Niobium Oxide-Based Threshold Switching and Dynamic Memristors for Spiking Neural Network

Inspired by the human nervous system, neuromorphic computing based on spiking neural networks (SNNs) can offer numerous advantages including parallel processing, adaptive learning, energy efficiency, high fault tolerance, and high plasticity. However, implementing neuromorphic computing on traditional CMOS requires a substantially complex and extensive assemblage of components. The presence of the volatile threshold switching (TS) memristors simplifies the hardware structure of neuromorphic computing system, making hardware implementation more straightforward.<br/><br/>Niobium oxide (NbO_x) is a typical Mott insulator characterizing by its temperature-triggered insulator-metal transition. NbO_x-based memristors exhibit S-type negative differential resistance (NDR) during current-controlled I-V sweep measurement and show the volatile threshold switching (TS) characteristics when conducted voltage-controlled sweeping. Connected in parallel with a capacitor, TS devices can follow a Pearson-Anson-like relaxation oscillator circuit model, serving as a prototype for the Leaky Integrate-and-Fire (LIF) model. Due to the MIM (Metal-Insulator-Metal) structure of the TS device, the parasitic capacitance inherent in these components also simplifies the hardware implementation of the LIF model. When pulse strain is being applying to the NbOx-based LIF circuit, it generates spiking signals of various frequencies when pulse voltage and input frequency exceed a certain value (spiking from 4kHz to 15kHz depending on the input conditions). However, when the voltage level or input frequency falls below a specific threshold, no spike signals are generated, demonstrating an all-or-nothing characteristic.<br/><br/>With varying the thickness of the niobium oxide active layer, NbOx-based devices can exhibit dynamic resistive switching (DS) as well. DS devices possess short-term memory characteristics that depend on operational history and duration, leading to dynamic resistance changes. After electrical stimulation with a 5V amplitude, 3ms pulse width square pulse, the DS device current value measured with 2V square pulse increased around 20 times larger than value measured before stimulation (from 10μA to 209μA). Applied the gradual increasing electrical wave from 0V to 3V, current level at 1.5V is 1.6μA and 3V is 590μA. This nonlinear growth in current value imparts DS devices with filtering capabilities.<br/><br/>DS devices also exhibit diode-like characteristics, enhancing their functionality within circuits. The combination of TS devices with DS devices allows for the realization of more complex LIF models, enriching the computational capabilities of SNNs, whether for controllable leakage behavior, gradual potential rise buffering, noise filtering or spatial summation characteristic. Both DS and TS devices utilize niobium oxide as active layer’s material offering advantages in manufacturing, make the hardware implementation of neural networks more streamlined.