Speaker
Description
Can a generative model learn the structure of the Standard Model directly from collider data, without being told which resonances or correlations to expect? We present ShellFlow, a transformer-based Riemannian conditional flow-matching model trained on approximately one billion proton-proton collision events from the ATLAS 13 TeV Open Data release. Conditioned on the observed event composition, the model generates particles directly on their physical mass shells, using only the on-shell constraint and invariant-mass relation as explicit physics priors. A single model spans five orders of magnitude in invariant mass and reproduces single-particle kinematics, the J/ψ, Υ, and Z resonances, the leptonic Weinberg angle, the W and top-quark masses, and nontrivial inter-particle correlations absent from its training objective. These results show that a substantial fraction of Standard Model structure can be learned directly from recorded LHC data, motivating new approaches to data-driven simulation, representation learning and anomaly detection. Paper: arXiv:2607.16144