Interpreting exoplanet observations using machine learning.

Observations of exoplanet atmospheres are currently interpreted using Bayesian retrieval techniques, which typically require hundreds of thousands of model evaluations. This means that model complexity is typically compromised in favour of faster retrievals. The analysis of upcoming data from JWST and other future facilities like ARIEL will require more complex models, becoming impractically time consuming with current techniques.

A possible solution to this problem would be to use machine learning, needing to compute models only once to train an algorithm, and being able to perform retrievals within seconds after that. Our aim is to understand if machine learning can provide reliable results and compare it to nested sampling for a large sample of spectra.

We train convolutional neural networks (CNNs) to perform retrievals on transmission spectra of both Hubble Space Telescope Wide Field Camera 3 (HST/WFC3) and James Webb Space Telescope Near-Infrared Spectrograph (JWST/NIRSpec) using free and equilibrium chemistry models. We compare the predictions of the CNNs with nested sampling for large samples of simulated spectra. We also compare the predictions of the two methods for real HST/WFC3 observations of 48 exoplanets. Finally, we test how well both methods deal with incomplete complexity or systematic uncertainties in the underlying atmospheric models.

We find that the CNNs reach a lower coefficient of determination between the predicted and true values of the parameters. Nevertheless, the predictions from the CNNs are more reliable, with only ~0.3% of predictions more than 3σ away from the true value, a fraction that is ~8% for nested sampling. For real HST/WFC3 data, both methods are in good agreement, being within 2σ of each other for 41 out of the 48 spectra. Finally, when performing retrievals with systematic changes in the underlying atmospheric model, the CNN predictions remain within 3σ of the truth for more than 90% of spectra, while for nested sampling this fraction is between ~59% and ~88%.

In conclusion, we show that machine learning retrievals are very reliable and should be seriously considered and further researched as a fast alternative to nested sampling retrievals.