Physics-Informed Neural Networks Advance Pulsar Magnetosphere Modeling with Improved Accuracy and Speed

A preprint submitted to the Journal of Computational Physics introduces PulsarX, an open-source framework that applies Kolmogorov-Arnold networks (KANs) to the problem of modeling axisymmetric pulsar magnetospheres using physics-informed machine learning. Prior PINN-based approaches required extensive manual hyperparameter tuning, achieved limited accuracy, and demanded several hours of training; the new framework replaces these with domain-specific neural architectures, an automated adaptive training pipeline, and a physics-based convergence criterion. The result is mean squared errors of PDE residuals at O(1e-6) in double precision — a two-orders-of-magnitude improvement — while converging in under 20 minutes in single precision. The method also reliably resolves stellar radii reduced by up to 80% relative to the baseline, overcoming severe spatial scale disparities that challenge both neural and traditional solvers. Additionally, the authors provide a correction to the equation relating the open magnetic flux to the position of the equatorial T-point, a physically important relationship in pulsar magnetosphere theory. The work sits at the intersection of computational astrophysics and machine learning, and the full codebase has been made publicly available. As a preprint, the findings have not yet undergone formal peer review.