Real-time surrogate modeling of nonlinear pulse evolution in multimode fibers

Modeling nonlinear pulse propagation in multimode fibers is challenging due to the large number of interacting modes and the resulting spatiotemporal complexity. Traditional optimization methods often become intractable, while learning-based approaches, such as recurrent neural networks, suffer from high computational cost and long inference times. We present a U-Net architecture as a fast, accurate surrogate for modeling nonlinear pulse propagation in multimode fibers. This approach overcomes the intractability of traditional methods while offering low computational cost. Trained on data generated by beam propagation method, our approach achieves an $\sim$88% average structural similarity index with simulations. The model can generalize to untrained propagation distances, demonstrating convolutional architectures as efficient tools for simulating complex spatiotemporal dynamics in multimode fibers and offering potential for applications like mode decomposition.