# Efficient Bayesian analysis of kilonovae and GRB afterglows with fiesta

**Authors:** Hauke Koehn, Thibeau Wouters, Peter T. H. Pang, Mattia Bulla, Henrik Rose, Hannah Wichern, Tim Dietrich  
**arXiv:** [2507.13807](https://arxiv.org/abs/2507.13807)  
**Date:** 2025-07-18  
**Summary page:** https://www.summarizepaper.com/en/paper/2507.13807v1/

## Key points

- Koehn et al. explore the use of machine learning (ML) surrogates to accelerate likelihood evaluations for gamma-ray burst (GRB) afterglows and kilonovae (KNe) from binary neutron star (BNS) mergers.
- Traditional methods involve sampling a Bayesian posterior over physical model parameters, which is computationally intensive due to numerous likelihood evaluations.
- The authors introduce fiesta, a python package that leverages ML surrogates for state-of-the-art GRB afterglow models and a new surrogate for KN emission modeled by possis code.
- Fiesta enables rapid evaluation of lightcurve posteriors within minutes, significantly reducing computational costs.
- Fiesta offers built-in posterior sampling capabilities using the flowMC package, efficiently scaling to higher dimensions with multiple nuisance sampling parameters.
- Utilizing the JAX framework, fiesta allows for GPU acceleration during both surrogate training and posterior sampling.
- The authors apply their framework to reanalyze AT2017gfo/GRB170817A and GRB211211A using the new pyblastafterglow model in Bayesian inference.
- While surrogates can expedite likelihood evaluations, it's essential to conduct sanity checks by running base models with best-fit parameters from the posterior to validate results in certain parameter regions where surrogate accuracy may be uncertain.

## Summary

In their article "Efficient Bayesian analysis of kilonovae and GRB afterglows with fiesta," Koehn et al. explore the use of machine learning (ML) surrogates to accelerate likelihood evaluations for gamma-ray burst (GRB) afterglows and kilonovae (KNe) that accompany binary neutron star (BNS) mergers. These electromagnetic transients play a crucial role in understanding cosmological parameters and ultra-dense matter behavior. Traditional methods involve sampling a Bayesian posterior over physical model parameters, but this process is computationally intensive due to numerous likelihood evaluations. To address this challenge, the authors introduce fiesta, a python package that leverages ML surrogates for state-of-the-art GRB afterglow models like afterglowpy and pyblastafterglow, as well as a new surrogate for KN emission modeled by possis code. These surrogates enable rapid evaluation of lightcurve posteriors within minutes, significantly reducing computational costs. Additionally, fiesta offers built-in posterior sampling capabilities using the flowMC package, which efficiently scales to higher dimensions with multiple nuisance sampling parameters. Utilizing the JAX framework, fiesta allows for GPU acceleration during both surrogate training and posterior sampling. The authors apply their framework to reanalyze AT2017gfo/GRB170817A and GRB211211A using the new pyblastafterglow model for the first time in Bayesian inference. They highlight that while surrogates can expedite likelihood evaluations, it's essential to conduct sanity checks by running base models with best-fit parameters from the posterior to validate results in certain parameter regions where surrogate accuracy may be uncertain. Overall, Koehn et al. 's work demonstrates how ML surrogates implemented in fiesta can enhance efficiency in analyzing GRB afterglows and KNe data, offering a promising approach for future studies in astrophysics research.