Method for proposing sort screen thresholds based on modeling etest/sort-class in semiconductor manufacturing

We propose a novel method of using machine learning algorithm for modeling and selecting important features, and a novel gradient threshold scheme for suggesting sort thresholds for filtering units that are likely to fail at class in a semiconductor manufacturing environment. Using machine learning such as gradient boosting tree enables the sort and class data to be modeled such that the effect input parameters could be ranked based on importance and each single input factor can be considered by the auto-screening algorithm. This is different from conventional statistical methods such as one-way plots, ANOVA analysis etc. which do not efficiently address the problem of confounding effect from multiple factors. The method employs the maximum gradient algorithm to select the threshold and then simulate the yield, overkill (potential good units being killed at sort) and underkill (potential bad units escaping sort screen).