Soutenance de thèse : Hugo Vivien

« Development and evaluation of a Convolutional Neural Network for the detection of planetary transits« .

The defense will take place on Friday, the 16th of May at 2pm (french time), in the LAM amphitheater, and will be in English.
Abstract:
The population of known exoplanets has been rapidly grown in the past couple of
decades. The next challenges are now twofold: (i) detect small, rocky, and temperate
exoplanets, and (ii) assess planetary systems physical and dynamical properties in
detail, to better understand planetary formation and evolution.

In this work, I propose a deep learning detection method, that is able to directly
pinpoint transits in unfiltered light curves. The model I create, PANOPTICON, is trained
on simulated PLATO light curves, which includes stellar variability signals, as well
as eclipsing binaries, and background eclipsing binaries. The exoplanet population
included is diverse, and biased towards small planets, to explore the efficiency of this
method in detecting difficult planets. I also test the model on other datasets, and train
dedicated versions for real-world TESS data. On the simulated PLATO data, I report
a recovery rate of planets of ≈ 86 to 90 % (~100% for SNR>3), for a respective false alarm rate of 0 to 1 %.
When applied to a new dataset, still for the specific case of the future PLATO light
curves, but simulated using a completely different approach, the model is still able
confidently recover ≈ 99 % of the planets in the dataset. The versions trained on TESS
data shows similar performances, with a 96 % recovery rate for a 0.44 % false alarm
rate.

Jury members:
Anthony Boccaletti (LESIA) — Reviewer
Shay Zucker (Tel Aviv university) — Reviewer
Claire Moutou (IRAP) — Examiner
Juan Cabrera (DLR) — Examiner
Thierry Artières (LIS) — President
Magali Deleuil (LAM) — Supervisor
Suzanne Aigrain (Oxford university) — Invited member