Over the past few years, noise2noise, noise2void, noise2self, and unsupervised deep-learning (DL) denoising techniques have achieved great success, particularly in scenarios where ground truth data is not available or is difficult to obtain. For semiconductor SEM images, ground truth or clean target images with lower noise levels can be obtained by averaging hundreds of frames at the same wafer location, but it is expensive and can result in physical damage to the wafer. This paper’s scope is to denoise SEM images without clean target images and with limited image counts. Inspired by noise2noise, we proposed an additive noise algorithm and DL U-net. We achieved good denoising performance using a limited number of noisy SEM images, without the clean ground truth images. We proposed the “denoise2next” and “denoise2best”. We compared generative adversarial network(GAN) generated images and Additive noise images for data augmentation. This paper further quantified the impact of image noise level, pattern diversity, and continuous (aka transfer) learning. The data sets used in the work include both line/space and logic pattern.
In the semiconductor manufacturing industry, Automatic Defect Classification (ADC) plays an important role in maintaining high wafer inspection quality and reducing yield loss. ADC performance has benefitted from using machine learning (ML) algorithms; however, performance is negatively affected by the data imbalance and limited amounts of training data. Synthetic Minority Oversampling Technique (SMOTE) is an oversampling technique to adjust the skewed class distribution of a dataset so that the bias of the majority class is reduced. This paper shows that applying SMOTE achieved higher accuracy and purity on two imbalanced datasets, consisting of scanning electron microscopy (SEM) images collected with ASML-HMI eP™ and eScan® series inspection tools. The ML models are also less sensitive to the selection of hyperparameters when SMOTE is applied. We also show that better classification results can be obtained with less training samples with SMOTE; we conducted an experiment where a ML model trained on only 25% of samples with SMOTE achieved a higher ADC accuracy and purity performance compared to the same ML model trained on all samples but without SMOTE. In another experiment using a highly imbalanced SEM dataset with very few counts of the defect-of- interest (DOI), the combination of SMOTE and random undersampling of the majority class improves the accuracy by up to 5x while maintaining the same level of purity.
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.