While extreme ultraviolet lithography has contributed to sub-10nm microfabrication, there are concerns about stochastic defects. Thus, the process evaluation requires fast and precise inspection of entire wafers. To do this, large field-of-view (FoV) e-beam inspection has been introduced. However, large FoV inspection sometimes suffers from image degradations due to aberrations and/or charged wafers that cause false detections during image comparison inspection. To reduce these false detections, we developed a deep learning-based image adaptation method to reduce the difference between the reference image and degraded inspection image. Here, the adapter that simply minimizes the difference often falls into over-adaptation that eliminates the difference in defect characteristics and decreases detection sensitivity. To address this, we introduced a patch-wise blind-spot network (PwBSN) that recognizes only the image degradation by leveraging the property that the defect region is smaller than the image degradation region. Since the PwBSN can only use surrounding regions due to its architectural constraints, it only minimizes the difference in degradations except for defects smaller than patches. We applied this method to deep learning-based die-to-database defect inspection. The evaluation on SEM images showed that the proposed method detects only defects, while a conventional method detects both defects and image degradation regions.
The measurement process is important in managing semiconductor device yield, which is affected by the availability of measurement equipment such as critical dimension scanning electron microscopes (CD-SEMs). Here, decreasing CD-SEM availability is caused by measurement errors when inappropriate measurement recipes are used. To improve CD-SEM availability, we developed a machine-learning-based error analysis method to identify error causes and x measurement conditions by using accumulated CD-SEM data. However, a single error analysis model can be applied to only a single semiconductor product because different semiconductor products have different data distribution even if they use the same recipe. Additionally, manufacturers often modify recipes every few months. As a result, an error-cause analysis method needs to be able to easily adapt to new recipes. Therefore, we developed a three-stage method that consists of error detection modeling, feature scoring, and error cause estimation on the basis of high-scoring features. Because we found the top scoring features do NOT change as the feature distribution changes when the error causes are the same, the error cause estimation on the basis of high-scoring features enable to be applied to different semiconductor products and new recipes. We evaluated our method with actual operational data, and estimated error causes that often correspond with the results of manual analysis by skilled engineers.
Extreme ultraviolet lithography has advanced microfabrication of semiconductor devices toward the sub-10-nm generation. In this situation, stochastic defects increase and hence process evaluation requires an entire wafer inspection at high speed. To satisfy this requirement, a large field of view (FoV) inspection with low-resolution enables us to inspect an entire wafer within an acceptable time because the throughput of e-beam inspection depends on imaging resolution. However, low-resolution images are difficult to inspect at high precision using conventional methods because of a smaller photographed defect size and worse signal-to-noise ratio. Moreover, deformation caused by the manufacturing process and larger distortion caused by large FoV result in false detections when we apply die-to-database (D2DB) inspection. To solve these issues, we propose trainable D2DB inspection, which predicts a pixel-value distribution of normal images from a corresponding design layout. The proposed method is robust to low-resolution images because it considers noise and acceptable deformation as variance of the learned distribution. In addition, by introducing a model to predict a misalignment between a design layout and inspection image, trainable D2DB becomes robust to image distortion. Experiments show that trainable D2DB can perform high-precision inspection on images with large noise and image distortion.
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.