Localization of gravitational waves using machine learning

An observation of gravitational waves is a trigger of the multimessenger search of an astronomical event. A combination of the data from two or three gravitational wave detectors indicates the location of a source and low-latency data analysis is key to transferring the information to other detectors sensitive at different wavelengths. In contrast to the current method, which relies on the matched-filtering technique, we proposed the use of machine learning that is much faster and possibly more accurate than matched filtering. Our machine-learning method is a combination of the method proposed by Chatterjee et al. and a method using the temporal convolutional network. We demonstrate the sky localization of a gravitational-wave source using four detectors; LIGO H1, LIGO L1, Virgo, and KAGRA, and compare the result in the case without KAGRA to examine the positive influence of having the fourth detector in the global gravitational-wave network.

Figure 1

Illustration of divided celestial spheres. The left sphere is divided into 18 sectors, and the right is divided into 200.

Figure 2

The structure of our models: (a) ANN architecture for method I, (b) Temporal block in a TCN, and (c) TCN architecture for method II.

Figure 3

The localization accuracy using three detectors.

Figure 4

The localization accuracy using three and four detectors. HLV means using LIGO H1, L1, and Virgo. HLVK means using LIGO H1, L1, Virgo, and KAGRA.

Figure 5

The localization accuracy for 200 sectors vs SNR.

Figure 6

The localization accuracy for 200 sectors vs declination.