Princeton's Sengupta lab used reinforcement learning and inverse design — specifying the target electromagnetic behavior and letting the model generate the layout — to synthesize millimeter wave chips that match hand tuned benchmarks in orders of magnitude less time, exposing…
For decades, the people who design the radio chips inside 5G base stations, automotive radar modules, and satellite payloads have guarded their craft like a guild secret. You learn by apprenticeship, lean on proprietary simulation datasets, and trust intuitions that took senior engineers twenty years to build. The Princeton lab of Kaushik Sengupta has spent the last several years trying to make that entire apprenticeship unnecessary. In a 2024 Nature Communications paper and follow-up results documented by IEEE Spectrum, his group has shown that reinforcement learning and diffusion-based generative models can design millimeter-wave radio-frequency integrated circuits (RFICs) from scratch, producing layouts that hit performance numbers the human experts were chasing for years.
The numbers in the most recent Princeton dataset, summarized in the lab's publications page and a 2023 ResearchGate paper on deep-learning-based inverse-designed millimeter-wave passives and power amplifiers, are not incremental. AI-designed power amplifiers across a 30 to 94 gigahertz band reached power-added efficiency between 16 and 24.7 percent, with saturated output power between 16.7 and 19.5 decibel-milliwatts. For context, those figures sit at or above what the same lab's human-designed reference circuits achieved, and the AI produced them in orders of magnitude less time.
That second point is the real story. The traditional pipeline for an RFIC moves from system specification to schematic design to electromagnetic simulation to layout to tape-out, with feedback loops that can stretch a single design cycle from weeks to months. Sengupta's framing, repeated in a Princeton Engineering interview from early 2025, is that the field is "not slowed by physics, it is slowed by iteration." Reinforcement learning collapses that iteration loop. The model proposes a layout, the simulator grades it, and the next proposal is informed by the score. Over thousands of cycles in software, the agent converges on structures that look unfamiliar on a microscope slide: fractal-like inductor geometries, asymmetric capacitor banks, routing patterns no human floorplan would suggest.
IEEE Spectrum describes the work as a shift from handcraft to inverse design, where the engineer specifies the desired electromagnetic behavior and the model solves for a physical layout that achieves it. The diffusion-model angle is newer and more striking. In that variant, the network is trained on a corpus of existing RF layouts and learns to sample entirely new ones, the same way text-to-image generators produce pictures that do not exist in their training set. The novelty is the point: a circuit optimized purely against a simulation objective often settles into a local minimum that matches human conventions, while a generative model can wander into structural territory human designers would never explore.
There is a real caveat, and it lives in the data layer. As the Princeton team has acknowledged in the collaborate.princeton.edu publication record for the millimeter-wave passive and amplifier work, RFIC design lacks the equivalent of ImageNet. There is no large, public, well-labeled corpus of radio-frequency chip layouts and measured performance. Every lab trains against its own private simulations and its own proprietary measurement rigs. That means the gains reported today come from a few well-resourced groups, and the field is reproducing their results more slowly than the underlying method would suggest. A community-scale electromagnetic-circuit ecosystem, the kind that already exists for vision and language, would be the next unlock, and it does not exist yet.
What changes in the near term is which skills become valuable inside a radio-chip team. Schematic intuition still matters for setting system-level targets, and electromagnetic expertise still matters for verifying that a fabricated chip matches simulation. The middle of the design chain, where a senior engineer spends weeks tuning matching networks and walking layouts around parasitic losses, is where the apprenticeship premium is being repriced. The "dark art" framing that has followed RFIC work for a generation was, in retrospect, a description of an unsolved search problem. The Princeton results suggest the search is now tractable, and the people who learn to direct the search will replace the people who perform it by hand.
The next move to watch is whether the same inverse-design pattern generalizes from passive components and amplifiers to full transceivers, and whether a major foundry or fabless radio chip vendor publishes an AI-designed part in a shipping product within the next design cycle. The technique is ready. The organizational shift is not.