基于局部与非局部线性判别分析和高斯混合模型动态集成的晶圆表面缺陷探测与识别  

Wafer Defect Detection and Recognition Based on Local and Nonlocal Linear Discriminant Analysis and Dynamic Ensemble of Gaussian Mixture Models

作　　者：余建波[1] 卢笑蕾 宗卫周 

机构地区：[1]同济大学机械与能源工程学院,上海201800

出　　处：《自动化学报》2016年第1期47-59,共13页Acta Automatica Sinica

基　　金：国家自然科学基金项目(51375290;71001060);上海市教育委员会科研创新项目(13YZ002);中央高校基本科研业务费专项资金资助~~

摘　　要：在复杂的半导体制造过程中,晶圆生产经过薄膜沉积、蚀刻、抛光等多项复杂的工序,制造过程中的异常波动都可能导致晶圆缺陷产生.晶圆表面的缺陷模式通常反映了半导体制造过程的各种异常问题,生产线上通过探测和识别晶圆表面缺陷,可及时判断制造过程故障源并进行在线调整,降低晶圆成品率损失.本文提出了基于一种流形学习算法与高斯混合模型动态集成的晶圆表面缺陷在线探测与识别模型.首先该模型开发了一种新型流形学习算法—局部与非局部线性判别分析法(Local and nonlocal linear discriminant analysis,LNLDA),通过融合数据局部/非局部信息以及局部/非局部惩罚信息,有效地提取高维晶圆特征数据的内在流形结构信息,以最大化数据不同簇样本的低维映射距离,保持特征数据中相同簇的低维几何结构.针对线上晶圆缺陷产生的随机性和复杂性,该模型对每种晶圆缺陷模式构建相应的高斯混合模型(Gaussian mixture model,GMM),提出了基于高斯混合模型动态集成的晶圆缺陷在线探测与识别方法.本文提出的模型成功地应用到实际半导体制造过程的晶圆表面缺陷在线探测与识别,在WM-811K晶圆数据库的实验结果验证了该模型的有效性与实用性.The manufacturing process of wafers is very complicated, which involves thin film deposition, chemical mechanical polishing, etching, etc. Abnormal variation in the manufacturing processes could result in the generation of defect patterns on wafers. On the other hand, the defect patterns on the wafer usually indicate the abnormal sources existing in the manufacturing process. Therefore, wafer defect recognition plays a key role in finding the root causes of the outof-control process and improving yield. This paper develops a model for wafer defect detection and recognition based on manifold learning and dynamic ensemble of Gaussian mixture models（GMMs）. A new manifold learning algorithm called local and nonlocal linear discriminant analysis（LNLDA） is developed, which aims to preserve the geometric structure of samples that belong to the identical cluster in the low dimensional space while maximizing the distance of the projected samples from different clusters. Due to the randomness and complexity of wafer defect generation, a dynamic ensemble scheme for GMMs is proposed to detect and recognize various defect patterns occurring in wafers. The proposed model is applied to a real-world semiconductor manufacturing process. Experimental results on WM-811 K wafer bin map database approve the effectiveness of the proposed model.

关 键 词：半导体制造 晶圆缺陷 模式识别 流形学习 高斯混合模型 

分 类 号：TN305[电子电信—物理电子学]