The world’s largest SETI effort is scanning the skies with AI
In 1960, astronomer Frank Drake turned the radio telescope at the National Radio Astronomy Observatory in Green Bank, W.Va., toward two nearby sunlike stars to search for transmissions from intelligent extraterrestrial societies. With a pair of headphones, Drake listened to Tau Ceti and Epsilon Eridani for a total of 150 hours across multiple sessions. He heard nothing, except for a single false positive that turned out to be an airplane.
Statistically speaking, Drake's pioneering project had little chance of success. Estimates from both NASA and the European Space Agency (ESA) suggest that our galaxy alone contains roughly 100 billion stars. (And there are perhaps 2 trillion galaxies in the universe).
Nevertheless, Drake's effort, called Project Ozma, launched the modern search for extraterrestrial intelligence (SETI) efforts that continue to this day. But given the universe's size, a SETI project trying to make even the smallest dent in searching for intriguing signals needs to be much more than one astronomer with a pair of headphones. The project will need more telescopes, more data, and more computational power to find such signals. Breakthrough Listen has all three.
Breakthrough Listen is the most comprehensive and expensive SETI effort ever undertaken, costing US $100 million, using 13 telescopes, and bringing together 50 researchers from institutions around the globe, including me as part of the Parkes telescope team. The project is making use of the newest advances in radio astronomy to listen to more of the electromagnetic spectrum; cutting-edge data processing to analyze petabytes of data across billions of frequency channels; and artificial intelligence (AI) to detect, amid the universe's cacophony, even one signal that might indicate an extraterrestrial intelligence. And even if we don't hear anything by the time the project ends, we'll still have a wealth of data that will set the stage for future SETI efforts and astronomical research.
Breakthrough Listen is a 10-year initiative, launched in July 2015, and headquartered at the University of California, Berkeley. As my colleague there, J. Emilio Enriquez, likes to put it, Breakthrough Listen is the Apollo program of SETI projects.
It was the first of several Breakthrough Initiatives founded by investor Yuri Milner and his wife, Julia. The initiatives are a group of programs investigating the origin, extent, and nature of life in the universe. Aside from Listen, there's also Message, which is exploring whether and how to communicate with extraterrestrial life; Starshot, which is designing a tiny probe and a ground-based laser to propel it to the star Alpha Centauri; and Watch, which is looking for Earth-like planets around other stars. There are also projects in earlier stages of development that will explore Venus and Saturn's moon Enceladus for signs of life.
Supersize SETI: The 64 radio antennas of South Africa's MeerKAT array produce 2.2 terabits of data per second. Combing through that amount of data requires a complement of 128 servers and anomaly-detection artificial intelligence. Without MeerKAT, Breakthrough Listen wouldn't be able to survey millions of stars for potential extraterrestrial signals. Photo: Enrico Sacchetti/INAF
Listen was started with three telescopes. Two are radio telescopes: the 100-meter Robert C. Byrd Green Bank Telescope, in Green Bank, W.Va., and the 64-meter Parkes Observatory telescope, in New South Wales, Australia, for which I'm the project scientist. The third is an optical telescope: the Automated Planet Finder at Lick Observatory, in California. We're using these telescopes to survey millions of stars and galaxies for unexpected signals, and powerful data processors to comb through the collected data at a very fine resolution.
In 2021, Listen will also begin using the MeerKAT array in South Africa. MeerKAT, inaugurated in July 2018, is an array of 64 parabolic dish antennas, each 13.5 meters in diameter. MeerKAT is located deep in the Karoo Desert, in a federally protected radio-quiet zone where other transmissions are restricted. Even better, while at the other three telescopes we have to share time with other research efforts, at MeerKAT our Listen data recorders can “eavesdrop" on antenna data 24/7. Listen has also partnered over the last couple of years with several other observatories around the globe.
The anomalous signals we're searching for fall under a broad umbrella called “technosignatures." A technosignature is any sort of indication that technology exists or existed somewhere beyond our solar system. That includes both intentional and unintentional radio signals, but it can also include things like an abundance of artificial molecules in the atmosphere. For example, Earth's atmosphere has plenty of hydrofluorocarbons because they were used as refrigerants before scientists realized they were damaging the ozone layer. Breakthrough Listen isn't searching for anything like artificial molecules, but we are looking for both intentional communications, like a “We're here!" message, and unintentional signals, like the ones produced over the years by the Arecibo Observatory in Puerto Rico.
Arecibo, before cable breaks destroyed the dish in November 2020, was a telescope with four radio transmitters, the most powerful of which effectively transmitted at 22 terawatts at a frequency of 2,380 megahertz (similar to Wi-Fi). The telescope used those transmitters to reflect signals off of objects like asteroids to make measurements. However, an artificial signal from Arecibo that made it into interstellar space would be a clear indication of our existence, even though it wouldn't be an intentional communication. In fact, Arecibo's signal was so tightly focused and powerful that another Arecibo-like radio telescope could pick up that signal from halfway across the galaxy.
Looking for technosignatures allows us to be much broader in our search than limiting ourselves to detecting only intentional communications. In contrast, Drake limited his search to a narrow range of frequencies around 1.42 gigahertz called the 21-centimeter line. He picked those frequencies because 1.42 GHz is the frequency at which atomic hydrogen gas, the most abundant gas in the universe, spontaneously radiates photons. Drake hypothesized that an intelligent society would select this frequency for deliberate transmissions, as it would be noticed by any astronomer mapping the galaxy's hydrogen.
Drake's reasoning was sound. But without a comprehensive search across a wider chunk of the electromagnetic spectrum, it's impossible to say whether there aren't any signals out there. An Arecibo-like 2,380-MHz signal, for example, would have gone completely unnoticed by Project Ozma.
Meet the Gang: Breakthrough Listen started with two radio telescopes, Green Bank (top) in West Virginia and Parkes (center) in New South Wales, Australia, as well as an optical telescope, the Automated Planet Finder in California (bottom). Photos, from top: Andrew Caballero Reynolds/AFP/Getty Images; Torsten Blackwood/AFP/Getty Images; Laurie Hatch
Ground-based radio telescopes can't scan the entire electromagnetic spectrum. Earth's atmosphere blocks large swaths of bandwidth at both high and low frequencies. For this reason, gamma ray, X-ray, and ultraviolet astronomy all require space-based telescopes. However, there is a range of frequencies from roughly 10 MHz to 100 GHz that can pass easily through Earth's atmosphere and reach ground-based telescopes called the terrestrial microwave window. Breakthrough Listen is focused on these frequencies because they're probable carriers for technosignatures from other planets with Earth-like atmospheres. But it's still a huge range of frequencies to cover: Drake searched a spectrum band of only about 100 hertz, or one-billionth of that window. In contrast, Breakthrough Listen is covering the entire range of frequencies from 1 GHz to 10 GHz.
Each radio telescope is unique, but they all function according to the same basic principles. A large reflective dish gathers radio signals and focuses them on the same point. At that point, a receiver converts those radio signals into signals on a coaxial cable, just like a well-aligned TV antenna. Then computers digitize and process the signals.
The telescopes we're using for Breakthrough Listen are wideband telescopes, which means that they can record signals from across a wide range of frequencies. Parkes, for example, covers 3.3 GHz of spectrum, while Green Bank can detect signals across 10 GHz of spectrum. Wideband telescopes allow us to avoid making difficult choices about which frequencies we want to focus on.
However, expanding the search band also greatly increases the amount of data collected. Fortunately, after the receivers at the telescopes have digitized the incoming signals, we can turn to computers—rather than astronomers with headphones—for the data processing. Even so, the amount of data we're dealing with required substantial new hardware.
Each observatory has only enough memory to store Listen's raw data for about 24 hours before running out of room. For example, the Parkes telescope generates 215 gigabits per second of data, enough to fill up a laptop hard drive in a few seconds. The MeerKAT array produces even more data, with its 64 antennas producing an aggregate of 2.2 terabits per second. That's roughly the equivalent of downloading the entirety of Wikipedia 10 times every second. To make such quantities of data manageable, we have to compress that raw data to about one-hundredth of its initial size in real time.
After compression, we split the frequency-band data into sub-bands, and then process the sub-bands in parallel. Individual servers each process a single sub-band. Thus the number of servers at each telescope varies, depending on how much data each one produces. Parkes, for example, has a complement of 27 servers, and Green Bank needs 64, while MeerKAT requires 128.
Each sub-band is further split into discrete frequency channels, each about 2 Hz wide. This resolution allows us to spot any signals that span a very narrow range of frequencies. We can also detect wider-frequency signals that are short in duration by averaging the power of each channel over a few seconds or minutes. Simultaneous spikes for those seconds or minutes over a contiguous band of channels would indicate a wide-frequency, short-duration signal. We're focused on these two types of signals because concentrating signal power continuously into a single frequency or into short pulses over a range of frequencies are the two most effective ways to send a transmission over interstellar distances while minimizing dissipation. Therefore, these signal types are the most likely candidates for extraterrestrial technosignatures.
Humble Beginnings: Frank Drake searched two nearby stars for signs of extraterrestrial life in 1960 using the Howard E. Tatel radio telescope. Decades later, Breakthrough Listen is using Tatel's successor, the Green Bank telescope, as part of the largest SETI project in history. Photo: NRAO/AUI/NSF
Fortunately for us, these kinds of signals are the easiest to spot, too. Narrowband signals are easily distinguishable from ones caused by natural astrophysical processes, because such processes do not produce signals narrower than a few thousand hertz. To understand why, let's look at the example of a cloud of hydrogen-gas molecules in space emitting radiation in the 21-cm line. When Drake searched that band, how would he have known the difference between an intentional signal from the noise created by hydrogen gas?
Here's how: Even though all the hydrogen gas in a cloud radiates at that frequency, each molecule is moving in a different, random direction. This movement causes each emitted photon to arrive at a radio telescope with a slightly different frequency. It's the Doppler effect at work: An emitted photon will have a higher frequency if the hydrogen molecule was moving toward you and a lower frequency if the molecule was moving away. In the end, the hydrogen cloud's signal is smeared across a frequency band centered around 1.42 GHz. If Drake had heard a narrowband signal at that frequency, it would have meant he had detected a physical impossibility: a hydrogen cloud in which every molecule was stationary relative to Earth. Alternatively, the signal was artificial.
We can make a similar assumption about short-duration signals. While there are some natural short-duration signals, namely fast radio bursts and pulsars, these have other characteristics that single them out. A signal emitted by a pulsar, for example, has a long “tail" caused by the lower frequencies lagging behind higher frequencies over interstellar distances.
Also, I should mention that artificial signals generally don't turn out to be extraterrestrial. In 2019, we published a paper in the Astronomical Journal, detailing our search across 1,327 nearby stars. While we detected tens of millions of narrowband signals, we were able to attribute all of them to satellites, aircraft, and other terrestrial sources. To date, our most promising signal is one at 982 MHz collected in April 2019 from the nearby star Proxima Centauri. The signal is too narrow for any known natural phenomenon and isn't in a commonly used terrestrial band. That said, there remains a lot of double-checking before we rule out all possibilities other than an extraterrestrial technosignature of some kind.
It's also entirely possible that an alien society might give off radio technosignatures that are neither narrowband nor short duration. We really don't know what kinds of communications such a society might use, or what kinds of signals it would give off as a by-product of its technologies. Regardless, it's likely we couldn't find such signals simply by finely chopping up the telescope data and analyzing it. So we're also using AI to hunt for more complicated technosignatures.
Specifically, we're using anomaly-detection AI, in order to find signals that don't look like anything else astronomers have found. Anomaly detection is used in many fields to find rare events. For example, if your bank has ever asked you to confirm an expensive purchase like airline tickets, odds are its system determined the transaction was unusual and wanted to be sure your credit card information hadn't been stolen.
A major challenge for any SETI project is to distinguish natural, artificial, and potentially extraterrestrial signals from one another. A signal's frequency and shape can reveal a lot of information about the source of that particular signal.
Terrestrial signals show up as vertical lines, because the Earth is stationary in the frame of reference for both telescope and signal source.
Diagonal lines indicate deep-space signals. Satellites and probes are identified by comparing signals with known frequencies and directions.
Some natural phenomena (like fast radio bursts) appear as curved lines because the high and low frequencies in their emissions spread out while crossing interstellar space.
A potential extraterrestrial signal must not match any known artificial or natural source, and it must be studied to ensure it is not a previously unknown natural phenomenon.
We're training our AI to spot unexpected signals using a method called unsupervised learning. The method involves giving the AI tons of data like radio signals and asking it to figure out how to classify the data into different categories. In other words, we might give the AI signals coming from quasars, pulsars, supernovas, main-sequence stars like the sun, and dozens of other astronomical sources. Crucially, we do not tell the AI what any of these sources are. That requires the AI to figure out what kinds of characteristics make certain signals similar to one another as it pores through the data, and then to lump like signals into groups.
Unsupervised learning causes an AI classifying radio signals to create a category for pulsar signals, another category for supernova signals, a third for quasar signals, and so on, without ever needing to know what a pulsar, supernova, or quasar is. After the AI has created and populated its groups, we then assign labels based on astronomical objects: quasars, pulsars, and so on. What's most important, though, is that the AI creates a new category for signals that don't fit neatly into the categories it has already developed. In other words, it will automatically classify an anomalous signal in a category of its own.
This technique is how we hope to spot any technosignatures that are more complicated than a straightforward narrowband or short-duration signal. That's precisely why we needed an AI that can spot anything that's out of the ordinary, without having any preconceived notions about what an odd signal should look like.
The Listen program is dramatically increasing the number of stars searched through SETI projects, by listening in on thousands of nearby sunlike stars over the program's duration. We're also surveying millions of more-distant stars across the galactic plane at a lower sensitivity. And we're even planning to observe nearby galaxies for any exceptionally energetic signals that have crossed intergalactic space. That last category is a bit of a long shot, but after all, the project's purpose is to listen more than ever before.
Breakthrough Listen is a hugely ambitious program and the most comprehensive search for intelligent life beyond Earth ever undertaken. Even so, odds are we won't find a single promising signal. Jill Tarter, the former director of the SETI Institute, has often likened the cumulative SETI efforts so far to scooping a glass of water out of the ocean, seeing there are no fish in the glass, and concluding no fish exist. By the time Breakthrough Listen is finished, it will be more like having examined a swimming pool's worth of ocean water. But if no compelling candidate signals are found, we haven't failed. We'll still be able to place statistical constraints on how common it is for intelligent life to evolve. We may also spot new natural phenomena that warrant more study. Then it's on to the next swimming pool.
In the meantime, our data will help provide insights, at least, into some of the most profound questions in all of science. We still know almost nothing about how often life emerges from nonliving matter, or how often life develops into an intelligent society, or how long intelligent societies last. So while we may not find any of the signs we're looking for, we can come closer to putting boundaries on how often these things occur. Even if the stars we search are silent, at least we'll have learned that we need to look farther afield for any cosmic company.
This article appears in the February 2021 print issue as “Seti's Million Star Search."