Cluster-Type Conductive Path-Based Selector-Less 1r Memristor Array for Spiking Neural Networks

Abstract

Memristors hold great promise as next-generation devices, but their practical application faces challenges such as achieving low power consumption, multi-level resistance states, and efficient crossbar array construction. The switching characteristics and performance of memristors depend largely on the mobile species and the matrix through which they move, yet controlling ion dynamics remains difficult. In this study, we employed ruthenium (Ru) as the active electrode and utilized a SiO2 matrix in a nanorod structure, which reduces the activation energy for Ru ion diffusion and enhances redox reactions. Precise control of Ru ion dynamics enabled us to develop novel conduction paths and mechanisms. The Pt/SiO2 nanorods/Ru structure exhibits improved switching characteristics, including electroforming-free operation, low power consumption, highly linear conductance modulation, and inherent nonlinearity in the on-state. To demonstrate operational potential in large-scale crossbar arrays, we introduced a novel Spiking Neural Network (SNN) simulator that incorporates both device-level switching behaviors and key array-level parameters such as line resistance and sneak currents. Using this simulator, we successfully implemented a 16 × 16 selector-less crossbar array, achieving 80% accuracy on the MNIST dataset.

Keywords: memristor, ion dynamic modulation, non-linearity in On-state, 1R-array, array-based spiking neural networks