Automated Defect Classification and Metrology via Frequency-Domain Convolutional Neural Networks on Terahertz Scanning Systems for Advanced Wafer Surface Inspection

**Abstract:** This paper introduces a novel approach to automated defect classification and metrology within advanced wafer surface inspection utilizing Terahertz (THz) scanning systems. Current inspection techniques often struggle with accurate identification and measurement of subtle defects, particularly those affecting ultra-thin films and nanoscale features. We propose a Frequency-Domain Convolutional Neural Network (FD-CNN) architecture combined with a sophisticated pre-processing pipeline to extract precise metrological data directly from raw THz interferometry data. This approach leverages the inherent wavelength multiplexing capabilities of THz waves to simultaneously extract both spatial and spectral information, achieving significantly improved accuracy and resolution compared to traditional spatial-domain CNN-based methods and traditional metrology approaches. This method delivers immediate, quantifiable improvements in yield prediction and defect characterization, impacting semiconductor manufacturing processes directly.