A new AI technique designs circuit boards by having a fast, rough simulator and a slow, precise one take turns steering the process.
Researchers built Simulator-Refined Diffusion (SRD), a method for generating printed circuit board layouts that meet target electromagnetic specs, measured by S-parameters, the numbers describing how a circuit sends and reflects radio signals. Diffusion models could already sketch plausible layouts, but hitting real electromagnetic targets was the hard part. Full-wave simulators check those targets accurately but are slow and not differentiable, so earlier approaches leaned on faster differentiable "surrogate" models that are less trustworthy. SRD splits the labor: the cheap surrogate's gradient suggests a direction to change the design, then the accurate simulator searches along that direction to pick the actual update.
PCB design, especially for antennas and wireless hardware, is normally a loop of guess, simulate, and adjust, where each simulation costs real time. A generation method that lands closer to spec on the first few tries could meaningfully shrink that loop instead of just producing one plausible-looking layout. The paper reports layouts landing up to 21.2% closer to target specs than prior methods on familiar design types, and up to 19.8% closer on designs unlike anything in training, the harder test of whether the approach actually generalizes.
Those numbers come from the paper's own simulated benchmarks, not a board that has been etched, soldered, and measured on a bench, so the real test is still ahead.