The R3B setup is an experimental setup used to study nuclear structure properties of short-lived isotopes. It is located at the high-energy branch of the Super Fragment Separator (SuperFRS) in Germany. The setup consists of a fixed target where a secondary beam generated in the Super-FRS is directed. The goal of the R3B experiment is to reconstruct all particles participating in the reaction in order to study the nuclear structure of the beam isotope. To accomplish this, the fixed target is surrounded by different detector systems, including silicon strip vertex detectors and the CALIFA system for measuring gamma-rays and light-charged particles. Heavier charged particles are analyzed using the superconducting dipole magnet GLAD, and a tracking system is used to measure their tracks. Neutrons produced at the target are detected by NeuLAND, a Time-of-Flight spectrometer. The problem addressed in this paper is the identification of the shower head, the first signal generated by a complex shower of secondary particles in NeuLAND. A Machine Learning algorithm is proposed to find the shower heads in NeuLAND. The methodology of the algorithm is discussed, as well as its performance and future possibilities for improvement. The paper concludes with a summary and conclusions. The text describes the implementation of a Deep Neural Network (DNN) algorithm for the classification of events in a particle detector called NeuLAND. The DNN is trained to determine the multiplicity (number of neutrons) and identify the shower heads in the detector. The DNN is equipped with input neurons that contain information about the energy deposition, time, and position of each scintillator hit in the detector. Additional inputs, such as total energy deposition and number of clusters per event, are also given to improve efficiency. The DNN is trained using supervised learning with a specific minimization algorithm and function. The network architecture and parameters are determined through optimization. The DNN algorithm is compared to two other algorithms, including one that relies on perfect tracking information from simulation data. The performance of the DNN algorithm is evaluated in terms of efficiency and compared to the other algorithms. The text also discusses the computational requirements and availability of the code for the DNN algorithm. The systematic uncertainties in the algorithm performance due to simulation data and physics list are also addressed. Overall, the DNN algorithm shows promising results for event classification in the NeuLAND detector.
