Few-shot channel estimation and adaptive precoding in mmWave MIMO using metasparse-DUNet and conditional GANs

Millimeter-wave (mmWave) massive MIMO systems offer high data rates and spectral efficiency but require accurate channel estimation and hardware-compliant hybrid precoding under dynamic conditions. Traditional compressed sensing and model-driven approaches often struggle with adaptability and robustness, while existing deep learning solutions lack generalization across varying propagation environments and channel sparsity levels. To address these limitations, this paper proposes a novel two-stage learning-driven architecture that combines a meta-learning-enhanced Deep Unfolding Network (MetaSparse-DUNet) for sparse mmWave channel estimation with an adaptive generative adversarial network (GAN)-based hybrid precoding mechanism for context-aware beamforming. MetaSparse-DUNet uses model-agnostic meta-learning (MAML) to enable fast adaptation to new channel conditions with limited data. In the second stage, a task-conditioned GAN generates hybrid precoders under hardware constraints, optimized via spectral-efficiency-aware loss and constraint projection. Extensive simulations demonstrate that the proposed method outperforms conventional techniques by achieving up to 9 dB lower NMSE, 5.6 bps/Hz higher spectral efficiency, and lower BER under various antenna configurations and mobility scenarios. The proposed framework provides a scalable and hardware-efficient solution for mmWave MIMO systems.