UCLA Researchers Advance Physical AI Technology

How Hardware Becomes the Neural Network

Digital computing using silicon chips has powered the growth of artificial intelligence, but scaled models demand increasingly heavy resources in energy, water, and infrastructure. At the same time, edge applications such as robots, satellites, and distributed sensors must process information directly where it is generated, often under strict power limits. According to researchers, software and hardware are no longer separate in this emerging platform. Instead of running a neural network program on traditional chips, computation occurs directly within the physical structure of a self-organizing material.

Over the last 15 years, scientists have harnessed the collective behavior of these materials to process complex information in real time with low power demands. This approach helps advance physical AI, embedding learning and computation straight into the physical hardware that interacts with the real world. Adam Stieg, a UCLA research scientist and associate director of the California NanoSystems Institute, and James Gimzewski, a distinguished professor of chemistry at the UCLA College, led early studies introducing this method. Stieg collaborated with review co-author Zdenka Kuncic, a physicist at the University of Sydney, to further develop the technology.

“Silicon-based electronics have shaped how we think about computing, but they’re not the only way to do it,” Stieg said, describing how models evolve inside the physical network.

Processing Data at the Edge for Satellites and Robotics

Inspired by the cortex of the human brain, self-organized networks process complex data streams with minimal power usage. According to the review paper in Nature Reviews Physics, these networks successfully perform machine learning benchmarks like speech and image recognition in real time. They achieve this by exploiting the physical behavior of the network rather than executing a conventional software model.

This capability proves vital for edge computing, which happens wherever sensors gather real-world data. Modern sensors generate massive volumes of information, but only a fraction is ultimately useful. Conventional systems must digitize these data and rely on algorithms to filter patterns. That pipeline becomes costly in environments like space exploration. Satellites often collect more data than they can transmit back to Earth, requiring filtering before transmission.

Self-organizing physical networks solve this by adapting their structure dynamically to incoming signals. Such systems allow artificial intelligence to operate continuously and locally in resource-constrained settings.

“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.”

Emulating Biological Efficiency Without Replicating the Brain

Brain-inspired computing dates back to the 1980s and continues in today’s neural network software. However, self-organized nanoparticle and nanowire systems physically embody that metaphor. Researchers aim to produce technology with brain-like capabilities that handle changing streams of information efficiently while drawing far less power than conventional hardware.

At the same time, the research team acknowledges the limits of the biological comparison. The goal focuses on extracting specific properties that make biological systems effective rather than building a literal artificial brain.

“The original inspiration was the brain, but we were never trying to build one,” Stieg said. “I don’t think of the brain as a computer, and I don’t think we can reproduce the full richness of a biological system.”

Realizing the full potential of hardware-based neural networks requires collaboration across multiple disciplines. Theoretical physicist Francesco Caravelli of the University of Pisa served as the first and corresponding author of the review. The work also included contributions from physicist Carlo Ricciardi of the Polytechnic University of Turin and experimental physicist Gianluca Milano from Italy’s National Institute of Metrological Research. Funding was provided by the European Research Council, the European Union’s Next Generation EU program, New Zealand’s Marsden Fund, New Zealand’s MacDiarmid Institute for Advanced Materials and Nanotechnology, and the U.S. Department of Energy.

Frequently Asked Questions

What is a hardware-based neural network?

It is an emerging computing platform where computation occurs directly within the structure of a self-organizing material, such as nanowires or nanoparticles, rather than executing software on a traditional silicon chip.

How do self-organized networks save energy?

By processing sensor data locally at the edge—such as inside a satellite or autonomous robot—these networks eliminate the need to transmit raw, uncompressed data back to centralized cloud servers.

Who leads research into physical AI at UCLA?

Adam Stieg, a research scientist and associate director at the California NanoSystems Institute, and James Gimzewski, a distinguished professor of chemistry at the UCLA College, helped pioneer this research approach.