Deep Learning-Based Channel Estimation for Stacked Intelligent Metasurface-Enhanced Multi-User Communications

Stacked intelligent metasurface (SIM) has emerged as a promising technique to implement holographic massive multiple-input multiple-output (MIMO) by incorporating multiple metasurfaces into the antenna transceivers with limited RF chains. Channel estimation is challenging in SIM-enhanced communication systems, as the dimension of the channel matrix to be estimated is much larger than the number of RF chains. Although the conventional minimum mean square error (MMSE) estimator performs well, it requires prior knowledge of channel statistics and frequent updates of phase shifts of the meta-atoms. This paper proposes a new deep learning-based channel estimator, termed as flattened channel estimation network (FlatCE-Net). Specifically, the proposed FlatCE-Net refines the coarse least squares (LS) channel estimation by introducing a flattening strategy that transforms the input into sequential data. This allows the use of deeper and wider dilated convolutions to extract richer features, which in turn enhances the denoising of the initial LS estimation and improves the channel estimation accuracy. Moreover, it requires no prior channel knowledge and the SIM only needs to be configured once. The simulation results demonstrate that FlatCE-Net outperforms the existing deep learning-based schemes.