Parts | Days | Selection | Search | Updates | Downloads | Help
T: Fachverband Teilchenphysik
T 56: Data, AI, Computing, Electronics VI
T 56.4: Talk
Wednesday, March 18, 2026, 17:00–17:15, KH 02.014
Cross-Geometry Transfer Learning in Fast Electromagnetic Shower Simulation — •Lorenzo Valente1, Gregor Kasieczka1, and Frank Gaede2 — 1Institut für Experimentalphysik, Universität Hamburg, Luruper Chaussee 149, 22607 Hamburg, Germany — 2Deutsches Elektronen-Synchrotron DESY, Notkestr. 85, 22607 Hamburg, Germany
Fast and accurate particle shower simulation is essential for high-energy physics, but traditional Monte Carlo methods like Geant4 are computationally expensive, while machine learning alternatives typically require complete retraining for each detector geometry. We present a transfer learning approach for generative calorimeter simulation using point cloud representations and diffusion/flow models. By pretraining on detector configurations and fine-tuning with limited target data, our method enables efficient adaptation across diverse geometries without geometry-specific preprocessing. We demonstrate significant performance improvements with minimal training samples through both full fine-tuning and parameter-efficient adaptation strategies. This work establishes transfer learning as a practical technique for geometry-flexible fast simulation, reducing computational requirements for detector design studies and physics analyses.
Keywords: Transfer learning; Fast simulation; Calorimeter simulation; Point cloud representation; Generative models