Here’s what you’ll learn when you read this story:

• Muons are one of the key subatomic particles for discovering new physics, but tracking them after particle collisions can be difficult and prone to error.
• A new study suggests that machine learning algorithms could streamline the process while improving detection efficiency.
• Although tested on a simple simulation of the ATLAS experience at the Large Hadron Collider (LHC), improved AI models like this one could come in handy as the collider receives more powerful upgrades.

When scientists at the Conseil Européen pour la Recherche Nucléaire, or CERN, in Switzerland want to study the subatomic world, they first need to sort through the mess. The world’s largest particle accelerator, the Large Hadron Collider (LHC), smashes particle beams together at nearly the speed of light, and the result is a shower of new particles, many of which decay nearly instantly. Within this subatomic smorgasbord are muons, fundamental particles some 200 times heavier than an electron which exist for only a couple microseconds (one millionth of a second) before decaying into an electron, a muon neutrino, and an electron antineutrino.

The muon is intensely studied, especially its anomalous magnetic moment (or g-2) where any deviation from experimental results versus theoretical predictions can lead to new physics not currently described by the Standard Model. But because muons only exist for a fleeting moment, tracking these particles in order to study them can be pretty difficult. Now, a new study published in the journal Machine Learning: Science and Technology uses a simplified simulation of the ATLAS [A Toroidal LHC Apparatus] detector at CERN to demonstrate how AI could one day help track these particles.

“Reconstructing the trajectories of charged particles in high-energy collisions requires high precision to ensure reliable event reconstruction and accurate downstream physics analyses,” the authors write. “With the forthcoming increase in luminosity and data complexity at the LHC, developing optimized methods for tracking has become an important research direction, also considering the important detector upgrades of the ATLAS and CMS [Compact Muon Solenoid] collaborations.”

Traditional methods of tracking muons are laborious and require a two-step process. First, software needs to separate the muons from the rest of the “noise,” and then an additional algorithm must accurately track the paths of muons perpendicular to the beam line, otherwise known as their transverse momentum. Of course, any errors introduced in the first step impact the second, so a team of scientists based in Italy wondered if machine learning could put this process into a single pipeline. Their method involves using a type of artificial intelligence called a Graph Attention Network (GAT) to mark each ping of the detector with a dot and simultaneously map possible trajectories. According to the study, the new AI detection method made noticeable improvements over traditional methods.

“Applied to a toy simulation based on the ATLAS muon-spectrometer geometry and noise, the model demonstrates promising results in both hit classification and transverse momentum estimation and outperforms a sequential baseline approach,” the authors write. “These achievements show the potential of differentiable, end-to-end techniques to become an alternative to traditional tracking algorithms.”

Because this study used a simplified simulation of ATLAS, there are still a lot of real world issues that need addressing, including overlapping particle tracks and other inevitable imperfections, but the idea aligns with CERN’s stated goal to integrate AI into its overall strategy. In a blog post in November 2025, Joachim Mnich, the director of research and computing at CERN, expressed the goal to use AI for scientific discovery as well as productivity and efficiency.

“Not only has AI transformed research, but it has also penetrated all sectors of the organization,” Mnich wrote. “Could CERN live without AI? The answer is no.”