Data Science in Scanning Probe Microscopy: Advanced Analytics and Machine Learning
Open Access
- Author:
- Dusch, William George
- Graduate Program:
- Physics
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- July 25, 2018
- Committee Members:
- Eric Hudson, Dissertation Advisor/Co-Advisor
Eric Hudson, Committee Chair/Co-Chair
Jorge Sofo, Committee Member
Mauricio Terrones, Committee Member
Roman Engel-Herbert, Outside Member - Keywords:
- scanning tunneling microscopy
data science
machine learning
python
data analysis
deep learning
scanning probe microscopy
condensed matter
software
vibration cancellation
automation
unsupervised learning
physics - Abstract:
- Scanning probe microscopy (SPM) has allowed researchers to measure materials’ structural and functional properties, such as atomic displacements and electronic properties at the nanoscale. Over the past decade, great leaps in the ability to acquire large, high resolution datasets have opened up the possibility of even deeper insights into materials. Unfortunately, these large datasets pose a problem for traditional analysis techniques (and software), necessitating the development of new techniques in order to better understand this new wealth of data. Fortunately, these developments are paralleled by the general rise of big data and the development of machine learning techniques that can help us discover and automate the process of extracting useful information from this data. My thesis research has focused on bringing these techniques to all aspects of SPM usage, from data collection through analysis. In this dissertation I present results from three of these efforts: the improvement of a vibration cancellation system developed in our group via the introduction of machine learning, the classification of SPM images using machine vision, and the creation of a new data analysis software package tailored for large, multidimensional datasets which is highly customizable and eases performance of complex analyses. Each of these results stand on their own in terms of scientific impact – for example, the machine learning approach discussed here enables a roughly factor of two to three improvement over our already uniquely successful vibration cancellation system. However, together they represent something more – a push to bring machine learning techniques into the field of SPM research, where previously only a handful of research groups have reported any attempts, and where all efforts to date have focused on analysis, rather than collection, of data. These results also represent first steps in the development of a “driverless SPM” where the SPM could, on its own, identify, collect, and begin analysis of scientifically important data.