Graph networks show great promise for HEP tracking on detectors ranging from silicon trackers to LAr TPCs. As the demonstrators are applied to increasingly realistic datasets, they face computing and physics performance challenges. This workshop aims to discuss these challenges and new ideas to address them. The list includes (but is certainly not limited to)
- Size/purity/efficiency tradeoff for the graphs presented as input to the GNNs
- Parallelization strategies for tracking graph networks (Region of Interest tracking, geometric partitioning, etc.)
- Graph (and model) compression, pruning, partitioning for online applications
- Model resilience to detector effects (noise, alignment, etc.) and new physics domains (e.g., large radius tracking).
- Heterogeneous graph networks operating on multiple detector types and at multiple abstraction levels (e.g., channel, cluster, view, track, particle)
- Physics-informed models incorporating symmetry constraints, etc.
- Object condensation, instance segmentation, and other approaches to extract directly track parameters
- Graph-level inference for event classification
This workshop will be a one day event held at Princeton University, following the Connecting the Dots workshop. In person and virtual participation is planned.