Instance segmentation pipeline for etch pit detection and prismatic slip characterization on silicon carbide substrates

Silicon carbide, a wide-bandgap semiconductor, is well-suited for high-frequency and high-temperature applications. The performance of devices fabricated from silicon carbide is negatively affected by extended defects such as dislocations. Therefore, it is crucial to measure and characterize these defects to ensure the reliability of the devices. The most commonly employed method for visualizing dislocations on the surface is defect-selective etching using molten potassium hydroxide. To detect and characterize various types of etch pits on the surface of single crystal substrates, a Mask Region-based Convolutional Neural Network (Mask R-CNN) is utilized. An inference pipeline based on slicing the images into overlapping tiles assures that large images of whole wafers can be processed. Masks generated by the network are further used to extract the dislocation line direction of basal plane dislocations, enabling the characterization of the extent of prismatic slip. The method performs well, even for overlapping etch pits, and can easily be integrated into existing characterization workflows, provided that segmentation masks of the etch pits are available.