physics.ao-phMar 12, 2026
SaveFloeNet: A mass-conserving global sea ice emulator that generalizes across climates
Organizations: Atmospheric and Oceanic Sciences Program, Princeton University, NJ, USA · Geophysical Fluid Dynamics Laboratory, NOAA, Princeton, NJ, USA · Allen Institute for Artificial Intelligence (Ai2), Seattle, USA · Courant Institute School of Mathematics, Computing, and Data Science, New York University, New York, NY, USA
Abstract
We introduce FloeNet, a machine-learning emulator trained on the Geophysical Fluid Dynamics Laboratory global sea ice model, SIS2. FloeNet is a mass-conserving model, emulating 6-hour mass and area budget tendencies related to sea ice and snow-on-sea-ice growth, melt, and advection. We train FloeNet using simulated data from a reanalysis-forced ice-ocean simulation and test its ability to generalize to pre-industrial control and 1% CO2 climates. FloeNet outperforms a non-conservative model at reproducing sea ice and snow-on-sea-ice mean state, trends, and inter-annual variability, with volume anomaly correlations above 0.96 in the Antarctic and 0.76 in the Arctic, across all forcings. FloeNet also produces the correct thermodynamic vs dynamic response to forcing, enabling physical interpretability of emulator output. Finally, we show that FloeNet outputs high-fidelity coupling-related variables, including ice-surface skin temperature, ice-to-ocean salt flux, and melting energy fluxes. We hypothesize that FloeNet will improve polar climate processes within existing atmosphere and ocean emulators.
Figures & tables
Figure 1: FloeNet mass and area budget anomaly climatologies (relative to Jan 1 00:00:00) for the period 2006–2022. ( A ) Arctic sea ice mass tendencies due to source terms ( LSRCi ), sink terms ( LSNKi ), and transport convergence ( XPRTi ). The sum of these terms gives the total sea ice mass ( simass ). ( B ) Same as ( A ) but for Antarctic. ( C ) The global sea ice mass budget conservation error, computed as the difference between the prognostic ice mass and the mass computed through a budget reconstruction. ( D – F ) Same as ( A – C ) but for snow mass. ( G – I ) Same as ( A – C ) but for sea ice concentration.
Figure 2: Arctic sea ice thickness and volume mean state and trends. ( A ) Time-mean sea ice thickness climatology from OM4, computed over 2006–2022. ( B ) Same as ( A ) but for a FloeNet rollout with OM4 forcing. ( C ) Same as ( A ) but for a full-state model rollout with OM4 forcing. ( D ) Same as ( A ) but for CM4 under piControl forcing. ( E – F ) Same as ( B – C ) but with CM4 piControl forcing. ( G – I ) Same as ( D – F ) but with 1% CO 2 forcing. ( J ) Seasonal climatology of Arctic sea ice volume, computed over 2006-2022. ( K ) Annual-mean Arctic sea ice volume between 1969–2022. ( L ) Same as ( J ) but for piControl and 1% forcing runs, computed over 211–350. ( M ) Same as ( K ) but for piControl and 1% forcing between 211–350.
Figure 3: Arctic sea ice volume anomalies under different forcings. ( A ) Anomaly correlation coefficient (ACC) of annual-mean sea ice volume anomalies between OM4 and FloeNet, computed at each grid cell over the period 1969–2022. ( B ) Same as ( A ) but ACC computed between OM4 and the full-state model. ( C ) Central Arctic annual-mean sea ice volume anomalies over the period 1969–2022. ( D – F ) Same as ( A – C ) but for piControl forcing. ( G – I ) Same as ( A – C ) but for 1% CO 2 forcing. ( J ) Same as ( C ) but now decomposing FloeNet and OM4 volume anomalies into thermodynamic and dynamic contributions. All central Arctic anomaly time series are computed as averages over the region shown by the contours in the spatial plots.
Figure 4: Same as Fig. 3, but for Antarctic. The contour in the spatial maps corresponds to the latitude 63 ∘ S.
| Name | Description | Type | Time def. | Loss weight |
|---|---|---|---|---|
| siconc | sea ice area fraction [0-1] | prognostic | 6-hour snapshot | 5 |
| simass | sea ice mass [kgm -2 ] | prognostic | 6-hour snapshot | 10 |
| sisnmass | snow mass [kgm -2 ] | prognostic | 6-hour snapshot | 5 |
| TS | ice-surface skin temperature [ ∘ C] | prognostic | 6-hour snapshot | 1 |
| LSRCi | sea ice mass source [kgm -2 s -1 ] | prognostic | 6-hour mean | 1 |
| LSNKi | sea ice mass sink [kgm -2 s -1 ] | prognostic | 6-hour mean | 1 |
Table S1: Details of the FloeNet input and output variables. Radiative fluxes are defined at the top of the sea ice.
| Activation functions | SiLU |
| Batch size | 32 |
| Bias parameters | True |
| EMA decay | 0.999 |
| Epochs | 50 |
| Latent/feature space dimension | 256 |
Table S2: Details of the FloeNet model and training configuration.
Figure S1: Same as Fig. 2 of main article, but for Arctic sea ice concentration and extent.
Figure S2: Same as Fig. 2 of main article, but for Antarctic sea ice concentration and extent.
Figure S3: Time-mean Arctic sea ice thickness bias under different forcings. ( A ) FloeNet’s Arctic sea ice thickness bias under present-day (OM4) forcing, for the period 2006–2022. ( B ) Same as ( A ) but under piControl forcing, for the period 211–350. ( C ) Same as ( B ) but under 1% CO 2 forcing, for the period 211–350. ( D – F ) Same as ( A – C ) but for the full-state model.
Figure S4: Summer and winter Arctic and Antarctic sea ice extent ( A – D ) and volume ( E – H ) from the 1% CO 2 forcing experiment. Arctic summer and winter are given as September and March, while Antarctic summer and winter are given as February and September. These months correspond to the respective sea ice maximum and minimum extent in each hemisphere.
Figure S5: Same as Fig. 2 of main article, but for Antarctic sea ice thickness and volume.
Figure S6: Same as Fig. 2 of main article, but for Arctic snow-on-sea-ice thickness and volume.
Figure S7: Same as Fig. 2 of main article, but for Antarctic snow-on-sea-ice thickness and volume.
Figure S8: Evaluation of FloeNet outputs which have potential application for atmosphere-ice-ocean coupling. ( A ) Ice-surface skin temperature seasonal cycle, averaged over grid cells north of 60 ∘ N and for the period 2006–2022. ( B ) A snapshot of ice temperature from OM4 on June 1 2018 12:00:00. ( C ) Same as ( B ) but generated from FloeNet. ( D ) Same as ( B ) but generated from the full-state model. ( E – H ) Same as ( A – D ) but for ice-to-ocean salt flux. ( I – L ) Same as ( A – D ) but for bottom melting energy flux. ( M – P ) Same as ( A – D ) but for bottom melting energy flux.
Figure S9: Same as Fig. S7, but for Antarctic. Seasonal cycles are averaged over grid cells south of 60 ∘ S and snapshots correspond to November 1 2018 12:00:00.
Figure S10: Validation of sea ice thickness and volume for different random seed initializations during training. ( A – C ) Arctic sea ice thickness bias. ( D – F ) Same as ( A – C ) but for Antarctic. ( G ) Pan-Arctic sea ice volume. ( H ) Annual-mean Arctic sea ice volume between 2006–2022. ( I , J ) Same as ( G , H ) but for Antarctic.
Figure S11: Same as Fig. S9 but for snow-on-sea-ice thickness and volume.
Figure S12: Receptive field of different multi-mesh resolutions, shown as contours of grid cells which are activated from a given seed node (red) in the Arctic ( A – C ) and Antarctic ( D – F ). ( A , D ) All resolutions M 0 through M 6 . ( B , E ) Resolutions M 4 through M 6 (FloeNet). ( C , F ) Only M 6 . For scale, the contours in ( B ) and ( E ) have a radius of approximately 1500 km, while those in ( C ) and ( F ) have a radius of approximately 700 km.
Figure S13: Validation of sea ice thickness and volume for different receptive fields; see Fig. S11 for definitions. ( A – C ) Arctic sea ice thickness bias for increased mesh receptive field, the default FloeNet mesh, and a smaller receptive field, respectively. ( D – F ) Same as ( A – C ) but for Antarctic. ( G ) Pan-Arctic sea ice volume. ( H ) Annual-mean Arctic sea ice volume between 2006–2022. ( I , J ) Same as ( G , H ) but for Antarctic.
Figure S14: Same as Fig. S13 but for snow-on-sea-ice thickness and volume.
Figure S15: Ablation test, removing rate of snowfall ( SNOWFL ) from the set of input forcing variables. ( A – D ) Arctic and Antarctic snow-on-sea-ice thickness biases. ( E , F ) Annual-mean snow volume time series between 2006–2022. ( G – L ) Same as ( A – F ) but for ice-surface skin temperature.
Figure S16: Ablation test, removing ice-ocean stress ( tauuo , tauvo ) from the set of input forcing variables. ( A – D ) Arctic and Antarctic sea ice thickness biases. ( E , F ) Annual-mean sea ice volume time series between 2006–2022. ( G – L ) Same as ( A – F ) but for snow-on-sea-ice thickness and volume.
Figure S17: Evaluating FloeNet rollouts when trained with a 6-hour vs 5-day timestep. ( A – D ) Arctic and Antarctic sea ice thickness biases. ( E , F ) Annual-mean sea ice volume time series between 2006–2022. ( G – L ) Same as ( A – F ) but for ice-surface skin temperature.
Figure S18: Comparing Graph- (FloeNet) and Convolution-based (ConvNeXt) architectures. ( A – D ) Arctic and Antarctic sea ice thickness biases. ( E , F ) Annual-mean sea ice volume time series between 2006–2022. ( G – L ) Same as ( A – F ) but for snow-on-sea-ice thickness and volume.
More in this topic
Climate Modeling
Climate modeling uses physical simulations and data-driven methods to understand, predict, and project Earth’s climate and its response to external forces.
See all 68 papers