Characterization and reconstruction of single molecules utilizing atomic force microscopy and machine learning

Atomic force microscopy (AFM) with functionalized tips has emerged as the primary experimental technique for understanding the atomic structure of organic molecules on surfaces. Additionally, when combined with atom manipulation, AFM provides the opportunity to explore the atomic structures of single molecules with high precision. As the molecule size increases, the need for accurate computational simulations becomes more burdensome. This thesis explores the capabilities of combining deep learning via neural networks with experimental non-contact atomic force microscopy (nc-AFM), addressing issues with tip functionalization, adsorption configuration, and molecular characteristics, such as electrostatics. To understand integration of machine learning methods with experimental nc-AFM, CO molecules were manipulated onto the end of a metallic tip after identification utilizing computer vision (CV) and convolution neural networks (CNN). This provided an automated way of preparing the AFM tips for later experiments and addressed a common issue in many experiments utilizing nc-AFM, particularly regarding the time involved to prepare a tip. Most molecules studied with nc-AFM are planar, but knowledge on the properties non-planar molecules was limited due to the challenges interpreting the resulting data. To address this, a deep learning infrastructure with a CNN was developed which matches a set of AFM images with a molecular configuration descriptor. This infrastructure was then applied to 1S-camphor on Cu(111), which demonstrated a rapid and reliable method for interpreting AFM images as well as reducing the potential adsorption configurations for future detailed studies. Approaching from a similar perspective, the conventional structure search process, even for small planar molecules, is computationally intensive. By employing Bayesian optimization, first-principles simulations with unbiased structure inference for multiple adsorption configurations became possible. To provide a reliable electrostatic characterization of individual molecules, experiments of PTCDA and benzene derivatives on Cu(111) were combined with a machine learning method to provide quantitative maps of the electrostatic potential directly from AFM data. Machine learning methods combined with atomic-scale AFM images provide a straightforward pathway to quantitatively understanding molecular properties with reduced bias.