Deep PACBED: Multitask analysis of PACBED images using deep neural networks

Modern scanning transmission electron microscopes (STEMs) can generate large image datasets of semiconductor samples. The recorded position-averaged convergent beam electron diffraction (PACBED) images are useful for measuring parameters such as sample thickness, rotation, or mistilt. Typically, these measurements are performed manually, which makes them time-consuming and often inaccurate. This is why machine learning methods are an attractive option for analyzing large datasets of PACBED images quickly and consistently. Previous approaches focus on machine learning models that analyze each sample parameter individually. This paper investigates multitask deep neural networks that simultaneously predict multiple sample parameters, outperforming models trained on a single parameter. Our deep learning models are trained using a combination of synthetic images of III-V semiconductor materials and silicon, simulated by a multi-slice algorithm and small amounts of experimental data. In addition, we explore how the amount of experimental training data impacts the practical performance of various neural network architectures. On our experimental test datasets, the best deep neural network achieves a mean absolute error of 4.19° for predicting sample rotation α, 0.43milliradians for predicting sample mistilt β, and 2.93nanometers for predicting sample thickness t. These values are within the range of experimental measurement uncertainty. For material classification m, the best models perform without errors on the trained materials.