Semiconductor Manufacturing Process Monitoring Based on Adaptive Substatistical PCA

Increasing yield and improving product quality are two important issues in the area of semiconductor manufacturing. The purpose of multivariate statistical process control is to improve process operations by quickly detecting process abnormalities and diagnosing the sources of the detected process abnormalities. The statistical-based multiway principal component analysis (PCA) method has drawn increasing interest in semiconductor manufacturing process monitoring. However, there are several drawbacks of this method, including future value estimation, limited number of batches, and non-Gaussian behavior of the process data. This paper proposes a new adaptive substatistical PCA-based method that can avoid future value estimation. By employing support vector data description, a new monitoring statistic is developed that has no Gaussian limitation of the process data. In addition, correlations among the new method, multimodel, and multiway PCA are detailed. Capabilities of the proposed method are demonstrated by an industrial example.