Long short-term memory for early warning detection of gravitational waves

Pre-merger detection of gravitational waves during the early inspiral of compact binary coalescences would enable electromagnetic observations of the earliest merger stages. This would significantly impact multi-messenger astronomy, giving astronomers potential access to rich new information. Here, we introduce a proof-of-concept deep-learning-based approach to produce early-warning alerts for binary black hole systems. We show the possibility of using a long short-term memory network trained on the whitened detector strain in the time domain to detect and classify compact binary events. In this work, we consider a single advanced laser interferometer gravitational-wave observatory detector at design sensitivity and make approximate sensitivity and early warning capability comparisons with approximations to traditional matched filtering approaches. We find that our model is competitive in both aspects, and when applied to a simulated test dataset was able to produce an early alert up to 5.3 s before the merger at a fixed false-alarm rate of one per day. These results demonstrate the feasibility of lightweight, low-latency recurrent neural networks for rapid gravitational-wave discovery, providing a pathway toward real-time early-warning systems for multi-messenger follow-up. This proof-of-concept in Gaussian noise for a single detector is readily extendable to real multi-detector observations.