Intelligent Read Framework With Meta-Learning for NAND Flash Memory Under Process Variation

The storage density of NAND flash memory has significantly increased due to multi-leveling and downscaling technologies. As a side effect of enhanced storage capacity, flash memory becomes vulnerable to circuit-level noise. To enhance data reliability, NAND flash memory employs error correction codes such as low-density parity-check (LDPC) code with soft-decision (SD) decoding based on log-likelihood ratio (LLR) information. Conventional SD read uses pre-defined uniform quantization, which is not optimized for varying threshold voltage distributions caused by physical noise, degradation, and process variation (PV). We propose an intelligent read framework employing a non-uniform quantizer that adjusts read reference voltages (RRVs) and LLR values according to the memory channel state. The optimization of RRVs and LLRs requires the information of threshold voltage distributions. To accurately estimate the threshold voltage distribution from limited channel information, we propose a machine learning-based estimation model. Moreover, we incorporate model-agnostic meta-learning, enabling precise estimation of threshold voltage distribution across different flash chips addressing PV. Numerical results demonstrate significant improvements in read performance, validating the effectiveness of the proposed framework.