|
Recently JGA, in cooperation with Brooks Instrument,
explored developing a software package to give end-users access
to the bells and whistles that digital mass flow products
provide. This project opened my eyes to the potential this
technology offers for routine monitoring of many different
types of critical sensing components in addition to MFCs.
In this article I present some sample data extracted from
a simulated Single-Wafer CVD System equipped with a bank of
digital MFC prototypes. I hope it will be illustrative of
the potential of digital technology for tighter process monitoring.
Digital processing has permeated our daily lives
in many ways. We communicate with each other digitally via
cell phones and e-mail; read digital news on the internet;
take digital photos and videos of our families; we will soon
be watching digital TV. A decade ago, all of these media were
analog - or didn't even exist. The reasons for the transformation
are obvious to those of us who appreciate the power and flexibility
of these toys. But "We," in Semiconductor Wafer Processing,
are a bastion of Luddites. We love our analog signal processing,
the noisier the better.
Take, for example, mass flow controllers. Analog MFCs are
found everywhere in wafer fabs. They are used for metering
precise quantities of reactive gases in many different types
of equipment. Their performance and reliability are important
determinants in our industry's ability to produce components
with in-spec characteristics. Future processing - and safety
- requirements will no doubt demand even more precise control
and reliability in the metering of gaseous reactive chemicals
than exists today.
Here are some reasons why I believe replacing analog components
with their digital counterparts in current generations of
fab equipment can help improve processes. First is "bandwidth
and access." You can generate all the data you need to understand
the subtleties of your equipment operation, without having
to depend on a plodding SECS interface for data acquisition.
This advantage becomes even more pronounced when process step
times are measured in seconds. As an example, consider the
many processes that depend upon maintaining a positive pressure
gradient between two chambers during critical operations.
If the equipment does not provide adequate monitoring and
remedial action when error conditions occur, superimposing
a digital pressure data monitoring system can.
A second reason is signal integrity (no noise), meaning you
can easily use the digital data to dynamically track what
a component thinks its setpoint is, and what it thinks its
output is. Often this is quite different from what the equipment
itself thinks. For example, MFC calibration problems frequently
reveal themselves via a significant non-zero flow output signal
when there is no gas flow. Yet, the process controllers on
most current generation fab equipment are rarely programmed
to look for this condition. In cases like this, where the
existing equipment infrastructure is not able to deal with
an identifiable problem, digital data monitoring systems can
help add the needed intelligence to look for these conditions
and initiate appropriate responses.
A third reason is that digital components come equipped with
lots of Gadgets - just like cameras and cell phones - that
improve their flexibility and usefulness over analog models.
For example, you can instantly change their ranges, change
the type of gas they can be used with, or automatically "tune"
them to improve their performance under local conditions.
I used a test fixture consisting of a generic process controller,
a bank of hybrid MFCs (equipped with both standard analog
and digital interfaces), a "process chamber," and a vacuum
system that crudely simulates a piece of process equipment
such as a CVD system.
In this group of experiments, I tried to simulate the short-duration
repetitive on-off cycling of gases that occurs in many different
types of process equipment in use today in order to see what
kinds of MFC problems high-bandwidth data capture might reveal.
Nitrogen gas was introduced into a vacuum system at two-minute
intervals, followed by an intervening minute with no flow.
Trace data from a typical run is shown in Figure 1. This sequence
was repeated several times.
Figure 2 shows some mean flow rate trends. There
are three panels in this figure, each containing data from
a different MFC. Each run is represented by three points (See
Figure 1).
Note that there is substantial variation in the average flow
rates of N2 from interval to interval. For the MFC in the
upper panel, the average flow rate is 777 SCCM, 1-sigma is
about 14 SCCM, and 3-sigma is 42 SCCM - about 5.4 percent.
For the MFC in the center panel, 3-sigma is about 6 percent,
and for the MFC in the lower panel - about 4 percent. Flow
variations of this magnitude in process gases such as TEOS
or SiH4 could possibly give rise to film non-uniformity of
similar magnitude.
Figure 3 provides some insight into one of the possible reasons
for the observed flow rate variations. This figure shows an
overlay of the MFC trace data from several intervals. There
is substantial variation. But does it account for the observed
large standard deviations seen in Figure 2? Figure 4 shows
the same trend results as in figure 2, but a 10-second delay
in the start of the calculation was inserted. The standard
deviations are lower in all cases, indicating that most of
the observed flow variation from interval to interval occurs
while the MFC is ramping up.Table I lists the means and standard
deviations for all three flow controllers with and without
the 10-second delay.
Table I
Flow Interval Mean and Standard Deviation Data
Calculated With and Without A 10-Second Delay
| MFC Name |
Mean Del
1 sec |
Std Dev Del 1 sec |
Mean Del
10 sec |
Std Dev Del
10 sec |
| N2 1500 |
592 |
12.3 |
603 |
6.2 |
| N2 1000 |
778 |
13.9 |
820 |
7.9 |
| N2 75 |
43.6 |
0.58 |
43.0 |
0.43 |
Figure 5 shows a composite trace data plot of 15 flow intervals.
The flow variations from interval to interval are striking.
If anybody's real CVD systems exhibit flow profiles like those
in the figure, I would think it might be worthy of some attention.
Figure 6 shows some other interesting trend data taken from
the N2 1000 MFC. The top panel contains flow mean data for
each interval, the lower panel contains valve voltage data.
The data are seen to shift abruptly at about points 13 and
25. The flow rate shift at point 13 is about 2 percent. Figure
7 shows the same data, but the x-axis is plotted on a time
scale. The abrupt changes are seen to occur after a time interval
of several hours during which no processing was done. The
explanation for the shifts is as follows: Since I'm using
bottled Nitrogen, after each session I turned off the gas
supply at the source to avoid having all my N2 leak away.
When I turned it back on, I readjusted the input pressure
to "something near 18 Psi", but the variations apparently
produced a big shift in the valve voltage, and a measurable
shift in flow.
Since time interval data are available, I thought
it might be instructive to see what sorts of variations exist
from interval to interval. Figure 8 shows some duration trend
data. The average interval duration is 1.96 minutes - about
four percent below the programmed interval duration of 2.0
minutes. Additionally, 1-sigma is .017 minutes, and 3-sigma
is .051 minutes - 3 seconds. This represents a +/- 2.5 percent
variation in interval duration among all the intervals measured.
If you're trying to control film thickness down to the proverbial
gnat's eyebrow, a 2.5 percent variation in the deposition
time interval is just one more thing contributing to film
non-uniformity.
Using a digital network, I was able to collect high-resolution
flow data that illustrated several common types of MFC problems
in gory detail. This was accomplished with a simple two-wire
twisted pair network and a PC, and without touching any of
our simulated tool's hardware infrastructure.
Whether the observed interval-to-interval flow variation
during ramp-up is an MFC problem or some other equipment problem
is immaterial - most MFCs, whether analog or digital, will
experience problems sometime during their lifetimes. The point
is that to fix problems you first have to be able to detect
them, and to detect them you need the DATA.
Consider the litany of potential MFC problems: flow overshoots
during setpoint changes, long settling times, unstable flow
(oscillation), zero drift and calibration drift. MFCs function
in a rugged environment; they are subjected to reactive and
corrosive gases, and unstable pressure environments. It shouldn't
surprise anyone that they break. The point is it's easier
to spot one that's broken if it is a digital. I believe digital
technology presents an opportunity to provide unprecedented
automatic monitoring and detection of all sorts of subtle
problems for hundreds of MFCs and other types of digitally
equipped sensors throughout a wafer fab. Early Response systems
are already in place in many fabs that automatically shut
down process equipment when problems are encountered. Fault
detection systems that utilize digital sensors can provide
an extra measure of fab-wide early failure detection and intervention.
There are two potential channels for implementing digital
technology in a wafer fab. One is through the OEMs who will
incorporate digital components in future new equipment designs.
This may happen on new-generation 300-mm equipment, but don't
hold your breath waiting for OEMs to offer digital sensor
retrofit kits. The other is for end-users to employ digital
technology for their own use. This is actually fairly straightforward,
because many components with digital interfaces are hybrids
that can replace existing analog components. The networks
are generally easy to install, usually requiring a simple
two-wire twisted pair strung daisy-chain fashion from component
to component, and a PC. There is no need to dig into the bowels
of the tool's control system to install the network.
I believe digital technology presents an opportunity to provide
unprecedented automatic monitoring and detection of all sorts
of subtle problems for hundreds of MFCs and other types of
digitally equipped sensors throughout a wafer fab. Early Response
systems are already in place in many fabs that automatically
shut down process equipment when problems are encountered.
Fault detection systems that utilize digital sensors can provide
an extra measure of fab-wide early failure detection and intervention.
* The Luddites were members of a Scottish movement opposed
to the introduction of new weaving technology into the textile
mills of the early 19th century. The name is sometimes loosely
applied in modern times to those who express misgivings about
today's new technology.
|
PRESS
CENTER 