Short answer: Princeton University researchers have developed AI systems that design radio-frequency chips with alien-looking, QR-code-like layouts — outperforming circuits crafted by human engineers while slashing design time from months to days.
Radio-frequency integrated circuits (RFICs) power everything from 5G antennas to autonomous vehicle radar and satellite communications. For decades, designing them has been a “dark art” — an intuition-driven craft taking highly specialized engineers months or years. Now, AI is cracking that code, and the chip layouts it produces look like nothing a human would ever draw.
The “Dark Art” of RFIC Design
Unlike digital chip design — highly automated with standard cells — RFIC design remains stubbornly manual. RF chips operate across multiple physical domains simultaneously: electromagnetics, device physics, thermal expansion, and heat dissipation — all interacting at frequencies from 30 to 120 GHz. This complexity creates a bottleneck constraining progress across 5G/6G, autonomous vehicles, and satellite communications.
“The design of an RFIC is an exercise in engineering across multiple physical domains,” explains Professor Kaushik Sengupta of Princeton in a first-person account for IEEE Spectrum.
Three AI Techniques That Design Chips Differently
Sengupta’s group uses three complementary AI methods, treating RFIC design not as an engineering problem but as a game to be learned and optimized.
Watch: Google DeepMind’s AlphaChip uses reinforcement learning to design chip layouts — the same approach Princeton applies to RF circuit design. (Credit: Google DeepMind)
Reinforcement Learning — The AlphaGo of Chip Design
The team formulated RFIC design as an RL problem where the “state” is the current design topology, “actions” modify the circuit, and “reward” is simulated performance. Like DeepMind’s AlphaGo, the RL agent explores the physics-driven design space without human bias, discovering topologies no engineer would conceive. At ISSCC 2025, the team demonstrated RL-based designs spanning 30 to 120 GHz.
Inverse Design — Specify Performance, Get Geometry
Instead of starting with geometry and simulating performance, inverse design flips this: engineers specify desired specs, and the AI computes the geometry. Deep neural networks predict electromagnetic fields, replacing slow EM simulators with instant inference. A December 2024 Nature Communications paper demonstrated this for multi-port RF passives at sub-terahertz frequencies.
Watch: Professor Kaushik Sengupta of Princeton University explains how AI-based inverse design and reinforcement learning are opening a new RFIC design space. (Credit: IEEE Solid-State Circuits Society)
Diffusion Models — Borrowed From Nobel-Winning Biology AI
The team adapted diffusion models — the same architecture behind Nobel Prize-winning protein-folding AI — for RF chip design, calling it RFdiffusion. It rapidly generates novel circuit layouts with a unique “dial” that controls spatial frequency, letting designers tune between alien-looking (max performance) and human-interpretable layouts.
Real Chips, Fabricated and Working
The results look like modern art. “The electromagnetic structures that came from these AI algorithms looked like very complicated QR codes,” Sengupta told GlobalFoundries. “By looking at it, nobody can tell what it does. However, once you add those circuit elements, the entire circuit works exceptionally well.”
These “alien” layouts have been fabricated on GlobalFoundries’ silicon germanium 9HP platform and demonstrated across power amplifiers, passives, antennas, and complete transmitters from 30 to 120 GHz — in many cases outperforming state-of-the-art human designs.
$30M Government Bet: The AIDRFIC Program
Through Natcast — the NSTC operator under the CHIPS Act — the $30 million AIDRFIC program funds three teams: Princeton (~$10M), Keysight, and UT Austin. Partners include RTX, Cadence, Qualcomm, Skyworks, TI, Nokia Bell Labs, and Ericsson.
“Embracing AI for radio frequency design is paramount for maintaining the United States’ leadership in technological innovation,” said Deirdre Hanford, CEO of Natcast, in Princeton Engineering’s announcement.
This funding arrives as the semiconductor industry faces Moore’s Law limits. TekMag recently covered how IBM pushes toward sub-1nm chips — AI-driven design offers a different path: not shrinking transistors further, but arranging them in ways humans never would.
From Months to Days: Industry Impact
Design time collapses from months to “orders of magnitude less time” — real days. This has profound implications:
Democratized RF design — less reliance on “decades of tribal knowledge” that limits the talent pool
Faster 5G/6G deployment — more efficient RF front-ends iterated in days
Cheaper satellite communications — high-performance RF chains at lower design cost
Better automotive radar — AI-optimized 77–79 GHz circuits
Advanced defense systems — next-gen radar and communications hardware
The commercial EDA market — dominated by Cadence, Synopsys, and Keysight — faces potential disruption. Notably, all three are AIDRFIC partners, suggesting they see where the industry is heading.
The Trust Paradox
AI-designed hardware faces a unique challenge: you can’t simply look at a QR-code-like layout and understand why it works. RFdiffusion’s “dial” between alien and interpretable layouts offers a practical bridge, letting engineers build trust gradually while still pushing beyond intuition. The IEEE Spectrum article notes that large, shared chip design datasets and open ecosystems are essential for the field to mature.
As custom chip design reshapes AI infrastructure and AI continues to dominate headlines, this Princeton breakthrough marks yet another frontier: AI is no longer just a software tool — it’s designing the very hardware it runs on.
More from TekMag: Best of TekMag June 2026 — catch up on the month’s biggest stories.
Frequently Asked Questions
Are these AI-designed chips actually being manufactured?
Yes. The circuits have been successfully fabricated on GlobalFoundries’ silicon germanium 9HP platform through their GlobalShuttle Multi-Project Wafer Program — real silicon, not simulations.
How much faster is AI chip design than human engineers?
The methods reduce design time from months to days — Sengupta describes the improvement as “orders of magnitude less time.”
Will AI replace RF chip engineers?
Not entirely. The AI excels at exploring unknown design spaces, but human engineers remain essential for defining specifications, integrating systems, and interpreting results.
Photo credit: Closeup of Princeton’s AI-designed RF microchip showing intricate gold circuitry patterns. Photo by Emir Ali Karahan, Princeton University.
Sources: IEEE Spectrum · Princeton Engineering · GlobalFoundries · Nature Communications
Originally published on TekMag — Your Front Row Seat to the Tech Revolution.
Top comments (0)