Development of a Deep Neural Network for the data analysis of the NeuLAND neutron detector

Abstract

A new Machine Learning algorithm for shower-head identification in the NeuLAND neutron detector is presented. The new algorithm uses densely-connected Deep Neural Networks (DNNs) to properly classify events and clusters, which allows accurate reconstruction of the 4-momenta of the detected neutrons. As data-events recorded with NeuLAND vary quite a lot in size, and not all emitted neutrons always produce signals in the detector, careful pre- and post-processing of the data turned out to be required for letting the DNNs be successful in their classifications. However, after properly implementing these procedures, the new algorithm offers a better efficiency than previously-used algorithms in virtually all investigated scenarios. However, the newly-developed algorithm (as well as previous ones) suffers from systematic uncertainties. These uncertainties mainly arise from the physics lists used in the Geant4 simulations to train the DNNs. They are particularly large for the neutron energy range around 200 MeV and for NeuLAND configurations of few double-planes (slimmed down version of the detector). The accuracy improves with a larger number of double-planes. Furthermore, both model improvements and accurate benchmarks are needed for the currently used Geant4 physics lists to reduce the systematic uncertainties of the new algorithm for high-precision studies. Further improvement of the present DNN algorithm is also needed, especially for experiments that require high precision in the neutron scattering angle reconstruction. However, it seems unlikely that this improvement can be realized using only NeuLAND data.