flarephysicsai

Reconstructing a collision

Turning millions of detector signals into a list of particles, one step at a time. This is the work CERN wants Flare to learn.

Updated

Protons cross inside the big LHC experiments 40 million times a second (CMS). Each crossing leaves signals in millions of detector channels: small hits in silicon near the beam, flashes of energy in the calorimeters further out, and a few long tracks in the muon chambers at the edge. Nobody sees particles directly. Reconstruction is the work of turning those signals back into the particles that made them.

The chain today

  1. Track finding joins the hits in the inner tracker into the curved paths of charged particles. The magnetic field bends them, and the curve gives their momentum.
  2. Vertexing traces tracks back to the points where collisions happened. Dozens of collisions pile up in every crossing.
  3. Clustering groups energy deposits in the calorimeters, where particles stop and shower.
  4. Particle flow combines all of it to decide, particle by particle, what passed through: an electron, a photon, a charged or neutral hadron, a muon (the CMS particle-flow paper).
  5. Identification and jets: algorithms tag what each object was and group sprays of particles into jets.

Today

  1. Detector hits
  2. Track finding
  3. Vertices
  4. Calorimeter clusters
  5. Particle flow
  6. Particle ID
  7. Jets and physics objects

Each step is its own algorithm, tuned by its own team.

The plan

  1. Detector hits
  2. One pretrained model, learned from billions of simulated collisions
  • Reconstruction
  • Particle identification
  • Unexpected signals
  • Detector design studies

The same model tuned for each job.

Our own simplified sketch. The steps differ between experiments, and CERN hasn't said how Flare will be built.

Each of those steps is its own algorithm, tuned by its own experts, and machine learning already sits inside many of them: CMS has used a graph network called ParticleNet and a transformer called UParT to recognise jets from bottom quarks, and ATLAS uses a transformer tagger called GN2 (ATLAS, arXiv 2505.19689).

Before any of that, the trigger

Most collisions are thrown away within microseconds. ATLAS's first-level trigger keeps up to 100,000 events a second and its software trigger about 1,000 of those (ATLAS). CERN says roughly one event in 20,000 is kept. CERN reported in May 2026 that both big experiments now run machine-learning models on chips in their first-level triggers (CERN).

What one learned model has already done

In February 2026 CERN reported the first fully machine-learned reconstruction in CMS: one model, trained on simulated collisions, replaced much of the hand-written particle-flow logic, matched the traditional algorithm, and measured jets 10 to 20% more precisely in some momentum ranges of simulated top-quark events. It runs on GPUs (CERN).

What Flare would change

CERN's release says a foundation model could learn across the whole chain, from hits to particles, in one model (CERN). The nearest published version of that idea is a September 2026 paper whose model takes any mix of tracks, calorimeter cells and particles and predicts the others (arXiv 2609.31862). How far Flare goes, and whether experiments would trust it in their main reconstruction, is not yet known.