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

T: Fachverband Teilchenphysik

T 93: Data, AI, Computing, Electronics VIII

T 93.4: Talk

Friday, March 20, 2026, 09:45–10:00, KH 00.024

OmniJet-alpha for Calorimeters - Autoregressive generation of Calorimeter Showers — Joschka Birk¹, Frank Gaede², Anna Hallin¹, Gregor Kasieczka¹, Martina Mozzanica¹, and •Henning Rose¹ — ¹Institute for Experimental Physics, Universität Hamburg Luruper Chaussee 149, 22761 Hamburg, Germany — ²Deutsches Elektronen-Synchrotron DESY, Notkestr. 85, 22607 Hamburg, Germany

We present an autoregressive approach for the generation of high-granularity calorimeter showers based on the OmniJetAlpha architecture. The proposed method directly embeds individual calorimeter shower hits, enabling end-to-end autoregressive generation without relying on discrete tokenization or vector-quantized codebooks. To model hit features efficiently, the architecture employs separate prediction heads for each feature dimension, allowing the overall model size to remain compact even at very high spatial and energy granularities, while avoiding codebook collapse and related representational bottlenecks. This design facilitates stable training and scalable generation in regimes where traditional token-based approaches become impractical. Our results demonstrate that generative pre-training can be performed directly at the data level for calorimeter shower modeling, removing the need for intermediate representations. This is a significant step toward leveraging transformer-based foundation models in high-energy physics, as autoregressive pre-training has proven exceptionally effective in generative modeling, as evidenced by recent advances in large language models.

Keywords: Autoregressive Generation; Foundation Models; Calorimeter Showers; Machine Learning; Simulation