Recent advances in the data acquisition process in the electron microscope allow us to collect multidimensional data of the sample under study at a rate of around 1TB/hour. In principle, with all of these experimental data, we should be able to retrieve all information about the specimen under study, such as three-dimensional atomic positions, atomic compositions, valence states, etc. However, due to the electron-specimen interaction is governed by quantum mechanics, all information about the sample is scrambled in the experimental data. Therefore, to recover the required information, we have to perform reverse engineering on the recorded data. This can be done by using machine learning, in which our features will be electron microscopy simulations, and labels will be the required sample properties.
In this talk, I will show some applications of machine learning in electron microscopy. In the first part, I will focus on how to mathematically model undistorted/distorted SEM/STEM/TEM data using accurate and/or fast electron-specimen interaction models. In the second part, a few neural networks are presented which are able to compensate for high levels and combinations of distortions present in experimental SEM, STEM, TEM, and CBED datasets.
Copyright. All Rights Reserved.