Metrology and segmentation in SEM/TEM imaging: accelerating semiconductor analysis through advanced deep learning techniques

In the semiconductor industry, accurate and fast metrology and segmentation are crucial for complex structures analysis observed in Scanning Electron Microscopy (SEM) and Transmission Electron Microscopy (TEM) images. Traditional image analysis methods often struggle with low contrast, intricate object boundaries, and the need to distinguish between different materials composing an object. To address these challenges, we propose a deep learning model that enhances segmentation accuracy and automates metrology processes to analyze objects composed of multiple materials at advanced steps of manufacturing. Our approach proposes a transformer-based feature extractor in combination with a multi-part segmentation model. We adapted the feature extractor with SEM/TEM public images and then trained the segmentation model. This model also incorporates a novel combination of different losses to increase robustness and accuracy. These enhancements enable the model to precisely detect and segment both inner and outer objects, even in low-contrast environments, while maintaining high sensitivity to edges and contours with minimal user inputs. By integrating these advanced loss functions, our model achieves robust segmentation performance across a variety of semiconductor materials and structures. The ability to accurately delineate complex features, such as inner voids and outer shell boundaries, significantly improves the reliability of automated metrology, reducing the need for exhaustive manual annotation. This work demonstrates that our deep learning model not only accelerates the analysis process but also enhances the precision of measurements. Enabling the use of this model in semi-automatic mode with user feedback or fully automatic once validated makes it an invaluable tool for semiconductor manufacturing and research. The model's adaptability to varying image conditions and its focus on both inner and outer object detection with the possibility to design measurement strategies mark a significant step forward in the field of automated microscopy image analysis. It provides a scalable solution for the semiconductor industry's growing demands.