Adaptive TTD Configurations for Near-Field Communications: An Unsupervised Transformer Approach

True-time delayers (TTDs) are popular analog devices for facilitating near-field wideband beamforming subject to the spatial-wideband effect. In this paper, an adaptive TTD configuration is proposed for short-range TTDs. Compared to the existing TTD configurations, the proposed one can effectively combat the spatial-wideband effect for arbitrary user locations and array shapes with the aid of a switch network. A novel end-to-end deep neural network is proposed to optimize the hybrid beamforming with adaptive TTDs for maximizing spectral efficiency. First, based on the U-Net architecture, a near-field channel learning module (NFC-LM) is proposed for adaptive beamformer design through extracting the latent channel response features of various users across different frequencies. In the NFC-LM, an improved cross attention (CA) is introduced to further optimize beamformer design by enhancing the latent feature connection between near-field channel and different beamformers. Second, a switch multi-user transformer (S-MT) is proposed to adaptively control the connection between TTDs and phase shifters (PSs). In the S-MT, an improved multi-head attention, namely multi-user attention (MSA), is introduced to optimize the switch network by exploring the latent channel relations among various users. Third, a multi-feature cross attention (MCA) is introduced to simultaneously optimize the NFC-LM and S-MT by enhancing the latent feature correlation between beamformers and the switch network. Numerical simulation results show that 1) the proposed adaptive TTD configuration effectively eliminates the spatial-wideband effect under uniform linear array (ULA) and uniform circular array (UCA) architectures, and 2) the proposed deep neural network can provide near-optimal spectral efficiency, and solve the multi-user beamformer design and dynamical connection problem in real-time.