The AI step that turns a particle collision into a usable inventory

A collision at the Large Hadron Collider leaves measurements in different detector layers. Reconstruction software has to decide which pieces of evidence belong together and what produced them.[2]

At CMS, the tracking detector records the paths of charged particles. Their curvature in a magnetic field helps determine momentum. The electromagnetic calorimeter measures the energy deposited by electrons and photons using crystals that emit light. These instruments provide complementary information about the same event.[3][4]

The combination creates a subtle accounting problem. A charged particle can leave a track and an energy deposit; treating those as two unrelated particles would count the same physical contribution twice. Conversely, an energy deposit without a matching track can still represent a real particle. Reconstruction has to reconcile the observations.

Learning which signals belong together

Particle flow performs that reconciliation through physics-informed rules. CMS’s machine-learning alternative, MLPF, learns from simulated collisions and produces the particle candidates together.[2]

CMS’s August paper describes a transformer that relates reconstructed tracks and calorimeter clusters, then assigns particle types, energy and momentum. Upstream software still builds those tracks and clusters. Downstream software can use the resulting particle list through its existing interfaces.[1]

This is a useful place to apply attention, the transformer mechanism that weighs relationships between inputs. The meaning of one detector measurement depends partly on its neighbours. A track and a nearby cluster may be evidence about the same particle; another combination may require a different interpretation. The model evaluates those relationships across the event.

Sharper jets, faster reconstruction

The physics result includes a 10–20% improvement in jet energy resolution in simulated top quark–antiquark events under Run 3 conditions. The comparison concerns jets with transverse momentum of 30–100 GeV. A jet is a cluster of particles analysed together; better resolution means less spread around its known simulated energy after reconstruction.[1][2]

The collaboration also checked reconstructed distributions against real 2024 collision data. Simulation provides known particles for training; detector data tests whether that reconstruction remains usable.[1]

In the reported setup with 64 parallel inference streams, the particle-flow step took about 20 milliseconds per event on an Nvidia L4 GPU, versus 110 milliseconds for conventional particle flow on the CPU, including transfers and preparation. This is not the complete reconstruction chain: particle flow occupies a relatively small share of Run 3 processing.[1]

The distinction makes the engineering result more informative. A faster component becomes valuable when it improves the throughput of the surrounding system; the model’s GPU work and the remaining CPU work have to coexist. Here, the team has demonstrated integration into CMS software as well as a favourable timing comparison.

Many later measurements depend on this intermediate particle inventory. Improving how signals are combined can sharpen the scientific use of an existing detector.[2]

Report an error or suggest a correction