Generative AI models capture realistic sea-ice evolution from days to decades

TL;DR

GenSIM is introduced, the first generative AI model to predict the evolution of the full Arctic sea-ice state at 12-hour increments and implicitly learns hidden signatures of multi-year ice-ocean interaction, so that generative AI can extrapolate from sub-daily forecasts to decadal simulations, while retaining physical consistency.

Abstract

Sea ice plays an important role in stabilising the Earth system. Yet, representing its dynamics remains a major challenge for models, as the underlying processes are scale-invariant and highly anisotropic. This poses a dilemma: physics-based models that faithfully reproduce the observed dynamics are computationally costly, while efficient AI models sacrifice realism. Here, to resolve this dilemma, we introduce GenSIM, the first generative AI model to predict the evolution of the full Arctic sea-ice state at 12-hour increments. Trained for sub-daily forecasting on 20 years of sea-ice-ocean simulation data, GenSIM makes realistic predictions for 30 years, while reproducing the dynamical properties of sea ice with its leads and ridges and capturing long-term trends in the sea-ice volume. Notably, although solely driven by atmospheric reanalysis, GenSIM implicitly learns hidden signatures of multi-year ice-ocean interaction. Therefore, generative AI can extrapolate from sub-daily forecasts to decadal simulations, while retaining physical consistency.

Figures (15)

• Figure 1: Schematic of the efficient generative AI-based sea-ice model GenSIM. Based on flow matching, GenSIM jointly predicts all important sea-ice states for a lead time of 12 hours. (a) Conditioned on initial states $\mathbf{x}_{t}$ and external atmospheric forcings $\mathbf{F}_{t:t+12\,\text{h}}$, the model iteratively generates a prediction $\widehat{\mathbf{x}}_{t+12\,\text{h}}$ 12 hours later. (b) Within an iterative flow step, at arbitrary pseudo time $s \in [0, 1]$, the full Arctic domain is decomposed into overlapping subdomains and processed in parallel with a scale-aware transformer, which predicts a velocity, used to update the fields to the next pseudo tome step $s' \ge s$.
• Figure 2: Physical consistency of GenSIM. Snapshots from a 30-year-long GenSIM simulation for (a--c) 2024-02-13 03:00 UTC and (d--f) 2024-02-13 15:00 UTC with shown sea-ice concentration (a&d), divergence rate (b&e), and change in the sea-ice thickness within 12 hours (c&f). Regions of interest for the (I) opening of leads, (II) healing, and (III) ridging are shown in black boxes.
• Figure 3: Overview over GenSIM's long-term projection qualities. (a) The simulated monthly averaged sea-ice volume in April (maximum) and September (minimum) for GenSIM, neXtSIM-OPA, PIOMAS, the merged CS2/SMOS product, and IceSAT2. Note that CS2/SMOS and IceSAT2 are unavailable for September. (b) The seasonal cycle of the average volume averaged for the period 2015--2018 with the same colouring as for (a). (c, d) The change in the winter thickness between 2020-2024 to 2000--2004 for GenSIM and PIOMAS. (e, f) The winter sea-ice thickness averaged between 2015--2018 for GenSIM and neXtSIM. We define the winter season as January to April.
• Figure 4: GenSIM exhibits long-term ocean-like patterns. (a--d) Volume change within 12 hours for (a) GenSIM, (b) neXtSIM, (c) net bottom growth - bottom melt from neXtSIM, and (d) average 2-metre temperature from the ERA5 reanalysis, averaged for 2015--2018. The black contour line indicate the $-1.8\,^{\circ}\text{C}$ isotherm from the averaged 2-metre temperature. The dynamical component due to advection has been removed from the changes in (a) and (b). (e--h) Speed (background) and streamlines (black lines) for (e) sea ice in GenSIM, (f) sea ice in neXtSIM, (g) ocean surface from the neXtSIM-OPA simulation, and (h) the 10-metre wind from the ERA5 reanalysis, averaged for 2015--2018.
• Figure S1: The different regions used throughout the manuscript with the following names: 1=central Arctic; 2=Beaufort sea; 3=Chukchi sea; 4=east Siberian sea; 5=Laptev sea; 6=Kara sea; 7=Barents sea; 8=east Greenland sea; 9=Baffin bay, Labrador seas, and gulf of St. Lawrence; 10=Hudson bay; 11=Canadian archipelago. The black hatched rectangle is the region in the central Arctic used to estimate the spectra and fractal scaling. The regions are extracted from NSIDC's regions on the EASE-Grid 2.0 with a 3.125 km resolution meier_arctic_2023.