Development of a tree-based model for the fast generation of large point clouds representing particle showers in calorimeters

In High-Energy Physics, detailed and time-consuming simulations are needed to describe particle showers in calorimeters. These particle showers are recorded as energy deposits (hits) in the cells of the detector. To mitigate the computational demands of such simulations, surrogate models are widely studied. In this thesis, Generative Adversarial Networks (GANs) are investigated as a fast and flexible approach. If the cells of the detector could be represented by a regular grid, a GAN model would usually use (De-)Convolution layers to up/down-scale the number of voxels. However, due to the often irregular geometry in modern high-granular calorimeters and the small fraction of cells with a hit in such detectors, a grid representation is often not feasible. By representing the shower as a point cloud (PC), i. e. , a set of real vectors, these issues can be addressed. In PCs, the complex dependencies between the points must be correctly modeled. Particle showers are inherently tree-like processes, as each particle is produced by the decay or the detector interaction of a particle of the previous generation. With this inductive bias, a GAN has been developed, that generates such PCs in a tree-based manner. For this model, numerous new components for Graph Neural Networks (GNNs) have been developed that allow up/down-scaling of PCs. This model is applied to two popular benchmark datasets, which both can be represented as PCs: JetNet, a dataset containing jet constituents, and CaloChallenge, a dataset containing particle showers in calorimeters. The novel model achieves a good fidelity on both datasets.