Characterising Exoplanet Atmospheres using Traditional Methods and Supervised Machine Learning
Since the discovery of the first extrasolar planets over 25 years ago, the field of exoplanet research has exploded. Today we have over 4000 confirmed exoplanets, with a wide variety of sizes, orbital separations, and host stars. The characterisation of this diverse population of objects has led to exciting discoveries about the conditions of alien worlds. Future technological advances are expected to provide an abundance of exoplanet spectra with a higher precision and sensitivity than ever before. This calls for a parallel advancement in the accuracy and speed of atmospheric models to interpret this influx of data.
In this thesis, my work on atmospheric retrievals is presented. Starting with traditional techniques, my first thesis paper applies a Bayesian retrieval in combination with an analytical atmospheric model to the Hubble transmission spectra of 38 different exoplanets. My second paper considers the theoretical model of the sodium doublet, and the effect of dropping the assumption of local-thermodynamic equilibrium. From here, I went on to develop a method that uses supervised machine learning to improve the speed and efficiency of the retrieval. This method was explained and tested in a collaborative paper with machine learning experts in Bern. The machine learning retrieval is then applied in several follow-up studies, covering a range of different scenarios. One of these was my final thesis paper, which further extends the new retrieval to high-resolution spectra using the cross-correlation function.
In addition to my own papers, I have contributed to a number of studies led by collaborators by running retrievals, assisting other students, and participating in scientific discussions. I have also worked on several observing proposals, both for high-resolution ground-based observatories and for the upcoming James Webb Space Telescope.
I plan to continue my work on exoplanet characterisation and machine learning in the future, using the technique to combine high- and low-resolution spectra to gain further insight into the atmospheres of these distant planets. The speed and efficiency of machine learning will also allow for statistical studies of exoplanets as the quantity of atmospheric spectra from new and upcoming telescopes escalates. Not only will these studies teach us about the conditions and potential habitability of exoplanets, but they will also answer questions about planet formation, diverse chemical processes, and the uniqueness of our solar system.