When Artificial Intelligence Meets Physics Beneath the French-Swiss Border

Deep beneath the French-Swiss border, the Large Hadron Collider sends protons crashing together at extraordinary speeds, allowing scientists to study the fundamental building blocks of matter and the forces that shape the universe. Those collisions help researchers investigate questions ranging from the behavior of subatomic particles to the possibility of new physics beyond what is currently known.

But the machine creates a practical problem as well as a scientific opportunity: it produces far more collisions than researchers can possibly save.

That challenge is at the heart of award-winning research by Assistant Professor Yuxin Chen and computer science Ph.D. student Zixin Ding, whose paper received a best paper award at the ICML 2026 Workshop on AI for Physics. The project is part of a broader collaboration bringing together Chen and Ding from the Department of Computer Science, Professor David Miller from the Department of Physics, and colleagues from the University of Michigan and FermiLab. Together, the team explored how artificial intelligence can help the collider experiments make faster, smarter decisions about which collision data to keep, while staying within the strict limits of a real scientific instrument.

“Making these decisions well, often with incomplete information and staggering operational constraints, can mean the difference between a Nobel Prize winning discovery and just another day at the office,” explained Miller. “Thankfully, we can work together with scientists studying complex computational approaches to exactly these kinds of problems in other contexts to develop novel and exciting ways to overcome these challenges and bring sophisticated solutions to bear for the discoveries therefore will not miss.“

It is a strong example of the kind of interdisciplinary work increasingly taking shape between artificial intelligence and physics — and of how computer science methods can be shaped by real-world scientific constraints.

The Race To Keep The Right Data

At the Large Hadron Collider, collisions happen at a staggering rate.

“The trigger system decides, in real time, which collisions to keep and which to throw away forever, and a discarded collision is gone for good,” Chen said.

The trigger system is essentially a rapid filtering mechanism. It uses preset thresholds to decide whether a collision looks interesting enough to store for later study. If the threshold is too loose, the system saves too much and overwhelms storage and readout capacity. If it is too strict, it may reject rare events that could lead to important discoveries.

Traditionally, those thresholds are set by experts and adjusted by hand. But collider conditions shift over time. Beam intensity changes, background patterns drift, and a setting that worked well at the start of a run may no longer be ideal hours later. The researchers asked whether that process could become more adaptive — something Miller describes as a self-driving trigger.

“By building a system that can learn and adapt to the experimental conditions, we can not only optimize our instruments more effectively and efficiently, we can also allow the scientists to focus on the broader questions about why and what our experiments should be measuring to make the next big discovery.“

A Computer Science Approach To A Physics Problem

That question brought a distinctly computer science perspective into a physics setting. Rather than treating the trigger as a fixed control problem, the researchers framed it as a sequential decision-making task, a common idea in artificial intelligence. Instead of responding only to whether the event rate has moved above or below a target, the system looks at recent conditions and tries to learn what kind of change has taken place.

As Ding explained, the goal was not just to react to drift, but to understand it: “The policy can then diagnose why the rate drifted, not simply react to the fact that it drifted.”

That distinction mattered because the Large Hadron Collider is not an environment where an artificial intelligence system can afford to be loosely right. Storage and readout capacities are hard physical limits. If the trigger accepts too much background data, valuable bandwidth is lost and the detector can effectively go blind for brief periods. In that setting, a method that performs well on average but breaks the rules too often is not merely imperfect. It is unusable.

That requirement shaped the team’s main technical insight. Existing reinforcement learning methods often work by comparing a group of possible actions and favoring the ones that seem best. But Chen and his collaborators found that when conditions shifted, those candidate actions could all be poor choices. In those moments, the system risked learning from actions that were unacceptable to begin with.

Their solution was a method called Group-Filtered Policy Optimization. The logic is simple: before deciding which action is best, first determine which actions are safe enough to consider.

In tests on simulated data and on real collision data from the Compact Muon Solenoid experiment, the method improved how often the trigger stayed within its target operating range while also preserving more of the unusual events physicists care about. Just as importantly, the system transferred from simulation to real collision data without needing extra fine-tuning during deployment, a hurdle that often trips up artificial intelligence systems in high-stakes settings.

For Chen, that result helped explain why the paper resonated. “The community has learned to discount RL results that live in simulation,” he said.

A Broader Collaboration With Wider Implications

The paper also builds on an ongoing line of interdisciplinary research. An earlier project from the same collaboration, Adaptive Anomaly Detection in LHC Triggers, led by University of Michigan Ph.D. student Shaghayegh Emami, received a Poster Award on Innovative Result at the Fast Machine Learning for Science Conference in 2024. That work helped lay the foundation for the newer paper by showing how adaptive methods could help trigger systems respond to changing conditions.

The implications may extend well beyond particle physics. The researchers point to similar threshold-based problems in areas such as fraud detection, network security and industrial monitoring, where systems must adapt to changing conditions while operating under strict limits on time, attention or capacity.

For now, the team sees the work as a way of giving scientists better tools, not replacing human judgment. The decisions about what kinds of physics matter most still belong to researchers. But if artificial intelligence can help a system adapt more intelligently in real time, it may help ensure that fewer important events slip away unseen.