An Auto-adjusting Weight Model for Imbalanced Wafer Defects Recognition

Abstract

Wafer defect recognition is popular research in the semiconductor industry. Generally, each defect pattern is related to a specific manufacturing problem. By identifying defect patterns correctly, manufacturing problems can be recognized and fixed in time, which improves the quality and production yield of wafers. However, due to the location, light and the increasing number of wafers, traditional recognition methods achieve unsatisfactory performance. Currently, convolutional neural network (CNN) based methods outperform traditional methods in accuracy and speed, but fail when training with imbalanced target classes. To address the imbalanced problem, a CNN-based knowledge distillation (KD) method is proposed. To improve the identification of different types of defects, a multi-head attention layer is applied to the proposed CNN model, which enriches local and global information of features. Besides, when training the CNN model, target features are constrained with Distillation Loss and Focal Loss, reducing the effect of dataset imbalance. Experiments on the public dataset WM-811K are conducted to verify the proposed methods, and experimental results showed that the accuracy, precision, recall, specificity, and F1 score of our method reached 97.7%, 96.9%, 97.2%, 99.7% and 97.0% respectively, and the classification accuracy of each class was above 93.0%, which indicates the proposed method was reasonable and effective on large-scale imbalanced wafer defect datasets.