Abstract
Simulating the Arctic Ocean at high resolutions is computationally intensive and often plagued by boundary errors. This paper presents a modified NEMO-LIM3 ocean-ice model utilizing a decoupled, two-grid approach with custom boundary conditions and an indirect ice-data assimilation scheme. This framework significantly reduces computational overhead and model spin-up time while enhancing the physical accuracy of regional Arctic hindcasts.
Introduction
As Arctic shipping and offshore development expand, the need for high-fidelity, long-term historical data (hindcasts) of ocean currents and sea ice becomes critical for operational risk assessment. However, running global ocean models at high spatial resolutions (e.g., 5 km) is computationally prohibitive. Regional models offer a solution but introduce specific challenges: nesting inflexibility, mathematical instabilities at open boundaries (especially under sea ice), and lengthy "spin-up" periods required to reach a stable physical state.
Methodology
The authors developed a custom simulation framework by introducing several key modifications to the NEMO-LIM3 source code to bypass traditional modeling bottlenecks.
Decoupled Multigrid Setup
Instead of running nested grids simultaneously—which requires strict synchronization of supercomputer nodes—the authors decoupled the process. A coarse pan-Arctic grid (14 km) is run first. Its output is then saved, interpolated, and used to provide boundary and initial conditions for a separate, fine-resolution grid (5 km) focused on the Russian Arctic seas.
Hybrid Open Boundary Conditions (OBC)
To eliminate non-physical artifacts at the edges of the regional model, the team implemented a hybrid approach:
- Tracers (Temperature/Salinity): A Flow Relaxation Scheme (FRS) was combined with an advection scheme (ADV) to prevent rapid gradient changes.
- Ice Drift: Default rules were augmented with FRS conditions that incorporate external satellite drift data, preventing unrealistic ice accumulation at the boundaries.
Flux-Based Ice Restoring
To accelerate the spin-up phase, an indirect data assimilation scheme was implemented. Rather than overwriting the model's ice state (which can break physical consistency), the system uses a feedback loop in the heat flux equations. If the model deviates from satellite observations, it adjusts solar melting or freezing fluxes to guide the model toward reality naturally.
Datasets and Forcing
The model integrates a variety of high-quality data sources:
- Ocean Conditions: GLORYS2V4 global ocean reanalysis.
- Atmospheric Forcing: Polar WRF, ERA-Interim, NCEP-DOE 2, and DFS 5.2 reanalyses.
- Satellite Ice Data: Daily OSI SAF (concentration) and weekly CS2SMOS (thickness) datasets.
- Riverine Inputs: Empirical datasets for Russian Arctic rivers, with direct overrides at river mouths to stabilize coastal salinity and temperature.
Key Results and Discussion
The decoupled approach proved highly efficient for High-Performance Computing (HPC) logistics, allowing independent scheduling of grid runs.
- Computational Efficiency: The 5 km fine-grid simulation required only 18 hours of wall-time per simulated year, a significant improvement over higher-resolution alternatives.
- Reduced Spin-up: The flux-based restoring scheme cut the required spin-up time from over four cycles down to just one or two.
- Improved Accuracy: The modified boundary conditions eliminated temperature artifacts. With ice restoring active, the fine-grid model's ice area Root Mean Square Error (RMSE) improved to 0.25 (10^6 km^2).
Conclusion
By decoupling the grid hierarchy and utilizing indirect data assimilation, the authors have created a viable path for generating long-term, high-resolution Arctic hindcasts. While the offline approach adds complexity to the data pipeline, the gains in computational flexibility and physical stability are substantial. Future work will focus on utilizing this configuration to produce a continuous 50-year regional hindcast.
Summary updated