Simulation-driven 3D line roughness reconstruction: A proof-of-concept deep learning framework for SEM metrology

Accurate 3D topographical metrology of nanoscale semiconductor line patterns is essential for advanced process control. However, inferring 3D depth from single top-down Scanning Electron Microscopy (SEM) images is an inherently ill-posed inverse problem due to the fundamental loss of spatial information. In this paper, we propose a proof-of-concept, physics-driven multi-channel 1D deep learning framework designed to infer 3D line roughness and surface morphologies using models trained exclusively on Monte Carlo electron yield simulations. By integrating a localized multi-line spatial context window into an efficient 1D U-Net architecture, the proposed computational approach successfully mitigates depth ambiguity while bridging the gap between 1D computational efficiency and 2D spatial awareness. Quantitative evaluations on synthetic SEM datasets demonstrate the model's extrapolation robustness across out-of-distribution roughness parameters. Furthermore, frequency-domain analysis via Power Spectral Density (PSD) confirms that our framework effectively mitigates over-smoothing artifacts, faithfully preserving the essential high-frequency stochastic roughness signatures. Finally, we demonstrate the qualitative feasibility of the framework by applying it to experimental SEM images. While generating physically plausible 3D surface profiles, we identify and discuss the synthetic-to-experimental domain shift caused by unmodeled physical phenomena, establishing a clear pathway for future calibration strategies. Ultimately, this research establishes a scalable, non-destructive, and high-throughput computational foundation for 3D morphological profiling, paving the way toward fully quantitative in-line 3D line roughness metrology in next-generation semiconductor manufacturing.