Defect analysis of silicon carbide wafers using optical measurement employing a newly developed algorithm

Abstract

Defect analysis is essential in the production of silicon carbide (SiC) semiconductor wafers. Crystallographic imperfections formed at the nanoscale during crystal growth can develop into macroscopic defects that lower manufacturing yield, increase costs, and affect device reliability. Optical inspection methods, such as Nomarski differential interference contrast microscopy and ultraviolet photoluminescence imaging, are widely used for defect detection. However, existing automated systems use proprietary algorithms that lack transparency, flexibility, and often misclassify defects or inaccurately define their boundaries. This thesis develops a new, transparent algorithm for defect detection and classification based on optical measurement data. The approach combines deep learning for object detection with rule-based refinement to allow more flexibility in adjusting the classification criteria. A detailed theoretical section reviews SiC properties, manufacturing processes, defect types, and optical inspection methods. This knowledge is critical because understanding the origin and behavior of defects directly influences the algorithm design and analysis strategy. The practical part describes the creation and training of the deep learning model using mapping images from Nomarski and photoluminescence inspections. The new algorithm is tested against a commercial solution, showing similar accuracy for identifying critical "killer" defects, better detection of carrot defects, and more precise defect boundaries. The developed method provides a strong foundation for future improvements and aims to be integrated into production lines to enhance quality control in SiC wafer manufacturing.