Enabling High-Accuracy Neuromorphic Computing via Precise Synaptic Weight Tuning in HfOx-Based 3D Vertical Memristors

Abstract

This study presents a multibit implementation strategy using a vertically stacked resistive random-access memory (VRRAM) that uses an HfO_x-based switching layer. The proposed VRRAM device operates via filamentary switching; however, by selectively forming and removing portions of the filament, it effectively mitigates the inherent issues of dispersion and nonlinearity typically associated with filament-based mechanisms. Furthermore, using an incremental step pulse with verify algorithm (ISPVA) measurement method where the device is allowed to reach a predetermined current level before transitioning to the subsequent target further enhances both the linearity and reduces the dispersion of the filamentary memory cell. In addition, the device demonstrates outstanding performance on modified national institute of standards and technology (MNIST) and fashion MNIST datasets, achieving accuracies of 96.65% and 76.50%, respectively, thereby surpassing current state of the art hardware-based implementations. These results collectively advance the scalability and practical feasibility of next-generation neuromorphic computing systems.