WorksPhysics and Astronomy
Interpretable attention-based neural posterior estimation for fast radio burst cosmology
Abstract
Fast radio bursts (FRBs) probe the integrated free-electron density along extragalactic sightlines and can be used to study late-time cosmic evolution. Conventional Bayesian analyses require repeated evaluations of computationally expensive likelihoods, making inference increasingly challenging as FRB catalogs grow. In this work, we develop an interpretable attention-based neural posterior estimator (AttNPE) for rapid posterior inference of the physical parameters from FRB catalogs with measured redshifts. The attention encoder compresses the catalog into a low-dimensional representation, and a conditional masked autoregressive flow maps samples from a base distribution to the seven-dimensional posterior. AttNPE yields posterior distributions in close agreement with Markov chain Monte Carlo (MCMC), recovering the parameter correlations in both simulated and observed FRB catalogs. For the observed catalog, AttNPE generates posterior samples in ~s after training, compared with ~h for the reference MCMC run, giving an elapsed-time ratio of . The attention-weight analysis shows that intermediate-redshift FRBs () account for the largest share of the total attention averaged over simulated catalogs following the observed redshift distribution, consistent with the combined influence of source abundance and redshift-dependent sensitivity. These results demonstrate the potential of interpretable neural posterior estimation for efficient and repeated inference in FRB cosmology.