Deep learning for anomaly detection in scanning transmission electron microscopy

Identifying point defects and other structural anomalies using scanning transmission electron microscopy (STEM) is important to understand a material's properties caused by the disruption of the regular pattern of a crystal. Thanks to the high spatial resolution of aberration-corrected transmission electron microscopes, atomic-resolution images with a field of view of several hundred nanometers can be taken. Such data, which often contains thousands of atomic columns need to be analyzed. This process has been done manually in the past, but recent developments in machine learning (ML) can be very useful to speed it up. In this contribution, we will utilize a convolutional variational autoencoder (VAE) which, after being trained with a set of bulk samples, generates an example (prediction) of given input images based on the trained features. We will demonstrate that the performance of a VAE in replicating an input image can be used to differentiate between bulk or defects. In the case of a bulk input, the VAE can replicate well the input within a threshold value that can be set by testing the predictions. For a defect input, the VAE will fail to output a prediction within the set threshold, allowing for a clear and automatic distinction of defects.