9.1 Simulating jets

We test the performance of our generative models on the simulation of high energy jets, from parton-level inputs through parton showering and hadronization up to emulating the detector response. As discussed in Chapter 4.1, high-energy proton-proton collisions at the LHC produce elementary particles like quarks and gluons, which cannot be isolated due to the QCD property of color confinement. These particles continuously radiate or into sets of particles, a process referred to as parton showering. Eventually, they cool to an energy at which they undergo the process of hadronization, where the fundamental particles combine to form color-neutral hadrons, such as pions and protons. The final set of collimated hadrons produced after such a process is referred to as a jet.

The task of simulating a single jet can be algorithmically defined as inputting an initial particle, which produces the jet, and outputting the final set of particles a.k.a. the jet constituents. Typically in HEP the parton shower and hadronization are steps that are simulated sequentially using MC event generators such as pythia [313] or herwig [314]. Simulating either process exactly is not possible because of the nonperturbative nature of QCD at low energies, and instead these event generators fit simplified physics-inspired stochastic models, such as the Lund string model for hadronization [142], to existing data using MC methods. The present work can be seen as an extension of this idea, using a simpler, ML-based model, also fitted to data, for generating the jet in one shot. Effectively, we are trading the interpretability of MC methods for the speed of GPU-accelerated ML generators.