Using Cycle-GANS to Generate Realistic STEM Images for Defect Identification

Identifying atomic defects in aberration-corrected scanning transmission electron microscopy (STEM) data is critical to understanding the structure and properties of 2D materials. Recent advances in machine learning techniques now allow very fast defect identification in STEM imaging. The training sets for these machine learning models are constructed with simulation codes that replicate realistic data via the manual addition of Gaussian noise, probe jittering, image shear, and background contamination. This procedure is not only time consuming, but the manual tuning of noise may not cover all the realistic factors of the experimental data. We present an alternative approach to generating realistic STEM images by employing a cycle-GAN to automatically add realistic noise to simulated data. In a cycle-GAN, the noise present in the experiment STEM images are "transferred" over to the simulated images. We train our defect-identification model using these generated images and evaluate the model on real STEM images to locate atomic defects within them. The application of Cycle-GAN removes the need for human intervention in the machine-learning workflow allowing for higher throughput results as well providing other machine learning models with more realistic data for any type of supervised learning.