High-fidelity surrogate modeling of high-harmonic generation using Fourier neural operator with high-frequency weighted loss

High-harmonic generation (HHG) is central to ultrafast nonlinear optics and attosecond science, yet conventional time-dependent Schrödinger equation simulations are computationally costly and poorly scalable under multi-parameter conditions, limiting HHG optimization. We introduce Weighted-PhysFNO (W-PhysFNO), a physics-informed surrogate model that combines Fourier neural operators with a high-frequency weighted loss to learn a nonlinear operator mapping from physical parameters (molecular orientation(θ), internuclear distance(R), laser pulse intensity(I)) to dipole acceleration, evaluated on symmetric diatomic molecules as a case. By emphasizing weak yet physically important high-order harmonics, W-PhysFNO accurately captures harmonic features and cutoffs. Numerical results show that, compared with conventional time-domain solvers, W-PhysFNO maintains high predictive accuracy while substantially reducing computational cost. Compared with Convolutional Neural Network (CNN)-based surrogates, it achieves superior accuracy and small-sample generalization, enabling rapid spectral prediction for unseen parameters and efficient large-scale parameter exploration.