Abdul Qadeer Rehan - Final.pdf

With decreasing feature dimensions, increasing layout complexity, and greater material variations, undetected semiconductor patterning defects can have both technological and financial consequences in nanoelectronics fabrication. Optical methods are ideal for the fast, non-destructive identification of defect locations, but separating defect-based signals from measurement noise is a daunting task. Data-driven Machine Learning has been successfully applied to classifying simulated images,1 and our project has optimized a dataset of experimental images collected at 193 nm wavelength for enhanced processing using supervised neural networks. Images of intentional defect arrays have been obtained using a scatterfield microscope and processed here to enhance the training of convolutional neural networks. Initial results are presented showing the successful binary classification of defect and no-defect examples with discussion of the imbalanced costs due to false identifications. 1) Mark-Alexander Henn, Hui Zhou, Richard M. Silver, Bryan M. Barnes, "Applications of machine learning at the limits of form-dependent scattering for defect metrology," Proc. SPIE 10959, Metrology, Inspection, and Process Control for Microlithography XXXIII, 109590Z (26 March 2019); doi: 10.1117/12.2517285