Reservoir computing in an artificial spin-vortex ice

Abstract

Sparked by the limitations of the Von-Neumann architecture, the burgeoning field of neuromorphic computation aims to develop systems that process information both ‘in-memory’ and in a brain-like manner. Neural networks comprise three types of layers – an input layer, hidden layers and an output layer each connected by numeric weights. Training such networks requires continuous updating of each weight requiring both time and energy. Reservoir computing (RC)1 is a subset of neuromorphic computation where training occurs only on the output layer, reducing time and energy cost significantly compared to traditional neural networks, while competing with traditional approaches for time-series classification / prediction task accuracy.Artificial spin networks comprising patterned nanomagnets serve as promising computational substrates for such applications, information can be stored in a single nanomagnet spin (i.e. a nanomagnetic ‘bit’) or the magnetic configuration (microstate) of the entire network. Strong dipolar coupling enables a highly non-linear history-dependent collective response to external stimuli, a key property for RC1. Artificial spin ice (ASI) is one such network where geometric frustration gives rise to a rich microstate space and emergent collective behaviour2 highly promising for RC3,4. However, in these systems individual nanomagnets are constrained to Ising-like states, severely limiting microstate richness. We have previously shown in simulation that combining Ising-like and vortex states allows powerful reconfigurable functionality5.Modifying a subset of bars6,7 in an ASI such that the Ising and vortex states are bistable gives rise to highly non-linear responses to global- and RF-fields allowing multiple reservoir readouts, detectable by the recently developed ‘spectral fingerprinting’ technique8. Here, we explore ASVI-based RC for forecasting and classifying time-series with varying complexity and assess the strengths and limitations of such systems compared against traditional ASI and competing neuromorphic hardware platforms.