The accurate prediction of water droplet collection coefficients is essential for icing analysis and the design of anti- and de-icing systems. Traditional high-fidelity numerical simulation methods, however, are often hindered by their computational complexity and time-intensive nature. Deep learning-based rapid prediction methods present a promising avenue to address these challenges. In this study, we propose a fast prediction approach that leverages proper orthogonal decomposition (POD) and Kolmogorov-Arnold networks (KAN) to accurately predict water droplet collection coefficients on three-dimensional spherical surfaces. Using POD, we extract its dominant intrinsic modes and corresponding fitting coefficients. A KAN-based deep learning model is then developed to map working condition parameters to the fitting coefficients. Experimental results demonstrate that the proposed POD-KAN model is well-suited for predicting water droplet collection coefficients on 3D spheres, delivering high accuracy with an average absolute error of 3.386 × 10⁻⁴. Moreover, after model training, the computational efficiency for obtaining the water droplet collection coefficients is improved by nearly 2.7 × 10⁵ times compared with traditional high-fidelity numerical simulations. This method provides efficient and reliable technical support for the rapid iterative optimization design of aircraft anti-icing/de-icing systems, and holds significant engineering application value for improving aviation flight safety.