Accelerating iterative ptychography with an integrated neural network
Arthur R C McCray1,2, Stephanie M Ribet2, Georgios Varnavides2,3
1Department of Materials Science and Engineering, Stanford University, Stanford, California, USA.
Journal of Microscopy
|April 8, 2025
Summary
We developed a neural network (NN) to speed up electron ptychography reconstruction. This method accelerates gradient descent algorithms by boosting low spatial frequencies, improving imaging of nanoparticles.
Area of Science:
- Materials Science
- Physics
- Computational Imaging
Background:
- Electron ptychography offers high-resolution, dose-efficient imaging.
- Iterative reconstruction algorithms are computationally intensive, especially for low spatial frequencies.
- Gradient descent methods in ptychography can exhibit slow convergence.
Purpose of the Study:
- To accelerate gradient descent-based iterative reconstruction in electron ptychography.
- To improve the speed and efficiency of reconstructing low spatial frequencies.
- To integrate neural networks into the iterative reconstruction loop.
Main Methods:
- Developed a neural network (NN) applied within the iterative reconstruction loop.
- The NN operates in Fourier space to selectively enhance low spatial frequencies.
- Employed iterative training to address challenges of NN integration into iterative algorithms.
Main Results:
- Successfully accelerated gradient descent-based iterative ptychography.
- Demonstrated significantly faster reconstruction of gold nanoparticles.
- Validated the method using both simulated and experimental data.
Conclusions:
- The proposed NN-based acceleration method enhances electron ptychography efficiency.
- This approach offers a viable solution for overcoming slow convergence in low spatial frequency reconstruction.
- The technique shows promise for advanced imaging applications, particularly for nanoscale materials.
