A neural network has helped astronomers catalog a whopping 27 million galaxies collected from one of the largest astronomical surveys probing the mysterious nature of dark energy.
The Dark Energy Survey (DES) kicked off in 2013, and aims to snap galaxies and supernovae across an eighth of the Earth's night sky. Although that may not sound like much, the number of objects filling such a region of space – giving us a window onto the universe as far back as when it was just six billion years old, less than half of its age today – is staggering.
If we know the age of each of the galaxies we observe, and know where they are, we can better understand the workings and evolution of our universe. However, cataloging tens or hundreds of millions of galaxies by hand is largely infeasible.
The DES is thus a prime example of scientists acquiring astronomical data at an astronomical rate, far beyond what they can manually process. Computers are sorely needed to automatically identify and catalog what's in each snap because it'll take humans far too long, notes a research paper published in the Monthly Notices of the Royal Astronomical Society (here’s the free arXiv pre-print).
“Applying automated classification methods to such large surveys is mandatory,” the study's large team of astronomers, spread across 53 academic institutions, wrote.
And so they proceeded.
Captain, the computer has identified 250 alien stars that infiltrated our galaxy – actual science, not science-fictionREAD MORE
The scientists turned to a convolutional neural network to sift through the pictures obtained by the DES project using the four-metre Victor M. Blanco Telescope in Chile. First, they trained an algorithm on tens of thousands of real and simulated galaxy images to classify whether a particular specimen was either an early-type galaxy or a late-types one. Younger objects are blobs of gas and bright stars, and are often in spiral-like structures, whereas older ones are more likely to have elliptical shapes.
Second, the algorithm was also taught to identify when a galaxy was snapped face-on or edge-on. For example, our own Milky Way looks very different from above compared to its side. The giant spiral arms are visible face-on, and disappear edge-on. Overall, the team’s algorithm was 97 per cent accurate, and had a false negative and false positive rate of three to four per cent, respectively in testing.
The trickiest samples to classify are ones where the galaxies are dim and have little to no discernible visual features. To overcome this challenge, and teach the software how to cope with faint objects, the astronomers simulated images in the training data by tweaking existing examples to appear less bright and further away.
"We simulate how local close galaxies with available classifications would look like if they were more distant, while keeping their original labels,” Jesús Vega-Ferrero, first author of the research paper and a former postdoc researcher at America's University of Pennsylvania, told The Register on Tuesday.
We increased the number of objects we can correctly classify by a factor of 50x
"By doing so we were able to classify fainter galaxies – 1000x less luminous galaxies – which are more common in the universe. We increased the number of objects we can correctly classify by a factor of 50x."
When the next batch of data from the survey, containing images of some 600 million galaxies, lands soon, the team will process them again with the help of neural networks. They hope that their work can help the wider scientific community figure out how dark energy is accelerating the expansion of the universe, and also how galaxies evolve over time, from their galaxy classifications.
"Some of the galaxies included in this catalog are as far as eight billion light-years, the light we receive today from these objects was emitted by them when the universe was less than half of its current age," Vega-Ferrero added.
"This means we can study how the fraction of the two types: early-type galaxies and late-type galaxies change with cosmic time and how the structural transformations are related to their stellar content. This is fundamental to understand the formation mechanisms of galaxies and their evolutionary paths." ®