Abstract
Because of the complex fabrication processes and the cost resulted from defects, analyzing wafer bin maps and mining potentially useful information from large such database become increasingly important in both research and application for semiconductor manufacturing. We develop a hybrid approach that combines spatial statistics and neural networks for wafer bin map (WBM) clustering to assist the product engineers in narrowing possible causes of manufacturing defects. We conducted an empirical study in a semiconductor fab and the results showed the practical viability of this approach.This study proposed a method to train the algorithm and estimate the validity. In particular, we have developed a hybrid approach that integrates the spatial statistics test, cellular neural network (CNN), adaptive resonance theory (ART) neural network and moment invariant (MI) concept for WBM clustering. The proposed method can cluster a large number of WBMs into a number of groups that show specific failure patterns and thus shorten the time and scope for troubleshooting to improve yield. However, ART is an unsupervised neural network without given target value for comparison. Thus, it is difficult to validate the results of WBM clustering. In particular, we design an experiment to train the algorithm. In this study, we designed an experiment with the generated WBMs based on domain knowledge and thus their true categories are available. We verify this approach with different configurations of ART by examining four indexes of clustering results. Then, we thus select optimal parameter configuration of ART for clustering the WBMs in real setting.