Automating high-resolution atomic force microscopy image interpretation

Atomic force microscopy (AFM) has become an important tool in nanoscale studies of matter, imaging and characterizing its properties, reaching from micrometer scale down to the sub-nanometer scale. In particular, high-resolution AFM performed in ultra-high vacuum with functionalized tips has achieved a resolution capable of identifying individual atoms in single molecules adsorbed on surfaces. However, applications so far have been mostly limited to simple planar molecules due to the difficult interpretability of the obtained images for more complicated sample molecules. At the same time, fast and accurate simulations of high-resolution AFM images have become available in the form of the Probe Particle Model (PPM). The PPM simulations make it easy to go from an atomic structure into an AFM image, but the reverse process is not as easy, often requiring a lot of manual labour with testing different candidate molecule geometries. An automated approach for solving this inverse problem would be a major benefit for the wider applicability AFM into imaging atomic scale systems. To this end, one can note the recent emergence of large-scale machine learning (ML) models, especially deep neural networks, enabling major progress in many fields of science. Utilizing a very large dataset, it is possible to train an ML model to perform a task where no known algorithmic approach works. Given the rapid data-generation capability of the PPM simulations, it should be possible to train an ML model to perform the inverse-imaging task in AFM, going from an AFM image into a molecule geometry or some other property of interest. The work in this thesis utilizes the PPM to generate simulations for a large database of molecules, and uses those simulations to train neural networks for predicting the molecule geometry as well as the electrostatic field from AFM images. The trained models are subsequently tested also on experimental AFM images, where generally good results are found, but in some cases differences in the details between the simulation and the real experiment make the predictions incorrect or ambiguous. At this stage full general automation of the AFM image interpretation process is not yet possible, but the results here present the first steps in this direction. Additionally, an ML method for automatizing the tip functionalization part of the AFM measurement preparation is presented.