edfas.org 9 ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 26 NO. 3 The U-Net 2D model with Dice loss is used for detection as it outperformed the other models tested. As a 2D model, it is applied individually to each slice of the test set; then these predicted slices are combined to create a 3D representation using Dragonfly software. The data showcased in the graph represent the 3D morphology obtained from the model’s predictions. The data by a human operator is obtained by manually marking the void area. The volume calculation is done by applying connected component analysis to the 3D prediction of the model. This comparison of human operator identification with deep learning model prediction allows evaluating how the deep learning segmentation model performed compared to human operators in predicting both the total number and volume of voids. The U-Net 2D model detected 71 voids, while the human operator identified 77 voids in the same sample. However as mentioned earlier, it’s important to acknowledge that human labeling may contain some degree of errors. Closer examination found that the voids not detected by the U-Net 2D model but present in the manual label data are primarily small voids falling within the range of 0.005e-8 – 2.0e-4 mm3 (in the manually labeled data). To assess the overall performance of the model, the overall volume of detected voids by the model and by the human operator were compared. The U-Net 2D model predicted a total volume of 2.05e-3 mm³, while the total volume of all voids identified by the human operator is 2.01e-3 mm³. This means the U-Net 2D model produces approximately +2% of the error rate compared to the human operator for the particular sample. To further validate the accuracy of void detection on an individual basis, five random voids were selected from the sample and their volumes were calculated. In comparison to the volumes of voids detected by the human operator, a mean error rate of approximately +3% was found for these voids. This side-by-side comparison offers valuable insights into the accuracy of void detection by the deep learning model, providing context for understanding its performance relative to human operators. CONCLUSION In this article, a deep learning-based nondestructive approach for void segmentation in BGA solder balls using 3D x-ray microscopy was presented. Various deep learning segmentation architectures with different loss functions were evaluated, including U-Net 3D. The results showed that U-Net 2D and U-Net++ outperformed the other models when combined with Dice loss and categorical cross entropy (CCE) loss respectively, achieving an intersection over union (IoU) score of 88% and an F1 score of 92%. Additionally, the comparison of the overall volume of detected voids by the deep learning model (U-Net 2D model with Dice loss) provided an error rate of +2% compared to the human operator. However, it is important to note that despite an IoU score of 88%, the models may encounter challenges in maintaining consistency across each slice of 3D scans, particularly in the presence of cracks within the solder balls. These cracks can also have a negative impact on the performance of electronic devices. Therefore, it becomes important to segment these cracks as well. Future work will address the current limitation arising by the presence of the cracks and focus on developing a model that can segment cracks independently, in addition to void segmentation in the XRM data of solder balls. Furthermore, the limitations of current 2D-based methodologies will be overcome by improving the current model for automated 3D void detection for precisely locating voids within solder balls and also providing a more accurate 3D analysis of these voids. This could eventually replace the current 2D bases IPC standard as 3D analysis provides a more comprehensive understanding of the structure and quality of solder balls. Overall, this method will contribute to the improvement of electronic device manufacturing processes. ACKNOWLEDGMENTS This work was funded by the Federal Ministry of Education and Research, Germany, in the scope of the NextGenNDT research project (grant no. 13FH566KX9). REFERENCES 1. K. Rathod, et al.: “Semantic Segmentation for Non-destructive Defect Detection in 3D X-ray Microscopy Data of Solder Joints,” 2023, unpublished. 2. PCB Assembly: Ball Grid Array Explained, Garner Osborne, 2024, www.garnerosborne.co.uk/insights/pcb-assembly-ball-grid- array-explained. 3. D.N. Vieira, et al.: “Prediction of Solder Joint Reliability with Applied Acrylic Conformal Coating,” J. Electron. Mater., 2022, Vol. 51, p. 273–283, doi.org/10.1007/s11664-021-09232-9. 4. M. Yunus, et al.: “Effect of Voids on the Reliability of BGA/CSP Solder Joints,” Micron. Reliab., 2003, Vol. 43, p. 2077-2086, doi.org/10.1016/ S0026-2714(03)00124-0. 5. K.J. Puttlitz and K.A. Stalter: Handbook of Lead-free Solder Technology for Microelectronic Assemblies, 2004. 6. A. Nee (ed.): Handbook of Manufacturing Engineering and Technology, 2020, doi.org/10.1007/978-1-4471-4976-7. 7. S.B. Dunkerton and J.M. Goward: “A Review of Ball Grid Arrays for Electronic Assembly,” 1998, www.twi-global.com/technical-knowledge/published-papers/a-review-of-ball-grid-arrays-for-electronicassembly-september-1998. 8. IPC-A-610: Acceptability of Electronic Assemblies, 2020, IPC.
RkJQdWJsaXNoZXIy MTYyMzk3NQ==