the 22nd Pacific−Asian Marginal Seas meeting (PAMS)

The commercialization of the Northern Sea Route (NSR) offers significant economic advantages for the Asia-Pacific region. However, the rapid variability of Arctic sea ice poses severe threats to safe vessel navigation, particularly near critical gateways like the Bering Sea and adjacent marginal seas. Traditional numerical climate models often struggle to provide the high spatiotemporal resolution and timeliness required for tactical, real-time ship routing. To address this, we propose a high-resolution short-term prediction model for Sea Ice Concentration (SIC) based on a Graph Neural Network (GNN), which effectively captures multi-scale spatial interactions and temporal rates of change.

Our framework constructs the Arctic and its adjacent marginal seas as a heterogeneous graph network, integrating 2D ocean grid nodes with real-world meteorological station nodes. This approach successfully models the non-linear spatial dependencies between atmospheric forcing (e.g., ERA5 temperature and wind stress) and sea ice dynamics. To overcome the limitations of static single-point data, we introduce momentum learning—combining past and present variables—to capture the acceleration of ice melting and drift. Furthermore, autoregressive rollout and residual learning techniques are applied to prevent error accumulation, enabling multi-step predictions for up to 30 days.

Crucially for understanding the connectivity between the Pacific and the Arctic, we employ a top-down downscaling approach. The model pre-trains on global atmospheric circulation and macroscopic advection before downscaling to local marginal seas and specific Arctic routes. This minimizes prediction biases caused by coastal topography and local physical effects while maximizing resolution. Evaluated using statistical metrics such as RMSE and ACC against ground-truth data, this AI-based framework demonstrates robust performance. Ultimately, by providing a reliable, real-time decision-support tool for shipping practitioners, this model minimizes navigational risks and optimizes fuel consumption, contributing to the safe integration of Pacific-Asian marginal seas with the NSR.