acceptodds
Under review as a conference paper at ICLR 2027

OceanFlow: Coherent Flow Matching for Probabilistic Ocean Modeling

14 views

Abstract

Probabilistic ocean modeling is essential for risk-aware marine operations, climate analysis, and scientific decision-making, where downstream decisions depend on the conditional distribution of future ocean states rather than a single deterministic trajectory. Recent data-driven ocean models have made deterministic global ocean modeling increasingly accurate and efficient, but probabilistic ocean modeling remains underdeveloped. Specifically, perturbation-driven ensembles can create spread without explicitly optimizing the joint distribution of their members, often yielding poor calibration and physical inconsistency. To address these issues, we propose OceanFlow, a coherent flow matching framework for probabilistic global ocean modeling. OceanFlow first establishes a sharp, stable mean anchor for the conditional ensemble. Flow matching then learns a coherent state-conditioned pathwise law that jointly reshapes the residual distribution around this anchor at each transition, preserving the predictive center while refining state-dependent uncertainty. We further introduce a Factorized Dynamics Operator (FDO), which combines factorized transitions with a learned state-conditioned response to parameterize the mean anchor and residual-flow velocity. Extensive experiments show that OceanFlow improves ensemble-mean accuracy, probabilistic reliability, and physical-consistency diagnostics over competitive baselines, supporting coherent flow matching as an effective route toward data-driven probabilistic ocean modeling. Code is available at https://anonymous.4open.science/r/Code_OceanFlow-25B9/.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Ocean Forecasting, Spatio-temporal Data Mining, AI for Science

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_VBBhOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes OceanFlow, a probabilistic data-driven model for global ocean simulation and forecasting. A deterministic backbone, the Factorized Dynamics Operator (FDO), is run on Perlin-perturbed initial states to produce a reference ensemble. A learned correction then adjusts the ensemble mean, and a flow matching step reshapes the residuals around it, with a velocity conditioned on ensemble statistics. A final recentering keeps the ensemble mean at the corrected value. The claimed contributions are this coherent flow matching framework, the FDO architecture, and a short theoretical result on the resulting conditional law. The model is trained on GLORYS12 coarsened to 1 degree with ERA5 forcing: five ocean variables, four of them on 23 levels, giving 93 channels. Experiments cover simulations up to 120 days with reanalysis forcing and forecasts up to 30 days with OneForecast forcing, on 2020 and 2024. OceanFlow is compared with SimVP, GraphCast, CirT, WenHai, NeuralOM and DSLCast, some of them turned into Perlin-noise ensembles. Evaluation uses RMSE, ACC, CRPS, spread-skill ratio, rank histograms, coverage, a Nino-3.4 analysis, spectra, physical diagnostics and ablations. The authors report improvements in accuracy, probabilistic reliability and physical consistency over all baselines.

Critical strengths

The paper addresses a relevant problem: moving from deterministic prediction to ensembles, which is a natural next step for data-driven ocean models, and uncertainty estimates are needed for most practical uses. Separating the ensemble center from the spread around it is a simple and reasonable design. The evaluation goes beyond RMSE, with CRPS, spread-skill ratio, rank histograms and coverage, which is the right set of tools for judging an ensemble. Beyond this, the contribution is limited, for the reasons below.

Critical weaknesses
  1. In Appendix E.8, the member conditions for the next day are taken from the deterministic reference ensemble, not from the transported members. The flow and the mean correction are applied to each day's reference output and then thrown away, so what remains is a Perlin-perturbed deterministic ensemble with a learned post-processing step on top. The residual noise is also redrawn at every transition and the context reset every day (E.4), so the learned part of member m on one day has nothing to do with member m on the next, and members can't be used as trajectories. Both learned components are also trained only on one-step transitions from true states, with a spread target taken from the one-day error. I don't see how this can represent uncertainty at 90 or 120 days, or how a one-step correction can fix drift built up over months. My guess is that long-lead spread comes mostly from the divergence of the Perlin reference and the hand-designed spread floor. None of this supports the "coherent state-conditioned pathwise law" claim, and no diagnostic looks at member trajectories or at where the spread actually comes from.

  2. Reading Appendix E.4, each member's training target is built from the same noise it starts from: mostly that noise scaled up or down, with only 5% coming from the actual error pattern. So the flow mainly learns how much to stretch or shrink the noise at each grid point. That's useful, but it isn't learning a distribution, and learned residual keeps the shape of the generic smooth noise. With only two flow steps, the continuous-time theory describes an idealized version, not the actual model. The ablation points the same way. Removing the learned flow moves CRPS from 0.371 to 0.379 and RMSE from 0.684 to 0.697, and most of the gain over FDO+Perlin comes from the calibration steps around it. The headline contribution adds a little on its own.

  3. Lemma 1 says that subtracting the ensemble mean and adding µ gives an ensemble with mean µ. This is true by construction, and the authors admit it says nothing about calibration or skill. The FDO is mostly standard components with new names: "factorized dynamics" is gated multi-branch mixing with FiLM-style conditioning, and the "state-conditioned response" is a learned residual connection. The deterministic FDO is also clearly weaker than NeuralOM (RMSE 0.825 vs 0.682 at 90 days), so it can hardly be called an accurate deterministic reference.

  4. Several relevant data-driven ocean models are neither discussed nor compared against, including Njord, Samudra (1° and 1/4°), GLONET, Neptune, ORCAst and the GNN-based model of Hirabayashi et al., as well as regional efforts such as SeaCast and MedFormer. Njord and Neptune already produce probabilistic ensembles. Others are trained on the same GLORYS data, and some are public models that could serve directly as baselines. As written, the related work gives a misleading picture of the field, and the claim that probabilistic ocean modeling is "underexplored" needs more context.

  5. OceanBench is built on the same GLORYS reanalysis, with defined metrics, protocols and published scores, and already includes models this paper compares against or cites (WenHai, XiHe, LangYa). The authors neither use nor mention it, and OceanForecastBench is only cited in passing. Their numbers can't be checked against published scores, and there's no way to tell whether their re-evaluation of WenHai matches what is reported there. Choosing a custom protocol when a standard one exists on the same data, with some of the same models, needs a clear justification.

  6. Samudra uses a convolutional UNet, is publicly available, and has shown stable and skillful long rollouts. That is hard to square with the convolutional results here: SimVP has ACC 0.088 at 90 days, and the plain convolutional ablation reaches RMSE 1.054. CirT also looks broken, with normalized RMSE around 2.8, ACC near zero and a member spatial-std error of 1.02. The retrained baselines seem poorly tuned, which inflates the gains.

  7. NeuralOM alone already matches OceanFlow's ensemble mean (0.682 vs 0.681, within the std), yet it's the only strong baseline never given a perturbation ensemble. Perturbation averaging lowered DSLCast's RMSE by about 0.13, so a NeuralOM ensemble would very likely beat OceanFlow. NeuralOM is also missing from Figure 4, and DSLCast is called "the strongest perturbation-based baseline" in Section 3.4 only because NeuralOM was never perturbed. More broadly, OceanFlow tunes its perturbation amplitude and spread against CRPS, while the baselines get untuned Perlin noise and are badly under-dispersed (SSR sown to 0.22). Any CRPS-trained calibration would improve CRPS and calibration error, so this says little about the method. FuXi-ONS (discussed but not run), a CRPS-trained noise-injection model, or simple spread inflation on the baselines would make a fairer comparison. It's also unclear how the pretrained models (WenHai, NeuralOM, DSLCast) were evaluated in this setting: different resolutions, forcing..

  8. The 20 initial conditions are 20 consecutive days starting January 1, 2020 (Appendix F), and the 2024 test uses 10. These aren't independent samples, they cover a single season, and the tiny standard deviations in the tables are misleading. The Nino-3.4 analysis suffers most: with only 10 initial conditions, whose dates are not given, nothing can be said about ENSO skill. The correlation metric is defined but never reported.

  9. The paper talks about "93 ocean variables", but these are five physical variables, four of them on 23 levels, so 93 channels. Every main table averages over all of them in normalized space, which hides a lot. Ocean variability is very uneven: some regions, such as western boundary currents, the equatorial band or the Southern Ocean, are highly energetic, while others change very little over a few months. Depth levels respond to forcing on different timescales, and the variables themselves differ widely in how much they vary. A single averaged score can therefore improve because of how a model does on part of the channels or regions, while saying little about the rest. The calibration numbers may be an example of this: an SSR of 0.71 indicates clear under-dispersion, yet a CalibErr of 0.032 indicates near-perfect coverage, which suggests the coverage statistic is dominated by a subset of points. per-variable results cover only five surface fields and three subsurface levels, only against the Perlin baselines, with no regional breakdown.

  10. Section 3.5 argues that the ensemble mean shouldn't be read as a physical state, then evaluates two of its three physical diagnostics on the ensemble mean. The diagnostics are thin anyway: a global-mean SST tendency, a correlation of correlations between SSH and currents, and an amplitude ratio. There are no budget checks for heat, salt or volume, no member-level balance checks and no stratification checks. The member spectra may also look closer to the truth partly because of the smoothing applied in the residual branch rather than better physics.

  11. The gain in the two-way coupling experiment (Table 5) comes from better atmospheric forcing produced by another system, and DSLCast isn't run with the same forcing, so it says little about OceanFlow. Section 3.7 is called "efficiency analysis" but gives no runtime, even though inference needs a deterministic step plus two flow evaluations per member and per day. The "fewest trainable parameters" claim probably excludes the frozen backbone and leaves out the pretrained baselines?

  12. Minor issues:

    • RMSE goes down from 90 to 120 days in several rows, unexplained.
    • The years used for the ACC climatology are not given.
    • The non-fair empirical CRPS is used with only 10 training members, which favors under-dispersion.
    • Ensemble statistics fed to the velocity use 10 members in training but 20 at evaluation.
    • The "w/o State Response" ablation has a huge variance (ACC 0.072 +- 0.090), suggesting unstable training or a single seed.
    • A more direct description of the method, with less specialized terminology, would help readability.

Some of the cited papers:

  • Holmberg, Daniel, et al. "Accurate Mediterranean Sea forecasting via graph-based deep learning." Scientific Reports 15.1 (2025): 45051.
  • Holmberg, Daniel, et al. "Njord: A Probabilistic Graph Neural Network for Ensemble Ocean Forecasting." arXiv preprint arXiv:2605.15470 (2026).
  • Garcia, Pierre, et al. "Orcast: Operational high-resolution current forecasts." Artificial Intelligence for the Earth Systems 4.4 (2025): e250002.
  • Epicoco, Italo, et al. "MedFormer: a data-driven model for forecasting the Mediterranean Sea." Earth Science Informatics 19.8 (2026): 129.
  • El Aouni, Anass, et al. "GLONET: Mercator's end‐to‐end neural Global Ocean forecasting system." Journal of Geophysical Research: Machine Learning and Computation 2.3 (2025): e2025JH000686.
  • Donno, Davide, et al. "Neptune: An AI model for Global Ocean Subseasonal Prediction." arXiv preprint arXiv:2609.08606 (2026).
  • Li, Jinxin, et al. "OceanGraph: Adaptive Mesh-Grid Coupling for Multiscale Ocean State Prediction." IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (2026).
  • Dheeshjith, Surya, et al. "Samudra: An AI global ocean emulator for climate." Geophysical Research Letters 52.10 (2025): e2024GL114318.
  • Yuan, Yuan, et al. "Samudra 2: Scaling Ocean Emulators across Resolutions." arXiv preprint arXiv:2606.02610 (2026).
Questions for authors

Please refer to the weaknesses summarized above.

9 comments
AuthorsOpenReview ↗

Response to Reviewer VBBh — Part 1/9

Dear Reviewer VBBh,

Thank you for the detailed review and for identifying points that require clearer definitions and evaluation context. We respond below with clarifications, the existing controlled comparisons, and an additional independent retraining experiment.

Q1. In Appendix E.8, the member conditions for the next day are taken from the deterministic reference ensemble, not from the transported members. The flow and the mean correction are applied to each day's reference output and then thrown away, so what remains is a Perlin-perturbed deterministic ensemble with a learned post-processing step on top. The residual noise is also redrawn at every transition and the context reset every day (E.4), so the learned part of member m on one day has nothing to do with member m on the next, and members can't be used as trajectories. Both learned components are also trained only on one-step transitions from true states, with a spread target taken from the one-day error. I don't see how this can represent uncertainty at 90 or 120 days, or how a one-step correction can fix drift built up over months. My guess is that long-lead spread comes mostly from the divergence of the Perlin reference and the hand-designed spread floor. None of this supports the "coherent state-conditioned pathwise law" claim, and no diagnostic looks at member trajectories or at where the spread actually comes from.

A1. Thank you for examining the inference procedure closely. OceanFlow learns a conditional ensemble law at each ocean transition, not a fully flow-driven joint law over an entire multi-month trajectory. The deterministic reference advances the member conditions in physical time; the learned mean correction and residual transport determine the ensemble that is saved and evaluated at that transition. Their outputs are therefore not discarded from the forecast, although they are not fed back as the next day's member conditions. Here, pathwise refers to residual evolution in auxiliary flow time within a transition. Coherent refers to member updates being jointly conditioned on the evolving ensemble, with final recentering preserving their common anchor. Resetting the residual noise and context does not remove this within-transition interaction. It does mean that a fixed transported-member index should not be interpreted as a validated stochastic trajectory across days. We will make this scope explicit in the main method and inference description.

The long-lead condition comes from the evolving reference rollout, rather than repeatedly restarting from the initial state. OceanFlow learns to transform that reference ensemble using its current state and forcing. Applying this one-step-trained transformation at long leads is an empirical generalization question; we do not claim that the training objective guarantees removal of arbitrary accumulated drift. The 90-/120-day experiments evaluate precisely whether this transformation remains useful under those rollout conditions. The same-reference ablations provide evidence beyond the presence of Perlin spread or a floor. At 90 days over 10 ICs:

MethodRMSE↓CRPS↓RankDev↓CalibErr↓
FDO+Perlin0.715±0.0050.403±0.0020.204±0.0010.225±0.001
w/o learned pathwise flow0.697±0.0040.379±0.0020.109±0.0010.081±0.002
OceanFlow0.684±0.0050.371±0.0020.088±0.0000.032±0.000

FDO+Perlin supplies the same deterministic backbone and perturbation reference without learned distributional refinement. The no-flow variant retains mean control and the reference-based output construction but removes learned residual transport. Its weaker calibration shows that the full result is not recovered by the reference and mean-control mechanisms alone. These comparisons establish the usefulness of learned ensemble refinement, although they do not constitute a variance-budget decomposition of every source of spread. Our contribution is thus reference-guided, mean-anchored ensemble transport with improved conditional reliability at the evaluated lead times. Temporally coherent stochastic trajectories and feedback of transported members are distinct extensions, rather than properties implied by the current coherent-law claim.

AuthorsOpenReview ↗

Response to Reviewer VBBh — Part 2/9

Q2. Reading Appendix E.4, each member's training target is built from the same noise it starts from: mostly that noise scaled up or down, with only 5% coming from the actual error pattern. So the flow mainly learns how much to stretch or shrink the noise at each grid point. That's useful, but it isn't learning a distribution, and learned residual keeps the shape of the generic smooth noise. With only two flow steps, the continuous-time theory describes an idealized version, not the actual model. The ablation points the same way. Removing the learned flow moves CRPS from 0.371 to 0.379 and RMSE from 0.684 to 0.697, and most of the gain over FDO+Perlin comes from the calibration steps around it. The headline contribution adds a little on its own.

A2. The endpoint construction is deliberately structured, but the observation enters both its amplitude and its error-pattern component, and training also scores the generated ensemble directly against observations. Specifically, yy is the next-day GLORYS12 reanalysis field and e=y−sg⁡(μ)e=y-\operatorname{sg}(\mu) is its error relative to the anchor. The 95% component uses a spread target derived from the channel-wise ocean RMS of this error; the remaining 5% introduces its spatial pattern. Thus, 5% is not the fraction of the target informed by reanalysis. The shared source noise defines a source-target coupling for flow matching, rather than an independent target sample. We agree that this is a constructed residual target, not an observed ensemble of alternative futures. Its purpose is to provide tractable transport supervision. The terminal ensemble is additionally optimized through empirical CRPS against yy, with gradients propagated through the integrated residual paths. Mean, member-error, and spread objectives provide complementary supervision. Consequently, the model is not trained only to reproduce a fixed rescaling of smooth noise. Even learned conditional amplitude changes alter a distribution, while the spatial velocity network and ensemble context allow departures from that restricted transformation. We will distinguish this modeling capacity from claims of recovering an unrestricted true conditional law.

Two numerical updates are used for both distributional training and inference. The implemented composition defines a finite-step conditional ensemble map; the mean-preservation and member-relabeling properties do not require it to equal the exact solution of a continuous ODE. The continuous formulation describes the transport construction, not a claim of negligible discretization error. The no-flow comparison is particularly informative for reliability. Using the original ablation table:

Metricw/o learned pathwise flowOceanFlow
CRPS↓0.379±0.0020.371±0.002
RankDev↓0.109±0.0010.088±0.000
CalibErr↓0.081±0.0020.032±0.000

Removing transport raises the calibration error from 0.032 to 0.081. Equivalently, retaining it reduces CalibErr by approximately 60.5% and RankDev by 19.3%, alongside a 2.1% CRPS improvement. The modest incremental CRPS change therefore does not capture the full reliability difference. We do not regard mean control as incidental machinery outside the contribution: the core idea is to learn the residual distribution around an explicitly controlled center. The no-mean-control and no-flow variants test these complementary roles. Because these components are optimized together, their gains need not add independently. The results support the complete mean-anchored transport formulation, rather than attributing all improvement to flow integration alone.

AuthorsOpenReview ↗

Response to Reviewer VBBh — Part 3/9

Q3. Lemma 1 says that subtracting the ensemble mean and adding µ gives an ensemble with mean µ. This is true by construction, and the authors admit it says nothing about calibration or skill. The FDO is mostly standard components with new names: "factorized dynamics" is gated multi-branch mixing with FiLM-style conditioning, and the "state-conditioned response" is a learned residual connection. The deterministic FDO is also clearly weaker than NeuralOM (RMSE 0.825 vs 0.682 at 90 days), so it can hardly be called an accurate deterministic reference.

A3. We appreciate this distinction between a structural statement and a performance guarantee. Lemma 1 characterizes the generated conditional ensemble law; it does not prove calibration or forecast skill. Centering preserves the empirical mean by construction, and existence under the stated regularity assumptions is standard. The useful connection is that ensemble-conditioned transport, reference composition, and recentering together define one conditional map on the member set, without dependence on arbitrary member labels. We will state this modest role directly and streamline the proof accordingly. FDO's architectural contribution lies in organizing factorized evolution and state-conditioned response within one conditional transition operator. Learned projections form factor subspaces that undergo separate transformations, followed by cross-factor interactions with state-dependent selection. The output response then uses the original condition to adapt the decoded transition. This organization provides separate parameterizations for the deterministic reference and residual-flow velocity within the mean-anchored framework. The deterministic ablations evaluate this design by replacing the complete factorized block, removing the state response, or replacing it with a fixed state addition. Their results support the combination of structured factor evolution and a learned response, beyond a fixed input shortcut.

We also agree that the standalone FDO does not outperform NeuralOM at 90 days. “Accurate reference” should describe its role in the ensemble construction, not imply that it is the strongest deterministic model. NeuralOM is a substantial comparator: its published configuration uses 0.5° data, staged residual modeling, and multi-step fine-tuning, whereas our deterministic stage uses daily 1° data and 50 training epochs. These are differences in training protocol, not grounds for dismissing its stronger result. Crucially, our probabilistic result does not rely on claiming a superior deterministic backbone. With the same FDO reference, CRPS improves from 0.403 to 0.371, RankDev from 0.204 to 0.088, and CalibErr from 0.225 to 0.032 when learned OceanFlow refinement is introduced. The main 20-IC comparison also reports essentially tied 90-day RMSE for NeuralOM and OceanFlow, while OceanFlow provides an evaluated ensemble distribution. We will distinguish the structural lemma, the supporting architecture, and the empirical contribution of mean-anchored distribution learning more clearly.

Reviewer_7GYWOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The submission proposes a flow matching approach to probabilistic Ocean modeling. The core idea is to sample an ensemble from a flow matching model that only models the residual around a fixed mean trajectory. This so called mean anchor is the empirical mean of an ensemble of deterministic forecasts. The distribution of the residual ensemble members is post processed to have (empirical) mean zero and trained via a complex compound loss function. The method is evaluated empirically according to a host of metrics on various settings up to reasonable complexity.

Critical strengths

One aspect I applaud is the empirical evaluation of the method in terms of showing a variety of relevant metrics. In weather and climate modeling, it is non-trivial to find good summary metrics that encapsulate relevant aspects of model performance. The authors here do a solid job in presenting a whole host of domain-specific relevant metrics without just throwing large walls of numbers at the reader. In particular, the addition of rank histograms, calibration curves and especially the spectral analysis really adds to understanding what is going on.

The assessment of how to interpret ensemble means in comparison to the individual member trajectories that constitute the ensemble was one of the key insights I took away (not necessarily core to this paper, but a general learning).

Critical weaknesses

The readability especially of the method part up until the experiments is really lacking. There are many sequences of words that all roughly make sense, but do not map to concise and clear concepts or mathematical statements. It is hard to pin this down, but even reading carefully and refering to the appendix, it was not easy to understand what the method actually does. I certainly would have great trouble implementing the method just given the paper. I'll try to give examples of unclear/empty writing are "establishes a sharp, stable, mean anchor". This becomes clear later, but at this stage, the words "sharp" and "stable" do nothing but confuse. Similarly the "Factorized Dynamics Operator" remains fairly opaque to me throughout (More details later).

This difficulty of understanding is corroborated by not explaining the introduced math in the main text well enough (in my view). Specifically, section 2 introduces many objects very briefly without sufficient explanation of what the symbols and objects do. In short, it's very hard to understand the bigger picture pipeline in terms of how we go "from data to trained model". The theoretical analysis in my view is entirely superfluous. I am not sure what the relevant statement is in Lemma 1. (See questions)

For me, the motivation is also not fleshed out well enough. Specifically, I am missing a stronger motivation for (a) why probabilistic modeling is so desirable and (b) why the deficencies of ensembling deterministic methods matter (I do understand they may have deficiencies, but why are they problematic). Here I'm looking for motivation that goes beyond "it empirically works better according to metrics that care about the distribution".

Questions for authors
  • One aspect that remained fairly unclear to be relatively deep into the paper is "where do we get the target distribution from to train a flow". I see how the mean just comes from averaging deterministic rollouts. But to train the residual ensemble, some desired target distribution must be known. Given the arguments about the deficiencies of simply ensembling deterministic rollouts from perturbed initial states, in my view this target should not come from the deterministic ensemble. I was trying to find this in "Coherent state-conditioned pathwise flow" at first, where there is only a cryptic description of "verifying center error" or later "where y is the verifying state" (one of these word constructions that aren't pinned down). I then tried to pin it down in terms of the training loss. But equation (13) suddenly drops a 4-component loss out of nowhere with just a single sentence describing in words what each term roughly intends to do. What justifies all these terms (beyond just empirical ablations), choosing the many parameters fells daunting, what exactly are the terms? Also, I still don't understand where the critical targets y come from.
  • As mentioned the factorized dynamics operator remains hard to understand. What is the "factorized" aspect, why is it a "dynamics operator"? This appears to be unusual wording for an autoregressive time-stepping. Equations (9) through (13) lack explanation and context in my view. (9) and (10) introduce a whole host of symbols that are not defined mathematically and the verbal description is fuzzy and ambiguous (e.g., "combines cross-factor interactions with state dependent selection"; what happens mathematically or in terms of implementation? or "integrates collective response into path state"; how does it "integrate"; what even is a patch state (we only have E^patch in (9) without any explanation), etc).
  • In Lemma 1, what is the relevant statement. Is the mere existence (i.e., that one obtains a valid conditional law) itself non-trivial (appears natural to me given that all required existence and uniqueness assumptions are dropped in). Or is the relevant statement just that it preserves the empirical mean? (Given the construction, this appears trivial as well?) The proof is not very human-friendly to me (artificial contrasts, additional unnecessary re-assurances about what's required when and where, etc). I don't see the relevance of all the mentioned equivariances.
  • Unfortunately, I am not familiar enough with ocean modeling specifically. Hence, it is a bit hard for me to judge the relevance of the empirical settings and results. Could the authors put the chosen settings and results into perspective? How far are we from full-fledged ocean model components in state of the art numerical weather prediction models? Is the long-term goal to replace/augment existing ocean component models with these learned models? What would be the main challenges there?

Side note: I did look into the extensive appendix and I know that many objects I am complaining about above are defined and explained there. However, I would expect to get a solid idea of what is going on from reading the main alone, wihch is currently not the case.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 7GYW — Part 1/4

Dear Reviewer 7GYW,

Thank you for the careful reading and specific questions. Your comments identify definitions that should be clearer in the main text. We respond point by point below.

Q1. The readability especially of the method part up until the experiments is really lacking. There are many sequences of words that all roughly make sense, but do not map to concise and clear concepts or mathematical statements. It is hard to pin this down, but even reading carefully and refering to the appendix, it was not easy to understand what the method actually does. I certainly would have great trouble implementing the method just given the paper. I'll try to give examples of unclear/empty writing are "establishes a sharp, stable, mean anchor". This becomes clear later, but at this stage, the words "sharp" and "stable" do nothing but confuse. Similarly the "Factorized Dynamics Operator" remains fairly opaque to me throughout (More details later).

A1. Thank you for the concrete examples. OceanFlow learns a conditional ensemble around a corrected deterministic center. At each ocean transition:

  1. A frozen deterministic model advances perturbed ocean states under shared atmospheric forcing. Averaging these reference outputs and applying a learned, bounded correction gives the mean anchor μ\mu.
  2. Residual particles evolve through a learned velocity field conditioned on their member states, ocean-forcing inputs, flow time, and the current ensemble mean, standard deviation, and mean absolute deviation. These summaries are recomputed during transport, making member updates responsive to the evolving ensemble.
  3. Transported residuals are combined with reference-member anomalies and calibrated. Writing the resulting anomalies as d(m)d^{(m)}, the final output is x^(m)=μ+d(m)−M−1∑jd(j)\hat x^{(m)}=\mu+d^{(m)}-M^{-1}\sum_jd^{(j)}. Its ensemble mean therefore equals μ\mu exactly.

Here state-conditioned identifies the ocean-forcing inputs; pathwise refers to residual evolution in auxiliary flow time; and coherent refers to the ensemble-conditioned law governing these updates together. The implementation uses two transport updates per ocean transition. The deterministic reference supplies the next day's member conditions; the transformed ensemble is the evaluated output.

“Sharp, stable mean anchor” was intended to describe an accurate center that residual transport cannot displace. We will replace these adjectives with the explicit construction and mean-preservation property, rather than suggest an additional mathematical guarantee. FDO supplies separate reference and velocity networks, whose operations are explained in A4. We will define these operations before introducing abstract terminology.

Q2. For me, the motivation is also not fleshed out well enough. Specifically, I am missing a stronger motivation for (a) why probabilistic modeling is so desirable and (b) why the deficencies of ensembling deterministic methods matter (I do understand they may have deficiencies, but why are they problematic). Here I'm looking for motivation that goes beyond "it empirically works better according to metrics that care about the distribution".

A2. The motivation is to distinguish possible future ocean states and their probabilities, not merely improve a distributional metric. Two forecasts can have the same mean temperature or current yet assign very different probabilities to unusually warm water or strong currents. A deterministic forecast cannot distinguish these cases. Likewise, averaging states with displaced fronts or currents need not produce a representative realization. An ensemble therefore needs both an informative center and plausible variability.

Perturbing a deterministic model is a useful starting point, but averaging its outputs does not establish whether the resulting uncertainty ranges are reliable. Too-narrow variability omits plausible outcomes; misplaced variability represents the wrong alternatives even when the mean is accurate. Independent sampling is not inherently invalid: the issue is that prescribed perturbations and deterministic training do not themselves optimize the resulting conditional distribution against verifying ocean states.

OceanFlow supervises both the center and transported ensemble against reanalysis, with residual updates conditioned on ensemble context. The original same-backbone, 90-day ablation illustrates this distinction (mean±std over 10 ICs):

MethodCRPS↓RankDev↓CalibErr↓
FDO+Perlin0.403±0.0020.204±0.0010.225±0.001
OceanFlow0.371±0.0020.088±0.0000.032±0.000

These scores assess agreement between the generated alternatives and verifying outcomes beyond central-state accuracy; the spectral and physical diagnostics additionally examine field structure. We will foreground this motivation before the architecture.

AuthorsOpenReview ↗

Response to Reviewer 7GYW — Part 2/4

Q3. One aspect that remained fairly unclear to be relatively deep into the paper is "where do we get the target distribution from to train a flow". I see how the mean just comes from averaging deterministic rollouts. But to train the residual ensemble, some desired target distribution must be known. Given the arguments about the deficiencies of simply ensembling deterministic rollouts from perturbed initial states, in my view this target should not come from the deterministic ensemble. I was trying to find this in "Coherent state-conditioned pathwise flow" at first, where there is only a cryptic description of "verifying center error" or later "where y is the verifying state" (one of these word constructions that aren't pinned down). I then tried to pin it down in terms of the training loss. But equation (13) suddenly drops a 4-component loss out of nowhere with just a single sentence describing in words what each term roughly intends to do. What justifies all these terms (beyond just empirical ablations), choosing the many parameters fells daunting, what exactly are the terms? Also, I still don't understand where the critical targets y come from.

A3. The critical target yy is the next-day GLORYS12 reanalysis ocean field, after the same anomaly conversion, normalization, and ocean masking as the model output. It is paired with ct=(Ot,Ft,Ft+1)c_t=(O_t,F_t,F_{t+1}) in the training data. “Verifying state” means this reference field, and “verifying center error” means y−μy-\mu. Neither denotes a prediction from the deterministic ensemble. The anchor itself is the reference ensemble mean plus a learned correction.

There is one verifying field per historical transition, not an observed ensemble of alternative futures for an identical condition. Conditional distribution learning uses many such input-target pairs. Our implementation combines an observation-informed residual transport target with direct scoring of the generated ocean ensemble:

Residual transport supervision. Let e=y−sg⁡(μ)e=y-\operatorname{sg}(\mu), where sg⁡\operatorname{sg} stops gradients. A scaled, bounded channel-wise ocean RMS of ee determines stargets_{\mathrm{target}}. With smooth normalized noise ϵ(m)\epsilon^{(m)}, the endpoints are

z_1^{(m)}=0.95s_{\mathrm{target}}\odot\epsilon^{(m)}+0.05r_{\mathrm{pat}}^{(m)}(e).$$ Here $r_{\mathrm{pat}}^{(m)}(e)$ is the scaled error field modulated by member coefficients that are centered across members and RMS-normalized. This is a constructed residual target informed by reanalysis error, not a claim that the true conditional distribution is observed. For $z_\tau=(1-\tau)z_0+\tau z_1$, flow matching regresses the velocity to $z_1-z_0$. **Generated-ensemble supervision.** During training, we also integrate the residual flow, form the output ensemble $\widehat X$, and compare it directly with $y$. In particular, the empirical fieldwise CRPS is $$S_M(\widehat{X},y)=\frac{1}{M}\sum_{m=1}^{M}\lVert\hat{x}^{(m)}-y\rVert_{1,o}-\frac{1}{2M^2}\sum_{m=1}^{M}\sum_{n=1}^{M}\lVert\hat{x}^{(m)}-\hat{x}^{(n)}\rVert_{1,o},$$ where the norm denotes the masked-grid absolute average. Its first term penalizes member-target discrepancy, while the second accounts for ensemble dispersion. Gradients pass through the integrated residual paths. Thus, training is not restricted to reproducing the constructed endpoint, and the deterministic ensemble is not treated as the desired truth. The four groups in the compact objective have the following definitions and roles; their component expressions are in the appendix subsection **Training Losses and Calibration**: |Term|Definition and purpose| |---|---| |Flow matching|Masked squared velocity error against $z_1-z_0$, supervising the residual transport.| |Mean supervision|Channel-averaged relative $L_2$ errors of the anchor and output mean, plus a penalty when mean squared error exceeds that of the reference center.| |Probability supervision|Empirical CRPS against $y$, plus $[S_M(\widehat X,y)-\operatorname{sg}(S_M(X_{\mathrm{base}},y))]_+$ to penalize degradation relative to the reference.| |Regularization|Member-wise relative $L_2$ error, spread errors against $s_{\mathrm{target}}$, and a penalty for falling below a prescribed fraction of reference spread.| The first three groups connect transport, center accuracy, and terminal distribution quality; the auxiliary penalties regularize member error and spread during finite-ensemble optimization. The weights balance these distinct quantities rather than introduce additional target distributions. We will state the source of $y$ and the endpoint construction before the first flow equation, and consolidate the implemented loss coefficients alongside their definitions so that the compact objective does not leave these choices implicit.
AuthorsOpenReview ↗

Response to Reviewer 7GYW — Part 3/4

Q4. As mentioned the factorized dynamics operator remains hard to understand. What is the "factorized" aspect, why is it a "dynamics operator"? This appears to be unusual wording for an autoregressive time-stepping. Equations (9) through (13) lack explanation and context in my view. (9) and (10) introduce a whole host of symbols that are not defined mathematically and the verbal description is fuzzy and ambiguous (e.g., "combines cross-factor interactions with state dependent selection"; what happens mathematically or in terms of implementation? or "integrates collective response into path state"; how does it "integrate"; what even is a patch state (we only have E^patch in (9) without any explanation), etc).

A4. Thank you for requesting a concrete explanation. “Factorized” refers to learned feature subspaces with separate transformations and explicit interactions. “Dynamics operator” denotes the conditional mapping used to model an ocean transition or a residual-flow velocity. It does not designate an additional numerical solver.

Patch representation and factors. In Eq. (9), Epatch\mathcal{E}^{\mathrm{patch}} is a patch encoder implemented by a strided convolution, producing a spatial feature grid H0∈RD×Hp×WpH_0\in\mathbb{R}^{D\times H_p\times W_p}. “Patch state” simply means this feature grid. The input XX contains ocean-forcing channels for the deterministic instance, and additionally residual and ensemble-summary channels for the velocity instance. The latter's context CC includes flow time and an ocean-forcing embedding; the deterministic instance uses a learned static embedding.

Within a block, normalized and context-modulated features UU undergo spatial processing M\mathcal{M} and a learned projection P\mathcal{P}, then split into KK equal-width factors. Each factor has its own transformation:

Fk=[P(M(U))]k,Gk=Tk(Fk),G~i=∑j=1KBijGj,B=I+0.1tanh⁡(A).F_k=[\mathcal{P}(\mathcal{M}(U))]_k,\qquad G_k=\mathcal{T}_k(F_k),\qquad \widetilde{G}_i=\sum_{j=1}^{K}B_{ij}G_j,\quad B=I+0.1\tanh(A).

Here AA is a learned matrix. Factors are learned subspaces, not predefined temperature, salinity, or current groups, and we do not assume that each factor identifies a particular physical process.

Interaction and state-dependent selection. A pointwise softmax gate Γ=softmax⁡k(G(U))\Gamma=\operatorname{softmax}_{k}(\mathcal{G}(U)) weights the interacting factors. The weighted factors are concatenated and projected back to the feature width, followed by residual feature updates. Thus, BB specifies learned cross-factor mixing, whereas Γ\Gamma makes each factor's contribution depend on the current input and spatial location. The mixing matrix itself is state independent. In Eq. (10), Pℓ,k\mathcal{P}_{\ell,k} collects factor induction, Tℓ,k\mathcal{T}_{\ell,k} the factor-specific transformation, Aℓ\mathcal{A}_{\ell} mixing and gating, and Φℓ\Phi_{\ell} the projected update and subsequent channel update. “Integrates” means incorporating these updates into HℓH_{\ell}, not performing a hidden time integration.

Output and state-conditioned response. The decoder R\mathcal R restores the field Y~\widetilde Y, and the final mapping is Y=Y~+γS(Y~,X).Y=\widetilde Y+\gamma\mathcal S(\widetilde Y,X). The response network takes the concatenation of the decoded field and original input, and γ\gamma is learned. This is a learned input-conditioned residual correction, rather than a fixed addition of the input state. Its purpose is to adapt the decoded factorized transition using information in the original condition. Separate networks with this architecture parameterize the deterministic reference and flow velocity; their weights are not shared.

The original deterministic ablations assess the complete factorized block and the output-response choices. Results below are mean±std across the same 10 ICs at 90 days:

VariantRMSE↓ACC↑
w/o Factorized Dynamics1.054±0.0080.188±0.009
w/o State Response0.934±0.0620.072±0.090
w/o Adaptive State Response0.844±0.0070.322±0.010
FDO0.825±0.0090.352±0.007

The response variants respectively remove the correction or replace it with a fixed ocean-state addition. The results support retaining a learned response alongside the factorized blocks. We will replace the ambiguous descriptions with these definitions in the main text, making clear which operations the abstract symbols represent.

AuthorsOpenReview ↗

Response to Reviewer 7GYW — Part 4/4

Q5. In Lemma 1, what is the relevant statement. Is the mere existence (i.e., that one obtains a valid conditional law) itself non-trivial (appears natural to me given that all required existence and uniqueness assumptions are dropped in). Or is the relevant statement just that it preserves the empirical mean? (Given the construction, this appears trivial as well?) The proof is not very human-friendly to me (artificial contrasts, additional unnecessary re-assurances about what's required when and where, etc). I don't see the relevance of all the mentioned equivariances.

A5. The lemma is a structural characterization of the ensemble generator, not a new general existence theorem. We agree that existence under the stated assumptions and mean preservation by centering are standard facts. Their role is to specify the claimed “coherent law.”

For a fixed transition condition, ensemble-conditioned updates define a map on the entire residual set. Transport, reference composition, calibration, and recentering together induce a conditional ensemble law with the learned anchor as its empirical mean for every realization. Dependence on S(Z)S(Z) specifies ensemble interaction; recentering separately constrains the output mean.

Permutation equivariance ensures that jointly relabeling members and their conditions only relabels the output: arbitrary member indices cannot change the generated ensemble. It is not independent evidence of physical validity. These structural properties also apply to the finite-step implementation, without requiring two updates to reproduce an exact continuous-time solution.

We will focus the statement and proof on these properties and their connection to the generator. Calibration and forecast skill are assessed empirically, not guaranteed by the lemma.

Q6. Unfortunately, I am not familiar enough with ocean modeling specifically. Hence, it is a bit hard for me to judge the relevance of the empirical settings and results. Could the authors put the chosen settings and results into perspective? How far are we from full-fledged ocean model components in state of the art numerical weather prediction models? Is the long-term goal to replace/augment existing ocean component models with these learned models? What would be the main challenges there?

A6. Our experiments study daily, global ocean evolution at 1° resolution, with 93 channels representing temperature, salinity, and two current components at 23 depths, plus sea-surface height. This is a conditional ocean component, not a complete operational coupled forecasting system.

The 90-/120-day simulations use prescribed future atmospheric forcing from reanalysis, testing ocean-state evolution without atmospheric forecasting error. They are not operational forecasts with unknown future forcing. The main 20-/30-day forecasting experiments instead use OneForecast atmospheric predictions. Additional 2024 experiments assess generalization beyond the main 2020 test year. We will make these distinct purposes explicit.

The longer-term goal is to augment numerical Earth-system modeling with efficient learned ensemble components. The current results support learning conditional ocean ensembles with a controlled center, rather than establish a drop-in replacement for a full numerical ocean component. Further requirements include higher-resolution and coastal evaluation, conservation and balance diagnostics, initialization through data assimilation, robustness under changing forcing, and stable ocean-atmosphere feedback. End-to-end probabilistic coupling must also propagate atmospheric uncertainty instead of conditioning all ocean members on shared forcing. These are concrete directions beyond the current conditional modeling setting.

Q7. Side note: I did look into the extensive appendix and I know that many objects I am complaining about above are defined and explained there. However, I would expect to get a solid idea of what is going on from reading the main alone, wihch is currently not the case.

A7. We agree that the main paper must stand on its own. The revisions described above will bring the essential definitions into the method section: the source of yy, the corrected mean anchor, the construction of residual endpoints, the ensemble-conditioned update, and the roles of the loss terms. We will also define the FDO symbols by their actual operations and state that the coherent law acts within each ocean transition. The appendix will retain detailed derivations and implementation settings, rather than carry definitions needed to understand the core method. These changes preserve the central idea while making its algorithmic content explicit.

We sincerely thank you again for your constructive suggestion. Your comments help clarify our motivation, method, and scope. If you have any questions, please feel free to let us know. We will gladly address them as thoroughly as possible.

Reviewer_GLYPOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes OceanFlow, a probabilistic ocean modeling framework that combines mean anchoring, ensemble-conditioned residual flow matching, and a Factorized Dynamics Operator (FDO). The method aims to preserve an accurate predictive center while learning state-dependent ensemble uncertainty. Experiments on global ocean simulation and forecasting demonstrate improvements in prediction accuracy, probabilistic calibration, and physical-consistency diagnostics, including evaluations on a held-out year.

Critical strengths
  1. I find probabilistic ocean modeling a meaningful research problem with clear practical relevance. The separation between mean prediction and uncertainty modeling is intuitive and well motivated, and the proposed framework provides a reasonable way to jointly address these objectives.

  2. I appreciate the broad evaluation across simulation and forecasting tasks, including probabilistic calibration, physical consistency, and out-of-sample testing. The generally consistent improvements across these metrics provide encouraging evidence for the effectiveness of the overall framework.

Critical weaknesses
  1. The main novelty lies in jointly modeling ensemble members through an ensemble-conditioned flow. However, the current ablations do not isolate the benefits of this coupling over independent conditional flow models. Moreover, the reported probabilistic and physical diagnostics provide limited direct evidence that OceanFlow better captures spatial and multivariate dependencies, leaving its central claim insufficiently validated.

  2. While OceanFlow achieves promising results, most probabilistic baselines rely on perturbations of deterministic models. Recent approaches such as FuXi-ONS already explore learned probabilistic ocean forecasting. A more direct comparison or clearer distinction would help establish the methodological advancement and empirical advantages of OceanFlow over these approaches.

Questions for authors
  1. How much improvement comes from ensemble conditioning itself? Would independent memberwise flow achieve similar results?

  2. Why are learned ensemble models such as FuXi-ONS not included in the comparisons?

  3. Beyond marginal calibration, what evidence shows that OceanFlow better captures spatial and multivariate dependencies?

5 comments
AuthorsOpenReview ↗

Response to Reviewer GLYP — Part 1/3

Dear Reviewer GLYP,

We are truly grateful for the time you have taken to review our paper and your insightful review. Here we respond to your questions point by point.

Q1. The main novelty lies in jointly modeling ensemble members through an ensemble-conditioned flow. However, the current ablations do not isolate the benefits of this coupling over independent conditional flow models. How much improvement comes from ensemble conditioning itself? Would independent memberwise flow achieve similar results?

A1. Thank you for this important question. To examine whether independent conditional generation can achieve comparable performance, we trained an additional Direct Conditional Flow Matching (Direct FM) variant.

(a) Comparison with independent conditional generation.

Direct FM uses a conditional U-Net to transport Gaussian noise directly to the next ocean state, conditioned on the same ocean-state and atmospheric-forcing input variables. Members share network weights but evolve without access to other members or ensemble summaries, and each member's output provides its ocean-state condition for the next autoregressive step. We train this variant from scratch for 53 epochs on the same training dataset, matching the cumulative data exposure of OceanFlow's 50-epoch deterministic stage and 3-epoch residual stage. Both methods are evaluated at the 90-day horizon on the same 10 ICs, with 20 ensemble members and two Euler integration steps per transition. Results are reported as mean and standard deviation across ICs.

MethodRMSE↓ACC↑CRPS↓RankDev↓SSR→1CalibErr↓
Direct FM1.616±0.0030.016±0.0010.482±0.0010.161±0.0000.230±0.0030.183±0.000
OceanFlow0.684±0.0050.425±0.0020.371±0.0020.088±0.0000.708±0.0030.032±0.000

Under this protocol, independent direct generation does not achieve comparable long-range performance: OceanFlow reduces RMSE by 57.7% and CRPS by 23.0%, improves rank and coverage calibration, and brings SSR closer to one. The gains therefore extend beyond the ensemble mean to the quality of the generated distribution.

(b) Ablations supporting mean-anchored distribution learning.

Our central motivation is that creating ensemble spread is not equivalent to learning a calibrated conditional distribution. OceanFlow addresses this distinction by establishing a controlled mean anchor and learning a coherent residual law around it. The existing 90-day, 10-IC ablations substantiate this design in three complementary ways:

  • FDO+Perlin directly tests the perturbation-driven alternative in our motivation. It uses the same deterministic backbone and structured perturbation reference, generating spread without explicitly training the resulting ensemble distribution. Its CRPS, RankDev, and CalibErr are 0.403, 0.204, and 0.225, compared with 0.371, 0.088, and 0.032 for OceanFlow in the original table. These improvements support the shift from prescribing ensemble diversity to learning how that diversity represents conditional uncertainty. Since the deterministic backbone is unchanged, the gains cannot be explained by a stronger deterministic model alone.
  • w/o mean correction/re-centering tests the role of the mean anchor in this distributional learning. Although learned residual transport is retained, removing mean correction and re-centering worsens ensemble-mean accuracy and increases CRPS to 0.376 and CalibErr to 0.059. This supports explicitly correcting and preserving the center around which residual uncertainty is learned: residual transport benefits from a controlled predictive center, rather than replacing the need for one.
  • w/o learned pathwise flow tests whether mean control alone resolves the limitations of the perturbation reference. It retains mean correction/re-centering but removes learned residual transport, yielding CRPS, RankDev, and CalibErr of 0.379, 0.109, and 0.081, respectively. Thus, establishing a mean anchor does not by itself recover the full method's distributional quality. Learning the residual law provides additional calibration gains beyond correcting the center of a perturbation-driven ensemble.

(c) Implication for the core idea.

Together, these results support the central premise of OceanFlow: probabilistic ocean modeling benefits from learning the conditional ensemble distribution around a controlled mean anchor, rather than relying on perturbation-induced spread alone. The Direct FM comparison supports this mean-anchored residual formulation over direct state generation, while the existing ablations show that both mean control and learned residual transport contribute to its performance. Coherent flow matching brings these roles together by jointly refining member residuals under a state-conditioned law while preserving the predictive center.

AuthorsOpenReview ↗

Response to Reviewer GLYP — Part 2/3

Q2. While OceanFlow achieves promising results, most probabilistic baselines rely on perturbations of deterministic models. Recent approaches such as FuXi-ONS already explore learned probabilistic ocean forecasting. A more direct comparison or clearer distinction would help establish the methodological advancement and empirical advantages of OceanFlow over these approaches. Why are learned ensemble models such as FuXi-ONS not included in the comparisons?

A2. Thank you for highlighting this important direction. We agree that learned probabilistic models are valuable comparisons. As detailed in A1, we have added Direct Conditional Flow Matching (Direct FM), which learns a conditional noise-to-state transport rather than generating ensembles through prescribed perturbations of a deterministic model. At the 90-day horizon, OceanFlow achieves a CRPS of 0.371 compared with 0.482 for Direct FM, together with better ensemble-mean accuracy and calibration diagnostics. This adds a learned generative comparison beyond the perturbation-based ensembles in the original experiments.

For FuXi-ONS specifically, we could not locate publicly available official code or pretrained checkpoints at the time of this response. Its public manuscript does not provide a code or checkpoint download link. This practical limitation prevented a direct evaluation of the authors' implementation under our protocol.

We recognize FuXi-ONS as an important advance in learned probabilistic ocean forecasting. It moves beyond prescribed noise by learning state-dependent, spatially structured perturbations and incorporating atmospheric conditioning for long-range ensemble prediction. Our work addresses a complementary question: how can the ensemble distribution itself be jointly refined while its predictive center remains explicitly controlled?

OceanFlow learns a coherent state-conditioned pathwise law for member residuals around a mean anchor. Each member's velocity responds to the evolving ensemble context, probabilistic objectives supervise the transported distribution, and final re-centering preserves the anchor. The distinction is therefore in the modeling focus: FuXi-ONS learns structured perturbations for ensemble generation, whereas OceanFlow emphasizes ensemble-conditioned residual transport and distributional refinement under explicit mean control. The Direct FM comparison in A1 and the existing component ablations provide empirical support for this formulation, complementing the methodological distinction from FuXi-ONS.

AuthorsOpenReview ↗

Response to Reviewer GLYP — Part 3/3

Q3. Moreover, the reported probabilistic and physical diagnostics provide limited direct evidence that OceanFlow better captures spatial and multivariate dependencies, leaving its central claim insufficiently validated. Beyond marginal calibration, what evidence shows that OceanFlow better captures spatial and multivariate dependencies?

A3. Thank you for emphasizing the distinction between marginal calibration and dependence structure. Our evaluation includes complementary evidence in Section 3.5, "Analysis of Physical Consistency," specifically Figure 7 (spectral analysis) and Figure 8 (physical diagnostics). The diagnostic definitions are provided in Appendix C, "Evaluation Metrics." For convenience, the corresponding figures can be viewed directly through these links:

Multivariate dependencies. The middle panel of Figure 8 compares latitude-weighted spatial correlations between zonal current anomalies and the negative meridional SSH-anomaly gradient, and between meridional current anomalies and the zonal SSH-anomaly gradient, against the corresponding correlations in the verifying fields. These correlations are computed from the ensemble-mean fields over valid ocean points away from the equator. OceanFlow achieves the lowest error among the methods shown, 0.023, compared with 0.032 for DSLCast+Perlin. This directly assesses a dynamically relevant cross-variable relationship, rather than separate field-wise errors or marginal calibration.

Member-level spatial structure. Figure 7 evaluates the spatial energy distribution across wavenumbers at the 120-day horizon. The dashed curves are computed from each member separately and then averaged over 20 members and 20 ICs, avoiding the smoothing introduced by averaging members before computing spectra. OceanFlow's member-averaged spectra are overall closer to the verifying spectra across the representative surface variables. The right panel of Figure 8 complements this analysis by evaluating each member's spatial standard deviation before averaging across members: OceanFlow obtains an error of 0.043, compared with 0.128 for DSLCast+Perlin. Together, these diagnostics examine both the distribution of spatial energy across wavenumbers and the overall amplitude of member-level spatial variability. The left panel additionally checks the temporal consistency of the ensemble center through SST-anomaly mean tendencies. Thus, our evidence extends beyond marginal calibration to multivariate dependencies in the ensemble center and spatial structure within individual members, supporting improvements in the structure of the generated ocean fields rather than only the amount of ensemble spread.

We sincerely thank you again for your constructive suggestion. Your comments have helped us clarify the role of ensemble conditioning, the comparison with learned probabilistic models, and the evidence for spatial and multivariate dependencies. If you have any questions, please feel free to let us know. We will gladly address them as thoroughly as possible.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.