Convolutional neural networks as an alternative to Bayesian retrievals for interpreting exoplanet transmission spectra

Exoplanet observations are currently analysed with Bayesian retrieval techniques. Due to the computational load of models used, a compromise is needed between model complexity and computing time. Analysis data from future facilities, will need more complex which increase retrievals, prompting search for faster approach interpreting exoplanet observations. Our goal compare machine learning retrievals transmission spectra nested sampling, understand if can be as reliable statistically significant sample while being orders magnitude faster. We generate grids synthetic their corresponding planetary atmospheric parameters, one using free chemistry models, other equilibrium models. Each grid subsequently rebinned simulate both HST/WFC3 JWST/NIRSpec observations, yielding four datasets in total. Convolutional neural networks (CNNs) trained each datasets. perform on 1,000 simulated combination type instrument sampling learning. also use methods real WFC3 spectra. Finally, we test how robust against incorrect assumptions our CNNs reach lower coefficient determination predicted true values parameters. Nested underestimates uncertainty ~8% whereas estimate them correctly. For agree within $2\sigma$ ~86% When doing assumptions, ~12% ~41% cases, this always below ~10% CNN.