A Robust, Non-Parametric Method to Identify Outliers and Improve Final Yield and Quality

This paper discusses the shortcomings of two widelyused outlier detection methods, and then introduces a more robust non-parametric algorithm that can better handle the wide variety of non-standard distributions seen in the semiconductor manufacturing industry. When applied to test data gathered at a wafer sort step, this algorithm allows for the removal of outlier die (that otherwise passed test limits) in order to improve quality and yield in the final product. The algorithm was designed to be as simple as possible, with a minimum number of assumptions about the underlying dataset. Its design was inspired by the human ability to identify outliers almost instantaneously from almost any distribution. It will be presented first in mathematically-informal, yet easily-understood terms, with a more rigorous mathematical definition to follow.