Endurance prediction and error Reduction in NAND flash using machine learning

NAND flash is rapidly becoming the media of choice for data storage, due in part to its speed and low power consumption. However, flash wears out through repeated program-erase (P-E) cycling, causing the raw bit error rate (RBER) to increase. Error correction codes (ECCs) are used to detect and correct errors in a sector of data called a codeword. An uncorrectable error occurs when the number of bit errors in a codeword exceeds a certain level, meaning the data cannot be recovered. This research uses machine learning to predict how far each sector of a NAND device can be cycled before it becomes uncorrectable by ECC, thereby enabling the controller to better manage the flash. We characterise a number of flash metrics to investigate their suitability for use as inputs to the prediction model, and determine the impact of each metric on the accuracy of the model. We show for the first time that it is possible to reliably make such predictions, achieving an overall accuracy of greater than 99% when the model was tested on unseen data.