Sign In

Communications of the ACM

ACM News

The Power of AI in the Sky


Artist's rendition of the Low Frequency Aperture Array of the planned Square Kilometre Array.

One of the largest big data projects on the horizon is an international effort to build the world's biggest radio telescope, with a collecting area measuring one square kilometer (one million square meters).

Credit: SKA

Picture the universe, an ethereal mix of planets, moons, stars, galaxies, dark matter, and, perhaps, civilizations other than our own. The picture is intriguing and the possibility of finding new objects in the cosmos is emerging through the implementation of sophisticated artificial intelligence and big data processing.

Many astronomy projects already use artificial intelligence (AI), machine learning, and big data to a limited extent, but that is about to change as new surveys of the universe get under way and deliver up to an exabyte of data a day that must be analyzed and classified.

One of the largest big data projects on the horizon is the Square Kilometre Array (SKA), an international effort to build the world's biggest radio telescope, with a collecting area measuring one square kilometer (one million square meters). The first of two phases of construction is due to start in 2021 or 2022. The project involves 12 member countries and will implement telescopes in areas of Australia and Africa with small populations and low radio interference. SKA could make a first survey of the entire sky within a year, thousands of times faster than any other system in existence.

At a macro level, the project's goals are to research where the universe came from and where it is going. More specific aims include understanding why expansion of the universe is accelerating, when expectations were that it would decelerate as gravity pulls things together;

And, looking for extraterrestrial life.

Precursors to SKA, like the MeerKAT radio telescope project in South Africa and ASKAP radio telescope project in Australia, are under way, with MeerKAT having set up its first seven dishes and produced early pictures, and ASKAP starting an early science program in late 2016 using an array of 12 antennas to image the sky.

Bruce Bassett, professor of mathematics at the University of Cape Town and head of data science for SKA in Africa, says AI will be critical to extracting the full value of data created by the SKA project. He envisages AI being used to discover new classes of objects that have never been seen before, while machine learning will classify images of celestial objects that can then be counted, and any changes observed. Using telescopes that are so sensitive they can see a cellphone on a star, machine learning will also be used to detect whether a signal is cosmic or manmade so that any radio frequency interference can be cleaned up.

With phase one of SKA expected to start with 64 telescopes and add to a total of 215, and phase 2, which has yet to be funded but may start in the second half of the 2020s, planned to add 10 times as many telescopes and deliver an exabyte of data a day, Bassett says, "More time, more telescopes, and more AI mean better analysis of astronomical images. There is also a case to find ways that allow AI to make a bigger contribution."

Looking forward, and questioning whether AI will ever be able to do fundamental research, he says: "On a 10-year timescale, AI could get so good it could contribute to research, perhaps being able to explore and come up with innovative ideas, investigate their viability, and disseminate them."

Alan Heavens, head of the Astrostatistics group at Imperial College London, agrees that the SKA project will be the biggest data challenge in astronomy, but says AI without human intervention has yet to be achieved in astronomy. Machine learning, however, is proving successful with algorithms able to find particular types of objects of interest in images relayed from cameras in space. Neural networks can also be trained to find specified objects.

Heavens explains, "We have a physical model of the behavior of the universe on a large scale. We know what we are looking for and guide data analysis rather than start with a blank slate and use AI to tell us what it has learned. The complexity of the model and amount of data we are exploring is massive, so we do need big data processing."

The big questions Heavens is addressing include: is Einstein's theory of gravity correct, what is the universe made of, and what is its future? Like Bassett, Heavens is taking part in a larger study of the universe: the European Space Agency's Euclid mission, that will map the geometry of the dark universe and is planned to start surveying the sky in 2020. The project will put a camera in space that is able to image billions of pixels and will stare at the sky for a few years, sending petabytes of data to researchers for analysis using big data techniques.

Says Heavens, "Part of the Euclid project will look at all the images created to find tiny distortions of image shapes created by gravity." These should support Einstein's theory of gravity. "The project will also investigate dark matter and should tell us more about what the universe is made of."

The SLAC National Accelerator Laboratory, a U.S. Department of Energy Office of Science Laboratories facility operated by Stanford University, is also studying such image distortion, or gravitational lensing, in this case to discover dark matter, which appears to be five times the amount of regular matter in the universe, can't be seen, but is known to be there as it has a gravitational effect on other objects. Until recently, analysis of gravitational lensing was a lengthy process that compared actual images of lenses with a large number of computer simulations of mathematical lensing models. Analysis of a single lens could take weeks or months.

By introducing neural networks, researchers have made the process millions of times faster. To train the neural networks on what to look for, the researchers showed them 500,000 simulated images of gravitational lenses for a day. Once trained, the networks were able to analyze new lenses almost instantaneously with a precision comparable to traditional analysis methods.

Yashar Hezaveh, a Hubble Fellow at the Kavli Institute for Particle Astrophysics and Cosmology at Stanford University, says the next generation of sky surveys is expected to discover about 200,000 gravitational lenses (around 300 are known now). One of the surveys will be carried out by the Large Synoptic Survey Telescope (LSST), a ground telescope with the largest digital camera ever made that will be installed in Chile. It is expected to be functional in 2022 and will be able to map the entire sky every night. It is also likely to produce more data than ever before.

The response will be increasing development and use of AI. Hezaveh concludes: "The possibilities of AI are amazing. It will change the way we work and revolutionize astrophysics in the next few decades."

Sarah Underwood is a technology writer based in Teddington, U.K.


 

No entries found

Sign In for Full Access
» Forgot Password? » Create an ACM Web Account