AI in PCB Layout Design

[dindjarin]

Hello there, I’m trying to better myself in understanding the influence of AI in terms of PCB design.

Have anyone tried InstaDeep’s deepPCB, or do you have any knowledge about it? Or have you seen any articles/papers discussing it or its functionality? I would be happy to hear!

[trophosphere]

I'll leave this here:

[Psi]

It's a problem that should be solvable with AI, but so far there's nothing that does even a passable job at it.

[n4u]

its solvable, but projects with electronics and simulations arent open source - you need a lot of data to teach ai something . Maybe companies may make own ai versions for more complex design without errors - but i dont think it will be free.

[dindjarin]

How come you say it should be solvable by AI? Are there consensus that AI should solve unsolved NP-hard problems?

[thm_w]

Its not so much that you need to "solve" the PCB, it doesn't need to be perfectly solved, it just needs to be good enough.

https://www.eevblog.com/forum/blog/eevblog-1535-deeppcb-ai-autorouting-tested!/

[Sagar]

let me give you a real life example, I am working on a class D project on Easyeda. I designed the schema and converted it into the PCB. Then I made all the connection according to the knowledge and expereienec I have gained in past 2 years. I made the very passionatly and got almost 12 DRC. I know my design is good and do not need anything else. But in AI designs you don't have those flexibilties. DRC is important but if you know that the design is good you don't have to remove it. Although I faced some DRC errors but I made the final product according to me but that may not possible with AI. That's why in an technical design I don't prefer to use AI tools. It is like autorouting a rf PCB without taking phase velocity and frequency consideration in mind.

[KE5FX]

How come you say it should be solvable by AI? Are there consensus that AI should solve unsolved NP-hard problems?

If a computer program can thrash a 9-dan Go player, a similar one can route a PCB. 

[Psi]

How come you say it should be solvable by AI? Are there consensus that AI should solve unsolved NP-hard problems?

- The problem with normal auto-routers is that they do things that are electrically correct but not how a human would ever do it.  AI wont have that problem, the AI WILL copy all the human designed boards and how humans design pcbs because that is all it has ever seen in training.

- You can couple the AI training loop with DRC, so it should learn how to make a PCB pass DRC very easily.

- AIs are excellent at detecting and copying complex links between things. The AI will quickly learn all the unwritten rules about what parts you place near other parts on the PCB.   eg, how DCDC chips always have the inductors and fets close to the chip.  There are so many things the AI will pickup on that you could never code a normal autorouter to do because there are just too many. 
You do so much on instinct when making a PCB and don't even know why.

It should be able to route a perfectly passable PCB just like it can create a perfectly passable photorealistic image.

[tszaboo]

How come you say it should be solvable by AI? Are there consensus that AI should solve unsolved NP-hard problems?

- The problem with normal auto-routers is that they do things that are electrically correct but not how a human would ever do it.  AI wont have that problem, the AI WILL copy all the human designed boards and how humans design pcbs because that is all it has ever seen in training.

- You can couple the AI training loop with DRC, so it should learn how to make a PCB pass DRC very easily.

- AIs are excellent at forming complex links between things. The AI will quickly learn all the unwritten rules about what parts you place near other parts on the PCB.   eg, how DCDC chips always have the inductors and fets close to the chip.  There are so many things the AI will pickup on that you could never code a normal autorouter to do because there are just too many. 
You do so much on instinct when making a PCB and don't even know why.

It should be able to route a perfectly passable PCB just like it can create a perfectly passable photorealistic image.

You know, I totally believe that AI will be able to do boards that look passable for the untrained eye.
The same way all these Arduino boards that are copied over and over are filled with mistakes and sub-optimal design choices. Since your AI will be trained with openly available hardware designs, that are made by enthusiasts, not professionals, the designs will be that quality level.
Unless someone mayor, like Altium breaks our trusts and uses our designs to train AI.

[Psi]

It's never going to be perfect, but it should be night and day better than current autorouters that are useless except in some limited situations where they are fined tuned for specific tasks on pcbs with a huge number of layers.

Another problem with current autorouters on like a 2 layer PCB.  If you let it autoroute the board you will end up with something that you cant even edit to make it useable. It's always so bad you can't fix it and have to start from scratch.

AI autorouted boards should be much better for doing a few manual fixes and getting something good out