Robustness of GNN-Based End-to-end Reconstruction Algorithms for HGCAL

End-to-end machine learning models are a promising approach for event reconstruction in particle detectors such as the CMS High-Granularity Calorimeter (HGCAL) at the High-Luminosity LHC. Because these models are trained entirely on simulation, an open question is how robust they remain under realistic, imperfect conditions. We present a case study of a trained graph neural network and introduce a toolkit to evaluate robustness against variations in shower modeling and detector operating conditions. The network shows general stability under moderate noise increases and missing layers, but care in training will be required to ensure translational invariance and to better capture subtleties in signal topology and detector noise modeling.