Evolutionary computation-based reliability quantification and its application in big data analysis on semiconductor manufacturing

Abstract

Big data analysis of wafer maps in semiconductor manufacturing is essential for process reliability assessment, it is an important means of fault diagnosis in manufacturing. Although many excellent wafer map defect pattern identification (WMDPI) methods have been proposed, such as deep convolutional neural network (DCNN) based methods, there is no method that can make completely correct decisions and avoid false detections. Decision reliability is critical to the training evolution process of DCNN. In this paper, we propose a reliability quantification algorithm for wafer big data analysis. We estimate the parameter evolution uncertainty of DCNN in the WMDPI process by mixed uncertainty and thus quantify the decision reliability. In particular, reliability assessment is achieved by quantifying hybrid uncertainties, including epistemic uncertainty and aleatoric uncertainty. For epistemic uncertainty, this paper uses Monte Carlo dropout and explores the effects of drop rate, action location, and uncertainty criteria on reliability quantification through an empirical study. For aleatoric uncertainty, this paper proposes a calculation method based on Monte Carlo augmentation. The experimental results show that the identification accuracy of unknown defect patterns is 97.04% when hybrid uncertainty is considered; the identification accuracy of known defect patterns can be improved by up to 7.89%. The proposed DCNN can effectively avoid false and missed detections in the process of wafer big data analysis.