How can machine learning help astronomers find Earth-like exoplanets? This is what a recently accepted study reveals. Astronomy and Astrophysics As an international team of researchers investigated how to use a new neural network-based algorithm to detect Earth-like exoplanets using data from the radial velocity (RV) detection method, the study has the potential to help astronomers develop more efficient methods for detecting Earth-like exoplanets that have traditionally been difficult to identify in RV data due to the intense stellar activity of their host stars.
The study states, “Machine learning is one of the most efficient and successful tools for processing large amounts of data in scientific fields. Many algorithms based on machine learning have been proposed to moderate stellar activity to better detect low-mass and/or long-period planets. These algorithms can be divided into two categories: supervised learning and unsupervised learning. The advantage of supervised learning is that the proposed model contains a large number of variables and can generate relatively accurate predictions based on the training data.”
In the study, the researchers applied the algorithm to three stars to see its ability to identify exoplanets in stellar activity data: the Sun, Alpha Centauri B (HD 128621), and Tau Ceti (HD 10700), about 4.3 light-years away from Earth, and about 12 light-years away from Earth. By inserting simulated planetary signals into the algorithm, they found that the algorithm successfully identified simulated exoplanets with potential orbital periods ranging from 10 to 550 days for the Sun, 10 to 300 days for Alpha Centauri B, and 10 to 350 days for Tau Ceti. It is important to note that while Alpha Centauri B currently has several potential exoplanets detected but not confirmed, Tau Ceti currently has eight exoplanets listed as “unconfirmed” in its system.
Additionally, the algorithm identified these results as indicating that Alpha Centauri B and Tau Centacles may host exoplanets roughly four times the size of Earth, and are within the habitable zones of those stars. After feeding the algorithm more stellar activity data, the researchers found that the algorithm successfully identified a simulated exoplanet roughly 2.2 times the size of Earth, while orbiting the same distance from the Sun as Earth.
The study concludes: “Here we develop a neural network framework that efficiently mitigates stellar activity at the spectral level to enhance the detection of low-mass planets in the time window corresponding to the habitable region of solar-type stars, from a few days to a few hundred days.”
While this study focuses on finding Earth-like exoplanets within the RV data, the researchers note that additional data such as transit times, phases, and space-based photometry could potentially be used to identify Earth-like exoplanets. They highlight that the European Space Agency's PLATO space telescope mission, currently under development and scheduled for launch in 2026, could achieve this. After launch, it will be deployed at Sun-Earth L.2 The Lagrangian point is located on the opposite side of the Sun from Earth and uses the transit method to scan up to one million stars for exoplanets, with an emphasis on terrestrial (rocky) exoplanets.
This study comes as the number of exoplanets confirmed by NASA reaches 5,632 at the time of writing, of which 201 are Earth-like exoplanets, and provides ample opportunity for the upcoming PLATO mission to discover even more Earth-like exoplanets in our Milky Way galaxy.
How will machine learning help astronomers discover Earth-like exoplanets in the coming years and decades? Only time will tell. And this is why we do science.
As always, keep on science and keep looking up!
