Can nanomagnetism be exploited for AI tasks?
14 May 2021
- Shikha Gianchandani



Emergent properties of a nanomagnetic system have potential to be harnessed for artificial intelligence tasks, like speech recognition.


​​​Image Credit: Pixabay


​In biology, cells group together to form tissues, which group to form an organ. The organs form an organ system, which then helps carry out the life functions of the complete organism.

Just like when cells group together and have a synergistic effect, connected magnetic nano-rings are more than the sum of their parts, with new emergent behaviour arising from their interactions. Scientists from the University of Sheffield, ALBA Synchrotron Light Facility and Diamond Light Source worked alongside ISIS scientists to exploit this property to perform speech recognition tasks, using these rings as a “reservoir" computer. 

 In their study, published in Advanced Functional Materials, the scientists demonstrated how the nano-rings magnetically interact with each other through their domain walls (regions in the ring where the magnetism abruptly changes direction). This allows the rings to “talk" to each other, but also, crucially, allows them to change how much they talk to each other, since the domain walls can actually be destroyed or created by repeated interactions with each other. In this study, these interactions were instigated using a rotating magnetic field of different strengths. The interactions are physical analogy of a neural network (the basis of artificial intelligence) “learning" which connections should be kept, and which are not important and should be ignored.

Before illustrating how this nanomagnetic system can be exploited for powerful automated tasks, its properties had to be understood. Polarised neutron reflectometry measurements were carried out using the ISIS instrument, Offspec. Offspec is an advanced reflectometer that allowed a quantitative measure of the array's magnetisation. This informed the researchers of the equilibrium between creating and destroying domain walls across the array, when changing the strength of the rotating driving field. Since the array “learns" by interacting, this understanding allows the sweet-spot in driving field to be found; too little interaction means the array will never change or learn, too much and it will destroy all the domain walls and never learn.

In normal applications, the differing strengths of interactions between neighbouring nano-rings has limited commercial applications, since they are hard to control. In this study however, the researchers were able to show that, because the interaction strength was able to vary, this opens the door for applications in this next generation “reservoir computing"

To explore the emergent properties, the researchers devised a simulation to perform a voice recognition task, classifying spoken digits of zero to nine by eight different speakers. The approach using magnetic nano-rings achieved close to a 100% success rate.

Thus, nanomagnetic systems can play a pivotal role in carrying out advanced artificial intelligent tasks, like speech recognition, and how instruments like Offspec play a powerful role in these developments.

Further information:
The full paper can be found online at DOI: 10.1002/adfm.202008389

Contact: Gianchandani, Shikha (STFC,RAL,ISIS)