Parts | Days | Selection | Search | Updates | Downloads | Help
T: Fachverband Teilchenphysik
T 102: Invited Topical Talks III-B
T 102.2: Invited Topical Talk
Thursday, March 23, 2023, 14:20–14:40, HSZ/0004
Enhancing the CMS Level-1 Trigger with real-time Machine Learning — •Artur Lobanov — Institut für Experimentalphysik, Universität Hamburg, Hamburg, Germany
The Level-1 Trigger (L1) is the first stage of the online event filter system of the CMS Experiment at the LHC. It reduces the event rate from 40 MHz to O(100) kHz by reconstructing, identifying and filtering collision events in real-time using dedicated processing hardware based on field-programmable gate arrays (FPGAs).
Following the success of machine learning (ML) in enhancing event selections in the offline analysis of recorded data, ML algorithms are finding their way into the real-time processing of the CMS L1 Trigger system. Contrary to current filters that rely on simple rule-based selection algorithms using the detected physics objects, ML allows to capture deeper correlations between and within the objects, improving the identification of the event.
In addition to the tight constraints on the processing latency of several microseconds, trigger algorithms also have to fit into the restricted processing resource budget of the FPGAs. This requires a dedicated optimisation of ML models for their use in hardware in these challenging conditions.
In this talk I will outline the basics of the CMS L1 Trigger system, the principles of ML inference in FPGAs, and present the current state-of-the-art developments of novel ML algorithms enhancing the trigger performance at the LHC and beyond.