On a clear night you might easily see Jupiter, Mars or other planets from our solar system. This is how astronomers from thousands of years ago discovered the first planets. We cannot directly observe planets that are farther out in our solar system, as are Neptune and Uranus, so, of course, we cannot see planets that orbit other stars. However, with modern telescopes and satellites like the Kepler mission, it was possible to find the first planets outside of our solar system, also known as exoplanets. These discoveries are essential for a better knowledge of our galaxy and universe as well as for the search for life elsewhere.
TESS (Transiting Exoplanet Survey Satellite) is NASA’s new planet hunter, and the first mission of this kind that will cover the whole sky. The satellite is a space telescope equipped with 4 cameras that will help astronomers by discovering thousands of exoplanets orbiting nearby stars. It will do that by monitoring hundreds of thousands of stars for two years and collecting information about their light. Every time a planet passes in front of a star, some of the star’s light is blocked. So, it appears dimmer to us. Analyzing the brightness of stars over time to identify these events is called the “transit method” and it is the most common technique for finding exoplanets. However, in practice it is hard to tell the difference between dimming due to real planets and dimming due to natural variations in brightness, not to mention variations in the instrument itself and other reasons.
Normally, these signals are analyzed by experts, who try to distinguish planets from other signal sources. Human decisions can be biased and, also, with the huge amount of data produced by TESS, we would need an army of scientists to analyze it. So, scientists are looking for a precise and automatic way of processing TESS data in order to detect these planet signals.
Deep Neural Networks are a type of artificial intelligence that was thought to have the potential of helping and even replacing humans at identifying new planet signals. Deep neural networks are computer processes called algorithms, and they are unique in that they work like our brains and can learn by themselves to do specific tasks. They do this by learning information from examples, the same way students try to learn math: at the beginning, they do not do very well, but the more problems they try the better they get at it. Deep neural networks can be used for face or speech recognition, cancer diagnosis, weather prediction, and autonomous cars, amongst other applications.
To make sure that the artificial intelligence successfully identified planets, the algorithms were put through something called “training.” This is like teaching a student how to solve a math problem by showing an example of one that was solved correctly. Now imagine doing this thousands of times. The computer will learn the rules for finding “correct” planets by looking at a dataset that contained signals from planets and other sources. In this first training stage, the signal examples were not real measurements but simulated data, and the computers could correctly identify planets in the data with a precision of 97.3%. The next and most important step was to test if the computer could identify planets within the real signals collected by the satellite of the TESS mission. It was able to successfully identify 95% of the planets known before the mission.
This is great news for astronomers. Once trained, these artificial intelligence algorithms can classify potential planets extremely fast, keeping up with the pace of data generated by TESS. This automatic system can give a list of planets that are interesting for follow-up studies with bigger ground-based telescopes during the next two decades. The goal is then to know more about their characteristics and perhaps answer the question: Could those planets support life?