A Machine Learning Approach to Exoplanet Atmospheric Retrieval: Application to Optical Filter Ranking

Le résumé fourni par la source

Abstract Transmission spectroscopy probes the composition and structure of exoplanetary atmospheres. However, traditional retrieval methods such as Markov Chain Monte Carlo and nested sampling are computationally expensive, often taking hours or days per spectrum. In this work, we employ the random forest regressor (RFR) machine learning approach to predict the atmospheric parameters of hot Jupiters. The RFR is trained on synthetic optical transmission spectra generated by Tau Retrieval for Exoplanets ( TauREx3 ) and binned according to the transmission profiles of the Johnson–Cousins and Sloan Digital Sky Survey filters. The trained model is used to predict the planetary radius ( R p ), the equilibrium temperature ( T p ), and the mixing ratios of titanium oxide ( X TiO ) and vanadium oxide ( X VO ). We find that spectral features induced by variations in R p are significantly more prominent than those associated with other parameters. We therefore adopt a two-stage modelling strategy: we first train a model to predict R p and then fix the predicted R p when training a second model to infer the remaining parameters. This approach achieves R 2 = 0.9989 for R p , while the R 2 values for T p , X TiO , and X VO are 0.886, 0.739, and 0.821, respectively. In the benchmark test, the RFR achieves very high accuracy compared to nested sampling, but with a computational speed-up of approximately 400,000× on the same dataset. Recursive feature elimination shows that five filters perform as well as 10. This highlights the RFR model’s utility in providing reliable prior parameter estimates, especially when data is limited or boundaries are poorly defined.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.