Physical neural network using skyrmion-based spin torque nano-oscillators

Abstract

Abstract Due to physical limitations on the miniaturization of traditional electronic devices, architectures based on emerging principles have become the focus of current research to meet the needs of rapidly developing information technologies in the post-Moore era. Neuromorphic devices hold huge potential for use in future artificial intelligence (AI) chips beyond conventional architectures. Benefiting from a wealth of nonlinear dynamic characteristics of spin torque nano-oscillators (STNOs), studies of neuromorphic computations and their applications based on STNOs are attracting growing attention. In this article, at first, we construct a magnetic skyrmion-based STNO and analyze its characteristics; on this basis, we propose a physical echo state network (ESN) including eight skyrmion-based STNOs, which is utilized to implement an image recognition task. Micromagnetic simulations of the nonlinear response of skyrmion-based STNOs to current pulses imply that such a physical neural network has remarked performance in handwritten digit recognition. The high precision, low energy consumption, and fast processing speed of STNO-based neuromorphic devices are desirable in multitudinous practical applications, possibly leveraging the use of STNO-based physical neural networks in the field of artificial intelligence.