Deep Learning Enhanced STM: Recognizing the Herringbone Surface Reconstruction and Atomic Lattice of Au(111)

Performing high quality scanning tunneling microscopy (STM) requires a clean, sharp, and stable STM probe tip. Many research groups use Au(111) to prepare and pre-characterize the STM tip for experiments. The preparations vary from imaging Au(111) until the images improve, applying voltage pulses while tunneling, to diving the tip into Au(111). Good tips are identified by imaging and performing tunneling spectroscopy of Au(111). All preparation techniques and the final inspection are time consuming and tedious, limiting the productivity of the microscope operator. Here we present a deep learning approach using a convolutional neural network (CNN) to recognize certain features in STM images of Au(111) to automatically identify tips which are suitable for high resolution imaging. These features include the herringbone surface reconstruction and atomic lattice of Au(111) samples. The dataset used to train our model comprised of STM images of Au(111) collected over several years, under a variety of imaging conditions. Even though the dataset was non-curated, our CNN was able to achieve 96.7% accuracy in identifying high quality tips. We will show that our technique potentially has broader utility than efforts based on more highly curated datasets.