Accurate Mediterranean Sea forecasting via graph-based deep learning

Daniel Holmberg; Emanuela Clementi; Italo Epicoco; Teemu Roos

REVIEW 3 major objections 5 minor 58 references

Accurate Mediterranean Sea forecasting via graph-based deep learning

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read SeaCast, a graph neural network for the Mediterranean Sea, produces 15-day, 1/24-degree forecasts that match or beat the operational MedFS system across all modeled variables and depth levels, in 20 seconds on a single GPU.

desk verdict SeaCast looks like a real step forward for data-driven regional ocean forecasting, but the headline 'outperforms MedFS everywhere' claim needs the evaluation window widened and the circularity of the subsurface reference addressed before it fully lands. read the letter →

arxiv 2506.23900 v1 pith:MJPKR25M submitted 2025-06-30 physics.ao-ph

Daniel Holmberg , Emanuela Clementi , Italo Epicoco , Teemu Roos This is my paper

classification physics.ao-ph

keywords regionaloceanforecastinggraphneuralnetworkslearnedsimulationMediterraneanSeaMedFSdata-drivenmarineheatwavesatmosphericforcing

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper introduces SeaCast, a graph neural network that forecasts the Mediterranean Sea on a 1/24-degree grid over 18 depth levels, predicting currents, temperature, salinity, and sea surface height up to 15 days ahead. The central claim is that SeaCast consistently beats the operational Mediterranean Forecasting System (MedFS) in forecast skill, across all modeled variables and depth levels, whether the reference is daily analysis fields or satellite observations, with the gap widening at longer lead times. The paper also reports a large computational advantage: a full 15-day forecast runs in about 20 seconds on a single GPU, versus roughly 70 minutes for a 10-day MedFS forecast on 89 CPU cores. If the claim holds, data-driven models become a practical option for high-resolution operational ocean forecasting, enabling rapid scenario testing, large ensembles, and earlier warnings of marine heatwaves.

What carries the argument

The central mechanism is a hierarchical graph neural network on a three-level mesh that conforms to the irregular coastline of the Mediterranean. Input sea states and atmospheric forcing are projected from the 1/24-degree grid onto the mesh, passed through GNN layers that exchange messages between progressively coarser levels, and decoded back to the grid; the model learns the one-day tendency as a residual added to the current state. At each step, lateral boundary nodes at the Strait of Gibraltar and the Dardanelles Strait are overwritten with external forecast data, and the mesh keeps only ocean grid nodes, dropping edges that cross land, which is what lets an architecture developed for global weather models handle a regional ocean with complex geometry.

What would settle it

Re-run both models with initializations spread across all four seasons and verify every subsurface variable against independent in-situ profiles (Argo floats and CTD casts not assimilated into, and withheld from, the reference system) over the full 15-day horizon; if SeaCast's normalized RMSE advantage over MedFS disappears or reverses once the reference is purely observational and spans winter, the claim of consistent superiority is refuted.

Watch

Extended reading notes

Core claim

SeaCast is an autoregressive model built on an encode–process–decode architecture with a hierarchical mesh over the Mediterranean basin; it predicts the one-day tendency of the sea state, which is added to the current state and fed back for multi-step forecasting. The paper claims it delivers improved forecast skill relative to MedFS for all modeled variables across all depth levels, with the largest gains at longer lead times, verified for subsurface fields against daily MedFS analysis fields and for SST and sea level anomaly against independent satellite observations. SeaCast also edges out MedFS at detecting SST extremes above the 90th percentile. Controlled ablations show that permuting wind stress degrades skill most for currents, temperature, and salinity, while mean sea level pressure dominates sea level anomaly, consistent with inverted-barometer physics; training-period experiments show that a variant trained on only 10 years of data matches MedFS on currents, temperature, and SST, while the full 35-year model with fine-tuning on recent analysis data is required to beat it on salinity and sea level anomaly.

Load-bearing premise

The claim that SeaCast beats the operational model below the surface rests on treating daily-mean MedFS analysis fields as accurate enough to serve as both the training target and the verification reference; if those fields carry systematic model biases, the measured advantage could be fidelity to the analysis system rather than genuine forecast skill, and the July–December 2024 evaluation window leaves winter and spring untested.

Editorial extensions

If this is right

A full 15-day regional forecast runs in roughly 20 seconds on a single GPU, against about 70 minutes for a 10-day MedFS forecast on 89 CPU cores, making many-member ensembles and rapid scenario testing operationally practical.
The 15-day horizon extends five days beyond MedFS, which the authors connect to earlier warnings of marine heatwaves and improved medium-range planning.
A version trained on only 10 years of reanalysis reaches MedFS-level skill for currents, temperature, and SST, so regions with limited historical archives and modest compute budgets could build competitive systems.
Wind stress is the dominant atmospheric forcing for currents, temperature, and salinity, so a data-driven ocean forecast inherits the accuracy of the surface wind product that drives it.
Fine-tuning on the analysis system used for initialization and verification consistently improves skill, so operational deployment must track upgrades to the data-assimilation system.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A decisive test would run both models from initializations spread across all four seasons, verified against in-situ profile data withheld from the analysis system; the July–December 2024 window does not exercise winter mixed-layer deepening, where model biases are largest.
The same hierarchy-and-boundary design should transfer to other semi-enclosed seas such as the Black Sea, Red Sea, or Baltic, where lateral open boundaries and irregular coastlines dominate; the authors leave this untested.
Because the training targets (Section 4.2) and the subsurface verification reference (Section 2.2) both come from the MedFS analysis system, a circularity the paper itself notes in Section 2.5, part of the measured advantage could be fidelity to the analysis rather than predictive skill; an observation-only benchmark would separate the two.
The ablation results suggest a concrete extension: feeding wind forcing at 6-hourly rather than daily resolution, which the paper lists as future work, should improve near-surface skill if wind-stress fidelity is the bottleneck.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: this is the first depth-resolved, 1/24-degree data-driven regional ocean forecasting system I know of, and over the tested period it appears to beat the operational MedFS on most skill metrics. The architecture is a sensible adaptation of hierarchical GNNs to an irregular ocean mesh, with lateral boundary forcing handled properly, and the evaluation is more careful than most ML-forecasting papers I see. Code, weights, and preprocessed data are all released, which deserves credit.

What's genuinely new: adapting the hierarchical GNN framework to a regional ocean with open boundaries and atmospheric forcing, and showing that even a 10-year training variant stays competitive with a physics-based operational system. The compute speedup is real—20 seconds on one GPU versus 70 minutes on 89 CPU cores—and the paper's ablations on atmospheric forcing and training length are thoughtful.

Soft spots, in proportion. The evaluation window is only July to December 2024. Six months, mostly the stratified warm season, tells you little about winter mixed-layer dynamics or spring bloom conditions. The paper's Discussion claim of improved skill for 'all modeled variables across all depth levels' is broader than the evidence. The subsurface reference is the MedFS daily-mean analysis, which is also the training target. That is partial circularity, not necessarily fatal—the supplementary in-situ comparison points the same direction, and the satellite SST/SLA checks are independent—but it does not justify the sweeping wording. Also, the MedFS baseline stops at day 10, so the 15-day skill is only compared against persistence, not against any physics model, at those leads. Minor, but worth noting.

Would I send this to peer review? Yes. The paper deserves serious refereeing. The main things I'd ask referees to push on: extend the evaluation into winter/spring if the data allow, or at least clearly frame the claim as season-specific; bring the in-situ subsurface comparison into the main text; and soften the 'all variables, all depths' phrasing to match what is actually verified. The core result is plausible and important, and the authors have done the reproducibility work that makes a referee's job honest.

Referee Report

3 major / 5 minor

Summary. SeaCast is an autoregressive graph neural network for Mediterranean Sea forecasting on a 1/24-degree grid with 18 depth levels and a 15-day lead time. The model is trained on 35 years of MedFS reanalysis and fine-tuned on two years of operational analysis, with atmospheric forcing from ERA5, ENS, or AIFS and lateral boundary forcing from MedFS. It is evaluated against the operational MedFS over initializations from July 3 to December 31, 2024, using RMSE, HSS, satellite SST/SLA references, and a supplementary in-situ comparison, with a persistence baseline and bootstrap confidence intervals. The central claim is that SeaCast consistently outperforms MedFS in forecast skill while reducing wall-clock time from about 70 minutes on 89 CPU cores to 20 seconds on one GPU.

Significance. If the central claim holds, this is a significant demonstration that a data-driven regional ocean model can match or exceed an operational physics-based system at high resolution, with large computational savings and an extended forecast horizon. The paper has several concrete strengths: the evaluation uses the same initial conditions as MedFS, includes a persistence baseline, provides bootstrap intervals, validates surface fields against independent satellite observations, includes an in-situ comparison in the supplement, reports ablations for atmospheric forcing and training-period length, and makes code, weights, and preprocessed data publicly available. These ingredients make the study reproducible and the comparison design largely fair. However, the strongest claims about subsurface skill rest on a partially circular verification setup and a single-season evaluation window, which must be addressed before the headline conclusion is fully supported.

major comments (3)

[§2.2, §4.2, and §3] The primary subsurface evaluation is circular in an important sense. Section 2.2 states that 'Subsurface variables are validated against daily mean MedFS analysis fields,' while Section 4.2 explains that SeaCast is trained on the same MedFS reanalysis/analysis system and fine-tuned on 2022–2023 daily operational analysis. The Discussion then claims 'improved skill for all modeled variables across all depth levels compared to the operational MedFS.' If the MedFS analysis fields carry systematic biases, SeaCast can obtain lower RMSE against those fields simply by reproducing analysis-like states, whereas the physically evolved MedFS forecast may drift away from the analysis toward the true ocean state. The satellite SST and SLA checks do not constrain the subsurface. The in-situ comparison in Supplementary D.1 is a step in the right direction, but it is not part of the main claim and is limited in spatial and temporal sampling. To support the headline claim, the authors should either promote and substantially expand the independent in-situ evaluation (e.g., against Argo, moorings, and gliders with coverage statistics) or explicitly qualify the subsurface claim as skill relative to the analysis reference rather than to the true ocean state.
[§4.2 and §2.2] The evaluation window covers only July 3 to December 31, 2024. The abstract and Discussion state that SeaCast 'consistently outperforms' MedFS without a seasonal qualifier, but the test period excludes winter and spring, when mixed-layer deepening, deep convection, and different atmospheric forcing regimes are most active. The reported skill may not generalize to those seasons. The authors should either extend the evaluation to a full year (or at least include boreal winter and spring dates) or restrict the claimed superiority to the late-summer/autumn period actually tested.
[§3 and Fig. 3] The Discussion's assertion of 'improved skill for all modeled variables across all depth levels' is stronger than what Fig. 3 appears to show. The figure displays depth- and lead-dependent normalized RMSE differences and the text acknowledges that at the lowest level (192 m) SeaCast 'generally does not outperform MedFS that much.' Please either quantify the fraction of level–lead combinations with negative differences, report the uncertainty on each cell, or soften the claim to specify the regions and depths where the improvement is robust. As written, the phrase 'all depth levels' outruns the evidence presented.

minor comments (5)

[Abstract and §3] The abstract and concluding statements should include the evaluation period ('July–December 2024') and the reference systems used for verification, so that readers do not infer year-round or fully independent validation.
[§2.2 and §4.2] The statement that lateral boundary conditions are 'persisted' for the last five days is an ad hoc extension from 10 to 15 days. Please note this as an explicit assumption or provide a sensitivity test showing that this choice does not materially affect the extended-lead skill.
[Eq. (2)] There is a typo in the text near Eq. (2): 'Nl is the the number of ocean grid nodes' should read 'Nl is the number of ocean grid nodes.'
[Fig. 2 and §2.2] The 50% bootstrap intervals are reported, but for the central 'outperforms' claim a more standard significance assessment (e.g., 95% intervals or paired tests across initializations) would strengthen the interpretation, especially for variables where the two curves are close.
[Supplementary D.1] The in-situ evaluation would be much more informative if it reported the number and geographic/depth distribution of profiles per lead time, and if it were summarized in the main text as an independent check on the subsurface claim.

Circularity Check

2 steps flagged · score 4.0 of 10

Subsurface skill claim is partially circular: SeaCast is trained and evaluated on the same MedFS analysis fields, and the reported RMSE is the training objective applied to a later interval.

other [Sections 2.1, 2.2, and 4.2; Eq. (2); Supplementary Eq. (1)]
"SeaCast is trained on 35 years (1987–2021) of Mediterranean reanalysis daily mean data [28] and fine-tuned using two additional years (2022–2023) of daily operational analysis [10]. ... Subsurface variables are validated against daily mean MedFS analysis fields."

The training loss (Eq. 2) and the evaluation RMSE (Supp. Eq. 1) are the same MSE/RMSE functional evaluated against the same X target: both compare predicted fields to MedFS reanalysis/analysis states. The depth-resolved 'consistently outperforms MedFS' claim therefore partly measures how well SeaCast reproduces the very product it was trained on, on a later interval, rather than independent skill against the true ocean. This is not identity-forced because MedFS is scored on the same reference and SeaCast could fail, but the evaluation reference is the same modeling system as the training target, so the subsurface comparison is statistically favored by construction of the metric.
other [Section 2.5]
"Fine-tuning on analysis data yields consistent performance improvements across variables, which is expected given that several evaluation targets are derived from the same analysis system."

The paper itself attributes the fine-tuning gains to the evaluation targets being derived from the same analysis system used for fine-tuning. This concedes that part of the reported improvement is an artifact of the shared reference and does not by itself demonstrate better fidelity to the true ocean state. It is an in-manuscript admission of the training/evaluation overlap that underlies the main subsurface skill claim.

full rationale

SeaCast is trained on 35 years of MedFS reanalysis plus two years of MedFS analysis, and the primary depth-resolved evaluation is against daily mean MedFS analysis fields (Section 2.2). The training loss (Eq. 2) and the reported RMSE (Supp. Eq. 1) are the same functional of the same product, so the subsurface 'outperformance' is in part a measure of how well the model reproduces its training product on a later period, not an independent measure of true-ocean skill. The paper explicitly acknowledges this in Section 2.5 ('expected given that several evaluation targets are derived from the same analysis system'). This is partial, not full, circularity: SST and SLA are checked against independent satellite observations (Fig. 2e,f), the supplement includes an in-situ comparison (Supp. D.1), MedFS is also scored against the same analysis fields, and the model weights and code are released for independent reproduction. The 'stable reference' assumption for subsurface analysis fields is a real limitation that should be weighed in correctness risk, but the model is not simply renaming its inputs. Score 4 reflects the load-bearing training/evaluation overlap while crediting the independent surface and supplementary in-situ checks.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

SeaCast is an empirical machine-learning system, so its central claim relies less on mathematical axioms and more on the validity of the training data, the fairness of the evaluation, and modeling choices for the Mediterranean basin. The entries above list the fitted or hand-set weights and the domain assumptions that would break the result if violated.

free parameters (4)

Learned network weights = 17.7 million trainable parameters
All model parameters are fitted to 35 years of reanalysis and 2 years of analysis. The central forecast skill is carried by these fitted values, so the headline result is an empirical fitting claim rather than a first-principles derivation.
Depth-level loss weights w_l = proportional to (200 - depth), normalized to unit mean
Hand-chosen in the loss function (Eq. 2) to prioritize the upper ocean. Different weights would change what the model optimizes and the resulting depth-resolved skill.
Per-variable loss scales lambda_i = inverse variance of time differences for each variable
Computed from training data in Eq. (2) to normalize error scales across variables. This is a data-derived normalization, not a physical constant.
Training schedule hyperparameters = warmup to 1e-3, cosine decay, 200 pretrain + 30 finetune epochs, rollout steps up to 3
Selected by hand and not swept; they affect the final model and the comparison, although they are standard choices in the ML weather forecasting literature.

assumptions (5)

domain assumption Daily-mean MedFS reanalysis and analysis fields are accurate enough to serve as training targets and verification truth for subsurface variables.
Section 4.2 describes the reanalysis/analysis system; Section 2.2 uses MedFS analysis fields as the reference for depth-resolved forecasts. Systematic biases in these fields would be inherited by the trained model.
domain assumption Lateral boundary forcing from MedFS can be treated as known during evaluation.
Equation (1) overwrites predictions inside boundary regions with MedFS forecast data. SeaCast therefore depends on the operational model at Gibraltar and the Dardanelles straits.
ad hoc to paper Persisting the last boundary state five times is a valid extension from 10 to 15 days.
Section 4.2 introduces the persistence of the boundary forcing without independent validation; the 15-day forecast claim relies on it.
domain assumption The July-December 2024 evaluation period is representative of year-round model skill.
The test set covers only summer-autumn conditions. Winter and spring dynamics, including mixed-layer deepening and different atmospheric regimes, are not covered by the headline comparison.
domain assumption The hand-built hierarchical mesh represents the relevant ocean connectivity.
Section 4.1 describes mesh construction with edges crossing land excluded by a threshold. The graph structure determines which spatial interactions the GNN can represent and is not validated against alternative meshes.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Accurate Mediterranean Sea forecasting via graph-based deep learning." pith.science (2026). https://pith.science/paper/MJPKR25M

@misc{pith2026250623900,
  author       = {Pith},
  title        = {Pith review of: Accurate Mediterranean Sea forecasting via graph-based deep learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MJPKR25M}},
  note         = {Machine review of arXiv:2506.23900}
}

read the original abstract

Accurate ocean forecasting systems are essential for understanding marine dynamics, which play a crucial role in sectors such as shipping, aquaculture, environmental monitoring, and coastal risk management. Traditional numerical solvers, while effective, are computationally expensive and time-consuming. Recent advancements in machine learning have revolutionized weather forecasting, offering fast and energy-efficient alternatives. Building on these advancements, we introduce SeaCast, a neural network designed for high-resolution regional ocean forecasting. SeaCast employs a graph-based framework to effectively handle the complex geometry of ocean grids and integrates external forcing data tailored to the regional ocean context. Our approach is validated through experiments at a high horizontal resolution using the operational numerical forecasting system of the Mediterranean Sea, along with both numerical and data-driven atmospheric forcings. Results demonstrate that SeaCast consistently outperforms the operational model in forecast skill, marking a significant advancement in regional ocean prediction.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

58 extracted references · 43 canonical work pages

[1]

Le Traon, P. Y. et al. From observation to information and users: The Copernicus Marine Service perspective. Frontiers in Marine Science 6, 234 (2019)

work page 2019
[2]

Sun, R. et al. SKRIPS v1.0: A regional coupled ocean–atmosphere modeling framework (MITgcm–WRF) using ESMF/NUOPC, description and preliminary result s for the Red Sea. Geoscientiﬁc Model Development 12, 4221–4244 (2019)

work page 2019
[3]

Sakamoto, K. et al. Development of a 2-km resolution ocean model covering the coastal seas around Japan for operational application. Ocean Dynamics 69, 1181–1202 (2019). 14

work page 2019
[4]

Ciliberti, S. A. et al. Monitoring and forecasting the ocean state and biogeochemical pro- cesses in the Black Sea: Recent developments in the Copernicus Mar ine Service. Journal of Marine Science and Engineering 9, 1146 (2021)

work page 2021
[5]

Nemo-Nordic 2.0: Operational marine forecast model for the Baltic Sea

K¨ arn¨ a, T.et al. Nemo-Nordic 2.0: Operational marine forecast model for the Baltic Sea. Geoscientiﬁc Model Development 14, 5731–5749 (2021)

work page 2021
[6]

Zhu, X. et al. Improvements in the regional South China Sea operational oceanography forecasting system (SCSOFSv2). Geoscientiﬁc Model Development 15, 995–1015 (2022)

work page 2022
[7]

Bruschi, A. et al. Indexes for the assessment of bacterial pollution in bathing waters fr om point sources: The northern Adriatic Sea CADEAU service. Journal of Environmental Management 293, 112878 (2021)

work page 2021
[8]

Liubartseva, S. et al. Modeling chronic oil pollution from ships. Marine Pollution Bulletin (2023)

work page 2023

Show all 58 references

[9]

L., Carelli, L., Petacco, N

Mannarini, G., Salinas, M. L., Carelli, L., Petacco, N. & Orovi´ c, J. VI SIR-2: Ship weather routing in Python. Geoscientiﬁc Model Development 17, 4355–4382 (2024)

2024
[10]

Coppini, G. et al. The Mediterranean forecasting system—Part 1: Evolution and performance. Ocean Science 19, 1483–1516 (2023)

2023
[11]

Bi, K. et al. Accurate medium-range global weather forecasting with 3D neural networks . Nature 619, 533–538 (2023)

2023
[12]

Nguyen, T. et al. Scaling transformer neural networks for skillful and reliable mediu m- range weather forecasting. Advances in Neural Information Processing Systems 37, 68740– 68771 (2024)

2024
[13]

Pathak, J. et al. FourCastNet: A global data-driven high-resolution weather model using adaptive Fourier neural operators. arXiv preprint arXiv:2202.11214 (2022)

2022 arXiv
[14]

Forecasting global weather with graph neural networks

Keisler, R. Forecasting global weather with graph neural networks. arXiv preprint arXiv:2202.07575 (2022)

2022 arXiv
[15]

Lam, R. et al. Learning skillful medium-range global weather forecasting. Science 382, 1416–1421 (2023)

2023
[16]

& Lindsten, F

Oskarsson, J., Landelius, T., Deisenroth, M. & Lindsten, F. Probab ilistic weather forecast- ing with hierarchical graph neural networks. Advances in Neural Information Processing Systems 37, 41577–41648 (2024)

2024
[17]

Dheeshjith, S. et al. Samudra: An AI global ocean emulator for climate. Geophysical Research Letters 52, e2024GL114318 (2025)

2025
[18]

Wang, C. et al. Coupled ocean-atmosphere dynamics in a machine learning Earth syste m model. arXiv preprint arXiv:2406.08632 (2024)

2024 arXiv
[19]

Guo, Z. et al. Data-driven global ocean modeling for seasonal to decadal prediction. arXiv preprint arXiv:2405.15412 (2024)

2024 arXiv
[20]

Wang, X. et al. XiHe: A data-driven model for global ocean eddy-resolving forecasting. arXiv preprint arXiv:2402.02995 (2024)

2024 arXiv
[21]

Aouni, A. E. et al. GLONET: Mercator’s end-to-end neural forecasting system. arXiv preprint arXiv:2412.05454 (2024). 15

2024
[22]

Cui, Y. et al. Forecasting the eddying ocean with a deep neural network. Nature Communications 16, 2268 (2025)

2025
[23]

Andersson, T. R. et al. Seasonal Arctic sea ice forecasting with probabilistic deep learning. Nature Communications 12, 5124 (2021)

2021
[24]

Chattopadhyay, A., Gray, M., Wu, T., Lowe, A. B. & He, R. OceanNet: A pr incipled neural operator-based digital twin for regional oceans. Scientiﬁc Reports 14, 21181 (2024)

2024
[25]

& Zanna, L

Subel, A. & Zanna, L. Building ocean climate emulators. arXiv preprint arXiv:2402.04342 (2024)

2024 arXiv
[26]

Sanchez-Gonzalez, A. et al. Learning to simulate complex physics with graph networks. International Conference on Machine Learning 8459–8468 (2020)

2020
[27]

Marchesiello, P., McWilliams, J. C. & Shchepetkin, A. Open boun dary conditions for long-term integration of regional oceanic models. Ocean Modelling 3, 1–20 (2001)

2001
[28]

Escudier, R. et al. A high resolution reanalysis for the Mediterranean Sea. Frontiers in Earth Science 9, 702285 (2021)

2021
[29]

Hersbach, H. et al. ERA5 monthly averaged data on single levels from 1940 to present. Copernicus Climate Change Service (C3S) Climate Data Store (CDS) (2023)

2023
[30]

Molteni, F., Buizza, R., Palmer, T. N. & Petroliagis, T. The ECMWF e nsemble predic- tion system: Methodology and validation. Quarterly Journal of the Royal Meteorological Society 122, 73–119 (1996)

1996
[31]

Lang, S. et al. AIFS - ECMWF’s data-driven forecasting system. arXiv preprint arXiv:2406.01465 (2024)

2024 arXiv
[32]

Plans for high resolution forecast (HRES) and ensemble forecas t (ENS)

ECMWF. Plans for high resolution forecast (HRES) and ensemble forecas t (ENS). In focus (2024). URL https://www.ecmwf.int/en/about/media-centre/focus/2024/ plans-high-resolution-forecast-hres-and-ensemble-forecast-en s

2024
[33]

B., Tronconi, C., Pisano, A

Nardelli, B. B., Tronconi, C., Pisano, A. & Santoleri, R. High and Ultra-High resolution processing of satellite Sea Surface Temperature data over Souther n European Seas in the framework of MyOcean project. Remote Sensing of Environment 129, 1–16 (2013)

2013
[34]

Hobday, A. J. et al. A hierarchical approach to deﬁning marine heatwaves. Progress in Oceanography 141, 227–238 (2016)

2016
[35]

& Schott, F

Marshall, J. & Schott, F. Open-ocean convection: Observations, th eory, and models. Reviews of Geophysics 37, 1–64 (1999)

1999
[36]

G., McWilliams, J

Large, W. G., McWilliams, J. C. & Doney, S. C. Oceanic vertical mixi ng: A review and a model with a nonlocal boundary layer parameterization. Reviews of Geophysics 32, 363–403 (1994)

1994
[37]

inve rted barometer

Wunsch, C. & Stammer, D. Atmospheric loading and the oceanic “inve rted barometer” eﬀect. Reviews of Geophysics 35, 79–107 (1997)

1997
[38]

Gill, A. E. Atmosphere—ocean dynamics (Academic Press, 1982)

1982
[39]

Clementi, E. et al. Coupling hydrodynamic and wave models: ﬁrst step and sensitivit y experiments in the Mediterranean Sea. Ocean Dynamics 67, 1293–1312 (2017). 16

2017
[40]

McDonagh, B., Clementi, E., Goglio, A. C. & Pinardi, N. The characteris tics of tides and their eﬀects on the general circulation of the Mediterranean Sea. Ocean Science 20, 1051–1066 (2024)

2024
[41]

R¨ uhling Cachay, S., Zhao, B., Joren, H. & Yu, R. Dyﬀusion: A dynami cs-informed dif- fusion model for spatiotemporal forecasting. Advances in Neural Information Processing Systems 36, 45259–45287 (2023)

2023
[42]

& Lindsten, F

Andrae, M., Landelius, T., Oskarsson, J. & Lindsten, F. Continuous e nsemble weather forecasting with diﬀusion models. International Conference on Learning Representations (2025)

2025
[43]

Nipen, T. N. et al. Regional data-driven weather modeling with a global stretched-grid. arXiv preprint arXiv:2409.02891 (2024)

2024 arXiv
[44]

Adamov, S. et al. Building machine learning limited area models: Kilometer-scale weather forecasting in realistic settings. arXiv preprint arXiv:2504.09340 (2025)

2025 arXiv
[45]

Nguyen, T., Brandstetter, J., Kapoor, A., Gupta, J. K. & Grover, A. C limaX: A foun- dation model for weather and climate. International Conference on Machine Learning (2023)

2023
[46]

Bodnar, C. et al. A foundation model for the Earth system. Nature 1–8 (2025)

2025
[47]

Price, I. et al. Probabilistic weather forecasting with machine learning. Nature 637, 84–90 (2025)

2025
[48]

& Lindsten, F

Larsson, E., Oskarsson, J., Landelius, T. & Lindsten, F. Diﬀusion-LAM : Probabilistic limited area weather forecasting with diﬀusion. arXiv preprint arXiv:2502.07532 (2025)

2025 arXiv
[49]

& Battaglia, P

Fortunato, M., Pfaﬀ, T., Wirnsberger, P., Pritzel, A. & Battaglia, P . Multiscale meshgraphnets. arXiv preprint arXiv:2210.00612 (2022)

2022 arXiv
[50]

Battaglia, P., Pascanu, R., Lai, M., Jimenez Rezende, D. et al. Interaction networks for learning about objects, relations and physics. Advances in Neural Information Processing Systems 29 (2016)

2016
[51]

Ramachandran, P., Zoph, B. & Le, Q. V. Searching for activation functions . arXiv preprint arXiv:1710.05941 (2017)

2017 arXiv
[52]

& Hinton, G

Ba, J., Kiros, J. & Hinton, G. Layer normalization. arXiv preprint arXiv:1607.06450 (2016)

2016 arXiv
[53]

& Hutter, F

Loshchilov, I. & Hutter, F. Decoupled weight decay regularization. International Conference on Learning Representations (2019)

2019
[54]

Madec, G. et al. NEMO ocean engine. Scientiﬁc Notes of Climate Modelling Center 27 (2017)

2017
[55]

User Manual and System Do cumen- tation of W A VEW ATCH III Version 6.07

W A VEW ATCH III Development Group (WW3DG). User Manual and System Do cumen- tation of W A VEW ATCH III Version 6.07. Tech. Note 333, NOAA/NWS/NCEP/MMAB, College Park, MD, USA (2019)

2019
[56]

& Pinardi, N

Dobricic, S. & Pinardi, N. An oceanographic three-dimensional variati onal data assimilation scheme. Ocean Modelling 22, 89–105 (2008). 17

2008
[57]

Weatherall, P. et al. A new digital bathymetric model of the world’s oceans. Earth and Space Science 2, 331–345 (2015)

2015
[58]

& Rosenstein, M

Hellerman, S. & Rosenstein, M. Normal monthly wind stress over the world ocean with error estimates. Journal of Physical Oceanography 13, 1093–1104 (1983). Supplementary information. We provide a Supplementary Materials document that con- tains additional methodological detail...

1983

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Le Traon, P. Y. et al. From observation to information and users: The Copernicus Marine Service perspective. Frontiers in Marine Science 6, 234 (2019)

work page 2019

[2] [2]

Sun, R. et al. SKRIPS v1.0: A regional coupled ocean–atmosphere modeling framework (MITgcm–WRF) using ESMF/NUOPC, description and preliminary result s for the Red Sea. Geoscientiﬁc Model Development 12, 4221–4244 (2019)

work page 2019

[3] [3]

Sakamoto, K. et al. Development of a 2-km resolution ocean model covering the coastal seas around Japan for operational application. Ocean Dynamics 69, 1181–1202 (2019). 14

work page 2019

[4] [4]

Ciliberti, S. A. et al. Monitoring and forecasting the ocean state and biogeochemical pro- cesses in the Black Sea: Recent developments in the Copernicus Mar ine Service. Journal of Marine Science and Engineering 9, 1146 (2021)

work page 2021

[5] [5]

Nemo-Nordic 2.0: Operational marine forecast model for the Baltic Sea

K¨ arn¨ a, T.et al. Nemo-Nordic 2.0: Operational marine forecast model for the Baltic Sea. Geoscientiﬁc Model Development 14, 5731–5749 (2021)

work page 2021

[6] [6]

Zhu, X. et al. Improvements in the regional South China Sea operational oceanography forecasting system (SCSOFSv2). Geoscientiﬁc Model Development 15, 995–1015 (2022)

work page 2022

[7] [7]

Bruschi, A. et al. Indexes for the assessment of bacterial pollution in bathing waters fr om point sources: The northern Adriatic Sea CADEAU service. Journal of Environmental Management 293, 112878 (2021)

work page 2021

[8] [8]

Liubartseva, S. et al. Modeling chronic oil pollution from ships. Marine Pollution Bulletin (2023)

work page 2023