The challenge facing experimental particle physics is the never-ending increase in data coming from detector measurements and from Monte Carlo simulations. As a result, machine learning is becoming a standard tool for solving a variety of tasks found in this field of science. This work explores the use of generative models for increasing the final stage statistics of standard simulations by generating synthetic data that follow the same kinematic distributions. We show the use of two types of generative algorithms, variational autoencoders and normalizing flows, which are capable of fast generation of an arbitrary number of new events. As an example of Monte Carlo simulated data we use a theoretical Higgs boson production beyond the Standard Model. In this work we investigate the applicability of different types of the two methods with different model parameters and numbers of initial events used in training. The resulting event distributions are compared with original Monte Carlo distributions using statistical tests, to evaluate their similarity and quality of reproduction.
|
