Soldering plays a critical role in connecting semiconductor devices to printed circuit boards (PCBs). The solder joint's quality directly impacts the functionality and reliability of electronic devices because it affects mechanical fixation, electrical contact, and effective device cooling.

One significant challenge faced in soldering is the presence of voids in the solder joints. These voids can negatively affect the electrical and thermal conductivity and may lead to device malfunctions or tilting issues. It can have more detrimental influence on the mechanical stability if it is located in the interface region towards the pad compared to a localisation in the middle of the solder volume. A quantitative analysis of the void distribution is required to enhance the soldering process. 2D inline x-ray projection has been used for this purpose as traditional method but it does not allow to judge the influence of the spatial position of a void with a given size. To address this, advanced non-destructive imaging techniques such as Computed Tomograpy (CT) and X-ray microscopy (XRM) have been applied. However, these imaging techniques produce a vast amount of 3D data. Manual evaluations are carried out to analyse the 3D data, which frequently poses challenges because there are single semiconductor components or e.g. stacked PCBs in the automotive control units with hundreds, sometimes thousands of solder balls and the manual analysis for such large amount of data is labour-intensive and time consuming therefore automated evaluation was needed. Data-driven deep learning methods are emerging as an alternative way for non-destructive 3D image analysis, which is fast and accurate. Escpecially, 3D image segmentation has proven to be a promising approach for defect detection [1].

This research investigates and implements a deep learning-based supervised 3D image segmentation approach for void detection in 3D CT images of PCB solder joints. Dragonfly, a widely used software for 3D data exploration that also offers a no-code and low-code based deep learning training module, is used to implement this approach.The 3D U-Net architecture is implimented for void segmentation and the performance of various models, trained with the different hyperparameters, are compared to achive optimal performance. The implemented supervised 3D image segmentation method has the potential to improve and automate the quality assurance process, resulting in greater efficiency, dependability, and eventuall improved product quality.