Optimizing automatic defect classification feature and classifier performance for post-fab yield analysis

In this paper we present a methodology for enhanced automatic defect classification (ADC) of defects optically detected during post fab inspection and present results from production wafers. We have developed a unique approach to statistical feature calculation that enables the selection of four possible input intensity bands (gray, edge, hue, saturation), three image types (defect, reference, difference), and three defect masks (interior, edge, surround). To achieve the greatest separation between defect classes the optimum subset of features for a given training set must be determined. We propose an approach for feature ranking based on a feature evaluation index (FEI). The final step in optimizing the ADC performance is the selection and training of a pattern classification algorithm. We have evaluated three non-parametric, supervised classifiers including the k-nearest neighbor (KNN), fuzzy KNN, and radial basis function (RBF). The described approach is applied to several sets of defects detected with Electroglas' QuickSilver^TM post-fab inspection system. The test results from this library were very good with the optimum accuracy of 84% (on testing set that was not seen during training). This level of performance was also seen in several other smaller libraries used during the development of the underlying algorithms. We believe that this approach of mask based descriptive feature calculation, feature ranking and non-parametric classifiers will enable reliable ADC in the post-fab environment. This post-fab ADC approach is complementary to in-line ADC and enables a more complete yield analysis process.