Harnessing Machine Learning for the Analysis of Erbium Dipolar Quantum Gas Experiments

Lucas Hofer - Christ Church College - University of Oxford
2023 - DPhil thesis

This thesis centers on the use of machine learning methods in an erbium dipolar
quantum gas experiment. The first part of the thesis outlines the experiment, which
consists of both the previously built erbium machine used to cool erbium atoms
to quantum degeneracy as well as an experimental addition, which adds a second
atomic species to create a dual erbium-potassium cold atom machine. The design
of the potassium addition's vacuum and optics systems are detailed, as well as the
initial construction.
The second part of the thesis describes the use of deep learning methods to analyze
experimental images. This include the first neural network trained to simultaneously
profile multiple laser beams on a single image, as well as a deep neural network which
can both locate and classify cold atom clouds in an image. Furthermore, this section
contains details on the cold atom imaging systems I built during my DPhil, as well
as the experiment's data systems which I developed to acquire, store and extract
experimental data.
The last part of the thesis describes our custom image analysis software for cold
atom images. This includes the development of an open-source, GPU accelerated
curve fitting library, JAXFit, as well as a second software library, AtomCloud, which
builds on JAXFit and analyzes images of thermal atom clouds and Bose-Einstein
condensates. Finally, using both the experimental and data systems described in the
thesis, we measure a dipolar dependent shift in the critical number of atoms required
to achieve a Bose-Einstein condensate of erbium atoms.