the semiconductor industry experiences rapid technological changes,
new products and processes are continuously developed while technology
becomes ever more complicated and precise. These changes give rise
to the need for increasing numbers of manufacturing steps. In addition,
with the introduction of 300-mm production, human intervention alone
can no longer control semiconductor processes. Consequently, automation
has become widespread in IC fabs, expanding new technology horizons.
Nevertheless, it is still necessary to reduce manufacturing costs
and increase return on investment. Among other things, the semiconductor
industry must decrease the use of nonproduct wafers, increase overall
equipment effectiveness, reduce defects, improve product quality and
yields, meet specifications for reduced feature sizes, improve production
efficiency, and achieve faster process development and yield learning.
Achieving these and other objectives requires the implementation
of process and equipment modeling, monitoring, and control, which
are used widely in many industries and are beginning to take hold
in the semiconductor industry, where such methods are commonly called
advanced process control (APC) and advanced equipment control.
As device features shrink to 0.15 µm and below and 300-mm fabs
proliferate, APC systems are becoming increasingly important. From
the outset of 300-mm manufacturing, productivity improvements and
fab effectiveness have been key considerations. To ensure the profitability
of these new fabs, productivity gains must reach 3035%.
Diagnosis, monitoring, and APC are needed to predict when tools require
maintenance and to avoid unscheduled downtime. These control strategies
also should reduce the mean time to repair. APC enables tools to perform
with greater consistency, repeatability, and precision, leading fabs
to increase production efficiency and reduce their total cost of tool
Even 200-mm fabs are discovering that the implementation of such
control strategies is necessary to extend the lifetime of their process
tools. In 200-mm fabs, implementing APC allows manufacturers to integrate
new components such as sensors so that they can ensure that their
tools keep pace with new technologies.
|Figure 1: A two-dimensional T2 control
chart showing error signatures for two detected errors with two
variables, one of which is a function of the other.
Given the complexity of semiconductor processes and the need to adapt
rapidly to technological changes, it is critical to choose optimal
process and equipment modeling, monitoring, and control methodologies.
This article, based on joint work between Si Automation (Montpellier,
France) and STMicroelectronics (Crolles, France), explores the pros
and cons of various process control strategies, focusing on global
process control (GPC), a system for performing fault identification
based on statistical and probabilistic criteria and for achieving
automatic fault correction. GPC can be applied to process tool control
through the use of data collector tools.
Comparing Process Control Methods
There are two major types of APC systems: statistical process control
(SPC) and model-based process control (MBPC). SPC can be broken down
into univariate and multivariate fault detection methods.
Fault Detection through Univariate SPC. Univariate SPC systems
are based on the idea that variations of a controlled variable (variations
that affect a process all the time and are essentially unavoidable
within that process) have a common cause. SPC allows the normal (common-cause)
variation of a controlled variable within acceptable limits and detects
any assignable cause of abnormal variation of the controlled variable
as soon as possible. The technique is characterized by the creation
of control charts with a target value and upper and lower control
limits that plot individual monitored variables.
Because univariate SPC leads to the creation of as many control charts
as monitored variables, it is of no value at the fab level, where
a huge number of variables must be monitored, and of little value
at the tool level because an engineer can monitor a maximum of only
four charts at the same time. Moreover, of all control methodologies,
univariate SPC results in the most undetected errors and false alarms,
leading to wasted time. Also, the monitored variables must theoretically
be independent, which is usually not the case in the semiconductor
Fault Detection through Multivariate SPC. Multivariate SPC
is superior to univariate SPC. Rather than create separate control
charts for each variable, multivariate SPC uses global control charts
that represent as many variables as needed. Moreover, multivariate
SPC is much more precise than univariate SPC in detecting errors and
avoiding false alarms. Finally, multivariate SPC takes into account
the dependencies among monitored variables.
Multivariate methodologies have their limitations, however. While
they generate almost no false alarms or undetected errors, they cannot
indicate the origin of errors and are thus unable to automatically
manage error occurrences. Detecting error origins with these methodologies
can be so complex that engineers either require a very high level
of expertise and intuition to use them or cannot use them at all.
Model-Based Process Control. MBPC directly links product variables
and process variables to models. This control method tries to determine
the variables of a process from the characteristics of a desired output.
Inversely, it also tries to predict product output from the process
variables. These two applications are the basis for feed-forward,
feed-backward, and real-time control. The power of the models is their
ability to compare desired and real output variables.
Building models for MBPC requires a good knowledge of equipment and
processes. While performing MBPC on one type of equipment may provide
good results, doing so on another may provide unsatisfactory results.
In all cases, performing MBPC requires that customers learn the method
or that suppliers learn the processes, resulting in slow, complicated,
and relatively inflexible implementation. But the main limitation
of MBPC is theoretical: it acts on measured variables to obtain a
desired output but gives no indication about the underlying causes
of observed deviations. Finally, MBPC corrects the symptoms of errors
but not the errors themselves, resulting in fewer out-of-control outputs
but more-significant problems.
An optimal control methodology must be able to perform automatic
fault identification and single out anomalies when detected errors
do not fit existing classifications. Few available tools indicate
error types, let alone errors themselves, and those that do require
that engineers analyze monitored profiles when faults are detected.
That procedure can lead to misjudgments because profiles are not always
obvious. Moreover, engineer intervention is time-consuming and ineffective
when new faults appear.
To identify faults, some methods associate a particular signature
with an out-of-control event. Each out-of-control event must be compared
with a set of signatures to identify its origin. Considering the great
number of candidate signatures, it is difficult to apply this method
in real time. Engineer subjectivity can also be a source of errors,
because the relationship between the signature and a multidimensional
observation often is not clear.
Because no single control solution provides adequate results, a combination
of different solutions is needed. Real-time control (starting with
run-to-run and wafer-to-wafer controls) requires automatic forward
or backward error correction. But an automatic correction system presupposes
the existence of automatic fault detection and classification. Fault
detection must be able to function as a plug-in module that fits the
central APC system and is easy to reuse. A modeling application necessarily
has limited relevance unless an efficient fault identification system
is established. On the other hand, only MBPC can efficiently automate
the correction of anomalies.
Global process control was designed to carry out these multiple tasks.
A subtle combination of multivariate SPC, MBPC, and mathematical procedures,
GPC combines the advantages of all these methodologies while avoiding
their inherent drawbacks.
The use of multivariate control charts for fault detection is essential
for extracting relevant information from the large amounts of data
generated by equipment and processes. This method reduces the frequency
of irrelevant alerts and detects faults that are invisible on common
SPC charts. Moreover, multivariate methods solve the problem of fault
identification. Because the origin of a fault is identified in real
time, fab personnel can take corrective action quickly and limit rejected
GPC is performed by following a sequence of steps. First, a GPC error
detection chart is created on the basis of Hotelling's T2 multivariate
SPC methodology. This chart enables engineers to monitor all error
occurrences in a single chart and takes new measurements into account.
Next, the GPC methodology adds signatures to already known errors.
These signatures are linear tracks on the T2 chart. The establishment
of signatures makes it possible to determine whether an error fits
an already known error type or whether it is a new type. Figure 1
shows error signatures in a two-dimensional T2 control chart for two
detected errors. The x- and y-axes depict two different variables,
both of which have absolute values. The red circles represent one
type of error, the red triangles the other; the blue crosses correspond
to controlled measurements. The ellipse is the graphical representation
of the T2 control limit. The numbered plots represent products. This
figure shows how linear tracks can be represented.
To show the intensity and the variability of an error, the next GPC
step is to create a specific error control chart for each error signature,
which enables engineers to monitor error types that have already been
GPC then uses a factorial methodology to determine the correlation
of each variable to these signatures (or the impact of these variables
on the signatures). This step has two major advantages: it automatically
determines the origin or cause of errors and also provides a model
to determine the impact of each variable on the errors, enabling engineers
to determine which parameters have to be modified to maintain a controlled
In the event of a new error type, the linear track or signature of
the newly encountered error is detected automatically. As illustrated
in Figure 2, an associated specific error control chart is then created
for this new error signature. This two-dimensional figure, an extension
of the data in Figure 1, demonstrates that new measurements will either
be in control (inside the ellipse) or out of control (outside the
ellipse) and that the out-of-control measurements will appear either
on the existing linear tracks or on other linear tracks. Each linear
track corresponds to an error type.
|Figure 2: Two-dimensional
chart extending the data in Figure 1. The new (larger) measurements
can either be in control (inside the ellipse) or out of control
(outside the ellipse). Out-of-control measurements appear on the
existing linear tracks and on other linear tracks; each linear
track corresponds to an error type.
|Figure 3: Control chart for two different variables
(on the x-axis and y-axis, respectively) indicate whether a measurement
should be viewed as in control or out of control in the T2 calculation;
95% of controlled variables fall within the small ellipse and
95% of uncontrolled variables fall within the large one.
GPC's probabilistic approach to factorial analysis can discriminate
between variables by identifying whether a variation associated with
a given measurement is exceptional or ordinary. GPC also provides
important information on which variables are anomalous. In Figure
3, a control chart for two different variables (on the x-axis and
y-axis, respectively) highlights specific problems. These data enable
engineers to determine whether a measurement should be viewed as out
of control in the T2 calculation. The truncation limits in the figure
represent measurement points that are outside the control limits.
By carrying out the different GPC steps, engineers can completely
automate the detection and identification of errors. Using this method,
models can be created so that when errors are detected, corrective
action can be performed manually after an alarm is sounded or automatically
with the aid of the obtained models.
Implementing GPC on the Fab Floor
To determine GPC's process control capabilities, a test involving
a chemical vapor deposition (CVD) tool was conducted at STMicroelectronics.
A SilverBox from Si Automation provided equipment and sensor integration
with the CVD tool and performed data collection. To obtain models
for automatic fault detection and identification, the GPC Scan software
module analyzed data used for automation and data collected during
To implement automatic fault detection, classification, and identification
(FDCI), information was first collected by the SilverBox through its
equipment, peripheral, and sensor interfaces, as well as from its
automation part. The SilverBox then prepared this information for
the GPC methodology by transforming it into homogeneous data following
Automatic FDCI was obtained from this rescaled data by using the
SilverBox's GPC Guard. The GPC Guard used the results obtained by
the GPC Scan and was responsible for fault detection by calculating
the T2, for fault identification by classifying the error signatures,
and for launching corrective action. The GPC Scan also prepared models
for the GPC Guard. On historical data, it carried out fault detection
by calculating the T2, fault classification by identifying the error
signatures, and fault identification by iterative factorial analysis.
|Figure 4: Global control
chart for error detection enabled engineers to monitor errors
globally while minimizing undetected errors and false alerts.
|Figure 5: Factorial control chart of two variables
before error classification. This chart enabled engineers to determine
whether measurements should be considered in control or out of
control in the T2 calculation step. The ellipse corresponds to
the control limit and truncation limits show out-of-range measurements.
Initial Univariate Analysis. The test was based on the GPC
analysis of 11 variables collected during the process. Initially,
11 univariate SPC control charts of the monitored variables were prepared
in which the variables were scaled on normal law for easier interpretation.
Each chart displayed the absolute values of a given variable over
a series of wafers and lots. The center of each graph showed the average
of all the measurements for the given variable.
Fault Detection. A T2 control chart was then built for the
11 variables, which enabled the engineers to monitor errors globally
while minimizing undetected errors and false alerts. Figure 4, the
global control chart for error detection, displays the wafer numbers
on the x-axis and the distance of the measurements from the center
of the controlled distribution on the left-hand side of the y-axis.
The control limit was set up to 5% (associated probabilities are shown
on the right-hand side of the y-axis). Variables in the range of 65
and above are shown on the truncation line.
Fault Classification. Next, the GPC Scan performed an iterative
factorial analysis to classify errors by isolating presumed out-of-control
observations at each iteration. Figure 5 is an example of a factorial
control chart for two variables before error classification. The chart
shows one compiled variable as a function of another compiled variable.
The ellipse corresponds to the control limit and truncation limits
show out-of-range measurements. This and other factorial control charts
enabled engineers to determine whether measurements should be considered
in control or out of control in the T2 calculation step.
The iterative factorial analysis step led to the error classification
step. As shown in Figure 6, the iterative process allowed engineers
to classify errors to obtain a stable classification. At first, each
iteration resulted in a new error class. But during the fifth iteration,
two classes were gathered into one when the iteration determined that
they belonged to the same class. Other error classes were then determined,
after which three of them were gathered into one and split again into
two. Since subsequent iterations did not produce different error classes,
the engineers determined that the classification procedure was concluded.
The result was six classes of errors.
Figure 7 presents a factorial control chart after error classification.
As in Figure 5, this chart represents a compiled variable as a function
of another compiled variable. Truncation limits are also shown. This
time, however, all out-of-control measurements were determined. Most
of them lay far from the controlled zone, which was much thinner than
at the beginning of the iterative process. The out-of-control measurements
that lay inside the controlled zone would have fallen outside it if
the corresponding relevant compiled variables had been used to draw
the chart. This chart is important because it shows that no controlled
measurements lie outside the controlled zone.
|Figure 6: Chart showing
the classification step, which allowed engineers to obtain a stable
|Figure 7: Factorial control chart after error
classification in which all out-of-control measurements, most
of which lay far from the controlled zone, were determined. The
chart shows error class INDHCA (the stabilized classification
in Figure 6).
The error classification process led to the creation of specific-error
control charts for each type of error. These charts were used to focus
on a given anomaly. From the useless initial univariate control charts
for individual variables, GPC Scan built control charts by error type.
Apart from focusing on a given anomaly, these control charts differentiated
between errors of the same type according to their intensity and variability.
Specific-error control charts for the first three detected error
types in this test are presented in Figure
8. These charts show the distance of the measurements from the
center of the controlled distribution. In each chart, the measurements
correspond to the linear track of the specific error type; the results
represent the distance of the measurements from the center of the
distribution of the linear track in question. In each chart, blue
crosses correspond to controlled measurements, red disks to the concerned
error, and red circles to other error types. The x-axis shows the
wafer number while the left-hand y-axis shows the absolute value of
the distance from the center and the right-hand the associated probabilities
Fault Identification. Based on the factorial analysis, the
GPC Scan then determined the origin of the detected errors for each
of the six error classes. That step led to the creation of sensitivity
charts for each detected error type. Figure
9 shows the sensitivity of each detected anomaly to the different
variables. On the y-axis, 0 means that the variable had no impactin
other words, that it was not responsible for the type of error shown
in the chart; 1 means that the error type was perfectly correlated
to the variable. These charts typically show the profile of the different
Now that a complete profile of errors existed, the GPC software was
ready to interpret error origins. Figure 10, profiles of out-of-control
variables, shows corresponding values of the variables for each wafer
detected to be out of control. Wafers with the same error types were
gathered together. Variables that fall between the red lines were
|Figure 10: Profiles of out-of-control variables
showing corresponding values of the variables for each wafer detected
to be out of control. Variables that fall between the red lines
were in control. (MOYVCR = centered and reduced mean of variable.)
Decision Matrix Construction. In addition to factorial analyses,
the sensitivity charts made it possible to create process control
models, or decision matrices, which associated actionswhether alarms
or automatic correctionswith identified errors. Each type of error
had a signature, which was its linear track on the T2 graph and which
corresponded to its sensibility chart. A decision matrix associates
out-of-control action plans with each signature. Depending on the
value of the measurements, these plans could be anything from alarms
to no-action commands to a complete set of automated actions. Decision
matrices allow engineers to control or correct errors automatically.
Applying GPC to Real-Time Control. After the GPC procedure,
from the creation of the global T2 control chart to the construction
of the decision matrix, had been performed off-line by the GPC Scan,
it could then be applied to process tool control in real time. The
GPC Guard software module uses the obtained process control models
and decision matrices for run-to-run or wafer-to-wafer control. This
reaction delay was chosen for the first versions of GPC Guard, even
though GPC can theoretically be applied in real time. But compared
with the huge delays for obtaining results from other available methodologies
(days sometimes), run-to-run and wafer-to-wafer control is very close
to real-time control.
Like any APC solution, GPC confronts the problem of implementation
in the real world. Although many standards established by SEMI for
equipment manufacturers and by Sematech for IC manufacturers regulate
fab life, many fab areas remain beyond the scope of these standards,
especially those that connect the SEMI world to the Sematech world.
Consequently, the semiconductor industry has experienced several APC
solutions, none of which has been sufficiently generic to offer an
efficient solution for a broad range of applications.
Problems associated with implementing APC are manifold. How should
relevant data be acquired and different data matched? On what platform
should data acquisition and matching take place and how should the
results be used? What type of control system has the least impact
on manufacturing and product yields?
The test discussed in this article demonstrates that the GPC approach
can offer a solution to many APC-related problems. While the test
was performed in a pilot project for a CVD application, the results
demonstrated that GPC could identify all out-of-control variables
as the causes of equipment or process deviations. The use of the GPC
classification tool permitted engineers to focus on relevant problems
and ignore the signatures of second-order ones. For example, the signature
of a first-wafer effect in dry etch, which was detected through GPC,
could be automatically ignored in-line.
The flexible configuration possibilities of the data treatment phase
of the GPC test adapted well to the customer's needs. Although the
test was performed on dry etch equipment, the customer could quickly
configure the GPC software to suit other types of tools.
This rapid reconfiguration capacity also allowed the customer to
adapt GPC to the changing needs of the dry etch equipment. GPC was
first applied to the critical phase of the dry etch process. After
it proved its ability to detect and analyze all types of known errors
in that phase, automatic corrective actions were integrated into the
model, decreasing the need for human intervention. GPC also detected
new error types, leading to its extension to all other process phases.
For example, GPC detected and automatically corrected an error in
the precritical phase caused by a problem in loading a particular
gas. An error caused by an unloading problem in the last phase of
the process also was detected and automatically corrected.
Although GPC allows real-time fault identification in an industrial
context, automatic fault detection is not enough in the semiconductor
industry; corrective action also must be taken. After identifying
faults, GPC can automatically correct them. Actions taken to correct
one fault can be used as a model to correct other faults of the same
type. Most importantly, the GPC approach leads to the creation of
corrective models based on the results of fault identification.
François Pasqualini joined STMicroelectronics's
central R&D department in 1999. He is SPC manager in charge of
statistical process control, advanced process control, and applied
statistics. From 1983 to 1988 he was a research engineer in gallium
arsenide technology at Philips Research Laboratories in Limeil-Brevannes,
France. From 1989 to 1999 Pasqualini worked at IBM Microelectronics
Division in Corbeil-Essonnes, France; his interests were in the field
of line monitoring, defect density, yield modeling, design of experiments,
and statistical process control. He is a member of the French Statistical
Society and of the Statistical Group of French Quality Organization.
He received engineer diplomas from the National Polytechnic Institute
of Engineering (Toulouse, France) in the physical chemistry of materials
in 1981 and in electronics in 1982. (Pasqualini can be reached at
+33 476 925035 or email@example.com.)
Martial Baudrier joined the central R&D
department of STMicroelectronics (Crolles, France) in 2000. He is
mainly involved in APC projects, especially in the field of multivariate
approach developments. He received an engineer diploma in statistics
and functioning safety from the Institute of Engineering Sciences
& Techniques of Angers. (Baudrier can be reached at +33 476 925438
Daniel Lafaye de Micheaux, PhD, joined Si Automation
(Montpelier, France) in 2001 when the company acquired a GPC license.
He is the company's statistical and probabilistic expert and has been
recognized for his contributions to quality and manufacturing. A professor
at Sofia Antipolis University (Nice, France), Lafaye de Micheaux patented
GPC in 1998. He received a PhD in statistics and has researched and
taught in that field for more than 25 years. (Lafaye de Micheaux can
be reached at +33 466 807288 or firstname.lastname@example.org.)
François Lemaire joined Si Automation
as product manager in 2000. Previously, he conducted research in R&D
process improvements in the fields of pharmaceuticals and aeronautics.
He received a diploma from the HEC School of Management (Paris) and
has specialized in statistics, probabilities, and econometrics. (Lemaire
can be reached at +33 467 033700 or email@example.com.)