September 24th, 2026
Key Findings
- FloeNet is a machine learning emulator of GFDL’s global Sea Ice Simulator version 2 (SIS2) that predicts changes in sea ice and snow on sea ice every six hours.
- Using a graph neural network, the model represented sea ice growth, melt, and transport while conserving sea ice mass and area at each timestep.
- When tested under present-day, pre-industrial, and increasing CO₂ conditions, the model successfully reproduced sea ice trends and year-to-year variability across climates beyond those used for training.
- Enforcing physical conservation improved FloeNet’s performance across different climate conditions compared with a model that does not conserve mass.
- The model also simulates variables needed for interactions among sea ice, the atmosphere, and the ocean, providing a potential sea ice component for coupled machine learning models.
William Gregory, Mitchell Bushuk, James P C Duncan, Elynn Wu, Adam Subel, Spencer K Clark, William J Hurlin, Oliver Watt-Meyer, Alistair Adcroft, Christopher S Bretherton, and Laure Zanna. Geophysical Research Letters. DOI: 10.1029/2026GL122981
Sea ice is an important part of the Earth system because it regulates exchanges of heat, moisture, radiation, momentum, mass, and salt between the atmosphere and the ocean. In this study, the authors introduce FloeNet, a machine learning emulator designed to reproduce the behavior of GFDL’s global sea ice model, SIS2, at a substantially lower computational cost. The research examines whether incorporating physical constraints into a machine learning model can improve its ability to simulate sea ice under conditions different from those used for training.
FloeNet uses a graph neural network to predict six-hour changes in sea ice and snow on sea ice caused by growth, melt, and horizontal transport. Rather than directly predicting only the resulting sea ice state, FloeNet predicts the individual processes that change ice and snow mass and sea ice area. These quantities are combined at every timestep in a way that conserves mass and area. This approach also makes the model more interpretable because changes in sea ice can be separated into contributions from thermodynamic processes, such as growth and melt, and dynamic processes, such as transport.
The authors trained FloeNet using output from a GFDL ocean–sea ice simulation driven by present-day atmospheric conditions and then tested it under present-day, pre-industrial, and increasing CO₂ conditions. Across these tests, FloeNet generally reproduced sea ice mean conditions, long-term trends, and year-to-year variability more accurately than a comparable machine learning model without the conservation framework. The results were particularly strong for sea ice volume variability: anomaly correlations exceeded 0.96 in the Antarctic and 0.76 in the Arctic across the tested forcing conditions. The model nevertheless retains some biases, including overestimation of Arctic Sea ice thickness under the pre-industrial and increasing-CO₂ experiments.
FloeNet can also reproduce variables involved in exchanges between sea ice, the atmosphere, and the ocean, including ice-surface temperature, ice-to-ocean salt flux, and melting energy fluxes. Once trained, a 140-year FloeNet simulation took about 4.75 hours on a single GPU; for comparison, the authors report that SIS2 required about 17 hours on 4,671 CPUs within a comparable 140-year coupled ice-ocean simulation. These capabilities provide a computationally efficient approach for studying sea ice variability and for future development of coupled atmosphere-ocean-sea ice machine learning models. The work also contributes to GFDL’s development and evaluation of modeling approaches for understanding and predicting changes in the Earth system.
