We're researching machine-learning systems that treat component placement and routing as optimization problems, learning the design intuition that today lives only in experienced engineers' heads.
Placement and routing are where most PCB design time goes. They're combinatorial, constraint-heavy, and deeply dependent on experience, the kind of tacit judgment that's hard to write down as rules.
Classic auto-routers optimize for a narrow objective and produce boards engineers rarely ship as-is. We think the gap isn't compute. It's that these tools don't learn from the space of good designs.
We frame layout as a sequential decision problem and apply reinforcement learning: an agent places and routes while a reward model captures the trade-offs engineers actually care about: density, signal integrity, manufacturability, and clean, reviewable results.
Learned component placement that balances thermal, signal, and mechanical constraints instead of raw wirelength alone.
RL-guided routing that produces layouts an engineer would recognize and trust, not just electrically valid ones.
Reward models tuned to real engineering objectives: manufacturability, review-ability, and design margin.
Landing inside ZIRO Designer, so the research reaches engineers as a tool, not a paper.