etch tools using real-time fault detection and classification
Chen and T. F. Yen, ProMOS Technologies; and Barry Coonan,
manufacturers use process control methods and an analysis of tool sensor
outputs to improve yields, increase tool productivity, and reduce manufacturing
costs. Statistical process control (SPC) utilizes statistical algorithms
to detect excursions. In contrast, this article presents a novel fault
detection and classification (FDC) approach from Straatum (Dublin, Ireland)
that is based on a pattern-recognition algorithm. Outputting a chamber-status
metric known as the plasma index, this real-time FDC method is in place
at ProMOS Technologies (Hsinchu, Taiwan), a 200-mm manufacturing facility
that runs a variety of semiconductor tools. Based on a number of case
studies, this article discusses the use of Straatum's FDC method in
conjunction with several DRM oxide etch tools from Tokyo Electron (Tokyo).
behavior of plasma etch tools is difficult to control because small
variations in equipment and process setup can have an unpredictable
impact on etch quality. Many parameters have a very small margin of
deviation (<1%) before yields decline. Historically, methods to control
etch processes have relied on end-of-line metrology and short-loop test
wafers. While these methods have been used to detect process chamber
faults, they are slow and sometimes provide an unrealistic picture of
etch chamber behavior. Determining the root cause of chamber faults
has relied on engineer experience and interpretations of tool data—clearly
not the most efficient way to detect and remedy tool faults. Given the
wealth of data provided by process tools, more-sophisticated methods
have been developed, which are known collectively as advanced process
SPC, engineers can systematically interpret data from many tool sensors
using statistical rules that can track tool performance and detect fault
conditions in real time. This technique can be used to analyze data
in univariate or multivariate modes, in which excursions are defined
deviations from a statistical mean outside the normal limits of variation.
A drawback to this method is that the tool parameters in question often
do not display a normal distribution, so that such concepts as average
and standard deviation can only be approximate. This uncertainty can
lead to compromises that broaden fault limits, reducing sensitivity,
or that allow benign statistical outliers to be processed, giving rise
to false alarms.
FDC methodology at ProMOS uses a pattern-recognition algorithm to provide
a real-time interpretation of the state of the etch chamber.3
This approach overcomes the limitations of standard SPC, since it does
not assume a priori knowledge of the statistical nature of the data.
By learning the characteristics of a chamber fault using tool data,
engineers can create a library of fault patterns.
fault library is generated using data from a design of experiment performed
on the chamber. There is no limit to the number of faults that can be
added. When a wafer is run in the chamber, a fingerprint, or data pattern,
associated with the data received from tool sensor outputs is recorded
and compared with the patterns in the fault library. When a match is
found between the wafer fingerprint and an entry in the fault library,
a verdict can be reached as to whether the chamber is in a fault state
and what the most probable root cause of the fault is. The go/no-go
chamber decision is presented as a single value—the plasma index. If
the plasma index is greater than 1 or less than –1, the tool is
in a fault condition.
if a wafer is statistically outside the defined variation limit of a
particular measured parameter, its fingerprint must first match a known
fault before an alarm is triggered. Thus, normal sensor data variations
do not trigger false alarms, and the number of false alarms caused by
benign outliers is reduced.
quality of an etch process is principally driven by the condition of
the plasma: Changes in chemistry resulting from variations in the gas-flow
rate, the deterioration of chamber parts, pressure deviations, or power
or electrode gaps all affect the plasma and, therefore, etch characteristics.
Nonintrusive measurements of the plasma state from instruments such
as impedance sensors or optical emission spectroscopes can provide rich
information that can be used for process monitoring. In the FDC technique
discussed here, the pattern-recognition model is based on data from
an impedance plasma sensor from Scientific Systems (Dublin, Ireland),
although the method can use data derived from other tool sensors or
combinations of sensor data.
impedance sensor measures the fundamental and harmonic voltage, current,
and phase of the radio-frequency power signal to the chamber. Positioned
between the match and the chamber, it can detect changes in plasma condition
without interference from the match or other components in the transmission
line. The sensor does not affect process characteristics and has the
same form factor as the tool component that it replaces at installation.
1: Impedance sensor measurements of the magnetic interaction of
twin etch chambers. The measurements were taken over a time period
of approximately 60 seconds.
sensitive measurement data from an impedance sensor presented in Figure
1 demonstrate that the magnetic fields of the two adjacent chambers
in an etch system interact with each other and affect the system's impedance.
When only one chamber magnet rotates, the character of the impedance
fluctuation changes. While the plasma is affected by the orientation
of the chamber magnets, this phenomenon has no impact on etch performance.
2: Fundamental voltage measurements from the etch tool taken over
several preventive maintenance cycles. The red vertical lines mark
when preventive maintenance actions occurred.
2 shows how preventive maintenance (PM) cycles performed on the etch-tool
chamber influence the tool's fundamental voltage signal when measured
at 13.56 MHz. Although PMs cause the tool to exhibit nonnormal behavior,
their impact is benign. Hence, a statistical approach to understanding
them would be inefficient, leading to a compromise between sensitivity
loss and false alarms. In contrast, the FDC approach used at ProMOS
learns only those patterns associated with faults and thus ignores data
from events such as PM cycles.
3: Schematic diagram of the tool-monitoring data flow.
illustrated in Figure 3, data acquired from the sensor are sent to a
Straatum real-time (SRT) controller, which uses ImPrint MX software.
After acquiring data from the sensor, the software tests the wafer fingerprint
against the fault library using the FDC algorithm and calculates the
plasma index. The SRT has a connection to the tool/host SECS stream
to gather data-labeling information and a network connection through
which data are sent to a central fab server to permit factorywide access.
4: Software engineering view illustrating that wafers processed
when the chamber was in a fault condition lie outside the plasma
index alarm limits (central shaded band).
have access to the software via a touch screen at the tool. Information
available to equipment operators includes whether or not the tool is
in a fault state, what the probable root cause of a fault is, and a
list of wafers affected by the fault. Engineers have access to more-detailed
information on recent wafers run on the tool, including a calculated
multivariate analysis index and raw sensor data, as presented in Figure
4. A Pareto chart, as shown in Figure 5, is used to indicate the probable
root cause of an excursion and the approximate magnitude of the parameter
5: Fault classification Pareto chart showing the most likely root
cause of an excursion (left-hand bar). The shaded bar indicates
that the fault library parameter lies outside the alarm limits as
defined by the user. That bar represents the most likely root cause
of the failure.
to the fab server enables engineers to analyze data over a long period
so that they can detect tool-behavior trends that may indicate fault
conditions and compare data from different tools. Plasma index alarm
limits can be defined by the engineer using the software, which also
allows them to test the robustness of the FDC by archiving data to be
tested along with the alarm limits. To set the plasma index limits accurately,
the engineer must build the fault library so that it indicates how much
variation is required for a parameter to trigger a fault condition that
results in yield loss.
6: Software engineering view showing that affected wafers lie outside
the plasma index limits and a Pareto chart (inset) showing the fault
classification. The red plots represent the C5F8
flow deviation for these wafers. Because the fault impacts the ratio
of O2 to C5F8, O2 is
also reported to be out of spec. Once the C5F8
fault is rectified, the chamber returns to a good state for all
the FDC System
Study 1: C5F8 Gas-Flow Excursion.
In the first case study, the plasma index classification indicated a
chamber fault whose root cause was a C5F8
flow-rate problem associated with a mass-flow controller, as presented
in Figure 6. The magnitude of the deviation was found to be approximately
5%, which was outside the safe limits as defined by the user and a potential
source of yield loss. Subsequent end-of-line yield data, reflected in
the wafer maps in Figure 7, showed this to be the case. Knowing the
specific cause of the excursion enabled equipment engineers to respond
rapidly, minimizing tool downtime.
7: End-of-line yield map indicating that wafers 1 and 2 (circled
in red) experienced a yield-loss excursion that was detected and
classified by the pattern-recognition algorithm.
this case, the FDC system also indicated that O2 flow was
outside the normal process limits, as illustrated in the inset in Figure
6. This effect was a consequence of the C5F8 flow
deviation, which caused O2 partial pressure to shift and
to operate outside safe limits. Once the C5F8
flow had been restored to the correct level, O2 partial pressure
returned to normal. This example demonstrates the importance of addressing
the most significant, or left-most, parameter in the Pareto chart first,
even if more than one root cause is indicated.
Study 2: Pressure Control Excursion. In the second case study,
a yield-impacting excursion resulted from a malfunction in the chamber-pressure
control system. Engineers replicated the excursion by reducing the valve
angle at the chamber pumping port, thus increasing the chamber pressure.
The plasma index detected and classified the fault correctly, calculating
it to be close to 1.5 (i.e., a pressure deviation 50% greater than the
alarm limit). Figure 8 presents the pressure data and Figure 9 classifies
8: Data from a fault in the chamber-pressure control system (represented
by the red plots) that was clearly outside the plasma index limit
and, therefore, a potential yield limiter.
general, when a fault is detected and classified, it is not necessarily
the parameter control device but the particular chamber system that
is determined to be the source of the problem. In this case, the fault,
a damaged manometer cable, was correctly detected and classified by
the equipment engineers. In other situations, the cause may be classified
as a power fault while the actual faulty device may be the generator,
the match, or a transmission line.
9: Classification of the excursion presented in Figure 8 shows that
the root cause of the failure was a pressure fault. The red line
is a user-defined confidence cutoff limit, below which the confidence
level is too low to classify a fault accurately.
article has dealt with an FDC system in a 200-mm semiconductor manufacturing
environment. Unlike SPC methods, the FDC technique utilized here is
based on a pattern-recognition algorithm that ignores nonnormally distributed
chamber data. The system is therefore immune to drifts in sensor parameters
caused by normal chamber cycling. The software and data presentation
are user-friendly and intuitive, aiding tool operators as well as equipment
and process engineers who perform troubleshooting operations. The data
can be accessed throughout the fab over an internal network, enabling
engineers to perform intensive analyses.
examples of the technique presented here illustrate that real-time,
accurate detection and classification of excursions can decrease tool
downtime. Furthermore, the fault library can be expanded to include
all process and chamber hardware faults, offering fab personnel a comprehensive
detection and classification tool.
article is based on a presentation given at the Taiwan Semiconductor
Industry Association's Semiconductor Manufacturing Technology Workshop,
held September 9–10, 2004, in Hsinchu, Taiwan.
Fiorletta, "Capabilities and Lessons from 10 Years of APC Success,"
Solid State Technology 47, no. 2 (2004): 67–70.
Edgar et al., "Automatic Control in Microelectronics Manufacturing:
Practices, Challenges and Possibilities," Automatica 36, no.
11 (2000): 1567–1603.
Scanlan, MB Hopkins, and K O'Leary, "Knowledge-Based Process Control
for Fault Detection and Classification," Semiconductor Manufacturing
4, no. 10 (2003): 132–136.
Chen is section manager of the etch department at ProMOS Technologies'
Fab I (Hsinchu, Taiwan). In 1999 he received a BS in mechanical engineering
from National Taiwan University of Science and Technology in Taipei.
(Chen can be reached at +886 3 5798308 or maoshiung_chen@
F. Yen, PhD, is a senior manager of the Fab III etch department
at ProMOS Technologies. From 2002 to 2004 he was manager of the company's
Fab I etch department. He received a PhD in materials science and engineering
from National Cheng-Kung University of Taiwan. (Yen can be reached at
+886 3 5798308 or firstname.lastname@example.org.)
Coonan, PhD, is a support engineer from Straatum (Dublin, Ireland)
who is stationed in Taiwan. Previously, he worked on low-k chemical
vapor deposition process development at Trikon Technology and the National
Microelectronics Research Centre in Cork, Ireland, on the design, fabrication,
and characterization of novel strained-silicon devices. He received
a PhD in plasma physics from Dublin City University in Ireland. (Coonan
can be reached at +353 1 8395122 or email@example.com.)