Upon Human's Unknown, Can AI Help to Address Uncertainty? A Practice on Package Substrate AOI Defect Detection

To recognize defects produced in manufacturing (MFG) process becomes a daily in-line and in-time challenge in today's semiconductor industry. This paper presents an AI-assisted defect inspection method for substrates, which are widely used in packaged IC products. With the help of Automated Optical Inspection (AOI) equipment and the adoption of Machine Learning (ML) and Deep Neural Network (DNN) technologies, manual efforts in substrate defect examination have significantly been reduced, and yet it is still insufficient for the high standard required by automatic defect classification task. One big challenge is that while MFG in-process conditions change, the defect variation also occurs. The defect variances may include some uncertain scenarios, such as compound of multiple known defects, or even out-of-domain (OOD), i.e., unknown defect. A question is then raised: can we build an AI model which has capability to address uncertainty just like human reaction when facing unsure situations? In this paper, we introduce a new ML technology AL-VAE which integrates Active Learning (AL) and Variational Autoencoder (VAE) techniques to enhance our field-deployed AOI defect detection flow. The AL can address the uncertainty and filter out the ambiguity, accelerate training convergence, and minimize the training dataset to reduce labeling efforts. The VAE aims at learning deep latent-variable distribution of image, which can reproduce the similar image's feature multiple times from latent space. Aggregating the reproduced samples can significantly increase the uncertainty information, especially for the unknown defects. To demonstrate the advantage of our ML method, the public MS COCO dataset and our MFG substrate-defect dataset are evaluated. The results indicate that our method selects around 50% samples from original training dataset, and the overall accuracy is still kept at the same level. In addition, the uncertainty information aggregating samples reproduced from VAE shows that our model can recognize the difference between trained classes and untrained scenarios. Our technology enhances the field-deployed AI-assisted defect inspection flow with human-like intelligence which can significantly narrows the gap between AI computing and human understanding.