Comment by kaipereira
2 days ago
Routing is honestly one of my favorite parts of making a PCB, there's something about the feeling of understanding exactly what each trace is doing that I find exhilarating.
The problem with most AI powered tools is that they don't understand the context of your board, which is absolutely crucial for routing. Like so-and-so trace could be switching really fast and an AI powered design tool wouldn't understand that and wouldn't route it properly. Or like power for example, different things have different power requirements and I feel like we're not at the point where these tools are referencing every datasheet on the planet to figure out trace/fill sizes and whatnot.
There's also some concepts I don't think these tools can wrap around quite yet. For example, minimizing loops, consistent ground planes, proper impedance control, and just all of these factors lead to an inferior board.
Not saying that humans are perfect and understand all of these too, I just feel like if you understand what's happening with your board, you can do a much better job.
At the end of the day, you'll be spending more time fixing the mistakes, which are actually pretty hard to spot without checking every single trace sometimes, then actually just routing it yourself, but maybe one day we'll get there!
I'd love to hear some other takes though ;)
You're right that physical context is crucial. However, mathematically, PCB routing is a Graph Theory optimization problem that falls into the NP-hard complexity class (related to the Steiner Tree problem). Traditional solvers rely on heuristics (like A* or annealing) because finding an exact mathematical solution is computationally intractable for complex boards. My point about AI—specifically Graph Neural Networks (GNNs)—isn't about bypassing the learning process. It's about using statistical approximation to handle these high-dimensional topological constraints more effectively than human intuition or standard heuristics can. It's not 'lazy'; it's applying a modern computational lens to an unsolvable deterministic problem.