Learning Standard Model structure from LHC data with Riemannian flow matching
A single generative model learned recognizable Standard Model features from raw ATLAS collision data.
The paper introduces ShellFlow, a transformer-based Riemannian flow matching model trained on about \(10^9\) real 13 TeV proton-proton collision events from ATLAS Open Data. With only the on-shell condition and invariant-mass formula built in, it reproduced particle kinematics, known dilepton resonances, the leptonic Weinberg angle, and the \(W\) and top-quark masses. The authors argue this shows a substantial fraction of Standard Model structure can be learned directly from recorded collider data. ArXiv · AI/CL/LG's note
The paper introduces ShellFlow, a transformer-based Riemannian flow matching model trained on about \(10^9\) real 13 TeV proton-proton collision events from ATLAS Open Data. With only the on-shell condition and invariant-mass formula built in, it reproduced particle kinematics, known dilepton resonances, the leptonic Weinberg angle, and the \(W\) and top-quark masses. The authors argue this shows a substantial fraction of Standard Model structure can be learned directly from recorded collider data. ArXiv · AI/CL/LG's note
score 5