The Position Dependence of Electron Beam Induced Effects in 2D Materials with Deep Neural Networks

The electron beam in the aberration corrected scanning transmission electron microscope (STEM) is routinely used to observe the nature of materials at the atomic scale, where the interaction of the electron beam with a specimen is generally a highly complex process, but it can also include structurally changing the material. Traditionally, any direct modification of a material by the electron beam is viewed as an adverse effect – aka, “beam damage” – and many efforts to avoid this entirely are actively being developed, such as more efficient direct electron detectors. A variety of damage mechanisms can be responsible for the outcome of the beam interacting with a specific material [1] – for example, knock-on damage or radiolysis, where these depend on several factors like accelerating voltage and material. It was observed in 2008 that single Au atoms could be influenced by the electron beam but uncontrollably [2]. The first efforts of controlled atom manipulation at the single atom level were the seminal works in 2017 of Dyck [3] and Susi [4] to guide a silicon impurity atom throughout the graphene lattice, but this was done by hand in both cases and consequently was a time consuming and laborious process. Significantly overlooked is the position-dependence of beam induced effects. If the atom-sized electron beam is placed directly on an atom, is the effect different than if it is instead placed directly between two atoms (i.e., on the bond), or on a different type of atom? It would appear intuitively obvious that a difference must exist, however this has been elusive to quantify experimentally because positioning the beam deterministically with respect to specific atoms has been very challenging. These questions are addressed in a data-driven approach using intelligent beam positioning to sample a large variety of different beam positions relative to specific atom columns and the possible state change that accompanies it (i.e., acquiring a fast image after beam placement). The ensemble neural network-based atomic identification and beam positioning framework developed by the present authors [5] was used for precise beam control relative to specific atomic targets (silicon), where deep neural networks were then used to extract probability maps portraying regions in space relative to a target atom (Si) that have high probability to cause a particular outcome, e.g., drive Si along a certain crystal direction. We recently demonstrated this approach to autonomously manipulate 3-fold coordinated Si throughout the graphene lattice [6] (Figure 1). We extend this further with MoS2 aiming to gain insights into beam positions that drive specific defect states. A prominent distinction from manipulating Si in graphene is that in MoS2 and other transition metal dichalcogenides (TMDs), the interaction with the electron beam can cause material ejection (S vacancy generation) from the system at almost any accelerating voltage, and so the process is both dynamic and irreversible. Here there are multiple defect generation pathways dictated by how and where S vacancies form relative to one another, which can result in different defect structures (Figure 2). Beam position probability maps are extracted for these complex and multi-step scenarios with the aim of understanding the idealities required to gain highly precise selectivity and control of defect formation. Finally, these results are experimentally validated to gain more fundamental understanding and control of beam-matter interactions in 2D materials with opportunities to engineer matter and defects at the atomic scale [7]. (Left) Sampling many possible positions with the electron beam relative to Si atom and collecting cause-effect relationships (acquiring image after positioning beam) allows to determine the statistical probability distribution map for generating different outcomes – here, moving Si to position 1, 2, or 3 (center). The model is evaluated by autonomously manipulating Si throughout graphene using the learned relationships and protocols (right). Multiple defect formation pathways. All begin from pristine MoS2 (left) where the orange shaded region is sampled to determine best position to create S vacancies (center). From a single S vacancy, this is again sampled in green shaded region to determine beam placement route for each defect type.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.