Ensemble convolutional neural networks with weighted majority for wafer bin map pattern classification

Wafer bin maps (WBM) provides crucial information regarding process abnormalities and facilitate the diagnosis of low-yield problems in semiconductor manufacturing. Most studies of WBM classification and analysis apply a statistical-based method or machine learning method operating on raw wafer data and extracted features. With increasing WBM pattern diversity and complexity, the useful features for effective WBM recognition are highly dependent on domain knowledge. This study proposes an ensemble convolutional neural network (ECNN) framework for WBM pattern classification, in which a weighted majority function is adopted to select higher weights for the base classifiers that have higher predictive performance. An industrial WBM dataset (namely, WM-811K) from a wafer fabrication process was used to demonstrate the effectiveness of the proposed ECNN framework. The proposed ECNN has superior performance in terms of precision, recall, F1 and other conventional machine learning classifiers such as linear regression, random forest, gradient boosting machine, and artificial neural network. The experimental results show that the proposed ECNN framework is able to identify common WBM defect patterns effectively.