Efficient prediction of vector optical soliton propagation in birefringent fibers via an enhanced Fourier neural operator.

Efficient prediction of vector soliton propagation is necessary for parameter sweeps and system design in birefringent optical fibers. However, repeated simulations with the split-step Fourier method are computationally expensive. Neural operators have been widely used to accelerate simulations of nonlinear optical dynamics, but existing architectures often emphasize either global spectral modeling or local multiscale reconstruction. Here, we develop a surrogate model based on the U-Net-enhanced Fourier neural operator (U-FNO) architecture for efficiently predicting vector soliton propagation. The model embeds lightweight U-Net branches into selected Fourier layers. We compare it with a convolutional neural network, U-Net, DeepONet, and FNO on a simulation benchmark. U-FNO substantially improves upon the FNO backbone. The power-weighted absolute percentage error, power mean absolute percentage error, and terminal power root-mean-square error (RMSE) decrease by 28%, 32%, and 29%, respectively. Among the evaluated models, U-FNO achieves the lowest field mean absolute error and terminal power RMSE, and it is substantially more robust to input noise and to unseen temporal grid resolutions than the other high-accuracy models. Per-sample diagnostics show that larger gains tend to occur for more concentrated and sharper target fields. These results demonstrate the potential of U-FNO for efficient vector soliton propagation prediction in birefringent optical fibers.