With the development of ever-improving telescopes capable of observing exoplanet atmospheres in greater detail and number, there is a growing demand for enhanced three-dimensional climate models to support and help interpret observational data from space missions such as CHEOPS, TESS, JWST, PLATO, and Ariel. However, the computationally intensive and time-consuming nature of general circulation models (GCMs) poses significant challenges in simulating a wide range of exoplanetary atmospheres. The aim of this study is to determine whether machine learning (ML) algorithms can be used to predict the three-dimensional temperature and wind structure of arbitrary tidally locked gaseous exoplanets in a range of planetary parameters. We introduced a new three-dimensional GCM grid comprising 60 inflated hot Jupiters orbiting A, F, G, K, and M-type host stars, which we modelled using . We defined four climate characteristics to characterise these planets: the dayside–nightside temperature difference, the evening–morning temperature difference, the maximum zonal wind speed, and the wind jet width. We trained a dense neural network (DNN) and an extreme gradient boosting algorithm (XGBoost) on this grid to predict local gas temperatures, as well as horizontal and vertical wind fields. To assess the reliability and quality of the ML models' predictions, we selected WASP-121 b, HATS-42 b, NGTS-17 b, WASP-23 b, and NGTS-1 b–like planets, all of which are targets for PLATO observations, and modelled them using and the two ML methods. For these test cases, we calculated the equilibrium gas-phase composition and transmission spectra to evaluate whether differences in local gas temperature between the general circulation model and ML approaches significantly affected the predicted chemical composition and transmission spectra. ExoRad ExoRad With the multi-layer neural network, that is DNN, predictions for the gas temperatures are to such a degree that the calculated spectra agree within 32 ppm for all but one planet, for which only one single HCN feature reaches a 100 ppm difference. The XGBoost predictions are somewhat worse but never exceed 380 ppm differences. Generally, the resulting deviations are too small to be detectable with the observational capabilities of modern space telescopes, including JWST. For the DNN, only the WASP-121 b-like planet, which is the hottest investigated planet, shows a general offset that is smaller than 16 ppm. Horizontal wind predictions are less accurate but can capture the most general trends. As with temperature, the DNN also outperforms XGBoost in this respect. Predicting vertical wind remains challenging for all ML methods that we explored in this study. The developed ML emulators can, within one second, reliably predict the complete three-dimensional temperature field of an inflated warm to ultra-hot tidally locked Jupiter around A to M-type host stars. They therefore provide a fast and computationally inexpensive tool to complement and extend traditional GCM grids for exoplanet ensemble studies. The quality of the predictions is such that no, or only minimal, effects on the gas-phase chemistry and, consequently, on cloud formation and transmission spectra are expected.
Plaschzug et al. (Tue,) studied this question.