26 June 2026

Photo: AG Haller

At the trigger station in the CMS control room at CERN: Dr. Artur Lobanov (far right) and Lukas Ebeling (3rd from right) together with other members of the UHH team (from left: R. Heine, B. Letzer, B. Weidner, L. Markus).

At CERN, millions of particle collisions occur every second—far too many to store them all. A clever filtering system therefore decides in real time which data are worth keeping. Researchers from Universität Hamburg in the Cluster of Excellence Quantum Universe have now integrated artificial intelligence directly into this system for the first time, making it possible to identify rare physical processes of two Higgs bosons more reliably.

The CMS experiment at the Large Hadron Collider (LHC) at CERN "photographs" millions of proton collisions every second. Since this volume of data cannot be stored in full, a fast electronic filter — the so-called "trigger system" — decides in real time which collisions are selected for later analysis.

For rare processes such as the simultaneous production of two Higgs bosons, however, the filtering algorithms used so far reach their limits: many interesting collisions go unrecognized and are lost. Novel artificial intelligence (AI) algorithms can deliver significantly better results here.

Dr. Artur Lobanov and Lukas Ebeling from the Department of Physics at Universität Hamburg, together with colleagues from the CMS experiment, have integrated the first AI-based filtering algorithms for such processes directly into the trigger system. Since the beginning of this year, they have been successfully recording collisions at CERN using this approach.

Photo of the electronic components of the
first trigger layer of the CMS detector,
where the AI-based algorithms
are applied. Photo: AG Haller

"We began working on the new AI algorithm more than three years ago here in Hamburg," explains Dr. Artur Lobanov, who leads the project at Universität Hamburg. "We had to implement the algorithm at both trigger levels of CMS: the first level, which consists of fast electronic components, as well as the computing farm made up of CPUs and GPUs. That was quite a technical challenge. It's fantastic that the algorithm now works so well at CERN—we were able to improve the filtering efficiency by more than 20%!"

"This shows that such AI algorithms can generally be used in the trigger systems of LHC experiments. In principle, AI can then be used to select all kinds of different physical processes," says Lukas Ebeling, a PhD student who played a key role in the project. "This will pay off above all in the future: in the upcoming high-luminosity phase of the LHC, the gain from the improved selection would correspond to a full year's worth of data. With AI, we are essentially saving the operating costs of an entire year."