How Fast Can Gravitational Wave Detection Get?

[gingerbread80]

Claudio Giovannini/AFP/Getty Images

On September 14, 2015, gravitational waves from the smashup of two black holes reached Earth. And for the first time, a scientific instrument was ready to detect this type of signal. This was a massive discovery: Astronomers had never before seen the undulations of spacetime—that stuff in which everything in the universe swims. The detection validated a prediction Albert Einstein had made a century before and resulted in a Nobel prize. But it took five months for scientists to vet, validate, and interpret that discovery confidently enough that they could go public, which they finally did in February 2016.

Gravitational waves ripple through the universe every time an object with mass accelerates. Jump up and down: Congratulations, you made some. Your puny waves may be too weak to register with LIGO, the Laser Interferometer Gravitational-Wave Observatory, but LIGO can pick up the stretching and squeezing of spacetime that happens when, say, black holes crash into each other.

And it has: Astronomers with LIGO and its new collaborating instrument Virgo have now caught waves from six violent cosmic collisions. Today, the team is optimizing detection algorithms, and developing new tools through machine learning, so that they can shrink the time between when a wave bumps into Earth and when earthly astronomers around the world—and you—know about it.

Identifying gravitational wave candidates more quickly will improve astronomers' ability to study the phenomena that produce them. After all, LIGO isn’t a “Can we find a gravitational wave? y/n” machine, or at least it’s not just that. Scientists want to know what those gravitational waves reveal about black holes, neutron stars, pulsars, cataclysmic collisions, quick orbits, and, you know, the nature of space and time.

Key to this understanding is interpreting LIGO’s observations in the context of electromagnetic data—visible light, infrared radiation, radio waves, X-ray emissions, etc. But to capture that light, astronomers have to know when a gravitational wave event is happening, so that they can tell telescopes to turn toward the right spot in space.

That kind of data collection—called “multi-messenger astronomy”—happened for the first time in August 2017. Two neutron stars spiraled toward and ultimately crashed into each other, releasing not just gravitational waves for LIGO but also gamma rays, X-rays, ultraviolet and infrared radiation, visible light, and radio waves for telescopes, which got word of the gravitational waves in time to collect data. Getting a quicker jump on these signals means scientists can see the phenomena in action, and then interpret that action, rather than just its aftermath (or nothing at all).

The LIGO collaboration has already accelerated how fast it can send “Hey, look over there!” alerts. For the first two discoveries, there was a more than day-long gap between when waves arrived and when notices went to a network of scientists. For the inspiraling neutron stars, though, that lag was just over 30 minutes.

It would happen a lot faster if it weren't for humans, who still weigh in on the software's analysis. “The automated part can be done as quickly as 15 seconds at this point,” says Chad Hanna, an astrophysicist at Penn State. Hanna works specifically on software that finds gravitational waves from “compact binary” mergers—objects like black holes or neutron stars smashing together. He and other LIGOers have created templates—models of how gravitational waves from these collisions should look—for a range of collision scenarios. “Then we go looking for that exact thing,” says Hanna (or, really, those many, many possible exact things). The technique is formally called "matched filtering."

Other groups develop data pipelines to search for the continuous waves, the chorus of background waves, and bursty waves of unknown origin. Sergey Klimenko, of the University of Florida, works on the unknowns. “We developed a method that does not require you to know in advance what are the properties of your source,” he says. “It makes sense because nature may have some surprises for us.” This method looks for spacetime signals that show up the same at the LIGO site in Louisiana, the one in Washington state, and, recently, a Virgo location in Pisa.

Once software finds a candidate, the algorithms give that event a ranking—how likely is it to be real? If a candidate exceeds a threshold of possibility, it wends its way into a database.

That’s the part that happens pretty fast. But after that comes the real slow-down: people.

LIGO and Virgo scientists on an alert channel first get a text, phone call, or email about the candidate, and then they convene for a telecon (searching for gravitational waves may seem special, but it turns out it’s a lot like any other job). Together, humans vet the data and confirm or deny the computers’ suspicions. That ups the official-alert time, right now, to around 30 minutes.

Maybe, though, the team doesn’t need that phone call anymore. Maybe the algorithms are good enough, and the stakes of misidentifying a wave here or there have lowered enough, to trust the software. The alerts could potentially go out automatically, and telescopes can slew to the spot if they want, knowing it might be nothing. “People who want to gamble can follow up right away,” says Hanna.

The already-quick algorithms are also getting faster—moving toward real-time detection—and they might someday alert more than just astronomers. “Everybody should be able to see them if they’re interesting,” says Klimenko.

Making that data pipeline flow better isn’t always pretty, though. When you want to update software that affects a huge fraction of the world’s astronomers, it’s a process. “I wouldn’t really call it bureaucratic,” says Hanna. “It’s academic—people getting on the phone and arguing about things and hashing it out.”

Which means that changing the game isn’t simple—especially when it’s working fine and getting better.

But a PhD student named Daniel George has a new possible way to identify gravitational wave candidates, which he says can be even faster than the current methods, while also sensing signals they might miss. “The idea is to take all the templates they’ve used for template matching and use that to train a neural network instead,” he says. Neural networks use examples—in this case, of gravitational wave data—to learn to perform some task—in this case, to recognize and glean information about gravitational waves.

George recently published a paper about his method, which he calls Deep Filtering. The research suggests that, using a graphical processing unit (GPU) and Deep Filtering, you can process data more than a thousand times faster than with matched filtering. And, while matching templates to the data requires that a template match the waves pretty well, Deep Filtering doesn’t.

Take, as one always does on the internet, pictures of cats. If you wanted to use matched filtering to identify feline portraits, your cat-templates would need to be pretty close to the catalog of images you were scanning for more cats. But if you train a neural network with a bunch of different kinds of cats shown at a bunch of different angles against a bunch of different backgrounds, the network can generalize: It can recognize a kind of cat it’s never seen in a situation it’s never encountered. The same, says George, is true of Deep Filtering and gravitational waves—so even if it were fed, for example, only black holes in circular orbits around each other, it could teach itself to recognize black holes in really elliptical orbits around each other.

Matched filtering, though, does provide more statistical information about a gravitational wave event and the system it came from. So George says the LIGO collaboration plans to use Deep Filtering in the next observing run, not instead of but in combination with the matched filtering.

If scientists use ultra-fast Deep Filtering to trigger an alert and get some initial ideas of where a wave came from—from two black holes around 30 times the mass of the Sun, for instance—in real time, then matched filtering can then kick in. And it can focus on templates that involve two black holes around 30 times the mass of the Sun, rather than running through the multitude of cosmic possibilities.

About 10 different groups within LIGO are working on machine learning, according to George. And the LIGO team already uses this kind of computation for some analyses—specifically for weeding out false signals, like jostles from a nearby logging operation in Louisiana or seismic activity. “[Machine learning] will play an ever increasing role in the coming years,” says Hanna. “While it may not revolutionize our ability to detect gravitational wave signals in the near-term—since our present statistical techniques are already quite sophisticated—it is important to conduct machine learning research now with the hope of future payoff."

And gravitational wave astronomy is just getting started, so the future is sure to be long, vibrant, and uncertain. Imagine a time when we only knew about six exoplanets! Or had only seen six supernovae! Our understanding of planets and dying stars has evolved hardcore since those first discoveries. The same will be true of gravitational waves, so the faster scientists find and identify them the better. “They’re precious things,” says Hanna. “Every one tells us something new about the universe.”

 

Source