Artificial Intelligence (AI) for Printed Circuit Board (PCB) Design

Comments

Created Friday 29 September 2023

Prompted from a one-off comment on HackerNews, I was inspired to start compiling a list of machine learning (ML) and artificial intelligence (AI) resources for PCB design. First though, what is AI/ML? These days, most people associate AI with ChatGPT, an interactive chat tool which answers queries using a complex AI model. In reality, AI/ML has been around for decades in many different forms, though the availability of cheap, powerful hardware for training models has contributed to the recent popularity. (I suspect that we can also leverage some of this hardware for direct improvements to traditional physics-based simulation engines, before we even start considering throwing AI into the mix). Seriously though, people have been considering the use of AI for PCBs going back to when I was a toddler!

//Image from Wikipedia:// https://en.wikipedia.org/wiki/Machine_learning#/media/File:AI_hierarchy.svg

AI is a broad term which encompasses any kind of machine intelligence. Machine learning as a subset of this corresponds to a computer program which learns its task on its own, over time. A subset of ML is deep learning, which is only one of the possible methods used to implement machine learning in PCB design.

Resources

The following list is a compilation of various options for AI/ML in electrical engineering and PCB design. This specific application for AI/ML is relatively new, so there are not many production-ready options (i.e., there is no ChatGPT for PCB design yet). Note that there are also offerings in this space from both Cadence and Synopsys on the chip-design side of things, but since this site is primarily focused on PCBs, we'll leave those off for now.

System Design

Circuit Mind

Circuit Mind focuses on the system architecture and schematic design, allowing a designer to start with a block diagram (which they are probably already doing). The tool creates a schematic with various tradeoffs such as cost, performance, etc. They specifically do not address the PCB layout portion of the design, but do work alongside existing schematic tools
https://www.circuitmind.io/

PCB Layout

Flux

Probably the most advanced of any of the options listed here, this tool assists with both schematic and PCB design.
https://www.flux.ai/p

Quilter

After designing your schematic (currently support for Altium, KiCAD and Eagle), Quilter creates a PCB layout using AI.
https://www.quilter.ai/

Cadence Allegro-X

Cadence is probably the biggest EDA company talking about ML/AI in the PCB space (that I've noticed) with their recently announced "Allegro X" stuff. The videos are mostly marketing fluff, but they seem to be on track with what I'd expect for "AI assisted PCB design".
https://www.cadence.com/en_US/home/tools/pcb-design-and-analysis/allegro-x-design-platform.html
https://www.youtube.com/watch?v=RlWfSQq0NkA

DeepPCB

Can't speak for this one as I have only ever heard of it via online search, but it offers fully automated PCB layout.
https://www.deeppcb.ai/

Circuit Mind

This one is neat because the input is a block diagram of the system architecture.
https://www.circuitmind.io/

JITX

These guys have been active on the podcasting circuit recently. Very interesting because it seems to be text-based logical design with AI-assisted routing.
https://www.jitx.com/

Design for Manufacturing (DFM) Checks

Cadstrom

Startup looking at a post-layout DFM check workflow: https://cadstrom.io/
See their docs site here: https://cadstrom.io/docs/en

Simulation Tools

Cadence Optimality

In addition to Allegro-X, Cadence also has a neat machine-learning optimization tool for their electronics simulations.
https://www.cadence.com/en_US/home/tools/system-analysis/optimality.html

Future Research Opportunities

Since this field is so new, there are many areas undergoing active research which show promise.

A Deep Neural Network Modeling Methodology for Extraction of RLGC Parameters in µ-wave and mm-wave Transmission Lines
This paper shows a neural net which replicates the results of a 2-D field solver, but with a 14x speedup in computation time.

A Novel Deep Neural Network Methodology for Fast Scattering-Parameter Extraction of Discretized Structures in Three-Dimensional Space
The paper here is only a proof-of-concept, but the first in its field. It trains a neural net on a generalized geometric model rather than design parameters. What this means is that the field solver has no concept of trace width, via diameter, etc. All it knows is whether a particular location in 3-D space is metal or dielectric. The results are impressive for a first-ever example of this technology. It also shows a 3,600x speedup compared to a traditional full-wave solver. I am hopeful that this kind of underlying technology can augment traditional physics-based field solvers for extremely quick simulations.

Other Resources

Thanks to Pallav Aggarwal for his list which helped me find some of these companies!
https://pallavaggarwal.in/automated-pcb-design-using-ai/