I developed a convolutional neural network (CNN) to identify different particle types in ProtoDUNE, including protons. I used test beam data and Monte Carlo (MC) simulation thereof to evaluate and train the network respectively. There were found to be differences between the data and Monte Carlo simulations. This is due to systematic differences in how we simulate the liquid argon detector and the actual microphysics of the detector. I extended this work to mitigate for these uncertainties by using a Domain Adversarial Neural Network (DANN). The concept of the DANN is to simultaneously learn the problem (classification of particles) and unlearn any domain differences (between data and MC). The classification was trained on the MC as before, and the domain differences were unlearned using images labelled as data or MC, but without class labels. This means the data class labels (from the beamline instrumentation) were still unseen during training and could be used for evaluation of the performance of the network. Now the performance not only matched between data and MC, but was significantly improved for proton identification. Protons from the ProtoDUNE test beam data were then used to perform a measurement of electron-ion recombination in liquid argon. This is a vital measurement for the DUNE experiment as it will be used to convert observed charge in the detector to an energy deposit, needed to deduce the neutrino’s energy and be able to measure a matter-antimatter asymmetry in neutrino oscillations.
I helped to install the new readout chamber for the High Pressure gaseous Time Project Chamber (HPgTPC) for DUNE and made preparations for its beam test which will eventually take place after the end of this fellowship. In addition, I participated in the preparation for the liquid argon near detector prototype neutrino beam test run. I participated in the reinstallation of the MINERvA planes which will for the muon range detector for the test run. I was responsible for getting the nearline data monitoring system working, which was successfully completed. The nearline monitoring first needed to unpack the raw data files into a standardised data format that could be used for analysis. The monitoring then checks that the voltages read out are close to those set, and that there are no significant time variations or light leaks. I also helped to install the photo-multiplier tubes and cable up the detector.
I used the neutrino interaction generator, GENIE 3.2 to simulate neutrino interactions with argon and other elements for reference. GENIE 3.2 has a correlated Fermi gas model that tries to account for short-range correlations (SRCs) between pairs of nucleons. These nucleons in SRC pairs then occupy high momentum states. For heavier nuclei, fractionally more protons should be in high momentum states compared to neutrons, so a bigger modification is observed for lead than argon. To see if GENIE 3.2 was reproducing the correct behaviour, I compared the struck nucleon momenta distributions to electron scattering data from CLAS. From this work it has become clear that the correct behaviour is not being produced for heavier nuclei, in fact GENIE is simulating a higher fraction of neutrons in high momentum vs low momentum states than for protons. This means the predictions for the modification to neutrino/antineutrino interaction cross-sections are likely to be incorrect in this simulation.