Through silicon vias (TSVs) are a key enabling technology for interconnection and realization of complex three-dimensional integrated circuit (3D-IC) components. In order to perform failure analysis without the need of destructive sample preparation, x-ray microscopy (XRM) is a rising method of analyzing the internal structure of samples. However, there is still a lack of evaluated scan recipes or best practices regarding XRM parameter settings for the study of TSVs in the current state of literature. There is also an increased interest in automated machine learning and deep learning approaches for qualitative and quantitative inspection processes in recent years. Especially deep learning based object detection is a well-known methodology for fast detection and classification capable of working with large volumetric XRM datasets. Therefore, a combined XRM and deep learning object detection workflow for automatic micrometer accurate defect location on liner-TSVs was developed throughout this work. Two measurement setups including detailed information about the used parameters for either full IC device scan or detailed TSV scan were introduced. Both are able to depict delamination defects and finer structures in TSVs with either a low or high resolution. The combination of a 0.4 objective with a beam voltage of 40 kV proved to be a good combination for achieving optimal imaging contrast for the full-device scan. However, detailed TSV scans have demonstrated that the use of a 20 objective along with a beam voltage of 140 kV significantly improves image quality. A database with 30,000 objects was created for automated data analysis, so that a well-established object recognition method for automated defect analysis could be integrated into the process analysis. This RetinaNet-based object detection method achieves a very strong average precision of 0.94. It supports the detection of erroneous TSVs in both top view and side view, so that defects can be detected at different depths. Consequently, the proposed workflow can be used for failure analysis, quality control or process optimization in R&D environments.
