A team led by University of Chicago computer scientist Asst. Prof. Yuxin Chen has developed an AI method to help CERN's Large Hadron Collider make smarter, real-time decisions about which particle collision data to preserve. The research, which earned a best paper award at the ICML 2026 Workshop on AI for Physics, addresses a fundamental bottleneck: the collider generates far more collisions than can be stored, and discarding the wrong ones means losing potentially discovery-critical events forever.
The project merges computer science and experimental physics through a collaboration with Prof. David Miller from UChicago's Department of Physics, along with colleagues from the University of Michigan and Fermilab. Ph.D. student Zixin Ding worked alongside Chen to frame the collider's trigger system as a sequential decision-making problem rather than a fixed control threshold.
The trigger problem
At the Large Hadron Collider, proton collisions occur at an extraordinary rate. A hardware and software system called the trigger acts as a rapid filter, deciding within microseconds which events to record and which to discard permanently. "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.
These decisions carry high stakes. Set thresholds too loosely, and the system drowns in background noise, overwhelming storage and readout capacity. Set them too strictly, and rare events that could point to new physics disappear. Traditionally, expert physicists set these thresholds by hand and adjust them periodically, but collider conditions drift throughout a run-beam intensity shifts, background patterns change, and a setting that worked at noon may fail by evening.
Learning why conditions shift, not just that they shifted
The team took a distinctly computer science approach. Instead of treating the trigger as a control problem that simply reacts when event rates exceed a target, they built a system that tries to diagnose the underlying cause of the drift. "The policy can then diagnose why the rate drifted, not simply react to the fact that it drifted," Ding explained.
This matters because the collider imposes hard physical constraints. If the trigger accepts too much background data, the detector can go blind for brief periods-lost time that cannot be recovered. A method that performs well on average but occasionally violates those constraints is unusable in practice. The researchers found that standard reinforcement learning techniques often failed under shifting conditions because all candidate actions could become poor choices, leaving the system learning from unacceptable options.
Filtering actions before choosing the best one
Their solution, called Group-Filtered Policy Optimization, adds a safety step before decision-making: first determine which actions are safe enough to consider, then pick the best among them. In tests on both simulated data and real collision data from the Compact Muon Solenoid experiment, the method kept the trigger within its target operating range more reliably while preserving more of the unusual events physicists seek.
Notably, the system transferred from simulation to real collision data without requiring additional fine-tuning during deployment-a hurdle that often derails AI systems in high-stakes settings. "The community has learned to discount RL results that live in simulation," Chen said, which helps explain why the paper resonated with reviewers.
From particle physics to broader monitoring problems
The work builds on an earlier project from the same collaboration, led by Michigan Ph.D. student Shaghayegh Emami, which received a Poster Award on Innovative Result at the Fast Machine Learning for Science Conference in 2024. That foundation showed how adaptive methods could help trigger systems respond to changing conditions.
The implications extend beyond particle physics. The researchers point to similar threshold-based challenges in fraud detection, network security, and industrial monitoring-any domain where systems must adapt to shifting conditions while operating under strict limits on time, attention, or capacity. Miller described the vision 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."
The work is part of a growing intersection between AI for Science & Research, where computational methods are shaped by the real-world constraints of scientific instruments. For research scientists looking to apply similar techniques, structured training paths like AI for Research Scientists provide grounding in the sequential decision-making and reinforcement learning approaches that underpin this work.
Why this matters for science and research professionals
The paper demonstrates a practical pattern that transfers across scientific domains: when instruments generate data faster than it can be stored, AI can act as an adaptive filter rather than a static rule. The key insight is not that AI makes decisions faster-it is that the system was designed to respect hard physical constraints first and optimize second. For researchers managing high-throughput experiments, telescopes, or sensor networks, the approach offers a template for building trigger and filtering systems that remain safe under drift while losing fewer rare events. The decisions about what physics to prioritize still belong to scientists; the AI handles the real-time adaptation that human operators cannot match.
Your membership also unlocks:
