Atomic resolution scanning probe microscopy, and in particular scanning tunnelling microscopy (STM) allows for high-spatial-resolution imaging and also spectroscopic analysis of small organic molecules. However, preparation and characterisation of the probe apex in situ by a human operator is one of the major barriers to high-throughput experimentation and to reproducibility between experiments. Characterisation of the probe apex is usually accomplished via assessment of the imaging quality on the target molecule and also the characteristics of the scanning tunnelling spectra (STS) on clean metal surfaces. Critically for spectroscopic experiments, assessment of the spatial resolution of the image is not sufficient to ensure a high-quality tip for spectroscopic measurements. The ability to automate this process is a key aim in development of high resolution scanning probe materials characterisation. In this paper, we assess the feasibility of automating the assessment of imaging quality, and spectroscopic tip quality, via both machine learning (ML) and deterministic methods (DM) using a prototypical tin phthalocyanine on Au(111) system at 4.7 K. We find that both ML and DM are able to classify images and spectra with high accuracy, with only a small amount of prior surface knowledge. We highlight the practical advantage of DM not requiring large training datasets to implement on new systems and demonstrate a proof-of-principle automated experiment that is able to repeatedly prepare the tip, identify molecules of interest, and perform site-specific STS experiments using DM, in order to produce large numbers of spectra with different tips suitable for statistical analysis. Deterministic methods can be easily implemented to classify the imaging and spectroscopic quality of a STM tip for the purposes of high-resolution STM and STS on small organic molecules. Via automated classification of the tip state, we demonstrate an automated experiment that can collect a high number of spectra on multiple molecules without human intervention. The technique can be easily extended to most metal–adsorbate systems and is promising for the development of automated, high-throughput, STM characterisation of small adsorbate systems.
Barker et al. (Mon,) studied this question.