acceptodds
Withdrawn from ICLR 2027

EddyCast: Physics-as-Bypass for 3D Mesoscale Eddy Forecasting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Existing physics-informed ocean forecasting often incorporates physical knowledge through loss penalties or prescribed physical structures, which can limit the flexible integration of physical priors with data-driven representations. We propose Physics-as-Bypass, a physical-embedding paradigm for medium-range 3D mesoscale eddy forecasting. Its zero-initialized residual bypass preserves the data-driven baseline at initialization while enabling adaptive integration of physical priors through a learnable pointwise projection. Based on this paradigm, EddyCast uses fixed analytical operators to provide physically grounded guidance, while EddyCast-VMoD employs learnable vertical modal decomposition to quantify prior-data alignment and modal saturation. We further introduce EddyCast-Prob, which equips the shared backbone with an EDM head for probabilistic ensemble forecasting. Experiments show that EddyCast consistently outperforms strong deterministic baselines. Modal ablations and subspace analysis reveal a strong low-rank vertical structure in 3D eddy temperature-salinity fields and show that optimization moves learned representations toward data-aligned physical subspaces. We further explain modal saturation through residual-energy coverage and basis alignment, showing that forecasting-optimal physical dimensionality need not coincide with the effective rank of the data.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Mesoscale eddy forecasting, 3D ocean forecasting, Physical representations

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_P6JZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

EddyCast forecasts daily upper-ocean temperature (10 levels) and salinity (15 levels) in the Northwest Pacific for seven days, trained on GLORYS12V1 reanalysis. A compact 3D convolutional encoder–decoder is extended with a "Physics-as-Bypass" branch. Physical features, either EAG (SST Laplacians at t and t−1 and their difference) or VMoD (learnable vertical modes), pass through a small network and are added to the output through a zero-initialized 1×1×1 projection. An EDM head gives a probabilistic variant. EddyCast beats seven baselines (RMSE 0.0106 vs 0.0118 for WenHai) with about 3M parameters, and with the same EAG features and backbone it beats feeding the features as inputs or as an auxiliary loss.

Critical strengths
  1. Table 2 compares the same EAG features used as input, as loss and as bypass, with the backbone and training fixed. This is cleaner evidence for the integration mechanism than the cross-architecture comparison.
  2. The accuracy–cost trade-off is good: about 3M parameters, 25.7 ms for a seven-step rollout, and the lowest PhysRMSE at every lead.
Critical weaknesses
  1. The task is not forecasting as presented. Per App. A.1 and Table 3, every rollout step receives the target-day SSH, SST, SSS, surface currents and mixed-layer depth, taken from the reanalysis. These are not available when a forecast is issued, and they carry much of the skill: removing them (w/o Surf) gives RMSE 0.0126, worse than the no-physics model (0.0124). The results therefore measure subsurface estimation given future surface truth, not seven-day forecast skill. Persistence does not receive this information.
  2. A physics-specific benefit is not shown. Zero-initialized residual branches are a standard device (e.g. ControlNet, ReZero), and EAG is three simple SST-derived maps, so the gain over input concatenation (0.0114 → 0.0106) may come from the extra trainable branch. There is no parameter-matched generic bypass (the same branch fed learned or random features, or without zero initialization). The vertical analysis is confounded by the channel layout: the "vertical" axis concatenates 10 temperature and 15 salinity levels with different depths (Table 4), and the depth-resolved trust curves in Fig. 3(c, e) jump exactly at layer 10, the temperature/salinity boundary. In Fig. 6 the best model (K = 2) also has the lowest EOF alignment (about 0.45 vs 0.64 for K = 4 and 8), which does not support the claim that forecasting utility follows prior–data alignment.
Questions for authors
  1. How does performance change when the target-day surface fields are replaced by persistence or by a forecast available at initialization? Did all baselines receive the same surface inputs?
  2. Does the advantage hold against a parameter-matched generic bypass, and against the same bypass without zero initialization?
  3. What is the spatial extent of each sample? Fig. 3(d) shows a grid of about 36 × 36 points (about 3° × 3°), and training uses ±18-pixel crop jitter. Are training and evaluation done on small windows or on the full 120–155°E, 10–40°N domain, and how many test cases are there? WenHai and Swin3D show strong block artifacts in Fig. 9; how were these global models adapted and tuned for this setting?
  4. The manuscript does not contain the AI use statement required by the ICLR 2027 Author Guidelines. Please add it.
Reviewer_M864OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes EddyCast for seven-day prediction of upper-ocean temperature and salinity. Its main idea is to introduce analytical features through a residual bypass with a zero-initialized projection. The features describe spatial curvature and recent changes in sea-surface temperature. The paper also studies a learnable modal representation and a diffusion extension. Experiments on Northwest Pacific reanalysis data show lower errors than the evaluated baselines, with ablations comparing several ways of incorporating the same analytical features.

Critical strengths

Table 2 provides a direct comparison of ways to introduce the same EAG features into the same backbone. RMSE falls from 0.0114 with input concatenation and 0.0111 with the auxiliary loss to 0.0106 with the bypass. This supports the bypass design. EddyCast also improves on WenHai's RMSE of 0.0118 with fewer parameters (2.984M versus 6.823M, Table 1), within the evaluated setting.

Critical weaknesses
  1. The surface-input protocol needs clarification before the forecasting results can be interpreted. Appendix A.1 supplies target-day SST, SSS, SSH, currents, and mixed-layer thickness at every rollout step. These fields are prescribed, but their source at test time is not specified. If they come from contemporaneous reanalysis, the model is predicting subsurface states given future surface information. This does not establish seven-day performance using only information available at initialization. The distinction is especially relevant for targets confined to the upper 55.76 m for temperature and 29.45 m for salinity. Removing surface inputs increases RMSE to 0.0126. Giving the same inputs to the baselines makes the relative comparison informative, but leaves the forecasting task itself unclear.
  2. The ablations support the bypass more clearly than the particular physics features. EAG uses SST Laplacians and a temporal difference. Table 2 tests whether the branch helps and how to introduce it, but does not compare these operators with learned local filters applied to the same SST inputs. Keeping the bypass and capacity comparable in that test would show whether the prescribed operators add value beyond an extra feature branch.
Questions for authors
  1. Where do the target-day surface fields come from at evaluation time, and are they available at forecast initialization? If they are reanalysis fields, are there results using persisted or independently forecast surface inputs?
  2. Is there a comparison replacing EAG with learned local features from the same SST inputs, while retaining the bypass and comparable parameter count?
Reviewer_vpxoOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper suggests a new method to include physical constraints in neural eddy forecasting. Instead of imposing them as physical priors or losses, the authors design two embeddings with respectively surfacic and vertical representation of ocean states that are added to the predictor with a learned projection. This set-up can also be extended for probabilistic forecasting. The method is tested on ocean reanalysis and the results highlight the improvement of mesoscale eddy-forecasting through several metrics compared to different forecasting baseline.

Critical strengths
  • Interesting application of feature embedding to ocean forecasting
  • It seems to give good results compared to the baseline on the eddy forecasting task.
Critical weaknesses
  • The paper does not consider previous work in computer vision on feature embedding, as in [1,2]. The distinction between the a. and c. boxes in Fig. 1 is mostly about at which level we inform the network of the physical features (input or through latent features), and this has been widely studied in ML literature for years. It feels like, in that aspect (i.e. the aspect illustrated in Fig1 a vs c and all the downstream justification in 3.3), the paper is not introducing novelty on the machine learning standpoint.

  • Grounds behind analytical, physical guidance term and vertical modal decomposition is unclear. The analytical terms have a quite naive formulation (Laplacian and temporal difference of SST) and it’s hard to understand why the network couldn’t learn those features by itself having the SST as input. Both guidance terms are brought like heuristics, with no real foundation on why we choose those ones and not others,

  • Most of the metrics directly depend on RMSE (MAE is the same evaluation with a different norm, PSNR is literally the log of RMSE, Forecasting Skill is the RMSE divided by the RMSE of the baseline…). It seems off to present so many metrics dependent on the same evaluation, especially given that geophysical science is offering many evaluations beyond RMSE (frequency spectrum to assess for turbulence, for instance).

  • Minor : Main schematics in the paper, and especially Fig.2, are crowded, and it’s written so small that is impossible to read without zooming a lot.

Questions for authors
  • How does the network perform on long term simulation ? At least running simulation to assert that the rollout is not diverging. Authors mention that “These results suggest that the physical bypass contributes not only to short-term correction but also to more stable long-horizon 3D forecasting.” (L451) but don’t support this claim and only test their model on 7-day rollout which is fairly short for ocean simulation (looking at the snapshots in Fig8 in Appendix, the state barely moves)

  • How do the results presented on Table 2 vary w.r.t model initialisation ? In some cases like “EddyCast w/o 3D” the difference of performance with the full version is so thin that we could wonder if it doesn’t land in the variability due to initialisation which would mean the 3D representation has no significant (in the statistical sense) impact on the results.

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

Related papers

Discussion (0)

Sign in to comment.