At the Large Hadron Collider, protons smash into each other like two trucks full of confetti. The debris, dozens of new particles, scatters in all directions. The detector records their tracks, and then the real work begins: piecing together a complete picture of the collision from those tracks.
Previously, this was done by a set of rules devised by physicists. Each particle type was processed separately, according to its own laws. It worked, but slowly and not always accurately. Now, instead of a set of rules, there's a single neural network. It was trained on millions of simulated collisions, and it learned on its own to reconstruct the entire picture.
The neural network runs on a graphics card — the same kind that renders graphics in games. It takes 20 milliseconds per collision. The old method took 110 milliseconds. And for jets, streams of particles flying in one direction, the accuracy of energy measurement improved by 10–20 percent.
This approach is like a translator who used to translate word by word using a dictionary, but now understands the whole phrase. As a result, the collision picture becomes clearer, like a photograph taken in good weather.
as written by the authors arXiv:2601.17554
The particle-flow (PF) algorithm constructs a global description of each particle collision by producing a comprehensive list of final-state particles, and is central to event reconstruction in the CMS experiment at the CERN LHC. The existing PF implementation relies on physics-motivated heuristics and assumptions that can be replaced by machine-learning (ML) models trained directly on simulated data and naturally suited to modern graphics processing units (GPUs). A state-of-the-art ML-based PF (MLPF) reconstruction algorithm, implemented within the CMS software framework, is presented. The MLPF algorithm performs a learnable full-event reconstruction on GPUs, generalizes across detector conditions and collision energies, and replaces multiple modular reconstruction steps with a single unified model. Physics performance comparable to standard PF reconstruction is achieved in both simulation and data, with improved jet energy resolution and inference time. In simulated top quark-antiquark events under LHC Run-3 (2023$$-$$2024) conditions, the jet energy resolution improves by 10$$-$$20% for jets with transverse momentum between 30$$-$$100 GeV. Inference time is evaluated using simulated multijet events, with a median of 20 ms per event on an Nvidia L4 GPU, compared to approximately 110 ms for the standard CMS PF reconstruction.
The neural network processes one collision in 20 milliseconds — faster than a human blink.