ShowDaily Day3 - Flipbook - Page 14
sponsored by
More Pumping Speed, Same Footprint:
Edwards Previews Its Next-Gen Turbopumps
As chip features shrink and
device architectures grow
more three-dimensional, the
turbomolecular pumps that
clear etch byproducts out
of process chambers have to
move faster just to keep pace.
Ahead of SEMICON West,
Dawn Stephenson, Business
Development Manager,
Chamber Solutions at
Edwards Vacuum, discussed
how her team is pushing turbopump speeds past 5,000 liters
per second — and soon toward
7,000 — without growing the
pump’s physical footprint, and the
unlikely role artificial intelligence is
playing in that design process.
“The need for pumping speed has
crept up gradually,” Stephenson said.
“Ten years ago we were at 2,000 to
3,000 liters per second, then it moved
to 4,000. Now we’re at 5,000, and
customers still want more. It used
to be the top three OEMs pushing
for that; now it’s the top five. But
the real change is that there’s no
more physical space to grow into.
Anyone can increase pumping speed
by making the blades bigger — the
challenge now is getting more speed
in the same footprint.”
That challenge produced Edwards’
5K turbopump (FIGURE 1), officially launching at SEMICON West
this year even though it has been
shipping for close to a year, with
prototypes dating back further
still. It’s now qualified by several
of the company’s major customers,
Stephenson said. A 7K pump is next:
still in development, with prototypes
14 | Thursday, October 15
pumping speed, but now advanced
logic nodes are too, in no small part
because of AI.”
Designing with an AI assist
Figure 1. Edwards’ 5K turbopump
(STP-iXA5507), launching at
SEMICON West, pushes pumping
speed past 5,000 l/s without
increasing the pump’s footprint.
Source: Edwards Vacuum.
targeted by the end of this year, built
in roughly the same envelope as the
4K and 5K generation before it.
Why space, and why now
Faster pumping shortens what engineers call residence time — how
long etch byproduct gases linger in
the chamber before being cleared —
which in turn improves uniformity,
yield, and throughput. “FinFETs,
gate-all-around transistors, deep
memory structures, advanced
interconnects — these all involve
difficult etch and deposition steps,”
Stephenson said. “It used to be
memory driving the need for higher
Getting more pumping speed out of
the same physical envelope meant
rethinking the pump’s blade-path
and drag-stage geometry — and
Edwards turned to an AI- and
machine-learning-assisted
modeling tool to do it. The tool
searches across competing design
parameters, such as pumping
speed versus compression ratio,
running through generations of
design candidates and re-optimizing from the best performers
until it converges on an optimized
geometry.
“Doing that manually, through
repeated modeling alone, would have
been hit or miss,” Stephenson said.
“It’s produced a real step change in
what we’ve been able to achieve,
even after decades of turbo pump
design experience in this industry.”
The irony isn’t lost on her: the
AI-driven chip demand that’s
pushing pumping-speed requirements is the same technology now
helping design the pump that meets
it. “Ironically, it’s the use of AI and
machine learning that has enabled
us to optimize turbine blade-path
efficiencies better than ever before,”
Stephenson said. “It’s a case of AI
paying it back.”
Borrowing from aerospace
Edwards calls the resulting design
work “Rocket Technology,” a nod to
how much of it
Continued on page 15
www.semiconductordigest.com
