THERMAL MANAGEMENT GENERATIVE DESIGN
simulation if a large number of variants
need to be analysed in parallel before
further progress can be made.
Training a deep-learning model with
the results of extensive simulations
has been used in areas such as
materials and particle physics research
because the trained models can run
faster than the original simulations,
which may take days to complete even
on a supercomputer. They do not offer
full accuracy but can be used on a fast
path to home in on parameters before
running a nal detailed simulation.
GPU maker nVidia, which is already
a strong proponent of machine learning
has applied the same technique to
a project called SimNet to do the
same for the kinds of computational
uid dynamics (CFD) models used for
thermal engineering.
A key problem is that deep learning
itself is computationally intensive and
requires weeks of simulation to build
a big enough training set to let an AI
model learn how parameters affect air
and heat ow in space.
Vervecken says the range of
projects in which they are involved,
which range in scale from mobile-phone
processor heatsinks to truck-scale
battery cold plate, increases the
complexity of the problem.
”To train a single model to cover
that whole range? That’s dif cult,”
Vervecken says, saying that it makes
more sense to prioritise machine
learning for the design
tools rather than the
simulation engine.
Against a possible AI-based
speedup you need to weigh the
contributions that can come from
simply optimising the computational
uid dynamics (CFD) algorithms
directly.
Specialist vendors such as Future
Facilities have employed multicore
execution and other optimisations to
cut overall turnaround time. In a study
with Rohde & Schwarz, the time taken
from run a simulation from CAD import
through mesh generation to analysis
was cut to below 15 hours from more
than 40.
As well as heatsinks, simulationdriven
machine learning can be applied
to core device and system design.
“One area that Future Facilities is
working with machine learning research
groups is in using CFD simulation to
train machine learning algorithms,”
Gregory notes.
There are a number of areas where
the machine learning can prove useful.
One is to help build simple models
of heat generation and transfer that
can guide algorithms used to decide
when to power down processors on a
multicore SoC.
“Thermal simulation can easily
train the machine learning algorithm
how the temperature of the device
would change based on processor
usage. This can be done quickly in
parallel before the device has
even been manufactured,”
Gregory says.
Another use lies
in R&D for new types
of device, such as those
used for power electronics,
where excessive heat
can cause thermal
runaway and other safety issues.
“Comprehensive training data is
essential for an effective machinelearning
algorithm, but it’s often not
practical or safe to obtain training
data from physical devices. If you are
optimising a new device or system,
a physical device may not have been
created,” Gregory says.
Performance is not the only
motivation for using AI-based models.
“Very often it’s a cheaper solution
they are looking for,” adds Vervecken.
“There’s a misconception that exists
with our technology. We are not
always looking for the most ef cient
heatsink. We have a project on-going
where the customer is instead looking
for the most affordable solution. The
customer said performance is not an
issue: the existing design is already
overprovisioned by 30 per cent but
they can’t get that type of design to be
cheaper.”
It is still early days for machine
learning for thermal engineering but
generative design techniques are
demonstrating the value of simulation
and rapid iteration in coming up with
designs that work better, work out
cheaper or both.
Above: Examples
of 3D designs from
Diabatix
Below: The result of
generative design –
the HP duct tongue
profusion
www.newelectronics.co.uk 9 March 2021 23
/www.newelectronics.co.uk