Research On Spatial Signature Analysis Of Wafer Defects

With the development of Microelectronics technique, increased wafer dimension and decreasing line width are creating huge quantities of defects from the fab environment. As the fabrication processes increase in complexity, there is a great probability of defects on the wafer. Billions of dollars are invested in equipments for wafer fabrication and a rapid return on investment is crucial for a semiconductor manufacturer. To maintain certain yield levels, amount of time must be spent for finding and correcting manufacturing problems. It has been shown that a defect in processes is one of key causes which make yield decreases. Flexible defects analysis system is a critical component in a proactive integrated yield management system. This dissertation concerns on some basic problems in wafer spatial signature analysis. Generally, there are three kinds of plot in defects detection: Wafermap, Binmap, and Bitmap. This dissertation is focused on the three issues of wafermap analysis as follows:Firstly, basic method of defects pattern analysis is studied, and some issues in defects spatial signature are discussed. After essential theories and algorithms in spectral clustering are presented, a new method of noise cancellation of linear defects is proposed by using this clustering algorithm. The simulation and analysis in two samples show that new method can remove the noise of defects pattern effectively, which forms a basis to identify linear defects pattern in next step.Secondly, an improved method to compute the local outlier factor is introduced. The mean k-nearest neighbor (k-NN) distance for each current defect in wafermap is calculated by detecting k-NN defects of each defect, then the numbers of defects within this mean k-NN distance are recorded, and the ratio of mean number of defects to number of current defect is used as a k-NN outlier factor. A defects clustering method based on k-NN outlier factor is proposed, and some experiment results show that new algorithm makes the defects spatial clustering more effectively.Thirdly, the principal curve and polygonal line algorithm are introduced, and an implementation of polygonal line algorithm based on Quasi-Newton numerical optimization is illustrated. The simulation results show that when this algorithm isused in the detection of curvilinear defects on wafer map, it can converge fast, compute accurately and meet the demand of application of defects pattern analysis, and with the number of line segment increasing, the polygonal line will approximate to the Kegl principal curve.