In the rapidly evolving semiconductor industry, the motivation for technological advancement has led to a trend towards process micronization, resulting in smaller, more complex designs. This advancement has not only increased the complexity of wafer fabrication processes and the required equipment but has also led to a higher rate of defects. Furthermore, the complexity is compounded by the fact that most manufacturing processes are multi-staged, where raw materials are transformed through various pieces of equipment and phases. Identifying the origins of defects in such a detailed and layered process becomes a daunting task, heavily dependent on the expertise of operators and prone to time-consuming analyses and subjective judgments. This scenario underscores the critical need for more efficient, objective, and reliable methodologies for defect identification and analysis, moving away from traditional, manual techniques that struggle to keep pace with the demands of modern semiconductor manufacturing.
In contemporary semiconductor production environments, the situation is further complicated by the extensive volumes of sensor data generated in real-time by the array of advanced equipment in use. This data encompasses a wide array of critical information, including the current state of the wafers, detailed insights into the progress of various fabrication processes, and the operational conditions of the manufacturing equipment. These comprehensive data sources, if leveraged correctly, have the potential to revolutionize defect analysis in semiconductor manufacturing.
The focus of this study is to investigate the potential applications of this extensive sensor data in developing a deep learning-based model that can effectively utilize the abundance of information produced across multi-stage manufacturing processes. The deep learning model proposed in this research aims to automate the classification of wafers based on their defect status and, crucially, to provide insights into the underlying causes of these defects. To validate our approach, we utilized sensor data collected from two distinct pieces of process equipment used in semiconductor manufacturing. Our experiments demonstrate that our deep learning model outperforms existing methods for deep learning-based defect classification, offering a more accurate and reliable means of detecting and categorizing manufacturing defects.
Moreover, by analyzing the model's classification results, we are able to pinpoint specific sensors and pieces of equipment that have a significant influence on the defect detection result. This capability not only enhances our understanding of the manufacturing process but also identifies potential areas for improvement in equipment performance and maintenance.
The methodology proposed in this study is expected to reduce the reliance on manual analysis and lead to a more efficient, objective, and consistent approach to defect analysis. This change is anticipated to enhance production efficiency, lower manufacturing costs, and improve the overall quality of semiconductor products.