In recent years, cosmologists have imposed tight constraints on dark matter and dark energy densities, even as their fundamental nature remains elusive. A key goal is creating precise maps of dark matter distribution across and within galaxies. In a groundbreaking study, researchers at ETH Zurich merged cosmology with AI expertise to build a neural network that quantifies dark matter far more accurately than traditional human-led methods.
Physicists and computer scientists at ETH Zurich in Switzerland collaborated to refine dark matter estimation techniques using cutting-edge machine learning algorithms—similar to those powering facial recognition on platforms like Facebook. Their peer-reviewed findings appear in Physical Review D.
"While there's no faces to recognize in cosmic images, we're hunting for analogous signatures," explains Tomasz Kacprzak from Alexandre Refregier's group at the Institute of Particle Physics and Astrophysics. "Facebook algorithms spot eyes, noses, and mouths; ours detect hallmarks of dark matter and dark energy."
Dark matter evades direct telescope observation, so experts leverage its gravitational influence—like ordinary matter—to subtly bend light from distant galaxies. This weak gravitational lensing effect distorts galaxy images ever so slightly, enabling cosmologists to map dark matter distributions.
These maps are then benchmarked against theoretical models to identify the best-fitting cosmology. Traditionally, this relies on human-crafted statistics like correlation functions, which struggle to capture intricate patterns.
"Our approach marks a paradigm shift," says Alexandre Refregier. "Rather than devising statistics manually, we empowered computers to discover them autonomously." Enter Aurelien Lucchi and team from the Data Analytics Lab in ETH's computer science department.
They harnessed deep neural networks, training them on simulated universe data to maximize insights from dark matter maps. With known "ground truth" cosmological parameters (e.g., dark matter-to-dark energy ratios) for each simulation, the networks iteratively honed in on key features.
Related topic: New map of dark matter distribution suggests revising the speed of evolution of it
The outcome? Neural networks delivered 30% greater precision than conventional human statistics—a leap that would otherwise demand twice the telescope observation time, at significant cost.
Applying their trained model to the KiDS-450 dataset yielded superior results. "This is the first application of such deep learning in cosmology, extracting richer data than prior methods," notes Daniel Fluri. "We anticipate broad future impacts."
Next, the team aims to scale up with datasets like the Dark Energy Survey, incorporating additional parameters such as dark energy properties.
