<title>Comprehensive methodology for integrated circuit in-line defect classification</title>

Abstract

The earliest attempts by human inspectors to classify defects found during in-line inspection of integrated circuits were fraught with difficulties in clarifying defect definitions and in training a diverse and changing inspector staff. These deficiencies were exacerbated by the challenges of expanding classification categories as new defects were discovered. Our diversified product mix had accumulated a knowledge base of approximately seventy defect types, posing a formidable learning challenge for even the most knowledgeable inspector. Not surprisingly, the average accuracy of the group in classifying defects was -55%, and even the best inspector scored around 70%. To address these issues, we developed a comprehensive methodology for classifying defects. This methodology includes both word descriptions of the physical appearance of defects and a hierachical questionnaire leading to precise defect classification. After adopting this methodology and implementing strong training programs, our team significantly improved its defect review process, ultimately reaching 8O% classification accuracy. With this degree of accuracy, we were able to implement defect specific statistical process control (SPC) charts, together with formalized "decision tree" procedures for correcting defect excursions. These formalisms then became an effective part of the fab's yield improvement program. Today, as technology advances into the realm of automatic defect classification (ADC), the lessons learned from human defect inspection form a strong foundation by establishing a comprehensive set of defect categories uniquely related to causality and supporting defect identification standards that can be used by the entire community ofADC training engineers.

Automatic classification of microlithography macrodefects using a knowledge-based system

“…However, human classifi cation can be inconsistent because it is diffi cult to clearly defi ne defect types and then to train individuals to use the defi nitions systematically. 2 In addition, human inspectors are relatively slow compared to the number of wafers being processed through lithography cells in high-volume production. This is particularly problematic with 300 mm wafers, which have approximately 2.5 times more surface area to inspect than 200 mm wafers, and yet can be processed at the same speed or faster than 200 mm.…”