A fast deep-learning approach to probing primordial black hole populations in gravitational wave events

Abstract

Primordial black holes (PBHs), envisioned as a compelling dark matter candidate and a window onto early-Universe physics, may contribute to some of the gravitational-wave (GW) signals detected by the LIGO-Virgo-KAGRA network. Traditional hierarchical Bayesian analysis, which relies on precise GW-event posterior estimates to extract information on potential PBH populations from GW events, becomes computationally demanding for catalogs with a large number of events. Here, we present a fast deep-learning framework, leveraging Transformer and normalizing flows, that maps GW-event posterior samples to joint posterior distributions over the hyperparameters of the PBH population. Our approach yields credible intervals with acceptable accuracy while delivering an order-of-magnitude speedup. These results highlight the potential of deep learning for fast and accurate PBH population studies, and its applicability to next-generation GW detectors when combined with appropriate event-level inference models.