uncertainty in the standard deviation is greater than
the uncertainty of the mean.

95% confidence intervals

Mean Standard deviation
10 measurements X —6.3 < µ < X + 6.3 17.2 < 3a < 45.6
25 measurements X —4.8 < µ < X + 4.8 19.5 < 36 < 34.8
100 measurements X —1.7 < µ < X+1.7 22.0 < 36 < 29.0

Table 2.1. Confidence intervals for the mean and standard deviation from the
sample statistics of Eqs. 2.19 and 2.20.

These uncertainties are significant when establishing control charts for the
first time, which requires that the process mean and the ±36 limits be
determined. Errors in determining the mean and standard deviation will result
in control charts that either fail to signal out-of-control conditions or produce
numerous false alarms. While the sample standard deviation was given by Eq.
2.20, the actual value for 36 could have been as high as 45.6 or as small as
17.2, to the 95% level of confidence, if there had been only ten measurements.
Control charts established with the value for the standard deviation calculated
from the first 10 measurements could have limits much too loose or too tight.
The number of measurements needed to establish the control limits is driven
primarily by the uncertainty in the standard deviation. In terms of correctly
determining control limits, the uncertainty in the standard deviation is more
significant than the uncertainty in the mean. However, for the very reason that
the "eight-in-a-row" test is a sensitive indicator of a shift in the process mean,
small errors in estimates of process means will result in control charts that
indicate out-of-control conditions for the "eight-in-a-row" test.
The proper method for establishing control charts for newly established
processes has been the subject of discussion in the quality control literature.'' 2
Frequently it is recommended that control charts should not be implemented
until they contain at least 25 data points. As one can see from Table 2.1, there
will be considerable uncertainty in the control limits even for control charts
with 25 entries, and out-of-control indications by the Western Electric Rules
should be interpreted with this in mind. Incorrect control limits can lead to
excessive numbers of false alarms or insensitivity to real process shifts, either
of which is possible when control charts are set up using small amounts of data.
A final problem in which sample size is an important consideration
involves estimates of variation from nested components. Consider, for example,
the data in Table 1.2. The resist thickness across individual wafers varies with
6 = 9 A. If this thickness variation is random, then there will be an apparent
wafer-to-wafer thickness variation, as a consequence of Eq. 2.3, for small
values of n. That is, the across-wafer variation leads to an apparent wafer-to-
wafer variation due to statistical fluctations inherent in small sample sizes. This
effect is particularly pronounced when across-wafer variations are large
compared to wafer-to-wafer or lot-to-lot variations. When calculating the total

process variation from nested components, the effects of small sample sizes
must be taken into account. If we consider the total variance to be given by:
2 _2 2 2
ó total — ó within—wafer + Cr wafer—to—wafer + ólot—to—lot 2'21

then the total variation can be overestimated because sample estimates of the
wafer-to-wafer and the lot-to-lot variations will contain contributions from the
within-wafer variations. The wafer-to-wafer variation arises from the difference
in the average resist thickness from one wafer to another. Because of Eq. 2.3,
there will be a measured wafer-to-wafer variation, even if the actual averages of
resist thickness change little from one wafer to another.

2.2 MEASUREMENT LOCATION CONSIDERATIONS

Consider a stepper with a magnification error. If overlay is measured at all

numbered sites in Fig. 2.3 then it will be possible to identify the problem with
magnification. On the other hand, measurements taken only at the sites
numbered one will confound magnification and translation. The problem
illustrated in Fig. 2.3 is a consequence of the non-random properties of overlay.
Because of the physics of wafer steppers, overlay will vary in particular ways
across exposure fields and between fields. There is a decided element of non-
randomness associated with step-and-repeat and step-and-scan systems.
Intrafield sources of variation — magnification in this example — repeat from
field to field. Measurements taken in several fields across a wafer, but at a
single intrafield location, will share an offset from the process mean that is
characteristic of the intrafield error at those sites within the fields. Further
discussion of the issue of measurement locations in the context of overlay will
be deferred until Chapter 5, where the subject of overlay will be covered in
detail.
Linewidths will also vary characteristically across exposure fields, as a
consequence of lens aberrations and linewidth variations on the reticles.
Measurements taken at only a few points within each exposure field, and
involving only one or two feature types, is insufficient to characterize
linewidths well. A particular feature on a reticle may have a linewidth in the
middle of the overall linewidth distribution on one lens, but could, as a
consequence of a different aberration signature, have relatively low or high
dimensions when printed using a different lens. Although linewidth
characterization is usually inadequate when few points are measured with
exposure fields, it is impractical and unnecessary to measure many points per
field on production wafers. Linewidth variations caused by lens aberrations are
stable over time, so lenses can be characterized off-line, and the resulting data
can be used in conjunction with in-line data collected during production.
Electrical linewidth measurement methods are useful for collecting the large
amounts of data required for detailed characterization of linewidth variations
across exposure fields. 3 These techniques will be discussed further in Chapter 8.

Several resist processes, such as coating and develop, involve radially

symmetric geometries. It can therefore be expected that linewidths should vary
as a function of wafer radius . 4 Measurement sites should be distributed across
wafers in a layout that will capture radial variation. An example is given in Fig.
2.4. However, the layout shown in Fig. 2.4(a) has a particular inadequacy. The
area of an annulus on a wafer increases linearly with the radius. Sampling plans
which are weighted uniformly to the wafer area they sample must have a linear
increase in the number of sampling points as a function of the radius [Fig.
2.4(b)].

substrate pattern

- - - 4 .---- 3'____
-

overlaying pattern

i i 1 ,

Figure 2.3. Overlay measurement sites.

•R9 n 01 0fl S ill

^ DnO i

a) b)

Figure 2.4. a) Measurement sites (marked X) on different radii. b) A sampling

plan with improved weighting proportional to wafer area.

2.3 CORRELATIONS

At Sierra Semiconductor a number of years ago, the monitoring of resist

thickness was assigned to third shift, the hours of midnight to 7 AM.
Frequently, when the engineer responsible for the resist coaters arrived in the
morning she would find the tracks disqualified for operation because the resist
thickness was out of specifications. After taking care of important lots which
were on hold, the engineer would process another resist thickness monitor to
confirm the status of the coaters and would often find that the resist thickness
was within specifications. This went on for several months, when the engineer
was usually unable to reproduce the problem with resist thickness found by the
night shift operator. At one point the engineer came into the fab at 5 AM to
retrain the operator who monitored the resist thickness but found that she was
doing everything correctly. Having eliminated human error as the source of the
problem, attention was turned to possible physical mechanisms. Eventually, it
was determined that the problem was related to the temperature control of the
facility. The temperature in the fab was appreciably colder at night, when the
operator on third shift monitored the resist thickness, compared to the
temperature mid-morning when the engineer processed her monitor. This was
an example of a correlation in time.
Correlations in time can be identified by calculating the sample
autocorrelation function.' Suppose that a process is measured sequentially in
time, resulting in values for a particular parameter: x1 , x2 , x 3 ,... Then the
sample autocorrelation function is given by

n-k
1 (x — x Xxt+k — x )
t
rk = t-1 n
, k = 0,1,... (2.22)
_)2
(x
t=1

where x is the measured process average and n is the number of events which
are considered. The problem observed at Sierra Semiconductor would have
been observed with sampling intervals (the time between measuring
xt and xt+1 ) equal to the duration of a single shift or shorter. The sample
autocorrelation function is an estimate of the autocorrelation function p k . Some
properties of the autocorrelation function are:

Ip k l<_ 1 for all k (2.23)

and

po =1. (2.24)

For a random function,

pk =0 fork#0. (2.25)

Significant deviations of Pk from zero indicate a non-random process. For a

random process, the sample autocorrelation will vary with standard deviation
so values of rk I significantly greater than 3 indicate that there are
V
correlations in the process.
The utility of Eq. 2.22 is illustrated by an example. A stepper engineer in
one fab was particularly concerned about focus errors, and he instituted a focus
check twice per shift for 50 consecutive shifts, generating a total of n = 100
measurements. The control chart which resulted is shown in Fig. 2.5. The
engineer found that the "eight-in-a-row" Western Electric Rule was violated on
several occasions. Further analysis indicated that this was the consequence of a
situation in which successive measurements were correlated. In Table 2.2 the
correlation coefficient (Eq. 2.22) is shown for the data of Fig. 2.5. Values of
rk for k = 1,2,3 are all greater than = 0.3, indicating that the data are
n
correlated. This correlation leads to consecutive measurements that are close in
value, thus leading to repeated violations of the "eight-in-a-row" Western
Electric Rule.

k rk
1 0.657
2 0.461
3 0.302
4 0.141
5 0.153

Table 2.2. The sample autocorrelation function for the data shown in Fig. 2.5.
0.3
8 in a row greater
than the mean Upper control limit
0.2
0
Q 0.1

k 0
Z -0.1
0 ^_ 8 in a row less
w -0.2 Lower control limit

-0.3 0
10 20 30 40 50 60 70 80 90
Focus check #

Figure 2.5. The control chart for frequently monitored stepper focus.

2.4 MEASUREMENT FREQUENCY

In order to lower the expense of control, the number of measurements should be

minimized. At the same time, there must be a sufficient number of
measurements to maintain control. The two requirements, expense of control
and the costs that result from a loss of control, must be balanced. Accordingly,
quick and easy measurements using inexpensive equipment can be done more
frequently than measurements that take a long time and require costly tools. For
example, defects on unpatterned resist films can be detected using laser
scattering systems, in which a wafer can be measured in a few minutes. Tools
which are capable of detecting defects on patterns are much more expensive
than simple laser scattering systems, and it can take an hour or more to measure
an entire wafer. Accordingly, resist defects can be monitored frequently, but
economically, using laser scattering systems. On the other hand, there are many
defect mechanisms that can only be detected on patterned wafers, so some level
of inspection using the more expensive equipment will be required.
In general, measurements need not be made when processes can be
expected to be unchanged. For example, the data in Table 1.2 show that resist
thicknesses vary little during the time required to coat 10 consecutive wafers.
Over longer periods of time considerably more variation is observed. Clearly,
resist thickness is something that need not be monitored hourly, but significant
variations may be observed with daily monitors.
The cost of losing process control depends on the type of circuits that are
being manufactured. In a facility in which custom integrated circuits are being
fabricated, a scrapped lot could represent 100% of the parts in the line for a
particular customer. The cost of scrapping that lot is much more significant
than the cost of a single scrapped lot in a DRAM fabricator. In general, one
should not fear scrapping wafers, to the extent that preventive measures result
in costs which are higher than the value of the scrapped wafers. Very often
100% inspection and measurement plans have been instituted to address
problems that occur only once every few years. The expense of these
inspections and measurements usually far exceeds the costs of the wafer scraps
that were avoided. The objective should be to minimize costs.

2.5 SYSTEMATIC SOURCES OF VARIATION

Some parameters vary non-randomly. As noted previously, because of the

geometry of resist processes, radial systematic variations across wafers are
common. In such circumstances, statistics that characterize the process, such as
the standard deviation, must be interpreted carefully. For example, consider the
situation depicted in Fig. 2.6. A parameter varies linearly from 0 to A, and
measurements are taken at equal intervals. The calculated sample standard
deviation varies by nearly a factor of two as the number of measurements varies
from three to more than 20. A similar situation occurs for parameters that vary
quadratically (Fig. 2.7). Quadratic systematic variations are often observed

when measurements are sampled along the diameter of a wafer that has
parameters that vary radially.

Sample standard deviation

Linearly varying parameter for a linearly varying parameter

A A

ó R
19

Ii
. •L

A/2 Asymptotic value

0
2 3 4 5 6 7 8 9 0 5 10 15 20 25 30
Measurement site Number of measurements

Figure 2.6. A linearly varying parameter, and the change in the sample standard
deviation as a funtion of the number of measurements. The graph on the left
illustrates a parameter sampled at nine evenly spaced sites.

Sample standard deviation

Quadratically varying parameter for a quadratically varying par a meter
A I
3 0.6 A
. ó
N g
a
0.3 A Asymptotic value ^^+
01
0 2 4 6 8 0 5 10 15 20 25 30
Measurement site Number of measurements

Figure 2.7. A quadratically varying parameter, and the change in the sample
standard deviation as a function of the number of measurements. As illustrated
on the left of the figure, the parameter is sampled at nine sites.

There can be significant consequences if systematic sources of variation are

handled carelessly. For example, the number of measurements is frequently
reduced as processes progress from development to the mature stages of
manufacturing. As seen in Fig. 2.6 and Fig. 2.7, in the presence of systematic
sources of variation it may appear that the variability of the process increases
when the number of measurements is reduced. This would not be case if there
was only random variation. In the absence of systematic sources of variation
the average measured variation is independent of sample size. More than one
lithography engineer has been asked to investigate apparent increased process
variability that was only the consequence of a reduced sampling plan for a
process with a significant systematic component.

As discussed in Section 2.2, reticles and lenses are sources of systematic

variation. Linewidth variations and overlay signatures will repeat in every
exposure field, but there will also be random components. The systematic
errors complicate the statistical characterization of variation. One estimate of
the distribution of errors, including the random and systematic contributions,
can be obtained as follows. Suppose the random error g(E) at any point z =
(x, y) on the wafer is approximately represented by a normal distribution about
the mean error at z:

(c-µ(Z)r
1 262
8(E, µ(z)) = 6 2rt e (2.26)

The distribution of errors overall is then given by

h (E)= 1 f g(F-,µ(z))dxdy, (2.27)

where the integrations are extended over the relevant areas of the wafers. In
general, a non-normal distribution will result, but the fraction of sites 0 E L )
that do not conform to specifications [ E L ,E L j can be computed using Eq.
2.28. 6

—EL o0

(E L )= f)dc + f h(E)dE (2.28)

-^ s

=1— ƒ h(E)ds . (2.29)

— EL

As an example, suppose that the mean linewidth varies radially across a

wafer:

µ(r)= A r2 , (2.30)
ro

where ro is the radius of the wafer, and r is the radial distance from the center
of the wafer. In addition to this systematic linewidth variation, at each point on
the wafer the linewidth varies normally about the mean µ(r) with standard
deviation c. The resulting linewidth distributions are shown in Fig. 2.8 for
various values for A (in units of 6) along a diameter. As A is increased the

distribution is shifted to the right, since the mean is increased, and the
distribution in broadened. The distributions also become skewed and flattened,
but the values for skewness and kurtosis will remain within the confidence
intervals shown in Tables 1.4 and 1.5, except for values of A much larger than
shown in Fig. 2.8.
As discussed in this chapter, there are several issues that must be
considered when determining the number and frequency of physical
measurements, such as linewidths, resist thicknesses, and overlay. Having
established some of the basic statistical issues with respect to process control, it
is time to discuss the nature of process monitors specific to lithography in more
detail. This discussion begins in the next chapter. The interplay between
physical measurements and processing parameters will also be discussed.
Further discussions of process drift and how to correct for it will be found in
Chapter 7.
Values for A
0.5
0
0.4
1
0.3 2

0.2

0.1

0
-4 -3 -2 -1 0 1 2 3 4
Variable (in units of sigma)
Figure 2.8. The probability density functions for three values of A in Eq. 2.31.
There is a shift in the mean as a consequence of the quadratic variation in the
mean across the wafer. The distribution also acquires a slight skew and becomes
somewhat flatter for non-zero values of A.

1F. S. Hillier, " X - and R-Chart Control Limits Based on a Small Number of
Subgroups," Journal of Quality Technology, Vol. 1(1), pp. 17-26 (1969).

2C. P. Quesenberry, "SPC Q Carts for Start-Up Processes and Short or Long
Runs," Journal of Quality Technology, Vol. 23(3), pp. 213 —224 (1991).

3 C. Yu, T. Maung, C. Spanos, D. Boning, J. Cheung, H-Y Liu, K-J. Chang, and
D. Bartelink, "Use of Short-Loop Electrical Measurements for Yield
Improvement," IEEE Trans. Semicond. Manuf., Vol. 8(2), pp. 150 — 159
(1995).

4R. C. Elliott, R. R. Hershey, and K. G. Kemp, "Cycle-time Reduction of CD

Targeting using Automatic Metrology and Analysis," SPIE Vol. 2439, pp.
70-77(1995).

5G. E. P. Box, G. M. Jenkins, and G. C. Reinsel, Time Series Analysis, 3rd

Edition, Prentice Hall, Englewood Cliffs, NJ (1994).

6T. R. Groves, "Statistics of Pattern Placement Errors in Lithography, " J. Vac.

Sci Technol., B9(6), pp. 3555 — 3561 (1991).

3.1 DEFINITIONS

The purpose of statistical process control is the identification of abnormal

variations in the materials, equipment, parameters or procedures used as inputs
for particular processes (see Table 3.1). This is usually accomplished by
measuring physical objects on the output, such as linewidths or overlay
structures. When processes require test wafers, which often occurs during
process development or the initiation of a new manufacturing process, the
output of the process is decoupled from the input, and statistical process control
cannot fulfill its primary purpose.

Input parameter Examples

Materials Photoresist, developer, anti-reflection coating,
substrate films
Equipment Particular stepper model, resist track
Procedure Operating specifications, linewidth target and
allowable range
Parameters Exposure dose, overlay offset, focus setting

Table 3.1 Process input parameters and typical examples.

This situation is shown schematically in Fig. 3.1. The inputs collectively

comprise the process. A single wafer is taken from a lot of wafers and
processed through the lithography operation. After the processing is complete,
this test wafer is measured for parameters of interest, such as linewidths or
overlay. From the values of these measurements, in comparison to the process
targets, the remaining wafers in the lot are processed through the lithography
operation with adjusted process parameters. For example, the exposure dose
might be adjusted to bring linewidths to the process target. Changes in the dose
might be required to compensate for drift in the stepper's dose control system
or the changes in the resist process. Measurements of the linewidths of the lot,
except for the test wafer, will not reveal a drift in the stepper's dose control
system or a change in the resist process, because the exposure dose has been
adjusted to compensate for these instabilities. In order for statistical process
control to reveal variations and instabilities in the inputs, it must be applied to a
simple process, where the input variables are directly coupled to the measurable
output.' In this chapter a process control methodology applicable to situations
in which test wafers are used is presented.

Materials —

Equipment — Lithograph. Measure Parameter

test Operation --► Output
Procedure — operation adjustment
wafer
Parameters —

Figure 3.1. Material flow in a complex process involving test wafers.

3.2 WHY TEST WAFERS ARE USEFUL

The use of test wafers is common in lithography, particularly for processes in

development or the early stages of manufacturing. Exposure doses are changed
to adjust critical dimensions, stepper alignment offsets are adjusted to improve
overlay, or a spin speed might be changed to adjust resist thickness. This
practice of making adjustments based upon a small set of data, while common,
is contrary to the principles of statistical process control. The idea that
adjustments to processes are in conflict with statistically sound practices was
preached by the great W. Edwards Deming, who illustrated the problem
through his funnel experiment. 2
The funnel experiment is conducted by dropping a marble through a funnel,
which is held above a table which has a target marked at a single point (Fig.
3.2). A marble is dropped through the funnel. After the marble lands on the
table it will bounce, roll and eventually come to rest. The distance between the
target and the final resting place of the marble is measured. This process is
repeated, with the objective of producing the tightest cluster about the target of
finally resting spots. Four rules for the process are proposed:

Rule 1. Leave the funnel fixed, aimed at the target, and make no
adjustments.
Rule 2. At drop k, the marble will come to rest at point xk . Move the
funnel to — xk from its last position for drop k+1.
Rule 3. Set the funnel over point — k' measured from the target, for drop
k+ 1.
Rule 4. For drop k+1, set the funnel over the previous spot where the
marble came to rest after drop k.

Rules 2 — 4 are procedures for adjusting the process in attempts to tighten the
distribution about the target. The results of operating according to these rules
are:

Rule 1. This rule produces the tightest cluster of resting spots about the
target.
Rule 2. The output from this process is stable, but the variance is twice that
of Rule 1.
Rule 3. This system is unstable. The marble will eventually move further
and further from the target.
Rule 4. This system is also unstable.

i .

Figure 3.2. The funnel experiment.

The funnel experiment was designed to illustrate the futility of making

adjustments based on the outcome of a randomly varying process, since all
attempts to improve the process had the opposite effect. On the other hand,
experience with test wafers has shown that improvement is possible using data
from only one wafer. This apparent contradiction can be understood from the
nature of the sources of variation that occur in lithography. Adjustments
typically are applied soon after test wafers are processed and measured.
Consequently, long-term drift is not a factor which affects the applicability of
test wafer data to the rest of the wafers in the lot. 3 There are several physical
phenomena that lead to long term drift. For example, the overlay of wafer
steppers is affected by the temperature of the stepper. In spite of efforts to
control the equipment, the stepper's temperature will vary by small amounts.
However, because of the large thermal mass involved, this temperature change
can occur only very slowly. Effects which result from changes in barometric
pressure will also vary at the rate the pressure changes significantly, which is
typically over the course of several hours.
Linewidths and overlay are also affected by the substrates, which are often
more repeatable within a lot of wafers, compared to the lot-to-lot variation. Test
wafers capture the substrate signature of the lot. There are also events that are

identifiable, which indicate that the process has changed. This includes new
batches of photochemicals, equipment maintenance, and new reticles.
The funnel experiment was designed to demonstrate the futility of making
adjustments to processes in a crudely reactive mode. Sources of variation that
are not identified or subjected to control will apparently lead to "natural" levels
of variation. There is considerable difference between identifying and
controlling sources of variation and adjusting processes based solely upon prior
outcomes. There are no sources of variation, at the level of significance for
semiconductor processing, that are inevitable and cannot be eventually
subjected to control. The decision to establish control is based upon knowledge
of the sources of variation and an assessment of cost-effectiveness.

3.3 HOW TO ADDRESS COMPLEX PROCESSES IN LITHOGRAPHY

In lithography, the use of test wafers is usually justifiable, and often helpful.
However, when they are used, statistical process control cannot be applied
conventionally for controlling the process. Without one of our most powerful
quality improvement tools, it becomes difficult to improve the process so that
test wafers can be eliminated and efficiency improved. In this section, a method
will be presented for applying statistical process control to complex processes.
This method is a tool for process monitoring and improvement.
Conventionally, processes are monitored and controlled by trending and
analyzing physically measurable data such as linewidths or overlay. We have
seen the difficulties in applying this method to complex processes. Instead of
monitoring physically measurable data, complex processes may be controlled
by trending and analyzing process parameters 4 (Fig. 3.3). In this approach, the
process parameters used to process each lot are trended. For example, one of
these parameters could be the exposure dose. For a completely stable process
the exposure dose would be constant. Variations in resist materials, exposure
tools and substrates would require changes in the exposure dose in order to
achieve linewidths that are close to the target value. Fluctuations in the
exposure doses used to process lots would reflect these variations.
The basis for this approach can be appreciated analytically. Consider a
situation involving critical dimensions. Linewidths are determined by a number
of factors, such as exposure dose (E), resist thickness ( t reslst ), post-exposure
bake temperature (TPEB ), and so forth. Conceptually, the linewidth is a function
of these factors:

linewidth = x = f (E,tresist,TPES,...) (3.1)

even though the exact form of the function f may not be known. It should be
noted that to the level of precision relevant to microelectronics processing, this
equation is deterministic, a consequence of the natural laws of physics and
chemistry.

Materials

Equipment vteasurable

Procedure output

Parameter

Lot

Figure 3.3. Process parameter monitoring. Shown in this example is a feedback

loop.

If all of the independent variables (E, tresist TPEB •••) could be held I ,

exactly at their set points, the linewidth x would equal its target value x0 .
Variations in x are a consequence of fluctuations in the values of E, trist
TPEB , and the other independent parameters about their set points. For example,
a hotplate may be set to a particular temperature, but the actual temperature will
vary about its set point, and linewidths will vary as a consequence of this
variation in temperature. Monitors of the linewidth, where the equipment
settings for all process parameters are set to fixed nominal values, will show
variations that indicate that one (or more) of these independent variables is
fluctuating. Monitors of the linewidth cannot reveal uncontrolled fluctuations of
independent variables if one of these parameters is varied intentionally, such as
when the exposure dose is changed on the basis of test wafer measurements,
unless one can predict exactly how the change in dose will vary the linewidth
and compensate accordingly. On the other hand, Eq. 3.1 can be inverted, at
least conceptually, to express the exposure dose in terms of a function f' of the
other independent parameters and the linewidth.

E= f x, t resist , TPEB 1 "'J

(3.2)

When the linewidth x in Eq. 3.2 is held at its target value x 0 , the right side of
this equation consists entirely of factors which are nominally constant, i.e.,
none of these parameters are varied intentionally. This is the situation that
arises when test wafers are used. The linewidth target x0 is held constant, as

are the settings for all of the other parameters, and it is the exposure dose on the
left-hand side of the equation that is varied. Consequently, variations in the left
side of the equation indicate real fluctuations in the factors on the right-hand
side about their set point values. Monitors of E can thus be used to monitor and
control the process.
Alternatively, the control methodology can be based on physical
measurements from the test wafers rather than the processed lots, so long as the
test wafers are always processed using the same equipment set point values. In
this case the measurements do reflect real variations of the inputs. However,
there is an advantage to the approach where process parameters, such as
exposure doses, are trended. Because it is desirable to minimize rework, the
monitoring of process parameters provides the feedback for adjusting the
process inputs to maximize the likelihood that the test wafer will be within
specifications 5 (shown in Fig. 3.3). In fact, it was in the context of reducing
rework and test wafers that this methodology was first used. 6
This approach to process control, by monitoring operating parameters such
as exposure doses, addresses the fundamental problem introduced by test
wafers, the decoupling of the quantity that is tracked from the process inputs.
However, there are a number of problems with this approach that need to be
addressed. First, adjustments to operating parameters are not always made.
When the results of test wafers are very close to target, there may be a
reluctance to adjust the process parameters. As a consequence, the distribution
of process parameters used on product lots will look like the histogram shown
in Fig. 3.4. Many lots are processed at a single value, while there are no lots
that are processed with the process parameter near this value. The gaps result
from lots which have test wafer results near the target. The distribution shown
in Fig. 3.4 is very non-normal, and the data need to be analyzed accordingly. In
particular, the Western Electric Rules cannot be applied to these data directly.

U
a..
w

Parameter value

Figure 3.4. Histogram of a process parameter used to process lots. The ordinate is
the process parameter used to process the lots.

There are also some problems in targeting the parameters used for
processing lots. When single test wafers are used there is inevitable statistical
noise which results from wafer-to-wafer variation. Moreover, even though the
short-term drift is significantly less than the long term drift, there still may be
sufficient drift to affect the targeting of operating parameters. Finally, the
relationship between the physical parameters and the operating parameters must
be taken into account properly. Consider, for example, critical parameters as a
function of exposure doses (Fig. 3.5). Over a sufficiently large range, this
function is non-linear. In manufacturing, there is a trade-off between simplicity
and accuracy, and a linear approximation is often used. The relationship
between linewidth and exposure dose is affected by other parameters, such as
focus. Consider the curves in Fig. 3.6. A smaller change in dose is required to
change the spacewidth by a certain amount at non-zero defocus, compared to
the dose change required at zero defocus. If a test wafer is processed with some
amount of defocus, the remainder of the lot will not be targeted correctly if the
amount of change in spacewidth is estimated on the basis of data collected at
zero defocus.

r_
0

ca?
6 . ....................... ...._.................•.
U

Exposure dose

Figure 3.5. Linewidth as a function of exposure dose, showing typical non-

linearity.
Bossung curves
280
Dose
260 0.9

9
2 240

1: 180
1.1
0

t
-

1.2

160
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Defocus (microns)
Figure 3.6. Spacewidth variations as functions of dose and defocus. The
spacewidth varies more with dose at larger values of defocus.

Because of the various problems with the method of monitoring process

parameters, a more sophisticated form of the method is used where there are
good data handling capabilities. In this approach, the lot data are also used to
refine the estimate of what the process parameter should have been (Fig. 3.7). ?

Highly automated data handling is required, since the process parameters that
are trended are not simply the parameters used for the lot, but they are corrected
using measured data after the lot is processed. This increases the amount of data
on which the trended parameter is based, since multiple wafers are often
measured after the lot has completed processing. This method involves
significant quantities of data and must be automated, particularly since process
parameters can be specific to each combination of product, masking layer, and
product. Small corrections will be included automatically, correcting the
deficiency illustrated in Fig. 3.4 without making the job more complicated for
operators.
Automatic feedback should always be applied with caution, since
adjustments are being made in one parameter to compensate for variations in
another. This is generally acceptable, within limits, for certain types of
corrections, such as using exposure dose to compensate for variations in the
resist processing. On the other hand, it would be best to correct a change in
focus rather than compensate for its effect on linewidth by changing the
exposure dose. Active process control will be revisited in Chapter 7.

3.4 DISTINGUISHING BETWEEN LAYER-SPECIFIC AND EQUIPMENT-SPECIFIC EFFECTS

Because of the immaturity of film deposition operations, anneals, etch

processes, and polishing, substrates may be quite variable when bringing up
new manufacturing facilities. This will lead to large fluctuations in linewidth
and overlay. At the same time, there will be little confidence in the stability of
the lithography process itself, because the lithography operation will be
similarly immature. This problem is also quite common on development pilot
lines, where non-lithography processes, which can have profound effects on
lithography, are changed frequently. Even in mature manufacturing operations
there can be losses of process control that can cause substrate variability.
Lithographers need a method to distinguish between variations in substrates and
fluctuations caused by variable lithography processes and equipment.
A method capable of accomplishing this involves monitors of the
lithography equipment and processes that employ extremely reproducible
substrates. For most applications, bare silicon wafers work fine. For example, a
silicon wafer can be processed through the lithography operation daily, or some
other consistent interval, and the linewidths can be measured. These
measurements provide a baseline indicating the fundamental stability of the
lithography operation. Variations seen on such monitors can be expected on
product wafers. If these monitors are relatively stable, then large variations on
product wafers can be attributed to some mechanism involving the substrate.

U
cici
`^.á
R r.
^ o O
10
a)
rh
Cd
.a
0 y C

O r"eó E
ia

J O
I.

Ero
s..

b1D
I SS Oid

U V]

ó O m
v vW
o - -

UI vi ,^ C M

é7 £ ,
o. b E .
ri..^PP- z aEi
ce ó. ó
w a w

An example of such a monitor is shown in Fig. 3.8, which was used to

monitor linewidth control on early deep ultraviolet (DUV) excimer laser
steppers. 24 These steppers were being used on a development pilot line. At the
same time, the photoresists were undergoing development, with large variations
in photospeed between batches. The steppers were very immature, and there
was a steep learning curve that needed to be climbed. Because the substrate
composition was changing frequently as a result of the development function of
the pilot line, large linewidth variations could be expected. These needed to be
distinguished from variations in resist materials and lithography equipment.
The bare silicon image process monitor enabled the engineers to monitor the
process, identify changes, and correct them prior to processing "product"
wafers.

0.6

0.55
^— New lot of photochemicals

0.5

^c 0.45

: 0.4
U
0.35
^-- Adjustment to excimer laser
0.3
6 11 16 21 26 31 36 41
Day
Figure 3.8. Linewidth monitor for an early excimer laser photocluster.

Patterned wafers that are measured on linewidth measurement tools provide

good monitors for the entire imaging process. The measurements often require
expensive equipment and can take a long time. Photoresist processing
equipment and steppers can be left idle while measurements are being taken.
Measurements of E 0 , the minimum exposure dose to clear positive resists to
the substrates, provide quicker monitors that do not require expensive and time-
consuming metrology, and which provide most of the information needed to
control resist processes. $ For these types of monitors, "open frame" exposures
are used. These are large area exposures, typically several millimeters per side,
often accomplished by having no reticle in the stepper. The dose at which the
resist is first cleared from the substrate can be determined from visual
inspection. Measurement noise can be reduced and interpolation between doses
is possible by measuring resist thickness at each exposure and fitting the data to
a common resist contrast curve, with only a small increase in time. Nearly all

positive photoresists have similar resist contrast curves, where resist thickness
varies linearly as the logarithm of the dose, in the vicinity of E0 :

thickness = Tyln EE , (3.3)

where To is the initial resist thickness, 'y is the resist contrast, and E is the
exposure dose. Measurements of E0 have been used to identify fluctuations in
materials and processes. 9 Linewidth measurements provide better monitors for
very high contrast resists, such as many DUV materials, because E. is very
hard to measure for these resists.
Similarly, "golden" or "holy" wafers can be used for controlling
overlay. 10 '" Like the wafers used for monitoring linewidths, these should be
well controlled substrates. Generally, they consist of patterns etched into
substrates. The substrates should be chosen so that the alignment targets and
overlay measurement structures are free from noise factors, such as grains in
films, that will degrade the utility of these wafers for monitoring the status of
the equipment. Quite frequently, patterns etched into bare silicon are ideal. Bare
silicon substrates also do not expand or contract from being etched, as might
occur from etching deposited silicon nitride or oxide films, and this enhances
the reproducibility of the substrates. Wafers will expand or contract because of
changes in temperature, regardless of composition, and any overlay control
scheme must take this into account. Silicon has a coefficient of thermal
expansion of 3 ppm/°C. 12 Points on opposite sides of a 200 mm wafer will
move 60 nm apart for every 0.1 ° C change of temperature. This represents a
substantial fraction of overlay budgets for 250 nm technologies and smaller.
Equipment and processes may drift, and these changes may not be
observable on monitors using ideal substrates. For example, an alignment
system may drift, causing difficulty aligning wafers whose alignment targets
provide low signal-to-noise. This drift may not be observable for targets that
provide strong alignment signals, as should exist on the "golden" wafers.
Similarly, problems in resist processing may occur only for certain substrates,
or wafers with topography, and therefore may not be observed on linewidth
monitors that use bare silicon substrates. Although there may be process
problems that may go undetected using process monitors that employ particular
ideal standardized substrates, these types of monitors remain extraordinarily
useful. When there are problems with overlay, there is no substitute for the
knowledge that the steppers are (or are not) providing good overlay on
standards which, if not etched into stone, are at least etched into silicon.
Processes change when operating parameters are adjusted, and these
changes must be taken into account when monitoring processes. This is
particularly important in lithography, where test wafers are often used, and
process parameters are adjusted on the outcomes of test wafer measurements. A

method for addressing alterations in process parameters was presented in this

chapter. It is also important to have a capability for distinguishing between
changes in the non-lithographic parts of the semiconductor manufacturing
processes and those that occur in the lithography process itself. The use of
"golden" wafers provides this capability. One can take greater advantage of
these techniques by going beyond purely mathematical analyses of the data and
using knowledge of the physics and chemistry specific to lithography to more
fully interpret the results. We now turn to the science of lithography.

' H. J. Levinson, "Control and Improvement of Complex Processes," Quality

Engineering, Vol. 5(l), pp. 93 —106 (1992).

2 W. Edwards Deming, Out of the Crisis, MIT Center for Advanced

Engineering Study, Cambridge, MA (1982).

3 J. F. MacGregor, "A Different View of the Funnel Experiment," Journal of

Quality Technology, Vol. 22, pp. 255 — 259 (1990).

4 L. C. Mantalas and H. J. Levinson, "Semiconductor Process Control," in

Handbook of Critical Dimension Metrology and Process Control, K. M.
Monahan, ed., SPIE Press, Bellingham, WA (1994).

5M. Drew and K. Kemp, "Automatic Feedback Control to Optimize Stepper

Overlay," SPIE Vol. 1926, pp. 422 — 428 (1993).

6 H. J. Levinson and C. DeHont, "Leading to Quality," Quality Progress, pp.

55 — 60 (1992).

7C. P. Ausschnitt, A. C. Thomas, and T. J. Wiltshire, "Advanced DUV

Photolithography in a Pilot Line Environment," IBM J. Res. Develop., Vol.
41(1/2), pp. 21— 36 (1997).

8D. Heberling, "Litho Equipment Matching with E 0 ," Proceedings of the KTI
Microelectronics Seminar, pp. 233 — 243 (1990).

9C. Takemoto, D. Ziger, W. Connor, and R. Distasio, "Resist Tracking:

A Lithographic Diagnostic Tool," SPIE Vol. 1464, pp. 206 — 214 (1991).
10
M. A. van den Brink, C. G. M. de Mol, H. F. D. Linders, and S. Wittekoek,
"Matching Management of Multiple Wafer Steppers Using a Stable Standard
and a Matching Simulator," SPIE Vol. 1087, pp. 218 — 232 (1989).

" K. Kemp, C. King, W. Wu, and C. Stager, "A "Golden Standard" Wafer
Design for Optical Stepper Characterization," SPIE Vol. 1464, pp. 260 — 277
(1991).
12
Handbook of Chemistry and Physics, 61st Edition, CRC Press, Boca Raton,
Florida (1981).

In the preceding chapters the focus was primarily on some of the mathematical
tools available for controlling processes. Statistical process control is a
methodology that enables people to identify abnormal levels of variation. Once
it has been determined that a process is out of control, the sources of the
excessive variation or process drift need to be identified and corrected.
Moreover, it is preferred to have processes that normally have low levels of
variation. In order to design processes which are intrinsically stable or to take
corrective action when processes go out of control, it is necessary to understand
and apply the science of lithography. For example, when linewidths drift, one
needs to understand what parameters can potentially cause changes in the
process.
Lithographers need to control linewidths, overlay and defects. Aspects of
lithographic science relevant to linewidth control will be discussed in this
chapter. Similar discussions for overlay will follow in Chapter 5, and for yield
in Chapter 6.

4.1 CAUSE AND EFFECT

There are a large number of variables that can affect linewidths, many of which
will be discussed in the next section of this chapter. If control of a process is
lost, it is the job of the process engineer to determine what specifically caused
the abnormal change, among all possible causes, so that appropriate corrective
action can be taken. A technique for identifying probable causes, particularly
useful for brainstorming with groups, is the cause-and-effect diagram.' The
underlying principle of this technique is causality, i.e., there is a cause for every
observed phenomenon.
An example of a cause-and-effect diagram is shown in Fig. 4.1. The cause-
and-effect diagram is often called the Ishikawa diagram, named for its inventor,
and is sometimes referred to as a fishbone diagram, because of its appearance.
In the cause-and-effect situation, the effect is the loss of process control. There
can be a number of possible causes for this. For manufacturing, a universally
applicable cause-and-effect diagram has been constructed, shown in Fig. 4.2.
All losses of control can be traced to one of the six categories represented in
Fig. 4.2.
Having identified certain general classes of causes, it remains to determine
the cause of poor process control more specifically. To this end, more detailed
causes can be linked to the general category, in the graphical format shown in
Fig. 4.3. Suppose post-exposure bake (PEB) temperature is suspected as being
the cause of a change in linewidths. There are a number of reasons that the PEB

Primary Primary
cause im #2
( cause an
cause

Control

Primary Primary Primary

cause #4 cause #5 cause #6

Cause Effect
Figure 4.1. The basic form of the cause-and-effect diagram.

Control

Cause Effect
Figure 4.2. A universal cause-and-effect diagram for manufacturing.

PEB I Exhaust
temperature
Probe
calibration
Hot-plate feedback
control Procedure

cleanliness \ _/ Frequency

Linewidth
control

Figure 4.3. A branch of the cause-and effect diagram, for the problem of
-

linewidth control.

temperature may have changed, and some possibilities are shown. Some of the
suspected causes may have more specific underlying origins, and these can be
shown as well.
The cause-and-effect diagram is qualitative, in that it does not distinguish
relative magnitudes of effects due to specific causes. After collecting possible
causes, engineers should decide which causes are most likely and most
significant, and take appropriate measurements to determine if these are indeed
the primary reasons that control has been lost.
Process control is often lost in high stress situations. Under such
circumstances there is a tendency for fingers to be pointed at people rather than
technical factors. The cause-and-effect diagram is useful for refocusing
engineers and managers on the technical problems instead of people. The
human aspect of process control will be considered further in Chapter 9.

4.2 INDEPENDENT VARIABLES

As expressed in Eq. 3.1, linwidths are functions of particular process variables.

The most significant of these process factors are listed the Table 4.1. In
addition to these variables, there are other factors that are more specific to
particular processes. Consider, for example, a puddle develop process. A
puddle process is one in which developer is dispensed over the photoresist,
thereby forming a puddle. After a preset period of time the developer is
removed from the wafer, usually by using the centrifugal force of spinning. The
resulting linewidths will depend upon the volume of developer dispensed,
whether it is a single or multiple puddle process, the speed at which developer
is applied, and the dispense nozzle configuration. The remainder of this chapter

will be devoted to a more detailed discussion of the parameters in Table 4.1,

and the ways in which they are known to affect linewidths.
Exposure dose Developer normality
Focus Developer time and temperature
Resist thickness Substrate reflectivity
Softbake time and temperature Resist composition
Post-exposure bake time and temperature Ambient amine levels (DUV resists)

Table 4.1. Important variables which affect linewidths.

It is essential to know quantitatively how each process parameter affects

linewidths and what the normal levels of variation are for each parameter. For
example, suppose that the post-exposure bake changes linewidths by 2 nm/ ° C.
With this knowledge it is possible to conclude that an 0.5 ° C change in the
temperature of the hotplate used for the post-exposure bake cannot account for
an observed 10 nm shift in the mean linewidth, i.e., even though the hotplate
temperature did change, the magnitude of the change was insufficient to
account for an observed linewidth shift. It is also essential to know facts such as
the normal 36 variation of hot-plate temperatures. In general, the relationship
between the change in a critical process parameter, such as linewidth, and
changes in parameters, such as bake temperatures, can be expressed using
Taylor's theorem. Let the function f represent the critical process parameter,
and x the independent variable. If x o is the nominal operating point for x, then:

f(x)= f(x0) +_(x—x°)+ 1 dx

2
2 (x — xo 2 +... ,
) (4.1)

where the derivatives are evaluated at x = x0 . The change in the critical

parameter, as a consequence of a change in the independent variable from its
nominal value, is:
2
Í(x) — f(x0 )=-(x — x°f (x—x 0 ) 2 +... (4.2)

In many circumstances only the first term on the right-hand side needs to be
retained, and there is a linear relationship between the linewidth and the
independent variable. To understand the change in the linewidth, one needs to
know the magnitude of df , which is the rate of change (e.g., 2 nm/ C), while
(x — x0 is the amount the independent variable changes (e.g., 0.5 °C). If x varies
)

normally about x o with standard deviation a, then f will also vary normally,

with standard deviation ^ • 6 , when f and x are related linearly. In a number of

circumstances, f
d = 0. This occurs, for example, when x represents focus. 2 In
dx
this situation the relationship between the linewidth and the independent
variable is more complicated, but tractable, since the methodology just
described can be applied using the next higher order terms in the expansions
Eq. 4.1 or 4.2. This more complicated relationship arose in the earlier
discussion on skewed distributions (Section 1.4). A software package,
NORMAN, has been developed at IMEC 3 that combines the deterministic
relationship between linewidth and process parameters — such as focus, dose,
resist thickness, and reticle dimensions — and the probabilistic variation in
these very parameters, to produce histograms of linewidth.4, 5
Such modeling software is very useful for process improvement, because it
can be used to rank the sources of variation quantitatively. 6 Plots of the sources
of variation in the format shown in Fig. 4.4 are referred to as Pareto Charts.' In
the Pareto Chart, contributors are ordered from left to right according to the
magnitude of their contribution. In order to do this, contributions must be
characterized according to a common metric. This type of analysis is very
useful, because it shows clearly the improvements that will reduce linewidth
variation most significantly. Those who manage lithography operations or
engineering projects should use Pareto Charts in order to focus resources in
those areas where improvement will have the biggest impact.

40
.c
12 30
0

ó 20

1 10

y ^^^octo^ -
^^1oß

Process factor

Figure 4.4. An example of a Pareto chart.

There is one type of process change that is highly desirable —

improvement. Better processes result from identification of the causes of

variation, and the application of controls to the sources. Because resources are
limited, improvement efforts should be focused on the largest sources of
variation. Controlling these will provide the greatest improvement to the overall
process.
Models can also be used for improving quality by indicating that the
accounting of sources of variation is incomplete. Independent measurements of
sensitivities to parameters such as resist thickness and focus, and the amount of
variation normally seen for these parameters, can be used as inputs to a model
for calculating total linewidth variation. 7 If the model produces less variation
than the actual process, then some significant source of variation has been
overlooked. Processes may be improved by identifying and controlling this
factor.
Additional engineers and technicians and new equipment are often
required in order to improve process control. Since these cost money, it is
essential that managers be prepared to justify increased expenditures. The cost
of inadequate process control, due to rework and lost yield, can be compared to
the costs of hiring additional staff or purchasing new equipment, if the quality
costs have been estimated in monetary terms. 8
If sufficiently large variations in parameters occur, the utility of Taylor
expansions, such as Eqs. 4.1 and 4.2, is reduced. Consider the variations in
linewidth shown in Fig. 4.5, which were calculated using the lithography
simulation program, PROLITH2. 9 If the resist thickness varies by an amount in
excess of the quarter-wavelength of light, the functional form of linewidth
variation goes considerably beyond the first few terms in a Taylor series
expansion. If a particular process represents such a situation, then the concepts
in this section can be extended by accounting for the more complicated
functional dependence of linewidth on particular parameters. Processes which
are centered at a minimum or maximum of the swing curve, and where the
resist thickness varies less than a quarter-wavelength of light, will have skewed
distributions as a consequence of thin film thickness variation.

0.65
2800 A
2 0' 6 oxide
0.55
2500A
0.5 oxide

0 0.45

0.4
8000 9000 10000 11000 12000
Resist thickness (A)

Figure 4.5. Variations in linewidths as a function of resist thickness, calculated

using PROLITH 2. The nominal linewidth was 0.5 microns, and the optics were i-
line. The substrate was oxide on silicon.

While each of the process factors summarized in Table 4.1 affect linewidths
in quantifiable ways, each parameter is itself the consequence of additional
parameters, most of which are intimately related to equipment and materials.
For example, bake time and temperatures will be functions of hotplate
temperature control, its response time and calibration, and drift. Ambient
temperatures and exhaust, as well as wafer backside contamination also affect
the effective bake temperature. Process control requires an understanding of the
specific equipment and materials used.
4.2.1 Exposure dose

The amount of light incident on resist-coated wafers is controlled on steppers,

using closed-loop feedback control (Fig. 4.6). Dose repeatability on g- and i-
line wafer steppers is typically <± 1%, and <± 1.5% on DUV systems. The
absolute control of dose is limited by calibration errors, which are currently
about 2% for 248 nm excimer laser illumination. 10 Generally, linewidths remain
within specifications as long as the amount of light absorbed in the resist varies
by no more than ± 10%, so the level of dose control provided by wafer steppers
is quite adequate for 250 nm and older technologies. Improved levels of dose
control will be required for optical lithography well below the wavelength of
light.
Conditioning Light
Beamsplitter optics Shutter source
^I
To
UU 99%
reticle V

Detector Electronics --►

I Shutter
control

Figure 4.6. The amount of light exiting the illuminator is monitored constantly,
and a closed-loop system ensures that the stepper produces a controlled exposure
dose independent of lamp or laser brightness.

The amount of light absorbed by resist is determined by the light reflected

by the substrate as well as the incident light. Consider a situation involving
silicon wafers with uniform oxide films on them. If the wafers are coated with a
photoresist film of varying thicknesses and then illuminated with uniform
exposures of light, the total amount of light absorbed by the resist films is
shown in Fig. 4.7. Because of thin film optical interference effects, the amount

of light absorbed varies by ± 24% as the resist thickness varies over the range
of thicknesses shown in the graph. Because of the manner in which the
dependent variable "swings" up and down as a function of the film thickness,
graphs such as those in Fig. 4.7 are referred to as swing curves.

0.9
2800 A
ó 0.8
Gn
oxide
0.7
2500 A
- C
°
=, 0.6 oxide
0.5
ao
0.4
8000 9000 10000 11000 12000
Resist thickness (A)

Figure 4.7. The amount of light absorbed by a resist film under uniform i-line
illumination. This is the same configuration for the curves shown in Fig. 4.5.

It should be noted that swing curves obtained with particular numerical

aperture and illumination settings may not apply in situations where different
values for these parameters are used." This may be understood as follows. The
maxima and minima in Fig. 4.7 occur when particular phase relationships are
satisfied between the incident and reflected light (Fig. 4.8). These relationships
will depend on the optical path lengths of the light in the photoresist film, and
these will be different for non-normal light at different angles of incidence (Fig.
4.8b).
Incident Reflected Incident Reflected
light light Ii\ lit

Photoresist Photoresist
Substrate Substrate

a) Normally incident light b) Non-normally incident light

Figure 4.8. Incident and reflected light in photoresist films. Light has different
path lengths for normally and non-normally incident light. Consequently, the
incident and reflected light have different phase relationships depending upon the
angle of incidence.

High numerical aperture, off-axis illumination, and the pattern pitch will all
affect the angles of incidence and hence the resist thicknesses at which swing
curve extrema will occur. This can be understood by considering the imaging of
a simple diffraction grating. Consider a beam of light that passes through a
grating consisting of long, parallel lines and spaces. If the light comprising this
beam is all of one wavelength, and all of the waves in the beam are in phase,
then the light will diffract into well-defined beams 12 (Fig. 4.9). Suppose the
incident beam is in a plane perpendicular to the lines of the grating. Then the
diffracted beams will be transmitted at angles 0 given by

sin8—sin8 0 = M
X (4.3)
p

where p is the pitch of the grating, 0 0 is the angle of incidence, m is an integer,

and X is the wavelength of the light. This physics has implications for imaging
in lithography. Depending on the angle of incidence of the illumination and the
pitch of the pattern, the light from a grating pattern projected through a lens
will have different angles of incidence 0 on the wafer. The path length through
the photoresist will depend on 0, as will the exact positions of the minima and
maxima on the swing curves. Consequently, the positions of swing curve
extrema will be functions of the pitch, illumination conditions, and the
numerical aperture.

Condenser
lens

Reticle

m = +1

Projection
lens

Wafer

Figure 4.9. A beam of light passing through a grating pattern.

Large linewidth variations can also occur when the thickness of the oxide
underlying the photoresist is changed by only a few hundred angstroms. From
the graph shown in Fig. 4.5 it is evident that a process will have, for a particular
oxide thickness, the least linewidth variation as a consequence of changes in
resist thickness if the target value for the thickness of the resist is either a
maximum or minimum of the appropriate curve. If the oxide thickness is well
controlled, then this is sufficient. There are circumstances, often the result of
chemical-mechanical polish, where the oxide thicknesses can vary
considerably, and there is no stable operating point for the resist thickness. In
such a circumstance, something is needed to address the fundamental cause of
the swings in the graphs of Figs. 4.5 and 4.7, which is the reflectivity of the
underlying silicon substrate.
Bottom anti-reflection coatings reduce the reflectivity of the effective
substrate, while top anti-reflection coatings reduce the phase-coupling of
incident and reflected light, 13 thus reducing the magnitude of the swings shown
in Fig. 4.5. Anti-reflection coatings do increase process complexity and cost, so
there has been some reluctance in using them. However, it is often true that the
resulting improvement in linewidth control provides greater value than this
cost, and the optimal solution is often now found to involve the use of anti-
reflection coatings.
4.2.2 Resist thickness

In order to maintain operation at a swing curve extremum, resist thickness must

be well controlled. A number of parameters modulate resist thickness and
uniformity, and these are listed in Table. 4.2. The specific values for these
parameters will depend upon the specific equipment and resists that are used.
Nevertheless, it is very useful to know which parameters can be relevant to
resist coating uniformity and control.
Temperatures play a significant role in resist coating process. Through the
use of a chill plate prior to resist coating, wafers are brought to a controlled
temperature. Resist thickness non-uniformity characterized by thick resist in the
middle of the wafer can often be improved by raising the temperature of this
chill plate. The ambient air will be at a certain temperature, which is usually
controlled and independent of the chill plate temperature. The resist that is
applied to the wafer can be at yet a third temperature. Resist thickness and
uniformity are functions of these temperatures,' 4' 15 though the magnitudes of
sensitivities also depend on the particular resist solvent. For example, reported
values for changes in mean resist thickness range as low as 15 A, 16 and as high
as 65 A 17 to 153 A, 51 for a degree change in the temperature of the ambient air.
The temperature of the chuck, and the materials it is composed of, can affect
resist thickness uniformity. 18 Temperatures for the coating processes are usually
chosen to minimize thickness variation across wafers and wafer-to-wafer, while
the process average is modulated through the spin speed and resist viscosity.

Parameters which affect resist thickness and uniformity

Temperatures Resist
Ambient air
Wafer
Chuck
Softbake hot late
Spin speeds Speed during dispense
Final spin speed
Accelerations
Dispense Resist volume
Dispense rate
Arm movement
Timing Casting time
End of dispense to cast delay
Exhaust Velocity

Table 4.2. The parameters that affect resist thickness and uniformity.

Besides temperature, there are a number of other parameters that determine

resist thickness uniformity and reproducibility. Humidity will also affect resist
thickness. Sensitivities of 11 — 25 A for a 1% change in relative humidity have
been reported.27, 19 Sensitivity to humidity is very dependent upon the resist
solvent.
Exhaust is a critical parameter, which needs to be controlled dynamically, 20
since the exhaust provided to most semiconductor equipment will fluctuate as a
consequence of varying loads from other equipment in the wafer fab.
Equipment designers have recently paid closer attention to exhaust,21, 22 which
also can affect defect levels significantly as well as resist thickness.
For a given track and resist dispense system, there is a dispense volume
which optimizes resist uniformity. 23 To minimize resist costs, this optimum
volume should be small for well-designed equipment. Typical dispense
volumes for coating 200 mm wafers range from 1 to 5 cc per wafer. This
volume is very sensitive to the surface material composition and preparation, as
well as the capability provided by the coating hardware. Resist spin speed and
dispense arm movement are other parameters which affect the across-wafer
resist thickness uniformity?' 25
4.2.3 Focus

Focus is controlled by the wafer stepper. The separation between the lens and
the wafer is measured prior to each exposure or during each scan, and the wafer
height is adjusted with the objective of maintaining a constant lens-to-wafer
distance. (On some older steppers it was the lens height which was changed.)
The tilt of the exposure area relative to the optical axis can be determined by
measuring the lens-to-wafer separation at several points on the wafer. There are
a number of factors that influence the degree of focus control, many of which

relate to the basic mechanism of the focus system. Three methods for focusing
wafers are used on steppers:

• Optical.
• Capacitance.
• Pressure.

The optical method, illustrated in Fig. 4.10(a) is the technique used most
commonly. In this approach, light is reflected from the substrate with a
glancing angle of incidence. The reflected light will hit the detector at different
positions, depending upon the vertical position of the substrate. In Fig. 4.10(a),
light which reflects off a wafer at the position indicated by the solid line will hit
the detector at a different position than light which reflects from a wafer at the
position indicated by the dashed line. Detectors that can distinguish the position
at which the light strikes the detector will be able to measure focus. The degree
to which the optical method is insensitive to substrate film type and topography
is critically dependent upon the system design. 26 For example, the optical
system may detect the most strongly reflecting surface—the position of the
resist film will be found at a different distance from the lens if it is sitting on
top of a thick oxide film versus a bare silicon substrate. Process engineers must
determine empirically if their combination of steppers and substrates creates
such a sensitivity. Metal and thick oxide films are the substrates most
commonly susceptible to focus errors. Minimizing the angle 0 and using
multiple wavelengths will provide the least sensitivity to films on the wafer
surface.27' 28 29
'

Two other methods of locating the wafer surface are shown in Fig. 4.10. In
one method pressurized air is forced into a block which contains a pressure
sensor. The pressure in the block will depend upon the gap between the block
and the wafer surface. By mounting such blocks on the sides of lenses, the
separations between lenses and wafers can be measured. This method is
sensitive only to the top surface of the resist film and is independent of
substrate composition. It does require calibration to the ambient barometric
pressure, and this type of focus sensor cannot measure the height of the wafer
surface directly below the lens during exposure, since the sensor would
interfere with the imaging.
Capacitance sensors are also used to measure the lens-to-wafer distance. A
capacitor is created where one plate is the wafer [Fig. 4.10(c)]. Capacitance C is
given by:

C= d (4.4)

where A is the area of the capacitor, d is the separation between the plates, and
E is the dielectric constant of the material between the plate. The separation
between the lens and wafer can be determined by a measurement of capacitance

-
I

. I)

1/)

rn
o (#
O

0
cn

ri
2
U

rI
2
n.

ho
/
/
II+I

I
I-

I
ri
I 01
hI C-)

Li
C
cl

because of its dependence on d. Like the optical method, this technique is

sensitive to the composition of the films on the silicon wafers.
All focus systems need calibration, and focus metrology has long been a
problem. For many years best focus has been determined by patterning focus
and exposure matrices. Inspection using optical microscopes was then used to
identify best focus. Various patterns were used, including minimum lines and
checkerboard patterns. 30 These methods suffer from inaccuracy and degrees of
subjectivity, causing variation among operators. For linewidths > 0.5 µm
optical linewidth measurement systems can be used to eliminate the subjective
element. 3 ' As linewidths have fallen below 0.25 gm, optical inspection and
linewidth measurement systems have become inadequate. Scanning electron
microscopes can be used, but these systems are typically very slow in
processing the large numbers of images required to determine focus accurately.
Two methods have been introduced which provide greater accuracy and speed
for measuring focus: phase-shifting focus and stepper self-metrology.
The phase-shifting focus monitor is based upon asymmetries in imaging
which occur when patterns with 90 ° phase shifters are defocused. 32'33 Consider
the pattern shown in Fig. 4.11. On the mask there is a chrome line, with 0°
phase shifting on one side and 90° phase shifting on the other side. The aerial
image of the line is shown in Fig. 4.12, at best focus, and out of focus for an
ideal lens. At best focus, the aerial image is symmetric, but when defocused,
the image becomes asymmetric. More importantly, the asymmetry moves in
opposite directions, depending upon whether the wafer is moved closer or
further from the lens. When patterning wafers, the resulting resist feature will
move as a consequence of this asymmetry. Equipment that is normally used for
measuring overlay can be used to measure defocus by the use of the structure
shown in Fig. 4.13, for which defocus will cause effective shifts in overlay.
This monitor has demonstrated the ability to measure focus to a precision of 50
nm.
The image shifts shown in Fig. 4.12 represent ideal optics. In practice,
centering of the structures in Fig. 4.13 does not always represent best focus, and
the calibration of the monitor has been found to be dependent on the numerical
aperture and partial coherence settings of the stepper. 34 Nevertheless, the phase-
shifting focus monitor can be calibrated, and thereafter provide a high precision
(50 nm) tool for measuring defocus. Once calibrated, this monitor can be used
to measure focus at many points across wafers, and the data can be collected
quickly using the tools normally used for automatically measuring overlay.
This capability has enabled focus problems to be identified, such as poor
chucking. 35
The determination of a stepper's best focus setting by processing wafers on
the stepper and measuring the wafers on scanning electron microscopes or
overlay measurement tools is time consuming and requires coordination of the
processing equipment and metrology tools. More efficient and operationally

Figure 4.11. A chrome line with different phase light transmitted on its two sides.

Figure 4.12. The light intensity profile of the line shown in Fig. 4.11, for different
defocus.

Figure 4.13. An overlay measurement structure which can be used to measure

defocus.

simpler is stepper self-metrology,36' 3" 3s 39 where steppers can determine their

own best focus setting. In some situations patterns are imaged into resist, 4°
while other self-metrology systems do not require wafers at all.
The glasses used for i -line lenses absorb small amounts of light. As a
consequence the lenses heat, resulting in shifts of focus. 41 Modern steppers will
automatically shift focus as wafers are exposed and the lenses are heated. The
amount of heating that takes place is a function of the quantity of light which is
transmitted through the reticle and into the lens, which is a function of pattern
density, types of features, illumination, and the numerical aperture of the
projection optics. Steppers often have sensors on the stage that can be used to
measure the amount of transmitted light directly. As steppers sit idle the lenses
will cool, and the stepper software must have the correct time constants to
account properly for lens cooling as well as heating.
Focus is also a function of barometric pressure.2' 42 Modern lenses are either
pressurized to maintain a constant pressure inside, or the stepper software is
used to adjust focus as barometric pressure varies. The software parameters
required tend to vary from lens to lens, even for lenses of the same design.
These parameters do not require changing, but software is always susceptible to
corruption, and focus errors can result from improper compensation for
barometric pressure, even for systems initially set up correctly. Good control of
steppers, and all other equipment, requires good records and back-up of
software parameters.
4.2.4 Bake temperatures

Much of resist processing involves thermal processes and chemical reactions.

Reaction rates are typically temperature dependent:
E
rate = A0 e kr , (4.5)

where k is Boltzmann's constant, T is the absolute temperature, Ea is an

activation energy, and A 0 is a rate constant. The specific requirements for
control depend upon the particular resist materials, which determine the rate
constants, but generally it is important to control the temperatures used for
processing photoresist.
Bake times also need to be controlled. Consider the dose-to-clear data
shown in Fig. 4.14. 43 A one second change in a 20 second bake changes E0 by
1.2%, and changes E0 by only 0.5% for a 60 second bake. The bake time should
be long enough so that the time over which the bake can be controlled does not
change the process significantly.
Effective bake times are determined by more than the length of time wafers
are on hotplates. Once wafers are removed from hotplates they will begin to
cool, and this cooling needs to occur in a controlled way. Consider the data

shown in Fig. 4.15, which show the temperatures of wafers as they are placed
on a hotplate, removed, and finally set on a chillplate." When the wafers are
placed at time = 0 on the hotplate their temperatures rise quickly to the hotplate
temperature of 105° C. After 60 secs on the hotplate the wafers are lifted above
the plate on pins, so that the wafers can then be picked up by a wafer transfer
arm and moved to a chill plate. The temperature of the wafers declines slowly
while the wafers are sitting on the pins. Because the wafer transfer arm may or
may not be occupied moving other wafers, the wafers may sit on the pins for
variable times. The change in critical dimensions was measured for a 10 second
variation in times on the pins and was found to be 7.4 nm when the relative
humidity was 42% and 10.8 nm when the relative humidity was 10.8% for a
conventional g-line resist. 44 Chillplates were introduced to semiconductor resist
processing in order to provide reproducible bake processes, 45 and this example
shows that even subtle aspects of the bakes need to be considered in order to
squeeze out the last few nanometers of process control.

EO versus PEB time for XP-9402

Acid-catalyzed DW photoresist
80

p 50

W 40
30
20 0
20 40 60 80 100
PEB time (secs)

Figure 4.14. Changes in dose-to-clear as a function of post-exposure bake (PEB)

time.
Bake profile
C' 2°
Long
l0U delay
I 80 Short
60 delay

40
20

-20 0 20 40 60 80 100
Time (secs)

Figure 4.15. Changes in effective baking as a consequence of variable times

between bake and chill.

The hotplate needs to be well designed in order to ensure good process

control. The hotplate receives a thermal shock when a wafer is first placed on it,
and the heating elements will respond. If the hotplate controller is not well
designed, there may be a temperature overshoot before the wafer temperature is
returned to the desired value. This overshoot can cause statistically measurable
shifts in linewidth 46 In order to have a well-controlled process, the wafer
temperature needs to be controlled across each wafer, wafer-to-wafer, and
during the entire heating and cooling cycle for each wafer.
In order to verify that bake temperature is well controlled, wafer
temperatures need to be measured. This is accomplished typically by attaching
temperature sensors to many points on a wafer and placing the wafer on a
hotplate. Temperature readings for each sensor can be sampled in time, and a
history of the wafer temperature can be collected. This method requires that
accurate temperature sensors be used. Today, these thermal probes are usually
platinum resistance temperature detectors (RTDs), with precision of ± 0.05 °C
and an absolute accuracy of ± 0.1 °C. 47 Use of wafers covered by temperature
sensors is not straightforward (because of the delicate wires which run between
the probes and the associated electronics), but can nevertheless be a very useful
diagnostic tool.

4.2.5 Resist development

Resist development also involves a chemical reaction. Parameters that are

expected to be critical are temperature and developer concentration. The effect
of developer temperature on linewidth control was studied extensively in the
early and mid-1980s. 48 During that time it became recognized that developer
temperature can affect linewidths significantly, and good temperature control
was implemented on the equipment used for developing wafers. Puddle develop
on a track system is the method used most commonly today. The tubes carrying
the developer for the wafer are jacketed, and temperature is controlled actively
to within ±0.2 °C. Once the developer is on the wafer it will begin to cool
because of evaporation. Because the rate of evaporation is greatest on the
outside of the wafer, thermal gradients are generated, leading to different rates
of development from the center to the edge of the wafer [tetramethyl
ammonium hydroxide (TMAH) developers become more aggressive at lower
temperatures]. Evaporation will also cause developer concentration to increase.
These effects can be minimized by using developer that is dispensed at a low
temperature to begin with and by having a low air flow around the wafer while
the resist is developing. 49
Developer composition also needs to be tightly controlled. The most
common resists used for semiconductor lithography use aqueous basic
developers. Early developers used sodium hydroxide or potassium hydroxide
for the base. Because sodium and potassium can lead to reliability failures in
MOS transistors, the most commonly used developer today is TMAH. To
improve wetting of resist surfaces, surfactants are often added to TMAH

developers, and buffers have been used to maintain constant developer pH of

metal-ion developers during the development process. The chemical activity of
multi-component developers is characterized by multiple inflection points 5o
(Fig. 4.16). For this reason, many lithographers prefer to use developers which
do not contain surfactants or buffers, since only a single component is then of
concern, and the composition is easier to control. Developer is produced in
batches, and the developer composition is very uniform within the batch. The
introduction of a new batch of developer is a special event, and changes in
linewidths can be correlated to transitions between batches. If no measurable
changes in linewidths are observed when developer batches are changed, then
the developer composition is being controlled adequately.
14
12
10

x
a g
6
4
2
0
0 10 20 30 40 50
Meq H+ added
Figure 4.16. Titration curve for developer.

4.2.6 Humidity

Water is a required reactant in the chemistry of napthaquinone diazide

photoresists, i.e., nearly all positive g- and i -line resists. It might therefore be
expected that linewidth control depends on the relative humidity. In one
experiment, linewidths were measured over a range of 30 — 60% relative
humidity. Changes in linewidth were observed, as a function of relative
humidity in the resist coater (Fig. 4.17), but after pursuing the problem further,
it was found that the dominant effect was a change in resist thickness 51 (Fig.
4.18). The linewidths changed because of thin film optical effects. 52 In another
experiment, the humidity was controlled in the exposure tool, and the resulting
level of photoacid was measured in several resists using Fourier transform
infrared spectroscopy. 53 A typical result from that work is shown in Fig. 4.19.
While water is required for the chemical reaction in napthaquinone diazide
photoresists, it appears that the photochemistry will be stable as long as the
relative humidity does not fall below 35%. In order to control resist thickness,
however, the humidity must be controlled. For DUV resists that do not require
water as part of their chemistries, this latter fact remains significant.

Resist A Resist B
Ethyl-3-ethoxyproprionate solvent Ethyl lactate solvent
0.56 22 0.51
0.55 .0 0.50

t::
0

i:_ — €_
30 35 40 45 50 55 60 0.47 30 35 40 45 50 55 60
Relative humidity (%) Relative humidity (%)

Figure 4.17. Changes in linewidth found as a fuction of humidity. The change in

linewidth is 1.1 nm per 1% humidity for Resist A and 0.4 nm per 1% humidity for
Resist B.

Resist A Resist B
^ Ethyl-3-ethoxyproprionate solvent Ethyl lactate solvent
920 980
910
970
900
890 960
880 950
870
30 35 40 45 50 55 60 30 35 40 45 50 55 60
Relative humidity (%) Relative humidity (%)

Figure 4.18. Resist thickness as a function of humidity. The thickness of Resist A

changes by 0.8 nm per 1% of humidity, and the thickness of Resist B changes 1.1
nm for every 1% change in relative humidity.

0.035

0.0345

o 0.034
03
0.0335
ta
0.033
X
0.0325

0.032'
I
10 20 30 40 50
Humidity (%)

Figure 4.19. Infrared absorbance of the acid produced by exposure of a

novolac/diazonapthaquinone resist.

4.2.7 DUV resists - special considerations

Acid catalyzed, chemically amplified resists are susceptible to poisoning by

part-per-billion levels of ambient amines 54,55 The photochemistry of these
resists involves the generation of an acid upon exposure to DUV light. Amines
that diffuse into the resist film can neutralize these photoacids and prevent
deprotection from taking place. The resist will not fully develop where this
occurs. One of the most common sources of amine contamination is the
ambient air, so the top of the resist is where the photoacid is first neutralized. A
typical result of this poisoning of the resist chemistry is shown in Fig. 4.20. For
slight poisoning the photoresist features are slightly flared at the top, but the
resist pattern can be completely bridged following longer exposures to ambient
amines. Contamination at lower levels may not have such dramatic
consequences, but there will be significant changes in the dimensions of critical
linewidths. 56

Figure 4.20. Apex E poisoned by ambient amines. The example on the left is
slight poisoning, while the example on the right is the result of longer exposure to
amines.

Special precautions need to be taken in order to prevent, or at least

minimize, the consequences of resist poisoning. The most common method of
reducing the level of ambient amines in the air around resist-covered wafers is
filtration. Activated charcoal is used to remove complex organic amines, such
as N-methyl-pyrilidone, while weak acids, such as citric acid, are used to
neutralize ammonia. The air in the steppers and resist process equipment is
often recirculated to maximize the removal of amines.
Substrates can also cause difficulties with chemically amplified resists,
including titanium nitride, silicon nitride, and various forms of silicon

dioxide. 57 '58,59 These substrates can cause footing at the bottom of resist
profiles. Processes should be set up initially to be reasonably free of such
footing, but resist poisoning by substrates is sensitive to deposition conditions.
If the deposition of substrate films changes, the lithography process may be
affected. Under such circumstances the linewidth monitors described in the
previous chapter are invaluable, providing independent tests that the
lithography process has remained stable. Sensitive monitors of the deposition
process can also prove useful. Substrates may also be treated to reduce the
extent of resist poisoning. 60
Chemically amplified resists are susceptible to poisoning by amines
between the time of exposure, when the photoacids are generated, and post-
exposure bake, which is when the deprotection reaction takes place. It is
essential to minimize this time between exposure and bake in order to keep the
effects of poisoning to a minimum. Moreover, it is important to maintain
consistent times between exposure and bake. This can be accomplished through
the design and programming of resist processing equipment. While controlling
the time between exposure and the post-exposure bake is of particular
importance for DUV resists, it should be noted that similar sensitivities have
been observed for novolac resists, 61 though as a consequence of a different
mechanism.
The newest generations of DUV resists have been designed to reduce this
poisoning effect. 62 Processes that use these resists are easier to control.
Sensitivity to amines is a significant parameter when choosing a DUV
photoresist, and this factor needs to be considered during resist evaluations.
4.2.8 Contributions from reticles

Reticles represent sources of variation in a manufacturing facility or pilot line,

because the average dimension and the pattern of variation changes from reticle
to reticle. The consequence of this was observed in a facility producing
application-specific integrated circuits, where there were multiple new products
introduced every month, and it was found that the reticles were the largest
source of linewidth variation [Ch. 3, Ref. 6]. As linewidths shrink, a point is
reached where the dimensions on the wafer change faster than on the reticle,
with appropriate scaling for lens reduction. Consider the data in Fig. 4.21 63 For
features larger than 300 nm the image is transferred from the reticle to the
wafer at a ratio of 4:1, the nominal reduction of the 0.5 NA Micrascan II lens.
Consequently, deviations in size for features on the reticle are reduced in
magnitude by a factor of four. For small features, this error reduction decreases
to a factor of only two. Processes that operate at or near the diffraction limit of
the optics will have smaller mask error reduction factors than those which are
far from the diffraction limit. For this reason, the nominal reduction factor of
4x for the lenses used for patterning the most critical layers is being
reconsidered by lithographers.

550

500

450

400

350

300
cd
s 250
vi

200
200 250 300 350 400 450 500 550
Dimension on the reticle (nm)

Figure 4.21. Wafer dimension versus reticle feature size, (scaled for the
4x reduction of the lens). The imaging system had a numerical aperture of 0.5 and
a wavelength of 248 nm.

4.3 MAXIMIZING THE PROCESS WINDOW

The best process is one that varies the least for the normal variations in the
equipment and materials that are used. The most effective techniques for
determining the optimum process are referred to as "design of experiments,"
which include factorial designs, 64 surface response techniques, 65 and Taguchi
methods. 66 Each of these techniques is legitimately a topic for an entire book
and is therefore beyond the scope of this text. The power of these methods lies
in the ability to efficiently identify sources of variation and process optimums
without a detailed understanding of the physical or chemical mechanisms
driving the variation.
The purpose of optimization in the context of process control is the
minimization of variation. As discussed in Chapter 1, many wafers are required
to determine the level of variation, which is not consistent with efficiency. It is
more effective to identify a parameter whose mean must be either maximized or
minimized. Consequently, in the context of linewidth control, most
optimization programs maximize the focus-exposure window. 67 There are a
number of reasons why this is an effective approach to minimizing linewidth
variation. First, focus is a critical parameter for controlling sub-micron
linewidths, particularly those smaller than 0.5 tm. The maximization of focus-
exposure windows therefore includes this critical parameter. Second,
maximizing exposure latitude addresses a number of problems. As discussed
earlier in this chapter, even though steppers control exposure doses reasonably
well, there are many phenomena, such as thin film optical effects, which
modulate the amount of actinic light coupled into photoresist films. Large

exposure latitude reduces sensitivity to these effects. Finally, focus-exposure

matrices can be patterned on single wafers, enhancing the efficiency of this
approach. Modern steppers have variable numerical aperture and illumination.
Process optimization must include the resist process as well as these stepper
parameters.
It has been observed by the author of this text and others68' 69 ' 7° that process
latitude can often be maximized when the dose to replicate mask features at
their nominal dimensions (E 1 _ 1 ) is much larger than the dose-to-clear (E 0 ):

E1-1 » 1, (4.6)
Eo

particularly on reflective substrates. (A different process optimum may occur

on substrates with low reflectivity.) For example, if the reticle features consist
of 250 nm lines and spaces (times the reduction factor of the lens), then E 1 _ 1 is
the exposure dose which produces 250 nm lines and spaces in the resist. This
dose is typically larger than the minimum dose required to clear a large exposed
area (> 1 tm on a side) of positive photoresist.
An overview of the physics and chemistry relevant to the formation of
patterns in photoresist was presented in this chapter. Such scientific
understanding is necessary for the efficient optimization and control of
linewidths. The technical aspects of overlay are the subjects of the next chapter.

K. Ishikawa, Guide to Quality Control, Asian Productivity Organization,

White Plains, New York (1991).

2H. J. Levinson and W. H. Arnold, "Focus: The Critical Parameter for Sub-
micron Lithography," J. Vac. Sci. Technol., B5(1), pp. 293 — 298 (1987).

3 IMEC is the acronym for Interuniversity Microelectronics Centre, Leuven,

Belgium.

4 K. Ronse, R. Pforr, L. van den Hove, and M. Op de Beeck, "CD Control: the
Limiting Factor for i -line and Deep-UV Lithography?" OCG Microelectronics
Seminar, pp. 241— 254 (1995).

K. Ronse, M. Op de Beeck, A. Yen, K-H. Kim, and L. van den Hove,

"Characterization and Optimization of CD Control for 0.25 um CMOS
Applications," SPIE Vol. 2726, pp. 555 — 563 (1996).

6 p. Schoenborn and N. F. Pasch, "Process Sensitivity Analysis: Applications to

Photolithography," SPIE Vol. 1087, pp. 290-298 (1989).

7 Z. Krivokapic, W. D. Heavlin, and D. Kyser, "Process Capabilities of Critical

Dimensions at Gate Mask," SPIE Vol. 2440, pp. 480 — 491 (1995).

8 Guide for Reducing Quality Costs, 2nd Edition. ASQ Press, Milwaukee
(1987).

9 Finle Technologies, Austin, Texas.

R. W. Leonhardt and T. R. Scott, "Deep-UV Excimer Laser Measurements at

NIST," SPIE Vol. 2439, pp. 448 — 459 (1995).

11K. H. Kim, W. S. Han, C. H. Kim, H. Y. Kang, C. G. Park, and Y. B. Koh,

"Characteristics of Standing Wave Effect of Off-axis Illumination Depending
on two Different Resist Systems and the Polarization Effect of Stepper," SPIE
Vol. 2197, pp. 42 — 53 (1994).

12 M. Bom and E. Wolf, Principles of Optics, 6th Edition, Pergamon Press,

New York (1990).

13
J Sturtevant and B. Roman, "Antireflection Strategies for Advanced
Photolithography," Microlithography World, pp. 13 —21(1995).
14
G. MacBeth, "Thermal Effects in Photoresist Coating Processes," KTI
Microelectronics Seminar, pp. 327 — 340 (1988).
15
S. Dick and B. Greenstein, "Improved Photolithography Process Performance
Through the Use of an Integrated Photosector," KTI Microelectronics Seminar,
pp. 1— 8 (1989).
16
M. Reihani, "Environmental Effects on Resist Thickness Uniformity,"
Semiconductor International, pp. 120 — 121 (June, 1992).
17
Batchelder, M. Ha, R. Haney, and W. Lee, "Sub-micron Linewidth
Control with Automatic Optical Monitors," KTI Microelectronics Seminar, pp.
231 — 242 (1989).

D. P. Birnie III, B. J. J. Zelinski, S. P. Marvel, S. M. Melpolder, and R. L.

Roncone, "Film/Substrate/Vacuum-Chuck Interactions During Spin-Coating,"

Optical Engineering, Vol. 31(9), pp. 2012-2020(1992).

19 M. Watts and S. Williams, "A Novel Method for the Prediction of Process
Sensitivity in Photolithography," SPIE Vol. 1261, pp. 345 — 359 (1990).

20 R. Murray, P. T. Edwin, V. Taburaza, and J. Olin, "Airflow Controller

Improves Photoresist Spin/Coat Uniformity," Semiconductor International, pp.
172 —173 (April, 1987).

21 L. Matter, J. Zook, M. Hinz, J. Banas, and S. Ibrani, "New Coat Bowl Design
Improves Photoresist Uniformity and Decreases Particle Contamination," Olin
Microlithography Seminar Poster Session (1997).
22
X. Zhu, F. Liang, A. Haji-Sheikh, and N. Ghariban, "A Computational and
Experimental Study of Spin Coater Air Flow," SPIE Vol. 3333, pp.
1441 — 1451 (1998).
23
W. J. Daughton, P. O'Hagan, and F. L. Givens, "Thickness Variance of
Spun-On Photoresist, Revisited," Proceedings of the Kodak Microelectronics
Seminar, pp. 15 — 20 (1978).
24
K. Kemp, D. Williams, J. Daggett, J. Cayton, S. Slonaker, and R. Elliott,
"Critical Dimension Performance Characterization of an Advanced DUV
Process Cell," Olin Microlithography Seminar, pp. 99 — 108 (1996).
25
D. Boutin, A. Blash, J. P. Caire, D. Poncet, P. Fanton, M. Danielou, and
B. Previtali, "Resist Coating Optimization on Eight Inches Deep UV Litho Cell
Modelization and Application to 0.25 pm Technology, SPIE Vol. 2439, pp. 495
— 502 (1995).
26
T. O. Herndon, C. E. Woodward, K. H. Konkle, and J. I. Raffel,
"Photocomposition and DSW Autofocus Correction for Wafer-Scale
Lithography," Proceedings of the Kodak Microelectronics Seminar, pp.
118 — 123 (1983).

27A. Suzuki, S. Yabu, and M. Ookubo, "Intelligent Optical System for a New
Stepper," SPIE Vol. 772, pp. 58 — 65 (1987).

28 J. E. van den Werf, "Optical Focus and Level Sensor for Wafer Steppers,"
J. Vac. Sci. Technol., Vol. 10(2), pp. 735 — 740 (1992).
29
M. A. van den Brink, J. M. D. Stoeldraijer, and H. F. D. Linders, "Overlay
and Field by Field Leveling in Wafer Steppers Using an Advanced Metrology
System," SPIE Vol. 1673, pp. 330 — 344 (1992).
30
J. W. Gemminck, "Simple and Calibratable Method for Determining Optimal
Focus," SPIE Vol. 1088, pp. 220 — 230 (1989).

31
S. Venkataram, C. Olejnik, G. Flores, and D. Tien, "An Automated
Technique for Optimizing Stepper Focus Control," SPIE Vol. 2725, pp.
765 — 778 (1996).
32
T. A. Brunner, A. L. Martin, R. M. Martino, C. P. Ausschnitt, T. H. Newman,
and M. S. Hibbs, "Quantitative Stepper Metrology Using the Focus Monitor
Test Mask," SPIE Vol. 2197, pp. 542 — 549 (1994).
33
R. D. Mih, A. Martin, T. Brunner, D. Long, and D. Brown, "Using the Focus
Monitor Test Mask to Characterize Lithographic Performance," SPIE Vol.
2440, pp. 657 — 666 (1995).
3a
E. R. Sherman and C. Harker, "Characterization and Monitoring of Variable
NA and Variable Coherence Capable Photo Steppers Utilizing the Phase Shift
Focus Monitor Reticle," SPIE Vol. 2439, pp. 61— 69 (1995).
3s
T. A. Brunner and R. D. Mih, "Simulations and Experiments with the Phase
Shift Focus Monitor," SPIE Vol. 2726, pp. 236 — 243 (1996).
36
T. A. Brunner and S. M. Stuber, "Characterization and Setup Techniques for
a 5X Stepper," SPIE Vol. 633, pp. 106 — 112 (1986).

37M. van den Brink, H. Franken, S. Wittekoek, and T. Fahner, "Automatic On-
Line Wafer Stepper Calibration System," SPIE Vol. 1261, pp. 298 — 314
(1990).

38T. A. Brunner, S. Cheng, and A. E. Norton, "Stepper Image Monitor for

Precise Setup and Characterization," SPIE Vol. 922, pp. 366 — 375 (1988).
39
R. Pforr, S. Wittekoek, R. van den Bosch, L. van den Hove, R. Jonckheere,
T. Fahner, and R. Seltmann, "In-process Image Detecting Technique for
Determination of Overlay and Image Quality for the ASM-L Wafer Stepper,"
SPIE Vol. 1674, pp. 594 — 609 (1992).

40P. Dirksen, W. de Laat, and H. Megens, "Latent Image Metrology for

Production Wafer Steppers," SPIE Vol. 2440, pp. 701 — 711 (1995).

T. A. Brunner, J. M. Lewis, and M. P. Manny, "Stepper Self-Metrology

41
Using Automated Techniques," SPIE Vol. 1261, pp. 286 — 297 (1990).
42
K. Hale and P. Luehrman, "Consistent Image Quality in a High Performance
Stepper Environment," Proceedings of the Kodak Microelectronics Seminar,
pp. 29 — 46 (1986).

a3.
J. S. Petersen, C. A. Mack, J. W. Thackeray, R. Sinta, T. H. Fedynyshyn,
J. J. Mori, J. D. Byers, and D. A. Miller, "Characterization and Modeling of a
Positive Acting Chemically Amplified Resist," SPIE Vol. 2438, pp. 153 — 166
(1995).

J. M. Kulp, "CD Shift Resulting from Handling Time Variation in the Track
44
Coat Process," SPIE Vol. 1466, pp. 630 — 640 (1991).
as
G. MacBeth, "Prebaking Positive Photoresists," Proceedings of the Kodak
Microelectronics Seminar, pp. 87 — 92 (1982).
46
O. D. Crisalle, C. L. Bickerstaff, D. E. Seborg, and D. A. Mellichamp,
"Improvements in Photolithography Performance by Controlled Baking," SPIE
Vol. 921, pp. 317-325 (1988).

47 J. Parker and W. Renken, "Temperature Metrology for CD Control in DUV

Lithography," Semicond. International, pp. 111— 116, (September 1997).

48C. M. Garza, C. R. Szmanda, and R. L Fischer, "Resist Dissolution Kinetics

and Submicron Process Control," SPIE Vol. 920, pp 321— 338 (1988).
a9
M. K. Templeton, J. B. Wickman, and R. L. Fischer, Jr., "Submicron
Resolution Automated Track Development Processes, Part 1: Static Puddle
Development," SPIE Vol. 921 pp. 360 — 372 (1988).
50
W D. Hinsberg and M. L. Gutierrez, "Effect of Developer Composition on
Photoresist Performance," Proceedings of the Kodak Microelectronics Seminar,
pp. 52 — 56 (1983).

51E. Bokelberg and W. Venet, "Effects of Relative-humidity Variation on

Photoresist Processing," SPIE Vol. 2438, pp. 747 — 752 (1995).
52
W. H. Arnold and H. J. Levinson, "High Resolution Optical Lithography
Using an Optimized Single Layer Photoresist Process," Proceedings of the
Kodak Microelectronics Seminar, pp. 80-92 (1983).
53
J. A. Bruce, S. R. DuPuis, and H. Linde, "Effect of Humidity on Photoresist
Performance," Proceedings of the OCG Microlithography Seminar, pp. 25 — 41
(1995).
s4
S. A. MacDonald, C. G. Wilson, and J. M. J. Frêchet, "Chemical
Amplification in High-Resolution Imaging Systems," Accounts of Chemical
Research, Vol. 27(6), pp. 151— 158 (1994).

55
K. R. Dean, D. A. Miller, R. A. Carpio, and J. S. Petersen, "Airborne
Contamination of DUV Photoresists: Determining the New Limits of
Processing Capability," Proceedings of the Olin Microlithography Seminar, pp.
109— 125 (1996).
56
K. van Ingen Schenau, M. Reukman, and S. Slonaker, "Investigation of DUV
Process Variables Impacting Sub-Quarter Micron Imaging," Proceedings of the
Olin Microlithography Seminar, pp. 63 — 80 (1996).

N. Thane and G. A. Barnes, "Process Compatibility of Anti-reflection

57
Coatings with Various Deep UV Resists," Proceedings of the OCG
Microlithography Seminar, pp. 61— 76 (1993).

58J. S. Sturtevant, S. Holmes, S. Knight, D. Poley, P. Rabidoux, L. Somerville,

T. McDevitt, A. Stamper, E. Valentine, W. Conley, A. Katnani, and J. Fahey,
"Substrate Contamination Effects in the Processing of Chemically Amplified
DUV Photoresists," SPIE Vol. 2197, pp. 770 — 780 (1994).
59
K. R. Dean, R. A. Carpio, and G. K. Rich, "Investigation of Deep Ultraviolet
Photoresists on TiN Substrates," SPIE Vol. 2438, pp. 514-528 (1995).
60
A. Usujima, K. Tago, A. Oikawa, and K. Nakagawa, "Effects of Substrate
Treatment in Positive Chemically-Amplified Resist," SPIE Vol. 2438, pp.
529 — 539 (1995).

61 Y-K. Hsiao, C-H. Lee, S-L. Pan, K-L. Lu, and J-C. Yang, "A Study of the
Relation Between Photoresist Thermal Property and Wettability," Proceedings
of the Olin Microlithography Seminar, pp. 315 — 323 (1996).
62
H. Ito, W. P. England, H. J. Clecak, G. Breta, H. Lee, D. Y. Yoon,
R. Sooriyakumaran and W. D. Hinsberg, "Molecular Design for Stabilization of
Chemical Amplification Resist Toward Airborne Contamination," SPIE Vol.
1925, pp. 65 — 75 (1993).
63
W Maurer, K. Satoh, D. Samuels, and T. Fischer, "Pattern Transfer at k l =
0.5: Get 0.25 pm Lithography Ready for Manufacturing," SPIE Vol. 2726, pp.
113— 124 (1996).
64
G. E. P. Box, W. G. Hunter, and J. S. Hunter, Statistics for Experimenters,
John Wiley & Sons, New York (1978).
65
G. E. P. Box and N. R. Draper, Empirical Model-building and Response
Surfaces, John Wiley & Sons, New York (1987).

66
G. Taguchi, System of Experimental Design, Vols. 1 and 2, American
Supplier Institute Press, Dearborn, Michigan (1987).

67For example, K. Petrillo, "Process Optimization of Apex-E," SPIE Vol. 1926,

pp. 176 — 187 (1993).

68S. G. Hansen, G. Dao, H. Gaw, Q-D. Qian, P. Spragg, and R. J. Hurditch,

"Study of the Relationship Between Exposure Margin and Photolithographic
Process Latitude and Mask Linearity," SPIE Vol. 1463, pp. 230 — 244 (1991).
69
J. S. Petersen and A. E. Kozlowski, "Optical Performance and Process
Characterization of Several High Contrast Metal-Ion Free Developer
Processes," SPIE Vol. 469, pp. 46 — 56 (1984).

70J. S. Petersen, "An Experimental Determination of the Optical Lithographic

Requirements for Sub-Micron Projection Printing," SPIE Vol. 188, pp.
540-567 (1989).

5.1 OVERLAY MODELS

A discussion of overlay should begin with definitions. Overlay is defined in

SEMI Standard P18-92, "Specifications for Overlay Capabilities of Wafer
Steppers," as':
Overlay - A vector quantity defined at every point on the wafer. It is the
difference, Ö, between the vector position, Pl , of a substrate geometry
and the vector position of the corresponding point, P Z in an overlaying ,

pattern, which may consist of photoresist:

O = PZ — P1 . ( 5.1)

A related quantity, registration, is defined similarly:

Registration - A vector quantity defined at every point on the wafer. It
is the difference, k, between the vector position, Pl , of a substrate
geometry and the vector position of the corresponding point, i, in a
reference grid:

R = P1 — Po . ( 5.2)

Note that overlay is a relative quantity, while registration is an error compared

to an absolute standard, fin .
Overlay errors can be considered in terms of a hierarchy (Fig. 5.1). Most
fundamental are those errors which occur when only a single stepper and ideal
substrates are used, and where the latter provide high signal-to-noise alignment
signals. This basic set of overlay errors is well described by overlay models,
which will be discussed in detail shortly. When more than a single stepper is
used, an additional set of overlay errors is introduced, referred to as matching
errors. Finally, there are process-specific contributions to overlay that can result
in non-ideal alignment targets. A full accounting of overlay errors in terms of a
hierarchy is useful for assessing and improving overlay. 2
The control of overlay requires the use of models, because these models are
employed directly in the software of wafer steppers when aligning wafers, and
the model parameters are changed when adjustments are made to wafer
steppers. Overlay models conform to the physics of wafer steppers. There are
two categories of overlay errors: intrafield, which involve the variation of

overlay errors within exposure fields, and interfield errors, which are the errors
which vary from field to field across wafers.

Process dependent
effects

Matching errors: factors which occur

only when multiple steppers are used.

Fundamental overlay errors: stage precision, wafer

non-flatness effects, and factors incorporated into the
overlay models.

Figure 5.1. The hierarchy of overlay errors. The foundation includes those
factors common to all situations.

Interfield errors can be understood by considering the overlay error at

equivalent points in each exposure field. Conventionally this is the center of the
exposure field. Let (X,Y) be the coordinate of a point on the wafer
corresponding to the center of some exposure field. The center of the wafer is
the most convenient location for the origin of the coordinate system. The
overlay error at the point (X, Y) is given by:

AX = ff (X,Y) (5.3)

AY= ff (X,Y) (5.4)

where AX is the overlay error in the X direction and AY is the error in the Y
direction. The overlay error AX can be expressed in a MacLaurin series of the
function fx :

AX = Tx +Ex X —R x Y+ higher order terms + e X . ( 5.5)

The reason for the negative sign prior to the R x Y term will be discussed
shortly. There is a similar expression for AY:

AY = Ty +EY Y+RY X + higher order terms + e. (5.6)

In addition to the higher order terms there are residual errors, ex and e y . , at
every measured point which do not conform to the model. The most common
source of residual error is the imprecision of the stepper's stage, which is on the

order of 15 nm (36) for the current generation of wafer steppers. 3 Another

source of non-modeled error arises when multiple steppers are used. This is the
issue of matching, which will be discussed in the next section.
The linear terms in Eqs. 5.5 and 5.6 have physical meanings. Each layer
can be considered as patterns printed on a rectangular, or near-rectangular, grid
(Fig. 5.2).

Figure 5.2. Interheld overlay errors.

The parameters Tx and Ty represent translation errors in the X and Y

directions, respectively, and indicate an overall shift of one grid relative to the
other. The factors E X and E y are scale errors, which represent the errors made
by the stepper in compensating for wafer expansion or contraction. Scale errors
are dimensionless and are usually expressed in parts-per-million (ppm). The
coefficients RX and R y are rotation factors. When R X = Ry one grid is rotated
relative to the other, and this accounts for the sign convention chosen in Eq.
5.5. If the angle between the axes of the grids are not equal, then R X ^ R.
This latter error is referred to as an orthogonality error, since it represents the
situation in which the grid for at least one of the layers (substrate or overlaying)
has non-orthogonal axes. The rotation prefactors are tangents of angles, but
since

0 = tan 0 (5.7)

for small angles 0, the rotations are usually expressed in radian measure,
typically in units of microradians.
The type of alignment used with most wafer steppers is some form of
global or enhanced global alignment, in which the positions in space of at least
two points on the wafer are determined. When only two points are determined
by the alignment process it is possible to compute four of the parameters in
Eqs. 5.5 and 5.6. For such two point global alignment it is usually assumed that
EX = Ey and RX = Ry . By aligning the wafer at three or more (non-colinear)
points on the wafer it is possible to estimate all of the linear terms in Eqs. 5.5
and 5.6. The overlaying pattern is adjusted according to these linear corrections,
through the software of the stepper. This model is used directly in the stepper
software and is relatively unchanged from the original interfield model
introduced by Perloff. 4
Some overlay errors cannot be detected during alignment. For example, it
has been observed that wafers heat during i-line exposures, causing wafer
scaling errors that arise after the wafers have been aligned. 5 Non-linear errors
will not be corrected by the software of most steppers, which use a linear
model. Some errors which are not included in the model of Eqs. 5.5 and 5.6
will be discussed in the next section.
The overlay errors may also vary across each exposure field. These types of
errors are referred to as intrafield errors, examples of which are shown in Fig.
5.3. Consider, for example, a magnification error. The lens reduction typically
has nominal values of 4:1 or 5:1, but the magnification will deviate from these
nominal values by some small amount. When there is wafer expansion and
contraction between masking steps, it is necessary for the stepper to be
programmed to measure this change and compensate for it, not only in the grid
terms, but in the size of each field as well. Errors in doing this correction result

in magnification errors. The other primary intrafield errors are shown in

Fig. 5.3.

Figure 5.3. Intrafield overlay errors.

Intrafield overlay models for wafer steppers were introduced by MacMillan

and Ryden in 1982. Since then, the most significant change has been the
introduction of models for step-and-scan systems. The models for step-and-scan
systems include parameters for asymmetric magnification and skew, which are
not relevant for step-and-repeat systems. Suppose the same stepper is used for
printing two layers, one overlaying the other. Let (x, y) be the coordinates of a
point on the wafer, relative to the center of the exposure field in which it is
contained. For a step-and-repeat system the intrafield overlay errors are
modeled:

Sx=mx —ry+T xy+Ty x 2 +ex(5.8)

8y=my+rx +T xy +T y 2 +e y( 5.9)

In these equations, the parameter m represents a magnification error. It is

dimensionless and is usually expressed in parts-per-million. The magnification
error can be interpreted as follows. The error in overlay increases linearly with
distance from the center of the exposure field. The rate at which this error
increases with distance from the center of the exposure field is the

magnification error coefficient m. For step-and-repeat systems the

magnification error is the same in x and y directions. The coefficient r
represents reticle rotation. As with wafer rotation, this is usually expressed in
microradians. Reticle rotation error also increases linearly with the distance
from the center of the exposure field, and the error coefficient is the same for
errors in x and y. The factors T and T are trapezoidal errors. These types of
overlay errors occur on lenses that are not telecentric on the reticle side. The
terms ex and e are residual errors that represent the intrafield overlay that does
not conform to the model.
Interfield models are identical for step-and-repeat and step-and-scan
systems, but the appropriate intrafield models differ. For step-and-scan systems,
the intrafield model is:

Sx=mXx—rxy +ex (5.10)

Sy=m y y+ Yx +e y ( 5.11)

The step-and-scan model is similar to the step-and-repeat model for doubly

telecentric lenses, with some differences. Most notably, the magnification
coefficents (mx and m y ) and rotation coefficients (x and ry ) in Eqs. 5.10 and
5.11 are independent in the two equations, and not coupled, as in Eqs. 5.8 and
5.9. The magnification for step-and-scan systems can be different in the x and y
directions, because the magnification in the x direction has a different origin
than the magnification in the y direction. The magnification in the direction of
the scan is determined by the relative speeds of the reticle and wafer stages.
Ideally, the reticle stage should move N times faster than the stage, where N: 1 is
the reduction ratio of the lens (typically, N = 4). When the reticle stage speed
moves faster than N times the wafer stage speed, the printed field will be
shorter in the scan direction. The magnification in the direction perpendicular to
the scan is determined only by the lens reduction factor (N: 1). For step-and-
scan systems, pure reticle rotation occurs when the scans of the reticle and
wafer stages are parallel, but the reticle is rotated relative to both. Field skew
results when the reticle is rotated properly with respect to the wafer, but the
reticle stage is rotated.
For any point on the wafer the overlay error is the sum of the interfield and
intrafield errors. In the X direction, the total overlay error is O X x = &X+Sx, and
the overlay error in the Y direction is OY y = AY+Sy, at the point (X + x, Y + y).
For step-and-scan systems:

°x,x= TX+EX X—RXY +mxx — xy +pXx (5.12)

OY y =TY +EY Y+RY X+m y y+ry x+p Y y(5.13)

where p X x and p Y y are total residual errors. There are similar equations for
step-and-repeat systems.
Classifying errors is important for controlling overlay, because intrafield
and interfield overlay errors generally arise from different causes. For example,
reticle rotation errors involve the alignment of the reticle, while wafer rotation
issues involve water alignments. In order to classify overlay errors the
coefficients in Eqs. 5.12 and 5.13 need to be determined. The most common
method for extracting these coefficients from measured data is the method of
least-squares. 6'' In this technique, more data are needed than there are
parameters. For Eqs. 5.12 and 5.13, measurements need to be made at two
points per exposure field, or more. (Each measurement site provides two
measurements, one in X and one in Y.) The least squares method determines the
set of coefficients {TX ,T,,m x ,...} which minimizes the sum of squares of the
residual errors:

(5.14)
X,x,Y,y

where the sum is extended over all measured points. Computational speed and
insensitivity to noise are among the reasons why the least-squares method is the
method most commonly used. The choice of mathematical methods used to
determine the model coefficients will be discussed further later in this chapter.
In Section 3.3 it was noted that the trending of process parameters was
useful in circumstances in which test wafers were required. For overlay, these
process parameters are the coefficients in the overlay model. By examining the
variation in these coefficients, the sources of variation can often be identified.
For example, large excursions might be correlated to the stepper used for a
particular prior layer.

Example: An engineer wanted to compensate for reticle registration errors that

had been measured on a product reticle. Errors on the reticle that fit the model
of Eqs. 5.10 and 5.11 could be corrected on step-and-scan systems. Constants
were added to each of these equations as well, to correct for translation shifts.
The reticle errors, as measured on the product reticle, are given in Table 5.1.
The model coefficients are given in Table 5.2. As can be seen, a substantial
fraction of the reticle error was eliminated by applying magnification, rotation,
and translation offsets. For example, the registration error at the position (-69.5,
0) is given by:

bx = X^anslation +m x x — z y + e x(5.15)

=-10.6+0.963x69.5-0.329x0.0+e x(5.16)

= 56 . 3 +e (5.17)

The measured error was 63 nm, so the model was able to account for most of
the registration error at that point, and most of the error was therefore
correctable.

Measurement Measured Residual registration

positions registration errors errors
Step-and-scan model
x (mm) y (mm) 6x (nm) Sy (nm) e., (nm) ey (nm)
-69.5 0.0 63.0 0.0 6.7 -9.7
69.5 0.0 -63.0 0.0 14.5 -12.9
-30.2 51.0 17.0 -15.0 15.3 -2.2
30.2 51.0 -14.0 1.0 42.4 12.4
30.2 -51.0 -21.0 30.0 1.9 -5.4
-30.2 -51.0 48.0 29.0 12.8 -5.0
-29.9 50.8 -41.0 -16.0 -42.5 -3.3
29.9 50.8 -82.0 -7.0 -25.9 4.3
29.9 -50.8 -61.0 39.0 -38.3 3.7
-29.9 -50.8 48.0 52.0 13.1 18.1

Table 5.1. Reticle registration errors.

The first two columns are the positions where registration was measured. The
third and fourth columns are the registrations errors, as measured, and the last
columns are the residual errors, after the modeled part of the errors was
subtracted from the raw data.

x-translation -10.6 nm -translation 11.3 nm

m -0.963 ppm m 0.023 ppm
x y
x 0.329 µrads ry -0.459 µrads

Table 5.2. The model coefficients, determined by least squares, for the reticle mis -
registration data in Table 5.1.

5.2 MATCHING

A set of non-random overlay errors that are unaccounted for in the overlay
models are found in situations where more than one stepper is used. These
errors are referred to as matching errors, and there can be grid and intrafield
matching errors. Grid matching errors arise from absolute stepping errors.
While stepper stages are extremely repeatable, their stepping may deviate at
each stage position, on average, from a perfect grid. When a single stepper is

used for both the substrate and overlaying layers, these average deviations
cancel out. A different situation arises when different steppers are used for the
two layers.
Grid registration errors originate in the non-flatness of the stage mirrors.
Consider the stage mirror drawn in Fig. 5.4. The branch of the interferometer
that measures the stage's position in the Y-direction will measure apparent
movement in the Y-direction as the stage is moved from left to right, because of
the non-flatness of the mirror. When a single stepper is used, the resulting
stepping errors, relative to a perfect grid, usually do not lead to overlay errors,
because these stage stepping errors occur for all layers and cancel out. In
general, the mirror non-flatness varies from stepper to stepper, and stage mis-
match occurs. For modern wafer steppers the stage mirrors are flat to
approximately X/40, where X = 632.8 nm is the wavelength of light of the
stage's interferometer, or about 16 nm. In addition to not being flat, the mirrors
used for the interferometry in the X and Y directions will not be exactly
orthogonal, with errors on the order of one arc second.

Stage mirrors

Laser —►

Laser

Figure 5.4. Non-flat stage mirrors, which lead to matching errors.

Grid registration errors can be corrected using software. Typically,

reference matching wafers are used for matching all steppers in a facility to the
same positions. Look-up tables can be used for these corrections, or the grid
models can be expanded to higher orders. The second order interfield term is
wafer bow. 8 Matching grids to less than 5 nm is not uncommon. Laser
interferometers for wafer steppers usually have more than one beam for each
direction, in order to measure stage rotation during stepping. Mirror non-
flatness can then lead to intrafield rotation as fields are stepped across wafers, 9
which will not be captured by models which treat the reticle rotation the same
in every field.
Lenses do not always place geometries where they belong. For example,
suppose the pattern on the reticle consisted of a perfect rectangular grid (Fig.
5.5). Because of the lens placement errors the vertices of the pattern will be

slightly moved. Placement errors run from 200 nm on older lenses to < 20 nm
on contemporary machines. These errors may be inherent in the design or result
from fabrication imperfections. The design contribution is referred to as lens
distortion, and has a known functional form:

Or = D3 r 3 +D5 r 5(5.18)

dx = D3 xr 3 + D5 xr 4(5.19)

Ay = D3 yr 2 + DS yr 4(5.20)

where r is the radial distance from the center of the lens field:

r= x 2 +y 2 . ( 5.21)

Because of the magnitude of the exponents of terms in Eq. 5.18, D3 and D5 are
referred to as third and fifth order distortion, respectively.

nMnnnnnn s1ZWIRS
MEENnnnn
nnnnnnnn

Pattern on the wafer,

distorted as a consequence
Perfect pattern on the reticle. of lens placement errors

Figure 5.5. Pattern distortion as a consequence of lens placement errors. The

movement of the vertices is exaggerated in this figure for the purpose of
illustration.

Distortion tends to be fairly constant within a particular lens family.

However, there are strong economic reasons to minimize the numbers of very
high performance steppers, and to mix-and-match steppers of different types.'°

Consequently, lenses of more than one type will usually be used to fabricate
semiconductor devices. Older generations of projection optics had non-zero
levels of distortion in the design, where the magnitude of Ox and Ay due to
distortion alone could be as large as several tens of nanometers, up to 200 nm.
In such situations, good matching could be achieved only within given lens
families, particularly when different lens designs might have distortions with
opposite signs for D3 and D5 . The designs of modern lenses have negligible
levels of radial third and fifth order distortion. For these newer lenses,
placement errors are due primarily to manufacturing imperfections, which
produce intrafield errors that are not systematic, as in the case of third and fifth
order lens distortion. Polishing variations across the surfaces of lenses and
mirrors and inhomogeneity in optical materials can lead to intrafield
registration errors, and these will tend to have random patterns.
In the presence of significant radial lens distortion it is impossible to
distinguish between third and fifth order distortion and simple magnification
errors, unless a sufficient number of points are sampled within the exposure
field. In-line sampling plans usually consist of measurements at the four corners
of the exposure field, perhaps with measurements also near the center of the
field, and this is insufficient for distinguishing between magnification and
higher order radial distortion. The problem of small samples for in-line process
control will be considered shortly in more detail.
While distortion is extremely stable for i-line systems, distortion varies
with very small shifts — less than a picometer — in wavelength for DUV
systems.' 1, 12 Consequently, additional controls are needed on DUV machines,
the most critical being wavelength control. Modern KrF excimer lasers, the
light source for 248 nm exposure systems, contain internal wavelength
references that provide good wavelength control. 13 As with anything, this
calibration system can malfunction or drift, and there is always the potential for
changes in third and fifth order distortion as a consequence of a small shift in
wavelength. Also, changes in barometric pressure will change the wavelength
of light in the air, and excimer steppers must adjust for this correctly, or third
and fifth order distortion may be introduced.
For step-and-scan systems the intrafield placement errors will also have a
scanning component. There will be random and systematic contributions to the
scanning errors. This random component does not exist in step-and-repeat
systems, so there is more potential for overlay errors with scanners, even for a
stepper-to-itself. Moreover, the stage which scans the reticle can malfunction,
another control problem unique to step-and-scan systems.
The intrafield overlay will contain some theoretically correctable
contributions which can be fit to the overlay model (Eqs. 5.8 - 5.11), and other
components that arise from differences between lenses which do not conform to
the model. The matching of the lenses is the set of residual overlay vectors
which remains after subtracting out the correctable contributions. These sets of
vectors are dependent upon the criterion used for determining the model
parameters. The least squares method is the one used most commonly, but it is

not the only criterion that is reasonable to use. For example, since devices are
usually designed to yield so long as overlay is less than a particular value and
fail when this value is exceeded, yield can be maximized when the worst
overlay errors are minimized. 14 This criterion consists of the minimization of:

I,
m,x(le x le y I ), (5.22)

where e x and e y are the residual errors in Eqs. 5.8 — 5.11. This differs from the
least-squares criterion, which consists of the minimization of the quantity in
Eq. 5.14 and will typically lead to different model coefficients. We say that
lenses are matched when correctable parameters have been set to minimize the
selected criterion, either Eq. 5.14, 5.22, or another that may have been chosen.
The least-squares method has two key advantages. The first is
computational simplicity, since model coefficients can be found through simple
matrix calculations. The least-squares method also provides transitivity, 15 that
is, if Stepper A is matched to Stepper B, and Stepper B is matched to Stepper
C, then Stepper A will be matched to Stepper C. Minimization of maximum
error does not share this characteristic. 16
When only a single stepper is used, the intrafield overlay model accounts
for nearly all of the intrafield overlay errors, and the overlay errors throughout
the entire field can therefore be inferred from measurements at only a few
points with the exposure fields. It is quite a different situation when more than
one stepper is used. Matching methodology must account for overlay in the
areas of the exposure field that are not accessible when measuring product
wafers. Specialized structures are normally used for measuring overlay, and
these are usually placed in the scribe lanes between product dies. This limits the
amount of overlay data collected on product wafers, since most of the exposure
field is occupied by product devices, not overlay measurement structures. This
limitation on the number of points within each exposure field at which overlay
can be sampled has significant consequences, since overlay is measured in
areas where the product is not placed.
Suppose overlay is measured on several fields on the wafer, and within
each field at four sites with the same intrafield positions. One can fit the
acquired data to the overlay model (Eqs. 5.12 and 5.13) by minimizing the
appropriate metric (least squares, minimum error, etc.). This approach may not
optimize the overlay over the entire exposure field, particularly the parts of the
field in which the product is located and overlay is not measured. This was
demonstrated in the following way." Between two steppers the overlay was
measured on a 12 x 11 grid with 1.95 mm spacings in the x and y directions
within the exposure field. The resulting lens matching is shown in Fig. 5.6. In a
gedanken ( thought) experiment the overlay measurements were considered at
only four points among these 132 sites, in a pattern that resembled a typical
overlay measurement plan. Several subsets of four points were considered

(Fig. 5.7). For each set of four sites the overlay model coefficients were re-
calculated and plotted in Fig. 5.8. As one can see, the resulting model
coefficients varied significantly among sampling plans. Recall that the baseline
set of coefficients was the one that optimized the overlay over the entire
exposure field, not just at the four corner points. By measuring at only four
points and adjusting the stepper to minimize overlay at just the measured
points, overlay is moved away from the optimum, overall. This is a problem
that occurs when different steppers are used for the overlaying patterns.
In earlier discussions it was mentioned that the production of good devices
requires that critical parameters remain within specifications at all locations
within the dies. Lens mismatch may result in particular points at which overlay
is very poor. Overall overlay may be very good, but the die at the field location
where the mismatch occurs will yield poorly. An example of this is shown in
Fig. 5.6, where the upper left corner of the field has a point with particularly
bad overlay. Overall, the overlay is good, but the die located in that corner will
consistently have poor overlay. With 1 — 8 die per field being typical, a single
bad site within the exposure field can degrade the yield significantly. The
concept that yield requires all points within a die to be within specifications
was first introduced in the context of overlay.' 4
To complicate matters further, it has been observed that lens placement
errors vary with illumination conditions, such as partial coherence' 8,19,20 or off-
axis illumination. 2' 22 For purposes of maximizing depth of focus or exposure
'

latitude, different illumination conditions are often used at different layers. For
example, large values for the partial coherence are used for gate and metal
layers, at least when binary masks are used. Small values of partial coherence
are used for contact layers and phase shifting masks. Consequently, there will
be intrafield errors, even when the same stepper is used for all layers, if
different illumination conditions are used for different layers. The overlay
errors between a first layer printed with standard illumination (NA = 0.63, a =
0.65) and a second layer exposed using the same numerical aperture, but a
partial coherence = a = 0.3, are shown in Fig. 5.9. 20 Even though both layers
were printed using the same lens, there is considerable intrafield error. Each
vector is the overlay error at the point located at the tail of the vector. It should
also be noted that the use of a single stepper does not guarantee the non-
existence of grid and intrafield matching errors, because drift, database
corruption, and stepper malfunction are possible. However, such events are
rare, and the matching errors discussed here occur infrequently when single
steppers are used with fixed operating parameters.

^
,, y rA 4'

1 i ^t
,
/` ''

^ 4;.
r ßv. •
I .
C F -*. _ R, -,\ \ R'\ 4

80nm

Figure 5.6. The lens matching between a step-and-scan system and a step-and-
repeat system. The vectors are the residual errors after subtracting correctable
errors, at the points located at the tails of each vector.

1 3 5 8 10 139 8 5 3 1
2 6 12 15 18 18 15 12 6 2
4 11 17 20 22 22 20 17 11 4
7 14 19 23 25 25 23 19 14 7
9 16 21 24 26 2624 2116 10
13 13
10 16 21 24 26 26 24 21 16 9
7 14 19 23 25 25 23 19 14 7
4 1117 20 22 22 20 17 11 4
2 6 12 15 18 18 15 12 6 2
1 3 5 8 9 13 10 8 5 3 1

Figure 5.7. Sampling of the measurement site of the data shown in Fig. 5.6. For
example, the four corner sites were chosen for Sampling Plan 1.

X translation Y translation
20 30
15
10 20
ó 5 10
0
-5
10
1 ° -10
X -15
20 20
0 5 10 15 20 25 0 5 10 15 20 25
Sampling subset number Sampling subset number

Asymmetric magnification Field isotropic magnification

8 8
6 6
4 4

á 0, 2
0
-2 -2
-4 -4

6 0 5 10 15 20 25 0 5 10 15 20 25
Sampling subset number Sampling subset number

Reticle rotation Field skew

8 8
10 46 4
14
'o 2 2
c0
0
1 0
-2 -2
oG -4
6
"O 5 10 15 20 25 0 5 10 15 20 25
Sampling subset number Sampling subset number

Figure 5.8. Variations in model parameters for the sampling plans shown in
Fig. 5.7.

5 :j«

0
50 nm

-10

-15
-15 -10 -5 0 5 10 15
Field location (mm)

Figure 5.9. Intrafield overlay errors, where the first level was printed on a stepper
with NA =0.63 and a = 0.65, and the second level was printed with the same NA
and a = 0.3.

The changes in lens placement errors can be understood by considering the

imaging of a simple diffraction grating, discussed in Chapter 4. Depending on
the angle of incidence of the illumination, numerical aperture and the pitch of
the pattern, the light of a grating pattern projected through a lens will pass
through different parts of the lens. Many aberrations are a result of variations in
the polishing of optical surfaces and inhomogeneity in glass materials. Light
rays going through one part of the lens will have different errors than rays
going through different parts of the lens. As a consequence, aberrations will
vary for light passing through different parts of the lens. For a given feature, the
aberrations of its image will depend upon the particular imperfections in those
parts of the lens through which the light from that feature passes. Hence,
aberrations will vary with pitch, numerical aperture, and the illumination
conditions. The light from patterns other than gratings will also be distributed
throughout the optics with dependence upon feature size and proximity to other
features.
As feature sizes shrink, these subtle issues will generate overlay errors of
significance. For example, certain aberrations, such as coma, result in a feature-
size dependency for intrafield registration. This is shown in Fig. 5.10, where
simulated image placement shifts of isolated clear lines are plotted versus
spacewidth and where the lens has a7 = 0.035 waves of coma. 23 Overlay
measurement structures with large features will measure different overlay than
actually occurs for critical fine-linewidth features in the product, in the
presence of coma. This implies that overlay measurements will not represent
the overlay of critical features in the circuits. Measuring overlay using small
features directly is a problem. Overlay is typically measured using optical tools
(Chapter 8) that can measure features reliably only if their size is greater than
about one micron. However, optical tools have throughput and cost advantages

over measurement equipment capable of measuring smaller features. Subtle

issues such as these will need to be addressed when 10 nm becomes a
significant fraction of the total overlay budget.

60
Sigma
1 50 0.3
•
40

30
0.6
a^
U
20

Q110

0 200 400 600 800 1000

Nominal space linewidth (nm)

Figure 5.10. Simulated image shifts for 0.5 NA optics at a wavelength of 248 nm,
and a 7 = 0.035 waves of coma.

In order to enhance productivity, high throughput steppers with large

exposure fields are often used for non-critical layers, while smaller field
steppers are used for the critical layers. The mixing of these different types of
steppers often results in non-concentric exposure fields (Fig. 5.11). Control of
overlay requires extensions of the models discussed thus far, that assume that
the centers of exposure fields sit approximately on top of each other. 24,25
Overlay needs to be measured at the corners of the critical fields, in order to
identify overlay errors associated with intrafield errors associated with the
smaller field. Mathematical models describing field rotation and magnification
must take into account the non-concentric nature of the fields.
Misregistration of geometries on reticles will result in overlay errors on
wafers. This misregistration can have random components, as well as other
contributions which vary systematically across the reticles. An example of a
systematic variation is one which varies linearly across the reticle. Such linearly
varying misregistration will appear as a magnification error. It is possible to
correct for systematically varying reticle errors which correspond to adjustable
intrafield parameters. 122, 26 For example, reticle misregistration which results
from orthogonality errors of the reticle beam writer will correspond to field
skew. Such an error is correctable on a step-and-scan system, but not a step-
and-repeat machine. Correction of reticle errors was described in an example
shown in the prior section.

Figure 5.11. Wafer layout when a wide field stepper is mixed and matched with a
smaller field critical layer stepper.

There remain other overlay errors that are not accounted for in the overlay
model. Wafers can experience non-linear plastic distortions, often caused by
rapid thermal processing, 27 which will not be corrected by alignment methods
that account only for linear errors. Wafer flexure, which could be caused by a
particle under the wafer, will also cause overlay errors which do not conform to
the overlay models discussed, but which are nevertheless quite real. Consider
the situation depicted in Fig. 5.12. Flexure of the wafer will result in lateral
pattern displacement, according to the following equation 28 :

AL = t (5.23)

where t is the thickness of the wafer. We can estimate the magnitude of this
overlay error from elasticity theory. A circular plate clamped at the edges and
displaced by an amount dcenter at its center will be displaced vertically at
radius r, measured from the center of the circle, by the amount 29 :

^ (r) — d eenter a r 2 In a + 2 (a 2 — r 2 ) (5.24)

where a is the radius of the circle. The maximum slope for this displacement is

8 = 4Acenter( (5.25)
max ae

where e is the base for natural logarithms. Suppose a particle on the chuck
causes a maximum vertical displacement of 0.2 gm, affecting a circular area of

0 = 5.89 x 10 -5 radians. For SEMI standard 200 mm

radius a = 5 mm. Then max
wafers, t = 750 µm. From Eq. 5.23, the resulting overlay error is 22 nm. For
180 nm technology, where 65 nm overlay is required, this is a large fraction of
the overlay budget.

Figure 5.12. Wafer deformed because of a particle between the wafer and the
chuck of the stepper.

5.3 CONTRIBUTIONS FROM PROCESSING AND ALIGNMENT MARK OPTIMIZATION

The quality of the overlay is dependent upon the ability of the stepper's
alignment system to acquire alignment targets accurately. There is an interplay
between the alignment signal and the nature of the alignment targets, which
depends upon the overall process for making semiconductors, including film
depositions, resist coatings, etches, and polishes. Consequently, lithography
engineers need to optimize alignment targets. This optimization will include the
dimensions of targets, etch depths, resist coatings, and polishing processes.
The spin coating of resist over topography does not produce symmetrical
films,30' 3' as illustrated in Fig. 5.13. Optical alignment systems will be
susceptible to thin film interference effects which result in low-contrast
signals 32 or misalignment of non-planar alignment marks. 33 ' 34 Because spin
coating is a radial process, the misalignment will go in opposite directions on
diametrical sides of the wafer and result in a scaling error. Resist coating
processes can be modified to reduce the impact of thin film optical effects.30' 35
Optical modeling has shown that dark field alignment systems are less sensitive
to these thin film optical effects than bright field or diffraction alignment
systems. 36 On the other hand, the graininess of highly reflective metal,
polysilicon or silicide surfaces introduces considerable noise into dark field
alignment systems. For metal surfaces, bright field alignment systems are
preferable. The influence of grains can be minimized by the use of low
coherence illumination (large 6) in the alignment optic or interference
techniques, which involves the design of the alignment system and is generally
beyond the control of the user.37

center of the wafer

resist

silicon

wafer topography

Figure 5.13. Resist asymmetry as a consequence of resist spin coating.

Sputter deposition of metals can cause asymmetries in overlay targets.

Metal ions will be partially shadowed by alignment mark topography, leading
to apparent shifts in the position of the alignment mark (Fig. 5.14). Because
sputtering geometries are usually radially symmetric, the resulting overlay
errors will appear as wafer scaling errors. The overlay measurement structures
may be affected in the same way as the alignment targets, and the overlay errors
may be apparent only after etching the metal. The asymmetry is affected by the
geometries of substrate features, and it may be different for large structures
used to measure alignment and critical features in the product. If the
asymmetries are consistent within a lot, and lot-to-lot, then overlay offsets may
be determined off-line and used for correction.
Chemical-mechanical polish (CMP) represents a particular challenge for
overlay.38, 39 The purpose of CMP is to produce highly planarized surfaces.
While this helps lithography by reducing thin film effects and increasing depth
of focus, it makes overlay difficult because it reduces alignment target contrast.
Consider the situation depicted in Fig. 5.15a. Alignment targets are invisible
when planarization is complete and wafers are covered completely by metal
films. The polishing process can be adjusted in order to produce some degree of
topography (Fig. 5.15b), 4° but the polishing often produces asymmetries (Fig.
5.15c), particularly when the width of the pattern is large. Alignment target
deformation can occur from all CMP processes, such as for shallow trench
isolation, not just those involving metal layers.
There are several things that can be done to improve overlay in the context
of CMP. Alignment mark dimensions can be varied or segmented in order to
determine the best mark. Polishing engineers should participate in any
programs to improve overlay, since modifications of the polish process can
often reduce overlay errors considerably, particularly when such errors result
from highly variable polish processes. For example, the use of harder polishing
pads may lead to reduced overlay errors. Stepper manufacturers have also

responded with modifications to their alignment systems, using modified

illumination4 ' and new algorithms for interpreting alignment signals 42 It is also
important to work cooperatively with one's stepper supplier, since alignment
system redesigns can result as the stepper manufacturer learns more about the
actual use of the system.

metal from
sputter target

thinner metal
because of
shadowing by the
wafer topography

metal

substrate

—i center of the wafer

Figure 5.14. Asymmetry as a consequence of the geometry of metal deposition.

Figure 5.15. Alignment targets following chemical-mechanical polish.

5.4 ADDRESSING THE PROBLEM OF NON-NORMAL DISTRIBUTIONS

Distributions of overlay measurements from individual lots are typically very

non-normal. Characterization of overlay and the dispositioning of lots must
take this into account. Consider, for example, the overlay distribution shown in
Fig. 5.16. The overlay had been measured at the four corners of five exposure
fields. In this situation the overlay specification was 120 nm. It can be seen that
all of the measurements are well contained within these specification limits.
Nevertheless, the +/- 36 points are approximately 150 nm, significantly larger
than the specification limit. Clearly the 3s points cannot be compared to the
specification limits in order to disposition lots, and it is therefore not obvious
how process capability can be assessed. The implementation of statistical
process control also becomes complicated because of non-normality, as noted
earlier. The issues of lot dispositioning, SPC, and process capability in the
context of overlay will be discussed further in this section.

Figure 5.16. Overlay distribution from a lot of wafers with a 3 ppm field
magnification error. Vertical lines are located at the specification limits of +/- 120
nm. The sampling plan is shown on the right.

When measurement distributions are normal, lot dispositions and the

implementation of statistical process control is straightforward. This is because
we know the functional form of the distribution, and it is completely
characterized by only two parameters, the mean and the standard deviation.
With overlay we also know the functional form, given by the overlay models
discussed in the previous section. It is, however, much more complicated than a
normal distribution (Table 5.3).

Normal distribution Step-and-repeat overlay Step-and scan overlay

model model
Mean X- & Y-translation X- & Y-translation
Standard deviation X- & Y-scale X- & Y-scale
Wafer rotation Wafer rotation
Orthogonality Orthogonality
Magnification Magnification
x- & y- trapezoid Asymmetric magnification
Reticle rotation Skew
Reticle rotation

Table 5.3. Parameters of different distributions.

Lot dispositioning can be accomplished in several ways. The simplest

method is based upon the largest overlay error, which is either the maximum
overlay error or the absolute value of the minimum error, whichever is biggest.
This approach is simple and works reasonably well. Clearly, if there are
measurements that exceed specification limits, then reworking the lot should be
considered. This method does have some drawbacks that should be noted. First,
the sampling must cover enough of the wafer. If there are wafer scaling errors,
then the worst overlay will be on the outermost portions of the wafer. However,
measurements near the edges of wafers are often unreliable, as a consequence
of non-uniformities in etch, film deposition, and polish. Bad measurements,
among other problems, can lead to outliers in the data, to which criteria based
upon extreme values are very sensitive. Methods for handling outliers will be
discussed later in this chapter.
A more sophisticated method for lot dispositioning directly involves the
parameters of the overlay model. It is good practice to calculate and track the
overlay model parameters, and this is easily accomplished with readily
available measurement equipment and overlay analysis software. Limits can be
placed on all of the individual parameters, and lots will be considered to have
failed to meet specifications if any parameter exceeds its limits.
Appropriate limits for the individual model parameters can be determined
through overlay simulations. Let us assume that each parameter is normally
distributed within each lot (wafer-to-wafer), and that the mean parameter value
varies normally lot-to-lot. One can then run a Monte-Carlo simulation to see
what the resulting overlay distribution might be. 43 The results of one such
simulation are shown in Fig. 5.17, using variations in the model parameters
shown in Table 5.4. This simulation was performed assuming that each
parameter in the overlay model (Eqs. 5.12 and 5.13) varied according to a
normal distribution. The "lot-to-lot" variation provided the mean error for the
wafers within each "lot," for each parameter. The stage precision listed in Table
5.4 gave the residual errors for Eqs. 5.12 and 5.13. For example, the average
translation error for the first "lot" was selected randomly, assuming that the
translation error varied normally, with a mean of zero and a 3 6 variation of

30.0 nm. The "wafers" in this "lot" had this randomly selected translation error,
plus additional translation error which varied "wafer-to-wafer" (15.0 nm, 3a).
Every parameter in the overlay model was determined similarly. The simulation
was repeated over thousands of "lots" and "wafers."

Simulated overlay for the parameters in Table 5.4

a^r.
3~
y
a^
w0
L
N
.a
G
z
-200.0 -150.0 -100.0 -50.0 0.0 50.0 100.0 150.0 200.0
Overlay (nm)

Figure 5.17. Simulated overlay, using the parameters of Table 5.4. The
"sampling" plan is the same one shown in Fig. 5.16, with the addition of
measurements in the centers of the sampled exposure fields.

In this simulation, the overlay was "measured" at five points per 20 mm x

20 mm field (the corners and center), on five fields on the 200 mm diameter
wafers. This distribution of measurement sites determined the values for X, Y, x
and y in Eqs. 5.12 and 5.13. The results of this simulation are shown in Fig.
5.17. The resulting overlay has only a few ppm measurements that exceed 120
nm, so the values in Table 5.4 could be used as specification limits for the
overlay parameters, for a process with 120 nm overlay requirements. There are
numerous combinations of specification limits on the individual overlay
parameters that are consistent with a total overlay budget. Assigning limits to
each parameter is an interative process, where equipment capability is used to
balance parameters in order to minimize rework.
To generate Pareto charts for overlay, contributions need to be converted to
common units. Even though translation errors typically have units of
nanometers, rotations are measured in micro-radians, and scale factors are in
terms of parts-per-million. There are several ways to compare them. The
simplest method is to convert the rotation and scale factors to equivalent
nanometers. A reticle rotation error of 1.2 µrad across a 20 mm square field
contributes up to ± 12 nm overlay error in x and y. A more sophisticated
approach involves the use of overlay models. An overlay simulator can be run
to replicate an existing process. One could then change each parameter in Table
5.2, one at a time, to zero. The contribution of each parameter to overlay is the

difference between the total overlay obtained with the zero and non-zero
parameters.

Parameter Wafer-to-wafer Lot-to-lot

(36) (36)
Grid scale 0.1 0.3 ppm
Orthogonality 0.1 0.3 Microradians
Wafer rotation 0.1 0.3 Microradians
Translation 15.0 30.0 Nanometers
Field magnification 3.0 3.0 ppm
X-magnification 3.0 3.0 ppm
Reticle rotation 3.0 3.0 Microradians
Field skew 3.0 3.0 Microradians
Stage precision 15.0 0 nm

Table 5.4. Parameters for the simulated results shown in Fig. 5.17.

If overlay parameters are going to be used for critical decisions, then the
overlay needs to be measured in ways that minimize the uncertainties in the
model parameters. How this is accomplished can be understood from
consideration of a simple example. Suppose one is fitting to a linear model
involving a single independent variable X:

Y = AX + B, (5.26)
using data collected in pairs:

(X 1 ,YI ), (X 2 ,Y2 ),...,(X n ,Y) (5.27)

Suppose that the Y are independent, and that the residual errors

SY = Y —AX E — B (5.28)

vary with the standard deviation a. If the fit of the model is optimized by the
criterion of least squares, then":

62
Standard deviation of A =
n _ (5.29)
1(Xi —Xy

n
X2
621

Standard deviation of B = L1 X - )—^. (5.30)

n
ny(X i -XY

From these expressions, two rules for obtaining good model parameter
estimates can be seen:
1) The best estimates are obtained by maximizing the separations among the
sampling points X i .
2) The uncertainty in the estimates of the model parameters decreases as the
square root of the number of measurements.
These results are similar for situations involving more than one independent
variable, and have been observed on measurements on products. 45
In order to calculate all linear grid parameters, one needs to measure
overlay in at least three non-collinear fields. This provides six pieces of data (X
and Y measurements), sufficient for computing the six grid parameters, X and Y
translation, X and Y scaling, wafer rotation, and orthogonality. By measuring
equivalent field locations one can avoid confounding interfield and intrafield
contributions. However, measurements in only three fields are usually
inadequate, because of noise. Eq. 5.30 is a useful guide for determining the
minimum number of measurement fields needed. In order to calculate scaling,
rotation, and orthogonality parameters it is useful to locate the measurement
sites close to the edge of the wafer, while avoiding anomalies that may occur
very close to the wafer edge.
The calculation of field terms also requires a minimum of three sites. As
with the grid parameters, the accurate calculation of the model coefficients
requires that the measurement sites be located as close to the outside edge of
the exposure field as possible. As discussed earlier, the in-line measurements
are limited to locations in scribe lanes between product dies, and the results
need to be interpreted accordingly.
Multivariate control charts provide an alternative to charting every model
parameter, 46 which can provide a substantial reduction in the number of control
charts required and the number of false alarms. However, since an out-of-
control condition will require that the specific out-of-control parameters be
identified, it will still be necessary to maintain databases of the individual
overlay parameters.

5.5 OUTLIERS

The problem of outliers, where one or several data points do not appear to
belong to the same distribution as the other data, is illustrated in Fig. 5.18.
There are data which clearly lie outside of the distribution of the rest of the
data. This can result from a number of sources. Metrology can often be the

cause of outliers. In this case the anomalous data can be disregarded in terms of
assessing the true quality of the process. Defects can also be the cause of
measurements which exceed control or specification limits. Particles on the
backsides of wafers can cause high values for overlay, as discussed earlier. If
such particles are attached to the backsides of individual wafers, then the
number of lost dies due to the particles may be small, unless these particles
occur repeatedly. If the particle is attached to the chuck of the stepper, then the
poor overlay will be occurring on all wafers processed on that tool,
significantly degrading yield. In this latter situation, there would be a
significant penalty for ignoring the outliers.

0
0

1
c

Parameter value

Figure 5.18. Histogram of data containing outliers.

Clearly, outliers must be handled carefully. The first step toward dealing
with outliers is identifying them as such. Simply because data exceed control
limits does not mean that there are outliers, since the process may well have
gone out of control. Tests that identify the presence of outliers are referred to as
discordancy tests. 47 A number of tests have been devised to detect outliers in
common distributions. For example, the following statistic can be used to
identify the presence of an outlier in an otherwise normal distribution:

x x x min
T = max xmax —
(5.31)
s s

where s is the sample standard deviation. If T exceeds the values in Table 5.5,
then one can assume that there is an outlier present.
Eq. 5.31 and Table 5.5 provide the information necessary to identify an
outlier in an otherwise normal distribution. As discussed earlier, the data
obtained in overlay measurements often form non-normal distributions. For
overlay, outliers can be identified by analyzing the residual errors. That is the
data which remain after subtracting out the modeled components:

px x =Ox x —(Tx +EX X—R x Y+ nix —rx y), (5.32)

where O x X is the measured overlay error in the X direction. Since the residuals
px,x tend to be normal, discordancy tests such as Eq. 5.31 can be applied.
Many overlay analysis programs automatically reject outlier data when
computing model parameters. As mentioned previously, fully automatic
rejection of outliers can prevent the identification of real and persistent
problems, such as particles on stepper chucks.

n 95% confidence 99% confidence

3 1.15 1.15
4 1.48 1.50
5 1.71 1.76
6 1.89 1.97
7 2.02 2.14
8 2.13 2.28
9 2.21 2.38
10 2.29 2.48
12 2.41 2.63
15 2.55 2.81
20 2.71 3.00

Table 5.5. Values for the parameter T, for the purpose of identifying outliers.

1Semiconductor Equipment and Materials International, Mountain View,

California.

2N. Magome and H. Kawai, "Total Overlay Analysis for Designing Future
Aligner," SPIE Vol. 2440, pp. 902 — 912 (1995).

3 M. A. van den Brink, J. M. D. Stoeldraijer, and H. F. D. Linders, "Overlay

and Field by Field Leveling in Wafer Steppers Using an Advanced Metrology
System," SPIE Vol. 1673, pp. 330 — 344 (1992).

4D. S. Perloff, "A Four Point Electrical Measurement Technique for

Characterizing Mask Superposition Errors on Semiconductor Wafers," IEEE
Sol. St. Circ., Vol. SC-13, No. 4, pp. 436 - 444 (1978).

5T. Saito, S. Sakamoto, K. Okuma, H. Fukumoto, and Y. Okuda, "Mask

Overlay Scaling Error Caused by Exposure Energy Using a Stepper," SPIE
Vol. 1926, pp. 440-449 (1993).

6J. D. Armitage, Jr. and J. P. Kirk, "Analysis of Overlay Distortion Patterns,"

SPIE Vol. 921, pp. 207 — 222 (1988).

' T. E. Zavecz, "Lithographic Overlay Measurement Precision and Calibration

and Their Effect on Pattern Registration Optimization," SPIE Vol. 1673, pp.
191 — 202 (1992).

8V. Nagaswami and W. Geerts, "Overlay Control in Submicron Environment,"

Proceedings of the KTI Microelectronics Seminar, pp. 89 — 106 (1989).

9M. A. van den Brink, C. G. M. de Mol, and R. A. George, "Matching

Performance for Multiple Wafer Steppers Using an Advanced Metrology
Procedure," SPIE Vol. 921, pp. 180 — 197 (1988).

J. G. Maltabes, M. C. Hakey, and A. L. Levine, "Cost/Benefit Analysis of

Mix-and-Match Lithography for Production of Half-Micron Devices," SPIE

Vol. 1927, pp. 814 — 826 (1993).

11S. K. Jones, E. S. Capsuto, B. W. Dudley, C. R. Peters and G. C. Escher,

"Wavelength Tuning for Optimization of Deep UV Excimer Laser
Performance," SPIE Vol. 1674, pp. 499 — 508 (1992).

12 M. E Preil and W. H. Arnold, "Aerial Image Formation with a KrF Excimer

Laser Stepper," Polymer Engr. and Sci., Vol. 32(21), pp. 1583 —1588 (1992).
13
R. K. Brimacombe, T. J. McKee, E. D. Mortimer, B. Norris, J. Reid, and
T. A. Znotins, "Performance Characteristics of a Narrow Band Industrial
Excimer Laser," SPIE Vol. 1088, 416 — 422 (1989).
14
H. J. Levinson and R. Rice, "Overlay Tolerances for VLSI Using Wafer
Steppers," SPIE Vol. 922, pp. 82-93 (1988).

15M. A. van den Brink, C. G. M. de Mol, and J. M. D. Stoeldraijer, "Matching

of Multiple Wafer Steppers for 0.35 pm Lithography, Using Advanced
Optimization Schemes," SPIE Vol. 1926, pp. 188 —207 (1993).

16 J. C. Pelligrini, "Comparisons of Six Different Intrafield Control Paradigms

in an Advanced Mix-and-Match Environment," SPIE Vol. 3050, pp. 398 — 406

(1997).

17 H. J. Levinson, M. E. Preil, and P. J. Lord, "Minimization of Total Overlay

Errors on Product Wafers Using an Advanced Optimization Scheme," SPIE
Vol. 3051, pp. 362-373 (1997).

18 N. R. Farrar, "Effect of Off-axis Illumination on Stepper Overlay," SPIE Vol.

2439, pp. 273 — 280 (1995).

19 T. Saito, H. Watanabe, and Y. Okuda, "Effect of Variable Sigma Aperture on

Lens Distortion and Its Pattern Size Dependence," SPIE Vol. 2725, pp.
414 — 423 (1996).

20 A. M. Davis, T. Dooly, and J. R. Johnson, "Impact of Level Specific

Illumination Conditions on Overlay," Proceedings of the Olin Microlithography
Seminar, pp. 1— 16 (1997).
21
C. S. Lee, J. S. Kim, I. B. Hur, Y. M. Ham, S. H. Choi, Y. S. Seo, and S. M.
Ashkenaz, "Overlay and Lens Distortion in a Modified Illumination Stepper,"
SPIE Vol. 2197, pp. 2— 8 (1994).
22
C-M. Lim, K-S. Kwon, D. Yim, D-H. Son, H-S. Kim, and K-H. Baik,
"Analysis of Nonlinear Overlay Errors by Aperture Mixing Related to Pattern
Asymmetry,' SPIE Vol. 3051, pp. 106 — 115 (1997).
23
T. A. Brunner, "Impact of Lens Aberrations on Optical Lithography," IBM
J. Res. Develop., Vol. 41(1/2), pp. 57-67 (1997).
24
M. E Preil, T. Manchester and A. Minvielle, "Minimization of Total Overlay
Errors when Matching Non-concentric Exposure Fields, SPIE Vol. 2197, pp.
753 — 769 (1994).
W. W. Flack, G. E. Flores, J. C. Pellegrini, and M. Merrill, "An Optimized
25

Registration Model for 2:1 Stepper Field Matching," SPIE Vol. 2197, pp.
733 — 752 (1994).
26
R. Rogoff, S. S. Hong, D. Schramm, and G. Espin, "Reticle Specific
Compensations to Meet Production Overlay Requirements for 64 Mb and
Beyond," SPIE Vol. 2197, pp. 781 — 790 (1994).

27 G. Rivera and P. Canestrari, "Process Induced Wafer Distortion:

Measurement and Effect on Overlay in Stepper Based Advanced Lithography,"
SPIE Vol. 807, pp. 806 — 813 (1993).

28H. Izawa, K. Kakai, and M. Seki, "Fully Automatic Measuring System for
Submicron Lithography," SPIE Vol. 1261, pp. 470-481 (1990).
29
A. E. H. Love, A Treatise on the Mathematical Theory of Elasticity, 4th
Edition, Dove, New York (1944).

3o
K. Chivers, "A Modified Photoresist Spin Process for a Field-by-Field
Alignment System," Proceedings of the Kodak Microelectronics Seminar, pp.
44-51 (1984).
31
L. M. Manske and D. B. Graves, "Origins of Asymmetry in Spin-Cast Films
Over Topography," SPIE Vol. 1463, pp. 414 — 422 (1991).
32
• Kuniyoshi, T. Terasawa, T. Kurosaki, and T. Kimura, "Contrast
Improvement of Alignment Signals from Resist Coated Patterns," J. Vac. Sci.
Technol., B5(2), pp. 555 — 560 (1987).

33
G. Flores and W. W. Flack, "Photoresist Thin-Film Effects on Alignment
Process Capability," SPIE Vol. 1927, pp. 367 — 380 (1993).
3a
N. Bobroff and A. Rosenbluth, "Alignment Errors from Resist Coating
Topography," J. Vac. Sci. Technol., B6(l), pp. 403 — 408 (1988).
3s
R. Mohondro, S. Bachman, T. Kinney, G. Meissner, and D. Peters, "High
Contrast Eduction Spin Coat Process Effects on Uniformity and Overlay
Registration," Proceedings of the Olin Microlithography Seminar, pp.
131 — 140 (1995).
36
C. P. Kirk, "Theoretical Models for the Optical Alignment of Wafer
Steppers," SPIE Vol. 772, pp. 134 — 141 (1987).

37K. Ota, N. Magome, and K. Nishi, "New Alignment Sensors for Wafer
Stepper," SPIE Vol. 1463, pp. 304 — 314 (1991).

38 D. L. Meunier, D. Humphrey, B. Peck, P. Feeney, J. Paquette, and

J. Thibault, "Optimization of Metal Layer Overlay and Alignment Targets in a
Chemical-Mechanical Polishing Environment," Proceedings of the Olin
Microlithography Seminar, pp. 355 — 366 (1996).
39
S-W. Hsia, G. Miyagi, and M. Brongo, "Alignment Characterization of CMP
Tungsten Process," Proceedings of the Olin Microlithography Seminar, pp.
381 — 390 (1996).

40D. L. Meunier, B. Plambeck, P. Lord, and N. Knoll, "The Implementation of

Coherence Probe Microscopy in a Process Using Chemical Mechanical
Polishing," Proceedings of the OCG Microlithography Seminar, pp. 155 — 169
(1995).

41 N. Shirishi, A. Sugaya, and D. P. Coon, "Alignment Strategies for

Planarizing Technologies," SPIE Vol. 3051, pp. 836 — 845 (1997).
42
T. Kanda, K. Mishima, E. Murakami, and H. Ina, "Alignment Sensor
Corrections for Tool Induced Shift (TIS)," SPIE Vol. 3051, pp. 846 — 855
(1997).
43
W. H. Arnold, "Overlay Simulator for Wafer Steppers," SPIE Vol. 922, pp.
94 — 105 (1988).

M. H. DeGroot, Probability and Statistics, Addison-Wesley Publishing

Company, Menlo Park, CA (1975).
45
B. Rangarajan, M. Templeton, L. Capodieci, R. Subramanian, and A.
Scranton, "Optimal Sampling Strategies for Sub-100 nm Overlay," SPIE Vol.
3332, pp. 348 — 359 (1998).
46
G. E. Flores, W. W. Flack, S. Avlakeotes, and B. Martin, "Process Control of
Stepper Overlay Using Multivariate Techniques," Proceedings of the OCG
Microlithography Seminar, pp. 201— 215 (1995).

47V. Barnett and T. Lewis, Outliers in Statistical Data, 3rd Edition, John Wiley
and Sons, Chichester (1995).

6.1 YIELD MONITOR STRATEGY

Productivity is closely related to yield, the fraction of dies on the wafer that are
functional at the end of the manufacturing process. Yield loss results from:

• Parameters that are out of specifications or improperly specified.

• Process or equipment failures.
• Chemical contamination.
• Particulate defects.

The issues of parameters such as overlay and linewidths were discussed in the
prior two chapters. Process or equipment failures, which might occur, for
example, when a resist pump malfunctions and wafers are not completely
covered with resist, are more properly the subjects of a book on equipment
maintenance, and will not be covered in this text. Chemical contamination
typically results from contaminated photochemicals, and will not be considered
in detail. Particulate contamination is the primary source of yield loss that
lithography engineers need to address and will be the subject of this chapter.
Yield is a significant issue in lithography because of the number of
masking steps required to fabricate an integrated circuit. Suppose there is a 3%
yield loss at each masking step, due to random defects. While this may appear
to be a small number, the total yield loss from lithography, in a 30 masking step
process, is

Yield loss = 0.97 30 = 0.40. (6.1)

A small yield loss per masking step causes an overall large yield loss, because
of the large number of lithography operations involved.
Not all defects are caused by lithography:

Yield = LDLY x NLLY x PALY x DLY (6.2)

where

LDLY = lithography defect limited yield,

NLLY = non-lithography limited yield,
PALY = parameter limited yield, and
DLY = design limited yield.

119

All of these yield quantities are fractions between zero and one. Design limited
yield occurs when parts are not designed to always yield even when all
parameters are within specifications. Lithographers need to be able to measure
the lithography contribution to yield, separate from other possible confounding
factors.
It is also important that defects be detected before parts reach electrical test.
A high level of defects at gate mask could go undetected for weeks if one relied
only on electrical testing of fully fabricated devices, during which time millions
of dollars of product could be lost. Electrically measurable structures,' such as
those shown in Fig. 6.1, can be created with only a few process steps, reducing
the time required for detection. For these types of monitors, an electrically
conducting film deposited on an oxide-covered wafer is patterned
lithographically and etched. There should be nearly infinite electrical resistance
between points A and B, in the absence of defects, while the resistance between
points C and D should be very low. Differences from these expected resistances
indicate the existence of defects. For these types of monitors, defects may arise
from films and etch, as well as lithography, and the particular process
responsible for shorts or opens must still be determined after a defect is found
electrically. It is also essential to have a method that can quickly assess the
level of defects in particular processes, in order to see if corrective action was
successful. For all of these reasons, yield engineering requires systems that are
capable of inspecting wafers and identifying defects.

Figure 6.1. Structures for measuring defects electrically. The resistance between
pads A and B should normally be very high and can provide a measurement of
defects that could have been caused by particles bridging lines of resist. The
resistance between pads C and D normally should be very low and can provide a
measurement of defects which could have been caused by breaks in the line of
resist.

Inspection systems are classified according to whether they are designed to

measure patterned or unpatterned wafers. The most commonly used systems for
inspecting pattern wafers are optical comparators, which compare optical
signals from equivalent positions in different dies. 2 If the signals from the
different positions are different, one can assume that there is a defect present at
one of the positions. The optical signal is measured in a third location in order

to determine which position has the defect. Holographic systems have also been
used for measuring defects on patterned wafers, 3 4 5 but these systems are used
' '

less frequently than optical comparators.

Laser light scattering is used to detect defects on unpattemed wafers (Fig.
6.2). If the film shown in Fig. 6.2 were homogeneous, the light would consist
only of the incident and specularly reflected beams. Particulates and bubbles
result in scattered light. The physics of laser light scattering defect detection
systems will be discussed in more detail in a later chapter.

Figure 6.2. Configuration for defect detection systems using the scattering of light.

One of the biggest drawbacks of optical systems is their limited resolution.

Defects as small as 100 nm can cause yield loss in 250 nm technologies, and
this size of defect is falling below the detection limit of optical inspection
systems. A defect inspection tool based upon a scanning electron microscope
has been introduced recently, but it can only inspect for 0.1 µm defects at the
rate of 0.033 cm2/minute 6 At this rate it is impractical to perform routine defect
monitors. This is a problem that will have to be addressed in order to
implement future technologies.
The following questions need to be answered when setting up systems for
controlling yield:
1)Which processes will be monitored?
2)Which monitors will involve patterned wafers and which will use
unpatterned wafers?
3)How many wafers should be measured?
4) Should product wafers be monitored for defects or should one use
special defect monitors?
Each of these questions will be discussed in this chapter.
Lithography defects are caused almost exclusively by resist processes,
which include the spin coating of photoresists, anti-reflection coatings, and
developing. Particulates can be caused by mechanical systems, but these are
typically very clean and cause high levels of defects only when broken. While
this does happen on occasion, defects are caused far more frequently by the

resist processes. Consequently, defect monitors should focus on resist coaters

and developers.
Resist and anti-reflection coatings can be monitored using laser scattering
systems. These have the advantage of speed, since wafers can be measured in
only a few minutes, and the systems for inspecting unpatterned wafers are
usually much less expensive than those that can inspect patterned wafers. If
resist or anti-reflection coatings are known to have defects often, frequent
monitors of these coatings can be instituted relatively easily. A defect found in
early DUV processes involved a surface energy mismatch between the
photoresist and the underlying anti-reflection coatings, which resulted in
microbubble formation between the two films. 7 ' 8 It was possible to detect,
monitor, and control these defects using inspection systems on unpatterned
wafers.
Resist patterns are usually required to identify developer defects. Typical
defects associated with the develop process are residues or incomplete develop,
such as missing contacts. There has recently been considerable focus on the
develop process as the source of lithography defects. Systems capable of
inspecting patterned wafers are usually required to identify defects which result
from the develop process.
Truly random defects will appear as a Poisson process, in which the
probability that a wafer will have N defects is given by

N
P(N) = N^ e -d , (6.3)

where d is the average number of defects per wafer. The resulting probability
distribution is shown in Fig. 6.3. Control charts can be generated, but care must
be taken to take possible non-normality into account. In Fig. 6.3, skewness is
evident. A sufficient number of wafers must be measured to have statistical
significance, and this can be a challenge when there are only a few defects per
wafer.
The yield monitoring strategy will depend on process maturity. For
immature processes, frequent monitors of equipment and individual processes
should be used, often using unpatterned wafers. 9 This is particularly true when
non-lithography processes produce high levels of defects, or when the
underlying films are very grainy. Monitors that involve the use of bare silicon
wafers permit smaller detection limits without overloading the system with
signals from substrate defects and grains. A general recommendation is that
defect monitor wafers should have 1/4 of the defects of the control limit in
order to obtain good results. For example, if the control limit is 20, monitor
wafers should have fewer than five defects. This is often very difficult to find
on product wafers that have gone through several processes. The use of non-
product wafers for defect monitoring parallels the use of non-product wafers for
monitoring linewidth and overlay control that was discussed in Chapter 3.

0.14

0.12

0.1

w
0
0.08

0.06

• 0.04

0.02

0
3 5 7 9 11 13 15 17 19
2 4 6 8 10 12 14 16 18 20
Number of defects per wafer

Figure 6.3. Probability of having a certain number of defects/wafer for a process

where the defects are distributed randomly and the average number of defects per
wafer is 10.

For mature processes, one can measure defects directly on product wafers,
using automated inspection equipment. It is important to check the tools, not
the lots, since the tools are the sources of defects. In general, this implies that
far fewer than 100% of the lots need to be inspected. In some instances the
resist processing equipment will have multiple modules per operation, in order
to maintain throughput. For example, resist tracks often have two or three
developer modules and two or three coater modules. In these situations the
sampling plan must be designed to include wafers from every module.
Scheduling measurements to properly sample tools and modules is more
difficult with production wafers than with separate defect monitors. It is easier
to design defect monitors which use non-product wafers. Sensitivity to very
small defects is usually highest on bare silicon wafers, because of the low
background of defects and substrate graininess. Where very small defects are a
concern, bare silicon wafers should be used for defect monitoring.
Measurement of product wafers has the advantage that additional processing
time is not required, thereby maximizing photocluster availability for
production, and may be used for this reason.
Defect classification provides a means for improving control of defects.
The defects created by lithography processes often have signatures in their
images. 10 Monitoring each defect type individually increases sensitivity for
detection. For example, suppose that each defect type contributes 10 defects on
average per wafer, and with a standard deviation of nine defects,

wafer-to-wafer, for each of three particular defect types. Then the total defect
level is 30 defects average per wafer, and the wafer-to-wafer variation is
6 = 16. Suppose that the defect level for one particular type of defect increases
to 15 defects average per wafer, and the standard deviation increases to 12. This
is a mean shift of 0.566 and an increase in 6 of 33% for that particular defect.
Overall, the mean total defect level will shift only 0.31o, and 6 overall changes
by only 6%. The monitors of the individual defects are therefore more sensitive
than monitors of the total defect level. There are a number of defect analysis
systems that can perform defect classification automatically. Some of these are
integrated into the defect detection system, while others are available on stand-
alone defect review stations. Automatic defect classification not only improves
sensitivity, but it also facilitates the corrective action, since the particular source
of the defects is often immediately identified.

6.2 YIELD MODELS

Yield models are very important when more than one product, process or
technology is involved. Yield models relate measured defect densities to
product yield. It is then possible to compare yield between products to see if
there is a design-related issue limiting yield on a particular product. Yield can
be trended as the product mix evolves. Yield models also provide the means for
extrapolating yields to new die sizes and new products yet to be produced. Most
importantly, yield models can be used to facilitate improvement."
Lithography defects do not always result in non-functional dies. A
protrusion off a metal line will generally cause a functional failure only when
the protrusion is large enough to result in bridging to another metal line. Some
pieces of metal might be located in a field area where there are no relevant
devices, so no damage is done. The probability k that a physical defect will
result in a non-functional device is called the "kill ratio."
If we assume that the defects are random, that is, the probability of defect
formation is constant in time, then the yield Y is given by

Y _ e -11 , (
6.4)

where D is the defect density in units of defects per unit area, and A is the die
area. In addition to random defects there may be systematic yield losses. For
example, products that are very sensitive to gate length may have yield loss
non-randomly because of inadequate linewidth control. In the presence of
systematic yield loss, the expression for yield is given by:
Y = Yse-" DA,
(6.5)

where Y represents the systematic yield loss.12

Defects often form in clusters. In this case the probability of defect

formation depends on the existing level of defects. In this situation the yield
follows a slightly different model:

kDA -a
Y=Y
1+ a (6.6)

Note that Eq. 6.6 reduces to Eq. 6.5 as a • An example of cluster defects is
—> O0

shown in Fig. 6.4. These spoke defects were a consequence of a surface energy
mismatch between the resist as it was being applied and an organic bottom anti-
reflection coating. This mismatch resulted in the formation of microbubbles. As
the wafer was spun the bubbles broke up and formed a string of smaller
bubbles.
If defects are random, then yield should behave exponentially as a function
of chip size. To determine the parameters of yield models with additional
parameters, such as systematic yield factors or cluster parameters, data on more
than one die size are needed. Often one does not have access to yield data for
more than one device, so the "window method" may be used to determine
whether there are significant levels of cluster defects and to calculate the cluster
parameter a if the cluster model (Eq. 6.6) is appropriate. 13 In this method, yield
is determined as a function of chip multiples. On the wafer layout, chips are
grouped together, to form effectively larger "chips." If any real chip within the
larger window has a defect, then the windowed chip is considered to have a
defect.

Figure 6.4. Spoke defects, generated by bubble formation at the interface between
the resist and the anti-reflection coating. As the resist is spun, the bubbles break
up into smaller multiple microbubbles.

The effect of clustered defects can be seen in Table 6.1. The random and
cluster models do not produce significant differences in yield for small die sizes
or very low defect densities. However, the random model overestimates the
yield losses for larger dies.

D = 0.1 defects/cm , a = 1.0 D = 0.3 defects/cm 2 , a = 1.0

Die size Random model Cluster model Random model Cluster
(cm2) model
0.5 0.95 0.95 0.86 0.87
1 0.90 0.91 0.74 0.77
1.5 0.86 0.87 0.64 0.69
2 0.82 0.83 0.55 0.63
2.5 0.78 0.80 0.47 0.57
3 0.74 0.77 0.41 0.53
3.5 0.70 0.74 0.35 0.49
4 0.67 0.71 0.30 0.45

Table 6.1. The effect of clustered defects.

Defect distributions are frequently very non-normal. The number of defects

per wafer is always non-negative, creating a problem for the lower control
limit. For truly random defects, Eq. 6.3 provides the basis for determining
control limits, based on a user-selected rate for false alarms. However, defects
in lithography are often very non-random. Clustered defects-such as those just
described-or scratches, are common, and these result in very long "tails" to
the defect distributions. 14 Data are sometimes fit to lognormal or other
distributions,' 5 ' 16 or transformed to a normal distribution," from which control
charts can be generated.
Another way to reduce sensitivity to clustered defects is to monitor the
number of defective dies rather than the total number of defects, since clusters
of defects are often contained within a single die. This has the additional
advantage of transforming the problem to one in which conventional np control
charts can be used 18 [Chapter 1, Ref. 3]. In this case p is the probability of a
single die having at least one defect, and n is the number of dies measured each
time.

6.3 PARAMETERS WHICH AFFECT YIELD

Lithography defects are created predominately by the resist process. Coatings

of resist and anti-reflection coatings, and the develop process, have potential for
causing defects. Consequently, yield improvement efforts in lithography focus
primarily on the resist process.

Photoresist and developers are usually filtered at the point-of-use. As might

be expected, pore size is an important factor. For state-of-the-art processing,
0.1 µm pore size is typical. The interaction of the resist and developer with the
filter material is important. Air bubbles will form if the fluids do not wet the
filters well. These bubbles will detach from the filter surface over time,
representing a source of defects. 19 Good wetting can be accomplished by
"pre-wetting" with appropriately chosen solvents or by the choice of filter
materials. 20
The thin photoresist films on the wafers are formed by spinning the wafers.
Excess liquid resist is spun off the wafers, and this waste resist has the potential
for hitting the walls of the coater bowl and splashing back onto the wafers. A
similar problem occurs when developer is spun off wafers. Splash-back can be
controlled by the design of the coater or developer bowl, choice of accelerations
and spin speeds, and exhaust control. High spin speeds can easily Iead to
splash-back. 2 ' Keeping the coater bowl clean also reduces the number of
defects created by splash-back. Some resist coaters have automatic bowl
cleaning capabilities, while manual cleans are required periodically on systems
without this feature. Daily cleans have been found to lead to measurable yield
improvement. 22
Exposed resist will dissolve in basic aqueous solutions. If the pH becomes
too small the resist will not stay in solution. If water is applied to a developer
puddle the pH will be reduced, and some resist may fall out of solution.
Reduced defects have been found by spinning the developer off before
rinsing. 23
Steppers can be the source of yield loss, and these losses will usually
appear with non-random patterns, reflecting the non-random step-and-repeat or
step-and-scan process. While the defects created by resist processes are
typically particulate defects, yield loss caused by steppers are usually related to
parametric problems, such as loss of linewidth or overlay control. It is an
extremely unusual circumstance where the Class 1 (or better) environment of
the stepper is degraded to the point where there are problems with particulate
contamination. Other problems with steppers may occur, and these will usually
be manifested in the linewidth or overlay.
Consider, for example, the stepping pattern shown in Fig. 6.5. A
magnification error would affect those dies that are farthest from the center of
the exposure field. The outermost dies would have lower yield than the inner
dies for an overlay-sensitive process, and this pattern would repeat from stepper
exposure field to exposure field. A similar yield pattern would be evident if
there was a tilt in focus across the stepper field.

Figure 6.5. Wafer layout with multiple product dies in each stepper field.

' C. F. King, G. P, Gill, and M. J. Satterfield, "Electrical Defect Monitoring for

Process Control," SPIE Vol. 1087, pp. 76 — 82 (1989).

2G. Dickerson and R. P. Wallace, "In-line Wafer Inspection using

100-Megapixel -per-Second Digital Image Processing Technology," SPIE Vol.
1464, 584 — 595 (1991).

3 L. H. Lin, D. L. Cavan, R. B. Howe, and R. E. Graves, "A Holographic

Photomask Defect Inspection System," SPIE Vol. 538, pp. 110 — 116 (1985).

4S. P. Billat, "Automated Defect Detection on Pattern Wafers," Semiconductor

International, pp. 116 — 119, (May 1987).

5D. L. Cavan, L. H. Lin, R. B. Howe, R. E. Graves, and R. L. Fusek,

"Patterned Wafer Inspection Using Laser Holography and Spatial Frequency
Filtering," J. Vac. Sci. Technol., B6(6), pp. 1934 — 1939 (November 1988).

6D. Hendricks, J. Jau, H. Dohse, A. Brodie, and D. Meisburger,

"Characterization of a New Automated Electron-Beam Wafer Inspection
System, SPIE Vol. 2439, pp. 174 —183, (1995).

J. Sturtevant, M. Chaara, R. Elliot, L. Hollifield, R. Soper, D. Stark, N. Thane

and J. Petersen, "Antireflection Coating Process Characterization and
Improvement for DUV Lithography at 0.25 tm Groundrules," SPIE Vol. 2440,
pp. 582 — 593 (1995).

o E. H. Bokelberg and M. E. Pariseau, "Excursion Monitoring of

Photolithographic Processes," Olin Microlithography Seminar, pp. 85 — 100
(1997).

9K. Phan, R. Chiu, S. Punjabi, and B. Singh, "Efficient and Cost Effective
Photo Defect Monitoring," SPIE Vol. 3332, pp. 709 — 720 (1998).
10
J. R. Dralla and J. C. Hoff, "Automatic Classification of Defects in
Semiconductor Devices," SPIE Vol. 1261, pp. 173 — 182 (1990).

" D. Dance and R. Jarvis, "Using Yield Models to Accelerate Learning Curve
Progress," IEEE Trans. Semi. Manufact., Vol. 5(1), pp. 41— 45 (1992).
12
A. Y. Wong, "Statistical Micro Yield Modeling," Semiconductor
International, pp. 239 — 148 (November, 1996).

13 S. H. Stapper, F. M. Armstrong, and K. Saji, "Integrated Circuit Yield

Statistics," Proc. IEEE, Vol. 71(4), pp. 453 — 470 (1983).

14
D. J. Friedman and S. L. Albin, "Clustered Defects in IC Fabrication: Impact
on Process Control Charts," IEEE Trans. Semiconductor Manufacturing, Vol.
4(1), pp. 36-42 (1991).

15 D. Michelson, "A Data-Driven Method for Calculating Limits for Particle

Control Charts," SPIE Vol. 2876, pp. 38 —44 (1996).

16
S. T. Mandraccia, G. D. Halverson, and Y-M. Chou, "Control Chart Design
Strategies for Skewed Data," SPIE Vol. 2876, pp. 196 — 205 (1996).

Y-M Chou, A. M. Polansky, and R. L. Mason, "Transforming Non-Normal

17
Data to Normality in Statistical Process Control," Journal of Quality
Technology, Vol. 30(2), pp. 133 — 141 (1998).

18D. Bakker and D. Icke, " Lithography Defect Characterization Using

Statistical Process Control and Automated Inspection," Proceedings of the KTI
Microelectronics Seminar, pp. 361— 367 (1990).

19 J. A. Orth, K. A. Phan, D. A. Steele, and R. Y. B. Young, "A Novel

Approach for Defect Detection and Reduction Techniques for Submicron

Lithography," SPIE Vol. 3050, pp. 586 — 601 (1997).

20M. E. Clarke and K-S. Cheng, "New Photochemical Filtration Technology

for Process Improvement," Olin Microlithography Seminar poster session
(1997).

21
Y-K. Hsiao, C-H. Lee, and K-L Lu, "A Simple Method to Reduce Post
Develop Residue," Olin Microlithography Seminar poster session (1997).
22
Y-T. Fan, H-P. Lin, Y-C. Lo, C-H. Lee, and K-L. Lu, "A Study on Methods
to Reduce Metal Defects Caused by a Coating Process," Olin Microlithography
Seminar poster session (1997).
23
E. H. Bokelberg, J. L. Goetz, and M. E. Pariseau, "Photocluster Defect
Learning and Develop Process Optimization," Olin Microlithography Seminar,
pp. 127 — 139 (1996).

7.1 ADJUSTING FOR PROCESS DRIFT

There are times when processes drift, and the best course of action is to make
an adjustment. This situation occurs when overlay correctable parameters or
focus offsets are changed. Every lithography operation should have controls to
detect process drift. As discussed in Chapter 1, the ±36 rule is not particularly
sensitive to shifts in the process mean, compared to other Western Electric
Rules. Consequently, it is advisable to include at least one additional Western
Electric Rule to supplement the ±3a rule to increase sensitivity to process drift.
An alternative to the Shewhart control chart, with greater sensitivity to shifts in
the process mean, is the exponentially-weighted moving average (EWMA)
chart, discussed in the next section.
Suppose that a process has been under control until a Western Electric Rule
is violated. If, after investigation, no assignable cause can be identified, it may
be decided that the process should be adjusted. Determining the amount of
adjustment is the subject of this section. Each of the Western Electric Rules
involves a certain amount of data. The ±36 rule is violated when a single point
exceeds control limits, while the eight-in-a-row rule involves 8x more pieces of
data. As discussed earlier, the process mean .t is known to within a "±3o' level
of confidence (99.7%) following n measurements:

µ = x ± - ( 7.1)
n

With more measurements there will be a smaller uncertainty in the actual

process mean (Fig. 7.1). For this reason, it is useful to repeat measurements
when Western Electric Rules have been violated. This is particularly true for
the ±36 rule, for which there may be only a single measurement indicating a
loss of process control. As seen in Fig. 7.1, the violation of the ±3a rule could
have been caused by only a very small shift in the process mean. Changing the
operating point to move the single violation to the process mean would most
likely represent over-compensation. The exponentially weighted moving
average, discussed in the next section, provides a means of both detecting and
estimating the process mean, particularly in situations involving gradual drift.
Suppose only the ±36 rule has been violated, and it is proposed to adjust
the process assuming that the single measurement represents a new shifted
process. The process is then adjusted on the basis of this assumption to some
percentage of the difference between the single value and the process target.

131

Control limits
One point beyond a control limit
I +/- 3a
Process mean
x ....
y_y_y___-_____-______ - _____ +/_ 3v
F
A run of 8 points
below the mean

0 5 10 15 20 25
Day
Figure 7.1. X-bar chart, where the uncertainty in the process mean is greater
when based upon a single value, compared to eight values.

The operation is performed again and new measurements are taken, and the
process is again adjusted on the basis of a single determination of the state of
the process. Ultimately the mean of the process will converge to the desired
target. It has been shown that the most efficient way to conduct this procedure
is to correct for 100% of the difference between the measured value and the
target on the first iteration, 50%® on the second, and 100x 1 % on the nth
n
iteration.' The final result is identical to that obtained by performing the
operation n times, taking the measurements, and making a single adjustment,
but has the advantage that an adequate process may be reached more quickly.
The primary difficulty with any attempt to determine the true state of the
process from multiple tests and measurements is the need to account for
fluctuations which occur over time scales greater than a few hours. This can be
accomplished only by monitoring the process over the appropriate time scale.

Example: Consider a process where the linewidth target is 250 nm. Suppose
that, although the correct exposure dose is unknown, the exposure latitude is
known to give a 25 nm shift in linewidth for a 10% change in exposure dose.
Using the algorithm where the correction on the nth iteration is 100 x 1 %, one
n
obtains the following.

Iteration Exposure dose Measured mean 100% 1

linewidth correction 100 x — %
n
correction
0 20.0 mj/cm2 280 nm 2.4 m'/cm2 2.4 m'/cm2
1 22.4 mjlcm2 260 nm 0.8 m'/cm 0.4 mj/cm 2
2 22.8 mj/cm2 242 nm - 0.6 m'/cm - 0.2 m'/cm
3 22.6 m'/cm 248 nm

7.2. THE EXPONENTIALLY-WEIGHTED MOVING AVERAGE

The exponentially-weighted moving average (EWMA) for measurement t is

defined as t :

EWMAt = a,yt + 1 — a,)EWMAt _ 1(7.2) (

= Xy t + x(1— x )yt-1 + X (i — X )2 yt-2 +... (7.3)

where 0< A. <-1 and t = 1,2,..., and y t is the parameter of interest. To calculate
EWMAI the process target or some prior estimate of the process mean for
" EWMA0 " can be used. The exponentially-weighted moving average provides
a measure of the process mean with more recent data being weighted more
heavily than older measurements. Smaller values of A weight older data more
heavily than larger values. As A —*1, the EWMA weighs the most recent data
more heavily, and control charts of the exponentially-weighted moving average
begin to resemble a standard Shewhart chart, such as shown in Fig. 1.5. The
value of A is something that must be determined by the person setting up the
control chart. Considerations needed to set this value will be discussed shortly.
If the y t are independent normally distributed random variables with
standard deviation a, then the standard deviation of the exponentially-weighted
moving average is known in closed form a :

EWMAt
2 % J

—6 for large values of t. (7.5)

2—A
From this, control limits for an exponentially-weighted moving average control
chart can be generated:

)\
Upper control limit = t + L6 2 A— (1—)
fl J, (7.6)

Lower control limit = µ — L6 2 —(i _ A)2 ` ( 7.7)

where .t is the process center line and L is another parameter to be determined

by the person setting up the control chart. These control limits approach
constant values as t gets large.

The values for A and L can be determined once two questions have been
answered:
1)What is the desired average run length (ARL) for the process when it is
in control?
2)What is the minimum shift at which detection is desired?
Values which lead to optimized EWMA control charts have been calculated for
various values of the in-control ARL and process shifts, for normal and
exponentially distributed processes. Examples of values for A and L are given in
Table 7.1 for a normally distributed process . 4

Minimum shift In-control average run length

(in units of 6) 100 300 500
0.5 A 0.065 0.055 0.050
L 1.98 2.43 2.62
ARL at shift 17.3 24.9 28.7
1.0 A 0.175 0.145 0.135
L 2.32 2.71 2.71
ARL at shift 7.0 9.1 10.2

Table 7.1. Values for A and L for establishing EWMA control charts for minimum
detectable shifts of 0.5 and 1.0 6, for a various average run lengths for the in-
control process.

EWMA control charts are sensitive to small shifts in the process mean. As
can be seen from the data in Table 7.1 and Fig. 1.3, this sensitivity is
comparable to conventional Shewhart charts used in conjunction with the
Western Electric Rules, but with fewer false alarms for processes in control . 5
The EWMA charts also have the advantage of providing the user with an
immediate estimate of the process mean. As seen in the preceding section, this
is not an immediate result of the Shewhart control charts. This is illustrated by
an example. Consider the linewidth data shown in Fig. 7.2, for a process with
an in-control mean of 250 nm and control limits of ± 25 nm (a = 8.33 nm)
about the mean. After 50 measurements, drift was introduced into the process,
and the linewidths grew larger, on average, over time. After 104 data points the
"eight-in-a-row" Western Electric Rule was violated, and the "± 36" rule was
violated at the 115th point. The corresponding EWMA average control chart for
this process is shown in Fig. 7.3. Control limits were calculated to provide
sensitivity to small shifts and were assumed to be based on a large number of
prior data (t large). An in-control average run length of 300 and a minimum
detected shift of 0.5a were chosen, giving a value of 0.055 for A (Table 7.1).
From Eqs. 7.6 and 7.7, the control limits for the EWMA control chart are 250 ±
3.41 nm. The drift in the process was detected at point 77 with the EWMA
control chart. A trend upwards in the mean of the process can be seen in both
charts, but it is more evident in the EWMA control chart. The EWMA is

particularly useful for calculating process parameters in the feedback systems

discussed in Chapter 3.

280

270

260

. 250 Control
3 limits
240

230

220
0 10 20 30 40 50 60 70 80 90 100 110 120
Sample #

Figure 7.2. A Shewhart variables control chart for a process that experienced a
drift in the process after sample #50.

260
C

5 255
vC
ó 250 Control
limits

245

U 1U 20 30 40 50 60 70 80 90 100 110 120

Sample #

Figure 7.3 The EWMA control chart for the same process illustrated in Figure
7.2.

7.3. AUTOMATIC PROCESS CONTROL

With an estimate of the process mean, a target value, and a rule for making
adjustments, one can consider automatic process control. The equipment used
in lithography utilizes a considerable amount of automatic process control.
Consider, for example, a hotplate used for resist processing. The temperature of
the hotplate is measured with an appropriate temperature probe, and additional
power is provided to the heating elements if the temperature is too low, or the

power is reduced if the temperature is too hot. There are many other process
factors which are controlled automatically by the equipment, including
exposure dose, focus, overlay, coater spin speeds, and exhausts.
It is common practice in some industries to adjust the process based on
critical product characteristics. For example, if one is making a batch of resist
developer, more water would be added if the normality were too high, or more
base might be added if the normality were too low. This adjustment of
developer normality is possible because the critical parameter — normality —
can be measured and changed during processing. On the other hand, one cannot
directly measure parameters that are critical to lithography, such as linewidth
and overlay, and make corrections during processing. The necessary
measurements require highly specialized equipment. One cannot perform a
scanning electron microscope (SEM) measurement of linewidths while wafers
are being exposed or developed.
It has been proven that it is possible to monitor the lithography process
indirectly, and thereby improve process control. The most notable technique for
doing this monitoring is develop end-point detection. In develop end-point
detection the thickness of the resist is monitored during develop. This can be
accomplished using optical thin film interference. This technique was
developed originally for characterizing resist development in a laboratory, ó and
was later extended to real-time application. 7 ' 8 In this method, light is incident
on the resist coated substrate, and the intensity of reflected light is detected.
During development of the resist the optical signal will vary because of thin
film interference effects (Fig. 7.4). Eventually the resist is developed
completely to the substrate in large open areas (point C). This signals the
transition of one part of the development process to another. If the develop time
between point C and the desired end of the process is fairly constant, then most
of the variations in the process can be captured in the variations in the amount
of time for the develop cycle to reach point C. A system that can determine the
time when the resist has just been developed in the large open areas (point C)
can be used to control the process automatically.
In some ways this approach is very attractive. Process control can be
accomplished by equipment in real time. Numerous control charts do not need
to be maintained, and people are not needed to review the control charts and
adjust processes. However, there is no such thing as a free lunch. Develop end-
point controllers require some teaching, as every process behaves somewhat
different from the others. Some layers, such as contacts, often provide
interference signals that are too weak for develop endpoint detection to be
useable. Moreover, develop endpoint controllers are not available from the
same equipment suppliers who make the rest of the resist processing
equipment, and integration is often problematic. This approach is also difficult
to apply to multiple-puddle processes, which are often implemented because of
overriding concerns about yield. Finally, many resists for DUV lithography
develop very quickly, taking only a few seconds. This is too fast for effective
develop end-point control. For these reasons, develop endpoint detection has

never been used extensively, although it has been demonstrated to improve

linewidth control. 9 Develop monitors have not found widespread acceptance for
in-line use, but they have been essential tools for determining parameters for
image modeling 10 and in-line" and off-line 12 process monitoring.

Figure 7.4. The develop end-point detection system.

Recently, the use of latent images for process monitoring has been
explored. 13 A latent image is the pattern in the resist film, prior to development,
that results from the difference in optical constants or film thickness between
exposed and unexposed resist. For example, the latent image can be monitored
during post-exposure bake of chemically-amplified resists. Feedback could be
applied to the post-exposure bake time.'4' 15,16 With i -line resists the information
gathered while measuring the latent image can be fed forward to the develop
step. So far, these techniques have remained primarily in the laboratory, 17,1 ,19,
20, 21,22 for a number of reasons. A significant complication of these techniques
is the non-equivalence between resist development or post-exposure bake and
other factors which affect linewidth, when many different feature types are
considered. For example, if there is defocus, the focus should be corrected,
rather than having the develop or bake time varied.
Among those who study process control, the nature of automatic process
control and its relationship to statistical process control have become subjects
of considerable interest.23, 24, 25 Automatic process control can be applied when
the state of a process can be measured, and the process can be regulated through

a control algorithm on the basis of these measurements. The other extreme

situation is exemplified by Deming's funnel experiment, in which the process is
completely random and not amenable to any control. Most processes involve
some potential for regulation as well as a significant level of statistical noise
that must be addressed.
Ultimately, the problem for applying automatic process control to
lithography remains one of metrology. Obtaining good data on which to make
decisions is critical. As will be seen in the next chapter, when linewidths and
overlay are measured in nanometers, this is difficult even under conditions
optimized for the purpose of making measurements.

1 F. E. Grubbs, "An Optimum Procedure for Setting Machines or Adjusting

Processes," Journal of Quality Technology, Vol. 15(4), pp. 186 — 189 (1983).

2S. V. Crowder, "Design of Exponentially Weighted Moving Average

Schemes," Journal of Quality Technology, Vol. 21, pp. 155 —162 (1989).

3 J. S. Hunter, "The Exponentially Weighted Moving Average," Journal of

Quality Technology, Vol. 18, pp. 203 — 210 (1986).

4J. M. Lucas and M. S. Saccucci, "Exponentially Weighted Moving Average

Control Schemes: Properties and Enhancements," Technometrics, Vol. 32, pp.
1 — 12 (1990).

5 G. E. Flores, W. W. Flack, S. Avlakeotes, and M. Merrill, "Monitoring and

Diagnostic Techniques for Control of Overlay in Steppers," SPIE Vol. 2439,
pp. 40-60 (1995).

6 K. L. Konnerth and F. H. Dill, "In-situ Measurement of Dielectric Thickness

During Etching or Developing Processes," IEEE Trans. Electron Devices, Vol.
ED-22, pp. 452-456 (1975).

L. J. Lauchlan, K. M. Sautter, and T. Batchelder, "Automatic Process Control

for VLSI Linewidth," Solid State Technology, pp. 333 — 337 (April, 1985).

8L. J. Lauchlan, K. Sautter, T. Batchelder, and J. Irwin, "In-Line Automatic

Photoresist Process Control," SPIE Vol. 539, pp. 227 — 233 (1985).

9C. Nygren, J. Daggett, and J. Grambow, "The Use of Develop End Point
Detection to Eliminate Photolithography Process Variation," Proceedings of the
KTI Microlithography Seminar, pp. 315 — 348 (1990).

10
P. D. Flanner III, "Improved Methods of Photoresist Development
Characterization," Proceedings of the KTI Microelectronics Seminar, pp.
231 — 238 (1987).

11 K. M. Sautter, M. Ha, and T. Batchelder, "Development Process Control and

Optimization Utilizing an End Point Monitor," Proceedings of the KTI

Microelectronics Seminar, pp. 99 — 112 (1988).
12
J. A. Bruce and B. J. Lin, "Determination of Exposure Dose by Photoresist
Development Rate," Proceedings of the KTI Microelectronics Seminar, pp.
1-11(1987).

13 M. L. Miller and D. A. Mellichamp, "Development of an End-Point

Detection Procedure for the Post-Expose Bake Process," SPIE Vol. 2439, pp.
78-88 (1995).
14
• Zaidi, S. L. Prins, J. R. McNeil, and S. S. H. Naqvi, "Metrology Sensors
for Advanced Resists," SPIE Vol. 2196, pp. 341— 351 (1994).

15 J. L. Sturtevant, S. J. Holmes, T. G. Van Kessel, P. D. Hobbs, J. C. Shaw,

and R. R. Jackson, "Postexposure Bake as a Process-Control Parameter for

Chemically Amplified Photoresist," SPIE Vol. 1926, pp 106— 114 (1993).
16
J. L. Sturtevant, S. Holmes, T. G. Van Kessel, M. Miller, and D. MelIichamp,
"Use of Scatterometric Latent Image Detector in Closed Loop Feedback
Control of Linewidth," SPIE Vol. 2196, pp. 352-359 (1994).
17
K. C. Hickman, S. M. Gaspar, K. P. Bishop, S. S. H. Haqvi, J. R. McNeil, G.
D. Tipton, B. R. Stallard, and B. L. Draper, "Use of Diffracted Light from
Latent Images to Improve Lithography Control," SPIE Vol. 1464, pp.
245 — 251 (1991).

T. E. Adams, "Applications of Latent Image Metrology in

18
Microlithography," SPIE Vol. 1464, pp. 294 — 312 (1991).

19 S. L. Prins and J. R. McNeil, "Scatterometric Sensor for PEB Process

Control," SPIE Vol. 2725, pp. 710-719 (1996).

20 L. M. Milner, K. C. Hickman, S. M. Gaspar, K. P. Bishop, S. S. H. Haqvi,

J. R. McNeil, M. Blain, and B. L. Draper, "Latent Image Exposure Monitor

Using Scatterometry," SPIE Vol. 1673, pp. 274 — 283 (1992).

21 M. L. Miller and D. A. Mellichamp, "Development of an End-Point

Detection Procedure for the Post-Exposure Bake Process," SPIE Vol. 2439, pp.
78-88 (1995).
22
T. Koizumi, T. Matsuo, M. Endo and M. Sasago, "CD Control using Latent
Image for Lithography," SPIE Vol. 2439, pp. 418 — 426 (1995).
23
D. C. Montgomery, J. B. Keats, G. C. Runger, and W. S. Messina,
"Integrating Statistical Process Control and Engineering Process Control,"
Journal of Quality Technology, Vol. 26(2) pp. 79 — 87 (1994).
24
The April, 1997, Vol. 29(2) of the Journal of Quality Technology has a
discussion on statistically-based process monitoring and control, with
contributions from several authors and extensive references.
25
G. Box and A. Luceno, Statistical Control by Monitoring and Feedback
Adjustment, John Wiley & Sons, New York (1997).

8.1 THE NEED FOR UNDERSTANDING THE MEASUREMENT PROCESS: DEFECT

DETECTION

The information upon which operators, technicians and engineers make

decisions consists of measured data. Bad data usually lead to bad decisions. For
this reason quality control professionals have always focused on metrology.
The problem is particularly significant for lithographers, where the extremely
small features of state-of-the art processes have pushed measurement
capabilities to their limits, and, some might say, beyond. Some of the metrology
issues and challenges faced by lithographers will be discussed in this chapter.
Defect detection will be used as an example to illustrate the importance of
understanding the measurement process in order to properly use measurement
tools. Consider a laser light scattering system used to detect defects (Fig. 6.2).
A particle is detected because it scatters light. However, small particles scatter
light primarily in the forward direction (Fig. 8.1). When defects are sitting on,
or embedded in, thin films, optical interference will affect the strength of the
signal produced by the scattered light, because the forward scattered light is
detected only after it has reflected from the substrate (Fig. 8.2). Since particle
size is determined by the strength of the scattered light signal in these systems,
the "measured" particle size will depend upon the films on the substrate, which
can modulate the reflectance (Fig. 8.3). A particle will scatter a certain amount
of light, but the films on the substrate will affect the amount of light that
actually reaches the detector. Consider, for example, the "measured" particle
size in a controlled experiment. Polystyrene latex spheres were used to calibrate
a system which used a HeNe laser (wavelength = 632.8 nm) on bare silicon.
The resulting measured particle sizes for spheres on bare silicon and substrates
with different thickness of silicon nitride are shown in Table 8.1. 1 Because of
the thin film interference effect, the apparent particle size is generally smaller
for the nitride films, since a reduced signal can result from either small particles
or a low-reflectance substrate. In some instances, the signal was too small to
provide a measurement. The opposite effect occurs on an aluminum substrate,
where the highly reflective surface appears to enhance the particle size (Table
8.2). Interpretation of defects that consist of embedded particles or voids must
be made with caution, as well.
It has just been shown that considerable misinterpretation of the data would
occur if one did not take thin film optical interference effects into account. This
is of particular importance to lithographers, who are usually concerned about
defects in photoresist films or anti-reflection coatings. Correct decisions require
an understanding of the measurement process. This has been illustrated here by

141

examining defect detection, but it is generally good engineering practice to

understand all metrology processes in order to be able to properly interpret
acquired data. From this perspective, the measurement of linewidths by
scanning electron microscopes is discussed in the next section.
particle

Incident light

Back scattered Forward scattered

light light

Figure 8.1. The angular distribution of light scattered from a spherical particle.

Incident
light

defect
thin film
substrate

Figure 8.2. Geometry for detection of defects on thin films. Light is scattered
from the defect and then reflected from the substrate. It is the reflected light
which is detected. Because of interference effects the reflected signal will vary
with film thickness.
Reflectance from nitride/pad oxide on silicon
0.4

0.3

8 0.2
Cy
r

ßßi
0.1

0 1000.0 2000.0 3000.0 4000.0 5000.0

Nitride thickness (A)

Figure 8.3. Because of interference effects between incident and reflected light,
signal from particles on thin films will vary with film thickness. The calculated
reflectance is for normally incident light, also collected near - normal, with a
wavelength of 632.8 nm (HeNe laser).

Nitride thickness
Latex sphere diameter (.tm) 0A 825 A 1400
0.36 0.26 0.19
0.50 0.53 0.55
0.62 0.59 0.61
0.76 0.71 0.73
0.90 0.74 0.084 0.83
1.09 0.94 0.13 0.96
2.02 2.1 0.21 1.50
5.00 6.7 0.75 3.23

Table 8.1. Measured particle size for polystyrene latex spheres on nitride films of
different thicknesses.

Latex sphere diameter Scattering cross section on Scattering cross section

(pm) silicon on aluminum
0.36 0.26 0.78
0.62 0.59 1.54
1.09 0.94 2.66
2.02 2.1 6.33
5.00 6.7 20.46

Table 8.2. Measured particle size for polystyrene latex spheres on silicon and
aluminum surfaces.'

8.2 LINEWIDTH MEASUREMENT USING SCANNING ELECTRON MICROSCOPES

The most common tool for measuring linewidths is the scanning electron
microscope (SEM). Low voltage SEMs are capable of measuring resist features
in-line. As shown in the prior section, understanding the measurement process
is essential for applying a measurement technique correctly, and this is
particularly true of linewidth measurements. In the scanning electron
microscope, electron beams are scanned across patterns on wafers. The voltage
of the electron beams ranges between a few hundred volts to tens of thousands
of volts. For measuring resist features, a typical range is 400 - 1000 volts. The
incident beam is scattered, both elastically and inelastically, producing
secondary and back-scattered electrons. 2' 3 By commonly accepted definition,
secondary electrons are those with energy less than 50 eV, while the back-
scattered electrons are those with energies closer to the energy of the incident
beam. The secondary electrons created by inelastic scattering or the elastically
back-scattered electrons are detected synchronously with the scan of the
incident beam (Fig. 8.4). Because the number and direction of the scattered
electrons depends upon the material composition and topography of the features

over which the incident beam is scanned, the intensity of the detected signal
varies so that an image can be created.

Figure 8.4. Cross-sectional view of a scanning electron microscope.

As the incident beam is scanned across a line, the detected signal varies.
Algorithms are required to relate the edges of the line to points on the intensity
versus position curve, and these algorithms do not always provide complete
accuracy. The signal at the edge of a feature is shown in Fig. 8.5. A
mathematical algorithm is required to identify a point on the signal profile to
represent the edge of the line. A number of edge detection methods are used. 4
The simplest algorithm uses a threshold value for identifying this point:

Threshold level = (1— P)x minimum + (P)x maximum, (8.1)

where P is a value between 0 and 1, typically in the neighborhood of 0.5. Other

methods include maximum slopes, inflection points, linear approximations,
minima or maxima.
Secondary and back-scattered electron emissions are proportional to slopes
of features, 5 typically varying as 1/cos(6) (Fig. 8.6). Steep slopes appear
"bright" in scanning electron microscopes. Consequently, the signal depends on
the slope of the feature, something which can change easily with stepper
defocus and other process variations for resist features, and variations in etch
processes can change the slopes of etched features. By measuring linewidths
and slopes from cross sections, and comparing these results to in-line SEM
measurements, a careful study found that measurements of 180 — 250 nm lines
could vary by as much as 10 nm per degree change in line-edge slope 6
(Fig. 8.7).

oq
Maximum
slope -_
Threshold
i
Maximum

Point of
Baseline
inflection
Minimum

Position along the scan

Figure 8.5. The electron signal corresponding to one edge of a feature being
measured by a scanning electron microscope. Several attributes of the feature can
potentially be used to determine the location of the edge of the feature.

70
60
50
40
1/cos 0 30
20
10
0
0 20 40 60 80 100
Theta (degrees)
cross section of feature
being measured

Figure 8.6. Slope of a feature being measured, and the secondary electron
production as a function of this slope.
280
I Sidewall angle
88 degrees
260
90:
J 240
220

201 80
200 220 240 260 280
CD from SEM x-section (nm)
Figure 8.7. Top-down SEM measurements versus measurements from cross
sections of the same etched polysilicon features (Ref. 6). Measurement error bars
are not shown.

Another problem associated with SEM linewidth measurements—sample

charging—is a particularly significant problem for measuring features
comprised of photoresist. Consider an electron beam focused onto a
homogeneous substrate. At low incident beam voltages more electrons are
introduced into the sample than are emitted in the form of secondary and back-
scattered electrons. As the beam voltage is increased, this balance changes, and
samples acquire a net positive charge during imaging' (Fig. 8.8). At yet higher
voltages the net charge in the sample reverses yet again. Imaging and
measurement is usually most stable at the point where the samples remain
electrically neutral. Point E 2 is the more stable of the two, and scanning
electron microsopes are typically operated close to this voltage. For conductive
materials this voltage ranges from 2 – 4 kV, while it is only 0.5 – 1.0 kV for
non-conducting materials such as photoresist. 203 Since imaging is typically
superior at higher voltages, SEM metrology for lithography is at a distinct
disadvantage.

1.0

# emitted electrons
# incident electrons

E1 E2

log (incident energy)

Figure 8.8. At incident energies El and E2 the sample will be electrically neutral
while it will charge for other energies. The values of El and E2 are material
dependent.

The charging of features measured in lithography is a complex

phenomenon, and the idea of a single voltage at which there is charge neutrality
is inadequate. The incident electrons and emitted secondary and back-scattered
electrons are not the only sources of electrical charge that need to be considered
in order to understand sample charging. Samples in scanning electron
microscopes are typically grounded, and there will be a current between the

surface of the sample and ground. This current will be proportional to the
voltage difference between the sample surface and ground, and it will depend
upon the resistivity of the materials that comprise the sample. Highly
conductive materials will not charge, because there will be a steady current to
ground. This charge balance can be written as:

Esurrface
I B = (S+ 11)I B + (8.2)
R

where I B is the incident beam current, 8 is the secondary electron yield, 71 is

the back-scattered electron yield, Esurface is the surface potential, and R is the
effective resistance to ground. Consequently, the surface potential is given by 8 :

Esurface =I BR(1 -5-1) (8.3)

The surface potential is seen to be proportional to beam current, once the

current to ground is taken into consideration. Conductivity to ground can be
complex in insulating materials, where the excitation of electrons to the
conduction band will be a function of beam energy and current. Complicating
the situation even more is the ability of insulators to trap charge. The time it
takes to discharge the trapped charges as a consequence of thermal relaxation is
on the order of minutes.
Unfortunately, this picture is still incomplete for the types of measurements
that occur in lithography applications and typically involve features patterned
into photoresist. As noted previously, secondary electron emission depends on
the angle between the incident beam and the plane of the surface being probed.
Consequently, the electron beam energy at which there is charge neutrality will
vary between the top and sides of particular features, because the rate of
emission of secondary and back-scattered electrons from the tops of features
will differ from the rate from the sides. The photoresist is nearly always
developed to the substrate in exposed areas, and the substrate will have a
different value for E2 than the photoresist.
Moreover, the surface potential may be more important than overall charge
neutrality. It is the electrical potential at the surface of the features, relative to
the potential at the detector, that determines the flow of secondary electrons
from the object being measured to the detector. Overall charge neutrality is
important for electrons far from the object, but the distribution of electrons
within and near the object will determine the electrical potential at the object's
surface. The issue of sample charging remains a topic of current research.
Because photoresist is not electrically conducting it is difficult to dissipate
charge once accumulated. Charging also affects the linewidth measurement. A
negatively charged line, which will occur for higher voltages, will deflect the
electron beam and result in a narrower measurement 9 (Fig. 8.9).

Scanning beam electrons

Electrically charged photoresist

Figure 8.9. Scanning beam electrons are deflected away from the negatively
charged photoresist. As a consequence, the line measures narrower than it
actually is.

Once secondary electrons are generated they need to escape from the
sample and hit a detector in order to create a signal. SEM measurements of
contact holes have been difficult for this reason, since it is difficult for low-
energy secondary electrons to emerge from positively charged holes. 10 For this
reason, detection of back-scattered electrons is sometimes used for metrology
with lithography applications. Because back-scattered electron emission peaks
at or near the direction of the incident beam, detectors are sometimes placed
within the lens. 11
The conversion of a secondary or back-scattered electron intensity into a
linewidth number is not automatic. It may be argued that measurement
precision is the critical requirement, and that absolute accuracy is less
important. This is true to a certain extent, but facilities with more than one
SEM, or those that need to transfer processes to or from others, will require
standards to match SEMs. Consequently, some level of calibration is usually
required, and this gauge-setting is generally material-dependent. Algorithms for
calibrations on resist features will differ from those for polysilicon, and will
depend on the thickness and slopes of the resist features. Consequently, there
are very few standards available from organizations such as National Institute
of Standards & Technology (NIST) that can be used directly in lithography
applications. Current NIST linewidth standards, SRM 473, 475, and 476, are
chromium-on-quartz, intended primarily for measuring features on
photomasks. 12 SRM 473 has linewidths down to 0.5 pm. There have been

efforts to develop standards designed specifically for use in SEMs. 13 ' 14 An

interim standard, SRM 8090, is available from NIST, consisting of
palladium/titanium lines on silicon, and it has linewidths down to 0.2 gm. It is
intended that this interim standard will be superseded eventually by SRM 2090.
Since SEMs will "measure" different linewidths for features made from
different materials or thickness, these standards are useful only for ensuring
measurements on photomasks or controlling such parameters as the
magnification of the SEM. However, the magnification can be verified by
measuring the pitch of grating patterns,'s' 16 which will give measurements
independent of materials and measurement algorithms. Although commercially
available SEM-based linewidth measurement systems may not produce
measurements that are accurate (i.e., give the "real" linewidth), they may be
precise and repeatable. The difference between accuracy and precision is
illustrated in Fig. 8.10.

O O• t

Precise, but Accurate, but Precise

not accurate not precise and
accurate

Figure 8.10. Shots at a target, illustrating precision and accuracy.

One problem preventing the creation of good linewidth standards for

lithography is carbon deposition during SEM measurement.' 7 While the
electron beam is scanning the sample, carbon is being deposited. As one might
expect, this is a particularly significant problem for photoresist samples.
Repeated measurements lead to changes in the measured linewidth, even
though the pitch remains unchanged. 18 One approach to dealing with the change
in measured linewidth due to carbon deposition on the standard is to
predetermine the magnitude of deposition-induced linewidth change per
measurement, and subtract that from the monitor values. 19 SPC can then be
applied to the residuals.
While an SEM can reproduce the linewidth on a particular standard artifact,
there may be some concern that linewidth measurements of other materials,
such as photoresist, may not be properly calibrated. To this end, atomic force
microscopes can be used to measure photoresist linewidths for the purpose of
calibration. 20 Atomic force microscopes are purely mechanical and will not
have the type of material dependence that scanning electron microscopes have.

As with any measurement system, the proper interpretation of the numbers

produced by atomic force microscopes requires an understanding of how they
measure. For example, because of probe tip size, atomic force microscopes will
not provide reliable measurements of structures with very tight pitches.
Nevertheless, the atomic force microscope is another useful metrology tool.
Calibration usually requires measurements of multiple linewidths. If r is the
"real" linewidth, then the measurement linewidth y is given by:

y = ar + b (8.4)

to first order, where a and b are constants. Measurement of more than one
linewidth is required to determine the two calibration parameters, as well as
reduce errors from measurement noise.
The SEM used for metrology must be properly focused and stigmated in
order to obtain good measurements. In the past this was accomplished by using
a sample of very small grains, such as sputtered gold or grainy silicon, and
adjustments were made until the image appeared best to the operator. Recent
work has shown that the subjective element associated with this approach can
be overcome by taking a two-dimensional Fourier transform of the image. 21
Defocus will involve a decrease in the magnitudes of the higher frequency
components, and astigmatism will show asymmetries between axes.
8.3 ELECTRICAL LINEWIDTH MEASUREMENT

Linewidths can be measured electrically as a supplement to SEM

measurements. This method has the advantage of speed; considerable data can
be collected in a small amount of time, once wafers have been patterned and
etched. Off-line stepper lens setup and characterization often requires that large
numbers of linewidths be measured, and the delays incurred because wafers
must be etched is tolerable. For example, the determination of best focus and
field tilt requires that linewidths be measured at many points in the field, with
features of at least two orientations, and at several focus settings. It is useful to
measure linewidths in multiple fields in order to average out noise from factors
not related to the lens and optical column tilt. It is not practical to measure this
many linewidths with a scanning electron microscope, even an automated one.
Electrical linewidth metrology is a practical method for this type of setup. 22
Linewidths of electrically conducting material, such as polysilicon, suicide, or
metal, can be measured using structures such as those shown in Fig. 8.11 23 The
electrical linewidth Wis related to two measured resistances:

R
W=—'-L. (8.5)
b

Rs is the sheet resistance and it is determined as follows. A current I is forced

between pads 3 and 4, and a voltage V measured between pads 2 and 5.
The current is reversed, and a voltage V25 is measured. The current I is then
forced alternately between pads 2 and 3, and voltages V 45 and V45 are
measured. The sheet resistance is given by:

R —IV'+5I+IVZSI+V45I+IV45
(8.6)
S 2 41

Rb is determined by forcing current I b alternately between pads 1 and 3 and

measuring the voltages between pads 5 and 6.

R b = Iv56I+IV6I
2 1(8.7)
b

To avoid end effects, typically W « L.

^ L ^

Figure 8.11. The resistor test structure for measuring linewidths electrically.

Electrically measured linewidths, as determined from Eqs. 8.5 — 8.7, tend to

disagree in absolute value with measurements from other metrology methods,
such as SEMs and atomic force microscopes, and are generally smaller. This is
often thought to be the result of non-conducting material on the sides of the
pattern, following etch. From the perspective of lithography, there is also an
etch bias between the resist and final dimensions which is relevant, and non-
uniformities in the etch may introduce across-wafer variations, which need to

be taken into account. Despite these complications, electrical linewidth

measurements are very useful for characterizing linewidth control across lenses,
because the speed of measurement enables a lot of data to be collected.
The structure in Fig. 8.11 can be used for measuring the widths of isolated
lines. Additional structures can be added to provide measurements of
linewidths of grouped structures, for any pitch that is desired. A method has
even been developed for measuring contact sizes using electrical resistance
measurements, 24 by measuring the effective electrical linewidth W of a structure
with holes (Fig. 8.12), compared to the electrical linewidth WW of of an identical
reference structure without holes. The diameter d of the holes is given by:

i
L _1 L
d= WYef 1+48 (8.8)
121r NWW ef I 2,

where N is the number of contacts in the structure being measured.

Electrical metrology can also be used for monitoring SEM stability. 25 Even
though electrical linewidth measurements do not provide absolute
measurements, a correlation between electrical and SEM measurements
provides a method for checking the stability of the SEMs, through comparison
with electrical measurements.

Figure 8.12. Configuration for measuring contact diameter electrically.

8.4 MEASUREMENT ERROR BUDGETS

Following measurement there are four possible outcomes, indicated in Table

8.1. Alpha errors, in which good products are measured as bad products, result
in unnecessary rework, while beta errors cause bad products to be accepted as
good and will result in yield loss. Processes that are barely capable will have a

high level of alpha errors, because of measurement imprecision. This is a direct

cost of inadequate measurement capability, coupled with marginal process
capability. 26 These costs will occur whenever process capability is low, and
measurements are required to disposition material. From this perspective, good
metrology can result in cost reductions.

Actual product quality Measured product quality

1, Good Bad
Good Correct decision Alpha error
Bad Beta error Correct decision

Table 8.1. The types of outcomes from measurements of products.

The current level of repeatability for SEMs used for linewidth measurement
is 5 nm (36). 27 The level of adequacy for this value needs to be determined.
Consider metrology metrics used in other industries. For example, the
Automotive Industry Action Group defines a measurement method with good
precision as one in which:

5.156 meas < 0.3(USL — USL) , ( 8.9)

where ó meas is the standard deviation of the measurement, and USL and LSL
are the upper and lower specification limits for the process, respectively. 28
Similarly, a measurement method with very good precision is one with

5.156meas < 0.1(USL — USL). (8.10)

For a 250 nm lithography process requiring +/-10% linewidth control, current

SEMs have good, but not very good, precision. Recent analysis has shown that
measurement systems provide correct decisions according to 29 :

ómeas
arctan LSL Z—0.5 USL—µ z
—0.51µ—
1_ process e6 P.aes.,
correct J+ e l 6P.aess

( 8.11 )

where ó process is the process standard deviation and µ is the process mean. For
example, consider a 250 nm lithography process with 36 process = 23 nm
linewidth control, requirements of +/- 25 nm (10%) linewidth control about the
250 nm target value, and a measurement system with 5 nm (3a' meas ) precision
and no mean offset. In this example, the correct decision will be made 96% of

the time. This means that 4% of processed material is either reworked or has
degraded yield, because of the measurement imprecision. When processes are
marginal, as often occurs in advanced semiconductor processing, measurement
systems with very good precision (Eq. 8.10) are required to avoid high levels of
unnecessary rework and degraded yield.

8.5 MEASUREMENT OF OVERLAY

Overlay is usually measured using optical linewidth measurement systems.

Consider the structure shown in Fig. 8.13. The linewidths X I and X 2 can be
used to calculate the overlay in the X direction:

AX= 2(X 2 —X l ) (8.12)

The features shown in Fig. 8.13 are sufficiently large that optical linewidth
measurement systems can be used, with their throughput advantage over SEMs.
Sufficient data can be collected automatically in order to perform the overlay
modeling discussed in Chapter 5.

Figure 8.13. A structure for measuring overlay. The inner box is part of the
substrate pattern, and the outer box is created in the overlaying photoresist layer.

Overlay measurement can introduce apparent overlay errors. The most

common of these errors will be discussed in this section. This error is called
tool-induced shift (TIS), and it results from asymmetries in the optics of the
overlay measurement system. Because the materials of Layer 1 and Layer 2 in
Fig. 8.13 are not the same, the measurements of X 1 and X 2 are not equivalent
if there are asymmetries in the measurement equipment. Types of asymmetries
that can occur are tilted or decentered lenses, nonuniform illumination, lens
aberrations, and nonuniform detector response. 30 The presence of TIS is easily
verified by rotating wafers from 0° to 180 3132 In the absence of TIS
0.

t^X (o)—AX (iso) , (8.13)

where OX (o) is the measurement for the unrotated wafer and &X (180 ) is the
measurement for the wafer rotation 180 0 . A measure of TIS is therefore

TIS = AX (0) + 0X(180), (8.14)

which is ideally zero. The TIS value can be measured when overlay
measurement programs are first established for particular process layers, and
automatic compensation can be made to reduce the TIS error. 33 Asymmetries
in alignment measurement can also be induced when features are too close to
each other (X i or X 2 is too small), relative to the resolving power of the optics
in the measurement tool. 34 The resolution of optical tools using visible light is
typically in the range of 0.8 — 1.0 tm, while systems using ultraviolet light may
have somewhat lower resolution.
Wafer processing can induce asymmetries in overlay measurement marks.
Consider situations in which metal is sputtered onto wafers. If the deposition is
not collimated, metal may build up preferentially on one side of the overlay
measurement marks near the edges of the wafers (Fig. 5.14). Measurements of
overlay will be shifted relative to the overlay of the critical features. Errors in
overlay measurement caused by wafer processing are referred to as wafer-
induced shift (WIS). For the example illustrated in Fig. 5.14, the overlay error
will be in opposite directions on opposite sides of the wafer and will appear as a
wafer scaling error. 35 To the extent that asymmetries in step coverage are
repeatable, wafer-induced-shifts can be calibrated by comparing overlay
measurements before and after etch. 36
Chemical-mechanical polishing (CMP), which causes problems for correct
acquisition of alignment targets (Chapter 5), also creates difficulties for overlay
measurement. Overlay measurement structures need to be optimized to
minimize this effect. The bar or frame structures shown in Fig. 8.14 typically
lead to less of a problem than the box-in-box structures, though the exact
dimensions need to be determined for individual processes.
As discussed in Chapter 5, intrafield registration depends upon the
illumination conditions and the feature type. Lines and spaces may have
different registration at the different illumination settings. This will cause a
problem for metrology if the critical features are one type, such as contact
holes, while other types of features, such as lines, are used for overlay
measurement. 37 This type of subtle difference will be significant as overlay
requirements approach 50 nm or less.

Figure 8.14. Various structures for measuring overlay.

' L. Galbraith and A. Neukermans, "Contamination Detection on

Semiconductor Wafers," SPIE Vol. 774, pp. 13 — 20 (1987).

2K. M. Monahan, M. Davidson, Z. Grycz, R. Krieger, B. Sheumaker, and

R. Zmrzli, "Low-Loss Electron Imaging and its Application to Critical
Dimension Metrology," SPIE Vol. 2196, pp. 138 —144 (1994).

3L. Reimer, Image Formation in Low-voltage Scanning Electron Microscopy,

SPIE Press, Vol. TT12 (1993).

4R. R. Hershey and M. B. Weller, "Nonlinearity in Scanning Electron

Microscope Critical Dimension Measurements Introduced by the Edge
Detection Algorithm," SPIE Vol. 1926, pp. 287 — 294 (1993).

5J. I. Goldstein, D. E. Newbury, P. Echlin, D. C. Joy, C. Fiori, and E. Lifshin,

Scanning Electron Microscopy and X-Ray Microanalysis, 2nd Edition, Plenum
Press, New York (1984).

6J. Finders, K. Ronse, L. Van den Hove, V. Van Driessche, and P. Tzviatkov,
"Impact of SEM Accuracy on the CD-control During Gate Patterning Process
of 0.25 im Generations," Proceedings of the Olin Microlithography Seminar,
pp. 17-30(1997).

K. M. Monahan, J. P. H. Benschop, and T. A. Harris, "Charging Effects in

Low-voltage SEM Metrology," SPIE Vol. 1464, pp. 2-9 (1991).

8 D. C. Joy and C. S. Joy, "Low Voltage Scanning Electron Microscopy,"

Micron, Vol. 27, No. 3 — 4, pp. 247 — 263 (1996).

9M. Davidson and N. T. Sullivan, "An Investigation of the Effects of Charging

in SEM Based CD Metrology," SPIE Vol. 3050, pp. 226 — 242 (1997).
10
C. M. Cork, P. Canestrari, P. DeNatale, and M. Vasconi, "Near and Sub-Half
Micron Geometry SEM Metrology Requirements for Good Process Control,"
SPIE Vol. 2439, pp. 106— 113 (1995).

11 S. R. Rogers, "New CD -SEM Technology for 0.25 pm Production," SPIE

Vol. 2439, pp. 353 — 362 (1995).

12 NIST SRM standards are available from the Office of Standard Reference

Materials, NIST, EM 205, Gaithersburg, MD 20899. Their phone number is

(301) 975-6776. Information can be obtained from the NIST web site,
www.nist.gov.
13
M. T. Postek, A. E. Vladar, S. Jones, and W. J. Keery, "Report on the NIST
Low Accelerating Voltage SEM Magnification Standard Interlaboratory
Study," SPIE Vol. 1926, pp. 268 — 286 (1993).

14 M. T. Postek, "Scanning Electron Microscope-based Metrological Electron

Microscope System and New Prototype Scanning Electron Microscope

Magnification Standard," Scanning Microscopy, 3(4), pp. 1087 — 1099 (1989).

15 Y. Nakayama and K. Toyoda, "New Submicron Dimension Reference for

Electron-Beam Metrology System," SPIE Vol. 2196, pp. 74 — 84 (1994).

16
B. L. Newell, M. T. Postek, and J. P. van der Ziel, "Performance of the
Protoype NIST SRM 2090A SEM Magnification Standard in a Low-
Accelerating Voltage SEM," SPIE Vol. 2439, pp. 383 — 390 (1995).

17 T. W. Reilly, "Metrology Algorithms for Machine Matching in Different CD

SEM Configurations," SPIE Vol. 1673, pp. 48 — 55 (1992).

18K. Phan, J. Nistler, and B. Singh, "Metrology Issues Associated with

Submicron Linewidths," SPIE Vol. 1464, pp. 424 — 437 (1991).

E. E. Chain, M. G. Ridens, and J. P. Annand, "SPC Qualification Strategy for

CD Metrology," SPIE Vol. 2876, pp. 218 — 224 (1996).

20D. A. Chernoff, "Atomic Force Microscope (AFM) Analysis of Photoresist

Test Structures for use in SEM as In-House Linewidth Standards," SPIE Vol.
2439, pp. 392 — 400 (1995).

21
M. T. Postek, A. E. Vladar, and M. P. Davidson, "Fourier Transform
Feedback Tool for Scanning Electron Microscopes Used in Semiconductor
Metrology," SPIE Vol. 3050, pp. 68-79 (1997).
22
L. J. Zych, G. Spadini, T. F. Hassan, and B. A. Arden, "Electrical Methods
for Precision Stepper Column Optimization," SPIE Vol. 633, pp. 98 — 105
(1986).
23
L. W. Linholm, R. A. Allen, and M. W. Cresswell, "Microelectronic Test
Structures for Feature Placement and Electrical Linewidth Metrology," in
Handbook of Critical Dimension Metrology and Process Control, K. M.
Monahan, ed., SPIE Press, Bellingham, WA (1993).
24
B. J. Lin, J. A. Underhill, D. Sundling, and B. Peck, `Electrical Measurement
of Submicrometer Contact Holes," SPIE Vol. 921, pp. 164 — 169 (1988).
25
E. E. Chain and M. Griswold, "In-Line Electrical Probe for CD Metrology,"
SPIE Vol 2876, pp. 135 —146 (1996).
26
S. M. Kudva and R. W. Potter, "Cost Analysis and Risk Assessment for
Metrology Applications," SPIE Vol. 1673, pp. 2 — 13 (1992).

27R. R. Hershey and R. C. Elliott, "Procedure for Evaluating Measurement

System Performance: A Case Study," SPIE Vol. 2439, pp. 363 — 373 (1995).

28 Measurement Systems Analysis Reference Manual, Automotive Industry

Action Group, Detroit, MI (1990).
29
J. Engel and B. de Vries, "Evaluating a Well-Known Criterion for
Measurement Precision," J. Quality Technol., 29(4), pp. 469 — 476 (1997).
30
R. M. Silver, J. Potzick, and R. D. Larrabee, "Overlay Measurements and
Standards," SPIE Vol. 3429, pp. 262 — 272 (1995).

31D. J. Coleman, P. J. Larson, A. D. Lopata, W. A. Muth, and A. Starikov, "On

the Accuracy of Overlay Measurements: Tool and Mark Asymmetry Effects,"
SPIE Vol. 1261, pp. 139 — 161 (1990).
32
A. Starikov, D. J. Coleman, P. J. Larson, A. D. Lopata, and W. A. Muth,
"Accuracy of Overlay Measurements: Tool and Mark Asymmetry Effects,"
Optical Engineering, Vol. 31, pp. 1298 — 1310 (1992).

33
M. E. Preil, B. Plambeck, Y. Uziel, H. Zhou, and M. W. Melvin, "Improving
the Accuracy of Overlay Measurements through Reduction in Tool and Wafer
Induced Shifts," SPIE Vol. 3050, pp. 123 — 134 (1997).
34
N. Smith, G. Goelzer, M. Hanna, and P. Troccolo, "Minimizing Optical
Overlay Measurement Errors," SPIE Vol. 1926, pp. 450-462 (1993).
3s
J-S. Han, H. Kim, J-L. Nam, M-S Han, S-K. Lim, S. D. Yanowitz, N. P.
Smith, and A. M. C. Smout, "Effects of Illumination Wavelength on the
Accuracy of Optical Overlay Metrology," SPIE Vol. 3051, pp. 417 — 425
(1997).
36
Y. Tanaka, M. Kamiya, and N. Suzuki, "New Methodology of Optimizing
Optical Overlay Measurement," SPIE Vol. 1926, pp. 429-439 (1993).

37 T. Saito, H. Watanabe, and Y. Okuda, "Overlay Error of Fine Patterns by

Lens Aberration using Modified Illumination," SPIE Vol. 3051, pp. 687 — 696
(1997).

CONTROL OF OPERATIONS

Once a set of superior processes and control methodologies has been

developed, additional actions are required for these to be made operational.
Implementation ultimately involves a large number of operators, technicians,
engineers, and supervisors, all of whom need to apply techniques consistently
and must work in coordination with each other. It is a sad fact that excellent
technology has often failed to become a reality in manufacturing because of
inadequate implementation. Operational considerations, such as
documentation, are essential for successfully controlling processes, and
lithography engineers and managers must address these issues as well as purely
technical matters.
Unsuccessful implementation often results even when attention is paid to
operational matters, but inadequate methods are used. The first section of this
chapter covers the theory of self-control, which provides a framework for
engineers and managers to implement processes and procedures successfully. A
key implementation tool, documentation, is covered in the next section. This
chapter, and the text, is concluded with a discussion of ISO 9000, the quality
standard which identifies the most essential elements of any quality control
program.

9.1 SELF-CONTROL

It is essential to consider the human aspects of processes. The idea of

controlling the actions of people has acquired a bad reputation during the past
few decades, largely as a consequence of the implementation of specific
practices in particular industries. Most notable were the American automobile
manufacturers a few decades ago, who operated according to top-down
management principles and suffered from management-worker conflicts, high
manufacturing costs, and poor quality. The application of the scientific
management principles of Frederick Taylor were often blamed for the poor
competitive position of the American automobile industry. At the same time
that these criticisms were being made, some of the most acclaimed and
successful businesses were making use of strict work instructions, and well-
defined procedures were found to be a key component of some of the best
Japanese practices. For example, employee suggestions have been a key
Japanese business practice, and it was observed by Kenjiro Yamada, the
managing director of the Japan Human Relations Association, that the number
of suggestions was higher in situations in which work instructions were rigidly
fixed.' The control of operations is essential for the efficient manufacturing of

161

quality products, but failure can result if this control is implemented

incorrectly. This chapter will contain an overview of the critical elements for
controlling operations.
A key factor for a successful operation, where processes are controlled, but
the workers are nevertheless empowered to make necessary decisions, is the
extent to which operators have been provided with the means for controlling
operations by themselves. An excellent set of guidelines for establishing
operator self-control has been formulated by Joseph Juran. 2 According to
Juran, operator self-control exists when all people involved with a process have
been provided with the means for:
1)Knowing what they are supposed to do.
2) Knowing what they actually are doing.
3) Regulating the process if 1) and 2) disagree.
Managers and engineers who expect operators and technicians to make
decisions need to ensure that these three requirements are met, and further
discussion of these criteria is worthwhile.
Knowing what they are supposed to do.
This consists of product and process standards, which should be written
specifications. For clean room operations, documentation is rarely available in
printed-paper formats, and "written specifications" is meant to refer to
documentation in concrete text and/or graphics formats. It is also important to
ensure that responsibilities are clearly defined. All too often specifications
contain statements such as "Resist coater bowls should be cleaned weekly."
With three or four shift operations, it is easy for everyone to assume that
someone else is performing the specified task. Training is also essential for
satisfying this element of self-control, to ensure that people know how to
execute the tasks that they have been assigned.
Knowing what they actually are doing.
Closed feedback loops are required between inspection/measurement and
processing operations. The pace of the cleanroom is swift, so data need to be
formatted to allow operators to review the data quickly. Sufficient information
needs to be provided to guide corrective action. It is also important to report
only a digestible number of important issues; otherwise, it will be too
overwhelming to be effective. Statistical process control is a particularly
powerful feedback system, providing information about the state of the
process, as well as the parameters of the system when it is in control.
Ability to regulate.
The process must be capable of producing a product to specifications if
operators are expected to be held accountable for producing a good product.
Corrective action is most effective in situations where the people have
understandings of common failure modes, and the process is responsive to

regulatory action in a predictable cause-and-effect relationship. Resources must

be readily available to the worker who regulates the process. For example, in
one wafer fab, operators were required to routinely wipe down equipment with
isopropanol and water. Because this solution contained a flammable material,
management required that bottles containing the liquid be kept, when not in
use, in a cabinet designed for storing flammable materials. Unfortunately, only
one such cabinet was installed within a large facility, and no one was assigned
the task of ensuring that a supply of full bottles would be maintained inside the
cabinet. It was not surprising that busy operators who worked in the parts of
the wafer fab far from the cabinet frequently failed to wipe down their
equipment. Making it a requirement to wipe the equipment was insufficient,
because the means for performing the task was supported inadequately.
Resources needed to be made available which would have allowed the task to
be performed easily and quickly.
Whenever there is an operational failure, such as when an incorrect
procedure is performed, it is useful to review the extent to which the lack of
operator self-control played a role. This review should follow a checklist based
on the three conditions that collectively constitute operator self-control. Very
often inadequacies in documentation, training, or the process itself can be
identified as factors contributing to the problem. Identifying and then
correcting these factors can prevent future failures.

9.2 DOCUMENTATION

Written instructions are a powerful quality tool. One engineer, who transferred
from one location where written specifications were well supported to another
location where documentation was not employed rigorously described the
problem he was having with implementing improvements: "I don't know what
they are really doing (on the fab floor), so I don't know what there is to change.
It's like trying to get a solid grip on oatmeal." As noted previously, the number
of suggestions in Japanese companies was higher when the working
instructions were rigidly specified. Documenting work instructions has several
advantages:

1) They are a centralized source of information, available to operators,

engineers, and managers.
2) They provide a communications link. In a well organized company
there is a level of assurance that operators are made aware of changes to
practices on a timely basis. Documented work instructions provide a well-
defined method for implementing quality and productivity improvement.
3) Closure for statistical process control is provided. SPC provides the
means for identifying a loss of process control, but does not provide
operators with guidance on what to do when control is lost. This is an
essential part of the "ability to regulate."

Part of the discipline of documentation is a requirement for clarity in work

instructions, and the document control system should contain training for those
who write work instructions. Documented work instructions should inform
those who perform operations as to what should be done. The reasons and
motivations for the particular instructions should not be included in the work
instructions; otherwise the documents will become too large, and it will
become difficult to find information. It is usually worthwhile to explain the
basis for work instructions, but this information should be provided to
operators in special off-line training classes.
It has often been asserted that well-defined work instructions are
appropriate for manufacturing organizations, but a development operation
represents a different situation. Examples show that documentation is
important in development. 3 In one instance, a company had produced working
circuits after working for two years to develop the newest generation of
DRAMs. Unfortunately, processes were frequently being changed, but not
documented, and the company was unable to determine what processes had
actually been used to produce the good chips. It was almost another year before
they were able to replicate their initial success. In a different company the
development organization waited until a new process was ready to be
transferred to manufacturing before documentation was generated. The writing
process took several months, delaying the introduction of the new technology.
Part of the reluctance to make use of a potentially powerful quality tool
stems from how difficult it is to use many document control systems. All too
often approvals take too long to obtain, or engineers encounter unexpected
questions concerning their new processes. These problems can be avoided by
establishing approval systems that involve pre-defined requirements and
minimum levels of approval. The requirements for making changes during
process development should be considerably different from those used during
volume manufacturing, allowing for greater flexibility where it is needed.
Management support is essential for documentation to be a useful quality
tool. First, people within an organization need to believe that adherence to
specifications is expected. Document control systems will quickly become
irrelevant if management does not insist that process changes be implemented
through the document control system. E-mail is a particularly insidious
technology, because it has the appearance of control ("I wrote the instructions
exactly"), but rarely are records maintained of process changes "implemented"
by e-mail.
The difference between a well-controlled and poorly-controlled process is
often just a detail. Documented work instructions provide the means for
ensuring that the details are correctly addressed. As mentioned earlier,
engaging operators and technicians into quality improvement programs is
facilitated by providing a concrete system for implementing changes.
Documentation is a key element of any program to improve process control.

9.3 ISO 9000

ISO 9000 generally refers to a set of standards to which quality systems are
expected to perform. These are not technical standards, which are necessarily
specific to individual industries and products, but standards which apply to the
management of organizations. The most comprehensive of these standards, ISO
9001, which is applicable to organizations which design, develop, manufacture,
install, and service products, will be discussed here (see Table 9.1). Some
companies or plant sites may be registered to ISO 9002, which is a subset of
the ISO 9001 standard and is intended for organizations that only manufacture
and install products. It is possible for one site of a company, where
development is done, to be registered to the ISO 9001 standard, and another
site, which does only manufacturing, to be registered to ISO 9002. A
comprehensive discussion of the ISO 9000 standards would require an entire
book, not just a chapter. It is the purpose of this section to highlight those
elements of ISO 9001 of greatest significance for lithographers.

Elements of ISO 9001

1) Management Responsibility
2) Quality System
3) Contract Review
4) Design Control
5) Document and Data Control
6) Purchasing
7) Control of Customer-Supplied
Product
8) Product Identification and
Traceability
9) Process Control
10) Inspection and Testing
11) Control of Inspection,
Measuring, and Test Equipment
12) Inspection and Test Status
13) Control of Nonconforming
Product
14) Corrective and Preventive Action
15) Handling, Storage, Packaging,
Preservation, and Delivery
16) Control of Quality Records
17) Internal Quality Audits
18) Training
19) Servicing
20) Statistical Techniques

Table 9.1. The twenty elements, or sections, of ISO 9001.

The need to reduce redundant quality audits, in which customers tour

facilities and peruse quality records, motivated the creation of the ISO 9000
standards. Such audits could occur several times in a single week. 4 Because
there are many characteristics common to all quality systems, there is
considerable redundancy among such audits. Under ISO 9000, companies
conform to certain quality standards and are audited by accredited auditors. A
registry of companies conforming to ISO 9000 addresses the problem of audit
redundancy.
There are few requirements to conform to the ISO 9000 standards, outside
of a handful of classes of regulated products in the European Union, relating
primarily to safety concerns. 5 However, a standard for quality systems that has
wide applicability must contain only the most essential elements for managing
quality. Few customers want to buy from suppliers who are unable to
demonstrate that they can meet these most basic of standards. For this reason,
ISO 9000 has become a de facto necessity of doing business.
The basic elements of ISO 9001 are shown in Table 9.1. Only a few of
these, of specific relevance to lithography, will be discussed in this text. Part of
the Process Control section of the ISO 9001 standard reads:

4.9 PROCESS CONTROL

The supplier shall identify and plan the production, installation, and
servicing processes which directly affect quality and shall ensure that
these processes are carried out under controlled conditions. Controlled
conditions shall include the following:
a) documented procedures defining the manner of production, installation,
and servicing, where the absence of such procedures could adversely
affect quality;
b) use of suitable production, installation, and servicing equipment, and a
suitable working environment;
c) compliance with reference standards/codes, quality plans, and/or
documented procedures;
d) monitoring and control of suitable process parameters and product
characteristics;
e) approval of processes and equipment, as appropriate;
f) criteria for workmanship, which shall be stipulated in the clearest
practical manner (e.g., written standards, representative samples, or
illustrations);
g) suitable maintenance of equipment to ensure continuing process
capability.

Documented work instructions are mentioned explicitly in this element of ISO

9001. Further details of the document and records control systems are
contained in elements 5 and 16. Section 4.9 of ISO 9001 also calls out
explicitly for process monitoring. Statistical process control and the methods

described in Chapters 3 and 7 of this text provide the means for satisfying this
requirement of ISO 9001.
The need for suitable measurement capability is stated explicitly in element
4.11, Control of Inspection, Measuring, and Test Equipment:

4.11.2 Control procedure

The supplier shall:
a) determine the measurements to be made and the accuracy required, and
select the appropriate inspection, measuring, and test equipment that is
capable of necessary accuracy and precision;
b) identify all inspection, measuring, and test equipment that can affect
product quality, and calibrate and adjust them at prescribed intervals, or
prior to use, against certified equipment having a known valid relationship
to internationally or nationally recognized standards. Where no such
standards exist, the basis used for calibration shall be documented;
c) define the process employed for the calibration of inspection,
measuring, and test equipment...

This section clearly addresses the standards dilemma in critical dimension

metrology. In order to meet the requirements of ISO 9000, metrologists must
implement a suitable program for controlling the metrology tools. Further
guidance for conforming to the metrology requirements of ISO 9000 can be
found in another standard, ISO 10012, "Quality assurance requirements for
measuring equipment."
ISO 9000 does explicitly mention the need for the use of appropriate
statistical techniques:

4.20 STATISTICAL TECHNIQUES

4.20.1 Identification of need

The supplier shall identify the need for statistical techniques required for
establishing, controlling, and verifying process capability and product
characteristics.
4.20.2 Procedures
The supplier shall establish and maintain documented procedures to
implement and control the application of the statistical techniques
identified in 4.20.1.

Because it is intended that ISO 9000 have wide applicability to numerous

industries, it is left to the technical experts within each industry to determine
where statistical methods, such as statistical process control, should be
implemented. This element of ISO 9000 brings us full circle to the preceding
subjects covered in this text: the identification of processes where controls
need to be inserted and the technical and statistical tools that can facilitate

implementation. ISO 9000 recognizes that the expertise of engineers and

technicians forms the foundation of quality control. This is a most certain truth
for microlithography.

1 M. Imai, Kaizen: The Key to Japan's Competitive Success, Random House,

New York, p. 113 (1986).

2 Juran 's Quality Control Handbook, 4th Edition, J. M Juran and F. M. Gryna,
eds., McGraw Hill, New York (1988).

3 H. J. Levinson and J. Ben-Jacob, "Managing Quality Improvement on a

Development Pilot Line," Quality Management Journal, Vol. 3(2), pp. 16 — 35

(1996).

4 C. A. De Angelis, "ICI Advanced Materials Implements ISO 9000 Program,"

Quality Progress, Vol. 24, pp. 49 —51 (1991).

5H. J. Levinson, "ISO 9000: What Every Microlithographer Should Know,"

SPIE Vol. 2196, pp. 536 — 550 (1993).

GENERAL REFERENCES ON QUALITY CONTROL:

• W. Shewhart, Economic Control of Quality of Manufactured Product, D.

van Nostrand Co., New York, (1931).

• Edwards Deming, Out of the Crisis, MIT Center for Advanced Engineering
Study, Cambridge, MA (1982).

• D. C. Montgomery, Introduction to Statistical Quality Control, John Wiley

& Sons, New York (1996).

• Western Electric's Statistical Quality Control Handbook, Delmar Printing

Company, Charlotte (1985).

• A. J. Duncan, Quality Control and Industrial Statistics, 5th Edition, Irwin,

Homewood, Illinois (1986).

• M. G. Natrella, Experimental Statistics, U.S. Government Printing Office,

Washington, D. C. (1966).

• S. S. Shapiro, How to Test Normality and Other Distributional

Assumptions, American Society for Quality Control, Milwaukee (1986).

• G. E. P. Box, G. M. Jenkins, and G. C. Reinsel, Time Series Analysis, 3rd

Edition, Prentice Hall, Englewood Cliff, NJ (1994).

• Guide for Reducing Quality Costs, 2nd Edition, ASQ Press, Milwaukee
(1987).

• V. Barnett and T. Lewis, Outliers in Statistical Data, 3rd Edition, John

Wiley and Sons, Chichester (1995).

• M. Imai, Kaizen: The Key to Japan's Competitive Success, Random House,

New York, p. 113 (1986).

• Juran's Quality Control Handbook, 4th Edition, J. M Juran and F. M.

Gryna, eds., McGraw Hill, New York (1988).

169

• K. Ishikawa, Guide to Quality Control, Asian Productivity Organization,

White Plains, New York (1991).

• G. E. P. Box, W. G. Hunter, and J. S. Hunter, Statistics for Experimenters,

John Wiley & Sons, New York (1978).

• G. E. P. Box and N. R. Draper, Empirical Model-building and Response

Surfaces, John Wiley & Sons, New York (1987).

• G. Taguchi, System of Experimental Design, Vols. 1 and 2, American

Supplier Institute Press, Dearborn, Michigan (1987).

PROCESS CAPABILITY:

• V. E. Kane, "Process Capability Indices," Journal of Quality Technology,

Vol. 18, pp. 41— 52 (1986).

• L. K. Chan, S. W. Cheng, and F. A. Spiring, "A New Measure of Process

Capability: Cpm Journal of Quality Technology, Vol. 20, pp. 162 — 175
,"

(1988).

• R. A. Boyles, "The Taguchi Capability Index," Journal of Quality

Technology, Vol. 23, pp. 17 — 26 (1991).

• D. J. Wheeler, A Japanese Control Chart, SPC Press, Knoxville (1986).

• R. McFadden, "Six-Sigma Quality Programs," Quality Progress, pp. 37 —

42 (June, 1993).

• Y-M. Chou, D. B. Owen, and S. A. Borrego, "Lower Confidence Limits on

Process Capability Indices," Journal of Quality Technology, pp. 223 — 229
(1990).

• F. S. Hillier, " X - and R-Chart Control Limits Based on a Small Number of

Subgroups," Journal of Quality Technology, 1, pp. 17-26 (1969).

• C. P. Quesenberry, "SPC Q Carts for Start-Up Processes and Short or Long

Runs," Journal of Quality Technology, 23(3), pp. 213 — 224 (1991).

• P. P. Ramsey and P. H. Ramsey, "Simple Tests of Normality in Small

Samples," Journal of Quality Technology, pp. 299 — 313 (1990).

• J. F. MacGregor, "A Different View of the Funnel Experiment," Journal of

Quality Technology, Vol. 22, pp. 255 — 259 (1990).

• S. T. Mandraccia, G. D. Halverson, and Y-M. Chou, "Control Chart Design

Strategies for Skewed Data," SPIE Vol. 2876, pp. 196 — 205 (1996).

LINEWIDTH CONTROL:

• D. Heberling, "Litho Equipment Matching with E 0 ," Proceedings of the

KTI Microelectronics Seminar, pp. 233 — 243 (1990).

• C. Yu, T. Maung, C. Spanos, D. Boning, J. Cheung, H-Y Liu, K-J. Chang,

and D. Bartelink, "Use of Short-Loop Electrical Measurements for Yield
Improvement," IEEE Trans. Semicond. Manuf., Vol. 8(2), pp. 150 — 159
(1995).

• R. C. Elliott, R. R. Hershey, and K. G. Kemp, "Cycle-time Reduction of

CD Targeting using Automatic Metrology and Analysis," SPIE Vol. 2439,
pp. 70 — 77 (1995).

• C. Takemoto, D. Ziger, W. Connor, and R. Distasio, "Resist Tracking: A

Lithographic Diagnostic Tool," SPIE Vol. 1464, pp. 206 — 214 (1991).

• M. van den Brink, C. G. M. de Mol, H. F. D. Linders, and S. Wittekoek,

"Matching Management of Multiple Wafer Steppers Using a Stable
Standard and a Matching Simulator," SPIE Vol. 1087, pp. 218 — 232
(1989).

• K. Kemp, C. King, W. Wu, and C. Stager, "A "Golden Standard" Wafer

Design for Optical Stepper Characterization," SPIE Vol. 1464, pp. 260 —
277 (1991).

• H. J. Levinson and W. H. Arnold, "Focus: The Critical Parameter for Sub-

micron Lithography," J. Vac. Sci. Technol., B5(1), pp. 293 — 298 (1987).

• K. Ronse, R. Pforr, L. van den Hove, and M. Op de Beeck, "CD Control:

the Limiting Factor for i -line and Deep-UV Lithography?" OCG
Microelectronics Seminar, pp. 241— 254 (1995).

• K. Ronse, M. Op de Beeck, A. Yen, K-H. Kim, and L. van den Hove,

"Characterization and Optimization of CD Control for 0.25 mm CMOS
Applications," SPIE Vol. 2726, pp. 555 — 563 (1996).

• P. Schoenborn and N. F. Pasch, "Process Sensitivity Analysis:

Applications to Photolithography," SPIE Vol. 1087, pp. 290 — 298 (1989).

• Z. Krivokapic, W. D. Heavlin, and D. Kyser, "Process Capabilities of

Critical Dimensions at Gate Mask," SPIE Vol. 2440, pp. 480 — 491 (1995).

• R. W. Leonhardt and T. R. Scott, "Deep-UV Excimer Laser Measurements

at NIST," SPIE Vol. 2439, pp. 448 — 459 (1995).

• K. H. Kim, W. S. Han, C. H. Kim, H. Y. Kang, C. G. Park, and Y. B. Koh,

"Characteristics of Standing Wave Effect of Off-axis Illumination
Depending on two Different Resist Systems and the Polarization Effect of
Stepper," SPIE Vol. 2197, pp. 42-53 (1994).

• J. Sturtevant and B. Roman, "Antireflection Strategies for Advanced

Photolithography," Microlithography World, pp. 13 —21(1995).

• G. MacBeth, "Thermal Effects in Photoresist Coating Processes," KTI

Microelectronics Seminar, pp. 327 — 340 (1988).

• S. Dick and B. Greenstein, "Improved Photolithography Process

Performance Through the Use of an Integrated Photosector," KTI
Microelectronics Seminar, pp. 1— 8 (1989).

• M. Reihani, "Environmental Effects on Resist Thickness Uniformity,"

Semiconductor International, pp. 120 — 121 (June, 1992).

• J. S. Petersen and J. D. Byers, "Examination of Isolated and Grouped

Feature Bias in Positive Acting Chemically Amplified Resist Systems,"
SPIE Vol. 2724, pp. 163 — 171 (1996).

• J. S. Petersen, C. A. Mack, J. W. Thackeray, R. Sinta, T. H. Fedynyshyn, J.

J. Mori, J. D. Byers, and D.A. Miller, "Characterization and Modeling of a
Positive Acting Chemically Amplified Resist," SPIE Vol. 2438, pp.
153— 166 (1995).

• J. M. Kulp, "CD Shift Resulting from Handling Time Variation in the

Track Coat Process," SPIE Vol. 1466, pp. 630 — 640 (1991).

• G. MacBeth, "Prebaking Positive Photoresists," Proceedings of the Kodak

Microelectronics Seminar, pp. 87 — 92 (1982).

• O. D. Crisalle, C. L. Bickerstaff, D. E. Seborg, and D. A. Mellichamp,

"Improvements in Photolithography Performance by Controlled Baking,"
SPIE Vol. 921, pp. 317-325 (1988).

• C. M. Garza, C. R. Szmanda, and R. L Fischer, "Resist Dissolution

Kinetics and Submicron Process Control," SPIE Vol. 920, pp 321 — 338
(1988).

• M. K. Templeton, J. B. Wickman, and R. L. Fischer, Jr., "Submicron

Resolution Automated Track Development Processes, Part 1: Static Puddle
Development," SPIE Vol. 921 pp. 360 — 372 (1988).

• E. Bokelberg and W. Venet, "Effects of Relative-humidity Variation on

Photoresist Processing," SPIE Vol. 2438, pp. 747 — 752 (1995).

• W. H. Arnold and H. J. Levinson, "High Resolution Optical Lithography

Using an Optimized Single Layer Photoresist Process," Proceedings of the
Kodak Microelectronics Seminar, pp. 80 — 92 (1983).

• J. A. Bruce, S. R. DuPuis, and H. Linde, "Effect of Humidity on

Photoresist Performance," Proceedings of the OCG Microlithography
Seminar, pp. 25 —41(1995).

• S. A. MacDonald, C. G. Wilson, and J. M. J. Frêchet, "Chemical

Amplification in High-Resolution Imaging Systems," Accounts of
Chemical Research, Vol. 27(6), pp. 151-158 (1994).

• K. R. Dean, D. A. Miller, R. A. Carpio, and J. S. Petersen, "Airborne

Contamination of DUV Photoresists: Determining the New Limits of
Processing Capability," Proceedings of the Olin Microlithography
Seminar, pp. 109 — 125 (1996).

• K. van Ingen Schenau, M. Reuhman, and S. Slonaker, "Investigation of

DUV Process Variables Impacting Sub-Quarter Micron Imaging,"
Proceedings of the Olin Microlithography Seminar, pp. 63 — 80 (1996).

• Y-K. Hsiao, C-H. Lee, S-L. Pan, K-L. Lu, and J-C. Yang, "A Study of the
Relation Between Photoresist Thermal Property and Wettability,"
Proceedings of the Olin Microlithography Seminar, pp. 315 — 323 (1996).

• H. Ito, W. P. England, H. J. Clecak, G. Breta, H. Lee, D. Y. Yoon, R.

Sooriyakumaran, and W. D. Hinsberg, "Molecular Design for Stabilization

of Chemical Amplification Resist Toward Airborne Contamination," SPIE

Vol. 1925, pp. 65 — 75 (1993).

• W. Maurer, K. Satoh, D. Samuels, and T. Fischer, "Pattern Transfer at

k l = 0.5: Get 0.25 µm Lithography Ready for Manufacturing," SPIE Vol.
2726, pp. 113— 124 (1996).

• K. Petrillo, "Process Optimization of Apex-E," SPIE Vol. 1926, pp.

176 — 187 (1993).

• S. G. Hansen, G. Dao, H. Gaw, Q-D. Qian, P. Spragg, and R. J. Hurditch,

"Study of the Relationship Between Exposure Margin and
Photolithographic Process Latitude and Mask Linearity," SPIE Vol. 1463,
pp. 230-244 (1991).

RESIST THICKNESS CONTROL:

• D. P. Birnie III, B. J. J. Zelinski, S. P. Marvel, S. M. Melpolder, and R. L.

Roncone, "Film/Substrate/Vacuum-Chuck Interactions During Spin-
Coating," Optical Engineering, Vol. 31(9), pp. 2012-2020 (1992).

• R. Murray, P. T. Edwin, V. Taburaza, and J. Olin, "Airflow Controller

Improves Photoresist Spin/Coat Uniformity," Semiconductor International,
pp. 172 —173 (April, 1987).

• L. Matter, J. Zook, M. Hinz, J. Banas, and S. Ibrani, "New Coat Bowl

Design Improves Photoresist Uniformity and Decreases Particle
Contamination," Olin Microlithography Seminar Poster Session (1997).

• X. Zhu, F. Liang, A. Haji-Sheikh, and N. Ghariban, "A Computational and

Experimental Study of Spin Coater Air Flow," SPIE Vol. 3333, pp.
1441-1451 (1998).

• W. J. Daughton, P. O'Hagan, and F. L. Givens, "Thickness Variance of

Spun-On Photoresist, Revisited," Proceedings of the Kodak
Microelectronics Seminar, pp. 15 — 20 (1978).

• "Lithography Track Thermal Metrology Update," Semiconductor

International (1998).

• D. Boutin, A. Blash, J. P. Caire, D. Poncet, P. Fanton, M. Danielou, and B.

Previtali, "Resist Coating Optimization on Eight Inches Deep UV Litho

Cell Modelization and Application to 0.25 µm Technology," SPIE Vol.

2439, pp. 495 — 502 (1995).

FEEDBACK AND AUTOMATIC PROCESS CONTROL:

• H. J. Levinson, "Control and Improvement of Complex Processes," Quality

Engineering, Vol. 5(1), pp. 93 — 106 (1992).

• T. Batchelder, M. Ha, R. Haney, and W. Lee, "Sub-micron Linewidth

Control with Automatic Optical Monitors," KTI Microelectronics Seminar,
pp. 231 — 242 (1989).

• T. Batchelder, M. Ha, R. Haney, and W. Lee, "Sub-Micron Linewidth

Control with Automatic Optical Monitors," KTI Microelectronics Seminar,
pp. 231 — 242 (1989).

• M. Watts and S. Williams, "A Novel Method for the Prediction of Process
Sensitivity in Photolithography," SPIE Vol. 1261, pp. 345 — 359 (1990).

• K. Kemp, D. Williams, J. Daggett, J. Cayton, S. Slonaker, and R. Elliott,

"Critical Dimension Performance Characterization of an Advanced DUV
Process Cell," Olin Microlithography Seminar, pp. 99 —108 (1996).

• M. Drew and K. Kemp, "Automatic Feedback Control to Optimize Stepper

Overlay," SPIE Vol. 1926, pp. 422 — 428 (1993).

• C. P. Ausschnitt, A. C. Thomas and T. J. Wiltshire, "Advanced DUV

Photolithography in a Pilot Line Environment," IBM J. Res. Develop. Vol.
41(1/2), pp. 21— 36 (1997).

• F. E. Grubbs, "An Optimum Procedure for Setting Machines or Adjusting

Processes," Journal of Quality Technology, Vol. 15(4), pp. 186 — 189
(1983).

• S. V. Crowder, "Design of Exponentially Weighted Moving Average

Schemes," Journal of Quality Technology, Vol. 21, pp. 155 —162 (1989).

• J. S. Hunter, "The Exponentially Weighted Moving Average," Journal of

Quality Technology, Vol. 18, pp. 203 — 210 (1986).

• J. M. Lucas and M. S. Saccucci, "Exponentially Weighted Moving

Average Control Schemes: Properties and Enhancements," Technometrics,
Vol. 32, pp. 1— 12 (1990).

• K. L. Konnerth and F. H. Dill, "In-situ Measurement of Dielectric

Thickness During Etching or Developing Processes," IEEE Trans. Electron
Devices, Vol. ED-22, pp. 452 — 456 (1975).

• L. J. Lauchlan, K. M. Sautter, and T. Batchelder, "Automatic Process

Control for VLSI Linewidth," Solid State Technology, pp. 333 — 337 (April
1985).

• L. J. Lauchlan, K. Sautter, T. Batchelder, and J. Irwin, "In-Line Automatic

Photoresist Process Control," SPIE Vol. 539, pp. 227 — 233 (1985).

• C. Nygren, J. Daggett, and J. Grambow, "The Use of Develop End Point

Detection to Eliminate Photolithography Process Variation," Proceedings
of the KTI Microlithography Seminar, pp. 315 — 348 (1990).

• P. D. Flanner III, "Improved Methods of Photoresist Development

Characterization," Proceedings of the KTI Microelectronics Seminar, pp.
231-238 (1987).

• K. M. Sautter, M. Ha, and T. Batchelder, "Development Process Control

and Optimiziation Utilizing anEnd Point Monitor," Proceedings of the KTI
Microelectronics Seminar, pp. 99 — 112 (1988).

• J. A. Bruce and B. J. Lin, "Determination of Exposure Dose by Photoresist

Development Rate," Proceedings of the KTI Microelectronics Seminar, pp.
1-11(1987).

• M. L. Miller and D. A. Mellichamp, "Development of an End-Point

Detection Procedure for the Post-Expose Bake Process," SPIE Vol. 2439,
pp. 78 — 88 (1995).

• S. Zaidi, S. L. Prins, J. R. McNeil, and S. S. H. Naqvi, "Metrology Sensors

for Advanced Resists," SPIE Vol. 2196, pp. 341— 351 (1994). '

• J. L. Sturtevant, S. J. Holmes, T. G. VanKessel, P. D. Hobbs, J. C. Shaw,

and R. R. Jackson, "Postexposure Bake as a Process-Control Parameter for
Chemically Amplified Photoresist," SPIE Vol. 1926, pp. 106— 114 (1993).

• J. L. Sturtevant, S. Holmes, T. G. Van Kessel, M. Miller, and D.

Mellichamp, "Use of Scatterometric Latent Image Detector in Closed Loop
Feedback Control of Linewidth," SPIE Vol. 2196, pp. 352-359 (1994).

• K. C. Hickman, S. M. Gaspar, K. P. Bishop, S. S. H. Haqvi, J. R. McNeil,

G. D. Tipton, B. R. Stallard, and B. L. Draper, "Use of Diffracted Light
from Latent Images to Improve Lithography Control," SPIE Vol. 1464, pp.
245-251 (1991).

• T. E. Adams, "Applications of Latent Image Metrology in

Microlithography," SPIE Vol. 1464, pp. 294 — 312 (1991).

• S. L. Prins and J. R. McNeil, "Scatterometric Sensor for PEB Process

Control," SPIE Vol. 2725, pp. 710 — 719 (1996).

• L. M. Milner, K. C. Hickman, S. M. Gaspar, K. P. Bishop, S. S. H. Haqvi,

J. R. McNeil, M. Blain, and L. Draper, "Latent Image Exposure Monitor
Using Scatterometry," SPIE Vol. 1673, pp. 274 — 283 (1992).

• M. L. Miller and D. A. Mellichamp, "Development of an End-Point

Detection Procedure for the Post-Exposure Bake Process," SPIE Vol. 2439,
pp. 78 — 88 (1995).

• T. Koizumi, T. Matsuo, M. Endo, and M. Sasago, "CD Control using

Latent Image for Lithography," SPIE Vol. 2439, pp. 418 — 426 (1995).

• D. C. Montgomery, J. B. Keats, G. C. Runger, and W. S. Messina,

"Integrating Statistical Process Control and Engineering Process Control,"
Journal of Quality Technology, Vol. 26(2), pp. 79 — 87 (1994).

• The April, 1997, Vol. 29(2) of the Journal of Quality Technology has a
discussion on statistically-based process monitoring and control, with
contributions from several authors and extensive references.

• G. Box and A. Luceflo, Statistical Control by Monitoring and Feedback

Adjustment, John Wiley & Sons, New York (1997).

Focus:

• T. O. Herndon, C. E. Woodward, K. H. Konkle, and J. I. Raffel,

"Photocomposition and DSW Autofocus Correction for Wafer-Scale
Lithography," Proceedings of the Kodak Microelectronics Seminar, pp.
118-123(1983).

• A. Suzuki, S. Yabu, and M. Ookubo, "Intelligent Optical System for a New

Stepper, "SPIE Vol. 772, pp. 58 — 65 (1987).

• J. E. van den Werf, "Optical Focus and Level Sensor for Wafer Steppers,"
J. Vac. Sci. Technol., Vol. 10(2) , pp. 735 — 740 (1992).

• M. A. van den Brink, J. M. D. Stoeldraijer, and H. F. D. Linders, "Overlay

and Field by Field Leveling in Wafer Steppers Using an Advanced
Metrology System," SPIE Vol. 1673, pp. 330 — 344 (1992).

• J. W. Gemminck, "Simple and Calibratable Method for Determining

Optimal Focus," SPIE Vol. 1088, pp. 220 — 230 (1989).

• S. Venkataram, C. Olejnik, G. Flores, and D. Tien, "An Automated

Technique for Optimizing Stepper Focus Control," SPIE Vol. 2725, pp.
765 — 778 (1996).

• T. A. Brunner, A. L. Martin, R. M. Martino, C. P. Ausschnitt, T. H.

Newman, and M. S. Hibbs, "Quantitative Stepper Metrology Using the
Focus Monitor Test Mask," SPIE Vol. 2197, pp. 542 — 549 (1994).

• R. D. Mih, A. Martin, T. Brunner, D. Long, and D. Brown, "Using the

Focus Monitor Test Mask to Characterize Lithographic Performance,"
SPIE Vol. 2440, pp. 657 — 666 (1995).

• E. R. Sherman and C. Harker, "Characterization and Monitoring of

Variable NA and Variable Coherence Capable Photo Steppers Utilizing the
Phase Shift Focus Monitor Reticle," SPIE Vol. 2439, pp. 61— 69 (1995).

• T. A. Brunner and R. D. Mih, "Simulations and Experiments with the

Phase Shift Focus Monitor," SPIE Vol. 2726, pp. 236 — 243 (1996).

• T. A. Brunner and S. M. Stuber, "Characterization and Setup Techniques

for a 5X Stepper," SPIE Vol. 633, pp. 106 — 112 (1986).

• M. van den Brink, H. Franken, S. Wittekoek,and T. Fahner, "Automatic

On-Line Wafer Stepper Calibration System," SPIE Vol. 1261, pp. 298 —
314 (1990).

• T. A. Brunner, S. Cheng, and A. E. Norton, "Stepper Image Monitor for

Precise Setup and Characterization," SPIE Vol. 922, pp. 366 — 375 (1988).

• R. Pforr, S. Wittekoek, R. van den Bosch, L. van den Hove, R. Jonckheere,

T. Fahner, and R. Seltmann, "In-process Image Detecting Technique for
Determination of Overlay and Image Quality for the ASM-L Wafer
Stepper," SPIE Vol. 1674, pp. 594 — 609 (1992).

• P. Dirksen, W. de Laat, and H. Megens, "Latent Image Metrology for

Production Wafer Steppers," SPIE Vol. 2440, pp. 701 —711(1995).

• T. A. Brunner, J. M. Lewis, and M. P. Manny, "Stepper Self-Metrology

Using Automated Techniques," SPIE Vol. 1261, pp. 286 — 297 (1990).

• K. Hale and P. Luehrman, "Consistent Image Quality in a High

Performance Stepper Environment," Proceedings of the Kodak
Microelectronics Seminar, pp. 29 — 46 (1986).

OVERLAY:

• R. Groves, "Statistics of Pattern Placement Errors in Lithography," J. Vac.

Sci Technol., B9(6), pp. 3555 — 3561 (1991).

• N. Magome and H. Kawai, "Total Overlay Analysis for Designing Future

Aligner," SPIE Vol. 2440, pp. 902 — 912 (1995).

• M. A. van den Brink, J. M. D. Stoeldraijer, and H. F. D. Linders, "Overlay

and Field by Field Leveling in Wafer Steppers Using an Advanced
Metrology System," SPIE Vol. 1673, pp. 330 — 344 (1992).

• T. Saito, S. Sakamoto, K. Okuma, H. Fukumoto, and Y. Okuda, "Mask

Overlay Scaling Error Caused by Exposure Energy Using a Stepper," SPIE
Vol. 1926, pp. 440-449 (1993).

• J. D. Armitage, Jr. and J. P. Kirk, "Analysis of Overlay Distortion

Patterns," SPIE Vol. 921, pp. 207 — 222 (1988).

• T. E. Zavecz, "Lithographic Overlay Measurement Precision and

Calibration and Their Effect on Pattern Registration Optimization," SPIE
Vol. 1673, pp. 191-202 (1992).

• V. Nagaswami and W. Geerts, "Overlay Control in Submicron

Environment," Proceedings of the KTI Microelectronics Seminar, pp.
89 —106 (1989).

• M. A. van den Brink, C. G. M. de Mol, and R. A. George, "Matching

Performance for Multiple Wafer Steppers Using an Advanced Metrology
Procedure," SPIE Vol. 921, pp. 180 —197 (1988).

• J. G. Maltabes, M. C. Hakey, and A. L. Levine, "Cost/Benefit Analysis of

Mix-and-Match Lithography for Production of Half-Micron Devices,"
SPIE Vol. 1927, pp. 814 — 826 (1993).

• R. K. Brimacombe, T. J. McKee, E. D. Mortimer, B. Norris, J. Reid, and

T. A. Znotins, "Performance Characteristics of a Narrow Band Industrial
Excimer Laser," SPIE Vol. 1088, pp. 416 — 422 (1989).

• H. J. Levinson and R. Rice, "Overlay Tolerances for VLSI Using Wafer

Steppers," SPIE Vol. 922, pp. 82-93 (1988).

• M. A. van den Brink, C. G. M. de Mol, and J. M. D. Stoeldraijer,

"Matching of Multiple Wafer Steppers for 0.35 mm Lithography, Using
Advanced Optimization Schemes," SPIE Vol. 1926, pp. 188 — 207 (1993).

• J. C. Pelligrini, "Comparisons of Six Different Intrafield Control

Paradigms in an Advanced Mix-and-Match Environment," SPIE Vol. 3050,
pp. 398 — 406 (1997).

• H. J. Levinson, M. E. Preil, and P. J. Lord, "Minimization of Total Overlay

Errors on Product Wafers Using an Advanced Optimization Scheme," SPIE
Vol. 3051, pp. 362-373 (1997).

• N. R. Farrar, "Effect of Off-axis Illumination on Stepper Overlay," SPIE

Vol. 2439, pp. 273 — 280 (1995).

• T. Saito, H. Watanabe, and Y. Okuda, "Effect of Variable Sigma Aperture

on Lens Distortion and Its Pattern Size Dependence," SPIE Vol. 2725, pp.
414-423 (1996).

• A. M. Davis, T. Dooly, and J. R. Johnson, "Impact of Level Specific

Illumination Conditions on Overlay," Proceedings of the Olin
Microlithography Seminar, pp. 1— 16 (1997).

• C. S. Lee, J. S. Kim, I. B. Hur, Y. M. Ham, S. H. Choi, Y. S. Seo, and

S. M. Ashkenaz, "Overlay and Lens Distortion in a Modified Illumination
Stepper," SPIE Vol. 2197, pp. 2— 8 (1994).

• C-S. Lim, K-S. Kwon, D. Yim, D-H. Son, H-S. Kim, and K-H. Baik,
"Analysis of Nonlinear Overlay Errors by Aperture Mixing Related to
Pattern Asymmetry," SPIE Vol. 3051, pp. 106— 115 (1997).

• M. E. Prei!, T. Manchester, and A. Minvielle, "Minimization of Total

Overlay Errors when Matching Non-concentric Exposure Fields," SPIE
Vol. 2197, pp. 753 — 769 (1994).

• W. W. Flack, G. E. Flores, J. C. Pellegrini, and M. Merrill, "An Optimized

Registration Model for 2:1 Stepper Field Matching," SPIE Vol. 2197, pp.
733 — 752 (1994).

• T. A. Brunner, "Impact of Lens Aberrations on Optical Lithography," IBM

J. Res. Develop., Vol. 41(1/2), pp. 57-67 (1997).

• R. Rogoff, S. S. Hong, D. Schramm, and G. Espin, "Reticle Specific

Compensations to Meet Production Overlay Requirements for 64 Mb and
Beyond," SPIE Vol. 2197, pp. 781 — 790 (1994).

• G. Rivera and P. Canestrari, "Process Induced Wafer Distortion:

Measurement and Effect on Overlay in Stepper Based Advanced
Lithography," SPIE Vol. 807, pp. 806 — 813 (1993).

• H. Izawa, K. Kakai, and M. Seki, "Fully Automatic Measuring System for

Submicron Lithography," SPIE Vol. 1261, pp. 470 — 481 (1990).

• K. Chivers, "A Modified Photoresist Spin Process for a Field-by-Field

Alignment System," Proceedings of the Kodak Microelectronics Seminar,
pp. 44-51 (1984).

• L. M. Manske and D. B. Graves, "Origins of Asymmetry in Spin-Cast

Films Over Topography," SPIE Vol. 1463, pp. 414-422 (1991).

• S. Kuniyoshi, T. Terasawa, T. Kurosaki, and T. Kimura, "Contrast

Improvement of Alignment Signals from Resist Coated Patterns," J. Vac.
Sci. Technol., B5(2), pp. 555 — 560 (1987).

• G. Flores and W. W. Flack, "Photoresist Thin-Film Effects on Alignment

Process Capability," SPIE Vol. 1927, pp. 366 — 380 (1993).

• N. Bobroff, and A. Rosenbluth, "Alignment Errors from Resist Coating

Topography," J. Vac. Sci. Technol. B, pp. 403 — 408 (1988).

• R. Mohondro, S. Bachman, T. Kinney, G. Meissner, and D. Peters, "High

Contrast Eduction Spin Coat Process Effects on Uniformity and Overlay
Registration," Proceedings of the Olin Microlithography Seminar, pp.
131— 140 (1997).

• C. P. Kirk, "Theoretical Models for the Optical Alignment of Wafer

Steppers," SPIE Vol. 772, pp. 134 — 141 (1987).

• K. Ota, N. Magome, and K. Nishi, "New Alignment Sensors for Wafer

Stepper," SPIE Vol. 1463, pp. 304 — 314 (1991).

• D. L. Meunier, D. Humphrey, B. Peck, P. Feeney, J. Paquette, and

J. Thibault, "Optimization of Metal Layer Overlay and Alignment Targets
in a Chemical-Mechanical Polishing Environment," Proceedings of the
Olin Microlithography Seminar, pp. 355 — 366 (1996).

• S-W. Hsia, G. Miyagi, and M. Brongo, "Alignment Characterization of

CMP Tungsten Process," Proceedings of the Olin Microlithography
Seminar, pp. 381— 390 (1996).

• D. L. Meunier, B. Plambeck, P. Lord, and N. Knoll, "The Implementation

of Coherence Probe Microscopy in a Process Using Chemical Mechanical
Polishing," Proceedings of the OCG Microlithography Seminar, pp.
155 — 169 (1995).

• N. Shirishi, A. Sugaya, and D. P. Coon, "Alignment Strategies for

Planarizing Technologies," SPIE Vol. 3051, pp. 836 — 845 (1997).

• T. Kanda, K. Mishima, E. Murakami, and H. Ina, "Alignment Sensor

Corrections for Tool Induced Shift (TIS)," SPIE Vol. 3051, pp. 846 — 855
(1997).

• W. H. Arnold, "Overlay Simulator for Wafer Steppers," SPIE Vol. 922, pp.
94 — 105 (1988).

• B. Rangarajan, M. Templeton, L. Capodieci, R. Subramanian, and A.

Scranton, "Optimal Sampling Strategies for Sub-100 nm Overlay," SPIE
Vol. 3332, pp. 348 — 359 (1998).

• G. E. Flores, W. W. Flack, S. Avlakeotes, and B. Martin, "Process Control

of Stepper Overlay Using Multivariate Techniques," Proceedings of the
OCG Microlithography Seminar, pp. 201— 215 (1995).

• G. E. Flores, W. W. Flack, S. Avlakeotes, and M. Merrill, "Monitoring and

Diagnostic Techniques for Control of Overlay in Steppers," SPIE Vol.
2439, pp. 40-60 (1995).

• R. M. Silver, J. Potzick, and R. D. Larrabee, "Overlay Measurements and

Standards," SPIE Vol. 3429, pp. 262 — 272 (1995).

• D. J. Coleman, P. J. Larson, A. D. Lopata, W. A. Muth, and A. Starikov,

"On the Accuracy of Overlay Measurements: Tool and Mark Asymmetry
Effects," SPIE Vol. 1261, pp. 139 —161 (1990).

• A. Starikov, D. J. Coleman, P. J. Larson, A. D. Lopata, and W. A. Muth,

"Accuracy of Overlay Measurements: Tool and Mark Asymmetry Effects,"
Optical Engineering, Vol. 31(6), pp. 1298 — 1310 (1992).

• M. E. Prei!, B. Plambeck, Y. Uziel, H. Zhou, and M. W. Melvin,

"Improving the Accuracy of Overlay Measurements through Reduction in
Tool and Wafer Induced Shifts," SPIE Vol. 3050, pp. 123 — 134 (1992).

• N. Smith, G. Goelzer, M. Hanna, and P. Troccolo, "Minimizing Optical

Overlay Measurement Errors," SPIE Vol. 1926, pp. 450 — 462 (1993).

• J-S. Han, H. Kim, J-L. Nam, M-S Han, S-K. Lim, S. D. Yanowitz, N. P.
Smith, and A. M. C. Smout, "Effects of Illumination Wavelength on the
Accuracy of Optical Overlay Metrology," SPIE Vol. 3051, pp. 417 — 425
(1997).

• Y. Tanaka, M. Kamiya, and N. Suzuki, "New Methodology of Optimizing

Optical Overlay Measurement," SPIE Vol. 1926, pp. 429 — 439 (1993).

• T. Saito, H. Watanabe, and Y. Okuda, "Overlay Error of Fine Patterns by

Lens Aberration using Modified Illumination," SPIE Vol. 3051, pp.
687 — 696 (1997).

YIELD:

• C. F. King, G. P, Gill, and M. J. Satterfield, "Electrical Defect Monitoring

for Process Control," SPIE Vol. 1087, pp. 76 — 82 (1989).

• G. Dickerson and R. P. Wallace, "In-line Wafer Inspection using

100-Megapixel -per-Second Digital Image Processing Technology," SPIE
Vol. 1464, pp. 584 — 595 (1991).

• L. H. Lin, "A Holographic Photomask Defect Inspection System," SPIE

Vol. 538, pp. 110— 116 (1985).

• S. P. Billat, "Automated Defect Detection on Pattern Wafers,"

Semiconductor International, (May 1987).

• D. L. Cavan, "Patterned Wafer Inspection Using Laser Holography and

Spatial Frequency Filtering," J. Vac. Sci. Technol., (November 1988).

• D. Hendricks, J. Jau, H. Dohse, A. Brodie, and D. Meisburger,

"Characterization of a New Automated Electron-Beam Wafer Inspection
System, SPIE Vol. 2439, pp. 174 — 183, (1995).

• J. Sturtevant, M. Chaara, R. Elliot, L. Hollifield, R. Soper, D. Stark, N.

Thane, and J. Petersen, "Antireflection Coating Process Characterization
and Improvement for DUV Lithography at 0.25 ❑ m Groundrules," SPIE
Vol. 2440, pp. 582 — 593 (1995).

• E. H. Bokelberg and M. E. Pariseau, "Excursion Monitoring of

Photolithographic Processes," Olin Microlithography Seminar, pp.
85 — 100 (1997).

• J. R. Dralla and J. C. Hoff, "Automatic Classification of Defects in

Semiconductor Devices," SPIE Vol. 1261, pp. 173 — 182 (1990).

• D. J. Friedman and S. L. Albin, "Clustered Defects in IC Fabrication:

Impact on Process Control Charts," IEEE Trans. Semiconductor
Manufacturing, Vol. 4(1), pp. 36-42 (1991).

• D. Michelson, "Data-Driven Method for Calculating Limits for Particle

Control Charts," SPIE Vol. 2876, pp. 38 — 44 (1996).

• Y-M Chou, A. M. Polansky, and R. L. Mason, "Transforming Non-Normal

Data to Normality in Statistical Process Control," Journal of Quality
Technology, Vol. 30(2), pp. 133 —141 (1998).

• D. Bakker and D. Icke, " Lithography Defect Characterization Using

Statistical Process Control and Automated Inspection," Proceedings of the
KTI Microelectronics Seminar, pp. 361— 367 (1990).

• M. E. Clarke and K-S. Cheng, "New Photochemical Filtration Technology

for Process Improvement," Olin Microlithography Seminar poster session
(1997).

• Y-T. Fan, H-P. Lin, Y-C. Lo, C-H. Lee, and K-L. Lu, "A Study on
Methods to Reduce Metal Defects Caused by a Coating Process," Olin
Microlithography Seminar poster session (1997).

• E. H. Bokelberg, J. L. Goetz, and M. E. Pariseau, "Photocluster Defect

Learning and Develop Process Optimization," Olin Microlithography
Seminar, pp. 127 — 139 (1996).

• L. Galbraith and A. Neukermans, "Contamination Detection on

Semiconductor Wafers," SPIE Vol. 774, pp. 13 — 20 (1987).

LINEWIDTH METROLOGY:

• K. M. Monahan, M. Davidson, Z. Grycz, R. Krieger, B. Sheumaker, and R.

Zmrzli, "Low-Loss Electron Imaging and its Application to Critical
Dimension Metrology," SPIE Vol. 2196, pp. 138 — 144 (1994).

• R. R. Hershey and M. B. Weller, "Nonlinearity in Scanning Electron

Microscope Critical Dimension Measurements Introduced by the Edge
Detection Algorithm," SPIE Vol. 1926, pp. 287 — 294 (1993).

• J. I. Goldstein, D. E. Newbury, P. Echlin, D. C. Joy, C. Fiori, and E.

Lifshin, Scanning Electron Microscopy and X-Ray Microanalysis, 2nd
edition, Plenum Press, New York (1984).

• J. Finders, K. Ronse, L. Van den Hove, V. Van Driessche, and P.

Tzviatkov, "Impact of SEM Accuracy on the CD-control During Gate
Patterning Process of 0.25 mm Generations," Proceedings of the Olin
Microlithography Seminar, pp. 17 — 30 (1997).

• K. M. Monahan, J. P. H. Benschop, and T. A. Harris, "Charging Effects in

Low-voltage SEM Metrology," SPIE Vol. 1464, pp. 2-9 (1991).

• L. Reimer, Image Formation in Low-Voltage Scanning Electron

Microscopy, SPIE Vol. TT12, SPIE Press, Bellingham, WA (1993).

• M. Davidson and N. T. Sullivan, "An Investigation of the Effects of

Charging in SEM Based CD Metrology," SPIE Vol. 3050, pp. 226 — 242
(1997).

• C. M. Cork, P. Canestrari, P. DeNatale, and M. Vasconi, "Near and Sub-

Half Micron Geometry SEM Metrology Requirements for Good Process
Control," SPIE Vol. 2439, pp. 106 —113 (1995).

• S. R. Rogers, "New CD -SEM Technology for 0.25 mm Production," SPIE

Vol. 2439, pp. 353 — 362 (1995).

• NIST SRM standards are available from the Office of Standard Reference
Materials, NIST, EM 205, Gaithersburg, Md. 20899. Their phone number
is (301) 975-6776. Information can be obtained from the NIST web site,
www.nist.gov.

• M. T. Postek, A. E. Vladar, S. Jones, and W. J. Keery, "Report on the

NIST Low Accelerating Voltage SEM Magnification Standard
Interlaboratory Study," SPIE Vol. 1926, pp. 268 — 286 (1993).

• M. T. Postek, "Scanning Electron Microscope-based Metrological Electron

Microscope System and New Prototype Scanning Electron Microscope
Magnification Standard," Scanning Microscopy, 3(4), pp. 1087 — 1099
(1989).

• Y. Nakayama and K. Toyoda, "New Submicron Dimension Reference for

Electron-Beam Metrology System," SPIE Vol. 2196, pp. 74 — 84 (1994).

• B. L. Newell, M. T. Postek, and J. P. van der Ziel, "Performance of the

Protoype NIST SRM 2090A SEM Magnification Standard in a Low-
Accelerating Voltage SEM," SPIE Vol. 2439, pp. 383 — 390 (1995).

• T. W. Reilly, "Metrology Algorithms for Machine Matching in Different

CD SEM Configurations," SPIE Vol. 1673, pp. 48 — 55 (1992).

• K. Phan, J. Nistler, and B. Singh, "Metrology Issues Associated with

Submicron Linewidths," SPIE Vol. 1464, pp. 424-437 (1991).

• E. E. Chain, M. G. Ridens, and J. P. Annand, "SPC Qualification Strategy

for CD Metrology," SPIE Vol. 2876, pp. 218 — 224 (1996).

• D. A. Chernoff, "Atomic Force Microscope (AFM) Analysis of Photoresist

Test Structures for use in SEM as In-House Linewidth Standards," SPIE
Vol. 2439, pp. 392 — 400 (1995).

• M. T. Postek, A. E. Vladar, and M. P. Davidson, "Fourier Transform

Feedback Tool for Scanning Electron Microscopes Used in Semiconductor
Metrology," SPIE Vol. 3050, pp. 68 — 79 (1997).

• L. J. Zych, G. Spadini, T. F. Hassan, and B. A. Arden, "Electrical Methods

for Precision Stepper Column Optimization," SPIE Vol. 633, pp. 98 — 105
(1986).

• L. W. Linholm, R. A. Allen, and M. W. Cresswell, "Microelectronic Test

Structures for Feature Placement and Electrical Linewidth Metrology," in
Handbook of Critical Dimension Metrology and Process Control, K. M.
Monahan, ed., SPIE Press, Bellingham, WA (1993).

• B. J. Lin, J. A. Underhill, D. Sundling, and B. Peck, "Electrical

Measurement of Submicrometer Contact Holes," SPIE Vol. 921, pp.
164 — 169 (1988).

• E. E. Chain and M. Griswold, "In-Line Electrical Probe for CD

Metrology," SPIE Vol 2876, pp. 135 — 146 (1996).

• S. M. Kudva and R. W. Potter, "Cost Analysis and Risk Assessment for

Metrology Applications," SPIE Vol. 1673, pp. 2— 13 (1992).

• R. R. Hershey and R. C. Elliott, "Procedure for Evaluating Measurement

System Performance: A Case Study," SPIE Vol. 2439, pp. 363 — 373
(1995).

• J. Engel and B. de Vries, "Evaluating a Well-Known Criterion for

Measurement Precision," J. Quality Technol., 29(4), pp. 469 — 476 (1997).

• D. C. Joy and C. S. Joy, "Low Voltage Scanning Electron Microscopy,"

Micron, Vol. 27, No. 3 — 4, pp. 247 — 263 (1996).

CONTROL OF OPERATIONS:

• Juran's Quality Control Handbook, 4th Edition, J. M Juran and F. M.

Gryna, eds., McGraw Hill, New York (1988).

• C. A. De Angelis, "ICI Advanced Materials Implements ISO 9000

Program," Quality Progress, Vol. 24, pp. 49 —51 (1991).

• H. J. Levinson, "ISO 9000: What Every Microlithographer Should Know,"

SPIE Vol. 2196, pp. 536 — 550 (1993).

• H. J. Levinson and C. DeHont, "Leading to Quality," Quality Progress, pp.

55 — 60 (1992).

• H. J. Levinson and J. Ben-Jacob, "Managing Quality Improvement on a

Development Pilot Line," Quality Management Journal, Vol. 3(2), pp.
16 — 35 (1996).

Aberrations, 102 Expected frequency, 14

Alignment, 53, 90, 105-107 Exponentially weighted moving
Amine contamination, 77 average, 133-135
Antireflection coatings, 66, 122 Exposure dose, 63-66
Atomic force microscope, 149
Autocorrelation function, 35 False alarms, 5
Automatic process control, 135-138 Focus, 16-17, 67-72
Average run length, 5-6 Focus, capacitance, 68
Focus, pressure, 68
Bake temperature, 72-74 Focus metrology, 70
Barometric pressure, 72 Focus, optical, 68
Funnel experiment, 44-46
Capability index, 20-23
Causality, 57 Golden wafers, 53
Cause-and-effect diagram, 57-59 Grid registration errors, 95
Central-limit theorem, 2-3
Chemical-mechanical polishing, 106, Holy wafers, 53
155 Hotplate, 74
Chi-squared test, 13-15, 30-31 Humidity, 67, 75-76
Chillplate, 73
Cluster defects, 125-126 Interfield overlay errors, 88-94
Confidence level, 29-32 Intrafield overlay errors, 87-94, 98,
Control limits, 8-9 102, 155
Control of operations, 161-168 Ishikawa diagram, 57
Control charts, 32 ISO 9000, 165-168
Correlations, 35-36
Cumulative probability, 11 Kill ratio, 124
Kurtosis, 16-19
Defect detection, 141-143
Defect monitors, 122 Latent image, 137
Defects, 120-128 Least-squares method, 93, 98
Defocus, 16-17 Lens distortion, 96-97
Degrees of freedom, 29-30 Lens placement errors, 95, 99, 102
Deterministic quantities, 1 Lepokurtic distributions, 16
Dispense volume, 67 Light scattering, 121
Documentation, 163-164 Line, 17
DUV resists, 77-78 Linewidth, 17
Linewidth control, 57-86
Eight-in-a-row test, 32, 36 Linewidth monitor, 52
Electrical linewidth metrology, Linewidth measurement, 33, 53,
150-152 143-152
Exhaust, 67 Linewidth variations, 49

189

Registration, 87, 94, 103

Resist coatings, 122
Resist development, 74-75
Resist processing, 126-128
Resist thickness, 8-10, 35, 53, 66-67
Reticle errors, 78-79, 92, 94, 103

190

For Training Purposes Only: ASML Confidential
No ratings yet
For Training Purposes Only: ASML Confidential
56 pages
12 TT l3 l3 Me Book
No ratings yet
12 TT l3 l3 Me Book
120 pages
Reticles: For Training Purposes Only
No ratings yet
Reticles: For Training Purposes Only
14 pages
09 TT L3 WSbook
No ratings yet
09 TT L3 WSbook
144 pages
11 - Ap L1 Iqc
No ratings yet
11 - Ap L1 Iqc
65 pages
Ayout Aintenance AND Djustments: Training Module
No ratings yet
Ayout Aintenance AND Djustments: Training Module
50 pages
18f AP L1 MachineDiagnosticLogging
No ratings yet
18f AP L1 MachineDiagnosticLogging
36 pages
03 TT l3 WH - Book
No ratings yet
03 TT l3 WH - Book
146 pages
Twinscan Computer System: TT-L2-CS Periodic Maintenance
No ratings yet
Twinscan Computer System: TT-L2-CS Periodic Maintenance
26 pages
R H M A: Eticle Andling Aintenance AND Djustments
No ratings yet
R H M A: Eticle Andling Aintenance AND Djustments
134 pages
Microlithography: Science and Technology Third Edition Smith Online PDF
No ratings yet
Microlithography: Science and Technology Third Edition Smith Online PDF
123 pages
Advances in Optical Form and Coordinate Metrology
No ratings yet
Advances in Optical Form and Coordinate Metrology
229 pages
L 2 P M T - A: Evel Eriodic Aintenance Raining Irmounts
No ratings yet
L 2 P M T - A: Evel Eriodic Aintenance Raining Irmounts
28 pages
L 2 P M T - R H: Evel Eriodic Aintenance Raining Eticle Andler
No ratings yet
L 2 P M T - R H: Evel Eriodic Aintenance Raining Eticle Andler
74 pages
P M T - S R S: Eriodic Aintenance Raining Canning Eticle Tage
No ratings yet
P M T - S R S: Eriodic Aintenance Raining Canning Eticle Tage
84 pages
03 TT L2 LO Module
No ratings yet
03 TT L2 LO Module
86 pages
L 2 P M T - S R S: Evel Eriodic Aintenance Raining Canning Eticle Tage
No ratings yet
L 2 P M T - S R S: Evel Eriodic Aintenance Raining Canning Eticle Tage
84 pages
P M - R H: Eriodic Aintenance Training Eticle Andler
No ratings yet
P M - R H: Eriodic Aintenance Training Eticle Andler
78 pages
12d - AP-L1-SL - Directory Structure & UNIX Commands
No ratings yet
12d - AP-L1-SL - Directory Structure & UNIX Commands
79 pages
12b AP L1 SL MachBasedDiagnosticSystem
No ratings yet
12b AP L1 SL MachBasedDiagnosticSystem
18 pages
Mark Designs: For Training Purposes Only
No ratings yet
Mark Designs: For Training Purposes Only
18 pages
L 2 P M T - S L: Evel Eriodic Aintenance Raining Ystem Ayout
No ratings yet
L 2 P M T - S L: Evel Eriodic Aintenance Raining Ystem Ayout
74 pages
Stigmatic Optics
No ratings yet
Stigmatic Optics
308 pages
MONA: Merging Optics and Nanotech
No ratings yet
MONA: Merging Optics and Nanotech
273 pages
TT PM LAY 3v1
No ratings yet
TT PM LAY 3v1
86 pages
17 AP L1 HI v2 (Read Only)
No ratings yet
17 AP L1 HI v2 (Read Only)
254 pages
06 AP L1 Imaging Overlay
No ratings yet
06 AP L1 Imaging Overlay
57 pages
ASML C Coding Standard
No ratings yet
ASML C Coding Standard
103 pages
11 TT L3 AM - Book
No ratings yet
11 TT L3 AM - Book
70 pages
Towards Ultimate Optical Lithography With NXT:1950i Dual Stage Immersion Platform.
No ratings yet
Towards Ultimate Optical Lithography With NXT:1950i Dual Stage Immersion Platform.
12 pages
Semiconductor Measurements and Instrumentation
No ratings yet
Semiconductor Measurements and Instrumentation
288 pages
Numerical Aperture in Microlithography
No ratings yet
Numerical Aperture in Microlithography
74 pages
I12 RF Ercdlst 251 E1
No ratings yet
I12 RF Ercdlst 251 E1
341 pages
Julia Zugby Masters Report
100% (1)
Julia Zugby Masters Report
40 pages
MCSV 100
No ratings yet
MCSV 100
208 pages
MCSV 310
No ratings yet
MCSV 310
33 pages
Coolend 019013674
No ratings yet
Coolend 019013674
45 pages
SL45 - Ch1-How To Protect Reticles From Electrostatic Damage
No ratings yet
SL45 - Ch1-How To Protect Reticles From Electrostatic Damage
67 pages
ASML PAS 5500 User Manual
No ratings yet
ASML PAS 5500 User Manual
11 pages
Mask Design For Single and Double Exposure Optical
No ratings yet
Mask Design For Single and Double Exposure Optical
191 pages
Applied Materials - SPIE Advanced Lithography + Patterning 2024 - Investor and Media Presentation
No ratings yet
Applied Materials - SPIE Advanced Lithography + Patterning 2024 - Investor and Media Presentation
21 pages
12e - AP L1 SL Software Architecture
No ratings yet
12e - AP L1 SL Software Architecture
52 pages
Mram 2 Dieny PDF
No ratings yet
Mram 2 Dieny PDF
25 pages
I12 RF Elcirc R10 B1 2
No ratings yet
I12 RF Elcirc R10 B1 2
400 pages
Low-Cost LIDAR Development Report
No ratings yet
Low-Cost LIDAR Development Report
67 pages
5.etch - Part 3 (2) (21613)
No ratings yet
5.etch - Part 3 (2) (21613)
31 pages
3-Process-flow-2023-12-7-BEOL V1-Handout
No ratings yet
3-Process-flow-2023-12-7-BEOL V1-Handout
109 pages
Photomask Design: Ee-527: Microfabrication Ee 527: Microfabrication
No ratings yet
Photomask Design: Ee-527: Microfabrication Ee 527: Microfabrication
56 pages
3.lithography - Part 2 (3) (17508)
No ratings yet
3.lithography - Part 2 (3) (17508)
33 pages
Photolithography Process Overview
No ratings yet
Photolithography Process Overview
16 pages
TSMC PDK Custom Device Guide
No ratings yet
TSMC PDK Custom Device Guide
13 pages
CD Sem
No ratings yet
CD Sem
20 pages
Semiconductor
No ratings yet
Semiconductor
389 pages
Lock in Amp Redpitaya PDF
No ratings yet
Lock in Amp Redpitaya PDF
11 pages
Photo Lithography
No ratings yet
Photo Lithography
112 pages
Photoresist in Photolithography
No ratings yet
Photoresist in Photolithography
3 pages
25975904
No ratings yet
25975904
93 pages
ODB Format Description v7 PDF
No ratings yet
ODB Format Description v7 PDF
295 pages
Practical Appication of Infrared
83% (6)
Practical Appication of Infrared
193 pages
The Basics of Spectros
100% (7)
The Basics of Spectros
133 pages
G7 Math Q3 - Week 8 - Classification of Polygons
No ratings yet
G7 Math Q3 - Week 8 - Classification of Polygons
24 pages
Journal Homepage: - : Introduction
No ratings yet
Journal Homepage: - : Introduction
11 pages
Keygen 2014
33% (12)
Keygen 2014
6 pages
Expansion Joints Data
No ratings yet
Expansion Joints Data
52 pages
Conveyor Belt Specs for Engineers
No ratings yet
Conveyor Belt Specs for Engineers
1 page
Africa 2030 PDF
No ratings yet
Africa 2030 PDF
212 pages
Lessons For The Pharmacy Technician
No ratings yet
Lessons For The Pharmacy Technician
129 pages
Topic Two: Knowledge Representation, Reasoning and Inference
No ratings yet
Topic Two: Knowledge Representation, Reasoning and Inference
72 pages
Architecture of Cape Verde - Wikiwand
No ratings yet
Architecture of Cape Verde - Wikiwand
11 pages
3D Molecular Structure Analysis Tools
No ratings yet
3D Molecular Structure Analysis Tools
69 pages
Definity: Enterprise Communications Server
No ratings yet
Definity: Enterprise Communications Server
2,350 pages
L5 Flew and Mitchell
No ratings yet
L5 Flew and Mitchell
4 pages
Rhodoline 642 PDF
No ratings yet
Rhodoline 642 PDF
2 pages
Family Law Post-Judgment Guide
100% (1)
Family Law Post-Judgment Guide
28 pages
Business Research Methods Guide
No ratings yet
Business Research Methods Guide
9 pages
MaaS PPT. 27.08.2025 - Rev
No ratings yet
MaaS PPT. 27.08.2025 - Rev
51 pages
PDHID Manual
No ratings yet
PDHID Manual
25 pages
TOEIC Part 5 Detailed Solutions
No ratings yet
TOEIC Part 5 Detailed Solutions
32 pages
National Math Competition 2018 Problems
No ratings yet
National Math Competition 2018 Problems
4 pages
100r06 FUNTIONAL CHECK
No ratings yet
100r06 FUNTIONAL CHECK
26 pages
Homework Psychology Definition
100% (1)
Homework Psychology Definition
5 pages
27week Hswalk
No ratings yet
27week Hswalk
6 pages
IGCSE English Revision Guide Extended
100% (1)
IGCSE English Revision Guide Extended
30 pages
Danfoss Series 90 Pump and Motor Guide
100% (1)
Danfoss Series 90 Pump and Motor Guide
34 pages
Characteristics of Indian Philosophy
No ratings yet
Characteristics of Indian Philosophy
13 pages
321B Excavator Hydraulic System: Kga1-Up AKG501-UP 9CZ1001-UP
No ratings yet
321B Excavator Hydraulic System: Kga1-Up AKG501-UP 9CZ1001-UP
2 pages
Frederick Schauer, The Proof - Zebras, Horses and The Nature of Inference
No ratings yet
Frederick Schauer, The Proof - Zebras, Horses and The Nature of Inference
31 pages
The Role of HR in Building Sustainable Organizations
No ratings yet
The Role of HR in Building Sustainable Organizations
1 page
Exotic Magazine June 2011 PDF
No ratings yet
Exotic Magazine June 2011 PDF
2 pages
12 Year Curriculum Plan (Scope and Sequence)
100% (14)
12 Year Curriculum Plan (Scope and Sequence)
3 pages