Karlsruhe 2024 – scientific programme

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

T: Fachverband Teilchenphysik

T 77: Invited Topical Talks 3

T 77.1: Invited Topical Talk

Thursday, March 7, 2024, 11:00–11:30, Geb. 30.21: Gerthsen-HS

Track reconstruction for the ATLAS Phase-II Event Filter using GNNs on FPGAs — •Sebastian Dittmeier — Physikalisches Institut, Universität Heidelberg, Heidelberg, Germany

The High-Luminosity LHC poses new challenges for the trigger and data acquisition system of the ATLAS experiment. The reconstruction of charged particle tracks is already now the computationally most intensive task of the trigger. It becomes even more expensive once the new tracking detector, called the Inner Tracker, is installed and the luminosity reaches HL-LHC target levels. To keep the computing resources within their given power, space and cost constraints, a heterogeneous server farm is proposed for the Event Filter, and novel algorithms are investigated.

Over the last years, it has been shown that Graph Neural Networks have great potential to efficiently solve the combinatorial challenge of finding track candidates in dense environments with hundreds of thousands of hits per event. Recent studies conducted for the ATLAS experiment come close to the physics performance of current tracking methods, while offering potential speed-ups. GNNs are well-suited to be implemented on FPGAs because of their intrinsic message passing algorithms, which lead to highly irregular computations and memory access patterns. This talk summarizes the development of the ATLAS Event Filter for HL-LHC, the most recent results of tracking with GNNs for ATLAS, and the translation of these models to FPGAs.

Keywords: ATLAS; Track reconstruction; Graph Neural Networks; FPGAs; Trigger