Joint Research Conference

June 24-26, 2014

A regularized regression procedure for virtual metrology application

Abstract:

Modern semiconductor tools are equipped with many sensors for tracing processing variables. This has prompted the semiconductor industry to build models to predict process outputs. The prediction, named Virtual Metrology, can enhance process control that traditionally is based on measurements from sampled wafers and can improve the percentage of wafers skipped for metrology. Wafers are usually produced in multiple chambers, each of which has multiple sides. A single regression model is built for each side of a chamber to predict the quality of wafers produced therein. In this paper, we proposed a regularized regression model which uses the relatedness among the parameters of various models to improve the prediction performance. This method also compensates for the lack of label information in building individual models and predicts well on future unseen examples. We propose an optimization algorithm to evaluate the model formulation and demonstrate its effectiveness on both synthetic and real data sets. Joint work with J. He and E. Yashchin (IBM).