Autonomous Scanning\nTunneling Microscopy Imaging via\nDeep Learning

Scanning tunneling microscopy (STM) is a powerful technique\nthat\nprovides the ability to manipulate and characterize individual atoms\nand molecules with atomic-level precision. However, the processes\nof scanning samples, operating the probe, and analyzing data are typically\nlabor-intensive and subjective. Deep learning (DL) techniques have\nshown immense potential in automating complex tasks and solving high-dimensional\nproblems. In this study, we developed an autonomous STM framework\npowered by DL to enable autonomous operations of the STM without human\ninterventions. Our framework employs a convolutional neural network\n(CNN) for real-time evaluation of STM image quality, a U-net model\nfor identifying bare surfaces, and a deep Q-learning network (DQN)\nagent for autonomous probe conditioning. Additionally, we integrated\nan object recognition model for the automated recognition of different\nadsorbates. This autonomous framework enables the acquisition of space-averaging\ninformation using STM techniques without compromising the high-resolution\nmolecular imaging. We achieved measuring an area of approximately\n1.9 μm<sup>2</sup> within 48 h of continuous measurement and\nautomatedly generated the statistics on the molecular species present\nwithin the mesoscopic area. We demonstrate the high robustness of\nthe framework by conducting measurements at the liquid nitrogen temperature\n(∼78 K). We envision that the integration of DL techniques\nand high-resolution microscopy will not only extend the functionality\nand capability of scanning probe microscopes but also accelerate the\nunderstanding and discovery of new materials.