Anomalous and Topological Hall Resistivity in Ta/CoFeB/MgO Magnetic Systems for Neuromorphic Computing Applications

TL;DR

This study investigates the electrical and magnetic properties of Ta/CoFeB/MgO heterostructures with skyrmions, demonstrating their potential as energy-efficient synaptic devices for neuromorphic computing with high accuracy.

Contribution

The paper combines experimental measurements and micromagnetic simulations to propose and evaluate a skyrmion-based synaptic device for neuromorphic applications.

Findings

Topological Hall resistivity peaks at zero magnetic field

Resistivity decreases linearly with magnetic field, matching simulations

Device achieves ~90% accuracy in MNIST classification

Abstract

Topologically protected spin textures, such as magnetic skyrmions, have the potential for dense data storage as well as energy-efficient computing due to their small size and a low driving current. The evaluation of the writing and reading of the skyrmion's magnetic and electrical characteristics is a key step toward the implementation of these devices. In this paper, we present the magnetic heterostructure Hall bar device and study the anomalous Hall and topological Hall signals in the device. Using the combination of different measurements like magnetometry at different temperatures, Hall effect measurement from 2K to 300K, and magnetic force microscopy imaging, we investigate the magnetic and electrical characteristics of the magnetic structure. We measure the skyrmion topological resistivity at different temperatures as a function of the magnetic field. The topological resistivity is maximum around the zero magnetic field and it decreases to zero at the saturating field. This is further supported by MFM imaging. Interestingly the resistivity decreases linearly with the field, matching the behavior observed in the corresponding micromagnetic simulations. We combine the experimental results with micromagnetic simulations, thus propose a skyrmion-based synaptic device and show spin-orbit torque-controlled potentiation/depression in the device. The device performance as the synapse for neuromorphic computing is further evaluated in a convolutional neural network CNN. The neural network is trained and tested on the MNIST data set we show devices acting as synapses achieving a recognition accuracy close to 90%, on par with the ideal software-based weights which offer an accuracy of 92%.