Master thesis : Exoplanet Orbital Characterization Using Simulation-Based Inference

Abstract

This thesis aims at leveraging advances in deep learning, particularly simulation-based inference, to enhance the orbital parameter characterization of exoplanets. The current methods, like MCMC, are computationally expensive and slow to converge. Using Normalizing Flows and the expected forward Kullback-Leibler divergence as a loss function to train the model, we reproduced the results of the state-of-the-art method, $\alpha$-DPI.

However, the non-amortized nature of this approach limited its generalizability, necessitating retraining for new datasets or additional observations of the exoplanet beta-Pic b. To address these limitations, a generic model for exoplanet astrometry was developed using a ResMLP as an embedding network. Using different experiments, we showed that this generic model was able to infer the posterior of the orbital parameters of all four planets of the HR 8799 system, significantly reducing the computational effort compared to MCMC.

Despite these advancements, challenges remain, particularly in generalizing the model across exoplanets from different systems, as this generic model could not infer the posterior of the orbital parameters of beta-Pic b.