Automatic defect detection on semiconductor wafer surfaces using deep learning

Solutions

Our team conducted thorough consultations with the client’s team. We proposed developing a backend solution using deep learning technology to identify defect patterns and perform image classification effectively. Our solution development process included:

Data analyzation

Our team analyzed the data to understand the root cause of the problem. We observed the position of defects in the wafer images at various locations which gave us insight into the common defect scenarios.

Preparing the dataset

We carried out the pre-processing of the data and applied the grayscale image segmentation technique. Our approach was to carry out detection as a classification task at the pixel level. Therefore, we created data models that examine background pixels and foreground pixels, which helped us to extract pattern features. However, the client provided us with a limited set of data, and deep learning networks require a large amount of data for training. Moreover, in the case of semiconductor wafers where defects must not be frequent, the number of non-defected images outweighed the defected images. Therefore, to resolve the imbalanced dataset problem and to enlarge the dataset, we carried out image augmentation with Keras Python Library. This helped us to increase the number of samples in our data set using the small data set provided by the client. Further, we divided the enlarged dataset into training and validation data.

Building the neural network model

We decided to use Google’s pre-trained TensorFlow Inception CNN Models instead of creating and training the deep neural nets from scratch to save time and data. Then we fine-tuned the weight parameters using the feature extraction mechanism which included feature creation, feature ranking and dimensionality reduction. We trained the new classification model written in Python on our dataset to identify defect patterns and predict the label of unseen wafer images in the future.

Validation and deployment

We evaluated the trained Python model on the new wafer images to calculate its predictive accuracy. Our solution achieved 99% defect detection accuracy, and we deployed it successfully to production on Nvidia GTX 1060 8GB GPU.