Spintronic FitzHugh–Nagumo spiking neuron device for spiking neural networks

Spintronic-based neuron devices for neural network hardware are gaining increasing attention. However, the majority of these devices are designed to replicate a highly simplified neuronal model known as the leaky integrate-and-fire (LIF) neuron. We present a domain wall motion-based magnetic tunnel junction (DW-MTJ) device that emulates the more bio-plausible FitzHugh–Nagumo neuron model. The neuron characteristics are realized using spin–orbit torque (SOT) driven DW motion in a circular nano-pillar. We obtain sustained magnetization relaxation oscillations in the presence of a DC charge current. The shape, frequency, and phase of the FitzHugh–Nagumo oscillations are controlled by SOT and voltage. A thorough parametric analysis of the proposed neuron device structure is done to assess the device viability with different material systems and environmental conditions. The device consumes energy per spike in the range from 9 to 47.5 fJ, depending upon the parameters. Furthermore, we map the device characteristics to the FitzHugh–Nagumo neuron model by tuning the parameters. The device model is integrated into a fully connected, 3-layer spiking neural network (SNN) framework to classify the MNIST handwritten digit images dataset. We train and test the network under different device conditions. The SNN achieves a classification accuracy of more than 98% in all cases, showcasing its potential for efficient, bio-plausible neuromorphic systems.