Multilayer perceptron approach for determining the centroid position of the electromagnetic shower in laterally segmented calorimeter

Abstract This paper presents a method employing a Multi-Layer Perceptron (MLP) neural network to enhance the spatial resolution of laterally segmented electromagnetic calorimeters. The study, based on GEANT4 simulations, explores the impact of different MLP architectures and hyperparameters on spatial resolution across a range of calorimeter block sizes and incident photon energies from 1 GeV to 10 GeV. Optimal configurations are identified, including the number of input neurons, hidden layers, and neurons per layer. The optimized MLP model consistently outperforms the widely used center of gravity method, achieving a significant improvement in spatial resolution that ranges from 8% for 1 GeV photons and a block size of half the Molière radius to 22% for 10 GeV photons and a block size of twice the Molière radius.