Key points are not available for this paper at this time.
The rapid growth of artificial intelligence (AI) has placed stringent demands on fiber-optic backbone networks supporting data transmission for intelligent computing centers. In optical fiber system design, achieving channel simulations that are both highly accurate and efficient is crucial. Building on the separable operator network framework, we introduce the Kolmogorov-Arnold Network (KAN), which outperforms traditional multi-layer perceptron (MLP) in modeling physical equations. Leveraging this advantage, we construct what we believe to be a novel separable operator KAN (SepOKAN) architecture. Furthermore, by incorporating physical constraints into SepOKAN, we develop a separable physics-informed KAN (SPIKAN) architecture that offers greater interpretability and improved simulation accuracy. We validate the proposed models in a stepwise manner: first in a single-channel dual-polarization scenario for fundamental verification, and then in an ultra-wideband (UWB) multi-channel scenario. We evaluate fitting accuracy and generalization performance through multi-dimensional comparisons. Finally, experiments on real-world single-channel and wavelength-division multiplexing (WDM) fiber link systems demonstrate that the SPIKAN model achieves high accuracy and robust performance, highlighting its practical value for optical fiber communication system design.
Yang et al. (Tue,) studied this question.