Simulation-based Inference for Exoplanet Atmospheric Retrieval: Insights from winning the Ariel Data Challenge 2023 using Normalizing Flows

Outline of the Ariel Challenge dataset era. The enter parameters are used to generate the artificial exoplanet spectra and noise arrays. A Nested Sampling is carried by the Ariel Data Challenge Organising workforce on these artificial preferrred spectra, producing samples of the goal parameters. Our Machine Learning mannequin goals to breed the samples of parameters of the Nested Sampling. See textual content for extra particulars. — astro-ph.EP

Advancements in house telescopes have opened new avenues for gathering huge quantities of information on exoplanet environment spectra. However, precisely extracting chemical and bodily properties from these spectra poses important challenges on account of the non-linear nature of the underlying physics.

This paper presents novel machine studying fashions developed by the AstroAI workforce for the Ariel Data Challenge 2023, the place one in all the fashions secured the prime place amongst 293 rivals. Leveraging Normalizing Flows, our fashions predict the posterior chance distribution of atmospheric parameters beneath totally different atmospheric assumptions.

Moreover, we introduce an alternate mannequin that displays greater efficiency potential than the winning mannequin, regardless of scoring decrease in the problem. These findings spotlight the have to reevaluate the analysis metric and immediate additional exploration of extra environment friendly and correct approaches for exoplanet environment spectra evaluation.

Finally, we current suggestions to boost the problem and fashions, offering priceless insights for future functions on actual observational information. These developments pave the manner for simpler and well timed evaluation of exoplanet atmospheric properties, advancing our understanding of those distant worlds.

Mayeul Aubin (1,2), Carolina Cuesta-Lazaro (1), Ethan Tregidga (1,3), Javier Viaña (4), Cecilia Garraffo (1), Iouli E. Gordon (1), Mercedes López-Morales (1), Robert J. Hargreaves (1), Vladimir Yu. Makhnev (1), Jeremy J. Drake (1), Douglas P. Finkbeiner (1), Phillip Cargile (1) ( (1) Center for Astrophysics | Harvard & Smithsonian, (2) Ecole Polytechnique, (3) University of Southampton, (4) Kavli Institute for Astrophysics and Space Research | Massachusetts Institute of Technology)

Comments: Conference continuing for the ECML PKDD 2023Subjects: Earth and Planetary Astrophysics (astro-ph.EP); Instrumentation and Methods for Astrophysics (astro-ph.IM); Machine Learning (cs.LG)Cite as: arXiv:2309.09337 [astro-ph.EP] (or arXiv:2309.09337v1 [astro-ph.EP] for this model)Submission historyFrom: Mayeul Aubin[v1] Sun, 17 Sep 2023 17:59:59 UTC (1,787 KB)

Recommended For You