Supervised Machine Learning Methods with Uncertainty Quantification for Exoplanet Atmospheric Retrievals from Transmission Spectroscopy jupyter ( Kluyver et al. 2016 ) , matplotlib ( Hunter 2007 ) , numpy ( van der Walt et al. 2011 ) , plotly ( Inc. 2015 ) , scikit-learn ( Pedregosa et al. 2011 ) ,
Roy T. Forestano, Konstantin T. Matchev, Katia Matcheva, Eyup B. Unlu · 2025
Read Supervised Machine Learning Methods with Uncertainty Quantification for Exoplanet Atmospheric Retrievals from Transmission Spectroscopy jupyter ( Kluyver et al. 2016 ) , matplotlib ( Hunter 2007 ) , numpy ( van der Walt et al. 2011 ) , plotly ( Inc. 2015 ) , scikit-learn ( Pedregosa et al. 2011 ) , by Roy T. Forestano, Konstantin T. Matchev, Katia Matcheva, Eyup B. Unlu (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-04982.txt.
Introduction
Roy Forestano, Konstantin Matchev, Katia Matcheva, and Eyup Unlu benchmark machine learning for reading exoplanet air. Standard Bayesian fits of transit spectra are sound but slow. New space scopes will flood teams with spectra that need fast reads. The team tests eight regression methods on the same retrieval task.
Methods span linear fits, neighbor rules, trees, forests, and boosting stacks. Data prep steps were varied to see which pairing works best. Scores weigh accuracy, spread of errors, and run speed together. Doubt bands are mapped across the full range of planet traits. The best pairing is checked on a real scope read of the planet WASP-39b.
Catalog record
| CREATOR | Roy T. Forestano, Konstantin T. Matchev, Katia Matcheva, Eyup B. Unlu |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 55 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2508-04982
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Supervised Machine Learning Methods with Uncertainty Quantification for Exoplanet Atmospheric Retrievals from Transmission Spectroscopy jupyter ( Kluyver et al. 2016 ) , matplotlib ( Hunter 2007 ) , numpy ( van der Walt et al. 2011 ) , plotly ( Inc. 2015 ) , scikit-learn ( Pedregosa et al. 2011 ) , scipy ( Virtanen et al. 2020 ) . Roy T. Forestano, Konstantin T. Matchev, Katia Matcheva, Eyup B. Unlu Abstract Standard Bayesian retrievals for exoplanet atmospheric parameters from transmission spectroscopy, while well understood and widely used, are generally computationally expensive. In the era of the JWST and other upcoming observatories, machine learning approaches have emerged as viable alternatives that are both efficient and robust. In this paper we present a systematic study of several existing machine learning regression techniques and compare their performance for retrieving exoplanet atmospheric parameters from transmission spectra. We benchmark the performance of the different algorithms on the accuracy, precision, and speed. The regression methods tested here include partial least squares (PLS), support vector machines (SVM), k k nearest neighbors (KNN), decision trees (DT), random forests (RF), voting (VOTE), stacking (STACK), and extreme gradient boosting (XGB). We also investigate the impact of different preprocessing methods of the training data on the model performance. We quantify the model uncertainties across the entire dynamical range of planetary parameters. The best performing combination of ML model and preprocessing scheme is validated on a the case study of JWST observation of WASP-39b. I Introduction Over the last three decades, the study of extrasolar system planets has shifted from discovery to inference with particular interest in the characterization of their chemical compositions and temperature profiles. The chemical inventory of an exoplanet atmosphere is impacted by the planet formation processes, evolutionary modifications, and its interactions with the local space environment, thus allowing us to place constraints on the existing evolutionary models from the retrieved atmospheric composition. Transit spectroscopy is currently the most widely used observational technique to study the chemical composition of transiting exoplanets ( Schneider 1994 ; Charbonneau et al. 2000 ) . During transit, the planet atmosphere is observed in transmitted light when a planet passes in front of its host star, i.e., the primary eclipse, and in emitted and/or reflected light when a planet travels behind its host star, referred to as the secondary eclipse. Both transmission and emission spectra can be obtained during a transit event. In this work we focus on the analysis of transmission spectra from transiting gas giant exoplanets. As the stellar light passes through the atmosphere of the transiting planet, it is subjected to absorption and scattering from gas molecules and other atmospheric particulates including cloud particles, haze particles, etc. These particulates leave a characteristic spectroscopic signature in the observed spectrum. Thus, the spectral analysis of the transmitted stellar light allows us to determine key characteristics of the exoplanet’s atmosphere, including its temperature profile, chemical abundances, cloud opacity, and further, have the ability to uncover signs of life. The mathematical procedure for deriving the atmospheric properties from the observed spectrum is known as inversion or atmospheric retrieval. Atmospheric retrieval models are complex computational models which aim to determine the thermal structure and chemical composition of the planet atmosphere from the observed spectrum by exploring the high-dimensional parameter space to best fit the observed data ( Madhusudhan 2018 ) . The primary component of a retrieval model is a radiative transfer model (RTM) ( Waldmann et al. 2015 ; Kitzmann et al. 2020 ; Harrington et al. 2021 ; Cubillos et al. 2021 ; Blecic et al. 2021 ; Welbanks & Madhusudhan 2021 ) . RTMs take in information about the geometry of the planet-star system, including masses, radii, orbital parameters, etc., and specify the atmospheric properties, such as chemical abundances, cloud coverage, and pressure-temperature profile. Some models have the ability to incorporate more advanced aspects, such as large-scale dynamical effects ( Pluriel 2023 ) , day/night asymmetries ( Pluriel et al. 2022 ; MacDonald & Lewis 2022 ; Welbanks & Madhusudhan 2022 ) and latitudinal/longitudinal variations ( Falco et al. 2022 ) . More realistic atmospheric models are inherently more complex, which inevitably leads to a larger number of unconstrained parameters and increased computational cost. Traditional atmospheric retrieval methods are based on statistical inference through the use of sampling approaches, such as Markov chain Monte Carlo (MCMC) ( Madhusudhan & Seager 2009 ; Cubillos et al. 2013 ; Line et al. 2013 ) or Nested Sampling ( Benneke & Seager 2012 ; Waldmann et al. 2015 ; Oreshenko et al. 2017 ; Gandhi & Madhusudhan 2018 ) , to perform Bayesian parameter estimation ( Nixon & Madhusudhan 2020 ) . The complexity of Bayesian retrievals typically scales with the number of model parameters, therefore, by increasing the number of chemical constituents or adding geometrical dimensions such as latitudinal or longitudinal variability, run-times can increase dramatically. Furthermore, due to possible degeneracies in the atmospheric parameters one needs to fully explore the parameter space and find all viable solutions, which adds to the computational challenge. With the James Webb Space Telescope (JWST) ( Greene et al. 2016 ) online and with its successor, the Roman Telescope, due to launch in 2027, as well as a number of exoplanet-specific prospective launches, including the Twinkle Space Telescope ( Edwards et al. 2019b ) and the European Space Agency (ESA) Ariel mission ( Tinetti et al. 2021 ) , the amount of spectral data observed from transiting exoplanets is expected to increase rapidly. The large number of observations and the improved spectral resolution pose further computational demands on the existing atmospheric retrieval methods and call for the development of alternative, more efficient atmospheric retrieval techniques. Recently, supervised machine learning (ML) techniques have been employed as alternatives to the standard atmospheric retrieval methods. A number of supervised ML approaches have been tried to obtain atmospheric parameters from the observed spectra: deep belief networks ( Waldmann 2016 ) , deep neural networks ( Yip et al. 2021 ) random forests ( Márquez-Neila et al. 2018 ; Oreshenko et al. 2020 ; Fisher et al. 2020 ; Nixon & Madhusudhan 2020 ; Guzmán-Mesa et al. 2020 ) , generative adversarial networks (GANs) ( Zingales & Waldmann 2018 ) , convolutional neural networks (CNNs) ( Soboczenski et al. 2018 ; Yip et al. 2021 ; Ardevol Martinez et al. 2022 ) , recurrent neural networks (like LSTMs) ( Yip et al. 2021 ) , an ensemble of Bayesian Neural Networks ( Cobb et al. 2019 ) , transformer-nspired deep learning architectures ( Unlu et al. 2023 ) , conditional invertible neural networks (cINNs) ( Haldemann et al. 2023 ) , normalizing flows ( Aubin et al. 2023 ; Yip et al. 2024 ) , sequential neural posterior estimation ( Ardévol Martínez et al. 2024 ) and simulation-based inference ( Lueber et al. 2025 ) . While this plethora of applicable approaches is a virtue, it also introduces the issue of epistemic uncertainty, which reflects how well the model has learned the underlying patterns in the data. Different ML architectures in principle give slightly different answers, thus contributing to the overall uncertainty of the retrievals ( Yip et al. 2021 ) . At