Title:Residual neural networks to classify the high frequency emission in core-collapse supernova gravitational waves

Abstract:We present the results of a detailed study on the detectability of the High Frequency Feature (HFF) in core-collapse supernova (CCSN) gravitational wave (GW) signals. We applied Residual Neural Networks (ResNet50), one of the state-of-the-art deep learning architectures in computer vision, to perform a multi-class classification of image samples built from time-frequency Morlet wavelet scalograms of LIGO-Virgo noise plus CCSN GW signals. We consider three target labels for three consecutive and mutually exclusive intervals in which the (first-order) slope of the HFF can be located. We optimized, trained, and tested the ResNet50 model with phenomenological waveforms. Next, we tested the optimized ResNet50 model with GW signals from CCSN simulations. At galactic distances of $1$Kpc and $5$Kpc with H1 and L1 data and $1$Kpc with V1 data, we obtained highly accurate results (test accuracies from $0.8933$ to $0.9867$), which show the feasibility of our methodology. In the case of further distances, we observed declines in test performance until $0.8000$ with H1 and L1 data at $10$Kpc and until $0.5933$ with V1 data at $10$Kpc. Without assuming the continuity and/or discontinuity of the HFF slope values, our methodology is general enough to address, at an early stage, the characterization of the HFF.