Google Open Sources Exoplanet-Hunting AI

This site may earn affiliate commissions from the links on this page. Terms of use.

NASA’s Kepler satellite has been observing the stars for nearly a decade, and it’s produced a mountain of data in that time. Late last year, Google showed how machine learning could help astronomers dig through the Kepler backlog, and it discovered a few new exoplanets in the process. Google has now open sourced the planet-spotting AI so anyone can give it a shot.

We can’t observe exoplanets directly yet, but we can spot the telltale dip in luminance when an exoplanet passes in front of its host star. So, Kepler was designed to scan large swaths of the sky in search of these signals. However, there are a substantial number of candidate signals that might point to an exoplanet. We don’t know for sure until astronomers can examine the data more closely, but there’s too much to check every reading. That’s why astronomers have only checked the 30,000 or so best candidates, which has resulted in the discovery of around 2,500 exoplanets.

Google’s solution to the data backlog was to unleash the power of neural networks on the problem. That seems to be Google’s go-to idea lately, whether you’re talking about self-driving cars or smartphone photography. Using 15,000 examples of exoplanets in the data, Google trained its network to sift through Kepler data and identify other exoplanet signals. After the training period, Google’s AI checked 670 stars that were known to have exoplanets. And wouldn’t you know it, Google spotted two new exoplanets that would have otherwise gone unnoticed.

A neural network is adept at recognizing patterns, but you need a lot of data to train the network. Google is saving everyone the time of training a neural network on Kepler data by releasing its code freely. You can get the TensorFlow code on GitHub right now, and Kepler’s data is also freely available. It’s not as easy as grabbing both piles of code and launching an app, though. Experience with Google’s TensorFlow platform and Python will help. Google’s blog post provides instructions on how you can test the network by re-discovering Kepler 90i, one of the planets Google found in the original test.

The Googlers behind this project hope that the community will make use of the code to go on the hunt for planets in Kepler’s data. There are thousands more signals to analyze, and that could be just the start. Googlers hope to develop more advanced neural networks to dig through NASA data for Kepler and future missions like the upcoming Transiting Exoplanet Survey Satellite.

ExtremeTechExtreme – ExtremeTech