Cornell researchers teach ‘microwave brain’ a new language

One year after unveiling a first-of-its-kind "microwave brain" microchip capable of computing on ultrafast data and wireless signals, researchers from the Cornell Duffield College of Engineering have shown how the chip can encode information into its own language.

The work builds on the world's first integrated microwave neural network designed by Bal Govind, M.S. '24, Ph.D. '26, and experimentally demonstrated with Maxwell Anderson '20, M.S. '24. Together, they showed that the low-power chip could harness the physics of microwaves to emulate the pattern-finding abilities of the brain and perform computations almost instantaneously.

In a new study published July 29 in Nature Communications, the researchers found that the device can now use what they describe as microwave token embeddings - similar to the tokens used in large language models - to encode messages into radio signals and compress data, capabilities that could enable faster, more secure communications for satellites, drones and other technologies.

"I like to think of it as establishing a microwave lexicon," said Govind, who led the study in the laboratory of senior author Alyssa Apsel, the IBM Professor of Engineering in the School of Electrical and Computer Engineering. "You could have one microwave neural network send messages that only another of the same kind could understand."

The chip's nonlinear microwave physics can transform information - such as navigation commands for a drone or satellite - into distinctive microwave pulse "tokens" that preserve relationships between pieces of information. This requires far less bandwidth and energy than conventional communications systems that first convert analog radio signals into digital data before extracting useful information.

"We're letting the physics do the work," Govind said. "Instead of transmitting something like a line of computer code, those instructions could be represented by just a few microwave pulses that another microwave neural network could immediately interpret."

Because every microwave neural network has its own physical characteristics and produces a large array of frequencies, it can be reconfigured for different sensing and computing tasks.

The researchers believe the technology could provide a new form of hardware-based cybersecurity. Decoding a transmission would require not only another microwave neural network, but also the correct sequence used to configure it, "almost like a public-private key scheme," Govind said.

The researchers also found that feeding gigabit-per-second data streams into the chip causes it to naturally generate probabilistic bits, or "p-bits," whose values depend on the incoming data rather than remaining fixed as zeros or ones. To demonstrate the capability, the researchers reconstructed a satellite image of a tropical storm system that preserved many of its key features while reducing the amount of transmitted data by about eightfold.

"Small satellites often can't transmit massive image files back to Earth because of power limitations or bandwidth regulations set by the Federal Communications Commission," Govind said. "They often have to transmit simple things like GPS coordinates that are going to lose a lot of the main features, and that's something we think our chip can help with."

The researchers have a patent pending and are participating in the Ignite Innovation Acceleration program through the Cornell Center for Technology Licensing, which is helping to advance the technology toward commercialization for low-power satellite communications and edge computing. The research is also supported through a long-standing collaboration with defense and information technology company L3Harris, and through a Kavli Institute at Cornell Engineering Graduate Fellowship.

Syl Kacapyr is associate director of marketing and communications for Duffield Engineering.