Statistical Process Control is widely used in Semiconductor Manufacturing. Univariate control charts exist for Machine data and for measurements of the processed wafers. They in general work very well but nevertheless not all faults can be detected. Especially the interaction between parameters is not taken into account.
The advantages of multivariate charts are:
- Data reduction by the use of only one chart for many parameters - Detection of faults that arise in the small drift of many univariate parameters - The immediate classification of the cause of failure in the chart This work tries to approach multivariate control charts for detecting and classifying faults of the etching equipments in one single step.
The idea of the work is based on a paper by Goodlin et. al.
('Simultaneous Fault Detection and Classification for Semiconductor Manufacturing tool', Journal of the Electrochemical Society G778-G784(2003)). They created an experiment, where several parameters were varied, one parameter per wafer. Their method focuses on the resulting variation of the parameters which were not changed by the experiment. Normally processed wafers and the experimental wafers were used to construct univariate, T 2 and Faultspecific control charts.
Especially the faultspecific charts detected most of the experimental simulated faults correctly.
In this work the idea of the paper was applied to real-time data. A similar experiment was constructed on a Poly-etching machine. The experiment was designed to reflect the most important known errors of the equipment.
A mix of expert know how and methods like Principal Components Regression or Partial Least Squares Regression were used to select the appropriate parameters for constructing the multivariate charts. In the next step the control charts were built for the real time data with the main focus on faultspecific charts. The next step was to control the etching of new wafers. Challenges in controlling these newly processed wafers were Wet Cleans and Process Changes. Therefore the original Faultspecific Models had to be extended to Dynamic Linear Models.