Self-Training-Based Wafer Map Defect Pattern Labeling for Classification Accuracy Improvement

Accurately classifying wafer map defect patterns is essential for tracking their occurrence and supporting root-cause analysis of systematic defects arising from manufacturing processes. While AI-based approaches have recently gained significant interest, they typically demand large volumes of labeled samples to achieve high learning accuracy. In this work, we introduce a self-training labeling approach that expands the training set by assigning labels to unlabeled data. Our method first trains an AI model using a limited amount of labeled data. The trained model then generates predictions for the unlabeled samples. We modify the prediction confidence for data that may be prone to misclassification and subsequently assign labels to a subset of the unlabeled samples based on this adjusted score. We present the results of a simulated application of the method on labeled data, demonstrating a labeling accuracy of 96.7%.