AI is finding radio-chip designs engineers would not draw
Princeton-led work shows machine learning can produce unusual RF circuit layouts in minutes, with implications for 5G, radar and satellite communications.
The radio chips inside phones, vehicles, radar systems and satellites are still designed through a slow, highly specialised process. Their layouts often reflect patterns refined by generations of engineers. A Princeton-led research effort is showing what happens when an algorithm is allowed to ignore those visual conventions and search more freely.
Professor Kaushik Sengupta's group has used reinforcement learning, inverse design and diffusion models to build radio-frequency circuit layouts from performance requirements. The results can look irregular—closer to abstract art than a conventional chip—but physical prototypes have matched or beaten strong human-designed circuits.
From a specification to a layout in minutes
In one workflow, a diffusion model produced a structure in roughly six minutes. An earlier AI-designed silicon power amplifier covered 30 to 100 gigahertz and, at the time, delivered a leading combination of bandwidth, power and efficiency for its class.
That speed could matter well beyond smartphones. Advanced RF chips are needed for 6G research, automotive radar, satellite links and quantum communication. Algorithms may also uncover workable designs that an engineer would reject simply because they look unfamiliar.
Engineers are still part of the circuit
A generated layout still has to survive fabrication, heat, packaging and verification. The field also lacks the large shared datasets needed for models to work reliably across manufacturing processes. For now, the practical model is collaboration: engineers set the physics and constraints, while software explores more possibilities than a human team could reasonably draw.
Source: IEEE Spectrum, June 24, 2026.