Listening to Black Holes with Neural Networks

Just over a decade after the first detection, the LIGO-Virgo-KAGRA Collaboration has reported 390 observations of merging black holes or neutron stars. Each is a faint chirping signal buried in detector noise. By comparing the data against precise signal models from general relativity, one can obtain the masses, spins, distance, and sky localization of the source. With standard Bayesian inference techniques, this can take from hours to days per event. This is becoming a bottleneck, with next-generation detectors expected to see hundreds of thousands of events per year. I will describe our approach to this problem using simulation-based inference (SBI): neural networks trained on simulated data that can return the full posterior distribution for a real event in seconds. I will explain how we make these results reliable by correcting the network against the exact likelihood. The speed of SBI enables real-time localization for multi-messenger astronomy and large-scale analyses that are otherwise impractical. SBI needs only the ability to simulate the data, not a closed-form likelihood, so it also opens a path to inference in realistic detector noise, including transient glitches. I will conclude by showing how SBI extends to populations, combining entire catalogs of events to learn the astrophysical distribution of binaries and measure the expansion of the universe.