A long short-term memory neural network-based approach for retention lifespan prediction in 3D NAND flash memory

Abstract A Long Short-Term Memory (LSTM) neural-network framework is presented to predict the temporal evolution of threshold-voltage ( V t ) distributions and estimate retention-induced lifespan in 3D NAND Flash. Rather than modeling individual mechanisms, the model learns percentile-wise V t trajectories from large-scale simulation data. The 3% V t of PV7—effective in determining page-level lifespan—is predicted up to four steps ahead with high R2 and negligible deviation from ground truth. Full-distribution prediction is obtained by grouping percentiles in 5% increments, enabling construction of the complete CDF. Lifespan is then computed for arbitrary failure-defining percentiles and reference voltages through error-rate evaluation. Page-level lifetime distributions from the model reproduce the simulation mean and standard deviation with high fidelity. Although accuracy modestly degrades with increasing prediction horizon and at distribution tails due to recursive rollout, performance is sufficient for practical reliability assessment.