An Empirical Comparison of Spatial Randomness Models for Yield Analysis

Abstract

Yield analysis is an important activity in the assessment and control of semiconductor fabrication processes. Tests of spatial randomness provide a means of enhancing yield analysis by considering the patterns of good and defective chips on the wafer. These patterns can be related to the likely sources of defects during production. This paper compares two approaches for determining spatial randomness based on join-count statistics. The first assumes that a random distribution of defects can be modeled as a spatially homogenous Bernoulli process (SHBP). The second uses a Markov random field (MRF) as the null distribution. While both methods are shown to have good performance, the MRF outperforms the SHBP on both clustered and random defect data.