Abstract: To address the low computational efficiency of traditional ray tracing methods in broadband, multi-frequency, and batch event analysis, this paper proposes a fast ray-tracing surrogate method for whistler waves based on physics-informed neural networks. This method uses the residual of the Haselgrove equation as the core physical constraint, combined with boundary conditions and propagation priors, to establish an end-to-end surrogate mapping model from initial propagation conditions to group delays and spatial propagation trajectories. The model was systematically validated under the background medium conditions constructed using the IRI-2020 ionospheric model and dipole magnetic field model. The results show that among 275 independent test samples, the root mean square error in group delay prediction is 1.44×10⁻² s, and the average absolute error in spatial propagation trajectory is 22.66 km. Using the traditional ray tracing method adopted in this paper as a reference baseline, the average single-sample computation time corresponding to 14 frequency points was reduced from 310.39 s to 0.0228 s. The geomagnetic colatitude extrapolation experiment shows that the model still has a certain extrapolation ability under the condition of no propagation, but the prediction error increases in the complex propagation environment of high geomagnetic colatitude. Analysis of measured events indicate that the predictions of this surrogate model align well with results from traditional ray tracing regarding group delay and propagation trajectories, and are capable of reproducing the typical dispersion patterns of observed whistler waves.