Pilot-efficient channel estimation for 6G MU-mMIMO using LS initialization and GAN-augmented CNN refinement

Accurate channel estimation with limited pilot overhead is a key bottleneck for 6G MU-mMIMO, particularly at mmWave (with implications for THz) where path loss and multiuser interference are severe. We propose a two-stage least squares convolutional neural network (LS-CNN) framework in which a robust LS estimator provides a stable coarse channel estimate, avoiding the sparsity assumptions of compressed sensing algorithms that often fail in rich-scattering outdoor environments. In the second stage, this estimate is refined using a CNN trained on a hybrid dataset expanded through controlled augmentation of a generative adversarial network (GAN), improving generalization while minimizing pilot overhead. Training data is refined using mean squared error (MSE) thresholds to ensure high-quality samples, and a robust standardization method is used to stabilize learning with complex-valued input. Simulations on the DeepMIMO ray-tracing dataset demonstrate that the proposed approach provides an up to 8 dB normalized MSE (NMSE) improvement at low signal-to-noise ratio (SNR) over classical LS and MMSE estimators. In our experiments, it achieves performance comparable to that of the ideal MMSE estimator while requiring 60% fewer pilot symbols. Complexity analysis confirms that the proposed hybrid architecture is significantly more efficient than state-of-the-art Vision Transformers, making it practical for real-time 6G deployments.