Joint Subarray Selection and Precoding Design for mmWave MIMO Systems: Deep Unfolding With GCN Approach

Millimeter-wave (mmWave) communication has been identified as a key enabler for future wireless systems due to its potential to offer high spectral efficiency with large antenna arrays. However, the high hardware cost and power consumption associated with the full digital architecture pose challenges in practical deployment. Hybrid precoding with antenna or subarray selection (AS/SS) offers a cost-effective solution but introduces challenging non-convex optimization problems. Traditional iterative algorithms are often computationally intensive and difficult to scale, while data-driven deep learning (DL) methods suffer from poor interpretability and limited generalization. To address these limitations, we propose a model-driven deep unfolding framework for joint precoding and AS/SS in mmWave systems. By using the equivalence between sum-rate maximization and mean square error (MSE) minimization, we unfold the weighted minimum MSE (WMMSE) algorithm into a trainable network with embedded sparsity modules. For AS, we incorporate a sparse mapping function with gradient-based matrix inversion to reduce complexity. For SS, a graph convolutional network (GCN) is integrated to exploit the spatial correlation among subarrays. Simulation results show that the proposed methods achieve near-optimal sum-rate with significantly lower complexity and good generalization, outperforming conventional optimization and learning approaches.