The two standard reticle defect inspection methods are die-to-die and die-to-database. The die-to-die inspection method compares images from the two dice on the same reticle to identify any defect. However, the die-to-database inspection method compares images from the reticle with the design data (CAD). The previous year, we built an SEM-based VSB writer classification system for die-to-die inspection that used state-of-the-art deep learning models to identify errors such as shape, position, and dose [1]. Using the deep neural networks and DL-based SEM digital twins [2], we showed better accuracy than the average human expert in classifying SEM-based defects. However, a limitation remained that the DL model wasn’t aware of chrome and glass regions, just from the input SEM. This information is helpful to make better decisions in classifying some typical errors achieving higher accuracy. In the current paper, we improve the accuracy of the existing classifier by enhancing the underlying deep learning model and supplementing it with the recognition of chrome and glass (exposed and unexposed) regions further. We make it possible with yet another DL-based SEM2CAD digital twin to automatically identify exposed/unexposed areas from the SEM and augment manual input by the expert to it. We feed this new information into the SEM classifier that currently takes a reference and error SEM image for more accurate results. In addition, we also built an SEM-based defect classification system for the die-to-database inspection to categorize various types of VSB mask writer defects, which requires defect SEM images and the reference CAD. Using several deep neural network models and digital twins, in this paper, we provide a production-grade system for the VSB writer’s SEM-based defect classification that works for both die-to-die and die-to-database inspection methods.
Sub-nanometer accuracy attainable with electron micrograph SEM images is the only way to “see” well enough for the mask analysis needed in EUV mask production. Because SEM images are pixel dose maps, deep learning (DL) offers an attractive alternative to the tedious and error-prone mask analysis performed by the operators and expert field application engineers in today’s mask shops. However, production demands preclude collecting a large enough variety and number of real SEM images to effectively train deep learning models. We have found that digital twins that can mimic the SEM images derived from CAD data provide an exceptional way to synthesize ample data to train effective DL models. Previous studies [1, 2, 3, 4] have shown how deep learning can be used to create digital twins. However, it was unclear if SEM images generated with digital twins would have sufficient quality to train a deep learning network to classify real SEM images. This paper shows how we built three DL tools for SEM-based mask analysis. The first tool automatically filters good quality SEM images, particularly for test chips, using a DL-based binary classifier. A second tool uses another DL model to align CAD and SEM images for applications where it is important that features on both the images are properly aligned. A third tool uses a DL multi-class classifier to categorize various types of VSB mask writer defects. In developing the three tools, we trained state-of-the-art deep neural networks on SEM images generated using digital twins to achieve accurate results on real SEM images. Furthermore, we validated the results of trained deep learning models through model visualization and accuracy-metric evaluation.
Sub-nanometer accuracy attainable with electron micrograph SEM images is the only way to “see” well enough for the mask analysis needed in EUV mask production. Because SEM images are pixel dose maps, deep learning (DL) offers an attractive alternative to the tedious and error-prone mask analysis performed by the operators and expert field application engineers in today’s mask shops. However, production demands preclude collecting a large enough variety and number of real SEM images to effectively train deep learning models. We have found that digital twins that can mimic the SEM images derived from CAD data provide an exceptional way to synthesize ample data to train effective DL models. Previous studies [1, 2, 3, 4] have shown how deep learning can be used to create digital twins. However, it was unclear if SEM images generated with digital twins would have sufficient quality to train a deep learning network to classify real SEM images. This paper shows how we built three DL tools for SEM-based mask analysis. The first tool automatically filters good quality SEM images, particularly for test chips, using a DL-based binary classifier. A second tool uses another DL model to align CAD and SEM images for applications where it is important that features on both the images are properly aligned. A third tool uses a DL multi-class classifier to categorize various types of VSB mask writer defects. In developing the three tools, we trained state-of-the-art deep neural networks on SEM images generated using digital twins to achieve accurate results on real SEM images. Furthermore, we validated the results of trained deep learning models through model visualization and accuracy-metric evaluation.
Deep learning has an increasing impact on our personal and professional lives. Deep learning has the potential to transform mask, semiconductor and electronics manufacturing. This paper reviews key results from the Center for Deep Learning in Electronics Manufacturing’s (CDLe’s) first year of operation. We consider results from adapting five common types of deep learning recipes to solve key challenges in the manufacture of photomasks, printed circuit boards (PCBs), and flat panel displays (FPDs). These deep learning applications include 1) grouping similar items to automatically categorize mask rule errors; 2) using U-Net architecture to construct fast mask designs; 3) using vision-based object classification to find and classify pick-and-place (PnP) errors on PCB assembly lines; 4) using anomaly detection to improve the quality of FPDs; and 5) using digital twins to create SEM images and optimize Inverse Lithography Technology (ILT). While we compare the relative benefits of these techniques, all show the importance of data to improve the success of deep learning networks and of electronics manufacturing. These applications rely on varying neural network architectures such as autoencoders, segmentation networks, deep convolutional networks, anomaly detection, and generative adversarial networks (GANs).
Deep learning (DL) is one of the fastest-growing fields in artificial intelligence (AI). While still in its early stages of adoption, DL has already shown it has the potential to make significant changes to the lithography and photomask industries through the automation or optimization of equipment and processes. The key element required for application of DL techniques to any process is a large volume of data to adequately train the DL neural networks. The accuracy of the classification or prediction of any DL system is dependent on the depth and breadth of the training data to which it is exposed. For semiconductor manufacturing, finding adequate data – especially for corner cases – can be difficult and/or expensive. In this paper, we will present two digital twins that are themselves built from DL as a part of a DL Starter Kit. We will demonstrate the creation of DL-based digital twins for a mask scanning electron microscope (SEM) and for curvilinear inverse lithography technology (ILT).
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.