Convoluted Events Neutron Reconstruction using Neural Networks

Abstract

The R3B experiment at FAIR will study properties of unstable nuclei through detection of reaction products of projectile-target interactions. It is essential that these reaction products can be measured with sufficient accuracy. Uncharged neutrons do not excite the scintillator material in the detector, and therefore can only be detected indirectly via charged products from neutron-nucleus interactions. These interactions can create multiple new particles including neutrons, in turn, interacting with other nuclei. Reconstructing the multiplicity and momenta of the neutrons entering the detector from these shower patterns is not trivial and requires sophisticated algorithms. This thesis explores the possibility of reconstructing neutron events with 3-dimensional image recognition using Convolutional Neural Networks, focusing mainly on neutron multiplicity. When a passing charged particle excites the scintillator material in the detector, it outputs the spatial coordinates of the excitation point along with the time and the energy the particle has deposited. The output can be converted into a sparse 3-dimensional image with time and deposited energy as pixel values. The 300 000 pixel values in each image make the required amount of parameters, even in the smaller networks, very large. It is shown that training these large but simple networks using a Central Processing Unit is not practically feasible, requiring months to train a single network. The use of a Graphics Processing Unit introduced a speed up in training with a factor of up to 185. By accounting only for the total deposited energy and number of hits in the detector, 72 % correct predictions were achieved on a large test set. Accounting also for the image of each event, an accuracy of 78 % correct predictions was achieved, showing that the networks are able to extract important features from the images.