Abstract: To recognize the wafer defects of semiconductor wafer manufacturing process for finding the root causes of out-of-control process,a detecting system for wafer defects recognition was developed.The hierarchical clustering method was applied to group the local defects into clusters,and the number of clusters was determined by silhouette coefficient.To best describe three common defect patterns of wafer (linear,curvilinear,ellipse),Gaussian Mixture Model with Local Consistency(LCGMM) was proposed with respect to the data's distribution on a submanifold of ambient space.Combined with Principal Curve (PC) model,the local defect patterns could be well modeled.On this basis,a hybrid model by integrating LCGMM and PC was constructed,which not only detected more wafer defect patterns but increased the accuracy of recognition rate.The proposed system was applied to real-world and simulate cases,and the effectiveness was proved.

Key words: semiconductor manufacturing, wafer defect, manifold regulation, hybrid models, pattern recognition, fault diagnosis

摘要： 为了探测和识别半导体晶圆生产线上的晶圆表面缺陷,及时诊断出半导体晶圆制造过程的故障源,提出一套晶圆表面缺陷检测与识别系统。该系统首先采用层次聚类法将晶圆表面的局部缺陷划分为缺陷簇,并提出一种基于轮廓系数标准的最优缺陷簇数目判定方法,提升了缺陷簇识别性能。针对晶圆表面常见的线形、曲线形和椭球形缺陷模式,该系统充分考虑数据在空间子流形上的分布,采用基于流形调节的局部连续高斯模型(LCGMM),同时加入主曲线模型,实现了对晶圆表面局部缺陷模式分布的统计描述建模。在完成初始建模识别的基础上,进一步提出集成LCGMM和主曲线模型的混合模型,对晶圆表面所有的缺陷模式进行建模识别,以提高缺陷模式识别的准确性。通过仿真案例和工业案例的实验结果,证明了该系统的有效性与实用性。

关键词: 半导体制造, 晶圆缺陷, 流形调节, 混合模型, 模式识别, 故障诊断