Deep-Learning Electron Diffractive Imaging.

Area of Science:

• Materials Science
• Physics
• Data Science

Background:

• Coherent electron diffraction imaging is crucial for nanoscale material analysis.
• Conventional methods like ptychography require iterative reconstruction, which can be time-consuming.
• Achieving sub-angstrom resolution is a key goal for atomic-level characterization.

Purpose of the Study:

• To develop a deep-learning approach for rapid and high-resolution phase retrieval in coherent electron diffraction imaging.
• To demonstrate the efficacy of convolutional neural networks (CNNs) trained on simulated data for experimental imaging.
• To enable real-time atomic imaging by replacing traditional iterative algorithms.

Main Methods:

• Utilized convolutional neural networks (CNNs) trained exclusively on simulated electron diffraction data.
• Applied trained CNNs to experimentally acquired diffraction patterns from various materials (hBN, graphene, gold nanoparticle).
• Developed CNNs for simultaneous recovery of the probe function from experimental data.

Main Results:

• Achieved sub-angstrom resolution (0.70–0.55 Å) comparable to conventional ptychography.
• Successfully reconstructed phase images from experimental diffraction patterns.
• Demonstrated CNNs' capability to recover the probe function, enhancing imaging accuracy.

Conclusions:

• Deep learning, specifically CNNs, offers a powerful alternative to iterative algorithms for coherent electron diffraction imaging.
• The developed method enables high-resolution, real-time atomic imaging from electron diffraction data.
• This technique holds significant potential for applications in physical and biological sciences requiring atomic-level resolution.