For decades, designing radio-frequency integrated circuits has been described as a "dark art" — a skill mastered only through years of painstaking experience. Now, researchers at Princeton University say artificial intelligence is learning that art, and in some cases surpassing the humans who once held a monopoly on it.
In a feature published by IEEE Spectrum on June 24, 2026, the team detailed how machine-learning methods are being used to generate radio-frequency integrated circuits (RFICs) — the tiny chips that let phones, cars, and satellites send and receive wireless signals — from scratch. Some of the resulting layouts, the researchers note, "look more like modern art than circuit layouts," yet their physical prototypes have bested state-of-the-art human-designed circuits on performance. For more context on this story, see our ongoing AI news.
The achievement, however, is not purely aesthetic. The AI conceived working designs in orders of magnitude less time than a human engineer would require.
Why RF Chip Design Is So Hard
Unlike the central processing units (CPUs) and graphics chips (GPUs) found in computers, RF chips cannot be designed through a standardized, algorithmic process. RFIC design is an exercise in engineering across multiple physical domains at once.
Maxwell's equations govern how electromagnetic fields interact with the active and passive devices on a chip. At the same time, the laws of thermodynamics determine how heat is generated and removed during operation, while the mechanics of thermal expansion dictate how reliably a chip and its packaging survive temperature swings. Accounting for all of these constraints simultaneously makes the design space almost impossibly large, with every decision involving competing priorities.
The result is that a single new RF chip design can take years and tens to hundreds of millions of dollars to complete.
The Frequencies at Stake
The stakes are high because RF chips underpin nearly every wireless technology. A typical CPU's transistors overheat at operating frequencies of just a few gigahertz. RF chips, by contrast, operate at far higher frequencies: 28 and 39 gigahertz for 5G signals, 26.5 to 40 gigahertz and beyond for satellite communications, and 77 gigahertz for automotive radar.
To survive those frequencies, RF chips cleverly manage a signal's energy through careful electromagnetic design. That takes the form of intricate networks of metal elements — geometrically regular, often symmetrical, and so elaborate they can resemble lacelike filigree. These passive structures, such as inductors and transmission lines, dominate the chip's real estate, taking up far more space than the transistors themselves.
How the AI Approaches the Problem
According to the IEEE Spectrum report, the Princeton group began exploring AI-driven RF design roughly seven years ago, in the wake of DeepMind's AlphaGo victory over world Go champion Lee Sedol in 2016. That breakthrough prompted the team to ask whether AI could be taught the "dark art" of RF design as well.
The answer, increasingly, is yes. The researchers use reinforcement learning combined with inverse design to rapidly create RFICs from scratch. Rather than starting from a human template, the system learns which structures produce the desired electromagnetic behavior.
More recently, diffusion models — the same family of generative AI behind modern image synthesis — have been used to rapidly generate novel RF layouts. These models can produce designs that are either deliberately wild or more human-interpretable, and they have achieved record performance while drastically cutting design time.
A concrete example illustrates the shift. An engineer designing a 28-gigahertz power amplifier for a 5G millimeter-wave handset would traditionally begin from established templates and follow symmetric, understandable patterns. Freed from those constraints, the AI can explore designs a human would never draw — and that do not need to be human-understandable to work.
What Comes Next
The researchers caution that the field is not yet at the point of fully automated, push-button RF design. Future progress, they argue, will depend on large, shared chip-design datasets and more open ecosystems, so that AI systems can learn universal electromagnetic and circuit behaviors rather than narrow, lab-specific patterns.
If those building blocks fall into place, the implications extend well beyond radio chips. The Princeton team suggests that AI-enabled design could become the future of all RF design — and possibly much more, pointing toward next-generation wireless technologies including widespread autonomous vehicles, quantum communications, 6G mobile service, and advanced satellite networks.
For an industry where a single design cycle can consume years and nine-figure budgets, even shaving a fraction of that time could be transformative. And as the AI-designed circuits increasingly outperform their human-made counterparts, the long-standing "dark art" of RF engineering may finally be turning into a science.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
