Second-scale atmospheric turbulence predicting with the Kolmogorov–Arnold network

Free-space optical communications are susceptible to atmospheric turbulence. Conducting high-temporal-resolution atmospheric turbulence prediction can provide strategic support for the optimization and improvement of free-space optical communication systems, especially satellite-to-ground laser communication systems. This paper presents a new, to our knowledge, artificial neural network which is known as the Kolmogorov–Arnold network (KAN), for the prediction of atmospheric turbulence with a time resolution of 1 s using meteorological parameters. Based on the Kolmogorov–Arnold theorem, KAN establishes a nonlinear mapping relationship between meteorological parameters and the atmospheric turbulence strength. Compared with traditional neural networks, KAN is more effective at capturing complex correlations among meteorological parameters by introducing learnable nonlinear basis functions. Under the same meteorological conditions, when using the KANs to predict the atmospheric refractive-index structure constant ( C n 2 ) at a second-level time scale, the resulting mean absolute percentage error (MAPE) and symmetric mean absolute percentage error (SMAPE) are 31.54% and 31.44%, respectively. These two metrics are reduced by 6.74% and 4.35% compared to the traditional multi-layer perceptron (MLP) architecture.