Modeling of random telegraph noise decay constant in 3D NAND flash memory using machine learning-based Monte Carlo simulation

Abstract

Random telegraph noise (RTN) has emerged as a major reliability concern in high-density 3D NAND flash memory, causing significant threshold voltage (V_t) distribution distortion. While technology computer-aided design (TCAD) simulations provide physical insights, they face fundamental limitations in capturing the statistical randomness of grain boundaries and traps due to high computational costs. In this study, we design the RTN decay constant (λ) model in 3D NAND using artificial neural networks (ANN) and Monte Carlo simulation. Our analysis reveals that unlike 2D NAND or single-Si 3D NAND, poly-Si channel 3D NAND exhibits average trap number dependent (μ_trap) scaling behavior due to percolation effects induced by GBs. Specifically, we find that the scaling sensitivity of cell variations directly influencing channel volume, such as word-line length (L_WL) and poly-Si channel thickness (T_poly), decreases as the μ_trap increases. Furthermore, we demonstrate that λ in 3D NAND is proportional to the cubic root of μ_trap (λ \(\alpha \sqrt[3]{{\mu}_{\text{t}\text{r}\text{a}\text{p}}}\)), a deviation from the square-root dependency observed in 2D NAND. These results provide a robust numerical model and essential design guidelines for predicting the reliability of 3D NAND flash memory devices.