Accurate channel estimation in extra-large multiple-input multiple-output (XL-MIMO) systems is fundamentally challenged by hybrid-field propagation, where the coexistence of near-field and far-field components induces spatially nonstationary and g...
Accurate channel estimation in extra-large multiple-input multiple-output (XL-MIMO) systems is fundamentally challenged by hybrid-field propagation, where the coexistence of near-field and far-field components induces spatially nonstationary and geometry-dependent channel characteristics across the antenna array. Conventional model-based estimators rely on restrictive assumptions that become ineffective under such conditions, while existing learning-based approaches typically employ generic tokenization schemes that fail to preserve antenna-domain structure, leading to a representation-structure mismatch. This thesis formulates hybrid-field channel estimation as a geometry-consistent representation learning problem, in which the input representation is explicitly aligned with antenna-domain structure. To address this problem, a MIMO-aware patching Transformer is proposed, where received pilot observations are partitioned into receive-subarray and transmit-group tokens, augmented with dual-domain spatial-spectral encoding, and processed through Transformer blocks to capture dependencies across geometrically related antenna-domain regions. The proposed framework is evaluated under hybrid-field XL-MIMO settings across varying signal-to-noise ratios (SNR). Controlled evaluations are conducted under fixed model capacity and dataset size to ensure fair benchmarking against classical and learning-based baseline methods, while additional analysis investigate scalability with increasing model capacity and training data. Simulation results demonstrate that the proposed framework consistently achieves lower normalized mean square error (NMSE) and higher achievable rate compared to the considered baseline methods, while maintaining stable performance across different operating conditions. The results further indicate that the effectiveness of attention-based channel estimation is strongly influenced by representation design rather than architectural complexity alone. In particular, preserving antenna-domain structure during tokenization enables more effective modeling of spatially nonstationary channel dependencies, highlighting the importance of geometry-consistent representations for hybrid-field XL-MIMO systems.