| Abstract Scope |
As the global demand for microelectronics continues to surge, methods must keep pace to detect small but critical manufacturing defects with high accuracy and throughput. We propose the use of coherent diffraction coupled with unsupervised machine learning techniques to learn the subtle changes in diffraction intensity that indicate a nanometer scale defect within a multi-micron imaged region. To complement standard unsupervised architectures such as the autoencoder, we propose a novel semi-supervised technique that, given a set of imperfectly labelled training data, learns to improve upon the input labels. This allows a sequence of increasingly powerful discriminator models to be trained to amplify the initial defect detection ability of, for example, an autoencoder alone. Key challenges, such as noisy diffraction, variable sample-beam positions and, most of all, the infrequency of defects within the diffraction images must be directly addressed for robust and accurate defect detection. |