3D reconstruction of SEM images based on point cloud sampling and conditional generative adversarial networks

Area of Science:

• Materials Science
• Nanotechnology
• Computer Science

Background:

• Conventional 2D scanning electron microscopy (SEM) struggles with characterizing complex 3D micro-nano structures.
• Advanced process nodes require more sophisticated metrology techniques for accurate dimensional analysis.

Purpose of the Study:

• To develop a novel 3D reconstruction method for micro-nano structures.
• To overcome the limitations of 2D SEM in characterizing complex 3D features at advanced process nodes.

Main Methods:

• A 3D reconstruction method combining random point cloud sampling and conditional generative adversarial networks was proposed.
• High-quality training datasets were generated using a custom 3D modeling toolkit and Monte Carlo simulation.
• A deep-learning model, P2PSEM1024, was developed for 3D reconstruction from top-view SEM images.

Main Results:

• The P2PSEM1024 model achieved precise 3D reconstruction from top-view SEM images.
• Satisfactory performance was demonstrated across multiple quantitative metrics, with a mean RMSE of 0.94 nm on unseen test samples.
• The method proved effective for complex micro-nano structures.

Conclusions:

• The proposed method offers an effective pathway for accurate and rapid 3D characterization of complex micro-nano structures.
• This deep learning-based approach enhances the capabilities of SEM for nanoscale metrology.