ShowDaily Day1 - Flipbook - Page 19
sponsored by
Machine Learning Could Make Gas Purity Data
Do More Than Pass or Fail
At the point where ultra-high-purity (UHP) gas changes
hands from the gas supplier to the fab, a process analyzer
has one main job: confirm that the gas meets specification. Most of the time, the answer is yes. That makes
the resulting data stream reliable, and as Jeff Wuenschell
of AMETEK Process Instruments puts it, “intentionally
and ideally kind of boring.”
Wuenschell, Gas Analysis Applications Scientist for
the company’s trace and UHP moisture process analyzers, will argue at SEMICON West that this boring
data could be worth a lot more. In “Beyond Certification:
Machine Learning for Gas Purity Monitoring in
Semiconductor Fabs,” part of the Semiconductor
Equipment and Components Forum on Tuesday, October
13, from 10:55 to 11:15 a.m. in Room 154 (South Hall,
Upper Mezzanine), he will look at how machine learning
techniques might extract added value from analyzer data,
from diagnosing problems upstream of the fab to scheduling maintenance on the analyzer itself.
The visit also marks a return. AMETEK’s semiconductor gas analyzer business has been heavily involved
in the Asian market for years, and this is the first time in
a while the group has come to SEMICON West, with the
goal of building more relationships with chipmakers and
gas suppliers in the U.S.
Beyond certifying the spec
AMETEK Process Instruments, which has supplied
process analyzers to industry since 1967, covers two
measurement problems that matter most to the fab. Its
trace analyzer line uses gas chromatography to monitor
bulk gas purity, targeting impurities such as H2, CO,
CO2, CH4 and non-methane hydrocarbons. Its UHP
moisture analyzers detect water at ppb and sub-ppb
levels.
These instruments typically sit at the custody transfer
point, where they verify the gas before it moves into
the fab’s distribution system. “The primary job of our
analyzers is to just certify that it’s meeting the specifications,” Wuenschell said. His talk asks what else that
position makes possible.
He sees three areas. The first is diagnosing upstream
conditions, so a change in the readings points to what
is happening at the gas source or in the purifier. The
www.semiconductordigest.com
second is providing more information about downstream
processes. The third is the analyzer itself: using its own
operating data to predict maintenance needs and help
fabs do a better job of scheduling downtime.
Training a model on rare events
The obstacle is the data itself. A well-run gas system
produces years of readings that show no impurities,
which is exactly what the fab wants to see but not what
a machine learning model needs. The events worth
detecting are, by design, rare.
Figure 1. The AMETEK ta7000 gas purity monitor
measures trace H2, CO, CO2, CH4 and NMHC
impurities in UHP semiconductor gases at low-ppb
and ppt levels.
“If we have an analyzer that’s sitting there that’s operating for years and years, and just kind of showing no
issues, but say, for example, someday you may have, like,
a valve failure upstream that’s introducing some contamination, or just say a purifier breakthrough,” Wuenschell
said. “If it only happens, say once a year or once every
couple of years, that’s a very small amount of data” for
training a model. “It’s better to have lots of data.”
A second barrier is ownership. Ideally, a model would
learn from many analyzers across many processes
and customers. In practice, fabs may not want to share
process data with an instrument maker, and customers
are even less likely to share it with each other, even
indirectly. “The fact that a lot of this may be proprietary
is another issue that could make this data hard to come
by,” he said.
His talk will cover the machine learning techniques
AMETEK has been evaluating, along with approaches
for working around sparse data, either Continued on page 20
Tuesday, October 13 | 19
