Silicon carbide (SiC) is a wide band-gap semiconductor that has gained popularity in various industries due to its unique properties, including higher voltage handling capacity and lower switching losses. This makes SiC an ideal candidate for use in high-performance power devices, which could play a significant role in energy efficiency, reduced power consumption, and lower carbon emissions. A key to utilize this high performance in power electronics is a low density of defects such as micropipes (MP), basal plane dislocations (BPD), threaded edge dislocations (TED), and threaded screw dislocations (TSD) in manufactured single crystals.

One of the most common ways to reveal these defects is molten KOH etching. Counting and characterization of these defects, however, still poses a challenge, as size and shape of the corresponding etch pits vary depending on etching duration, etching temperature, and doping level of the crystal. Both MPs and TSDs are screw dislocations. MPs have a relatively large burgers vector. Since the energy of a dislocation is proportional to the square of its burgers vector, the MP reduces its energy by removing the dislocation core resulting in the creation of a hollow tube. The geometry of the MP, therefore, causes a bottomless etch pit after KOH etching. TSDs on the other hand develop regular etch pits with bottoms, as they only possess a comparably small burgers vector. TEDs manifest in, depending on the etching conditions, smaller round etch pits, and BPDs as elliptically shaped “comet” like etch pits. In some cases, dislocations can exhibit a mixed edge and screw character, making their characterization more complex.

As part of machine learning, deep learning has already been successfully utilized in material science to detect features of interest in microscopic images. Deep convolutional neural networks (DCNN) have been proven to be very capable in computer vision tasks such as image classification, object detection, and instance segmentation. This often even superhuman-like performance can be achieved due to the DCNN mimicking to some degree the function of the human visual cortex. By subsequent application of filters and shrinking of the image (“pooling”), it is possible to detect features of different complexity.

In this work we show the use of Mask R-CNN, a region based CNN, to automatically detect and characterize the defects in SiC single crystals etched by molten KOH. Instance segmentation is implemented trough transfer learning using the open-source deep learning library PyTorch. The main challenge for the network is posed by the correct distinction between overlapping defects, especially in Regions with high defect densities. The focus lies on achieving a high confidence score for detected defects, to minimize reliance on manual inspection. Additionally, it will be possible to identify sections within the observed images that fail to meet predetermined criteria for defect density or confidence, and flag them for manual review.