Creating a deep learning algorithm to detect unexpected gravitational wave events

Since the first direct detection of gravitational waves in 2015, scientists have resorted to a bit of a shortcut: they can only detect waves that match theoretical predictions, which is actually the opposite of how science usually works.

A group of physicists have presented a computational model that can capture not only expected gravitational waves, but all gravitational waves passing through the Earth. The paper states: arXiv Preprint server.

Decades after Einstein discovered that general relativity predicted gravitational waves—ripples that travel through the fabric of space-time—physicists had calculated the expected characteristics of several simple scenarios. One of these, a passing waveform from a black hole merger, was the first gravitational wave detected in interferometric data received on September 14, 2015 (though the paper wasn't published until February of the following year).

By imagining the event that generated the waves, gravity scientists were able to predict the exact signals that would appear at LIGO (two in the US), VIRGO in Italy, and several other long-arm laser interferometer facilities around the world.

Observers needed to know what to expect in order to train the interferometer on what to look for: a passing wave would only move the interferometer's arms by a thousandth of a proton's width. Environmental noise, even a passing truck, could easily cause the arms to move, and that needed to be filtered out in order to distinguish it from real gravitational waves.

Calculations were also performed for neutron star-black hole mergers and neutron star-neutron star mergers. From the data, it was also possible to glean signatures of continuous gravitational waves generated by rapidly rotating symmetric neutron stars, as well as stochastic gravitational waves, for example from the Big Bang. In total, over 70 gravitational wave events were detected using these models.

But this method misses gravitational waves that don't show up in any of the known predictions – so-called “transients” or “gravitational wave bursts” that arise from unexpected events based on different physics. What's more, current detection methods are too slow.

After a gravitational wave passes through, astronomers hope to quickly pinpoint its source and alert other observatories to look for accompanying electromagnetic waves or particle phenomena emanating from the same source, a process known as multi-messenger astronomy.

Large-scale, intense astrophysical activity, including typical binary star mergers, is expected to produce electromagnetic radiation, including visible light, and neutrinos. Receiving a potential gravitational wave train would require hundreds of dedicated processing units to process and communicate with other instruments, and could take tens of seconds or even minutes – far too late for any “advance warning.”

In recent years, physicists have sought to improve on the waveform constraints by using a specialized type of deep learning algorithm, convolutional neural networks (CNNs), to circumvent detectors that are trained to only recognize certain events.

However, CNNs programmed to date require accurate models of the target signal for training, and therefore cannot notice unexpected sources, such as those expected from supernova core collapses or long gamma-ray bursts. Both unknown physics and computational limitations could undermine the possibility of multi-messenger detection.

Here, the researchers set a goal of reporting gravitational wave events in about one second using a single processor. They developed a multi-component architecture in which one CNN detects transients occurring simultaneously across multiple detectors, and a second CNN looks for correlations between the detectors to remove background noise and glitches that occur simultaneously.

“Thus, our search aims to harness machine learning to point 'conventional' telescopes at such sources within a matter of seconds,” said Vasileos Skurilis of the Gravitational Probes Laboratory in the School of Physics and Astronomy at Cardiff University in Wales, UK. “In this way, we will be able to extract as much information as possible from these unexpected events.”

The research group's deep learning approach was crucially different from previous methods: instead of teaching a CNN to identify specific signal shapes in the data, they created a CNN that could detect consistency in intensity and timing between two or more data streams.

The CNN was then trained using simulated signals and random noise bursts with similar characteristics. Using the same waveform patterns for both the signal and the noise prevented the CNN from relying on the signal's pattern to make its decisions. Instead, the CNN learned to evaluate how well the detectors matched each other, allowing the model to detect gravitational wave transients in true real-time.

As a test, they ran the data observed in the first two runs of LIGO and VIRGO and found good agreement.

“In the 1960s, when gamma-ray astronomy took its first steps, gamma-ray bursts were a new marvel in astrophysics,” Skrillis said. “Gravitational wave astronomy is similarly in its infancy, and an exciting future may lie ahead of us.”

© 2024 Science X Network