Deep reinforcement learning for data-driven adaptive scanning in ptychography.

Area of Science:

• Materials Science
• Microscopy
• Computational Imaging

Background:

• Electron ptychography requires high electron doses for accurate specimen reconstruction.
• Minimizing radiation exposure is crucial for sensitive or dynamic samples.
• Existing low-dose methods often compromise resolution or data completeness.

Purpose of the Study:

• To develop a novel adaptive scanning method for electron ptychography.
• To reduce the required electron dose while maintaining high-quality reconstructions.
• To leverage deep learning and reinforcement learning for dose optimization.

Main Methods:

• Implementing an adaptive scanning strategy based on specimen importance.
• Utilizing a deep learning model trained with reinforcement learning.
• Incorporating prior knowledge of specimen structure from training datasets.

Main Results:

• The adaptive scanning method significantly lowers the electron dose needed for reconstruction.
• Achieved superior reconstruction resolution and quality compared to alternative low-dose techniques.
• Demonstrated effective spatial information redundancy in critical specimen regions.

Conclusions:

• Adaptive scanning is a highly effective strategy for low-dose electron ptychography.
• Deep learning-based adaptive scanning offers a promising path for advanced microscopy.
• This method enables high-fidelity imaging with reduced radiation damage.