Using a Neural Network to Find Exoplanets in K2 Data

Title: Identifying Exoplanets with Deep Learning II: Two New Super-Earths Uncovered by a Neural Network in K2 Data

Authors: Anne Dattilo, Andrew Vanderberg, Christopher Shallue et al.

First Author Institution: University of Texas-Austin Astronomy Department

Status: Accepted to The Astronomical Journal, open access on arXiv

*Hey Siri*

*Find me an exoplanet from the K2 mission*

*Okay. But not without your help*

In today’s astrobite, we explore how to discover exoplanets using a neural network. Now let’s break that down for a hot second.

This image is currently unavailable due to a known server malfunction.

Let’s say I want to create an album of all known pictures of goats. I can create a neural network which is basically a fancy computer algorithm that is based off of the neural network in a human’s brain. We’re trying to get a computer to learn on it’s own. First, I give my neural network a ‘training set’ of pictures that are labeled ‘goat’ or ‘not goat’ corresponding with the correct picture.   

The neural network is like “cool, pics. Here I tried to identify some on my own, how did I do?”

This image is currently unavailable due to a known server malfunction.

Then I can tell the neural network “You did ayight. Here are some corrections, let’s try again”. Depending on how strict or lenient you let the network be, it might find some ‘false positives’ (like the horse that is NOT a goat) or some ‘false negatives’ (like the space goat that is actually one of the best goats).  

And I can repeat this processes over and over until my neural net gets real good at identifying goats. #goals.

Now let’s replace goats with exoplanet light curves.

One way to find exoplanets is by looking at the light emitted from stars for long periods of time to see if the light dims periodically due to an exoplanet passing in front of the star (see figure 1).  If the light curve (a plot of light emitted vs time) looks anything like this, the dips in brightness could correspond to an exoplanet crossing in front of the star.

This image is currently unavailable due to a known server malfunction.

Figure 1: This is an example of a star’s light curve with dips in brightness caused by an exoplanet blocking out some of it’s light as it crosses in front of the star. (Source: Andrew Vandenberg, Center for Astrophysics)

If we can get computers to look for transit patterns, they may be able to find exoplanets better than humans. Or at least find exoplanets a lot faster than individual people. The first step to building a neural network is to create a training set of light curves. In order to do that, a human needs to go through and identify light curves as exoplanet or no exoplanet. The authors of today’s paper used data from the K2 mission.

The K2 Exoplanet Mission

K2 is an extended mission of the space-based Kepler exoplanet mission. The Kepler Space Telescope looked at one patch of the night sky for a long time, just observing the light from stars searching for exoplanets. The idea is that if you look at a patch of sky for long enough you will be able to identify exoplanets with longer periods. To obtain a more complete picture of exoplanet types and find Earth-like planets, it’s important to be sensitive to exoplanets with longer periods. K2 began when the second of four gyroscopes on Kepler failed and the telescope was not able point on the same patch of sky. Now, it moves instead, following the ecliptic. It’s still searching for exoplanets, but now there’s an added degree of motion in the data that comes through K2. This is both an added challenge to the observations, and a good thing, since K2 is going to see more of the sky.

This is the motivation for the neural network that this team created. Kepler had a computer program that identified exoplanets pretty well, but K2 did not because of this systematic error that wasn’t in the original mission. After creating a set of training light curves and a neural network that was sensitive to the telescope’s motion, the group examined how well the neural network performed.

The Neural Net

The neural network gave each light curve a value between zero and one. Zero meaning that it definitely was not an exoplanet and one meaning it definitely was. There are a few different ways you can measure the success of a neural network. I’ll just mention two that the paper talked about.

  • Precision — How many true positives did you get right?
  • Recall — How many planets did you identify correctly as planets?

The plot below shows the results of a few different models. The significance of a detection is categorized by a signal-to-noise ratio (SNR), comparing the depth of the dip in the light curve to the small wiggles (noise) that appear in the data (see Fig. 1). A larger SNR translates to a more confident detection. The neural net was told to be sensitive to different SNR values, so they were testing different strict vs. lenient rules for the neural network.

This image is currently unavailable due to a known server malfunction.

Figure 2: If a model had high precision, that means it identified many true positives and very few false positives. This is great! Except that it only identified light curves that it was very very sure about, so it missed a lot of exoplanets (it found false negatives). While if your Recall is high, that means it’s finding lots of planets. That also sounds great! However, it’s calling everything an exoplanet, so it finds a lot of false positives. We want both Recall and Precision to be high, and the purple group did a good job at that. They also find that they have a consistently high AUC. This means that overall, the non-planets usually had a lower planet prediction than the real planets (remember, each plot got a value between 0-1) (Figure 6 from paper, edited).

This team successfully created a neural network that was able to process K2 lightcurves. But the neural network isn’t quite ready to go off and work by itself. It still isn’t perfect at finding exoplanets. What it is good at is determining what is not a planet. The neural network assigned the vast majority of the light curves a very low likelihood of planet-ness. So this team argues that with this neural network, one can give it a set of light curves, it will assign them values of planet-ness or not, and any light curve with a very low value you can chuck out and be relatively sure that you didn’t chuck out any false negatives (exoplanets identified as non-exoplanets). This is immensely helpful! It was a powerful enough tool that when this team ran a set of unidentified light curves from K2, the were able to find 2 previously undiscovered exoplanets!  So even if this neural net can’t do an astronomer’s job entirely, it can reduce the workload by half or even more (which is nice, because as a grad student, I would rather not be replaced by a robot).

About Jenny Calahan

Hi! I am a forth year graduate student at the University of Michigan. I study protoplanetary disks, which set the stage for planet formation. For the past few years I've been using high resolution ALMA data to pull out the 2D thermal structure of different types of disks using thermo-chemical modeling. Outside of astronomy, I love belting showtunes, eating Thai food, I enjoy crafting, and I love to travel and explore new places. Check out my website:

Discover more from astrobites

Subscribe to get the latest posts to your email.

Leave a Reply