Speaker
Description
The High-Luminosity upgrade of the LHC will increase the collision rate by a factor of five, leading to extremely dense environments with a large number of overlapping proton–proton interactions. In this context, the LHCb Upgrade II and its next-generation electromagnetic calorimeter, PicoCal, face major challenges in the accurate reconstruction of photons, electrons, and neutral pions, which are essential for precision measurements. We present a novel Graph Neural Network (GNN)–based reconstruction approach in which clusters of calorimeter cells are represented as graphs. By learning directly from cell-level information, the model suppresses pile-up contributions and significantly improves the energy resolution with respect to current standard reconstruction techniques. By learning detector effects such as energy leakage and shower development directly from data, the approach removes the need for explicit, hand-crafted correction procedures typically used in baseline methods. In addition, a lightweight, attention-enhanced architecture based on GarNet is explored, achieving comparable performance while reducing the inference time by up to a factor of eight. Further acceleration strategies, including student–teacher training and graph-to-MLP compression, demonstrate the potential of these approaches for real-time reconstruction in future LHC runs, including the upcoming Run 4, and in future high-luminosity collider environments.
| I read the instructions above | Yes |
|---|