Domain Generalization with Machine Learning in the NOvA Experiment

NOvA is a long-baseline neutrino oscillation experiment. By observing the disappearance of muon-type neutrinos and the appearance of electron-type neutrinos NOvA can probe outstanding questions in neutrino physics including the mass ordering and the existence of leptonic CP violation. The usage of deep neural networks (DNNs) is becoming widespread in particle physics and beyond. In many cases, and in particle physics in particular, DNNs must be trained on simulated data because real data is reserved for analysis. Any simulation is inherently an imperfect representation of the real physical processes that these networks are meant to target. Here, I present the application of an adversarial technique to generalize DNNs to improve their applicability to the inaccessible target data in the context of estimating the energy of muon-type neutrino interactions in the NOvA experiment.