Neural Network Surrogate Models for Fast Bayesian Inference: Application to Exoplanet Atmospheric Retrieval

Abstract

Exoplanet atmospheres are characterized via retrieval, the inverse modeling method where atmospheric properties are determined based on the exoplanet's observed spectrum. To determine the posterior probabilities of model parameters consistent with the data, a Bayesian framework proposes atmospheric models, calculates the theoretical spectra corresponding to the models via radiative transfer (RT), and compares the spectra with the observed spectrum. This typically requires thousands to millions of evaluated models, with each taking on the order of a second for RT. While recent machine-learning approaches to retrieval reduce the compute cost to minutes or less, they do so at the cost of reduced posterior accuracy. Here we present a novel machine-learning assisted retrieval approach which replaces the RT code with a neural network surrogate model to significantly reduce the compute cost of RT simulations, while retaining the Bayesian framework. Using emission data of HD 189733 b, we demonstrate close agreement between this method and that of the Bayesian Atmospheric Radiative Transfer (BART) code (mean Bhattacharyya coefficient of 0.9925 between 1D marginalized posteriors). This approach is  9x faster per parallel evaluation than BART when using an AMD EPYC 7402P central processing unit (CPU), and it is 90–180x faster per parallel evaluation when using an NVIDIA Titan Xp graphics processing unit than BART on that CPU.