March 3, 2026

Advanced deep learning architectures for automated silicon wafer defect detection with synthetic data augmentation

Key Points

Deep learning frameworks achieved up to 92.42% accuracy for silicon wafer defect classification with innovative techniques.
VGG-16 reached 80% accuracy, while MobileNet-V2 displayed the highest accuracy and recall among the tested models.
Advanced CNN models, including GoogleNet and VGG-19, showcased effectiveness in detecting rare defects through synthetic data augmentation.
The findings highlight the promise of using GAN-based techniques to address class imbalances in wafer defect detection.

Abstract

The semiconductor industry needs defect-free silicon wafers to make integrated circuits, as even small flaws can reduce yield and cause financial losses. Traditional inspection methods, such as rule-based image processing and manual checks, are time-consuming, error-prone, and inflexible. This study proposes a deep learning framework for automatic wafer defect classification using advanced CNN models and generative data augmentation to fix class imbalance and improve accuracy. The WM-811K dataset with 811,457 wafer maps was reorganized into four classes: Redundant, Crystal, Mechanical, and Defect-Free. Three baseline models (WDD-Net, MobileNet-V2, and VGG-16) were tested, with VGG-16 reaching 80% accuracy. Further experiments using deeper models (VGG-19, GoogleNet) and StyleGAN-based augmentation improved performance, especially for rare defect types. GoogleNet achieved a good balance of accuracy and efficiency, while MobileNet-V2 gave the highest accuracy (92.42%) and recall (92.41%). VGG-19 also showed strong generalization (F1-score: 90.41%), proving that deep CNNs and GAN-based augmentation are effective for wafer defect detection.

Bookmark

Advanced deep learning architectures for automated silicon wafer defect detection with synthetic data augmentation

Key Points

Abstract

Cite This Study