FloeNet: A mass-conserving global sea ice emulator that generalizes across climates

2026

Read FloeNet: A mass-conserving global sea ice emulator that generalizes across climates (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-12449.txt.

Introduction

Gregory and co-authors introduce FloeNet, a mass-conserving emulator of global sea ice that generalizes across climates. The model learns from the GFDL SIS2 sea ice model in training runs. It emulates six-hour budget tendencies for ice and snow growth, melt, and advection. Training used a reanalysis-forced ice-ocean run, with tests on pre-industrial and rising CO2 climates. FloeNet beat a non-conservative rival on mean state, trends, and year-to-year variability in both poles. Its thermodynamic versus dynamic response also stays physically readable.

Catalog record

YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE63289
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.12449

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

FloeNet: A mass-conserving global sea ice emulator that generalizes across climates 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% CO 2 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. † † journal: Geophysical Research Letters † † authors: William Gregory, Mitchell Bushuk, James Duncan, Elynn Wu, Adam Subel, Spencer K. Clark, Bill Hurlin, Oliver Watt-Meyer, Alistair Adcroft, Chris Bretherton, Laure Zanna † † corresponding: Will Gregory, [email protected] † † authors: William Gregory, Mitchell Bushuk, James Duncan, Elynn Wu, Adam Subel, Spencer K. Clark, Bill Hurlin, Oliver Watt-Meyer, Alistair Adcroft, Chris Bretherton, Laure Zanna 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 keypoints FloeNet is a mass-conserving graph neural network sea ice emulator that learns both ice dynamics and thermodynamics FloeNet is trained on a present-day climate and can generalize to pre-industrial and increasing 1%/year CO 2 climates Emulating budget terms and enforcing mass conservation improves generalization over predicting full state fields Plain Language Summary Climate models have been used for several decades to make long-term climate projections. These models are very expensive to run, meaning that we typically do not run enough simulations to appropriately constrain uncertainty related to climate change emissions scenarios and internal climate variability. This cost also effectively limits the number of institutions that are contributing to the climate science discourse. In various areas of scientific research, we are seeing how machine learning models can identify patterns in computer model outputs and approximate the underlying physical processes. These machine learning emulators have an advantage that, once trained, they are fast to run, even on a home computer. For climate emulators, this means that users can run hundreds to thousands more experiments than they can with a traditional climate model. This study represents a contribution to the field of climate model emulation. We develop an emulator of one global climate model’s sea ice component and show that incorporating our physical understanding of sea ice behavior into the emulator improves its accuracy when evaluated under colder and warmer climates. Our ambition is to combine this sea ice emulator with other atmosphere and ocean emulators to provide a fully machine-learned model of Earth’s climate. 1 Introduction Machine learning (ML) models are driving major advances in numerous areas of Earth system science, including remote sensing Bastin et al. (2019) ; Gregory et al. (2024) ; Chen et al. (2025) , climate downscaling Brenowitz et al. (2025) ; Mardani et al. (2025) ; Perkins et al. (2025) , and environmental sensor placement Andersson et al. (2023) . In climate modeling, ML models are now being integrated into large-scale climate model components, leading to improved subgrid parameterization schemes Mansfield & Sheshadri (2024) ; Behrens et al. (2025) ; Zanna et al. (2025) and bias-corrected simulations Watt-Meyer et al. (2021) ; Chapman & Berner (2025) ; Gregory et al. (2026) . However, hybrid and conventional physics-based models remain subject to computational limitations, requiring thousands of CPU cores to achieve modest throughputs of 10-20 simulated years per day. In response to the computational challenges of numerical modeling, we are seeing the rapid development of auto-regressive ML-based climate model emulators. These models have built on the success of data-driven numerical weather prediction models Pathak et al. (2022) ; Lam et al. (2023) ; Kochkov et al. (2024) ; Allen et al. (2025) , now with global-scale climate emulators of the atmosphere Chapman et al. (2025) ; Watt-Meyer et al. (2025) ; Perkins et al. (2025) , ocean Dheeshjith et al. (2025) , coupled atmosphere-ocean Clark et al. (2025) ; Duncan et al. (2025) , and Arctic sea ice Durand et al. (2024) ; Finn et al. (2025) . Across these domains, we consistently observe that, once trained, ML emulators can accurately simulate entire Earth system components in a fraction of the time of their numerical model counterparts; this has enabled new research directions, such as the study of extreme events using very large ensembles Paciorek & Cooley (2025) . Sea ice is an integral component of the coupled climate system. By modulating exchanges of heat, moisture, radiation, and momentum between the atmosphere and ocean, sea ice plays a significant role in regulating Earth’s surface temperature and high-latitude climate feedbacks Hahn et al. (2021) . For climate emulation, sea ice concentration may be provided as a forcing variable to atmospheric models Watt-Meyer et al. (2025) or, in a coupled model, emulated prognostically within the ocean Duncan et al. (2025) . Still, some of the largest outstanding surface air temperature biases within these models exist within the polar regions, prompting broader questions about the representation of sea ice within data-driven emulators. Sea ice concentration alone cannot fully capture the complex exchanges of energy, mass, and salt across the atmosphere-ice-ocean interface, and indeed, Duncan et al. (2025) found better generalization by also prognosing sea ice thickness. We therefore advocate for a sophisticated sea ice emulator which can not only simulate the evolution of sea ice and its response to different forcings, but also provide critical information on coupled exchanges. To this end, we introduce FloeNet, a data-driven emulator architecture trained on the Geophysical Fluid Dynamics Laboratory (GFDL) global sea ice model, SIS2. FloeNet is a deterministic auto-regressive graph neural network (GNN) which marks a step forward in sea ice emulation as the first model to dynamically evolve sea ice and snow-on-sea-ice while conserving mass and area. Specifically, FloeNet receives mechanical and thermodynamic forcing inputs from the atmosphere and ocean, and predicts ice and snow mass tendencies due to growth, melt, and advection. This yields a mass-conservative and interpretable model, as timestep-to-timestep changes in sea ice area and mass can now be attributed to each term in their respective budget. In this study, we train FloeNet on simulated data from a present-day climate and test its ability to generalize to present-day, pre-industrial control (piControl) and 1% CO 2 climates. Across these different forcings we show that FloeNet outperforms a non-conservative model at reproducing sea ice trends and inter-annual variability. Our study therefore provides evidence that incorporating physical principles into ML emulators can improve generalization Falasca (2025) . 2 Data and methods 2.1 Training and validation data To train FloeNet, we generate a 65-year simulation between 1958–2022 using the GFDL fourth-generation global ocean-sea-ice model,

Read the complete text