Artificial brain scans the galaxy for speeding stars
Neural networks come to astronomy as a self-adapting algorithm digs through star maps to find rogue fast-moving stars, writes Andrew Masterson.
An artificial neural network capable of learning from its own observations is helping astronomers identify a rare type of star that might offer clues to both the formation of the Milky Way and the role of dark matter in governing its motion.
In research published in the Monthly Notices of the Royal Astronomical Society, a team led by Elena Maria Rossi from Leiden University in the Netherlands detail how the self-adapting algorithm is spotting rogue stars among millions mapped by the European Space Agency’s Gaia satellite.
The stars in question are all travelling at much higher than usual speeds through the galaxy – a rare phenomenon identified only a decade ago. While most stars, including the Sun, move at about 220 kilometres per second – and those on the Milky Way’s outskirts clock only around 150 kilometres per second – these high-speed variants zoom along at two or three times the speed.
Until recently, only 20 such stars had been identified – all of them young and massive. It was acknowledged, however, that this was likely to be an artifact of the search methods used, so Rossi and her team set out to see if other types also sometimes went rogue.
“These hypervelocity stars are extremely important to study the over all structure of our Milky Way,” she says.
“These are stars that have travelled great distances through the galaxy but can be traced back to its core – an area so dense and obscured by interstellar gas and dust that it is normally very difficult to observe – so they yield crucial information about the gravitational field of the Milky Way from the centre to its outskirts.”
To conduct their search, the scientists utilised the first dataset made available from the Gaia mission – a massive amount of information released in September last year.
The full data maps about one billion stars, but Rossi’s crew were particularly interested in a single subset: measurements of two million stars that combined both Gaia’s information with that of ESA’s Hipparcos mission, which charted the same area of sky more than 20 years ago.
By comparing the differences between the two measurements, the astronomers hoped to be able to identify any stars that were moving at above average speeds. To achieve this, they used a neural network architecture that had been trained for six months to spot the sort of speed anomalies that could indicate the targets.
Neural networks are at the forefront of much artificial intelligence development. Their particular advantage is to be able to learn rapidly from their own mistakes, without having to be reprogrammed or updated by human operators, constantly thus narrowing their focus and improving their accuracy.
“On the day of the data release, we ran our brand new algorithm on the two million stars of TGAS,” says Elena.
“In just one hour, the artificial brain had already reduced the dataset to some 20,000 potential high-speed stars. A further selection including only measurements above a certain precision in distance and motion brought this down to 80 candidate stars.”
Further refinements reduced the quarry to just six stars – all now set for much more detailed investigation.
One of them, moving at more than 500 kilometres per second, appears destined to eventually shoot out of the Milky Way altogether.
The other five, however, are moving at slightly slower speeds, and Rossi’s team is intent on discovering what exactly is slowing them down. A strong candidate is friction caused by dark matter.
The next info dump from Gaia will take place in 2018. Rossi and her colleagues intend to apply their algorithm to this much larger dataset, allowing it to compare readings with the 2016 billion-point download to see if more high-speed rogues can be detected.