AI-enabled cross-domain image analysis for high-NA via defect quantification in random logic via arrays

Abstract

High numerical aperture (NA) extreme ultraviolet (EUV) lithography, with NA values of 0.55, represents a critical advancement for sustaining Moore’s law trajectory into sub-3nm technology nodes. By enabling improved resolution and maintaining single-exposure patterning for critical layers, high-NA EUV offers substantial benefits including reduced process complexity, lower cost-per-layer, and enhanced pattern fidelity compared to multi-patterning schemes. These advantages are essential for manufacturing increasingly complex logic designs with aggressive pitch scaling and improved electrical performance. At the same time, the smaller pitches enabled by the improved resolution require unprecedented precision in defect detection and characterization. Traditional inspection methodologies struggle with the scale and complexity of defect analysis required for high-volume manufacturing qualification. In this work, we address these challenges through comprehensive analysis of random logic via arrays printed at high-NA, where missing via defects represent a critical yield indicator. Large field-of-view scanning electron microscopy (SEM) and voltage contrast SEM imaging were employed to capture entire via arrays; however, these techniques introduced significant geometric distortions, intensity variations, and imaging artifacts that complicated defect analysis. To overcome these limitations, we developed a novel style transfer deep learning approach that transforms SEM images into the GDS design domain, creating pseudo-GDS representations that preserve critical structural features while tolerating imaging artifacts. This domain transformation enables robust alignment between measurement and design data using Fast Fourier Transform (FFT) phase correlation algorithms, achieving high-precision registration despite inherent imaging challenges. Our approach enables rapid, high-accuracy quantification of missing via defects through direct comparison of aligned pseudo-GDS and original design data. Furthermore, we analysed the spatial distribution of missing defects, identifying design-specific hotspots and correlating missing via patterns with particular geometric configurations. This comprehensive approach provides critical insights into high-NA patterning performance and establishes a robust framework for automated defect analysis, offering significant advantages for process optimization and yield enhancement in next-generation lithography nodes.