Generative-model-based null-space iterative reconstruction for atomic electron tomography with sparse data

Area of Science:

• Materials Science
• Electron Microscopy
• Computational Imaging

Background:

• Atomic electron tomography (AET) is crucial for 3D atomic structure determination.
• Conventional AET requires extensive data acquisition (high dose, wide angular range), limiting its practical use.
• Existing methods struggle with dose and time constraints, hindering the analysis of delicate or small-scale materials.

Purpose of the Study:

• To develop a novel algorithm for low-dose, fast, and tilt-constrained AET.
• To overcome the limitations of conventional AET regarding dose, acquisition time, and angular range.
• To enable high-resolution 3D atomic structure determination of challenging materials.

Main Methods:

• Introduced Null-Space Iterative Reconstruction (NSIRE), an interpretable and universal algorithm.
• NSIRE employs an unsupervised diffusion model to compute null-space solutions of projection equations.
• The algorithm integrates projection constraints with atomic potential priors for tomogram generation.

Main Results:

• NSIRE successfully resolves complex materials using sparse projections (6-12°) or small tilt ranges (±29°) without retraining.
• Achieved high-resolution 3D atomic structures of a 4-nm Pt nanoparticle and a 3-nm PtCo nanoalloy (smallest resolved to date).
• Demonstrated sub-20-pm root-mean-square displacement and high projection consistency, surpassing conventional AET limitations.

Conclusions:

• NSIRE significantly advances AET by enabling low-dose, fast, and tilt-constrained 3D atomic structure determination.
• The method overcomes critical scale, dose, and time limitations, broadening AET's applicability.
• NSIRE provides a powerful tool for analyzing nanoscale materials with unprecedented detail and efficiency.