AI Generates Bizarre RF Chip Layouts That Outperform Human Designs

A 30-to-100 GHz silicon power amplifier designed entirely by a reinforcement-learning agent hit the best combination of bandwidth, output power, and efficiency ever reported for a silicon-based millimeter-wave PA. The layout looks like a QR code someone dropped in a blender. That's not a bug. That's the point.

Kaushik Sengupta's group at Princeton published the results in 2023 and has since extended the approach to multiport circuits, low-noise amplifiers, and subterahertz designs. The method tosses out every human-generated circuit template and starts from nothing. The AI learns the game of electromagnetic design by playing against itself, much like AlphaGo Zero mastered Go without human games.

Reinforcement Learning Designs Circuits from Scratch

Traditional RFIC design is a dark art. Months of simulation, reliance on decades-old templates, and impedance-matching puzzles that make even seasoned engineers curse. Sengupta's team replaced that with a two-stage pipeline. First, a reinforcement-learning agent explores the combined space of architecture, topology, and device parameters. It discovers trade-offs by iterating thousands of candidate circuits, mapping performance without human bias.

Once the agent settles on a topology, inverse design takes over. The system uses a convolutional neural network emulator trained on millions of random pixelated structures labeled with their scattering parameters. The emulator predicts electromagnetic behavior in milliseconds instead of the hours a full-wave simulator would need. That speed lets the algorithm solve the inverse problem: given a desired set of S-parameters, what physical layout produces them?

The result? A power amplifier whose passive network looks nothing like the symmetric, lacelike filigree humans draw. It works because the AI optimizes across all physical constraints simultaneously, something a human juggling heuristics cannot sustain.

Diffusion Models Make AI Designs Interpretable

Nonsense-looking layouts are fine for performance, but engineers need to debug chips. If a layout is a black-box mess, troubleshooting becomes a nightmare. So the Princeton team added a diffusion model that allows designers to dial in spatial frequency.

Set the dial low, and the model generates classical, interpretable structures that look like traditional RF passives. Crank it up, and you get the wild, pixelated shapes that push performance. The model takes S-parameters as input and outputs a layout in about six minutes. Sengupta calls it a vending machine for electromagnetic structures: specify physically realizable parameters, get a chip back.

This hybrid approach means AI can both discover novel architectures and accelerate the creation of conventional ones. The same framework that produced the record-setting PA can also churn out a textbook-compliant matching network on demand.

What It Takes to Scale: Open Data and Trust

The current method works, but it's not a silver bullet. Hallucinations happen: the AI sometimes generates designs that pass the emulator but fail when simulated with real physics. Human verification remains mandatory. Worse, the data problem looms large.

ImageNet-style datasets for RFICs don't exist. Every company and lab simulates similar structures behind NDAs. Natcast, the U.S. CHIPS Act program that aimed to build shared infrastructure, shut down before it could deliver an RFIC dataset. Sengupta argues the community must build open repositories of electromagnetic simulation data to train foundational models that learn the universal physics of circuit behavior.

If that happens, this approach scales beyond RFICs to any domain where inverse design applies: antennas, metamaterials, even optical circuits. The genie is already out of the bottle; the question is whether the field will pool enough data to keep it working reliably.

Source: AI Is Designing Radio Chips That Humans Couldn't Even Imagine
Domain: spectrum.ieee.org