Material Removal Rate Prediction with Phase Sensitive Variables Selection and Phase Partition.

In the Chemical Mechanical Planarization (CMP) process of the semiconductor industry, high accuracy prediction of Material Removal Rate (MRR) is essential to achieve wafer-to-wafer (W2W) control. Therefore, data-driven Virtual Metrology (VM) methods for MRR prediction have received much attention. Traditional methods focus on extracting features from raw data along the time dimension. However, a CMP process is a batch process with multi-phase and uneven-length characteristics, and different phases have different operating points and data correlations. In this paper, a novel VM method based on phase partition and feature selection is proposed. The phase-sensitive variables are selected by a wrapped approach and combined with a clustering algorithm for phase partition. After extracting the phase features, a suitable subset of features is selected using multiple feature selection methods and input to a regressor for prediction. The results in the PHM16 Dataset validate the effectiveness of the proposed method, which has higher prediction accuracy compared with the methods without phase partition.