Speaker
Description
Liquid Argon Time Projection Chamber (LArTPC) detectors provide excellent spatial resolution, offering the potential to reconstruct a broad range of complicated topologies such as tau neutrino interactions, atmospheric neutrinos, nucleon decay events and others. However, realizing this potential requires sophisticated reconstruction algorithms capable of leveraging the wealth of information provided by the detector.
The Exa.TrkX project developed sophisticated Graph Neural Network (GNN) architectures to meet this challenge, utilizing an attention message-passing network over a heterogeneous graph structure of 2D planar detector hit nodes connected by 3D nexus spacepoint nodes. The NuGraph2 architecture performed filtering and semantic segmentation on detector hits in the MicroBooNE open data release, filtering cosmic background hits with 98% efficiency and labelling particle type with 95% efficiency. This mechanism was expanded by the NuGraph3 architecture to form a fully hierarchical GNN, adding decoders that produce outputs at the event and particle levels under a fully consistent representation. The NuGraph architectures have been deployed across a variety of LArTPC detector geometries including MicroBooNE, ICARUS, SBND and the DUNE Far Detector.
This poster will present recent developments to NuGraph3's clustering task, which utilizes the Object Condensation loss function to group detector hits into the reconstructed particle nodes that form the intermediate level of the graph's hierarchy. Moving the generation of the object condensation embeddings from a decoder step into the core message-passing loop yields improved performance, as the model is able to refine these embeddings during the forward pass. Results will be presented on the MicroBooNE open data release, with clustering performance quantified in terms of Adjusted Rand Score, along with time and memory benchmarks for model training and inference.