Magnetic soliton-based LIF neurons for spiking neural networks (SNNs) in multilayer spintronic devices

Abstract

Listen

Translate article

Neuromorphic computing, inspired by biological nervous systems, is gaining traction due to its advantages in latency, energy efficiency, and algorithmic complexity compared to traditional artificial neural networks. This has spurred research into artificial synapses and neurons that replicate brain functions. Spintronic-based technologies, particularly domain walls (DWs) and skyrmions (SKs), have shown remarkable potential for brain-inspired computing, facilitating energy-efficient data storage and advancing beyond CMOS computing architectures. Researchers have proposed various DWs- and Sks-based neuromorphic architectures for neurons and synapses. Leveraging magnetic multilayer structures, we propose a magnetic soliton that incorporates both DWs- and Sks-based magnetic tunnel junction (MTJ) device structures to emulate leaky integrate-and-fire (LIF) characteristics. These characteristics are controlled by spin–orbit torque (SOT)-driven motion within ferromagnetic thin films. By strategically placing the reading block and utilizing a combination of SOT and varying demagnetization energy, we achieve modified LIF neuron characteristics in both DW and Sks MTJ devices. The co-action of soliton dynamics across the nanotrack during the application of the current pulse, along with edge repulsion and variations in demagnetization energy, exploits LIF spiking behavior. Theoretical and micromagnetic analyses reveal that the transitory tunable positions of Sks and the total magnetization of the free layer for DWs mimic the membrane potential of biological neurons. Initial studies on multilayer DW-based LIF characteristics showed promise; however, maintaining leaky behavior required a constant negative current, which is energy inefficient. By incorporating the non-volatile properties of skyrmions and adding a chiral Dzyaloshinskii–Moriya interaction term, we further explored LIF dynamics, yielding encouraging results. Our proposed neuron model, implemented in fully connected and convolutional layers, achieves over 95% classification accuracy on the MNIST and Fashion MNIST datasets using a modified spike-based backpropagation method. With nanosecond latency, these spiking neuron devices, when integrated with CMOS, pave the way for high-density, energy-efficient neuromorphic computing hardware.