Exoplanet atmospheric characterization using amortized simulation based inference

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. These limitations become more pronounced with the anticipated influx of spectral observations from JWST and future missions such as ARIEL, particularly in the context of population studies. 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 thousands of spectral sources in only a few hours, its other key advantage lies in enabling statistical tests (such as coverage tests and L-C2ST tests) to assess the validity of the retrieved posteriors, something which is otherwise not possible using conventional algorithms. In this thesis, we use NPE to perform spectral retrievals of six brown dwarfs (aka exoplanet analogs) 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 their archival near-IR spectra. We additionally perform systematic retrievals on five brown dwarfs (including the previous two) using an amortized approach, and compare the results to identify trends, thus setting a precedent for future population studies using SBI. Lastly, we perform a pilot NPE retrieval study on a high-resolution near-infrared spectrum of an L type dwarf. We repeatedly perform validation 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 dissect the sources of Bayesian model uncertainty through a single retrieval, and suggest that nested sampling may produce overconfident posterior estimates. With these results, we also identify the challenges and opportunities for SBI in exoplanet retrievals going forward.