Benchmark Data for AI Safety for High Energy Physics

Datasets for the paper "AI Safety for High Energy Physics" by Ben Nachman and Chase Shimmin (arXiv:1910.08606)

This record contains two files: particles_jj.npz and particles_yz.npz, which contain simulated events of dijet and Z+photon production, respectively, from proton-proton collisions at sqrt(s)=13 TeV.

The parton-level events are generated with MadGraph5 aMC@NLO, which are then passed to Pythia 8 for parton showering and hardonization, and then finally to Delphes3 for ATLAS-like detector simulation. Reconstructed calorimeter towers are clustered using the anti-kT algorithm with radius parameter R=1.0. The highest-pT jet from each event is selected, and only events with jet pT > 300 GeV are saved.

The Npz files contain three dictionary keys:

jets: (N, 4)-shape array containing the pT, eta, phi, and mass of the leading R=1.0 jet for each event

constituents: (N, 128, 3)-shape array containing the pT, eta, phi of up to 128 highest-pT constituent momenta from the leading jet cluster. Jets with fewer than 128 constituents are padded with zero values.

photons: (N, 3)-shape array containing the pT, eta, phi of the leading reconstructed photon (if any) of the event. Events with no photon are filled with zeros.

pT and mass values are stored in units of TeV.