Google continues to make advances in artificial intelligence and machine learning, and now, it has put its expertise to use in searching for exoplanets in data collected by NASA’s Kepler telescope. The company even wants to enable anyone to find exoplanets using its AI tool, so it has made the code open source.
The Google Brain Team made the announcement about open sourcing their code in a blog post last week. Software engineer Chris Shallue explained how they trained their neural network to dig through the Kepler data and pick out the signals that present the most promise of being exoplanets. Their analysis covered only about 700 stars, but they believe it was enough to prove that their artificial intelligence can indeed be used to identify exoplanets using the Kepler data.
Scientists who are digging through the Kepler data looking for signs of exoplanets use the automated Kepler data processing pipeline software to identify signals that could be coming from planets. When a potential exoplanet is identified, scientists then check it manually to see whether it really is a planet or whether it’s a false positive. The scientists also set different restrictions on the signals the automated software detects so that they don’t come up with far more signals than what they can reasonably check within a given timeframe. Still, despite the restriction, scientists have manually reviewed more than 30,000 different signals gleaned from the Kepler data, and of all those signals, approximately 2,500 have turned out to be planets.
Of course, scientists are probably missing quite a lot of exoplanets because of the restriction on the signal-to-noise ratio. The problem has been the manual follow-up that’s required through this process to validate or invalidate a signal as a planet, as there are limits to how much work scientists can reasonably do. If they were to loosen this restriction, they would end up with a lot more false positives, making their job even more difficult.
However, the Google Brain team also considered that there could be some planets similar to Earth which are “relatively small and orbit around relatively dim stars … hiding just below the traditional detection threshold.” For this reason, they decided to see whether they could use machine learning to could check these signals much faster than human scientists can check them manually.
To build a neural network that can do this, they used 15,000 of the 30,000 signals from the Kepler data that had already been checked manually. Of the 15,000 signals, about 3,500 were verified as planets or as “strong planet candidates,” and they used this collection of manual inputs to train their neural network to identify exoplanets.
In case you’re interested in trying out Google’s open source code for analyzing the Kepler data, the company has made it available for you to access it in Github here. The company used the code to not only process the data from the Kepler space telescope but also to train its neural network. The team then went even further and used it to make predictions regarding “new candidate signals.”
Shallue said they hope the release of the code they developed will enable others to develop other models like it using data from other NASA missions, such as the second Kepler mission and the future Transiting Exoplanet Survey Satellite mission, a “two-year survey of the solar neighborhood” which seeks to “discover thousands of exoplanets in orbit around the brightest stars in the sky.”