Low-power and non-volatile multi-level synaptic weight update characteristics in self-selecting Nb2O5/CeO2 memristor crossbar array for in-memory computing system

Abstract

Non-filamentary valence change memory (VCM)-type memristors are considered the promising candidates as artificial synapses for in-memory computing systems to achieve high energy-efficient computing due to their analog or multi-level conductance change for training operations. However, their poor retention properties originating from unintended diffusion of redistributed oxygen ions limit their application only to the training operation, because the inference operation requires non-volatile retention of updated weights. In this study, nonvolatile retention with multi-level synaptic weight update characteristics is demonstrated in non-filamentary bilayered memristor with cerium oxide (CeO2) and niobium oxide (Nb2O5), i.e., Nb2O5/CeO2, where CeO2 acts as switching layer and Nb2O5 serves as an oxygen ion-holding layer. Due to their high oxygen-ion conductivity and active oxygen-ion exchange property, the memristor enables polarity-dependent, linear, and symmetric analog conductance changes as weight updates, while operating stably at low power with a current range of 0.5-500 nA at + 1.5 V or below. Notably, the device exhibits long-term retention of updated weights enabling discrimination of multi-level states via stably holding oxygen ions in Nb2O5 layer. In addition, by employing Nb2O5/CeO2 bilayered structure, it achieves self-selecting characteristics from asymmetric Schottky contacts at oxide/electrode interfaces as well as high non-linearity ratio in half-bias operation scheme, which minimizes write disturbance and sneak current in selector-free crossbar array architecture. Using the obtained weight update characteristics, pattern recognition accuracy is simulated to be 96.6 % for MNIST handwritten patterns using CrossSim program. These linear, symmetric, and analog non-volatile conductance change with low power consumption as well as self-selecting behaviors of bi-layered Nb2O5/CeO2 memristor crossbar array confirms the potential of the proposed device to be applicable to integrated in-memory computing systems that implement both efficient training and inference operations.