Abstract : All physics analyses at CMS rely on precise reconstruction of physics objects in the experiment’s calorimeters. To achieve the best performance, this reconstruction is performed by a mixture of rule-based algorithms and multivariate regressions, both of which typically rely on human-engineered high-level features to describe the event. However, recent advances in machine learning technology have enabled next-generation approaches to these tasks wherein more powerful ML architectures are trained directly on low-level detector signals to obtain substantially improved physics performance. One particularly promising avenue is in the use of graph architectures for reconstruction of particle properties from the pattern of energy deposits across the CMS calorimeters. This talk will discuss how this new approach has been applied to both hit clustering and energy correction of electromagnetic objects in the CMS crystal ECAL, where we obtain improved stability and energy resolutions with respect to the previous state-of-the-art approaches. Additionally, we will present recent efforts to develop similar approaches for future LHC running, including reconstruction in the planned CMS HGCAL.
Sunrise at FNAL (WH11NW)