Neural posterior estimation for exoplanetary atmospheric retrieval

Abstract

Context. Retrieving the physical parameters from spectroscopic observations of exoplanets is key to understanding their atmospheric properties. Exoplanetary atmospheric retrievals are usually based on approximate Bayesian inference and rely on sampling-based approaches to compute parameter posterior distributions. Accurate or repeated retrievals, however, can result in very long computation times due to the sequential nature of sampling-based algorithms. Aims. We aim to amortize exoplanetary atmospheric retrieval using neural posterior estimation (NPE), a simulation-based inference algorithm based on variational inference and normalizing flows. In this way, we aim (i) to strongly reduce inference time, (ii) to scale inference to complex simulation models with many nuisance parameters or intractable likelihood functions, and (iii) to enable the statistical validation of the inference results. Methods. We evaluated NPE on a radiative transfer model for exoplanet spectra (petitRADTRANS), including the effects of scattering and clouds. We trained a neural autoregressive flow to quickly estimate posteriors and compared against retrievals computed with MultiNest. Results. We find that NPE produces accurate posterior approximations while reducing inference time down to a few seconds. We demonstrate the computational faithfulness of our posterior approximations using inference diagnostics including posterior predictive checks and coverage, taking advantage of the quasi-instantaneous inference time of NPE. Our analysis confirms the reliability of the approximate posteriors produced by NPE. Conclusions. The inference results produced by NPE appear to be accurate and reliable, establishing this algorithm as a promising approach for atmospheric retrieval. Its main benefits come from the amortization of posterior inference: once trained, inference does not require on-the-fly simulations and can be repeated several times for many observations at a very low computational cost. This enables efficient, scalable, and testable atmospheric retrieval.