Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

T: Fachverband Teilchenphysik

T 71: Data, AI, Computing, Electronics VII

T 71.5: Vortrag

Donnerstag, 19. März 2026, 17:15–17:30, KH 00.024

Finding Symbolic Representations of Graph Neural Networks used for Track Finding — •Urs Fischer, Sebastian Dittmeier, and Andre Schöning — Physikalisches Institut, Universität Heidelberg, Germany

Graph Neural Networks (GNNs) have been shown to efficiently solve the combinatorial challenge of track finding at high luminosity collider experiments [1]. The available logic resources of Field-Programmable Gate Arrays (FPGAs) limit the size of possible neural networks that can be deployed in hardware triggers or accelerator cards.

In this study, we apply symbolic regression to the different steps of the GNN track finding process. Symbolic regression fits analytic expressions by combining algebraic operators stochastically to find the best representation in terms of simplicity and accuracy. The resulting algebraic functions have the potential to reduce computational costs of the Multi-Layer Perceptrons for FPGA deployment and allow for interpretation of the internal structure of the neural network, by offering an explicit representation of it.

This talk introduces this method and presents first results on its application to tracking at the ATLAS experiment.

[1] S. Farrell et al., "Novel deep learning methods for track reconstruction", in 4th International Workshop Connecting The Dots 2018. 2018. arXiv:1810.06111.

Keywords: Symbolic Regression; Machine Learning; Graph Neural Networks; Track Finding