guest ::
| Home > Publications database > Application of Generative Neural Networks to High Fidelity Simulation of Particle Detectors in High Energy Physics |
| Home > Publications database > Application of Generative Neural Networks to High Fidelity Simulation of Particle Detectors in High Energy Physics |
| Book/Dissertation / PhD Thesis | PUBDB-2026-01015 |
; ;
2026
Verlag Deutsches Elektronen-Synchrotron DESY
Hamburg
This record in other databases:
Please use a persistent id in citations: doi:10.3204/PUBDB-2026-01015
Report No.: DESY-THESIS-2026-007
Abstract: Monte Carlo simulation of particle interactions underpins nearly every aspect of modern collider experiments, yet its computational cost will exceed available resources as luminosities and detector complexities increase.Calorimeter shower simulation accounts for the dominant share of this cost, particularly for the highly granular designs planned for future facilities.Deep generative Machine Learning models offer a promising path toward addressing this bottleneck by learning from relatively small amounts of Monte Carlo data and then sampling new showers more efficiently.This work introduces CaloClouds, a generative model for fast electromagnetic shower simulation in highly granular sampling calorimeters, developed and evaluated using the silicon-tungsten electromagnetic calorimeter of the International Large Detector (ILD) as a case study.In contrast to previous approaches built on fixed voxel grids, CaloClouds represents showers as point clouds -- sets of energy-weighted coordinates with continuous spatial positions.This formulation minimizes the projection artefacts that arise when mapping grid-based outputs onto irregular detector readout geometries and enables the use of a single model across different detector positions with the same longitudinal structure.The model combines two complementary components: ShowerFlow, a normalising flow that captures global properties such as per-layer energies and number of energy depositions, and a diffusion model that generates individual energy depositions.A dedicated preprocessing pipeline converts Geant4 simulation steps into high-resolution point clouds, providing training data at finer granularity than the physical readout cells.To disentangle the contributions of data representation from model performance, this work introduces optimal shower generators constructed by projecting Geant4 steps onto virtual grids of varying resolution without generative modelling.Benchmarks against these references show that CaloClouds achieves fidelity close to the finest-grained optimal generator, indicating that the model reaches its maximum potential given the data representation.Simultaneously, comparisons with the coarse-grained optimal shower generators reveal fundamental limitations of fixed-grid representations independent of the generative model architecture.Full integration into the ILD software chain via the DDML library enables evaluation with standard reconstruction tools, including Pandora particle flow.Physics benchmarks span single-photon observables, di-photon separation, and $\pi^0$ reconstruction in $e^+e^-\rightarrow~\tau^+\tau^-$ events, with the majority demonstrating state-of-the-art agreement with Geant4.On CPU hardware, CaloClouds achieves more than 120-fold acceleration relative to full Geant4 simulation, while on GPU hardware, speedups of several thousand-fold are achievable.Taken together, these results establish CaloClouds as an effective solution that balances speed and accuracy for fast calorimeter simulation, providing a foundation for meeting the simulation demands of next-generation collider experiments.
|
The record appears in these collections: |
PDF 
PDF (PDFA) PDF PDF (PDFA)