Machine Learning Neutrino-Nucleus Cross Sections II: More Leptons

Abstract

Neutrino-nucleus cross sections are essential inputs for long-baseline neutrino oscillation experiments, yet remain challenging to model with high precision. Recent work demonstrated that fully black-box data-driven cross-section models cannot accurately generalize between near- and far-detector fluxes, but imposing Standard-Model symmetries via structure-function parameterizations provides a well-posed learning problem in a simplified DUNE-inspired muon-neutrino disappearance analysis. Here, we extend this framework to multiple lepton flavors and both neutrino and antineutrino beam modes, incorporating the electron appearance channel and enabling a simultaneous determination of $\sin^2θ_{23}$, $Δm^2_{31}$, and the CP-violating phase $δ_{\rm CP}$. We find that the learned cross section model recovers the injected oscillation parameters without significant bias, with constraints comparable to those obtained using the exact underlying cross section. Flux uncertainties have more significant effects in the multi-channel case, and incorporating flux nuisance parameters into the learning problem is required for accurate results. Our results demonstrate that this data-driven approach can be extended to the multi-channel measurements needed for long-baseline neutrino oscillation analyses.