A new study led by Dr. Jonas Glombitza demonstrates how artificial intelligence (AI) significantly enhances the analysis of ultra-high-energy cosmic radiation, reducing the time required for mass estimations from 150 years to just a fraction of that.
Image Credit: Steven Saffi /Pierre-Auger-Collaboration
AI often raises concerns due to the complexity of neural networks, which even experts struggle to fully understand. However, the societal impact of opaque algorithms depends largely on their application.
AI poses numerous risks, and in astrophysics, its worst consequence is an incorrect interpretation of the cosmos, explains Dr. Jonas Glombitza from the Erlangen Center for Astroparticle Physics (ECAP) at Friedrich-Alexander Universität Erlangen-Nürnberg (FAU).
Initial Skepticism
Dr. Glombitza employs AI to accelerate the analysis of data from an observatory studying cosmic radiation.
The results suggest that the most energetic particles hitting the Earth are usually not protons, but significantly heavier nuclei such as nitrogen or iron atoms.
Jonas Glombitza, Friedrich-Alexander Universität Erlangen-Nürnberg
His recent study, published in Physical Review Letters, marks only the second instance in astroparticle physics where machine learning (ML) has been applied.
“I found the use of machine learning in astrophysics fascinating,” Glombitza said.
He began programming ML tools at RWTH Aachen in 2017, later moving to FAU in 2022. In 2025, he was honored with the ETI Award, which recognizes promising talent at the university. While the term “artificial intelligence” is often debated, Glombitza prefers “machine learning,” as it is more widely understood. Initially, he faced skepticism from colleagues due to ML’s black-box nature. The turning point came when AI-generated results aligned with telescope observations, validating its effectiveness.
Radiation from Distant Galaxies
Ultra-high-energy cosmic radiation likely originates from galaxies beyond the Milky Way. These energetic particles, with charges ranging from 1018 to 1020 electron volts, are the most powerful found in nature. Upon entering Earth’s atmosphere, they interact with air molecules, triggering a cascade of secondary particles, including electrons, positrons, photons, and muons.
Some are absorbed by the atmosphere, while others reach the Earth's surface across a vast area. During this interaction, fluorescence light is emitted, which is detected by specialized telescopes like those at the Pierre Auger Observatory, the world’s largest cosmic radiation research facility.
“The measurements there have been running for 15 years,” said Glombitza.
Theoretical models suggest that these primary particles could be composed of elements ranging from hydrogen to iron. Due to its larger mass, an iron atom generates a more complex particle cascade upon entering the atmosphere compared to a proton. This results in the maximum fluorescence light appearing higher in the atmosphere for heavier elements, whereas lighter particles penetrate deeper before reaching their peak.
Only on Clear Moonless Lights
The Pierre Auger Observatory, spanning 3,000 km2 in Argentina’s Mendoza province, uses 27 telescopes positioned on four hills to capture fluorescence light from particle showers. Additionally, 1660 surface detectors, housed in evenly spaced water tanks, measure the distribution and number of incoming particles.
While fluorescence light analysis provides valuable insights into the primary particle’s mass, the telescopes can only operate on clear, moonless nights, significantly limiting data availability. In contrast, surface detectors function continuously, but until now, reconstructing the light maximum from their complex distribution patterns had been an unresolved challenge.
AI is now addressing this issue. Trained on extensive simulations of particle showers, it can infer the primary particle’s mass based on surface detector data. The AI models are then calibrated using real telescope observations, enabling mass estimations from 60,000 recorded particle showers.
“To achieve the same results without AI, we would have had to observe with the telescopes for 150 years. This is the breakthrough I have achieved,” concludes Glombitza.
Journal Reference:
Halim, A. A., et al. (2025) Inference of the Mass Composition of Cosmic Rays with Energies from 1018.5 to 1020 eV Using the Pierre Auger Observatory and Deep Learning. Physical Review Letters. doi.org/10.1103/physrevlett.134.021001.