Deep-Learning for Midband Gravitational-Wave Data Analysis

Proposed mid-band gravitational-wave detectors will help bridge the gap between LIGO and LISA, thus providing more opportunities for gravitational-wave science. Inspirals that may take place in the decihertz band may also be eccentric if a binary forms dynamically. Here, we investigate the efficacy of machine learning in the detection and parameter estimation of mid-band gravitational-wave signals. We train deep neural networks on a catalog of waveforms for different chirp masses and eccentricities with durations of order several minutes. We then carry out injection studies in Gaussian noise, where we evaluate both a network that predicts whether or not a given data segment contains a gravitational-wave signal and a network that estimates the parameters of an injected gravitational-wave signal. We find that our trained networks achieve few-percent-level accuracy, showing promise as a practical method for the efficient detection of lower-frequency gravitational waves.