NASA’s Kepler discovers a new exoplanet using Google’s Machine Learning

2 min read

Earlier this week, we wrote that NASA’s Kepler has made a major breakthrough discovery and is going to announce it in a press conference soon.

Well, The day has finally arrived!
NASA scientists have discovered a new planet outside our solar system. The Kepler-90i is a part of the eight-planet solar system called The Kepler-90 star system. Like Earth, the new planet is the third rock from its sun.

Kepler-90i orbits the star Kepler-90 and is 2,545 light years away from earth. However, it is much closer to its sun and so it has a temperature of 427 degrees Celsius at the surface, making it unlikely that life could exist there.

An interesting thing to note here is that this discovery was made by machines and not humans!

NASA has partnered with Google to harness Machine learning to crunch data collected from the Kepler’s telescope. The experts from Google taught the ML system how to identify planets around faraway stars.

The AI behind discovering Kepler 90i

NASA’s Kepler observed about 200,000 stars for four years, taking a picture every 30 minutes, creating about 14 billion data points. Those 14 billion data points translated to around 2 quadrillion possible planet orbits. To analyze this huge amount of data points, Google experts turned to Machine Learning.

They created a TensorFlow model to distinguish planets from non-planets. This model was created using a dataset of more than 15,000 labeled Kepler signals.


The model is a deep convolutional neural network which recognizes patterns created by light curves of actual planets, versus light curves and patterns caused by other objects like starspots or binary stars.

They considered 3 types of neural networks for classifying Kepler’s data points as either “planets” or “not planets”. A linear architecture as the base model, a fully connected architecture as the second layer and finally, a 1-dimensional CNN with max pooling.

The model was successful in distinguishing between a planet and a non-planet with 96% accuracy, when it was fed with signals never seen before.

This model was then run over the data of 670 stars, known to have two or more exoplanets. The model was successful in discovering two new planets: Kepler 80g and Kepler 90i.

Google Scientists now plan to use their model to search for other 200,000 stars in the Kepler data for unfound exoplanets. They also plan on incorporating new machine learning techniques to help fuel celestial discoveries for many years to come.

Jessie Dotson, a NASA project scientist for the Kepler Telescope, said “As the application of neural networks to Kepler data matures, who knows what might be discovered, I’m on the edge of my seat.”


Please enter your comment!
Please enter your name here