Artificial intelligence Analyses Dark Matter
TDark matter comprises most of the universe’s matter. But understanding its nature remains a mystery.
Understanding how our universe turned into what it is, is one of the most significant challenges of contemporary science. The deeper issues are mostly in what we cannot directly see: dark matter and dark energy. Dark matter pulls the universe together, and dark energy accelerates its expansion. But there’s still much unknown about them.
At least until today. A group of researchers from the Physics Department and Computer Science Department of ETH Zurich joined forces to improve the methods to approximate the dark matter content of the universe using AI, to help us understand better the nature of dark matter and the universe as a whole.
They used top-notch machine learning algorithms for cosmological data analysis. This method is quite similar to the one Facebook uses for facial recognition.
You may ask yourself how this is possible if there are no faces to recognize in the sky. Well, where Facebook uses the algorithms to find eyes or mouths, the researchers used it for the revealing signs of dark matter and dark energy.
Dark matter cannot be seen through telescope images, which means it bends the path of light rays that come to the Earth from other galaxies, what scientists call “weak gravitational lensing.” It’s a similar effect that what happens on a hot day: far-away objects seem blurred because light travels through different air layers at different temperatures.
Scientists use that bending to create mass maps that show where the dark matter is. After that, they relate the maps to theoretical predictions to find what model resembles the data. Usually, that is conducted with stats such as correlation functions that describe the relations of the different parts of the maps, but those stats are relatively limited. They are good as long as they find intricate patterns in the maps.
That’s when AI comes to the game. The lead investigator, Alexander Refregier, said that instead of creating the statistical analysis, they let the computers do it. They used machine learning algorithms (deep artificial neural networks) and taught them to take the most significant amount of information they could from dark matter maps.
How did they achieve this? First, they trained the neural networks with computer-created data, so the system knew the correct answers for different cosmological parameters. By repetitively analyzing dark matter maps, the machine learning network taught itself to look for the right set of patterns to extract a larger batch of the desired information. If we follow the Facebook’s analogy, that means the neural networks got better at distinguishing oval figures from eyes or mouths.
When the neural network is trained, it can be used to process cosmological data more effectively. Source: ETH Zurich
This AI-based analysis came up with results 30 % more accurate than those based on human-made stat analysis. The AI analysis is an enormous improvement considering that to obtain the same accuracy with human analysis would require an increase of telescope images that would double observation time and costs.
After training the neural network, they used it to analyze real dark matter maps and found it allows them to extract much more information from the data than before.
Now, as a next step, the ETH Zurich research group is planning to use their method to larger image sets, like the Dark Energy Survey. Likewise, they’ll feed the neural network with further cosmological bounds and enhancements such as details about the nature of dark energy.
The ETH Zurich research group, comprised of physicists and computer scientists seem to have opened a new pathway for AI and IT in general. The possibilities of expansion in the cosmological field are simply infinite.