SPPE-GAN: A novel model for Die-to-Database alignment and SEM distortion correction framework

In the inspection stage of semiconductor manufacturing processes, hotspots detection is a critical task for enhancing yield. Die-to-Database (D2DB) technology allows for the comparison of actual chip images from electron beam inspection (EBI) with design databases, enabling quick identification of potential hotspots. However, D2DB faces challenges with imprecise localization on high-density chips, frequently necessitating intricate manual corrections that diminish detection efficiency. Additionally, scanning electron microscope (SEM) images can suffer from field distortion, compromising detection accuracy. To address these challenges, we propose an unsupervised SEM pattern extraction model (SPPE-GAN). This model adeptly integrates local, global, and keypoint information to precisely identify pattern locations within SEM images, subsequently transforming them into graphic data system (GDS) style images. Building on SPPE-GAN, we employ scale-invariant feature transform (SIFT) detection and the fast library for approximate nearest neighbors (FLANN) algorithm for precise matching between SEM images and their corresponding design layouts. After matching, optical flow methods are applied to indirectly correct inherent distortions in SEM images. In experiments on industrial datasets, our framework achieved over a 10% improvement in contour intersection over union (IoU) compared to assessments by experienced engineers. Moreover, under the same matching algorithm conditions, the SPPE-GAN outperformed widely used supervised methods like pix2pix and current state-of-the-art unsupervised style transfer techniques.