| dc.description.abstract | Ensuring the quality of wafers in semiconductor manufacturing is crucial for the performance and reliability of electronic devices. Traditional defect sampling strategies, such as static and dynamic sampling strategy, often fail to cope with the increasing complexity of defect detection, leading to inefficiencies and high labor costs. To address this issue, this study proposes an advanced wafer defect sampling strategy leveraging machine learning techniques. The proposed strategy employs Random Forest (RF) and Extreme Gradient Boosting (XGBoost) models, combined with the Synthetic Minority Over-sampling Technique (SMOTE), and utilizes the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) method to classify defects into non-clustered and clustered categories, thereby enhancing the accuracy and efficiency of defect identification. For non-clustered defects, the Modified Sampling Gains Chart is used to determine the probability threshold required for defect identification. For clustered defects, the defects are ranked based on the probability of the dies within the cluster being defective, and the top five dies with the highest defect probability are sampled.
Validation experiments demonstrate that the proposed strategy significantly reduces the number of sampling dies while maintaining a high recall rate of defect. For non-clustered defects, when sampling dies with a defect probability exceeding 20%, the recall rate for nonclustered defects reaches 90%. This indicates that the proposed strategy effectively identifies individual defects across the wafer. Additionally, for clustered defects, the proposed strategy successfully identifies all defects within the clusters with minimal false positives, using a low number of sampling points. Compared to the sampling algorithms developed by the collaborative semiconductor manufacturer, the machine learning-based sampling strategy proposed in this study achieves a 90% defect recall rate while reducing the total number of sampling points by nearly 80%. This not only significantly enhances detection efficiency but also substantially reduces production costs. Consequently, this sampling strategy is more efficient than traditional methods and demonstrates practical applicability in production lines. | en_US |