A Novel End-to-End Deep Learning Framework for Chip Packaging Defect Detection.

Area of Science:

• Semiconductor Manufacturing
• Artificial Intelligence
• Materials Science

Background:

• Increasing semiconductor complexity raises void defect risks during packaging.
• Identifying these defects is challenging due to complex backgrounds and varied defect characteristics.

Purpose of the Study:

• To develop a deep learning framework for precise void defect segmentation in chip packaging.
• To improve the accuracy and efficiency of void defect detection in semiconductor manufacturing.

Main Methods:

• A novel framework combining solder region extraction and a void defect segmentation network.
• Utilized a Mamba model-based encoder with a visual state space module for multi-scale feature extraction.
• Employed an interactive dual-stream decoder with a feature correlation cross gate module for enhanced segmentation.

Main Results:

• The framework demonstrated effectiveness in quantitative and qualitative experiments on a custom X-ray chip dataset.
• Achieved 93.3% accuracy in chip qualification when applied to a real factory inspection line.
• Successfully addressed challenges of complex backgrounds, varying defect sizes, and blurred boundaries.

Conclusions:

• The proposed deep learning framework offers a robust solution for void defect segmentation in semiconductor packaging.
• This technology significantly enhances the reliability of automated inspection in chip manufacturing.
• The framework's successful real-world application highlights its practical value in quality assurance.