Speaker
Description
The Compressed Baryonic Matter (CBM) experiment at FAIR is designed to explore the phase diagram of strongly interacting matter at high net-baryon densities with unprecedented interaction rates of up to 10 MHz. A key technical challenge in this environment is the reconstruction of charged particle trajectories in real time, under extreme track multiplicities and high background levels. Track finding consumes the majority of CBM’s total online computing resources and plays a critical role in enabling rare signal reconstruction and physics analysis.
We present a Graph Neural Network (GNN)-based algorithm for track finding in the CBM experiment. We use GNN, Multi-Layer Perceptron (MLP) and the Kalman Filter (KF). We use metric learning to construct a sparse graph on which a GNN is applied to filter edges. Adjacent edges are then combined to form triplets which are fitted with the KF and filtered further. Next, overlapping triplets are combined to form track candidates. Finally, we use a MLP classifier to rank and select tracks. This new track competition based on classifier score shows improved performance compared to the widely-used Cellular Automaton (CA).
This GNN algorithm has been implemented in CBMROOT, enabling direct comparison with CA Track Finder. Our GNN algorithm achieves 97% overall tracking efficiency compared to 95.5% of the CA Track Finder. Crucially, our GNN algorithm increases reconstruction efficiency of low-momentum secondary tracks by 7% while maintaining low fake rates. This greatly impacts physics analysis involving the reconstruction of short-lived particles.
Details
Prof. Dr. Ivan Kisel, Uni-Frankfurt, FIAS, Germany
https://www.goethe-university-frankfurt.de/en?locale=en
| Internet talk | No |
|---|---|
| Is this an abstract from experimental collaboration? | No |
| Name of experiment and experimental site | CBM (FAIR/GSI) |
| Is the speaker for that presentation defined? | Yes |