April 2, 2019, 2:44 p.m.
Read time: 5 minutes
Author: Gueter Josmy Faure
NASA’s Kepler Space Telescope is now retired but not long ago it was part of the NASA’s Discovery program whose aim is to discover Earth-size exoplanets orbiting other stars. Thanks to this program, NASA’s astronauts have discovered 2,662 planets from 2009 to 2018.
Everything didn’t go as expected and this is the exact reason why the program took so long before being terminated. In fact, it was originally planned to be a 3.5 years program. But in 2012, there was too much noise in the data from the stars and the spacecraft, so NASA decided to extend its lifetime in order to meet their goal.
In 2013, a second reaction wheel failed, disabling the collection of science data and threatening the continuation of the mission. Far from discouraging the scientist, they decided to keep the mission alive with the two reaction wheels remaining that’s why K2 was proposed. Astronomers continued to find new exoplanets in the K2 data, but at a much lower rate than before, because the telescope wasn’t working as well as before. However, K2 had some great advantages as it could observe fields across the ecliptic plane, look at different regions of the sky and observe stars and planets that formed in different galactic environments.
The mission ended in 2018, 2662 planets were discovered but scientists were still not satisfied with the result. Maybe noise has prevented some of the planets from being discovered. How could we remove that noise and find potentially missed planets? That is when some researchers from the university of Texas at Austin come into the game. And their tool? Deep Learning.
In a paper recently published (March 25, 2019) the researchers from The University of Texas, Austin, created an automatic system for identifying planet candidates in K2 data using deep learning. They call the system AstroNet-K2.
“Our training set consists of possible planet signals that, following the naming convention in the literature, we refer to as “Threshold Crossing Events” or TCEs. These are potentially periodic signals (decreases in the brightness of a star) that have been detected by an algorithm designed to search for transiting exoplanets in a light curve. A TCE is characterized by the star on which it is observed, the period on which the signal appears to repeatedly cause the star to dim, the time at which the first repetition of the dimming signal is observed, and the duration of the dimming signal.”, we read in the paper.
Their goal was to sort the TCEs into two categories: planet candidates and false positives (non-planets which are identified as planets). They first did a triage in order to discard the TCEs that are less likely to be planets. Their sample is then reduced by 90% before they proceed to the next step, vetting. The vetting was a more thorough process. All of the samples are assumed to be planets and to prove themselves wrong, the researchers had to do an in-depth analysis on the remaining candidates in order to filter the most likely to be planets. Do do the vetting, they considered things like the TCE transit depth, their rotation shape… Any ambiguous TCE are removed from the training set at this step.
They were now left with 27,634 TCEs for their training dataset which were randomly shuffled into 80% training set, 10% validation set and 10% testing set.
After training, the model attained an accuracy of 97.84% on the testing data when predictions greater than 0.5 are classified as planet candidates and predictions less than 0.5 are classified as false positives. The Area Under the Curve (AUC) was 98.83%, meaning that there is 99% chance that the model will be able to distinguish between planets and non-planets.
Using the Neural network method, the researchers identified 2 new exoplanets which were not previously identified as planets. “Both planets are super-Earths orbiting G-dwarf stars”.
Despite being better than the method used before, AstroNet-K2 is not perfect. In fact, the researchers pointed out that planets showing significant transit timing variations could be misclassified as false positive. The triage was also not perfect. They admitted, “There are almost certainly incorrectly labeled signals in the training set.”
The deep learning method performed better than human astronauts but they don’t have to worry about their job(thus far). “The fact that our neural network did not identify the two disintegrating planets WD 1145+017 b and K2-22 b and the TTV system K2-146 as planet candidates shows the enduring value of humans classifying planet candidates by eye. Humans are good at recognizing unusual signals that machines will mis-classify or not recognize as interesting, which is crucial for discovering interesting and odd facets of the universe.” Says Datilo and Co.
Image Source: “planet” byfotomanu_93 – Under Creative Commons license
Computer science student and machine learning enthusiast. He writes and has strong opinions about a broad range topic including Technology (mostly covering Deep learning and AI related stuff) and Psychology.