Wafer composite defect recognition framework based on residual dynamic perception network with asymmetric multi-label...

Area of Science:

• Semiconductor Manufacturing
• Artificial Intelligence
• Computer Vision

Background:

• Semiconductor manufacturing demands high accuracy in wafer pattern recognition (DPR) for improved production yield.
• Existing DPR methods struggle with mixed-patterned defects, noise, and complex defect category variations common in industrial settings.
• The limitations of current DPR necessitate advanced automated solutions.

Purpose of the Study:

• To propose a novel deep-learning-based approach, the Residual Dynamic Perception Network (RDP-Net), for automated wafer pattern recognition.
• To address the challenges of mixed-patterned defects, noise, and inter/intra-class variations in defect detection.
• To enhance the accuracy and interpretability of defect pattern recognition in semiconductor manufacturing.

Main Methods:

• Development of the Residual Dynamic Perception Network (RDP-Net) incorporating a Dynamic Perception Mechanism (DPM).
• DPM adaptively fuses multi-source information (kernels, layers, channels, resolutions) for robust feature extraction and fusion.
• Introduction of Asymmetric Multi-Label Loss (ASL) to balance sample probabilities and mitigate mislabeling impacts.

Main Results:

• RDP-Net achieved a high accuracy of 99.13% on the challenging MixedWM38 dataset.
• The proposed method significantly outperforms current state-of-the-art DPR techniques.
• Visualized experiments confirmed the interpretability of the DPM's feature learning mechanism.

Conclusions:

• RDP-Net offers a feasible and effective solution for automated defect pattern recognition in semiconductor manufacturing.
• The DPM and ASL components successfully address complexities like mixed defects and noisy data.
• The method demonstrates robustness to noise, mislabeling, and complex defect category characteristics.