Morello, G;
(2016)
A machine learning approach to exoplanet spectroscopy.
Doctoral thesis , UCL (University College London).
Preview |
Text
final_thesis_Giuseppe_Morello.pdf - Accepted Version Download (7MB) | Preview |
Abstract
The characterization of exoplanetary atmospheres is the new frontier in the field of exoplanets. Transit and eclipse spectroscopy are invaluable sources of information, as they may reveal the chemical composition, the presence of clouds, and the temperature and pressure profiles of the atmospheres of exoplanets. A photometric precision of about one part in 104 is necessary to make statistically significant inferences. The native calibration of current observatories, except Kepler, is not sufficient to reach this precision. In the past, parametric models have been used extensively by most teams to remove instrumental systematics. This approach has caused many debates regarding the use of different parametric choices for the removal of systematic errors. Parametric models decorrelate the systematic noise with the aid of auxiliary information on the instrument: the so-called optical state vectors (OSVs). Such OSVs can include inter- and intra-pixel position of the star or its spectrum, instrument temperatures and inclinations, and/or other parameters. The choice of the parameters to include in the OSVs is somewhat arbitrary, as is the choice of the functional forms to approximate the dependence of systematic noise on those parameters. The solution to many of the issues deriving from the use of OSVs lies in the use of `blind', non-parametric techniques. Such methods do not require a model for the systematics, and for this reason, they can be applied to any instrument with few changes (if any). In this Thesis, I focus on the Independent Component Analysis (ICA) of multiple time series, which performs a linear transformation of those series into maximally independent components. The use of ICA to detrend instrument systematics in exoplanetary light-curves was first proposed by Waldmann (2012). They experimented with spectroscopic light-curves taken with HST/NICMOS and sequential Kepler observations as input light-curves for the ICA. In this Thesis, I present two novel approaches to detrend single photometric observations in a self-consistent way (pixel-ICA), and scanning-mode spectroscopic observations without mixing the signals at different wavelengths (stripe-ICA). The two techniques that I pioneered extend the applicability of ICA to single observations with different instrument design. Some unsupervised preprocessing steps are also tested. The better performances of these algorithms compared to other ones in the literature are demonstrated over a series of Spitzer and Hubble observations, and synthetic data sets. The (re)analysis of archive and new data with similar techniques will cast new light on the characterization of exoplanets.
Type: | Thesis (Doctoral) |
---|---|
Title: | A machine learning approach to exoplanet spectroscopy |
Open access status: | An open access version is available from UCL Discovery |
Language: | English |
Keywords: | methods: data analysis, planets and satellites: atmospheres, techniques: photometric, techniques: spectroscopic |
UCL classification: | UCL UCL > Provost and Vice Provost Offices > UCL BEAMS UCL > Provost and Vice Provost Offices > UCL BEAMS > Faculty of Maths and Physical Sciences UCL > Provost and Vice Provost Offices > UCL BEAMS > Faculty of Maths and Physical Sciences > Dept of Physics and Astronomy |
URI: | https://discovery.ucl.ac.uk/id/eprint/1531000 |
Archive Staff Only
View Item |