PG-ARUNet: A Physics-Guided Attention Residual Network for Arctic Sea Ice Concentration Prediction

Accurate prediction of Arctic sea ice concentration (SIC) is critical for understanding climate change and ensuring safe Arctic shipping. However, conventional U-Net architectures often degrade deep features and lack explicit physical constraints. To address this, we propose a physics-guided U-Net with attention residual connections, in which physical forcing steers feature learning at the architecture level rather than entering solely as loss-level constraints. An attention residual mechanism enables adaptive multi-scale feature fusion across layers, while a physics-constrained module is embedded throughout the encoder–decoder to enforce physical consistency. Thermodynamic and dynamic variables (sea surface temperature, 2 m air temperature, and 10 m wind components) are fused via cross-modal attention and physics-guided skip connections. A dual-input framework captures local fine-scale ice edge details and large-scale Arctic background fields simultaneously. The composite loss combines a weighted Charbonnier term and a freezing loss to better penalize ice edge mismatches and enforce freeze/melt physics. Experiments using National Snow and Ice Data Center (NSIDC) and ERA5 data demonstrate that our model outperforms six representative deep learning baselines at the 1-month lead under a unified multi-seed protocol, with all pairwise improvements statistically significant, and retains substantial skill at a 6-month lead, accurately reproducing the September minimum, with ice edges closely tracking observations. These results confirm that the physics-guided attention residual U-Net enhances spatial fidelity and physical consistency of subseasonal-to-seasonal SIC predictions, thereby supporting robust, medium-to-long-term assessment of Arctic navigability windows.