Physics-Informed Prediction of Arctic Sea Ice Drift

The most visible effect of climate change in the Arctic is the melting of sea ice. This has adverse implications for ship route planning and safe navigation, as melting ice fractures into floes. The dynamics of these floes are governed primarily by the near-surface wind, but also by the ocean surface current, sea ice properties (e.g., thickness and salinity), and overall sea ice distribution. Given the highly dynamic nature of sea ice, accurate short-term Arctic sea-ice motion forecasts are required; however, accurately predicting sea ice drift on short time scales remains a challenging problem.

Current methods for determining sea-ice motion rely on satellite data or climate model forecasts, both of which have limitations. The satellite data used include optical imagery, which is high-resolution, but, blocked by clouds for days at a time. Passive microwave data can penetrate clouds and provide decades-long time series, but are low-resolution, between 26 km and 10 km, depending on frequency and processing. SAR (Synthetic Aperture Radar) satellites have high resolution, but exact overpasses of areas of interest can be days apart. Further, the maximum cross correlation (MCC) technique, which relies on feature tracking, is problematic in the Marginal Ice Zone (MIZ) close to the ice edge or during summer melt conditions. During ice melt seasons or at locations in which sea ice dynamics are highly variable (e.g., the MIZ), decorrelation of the SAR signal precludes accurate feature tracking from one image to the next. In these very dynamic regimes, the majority of ice floes are small and tracking ice motion requires spatial resolutions on the order of only several hundred meters and temporal resolutions of at least a day. Large-scale climate models are designed to predict sea ice motion at seasonal time scales and are computationally expensive.

Machine learning (ML) has the potential to solve the prediction problem; however, classical ML techniques require large amounts of data from which to learn the best model. Further, there is no guarantee that these models obey the laws of physics that govern the modeled dynamics. Physics-informed neural networks (PINNs) are a recent innovation that reduce the amount of data required for learning by constraining the model to generate solutions that satisfy a given partial differential equation (PDE).

We present a PINN that predicts Arctic sea-ice motion and present a case study within the MIZ along the East Greenland current. Our model enforces a form of the momentum equation for sea ice that makes no assumptions about ice rheology and does not explicitly consider internal sea ice stresses. The model is trained using a small set of data including HYbrid Coordinate Ocean Model (HYCOM) ocean surface currents, Global Forecast System (GFS) wind vectors, Soil Moisture and Ocean Salinity (SMOS)-CryoSat2 merged sea ice thickness measurements, and Advanced Microwave Scanning Radiometer 2 (AMSR2)-derived sea ice concentration. The results are validated against the 25 km daily sea ice motion product from the National Snow and Ice Data Center (NSIDC).