AI placement in PCB design is both possible and could be a road that brings designers into a new era of innovation.
Intelligence has been available in most EDA tools including PCB layout for some time now. Though the potential for machine learning exists in
EDA, PCB designers have been slow to adopt a technology that currently
auto-places and auto-routes for silicon. Most PCB designers manually
route and design their boards, a time consuming and intricate process.
As early as the 1980’s, neural networking was an established
theoretical concept in EDA. By the 1990’s, there were tools in place
that could use the concepts.
Neuroroute – a product based on neural networks – was given
50 to 60 human-made PCB designs. These designs were relayed to an AI
routing engine that used supervised machine learning to create an
auto-router that makes decisions like a human.
Neuroroute paved the way for modern topological techniques. For
example, it mapped out board space in triangles or other non-rectangular
polygons, allowing for more efficient channel utilization and any-angle
routing. This was a big development over the rectilinear “shape-based”
routers of the time, but these early implementations lacked the
computing power as well as the quantity and depth of training sets to be
Part of my role as a product manager in EDA is to gain feedback from
PCB designers and understand their needs and uses for our tools. I hear
three common reasons why designers don’t like to use auto-routing:
- “It makes a mess”
- “The software doesn’t work.”
- “It may reduce the need for human PCB designers.”
What designers really mean when they say that auto-routing “makes a
mess” is that the finished PCB doesn’t look aesthetically pleasing after
the auto-routing process. PCB designers are a group that takes pride in
the utility of their designs and their designs’ ornamental value.
Having been a designer myself since I was eight-years-old, I
understand the desire to create something neat and polished. Yet it’s
the most pragmatic designers among us who understand that it doesn’t
really matter how a PCB looks, as long as it works and can be produced
at reasonable cost.
Designers who say the software doesn’t work are actually referencing
the default settings on their design software. When auto-routing is
paired with their designs, the software’s default settings often
complete anywhere from 50 to 70 percent of the design before giving up.
Luckily, this is an easy problem to solve. It may just be that
designers need to tweak the default settings, constraints and design
rules to match their designs. EDA software makers can improve
out-of-the-box functionality by readjusting these defaults.
Finally, it has been my experience that many PCB designers are
skeptical of automation in PCB design because it may one day replace
designers and leave them out of a job. To this I say that designers are
more involved in the design process than ever. Humans still need to add
flare and artistic integrity to each design.
Today, there is enough computing power and internet capacity to
enable this technology. AI libraries and infrastructure enable a
streaming data approach.
Given permission by users, AI could be doing deep learning of the PCB
design processes right now. And this wouldn’t replace the need for
human designers –it would only empower them to create the best designs
possible, like Iron Man putting on his special suit.