Advancements in The Simulation of High Granular Calorimeters for High Energy Physics using Generative Machine Learning Techniques

Abstract

The field of Particle Physics heavily relies on simulated data in order to perform analyses. The increase in the amount of measured data translates in the need for more simulations used to compare out current knowledge of Nature to actual measurements. One reason for a drastic increase of measured data in the future is the High Luminosity upgrade at the LHC, which will feature collisions at a much higher rate thus drastically increasing the number of measured collisions. Of particular interest for the scope of this work is the CMS High Granular calorimeter (HGCal), which will replace CMS’s current endcap calorimeters. This detector comprises circa 3 million readout hexagonal cells per endcap, making it a machine capable of producing fine-grained showers. It will also implement a system capable of recording the time of a hit measurement with a resolution of circa 30 ps, which will help with pile-up rejection and track reconstruction. Generative Machine Learning has risen recently as it has the potential to augment standard simulation techniques. This thesis focuses on multiple generative models that bring us closer to the goal of faster and more accurate simulation. The first study was performed on Graph Neural Networks, as graphs are a very natural way to describe electromagnetic showers, but this model architecture lacks in terms of scalability. We found that there is value in utilizing already given information like the geometry of the calorimeter to train such a network, but the high cardinality led us toward the direction of graphs that could grow using that information but only until the needed shower size instead of using the whole number of cells available at all times. As this first attempt proved to be too challenging and the technology evolved, we then moved on with the EPiC GAN model, which showed good fidelity and high generation speed on showers with reduced complexity but failed to scale up to the cardinality of the HGCal. Finally, we implemented CaloClouds II, a model that is a combination of a continuous-time diffusion model and normalizing flow, to not only be able to successfully simulate the HGCal calorimeter but to do so by also including the time-of-hits feature which will be a crucial integration in this detector upgrade.