acceptodds
Under review as a conference paper at ICLR 2027

Physics-Aligned Neural Modeling of Ocean Dynamics

Abstract

Data-driven models have achieved strong performance in geophysical forecasting, particularly in the atmosphere, but transferring these approaches to the ocean remains challenging due to fundamental differences in geometry and dynamics. The ocean is defined on a discontinuous domain with complex coastlines, strong stratification, and highly anisotropic, surface-driven transport, which are poorly aligned with standard learning architectures. As a result, existing models often exhibit physically inconsistent behavior, including land-sea mixing, unrealistic vertical transport, and incorrect propagation of surface forcing. We propose a physics-aligned neural architecture for ocean forecasting that embeds these structural constraints directly into the model design. The framework operates exclusively on the wet domain, respects boundary-driven forcing, decouples horizontal and vertical transport mechanisms, and incorporates geometry-aware interactions adapted to the structure of the ocean system. The resulting model produces stable long-horizon autoregressive simulations with coherent spatiotemporal structure across physical, sea ice, and biogeochemical variables, while remaining robust under perturbed forcing conditions. Additional analyses further demonstrate stable spectral behavior, physically consistent cross-variable coupling, and improved long-term stability relative to architectures that rely on generic feature-space interactions. These results highlight the importance of incorporating domain-specific geometry and inductive biases directly at the architectural level, rather than relying solely on data scale or loss-based physical constraints.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Geophysical forecasting, Geometry-aware learning, Long-horizon prediction, Coupled ocean dynamics

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_niA4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper propose a physics aligned neural architecture designed for ocean forecasting. It embeds structural constraints into the modules. It focus on physical inconsistent scenarios such as land sea mixing and unrealistic vertical transport. The experiments shows that the depth is an important feature dimension about representations.

Critical strengths
  1. The experiments includes coastal errors, vertical profiles and long-rollout diagnostics with standard prediction errors.
  2. The paper models explicitly depth-dependent ocean support and adjacent-layer interactions in its arthitecture.
Critical weaknesses
  1. I am not convinced that all five inventories in Table 8 should have zero residual and assumption in Sec. 5.2. Because dissolved oxygen and nutrients have biogeochemical transformations. Please clarify which source or boundary-flux terms are included in the budget and justify this assumption about interpreting the residuals as conservation errors.

2, I have a concern about Table 8 and Table 10. Both tables report different (positive and negative) drift sign for the same held-out interval, i.e., 2D-GNN nitrate drift is +152% per century in Table 8 but -1.25*10^18 per century in Table 10. What's the reason?

  1. The 2D-GNN ablation study changes depth represenation, geometry and vertical interactions. Thus it didn't isolate the proposed physical constraints.
Questions for authors

Please see the Weaknesses section and consider each point as a question.

1 comment

Thank you for your time and the constructive review. Each point is answered below, with new ablations in point 3.

1. The zero-residual assumption: oxygen and nutrient inventories are not conserved on their own. Oxygen has biological sources and sinks and air-sea exchange, nutrients cycle through biological pools, and our column ends at 3220 m. The Table 8 residual included none of these terms, so it measured the monthly inventory tendency, not conservation error. We now evaluate it against the reference budget: at each monthly step, the change in the model's global inventory minus the change in the GLORYS12/BIOMER inventory over the same domain.

InventoryOurs2D-GNN
Salt0.160.11
PO₄0.762.38
NO₃0.233.46
O₂13.153.42
Silicate3.380.40

Mean |R - R_ref|, % of inventory per century, 191 monthly intervals of the 2010–2025 held-out integration, all 20 levels.

Salt is negligible in both models. For phosphate and nitrate, our model tracks the reference budget 3x and 15x more closely. For oxygen and silicate, BIOMER gains inventory over the period while our model's inventory declines. BIOMER is a free-running hindcast that the producer's quality document (Copernicus Marine, GLOBAL_MULTIYEAR_BGC_001_029) describes as not yet at steady state below the surface, so this residual measures agreement with the hindcast rather than with the ocean. Observational constraints are the reference that would settle it, which we state as a limitation.

2. Opposite signs in Tables 8 and 10: the nitrate case you cite (+152% per century) traces to post-processing, not to the model. The 2D-GNN stores its 20 levels as feature channels, and the Table 8 calculation integrated only the surface layer. Over the full column, its nitrate drift is -5.08% per century and its phosphate drift -2.04%, both now matching Table 10 in sign. A variable-indexing error in post-processing also affected our own oxygen and silicate entries in Table 10; recomputed, they match Table 8 in sign. Table 8 reports the endpoint change of the inventory and Table 10 a linear trend fitted to the monthly series; with the corrections, the two agree in sign for every inventory except salt, whose endpoint drift is below 0.1% per century in both models. The revision reports both, defined in the captions, from the same integration and domain.

3. What the 2D-GNN isolates: the depth-dependent geometry, the interface fluxes and the surface-only forcing all require an explicit depth axis. Without one, there is no per-level wet domain to follow the bathymetry, no interface on which to place a flux, and no surface layer to which forcing can be restricted. Removing depth as a structural axis therefore removes all three together, so the 2D-GNN tests the depth-resolved design as a whole. It also follows the design the field already uses, from global weather prediction (Lam et al., 2023) to its ocean adaptation (Holmberg et al., 2024), in which depth levels are stacked as feature channels.

To isolate the constraints within the depth-resolved model, we retrained three variants under the same protocol as the full model:

Variantθ 200 mθ 2000 mVar. θ 110 mPO₄ / NO₃ residual+3 °C e-folding
Full model0.1830.2340.960.76 / 0.23201 m
No vertical exchange (V = 0)0.0900.1200.200.96 / 0.310.49 m
Forcing at every level0.1030.1401.190.91 / 0.29541 m
Unconstrained vertical operator0.1530.2010.830.82 / 0.26155 m

θ: temperature skill against persistence at 192 months; variance ratio to reanalysis (1 is ideal); residual as in point 1; e-folding: depth at which the warming falls to 1/e of its surface value.

Each constraint has a measurable cost when removed. Without vertical exchange, a +3 °C surface warming never leaves the top layer and skill halves; forcing every level drives the warming 2.7x deeper and inflates variance; an unconstrained vertical operator loses about 15% of the skill and damps variance. The full model leads on skill, variance and budget tracking.

The conclusions rest on internally consistent, measured numbers: positive skill against persistence through 16 years of held-out integration (Table 7), preserved variance and spectra (Table 6), closer tracking of the phosphate and nitrate budgets than the 2D-GNN, and a measurable contribution from each constraint (point 3). These address the three points you raised, and we would be grateful if you would consider whether they change your assessment.

Reviewer_UaqLOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper develops a graph-based ocean model that represents depth explicitly, restricts computation to wet points, and separates horizontal interactions from vertical exchange. It is trained on monthly physical and biogeochemical fields with prescribed atmospheric forcing. The evaluation emphasizes long integrations: held-out results cover 2010–2025, with additional warming experiments and a 200-year rollout under repeated forcing. The proposed model remains considerably more stable than the 2D GNN and retrained GLONET comparisons.

Critical strengths

The explicit treatment of depth and bathymetry is well motivated. Surface and deep wet points have different spatial support; the graph represents that distinction directly instead of placing all depths in a feature vector. The long integrations are the strongest empirical result. Table 7 reports positive skill against persistence through 192 months. Table 6 also checks variance and spectra, which helps rule out simple suppression of variability as the explanation for stability.

Critical weaknesses

The conservation diagnostics need a more careful physical definition. Section 5.2 and Table 8 assign zero expected residual to oxygen and individual nutrient inventories because there is no surface source. Dissolved oxygen has biological sources and sinks as well as air–sea exchange; dissolved nutrients exchange with biological pools. Neither inventory is generally conserved on its own. The modeled column also ends near 3220 m, leaving a lower-boundary exchange term to consider. Small inventory drift is desirable for long integrations, but it does not establish budget closure without these terms.

The two drift tables also need to be reconciled. Table 8 reports oxygen drift of -4.8% per century for the proposed model, while Table 10 reports a positive value of 1.09 in its stated units. The 2D GNN phosphate drift likewise changes sign. Both tables refer to the held-out integration. Different drift estimators may explain this, but a change of units alone cannot.

Questions for authors
  1. What source, sink, and lower-boundary terms are included in the tracer budgets? Could the authors distinguish inventory drift from conservation residual and evaluate the latter against the corresponding reference budget?
  2. How are the drift quantities in Tables 8 and 10 calculated, and what explains the opposite signs for oxygen and phosphate noted above?
1 comment

Thank you for your time and the constructive review. Both questions are answered below with recomputed numbers.

Budget terms. Oxygen and nutrient inventories are not conserved on their own. Oxygen has biological sources and sinks and air-sea exchange, nutrients exchange with biological pools, and our column ends at 3220 m. The Table 8 residual included none of these terms, so it measured the monthly inventory tendency, not closure. We now evaluate the residual against the reference budget, as you propose: at each monthly step, the change in the model's global inventory minus the change in the GLORYS12/BIOMER inventory over the same 0-3220 m domain.

InventoryOurs2D-GNN
Salt0.160.11
PO₄0.762.38
NO₃0.233.46
O₂13.153.42
Silicate3.380.40

Mean |R - R_ref|, % of inventory per century, 191 monthly intervals of the 2010-2025 held-out integration, all 20 levels.

Salt is negligible in both models. For phosphate and nitrate, our model tracks the reference budget 3x and 15x more closely. For oxygen and silicate, the residual is dominated by opposite-sign drifts: BIOMER gains inventory over 2010-2025 (O2 +1.33 x 10^{19} per century), while our model's inventory declines (-0.77 x 10^{19}). BIOMER is a free-running hindcast, which the producer's quality document (Copernicus Marine, GLOBAL_MULTIYEAR_BGC_001_029) describes as not yet at steady state below the surface layers. A residual against it therefore measures agreement with the hindcast, not with the ocean. Observational constraints on subsurface biogeochemistry are the reference that would settle this, and the revision states it as a limitation and a direction for future training.

Tables 8 and 10. As you note, a change of units alone cannot explain the sign changes. Table 8 reports the endpoint change of each inventory, and Table 10 a linear trend fitted to the monthly series. Rechecking both tables identified two post-processing issues that explain them, both now corrected:

  • Table 8's 2D-GNN entries integrated only the surface layer of its channel-packed output. Over all 20 levels, its drifts are PO4 -2.04, NO3 -5.08, O2 +11.77 and silicate +1.98% per century, now matching Table 10 in sign; salt is -0.03%.
  • Table 10's entries for our oxygen and silicate were affected by a variable-indexing error in post-processing. Recomputed, the trends are -0.77 x 10^{19} (O2) and -0.79 x 1018 (silicate) per century, or -5.0% and -1.1% of inventory per century, matching Table 8's endpoint values (-4.8% and -1.0%) in sign and magnitude.

With these corrections, the two estimators agree in sign for every inventory and model except salt, whose endpoint drift is below 0.1% per century in both models. The revision defines both in the captions and replaces the zero-residual framing with the reference budget. The evidence that the long integrations are stable rather than damped does not depend on these tables. Kinetic-energy drift matches GLORYS12 (-3.83 vs -3.88 x 10^{-4} m^{2} s^{-2} per century, against +0.07 for the 2D-GNN; Table 10); our model retains 96-100% of reanalysis interannual variance; and its spectral distances are 0.017-0.175, against 0.94-2.57 for the 2D-GNN (Table 6).

The paper's evidence rests on measured, internally consistent numbers: positive skill against persistence through 16 years of held-out integration, preserved variance and spectra, and closer tracking of the phosphate and nitrate budgets than the 2D-GNN. We hope these answers, and the corrected tables, strengthen your assessment of the paper.

Reviewer_2CEjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a physics-aligned neural architecture for global ocean forecasting that explicitly incorporates ocean-specific geometric and dynamical constraints into a graph-based framework. The proposed model operates on wet ocean points, separates horizontal and vertical transport, models vertical exchange through interface-based fluxes, and restricts atmospheric forcing to the surface layer. It further integrates stochastic latent dynamics and progressive depth-wise training to model coupled physical, sea ice, and biogeochemical variables. Experiments are conducted using GLORYS12, BIOMER, and ERA5 datasets, with comparisons against a 2D GNN and GLONET, as well as persistence and climatology baselines. The authors evaluate long-term autoregressive forecasting performance through predictive skill, spectral consistency, conservation diagnostics, and coastal errors. Additional experiments investigate responses to perturbed atmospheric forcing and stability over extended 200-year simulations. The reported results suggest that incorporating ocean-specific physical structures improves long-horizon forecasting accuracy and stability compared with generic learning architectures.

Critical strengths
  1. The paper incorporates ocean-specific physical characteristics, including complex coastlines, anisotropic transport, vertical stratification, and surface-driven forcing, directly into the neural network architecture. This physically motivated design provides a meaningful alternative to generic learning architectures and offers a promising direction for improving the physical consistency of data-driven ocean forecasting.
  2. The authors evaluate the proposed model through long-horizon autoregressive predictions and provide multiple diagnostic analyses, including spectral consistency, conservation residuals, and responses to perturbed atmospheric forcing. The additional 200-year simulation further examines the model's numerical stability. These experiments provide valuable evidence of the model's ability to maintain coherent ocean dynamics over extended integration periods.
  3. The proposed framework jointly models ocean physical variables, sea ice, and biogeochemical processes within a unified architecture. Evaluating these interconnected components goes beyond forecasting individual ocean variables and demonstrates the potential applicability of the framework to more comprehensive ocean system modeling.
Critical weaknesses
  1. The paper introduces several architectural designs, including horizontal–vertical operator factorization, interface-based vertical coupling, surface-only forcing, and progressive depth-wise training. However, the ablation experiments mainly compare the complete model with a 2D GNN, without independently evaluating the contribution of each component.

  2. The experimental evaluation primarily compares the proposed model against a 2D GNN and GLONET, while several other relevant ocean forecasting architectures are discussed in the related work but not included in the experiments. The limited diversity of baselines makes it difficult to establish the superiority and generalizability of the proposed framework over existing approaches.

  3. Although the proposed architecture incorporates physically motivated constraints, it remains unclear whether these mechanisms strictly enforce the intended physical properties. For example, the stochastic latent prior is conditioned on atmospheric forcing and can modulate the learned dynamics, potentially introducing indirect pathways for forcing information to influence subsurface layers.

  4. The authors acknowledge that the flux-based formulation does not guarantee exact conservation of the decoded variables. More rigorous theoretical analysis or targeted experiments are needed to substantiate the claimed physical consistency of the model.

Questions for authors

Please responds to my concerns in the weaknesses.

1 comment

Thank you for your time and the constructive review. Each point is answered below, with new ablations in point 1.

Contribution of individual components: the horizontal-vertical factorisation, the interface fluxes and the surface-only forcing all require an explicit depth axis: without one there is no vertical operator to separate, no interface on which to place a flux, and no surface layer to which forcing can be restricted. The 2D-GNN removes that axis and therefore tests the depth-resolved design as a whole. It follows the graph-based design used across the field, from global weather prediction (Lam et al. 2023) to its ocean adaptation (Holmberg et al. 2024)”, in which depth levels are stacked as feature channels. Progressive training is different in kind: the full 20-level graph (741,006 wet nodes, 1.73M edges) does not fit in memory end to end on a single H100, so it is what makes the model trainable, not a removable component.

To isolate the constraints within the depth-resolved model, we retrained three variants under the same protocol as the full model:

Variantθ 200 mθ 2000 mVar. θ 110 mPO₄ / NO₃ residual+3 °C e-folding
Full model0.1830.2340.960.76 / 0.23201 m
No vertical exchange (V = 0)0.0900.1200.200.96 / 0.310.49 m
Forcing at every level0.1030.1401.190.91 / 0.29541 m
Unconstrained vertical operator0.1530.2010.830.82 / 0.26155 m

θ: temperature skill against persistence at 192 months; variance ratio to reanalysis (1 is ideal); residual as in point 4; e-folding: depth at which the warming falls to 1/e of its surface value.

Each constraint has a measurable cost when removed. Without vertical exchange, a +3 °C surface warming never leaves the top layer and skill halves; forcing every level drives the warming 2.7x deeper and inflates variance; an unconstrained vertical operator loses about 15% of the skill and damps variance. The full model leads on skill, variance and budget tracking.

Baselines: the comparison covers three reference points: the 2D-GNN, which follows the graph-based family above; GLONET a short-range global ocean forecaster retrained on our data for long-horizon integration; and persistence and climatology as null baselines. Systems such as XiHe, WenHai, FuXi-Ocean and LangYa solve a different problem: daily to sub-daily forecasts of physical variables over short rollouts. Our target is long-horizon stability, where learned emulators are known to fail: they collapse or blow up through artefacts that surface only over long rollouts. The model is built to behave like a numerical ocean model integrated for decades, not to maximise short-range skill, so a short-range benchmark would not test what the paper claims. Of the three learned models, ours is the only one that keeps positive skill against persistence for every variable, depth and lead up to 192 months (Tables 7 and 9), and its skill does not depend on the start date (Table 14).

Forcing reaching the interior through the latent: the latent is not three-dimensional: its prior is conditioned on the surface state and the forcing, and z enters the update only at the surface layer. Its influence therefore reaches the interior through the same interface fluxes as the surface forcing itself, so it adds no direct pathway to depth, and the surface-only constraint holds for the full update.

Physical constraints: as the paper states, conservation is not exact by construction: the vertical coupling passes learned upward and downward messages across each interface, and the decoder is unconstrained. Closure is therefore measured rather than assumed. The revision evaluates closure against the reference budget, i.e. each model's monthly inventory change minus that of GLORYS12/BIOMER, so that biological sources and sinks, air-sea exchange and the 3220 m boundary enter through the reference. On this measure, our model tracks the reference 3x more closely than the 2D-GNN for phosphate and 15x for nitrate (0.76 vs 2.38 and 0.23 vs 3.46% of inventory per century), and salt is negligible in both. For oxygen and silicate, the free-running BIOMER hindcast gains inventory over the period while our model's inventory declines; since the producer's quality document (Copernicus Marine) describes BIOMER as not yet at steady state below the surface, we report this as a limitation of the hindcast reference. Kinetic-energy drift matches GLORYS12 (-3.83 vs -3.88 x 10^{-4} m^{2} s^{-2} per century, against +0.07 for the 2D-GNN; Table 10), and the model retains 96-100% of reanalysis interannual variance (Table 6), so its stability does not come from damping.

The ablations thus give each constraint a measured role, and the diagnostics show that the model’s long-horizon stability is not a product of damping. We would be grateful if you would consider whether these answers change your assessment of the paper.

Comments on the paper

1 comment

Dear authors and reviewers,

The author-reviewer discussion window will close tomorrow on November 18.

@authors: If not done already, please respond to the reviewers' comments before the discussion window closes. I would advise short and concise responses over lengthy explanations.

@reviewers: Please check the authors' responses and provide any additional comments before the discussion window closes. Consider updating your initial review if necessary, and in particular your score if your opinion has changed. You may increase your score if you feel the authors' responses warrant it, maintain your score if your opinion remains the same, or even decrease your score if you feel the other reviews or the authors' responses reveal new shortcomings. At the very least, please acknowledge the authors' responses.

Thank you for your attention and cooperation.

The AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.