Prediction of Material Removal Rate Using Phase Partition and Semi-supervised Regression

Abstract

In the Chemical Mechanical Planarization (CMP) process, the Material Removal Rate (MRR) is an important indicator of polishing performance. However, MRR is difficult to be measured. Therefore, Virtual Metrology (VM) methods to predict MRR by process variables have become a hot topic. Conventional VM methods extract the global statistical characteristics of the CMP process, ignoring its multi-phase characteristics. Meanwhile, a large amount of wafer data lacking MRR annotation is discarded, resulting in information waste. In this paper, a novel semi-supervised VM method for MRR is proposed. Through the clustering of sampling points, the CMP process is divided into multiple phases, and phase features are extracted. Since the information of unlabeled wafer samples needs to be utilized, two evaluation metrics, the stability constraint, and the accuracy gain are proposed to evaluate the confidence of pseudo-labels. The samples with high confidence will be added to the labeled sample set. Finally, the MRR is predicted by a semi-supervised regression algorithm. The proposed method is validated on a PHM Data Challenge dataset. The experimental results show that the application of phase partition and semi-supervised method is helpful to improve the prediction performance of the model and reduce the prediction error. The prediction performance of the proposed method outperforms most existing methods in the case of significantly fewer labeled samples.