Artificial Intelligence for a Large Neutrino Experiment

The Deep Underground Neutrino Experiment (DUNE) is under construction. It will make use of a neutrino beam created at Fermilab and a far detector 1300 km away at the Sanford Underground Research Facility (SURF). The physics goals of DUNE are to study CP violation in long-baseline neutrino oscillations, neutrino astrophysics/supernovae and nucleon decay. The far detector will be composed of 40 kton of liquid Argon (LAr) time projection chambers (TPC). The particles created by such events including background from e.g. 39Ar will produce large quantities of data from particles created in interactions in the LArTPCs which travel through the detector. Both the reconstruction of events and the identification of particles pose exciting challenges for which artificial intelligence techniques are now under consideration. The PhD project would study event reconstruction, where neural networks can be used for pattern recognition. An additional challenge is that these algorithms have to be implemented in a time-efficient way such that they can be used as a trigger for data reduction within the DUNE DAQ, where the maximum drift time is 2 ms. The PhD project would also design an Online event classification for the trigger. Here the challenge is to discriminate muon tracks originating from the Fermilab neutrino beam from cosmogenic activities and supernovae or proton decay events from background which is mainly 39Ar. In particular, a robust online triggering mechanism is required in order to trigger on supernovae events. The image-like nature of the data produced from DUNE makes it an ideal target for deep learning approaches. Convolutional Neural Networks and Sparse Convolutional Networks will be studied for classifying supernova neutrinos from backgrounds enabling a High Level Trigger decision to readout the data in DAQ. While convolutional Neural Networks (CNN) have great effectiveness in image like data with spatial correlations, these are typically used in high-density information environments, which is different from DUNE raw data. This will make training and evaluation more computationally intense. Sparse CNNs will significantly reduce the computational requirements, making them an excellent candidate to be studied for online identification and classification in the large DUNE data set.