Generative Models for the Cherenkov Detector Description Speed-Up

Modern experiments at the Large Hadron Collider require a large amount of simulated data to find deviations from the Standard model of elementary particle physics. Creating such data takes a lot of CPU time and resources in general. Data simulation at the Collider is performed using computationally expensive physically motivated models. The Cherenkov detector is an essential part of the LHCb detector, one of the four main LHC experiments. The physical simulation of this detector is complicated by the need to account for the propagation of each photon in the substance, which leads to a long simulation time. The new method \cite{main} suggests using generative-adversarial neural networks instead of such models. It provides a comparable quality of simulation, while significantly reducing the time spent on calculating a single event compared to existing methods. This work is aimed at further improving the new approach. The key idea is to use knowledge distillation techniques to reduce complexity and speed up event generation by neural networks. Distillation was applied to the network for this task.