• astronomy
• deep learning
• early warning
• gravitational waves
• machine learning
• multi messenger astronomy
• physics

This thesis aims at building a novel early warning pipeline for BNS signals based on neural net- works. In fact, deep learning is a promising tool for fast processing of big amounts of data, as those produced by Advanced gravitational wave interferometers. The neural network I realized is a one-dimensional Convolutional Neural Network to perform predictions on interferometers’ strain time series. I have trained it on a simulated LIGO-Livingston (L1) dataset I produced using custom simulations of GW signals embedded in a realistic, colored noise. I trained the network to recognize signals within a “sliding window” that approaches the merger with discrete steps. The idea is that the network can detect chunks of the signal, potentially early enough to allow telescopes to point to the source in time (eventually combined with a localization algorithm). The project is innovative in that training the neural network with the sliding window adds a “dynamic” layer to signal detection, making it more easily adaptable to a possible future real- time pipeline. The results show how deep learning is a viable approach to tackle the problem of early warning