In advanced semiconductor manufacturing, inspection speed is a crucial factor for manufacturers looking for a competitive edge—it can directly affect time-to-market, yield, and process control. Yet the structures that need to be inspected are becoming smaller, denser, and more complex. Through-silicon vias (TSVs) illustrate this challenge well: millions of nearly identical features may be packed onto a single wafer, while the critical defects that matter (such as voids) can be only micrometers in size.

Conventional computed tomography (CT), or even laminography (CL) can provide the required three-dimensional detail, but acquiring a volume with sufficient resolution, at a dose that will not damage the component, can often take too long for realistic inline use. To close that gap, our application team in Taiwan tested a workflow that combines continuous-rotation laminography with deep-learning image denoising and automated defect analysis. The result is a four-minute 3D inspection cycle that retains the structural information needed for reliable TSV void detection. Here is how the approach works—and why the combination matters.

Laminography can be performed using two main scanning strategies. In stop-and-go scanning, rotation pauses at every angular position while multiple frames are acquired. This improves the SNR of each projection, but the repeated stopping, settling, and restarting adds substantial mechanical overhead.

Continuous-rotation laminography keeps the sample moving while motion and acquisition remain synchronized. Removing the start-stop cycle sharply reduces total scan time. The trade-off is that motion blur and aliasing become possible, so precise motion control and projection synchronization are essential. With enough angular projections, however, the system can reconstruct TSV volumes accurately without introducing significant aliasing artifacts.

Faster, single-frame, low-dose acquisition naturally produces data with a lower SNR. The second stage addresses this limitation with an in-house deep learning denoising model trained for the specific device design. A high-projection stop-and-go laminography scan of one selected die provides the target reference for training.

Once trained, the model can be applied to newly acquired datasets from other dies of identical design—with no additional retraining required. This design-specific, cross-die deployment strategy supports accurate restoration while keeping inference fast enough for an inline environment.

The hardware and software accelerators can operate independently, but their real value appears when they are used in tandem. Continuous rotation first reduces acquisition time; deep learning then restores the image quality lost through faster, lower-dose capture. This “fast-on-fast” combination delivers high-quality volumetric data in a fraction of the time required by high-projection stop-and-go scanning.

The comparison shows why acquisition speed alone is not enough. Reducing the number of projections in a stop-and-go scan cuts the cycle from 60 minutes to four, but the relative SNR falls to 0.43 and the measured void ratio becomes less consistent. Continuous rotation provides a stronger four-minute baseline, with a relative SNR of 0.69 and a void ratio closer to the high-projection reference.

Adding deep learning denoising changes the outcome more decisively. The relative SNR rises to 1.08, slightly above the high-projection reference value of 1.00, while the measured void ratio reaches 8.72 ± 0.38% compared with 9.03 ± 0.18% for the 60-minute reference scan. The method therefore preserves measurement reliability while reducing acquisition time significantly. 

The interposer TSV samples analysed in this research had already undergone metal filling and chemical mechanical polishing (CMP) and were therefore ready for integration with micro-bumps and redistribution layers (RDLs). At this point in advanced packaging, voids can have a direct impact on downstream reliability, making rapid and dependable inspection especially valuable.

By combining continuous-rotation laminography, design-specific deep learning denoising, and automated segmentation, the team achieved high-quality 3D volumetric inspection within a four-minute cycle. That speed can support real-time process monitoring, faster yield improvement, and earlier detection of systematic process drift. Just as importantly, training the denoising model on a single die and applying it across an identical wafer design minimizes setup overhead. The broader lesson is clear: production-speed 3D inspection does not depend on a single breakthrough, but on coordinated acceleration across acquisition, reconstruction, and analysis.

Acknowledgements: 

This article would like to acknowledge the original authors of this research, Hao-Che Chien and Chien-Yu Kuo. Furthermore, This research would not have been possible without the valuable technical support in system integration, experimental setup, and data acquisition from the application team and R&D colleagues at Comet Solutions Taiwan. Special thanks are extended to the analysis software team for the contributions to the deep learning denoising model training, validation, and deployment. We would also like to acknowledge the collaboration of the process engineering partners who provided the TSV interposer samples and valuable feedback on inline inspection requirements. Their support was essential to the successful completion of this work.