AI-enabled Automation of Atomic Manipulation and Characterization in the STEM

Manipulation of matter at the atomic scale has been carried out for several years. The atom sized electron beam in the scanning transmission electron microscope (STEM) provides a useful analytical probe, but if used appropriately, it can control the motion or position of single atoms. The seminal works by Dyck [1,2] and Susi [3] show that single silicon atoms – as well as other impurity atoms – in graphene could be guided throughout the lattice by clever use of the electron beam. Several schemes exist now to garner control of atomic structures; however, many do not consider the coordinates of the atoms. In the initial implementations of using an electron beam to control, e.g., dopant motion in graphene, the human operator would manually position the beam nearby the dopant to cause it to exchange positions with a neighbor carbon. In these cases, the atom positions are considered but this is entirely a manual procedure. To scale this up and generalize to other systems, automation of beam positioning relative to specific atomic groups is required. In other words, atomic coordinates and types must be known in as close to real-time as possible and then followed by specific beam positioning. Recently, it was shown that ensemble neural networks can handle the real-time atomic segmentation of STEM images [4,5]. Perhaps even more crucial is the fact that this atomic classification scheme must be robust, as it is performed actively during the experiment, meaning that model hyperparameters cannot be constantly changed to provide a reasonable coordinate extraction. Regardless, ensemble networks provide both a fast and robust solution for atomic segmentation in real time. Provided the atomic coordinates and classes, the beam positions must be chosen. For some materials, it may be obvious where the beam should be positioned in order to elicit the desired response, i.e., formation of intended defect structure. In other materials, it may be considerably more complex and requires, for instance, a large collection of state-action pairs where the beam is positioned within a distribution relative to atomic classes and the resulting structures are imaged; theoretical calculations can alternatively be performed to aid in this beam-matter understanding. In this talk, however, we will focus primarily on the former, where we assume we have a high degree of confidence in what the resulting structure will be for a given beam placement. Here we discuss the AI-enabled automated beam manipulation experiments for a variety of both 2D and 3D materials, where the entire process is carried out using a Nion UltraSTEM instrument within the Nion Swift control software. Nion Swift allows, via Python, seamless control of the microscope parameters such as beam position and dwell time, as well as access to detector feedback, all of which are critical for these experiments. A custom plugin is used to specify the automated experiment details, e.g., placement of the electron beam on a chain of 10 S atoms in MoS2, while avoiding Mo atoms; or a checkerboard pattern targeting only Cr columns in CrSBr, another candidate for beam induced transformations [6]. Additionally, automatic drift correction and feedback informed process halting are incorporated into the workflow. In this talk, I will discuss how these controls and AI-enabled workflow are used in the creation of custom atomic defect classes. Finally, the created defects can be characterized in-situ with STEM-EELS without destruction using the same atomic detection scheme presented. AI-enabled atomic manipulation in MoS2. Single vacancy line formation in MoS2 captured in annular dark field with real-time atomically segmented image shown below each annular dark field (ADF) image. (a) through (c) illustrate consecutive steps of the process, with targeted sulfur atoms highlighted in red squares in the segmented images on the bottom row. Scalebars 1 nm. Atomic segmentation performed live during experiment with CrSBr. The ADF (a) and semantically segmented image (b) are presented. Selected Cr columns are targeted with the beam to create checkerboard pattern. AI parameters and beam controls accessible as custom plugin within Nion Swift shown in (c). Scalebar in (a) 1 nm.