Magnetic Soliton MTJ Devices for Neuromorphic Computing Applications

Magnetic soliton-based devices, including domain walls and skyrmions, hold significant promise for applications in energy-efficient storage and computing paradigms, both conventional and non-Von Neumann. Using micromagnetic simulations in multilayer ferromagnetic systems, we present a magnetic soliton-based devices that exhibit notable synaptic and neuron-like characteristics. The device characteristics can be tuned to act as a synapse or as a neuron. We demonstrate the utilization of spin-orbit torque (SOT) for controlling synaptic potentiation and depression. By simple device structure modifications, the devices work as a synapse and or leaky integrate and fire (LIF) neuron. Leveraging the synaptic characteristics of the device, we show numerical evidence of the system-level performance of these devices in 3-layer feedforward artificial neural network architectures and 3-layer spiking neural networks. Upon training and testing on the MNIST dataset, the system achieves a recognition accuracy of around 90%, in the ANN (as synapse) and 97% in the SNN (as neuron), which highlights the applicability of these spintronic devices in neuromorphic computing.