Abstract: To address the communication demands in complex scenarios such as satellite-terrestrial collaborative communication, low-altitude economy, and smart cities, this study proposes a convolutional neural network (CNN) for fast beamforming with coding metasurface arrays, enabling a direct inverse design process. The CNN is primarily based on the VGG network architecture, incorporating channel attention mechanisms to further enhance its prediction accuracy. Dataset is collected through phase compensation and multi-population genetic algorithm (MPGA). The trained network can generate coding matrices for both single-beam and dual-beam within milliseconds, demonstrating improvement in both training convergence speed and prediction accuracy compared to existing benchmark networks.This network significantly improves the design efficiency of large-scale coding metasurface arrays, offering a feasible solution for fast and real-time control of coding metasurface arrays.