Leveraging generative adversarial networks to create realistic scanning transmission electron microscopy images

TL;DR

This paper introduces a CycleGAN-based method to generate realistic STEM images from simulated data, enabling improved ML models for defect identification in large, variable datasets.

Contribution

The authors develop a CycleGAN with a reciprocal space discriminator to produce highly realistic electron microscopy images, enhancing ML training and adaptability.

Findings

Generated images are nearly indistinguishable from real data

The approach enables training of adaptable ML models for defect detection

Facilitates autonomous analysis of large microscopy datasets

Abstract

The rise of automation and machine learning (ML) in electron microscopy has the potential to revolutionize materials research through autonomous data collection and processing. A significant challenge lies in developing ML models that rapidly generalize to large data sets under varying experimental conditions. We address this by employing a cycle generative adversarial network (CycleGAN) with a reciprocal space discriminator, which augments simulated data with realistic spatial frequency information. This allows the CycleGAN to generate images nearly indistinguishable from real data and provide labels for ML applications. We showcase our approach by training a fully convolutional network (FCN) to identify single atom defects in a 4.5 million atom data set, collected using automated acquisition in an aberration-corrected scanning transmission electron microscope (STEM). Our method produces adaptable FCNs that can adjust to dynamically changing experimental variables with minimal intervention, marking a crucial step towards fully autonomous harnessing of microscopy big data.