A microwave neural network chip compresses wireless data and embeds messages directly into radio signals, enabling lower-power, bandwidth-efficient, and more secure communications for satellites, drones, and future 6G systems.
Researchers at Cornell University have demonstrated a microwave neural network (MNN) chip that can compress wireless data and encode information directly into microwave signals, potentially reducing communication bandwidth and power consumption while improving security. The technology could benefit satellites, drones, edge AI devices, and future 6G wireless networks by processing radio signals in the analog domain instead of relying on conventional digital communication methods.
The work builds on the team’s previously developed integrated microwave neural network, a low-power analog AI chip that performs computations using the physics of microwave signals. Instead of first converting analog radio waves into digital data for processing, the chip manipulates microwave frequencies directly, allowing computation and communication to occur simultaneously with significantly lower latency and energy requirements.
A key innovation is the introduction of microwave token embeddings, inspired by the token representations used in large language models (LLMs). The chip converts information, such as navigation commands or sensor data, into unique microwave pulse patterns while preserving relationships between data elements. These compact microwave “tokens” require much less bandwidth than transmitting conventional digital bitstreams, enabling efficient communication between compatible microwave neural networks.
The architecture also introduces a hardware-based layer of security. Since each microwave neural network possesses unique physical characteristics and frequency responses, only another similarly configured chip with the correct initialization sequence can accurately decode the transmitted information. This creates a communication mechanism resembling a public-private key system without depending solely on software-based encryption.
The researchers further demonstrated high-speed probabilistic bit (p-bit) generation by feeding gigabit-per-second data streams into the chip. Unlike conventional binary bits that remain fixed as either 0 or 1, p-bits vary probabilistically according to the incoming data, enabling efficient data compression and probabilistic computing directly in hardware.
To validate the approach, the team compressed and reconstructed a satellite image of a tropical storm, reducing the transmitted data volume by approximately eight times while preserving critical image features. Such capability could help small satellites operating under strict bandwidth and power constraints transmit richer information back to Earth instead of only simplified telemetry.
Published in Nature Communications, the research demonstrates how analog AI hardware can merge computation, compression and secure communications within a single microwave chip. Beyond satellite communications, the technology could support low-power wireless sensing, autonomous systems, edge AI, and future high-speed communication infrastructure where bandwidth, latency and energy efficiency are increasingly critical.
