GRAPH NEURAL NETWORK ARCHITECTURES FOR FAST SIMULATION AND MUON MOMENTUM INFERENCE AT THE CMS DETECTOR

Abstract

Accurate and fast simulation of particle physics processes is crucial for the high-energy physics community knowing that simulating the particle showers and interactions in the detector is both time consuming and computationally expensive. Classical fast simulation techniques based on non-parametric methods can improve the speed of the full simulation but suffer from lower levels of fidelity. For this reason, alternative methods based on machine learning can provide faster solutions, while maintaining a high level of fidelity. The main goal of a fast simulator is to map the events from the generation level directly to the reconstruction level. In this thesis work, we present novel approaches for Fast Simulation and Muon Momentum inference in High Energy Physics. More specifically, we explore the potential of graph neural networks in various applications including fast simulation of boosted jets and muon momentum estimation in the CMS detector. We introduce a graph neural network-based autoencoder model that provides effective reconstruction of calorimeter deposits using the earth mover distance metric. We also propose to use graph networks to infer the momentum of muons in the Cathode Strip Chambers given their ability to account for the several features affecting the particles' trajectories. We show that graph-based architectures outperform conventional deep learning baselines in terms of accuracy and result in relatively competitive inference and training time. In addition, we optimize our code for training using NVIDIA Nsight Tools and later investigate the scalability of our Fast Simulation graph model on multiple GPUs for which we get speedups of 1.62, 2.19 and 2.73 while scaling from 2 to 4 devices, respectively.