Identification of Lensed Gravitational Waves with Deep Learning. (arXiv:2010.12093v2 [gr-qc] UPDATED)
<a href="http://arxiv.org/find/gr-qc/1/au:+Kim_K/0/1/0/all/0/1">Kyungmin Kim</a>, <a href="http://arxiv.org/find/gr-qc/1/au:+Lee_J/0/1/0/all/0/1">Joongoo Lee</a>, <a href="http://arxiv.org/find/gr-qc/1/au:+Yuen_R/0/1/0/all/0/1">Robin S. H. Yuen</a>, <a href="http://arxiv.org/find/gr-qc/1/au:+Hannuksela_O/0/1/0/all/0/1">Otto Akseli Hannuksela</a>, <a href="http://arxiv.org/find/gr-qc/1/au:+Li_T/0/1/0/all/0/1">Tjonnie G. F. Li</a>
The propagation path of gravitational waves is expected to be bent near
massive astrophysical objects. The massive object acts as a lens. Similarly to
the lensing of electromagnetic waves, the lens amplifies gravitational waves’
amplitude and can produce multiple gravitational-wave images. If we suppose the
positions of lens and source of a gravitational wave deviate from the line of
sight, the gravitational-wave images arrive at different times because they
have traveled different trajectories around the lens at the same speed.
Depending on the difference in their arrival times, multiple gravitational
waves can be detected as repeated, near-identical events, or superposed
gravitational waves with characteristic “beating patterns”. In particular, when
the lens is small, $lesssim 10^5 M_odot$, the lens produces images with short
time delays that result in the beating patterns. We utilize deep learning to
study the lensing signature. It is known that many state-of-the-art deep
learning models are excellent at recognizing foreground images, similar to
spectrograms, from background noises. In this work, we study the feasibility of
applying deep learning to identify lensing signatures from the spectrogram of
gravitational-wave signals detectable by the Advanced LIGO and Virgo detectors.
We assume the lens mass is around $10^3 M_odot$ — $10^5 M_odot$ which can
produce the order of millisecond time delays between two images of lensed
gravitational waves. We discuss the feasibility of two aspects: distinguishing
lensed gravitational waves from unlensed ones and estimating the parameters
related to not only the lensing factor but also the source binary system and
lens. We suggest that the approach of this work would be of particular interest
for more complicated lensings for which we do not have accurate waveform
templates.
The propagation path of gravitational waves is expected to be bent near
massive astrophysical objects. The massive object acts as a lens. Similarly to
the lensing of electromagnetic waves, the lens amplifies gravitational waves’
amplitude and can produce multiple gravitational-wave images. If we suppose the
positions of lens and source of a gravitational wave deviate from the line of
sight, the gravitational-wave images arrive at different times because they
have traveled different trajectories around the lens at the same speed.
Depending on the difference in their arrival times, multiple gravitational
waves can be detected as repeated, near-identical events, or superposed
gravitational waves with characteristic “beating patterns”. In particular, when
the lens is small, $lesssim 10^5 M_odot$, the lens produces images with short
time delays that result in the beating patterns. We utilize deep learning to
study the lensing signature. It is known that many state-of-the-art deep
learning models are excellent at recognizing foreground images, similar to
spectrograms, from background noises. In this work, we study the feasibility of
applying deep learning to identify lensing signatures from the spectrogram of
gravitational-wave signals detectable by the Advanced LIGO and Virgo detectors.
We assume the lens mass is around $10^3 M_odot$ — $10^5 M_odot$ which can
produce the order of millisecond time delays between two images of lensed
gravitational waves. We discuss the feasibility of two aspects: distinguishing
lensed gravitational waves from unlensed ones and estimating the parameters
related to not only the lensing factor but also the source binary system and
lens. We suggest that the approach of this work would be of particular interest
for more complicated lensings for which we do not have accurate waveform
templates.
http://arxiv.org/icons/sfx.gif
