Applying the DUNE convolutional neural network to ICARUS liquid argon TPC data

The description of neutrinos that interact (flavor states) is different from the description of neutrinos that traverse time and space (mass states). To study these properties, we need to understand how many neutrinos go through our detectors, and how those neutrinos interact within them. ICARUS is a 430t liquid-argon neutrino detector located at Fermi National Accelerator Laboratory and serves as the far detector for the Short Baseline Neutrino program. It poses a unique opportunity to measure a variety of electron and muon (anti-)neutrino interaction rates with Ar nuclei, which will be an important input to the Deep Underground Neutrino Experiment (DUNE).

Computer scientists working in collaboration with High-Energy physicists developed a Convolutional Neural Network (CVN) for DUNE intending to achieve highly efficient and pure selections of electron and muon neutrino Charged-Current (CC) interactions. The network should provide considerable gains in time saved and overall identification accuracy. We believe the network code is flexible enough to be adapted to work with ICARUS datasets after minor changes. We trained a modified CVN in a variety of test scenarios with smaller data sets and shorter training periods. The next step will be full training using high-performance computing resources. I will describe the CVN inputs and network architecture, compare results between the published CVN results and our test runs, and present plans for our upcoming full training run.