A physics-guided surrogate modeling framework for RDF-dominated variability and bit-error rate analysis of 3D NAND flash memory

Abstract Three-dimensional (3D) NAND flash memory has become the dominant non-volatile storage technology due to its high density and scalability. However, aggressive vertical stacking introduces significant threshold-voltage (V T ) variability that affects device reliability and bit-error rate (BER). In this work, a physics-guided surrogate modeling framework is proposed to analyze variability mechanisms in vertical channel 3D NAND flash memory. A dataset generated from Synopsys Sentaurus TCAD simulations is used to train machine learning models including Random Forest, Gaussian Process Regression, and a heteroscedastic neural network capable of predicting both the mean and variance of the V T distribution. Physics-derived features such as active channel volume and dopant population are incorporated into the learning process to improve model interpretability. The proposed framework accurately reproduces variability trends with high predictive accuracy (R2 > 0.95, RMSE < 0.08 V) and enables rapid exploration of process parameters including taper angle and channel doping concentration. The results show that larger taper angles increase V T variability due to electric-field inhomogeneity, while increased channel doping statistically reduces random dopant fluctuation effects. This physics-guided surrogate framework provides a computationally efficient tool for variability-aware design exploration within RDF-dominated conditions, offering practical process guidelines for taper and doping optimization in vertical channel 3D NAND devices. The scope of the analysis is limited to RDF-induced threshold-voltage variability and does not encompass full-system reliability mechanisms such as retention, endurance cycling, or ECC interaction.