A stack fusion model for material removal rate prediction in chemical-mechanical planarization process

Abstract

In chemical-mechanical polishing process of wafers, the accurate prediction of the average material removal rate is vital for the estimation of the polishing time, which may significantly optimize the production efficiency while maintaining the acceptable quality. In this study, a new stacking fusion model is proposed, which offers a precise way to predict the material removal rate based on the indirect sensor data collected from the wafer polishing process. Through a procedure of feature creation, feature expansion, and feature encoding, the data from the wafer polishing process were transformed into multi-dimensional information. Then, through feature importance evaluation and by following steps of feature selection, a feature subset that is effective for the material removal rate was selected. The accuracy of the primary learner was optimized via the stacking fusion model to establish a highly non-linear mapping relationship between the features and the material removal rate. Compared with other weighted average models and neural network fusion models, this method presented improved precision under several working conditions. The method is promising in terms of becoming embedded into chemical-mechanical polishing machines to enable an online real-time accurate prediction of the material removal rate.