Thinking about the search for planets in space might make us think only of large telescopes pointed towards stars, the search for new worlds that are unknown and at enormous distances from Earth. In reality, many of the planets that astronomers are able to identify cannot actually be directly observed. Their presence can be discovered by analysing some small changes in the behaviour of the star around which they orbit.
One of the most widely used methods for detecting an exoplanet, or extrasolar planet, is the transit method. In fact, when a planet passes in front of its star from the position of Earth, it can block a very small part of its light. The telescope therefore records a slight decrease in the brightness of the star. If this decrease is repeated over time at regular intervals, it could represent a signal of the presence of a planet orbiting the star.
The data relating to the brightness of the star can be represented through a light curve, that is, a graph showing how the brightness changes over time. By observing this curve, astronomers can look for these small decreases in brightness. The problem is that these signals can be very weak. In addition, not all variations are caused by the presence of a planet.
It is precisely in this type of situation that artificial intelligence can be used, in particular, some machine learning systems which can be trained using examples of signals that astronomers have already classified. In this way, the system can learn to recognise some of the typical characteristics of signals associated with planetary transits and distinguish them from those that may have a different origin.
In the case of the search for exoplanets, a model can receive numerous light curves and learn to recognise some characteristics associated with possible planetary transits. It can also be trained using examples of false positives, that is, signals that may initially appear to be connected to a planet but which actually have a different origin. Thanks to this, the system can learn to distinguish some patterns present in the data and identify signals that deserve further analysis.
One type of model used for this purpose is the convolutional neural network, an artificial intelligence network particularly suited to recognising characteristics and patterns within organised data. In the case of light curves, it can analyse the behaviour of the signal and identify elements that could be compatible with a planetary transit.
This does not mean that the neural network independently understands what a planet is. The system recognises characteristics present in the data based on the examples with which it was trained. The scientific interpretation of the result remains the responsibility of the researchers.
A concrete example of the use of AI in astronomical research comes from NASA’s Kepler mission.
The Kepler space telescope was designed to observe the brightness of a large number of stars and search for variations that could be caused by the passage of planets. During the mission, it collected a large amount of data, within which signals of planetary transits that had not yet been recognised could be present. Analysing all this information manually was a very complex task. For this reason, astronomers developed and used automatic tools capable of examining the signals and identifying the most promising ones.
In 2017, Christopher Shallue, a researcher at Google, and astronomer Andrew Vanderburg used a deep convolutional neural network to analyse some signals from Kepler data.
The model had been trained to distinguish signals compatible with a planetary transit from false positives. In practice, the artificial intelligence analysed the light curves and looked for characteristics that could indicate the passage of a planet in front of its star.
The aim was not to obtain an automatic discovery, but to help astronomers examine a large amount of data and identify signals that might have been overlooked or difficult to recognise through a manual search. The researchers therefore used the neural network to analyse again some data collected by the Kepler mission, focusing on star systems that were already known to have multiple planets. The network was applied to 670 such systems and identified some signals that could be interesting. After further analysis by the researchers, two of these signals led to the identification of two new planets: Kepler-90i and Kepler-80g.
The case of Kepler-90i is particularly interesting because, before this analysis, the Kepler-90 system was already known to have seven planets. The signal identified by the neural network therefore allowed researchers to identify an eighth planet around the same star. Kepler-90i is a planet larger than Earth and completes its orbit around the star in about 14.4 days. Its proximity to the star means that it has very high temperatures and conditions that are very different from those on our planet.
What makes this case interesting is not only the discovery of these two planets, but above all the role that artificial intelligence played in allowing researchers to identify them within a large amount of data. AI did not independently determine that they were planets, but recognised signals that could be important and that deserved to be analysed more carefully by astronomers.
This example therefore makes it possible to understand how artificial intelligence can be used in scientific research as a support tool. When the amount of data to be analysed becomes very large, a machine learning system can help researchers recognise certain patterns and select the most interesting information, leaving scientists with the task of verifying and interpreting the results.
Artificial intelligence therefore does not replace the telescope or the work of astronomers. The telescope collects the data, AI can help analyse it, and scientists verify whether those signals can actually be explained by the presence of a planet. The search for new worlds thus becomes a concrete example of how artificial intelligence can enter scientific research and help scientists find, even within data collected years earlier, information that might be difficult to identify through a simple manual analysis.
Author: Ms Maria Lagani, psychology graduate- Master’s student. Research Team for JUMP staff (Italy)
Source:
1. NASA. Artificial Intelligence, NASA Data Used to Discover Eighth Planet Circling Distant Star. 2017.
2. NASA. Briefing Materials: Artificial Intelligence and NASA Data Used to Discover Eighth Planet Circling Distant Star. 2017.
3. Shallue, C. J., & Vanderburg, A. (2018). Identifying Exoplanets with Deep Learning: A Five-planet Resonant Chain around Kepler-80 and an Eighth Planet around Kepler-90. The Astronomical Journal, 155, 94. DOI: 10.3847/1538-3881/aa9e09.