Deep Generative Models for Fast Simulation of the Tracker et the MPD Experiment at NICA

In high energy physics complex and expensive equipment is required to conduct an experiment. Thus it is useful to be able to run a fast simulation. There exist simulations based on exact event reconstruction such as Geant4 (GEomentry And Tracking) which uses Monte Carlo methods. However such simulations are complex and slow therefore cannot be used in runtime applications. In previous works machine learning methods proved able to generate a particle detector response. The general idea is to think about these responses as samples from multi-dimensional distribution and to learn sampling from this distribution. Since our data is structured as two-dimensional matrices we can consider this as image generation task. GANs which learn to convert some simple distribution into a goal distribution are well-suited for such problems. Earlier works showed that GANs are able to achieve high generation quality for this task and are potentially much faster than exact event reconstruction methods. In this work we study such models for the TPC detector of the MPD experiment at the NICA accelerator complex and try optimizing them without quality loss. Simulation speed is crucial since reverse task is solved using simple iteration methods. More specifically generation is done for many different input parameters and parameters corresponding to generated signals which are the closest to real (derived from experiment) are chosen as reverse problem solution. Since our problem is not a simple image generation there is no native metric to evaluate the results. Therefore in this work we come up with our own unique metrics and are trying to justify them. In the end we achieve quite fast generative model with satisfying quality metrics values.