Application of spatial signature analysis to electrical test data: Validation study

This paper presents the results of the Spatial Signature Analysis (SSA) Electrical-test (e-test) validation study that was conducted between February and June, 1998. SSA is an automated procedure developed by researchers at the Oak Ridge National Laboratory to address the issue of intelligent data reduction while providing feedback on current manufacturing processes. SSA was initially developed to automate the analysis of optical defect data. Optical defects can form groups, or clusters, which may have a distinct shape. These patterns can reveal information about the manufacturing process. Optical defect SSA uses image processing algorithms and a classifier system to interpret and identify these patterns, or `signatures'. SSA has been extended to analyze and interpret electrical test data. The algorithms used for optical defect SSA have been adapted and applied to e-test binmaps. An image of the binmap is created, and features such as geometric and invariant moments are extracted and presented to a pair-wise, fuzzy, k-NN classifier. The classifier itself was prepared by manually training, which consists of storing example signatures of interest in a library, then executing an automated process which treats the examples as prototype signatures. The training process includes a procedure for automatically determining which features are most relevant to each class. The evaluation was performed by installing the SSA software as a batch process at three SEMATECH member company sites. Feedback from member company representatives was incorporated and classifiers were built to automatically assign labels to the binmap signatures. The three sites produced memory devices (DRAM) and microprocessors in a mature process fabrication environment. For all of these products, 5,620 signatures that encompassed approximately 552 wafers were human-classified and analyzed. The performance of the SSA E-test system indicates that the approach was successful in reliably classifying binmap signatures in a manner similar to the human expert.