An IP2I team evaluated the effectiveness of different neural network architectures in discriminating between neutron and gamma radiation signals on the NEDA detector.

NEDA (NEutron Detector Array) is a neutron detector designed for use with high-resolution gamma-ray spectrometers, such as AGATA (Advanced GAmma Tracking Array). The scintillating liquid used in NEDA is not only sensitive to neutrons, but also to gamma radiation: it is therefore imperative to differentiate between neutron and gamma interactions.

The IP2I team used different neural network architectures (densely connected, convolutional and recursive) to analyze and classify signals generated by each type of particle during the AGATA/NEDA/DIAMANT campaign at GANIL. Thanks to the coincidence information obtained on AGATA, they were able to measure the performance of each architecture in real situation.
For the simplest networks, the discrimination could be done online during the data collection, with a processing time of the order of a microsecond per signal. For the others, the results were obtained offline on the CCIN2P3 GPU farm.

This work has also allowed them to study the possibilities of neural networks in signal processing: compression, decorrelation of stacked signals (see illustration) or reconstruction of incomplete signals for example.

Their results are the subject of an article published in Nuclear Instruments and Methods in Physics Research section A.