Karlsruhe 2024 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

T: Fachverband Teilchenphysik

T 67: Data, AI, Computing 5 (normalising flows)

T 67.5: Talk

Wednesday, March 6, 2024, 17:00–17:15, Geb. 30.33: MTI

Generating Accurate Showers in Highly Granular Calorimeters Using Convolutional Normalizing Flows — •Thorsten Buss — Institut für Experimentalphysik, Universität Hamburg, Germany

Monte Carlo MC simulations are vital for collider experiments. They allow us to compare experimental findings with theory predictions. These simulations have a high computational demand, and future developments, such as higher event rates, are expected to push the computation needs beyond availability. Generative neuronal networks can alter MC simulations, speeding them up and mitigating the problem.

Last year, we presented a masked auto-regressive flow (MAF) based generation of particle showers. While generating highly accurate showers in highly granular calorimeters is possible, the generation on CPUs is slower than MC simulations. Also, the architecture does not scale well with input dimensions. Therefore, we change the MAF to a coupling-based flow with convolutional sub-networks. This speeds up the model by a significant factor and improves the model's accuracy further.

Lately, point-cloud-based generative models have become popular. These models represent showers as unordered sets of energy depositions characterized by their position in the detector and the amount of energy. Since fixed grid models, such as our flow, are usually trained on the irregular detector geometry, point-cloud-based models are expected to generalize better to different detector geometries. We conduct a systematic comparison between point cloud and fixed gride models.

Keywords: Generative AI; Calorimeter; Particle Shower