yy-ssang / ml

Convolutional Neural Networks for Anomaly Detection in Scanning Transmission Electron Microscopy

Enea Prifti, James P. Buban, Arashdeep Thind, Robert F Klie

University of Illinois Chicago, Department of Physics, 845 W Taylor Street, Chicago, IL 60607

The codes in this repository are used in the paper mentioned above to detect anomalies in the pattern of a crystal lattice using a Convolutional Variational Autoencoder (CVAE) trained to reproduce that pattern.

These codes are written in Google Colab using Python 3.8.15. Each code is provided in two versions .py and .ipynb. The requirements to run the codes are listed in requirements.txt.

CVAE_training - Code for training the CVAE and saving the trained model. It requires a bulk sample, which is then divided into smaller sections to obtain a training set (to train the CVAE) and a testing set (to test the behavior of the CVAE on bulk samples for future comparisons to distinguish bulk or anomalies). Parameters like SIZE, SIGMA, epochs, and latent_dim can be modified by the user. By default, they are set to the optimal values that we found after various tests.

CVAE_testing - Code we used to obtain the plots shown in the paper. It reads sets of various anomalies we created manually to show the efficiency of this method in finding point defects.

analysis_example - Provides a sample code for the reader to apply our method. After running CVAE_training.py with your bulk image, use analysis_exaple.py to look for anomalies in other image/images of the same crystal. It will provide a list of booleans labeling
if an anomaly was found in a section of the image/images inputted