Improving Semiconductor Defect Classification in SEM Images via Domain Adaptation

Automated defect classification in semiconductor manufacturing is a critical task, where various types of defects need to be identified and classified. With the increasing complexity of semiconductor devices, defect classification has become a challenging problem, especially when dealing with limited labeled data. However, in real-world scenarios, labeled data is often imbalanced and limited, making it challenging to develop accurate models. Domain Adaptation is a technique that enables models to adapt to new domains with limited labeled data. In this paper, we propose the Margin-Margin Discrepancy Distance (M$^{2}$DD) algorithm, which replaces cross-entropy with margin cross-entropy, allowing the minority class's decision boundary to be expanded and robustly separated from the majority class. We evaluate M$^{2}$DD on four domain adaptation tasks and demonstrate its superiority in defect classification tasks with sparse labeled data in the target domain. Our results show that M$^{2}$DD outperforms traditional ERM, which relies on spurious features and existing MDD, by effectively capturing core features. This makes M$^{2}$DD a reliable approach for real-world applications in semiconductor defect classification.