Wafer Defect Image Generation Method Based on Improved Styleganv3 Network.

Area of Science:

• Semiconductor Manufacturing
• Artificial Intelligence
• Computer Vision

Background:

• High-fidelity wafer defect image generation is crucial but challenging due to limited real-world data.
• Existing methods often lack physical authenticity and struggle with small datasets.

Purpose of the Study:

• To develop an enhanced generative model for high-fidelity wafer defect image synthesis.
• To improve the reconstruction ability of wafer datasets using limited data.
• To address the challenges of data scarcity and physical authenticity in generative models.

Main Methods:

• An enhanced StyleGANv3 framework incorporating a Heterogeneous Kernel Fusion Unit (HKFU) for multi-scale feature refinement.
• Integration of a Dynamic Adaptive Attention Module (DAAM) to enhance discriminator sensitivity.
• Training and evaluation on the Mixtype-WM38 and MIR-WM811K datasets.

Main Results:

• State-of-the-art performance achieved on benchmark datasets.
• FID scores of 25.20 and 28.70, and SDS values of 36.00 and 35.45 demonstrate high-fidelity generation.
• Successful generation of realistic wafer defect images from limited data.

Conclusions:

• The proposed method effectively alleviates the problem of limited datasets in generative modeling.
• The enhanced StyleGANv3 framework contributes significantly to data preparation for wafer defect classification and detection.
• This work advances the field of synthetic data generation for semiconductor quality control.