Efficient Physics-Informed Neural Network for Ultrashort Pulse Dynamics in Optical Fibers

Simulating the propagation of ultrashort pulses in optical fibers is vital for photonic technologies such as laser design, high-speed telecommunications, and high-resolution imaging. The conventional approach using the nonlinear Schrödinger equation (NLSE) is time-intensive and complex, creating a hurdle for real-time experimental design and pulse optimization. While recurrent neural networks (RNNs) have been explored to mitigate these issues, they often require extensive NLSE simulations for training, presenting challenges related to time and cost. To overcome these limitations, we propose a physics-informed neural network (PINN) that efficiently captures ultrashort pulse dynamics, reducing the computational burden and the need for extensive training data. We examine the model's applicability for initial pulse widths above and below 1 ps in optical fibers, evaluating its prediction accuracy, training duration, and speed of prediction. Our findings demonstrate that PINN offers a precise and efficient solution for predicting intricate pulse behaviors. With its adaptability to various input conditions and high predictive accuracy even with limited training data, PINN shows great promise for widespread use in experimental settings.