On the Accurate and Robust Prediction of Optimal Read Voltages in 3D NAND via Data Augmentation

With the exponential growth of data, the demand for storage capacity is constantly increasing, leading to the widespread adoption of 3D NAND flash memory in enterprise storage and data centers. However, 3D NAND flash memory faces reliability issues, necessitating the accurate calibration of read voltages to reduce the raw bit errors of read data. For a new chip model, extensive testing is typically required to obtain the optimal reading voltage (ORV) under various disturbance conditions, such as program/erase (P/E) cycles, to adapt to complex real-world scenarios. To reduce such costs, existing research has proposed applying machine learning models to predict the ORV with certain testing data or to fit the changing trend of ORV. In this paper, we first conduct a study on the prediction performance of these models, which shows that they all fall short in accuracy when the training data is limited. To deal with the challenge of limited test data, we propose to accurately and robustly predict the ORV of high-density 3D NAND flash memory through data augmentation in data-scarce scenarios. Specifically, with limited test data, we employ fitting techniques to estimate the ORV under P/E cycles that are not tested for data collection. With the test data and augmented data, we utilize machine learning models, including random forest (RF), support vector machine (SVM), and neural network (NN), for training and prediction. Experimental results on real-chip test data demonstrate that the proposed method significantly improves the prediction performance, where the predicted ORV is very close to the ideal ORV.