심층신경망을 이용한 3D Vertical SONOS NAND Flash의 Grain Boundary Distribution과 Geometrical Variation에 의한 전기적 특성 변동 분석 (Deep Learning Approach for Electrical Characteristics Analysis of 3D Vertical SONOS NAND FLASH by Grain Boundary Distribution and Geometrical Variation)

Planar 형태의 NAND Flash memory의 Scaling down의 한계로 3D Vertical Silicon-Oxide-Nitride-Oxide-Silicon (SONOS) NAND가 개발되었다. 3D Vertical SONOS NAND로 한정된 크기의 wafer에서 많은 transistor를 적층하여 planar type보다 더 많은 memory capacity를 확보할 수 있다. 하지만 vertical 구조로 변경되면서 공정 난이도 상승과 함께 공정에 소모되는 비용이 증가했다. 그래서 공정에 투입되는 비용을 줄이고 소자의 전기적 특성을 빠르고 정확하게 예측하는 기술의 필요성이 대두되었다. 본 논��에서는 TCAD simulation과 딥러닝을 통해 3D Vertical SONOS NAND의 polysilicon grain boundary distribution (Max-angle, Ycut, Xseed, Yseed, Aseed)과 geometrical variation (Width, Lcg)에 따른 전기적 특성 (Vtgm, Vti) 변동을 예측하고 분석했다. TCAD simulation 결과값을 바탕으로 학습한 딥러닝을 통해 전기적 특성을 예측했고 매우 높은 수치의 R2 score (Vtgm R2 score = 0.997, Vti R2 score = 0.999)로 TCAD simulation 결과값에 수렴한다는 것을 알 수 있다. 또한 SHapley Additive exPlanations (SHAP) value를 통해 input parameter의 중요도를 평가한 결과 Ycut과 Xseed parameter가 전기적 특성 변동에 가장 많은 영향을 준 것을 확인했다.

3D Vertical Silicon-Oxide-Nitride-Oxide-Silicon (SONOS) NAND was developed as a solution to the scaling down of planar-type NAND Flash memory. With 3D Vertical SONOS NAND, it is possible to stack many transistors on a limited-size wafer to secure more memory capacity than the planar type. However, with the change to the vertical structure, the cost of the process has increased along with the increase in process difficulty. Therefore, there is a need for a technology that reduces the process's cost and predicts the device's electrical characteristics quickly and accurately. In this paper, we used TCAD simulation and deep learning to predict and analyze the variation of electrical characteristics (Vtgm and Vti) of 3D Vertical SONOS NAND due to polysilicon grain boundary distribution (Max-angle, Ycut, Xseed, Yseed, Aseed) and geometrical variation (Width, Lcg). The electrical characteristics were predicted using deep learning trained based on TCAD simulation results and converged to TCAD simulation results with very high R2 scores (Vtgm R2 score = 0.997, Vti R2 score = 0.999). We also evaluated the importance of the input parameters through the SHapley Additive exPlanations (SHAP) value. We found that the Ycut and Xseed had the most influence on the variation of electrical characteristics.