Domain wall device with a graded anisotropy field as artificial neuron for neuromorphic computing

Abstract

Demand for Artificial Intelligence (AI) has been growing due to their usefulness in commercial electronics as well as tackling global issues, such as a pandemic [1]. Current devices that perform AI rely on the conventional von Neumann architecture, which requires extensive programming and power [2]. Therefore, significant attention is dedicated to the brain-inspired neuromorphic computing (NC) as an alternative that is more power-efficient and a more suitable platform for AI [3]. Neuromorphic architecture consists of a network of interconnected synapses and neurons that mimic their biological counterparts [4]. Sengupta et al. have proposed a neuron design based on a domain wall device that uses a magnetic tunnel junction (MTJ) [5] at one end. The tunneling magnetoresistance (TMR) depends on the position of the DW. A current is required to move the domain wall forward or backward (integrate or leak functions respectively). In this work, we place emphasis on designing a domain wall (DW) based artificial neuron, which displays the leaky function without the requirement for a driving current.To achieve leaky function and self-reset process of neuron, we propose a domain wall device, which has an anisotropy field (Hk) gradient in the region of interest (ROI), as shown in Fig. 1(a). This anisotropy gradient is crucial in inducing the automatic returning motion of DW in the absence of current. The design is simulated using Mumax3, and the DW in the free layer nanowire (NW) is driven by spin-transfer-torque (STT). As shown in Fig. 1, current is first applied to propagate the DW towards the threshold position to “fire” the neuron device; current is subsequently removed and DW motion is observed. The micromagnetic parameters used here represent the magnetic properties of CoFeB film with a perpendicular magnetic anisotropy. The magnetic anisotropy constant (Ku) and saturation magnetization (Ms) are graded in ROI, so that Hk gradient is achieved. Some results are presented in Fig. 2(b), showing DW position vs time graph of the reference NW, without any gradient. One may notice that the DW does not return towards the initial position in these devices. Fig. 2(c) corresponds to a device with an Hk gradient. We can notice that the device enables natural return of the domain wall towards the starting position in the absence of current. This suggests that the device displays the leakage function, as well as the ability to reset after the neuron has fired. Fig. 2(d) corresponds to the same device as that of Fig. 2(c), except in this case, three current pulses are consecutively applied with a time interval of 60 ns. These results indicate that the neurons can function with a reset time of about 60 ns. The proposed design has potential applications for DW-based artificial neuron since it displays the integration and leakage functions, while the firing function can be achieved with the utilization of MTJ during the fabrication process. **