Google has opened a code for machine search for exoplanets according to data from Kepler


A fragment of the solar system (left) in comparison with the eight-planet system Kepler 90 (right). An exoplanet Kepler 90i discovered by Google’s neural network with a circulation period of 14 days is marked in orange.
Researchers are finding new applications of neural networks for image processing. Theoretically, the possibilities of machine learning with reinforcement (DL) are truly endless, but when you look at the real achievements of these programs, there are not so many really useful ones in practice. Most often, neural networks achieve success in a game, giving out a quantitative result in points as a positive incentive. But what is the point that your AI has learned to play Counter-Strike perfectly if it is not able to put its knowledge into practice and neutralize real terrorists?

But in some cases, DL is still used not only to study DL itself, but also to solve practical problems that are important for humanity. For example, they are used in healthcare , quantum chemistry and nuclear physics (links to Google research are given everywhere). Now astrophysics has joined this list. Engineers from the Google Brain division found two new exoplanets , training the neural network to analyze data from the Kepler space telescope . Although these are just preliminary results after processing 670 star systems, they prove the applicability of machine learning in this area.

On March 8, 2018, Google developers posted on GitHub the source code of a program for processing data from Kepler, training programs for the neural network and issuing forecasts for the most promising star systems.

Google hopes the publication of the source code will help other researchers develop similar neural networks and learning algorithms for processing data from other space missions, including K2 (Kepler’s second mission) and Transiting Exoplanet Survey Satellite .

As you know, on May 12, 2013 the Kepler telescope failed; the orbital observatory lost its orientation in space, the on-board computer went into "safe mode". But NASA engineers did not give up and proposed an innovative way to stabilize a space station using solar wind. A schematically invented technique is shown in the illustration.



Since the start of the K2 mission, more than 20 exoplanets and more than 300 candidates have been discovered - and Kepler continues to supply new data for analysis.

The main scientific instrument of the space observatory is a photometer , composed of 42 CCD sensors measuring 50 × 25 mm and a resolution of 2200 × 1024 pixels each. Every three seconds, a read from the CCD. A photometer continuously records the brightness of a star, and if a planet passes in front of it, it captures a characteristic signal in the form of U.
The problem is that other astronomical objects can give a similar signal, not just planets. For example, it can be double stars, star spots , cosmic rays, noise from instruments and other phenomena.



NASA Automated Data Processing Pipeline tries to filter out signals that may correspond to exoplanets. But even after such filtering, there are too many candidates, and all require verification. At the moment, the conveyor has issued more than 30,000 candidates, and about 2500 in reality turned out to be planets.

Google engineers trained the neural network on the existing data set - the very 30,000 candidates who are already really tested by people (training with a teacher). So she learned to confidently distinguish planets from false positive candidates. After that, it was launched to test on a small sample of 670 star systems - and it was possible to find two previously unknown exoplanets Kepler-90i and Kepler-80g , which had previously escaped the attention of astronomers in manual verification of candidates.



Now, after the publication of the source code, anyone can start searching for new exoplanets on their own computer, by training the Google neural network according to the published instructions and downloading data from Kepler into it.

Processing is waiting for another 200,000 star systems. Around each of them, terrestrial planets can rotate.