arXiv2026
Neutrino oscillations provide crucial insights into fundamental particle physics, with two-flavor approximations offering a robust framework for analyzing reactor and atmospheric phenomena. Currently, standard numerical integrators represent the gold standard for solving the associated evolution equations, achieving machine precision at negligible computational cost. However, Physics-Informed Neural Networks (PINNs) have recently emerged as a powerful surrogate model for handling high-dimensional systems, complex geometries, and inverse parameter estimation. In this paper, we establish a foundational proof-of-concept by applying PINNs to solve the coupled ordinary differential equations governing neutrino evolution in vacuum and constant-density matter profiles. We also discuss the theoretical framework encompassing vacuum mixing and the matter enhanced neutrino oscillations and rigorously benchmark the PINNs against standard NuFast and RK4 (Runge-Kutta 4) solutions. By carefully optimizing our network architecture, we demonstrate that PINNs can achieve high-precision agreement with experimental standards for reactor (low-energy) and atmospheric (high-energy) neutrinos, which we use as our case study. This paper evaluates the robustness of the PINNs approach through a comprehensive error analysis, establishing a baseline validation for its application to more challenging problems in neutrino physics, such as variable density profiles.