CERN experiments deployed and benchmarked ultra-fast neural network models on hardware triggers to filter 40 million collisions per second.
CERN puts artificial intelligence on chips to catch rare physics
Real-time neural networks on custom silicon let particle detectors spot unexpected collisions within a microsecond.
In a nutshell
CERN is installing ultrafast neural networks directly onto detector microchips to prevent rare particle discoveries from being discarded by accident. With the Large Hadron Collider generating 40 million collisions a second and tens of terabytes of data every second, computers must trash over 99.9% of interactions in real time. By burning compressed AI models into hardware that decides what to keep in under a microsecond, physicists can now hunt for unexpected physics, like dark matter, without relying on rigid theoretical assumptions.
Highlights
- LHC proton beams collide 40 million times every second inside the 27-kilometer ring.
- Collisions generate tens of terabytes of raw data per second, forcing detectors to discard over 99.9% of events instantly.
- Staged neural networks running on custom FPGA chips achieve decision times under one microsecond.
- Research teams benchmarked transformer models and foundation model distillation for Level-1 hardware triggers.
From the Editor’s Diary
When raw data outpaces storage, moving decision-making intelligence directly onto the sensing silicon is the only way to avoid throwing away breakthroughs.
Who's involved
CERN
The European nuclear physics laboratory operating the Large Hadron Collider
goal → Wants to maximize discovery potential while handling extreme data throughput
CMS Collaboration
An international team operating the Compact Muon Solenoid detector
goal → Aims to deploy sub-microsecond AI models on hardware triggers to capture rare physics
ATLAS Collaboration
The research team managing the ATLAS general-purpose particle detector
goal → Seeks to optimize neural networks for on-detector FPGA calorimeter data filtering
Maurizio Pierini
CERN research physicist leading machine learning and trigger development
goal → Wants to advance model-independent anomaly detection inside real-time LHC data acquisition
In short
CERN is embedding artificial intelligence directly onto detector microchips to prevent rare physics from being discarded forever. The European Organization for Nuclear Research, based near Geneva, Switzerland, faces an immense computing bottleneck. Inside its 27-kilometer (17-mile) Large Hadron Collider ring, proton beams smash together 40 million times every second, unleashing tens of terabytes of raw data each second. Because no computer network on Earth can store such a flood, automated trigger systems must throw away more than 99.9% of collisions instantly.
The shift makes it likely that CERN will capture subtle, unpredicted phenomena that traditional rules missed, including long-lived exotic particles and dark matter candidates.
Whether that discovery happens remains uncertain, because it depends on whether novel physics actually exists within reach of the collider and whether unsupervised models can isolate genuine anomalies during future high-luminosity runs without being overwhelmed by ordinary collision noise.
How it unfolded
CERN reveals microsecond machine learning for collider triggers
CERN published an overview showing how machine learning is overhauling real-time data filtering across the ATLAS and CMS experiments. Physicists demonstrated microsecond-latency neural network inference running directly on field-programmable gate arrays, proving custom silicon could filter the onslaught of particle collisions without causing data bottlenecks.
Physicists push AI onto edge chips to flag unexpected physics
Following internal tests, CERN physicists showed that edge algorithms running on detector electronics, nicknamed trigger AI, can flag anomalous particle signatures without pre-set theoretical biases. Research teams accelerated model compression and transformer architectures targeting Level-1 FPGA electronics ahead of High-Luminosity LHC operations.
Where things stand
Real-time neural filtering is operating in staged setups across the CMS and ATLAS Level-1 trigger electronics, with algorithms clocking inference times under one microsecond on dedicated FPGA boards. Teams are training unsupervised anomaly detection models to ensure unpredicted exotic interactions are saved during future high-luminosity runs.