학술논문
심층신경망을 이용한 Total Ionizing Dose 및 Displacement Defect에 의한 Saddle Fin DRAM의 열화 특성 예측
Prediction of Degradation Characteristics in Saddle Fin DRAM Due to Total Ionizing Dose and Displacement Defect using Deep Neural Network
Prediction of Degradation Characteristics in Saddle Fin DRAM Due to Total Ionizing Dose and Displacement Defect using Deep Neural Network
Document Type
Article
Source
전자공학회논문지, 60(11), pp.29-37 Nov, 2023
Subject
Language
한국어
ISSN
2288-159X
2287-5026
2287-5026
Abstract
본 논문에서는 saddle fin dynamic random access memory (DRAM) 에 대한 total ionizing dose (TID) 와 displacement defect (DD) 영향을 Technology Computer-Aided Design (TCAD) simulation과 deep neural network (DNN) 를 사용해 조사하였다. Trap의 energy level, 농도, 위치 그리고 면적을 변수로 설정하였고, TCAD를 사용하여 saddle fin DRAM의 전류-전압 특성 dataset을 생성하였다. TCAD dataset을 전처리 과정을 거친 경우와 전처리를 하지 않은 경우로 나누어 DNN의 예측 정확도를 비교하였다. 그 결과 전처리 과정을 거쳐 훈련된 모델은 전처리 과정을 하지 않은 훈련 모델보다 mean square error (MSE) loss가 80 % 증가함과 동시에 R2 score가 37 % 증가하였다. 따라서 DNN을 활용한 정확한 예측을 위해서는 전처리 과정이 필수적이다.
In this paper, the effects of total ionizing dose (TID) and displacement defect (DD) in saddle fin dynamic random access memory (DRAM) are investigated using technology computer-aided design (TCAD) simulation and deep neural network (DNN). TCAD is used for generating the current-voltage characteristic data of the saddle fin DRAM and the energy level, concentration, location, and area of the trap are utilized for variables. The TCAD dataset is divided into preprocessed and un-processed cases to compare the prediction accuracy of DNN. The result shows that the model trained with preprocessing has an 80 % increase in mean square error (MSE) loss and a 37 % increase in R2 score compared to the training model without preprocessing. Therefore, preprocessing of a dataset is necessary for high prediction accuracy using DNN.
In this paper, the effects of total ionizing dose (TID) and displacement defect (DD) in saddle fin dynamic random access memory (DRAM) are investigated using technology computer-aided design (TCAD) simulation and deep neural network (DNN). TCAD is used for generating the current-voltage characteristic data of the saddle fin DRAM and the energy level, concentration, location, and area of the trap are utilized for variables. The TCAD dataset is divided into preprocessed and un-processed cases to compare the prediction accuracy of DNN. The result shows that the model trained with preprocessing has an 80 % increase in mean square error (MSE) loss and a 37 % increase in R2 score compared to the training model without preprocessing. Therefore, preprocessing of a dataset is necessary for high prediction accuracy using DNN.