Yu, Y., Wang, J., Chang, R., Chong, S. C., Pahwa, R. S., & Yang, X. (2025). An Efficient Data Augmentation and Semantic Segmentation Framework for 3D Defect Detection of HBMs. 2025 IEEE 75th Electronic Components and Technology Conference (ECTC), 1628–1635. https://doi.org/10.1109/ectc51687.2025.00277
Abstract:
Defect detection in semiconductor packaging, particularly for 3D High Bandwidth Memory (HBM) packages, is crucial for ensuring high manufacturing quality and reliability. Traditional destructive testing (DT) methods are inefficient and costly, while non-destructive testing (NDT) using 3D X-ray machines offers a promising alternative, especially when coupled with advanced deep learning methods. However, current 3D deep learning-based segmentation techniques rely heavily on large, densely annotated datasets, making them impractical for routine use due to the extensive time and effort required for annotation. To address this challenge, we propose an end-to-end solution that synthesizes 3D X-ray Microtomography (XRM) scans using a limited amount of labeled data, enabling advanced 3D semantic segmentation for defect detection. Our method significantly reduces annotation efforts by generating artificial training data, mitigating the impact of limited labeled samples on segmentation accuracy. Additionally, it enhances model efficiency by optimizing the training process, reducing data hunger, and improving adaptability to new designs. Experiments demonstrate that our approach reduces annotation requirements up to 90% while maintaining comparable performance up to 90% in bump segmentation and even better performance in defect detection. Our proposed method provides a practical and efficient solution for semiconductor defect detection by minimizing annotation costs while ensuring high segmentation accuracy. By reducing data and computational requirements, our work promotes the broader adoption of AI-driven solutions in semiconductor manufacturing.
License type:
Publisher Copyright
Funding Info:
This research / project is supported by the Agency for Science, Technology and Research (A*STAR) - Manufacturing, Trade, and Connectivity Programmatic Funds
Grant Reference no. : M23L7b0021