Autonomous Atomic Manipulation and Characterization in the STEM

Control over matter has allowed physics, condensed matter, and materials science communities to realize a wide range of properties useful in many fields and applications. Accessing control of nanometer and smaller scales allows to shape materials properties in fundamentally new and exciting ways. The scanning transmission electron microscope (STEM), like scanning probe microscopy (SPM), routinely enables direct visualization of matter at the atomic scale. The atom-sized electron probe in the STEM, however, also contains enough energy and momentum to structurally and chemically alter matter at the same atomic scale, where this electron-matter interaction is typically considered “beam damage.” It is argued that understanding and harnessing this interaction to manipulate matter at atomic scales will allow a more complete design and control of materials.<br/><br/>While manipulation of matter at the level of single atoms has been demonstrated in both STEM and SPM, this has primarily been performed manually without reproducibility or useful precision. Beam-induced effects in the STEM have been notoriously difficult to control, and the ordinarily stochastic changes have also been challenging to characterize analytically via electron energy loss spectroscopy (EELS) or 4D-STEM methods.<br/><br/>Artificial intelligence is leveraged to guide the microscope towards both atomic fabrication and autonomous characterization. The first step towards accurate atomic manipulation is to obtain knowledge of the atomic landscape in as close to real time as possible, while the second step is instructing the electron beam where to go relative to these positions. The atomic coordinate extraction – from image data - must be performed both quickly and reliably. Several methods exist that can detect atoms, but here we use deep convolutional neural networks. However even though a trained neural network is fast, it, like other methods, usually requires fine tuning of several parameters to ensure a reliable extraction. The extra time this incurs cannot be exhausted during a live experiment; further, each new image might require different parameter tuning. To overcome this and ensure robust predictability, deep ensembles are used, which handle the issues - also encountered in self-driving vehicles - of so-called “out of distribution” shifts.<br/><br/>Given atomic coordinates, one can classify different atomic or defect classes based on mixture models, intensities, or graph analysis, which allows to experiment with beam-induced effects precisely at the single-defect level. Different strategies of beam control are demonstrated in both graphene and MoS₂ and are shown to produce structures that cannot be fabricated by any other means.<br/>Meanwhile, characterization of atomic and mesoscale features in the STEM is largely carried out by operator intuition and expertise such that a small region of interest is selected for detailed analyses, e.g., EELS or 4D-STEM, by inspection of the structural image, i.e., the annular dark field (ADF) signal. For several reasons, including handling beam sensitive materials, autonomous exploration of a larger sample space via EELS or 4D-STEM is useful. Deep kernel learning is utilized here to develop and learn a relationship between image data (structure) and a physical response by either EELS or 4D-STEM (property). These structure-property relationships are constructed from scratch and on-the-fly, with no prior information being supplied to the model. In this way, a material system can be explored in a completely autonomous fashion, without operator bias and without imparting too much unnecessary dose to the specimen.<br/><br/>Therefore, in leveraging AI and deep learning routines, atomic fabrication and autonomous characterization are made possible in the STEM. In the former, beam-matter interaction is exploited, while in the latter, it is minimized to only what is necessary. Both schemes are explored, and the challenges and practical limitations are discussed.