Multi-channel discrimination of crystallographic defects in 4H-SiC epitaxial wafers using optical inspection, photoluminescence and X-ray topography

Accurate identification of crystallographic defects in 4H-SiC epitaxial wafers is essential for ensuring the yield and reliability of SiC power devices. Although a variety of inspection techniques have been developed, defect visibility varies substantially among characterization methods, making reliable automated defect classification challenging. In this study, a multi-channel defect identification framework that integrates optical inspection, photoluminescence, and X-ray topography using a convolutional neural network was developed. Images acquired by the three techniques were spatially registered and combined into multi-channel datasets for analysis. A Faster R-CNN was trained using annotated data collected from 18 epitaxial wafers. The trained model was evaluated using an independent test wafer containing reliability-relevant defects, including micropipes, basal plane dislocations, threading screw dislocations, stacking faults, stacking-fault complexes, polytype inclusions, and particle inclusions. Compared with models trained on individual inspection modalities, the multi-channel model exhibited improved classification performance through the integration of complementary morphological, luminescence, and crystallographic information.