3 March 2021

Magnetic materials attract attention for reservoir computing

The study of domain walls in ferromangnetic nanorings has demonstrated behavior which has the potential to be useful in the development of reservoir computing, a brain-inspired computing method.

Murmuration of Starlings
Murmurating starlings appear to behave as one despite there being many hundreds of birds. "Starlings murmurate around the i360" by susie2778 is licensed with CC BY-NC 2.0 (https://creativecommons.org/licenses/by-nc/2.0/)

There are times in life when good things seem to happen for no apparent reason. This is true across science and nature. Seemingly random instances coming together in what seems like an organised manner.

In nature, we are familiar with the way that a murmuration of starlings form patterns in the sky as they flock, despite there being hundreds or even thousands of individual birds. They change direction together and appear to be absolutely in sync with each other. Which one decides the direction and timing of their movements? It’s a phenomenon which scientists are working to understand.

But systems where apparent randomness leads to predictable behaviour are also observed in technology, and a group of interdisciplinary scientists and engineers at the University of Sheffield have been investigating how unpredictable properties in certain magnetic materials can be harnessed in the world of computing.

Traditionally, computing has been based on flows of digital information - streams of 1s and 0s flowing through circuits to present a final outcome. Granted, this is a very simplistic description of something that is already highly complex. But animals are able to solve a great variety of complex problems in uncertain situations, very efficiently. Can we gain an understanding of how brains function and learn to create alternative types of computation capable of a commensurate level of problem solving?

Neuromorphic computing is the research area that tries to address this question. Professor Eleni Vasilaki in the Machine Learning group of the department of Computer Science has been seeking inspiration from brain functionality to develop alternative computing methods. The hope is that such methods could be implemented in energy efficient hardware: the brain consumes approximately 10W of power to perform everyday tasks, while for a digital computer to perform the same level of computation, it would require in the region of one million times more power.

Reservoir computing is a brain-inspired method that utilises a number of units representing neurons, i.e. the “reservoir”, that are randomly connected, and yet can solve tasks via relatively simple learning methods. This is due to the fact that the complex form provided by the random connectivity of the reservoir can, perhaps against our intuition, simplify the problem to be solved. Most importantly, it is not necessary to use artificial neurons per se; the reservoir can be replaced with a magnetic material that is able to provide a similar complexity. Eleni and her team work towards improving the learning methods in reservoir computing. 

How can we then build efficient hardware for reservoir computing? Here we turn to the Department of Materials Science and Engineering, and Professor Dan Allwood’s Functional Magnetic Materials Research Group.

Reservoir computing has been in development for a few years now, and the concept has been proven using a variety of different systems, from piles of carbon fibres to buckets of water. However, these are not reproducible nor scalable to an industrial level.

The study of magnetic materials, and the mobile domain walls within these materials in particular, has demonstrated the potential for these materials to be used to advance reservoir computing technology. A significant barrier to the development of magnetic nanowire computer devices has been that the unpredictable behaviour of domain walls.

Applying magnetic fields to arrays of ferromagnetic nanorings has led to observation of emergent behaviour of domain walls in the array overall, similar to what is seen in murmurating starlings. This is despite individual domain walls still behaving in an unpredictable way. Nanorings are used as they restrict the number of domain walls which are adjacent to each other, and thus there is greater probability that neighbouring domains will be affected by the incoming signals.

Nanorings are being investigated for use in reservoir computing
Images of nanorings taken by magnetic force microscopy show how the nanorings interact. The domain walls appear at the intersection of the rings. The coloured circles in the right hand image represent different ring states.

Research published in Advanced Functional Materials (https://doi.org/10.1002/adfm.202008389) describes how these materials have been developed into reservoir computing systems capable of voice recognition tasks, with the ability to recognise nuances in up to eight different voices.

While initial investigations into the use of magnetic nanorings have been encouraging with the ring arrays exhibiting the characteristics required for reservoir computing and are likely to be part of the landscape as the technology advances further. One thing is for sure though: this recent work marks a significant step forward in the development of repeatable and predictable materials systems.

Taking these initial findings further, scientists at the University of Sheffield have started projects in collaboration with materials scientists, computer scientists and electrical engineers from the University of York and institutions around Europe to create a robot which responds to voice commands. We’re not quite at the stage of recreating a murmuration of starlings, but we will follow progress with interest.

Flagship institutes

The University’s four flagship institutes bring together our key strengths to tackle global issues, turning interdisciplinary and translational research into real-world solutions.