Defect detection of semiconductor wafer EB-SEM images based on convolutional neural networks

Abstract To enhance manufacturing yield and product reliability, it is necessary to detect nanoscale defects in semiconductor wafers with high accuracy. Nevertheless, the automated analysis of Electron Beam-Scanning Electron Microscope (EB-SEM) images is still a problem because of noise, sparse labeled data, class imbalance, and subtle defect features. This paper suggests a hybrid deep learning model to detect and classify wafer defects robustly. Wiener filtering is used to remove noise in EB-SEM images, whereas Conditional Generative Adversarial Network (C-GAN)-based augmentation is used to address data scarcity and imbalance. Several deep convolutional neural network architectures such as GoogleNet, EfficientNet, ResNet, and VGG are trained with a spider monkey-inspired optimization algorithm to improve feature learning and classification. A Graph Convolutional Network (GCN) with an XGBoost classifier is used to make final predictions to further utilize structural relationships between features. Experimental testing of wafer SEM image datasets shows that it is more accurate and robust than traditional methods. The suggested framework is quite effective in identifying various types of defects and is generalized well in different conditions. These findings suggest that the suggested approach is a viable and effective way to conduct automated inspection of defects in semiconductor production facilities.