Beamforming generation via conditional variational autoencoder and adversarial learning

Abstract To address the limitations of current array antenna beamforming methods in terms of prediction accuracy and generalization capability, this paper proposes an efficient deep learning-based beamforming approach. The method first employs high-precision numerical computations in conjunction with an improved genetic algorithm to optimize the design and generate datasets that closely resemble practical scenarios, thereby providing a solid foundation for neural network training. On this basis, a generative model architecture combining a conditional variational autoencoder and a generative adversarial network is designed to enable accurate and efficient beamforming prediction. Experimental results demonstrate that the proposed method achieves superior performance in planar array scenarios, with improved beam control accuracy compared to typical deep learning models, while significantly reducing computational complexity relative to conventional optimization algorithms.