Deep learning based channel estimation for THz band ultra massive MIMO systems

Official URL: https://dx.doi.org/10.1109/SIU71813.2026.11637088

THz ultra-massive MIMO is a key technology for 6G systems, but hybrid near- and far-field propagation, limited RF chains, and severe path loss make channel estimation challenging. Model-driven deep learning approaches, such as FPN-OAMP, address this problem by combining a physics-driven linear estimator with a learnable nonlinear estimator (NLE) within a contractive fixed-point framework with convergence guarantees. In this paper, we redesign the NLE of the FPN-OAMP framework through three deep-learning schemes: Differentiable Architecture Search (DARTS), DARTS with Block Recurrent Transformer (BRT) and Vision Transformer (ViT). DARTS optimizes multiscale convolutional operations, BRT provides iterative global refinement, and ViT enables attention-driven spatial modeling in the FPN-OAMP framework. Our simulations show consistent performance improvements for FPN-OAMP across a wide range of signal-to-noise ratio levels, while retaining similar asymptotic complexity.