Exoplanets: deep learning reduces some atmospheric analyses from several hours of computation to a few seconds

A literature review takes stock of the use of machine learning for the atmospheric detection and characterization of exoplanets with the JWST and the future Ariel mission.

The James Webb Space Telescope and the upcoming European Ariel mission produce volumes of light curves and spectra that saturate conventional processing chains. A review submitted to arXiv by Muallim Yakubu and Vwavware Oruaode Jude takes stock of the machine learning methods used to absorb this flux, on two fronts: detection (identification of transits, sorting of candidates, rejection of false positives) and atmospheric characterization (spectral inversion, correction of instrumental noise, modeling).

The overview covers classical approaches — random forests, convolutional networks — as well as more recent architectures, transformers and recurrent networks, as well as simulation-based inference: neural estimation of the a posteriori distribution, normalization flows. The authors note that these methods equal or exceed traditional pipelines, and above all that they radically compress inference times, going from hours of computation to a few seconds; nested sampling algorithms are accelerated by a factor of 3 to 8 without degrading statistical reliability.

Three obstacles remain, and the authors present them as the real projects of the discipline: the interpretability of models, the calibration of uncertainties and the generalization from one instrument to another — a sine qua non condition for a model trained on JWST data to remain usable on Ariel's data.