Abstract
A spatial pattern recognition algorithm is proposed for the effective sampling method to characterize defect occurrence of an entire wafer by reviewing a small number of defect samples. The algorithm classifies defects into clustered defects or regional defects or random defects based on analysis of defect densities and distances between neighboring points. Voronoi diagrams are applied to calculate the densities and the distances. Regional defects are classified into four major classes: rings, blobs, lines and arcs. Ring and blob patterns are detected by template matching techniques, while line and arc patterns are detected by utilizing their geometric properties. The Hough transform is used to detect line patterns and the detected pattern is verified by analyzing features. The proposed algorithm was evaluated using 916 sample wafers obtained in a real semiconductor process. 93.3% of the sample wafers with ring or blob patterns (250/268) were processed correctly while 3.9% of the sample wafers on which ring or blob patterns are detected (13/332) were false. 95.1% of the sample wafers with line patterns (117/123) were processed correctly while 18.7% of the sample wafers on which line patterns are detected (70/375) were false.
