guest ::
| Home > Publications database > Generating Calorimeter Showers as Point Clouds |
| Home > Publications database > Generating Calorimeter Showers as Point Clouds |
| Book/Dissertation / PhD Thesis | PUBDB-2025-01839 |
; ;
2025
Verlag Deutsches Elektronen-Synchrotron DESY
Hamburg
This record in other databases:
Please use a persistent id in citations: doi:10.18154/RWTH-2025-06385 doi:10.3204/PUBDB-2025-01839
Report No.: DESY-THESIS-2025-011
Abstract: Simulating particle interactions within high-energy physics detectors is essential for interpretingexperimental data and advancing our understanding of fundamental physics. Calorimetersmeasure particle energies through cascades known as showers, but their complex responsesand the computational intensity of traditional simulation tools like Geant4 posesignificant challenges. These are further compounded in high-granularity calorimeters, whichconsist of millions of cells yet register energy deposits in only a sparse subset, renderingfull-scale simulations impractical.This thesis addresses the need for efficient and accurate calorimeter simulations by developingnovel generative machine learning models that leverage the inherent sparsity and pointlikenature of calorimeter data. Initial efforts using voxel-based generative adversarial networks(GANs)—which model data in a discrete grid structure—encountered scaling issuesdue to data sparsity and high dimensionality. By shifting to a point cloud representation, theCaloPointFlow model was developed, marking the first application of point cloud generativetechniques to calorimeter simulation. This approach reduces data complexity by treatingshowers as collections of points in space rather than densely populated grids.Although CaloPointFlow marked a significant advancement, limitations such as inadequatepoint-to-point information exchange and difficulties in modeling discrete coordinate positionswere observed. To overcome these challenges, CaloPointFlow II introduced DeepSet-Flow, a novel normalizing flow architecture that enables direct interactions between points,capturing complex dependencies within the data. Additionally, a new dequantization strategycalled CDF-Dequantization was implemented to better map discrete cell positions tocontinuous space, along with a mitigation strategy to handle multiple energy deposits percell.Despite these improvements, purely point cloud-based models struggled to ensure one hitper calorimeter cell. To resolve this, CaloHit was introduced, a hybrid approach that combinesvoxel-based and point cloud methodologies. CaloHit first generates a hitmap to identifyactive cells using a voxel-based method and then predicts the energies of these hits witha point cloud-based model. This two-stage process effectively addresses the primary limitationsof previous models by ensuring that all calorimeter cells are sampled without replacementand that the one-hit-per-cell constraint is maintained.Evaluations of these models demonstrated their potential to accurately reproduce calorimetershowers while significantly reducing computational resources compared to traditionalmethods. Preliminary tests of the CaloHit approach showed promising results, indicatingthe feasibility of scaling this method to more complex and higher-dimensional datasets.In conclusion,this thesis contributes to the advancement of calorimeter surrogate modeling byintroducing and refining innovative generative models that effectively handle the complexitiesof high-granularity, sparse data. The proposed methods lay the groundwork for scalable,efficient, and experimentally validated simulation tools. Future work will focus on furtherimproving these models, exploring more sophisticated techniques such as diffusion models orconditional flow matching, and validating their performance in real experimental settings.The developments presented here hold significant potential for enhancing the efficiency andaccuracy of particle physics simulations, ultimately aiding in the pursuit of new discoveriesin the field.
|
The record appears in these collections: |
PDF
PDF (PDFA)