Full-wave simulation is too slow to search with. We train a model on physics we generate ourselves, explore on it, and send only the survivors through to the solver.
Not a chat layer over a solver. A learned model of the fields, built so that what physics guarantees is guaranteed in the model too.
Our own label factory, drawn to cover the design space — not scraped from whatever got published.
Ask for a field at a coordinate or a port-to-port response. One model, whatever resolution the question needs.
Mirror a layout and the fields follow. Reciprocity holds exactly — structural, not learned and hoped for.
You stop searching with the solver and start checking with it.
A full-wave solve is minutes to hours, so exploration gets rationed to a handful of candidates picked by intuition. Scoring thousands changes which designs you consider at all — and nothing ships on a prediction alone.
00:00 → feasibility reachable
00:01 → model_sweep 4,096 candidates
00:09 laminate A: band-limited — dropped
00:10 → promote top 6 → full-wave
00:52 → confirm convergence gates
00:58 every spec line met · pass
Each loop takes a spec you already write down and returns a verdict saying which lines are met, on what evidence.
Planar single-element design with the laminate free to change — often it is the substrate, not the layout, that decides whether a spec is reachable.
Board-level exploration on the rail that matters, closed against a target impedance rather than a rule of thumb about decap counts.
High-speed channels end to end — the stackup, the via transitions and the routing that decide whether the eye is still open at the receiver.
Distributed filters, couplers and matching networks synthesised against a response, rather than tuned by hand from a textbook starting point.
Thirty minutes: the spec you are trying to close, and the constraint that keeps biting. We run it and show you what comes back, including where it falls short.