Accurate Machine-learning Atmospheric Retrieval via a Neural-network Surrogate Model for Radiative Transfer

Abstract

Atmospheric retrieval determines the properties of an atmosphere based on its measured spectrum. The low signal-to-noise ratios of exoplanet observations require a Bayesian approach to determine posterior probability distributions of each model parameter, given observed spectra. This inference is computationally expensive, as it requires many executions of a costly radiative transfer (RT) simulation for each set of sampled model parameters. Machine learning (ML) has recently been shown to provide a significant reduction in runtime for retrievals, mainly by training inverse ML models that predict parameter distributions, given observed spectra, albeit with reduced posterior accuracy. Here we present a novel approach to retrieval by training a forward ML surrogate model that predicts spectra given model parameters, providing a fast approximate RT simulation that can be used in a conventional Bayesian retrieval framework without significant loss of accuracy. We demonstrate our method on the emission spectrum of HD 189733 b and find good agreement with a traditional retrieval from the Bayesian Atmospheric Radiative Transfer (BART) code (Bhattacharyya coefficients of 0.9843–0.9972, with a mean of 0.9925, between 1D marginalized posteriors). This accuracy comes while still offering significant speed enhancements over traditional RT, albeit not as much as ML methods with lower posterior accuracy. Our method is ∼9× faster per parallel chain than BART when run on an AMD EPYC 7402P central processing unit (CPU). Neural-network computation using an NVIDIA Titan Xp graphics processing unit is 90×–180× faster per chain than BART on that CPU.

Neural posterior estimation for exoplanetary atmospheric retrieval

“…This comes however at the expense of posterior accuracy, as the resulting parameter distributions are either not true posteriors (Márquez-Neila et al 2018), or are enforced to follow a multivariate Gaussian distribution (Zingales & Waldmann 2018;Nixon & Madhusudhan 2020;Ardévol Martínez et al 2022). Another approach is to use machine learning to generate more informed, narrower priors (Hayes et al 2020), or even to replace the exoplanet atmosphere simulator by a surrogate model (Himes et al 2022). These methods have the potential of providing more accurate posterior distributions, but at the expense of a more modest improvement in terms of inference time.…”

Toward 3D Retrieval of Exoplanet Atmospheres: Assessing Thermochemical Equilibrium Estimation Methods

“…Recently, Himes et al (2022) presented a novel machinelearning approach to exoplanet atmospheric retrieval, where the radiative transfer forward model is replaced with a neural network (NN) surrogate model. We found that this approach achieves similar quantitative results as the classical approach but at a fraction of the computational cost.…”

Transverse Vector Decomposition Method for Analytical Inversion of Exoplanet Transit Spectra

“…Classical inversion methods extensively used in analyzing infrared spectra from remote sensing of the Earth and solar system planets are described in Hanel et al (2003) and are based on linearizing the problem. More recent numerical approaches include sampler-based retrievals (Lee et al 2012;Line et al 2013bLine et al , 2013aWaldmann et al 2015;Lavie et al 2017;Blecic et al 2022;Cubillos et al 2022;Harrington et al 2022) and various machine learning (ML) techniques (Waldmann 2016;Márquez-Neila et al 2018;Soboczenski et al 2018;Zingales & Waldmann 2018;Cobb et al 2019;Fisher et al 2020;Himes et al 2020a;Oreshenko et al 2020;Himes et al 2020b;Guzmán-Mesa et al 2020;Nixon & Madhusudhan 2020;Yip et al 2021;Ardevol Martinez et al 2022;Himes et al 2022; for a recent comparative review, see . The main disadvantages of the numerical ML approach are: (i) the ML model is in principle a black box that hides the relevant physics and is not easily interpretable (Nixon & Madhusudhan 2020;Yip et al 2021); (ii) the ML model does not learn directly the relevant physics, but only the (finite amount of) data generated by the forward model, which introduces additional uncertainties during the training process (Matchev et al 2022a).…”