This thesis is broken into three main sections tracing the steps of the development of a new framework to search for and characterize planets from the WASP survey. While all methods were developed specifically for the WASP project, the principles are easily transferable to any ground or space based survey. In the first part of the thesis, I discuss the development of two machine learning methods, a Random Forest Classifier and a Convolutional Neural Network, that are able to find new exoplanet candidates from WASP archival data and lightcurves. In preparing the training dataset, I also created a standardized catalog of 1,041 false positives from SuperWASP, the northern component of WASP, that were verified with additional observations from other instruments. The second part of the thesis begins by discussing the results of the machine learning methods. In the analysis of the resulting probabilities, several patterns began to emerge that indicated the differing predictions between the algorithms carry useful information in itself. This realization sparked the development of a new “stacking” framework where the predictions of a number of different machine learning methods areused as the input to a second-level classifier that makes the final prediction. This method is straightforward to implement and demonstrates an improved performance over any individual classifier, making the stacked approach ideal for future large surveys. I use the model to classify and rank more than 100,000 lightcurves in the WASP archive that do not yet have a disposition associated with them and discuss the candidates that are rated most favourably. Finally, in part three I discuss what to do once a candidate is confirmed to be a planet. In particular, I describe a new MCMC method that combines the likelihood fits of transit and radial velocity data with prior knowledge from several sources including optical and infrared spectrophotometric measurements and the parallax measurements from Gaia to constrain the stellar parameters. I apply the method to characterize two new hot Jupiter planets found by the WASP collaboration and confirmed with SOPHIE and TESS measurements. WASP-186b is a dense (4.22 ± 0.18MJ, 1.11 ± 0.03RJ ) planet on an eccentric (e=0.33 ± 0.01) 5-day orbit around a mid-F type star. While also in a ~5 day orbit, WASP-187b is puffed up (0.8 ± 0.09MJ, 1.64 ± 0.05RJ ) and orbiting a star that has begun evolving away from the main sequence.
Date of Award | 1 Dec 2020 |
---|
Original language | English |
---|
Awarding Institution | |
---|
Supervisor | Andrew Collier Cameron (Supervisor) |
---|
From candidate identification to planet characterization: a machine learning approach
Schanche, N. (Author). 1 Dec 2020
Student thesis: Doctoral Thesis (PhD)