Erlangen 2026 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
T: Fachverband Teilchenphysik
T 34: Data, AI, Computing, Electronics IV
T 34.6: Talk
Tuesday, March 17, 2026, 17:30–17:45, KH 02.014
Implementation of a reliable ML model life cycle for the CMS Phase-2 L1 Trigger Upgrade — •Leon Joel Kerner1,2 and Alexander Schmidt1 — 1Physics Institute IIIA, RWTH Aachen University, (DE) — 2CERN, (CH)
To achieve the ambitious goals of the High-Luminosity LHC upgrade, a new Level-1 trigger must be developed for the CMS experiment. Machine Learning based models will be deployed in the trigger system, which introduces a range of new challenges. The development of such models involves many individual steps. Any change in the configuration or the data can require repeating the entire workflow, and these steps are currently carried out manually by the model developers. In addition, the deployment of trained models in the trigger requires robust procedures that ensure long-term quality and stability. To address these issues, methods from machine learning operations (MLOps) must be integrated into the workflow.
To address this problem, a GitLab CI/CD pipeline and a training infrastructure consisting of a MLflow server and a WebEOS instance was created to log and manage model training sessions.
Keywords: MLOps; Machine Learning; Electron-ID; CMS Phase-2 Level-1 trigger; Hi-Lumi-LHC