Transformer-Hypernetwork-Controlled Deep-Unfolded Phase-Aware Channel Estimation Refinement for Phase-Drift-Robust Backscatter Links

This paper proposes a transformer-hypernetwork-controlled deep-unfolded phase-aware channel estimation refinement (THUNDER) for phase-drifting backscatter links. Residual carrier-phase drift across the pilot block renders the backscattered observation phase-nonstationary, and a closed-form phase-aware channel estimation (PACE) compensates only the first-order phase component, leaving a deterministic high signal-to-noise ratio (SNR) error floor. THUNDER suppresses this floor by initializing from PACE and refining the estimate through unfolded Gauss-Newton steps on the exact phase-exponential model. A transformer extracts pilot-wide phase context, and a hypernetwork generates bounded controls and pilot-reliability weights. Evaluations show an 8.9 dB normalized mean square error gain over the strongest learning-based channel estimation baseline.