Identification of the Nonrandom Defects in Volume Production Wafer Maps

In a recent International Technology Roadmap for Semiconductors (ITRS), the Test and Automatic Test Equipment reports that one of the difficulties of the challenge is to detect systemic defects. In Yield Enhancement, this report also pointed out that the identification of Non-Visual Defects and Process Variations was set to the most important key challenge in the future. The project "Identification of the Nonrandom Defects in Volume Production Wafer Maps" aims to determine whether there is a non-random defect pattern from the wafer map analysis, including over-clustering and anti-clustering. The number of defects is used to represent the random nature, and then from the mass production wafer maps, the use of single-wafer segmentation, block size adjustment and multi-wafer overlap to resolve more defect Pattern. What is the pattern of random defects? According to the method of statistic hypothesis test, the defect pattern of the complete spatially random defect wafer map is assumed to be null hypothesis, and the defect pattern of the wafer map to be inspected is compared. The comparison is deemed statistically significant if the relationship between the data sets would be an unlikely realization of the null hypothesis according to a threshold probability—the significance level. Therefore, to determine whether there is a non-random defect pattern (the alternative hypothesis), we must start from exploring the defect pattern of complete space random defect wafer map.The project consists of three topics: (a) “Deriving a robust nonrandom test”, I have to formulate the test model to express directly with wafer yield, die size and confidence level. (B) “Finding a representative”, wafer yield is a good choice, but it is not fully able to express the homogeneous defect distribution. A new number of features is formed. When the defect pattern of the wafer map is random, the number of features will approximate equal the average number of defects - the number of defects on each die. The smaller the number shows an over-clustering phenomenon, and larger that there is an anti-clustering phenomenon. And (c) “Data mining”, from the production of wafer maps, resolve a more diverse defect patterns and the combined effect of equivalent defects (non-visual defects and process variations).