CRFCAN: A Complex-Valued Cross-Domain Residual Network for Joint Channel and Phase Noise Estimation in Sub-THz OFDM Systems
The coupling between multipath propagation and phase noise (PN) poses challenges for joint channel and PN estimation in sub-terahertz (sub-THz) cyclic-prefix OFDM systems. This paper proposes CRFCAN, a complex-valued residual FFT convolutional attention network for single-input single-output systems using comb-type pilots only. Its shared backbone alternately refines time- and frequency-domain features through complex-valued residual channel attention and PN-inspired multiplicative updates, producing channel and PN estimates in their respective domains. An unfolded PN refinement tail combines least-squares updates with feature-guided complex-valued CNN corrections. Simulation results show that CRFCAN achieves lower channel estimation error and bit error rate than the evaluated baselines in both training TDL-A/B/E + 3GPP RAN4 and TDL-C/D + 3GPP RAN1 generalization settings, together with lower PN estimation error at medium-to-high signal-to-noise ratios. Further tests without fine-tuning demonstrate its applicability across different data modulations, OFDM parameters, channel models, and PN spectra.