The deep learning technology for predicting memory hole contact distortion profile with multiple in-line data in 3D NAND flash memory

A novel approach combining optical emission spectroscopy (OES) and information from charge trap flash (CTF) Oxide/Nitride/Oxide (ONO) deposition process with transformer-based deep learning techniques has been proposed to accurately predict the hole etching distortion profile of 3D NAND products. This methodology included extracting relevant features from OES data, reducing dimensionality, and performing signal processing to obtain useful insights regarding the hole profile. The ONO deposition process parameters which empirically impacts the final device cell performance were also considered. A transformer model was then utilized to effectively capture the complex interrelationships within the datasets, leading to improved prediction accuracy. In the given 8-th generation VNAND samples, the linear correlation coefficients (R2) of elemental hole distortion index prediction were successfully obtained as the best results were 0.69 for unseen wafer datasets and 0.58 for all unseen chip datasets, respectively.