Simulating Collision Events using Neural Networks

Abstract

In this work, we study the use of machine learning techniques to generate simulation of pp collisions at the CMS experiment at the LHC. Standard model processes such as W +jets form a background for beyond standard model searches such as the search for vector-like leptons in the one lepton and two jets final state. Full GEANT4 based simulation of these backgrounds is time and compute intensive. However, if one focuses on only the quantities of interest, then the task can be done in a simpler way by using modern tools such as Generative Adversarial Networks (GANs) or variational autoencoders (VAEs). In this work, we implement a basic GAN and a basic VAE to produce event distributions for processes such as Drell-Yan and W +jets. We test the dependence of the GAN and the VAE on the hyperparameters of the corresponding network. We demonstrate the efficacy of using the generated distributions by training a binary classifier to distinguish the W +jets process from semileptonic tt̄ production. We find that an equivalent performance to GEANT4 based simulation can be obtained by instead using the VAE generated output. This shows us that the usage of these algorithms can be used to speed up the generation of simulations for the LHC experiments.