Looking to the future, astronomers are excited to see how machine learning – aka. deep learning and artificial intelligence (AI) – will enhance surveys. One field that is already benefitting in the search for extrasolar planets, where researchers rely on machine-learning algorithms to distinguish between faint signals and background noise. As this field continues to transition from discovery to characterization, the role of machine intelligence is likely to become even more critical.
Take the Kepler Space Telescope, which accounted for 2879 confirmed discoveries (out of the 4,575 exoplanets discovered made to date) during its nearly ten years of service. After examining the data collected by Kepler using a new deep-learning neural network called ExoMiner, a research team at NASA’s Ames Research Center was able to detect 301 more planetary signals and add them to the growing census of exoplanets.
These newly-detected exoplanets and the ExoMiner algorithm were described in a paper that was recently accepted for publication in the Astrophysical Journal. The paper and project team were led by Hamed Valizadegan, a machine learning manager with the Universities Space Research Association (USRA) at NASA Ames’, and included multiple researchers from the USRA, the SETI Institute, and universities from all around the world.
As they indicate in their paper, all 301 of the machine-validated planets were originally detected by the Kepler Science Operations Center pipeline. These planets were also promoted to the status of planet “candidate” by the Kepler Science Office (in other words, not confirmed). However, before the Kepler Kepler archive was examined using ExoMiner, no one was able to verify that these potential signals were exoplanets.
Like all machine-learning techniques, this new deep neural network learns to identify patterns based on the data it has been provided. In the case of ExoMiner, researchers at NASA Ames designed it using various tests and properties that human experts use to confirm the presence of exoplanets. Combined with NASA’s Supercomputer (Pleiades), it uses this knowledge to distinguish between actual exoplanets and various types of “false positives.”
Also indicated in the paper is how ExoMiner is more precise and consistent in ruling out false positives and identifying signatures of planets while also showing science teams how it arrived at its conclusion. As Valizadegan explained:
“When ExoMiner says something is a planet, you can be sure it’s a planet. ExoMiner is highly accurate and in some ways more reliable than both existing machine classifiers and the human experts it’s meant to emulate because of the biases that come with human labeling. Now that we’ve trained ExoMiner using Kepler data, with a little fine-tuning, we can transfer that learning to other missions, including TESS, which we’re currently working on. There’s room to grow.”
ExoMiner was specifically designed to assist experts who search through the data gathered during the Kepler and K2 campaigns. The reason for this has to do with the exoplanet-hunting method used by Kepler and its successor, the Transiting Exoplanet Survey Satellite (TESS). This consists of monitoring thousands of stars for signs of periodic dips in luminosity, which could be caused by exoplanets passing in front of them (aka. transiting) relative to the observer.
Known as the Transit Method (aka. Transit Photometry), this technique is the most effective means of exoplanet-detection to date, accounting for over 75% of all discoveries made to date. However, it is also subject to a substantial rate of false positives, which can be as high as 40% in single-planet systems (based on a 2012 study of Kepler mission data). What’s more, it is only effective for about 10% of star systems since they must be edge-on relative to the observer for transits to be visible.
The primary way of getting around this is to monitor thousands of stars in a single field, which creates the data-mining burden (mentioned above). For all of these reasons, having an automated helper that can process the data reliably (by knowing exactly what to look for) is a huge game-changer. As Jon Jenkins, an exoplanet scientist at NASA’s Ames Research Center, said in a recent NASA press release:
“Unlike other exoplanet-detecting machine learning programs, ExoMiner isn’t a black box – there is no mystery as to why it decides something is a planet or not. We can easily explain which features in the data lead ExoMiner to reject or confirm a planet… These 301 discoveries help us better understand planets and solar systems beyond our own, and what makes ours so unique.”
Unfortunately, none of the newly confirmed planets are believed to be “Earth-like,” meaning they are not rocky in composition nor do they orbit within their parent stars’ habitable zone (HZ). But they have some characteristics in common with the overall population of confirmed exoplanets in our galactic neighborhood, making these 301 planets a fitting addition to the exoplanet census.
In the very near future, ExoMiner and other machine learning techniques will prove very useful to missions relying on Transit Photometry. This includes TESS, which is scheduled to remain in operation until Sept. 2022 (barring further extensions), but also the ESA’s PLAnetary Transits and Oscillations of stars (PLATO) mission and NASA’s Nancy Grace Roman Space Telescope (RST) – which are scheduled to launch in 2026 and 2027 (respectively).
Further Reading: NASA