Scientists have announced the discovery of two new exoplanets , Kepler-90i and Kepler-80g. Exoplanets, or planets outside our solar system, are routinely being discovered, with the number of those that have already been found now standing at 3,567. But this announcement by the National Aeronautics and Space Administration (NASA) of the U.S. is particularly significant. First, with the discovery of the planet Kepler 90i, orbiting the star Kepler 90, we now know of another star besides the Sun that has eight planets orbiting it. Second, Christopher Shallue, a software engineer at Google, and Andrew Vanderburg, of the University of Texas, Austin, have discovered it using a deep learning neural network — an artificial intelligence tool that mimics the workings of a human brain. They “trained” their computer to analyse light readings made by NASA’s Kepler Space Telescope, which are archived and made available for anyone to use. During its mission from 2009 to 2013, the Kepler Space Telescope surveyed nearly 200,000 stars, with 35,000 possible planet signals. The duo’s network was made to learn to identify true signals using 15,000 previously vetted signals. They then studied the weaker signals recorded from 670 star systems that had multiple known planets orbiting them, finally coming up with this discovery. The network also identified another Earth-sized exoplanet, Kepler 80g, orbiting the star Kepler 80. This is a very stable system in which Kepler 80g and four of its neighbours are locked together in a so-called resonant chain.
Deep learning and neural networks have been used in other applications successfully, as in the AlphaGo AI player of the Go game. This is not also the first time that automation has been used in identifying exoplanets. After the initial years of their discovery, when the number of known exoplanets grew, the need for automating the initial vetting process became clear. The preprint of the Shallue-Vanderburg’s paper, to be published in The Astronomical Journal , mentions the Robotvetter program, the first attempt at automating the process of rejecting false positives in the signal. The preprint describes the careful process of doing away with the false positives and systemic blips before coming up with the true signals — in this case, the two signals corresponding to Kepler 90i and Kepler 80g. It also indicates the caveats and failure modes in the model where it needs to be improved before it can be used to function independently. Here, then, is the takeaway — good science not only solves problems but also can take a hard look at itself, at where and how it can improve. This is a leap for humankind, a measured leap.
