Silicon Carbide, more specifically, its polytype 4H-SiC, has emerged as a compelling choice of semiconductor material well-suited to producing chips for power electronics, owing to its unique properties such as wide bandgap and high thermal conductivity. One of the primary challenges that plague the general, widespread adoption of Silicon Carbide is its susceptibility to defect formation during the boule-growth phase. This is primarily due to its more complex growth process involving higher temperatures, as compared to Silicon. As such, the study of defects is of great value, both in scientific and industrial contexts.

KOH-etching is a cost-effective and reliable method of visualizing defects present in a SiC wafer. In this work, we aim to quantify the general defect density as well as the defect density categorized by type using optical-microscopy images corresponding to a 4H-SiC wafer treated with KOH etching. This is achieved by using a UNet model for binary semantic segmentation of defects.

Simultaneously, we explore the critical problem of deep learning models requiring an immense amount of labelled instances for decent accuracy metrics. To this end, we employ an intra-dataset transfer learning pipeline wherein we initially pre-train our UNet model on the original image dataset coupled with weak-masks, obtained using classical image processing. Then, we fine-tune the pre-trained model on a limited set of images from the same dataset whose masks were manually annotated. Post obtaining segmentation masks for the entire wafer, we extract isolated defect instances and train a simple CNN-based classification model using semi-supervised learning with a limited number of labelled defect instances. Finally, we generate composite-views showing the location of defects throughout the wafer, the general defect density and the defect density distribution by type (TED, TSD and BPD).

Through our work, we demonstrate that weak masks offer a promising approach for training deep learning models with minimal manual annotation, validating this through the use case of defect segmentation for a KOH-etched 4H-SiC wafer.