Convolutional neural network for imbalanced data classification of silicon wafer defects

Abstract

Integrated circuit chip fabrication may induce defects on silicon wafers due to inadequate manufacturing environment, equipment malfunctioning and operational flaws. An identification and analysis of these defects facilitates the process engineering by backtracking and addressing their causes of generation. There exist various computer vision techniques to automate the silicon wafer defect inspection. Recently, deep learning has been employed for the task, with its diverse modeling approaches and network configurations, aiming at providing the best performing classifiers for wafer defect inspection. However, the data for wafer inspection is known to be highly imbalanced. For instance, WM-811K, having abundance of non-defective images and small sets of samples in defect classes. Deep networks trained on such uneven data tend to bias towards major classes leading to false classifiers. This research proposes a convolutional neural network while addressing class imbalance through data undersampling. The proposed method performance has been evaluated by a set of metrics and results are compared with an existing machine learning approach. The results comparison demonstrated 90.44% test accuracy of the method as compared to 78.48% of the existing work. This shows that the imbalance mitigation with the proposed strategy offers a better solution for the wafer defects classification.