AI based reconstruction for hybrid electromagnetic calorimeters

In this study, several AI methods were applied for the reconstruction of electromagnetic calorimeters, including hybrid calorimeters. A hybrid calorimeter combines multiple types of scintillating materials, such as lead-glass and crystals, to capitalize on the individual strengths of each material optimizing the overall performance and cost-effectiveness. Such calorimeters are used at present and upcoming Jefferson Lab experiments and were proposed for Electron Ion Collider (EIC) detectors. Reconstruction of clusters in hybrid calorimeters presents additional challenges due to the varying response characteristics of different scintillating materials, more complex geometry, and intricate transitional regions where different materials meet. Several AI approaches were developed and evaluated to aid in addressing these issues, enhancing the performance relative to traditional methods that were compared as a baseline. In particular, variational autoencoders have been utilized to identify and segregate merged clusters. Deep learning architectures, including convolutional neural networks, ResNets, and graph neural networks, were trained to interpret the data, thereby improving the accuracy of cluster reconstruction, noise reduction, and overall resolution. Geant4 and DD4hep simulations, based on setups proposed for EIC and Jefferson Lab experiments, provided the data for training and validation. The effectiveness of the algorithms, demonstrating transfer learning, was validated by their application to experimental data — initially trained on simulations, they successfully adapted to real data from beam tests conducted at Jefferson Lab's HallD. These algorithms are not limited to hybrid electromagnetic calorimeters, indicating potential for broader application across various calorimeter designs. Additionally, it has been demonstrated that these algorithms perform efficiently on a variety of computing platforms, specifically CPUs and GPUs, with certain algorithms further adapted and proven to function effectively on FPGAs