Automated Thickness Determination from 4DSTEM data us-ing Convolutional Neural Networks

Determining sample thickness along the beam direction is critical for many characterization techniques in the transmission electron microscope (TEM), including quantitative electron en-ergy loss spectroscopy (EELS), coherent image contrast interpretation, and the correlation of experimental data with multislice simulations. Position-averaged convergent beam electron dif-fraction (PACBED) is a widely used method for precise thickness determination in scanning TEM (STEM), particularly beneficial for momentum-resolved (4DSTEM) experiments. However, conventional PACBED analysis relies on manual comparison with simulated patterns, which is time-intensive and user-dependent [3, 1]. To overcome these limitations, we propose a server-based approach leveraging convolutional neural networks (CNNs) for automated PACBED analysis [2]. A database of pre-trained CNN models is hosted on a network service, accessible directly from any data acquisition and analysis software providing a Python interface. This en-ables real-time, reproducible thickness determination within seconds during a microscope ses-sion, eliminating the need for expertise in machine learning or multislice simulations. We demonstrate a working prototype featuring a shared CNN database covering multiple material systems.