Radio-frequency integrated circuit design is one of those fields where decades of expert intuition have hardened into craft that resists codification. An RFIC designer does not just apply a textbook formula; they develop, over years, a feeling for how RF signals behave in silicon — what parasitics will matter, how to lay out a circuit so the real-world electromagnetic environment cooperates with the intended schematic, when to break convention. This knowledge is expensive to acquire and almost impossible to transfer systematically. Which is why this week’s IEEE Spectrum feature by Princeton’s Kaushik Sengupta is worth reading carefully.
The system Sengupta’s group built works in two stages. The first is a reinforcement learning loop that explores circuit topologies without relying on human-designed templates — it chooses architecture, device parameters, and layout geometry autonomously. The second is an electromagnetic emulator: a convolutional neural network trained to predict how a given circuit will behave electromagnetically, replacing what would otherwise be a simulation step that takes minutes to hours per design iteration. With the emulator in place, the RL agent can evaluate thousands of candidate designs in the time a human designer would spend on one conventional simulation.
The results from a 2023 experiment illustrate the scope of what this enables. The system designed a 30–100 GHz broadband power amplifier — a wide-bandwidth, high-frequency device where silicon-based designs have historically bumped against hard limits — and achieved what the paper describes as the best combination of bandwidth, output power, and efficiency then reported for a silicon power amplifier at those frequencies. It also handled multiport circuits, where full electromagnetic simulation had previously taken days or weeks per design. The AI compressed that to minutes.
The more striking detail is what these circuits look like. AI-designed RFICs apparently resemble QR codes: densely irregular, without the structured symmetry that makes human-designed schematics readable. When an RL agent is not constrained to human-legible topologies, it exploits geometrical degrees of freedom that designers intuitively avoid — not because those geometries would not work, but because circuits that look like random noise are difficult to reason about, verify, or modify by hand. The AI is not designing schematics. It is discovering physical configurations that satisfy the performance objectives, and those configurations happen not to map onto anything in a textbook.
This is not entirely novel as a phenomenon. The same dynamic appeared in AlphaGo’s “alien” moves: the system produced valid, winning plays that professional players could not have derived from existing theory. In RFIC design the implications are more tangled, because a chip that works well but cannot be understood by the engineers who would produce, validate, or iteratively improve it has real costs downstream.
Sengupta’s group has added diffusion models specifically to give designers control over how “alien” the output looks. You can dial between fully novel topologies and layouts constrained to resemble human-interpretable designs, trading some performance for legibility. That lever is practically useful: a circuit that requires weeks of expert analysis to validate has a hidden cost that does not show up in the benchmark.
The honest limitation of the whole approach is data. RFIC design relies on electromagnetic simulation outputs generated in-house and protected by NDA. Unlike code repositories or scientific papers, there is no large public dataset of RF circuit simulation results available for training general models. Sengupta explicitly flags this: the field needs shared, contributed simulation datasets to build foundational models with the kind of coverage that would generalize across process nodes, frequency bands, and circuit types. This is not a problem that more compute or better algorithms can solve on their own — it requires the industry to decide that the collective benefit of shared training data outweighs the competitive advantage of keeping it proprietary. For a field built on NDAs and process-level trade secrets, that is a real cultural shift, not a technical one.
What has been demonstrated already is that RL-based RFIC design does not merely replicate human performance on easy cases. It consistently exceeds best-known human results for the hardest class of silicon RF circuits, and it does so without requiring a human to propose a starting topology. For an engineering domain where design cycles have historically been measured in months and chips in hundreds of millions of dollars, the case for AI-led design is getting harder to argue against on merit alone — the remaining arguments are about validation, interpretability, liability, and data access, which are real but have nothing to do with whether the circuits work.