Le vendredi 20 juin 2025,  Malavika VASIST présentera l'examen en vue de l’obtention du grade académique de Docteur en Sciences (Collège de doctorat en Sciences spatiales) sous la direction de Olivier ABSIL et Gilles LOUPPE.

Cette épreuve consistera en la défense publique d’une dissertation intitulée :

« Exoplanet atmospheric characterization using amortized simulation based inference ».

Le Jury sera composé de :

M. M. FAYS (Président), MM. O. ABSIL (Promoteur), Q. CHANGEAT (University of Groningen), V. CHRISTIAENS (Secrétaire), G. LOUPPE (Co-promoteur), P. MOLLIERE (Max-Planck-Institut für Astronomie).

Abstract

Around 30 years after the first exoplanet detection and over 5000 detections later, we have come a long way in characterizing a huge diversity of exoplanets to understand their formation, evolution and habitability. Thanks to modern instrumentation providing high-quality spectra (emission and transmission), it is possible to study their structure and composition by atmospheric retrieval. Although conventional retrieval algorithms such as MCMC and nested sampling are reliable, they are limited in terms of time efficiency, scalability and testability. This leads us to explore an alternative family of algorithms called simulation-based inference, specifically using a variational deep-learning-based retrieval algorithm called neural posterior estimation (NPE), to estimate the posterior distribution directly by sidestepping likelihood computations. This algorithm improves over the traditional algorithms in speed, scalability and testability. Particularly, it offers amortization, which involves training a posterior estimator once that can then be used to perform quasi-instantaneous retrievals on subsequent observations of a similar kind. While this is useful for the rapid characterization of many spectral sources, its other key advantage lies in enabling statistical tests (such as coverage tests and posterior predictive consistency checks) to assess the validity of the retrieved posteriors, something which is otherwise not possible using conventional algorithms. We repeatedly perform these tests across all retrievals in this thesis, and in various cases, compare these retrievals with that of nested sampling. We find that the NPE posteriors are valid and consistently broader than those obtained with nested sampling. We formalize these uncertainties and suggest that the latter may be overconfident in its estimation of posterior width. As an application of our methodological developments, we use NPE to study the atmospheres of several brown dwarfs. Brown dwarfs are exoplanet analogs with similar atmospheres in the overlapping regions of their physical parameters (such as surface gravity, mass and effective temperature). We perform NPE spectral retrievals of six brown dwarfs ranging from L to Y spectral types, using various spectral wavelength regions and resolutions, in order to characterize them. We conduct a detailed atmospheric retrieval of two Y-type brown dwarfs using their mid-IR spectrum obtained with JWST/MIRI, together with archival near-IR spectra. We additionally apply our NPE retrieval framework on three other brown dwarfs ranging from late T to Y, based on their MIRI mid-IR spectrum. Further, we perform a preliminary comparative study of these five brown dwarfs in an attempt to identify trends. Lastly, we perform a pilot NPE retrieval study on a high-resolution near-infrared spectrum of an L type dwarf. With these results we identify the challenges and opportunities for SBI in exoplanet retrievals going forward.

Lien Orbi