Fault detection for a via etch process using adaptive multivariate methods

Multivariate process control charts like Hotelling T² and squared prediction error are gaining acceptance in the semiconductor industry to monitor the increasing amount of data available by modern process tools. These methods require models built based on the covariance matrix of a trainings data set. Slowly drifting manufacturing processes degrade this estimation for the covariance matrix creating false alarms. To overcome the problem, adaptive modeling schemes are considered. The tradeoff between sensitivity and false alarms for static and adaptive models applied to a via etching process is demonstrated. Possible improvements by incorporating domain knowledge are shown.