Gravitational waves: a normalizing flood estimates the probability of detection of LIGO, Virgo and KAGRA without costly injection campaigns

Trained on public injection data from the O4 run, the emulator reproduces detection efficiencies and is sometimes more accurate than Monte Carlo sampling.

To study the population of merging black holes and neutron stars, astrophysicists need to know what fraction of the sources would have been detected by interferometers. This calculation is usually based on large-scale simulated signal injection campaigns, processed by Monte Carlo, which are very computationally intensive. A team including Antonella Palmese on arXiv proposes to entrust this task to a normalizing flow, a generative neural network, trained on the injection dataset published for the fourth observation run (O4) of the LIGO/Virgo/KAGRA network.

More precise on atypical populations

The emulator correctly finds the distribution of parameters of the detected events, and its detection efficiency estimates agree with those of conventional preferential sampling. For population models with abrupt variations, it can even be more accurate. The authors show that it fits into a complete population analysis and make the training code and trained models freely available.