In this study, we aimed to improve the prediction accuracy of wafer shape in the double-side polishing process by focusing on the consistency between the temporal granularity of process parameters and fixed parameters. A machine learning model trained on time-series data with temporal granularity aligned to the batch unit exhibited higher prediction accuracy compared to a model using data with minimum time unit, as confirmed by improvements in both R² score and RMSE. Furthermore, feature importance analysis revealed that the number of pad usages was the most dominant factor, and aligning the temporal granularity with that of the fixed parameters allowed this influence to be more accurately captured. These findings suggest that wafer shape variations over time caused by changes in pad condition can be predicted with higher accuracy. The results of this study are expected to contribute to the realization of data-driven quality management and predictive control utilizing equipment log data from the manufacturing floor, thereby supporting further optimization of the double-side polishing process.
TSUCHIDA et al. (Wed,) studied this question.