Locating disturbances in semiconductor manufacturing with stepwise regression

The ability to locate disturbances in semiconductor manufacturing processes is critical to developing and maintaining a high yield. Analysis of variance (ANOVA), the best current practice for this problem, consists of conducting a series of hypothesis tests at each individual processing step. This approach can lead to excessive false alarms and limited sensitivity when the process contains more than one disturbance. We describe how this problem can be framed as a subset selection problem and propose two new methods based on stepwise regression. Results of over 90 000 Monte Carlo simulations suggest that these new SWR methods locate disturbances with fewer false positives and false negatives than ANOVA. This means process engineers will spend less time responding to false alarms and will be able to locate real disturbances more often.