Neural posterior estimation for exoplanetary atmospheric retrieval

Retrieving the physical parameters from spectroscopic observations of exoplanets is key to understanding their atmospheric properties. Exoplanetary retrievals are usually based on approximate Bayesian inference and rely sampling-based approaches compute parameter posterior distributions. Accurate or repeated retrievals, however, can result in very long computation times due sequential nature algorithms. We aim amortize exoplanetary retrieval using neural estimation (NPE), a simulation-based algorithm variational normalizing flows. In this way, we (i) strongly reduce time, (ii) scale complex simulation models with many nuisance intractable likelihood functions, (iii) enable statistical validation results. evaluate NPE radiative transfer model for exoplanet spectra petitRADTRANS, including effects scattering clouds. train autoregressive flow quickly estimate posteriors compare against computed MultiNest. produces accurate approximations while reducing time down few seconds. demonstrate computational faithfulness our diagnostics predictive checks coverage, taking advantage quasi-instantaneous NPE. Our analysis confirms reliability produced by The accuracy results establishes it as promising approach retrievals. Amortization makes several computationally inexpensive since does not require on-the-fly simulations, making efficient, scalable, testable.