Three-dimensional (3D) integrated circuits (ICs) utilizing through-silicon vias (TSVs) are crucial for enhancing chip performance and miniaturization. The Cu-filling process in TSVs involves several steps, including etching, deposition of insulating/blocking layers, Cu filling, CMP, and annealing. However, defects arising from CMP and annealing, such as residual stress, extrusion, cracking, delamination, and Cu leaks, significantly impact reliability. These defects often result from the CTE mismatch between the silicon substrate and Cu filler. Current defect characterization methods are often destructive and time-consuming. This research focuses on developing a nondestructive, real-time method for characterizing defects in TSV-Cu structures arising from CMP and annealing processes, addressing the need for high-accuracy, nanoscale control of Cu filling height during manufacturing.

Existing literature extensively covers TSV manufacturing processes and the impact of defects on 3D IC reliability. Studies have investigated Cu-filling protrusion and CMP dishing, highlighting the critical role of process parameters such as CMP speed, annealing time, and temperature. Insufficient or excessive polishing can lead to leaks and shorts, while inappropriate annealing temperatures and times can cause residual stress, delamination, and cracking. However, real-time, nondestructive characterization methods for Cu protrusions and dishing remain limited, particularly for nanoscale precision required in the z-direction of TSV-Cu structures. This gap in existing methods motivates the development of the proposed ellipsometry and deep learning approach.

This study employed Mueller matrix spectroscopic ellipsometry, combined with rigorous coupled-wave analysis (RCWA), to characterize TSV-Cu structures. An optical model was developed based on atomic force microscopy (AFM) images, representing the TSV-Cu structure with layers of Ta, SiO2, and Cu. The RCWA algorithm calculated the reflection electric field at various wavelengths (400-1000 nm), from which the Mueller matrices were derived. Deep learning was applied to analyze the Mueller matrix datasets. Initially, a single MME (m12) was used for defect classification, followed by a multichannel approach using seven effective MMEs for quantitative height measurement. The datasets included simulated data with varying Cu filling heights (Hu) representing defect-free, defect-OD, and defect-P conditions, along with added noise to simulate real-world conditions. The performance of both single-channel and multichannel deep learning models was evaluated using metrics such as cross-entropy loss, accuracy, and t-distributed stochastic neighbor embedding (t-SNE) for visualization.

The single-channel deep learning model using m12 achieved a high accuracy rate (99.94%) in classifying the three defect types (defect-free, defect-OD, and defect-P). The multichannel approach using seven MMEs demonstrated remarkable accuracy (98.92%) in quantifying the Cu filling height variation (Hu) with a resolution of 1 nm. The method's robustness was validated by testing its performance under various noise levels (up to 30% random noise), demonstrating consistent accuracy. The impact of different TSV morphologies (aspect ratio, top-bottom critical dimension difference) and measurement conditions (incident light azimuth angle) were assessed, showing minimal influence on the accuracy of the proposed method.

The results demonstrate the effectiveness of combining Mueller matrix spectroscopic ellipsometry and deep learning for nondestructive defect inspection in TSV-Cu structures. The high accuracy and resolution achieved in both defect classification and quantitative height measurement highlight the potential of this approach for real-time process monitoring during TSV manufacturing. The robustness against noise and variations in TSV morphology suggests the method's suitability for industrial applications. This work addresses the long-standing challenge of rapid and precise characterization of nanoscale defects in TSV-Cu, paving the way for improved reliability and yield in 3D IC fabrication.

This study successfully developed a rapid, nondestructive method for monitoring annealing and CMP processes in TSV-Cu structures. The combination of Mueller matrix ellipsometry and deep learning provided high accuracy in defect classification and precise quantification of Cu filling height. The robustness of the method under noisy conditions makes it suitable for real-world applications. Future work could focus on extending this technique to other materials and defect types in advanced packaging technologies.

The study's conclusions are based on a specific TSV-Cu structure and optical model. The generalizability to other TSV designs and materials requires further investigation. The accuracy of the method relies on the accuracy of the optical model and the quality of the ellipsometry measurements. While the method demonstrated robustness to noise, extreme noise levels might still affect the accuracy.