Speakers
Description
Low-energy (MeV-scale) neutrino detection in the Deep Underground Neutrino Experiment (DUNE) is essential for supernova and solar neutrino physics, but poses significant reconstruction challenges due to sparse energy depositions radiological backgrounds, and detector noise in liquid argon time projection chambers. We present a graph neural network–based reconstruction framework for DUNE low-energy events using NuGraph, which encodes detector geometry and topology. The model performs key reconstruction tasks, including interaction classification and track reconstruction. As machine-learning–based reconstruction becomes increasingly central to DUNE workflows, scalable and efficient inference is critical. We report progress toward integrating an inference-as-a-service (IaaS) model into the NuGraph-based low energy pipeline, enabling GPU-accelerated inference through decoupled inference servers. This approach improves utilization of heterogeneous computing resources and reduces inference latency relative to node-local execution. Initial performance studies demonstrate reduced inference time and increased throughput. This work advances scalable, GPU-enabled reconstruction for DUNE and provides a path toward deploying machine-learning inference efficiently across distributed scientific computing facilities.