“…Neural networks have already revolutionized image recognition in fields such as medical diagnosis, weather forecasting, and facial recognition; recently, they have also been applied to identify atomic defects in atomic-resolution (S)TEM images. − Conventionally, defect detection has been a labor-intensive task which is often done by hand , or simple image processing such as Fourier filtering or intensity thresholding. , Neural networks offer an opportunity to automate defect identification, making it possible to efficiently locate large numbers of defects to generate class averages systematically while minimizing human intervention. We trained FCNs using simulated data generated via incoherent image simulations using Computem. , In order to make our simulations more realistic, we apply a set of postprocessing steps to the images, including the addition of Gaussian noise, probe jittering, image shear, and varying spatial sampling, to create our final training data. Similar methods are well-established in the literature, − though we found that we achieved the highest classification precision on experimental data by introducing low-frequency contrast variations in the simulated data to emulate surface contamination.…”