Mask SEM image contour extraction using a graphic-data-stream-integrated nnU-Net model

BackgroundAccurate contour extraction from mask scanning electron microscopy (SEM) images is critical for optical proximity correction model calibration and mask quality control. However, advanced resolution enhancement techniques have resulted in increasingly complex two-dimensional mask patterns, rendering traditional one-dimensional critical dimension (CD) measurements inadequate. Furthermore, SEM images typically suffer from noise and low contrast, making precise contour extraction challenging. Given the proven capabilities of deep learning in robust contour extraction, we apply a dedicated deep learning framework to mask SEM images to assess its feasibility and accuracy.AimWe aim to develop a graphic data stream (GDS)-enhanced nnU-Net model for precise mask SEM contour extraction and CD measurements.ApproachAn automated segmentation model based on nnU-Net is proposed to segment line space type and hole type mask SEM images. The model can automatically perform preprocessing, implement data augmentation techniques, and dynamically adjust network configurations based on different imaging parameters. To improve performance on target boundary accuracy, we integrate GDS information into the nnU-Net architecture. The offset loss term that quantifies deviations between predicted contours and target GDS patterns is added to the loss function. In addition, we compared the CD values obtained from the predicted image and calculated by the software.ResultsOur proposed GDS-Integrated nnU-Net Model achieves automatic and accurate segmentation in mask SEM images, with predicted contours aligning closely with original SEM images. Notably, we compellingly demonstrate the significant potential of using model-predicted contours in metrology for semiconductor manufacturing.ConclusionsThe GDS-enhanced nnU-Net model has demonstrated its effectiveness as a powerful tool for extracting contours from SEM images.