Rapid and robust simulation-based inference for kilonovae

TL;DR

This paper introduces a simulation-based inference framework for kilonova parameter estimation that is faster and more robust than traditional MCMC methods, especially under modeling uncertainties.

Contribution

The authors develop a Gaussian process emulator-based SBI framework that provides rapid, accurate inference for kilonovae, overcoming limitations of likelihood-based methods.

Findings

SBI accurately recovers injected parameters in simulated data.

SBI produces posterior predictive light curves consistent with data.

MCMC suffers from systematic bias due to likelihood misspecification.

Abstract

With the next generation of both electromagnetic and gravitational wave observatories beginning to come online, rapid analysis methods for kilonova data are becoming increasingly important in astronomy. Traditional Bayesian parameter estimation using Markov chain Monte Carlo (MCMC) is time-consuming and relies on explicit likelihood approximations that can break down when modeling uncertainties are significant. We develop a simulation-based inference (SBI) framework for kilonova parameter estimation using density-estimation likelihood-free inference. The framework uses a Gaussian process emulator trained on $\sim1300$ radiative transfer simulations generated with the POSSIS code. We demonstrate that SBI provides a rapid alternative to MCMC for inference with emulators or approximate likelihoods that is robust to emulator uncertainty and likelihood misspecification. On simulated data, the SBI method accurately recovers injected parameters and produces posterior predictive light curves consistent with the data, but the MCMC posterior recovery suffers from systematic bias caused by likelihood misspecification. When analyzing AT2017gfo, the SBI and MCMC methods yield similar light-curve predictions but different posterior distributions, with a subset of the MCMC posteriors piling up at prior boundaries. The likelihood in the MCMC fails to capture the non-Gaussian, correlated structure of the emulator uncertainty, but SBI learns the posterior directly from forward simulations that include the full predictive distribution. Once trained, the SBI framework generates $\sim2\times10^4$ posterior samples in seconds.