acceptodds
Withdrawn from ICLR 2027

OA-Bridge: One-Step Coupled Ocean-Atmosphere Forecasting with a Vision–Language Model

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Atmosphere–ocean coupling is a fundamental physical process of the Earth system, yet most AI-based forecasting models treat the atmosphere and ocean as independent components. Learning a unified forecasting model is challenging because atmospheric and oceanic variables evolve on substantially different timescales, while straightforward joint training can cause negative transfer and degrade the predictive skill of otherwise strong component models. We propose OA-Bridge, the first VLM-based framework for explicitly coupled global atmosphere–ocean forecasting at daily resolution. Given the current atmospheric and oceanic states and a specified lead time, OA-Bridge directly predicts their joint future states without recursively applying a fixed-step transition model. A shared frozen Qwen2.5-VL backbone jointly processes the two representations conditioned on the requested lead time. Component-specific encoders, output adapters, and decoders preserve specialized representations, while a coupling adapter uses bidirectional cross-attention to exchange information between the two systems. Experiments from 1 to 60 days against a jointly trained FuXi baseline show that OA-Bridge substantially improves long-range ocean RMSE and ACC across depths, extreme-event skill, and SSH-derived geostrophic currents, while also improving atmospheric skill at extended lead times.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Ocean and Atmosphere coupling, Weather Forecasting, Generative Model, Deep Learning

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_7ptBOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

OA-Bridge predicts atmospheric and oceanic states at a requested lead time between 1 and 60 days. Separately pretrained encoders and decoders are connected through cross-attention and a frozen Qwen2.5-VL backbone, with the lead time supplied in a text prompt. The experiments use ERA5 and GLORYS12 on a common 1-degree grid. Compared with a jointly trained autoregressive FuXi model, OA-Bridge gives better extended-range ocean forecasts and improves some atmospheric variables, while sacrificing some short-range atmospheric accuracy.

Critical strengths
  1. The strongest result is the long-range ocean forecast. OA-Bridge stays ahead of persistence in Figure 2, and the gains extend across depths in Figure 3. Beating persistence is particularly relevant here because many ocean variables evolve slowly.
  2. Predicting a requested target day directly is attractive for extended-range forecasting: the reported inference time is 2.68 seconds for one target day. The paper also reports the loss of short-range atmospheric accuracy, so the tradeoff is visible.
Critical weaknesses
  1. How much does the VLM add? The largest ocean gains over FuXi in Figure 2 appear after about day 30, when FuXi's temperature and salinity errors increase sharply. Both the backbone and the prediction strategy change in this comparison. The direct MLP variant in Figure 5 is more informative, but its ocean results are much closer to those of the VLM. The clearest atmospheric gains are expressed as percentage changes in ACC, although several long-range correlations are already low in Figure 2. Absolute scores and the MLP configuration are needed to judge the size of the VLM contribution. The current evidence makes a stronger case for direct forecasting than for the choice of a pretrained VLM.
  2. Appendix B establishes that the two components influence each other, but not that this exchange improves accuracy. Its perturbation results give sensitivity rankings rather than gains from coupling. Both the VLM and MLP variants already use the coupling adapter, so their comparison cannot settle this either. A comparable direct predictor without ocean–atmosphere exchange is needed to support the coupling claim.
Questions for authors
  1. What are the MLP architecture and parameter count, and the absolute RMSE and ACC of both variants at 30 and 60 days?
  2. Do you have a comparison with otherwise similar direct predictors trained without cross-component exchange? The control would need to prevent exchange through both the adapter and the shared backbone.
Reviewer_DXEiOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes OA-Bridge for coupled ocean-atmosphere forecasting, which integrates bidirectional cross-attention with a frozen vision-language model to enable lead-time-conditioned direct prediction. Experiments demonstrate the effectiveness of OA-Bridge.

Critical strengths

1.The design of OA-Bridge, which integrates bidirectional cross-attention for ocean-atmosphere coupling and a frozen VLM for lead-time-conditioned direct forecasting, is clear and easy to follow.

Critical weaknesses

1.The limited methodological novelty of OA-Bridge makes it largely an incremental work. Its core components, including Swin Transformers, cross-attention, and vision-language model, are well-established techniques.

2.The motivation for introducing a pretrained VLM is insufficiently justified. It remains unclear why a VLM is necessary and how it specifically addresses the two challenges identified in the paper.

3.The paper has some weaknesses in the experiments, which are not convincing enough: (1) The ablation studies are insufficient to validate the effectiveness of the atmosphere-ocean coupling mechanism and the VLM component.

(2) The baseline comparisons are inadequate. Joint FuXi uses autoregressive forecasting while OA-Bridge adopts direct forecasting. Moreover, more advanced forecasting models should be included for a comprehensive comparison, e.g., OneForecast [1].

(3)The spatial resolution of 1° is relatively coarse. Experiments at 0.25° are needed to further validate the model's effectiveness and scalability.

[1]Gao Y, Wu H, Shu R, et al. (2025). Oneforecast: A universal framework for global and regional weather forecasting. ICML 2025.

Questions for authors

In Equation (3), the authors use oceanic features as queries and atmospheric features as keys and values when updating atmospheric representations. Could the authors clarify the motivation for this design, rather than using atmospheric features as queries and oceanic features as keys and values?

Reviewer_h1xkOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes OA-Bridge, a model that forecasts the global atmosphere and ocean together at 1degree and daily resolution, directly at any lead time from 1 to 60 days in a single forward pass. Separate encoders and decoders for each component are first pretrained on one-day forecasts. In a second stage, the encoded atmospheric and oceanic features exchange information through cross-attention and are passed, together with a prompt giving the lead time, through a frozen Qwen2.5-VL-7B before being decoded into the forecast. The model is trained on ERA5 and GLORYS12 and tested on 2023 against a jointly trained autoregressive FuXi baseline and persistence, using RMSE, ACC, extreme-event scores, SSH-derived geostrophic currents, an ablation replacing the VLM with an MLP, and an input perturbation analysis. The authors report gains over the baseline at long leads for both components and a 15.83x speedup at day 60.

Critical strengths

The paper addresses a relevant problem: most data-driven models still treat the atmosphere and ocean separately, and a joint model is a natural step toward coupled forecasting. Applying direct lead-time-conditioned prediction to a coupled atmosphere-ocean setting up to 60 days is an interesting direction, as it avoids error accumulation through long rollouts. The ocean evaluation goes beyond surface fields, with scores across depths, spatial error maps, extreme events and SSH-derived currents, and the authors are open about the weaker short-range atmospheric skill. Beyond this, the contribution is limited, for the reasons below.

Critical weaknesses
  1. The paper is sold on atmosphere-ocean coupling, but almost nothing is shown about the coupling itself.

    • It isn't shown whether the model really couples the two systems or just learns a joint mapping from one snapshot to another. The most direct test, the same model without exchange between the branches, is missing, and the VLM ablation keeps the coupling adapters.
    • That leaves Appendix B as the only evidence, and it's weak: each row uses a different lead time without justification, the ranking by relative RMSE increase favors channels with small errors, and some top responses are hard to explain physically, such as SSH perturbations whose strongest responses are winds at 50-150 hPa.
  2. The frozen VLM's role is neither justified nor properly ablated, in practice it's the only place the lead time enters and the only component with global attention, but nothing explains why its language and vision pretraining should matter here.

    • The MLP ablation removes size, global attention and pretraining all at once, and its results are thin: small ocean differences, and atmospheric gains that are relative changes on ACC (0.1~0.3). A trainable transformer of similar size, the same architecture with random weights, and a learned lead-time embedding would be needed.
    • the paper doesn't state whether the VLM's attention is causal. If that is the case, atmospheric tokens may never see the oceanic ones inside it, making the "bidirectional" coupling one-directional.
  3. The idea of one shot forecast at any lead time is interesting, but it isn't new (MetNet, ClimaX, Stormer, and LangYa for the ocean), and also the least validated part:

    • No forecast fields are shown, and nothing checks continuity between consecutive lead times.
    • Day 60 only appears in the last 1,000 of 60,000 training iterations, and nothing shows what happens beyond the 60 days it was trained on.
    • The model is never run on its own outputs. Chaining its forecasts would show whether they're realistic states or smooth fields that drift once fed back. Long coupled runs are exactly what SamudrACE is built for.
  4. Variables evolve very differently over lead time, and nothing checks whether they stay physically consistent with each other: currents vs the SSH gradient, temperature and salinity vs SSH, winds vs pressure. The "physical consistency" section only compares SSH-derived currents with the reference, not with the model's own predicted currents. If the model can't keep its own variables coherent, the coupling is even harder to claim.

  5. The only baseline is exactly the joint-training setup the introduction warns about. It's also no longer the published FuXi: it was enlarged to 2B parameters but trained for the same number of iterations, with no evidence it was properly tuned and no rollout-training length given.

    • It blows up and falls below persistence for salinity, temperature and SSH. for these variables the two models are close for the first 20-30 days, and the large gap appears only once the baseline diverges, only the currents are the exception, better from day 15 onward.
    • The NRMSE maps measure error relative to this diverging baseline, so blue means "less bad than the baseline", not "good".
  6. There's no climatology reference, and persistence keeps the full initial state, including the seasonal cycle, so beating it at 60 days is a low bar. RMSE also rewards smooth forecasts: after about two weeks, atmospheric RMSE goes flat while ACC stays around 0.1~0.3 for most of variables.

    • ACC is uncentered and computed against a 1993/2021 climatology, so a shared anomaly in 2023 could raise it. This could explain why Z500 skill stays flat or rises at long leads?
    • In Fig. 2, the y-axes are scaled by the baseline's blow-up, which compresses the range where the models are comparable.
    • The number of initialisations is never given; there's only one test year, no uncertainty and no score table.
  7. The model trades short-range skill for long-range gains: for most atmospheric variables, ACC and extreme-event skill are worse than the baseline over the first 10-20 days, and the advantage only appears once ACC is already low. Since each lead time is predicted separately, nothing shows that these long-range forecasts follow from the weaker short-range ones; the continuity and chaining (see point 3) would reveal this. For the ocean, extremes are similar at the surface until about day 40, so the extreme-event claim mainly holds at depth, once the baseline degrades.

  8. SamudrACE , AIFS (which already uses a shared backbone for atmosphere and surface ocean), Ola, FengShun-CSM, and well-trained uncoupled component models are not compared against, and the "first VLM-based" claim is narrow given LangYa.

  9. The "fair comparison" only matches trainable parameters: at inference, OA-Bridge runs about 9B parameters against 2B, with no breakdown given. The speedup also counts only a single target day, while a full daily trajectory needs 60 forward passes of this larger model.

minor points:

  • "93 channels" in Section 3.1 vs. "93 oceanic variables" in Section 4.1.
  • The lead-time choices in Appendix B are not justified.
Questions for authors

Please refer to the Critical Weaknesses section

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

Related papers

Discussion (0)

Sign in to comment.