Does artificial intelligence make you concerned for your job? Are humans at imminent risk of being replaced by robots, or even software-driven functions?

Kyle Miller says no and no.

Miller is head of a team of product developers at Zuken in Bristol, England, that is working on AI-based place-and-route technology. In addition to the 20-plus years spent in CAD tools, he has a doctorate in artificial intelligence, which means he’s a lot better at math than me or you.

Speaking at Zuken Innovation World in mid-April, Miller outlined the headway Zuken is making in machine-learning tools. The short answer: quite a bit. Machine learning-based programs are very good at pattern recognition and converting data into usable forms. Everyday uses include Google’s Android-based speech-to-text tools. ML is also apparently superior in finding and exploiting bugs in software, to the extent the developers of the space-flight simulation role-playing game known as Elite Dangerous had to eliminate the function after their game’s ML exploited a glitch to create an unstoppable weapon.

Zuken is not alone. Last July 2018, the US Defense Advanced Research Projects Agency (DARPA) chose Cadence to support its Intelligent Design of Electronic Assets (IDEA) program, an effort to leverage machine learning techniques to develop a single platform for system-on-chip, system-in-package and PCBs.
And faced with challenging PCB interposer layout due to irregular routing patterns and signal integrity constraints, Intel engineers introduced a router using AI and ML at PCB West last year.

But before you worry about a lifetime of servitude to robot overlords, Miller is quick to note ML is designed to look for shapes, fields and textures, and can be fooled by simple alterations to common items, and it cannot “watch” someone executing an action and learn from their expertise (or lack thereof). Akin to a bad child, it learns from its own “mistakes” only after it makes countless ones.

Miller explains ML in terms of two parts: the generator and the evaluator (or discriminator). The discriminator determines whether the data it is being given are real or not. Thus, by improving the evaluator, the generator also improves.

The evaluator can measure quality in many ways. In PCB terms, since the generator has no preconceptions, it must be taught what “good” or “bad” design is. This can take many forms, such as constraint testing or internal analysis (e.g., is a component in a strange place?) or simulation (does the transmission line work?). “We can feed the system with known good designs and use the system to extract them, and use that to update the evaluator,” Miller explains.

For layout, AI would determine which order to go around obstacles. It learns on its own how to connect different geometries. Then the evaluator looks at the results and qualifies which is best. Similarly, AI can evaluate the quality of routed busses. Based on placement costs, the generator is provided suggestions for better component placement.

The Zuken team is attacking the problem on multiple levels, from bus/sub-bus planning, all the way to individual nets. By fall, Zuken expects to have a demo tool capable of cluster routing, or multi-signal quick-routing. Other gains may be coming in design planning and placement, trace ordering, and control of the autorouter.

Down the road, Miller expects ML-based technology to be able to make recommendations on components and package styles. Zuken has begun working on fan-out and escapes from high-pin-count devices. The evaluator will say a part doesn’t work or do what is intended, so try this alternative part.

“If it’s scalable, we will end up with something that’s far better than an autorouter,” says Miller. “It’s different than a traditional autorouter because it can learn styles and develop something that’s much more human-like. Results can be tuned based on user style and feedback.”

With any revolution comes pushback. Miller is unperturbed. “There will always be space for a human in design,” he told me. “Even in the models we’ve come up with, what we’ve done has shifted what’s designed. So instead of actually laying out individual traces on a design, you might be controlling some other aspect of the design. It’s shifting the parameters of what you are actually designing, but I think there will still be some human element over the design.”

And for those who think ML-based CAD is bound to fail, consider this: The Bristol office is the former Racal-Redac site, the group that conceived the original shape-based router. History is on their side.

P.S. Naveid Rahmatullah and Xiao-Ming Gao will present the latest developments on the Intel AI router at PCB West this fall. Check pcbwest.com for details.