Factories face a paradox in training AI for quality control: as production improves, fewer defective products are available for training. In semiconductor inspection, this scarcity of defects hampers automated detection since neural networks often have many normal samples but few defect examples. This patent addresses the issue by generating additional training evidence instead of relying solely on real defective chips. It extracts features from available defect images and introduces variation, using generative adversarial networks to produce more defect images. These synthetic samples complement original data, helping the defect-detection model learn from scarce authentic defects. The goal isn’t to replace real data but to enhance training where defect examples are rare, enabling AI to better recognize various defects even in imperfect conditions. This approach adapts AI to situations with limited real-world evidence, such as incompatible datasets or changing environments. Here, the challenge is simply the rarity of necessary examples. The invention transforms scarcity into a response mechanism for AI training. In manufacturing, this can lead to more effective automated inspections without intentionally producing many faulty items. Overall, it demonstrates how to enrich training data when real-world examples are limited.
Creating examples of chip-defects that are hard to find
Patent number: CN 117173105 A
Inventor(s): F. Deng, J. Luo, Y. Wang, Y. Huang, J. Zhong, N. Li, H. Wang, D. Zhong, and J. Xi
Citation: F. Deng, J. Luo, Y. Wang, Y. Huang, J. Zhong, N. Li, H. Wang, D. Zhong, and J. Xi, “芯片缺陷检测方法、装置、设备及介质 [Chip-defect detection method, apparatus, device and medium],” China Patent Application CN 117173105 A, Dec. 5, 2023.
