Wafer bin map inspection based on DenseNet  

基于密集连接网络的晶圆图检测

作　　者：YU Nai-gong XU Qiao WANG Hong-lu LIN Jia 于乃功;徐乔;王宏陆;林佳(Faculty of Information Technology,Beijing University of Technology,Beijing 100124,China;Beijing Key Laboratory of Computing Intelligence and Intelligent System,Beijing 100124,China;Engineering Research Center of Digital Community,Ministry of Education,Beijing 100124,China)

机构地区：[1]Faculty of Information Technology,Beijing University of Technology,Beijing 100124,China [2]Beijing Key Laboratory of Computing Intelligence and Intelligent System,Beijing 100124,China [3]Engineering Research Center of Digital Community,Ministry of Education,Beijing 100124,China

出　　处：《Journal of Central South University》2021年第8期2436-2450,共15页中南大学学报（英文版）

基　　金：Project(Z135060009002)supported by the Ministry of Industry and Information Technology of China;Project(KZ202010005004)supported by Beijing Municipal Commission of Education and Beijing Municipal Natural Science Foundation of China。

摘　　要：Wafer bin map(WBM)inspection is a critical approach for evaluating the semiconductor manufacturing process.An excellent inspection algorithm can improve the production efficiency and yield.This paper proposes a WBM defect pattern inspection strategy based on the DenseNet deep learning model,the structure and training loss function are improved according to the characteristics of the WBM.In addition,a constrained mean filtering algorithm is proposed to filter the noise grains.In model prediction,an entropy-based Monte Carlo dropout algorithm is employed to quantify the uncertainty of the model decision.The experimental results show that the recognition ability of the improved DenseNet is better than that of traditional algorithms in terms of typical WBM defect patterns.Analyzing the model uncertainty can not only effectively reduce the miss or false detection rate but also help to identify new patterns.晶圆图(WBM)检测是评估半导体生产工艺的关键手段,有效的检测方法能够提升生产效率与产品良率。本文提出了一种基于密集连接网络的晶圆图缺陷模式检测方法,并根据晶圆图特点对模型结构和损失函数进行了改进。此外,提出了一种受限均值滤波算法滤除噪声晶粒。在模型预测时,采用基于熵的蒙特卡洛Dropout算法来量化模型决策的不确定性。实验结果表明,对于典型的晶圆缺陷模式,改进模型的识别能力优于传统算法。通过分析模型不确定性,不仅可以有效地降低漏检率和误检率,还有助于发现新模式。

关 键 词：wafer defect inspection convolutional neural network DenseNet model uncertainty 

分 类 号：TN305[电子电信—物理电子学] TP391.41[自动化与计算机技术—计算机应用技术]