The semiconductor industry is critical in driving technological advancements in various fields. Ensuring the production of high-quality semiconductor devices is critical to maintaining reliability and performance. Traditional manual inspection methods are time-consuming and expensive, which leads to an increasing need to develop automated systems that can efficiently and accurately detect abnormalities in semiconductor devices. However, there are many challenges in developing machine learning inspection systems for real-world production, from ensuring high-quality training data to designing appropriate algorithms to solve specific problems in the data. When preparing datasets for data-driven solutions, manual labeling is prone to human labeling errors, directly affecting detection syste...[ Read more ]

The semiconductor industry is critical in driving technological advancements in various fields. Ensuring the production of high-quality semiconductor devices is critical to maintaining reliability and performance. Traditional manual inspection methods are time-consuming and expensive, which leads to an increasing need to develop automated systems that can efficiently and accurately detect abnormalities in semiconductor devices. However, there are many challenges in developing machine learning inspection systems for real-world production, from ensuring high-quality training data to designing appropriate algorithms to solve specific problems in the data. When preparing datasets for data-driven solutions, manual labeling is prone to human labeling errors, directly affecting detection systems’ accuracy. Even if the inspection system is perfectly trained on a particular batch of data, when the distribution of the test data changes, such as a new fault type that has yet to be learned, the model will not work correctly. Implementing machine learning involves iterative cycles, requiring continuous training and evaluation.

This paper comprehensively studies the existing inspection system and analyzes its difficulties. The existing inspection system has a rejection rate of 0.5%. Due to the enormous amount of semiconductors produced daily, more than one million are discarded, but only less than 0.01% of them are actually defective. We aim to reduce the false positive rate and collect those ”hard samples” rejected by existing inspection systems. We then address this problem using a variety of machine learning techniques, ranging from traditional supervised image classification, self-supervised contrastive learning, and semi-supervised self-training to unsupervised reconstruction-based learning, focusing on dealing with noisy labels caused by human factors and data drift caused by marking codes of different production batches.