VCU Engineering researcher is developing magnetic devices that improve efficiency of artificial networks

Jayasimha Atulasimha, Ph.D.
Jayasimha Atulasimha, Ph.D., Qimonda Professor in VCU’s Department of Mechanical and Nuclear Engineering.

Jayasimha Atulasimha, Ph.D., Qimonda Professor in VCU’s Department of Mechanical and Nuclear Engineering, is principal investigator on a $450,000 collaborative grant from the National Science Foundation for a new method that could drastically reduce the power needed to operate artificial neural networks that are used in smartphones, medical technologies and other devices that use embedded systems. VCU and the Massachusetts Institute of Technology (MIT) will each receive $225,000 under this award.

This work is in collaboration with Caroline Ross, Ph.D., Toyota Professor of Materials Science and Engineering at MIT.

Embedded systems comprise processors and other components that execute specific functions within a larger system. More and more, these embedded systems learn and make inferences using neural networks made up of complex, multilayered networks of neurons and synapses. Because smart devices are constantly expected to do more with less power, their neural networks must be able to learn in real time and be energy efficient.

The researchers’ new method could make it possible for smart devices to learn in real time — on their own hardware — bypassing the energy-hungry process of sending data out to an external server for processing.

Atulasimha, known for his work in nanoscale magnet-based memory and computing devices, is proposing to experimentally study and model a device concept that uses electric voltage to control the magnetic state of a nanowire or “racetrack” to update and store multiple memory  states in the device's hardware. 

“We are trying to demonstrate that we can store and update, let’s say, five states in five different magnetic states within the hardware on the device itself,” Atulasimha said. “If we can realize such a nanoscale material system experimentally, it could be used as an enabling element to create a neural network that would be energy efficient, but would not compromise accuracy significantly.”

Ross, an expert in material growth and fabrication of nanoscale magnetic devices, is creating the system’s specialty materials and fabricating the prototype.

“This work takes advantage of recent advances in materials design. We will be using magnetic garnet oxides whose properties can be precisely tuned via their composition, allowing the devices to be scaled to ultra-small dimensions,” Ross said.

Damien Querlioz, Ph.D., a research scientist at the Centre for Nanoscience and Nanotechnology from Centre National de la Recherche (CNRS) at the Université de Paris-Saclay and an expert on neuromorphic computing devices, will be an unfunded collaborator on this project. The current research builds on ideas developed in a paper by Atulasimha’s graduate students Md Ali Azam and Dhritiman Bhattacharya in collaboration with Querlioz, Ross and Atulasimha.

“In our paper, we have already shown the possibility of controlling magnetic states in a nanowire using simulations. Experimental realization of such a system would pave the way towards practical implementation,” Bhattacharya said.

Daniel Gopman, Ph.D., a physicist at the Materials Science and Engineering Division of the National Institute of Standards and Technology (NIST) in Gaithersburg, Maryland will serve as another unfunded external collaborator. His unique facilities for the development and characterization of novel magnetic devices have supported doctoral studies of two recent VCU graduates and plans are underway for the VCU students to gain valuable experience by carrying out a portion of the experimental characterization for this project under Gopman’s supervision at NIST.