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

Adam Subel; Alistair Adcroft; Bill Hurlin; Chris Bretherton; Elynn Wu; James Duncan; Laure Zanna; Mitchell Bushuk; Oliver Watt-Meyer; Spencer K. Clark

REVIEW 2 major objections 5 minor 1 cited by

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

T0 review · 2 major / 5 minor · reviewed 2026-07-14 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A mass-conserving sea ice emulator trained on present-day data generalizes to pre-industrial and warming climates, and correctly separates growth, melt, and transport.

desk verdict Solid, usable advance: budget-based mass conservation plus a matched baseline gives real gains in cross-climate volume variability and coupling fluxes; mean-state thickness biases remain and energy is not closed. read the letter →

arxiv 2603.12449 v1 pith:V5WFPK3G submitted 2026-03-12 physics.ao-ph cs.LG

William Gregory , Mitchell Bushuk , James Duncan , Elynn Wu , Adam Subel , Spencer K. Clark , Bill Hurlin , Oliver Watt-Meyer

show 3 more authors

Alistair Adcroft Chris Bretherton Laure Zanna

This is my paper · ORCID

classification physics.ao-phcs.LG

keywords seaiceemulatormassconservationgraphneuralnetworkclimategeneralizationbudgettendenciespolarcoupledemulators

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Climate models are too expensive to run at the ensemble sizes needed to map uncertainty, so machine-learning emulators that reproduce whole Earth-system components in a fraction of the time are becoming central. This paper shows that a global sea-ice emulator can be made both more accurate and more interpretable by predicting the budget terms that change ice and snow mass and area (growth, melt, and advection) and then enforcing mass and area conservation at every step. Trained only on a present-day reanalysis-forced ice-ocean run, the model, FloeNet, reproduces mean state, multi-decadal trends, and year-to-year volume anomalies under both a colder pre-industrial control climate and a steadily rising CO2 climate, with volume anomaly correlations above 0.96 in the Antarctic and 0.76 in the Arctic. It also attributes those anomalies correctly to thermodynamic versus dynamic processes and supplies the fluxes (skin temperature, salt flux, melt energy) needed to couple to atmosphere and ocean emulators. The result is evidence that hard physical constraints improve out-of-sample climate generalization and open a practical path to interactive polar components inside fully machine-learned climate models.

What carries the argument

Mass-and-area budget emulation with hard conservation: the network outputs source, sink, and transport tendencies for ice and snow; the prognostic states are overwritten each 6-hour step by the time-integrated budgets (with non-negativity redistribution), so the mass and area equations are satisfied to machine precision and every change remains attributable.

What would settle it

Train the same budget architecture on the present-day run, then roll it out under piControl or 1% CO2 forcing and check whether Arctic volume anomaly correlations fall below those of the non-conserving baseline or the thermodynamic-versus-dynamic partition of volume anomalies no longer matches the parent model.

Watch

Extended reading notes

Core claim

A graph-neural-network sea-ice emulator that predicts mass and area budget tendencies and enforces conservation at every timestep outperforms an otherwise identical full-state model at mean state, trends, and inter-annual volume variability across present-day, pre-industrial, and 1% CO2 climates, while correctly partitioning thermodynamic and dynamic contributions and producing high-fidelity coupling fluxes.

Load-bearing premise

Enforcing mass and area conservation alone is enough for out-of-sample skill even though energy is not conserved and the training climate has a different mean ice thickness from the colder and warmer target climates.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The thing worth knowing is that this is a working, coupling-ready global sea-ice emulator that actually generalizes on variability and trends, not just another full-state GNN. They train on OM4/JRA-55, then roll out multi-decade continuous trajectories under piControl and 1% CO2 from CM4, and beat a matched non-conserving full-state baseline on volume anomaly correlations (Arctic ~0.76, Antarctic ~0.96) while correctly partitioning thermo vs dynamic contributions. That is new relative to Durand/Finn/Duncan-style full-state work, and the budget residual plots close to machine precision.

What they do well is concrete. The architecture (GraphCast-style multi-mesh GNN restricted to M4–M6, 4 processor layers, land nodes removed) is sensible for polar grids. Emulating the source/sink/transport tendencies and overwriting the prognostic mass/area with the integrated budget (with a clean redistribution rule for negatives) is simple and effective; the full-state baseline uses identical data, architecture and training, so the performance gap is cleanly attributable. Coupling diagnostics (TS, SALTF, TMELT, BMELT) look high-fidelity at 6 h, and the code/weights are public. Sensitivity tests on seed, mesh, forcing ablations, timestep and ConvNeXt are in the SI and honest.

Soft spots are real but proportionate. Positive Arctic thickness/volume biases appear under both piControl (+2645 km^{3}) and 1% CO2 (+1389 km^{3}) even though spatial patterns are right; energy/enthalpy is not conserved, so mass closure alone cannot guarantee that growth rates stay consistent with the new basal heat fluxes. The authors flag this themselves and sketch the enthalpy fix. Loss weights, mesh refinements and training-rollout length are free parameters, but they are documented. Mean-state generalization is therefore only partial; the variability and trend claims hold up.

This is for people building or coupling climate emulators (ACE2, Samudra, SamudrACE) and for polar-process groups who need interpretable budget diagnostics. It is not a pure theory paper. The math is standard, the data and citations are solid, and the empirical evidence is strong enough that a serious editor should send it to referees. I would engage with it, cite the variability and coupling results, and watch for the energy-conserving follow-up.

Referee Report

2 major / 5 minor

Summary. FloeNet is a GraphCast-style multi-mesh GNN that emulates SIS2 by predicting 6-hour ice/snow mass and area budget tendencies (LSRC, LSNK, XPRT) under atmospheric and oceanic forcing, then overwriting the prognostic state with the time-integrated budget (eqs. 1–3 and §S1) so that mass and area close to machine precision. Trained only on a reanalysis-forced OM4 present-day run (1969–2005), it is evaluated in continuous multi-decade rollouts under OM4 present-day, CM4 piControl, and CM4 1% CO2 forcings. Against a matched full-state baseline that predicts concentration, mass, snow, velocities and salt flux directly, FloeNet shows higher volume anomaly correlations (ACC > 0.96 Antarctic, > 0.76 Arctic), cleaner thickness spatial patterns, correct thermo/dynamic partitioning of anomalies, and usable coupling fluxes (TS, SALTF, TMELT, BMELT).

Significance. If the results hold, the paper supplies a practical, mass-conserving global sea-ice component that can be dropped into existing 6-hour atmosphere (ACE2, CAMulator) and ocean (Samudra) emulators and into SamudrACE. The open weights, Ai2 codebase, and explicit budget diagnostics make the work immediately usable and falsifiable. The controlled full-state ablation cleanly isolates the benefit of the budget formulation for variability and interpretability, which is a concrete advance over prior sea-ice emulators that only predict concentration or thickness.

major comments (2)

The abstract and Key Points claim that FloeNet “outperforms a non-conservative model at reproducing … mean state, trends, and inter-annual variability” and “generalizes across climates.” Figs. 2E,H,L and S3 document systematic positive Arctic volume biases of +2645 km³ (piControl) and +1389 km³ (1% CO2) relative to CM4, even while spatial patterns and detrended ACC remain good. Section 4 correctly notes that energy/enthalpy is not conserved and that the OM4 training mean thickness differs from the CM4 targets. The mass-overwrite procedure (eqs. 1–3 + §S1 redistribution) therefore integrates growth rates that can be inconsistent with the energy available under the new forcings. The mean-state half of the generalization claim is consequently overstated relative to the evidence; the paper should either (i) qualify the claim to emphasize variability/trends/interpretability or (ii) add a quan
The full-state baseline is the sole control that isolates the budget formulation. Supplementary §S2 reports non-negligible sensitivity of Arctic thickness RMSE to random seed (9.81–10.99 cm) and to mesh resolution, yet only a single seed is used for the full-state comparison in the main figures. Because the performance gap on mean state is modest under piControl/1% (and partly compensating in the full-state model), a multi-seed or multi-checkpoint comparison is needed to confirm that the reported superiority is not seed-dependent.

minor comments (5)

Fig. 1 caption and panels C/F/I: clarify that the “budget residual” for FloeNet is after the overwrite step; otherwise readers may misread the machine-precision closure as an emergent property of the network rather than an enforced constraint.
Table S1 loss weights (siconc=5, simass=10, sisnmass=5) are stated to be “critical for stable training”; a one-sentence note on how they were chosen (or a brief ablation) would help reproducibility.
In §S1 the redistribution rule for negative mass is described only by example. A short algorithmic box or pseudocode would remove ambiguity about the priority order among LSRC/LSNK/XPRT.
Figs. S8–S9: the bottom two rows of S8 are both labelled “bottom melting energy flux”; one should be top melt (TMELT).
The hypothesis that FloeNet will improve polar biases in SamudrACE is reasonable but untested; a single offline coupling diagnostic (e.g., salt-flux consistency with Samudra’s salinity budget) would strengthen the discussion without requiring a full coupled run.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: mass conservation is a post-prediction hard constraint; skill claims rest on independent CM4 climates and a matched full-state baseline.

full rationale

FloeNet is trained to match SIS2 mass/area budget tendencies on OM4/JRA-55 data (1969–2005). At inference the network’s prognostic state is overwritten by the time-integrated budget (eqs. 1–3 and §S1 redistribution) so that mass and area close by construction; this is an architectural constraint, not a fitted free parameter that is later re-labeled as a prediction. All reported skill metrics (volume anomaly correlations >0.96 Antarctic / >0.76 Arctic, trends, thermodynamic-vs-dynamic partition) are evaluated on held-out present-day years and on entirely independent CM4 piControl and 1 % CO2 climates never seen during training. The full-state baseline uses the identical GNN architecture, forcing variables, and training data, so the performance gap is an empirical comparison, not a tautology. No uniqueness theorem, self-citation chain, or ansatz is load-bearing for the central claim. The acknowledged positive Arctic thickness biases under climate shifts and the absence of energy conservation are correctness/generalization limitations, not circularity. Score 0 is therefore the honest finding.

Assumptions & free parameters 3 free parameters · 3 assumptions · 1 invented entities

The central generalization claim rests on standard GNN message-passing, the SIS2 mass/area budget equations, the choice to overwrite state with integrated tendencies, and the empirical claim that present-day OM4 training transfers to CM4 climates. No new physical constants or particles are invented; free parameters are ordinary network hyperparameters and loss weights.

free parameters (3)

loss weights on siconc/simass/sisnmass = 5, 10, 5
Elevated weights (5/10/5) chosen to stabilize budget training; not derived from first principles.
multi-mesh refinements and processor depth = M4-M6, 4 layers
Restricted to M4-M6 and 4 GNN layers by hand to control non-locality and cost.
rollout length during training = 8 steps
8-step (2-day) unrolls selected for performance; longer not tested for cost reasons.

assumptions (3)

domain assumption Sea-ice mass and area evolve according to the SIS2 source-sink-transport budget (eqs. 1-3) and remain non-negative.
Taken directly from the host model; used both for training targets and for the conservation overwrite.
ad hoc to paper A multi-mesh GNN with local encoder/decoder and limited processor depth can represent basin-scale elastic stresses without global edges.
Architectural choice justified by physical intuition but not proven; ablated only partially.
domain assumption Present-day reanalysis-forced OM4 statistics are sufficiently close to CM4 piControl and 1%CO2 statistics for zero-shot transfer of the learned tendencies.
Implicit in the generalization experiments; residual thickness biases show the assumption is imperfect.

invented entities (1)

FloeNet mass-conserving budget overwrite
purpose: Guarantee non-negative mass/area and exact budget closure after each 6-hour step.
Post-network correction that redistributes excess negative tendency among LSRC/LSNK/XPRT; independent evidence is the reported residual plots, but the procedure itself is an engineering construct of this paper.

how reviews work

0 comments

Cite this review

Pith. "Pith review of FloeNet: A mass-conserving global sea ice emulator that generalizes across climates." pith.science (2026). https://pith.science/paper/V5WFPK3G

@misc{pith2026260312449,
  author       = {Pith},
  title        = {Pith review of: FloeNet: A mass-conserving global sea ice emulator that generalizes across climates},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V5WFPK3G}},
  note         = {Machine review of arXiv:2603.12449}
}

read the original 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.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Hard conservation correctors can hide a degrading model when training autoregressive emulators
physics.ao-ph 2026-07 conditional novelty 7.0 of 10

Supervising only the corrected output of a hard water-budget corrector hides raw-precipitation amplitude drift, letting the required correction grow from 2% to 24% while delivered fields stay exactly conservative.

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.