The concepts of dark matter and dark energy have been puzzling the physics world for many decades, which are believed to be vital in understanding the origin and future of our universe. Researchers from ETH Zurich's Department of Physics and the Department of Computer Science have now used incorporated AI to improve on standard methods for estimating the dark matter content of the universe. They use advanced machine learning algorithms for cosmological data analysis, which is similar to the algorithms used by Facebook and other social media platforms for facial recognition. The results of their works were published in the scientific journal Physical Review D.

"Facebook uses its algorithms to find eyes, mouths or ears in images; we use ours to look for the tell-tale signs of dark matter and dark energy," explains Tomasz Kacprzak, a researcher in the group of Alexandre Refregier at the Institute of Particle Physics and Astrophysics.

Since one cannot spot dark matter with telescopes, physicist rely on a phenomenon called "weak gravitational lensing", which happens when matter slightly bends the path of light rays arriving Earth from distant galaxies. As a result, the images captured of far way objects and galaxies appears subtly distorted. Cosmologists then use that distortion to create mass maps of where dark matter is located in the sky.

They then compare those dark matter maps to theoretical predictions to find which cosmological model most closely matches the data. This comparison task is often done by human-designed statistics, which are limited in finding intricate patterns in the dark matter maps. 

However, in this case, researchers compared those dark matter maps to theoretical predictions by using machine learning algorithms called deep artificial neural networks and trained them to extract the most substantial possible amount of information from the dark matter maps. Once the neural network is trained, it can be used to extract cosmological parameters from actual images of the night sky.

So far, the results of that training were encouraging, as the neural networks came up with values that were 30% more accurate than those obtained by traditional human-made statistical analysis. For cosmologists, this is a significant milestone as the improvements can provide the same accuracy they get by increasing the number of telescope images by twice.