Leveraging generative adversarial networks to create realistic scanning transmission electron microscopy images
This work addresses the problem of automating materials research through ML in electron microscopy, representing an incremental advance by enhancing data augmentation for specific experimental conditions.
The paper tackled the challenge of developing machine learning models that generalize to large electron microscopy datasets under varying conditions by using a CycleGAN with a reciprocal space discriminator to generate realistic images from simulated data. This enabled training a fully convolutional network to identify single atom defects in a 4.5 million atom dataset, achieving adaptable models with minimal intervention.
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.