Adjacency-Clustering for Identifying Defect Patterns and Yield Prediction in Integrated Circuit Manufacturing

Abstract Adjacency-clustering is a new concept of capturing phenomena in the presence of spatial dependencies, or Neighborhood Effect (NE). The technique is applied here to prediction problems in the presence of NE that arise in manufacturing system monitoring, quality control and yield prediction. This work is motivated by Integrated Circuit Manufacturing (ICM) process that involves multiple steps and is exceedingly expensive. Spatial variation of parameters across each wafer, where the circuits are positioned, result from equipment or process limitations, and a circuit is likely to be defective if its neighbors on the wafer are defective as well. The existence of this Neighborhood Effect, while recognized, is not well captured in traditional modeling approaches. The challenge is to extrapolate, from given samples, the patterns of the defects and predict accurately the yield of the process. The patterns are effectively identified using adjacency-clustering that is achieved with the graph-theoretic separation-deviation model, also known as the Markov Random Field (MRF) model. The use of the technique is shown to identify the defects’ patterns and provide dramatic improvements in the accuracy of yield prediction as compared to state-of-the-art methods.