Physical artificial intelligence hardware developed by UCLA researchers and international collaborators processes complex data locally through self-organizing networks of nanowires or nanoparticles, according to a forward-looking article published in Nature Reviews Physics. The system bypasses conventional software-driven neural networks by letting hardware compute directly within structures measured on the scale of billionths of a meter, offering a low-power alternative for edge computing in smart devices, autonomous vehicles, and satellites.
Hardware-Based Neural Networks Replace Software Models
Traditional digital computing relies on silicon chips that separate software instructions from physical hardware. In contrast, self-organizing physical networks integrate computation directly into the material structure, according to UCLA research scientist Adam Stieg. Over the past 15 years, Stieg and distinguished professor of chemistry James Gimzewski have helped lead investigations into collective behavior within these nanoscale systems. Stieg collaborated on the recent review alongside Zdenka Kuncic, a physicist at the University of Sydney in Australia.
“Silicon-based electronics have shaped how we think about computing, but they’re not the only way to do it,” Stieg stated according to UCLA materials. “In our systems, the model evolves in the physical network itself. It adapts and changes.”
Did you know? Unlike standard AI models that run static algorithms on passive silicon platforms, self-organizing physical networks physically reorganize their structure in response to incoming electrical signals to learn and compute locally.
Edge Computing and Physical AI Applications
Self-organizing networks process complex information in real time using minimal electrical power by exploiting the physical behavior of nanoparticles and nanowires. According to the review paper, these networks have successfully handled machine learning benchmarks such as speech and image recognition without executing conventional software algorithms.
This capability addresses data bottlenecks in edge computing, where sensors gather massive volumes of information but face strict limits on energy, computing power, and communications bandwidth. Satellites and remote industrial robots often collect more data than they can efficiently transmit to Earth. By embedding computation directly into the hardware at the point of data collection, these physical AI systems filter and process information locally before transmission.
“Most AI treats the hardware as a passive platform for running software,” Stieg said. “We’re asking what becomes possible when the hardware itself is adaptive—when the material reorganizes in response to information and becomes part of the learning process.”
Biological Inspiration and Cross-Disciplinary Research
The architecture of these self-organizing systems draws conceptual inspiration from the human brain’s cortex, which handles perception and reasoning with high energy efficiency. However, the researchers emphasize clear boundaries in their biological metaphor. Stieg noted that the team never set out to build an artificial brain or replicate the full complexity of biological systems.
Theoretical physicist Francesco Caravelli of the University of Pisa in Italy serves as the first and corresponding author of the review published in Nature Reviews Physics. Additional co-authors include experimental physicist Gianluca Milano from Italy’s National Institute of Metrological Research and physicist Carlo Ricciardi from the Polytechnic University of Turin in Italy.
Department of Energy, New Zealand’s MacDiarmid Institute for Advanced Materials and Nanotechnology, New Zealand’s Marsden Fund, the European Union’s Next Generation EU program, and the European Research Council.
Frequently Asked Questions
How do self-organizing physical networks differ from traditional computer chips?
Traditional silicon chips execute software-based neural network models on a passive hardware platform. Self-organizing physical networks perform computations directly within the adaptive structure of nanomaterials, bypassing the separation between hardware and software.
What are the primary use cases for physical AI hardware?
Physical AI is suited for edge computing applications—such as satellites, autonomous vehicles, distributed sensors, and industrial robots—where devices must process large amounts of sensor data locally under strict power and bandwidth constraints.
Who authored the recent review on self-organized computing networks?
The review article in Nature Reviews Physics was led by corresponding author Francesco Caravelli of the University of Pisa, alongside co-authors Gianluca Milano, Carlo Ricciardi, Zdenka Kuncic, Adam Stieg, and James Gimzewski.
Explore More Research
Want to stay updated on breakthrough developments in nanotechnology and physical AI? Subscribe to our newsletter or leave a comment below to join the discussion on the future of computing hardware.
Related reading