Machine learning for exoplanet characterisation in the JWST era

Abstract

Interpreting exoplanet observations to infer physical and chemical properties of their atmospheres is typically done using Bayesian inference to find the joint posterior probability distribution of model parameters. The posterior is traditionally found using sequential sampling-based inference methods, like Multinest. This process, called retrieval, is very computationally expensive, often requiring hundreds of thousands to millions of forward model evaluations to converge. The high computational expense essentially limits the use of complex atmospheric models in retrievals, making it necessary to reach a compromise between model complexity and compute time. The model complexity arises from including more realistic physics, for instance, computing self-consistently the temperature structure and cloud formation or including disequilibrium chemistry.

The aim of this thesis is to develop, test, and exploit machine learning inference methods that allow us to overcome these computational limitations and interpret JWST observations in a manner that exceeds the simplistic model interpretation currently done.

In the first chapter of this thesis I explored the performance of simple convolutional neural networks (CNNs) trained to infer atmospheric parameters from transmission spectra. For a large sample of simulated spectra, I tested the performance of the CNNs vs. that of Multinest, finding that the CNNs were very reliable and Multinest could be overconfident at times. Interestingly, I also found that for simulated spectra not well fitted by the retrieval setup (e.g. missing molecular species), the CNNs were significantly more reliable than Multinest. This method however was quite simplistic as the posterior was approximated by a gaussian distribution, which is often not the case for exoplanet atmospheres.

In the second chapter I developed a machine learning retrieval tool (FlopPITy ) combining sequential neural posterior estimation and normalizing flows. By using a normalizing flow to approximate the posterior, this method is completely flexible to predict any kind of posterior. FlopPITy is a tool that retains all the flexibility of traditional Bayesian retrieval methods for parameter inference (but it cannot provide the Bayesian evidence). In this chapter I showed that it is reliable and can significantly speed up retrievals, making retrievals with self-consistent models computationally feasible.

In the third chapter I used FlopPITy to perform self-consistent retrievals on the 0.5 – 18 μm emission spectrum of the cold brown dwarf WISE 1828+2650. Model grids have struggled to reconcile its NIR and MIR flux, and here we explored if a detailed parameter space exploration via retrievals could solve this. We showed that the 1 - 2μm is not reproduced by any of the retrievals and the basic properties of WISE 1828+2650 object (effective temperature, radius, and gravity) are dependent on the model assumptions. This highlights our incomplete understanding of Y dwarf atmospheres.

In the fourth and final chapter, I used FlopPITy to perform self-consistent retrievals on the 0.5 – 10 μm transmission spectrum of the hot Jupiter WASP-39 b. WASP-39 b was the first planet where photochemically produced SO2 was found. Some previous studies (focusing on smaller wavelength ranges) find a very high atmospheric metallicity which seems to be unphysically high. In this chapter we show that the full spectrum can be reproduced by a single self-consistent model assuming radiative-convective equilibrium, chemical equilibrium relaxed to account for vertical mixing-driven disequilibrium, self-consistent cloud formation, and parameterized photochemistry. We find a bulk chemical composition that is consistent with the expectation given the detection of photochemically produced SO₂. We also conclude that including CO₂ and CH₄ in our parameterised photochemistry could have some impact on the derived properties.