High-throughput rotational laminography

3D inline wafer-level advanced packaging inspection

Three-dimensional integrated circuit (3D IC) packaging offers superior performance and energy efficiency over 2D IC packaging but requires reliable, high-throughput in-line inspection to detect defects like misaligned interconnects and voids. X-ray computed laminography (CL) is an effective technique for high-throughput volumetric imaging of large, planar samples, which is well-suited for wafer-level 3D IC packaging inspection.

Due to the significantly shorter photon path, rotational laminography can drastically reduce scan time compared to standard circular trajectory cone beam CT.

However, due to stringent time requirements demanded by inline semiconductor manufacturing inspection applications, which typically require completing inspection per region-of-interest (ROI) within minutes, the shot-noise limited scans tend to resulting in reconstructions that are too noisy for robust downstream metrology and defect analysis. To meet this challenge, we developed a self-supervised workflow to create a generalist deep learning model for high-resolution (~2.3 gigavoxels per scan) volumetric CT image denoising by pre-training the model on a large number of semiconductor packaging datasets. We showcase that the generalist model has superior zero-shot denoising capability on a varity of semiconductor packaging samples.

Denoised reconstruction on samples containing A. memory stack, B. TSV arrays with programmed void from IMEC, and C. a semiconductor package with C4 bump cracks. Extracted from doi.org/10.1093/mam/ozaf048.984

In addition, because computed laminography geometry violates Tuy’s data sufficiency condition, reconstructions typically exhibit butterfly-like cone-beam artifacts. To address this, we present a self-supervised deep image restoration workflow that produces noise-free, artifact-free volumetric reconstructions. The core of our proposed pipeline is an unbiased progressive artifact removal algorithm designed to suppress laminographic artifacts. These artifact-corrected volumes are then used to train a deep image restoration network, improving image quality without increasing inference time. We demonstrate the efficacy of this method on a variety of samples scanned with an in-house prototype system.

Volumetric reconstruction of a dynamic random-access memory (DRAM) sample. Noise reduction and artifact removal are applied simultaneously to improve image quality without increasing inference time. Extracted from doi.org/10.1117/12.3028278

Further reading

Xu, Shiqi, et al. “Self-supervised deep image restoration for x-ray computed laminographic tomography.” Developments in X-Ray Tomography XV. Vol. 13152. SPIE, 2024.

Guo, Zijing, Xu, Shiqi, et al. “The Role of Pretraining in High-Throughput Laminography Restoration.” Microscopy and Microanalysis (2025): ozaf048-984.