• 목록
• 게시판 리스트 옵션
  • 검색

In semiconductor manufacturing quality control, analyzing defect patterns in the Wafer Bin Map (WBM) plays a crucial role. Recent applications of deep learning techniques in this field have primarily focused on single-modal problems, which fail to adequately reflect the complexities of real manufacturing environments. As a result, even if defect pattern classes in WBM are accurately classified, tracking the specific causes of defects incurs significant effort and cost. To address this issue, this study proposes a multi-modal contrastive learning-based WBM defect analysis model that combines WBM image data with text data describing defect causes. Proposed model utilizes contrastive learning techniques to simultaneously optimize the loss for defect class predictions derived from WBM images and the loss for detailed textual information about defect causes. Specifically, ResNet is used to predict defect classes from WBM images, while LSTM and attention modules generate detailed textual descriptions of the defect causes. Additionally, Grad-CAM++ is utilized to visually map the textual defect information onto the images, enhancing context and interpretability. The study employs WBM defect images from the WM-811K dataset and detailed defect cause text data created based on expert knowledge. The performance of the proposed model is evaluated through classification accuracy and BLEU scores, achieving over 90% accuracy in image classification and over 20% in BLEU scores for text generation. This study introduces the first multi-modal deep-learning based WBM defect analysis model that integrates text and image data, proposing a more comprehensive and effective approach to quality control in semiconductor manufacturing processes. It is expected to enable a more accurate identification of the root causes of defects and enhance the efficiency of quality management processes.