Hybrid AI Model Improves Arctic Sea Ice Prediction.

Accurate modeling and prediction of Arctic Sea Ice Concentration (SIC) are crucial for climate assessment and environmental forecasting. Traditional numerical models are often computationally intensive and require complex parameterizations, while purely data-driven approaches frequently fail to explicitly incorporate fundamental physical laws. The Physics-Informed Hybrid Ice Model (PIHIM) bridges this gap by integrating deep learning with explicit physical knowledge. Its network structure is organized according to the physical dependencies outlined in the sea ice continuity equation, directly accounting for dynamical transport, thermodynamic growth/loss, and unresolved local processes. This design allows PIHIM to retain the representation capacity of deep learning while providing a process-decomposed formulation of ice displacement, freeze-melt changes, and local error closure. Evaluations in both reanalysis-forced simulation and forecast-forced prediction settings demonstrate PIHIM's enhanced ability to preserve ice-edge integrity and control error growth. The model also maintains measurable short-range prediction skill, offering a more robust and physically consistent approach to sea ice modeling. The code is planned for public release.