This paper addresses the challenge of developing machine learning (ML) models that rapidly generalize to large datasets under varying experimental conditions in electron microscopy. The authors employ a cycle generative adversarial network (CycleGAN) with a reciprocal space discriminator to augment 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. The approach is showcased by training a fully convolutional network (FCN) to identify single atom defects in a large dataset, demonstrating adaptable FCNs that adjust to dynamically changing experimental variables with minimal intervention.

Abid Khan, Chia-Hao Lee, Pinshane Y. Huang, Bryan K. Clark