Reducing defects in SiC is key to improving device quality, and wafer scale techniques for identification of defects is critical for their mitigation. Although methods for imaging defects are well established [1-3], large image datasets are generated for wafer scale mapping. In order to classify each defect, creating an automated process is difficult as it requires a complex mathematical routine with improvement to the model requiring extensive research. Success using traditional programming techniques have been shown to detect basal plane dislocations in Ultra-Violet Photoluminescence (UVPL) images using their intense luminescence spectra [4]. However, some defects such as threading dislocations often show smaller, less consistent changes in intensity with a lower signal to noise ratio. In this work, we utilize advanced Machine Learning models to precisely detect and delineate dislocations in whole wafer images using a selective training dataset that also incorporates data where the model produces incorrect predictions. Thus, research into its effectiveness at defect mapping is valuable. This research focuses on studying the effectiveness of creating a machine learning model to count dislocations on SiC using UVPL imaging and x-ray topography (XRT) whole-wafer maps. The models were trained using a convolutional neural network (CNN) trained and tested with TensorFlow [5]. The training data consists of imperfect, manually classified defects on a small dataset (less than 25 MB). The CNN is used to pinpoint the probability of a defect being present with an example shown in Fig 1. The results of this study show that a CNN is a useful tool for mapping defects in SiC, as it can accurately locate dislocations defects and stacking faults and polytype inclusions with >95% accuracy using both UVPL and XRT images. The model was used to efficiently map the locations of the defects on a full 150 mm wafer (Fig 2), which allows for study of defect distribution, its formation in addition to providing allowing for direct study of a defects impact on device performance. Additionally, this work enabled identifying defects with higher precession than simple manual classification since our CNN model was able to detect defects missed during manual classification as highlighted in Fig 3. Thirdly, this project demonstrated that large datasets are not necessarily required to train a ML model, as only ~1500 defects in 10 images were needed for the training of the CNN model. Details of the CNN model and algorithms used for the classification of defects in the UVPL and XRT images will be presented. Work at the U.S. Naval Research Laboratory was funded by the Office of Naval Research. [1] Stahlbush, Robert E., Liu, Kendrick X., Zhang, Q., and Sumakeris, Joseph J., Materials Science Forum 556-557 (2007), pp. 295-298. [2] Berwian, Patrick, Kaminzky, Daniel, Rosshirt, Katharina, Kallinger, Birgit, Friedrich, Jochen, Oppel, Steffen, Schneider, Adria et al., , Solid State Phenomena 242 (2015), pp. 484-489.. [3] Chen, Po-Chih, Miao, Wen-Chien, Ahmed, Tanveer, Pan, Yi-Yu, Lin, Chun-Liang, Chen, Shih-Chen, Kuo, Hao-Chung, Tsui, Bing-Yue, et al. , Nanoscale Research Letters 17 (2022), pp. 30. [4] Harada, Shunta, Tsujimori, Kota, and Matsushita, Yosuke, " Automatic Detection of Basal Plane Dislocations in a 150-mm SiC Epitaxial Wafer by Photoluminescence Imaging and Template-matching Algorithm” Journal of Electronic Materials 51 (2022), pp. 243-248. [5] Abadi, Martin, Barham, Paul, Chen, Jianmin, Chen, Zhifeng, Davis, Andy, Dean, Jeffrey, Devin, Matthieu, Ghemawat, Sanjay, Irvi et al., (2016), pp. 44.