acceptodds
Under review as a conference paper at ICLR 2027

HelmGeoNet: Physics-Embedded Forecasting of Global Geostrophic Currents via Helmholtz Decomposition

Abstract

Accurate and stable long-horizon forecasting of global ocean currents remains difficult. Data-driven models typically learn the velocity field as a whole and forecast it autoregressively. Without explicit guidance on the evolution of its physically distinct components, they capture the overall trend but learn each component inaccurately, so forecasts of currents and kinetic energy lose accuracy and stability at long horizons. To address this, we propose HelmGeoNet for more energy-faithful forecasting of geostrophic currents. Specifically, we embed the Helmholtz decomposition into the network architecture, so that the model explicitly learns the evolution of each physically distinct component. We also introduce a physics-guided relaxation mechanism that filters out redundant information from the input field, mitigating the error accumulation. Experiments on daily geostrophic currents confirm that HelmGeoNet outperforms eight spatiotemporal backbones in currents forecasting. At the 20-day lead, its eddy kinetic energy RMSE is 0.0029 m/s lower than that of the best baseline, and its EKE remains stable across the forecast horizon.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Deep Learning, AI for science, Spatiotemporal Forecasting, Helmholtz Decomposition, Geostrophic Currents

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 2 reviews, above 27% of the others.
Reviewer_hPzNOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces a structured predictor for global surface geostrophic currents. It separates velocity into solenoidal, irrotational, and uniform components, predicts vorticity corrections followed by spectral Biot-Savart inversion, applies damped persistence plus a learned correction to the irrotational component, and persists the uniform flow. Models trained for one-day prediction are rolled out for 20 days on AVISO/DUACS data, with evaluation of velocity and eddy kinetic energy (EKE). The central contribution is a task-specific combination of physical representation, residual learning, and component-dependent forecasting priors.

Critical strengths
  1. The application contribution is meaningful.** Predicting global geostrophic currents on a 680 x 1440 grid, while explicitly examining EKE over 20-day rollouts, presents a materially different setting from small synthetic fluid benchmarks or regional scalar-field prediction. HelmFluid already uses Helmholtz-inspired dynamics, but its reported sea-temperature experiment uses four 64 x 64 regions. HelmGeoNet's global current-vector task and energy-focused formulation deserve independent credit; they are not rendered unoriginal merely because the decomposition is classical. See HelmFluid.

  2. The methodological contribution is a coherent domain-specific adaptation.** Vorticity-residual prediction, an explicit inversion, unequal correction capacity across components, and a persistence-based initialization form a sensible architecture around the properties of this dataset. These are established ingredients, so I view algorithmic originality as moderate rather than foundational. Nevertheless, their arrangement addresses a concrete forecasting difficulty and is more substantive than applying an unchanged generic backbone.

  3. The reported gains are measurable and extend beyond comparison with generic models.** Day-20 EKE RMSE decreases from 0.0411 for TAU to 0.0382, a 7.1% reduction, and from 0.0437 for persistence, a 12.6% reduction. EKE ACC improves from 0.3975 to 0.4194. Importantly, the physics-only variant at rho = 0.97 has RMSE 0.0435, so the full model improves over its own non-learned scaffold by 12.2%. This supports a useful contribution from learned corrections, even though it does not identify the contribution of every architectural choice.

  4. The statistical motivation is supported, with an appropriate practical interpretation.** Recomputing Table 1 gives solenoidal persistence-error shares of approximately 90.6%, 93.6%, 94.8%, and 95.3% at 1, 5, 10, and 20 days. Ignoring the reported near-zero uniform term, the remaining accounting discrepancy is approximately 0.4-0.9% of total error. These values support approximate separation and allocating most capacity to the solenoidal branch. They do not prove independence or exact orthogonality, but they provide useful empirical motivation.

Critical weaknesses
  1. The physical interpretation is stronger than the numerical formulation establishes.** Equations (2)-(4) describe planar spectral operators, whereas the application uses a masked latitude-longitude grid and a cos(latitude)-weighted loss. The treatment of spherical metric factors, latitude boundaries, and land filling is not sufficiently explicit. Orthogonality of a planar FFT projection does not automatically survive the change of inner product or masking. I verified this distinction with a periodic analytic example: an unweighted cross inner product numerically equal to zero becomes 1/3 under cosine weighting. This is a counterexample to an automatic guarantee, not evidence that the submitted implementation produces that error. The small discrepancy in Table 1 suggests the approximation may work well in practice. The paper should distinguish that empirical success from exact physical claims.

  2. The sources of improvement are only partly identified.** The physics-only comparison establishes that learning helps, but does not show how much the decomposition itself improves over a similarly initialized TAU-style velocity-residual predictor. Likewise, the comparison with rho = 0.8 and 1.1 omits rho = 1, the most natural no-damping control. At day 20 these settings have RMSEs 13.3 and 1.8 times the default, respectively, showing substantial sensitivity. A small set of targeted controls would clarify whether the benefit comes primarily from representation, initialization, damping, or their interaction. A direct HelmFluid comparison would be informative, but is not the sole basis for judging originality because its original task and implementation differ.

  3. The energy-stability diagnostics need a narrower interpretation.** For a hypothetical constant normalized energy error e(t) = 0.5, both reported variation measures are zero despite a persistent 50% error. Thus, these metrics measure variation across lead time, not energy fidelity by themselves. The EKE RMSE results still provide meaningful evidence of accuracy. However, the use of velocity clipping during rollout makes the clipping rate and its influence relevant to the stability interpretation; clipping is applied across models, so its presence alone does not establish an unfair comparison.

  4. Some technical statements and selection details require correction or clarification.** Small cross terms do not imply statistical independence; Eq. (8) is optimal within scalar-multiple predictors, not arbitrary linear predictors. Appendix A's Eq. (14) also appears inconsistent with the earlier definition of E_irr as persistence error. Finally, the text should reconcile estimating rho from historical statistics with the test-set comparison described in Appendix C. I would not infer test leakage from the wording alone. These issues limit the theoretical explanation and reproducibility, but do not by themselves invalidate the reported forecasting improvements.

Questions for authors
  1. Which spherical metric factors, spatial units, boundary conditions, and land-filling procedure are used? Can you report divergence/curl and weighted orthogonality residuals for the actual masked implementation?
  2. How does the model compare with a matched velocity-residual predictor using the same backbone, initialization, seasonal inputs, and training budget? What happens for rho = 1 or a learned rho?
  3. What fraction of predicted ocean points is clipped at each lead, and how do unclipped EKE errors and failure rates compare? Can mean energy bias accompany the variation diagnostics?
  4. Were rho, climatology, and all other design choices fixed without consulting test outcomes? Can repeated runs or a temporally appropriate uncertainty estimate establish the reliability of the improvement over TAU?
6 comments
AuthorsOpenReview ↗

Response to Reviewer hPzN [1/4]

Thank the reviewer for kind review and the valuable feedback!

[Response to Q1]: We answer the four asked items and then report the residuals for the masked implementation.

Metric factors and spatial units. In the revised manuscript, on Page 15, Line 792, we state that "The spectral operators use a uniform reference zonal spacing dxdx given by the mean of the local spacing Rcos⁡ϕΔλR\cos\phi\Delta\lambda over the latitude grid, and the meridional spacing dy=RΔϕdy = R\Delta\phi." On Page 15, Line 794, we state that "Velocities are in m s$^{-1}$ and vorticities and divergences in s$^{-1}$."

Boundary conditions and land filling. On Page 15, Line 787, we state that "Longitude is periodic by construction. The latitude boundaries bound land and sea ice, where the flow is already near zero, and are zero-padded. Before the FFT, land points are initialized to zero and updated by Laplace relaxation (five iterations, with ocean values fixed), producing a smooth extrapolation across the land-sea boundary. The ocean mask is re-applied after the inversion, so the filled values never enter the forecast or the evaluation."

Divergence and curl residuals. Let usol\mathbf{u}_ {\mathrm{sol}} and uirr\mathbf{u}_ {\mathrm{irr}} denote the divergence-free and the curl-free components. In our own verification, the spectral divergence of usol\mathbf{u}_ {\mathrm{sol}} and the spectral curl of uirr\mathbf{u}_ {\mathrm{irr}} on the filled periodic domain vanish with relative residuals below 10−1010^{-10}. Table 1 reports the finite-difference residuals on the ocean points of the masked grid with ∥ϕ∥≥5∘\|\phi\|\ge 5^{\circ}, over the daily fields spanning 2000 to 2009. The raw field is the geostrophic velocity before the decomposition and serves as the reference level.

Table 1. Finite-difference readings on the ocean points with ∥ϕ∥≥5∘\|\phi\|\ge 5^{\circ}, averaged over the daily fields spanning 2000 to 2009.

Weighted RMS (s$^{-1}$)Pointwise max (s^{-1}\)
FD divergence of the raw field6.36×10−76.36\times10^{-7}5.01×10−55.01\times10^{-5}
FD divergence of usol\mathbf{u}_ {\mathrm{sol}}6.43×10−76.43\times10^{-7}3.46×10−53.46\times10^{-5}
FD curl of uirr\mathbf{u}_ {\mathrm{irr}}3.17×10−73.17\times10^{-7}2.33×10−52.33\times10^{-5}

The readings are dominated by the discretization error of the finite-difference operator itself, so the divergence reading of usol\mathbf{u}_ {\mathrm{sol}} sits at the level of the raw field. When the equatorial band is included, the weighted RMS is 7.51×10−77.51\times10^{-7} s$^{-1}$ for usol\mathbf{u}_ {\mathrm{sol}} against 8.74×10−78.74\times10^{-7} s$^{-1}$ for the raw field. In short, the divergence residual of the decomposition therefore stays at the level already present in the raw data, and the curl reading of uirr\mathbf{u}_ {\mathrm{irr}} stays at the same numerical level.

Weighted orthogonality residuals. Table 2 reports the normalized cross inner products ⟨a,b⟩w/(∣a∣w∣b∣w)\langle\mathbf{a},\mathbf{b}\rangle_ {w}/(\\|\mathbf{a}\\|_ {w}\\|\mathbf{b}\\|_ {w}) under the masked, latitude-weighted inner product, with w=cos⁡ϕw=\cos\phi and the sum over the valid ocean grid points, on the same daily fields.

Table 2. Normalized cross inner products of the three components under the masked, latitude-weighted inner product. The value is zero under orthogonality. The energy share of a pair, a\mathbf{a} and b\mathbf{b}, is 2⟨a,b⟩w/⟨u,u⟩w2\langle\mathbf{a},\mathbf{b}\rangle_ {w}/\langle\mathbf{u},\mathbf{u}\rangle_ {w} in percent, with the sign retained. (Table S1 in the revised manuscript.)

Pairmeanrmsmax abs.share (%)
solenoidal, irrotational+0.020+0.0200.0200.0200.0360.036+0.81+0.81
solenoidal, uniform flow+0.002+0.0020.0080.0080.0270.027+0.05+0.05
irrotational, uniform flow−0.071-0.0710.0720.0720.0980.098−0.16-0.16
net (three pairs)+0.70+0.70

On Page 15, Line 780, we state that "Table S1 reports its statistics for the three component pairs on the daily fields spanning 2000 to 2009, and every daily value stays below 0.100.10 in absolute value." and "The three components are therefore approximately orthogonal on the masked and latitude-weighted grid."

Your periodic analytic example confirms that an unweighted orthogonality does not transfer under cosine weighting. In the revised manuscript, on Page 3, Line 124, we state that "Though the inner product may change on the masked and latitude-weighted grid used for training and evaluation, the components of the AVISO/DUACS geostrophic currents remain approximately orthogonal for component-wise forecasting." In our case, the weighted orthogonality is reported as a measured approximation. On Page 15, Line 767, we update that "Under the masked, latitude-weighted inner product of Eq. (11) the cross inner products are small but not exactly zero." This could also answer the concerns in W1.

AuthorsOpenReview ↗

Response to Reviewer hPzN [2/4]

[Response to Q2]: We answer the two parts in turn.

Matched velocity-residual predictor. In the revised manuscript we submitted, on Page 17, Line 913, we state that "In Table S4, we isolate the role of the Helmholtz structure with a residual control. The control keeps the backbone and the training configuration of HelmGeoNet, removes the Helmholtz decomposition and the relaxation mechanism, and predicts the velocity residual around persistence with the TAU modules alone." The control shares the three-level encoder-translator-decoder with TAU translator blocks, the zero-initialized final projection, the seasonal channels, and the parameter budget. Table 1 lists the day-20 results.

Table 1. Day-20 EKE anomaly correlation (ACC) and latitude-weighted RMSE on the test set spanning 2023 to 2024. The best value per column is bolded. (Table S4 in the revised manuscript.)

ModelEKE ACCEKE RMSE (m$^2$/s$^2$)
TAU (direct prediction)0.39750.0411
Residual control0.38100.0390
HelmGeoNet0.41940.0382

The control improves the day-20 EKE RMSE over the direct TAU baseline, from 0.04110.0411 to 0.03900.0390 m$^2$/s$^2$, and its EKE ACC of 0.38100.3810 stays below the TAU value of 0.39750.3975. On Page 17, Line 917, we state that "Residual learning with zero initialization thus accounts for much of the gain, mainly in EKE RMSE. The three-branch structure and the relaxation mechanism of HelmGeoNet provide the remaining gain in EKE ACC."

ρ=1\rho=1 and a learned ρ\rho. Both are added in the revised manuscript. On Page 8, Line 385, we state that "As the most natural no-damping control, ρ=1.0\rho=1.0 also gives a stable rollout, with a day-20 EKE RMSE of 0.03890.0389 m$^2$/s$^2$ and an EKE ACC of 0.42800.4280." On Page 8, Line 387, we state that "A control that learns ρ\rho as a free parameter converges to ρ≈0.993\rho\approx0.993, and every metric stays within 0.0030.003 of the fixed setting, with the better day-20 EKE forecasts at ρ=0.97\rho=0.97." Table 2 lists the full sensitivity study.

Table 2. Sensitivity of HelmGeoNet to the divergence damping coefficient ρ\rho. EKE ACC and latitude-weighted RMSE at leads 5, 10, and 20 days on the test set spanning 2023 to 2024. The best value per column is bolded and the second-best is italicized, with tied best values both bolded. (Table 2 in the revised manuscript.)

lead 5d ACClead 5d RMSElead 10d ACClead 10d RMSElead 20d ACClead 20d RMSE
ρ=0.8\rho=0.80.86640.01880.44450.07100.03750.5077
ρ=0.97\rho=0.970.87000.01850.68720.02830.41940.0382
ρ=1.0\rho=1.00.87320.01840.69470.02840.42800.0389
ρ=1.1\rho=1.10.87050.01870.67750.02970.25270.0683
learned ρ\rho (≈0.993\approx 0.993)0.87200.01840.68970.02820.41660.0383

On Page 8, Line 390, we state that "within this range the corrector compensates for the imposed damping, so the differences in skill remain small at all leads. We keep ρ=0.97\rho=0.97 because it is fixed from training data alone and keeps the rollout away from the unstable settings."

Comparison with HelmFluid. HelmFluid is an impressive work and the manuscript cites it. Our work applies the Helmholtz decomposition to the evolution of geostrophic currents specially, so that each component learns its own evolution and the kinetic-energy forecast becomes more accurate and stable. The two frameworks differ in the input structure. Our framework consumes one velocity field per step and forecasts the next field, while HelmFluid consumes a sequence of historical frames to capture the evolution, so a direct comparison requires adapting one of the two input protocols, and we leave it to future work. This could also answer the concerns in W2.

AuthorsOpenReview ↗

Response to Reviewer hPzN [3/4]

[Response to Q3]: We answer the five asked items in turn.

Clipping fraction. For HelmGeoNet, the count of 18,361 clamping events is accumulated over the 711 daily initializations of the two test years, each rolled out to 20 leads, counting every grid point at every lead once. The count rises from 3 events at lead 1 to 2,235 events at lead 20, about 3 grid points per rollout at the worst lead. All events lie inside the excluded equatorial band, at latitudes between 0.625∘0.625^{\circ}N and 4.625∘4.625^{\circ}N.

Unclipped comparison. Table 1 lists the day-20 EKE skill with and without the clamp for HelmGeoNet and for SimVP, the baseline with the largest change.

Table 1. Day-20 EKE skill on the 2023 to 2024 test set with and without the velocity clamp.

ModelEKE ACC (with clamp)EKE RMSE (with clamp)EKE ACC (without clamp)EKE RMSE (without clamp)
HelmGeoNet0.41940.03820.41950.0381
SimVP0.34220.05200.33930.1384

HelmGeoNet and the other baselines stay essentially unchanged with or without the clamp, with SimVP the exception shown in the table. On Page 9, Line 460, we state that "Notably, HelmGeoNet stays ahead of every baseline with or without the clamp. Removing the clamp gives a day-20 EKE ACC of 0.41950.4195 and an EKE RMSE of 0.03810.0381 m$^2$/s$^2$."

Mean energy bias. Yes. On Page 7, Line 345, we state that "measure the variation of e(t)e(t) across lead time, and we also report the mean normalized error eˉ\bar{e} ... to assess the energy fidelity." Table 2 lists the mean and the maximum absolute normalized energy error of every model over the rollout.

Table 2. Mean and maximum absolute normalized energy error over the 20-lead rollout, in percent. The maximum absolute error is reported with the sign of the error at the lead where the absolute error is largest. (Table S6 in the revised manuscript.)

ModelMean error (%)Max abs error (signed, %)
HelmGeoNet−4.8-4.8−5.7-5.7
ConvLSTM+13.3+13.3+36.5+36.5
SwinLSTM+4.6+4.6+14.3+14.3
TAU+3.6+3.6+6.1+6.1
SimVP+1.2+1.2+3.3+3.3
PredFormer−2.9-2.9+12.9+12.9
VideoMamba+4.7+4.7+26.4+26.4
FourCastNet+0.8+0.8+11.3+11.3
Earthformer−5.0-5.0−7.0-7.0

On Page 18, Line 967, we state that "SimVP is the only baseline that attains both a smaller mean error and a smaller maximum absolute error, yet its day-20 EKE RMSE of 0.05200.0520 m$^2$/s$^2$ far exceeds the 0.03820.0382 m$^2$/s$^2$ of HelmGeoNet." and "HelmGeoNet keeps both errors steady across leads and stays below 66\\% in absolute error at every lead, so it avoids the large energy drifts and swings of the baselines."

[Response to Q4]: We answer the two asked items in turn.

Fixed design choices. Yes for the two named choices. On Page 5, Line 263, we state that "The damping coefficient ρ\rho is set to 0.970.97 according to the historically measured one-day autocorrelation ρ(1)=0.9717\rho(1)=0.9717 of Eq. (8), computed on 2000 to 2009, a period entirely inside the training range." On Page 7, Line 331, we state that "where uclim\mathbf{u}^{\mathrm{clim}} is computed from the training period spanning 2000 to 2020." The 2023-2024 test set are used exclusively for evaluation and are not consulted at any stage of model development.

Reliability over TAU. We have not added independent training runs in this revision, since they exceed our current compute budget. We quantify the variability over the 711 daily initializations with a paired block bootstrap. On Page 18, Line 931, we state that "A paired block bootstrap over the test initializations uses a block length of 30 days and 2000 resamples. The 9595\\% confidence interval for the day-20 EKE RMSE difference to TAU runs from −0.0034-0.0034 to −0.0024-0.0024 m$^2$/s$^2$." and "Both comparisons give p<0.001p < 0.001." and "The ranking of the three forecasts is the same in each of the two test years, with no reversal at any lead." We will actively pursue multi-seed training in an updated version. This could also answer the concerns in W3.

Reviewer_yEjROpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper forecasts daily surface geostrophic currents by separating the velocity into solenoidal, irrotational, and uniform components. HelmGeoNet learns a vorticity correction for the first component, applies damped persistence with a learned potential correction to the second, and persists the mean flow. The motivation is that these components have different temporal statistics. Experiments on AVISO/DUACS currents evaluate 20-day autoregressive forecasts, with particular attention to eddy kinetic energy (EKE).

Critical strengths
  1. The component-specific models have an empirical motivation. Table 1 uses 2000–2009 currents to show that most evolution error lies in the solenoidal component, while damped persistence approximates the divergent component. This is a concrete reason to model the components differently.
  2. The EKE results stand out: day-20 RMSE is 0.0382, compared with 0.0411 for TAU and 0.0437 for persistence (Table 3). Several learned baselines improve velocity error without improving EKE, so this gain is worth examining.
Critical weaknesses
  1. The exactness claims for the decomposition are difficult to reconcile with the data domain. Equation (4) gives a Cartesian Fourier construction that is exact on a doubly periodic domain; the currents occupy a masked longitude–latitude grid. Appendix A mentions filling and masking without specifying the procedure, and claims orthogonality under the ocean-only, latitude-weighted inner product. That orthogonality does not follow from Equation (4). Section 4.2 and Figure 2 also retain nonzero cross terms. The paper needs to state which properties survive the masking and metric treatment, since its physical interpretation relies on this separation.
  2. The TAU comparison does not isolate the decomposition. HelmGeoNet uses TAU-style correctors with zero-initialized output projections to learn residuals around persistence (Section 3.4). These choices are not held fixed in the baseline. The damping study and physics-only results in Appendix C do not replace a comparison with the same backbone learning velocity residuals directly. Since TAU is already the strongest learned EKE baseline, that is an important control.
Questions for authors
  1. How are land points and latitude boundaries handled by the Fourier operators, and does orthogonality still hold under the ocean-only, latitude-weighted inner product?
  2. How does HelmGeoNet compare with the same backbone predicting velocity residuals around persistence, using the same initialization and training settings?
3 comments
AuthorsOpenReview ↗

Response to Reviewer yEjR [1/2]

Thank the reviewer for the valuable feedback!

[Response to Q1]: We answer the two parts of the question in turn.

Handling of land points and latitude boundaries. Land points are filled by Laplace relaxation before the FFT and the ocean mask is re-applied after the inversion, the latitude boundaries are zero-padded, and longitude is periodic by construction. In the revised manuscript we submitted, on Page 15, Line 788, we state that "Before the FFT, land points are initialized to zero and updated by Laplace relaxation (five iterations, with ocean values fixed), producing a smooth extrapolation across the land-sea boundary. The ocean mask is re-applied after the inversion, so the filled values never enter the forecast or the evaluation. Longitude is periodic by construction. The latitude boundaries bound land and sea ice, where the flow is already near zero, and are zero-padded." On Page 15, Line 792, the metric treatment reads "The spectral operators use a uniform reference zonal spacing dxdx given by the mean of the local spacing Rcos⁡ϕΔλR\cos\phi\Delta\lambda over the latitude grid, and the meridional spacing dy=RΔϕdy = R\Delta\phi."

Orthogonality under the ocean-only, latitude-weighted inner product. Strictly, no. Orthogonality holds only on the filled periodic domain on which the spectral operators are defined. Under the ocean-only, latitude-weighted inner product it holds approximately, with an energy departure below one percent that the revised manuscript quantifies. We summarize the measured properties of the decomposition in three points.

  • On the filled periodic domain, the orthogonality holds. In our double-precision verification, the spectral divergence of usol\mathbf{u}_ {\mathrm{sol}} and the spectral curl of uirr\mathbf{u}_ {\mathrm{irr}} vanish with relative residuals below 10−1010^{-10}, and the orthogonality of the three components in the L2L^2 inner product of that domain follows.
  • On the masked grid, the orthogonality is approximate, and the departure is small. Under the inner product ⟨a,b⟩w=∑oceanwa⋅b/∑oceanw\langle\mathbf{a},\mathbf{b}\rangle_ {w}=\sum_ {\mathrm{ocean}} w\mathbf{a}\cdot\mathbf{b}/\sum_ {\mathrm{ocean}} w with w=cos⁡ϕw=\cos\phi, where the sum runs over the valid ocean grid points, the normalized cross inner products of the three component pairs stay below 0.100.10 in absolute value on every day of 2000 to 2009. Table 1 lists the statistics and the energy shares over the global ocean.
  • In the evolution energy, the cross terms stay small. On Page 15, Line 807, we state that "Across lags 1 to 20 the net contribution of all cross terms in Eq.(12) stays between 0.40.4\\% and 0.90.9\\% of ErawE_ {\mathrm{raw}}, so the evolution energy is carried almost entirely by the component-wise terms."

Table 1. Normalized cross inner products of the three components under the masked, latitude-weighted inner product, computed on the daily fields spanning 2000 to 2009. The value is zero under orthogonality. The energy share of a pair is 2⟨a,b⟩w/⟨u,u⟩w2\langle\mathbf{a},\mathbf{b}\rangle_ {w}/\langle\mathbf{u},\mathbf{u}\rangle_ {w} in percent, with the sign retained. (Table S1 in the revised manuscript.)

Pairmeanrmsmax absolute valueshare of total energy (%)
solenoidal, irrotational+0.020+0.0200.0200.0200.0360.036+0.81+0.81
solenoidal, uniform flow+0.002+0.0020.0080.0080.0270.027+0.05+0.05
irrotational, uniform flow−0.071-0.0710.0720.0720.0980.098−0.16-0.16
net (three pairs)+0.70+0.70

The negative entry of the irrotational-uniform pair is not anomalous. Each entry in the last column is the signed contribution of a cross term to the total energy, and a cross term measures the overlap of two components, so it carries a sign. The weighted inner product of the irrotational component and the uniform flow is negative on average. The nonzero cross terms in Section 4.2 and Figure 2 of the manuscript are this departure measured in the evolution energy. In summary, the orthogonality under the ocean-only, latitude-weighted inner product is approximate, with an energy departure below one percent that the revision quantifies. This could also answer the concerns in W1.

AuthorsOpenReview ↗

Response to Reviewer yEjR [2/2]

[Response to Q2]: We agree that the damping study and the physics-only results of Appendix C do not isolate the contribution of the residual formulation, and the revised manuscript adds the control that does. On Page 17, Line 913, we state that "In Table S4, we isolate the role of the Helmholtz structure with a residual control. The control keeps the backbone and the training configuration of HelmGeoNet, removes the Helmholtz decomposition and the relaxation mechanism, and predicts the velocity residual around persistence with the TAU modules alone." The control shares with HelmGeoNet the three-level encoder-translator-decoder with TAU translator blocks, the zero-initialized final projection, and the parameter budget. Table 2 compares the direct TAU baseline, the residual control, and HelmGeoNet.

Table 2. Day-20 EKE forecast skill on the test set spanning 2023 to 2024. EKE anomaly correlation (ACC) and latitude-weighted RMSE. The best value per column is bolded. (Table S4 in the revised manuscript.)

ModelEKE ACCEKE RMSE (m$^2$/s$^2$)
TAU (direct prediction)0.39750.0411
Residual control0.38100.0390
HelmGeoNet0.41940.0382

The control improves the day-20 EKE RMSE over the direct TAU baseline, from 0.04110.0411 to 0.03900.0390 m$^2$/s$^2$, so the residual formulation with zero initialization accounts for a substantial part of the RMSE gain. Its EKE ACC of 0.38100.3810 stays below the TAU value of 0.39750.3975, so the residual formulation alone does not reproduce the correlation gain, and HelmGeoNet attains the best value on both metrics. On Page 17, Line 917, we state that "Residual learning with zero initialization thus accounts for much of the gain, mainly in EKE RMSE." In summary, with the backbone, the zero initialization, and the parameter budget held fixed, the residual control explains most of the EKE RMSE advantage over TAU, and the Helmholtz structure with the relaxation mechanism provides the remaining RMSE gain and the EKE ACC advantage. This could also answer the concerns in W2.

Dear reviewer, we thank you very much for the careful reading and constructive comments on our work. We have posted our responses to the questions and weaknesses, together with the experiments requested during the discussion. We would be grateful for any feedback or further questions on these responses, and we remain available to provide additional clarification during the remaining discussion period.

Comments on the paper

5 comments
AuthorsOpenReview ↗

Response to Reviewer BEDo [1/3]

Thank the reviewer for kind review and the valuable feedback!

[Response to Q1]: In the revised manuscript we submitted, " On Page 15, Line 787, we state that "Longitude is periodic by construction. The latitude boundaries bound land and sea ice, where the flow is already near zero, and are zero-padded. Before the FFT, land points are initialized to zero and updated by Laplace relaxation (five iterations, with ocean values fixed), producing a smooth extrapolation across the land-sea boundary. The ocean mask is re-applied after the inversion, so the filled values never enter the forecast or the evaluation. The spectral operators use a uniform reference zonal spacing dxdx given by the mean of the local spacing Rcos⁡ϕΔλR\cos\phi\Delta\lambda over the latitude grid, and the meridional spacing dy=RΔϕdy = R\Delta\phi."

We assess the possible decomposition artifacts with three checks. First, we quantify the divergence and curl residuals with finite differences over the daily fields spanning 2000 to 2009, on the ocean points with ∥ϕ∥≥5∘\|\phi\|\ge 5^{\circ}. Table 1 lists the reading for the raw geostrophic field before the decomposition, and the readings for the two decomposed components.

Table 1. Finite-difference readings on the ocean points with ∥ϕ∥≥5∘\|\phi\|\ge 5^{\circ}, averaged over the daily fields spanning 2000 to 2009. The raw field is the geostrophic velocity before the decomposition.

Weighted RMS (s$^{-1}$)Pointwise max (s^{-1}\)
FD divergence of the raw field6.36×10−76.36\times10^{-7}5.01×10−55.01\times10^{-5}
FD divergence of usol\mathbf{u}_ {\mathrm{sol}}6.43×10−76.43\times10^{-7}3.46×10−53.46\times10^{-5}
FD curl of uirr\mathbf{u}_ {\mathrm{irr}}3.17×10−73.17\times10^{-7}2.33×10−52.33\times10^{-5}

On the filled periodic domain, the spectral divergence of usol\mathbf{u}_ {\mathrm{sol}} and the spectral curl of uirr\mathbf{u}_ {\mathrm{irr}} vanish to machine precision, with relative residuals below 10−1010^{-10}. The finite-difference readings provide the meaningful test and the finite-difference operator uses the pointwise local spacing. Away from the equator the geostrophic flow is nearly divergence-free, and its divergence content is carried by uirr\mathbf{u}_ {\mathrm{irr}} by construction, so the divergence reading of usol\mathbf{u}_ {\mathrm{sol}} sits at the level of the raw field in Table 1. When the equatorial band, where the geostrophic divergence concentrates and which is excluded from the evaluation, is included, the weighted RMS of the divergence reading is 7.51×10−77.51\times10^{-7} s$^{-1}$ for usol\mathbf{u}_ {\mathrm{sol}} against 8.74×10−78.74\times10^{-7} s$^{-1}$ for the raw field, so the decomposition lowers the divergence level of the data. In short, the divergence residual of the decomposition therefore stays at the level already present in the raw data, and the curl reading of uirr\mathbf{u}_ {\mathrm{irr}} stays at the same numerical level.

Second, we quantify the mixing among the solenoidal, irrotational, and uniform components on the same grid. The accounting identity of Eq. (12) in the revised manuscript splits the total persistence error into the errors of the three components and their cross terms,

Eraw(Δ)=Esol(Δ)+Eirr(Δ)+Eunif(Δ)+2∑a<b⟨Δxa,Δxb⟩w,E_ {\mathrm{raw}}(\Delta)=E_ {\mathrm{sol}}(\Delta)+E_ {\mathrm{irr}}(\Delta)+E_ {\mathrm{unif}}(\Delta)+2\sum_ {a < b}\langle\Delta\mathbf{x}_ {a},\Delta\mathbf{x}_ {b}\rangle_ {w},

where Eraw(Δ)=⟨xΔ−x0,xΔ−x0⟩wE_ {\mathrm{raw}}(\Delta)=\langle\mathbf{x}_ {\Delta}-\mathbf{x}_ {0},\mathbf{x}_ {\Delta}-\mathbf{x}_ {0}\rangle_ {w} with x0=x(t)\mathbf{x}_ {0}=\mathbf{x}(t) and xΔ=x(t+Δ)\mathbf{x}_ {\Delta}=\mathbf{x}(t+\Delta), EsolE_ {\mathrm{sol}}, EirrE_ {\mathrm{irr}}, EunifE_ {\mathrm{unif}} are the solenoidal, irrotational, and uniform persistence errors, and Δxa\Delta\mathbf{x}_ {a} denotes the persistence increment of component aa. On Page 15, Line 807, we state that "Across lags 1 to 20 the net contribution of all cross terms in (12) stays between 0.40.4\\% and 0.90.9\\% of ErawE_ {\mathrm{raw}}". An artifact of the decomposition on the weighted grid would enlarge these cross terms, and their share stays below one percent at every lag, so the three components remain well separated in the energy accounting.

Third, we test the robustness to the filling procedure. We decompose each original daily field and sum the three recovered components back into a single velocity field. When the number of relaxation iterations varies from 0 to 10, the RMSE between this reconstructed field and the original field changes by less than 0.70.7\\%, so the decomposition is insensitive to the filling strength.

Taken together, the three checks show that the spectral decomposition and the land masking introduce no artifacts beyond the level of the raw data, and that the decomposition is robust to the filling procedure. This could also answer the concerns in W2.

AuthorsOpenReview ↗

Response to Reviewer BEDo [2/3]

[Response to Q2]: For HelmGeoNet, the count of 18,361 grid-point events is accumulated over the full test set, over the 711 daily initializations of the two test years, each rolled out to 20 days, and every grid point at every lead is counted once. The count rises from 3 events at lead 1 to 2,235 events at lead 20, which averages to about 3 grid points per rollout at the worst lead. All of them lie inside the excluded equatorial band, at latitudes between 0.625∘0.625^{\circ}N and 4.625∘4.625^{\circ}N, so no evaluation point is clamped.

Yes, the main results remain consistent without clamping. In Table 2, we report the day-20 EKE skill with and without the clamp for HelmGeoNet and SimVP.

Table 2. Day-20 EKE skill on the 2023 to 2024 test set with and without the velocity clamp (EKE RMSE in m$^2$/s$^2$).

ModelEKE ACC (with clamp)EKE RMSE (with clamp)EKE ACC (without clamp)EKE RMSE (without clamp)
HelmGeoNet0.41940.03820.41950.0381
SimVP0.34220.05200.33930.1384

The other baselines also stay essentially unchanged with or without the clamp. On Page 9, Line 461, we state that "Notably, HelmGeoNet stays ahead of every baseline with or without the clamp. Removing the clamp gives a day-20 EKE ACC of 0.41950.4195 and an EKE RMSE of 0.03810.0381 m$^2$/s$^2$. For HelmGeoNet, all clamping events lie inside the excluded equatorial band, where geostrophic balance fails and the derived velocities carry large errors, so a few such points can grow to unphysical magnitudes during the rollout." This could also answer the concerns in W4.

[Response to Q3]: For a separate temporal period, yes. The test set spans 2023 to 2024, with training on 2000 to 2020 and validation on 2021 to 2022, so all reported skill is measured on years unseen during training. Within the test period, the year-wise evaluation added in the revised manuscript (Table S5) shows that HelmGeoNet outperforms the representative baseline at every lead in each of the two test years (see the response to Q5). For an independent ocean-current product, we have not evaluated in this revision, since retraining the model and all baselines under the same protocol exceeds our current compute budget. We will actively pursue this evaluation in an updated version. This limitation is stated in the conclusion. On Page 10, Line 535, the conclusion states that "Third, the present evaluation is limited to the current resolution and altimetry product. Extending it to higher-resolution datasets and independent altimetry products would provide a broader assessment of the model's generalizability and robustness." This could also answer the concerns in W1.

[Response to Q4]: The eight baselines differ widely in their standard configurations, since several of them were built as large Earth-system forecasters. The original FourCastNet, for example, forecasts a large set of atmospheric variables on multiple pressure levels and has hundreds of millions of parameters. Using such a configuration for a single-level single-variable ocean field would leave far more capacity than the task requires and would risk overfitting. We therefore scale all baselines to a shared budget of roughly four million parameters, which generally fits our forecasting task and the characteristics of our dataset. It also keeps the comparison across the eight backbones focused on the architectures. The baselines at their standard configurations have not been evaluated in this revision, since the retraining exceeds our current compute budget. We will actively pursue this comparison in an updated version. This could also answer the concerns in W3.

AuthorsOpenReview ↗

Response to Reviewer BEDo [3/3]

[Response to Q5]: Year-wise performance, yes. The revised manuscript adds Table S5 in Appendix D. We reproduce it here as Table 1.

Table 1. Year-wise EKE forecast skill for each of the two test years separately. EKE anomaly correlation (ACC) and latitude-weighted RMSE at leads 5, 10, and 20 days. The best value per column is bolded. (Table S5 in the revised manuscript.)

YearModellead 5d ACClead 5d RMSElead 10d ACClead 10d RMSElead 20d ACClead 20d RMSE
2023HelmGeoNet0.87120.01860.68950.02850.42820.0383
2023TAU0.86900.01940.68510.03050.40850.0412
2023persistence0.81050.02310.56580.03490.32650.0435
2024HelmGeoNet0.86880.01840.68470.02820.41000.0380
2024TAU0.86540.01910.67700.03010.38600.0410
2024persistence0.79870.02330.54180.03510.28370.0439

HelmGeoNet stays ahead of both baselines at every lead in both years. For variability of training runs, we have not added independent training runs in this revision, since they exceed our current compute budget. We will actively pursue independent training in an updated version. We quantify the variability of the skill over the 711 daily initializations of the two test years with a paired block bootstrap. On Page 18, Line 931, "A paired block bootstrap over the test initializations uses a block length of 30 days and 2000 resamples. The 9595\\% confidence interval for the day-20 EKE RMSE difference to TAU runs from −0.0034-0.0034 to −0.0024-0.0024 m$^2$/s$^2$. The interval for the difference to persistence runs from −0.0059-0.0059 to −0.0052-0.0052 m$^2$/s$^2$. Both comparisons give p<0.001p < 0.001. The 9595\\% confidence interval for the day-20 EKE ACC difference runs from 0.01530.0153 to 0.02880.0288 against TAU, and from 0.10390.1039 to 0.12530.1253 against persistence. The ranking of the three forecasts is the same in each of the two test years, with no reversal at any lead." This could also answer the concerns in W5.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.