GPU-accelerated single side-band ptychography

Python code to calculate a phase image from the bright field disk of atomic resolution 4D-STEM data using the single side-band method.

Acknowledgements

Philipp Pelz authored the original GPU-accelerated single side-band code with post-processing aberration correction (Github repo, paper). Peter Ercius helped me to understand and use this code.

I built upon this foundation by reorganizing the code, rewriting the data pre-processing part, and changing the aberration correction user interface somewhat.

Features

• Fast: The code takes less than a minute to import and preprocess a 4D-STEM data set, and less than 10 seconds (on my computer at least) to compute a single side-band reconstruction.
• I've written pre-processing code to support data from both the EMPAD and the 4D Camera from NCEM (see the provided Jupyter notebooks).
• Post-processing aberration correction. You can manually correct for defocus in your 4D-STEM data, allowing for you to quickly check your defocused 4D-STEM data sets. (Can also manually correct for geometric aberrations up to 3rd order.)
• Plots slices of $G(K_f, Q_p)$, which may let you better understand the probe aberrations in your 4D data set.

Limitations

• You have to manually put in experimental parameters like convergence angle, probe step size, etc.
• Limitations of single side-band ptychography: Assumes a weak phase object. Spatial resolution is limited by convergence angle; it can be a bit better than ADF-STEM but is not as good as iterative, full-field ptychography.
• Though the crop_bin_sparse_to_dense function uses less GPU memory, it still requires a lot of RAM (10s of GB for some 4D Camera scans) -- there is room for improvement here.

Requirements

• Needs a Nvidia GPU and CUDA installed.

Installation

1. Install cupy. The command you need to run likely depends on the version of CUDA you have, e.g.,

pip install cupy-cuda117

for CUDA v11.7.
2. Install ipywidgets.
3. Install this package using pip. Open a terminal in this folder, and run

pip install .

or

pip install -e .

This should automatically install this code as a Python package, along with all dependencies not already listed.