| Thesis | |
| Report number | CERN-THESIS-2023-222 ; 10.18154/RWTH-2023-09302 ; https://publications.rwth-aachen.de/record/969743/files/969743.pdf |
| Title | Deep Learning and Quantum Generative Models for High Energy Physics Calorimeter Simulations |
| Author(s) | Rehm, Florian (Rheinisch Westfaelische Tech. Hoch. (DE)) |
| Publication | 306. |
| Thesis note | PhD : RWTH Aachen University : 2023-04-19 |
| Thesis supervisor(s) | Borras, Kerstin ; Pooth, Oliver |
| Note | Presented 25 Aug 2023 |
| Subject category | Detectors and Experimental Techniques ; Quantum Technology |
| Abstract | The aim of this thesis is to explore Deep Learning and Quantum Computing approaches to reduce the challenging computing resource requirements of future High Energy Physics detector simulations while maintaining the necessary accuracy. In the forthcoming High Luminosity phase at the Large Hadron Collider at CERN, the number of collisions is increasing drastically. The demand for detector simulations to analyze the collision products is growing by the same scale. This causes an acute shortage in computing resources in both hardware and efficient software. Therefore, resource-optimizing simulation alternatives are needed, and novel technologies to be explored. In this thesis, Deep Learning Generative Adversarial Networks (GANs) are investigated as a first option to be deployed for a future representative high-granularity calorimeter. Novel neural network architectures are developed and reduced precision computing techniques applied to optimize the GAN models. The results are realistic and accurate models that are about five orders of magnitude faster than the current Monte Carlo simulation. The second study of this thesis addresses the novel and rapidly evolving field of Quantum Computing for a simplified calorimeter. The initial attempt employs a model available in the IBM software suite that has been successfully adapted and modified. It runs in a hybrid GAN mode employing a quantum generator and a conventional discriminator. However, this model learned only average distributions and suffers from fluctuations. Therefore, a new fully quantum model was designed to produce individual images. The in this thesis developed Quantum Angle Generator achieves extremely precise results. Due to its noise robustness and lightweight training, it is successfully executed on current noisy quantum devices with a limited number of qubits. The superior expressive power of quantum circuits compared to classical neural networks, combined with a potential exponential scaling, makes Quantum Computing a promising future computing approach. This thesis has pioneered and conducted studies which confirm that this potential has a realistic chance to materialize. |
Corresponding record in: Inspire
Email contact: [email protected]
記錄創建於2023-10-30，最後更新在2023-11-07
