Machine learning at the nanoscale
Although scanning probe microscopy (SPM) techniques have allowed researchers to interact with the nanoscale for decades now, little improvement has been made to the incredibly manual, time consuming process of setting up, running, and analysing the results of these experiments, oft...
| Main Author: | |
|---|---|
| Format: | Thesis (University of Nottingham only) |
| Language: | English |
| Published: |
2022
|
| Subjects: | |
| Online Access: | https://eprints.nottingham.ac.uk/67430/ |
| _version_ | 1848800418917777408 |
|---|---|
| author | Gordon, Oliver Miles |
| author_facet | Gordon, Oliver Miles |
| author_sort | Gordon, Oliver Miles |
| building | Nottingham Research Data Repository |
| collection | Online Access |
| description | Although scanning probe microscopy (SPM) techniques have allowed researchers to interact with the nanoscale for decades now, little improvement has been made to the incredibly manual, time consuming process of setting up, running, and analysing the results of these experiments, often arising due to the constantly varying shape of the probe apex. Unlike traditional computing methods, machine learning methods (with neural networks in particular) are considerably more capable of automating subjective tasks such as these, and we are only just beginning to explore the potential applications of this technology in SPM. In this thesis we explore a number of areas where machine learning could potentially massively change the way we go about SPM experimentation. We begin by discussing the history, theory, and experimental concepts of scanning tunnelling microscopy (STM), atomic force microscopy (AFM), and normal-incidence-x-ray standing wave (NIXSW). We then explore the makeup of a neural network and demonstrate how they can be applied to a variety of use-cases in SPM, including classification and policy prediction. Moving to the experimental chapters, we first discuss how we can successfully distinguish between STM tip states of the H:Si(100), Au(111) and Cu(111) surfaces. We also show that by adapting this network to work in real time, we improve performance while requiring on the order of 100x less data. We next discuss our attempts to combine these networks with expert examples to intelligently maintain tip apex sharpness during experimentation, envisioning an end-to-end automatic experiment. Because one of the main difficulties in applying machine learning is the frequent need to manually label data, we then show how we can use Monte Carlo simulations of self-organised AFM nanostructures to automatically label training data for a network, and then combine it with classical statistics and preprocessing to find specific structures in a mixed, messy dataset of real, experimental AFM images. As part of this, we also build a network to denoise experimental images. Finally, we present NIXSW results from an investigation into the temperature dependence of H20@C60, discussing the potential to use unsupervised clustering techniques to distinguish between noisy human-indistinguishable spectra to overcome limitations in data collection. |
| first_indexed | 2025-11-14T20:51:15Z |
| format | Thesis (University of Nottingham only) |
| id | nottingham-67430 |
| institution | University of Nottingham Malaysia Campus |
| institution_category | Local University |
| language | English |
| last_indexed | 2025-11-14T20:51:15Z |
| publishDate | 2022 |
| recordtype | eprints |
| repository_type | Digital Repository |
| spelling | nottingham-674302022-08-02T04:40:06Z https://eprints.nottingham.ac.uk/67430/ Machine learning at the nanoscale Gordon, Oliver Miles Although scanning probe microscopy (SPM) techniques have allowed researchers to interact with the nanoscale for decades now, little improvement has been made to the incredibly manual, time consuming process of setting up, running, and analysing the results of these experiments, often arising due to the constantly varying shape of the probe apex. Unlike traditional computing methods, machine learning methods (with neural networks in particular) are considerably more capable of automating subjective tasks such as these, and we are only just beginning to explore the potential applications of this technology in SPM. In this thesis we explore a number of areas where machine learning could potentially massively change the way we go about SPM experimentation. We begin by discussing the history, theory, and experimental concepts of scanning tunnelling microscopy (STM), atomic force microscopy (AFM), and normal-incidence-x-ray standing wave (NIXSW). We then explore the makeup of a neural network and demonstrate how they can be applied to a variety of use-cases in SPM, including classification and policy prediction. Moving to the experimental chapters, we first discuss how we can successfully distinguish between STM tip states of the H:Si(100), Au(111) and Cu(111) surfaces. We also show that by adapting this network to work in real time, we improve performance while requiring on the order of 100x less data. We next discuss our attempts to combine these networks with expert examples to intelligently maintain tip apex sharpness during experimentation, envisioning an end-to-end automatic experiment. Because one of the main difficulties in applying machine learning is the frequent need to manually label data, we then show how we can use Monte Carlo simulations of self-organised AFM nanostructures to automatically label training data for a network, and then combine it with classical statistics and preprocessing to find specific structures in a mixed, messy dataset of real, experimental AFM images. As part of this, we also build a network to denoise experimental images. Finally, we present NIXSW results from an investigation into the temperature dependence of H20@C60, discussing the potential to use unsupervised clustering techniques to distinguish between noisy human-indistinguishable spectra to overcome limitations in data collection. 2022-08-02 Thesis (University of Nottingham only) NonPeerReviewed application/pdf en cc_by https://eprints.nottingham.ac.uk/67430/1/OGordon_Thesis_Corrections.pdf Gordon, Oliver Miles (2022) Machine learning at the nanoscale. PhD thesis, University of Nottingham. CNN Machine learning STM AFM SPM Scanning probes Reinforcement learning Inverse reinforcement learning RL IRL Neural network H:Si(100) |
| spellingShingle | CNN Machine learning STM AFM SPM Scanning probes Reinforcement learning Inverse reinforcement learning RL IRL Neural network H:Si(100) Gordon, Oliver Miles Machine learning at the nanoscale |
| title | Machine learning at the nanoscale |
| title_full | Machine learning at the nanoscale |
| title_fullStr | Machine learning at the nanoscale |
| title_full_unstemmed | Machine learning at the nanoscale |
| title_short | Machine learning at the nanoscale |
| title_sort | machine learning at the nanoscale |
| topic | CNN Machine learning STM AFM SPM Scanning probes Reinforcement learning Inverse reinforcement learning RL IRL Neural network H:Si(100) |
| url | https://eprints.nottingham.ac.uk/67430/ |