Multiple spiking functionalities in annealing-optimized Ag/Hf$_{0.5}$Zr$_{0.5}$O$_2$-based memristive neurons

TL;DR

This paper presents a scalable, energy-efficient artificial neuron using Ag/Hf$_{0.5}$Zr$_{0.5}$O$_2$ memristors with multiple spiking functionalities, achieved through a novel annealing process.

Contribution

It introduces a two-step annealing method to enhance memristor parameters and demonstrates a neuron with multiple spiking modes without extra electronic components.

Findings

Achieved leaky integrate-and-fire behavior in multiple spiking modes.

Improved memristor crystallization and Ag diffusion control.

Neuron operates solely with a memristor and a series resistor.

Abstract

Rapid progress of artificial neural network applications in recent years has led to the issue of an unprecedented energy consumption. It can be solved by the implementation of energy efficient hardware based on non-von-Neumann architectures, which requires the development of electronic components emulating the behavior of synapses and neurons. While research of synaptic elements is vast, the technology for fabrication of scalable and highly reproducible neuronal elements is far less developed. In this paper, we demonstrate an artificial neuron with multiple functionalities based on filamentary switching Ag/Hf$_{0.5}$Zr$_{0.5}$O$_2$ (HZO) memristors. To improve the parameters of memristors, we propose a two-step annealing method, which allows for better control of the crystallization of the functional dielectric layer (HZO) as well as of the diffusion of active electrode (Ag) atoms. Furthermore, we demonstrate the leaky integrate-and-fire (LIF) neuronal behavior in multiple spiking modes: time-to-first-spike (TTFS), number of spikes and firing rate coding. Moreover, the neuron operation does not require the additional electronic overhead and is supported solely by a Ag/HZO memristor with a current limiting resistor connected in series. The presented results pave the way for the creation of next generation energy efficient neuromorphic hardware operating on the principles of spiking neural networks.