Read-reference voltage adaptation for 3D NAND flash memory via RBER-aware few-shot transfer learning

Program/erase (P/E) cycling and retention drift alter the threshold-voltage distributions of 3D NAND flash, causing previously calibrated read-reference voltages to become mismatched to new operating conditions. To update read references when target measurements are limited, we propose a raw bit error rate (RBER)-aware few-shot transfer framework that adapts a Gaussian-parameter predictor from a single characterized source condition without requiring a pre-calibrated multi-condition voltage grid. The predictor is pretrained on source measurements and fine-tuned using a small set of target records, with mean squared error (MSE) as the parameter-training objective. Candidate checkpoints are selected using a restricted RBER metric over the measured programmed-state boundaries, aligning model selection with read-reference quality. Experiments on measurements collected from 3D NAND flash cover six target conditions and five random seeds, accounting for both training and validation measurements. With a total budget of two target wordlines, the proposed framework achieves an average reduction of 83.0% in the evaluated RBER relative to the evaluated training-from-scratch configuration (95% confidence interval 67.2–98.8%). In a paired experiment using 32 target training records and 13 validation wordlines, RBER-based checkpoint selection reduces RBER by 42.1% compared with MSE-based selection on the same fine-tuning trajectories (95% confidence interval 34.7–49.5%). Target-wordline interpolation achieves lower mean restricted RBER than the proposed framework in all 24 matched-budget comparisons. The checkpoint-selection benefit therefore concerns improvement within the learned framework, while direct interpolation provides higher accuracy under the tested budgets. By combining source-model reuse with RBER-based checkpoint selection, the proposed framework enables read-reference voltages to be adapted to changing P/E and retention conditions using limited target measurements.