Context-aware deep learning for robust channel extrapolation in fluid antenna systems

Fluid antenna systems (FAS) offer remarkable spatial flexibility but face significant challenges in acquiring high-resolution channel state information (CSI), leading to considerable overhead. To address this issue, leveraging the inherent spatial continuity of electromagnetic fields, we propose CANet, a robust deep learning model for channel extrapolation in FAS. CANet combines context-adaptive modeling with a cross-scale attention mechanism and is built on a ConvNeXt v2 backbone to improve extrapolation accuracy for unobserved antenna ports. To further enhance robustness, we introduce a spatial amplitude perturbation strategy tailored to FAS channel extrapolation. Based on this perturbation, a Fourier-domain loss is incorporated to preserve spectral structure in the Fourier domain and improve learning stability under perturbations. Our simulation results demonstrate that CANet consistently outperforms benchmark methods without noise-variance side information in terms of estimation accuracy across a wide range of signal-to-noise ratio (SNR) levels and carrier frequencies, while also significantly reducing the outage probability with moderate computational complexity.