As semiconductor scaling transitions toward 3D integration, hybrid bonding has become a critical process for high-performance device stacking. However, severe wafer warpage and nanometer-scale alignment requirements present significant metrology challenges. This study proposes an integrated framework to ensure high-stability, high-efficiency overlay metrology by unifying robust algorithm, topography insights, and AI-driven compensation. We first introduce a Symmetry-based Grayscale Correlation (SGC) algorithm that significantly enhances overlay precision and reduces measurement uncertainty while maintaining robustness against process-induced variations. To investigate the root causes of misalignment, an in-house fringe reflection (FR) module captures 3D wafer topography, linking physical distortion to overlay fingerprints. Overlay model analysis is then performed to identify primary error sources, offering critical insights for process optimization. Finally, to address the pervasive defocus issues in high-volume manufacturing, we deploy a convolutional neural network (CNN) that restores blurred, out-of-focus images into sharp counterparts. This enables accurate measurement without mechanical refocusing, drastically improving throughput.
Paper
Full text
Advanced overlay metrology for wafer-to-wafer hybrid bonding in 3D integration
Semantic Scholar · Engineering · 2026
Abstract
As semiconductor scaling transitions toward 3D integration, hybrid bonding has become a critical process for high-performance device stacking. However, severe wafer warpage and nanometer-scale alignment requirements present significant metrology challenges. This study proposes an integrated framework to ensure high-stability, high-efficiency overlay metrology by unifying robust algorithm, topography insights, and AI-driven compensation. We first introduce a Symmetry-based Grayscale Correlation (SGC) algorithm that significantly enhances overlay precision and reduces measurement uncertainty while maintaining robustness against process-induced variations. To investigate the root causes of misalignment, an in-house fringe reflection (FR) module captures 3D wafer topography, linking physical distortion to overlay fingerprints. Overlay model analysis is then performed to identify primary error sources, offering critical insights for process optimization. Finally, to address the pervasive defocus issues in high-volume manufacturing, we deploy a convolutional neural network (CNN) that restores blurred, out-of-focus images into sharp counterparts. This enables accurate measurement without mechanical refocusing, drastically improving throughput.