Does This Count as Alien Technology?
Armed with AI-based tools, engineers are creating extraordinarily efficient wireless chips with highly unorthodox circuit designs.
As microchips shrink to ever smaller sizes, where features are just a handful of atoms across, the design of their circuits grows much more difficult. That difficulty is compounded exponentially in the world of wireless chips, where engineers must design not only standard electronic circuits, but also electromagnetic (EM) structures such as antennas and resonators. Every minor change introduces new properties that can either aid or hinder the function of the device. And since there is virtually an infinite number of possible designs, the challenge is very great.
While engineers have done an excellent job of designing cutting-edge wireless technologies in recent years, a team at Princeton University and the Indian Institute of Technology believes that they could do even better with some assistance. Their idea involves augmenting the design process with an artificial intelligence (AI)-based tool. Given a set of design parameters, this tool generates the necessary EM structures and supporting circuits required to fulfill them. Early experiments show that this system can come up with some very efficient — and sometimes very unexpected — designs.
Traditional design methods rely heavily on iterative processes and pre-defined templates for combining active circuit elements with passive EM structures. While effective to an extent, these methods are limited by their reliance on human intuition, time-intensive parameter sweeps, and preconceived design rules. As a result, the design space explored is inherently constrained, often falling far short of fundamental performance limits. The team’s novel AI-driven methodology overcomes these limitations by leveraging deep learning models for forward modeling and synthesis of arbitrary multi-port RF and sub-terahertz EM structures, unlocking new possibilities for functionality and efficiency.
The structures that can be designed include a range of components, such as filters, resonators, power splitters, combiners, antennas, and more. The AI-enabled design process eliminates the need for the resource-intensive electromagnetic simulations typically needed to fine-tune these designs, replacing them with deep learning models that can navigate the vast design space of arbitrary pixelated structures. For example, with a 25 × 25 grid representation, the design space encompasses a number of possible configurations that exceeds the number of atoms in the known universe, a complexity far beyond the reach of traditional optimization methods.
The models operate at a fine resolution, accounting for loss factors at the scale of each pixel (approximately one-hundredth of a wavelength). This capability is crucial for accurately designing compact, high-frequency structures that minimize energy losses. Additionally, the AI explores configurations that go beyond conventional symmetrical geometries, unlocking functionalities such as spectrally-dependent phase relationships or unequal power division, which are traditionally difficult to achieve.
By untethering themselves from traditional assumptions, the researchers have demonstrated the potential of this approach with applications ranging from filters and antennas to end-to-end mmWave circuits. And the team hopes that this is just the beginning. Future research will aim to scale the system for designing entire wireless chips and other complex systems, opening the door to a new era in high-frequency circuit design.