Testing in Chen’s lab at Duke shows an AI model that can identify thousands of images transmitted wirelessly with high accuracy in the blink of an eye. Credit: Duke University

As drones survey forests, robots navigate warehouses and sensors monitor city streets, more of the world’s decision-making is occurring autonomously on the edge—on the small devices that gather information at the ends of much larger networks.

But making that shift to edge computing is harder than it seems. Although artificial intelligence (AI) models continue to grow larger and smarter, the hardware inside these devices remains tiny.

Engineers typically have two options, neither are ideal. Storing an entire AI model on the device requires significant memory, data movement and computing power that drains batteries. Offloading the model to the cloud avoids those hardware constraints, but the back-and-forth introduces lag, burns energy and presents security risks.

Researchers at Duke University are exploring a third option, called WIreless Smart Edge networks (WISE), that bypasses the limitations of both approaches. They’ve shown that large AI model weights can be smartly embedded in the form of radio waves delivered over the air between devices and nearby base stations, opening a path to energy-efficient edge AI without the usual cost in energy, speed or size.

This work, published online in Science Advances on January 9, is led by Tingjun Chen, the Nortel Networks Assistant Professor of Electrical and Computer Engineering, alongside Dirk Englund’s team at the MIT Research Laboratory of Electronics (RLE).

At the heart of the approach is a concept called in-physics analog computing.

Traditional digital computing occurs through binary code. Devices convert data into ones and zeros, move those bits into a digital processor and compute long sequences of math operations. Even a simple task like unlocking a phone with biometrics triggers a rapid sequence of calculations. It’s reliable but not efficient for small, battery-powered devices.

In-physics computing works differently. Instead of shuttling ones and zeros from an edge device to a distant processor, the natural behavior of radio waves completes part of the math along the way.

In WISE, a base station stores the full AI model and broadcasts a radio frequency (RF) signal that encodes the model’s weight values—numbers required to complete those calculations. When the signal reaches a nearby device, radio hardware in the device mixes the broadcast signal with its own input data that can naturally perform computing directly in the RF or analog domain. One example is a passive frequency mixer that "approximates" the multiplication of two time-domain RF signals. That analog in-physics mixing process—directly taking place at RF—performs a key step in most deep learning models without the need of a digital processor.

"We’re taking advantage of computations that common, miniaturized electronics already gives us," Chen said. "Instead of running every step of the model on a chip designed for digital computing, the radio waves themselves help carry information in a way optimized for the computation."

Because the device doesn’t store the entire model or run it digitally, it overcomes the big memory and energy costs that limit edge AI today.

Zhihui Gao, a Ph.D. student in Chen’s lab and lead author on the paper, said the idea could benefit many kinds of devices. Drones, cameras and traffic sensors all generate data continuously, yet they struggle to run the advanced models that would help interpret those data.

"Technology is moving toward smaller devices that can do more than ever before," Gao said. "In order to achieve that, we need new improvements in edge computing. With WISE, we have shown how devices can run on powerful AI without relying on heavy chips or distant servers."

Gao noted another advantage of WISE is its ability to use existing infrastructure. Base stations already set up for 5G, emerging 6G or WiFi routers could be augmented to broadcast these AI models with relatively small adjustments. Plus, everyday wireless devices already contain the hardware, such as frequency mixers, needed to perform the in-physics computation.

"We’re not adding exotic components or creating entirely new hardware," Gao said. "We’re reusing features that are widely deployed and don’t consume extra energy."

In experiments, WISE achieved nearly 96% image classification accuracy while consuming more than an order of magnitude less energy than leading digital processors.

Although promising, WISE is still in its beginning stages. The current prototype works over short distances, but longer-range testing would require stronger transmission or integration with next-generation wireless gear. And while the approach is flexible, broadcasting multiple AI models simultaneously would require efficient multiplexing of the time-frequency-space resources or additional spectrum bandwidth.

Even so, the researchers see broad potential in applications. One base station could support a swarm of drones in a search and rescue mission or help traffic cameras coordinate intersection signals.

"This is the next step in wireless technologies becoming as powerful as wired ones," Chen said. "Beyond delivering data and information, these findings open a new direction, in which future networks may distribute intelligence by blending communication and computation to enable energy-efficient edge AI at massive scale."

More information: Zhihui Gao et al, Disaggregated Machine Learning via In-Physics Computing at Radio Frequency, Science Advances (2026). DOI: 10.1126/sciadv.adz0817. www.science.org/doi/10.1126/sciadv.adz0817

Citation: Radio waves enable energy-efficient AI on edge devices without heavy hardware (2026, January 10) retrieved 10 January 2026 from https://phys.org/news/2026-01-radio-enable-energy-efficient-ai.html

This document is subject to copyright. Apart from any fair dealing for the purpose of private study or research, no part may be reproduced without the written permission. The content is provided for information purposes only.