Evaluation of Memory Capacity and Time Series Prediction Using a Spin Hall Oscillator as Reservoir

Abstract

Spintronic devices, such as spin Hall oscillators and spin transfer torque oscillators, have become popular candidates for use as the reservoir in the reservoir computing architecture. The memory capacity of a reservoir quantifies the amount of information retained at any given time, indicating its significance for different reservoir computing tasks. In this work, we consider a spin Hall oscillator, consisting of a bilayer structure of platinum/permalloy, as the reservoir. Using micromagnetic simulations, we evaluate the change in memory capacity of the spin Hall oscillator system as the magnetization dynamics varies from transient state oscillations to limit cycle oscillations. We also perform a three-bit parity task to study the performance of the oscillator in carrying out a nonlinear task. A time series prediction task, namely the NARMA2 task, is also performed. The results of both tasks confirm a correlation between the evaluated memory capacity of the reservoir and its efficiency in performing temporal tasks. The best performance in reservoir computing tasks is observed when the output magnetization dynamics of the oscillator consists of both transient state and limit cycle oscillations.