|
Table of Contents
The semiconductor industry is going through yet another
business cycle. Wafer fabs are under intense pressure to become
more competitive, including those that have historically generated
the highest margins. The forces driving this cycle are the
"10X Changes" in the electronics industry described by Andy
Grove in Only The Paranoid Survive: Markets for new
chips emerge almost overnight, while today's cash cows often
end up as tomorrow's bargains at the local discount chip outlet.
Logic chips are on their way to becoming commodities. Memory
chip price deflation has entire national economies reeling.
Becoming more competitive requires cutting costs, shortening
turn-around times and increasing capacity, in short—improving
productivity. For equipment and process engineers, this translates
into delivering predictably high product uniformity and high
equipment uptime with low unscheduled downtime.
There is a disconnect between the forces driving equipment
productivity improvement and the means available to achieve
it. Consider the complexity of almost any piece of process
equipment, compared to the sophistication of the diagnostic
tools available for maintenance and troubleshooting.
Some of the most powerful, and in my self-serving opinion,
underutilized tools to achieve these productivity improvements
are software-based information systems. They can be developed
(relatively) quickly and inexpensively to suit a variety of
different special purpose applications. Because they are software
based, these tools are flexible and can be modified as needs
change, provided the software is well written. Data capture
and analysis software interfaced to manufacturing process
equipment can help predict system failures, determine maintenance
schedules, verify that maintenance operations were performed
successfully, reduce the number of test runs, and eliminate
test wafer usage.
These software tools can be easily retrofitted to existing
equipment with little or no impact on the existing manufacturing
process flow, but their impact on productivity can be enormous.
Manufacturing tools and processes that are running well have
identifiable characteristics that are different from tools
and processes that are not running well, but you can't guess
at what those characteristics are. The old-fashioned approach
of relying on a few individuals who, through experience and
a sixth sense, know the idiosyncrasies of the equipment in
their care is no longer viable. New people who have basic
technical skills, but little applicable experience, can do
a better job than wizened veterans, provided they are given
access to the relevant information characterizing the processes
and equipment.
Below I discuss two important classes of software tools
that can be used with data collected from almost any piece
of fab equipment: Filters and Miners. These engineering tools
can extract the real truth about process and equipment performance
from the collected data. The discussion focuses on challenges
to information system providers working directly for semiconductor
companies, as well to third party software developers serving
the semiconductor industry.
High Bandwidth Data Transformation via
Filters
Back to top
Consider the two broad categories of equipment data: One set
is fixed, static and immobile like a statue. The other is
fleeting, repeating and dynamic, like a fountain of water.
The fountain stays the same shape, not because it is made
of granite, but because all the pipes, valves and pumps that
deliver the water are stable (Metaphor borrowed from The
Fountain, by Charles Morgan). Our eyeballs and brains
do not have to collect and archive zillions of bytes of water
molecule location data to build the mental image that we see.
We only extract a small amount of the information available—limited
by our eyesight and our aesthetic curiosity. In other words,
we filter out extraneous information, and collect and process
only what we need. Similarly, the objective in operating a
piece of fab equipment should be to extract the image of the
process that best meets our production requirements, and then
do whatever is necessary to make sure we don't deviate.
A piece of fab equipment can have an incredibly high data
throughput. The first technical challenge for fab information
systems providers seeking to improve equipment productivity
is to give engineers high-level access to this low-level data
in a form they can use. Then let the engineers declare which
data should be sampled and how it should be analyzed. For
very simple technical reasons—mainly storage capacity
limitations—most data vanish into the bit bucket instantly,
or shortly thereafter at the point of consumption. Decisions
about what data to keep and what to throw away need to be
reversible as process knowledge improves.
The process tools in a fab all emit different sets of process
variables. Whether they purport to use a standard language
like GEM or a standard bus like RS488 they are still saying
different things. Software filters provide the means to collect
process variables, abstract and normalize them, then push
the data into its common forms: fields and timestamps in records;
common records in tables.
What kind of information should be allowed to pass through
the filter? As a wafer passes through a piece of process equipment,
it follows a series of steps. Each step has a function, some
important and some not so important. Each carries some potential
to damage the wafer, some steps more so than others. A filter
should distill out that which is critical for the successful
operation of a process from that which is not. Thus a successful
process becomes one in which the critical events are well-behaved.
For example, many processes are sensitive to gas-phase nucleation
problems caused by flow overshoots when an MFC ramps. Thus,
the critical event is the behavior of an MFC during turn-on.
As another example, consider the pumpdown and backfill steps
that are common to many types of equipment. Improper venting
procedures, poorly maintained automation, and a host of other
causes can generate particles during these steps. The critical
event is the successful loading and unloading of wafers without
introducing particles. Data filters should focus on these
critical events.
Critical Event Behavior Revealed
Through Data Mining
Back to top
The second technical challenge to fab IS providers is to deliver
and maintain data mining tools that can deliver the filtered
information in a useful form from its many sources to people
in a position to use it. The data must be accessible enough
that people will actually use the computers running the mining
utilities for their intended purpose, rather than for surreptitiously
running the computer games hidden away within the bowels of
even the most secure IT environments.
Effective data mining starts with ideas—mental models
and a nose for where the problems lie. You perceive there
is a problem. Based upon limited information, you suspect
something about a process or a piece of equipment is not operating
properly, or could be made to operate better. You must be
able to phrase a question, "If x is happening, then the data
should show me y." Good mining tools should be able to extract
the few kernels of useful information from the reams of raw
and filtered data available. Depending upon the complexity
of the point one is trying to prove, effective data mining
may use multiple software utilities and many disparate sources
of information: metrology data, maintenance logs, or input
parameter SPC data from equipment.
Effective filtering required a focus on critical events:
knowing the things that must happen correctly for a process
to be completed successfully. Effective mining similarly requires
being able to verify that they did indeed happen as advertised.
The two processes go hand in hand.
Actions Speak Louder than Successful
SPC Experiments
Back to top
Once experimentation uncovers information that can be used
to improve a process, the final challenge to the information
systems providers is two-fold. The first part is relatively
easy: helping engineering and manufacturing to implement systems
that integrate critical event monitoring with the routine
production process flow. In this way, the improvements become
institutionalized.
The second part is significantly more difficult. It involves
building the corporate storehouse of useful knowledge by archiving
the data analysis success stories in ways that make them available
to engineers at other sites, as well as future generations
of new engineering recruits destined to "rediscover" the same
issues on their own.
About the Author
Jon Goldman received his PhD degree in materials science from
MIT. He has been serving the semiconductor industry for nearly
25 years, receiving numerous patents, including one for developing
the LPCVD silicon nitride deposition process. After working
at Motorola SRDI and as technical director at Thermco Systems,
he founded Jon Goldman Associates in 1987. JGA provides software
solutions to the semiconductor processing industry. Email
him at jongoldman@compuserve.com.
For more information: Jon Goldman Associates, Inc.
2237 North Batavia St. Orange, CA 92865-3105. Tel: 714-283-5889,
fax: 714-283-2884.
|
PRESS
CENTER 
