The use of historical defect imagery for yield learning

Abstract

The rapid identification of yield detracting mechanisms through integrated yield management is the primary goal of defect sourcing and yield learning. At future technology nodes, yield learning must proceed at an accelerated rate to maintain current defect sourcing cycle times despite the growth in circuit complexity and the amount of data acquired on a given wafer lot El]. As integrated circuit fabrication processes increase in complexity, it has been determined that data collection, retention, and retrieval rates will continue to increase at an alarming rate.Oak Ridge National Laboratory (ORNL) has been working with International SEMATECH to develop methods for managing the large volumes of image data that are being generated to monitor the status of the manufacturing process [2, 31. This data contains an historical record that can be used to assist the yield engineer in the rapid resolution of manufacturing problems. To date there are no efficient methods of sorting and analyzing the vast repositories of imagery collected by off-line review tools for failure analysis, particle monitoring, line width control, and overlay metrology. In this paper we will describe a new method for organizing, searching, and retrieving imagery using a query image to extract images from a large image database based on visual similarity. Keywords yield management, yield learning, image management, datamining, content-based image retrieval, automated image retrieval, approximate nearest-neighbors searching, visual similarity