9.2 JetNet

We introduce the JetNet dataset as a benchmark dataset for studies of fast simulation techniques in high-energy physics. It is derived from Ref. [315],¹ and comprises simulated particle jets with transverse momenta $p_T^{\mathrm{jet}}\approx 1\text{TeV}$, originating from gluons, light quarks, top quarks, and W and Z bosons produced in $13\text{TeV}$ proton-proton collisions through a simplified detector (Figure 9.1).

As disucssed in Chapter 7, each jets is represented as a point cloud, or particle cloud, of particles, with each particle represented as a node in the cloud with the three kinematic $(p_T^{\mathrm{rel}},{\eta }^{\mathrm{rel}},{\varphi }^{\mathrm{rel}})$ features. $p_T,\eta ,\varphi$ are the transverse momentum, pseudorapidity, and azimuthal angle, respectively, commonly used in collider physics and defined in Chapter 6. $p_T^{\mathrm{rel}}$ is the transverse momentum of the particle relative to the jet $p_T$, and ${\eta }^{\mathrm{rel}}$ and ${\varphi }^{\mathrm{rel}}$ are the particle’s angular coordinates relative to the jet axis.

Details of the simulations are as follows. The parton-level events are first produced at leading-order using MadGraph5_aMC@NLO 2.3.1 [316] with the NNPDF 2.3LO1 parton distribution functions [317]. To focus on a relatively narrow kinematic range, the transverse momenta of the partons and undecayed gauge bosons are generated in a window with energy spread given by $\mathrm{\Delta }{p}_T∕p_T=0.01$, centered at $1\text{TeV}$. These parton-level events are then decayed and showered in pythia 8.212 [313] with the Monash 2013 tune [318], including the contribution from the underlying event. For each original particle type, 200,000 events are generated. Jets are clustered using the anti-$k_T$ algorithm [193], with a distance parameter of $R=0.8$ using the FastJet 3.1.3 and FastJet contrib 1.027 packages [319, 320]. Even though the parton-level $p_T$ distribution is narrow, the jet $p_T$ spectrum is significantly broadened by kinematic recoil from the parton shower and energy migration in and out of the jet cone. We apply a restriction on the measured jet $p_T$ to remove extreme events outside of a window of $0.8\text{TeV}<p_T<1.6\text{TeV}$ for the $p_T=1\text{TeV}$ bin. This generation is a significantly simplified version of the official simulation and reconstruction steps used for real detectors at the LHC, to remain experiment-independent as well as allow public access to the dataset.

Acknowledgements

This chapter is, in part, a reprint of the materials as they appear in NeurIPS, 2021, R. Kansal; J. Duarte; H. Su; B. Orzari; T. Tomei; M. Pierini; M. Touranakou; J.-R. Vlimant; and D. Gunopulos. Particle Cloud Generation with Message Passing Generative Adversarial Networks. The dissertation author was the primary investigator and author of this paper.