Repository for the preprint|paper "Exit Wavefunction Reconstruction from Single Transmisson Electron Micrographs with Deep Learning".

One-shot exit wavefunction with deep learning uses neural networks to recover phases of conventional transmission electron microscopy images by restricting the distribution of wavefunctions.

Figure: phases output by a neural network for input amplitudes are similar to true phases for In1.7K2Se8Sn2.28 wavefunctions.

In the wavefunctions directory, subdirectories numbered 1,2,3, ..., snapshot neural networks as they were developed. After ~20 initial experiments, architecture was kept almost uncharged for the GAN and direct prediction. Networks featured in the paper include

19: n=1, multiple materials
39: n=3, multiple materials

24: n=1, single material
38: n=3, single material

34: n=1, single material, generative adversarial network
37: n=3, single material, generative adversarial network

40: n=3, multiple materials, restricted simulation hyperparameters

A simple script to reconstruct focal series is in wavefunctions/hologram_reconstruction_old_code.py.

New datasets containing 98340 simulated wavefunctions, and 1000 experimental focal series available here.

n=3 datasets downsampled to 96x96 with antialiasing have beed added for rapid development.

Last saved checkpoints for notable files are here. Password: "W4rw1ck3m!" (without quotes). Note that the server is in a test phase, so it may be intermittently unavailable and the url will eventually change.

Multislice simulations were performed with clTEM. Simulations are GPU-accelerated, and are written in OpenCL/C++ for performance. Source code is maintained by Jon Peters and is available here with official releases.

Compiled versions of clTEM used in our research have been included

clTEM_files: n=1, Alternative physics
clTEM_file_0.2.4: n=3, Standard physics

Jeffrey M. Ede: j.m.ede@warwick.ac.uk - machine learning, general
Jonathan J. P. Peters: j.peters.1@warwick.ac.uk - clTEM
Jeremy Sloan: j.sloan@warwick.ac.uk
Richard Beanland: r.beanland@warwick.ac.uk