Automated Classification of Semiconductor Defect Density SEM Images Using Deep Learning

The early detection of defect density issues is a crucial part of product reliability in the semiconductor manufacturing industry, since it prevents failures of the final product. To assure reliability numerous inspections are performed. One of these inspections involves the classification of defects on wafers using images captured with scanning electron microscopes. Currently, experts perform manual classification of these images. However, this task is susceptible to errors and is not efficient. The goal of this work was to increase the reliability of the classification by developing a deep learning pipeline, using convolutional neural networks, for automatic defect classification. The basis for training the models is a set of historical images stored in the production’s database. For this work, we extracted from the database one less complex dataset and one complex dataset and applied data preparation methods to them to create our training datasets. The less complex dataset contains a few classes of defect images of one technology and one inspection step. The complex dataset contains a multitude of classes of defect images of different technologies and inspection steps. We performed and evaluated model-centric experiments on the less complex dataset, which we refer to as the Carinthia dataset, and we performed data-centric experiments on the complex dataset, which we refer to as Madrid dataset. Furthermore, we deployed the best model trained on the complex Madrid dataset, which achieved a validation accuracy of 92.7% for productive usage.