An Analytical Model of RRAM Relaxation Effect and Its Application for Neural Network Weight Refresh Strategy in Large-Scale RRAM Array

In this article, an analytical model for the retention behaviors of analog resistive random access memory (RRAM) is proposed. The model accounts for the diffusion of oxygen vacancies ( ${V}_{O}$ ), the recombination of ${V}_{O}$ , and the impact of programming pulsewidth on the number of metastable oxygen vacancies. It enables the analysis of the conductivity drift characteristics of RRAM under various resistance states, temperatures, and programming pulse widths. The model is in good agreement with our experimental results of analog RRAM arrays with high/low ${V}_{O}$ diffusion coefficients, confirming the accuracy and practicability of the model. Additionally, the model is integrated into a fully connected RRAM-based neural network to evaluate the reliability of the network. Furthermore, this article introduces a novel weight refresh strategy based on the accurate retention time (ART), defined as the period during which neural network accuracy degrades slowly, to balance the trade-off between neural network performance and power consumption. The prediction scheme of ART employs a two-stage machine learning framework. The predicted results on the neural network demonstrate that the strategy maintains high accuracy ( $\le 2$ % degradation) while minimizing refresh frequency. This work bridges physical mechanisms with neural network optimization, offering a scalable, low-power consumption solution for computation-in-memory (CIM) systems.