New method to observe gravitational waves emitted by core collapse supernovae

While gravitational waves have been detected from mergers of binary black holes and binary neutron stars, signals from core collapse supernovae, the most energetic explosions in the modern Universe, have not been detected yet. Here we present a new method to analyse the data of the LIGO, Virgo, and KAGRA network to enhance the detection efficiency of this category of signals. The method takes advantage of a peculiarity of the gravitational wave signal emitted in the core collapse supernova and it is based on a classification procedure of the time-frequency images of the network data performed by a convolutional neural network trained to perform the task to recognize the signal. We validate the method using phenomenological waveforms injected in Gaussian noise whose spectral properties are those of the LIGO and Virgo advanced detectors and we conclude that this method can identify the signal better than the present algorithm devoted to select gravitational wave transient signal.

Figure 1

Example of a realisation of a phenomenological template. We plot the strain (upper left), the corresponding spectrogram (upper right), the coefficient of the impulsive acceleration (lower left), and the frequency of the harmonic oscillator overplotted to the spectrogram (lower right).

Figure 2

The mechanism of additive color synthesis. LIGO Hanford is assigned to red, LIGO Livingston to green and Virgo to blue. A triple coincidence will appear in white, while a double coincidence in yellow, magenta or cyan, depending on which couple of detectors is involved.

Figure 3

From the top; the spectrogram of LIGO Hanford is red, then that of LIGO Livingston is green and Virgo is blue. At the bottom: the RGB image obtained by stacking the previous three spectrograms. In this case, the signal is present just in Hanford and Livingston so that the combined signal at the bottom is in yellow.

Figure 4

Sketch of the architecture of our model.

Figure 5

Distribution of the classifier output for noise (red line) and $\mathrm{signal}+\text{noise}$ (light blue line) in the case of $\mathrm{SNR}=8$.

Figure 6

${\eta }_{\mathrm{CNN}}$ (Eq. (8) for different SNRs (blue line) computed during the validation process at threshold ${\theta }_{\mathrm{CNN}}=0.5$. The red line is the false alarm probability [${\mathrm{FAR}}_{\mathrm{CNN}}$ Eq. (8)] associated to the various SNRs computed at the same threshold. The used data set is based on 20000 signals and noise events for each SNR, half used for training and half for validation.

Figure 7

Efficiency vs false alarm for cWB (continuous line) and our method (dashed line) in the three cases $\mathrm{SNR}=12$ (blue triangles), $\mathrm{SNR}=20$ (pink squares), and $\mathrm{SNR}=40$ (brown circles).

Figure 8

Efficiency vs SNR in the case of complete cWB (continuous) and our method (dashed) for all SNRs. We report also the curve that shows the ratio between the input events of the CNN and the total injected events in function of SNR (brown). This curve sets the maximum efficiency that our method can achieve.