Deep trenches (DT) in silicon are a source of mechanical stress that can eventually lead to the formation of crystal defects. Herein this work, it is shown that room-temperature micro-photoluminescence imaging can be applied for a nondestructive in-line monitoring of buried defects generated by DT. With the training of a convolutional neural network for automatic image segmentation, image analysis is automatized, allowing the inspection of large areas. The analysis of DT structures with various spacing and multiplicity reveals that the DT mutual distance is the main factor governing defects generation. Finally, a step-by-step analysis allows to detect the step responsible for defect generation and a further defect density increase along the process flow.
