Classification of Semiconductor Defects Using a Small Number of Training Data and Qualitative Knowledge

摘要

In semiconductor wafer manufacturing processes, defect candidates are usually extracted by an inspection system. The defect candidates are composed of true defects such as open circuits, contaminants, and bridges, as well as nondefect patterns, called nuisances, which predominate over true defects. The goal of this study is to classify the defect candidates as the various true defects and nuisances by using a small number of training data obtained by SEM inspection. It is shown that the accuracy of clustering is considerably improved by use of qualitative knowledge about the defects, given a priori by inspectors, in the clustering processes.