Dynamically nonlinear NbOx memristors for multifunctional reservoir and neuromorphic Computing

Abstract

Neuromorphic computing systems enable efficient and low-power information processing by emulating the structure and function of the human brain. Among various architectures, reservoir computing (RC) and spiking neural networks (SNNs) stand out as two promising paradigms for dynamic information processing. Yet, due to the differing performance requirements of the two systems, realizing and integrating both within a single memristor remains a significant challenge. Here, we pioneers a multifunctional TiN/NbOx/Pt memristor fabricated via a rapid (40 s), low-energy anodic oxidation process. The device exhibits excellent stability, strong nonlinearity, over 104 cycles of pulse programmability, and a high self-rectifying ratio of 2.5 × 104. Owing to its dynamic characteristics, the memristor emulates diverse synaptic plasticity behaviors and, when coupled with external circuits, implements a nonlinear leaky integrate-and-fire (LIF) neuron featuring dendritic signal filtering and nonlinear firing dynamics. Furthermore, a time-delay RC system based on the same device accurately predicts the chaotic Mackey–Glass time series, achieving a normalized root-mean-square error below 0.2. By integrating this RC system with a nonlinear SNN, we demonstrate efficient MNIST handwritten digit recognition, which reduces data complexity by ~75% and spike firing rate by ~29%. This work demonstrates a single-device platform for multifunctional neuromorphic computation, providing new pathways for designing energy-efficient and biologically realistic neuromorphic hardware.