Foundation deep learning model for accurate SEM image segmentation for CD-SEM measurements

As an inline metrology tool, CD-SEM (Critical Dimension Scanning Electron Microscope) needs to provide accurate segmentation of SEM images to ensure precise measurement of semiconductor layers. The segmentation task in CD-SEM typically requires manual tuning of parameters, which can be time-consuming and cumbersome. Additionally, the segmentation process heavily relies on information from recipe creation. In the case of process variation during runtime, the existing parameters and information may not be adapted, leading to segmentation failures. To overcome these limitations, we propose a deep learning-based segmentation algorithm. Our algorithm is trained on a large and broad dataset, making it a foundation model that can segment any new layer and does not require onsite training. The trained model possesses the ability to segment any type of layer without parameter tuning and can effectively handle process variation. For challenging cases, the model can be retrained, reducing time to market compared to entering a new development cycle as for classical algorithms. Our algorithm incorporates advanced techniques, including innovative loss functions and applicative augmentations, which effectively mimic real-world variations. It performs well under diverse SEM imaging conditions, even in scenarios with low signal-to-noise ratio (SNR), ensuring accurate identification and segmentation of nanoscale critical features. Here, we showcase the capabilities of our foundation deep learning model algorithm for segmentation of CD-SEM images.