301 new exoplanets in one fell swoop. A figure to add to the 4,569 whose existence has already been confirmed by astronomers. But how have researchers managed to discover so many new worlds at once? The answer is in the Artificial intelligence. And specifically, in a new deep neural network called ExoMiner.
Neural networks are machine learning algorithms capable of learning a task on their own when sufficient data is provided. And ExoMiner is exactly that: a deep neural network that harnesses the power of the supercomputer of the NASA, Pleiades, and can distinguish real exoplanets from different types of imposters or ‘false positives’. Its design is inspired by various tests and properties that human experts use to confirm the existence of new exoplanets.
ExoMiner learns by using worlds already confirmed in the past and from cases that turned out to be false positives. The system is extremely useful in complementing astronomers and helping them analyze data to find out what is and what is not a planet.
Specifically, the researchers applied the neural network to the enormous amount of data collected over the years by the space mission. Kepler, one of NASA’s biggest ‘planet hunters’. The probe, with thousands of different stars in its field of view, each with the potential to host one or more planets, produces a massive amount of data that takes a long time for astronomers to analyze. But not for ExoMiner, which can do it a thousand times faster.
“Unlike other exoplanet detection machine learning programs,” he explains. Jon Jenkins, from NASA’s Ames Research Center – ExoMiner is not a black box; there is no mystery as to why it decides that something is a planet or not. We can easily explain which characteristics of the data lead ExoMiner to reject or confirm a planet. ‘
Validated and confirmed
Before being totally sure that a new exoplanet has been discovered, they must go through two different states: validation and confirmation. A planet is ‘validated’ using statistics, that is, how likely, based on the available data, is that an object captured by the instruments is or is not a planet. In this phase, one speaks of ‘candidate planets’. And a candidate does not go to the status of ‘confirmed’ until different observation techniques reveal characteristics that can only be explained by the presence of a planet.
In an article published in ‘
Astronomical Journal‘, the Ames team shows how ExoMiner discovered all 301 planets using data from the set of candidate planets in the Kepler Archive. The 301 planets confirmed by the instruments were originally detected at Kepler’s Science Operations Center and promoted to planet candidate status by Kepler’s Office of Science. But until now no one had been able to validate them.
The article also demonstrates how ExoMiner is more accurate and consistent in ruling out false positives and is better able to reveal the genuine signatures of planets orbiting their parent stars. And all in a transparent way, since scientists can see in detail what led ExoMiner to its conclusions.
“When ExoMiner says something is a planet, you can be sure it’s a planet,” he says. Hamed Valizadegan, project leader-. ExoMiner is very accurate and somehow more reliable than existing machine classifiers and human experts, due to the biases that accompany human labeling.
None of the newly confirmed planets are believed to be Earth-like or in the habitable zone of their parent stars. But they share similar characteristics to the general population of confirmed exoplanets in our galactic neighborhood.
“These 301 discoveries,” Jenkins says, “help us better understand the planets and solar systems beyond our own, and what makes ours so unique.”
As the search for more exoplanets continues to expand with missions such as NASA’s Transiting Exoplanet Reconnaissance Satellite, or TESS, and the next mission PLAnetary Transits and Oscillations of stars, O PLATOfrom the European Space Agency, ExoMiner will have more opportunities to show that it is up to the task.
“Now that we have trained ExoMiner using data from Kepler,” concludes Valizadegan, “with a little tweaking we can transfer that learning to other missions, including TESS, which we are currently working on.”