An Attention-Augmented Convolutional Neural Network With Focal Loss for Mixed-Type Wafer Defect Classification

Abstract

Silicon wafer defect classification is crucial in improving fabrication and chip production. While deep learning methods have been successful in single-defect wafer classification, the increasing complexity of the fabrication process has introduced the challenge of multiple defects on wafers, which requires more robust feature learning and classification techniques. Attention mechanisms have been used to enhance feature learning for multiple wafer defects. However, they have limited use in a few mixed-type defect categories, and their performance declines as the number of mixed patterns increases. This work proposes an attention-augmented convolutional neural networks (A2CNN) model for enhanced discriminative feature learning of complex defects. The A2CNN model emphasizes the features in the channel and spatial dimensions. Additionally, the model adopts the focal loss function to reduce misclassification and a global average pooling layer to enhance the network's generalization by reducing overfitting. The A2CNN model is evaluated on the MixedWM38 wafer defect dataset using 10-fold cross-validation. It achieves impressive results, with accuracy, precision, recall, and F1-score reported as 98.66%, 99.0%, 98.55%, and 98.82% respectively. Compared to existing works, the A2CNN model performs better by effectively learning valuable information for complex mixed-type wafer defects.