Pushing the edge of computing, magneto-ionics imagines efficient processing for AI with reduced resource consumption

Artificial intelligence is a resource-intensive technology. A paper recently published in Nano Letters by collaborators at the Virginia Commonwealth University (VCU) College of Engineering and Georgetown University hopes to improve AI’s ability to parse the vast amounts of information it creates by applying magneto-ionics to the established concept of physical reservoir computing (PRC).

“Demonstrating we can make solid-state devices with magneto-ionic materials is an important step into further energy-efficient computing research, and this Nano Letters publication reinforces that,” said Muhammad (Md.) Mahadi Rajib, Ph.D., a postdoc with Jayasimha Atulasimha, Ph.D., Engineering Foundation Professor in the Department of Mechanical & Nuclear Engineering.

What makes a decision?

Our brains make countless complex decisions everyday. Input comes in, we weigh options and decide what to do. Within that simple path are countless identical loops of input, consideration and output as neurons fire in a chain that takes you from cause to effect.

For artificial intelligence, nodes within a neural network receive inputs and provide output, much like the neurons in our brains. These outputs can be sent to other nodes for continued processing, but those outputs need weight to have value. For AI, weight signifies one input or connection is more important than another.

Traditional neural networks have multiple layers consisting of countless nodes like this. Each node requires training in order to weigh things properly. Training consumes processing power, and processing power takes time and energy. Making tasks like analysis and prediction more efficient is how to continuously improve AI technology.

Less training, more efficiency.

Physical reservoir computing reduces the number of nodes an AI needs to train. Only the final output layer needs training in PRC, using a simple method for classification or prediction tasks. A physical “black box” replaces neural network nodes and synapses, like the ones used for AI inference, in PRC and processes inputs by implementing a nonlinear mathematical function with temporal memory.

To explain the inner workings of the black box, imagine two stones thrown into still water. One stone is thrown with high force and the other with low force, creating big and small ripples respectively. If the stones are thrown so the second stone lands before the previous ripples have dissipated, the new ripple is affected by the earlier one. This illustrates the concept of temporal memory. In this analogy, if multiple stones are thrown one after another into still water according to some complex trend, observing the ripples over time allows you to understand the trend and train a simple set of weights to predict the force of the next stone throw from the ripple pattern.

Repeatedly performing this cycle of input, interaction and observation is PRC. It reveals patterns over time that can predict chaotic systems, like market trends or the weather, using techniques like linear regression modeling to plot each output as a single point.

The magneto-ionic approach.

Using this same example above, the “water” in a magneto-ionic PRC is represented by a positive and negative electrode with solid-state electrolyte between them through which ions move when voltage is applied. The application of voltage is equivalent to throwing a stone and the ripple effect is comparable to the movement of oxygen ions in the system.

“In addition to its energy efficiency, a useful feature of the magnetoionic system is that time scales for ion diffusion can be controlled from microseconds to minutes,” Atulasimha said. “This leads to simple experimental demonstration, as no megahertz and gigahertz measurements are needed. One can work at the natural time scales of the target application in practical systems and remove the need for complex frequency conversion, which takes both energy and space due to complex electronics.”

Atulasimha imagines these energy-efficient reservoir systems have applications in edge computing devices like drones, automated vehicles and surveillance cameras. Tasks such as household energy load forecasting, weather prediction or processing hourly readings from wearable devices, which operate on hour-scale data, can also be performed using magneto-ionic PRC without additional preprocessing.

“We showed that the magneto-ionic physical reservoir has both memory and nonlinear behavior, two important properties necessary for using it as a reservoir block,” Rajib said. “Our system stands out because voltage-controlled ion migration is a highly energy efficient method of manipulating magnetization. We demonstrated the required reservoir properties in a physical system and did so using a very energy efficient approach.”

Two labs came together in order to pursue this research. Virginia Commonwealth University collaborators included Atulasimha, Rajib, and VCU Ph.D. students Fahim Chowdhury and Shouvik Sarker. The Georgetown University team included Kai Liu, Ph.D., Professor and McDevitt Chair in Physics, Dhritiman Bhattacharya, Ph.D., Christopher Jensen, Ph.D. and Gong Chen, Ph.D.

Atulasimha’s group illustrated physical reservoir computing using numerical models of spintronic devices and sought a material system to experimentally demonstrate PRC. Liu’s team worked with magneto-ionic materials and was intrigued by the possibility of using them for computing applications.