Autonomous experimentation in 4D-STEM and EELS with deep kernel learning

In the scanning transmission electron microscope (STEM), a variety of techniques exist to characterize matter at atomic and nanometer scales. Electron energy loss spectroscopy (EELS) and 4D-STEM allows to study electronic, vibrational, chemical, and structural properties of a specimen. The locations from which to acquire such data is chosen by the operator. There exists a vast and unexplored space within a given sample, where many features may be overlooked due to operator bias. The measurement location for EELS or diffraction data are chosen based on structure, typically in the form of high angle annular dark field (HAADF)-STEM images.



Here we show how deep kernel learning (DKL) can be utilized to build relationships on-the-fly between local structural information (in the form of HAADF-STEM image patches) and local analytical responses. In other words, structure-property relationships are formed between local structure and an EEL spectrum or diffraction pattern coming from the center of the local structure image patch. The microscope can then operate in an autonomous manner and collect EELS or 4D-STEM data by continuing to learn structure property relationships on the fly.