Machine Learning Approaches for Nuisance filtering in Inline Defect Inspection

Abstract

Broadband plasma (BBP) optical defect inspection systems are widely used for process monitoring. The outputs of inspection include the defects of interest (DOI) for that process step. It is important to not only detect the DOI, but also to separate them from other types of defects that are non- relevant to the process itself, i.e., nuisance defects. The process of separating DOI from nuisance is called nuisance filtering [1], [2]. Typical nuisance filtering algorithms used on BBP systems are user-created decision trees leveraging defect attributes assigned during inspection. As design nodes shrink and pattern density increases, nuisance filtering is becoming more difficult, leading to increased recipe setup time. Further, due to the increased complexity of the decision trees, user to user variation can affect inspection performance. To solve this problem, an innovative nuisance filtering algorithm is required. The key elements for such an algorithm are consistency and improved performance compared to user-created decision trees. This paper compares traditional decision trees as well as novel machine learning approaches for nuisance filtering in inline defect inspection tools, named inLine Defect Organizer™ 2.0 (iDO™ 2.0). The study achieved improvements in increased DOI capture rate, reduced nuisance defects and faster recipe setup time.