Spectrum-to-Structure Hybrid Metrology: SEM/TEM-Guided Machine Learning for 3D Reconstruction from X-ray Scattering Signals

Abstract

As three-dimensional (3D) memory and logic device architectures continue to increase in complexity, X-ray scattering–based inline metrology has become an essential technique due to its inherent sensitivity to high-aspect-ratio and deeply buried structures. Nevertheless, conventional X-ray structure reconstruction approaches remain fundamentally limited. Physics-based scattering models rely on simplified geometric parameterization, while purely data-driven machine learning approaches often lack direct linkage to physically verified device structures. Consequently, realistic structural variability observed in high-volume manufacturing is not sufficiently captured.

X-ray scatterometry has been widely adopted for the characterization of advanced three-dimensional semiconductor devices. Fan et al. demonstrated that physics-based X-ray scattering models can be used to extract detailed three-dimensional profiles of 3D NAND pillar structures [2]. Suenaga et al. further showed that transmission small-angle X-ray scattering enables precise reconstruction of high-aspect-ratio hole profiles through reciprocal-space analysis [3]. In parallel, machine learning has been introduced to improve robustness and throughput in X-ray-based metrology. Chouaib et al. reported that combining X-ray scatterometry with machine learning can enhance profile extraction performance for advanced memory structures by complementing conventional inverse modeling [1].

However, in most existing approaches, high-resolution reference metrology such as SEM and TEM is used primarily for post-validation or sparse calibration, leaving a gap between reconstructed structures and actual device geometries observed in high-volume manufacturing.

In this work, we address this limitation by proposing a Spectrum-to-Structure hybrid metrology framework that explicitly incorporates SEM and TEM measurements as physics-verified guidance within the machine learning process. When applied to production DRAM and VNAND manufacturing flows, the proposed approach achieves up to approximately 30% reduction in three-dimensional profile error relative to SEM/TEM references compared with conventional X-ray modeling or machine learning–only approaches, while enabling expanded wafer-level sampling.