Speakers
Description
Graph Neural Networks (GNNs) excel at modeling the complex, irregular geometry of modern calorimeters, but their computational cost poses challenges for real-time or resource-constrained environments. We present lightweight, attention-enhanced GNNs built on node-centric GarNet layers, which eliminate costly edge message passing and provide learnable, permutation-invariant aggregation optimized for fast inference and firmware deployment. Tailored for particle reconstruction in the proposed PicoCal for the LHCb Upgrade II, these architectures achieve up to 8× faster inference than traditional message-passing GNNs while maintaining superior energy-resolution performance compared to conventional reconstruction algorithms.
To further reduce latency, we evaluate two compressed variants: a compact GarNet student with ~40% fewer parameters that preserves the teacher’s performance, and a knowledge-distilled MLP trained on GarNet’s latent graph embeddings—a Graph(GarNet)-to-MLP approach—that provides an additional 2–6× speedup and even surpasses the GarNet teacher in energy resolution despite a ~95% reduction in model size. Together with ongoing firmware-level integration for real-time filtering in the LHCb trigger system, this work demonstrates a practical and scalable pathway for deploying high-performance, graph-based calorimeter reconstruction in future high-rate particle-detection pipelines.