Grant supports finding brain-inspired ways to develop low-energy computing

(Press-News.org) The human brain is an astonishing organ, as any neuroscientist can attest. And its ability to collect, store, analyze and use information is intriguing to physicists, engineers and computer scientists, too.

Benjamin Jungfleisch, associate professor of physics at the University of Delaware, is among them.

Jungfleisch, who joined UD’s faculty in 2018, is an expert in magnon spintronics. He uses lasers to explore the dynamics of magnetic nanostructures — tiny magnets that can be used to store and steer information through a circuit.

A primary focus of his work now is finding brain-inspired ways to develop low-energy computing, using interacting nanomagnets as the command center.

Neurons are the brain’s information processors, with electrical and chemical signals carrying information between neurons. In a similar way, magnons — the fundamental quantum excitations that make up “magnetic waves” or “spin waves” in a magnetic system — perform a similar process through arrays of magnetic nanostructures, carrying and processing information in ways that could lead to faster, more energy-efficient processing and even artificial intelligence (AI) devices.

This addresses a critical need, especially now as the energy consumption of AI is skyrocketing. AI has extraordinary potential for our world. But its complexity requires intensive computing power and an ever-increasing number of data centers to manage and meet the computational demand. Without innovative solutions, energy will be an increasing problem for society, industry and the climate.

The National Science Foundation has recognized the significance of Jungfleisch’s work with a 2024 CAREER Award, a five-year grant worth just over $798,000, to support his research team’s efforts to develop low-power computing and processing methods using these magnetic nanostructures. The project also is supported by the Established Program to Stimulate Competitive Research (EPSCoR).

Jungfleisch’s primary collaborator — Jack Gartside, a physicist at Imperial College in London — had a breakthrough in 2022, using effects Jungfleisch discovered in 2016 related to this hysteretic behavior and realized the network could be trained and predictions about the future could be made.

In the future, Jungfleisch said, training cycles that now require two or three hours to complete could take minutes, using novel phenomena such as spin torque.

Learning about these interactions and how to manipulate and tune them for specific purposes is critical to Jungfleisch’s CAREER project, which will address four important challenges:

Controlling magnons in two-dimensional arrays of nanomagnets

Manipulating magnon-magnon interactions

Increasing knowledge of the dynamics in magnetic nanostructures

Demonstrating these next-generation neuromorphic concepts experimentally

Two recent publications in Nature Communications by Jungfleisch and collaborators explain progress made recently in magnon-magnon coupling and nonlinear dynamics.

Three-dimensional work is underway and advancing quickly, Jungfleisch said.

The more recent publication describes a three-dimensional nanomagnetic structure that improves performance over two-dimensional arrays and requires simple fabrication and measurement techniques that are widely available. Among the advances are the demonstration that a three-dimensional magnonic material can be stacked with independently programmable magnetic nanostructured systems.

“You get many more states in your system and a much smaller footprint,” Jungfleisch said. “Storing more information in these networks is easier since you have more space available. Who doesn’t want more neurons?”

You also get more flexibility.

“This gives you a lot of reconfigurability,” Jungfleisch said. “You can change the dynamics of the synaptic behavior as well as the memory.”

Nonlinear dynamics also emerge as interesting phenomena in these networks. For example, Jungfleisch found that one magnon can split into two or two magnons can merge into one in these structures. This provides the foundation for parallel computing and information processing.

“We’ll use this phenomenon to improve the function of the neuromorphics,” he said.

Many questions remain, including how to create and manage randomness and disorder in these systems.

During his upcoming sabbatical, Jungfleisch plans to continue his work with research groups in Germany and India.

And later, as part of his CAREER Award project, he plans to develop a five-week introductory course on magnetism and electricity for adults in UD’s Osher Lifelong Learning Institute and create a wave machine demonstration that would make spin waves visible using an infrared camera. He also wants to develop a wave machine demonstration — using barbeque skewers, duct tape and gummy bears — that would be accessible to students with various disabilities. That device would be available for loan to science teachers in local middle schools and high schools.

Jungfleisch works with nanomagnetic arrays, which can be compared to the brain’s neural networks, the pathways used to move signals along. Magnon connections are akin to the “synapses” that transmit signals along specific circuits.

“These arrays of interacting nanomagnets are essentially just tiny bar magnets,” Jungfleisch said, “like the ones you have on your fridge and the ones children play with. They have a north and a south pole. And if you make them very small — on the nanometer scale — you can pattern them with state-of-the-art lithography, which we have available here.”

When Jungfleisch says “tiny,” he is talking about things you cannot see with your eyes. Nanoscale structures are measured in nanometers. It takes more than 25 million nanometers to make one inch. Much of the work is done in UD’s Nanofabrication Facility, headquartered in the Patrick T. Harker Interdisciplinary Science and Engineering (ISE) Laboratory.

“You can make lattices out of them and they interact,” he said. “They can store information — very similar to what the neurons do in our brain. And the neurons are all connected in a network. So we place these nanomagnets in a network and they feel each other.”

Traditional computers use a processor and memory.

“Data is constantly shuffled between the two and it’s highly inefficient,” Jungfleisch said.

Devices using interacting nanomagnets offer multiple advantages.

“These structures can do it all,” he said. “We do not need electrons, because we use magnetic excitations. And second, we can do processing and storage at the same time in the same unit.

“There are specific tasks such as artificial intelligence where this may be useful — what we do with ChatGPT, for example, or recently emerging chatbots for creating images.”

These nanomagnet networks can be trained, Jungfleisch said. They keep a history and remember the state they are in, but they also need to be susceptible to change and retraining — neuromorphic changes, they are called.

END