Automating Scanning Tunnelling Microscopy: A Comparative Study of Machine Learning and Deterministic Methods

Scanning probe microscopy (SPM) methods allow for atomic scale investigation of material surfaces in real space and provide the potential to construct atomically precise structures atom-by-atom. Although these techniques have been available for decades, full automation of the processes involved in these experiments has not yet been fully realised. Currently, the process of setting up, running, and maintaining the SPM remains a laborious, time-consuming process, often as a result of the constantly changing shape of the probe tip.

Manual identification and correction of the state of the probe tip in situ requires a human operator to compare the probe quality via manual inspection of topographical images after any change in the probe. Previous attempts to automate the classification of the scanning probe state have predominantly relied on machine learning (ML) techniques. However, training these models demands large, labelled datasets for each surface under study. These datasets are extremely time-consuming to create, and are not always available, especially when considering a new substrate or adsorbate system.

This thesis focuses on automating the classification of the probe tip in scanning tunnelling microscopy (STM), addressing both imaging and spectroscopic applications. The use of ML in this automation is compared to less data-intensive, deterministic techniques, exploring the broader need for ML in autonomous scripting. We find that using these deterministic methods, comparable results in classifications can be achieved to those obtained with the use of ML. ML models were trained when possible to demonstrate the efficacy of the deterministic methods, via direct comparisons between the two classification techniques. The applicability of deterministic approaches is further validated through the utilisation of these classifiers in automated experiments on various substrate systems. In addition to this, a scale-invariant method for surface mapping using Fourier space analysis is presented, which could aid in further automated experimentation.