Simulation Acceleration of Bit Error Rate Prediction and Yield Optimization of 3D V-NAND Flash Memory

When designing 3D V-NAND technologies with a gate induced drain leakage (GIDL) assisted erase scheme, many experiments must be conducted to determine the optimal GIDL design targets to achieve fast erase performance and secure yield characteristics. However, only a limited amount of data can be used since V-NAND processes are time-consuming and expensive in the early stage of development. TCAD and numerical methods also require a considerable amount of time and effort to calculate bit error rate (BER), and it is impossible to explore the entire design spaces in time. In this paper, we propose a novel simulation acceleration technique for bit error rate prediction and yield optimization in 3D V-NAND technology. This acceleration framework includes a machine learning (ML)-based compact model for the lognormal variability of GIDL currents and a physics-inspired slow cell model for the read margin reduction. Using a combination of these models with efficient Monte Carlo (MC) circuit simulations, we can accurately estimate threshold voltage (
$V_{th}$
) distributions to explore the entire design spaces using a limited amount of data. Based on the proposed technique, the predictive model achieves high accuracy in the current 176-layer V-NAND technology, and it also provides high scalability with respect to GIDL transistor geometries, temperatures, supply voltages, variabilities, and the number of stacking layers. Moreover, a contour map of bit error rate is newly introduced for the efficient design space exploration and read margin prediction. Therefore, the results indicate that the proposed framework can be further extended to large-scale experimental data and new architectures to accelerate the yield optimization in next-generation 3D V-NAND flash memory development.