Deep learning for sub-ångström-resolution imaging in uncorrected scanning transmission electron microscopy.

Area of Science:

• Materials Science
• Electron Microscopy
• Artificial Intelligence

Background:

• Sub-ångström resolution in electron microscopy has traditionally required complex aberration-corrected scanning transmission electron microscopy (AC-STEM).
• Computational super-resolution techniques offer alternatives but have limitations regarding sample thickness.
• There is a need for accessible methods to achieve high-resolution imaging in uncorrected STEM.

Purpose of the Study:

• To introduce SARDiffuse, a deep-learning diffusion model for enhancing spatial resolution and reducing noise in uncorrected STEM images.
• To demonstrate the capability of SARDiffuse to achieve sub-ångström resolution without aberration correctors.
• To provide a cost-effective solution for high-precision material characterization using conventional electron microscopes.

Main Methods:

• Development of SARDiffuse, a deep-learning diffusion model trained on experimental AC-STEM data.
• Application of SARDiffuse to uncorrected STEM images of various materials.
• Statistical analysis to validate the preservation of atomic positions and artifact correction capabilities.

Main Results:

• SARDiffuse successfully restored high-frequency information in uncorrected STEM images, achieving sub-ångström resolution (<1 Å) in silicon, strontium titanate, and gallium nitride.
• The model reliably preserved atomic positions, confirmed by detailed statistical analysis.
• SARDiffuse effectively mitigated spherical-aberration-induced artifacts and preserved background image information like thickness variations.

Conclusions:

• Deep learning, through models like SARDiffuse, can enable sub-ångström resolution imaging in uncorrected electron microscopes.
• SARDiffuse offers a powerful and cost-effective alternative to AC-STEM for imaging conventional single crystals.
• This approach democratizes high-precision material characterization by leveraging accessible instrumentation.