Artificial intelligence is developing at a rapid pace, necessitating that researchers discover innovative ways to improve the efficiency of computing power.

This includes research aimed at developing computer chips that behave like biological neurons. This would result in more efficient computing mechanisms and a better ability for AI to learn and adapt like a brain, with billions of neurons communicating efficiently.

Making this possible requires electronic devices that don’t operate like traditional computer chips, and the study of new materials provides avenues for possible future applications.

Dr. Timothy Brown, assistant professor in the School of Electrical and Computer Engineering in the College of Engineering, Architecture and Technology at Oklahoma State University, supported a research project titled “Electrothermally Induced Channel Formation in a Spin-Crossover Neuron,” while a post-doctoral researcher at Sandia National Labs in California, which was published in the journal ACS Nano. The work combined challenging experiments and simulations from a team of researchers from Sandia National Lab, Stanford University, and Texas A&M University under the Department of Energy-supported reMIND center.

He and his team studied lanthanum cobalt oxide (LaCoO3), a material that can foster neuronal behavior. When electrified, tiny conductive pathways form inside the material, allowing current to flow and causing a switch between different electrical states.

Brown provided expertise in Raman Microscopy, which involves illuminating a material with a laser and detecting how atoms are bonded by how the energy of the detected light changes. This allowed the team to investigate how the bonding changes during electrification, detailing how the channels form in LaCoO3.

This breakthrough resulted in two important discoveries. First, conduction pathways don’t stay in one fixed location, which could help prevent an area of a device from overheating and contribute randomness to the state switching that can be leveraged for efficient computation. The shape of a device and the surrounding electrical field suggest that future devices could be designed to control where these pathways form.

Secondly, the team discovered that these pathways are linked to small changes in the material’s atomic structure. The team observed different types of atomic bond stretching associated with pathway formation. This suggests that when a device operates at higher power, larger structural changes occur and pathways become more stable.

This provides possible advantages over traditional computer chips, which, although great for repetitive, high-accuracy , and logical tasks, have difficulty with learning and adaptation. Neuromorphic chips would allow the device to “remember past electrical activity,” similar to how the brain can store memories.

Tasks like seeing a picture of a flower and imagining its scent or predicting the next part of a novel based on the author’s tone require a different type of processing.

“One comparison that brings this into perspective is that training datacenters to classify images takes as much energy as it takes to run a small city for several days; whereas the brain does all this and more with 20 W of power, so 1/3 of a lightbulb,” Brown said.

“So, we ask what hardware the brain uses, and it’s not switches, but instead neurons that talk to each other by creating electrical spikes called action potentials, and the synapses that connect them get stronger or weaker depending on how they talk to each other. Our field is trying to build new computer hardware with artificial neurons and synapses to be able to train data centers with lightbulbs’ worth of energy instead of cities’ worth.”

The team’s discovery suggests that electrical fields create the conduction channels, while changes in the atomic structure of the material determine if those channels move or stay in one place. Because it can mimic neuron-like behavior and synapse-like functions, a single device could perform both functions.

“I’m always so grateful whenever any project finally gets to print because it is a long process, but a journal like ACS Nano is especially important because it publishes high-quality science and many people read it, so our message about conduction channels in LaCoO3 and how they can potentially be designed to advance neuromorphic AI can get out to many people,” Brown said.

By exploring materials that mimic neurological activity, he is advancing knowledge that could help shape the future of artificial intelligence and computing, pushing the boundaries of discovery in pursuit of solutions that impact society.