acceptodds
Under review as a conference paper at ICLR 2027

BridgeCast: Bridging Ocean Wave Forecasts to Reanalysis via Flow Matching with Exogenous Variables

Abstract

Ocean wave forecasting is essential for maritime safety, offshore operations, and coastal resilience, yet remains challenging due to systematic biases in physics-based models. Physical models, while widely used, rely on approximations and parameterizations that limit their accuracy under complex ocean-atmosphere conditions. To enhance ocean wave forecasting, we propose BridgeCast, within a physics-AI hybrid framework for bias correction. BridgeCast is a probabilistic model based on conditional flow matching (CFM) that learns to transform physical model forecasts into reanalysis-like fields. It treats physical forecasts as corrupted observations and employs a continuous-time generative process to bridge their distribution toward that of reanalysis data. BridgeCast is parameterized by a Transformer-based architecture that enables spatiotemporal modeling, incorporation of exogenous atmospheric variables, and flexible inference via both ordinary and stochastic differential equation formulations. Extensive experiments on real-world datasets demonstrate that BridgeCast consistently outperforms state-of-the-art baselines across regions and forecast lead times.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Ocean Wave Forecasting, Generative Models, Flow Matching

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_5REZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies statistical post-processing of physics-based ocean-wave forecasts. BridgeCast aims to transform GFS-WW3 or IFS-WAM forecast fields into fields closer to ERA5 reanalysis. It defines a stochastic interpolation from a Gaussian-perturbed physical forecast to the corresponding reanalysis target and uses a two-head Transformer to predict a residual velocity and Gaussian noise. The architecture combines spatial attention, temporal attention, atmospheric cross-attention, and adaptive layer normalization conditioned on flow time and forecast lead time. ODE and SDE inference modes generate forecast ensembles. Evaluation covers the Gulf of Mexico, North Sea and California coast, using two forecast systems and chronological data splits. The paper reports RMSE, CRPS, threshold-weighted CRPS, CSI and HSS for significant wave height, together with additional period and direction results, component ablations, noise sweeps and inference timings. The stated contribution is a generative forecast-correction framework with spatiotemporal and exogenous conditioning.

Critical strengths
  1. The paper addresses a practically important problem in ocean-wave forecasting through a correction module compatible with existing physical models.
  2. Evaluation across three distinct regions and two forecast systems provides useful evidence of the method’s applicability.
  3. The reported wave-height improvements are promising, with evaluation covering point accuracy, probabilistic performance and extreme-event detection.
Critical weaknesses
  1. The claimed equivalence between Eqs. (12) and (13) is invalid because antithetic pairing does not generally eliminate the noise dependent cross term. Consequently, the sampler omits a necessary drift correction, undermining the claimed consistency between training and inference.
  2. The unconditional FlowMatch baseline cannot isolate the benefit of forecast-anchored transport. The FlowCast and DiffCast adaptations also require clearer documentation.
  3. The DDIM replacement performs substantially worse than the raw physical forecast. Its configuration needs verification before attributing the gap to flow matching.
  4. The paper does not clearly establish a methodological advance beyond existing stochastic interpolants and standard Transformer components. The experiments do not isolate the contribution of the proposed transport formulation from architectural and conditioning choices.
Questions for authors
  1. Provide a consistent description of forecast horizons, window construction, split boundaries and IFS deduplication. Specify the actual tokenization, conditioning inputs and residual connections used in the reported experiments.
  2. Add matched conditional baselines using identical inputs, backbones and computational budgets to isolate the benefit of forecast-anchored transport. Document the FlowCast and DiffCast adaptations and verify the configuration underlying the unusually weak DDIM result.
  3. Explain the practical advantage of SDE inference over ODE ensembles generated through random initial conditions. Support this comparison with calibration, statistical uncertainty and complete inference costs at the ensemble size used for the main results.
  4. Provide independent observational validation where feasible, or limit operational accuracy claims to agreement with ERA5. Analyze the weaker California results to clarify when forecast correction helps and when the proposed design is less effective.
Reviewer_2bvTOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes BridgeCast, a flow-matching-based method for correcting ocean wave forecasts toward ERA5 reanalysis. The model combines a DiT-based architecture with atmospheric conditioning. Experiments on GFS-WW3 and IFS-WAM forecasts show improvements over the selected baselines across different regions and lead times.

Critical strengths
  1. The paper addresses a practical problem in ocean wave forecasting, and the results show clear improvements over the original physical forecasts.
  2. The experiments cover two forecasting systems and three regions, with both deterministic and probabilistic evaluation metrics.
Critical weaknesses
  1. Limited novelty. Bias correction for numerical forecasts is a well-established research area, with many related methods already developed for ocean wave forecasting. The paper does not clearly establish what is fundamentally new compared with existing conditional generative approaches.
  2. Insufficient support for the SOTA claim. Most baselines are general-purpose spatiotemporal models rather than methods specifically developed for ocean wave forecast correction. Several relevant approaches are discussed in the Related Work section but are not included in the experiments. It is therefore difficult to assess whether BridgeCast actually advances the state of the art in this field.
Questions for authors
  1. What is the main methodological novelty compared with standard conditional residual flow matching? Could the authors provide a controlled comparison using the same backbone and conditioning information?
  2. Could the authors provide stronger evidence to support their claim of outperforming the state of the art in ocean wave forecast correction?
  3. In the ablation study, atmospheric variables and cross-attention are removed together. Could the authors compare cross-attention with a simpler approach, such as input concatenation, using the same atmospheric variables?
  4. Could the authors compare the RMSE of individual ensemble members with that of the ensemble mean to quantify how much ensemble averaging improves forecast accuracy?
Reviewer_5MytOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes BridgeCast, a probabilistic model that post-processes the outputs of existing physical wave models (GFS-Wave and IFS-WAM). It takes their wave forecasts, together with the atmospheric fields from the same forecast, and maps them toward ERA5 reanalysis, the physical model itself is left unchanged. The correction is framed as conditional flow matching from the noise-perturbed physical forecast to the ERA5 field. It uses a small DiT-style transformer with spatial and temporal attention, cross-attention to the atmospheric fields, and conditioning on lead time. Two output heads allow deterministic and stochastic inference from the same trained model, and ensembles are generated from random perturbations of the input. Experiments cover three regional domains (Gulf of Mexico, North Sea, California coast) at 0.5degree, with lead times of up to one week. BridgeCast is compared against the raw physical forecasts and against several deterministic and generative architectures, most of which the authors retrained as correctors on the same data. All models are then evaluated with the same set of deterministic and probabilistic metrics. The authors report that BridgeCast performs best in most settings, and add ablations, a storm case study and inference-time measurements.

Critical strengths

The paper addresses a practical problem: operational wave forecasts carry systematic errors, and correcting them is directly useful for marine applications. Starting the correction from the physical forecast rather than from noise is a reasonable choice, and one trained model supports both deterministic and stochastic inference. The evaluation includes probabilistic and extreme-event metrics in addition to RMSE, considers two forecasting systems and three regions with different wave conditions, and the model itself is lightweight. Beyond this, the contribution is limited, for the reasons cited in critical weaknesses section.

Critical weaknesses
  1. The physics-AI hybrid framing is not supported: this is post-processing of physical-model outputs, and the physical model never sees the correction. "Plug-and-play for operational systems" is only shown for two physical models' outputs, on very small domains (about 33x47 points at 0.5degree), and never applied to the outputs of a data-driven model. In addition, comparing inference time with the physical model's 3-hour update cycle says nothing about operational value, since the correction adds cost rather than replacing anything.

  2. The comparison shows little, as most competitors are forecasting architectures retrained by the authors for correction, so the contribution amounts to a known generative framework applied to a simple task.

    • cited wave bias-correction methods (the BU-Net of Sun et al., the TrajGRU-based correction of Zhang et al., Waveformer) are not compared, nor is standard statistical post-processing (regression, quantile mapping, EMOS).
    • CorrDiff, a closely related generative correction model, is cited only for input design.
    • There is no comparison with ML wave forecasting models, so it isn't shown whether correcting GFS or IFS beats forecasting directly with ML.
  3. Since ERA5 serves as both the training target and the reference for every score, the evaluation measures closeness to ERA5 rather than to the real sea state. Part of the gain may simply come from learning ERA5's smoothness: ERA5 waves are provided at 0.5, while IFS-WAM is subsampled from 0.25 by taking every other point, and no smoothed or regridded physical forecast or spatial spectra are shown to rule this out. The model also learns ERA5's own biases, such as its underestimation of high waves at most buoys, and the extreme-event thresholds come from the same training data (presumably ERA5), so these biases cannot show up in the scores; the literature on reanalysis biases and their correction is not discussed. Independent verification against buoys or altimeters is therefore needed. The paper also states that ERA5 waves are assimilated against altimeter and buoy observations. As far as I can find, ECMWF's own documentation of ERA5 lists only altimeter wave heights as assimilated wave data, so this claim should be supported with a reference or corrected. If buoys are not assimilated, they would provide exactly the independent reference against which the corrected forecasts should be verified.

  4. Temporal continuity isn't addressed. The 24 frames of a 72-hour window are corrected together, but continuity within a window is never evaluated. Windows are corrected independently with no look-back, as the authors acknowledge, so nothing prevents jumps at boundaries, and ensemble member k in one window is unrelated to member k in the next. It's also unclear how test windows cover each lead band.

  5. The probabilistic evaluation is limited. Deterministic baselines get a CRPS equal to their MAE, so probabilistic models win on CRPS by construction. There are no calibration checks such as spread skill ratios or rank histograms. BridgeCast's noise parameters are tuned per region on validation, while baseline tuning isn't described.

  6. The ablation covers one region and one dataset, with no uncertainty. With sigma = 0 there is no ensemble, so the worse CRPS is trivial, and the degradation of up to -460% for the DDIM variant suggests a broken setup. The atmospheric inputs come from the same forecast cycle and help only slightly, as the authors state.

  7. The results are not robust enough for general claims: IFS has about ten months of training data and a 6-month test set of overlapping windows (median spacing 3h), GFS is tested on a single year, nothing reports uncertainty or repeated runs, and the storm example is a single case.

minor points:

  • The main text and appendix disagree on the lead-time range (up to 240 h vs. 168 h).
Questions for authors

The main questions are raised in the Critical Weaknesses section.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.