American astronomers, supercharged with a suite of leading-edge AI tools, have discovered a sensational archipelago of planets circling alien stars sprinkled across this sector of the Milky Way.
Just a generation after the first exoplanet orbiting a sun-like star was detected, in 1995, this group of astrophysicists has uncovered the greatest cache of planet candidates ever – more than 10,000 orbs around a spectrum of stars – in a fantastical find powered in part by their AI adjutants.
The team, headed by astrophysicist Joshua Roth at Princeton University, sketched out its findings in a fascinating paper titled “The T16 Planet Hunt: 10,000 New Planet Candidates from TESS Cycle 1.”
Roth, a graduate researcher at Princeton, told me in an interview his team ultimately found a treasure trove of exoplanet candidates, ranging from colossal gas giants locked in tight orbits around their stars, called “Hot Jupiters,” to “super-Earths” that might turn out to be habitable.
“Hot Jupiters are the easiest to detect,” he says, because their colossal size can eclipse a greater proportion of light when they cross in front of their star, and their short orbital periods create a rapid sequence of these light-blocking “transits,” all of which can be captured by high-power imaging telescopes.
Gas giants account for 9000-plus of the exoplanets his team discovered, he says, followed by more than 100 Neptune-size bodies and just 11 potential “super-Earths.”
The stars they tracked varied from M-dwarfs – small orange-colored suns that can live for more than a trillion years – to blue supergiants that burn bright but die young.
His alliance of astronomers, Roth says, spread across Princeton, MIT, UCLA, and Las Campanas Observatory in Chile, relied on an array of advanced AI tools to race through examining stellar imagery captured by the Transiting Exoplanet Survey Satellite (TESS) space-based telescope.
After these images were converted into light curves, or graphs charting the changing brightness of a star, potentially caused by an orbiting planet partially eclipsing its sun, the team deployed AI software including the Cambridge Exoplanet Transit Recovery Algorithm to examine the 80 million light curves derived from TESS to whittle down the top exoplanet candidates.
A machine learning system called Random Forest that his team trained further refined the pool of potential planets.
While astronomers around the world – stretching from Cambridge University in the UK to a Western European coalition of scholars, have been experimenting with combining different artificial intelligence breakthroughs to speed up their searches for planets across the galaxy, Joshua Roth’s team has taken the global lead in terms of finding new exosolar worlds.
Now, he says, his team is racing to parse through Cycle 2 of TESS telescope images, and he predicts they could find an astounding new trove of potential planets.
Changing the parameters and focus of the new search, he says, could turn up more Earth analogues, including some orbiting in the habitable zone of their stars, where life might take hold and multiply.
During the first search, whose astonishing advances in exoplanet detection are likely now being studied by astrophysicists worldwide, including across NASA and the European Space Agency, his team’s machine learning system sharply sped up the planet hunt.
“It is safe to say this search would not have been possible without the machine learning step,” Roth says.
“The best estimate I can give for how long this search would have taken without the machine learning step is based on the time it took us to manually vet 50,000 candidates, about 6 weeks of focused work.”
“The machine learning step,” he adds, “reduced our number of light curves from 80 million to about 2.5 million.”
“So in the worst case scenario we would have had to vet 32 times more light curves, which would have taken me about 192 weeks, over 3.5 years.”
Meanwhile, Roth, a doctoral candidate in astrophysics at Princeton, a worldwide powerhouse in the field with more than 20 winners of the Nobel Prize in Physics, says he aims to speed up the new planet search even more by incorporating convolutional neural networks to classify the full-frame-image cutouts produced by TESS.
“We are actively working on this.”
“I hope to have implemented convolutional neural networks by the end of the year, which will hopefully eliminate or at least minimize the manual vetting.”
“I think this will be feasible,” he predicts, “because we have a large [TESS telescope] Cycle 1 sample we can use as a training set now, which should make training the CNNs somewhat easier.”
He also says that with new space-based super-telescopes coming online and the expanding use of sophisticated machine learning and neural networks in astronomers’ searches for exoplanets, a new golden age of exploring the cosmos is unfolding.
Just over a hundred years after American astronomer Edwin Hubble, operating the world’s biggest telescope at the time, began proving that the Milky Way was just one island-galaxy in the cosmic expanse of the universe, a new epoch of discoveries has begun.
Over the course of the next decade, Roth says, an explosion of new candidate planets could be discovered.
One of the leading-edge telescopes set to be lofted by the European Space Agency, Plato, is designed to zoom in on small terrestrial planets that might resemble the Earth.
“The Plato mission, which is set to launch early next year, will likely find a lot of these small planets,” Roth says, and it features “better photometric precision.”
“The expressed goal of the Plato mission is to find an Earth-like planet around a Sun like star.”
Plato is also designed to seek out planets that feature liquid water on their surface, along with an atmosphere conducive to life.
Plato, he adds, “will allow for the detection of smaller radii planets and longer [orbital] period planets.
“So I definitely expect Plato to detect a much higher number of smaller planets on wider orbits, possibly even in the habitable zone, specifically around M-dwarfs, compared to TESS.”
“The Plato candidates should also be of quite high quality thanks to the extensive follow up program that is being planned.”
His team is “planning to use Plato to facilitate quick follow up of large numbers of our candidates.”
“Machine learning methods will be used extensively to search for planets in Plato data.”
Outracing the Plato observatory into orbit, Roth says, “The Roman space telescope (predicted launch date is August this year) is predicted to discover some 100,000 transiting planet candidates.”
“However, most of these will be around very faint stars (much fainter even than the ones we searched in our transit search), and thus, it will be exceedingly difficult to confirm how many of those candidates are true positives.”
“The Gaia space telescope, which collected data until last year, will also facilitate the detection of many more planets.”
“A recent prediction claims that Gaia’s final data release will include over 100,000 exoplanets, detected via the astrometry method.”
“So in short,” Roth forecasts, “I think it’s very likely that we will have 100,000 planet candidates within the next decade.”
While marking an astronomical leap in human discoveries of exo-worlds, that figure would still represent just a minuscule fraction of the total planet population that astrophysicists now predict inhabits the universe.
These days, scientists across NASA estimate that the Milky Way galaxy alone holds more than 100 billion stars, and that each solar system, on average, features at least one planet.
Zooming out, NASA astronomers say “the universe could contain up to one septillion stars – that’s a one followed by 24 zeros.”
