Quantum materials are making 'neuromorphic computing' a reality

Now, a new study shows a possible way to improve energy-efficient computing.

We often think of computers as more efficient than humans. After all, computers can complete a complex math equation in a flash and recall the name of that actor we always forget. However, the human brain can process complex layers of information quickly, accurately, and with little or no input energy: recognizing a face after seeing it only once, or instantly knowing the difference between a mountain and an ocean. These simple human tasks require a lot of processing and energy input from computers, and even then the accuracy varies.

The creation of brain-like computers with extremely low energy requirements would revolutionize every aspect of modern life. The Quantum Materials for Energy-Efficient Neuromorphic Computing (Q-MEEN-C), a national consortium led by the University of California, San Diego, with funding from the U.S. Department of Energy, has been at the forefront of this research.

Alex Frañó

Alex Frañó, assistant professor of physics at UC San Diego and co-director of Q-MEEN-C, sees the center's work as a phased approach. In the first phase, he worked closely with Robert Dynes, professor of physics and chancellor emeritus of the University of California, and Shriram Ramanathan, professor of engineering at Rutgers University, and together their teams succeeded in finding a way to create or mimic the properties of individual brain elements, such as neurons or synapses, in quantum materials.

Now, in a second phase, the latest Q-MEEN-C study, published in Nano Letters, shows that electrical stimuli delivered between neighboring electrodes also affect non-neighboring electrodes. This discovery, called "non-locality," is an important milestone in the development of a new type of device that simulates brain function, known as neuromorphic computing.

The second phase of the results, titled "Spatial Interactions in Hydrogenated Perovskite Nickelate Synaptic Networks," was published July 28 in Nano Letters.

Frañó, one of the paper's co-authors, said, "In the brain, it is known that these nonlocal interactions are nominal; they occur frequently and with minimal consumption. This is a key part of how the brain works, but there is very little similar behavior replicated in synthetics."

Like many research projects that are bearing fruit, the idea to test whether nonlocality in quantum materials is possible arose during the new crown. Physical lab space was closed, so the team performed calculations on arrays containing multiple devices to simulate multiple neurons and synapses in the brain. While performing these tests, they found that nonlocalization was theoretically possible.

After the lab reopened, they further refined the idea and brought in Duygu Kuzum, an associate professor in the Jacobs School of Engineering at the University of California, San Diego.

This involved inserting hydrogen ions into a thin film of nickelate, a quantum material with rich electronic properties, and then placing a metal conductor on top. A wire is attached to the metal to send an electrical signal to the nickelate. The signal causes the gel-like hydrogen atoms to move to a certain configuration; when the signal is removed, the new configuration is retained.

Demonstrating Signal Integration Using a Pt-HNNO Synaptic Array

That's the nature of memory," Frañó says. The device remembers that you perturbed the material. Now you can fine-tune where those ions go to create pathways that are more conductive and easier for current to pass through."

Traditionally, creating a network large enough to power a laptop would require complex circuits and successive connection points, which would be both inefficient and expensive. the Q-MEEN-C's design concept is much simpler, because the non-local behavior of the experiment means that all the wires in the circuit don't have to be connected to each other: think of a spider's web, where the movement of one part can be felt throughout the entire web.

This is analogous to the way the brain learns: not linearly, but in a complex, multi-layered way. Each learning session creates connections in multiple areas of the brain, allowing us to distinguish not only between a tree and a dog, but also between an oak and a palm tree, or a golden retriever and a poodle.

Until now, these pattern recognition tasks, performed brilliantly by the brain, could only be simulated by computer software. Artificial intelligence programs like ChatGPT and Bard use complex algorithms to simulate the brain's activities like thinking and writing; they do this very well. But software will one day reach its limits if it is not supported by the appropriate advanced hardware.

Frañó is eager for the hardware revolution to go hand in hand with the software revolution, and the possibility of reproducing nonlocal behavior in synthetic materials brings scientists one step closer. The next step will be to create more complex arrays, using more electrodes in finer configurations.

This is a very important step forward in our understanding and modeling of brain function," Dynes said. Demonstrating a system with non-local interactions can give us further insight into how our brains think. Of course, our brains are much more complex than that, but a physical system capable of learning must be highly interactive, which is a necessary first step. We can now think about greater coherence in space and time."

"It's widely recognized that for AI as a technology to really explode, we'll have to find ways to improve the hardware - the physical machines that can perform tasks in conjunction with software. The next phase will be to create highly efficient machines whose physical properties are those of performing learning. This will provide a new paradigm for our AI world."

Reference links:

[1]https://today.ucsd.edu/story/quantum-material-mimics-non-local-brain-function

[2]https://pubs.acs.org/doi/full/10.1021/acs.nanolett.3c02076

[3]https://qmeenc.ucsd.edu/