Deep Unrolling for Near-Field Channel Estimation: Challenges and Opportunities

As 5G enters full commercial operation, 6G research has accelerated, focusing on terahertz (THz) frequency bands to meet the escalating demands for transmission rates. Nevertheless, the use of higher frequency bands presents notable challenges, such as expanded near-field propagation regions and intensified path loss effects. Large-scale MIMO (Multiple-Input Multiple-Output) technology has emerged as a viable solution by increasing the number of antennas, yet it complicates traditional channel estimation and feedback mechanisms. Moreover, wideband THz communication systems face unique challenges from both near-field and far-field propagation, in which conventional far-field assumptions become invalid due to spherical wavefront effects and spatial non-stationarity. In this work, we propose a novel channel estimation framework for wideband THz large-scale MIMO systems based on deep unfolding. Specifically, we improve the compressed sensing-based Uniformly-Accelerated Message Passing-Sparse Bayesian Learning (UAMP-SBL) algorithm by substituting its conventional M-step with a trainable Multilayer Perceptron Mixer (MLP-Mixer) neural network module. This deep unrolling strategy effectively addresses convergence issues inherent in the original algorithm while preserving the computational efficiency of the E-step. By embedding structural priors that reflect sparsity and geometric channel features, the proposed model jointly optimizes spatial and sparsity-related characteristics via a composite loss function. Comparative simulations demonstrate that the UAMP-SBL unfolded with MLP-Mixer significantly outperforms convolutional neural networks (CNN) based counterparts in estimation accuracy and convergence speed under comparable training conditions, while maintaining scalability for large antenna arrays. The results highlight deep unrolling as a powerful, interpretable, and adaptable paradigm well-suited for next-generation near-field 6G communication systems.