Scientists working on SETI, the search for extraterrestrial intelligence, have turned a student project into a powerful AI-based tool for spotting radio signals that could have been generated by an alien civilization. When presented with a trove of real data, the algorithm flagged eight such examples that had previously been missed when the same data were analyzed using more conventional means.
The result suggests that AI can be used to expand the boundaries of what is detectable using current technology – and improve the odds that anyone broadcasting to us across the light years will be noticed.
“We demonstrated that it was able to pick up signals that it was never trained on,” said Peter Ma, a third-year undergraduate in math and physics at the University of Toronto. He initially created the algorithm in high school as a computer-science project during the pandemic.
It was then that Mr. Ma contacted SETI scientists at the University of California, Berkeley, to ask for technical details about their open-source data, which he had been working with for his project. The conversation led to a summer job and then further development of his machine learning (ML) SETI algorithm, which he continued while at U of T.
Along the way, Mr. Ma teamed up with Cherry Ng, a U of T Dunlap Institute radio astronomer, and others to refine the algorithm and test its ability to spot potential extraterrestrial transmissions hidden among the radio waves produced naturally in space, and also to distinguish them from all the radio noise that human activity is constantly generating on Earth.
“We are all impressed by how well ML has adapted to help sieve through tons of human-made interference and pick out unique signals of interest,” Dr. Ng said in an e-mail. “Peter is an expert in ML and he has brought tremendous progress to the field.”
A key challenge in hunting for signs of alien intelligence is that no one knows what a true signal will look like. That means the algorithm had to teach itself to see signals that simply look out of the ordinary when compared with the flood of radio energy that reaches the sensitive antenna from across the universe and from closer to home.
While there have been previous attempts to employ artificial intelligence in SETI in more limited ways, Mr. Ma’s approach is the first to do so using an end-to-end model that “allowed the algorithm to generalize and search for patterns that we might not have taught the program to look for in the simulated training data,” Dr. Ng said.
The algorithm was put to work on data from 820 stars that were recorded across a range of frequencies in 2016 and 2017 by a giant radio dish in Green Bank, W.Va., as part of a SETI project called Breakthrough Listen.
It identified eight candidate signals that had been overlooked in earlier analyses. The signals appear to originate from five different stars located between 30 and 90 light years from our solar system. Follow-up observations conducted last year did not confirm any repeat transmissions, which means the detections were likely false positives.
In general, SETI researchers are extremely conservative about claiming to have discovered anything until it is confirmed by other instruments. For that reason, the speed at which the algorithm was able to digest the data was another important aspect of the work.
“This scientific result shows that it is now possible to announce this kind of detection quickly enough to do the necessary follow-up,” said Franck Marchis, a senior astronomer at the SETI Institute in Mountain View, Calif., who was not a member of the project team. “A new era of SETI research is opening up, thanks to machine learning technology.”
A key strength of the approach is that the more data the algorithm is exposed to the better it gets. This turns one of the main obstacles on its head, by transforming the needle-in-a-haystack problem of searching through reams of data into a way of improving performance.
“You’re basically turning your worst enemy into your best friend,” Mr. Ma said.
Recent advances in AI have caused some researchers to speculate that humanity’s first contact with an advanced extraterrestrial civilization will be with alien-designed AI algorithms rather than with the aliens themselves. While Mr. Ma described himself as a fan of science fiction, particularly the popular Three Body trilogy written by Liu Cixin, he said he is still wrapping his mind around the possibility that an AI of his own creation could wind up hearing from another AI out there.
“That would be pretty crazy,” Mr. Ma said. “It’s like holding up a cosmic mirror.”
His algorithm is described in a research study published Monday in the journal Nature Astronomy with Mr. Ma as lead author alongside 16 SETI researchers based in North America, Australia and Britain.
The group is now planning to expand the method to other Breakthrough Listen projects, notably at the MeerKAT radio telescope array in South Africa and the Very Large Array in Socorro, N.M., which was featured in the film Contact.
The development is a sign of the growing utility of artificial intelligence in radio astronomy more generally. Researchers are already exploring how AI will be used to enhance the powers of the Square Kilometre Array, a large, multilocation international radio astronomy observatory that is under construction, and which the National Research Council said in an announcement last week that Canada plans to join.