AI programs handle enormous data volumes mainly by filtering, learning, and compressing, and the clearest documented case is particle physics. At CERN's Large Hadron Collider (LHC), trigger systems make a keep-or-discard decision 40 million times per second and retain roughly one event in 20,000, with neural networks now running on the electronics that make the first cut. The same case shows the limit of the approach: whatever the filter discards is gone, so the quality of the filter decides what can ever be discovered.
Key takeaways
- CERN's trigger systems decide 40 million times per second which LHC collision data to keep, about once every 25 nanoseconds.
- Neural networks have been used in particle physics since the 1990s, and since 2018 compressed models can be loaded onto FPGA chips that decide in well under a microsecond.
- CMS runs an autoencoder-based anomaly detector, AXOL1TL, in its first-level trigger to flag unusual events without a predicted signal.
- Filtering is a trade: fast decisions make big data manageable, but discarded events cannot be reanalyzed.
The scale problem at the LHC
CERN describes the LHC's four collision points as producing billions of proton collisions each second. Storing all of it is impossible, so trigger systems filter the stream in real time. In a May 2026 article on CERN's website, physicist Piotr Traczyk writes that trigger decisions are made 40 million times per second, roughly once every 25 nanoseconds, and that only about one in every 20,000 events is selected for readout and storage.
The data that survives is still large. CERN's computing pages say CERN's data center processed on average one petabyte (one million gigabytes) per day during LHC Run 2, and that preparations for Run 3 aimed at storing and analyzing more than 600 petabytes. For the High-Luminosity LHC, scheduled to begin operating in 2030, the same CERN article says ATLAS and CMS are expected to handle detector data at about a quarter of 2025 global internet traffic.
Where machine learning enters the pipeline
Particle physicists have used neural networks since the 1990s, chiefly to identify particle tracks and classify physics processes, according to the CERN article. What changed more recently is where the models run. The first trigger stage uses field-programmable gate arrays (FPGAs), chips that must respond in a fraction of a microsecond. A 2018 paper in the Journal of Instrumentation (JINST) by Javier Duarte and colleagues introduced hls4ml, a package that converts a trained neural network into FPGA firmware. Its example jet classifier fit within modern FPGA resources at a latency of roughly 100 nanoseconds.
CERN states that ATLAS and CMS now run this approach in their first-level triggers during data taking, and that a five-year project called Next-Generation Triggers, launched in January 2024, extends it. Models are compressed so they fit the hardware without a large loss in accuracy, which is why trigger networks are far smaller than the language or image models in the news.
Anomaly detection: filtering without a prediction
Classic triggers look for events matching a predicted signal, such as a known particle. An anomaly detector works the other way. It is trained on randomly selected collisions to learn what ordinary events look like and flags those that do not fit. CERN's article quotes CMS physicist Maurizio Pierini saying this allows physicists to scour LHC data for new phenomena without pre-judging them, and it explains that the technique has to run at trigger level because an interesting event that is discarded cannot be studied later.
CMS has put this into practice with AXOL1TL, an autoencoder-based algorithm. A CERN student report dated December 2, 2024 describes it as deployed in the global trigger of the CMS first-level system, and a CMS contribution submitted on November 29, 2024 describes the model being installed and monitored in a test crate during Run 3. The test crate receives the same input data but does not control readout, so a new algorithm can be checked on live collisions without risking physics data.
Stages of the filter
The table summarizes what the cited sources say about each stage. Cells state only what the sources give.
| Stage | Runs on | What the sources report |
|---|---|---|
| Level-1 trigger | FPGAs and fast optical links | Decides on every crossing, 40 million per second; sub-microsecond latency; neural networks in use |
| Anomaly detector (AXOL1TL) | Inside the CMS global trigger | Autoencoder trained on ordinary collisions; 50 ns latency constraint cited for the low-level data acquisition |
| Higher-level selection | Software on computing farms | Works on the events the first stage keeps; no latency figure used here |
| Storage and analysis | CERN data center, tape archive | About one petabyte per day processed in Run 2 |
As an illustrative example only: a keep ratio of one in 20,000 applied to 40 million decisions per second works out to about 2,000 retained events per second. This is arithmetic on the two quoted figures, not a published rate.
Checks for any claim that AI handles big data
- What is the model deciding: keep or discard, classify, predict, or summarize?
- What does it throw away, and can that data be recovered?
- Was it trained on typical data, and what would count as unusual to it?
- What is the time budget per decision, and does the hardware meet it?
- Is there a test environment, like the CMS test crate, that runs the model beside the live system?
Data volume alone is not the achievement. The LHC example works because the task is narrow, the decision time is fixed, the physics supplies a clear notion of ordinary background, and engineers can compare the filter against held-back data.
FAQ
Does AI analyze all the data the LHC produces?
No. Triggers select about one event in 20,000, and only that portion is stored for analysis. The rest is discarded in real time.
Are these models the same kind as chatbots?
No. Trigger models are small neural networks, including autoencoders, compressed to run on FPGAs within tight time limits. They do not generate text or images.
Why use anomaly detection instead of searching for a known particle?
A known-signal search can only find what theory predicts. An anomaly detector trained on ordinary collisions can flag events that disagree with expectations, whatever the cause.
Can the same approach work for business data?
The principles carry over: define the decision, fix the time budget, keep a test copy, and measure what is discarded. The LHC figures themselves do not transfer to other settings.
Related reading
- How AI handles and preserves data
- AI compared with data science as fields
- AI programs versus traditional software
- AI in data science work
More in AI Basics.
Last verified: October 2026
Sources
- Smarter decisions at the speed of collisions, Piotr Traczyk, CERN, May 7, 2026
- Storage, CERN Computing pages, CERN, 2022 and later
- Fast inference of deep neural networks in FPGAs for particle physics, Duarte et al., Journal of Instrumentation 13 P07027, 2018
- Real-time Anomaly Detection at the L1 Trigger of CMS Experiment, Gandrakota for the CMS collaboration, arXiv, November 29, 2024
- Improving AXOL1TL Trigger Anomaly Detection in CMS Level-1 Trigger, Paucar Mendoza, CERN Students Note, December 2, 2024