Ultra-high resolution pan-Arctic sea ice-ocean coupled simulation on a heterogeneous many-core supercomputer

Arctic sea ice has undergone dramatic changes in recent decades. The decline in sea ice thickness has resulted in more brittle ice, which is increasingly susceptible to deformation by wind and ocean currents. Small-scale features such as sea ice leads and ridges are frequently observed in the field but remain poorly understood. Accurately forecasting these features requires high-resolution sea ice modeling with a horizontal resolution of several kilometers. To address this, a pan-Arctic ultra-high-resolution (~500 m) sea ice-ocean coupled model has been developed. This model is based on the Massachusetts Institute of Technology General Circulation Model (MITgcm) but has been substantially refactored and enhanced to adapt to the heterogeneous many-core architecture of the computing system. The model's Pacific open boundary is positioned north of the Okhotsk Sea, away from the Aleutian Islands, while the Atlantic open boundary is set north of the Strait of Gibraltar to avoid the influence of deep convection processes. The model operates on a three-dimensional grid comprising approximately 15.1 billion points, with around 9 billion wet points. The sea ice component shares the same grid as the ocean model, enabling direct coupling between the two at each grid point. For sea ice thermodynamics, a zero-heat-capacity, one-layer model is employed, while sea ice dynamics are governed by viscous-plastic rheology. The highly nonlinear sea ice momentum equations are solved using a tridiagonal solver combined with a line successive relaxation method, achieving an accuracy of 1.0×10⁻⁵. The nonlinear integration is iterated 10 times, with each iteration allowing a maximum of 500 steps to ensure convergence of the high-resolution solutions. The model demonstrates significant improvements in simulating sea ice ridges compared to lower-resolution models. Validation against IceSAT-2 along-track data reveals strong agreement in both spatial distribution and probability density function, underscoring the model's enhanced capability to capture small-scale sea ice features.