Robust, Non-Redundant Feature Selection for Yield Analysis in Semiconductor Manufacturing

Abstract

Thousands of variables are measured in line during the manufacture of central processing units (cpus). Once the manufacturing process is complete, each chip undergoes a series of tests for functionality that determine the yield of the manufacturing process. Traditional statistical methods such as ANOVA have been used for many years to find relationships between end of line yield and in line variables that can be used to sustain and improve process yield. However, a large increase in the number of variables being measured in line due to modern manufacturing trends has overwhelmed the capability of traditional methods. A filter is needed between the tens of thousands of variables in the database and the traditional methods. In this paper, we propose using true multivariate feature selection capable of dealing with complex, mixed typed data sets as an initial step in yield analysis to reduce the number of variables that receive additional investigation using traditional methods. We demonstrate this approach on a historical data set with over 30,000 variables and successfully isolate the cause of a specific yield problem.