acceptodds
Under review as a conference paper at ICLR 2027

PE-TA-ViT: Dual-Resolution Residual Forecasting of Global Surface and Near-Surface Ocean States

Abstract

Accurate and computationally efficient forecasting of global surface and near-surface ocean states remains challenging in contemporary oceanography. High-resolution numerical models can better resolve mesoscale dynamics but incur prohibitive computational costs, whereas coarse-resolution models often suffer from biases associated with unresolved subgrid-scale processes. Here, we propose a dual-resolution deep learning framework for forecasting global surface and near-surface ocean states, with near-surface variables represented at 0.49 m depth. The framework adopts a residual reconstruction paradigm in which a Physics-Enhanced Time-Aware Vision Transformer (PE-TA-ViT), incorporating physics-guided inductive biases and adaptive temporal conditioning, predicts coarse-scale spatiotemporal increments at resolution, which are subsequently interpolated to and added to the evolving high-resolution background state. Over a 9-day forecast horizon, PE-TA-ViT reduces the average Root Mean Square Error (RMSE) relative to persistence by 6.0% to 24.8% across the evaluated ocean variables. It also extends the useful forecast window (ACC 0.6) for highly dynamic velocity fields by approximately one day. The resulting forecasts retain transient mesoscale structures and large-scale flow patterns, with the deviation in the global energy cascading fraction remaining below 0.009. This framework enables efficient high-resolution forecasting of global surface and near-surface ocean states.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Physics-Informed Deep Learning, Ocean State Forecasting, Sea Surface Temperature, Sea Surface Height

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_tgqXOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes PE-TA-ViT for global surface and near-surface ocean forecasting. The framework predicts daily increments at 1° resolution, bilinearly interpolates them to 1/4°, and adds them to the previous high-resolution state. The paper reports improvements over persistence and several learned baselines over a nine-day forecast horizon, together with vorticity, geostrophic-current, and spectral diagnostics.

Critical strengths
  1. The application is meaningful, and the framework is straightforward to understand and offers a plausible engineering approach to practical ocean forecasting problem.
  2. Promising reported forecasting results. Tables 1–2 show improvements over persistence across all six ocean variables, while Table 4 includes multiple learned baselines and GLONET. The evaluation also considers spatial and physical diagnostics beyond aggregate RMSE, which is useful for this application (Sec. 4; Appendix D).
Critical weaknesses
  1. While the application is meaningful, and the framework offers a plausible engineering approach. However, I am not convinced that the submission establishes a sufficient ML methodological contribution and provides sufficiently controlled evidence for its central claims. In particular, the experiments do not adequately distinguish preservation of initial fine-scale structure from successful prediction of its subsequent evolution. The work currently looks more like a specialized ocean-forecasting application than as a sufficiently validated contribution to machine learning community.
  2. Preserving high-resolution structure is not sufficient evidence of forecasting its evolution. Under Eqs. (1)–(3), every update to the high-resolution state is an interpolated coarse-grid increment. Consequently, changes remain restricted to the range of this fixed interpolation-and-masking operator, while components outside that range are inherited from the initial condition. Although Sec. 5 acknowledges limitations in generating new eddies, the issue also concerns the evolution of existing structures.
  3. Temporal cross-attention may reduce to simple additive conditioning. If each temporal embedding in Eq. (10) is a single key/value token, its attention weight is always one, making the output independent of the spatial query. Since the gating weights also depend only on time, TAFEv2 would reduce to a shared, date-dependent additive correction. The authors should clarify token dimensions and explain what this adds beyond standard temporal conditioning (Sec. 3.2.3; Eqs. 10–13). Please also see Question 2 below.
Questions for authors
  1. Can the authors quantify forecasting skill at different spatial scales and compare eddy displacement or temporal vorticity changes against persistence and simpler residual predictors?

  2. Can the authors provide individual module-removal ablations and compare TAFEv2 with simple additive calendar embeddings? In Eq. (10), does each temporal embedding constitute a single key/value token? If so, standard attention over one key has weight one and is independent of the spatial query. What additional functionality does the cross-attention formulation provide?

  3. How does Eq. (24) support a claim about cross-scale energy transfer rather than spectral-content retention?

  4. What causes the 33.792°C SST error of the direct-interpolation baseline? Does the advantage remain after using mask-aware interpolation and evaluating both approaches on the same coastal and offshore subsets?

Reviewer_gWg9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes a framework for forecasting surface ocean states, including sea surface temperature (SST), sea surface salinity, sea surface height (SSH), and velocity fields (i.e., U and V), using OSTIA SST, GLORYS12 ocean state variables (reanalysis data), and ERA5 atmospheric variables (wind field). The framework consists of two main components: (1) the Physics-Enhanced Time-Aware Vision Transformer (PE-TA-ViT), which predicts large-scale dynamical increments based on 1-degree inputs, and (2) a Dual-Resolution Residual Reconstruction mechanism, which projects these increments onto a higher-resolution ocean state. PE-TA-ViT consists of a multi-scale branch (i.e., two different kernel sizes) and a physics branch (gradients initialized with Sobel). The authors evaluate the proposed framework against persistence and several deep-learning baselines, including U-Net, Swin-Unet, ResNet, ConvLSTM, MUDDFormer, and GLONET.

Critical strengths

The paper proposes framework (Figure 2) that consists of (1) PEM-PE module; (2) Cross-Variable Aggregation; and (3) TAFEv2 module. The innovation of the framework employs a dual spatial resolution approach. It autoregressively evolve low resolution input and add the fixed high-resolution initial input to do the forecasting. The authors have demonstrated that their model can outperform persistence and some models on 0.25-degree sea surface tempeture, salinity, sea level high, and velocity field forecasting on a 9-day horizon.

Critical weaknesses
  1. Methodological innovation The proposed framework (Figure 2) consists of (1) PEM-PE module; (2) Cross-Variable Aggregation; and (3) TAFEv2 module. First, the PEM-PE module consists of multi-scale features (2 spatial kernels) and Sobel-initialized gradient filters to extract spatial-gradient features. Although this design may introduce a useful gradient-related inductive bias, it is not clear that it constitutes a meaningful incorporation of physical laws or ocean dynamics. The authors should clarify what physical knowledge is encoded in this module and how it differs from conventional multiscale feature extraction. The term “physics-enhanced” may overstate the extent to which the model is physically constrained. Second, the Dual-Resolution Residual Reconstruction mechanism is presented as the main contribution of the paper. The approach assumes that the temporal evolution of the high-resolution ocean state can be represented by a predicted low-resolution increment, which is then interpolated and added to the previous high-resolution state (equations 1-3). For a 9-day forecast with a spatial resolution of 0.25 degree, it is unclear whether this low-resolution representation is sufficient to capture the relevant temporal evolution, particularly for mesoscale and other fine-scale ocean features. The authors should provide stronger justification for this assumption and clarify the methodological distinction between their approach and conventional residual forecasting or coarse-to-fine reconstruction.

  2. Forecasting challenge and the proposed modeling strategy The framework uses assimilated ocean state estimates from GLORYS12, together with OSTIA SST and ERA5 atmospheric variables, to forecast ocean states over a 9-day horizon. Given the availability of these input fields, it is important to establish how challenging the forecasting task is. According to table 1 (persistence), as expected, it should be easy to forecast these variables on a spatial resolution of 0.25 degree on a 9-day horizon. Overall, although the paper presents a potentially useful deep-learning framework for ocean-state forecasting, the methodological contribution and the necessity of the proposed modeling strategy are not yet sufficiently established.

  3. Other comments In table 3, the performance of Direct on SST surprises me. Theoretically, SST should be easy to predict on a 9-day horizon and a spatial resolution of 0.25 degree.

Questions for authors

None

Reviewer_n2hwOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies efficient global forecasting of surface and near-surface ocean states. It propose PE-TA-ViT, predicts daily increments at 1° resolution and adds their interpolated values to a 1/4° background. The PE_TA-ViT achieves better RMSE with 6% to 24% on a 9 day prediction horizon.

Critical strengths
  1. The approach is simple and has low inference cost.
  2. The experiments cover several ocean variables and include useful ablations.
  3. The data split, normalization, and autoregressive evaluation are clearly described.
Critical weaknesses
  1. The references raise an ethical concern. "Linfei Cao, Jiachen Yang, Meng Xi, Jingyi He, Fei Gao, Jiabao Wen, and Zhijing Wang. M2f-pinn: A multi-scale frequency-domain multi-physics-informed neural network for ocean forecasting. In The Fourteenth International Conference on Learning Representations, 2026. " is listed as an ICLR 2026 publication, but its final decision was Reject.

  2. Appendix D.3–D.4 closely follow GLONET. GLONET is cited in D.1, but the source of the later material is unclear. Please explain what was adapted.

  3. Eq 2-3 make the initial fine-scale residual unchanged during the forecast. This residual cannot update independently. It is unclear how much of the reported detail comes from useful prediction. A comparison with persistence at different spatial scales would help.

  4. Table 1 and Figure 12 report different RMSEs for the proposed PE-TA-ViT, but the persistence values match. It raised my concern about the authenticity of the experiments results.

Questions for authors

Please see in the Weaknesses section for each weakness as a question.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.