Two extraterrestrial worlds found in the archives • The Register

A physics student explains to El Reg how she developed a machine learning code for the planet

An illustration of the Kepler Space Telescope

Machine learning algorithms were used to discover two previously unpublished exoplanets in the data archives collected by NASA's Kepler Space Telescope at retirement.

Launched in 2009, Kepler was sent to probe the darkness of the Milky Way. His job was to chase the extraterrestrial worlds by scrutinizing the light emitted by distant stars. Armed with a photometer, Kepler monitored the characteristic sag of brightness when a planet was crossing in front of its parent star. He retired last October.

The spacecraft has allowed scientists to discover more than 2,000 distant planets, and much remains to be discovered. A team of astronomers and engineers, led by the University of Texas, Austin and Google, collaborated to detect potential exoplanets with the help of a neural network at convolution. This software was designed using a set of star and planet data observed by Kepler. So, when you look at the readings of the brightness of other stars, it can predict the presence of an extraterrestrial planet for each of them.

The neural network discovered two worlds until then unknown, while he was browsing Kepler's data for signs of planets in orbit. Their existence has been confirmed with the help of telescopes in Arizona and Hawaii in the United States. Called K2-293b and K2-294b, the two bodies are close to each other and are 1,300 light-years and 1,230 light-years away respectively in the constellation Aquarius. They are slightly taller and warmer than the Earth. The complete results will be published in the next issue of The Astrophysical Journal (here is a free version of the abstract).

Thrust of the material

The data revealing the pair of planets comes from Kepler's K2 phase of his mission. In 2013, two of the four reaction wheels of the spacecraft broke down and could no longer remain oriented to a particular star, NASA reconfigured it so that its thrusters and other remaining wheels keep it in place.

Anne Dattilo, an undergraduate physics student at the University of Texas at Austin, says it was harder to process K2 data. The register. The team had to take into account a slight flickering and systematic noise in the spacecraft readings.

"I trained my modified algorithm using data from more than 27,000 K2 stars," she says, "It only took me 40 minutes to train my laptop, but we put months to understand how to make it work with K2 data. "

The convolutional neural network has been trained to look for periodic declines in the brightness of a star over time, suggesting the passage of a planet. Since it was formed only by using examples of stars with nearby planets, it does not catch all types of exoplanets, such as suns to distant worlds.

"The algorithm lacked" special "planets: those whose signals had a shape different from that of ordinary planets," said Dattilo. "Disintegrating planets are one example, and these planets have different forms of transits from" typical "planets, which means that human astronomers still need to find the most interesting planets," Dattilo said.

Despite this, Dattilo felt that neural networks would still be useful for finding new planets spying on NASA's TESS space telescope. The spacecraft was launched into the skies last year and is expected to discover tens of thousands of exoplanets over a two-year period.

"The same method should be applicable to TESS in the future because TESS works the same way as Kepler and K2 – they measure the change in brightness of the stars," Dattilo said. "However, I suppose there are a number of changes that need to be made – I've made some changes to Kepler's K2 code, and the TESS data is very different from the Kepler or K2 data because they look at the stars for a shorter period of time."

This is not the first time that AI helps scientists find new worlds. Last year, a similar team of researchers using machine learning came across the eighth planet orbiting the Kepler-90 star that had been missed before. This discovery made Kepler-90 the only other planetary system to have eight planets, like our own solar system. ®

Sponsored:
Become a leader in pragmatic security