Training on simulated collisions
CERN plans to train its model on billions of collisions that never happened. There are good reasons, and one real risk.
CERN says its Large Physics Model will be trained on billions of simulated proton-proton collisions at the LHC (CERN). Simulation is how particle physics has always worked: an analysis compares what the detector saw with what the known physics predicts, and the prediction comes from simulated collisions.
How a collision is simulated
Two steps. An event generator such as Pythia works out which particles a collision produces, using the Standard Model's rules (The Pythia 8.3 manual (arXiv 2203.11601)). Then a detector simulation, usually built on CERN's Geant4 toolkit, follows each particle through every layer of the detector and records the signals it would leave (Geant4). The output looks like real detector data, but with the answer attached.
Why not just use real collisions
- Real data comes without labels. In simulation you know which particle made every hit.
- You can make as much as you can afford to compute, including rare processes that real data barely contains.
- Simulation can model detectors that don't exist yet, which is how Flare could help assess future designs.
The risk
Simulation is very good, not perfect. A model trained only on simulated collisions can learn quirks of the simulator that real detectors don't share, and do worse on real data than its tests suggest. Physicists measure and correct this gap for every analysis today. One response from the foundation-model side is to pretrain on real data too: the Aspen Open Jets dataset gives 178 million real CMS jets for exactly that (arXiv 2412.10504). CERN hasn't said whether Flare will see real data.
Can you get the data?
Some of it. The CERN Open Data portal holds more than 5 petabytes from nine experiments, real and simulated (Ten years of the CERN Open Data portal). Whether Flare's training simulations will be published isn't known.