Neutron reconstruction using machine learning
The detection of high-energy neutrons requires large detectors with high granularity. In the present case, the NeuLAND detector will be composed of 3000 plastic scintillating bars that provide an active detector volume of 2.5 x 2.5 x 3 m^3. Neutrons typically interact more than once inside the detector, which makes it very difficult to identify how many neutrons hit the detector at the same time. Hence individual interactions have to be assigned to the different neutrons that create them. Machine-leaning software packages, such as tensorflow, are a possibility to improve existing reconstruction algorithms. Here the different interactions by one or several neutrons, and the corresponding signals they create within the detector, can be seen as a 3-dimensional image. Using simulated data as a training data set, tensorflow should be able to improve our ability to identify neutrons in experimental data. This project, which is part of a master thesis, attempts to test this approach with neural networks of growing complexity in a systematic way. As tensorflow works best when running on GPUs, a machine that provides those is needed.