REVIEW 4 major objections 5 minor 13 references
MedFormer: a data-driven model for forecasting the Mediterranean Sea
T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash
Pith's one-line read MedFormer is a fully data-driven neural model that, on 2022 forecasts, beat the operational Mediterranean forecasting system on temperature, salinity, and velocities at all depths and lead times.
desk verdict Credible engineering, overstated skill claim: MedFormer is fine-tuned on the same MedFS analysis used as its main verification target, so 'consistently outperforms' is not yet supported. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing piece is the 3D shifted-window attention U-Net: an encoder that progressively downsamples the (time, latitude, longitude) cube while widening feature channels, and a mirrored decoder with skip connections that reconstructs the next ocean state. Each layer block applies shifted-window self-attention over displaced windows, which lets the model learn local spatio-temporal structure at a manageable computational cost. The second load-bearing piece is the training objective: a weighted mean-absolute-error loss with surface-weighted vertical levels and variable-specific weights, extended during fine-tuning to accumulate error over multiple autoregressive steps so the model learns
What would settle it
Score MedFormer and the dynamical system over several years against withheld in-situ temperature and salinity profiles and satellite sea-surface temperature and sea-level observations, with those years excluded from MedFormer's training and fine-tuning; if the error advantage disappears or reverses outside the summer-autumn seasons, the headline skill gain is a training-target artifact.
Extended reading notes
Core claim
MedFormer is a U-Net whose encoder and decoder are built from 3D shifted-window (Swin) attention blocks, attending over time, latitude, and longitude. It ingests four consecutive daily ocean states plus seven atmospheric surface fields and predicts the next day's ocean state; longer forecasts are generated autoregressively by feeding predictions back as inputs. Training happens in two stages: pre-training on twenty years of daily ocean reanalysis (2000-2019), then fine-tuning on higher-resolution operational analyses (2018-2020) with a curriculum that increases the number of autoregressive steps from one to five. The paper's central claim is that, over the 2022 test year, this model has lowe
Load-bearing premise
The skill comparison is scored against the same operational analysis product that MedFormer was fine-tuned to match, so the reported advantage over the dynamical system may partly reflect the model having been trained on the verification target.
Editorial extensions
If this is right
- A fully data-driven model can sustain 9-day, 1/24° forecasts with lower error than a dynamical ocean system for the main 3D state variables, not just for surface fields.
- Because forecasts are initialized from the same analysis fields and use the same atmospheric forecasts as the operational system, MedFormer can be inserted into the existing weekly operational cycle as a fast alternative or as an additional ensemble member.
- The slower error-growth rate under autoregression implies the model retains useful dynamical information through the full 9-day range rather than collapsing toward persistence.
- Operational centres could run daily high-resolution forecasts at a fraction of the cost of a dynamical system, while keeping the dynamical model for sea-surface height and for seasons or regions where it is stronger.
Reading between the lines
- The main RMSD comparison uses the same operational analysis product that MedFormer was fine-tuned to reproduce, so the paper leaves open how much of the gap is genuine skill; verifying only on independent observations or on forecasts initialized from a different analysis would settle that.
- The appendix's seasonal scorecards show the advantage is strongest in summer and autumn and nearly disappears in winter and spring, especially near the thermocline; a season-aware loss or longer fine-tuning data is a natural next test.
- Since the dynamical system keeps the edge on sea-surface height, a hybrid forecast that takes MedFormer's temperature, salinity, and currents and the dynamical model's sea level is a concrete way to combine their strengths.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper presents MedFormer, a fully data-driven 3D Swin-UNet Transformer for Mediterranean Sea forecasting at 1/24° horizontal resolution. The model is pre-trained on 20 years of MEDREA24 reanalysis and fine-tuned on MedFS EAS6 operational analyses, producing 9-day autoregressive forecasts with atmospheric forcing. The authors benchmark MedFormer against the operational MedFS dynamical system using RMSD and ACC, claiming consistent superiority across temperature, salinity, and horizontal velocities, with SSH as a partial exception. Independent observational verification is provided for SST, SLA, and Argo T/S profiles. The paper includes detailed architecture, training protocol, and a GitHub code release.
Significance. If the central claim is robust, MedFormer would be a substantial advance: a high-resolution, data-driven regional ocean forecasting model that is computationally efficient and competitive with an operational numerical system. The paper's strengths include a detailed description of the architecture and training procedure, a reproducible code repository, and an evaluation against independent satellite and Argo observations for several variables. However, the headline analysis-vs-analysis evaluation is compromised by the use of the same MedFS EAS6 product as both the fine-tuning target and the verification reference, and the velocity skill (u/v) is never checked against independent observations. These issues bear directly on the paper's main claim and require additional evidence before the results can be fully accepted.
major comments (4)
- [Sec. 3 vs Sec. 5.4] The primary benchmark (Figs. 2–5) compares MedFormer and MedFS forecasts against the Mediterranean Sea Analysis as reference, but Section 5.4 states that fine-tuning uses 'the Mediterranean analysis as initial condition and as target' (Eq. 4). The model is therefore directly optimized to reproduce the same product used for verification, which inflates its apparent skill. This is especially problematic for u/v, where no independent observations are provided. Please provide a neutral evaluation reference (e.g., independent reanalysis or withheld observational data) or quantify the bias introduced by training on the verification target.
- [Sec. 3, Figs. 6–8; abstract; Sec. 4] The abstract and conclusion claim MedFormer 'consistently outperforms MedFS across key 3D ocean variables,' including horizontal velocities. However, the only evidence for u/v skill is the analysis-based RMSD/ACC of Section 3, which is the same product on which the model was fine-tuned. Independent checks exist only for SST, SLA, and Argo T/S. The velocity claim is unsupported without independent current observations (e.g., moored current meters, HF radar, or drifters). Please add at least one independent velocity dataset, or explicitly restrict the outperformance claim to variables verified against observations.
- [Sec. 3, Figs. 3–4; Sec. 5.5] The paper states that 95% confidence intervals demonstrate statistical significance, but the method used to compute these intervals is not described. The test set is a single year (2022) with weekly forecast cycles, so it is important to specify whether intervals are based on bootstrapping over forecast cycles or spatial blocks, and how temporal autocorrelation is handled. Without this, the significance claim cannot be assessed. Also, the seasonal analysis in Appendix A covers only one year, so its generalizability is limited.
- [Abstract; Sec. 3; Fig. 5; Appendix A] The repeated claim that MedFormer 'consistently outperforms MedFS' is contradicted by the paper's own results for sea surface height, where MedFS has lower RMSD beyond day 2 and across all seasons (Fig. 5, Fig. A5). Although SSH is a 2D variable, it is still a key forecast field. Please qualify the claim (e.g., 'for most 3D variables at most lead times') and discuss the SSH shortcoming explicitly in the abstract and conclusions.
minor comments (5)
- [Sec. 5.4, Eq. (3)] The formula for the level weights w_l is ambiguous: 'w_l = 1−0.01 L−1 · (1 − l) + 1' should be written with parentheses, e.g., w_l = 1 + (1−l)(1−0.01)/(L−1), to make clear that w_1=1 and w_L=0.01.
- [Sec. 1 and Sec. 5.4] The fine-tuning dataset description is inconsistent: Section 1 says '4 years of data spanning from 2018 to 2021,' while Section 5.4 says '2018-2020 for training and 2021 for validation' (three training years plus one validation year). Please clarify how many years are used for training and validation.
- [Sec. 3, Fig. 2] The caption and text state 'RMSD computed against the Mediterranean Sea Analysis' but do not specify whether the same MedFS analysis is used for both models as the reference and as the source of MedFS initial conditions. This should be stated explicitly.
- [Sec. 5.5, ACC definition] Anomalies are computed relative to 'the model climatology derived from MEDREA Reanalysis.' It is unclear whether the same climatology is used for MedFormer and MedFS, and whether the climatology is computed over the full available period or only the training period. Please specify.
- [General] There are minor typographical issues, e.g., 'T able 1' in the Table 1 caption, and inconsistent use of '1/24°' and '1 /24◦' in the text.
Circularity Check
Velocity skill is evaluated only against the fine-tuning target; the headline 'consistently outperforms' for u/v is not independently verified.
-
fitted input called prediction
[Section 5.4 (Training protocol) and Section 3 / Figure 2]
"The model is fine tuned in the second step of training by using the Mediterranean analysis as initial condition and as target. ... MedFormer and MedFS RMSD computed against the Mediterranean Sea Analysis for temperature, salinity, meridional and zonal velocities at the surface and SSH averaged across the 2022 year in the whole Mediterranean basin for all 9 forecast lead times."
MedFormer's fine-tuning loss (Eq. 4) is the average autoregressive MAE L( ˆX τ , X τ ) where X τ are the MedFS EAS6 analysis fields. The primary skill metric (RMSD, Eq. 5) is evaluated against the same MedFS EAS6 analysis (Fig. 2). Thus for horizontal velocities (u/v), the reported 'consistently outperforms MedFS' is a comparison of the model against the exact target it was optimized to reproduce during fine-tuning. Independent observational checks are provided only for SST, SLA, and Argo T/S (Figs. 6-8); no independent velocity observations are used anywhere. The u/v component of the headline claim therefore rests entirely on a training-target-as-truth evaluation, which is not neutral and partly forces the result.
full rationale
The paper's main evaluation for all four 3D variables uses the MedFS EAS6 analysis as ground truth (Fig. 2). The model is fine-tuned on that same analysis product as both initial condition and target (Sec. 5.4, Eq. 4). This makes the analysis-vs-analysis comparison for horizontal velocities circular in the sense that the model is directly trained to reproduce the reference; the low u/v error is partly an artifact of the training objective rather than independent evidence of skill. Independent observations are used for SST, SLA, and Argo T/S, which mitigates circularity for those variables and gives the paper genuine external grounding. However, the headline claim covers 'key 3D ocean variables' including u/v, for which no independent reference exists. This is a load-bearing gap: the 'consistently outperforms' assertion for u/v is not supported by any neutral benchmark. The circularity is partial (score 6) rather than total because the model is tested on a held-out year (2022) and compared against a separate dynamical system, and because independent observations corroborate the T/S and surface claims.
Assumptions & free parameters
free parameters (7)
- Variable loss weights =
T=3, S=2, u=1, v=1, SSH=1
- Vertical loss weights =
Linear from 1.0 at surface to 0.01 at 971 m
- Swin3D layer counts per block =
Encoder [2,8,4], decoder [4,8,2]
- Embedding dimension =
C=384, doubling after each downsample
- Window and patch sizes =
Window (2,8,8), patch (1,4,4)
- Autoregressive curriculum =
Increase autoregressive steps by 1 every 60 epochs, from 1 to 5
- Learning rates =
Pretrain 5e-4 to 5e-7; finetune 5e-5 to 5e-8 (cosine)
assumptions (5)
- domain assumption The daily-mean ocean state at t+1 is fully determined by the previous four daily states and the atmospheric surface forcing sequence through t+1 (Eq. 1).
- domain assumption The MEDREA24 reanalysis and the MedFS EAS6 analysis are mutually consistent enough to allow pre-training on one and fine-tuning on the other.
- ad hoc to paper The MedFS analysis used as both training target and verification reference is an unbiased estimate of the true ocean state.
- domain assumption Bilinear interpolation of atmospheric fields from 1/4 or 1/10 degree to the 1/24 degree ocean grid preserves the relevant forcing scales.
- domain assumption The 18 selected vertical levels are sufficient to represent the 3D ocean state for forecasting purposes.
Cite this review
Pith. "Pith review of MedFormer: a data-driven model for forecasting the Mediterranean Sea." pith.science (2026). https://pith.science/paper/7LVQZZKN
@misc{pith2026250900015,
author = {Pith},
title = {Pith review of: MedFormer: a data-driven model for forecasting the Mediterranean Sea},
year = {2026},
howpublished = {\url{https://pith.science/paper/7LVQZZKN}},
note = {Machine review of arXiv:2509.00015}
}
read the original abstract
Accurate ocean forecasting is essential for supporting a wide range of marine applications. Recent advances in artificial intelligence have highlighted the potential of data-driven models to outperform traditional numerical approaches, particularly in atmospheric weather forecasting. However, extending these methods to ocean systems remains challenging due to their inherently slower dynamics and complex boundary conditions. In this work, we present MedFormer, a fully data-driven deep learning model specifically designed for medium-range ocean forecasting in the Mediterranean Sea. MedFormer is based on a U-Net architecture augmented with 3D attention mechanisms and operates at a high horizontal resolution of 1/24{\deg}. The model is trained on 20 years of daily ocean reanalysis data and fine-tuned with high-resolution operational analyses. It generates 9-day forecasts using an autoregressive strategy. The model leverages both historical ocean states and atmospheric forcings, making it well-suited for operational use. We benchmark MedFormer against the state-of-the-art Mediterranean Forecasting System (MedFS), developed at Euro-Mediterranean Center on Climate Change (CMCC), using both analysis data and independent observations. The forecast skills, evaluated with the Root Mean Squared Difference and the Anomaly Correlation Coefficient, indicate that MedFormer consistently outperforms MedFS across key 3D ocean variables. These findings underscore the potential of data-driven approaches like MedFormer to complement, or even surpass, traditional numerical ocean forecasting systems in both accuracy and computational efficiency.
Reference graph
Works this paper leans on
-
[1]
Wang, X. et al. Xihe: A data-driven model for global ocean eddy-resolving forecasting (2024). URL https://arxiv.org/abs/2402.02995. 2402.02995
arXiv 2024
-
[2]
Aouni, A. E. et al. Glonet: Mercator’s end-to-end neural global ocean forecasting system (2025). URL https://arxiv.org/abs/2412.05454. 2412.05454
arXiv 2025
-
[3]
Dheeshjith, S. et al. Samudra: An ai global ocean emulator for cli- mate. Geophysical Research Letters 52, e2024GL114318 (2025). URL https://agupubs.onlinelibrary.wiley.com/doi/abs/10.1029/2024GL114318. E2024GL114318 2024GL114318
-
[4]
Xiong, W. et al. Ai-goms: Large ai-driven global ocean modeling system (2023). URL https://arxiv.org/abs/2308.03152. 2308.03152
arXiv 2023
-
[5]
Guo, Z. et al. Data-driven global ocean modeling for seasonal to decadal prediction (2024). URL https://arxiv.org/abs/2405.15412. 2405.15412
arXiv 2024
-
[6]
Coppini, G. et al. The mediterranean forecasting system – part 1: Evolution and performance. Ocean Science 19, 1483–1516 (2023). URL https://os.copernicus. org/articles/19/1483/2023/
work page 2023
- [7]
-
[8]
Hersbach, H. et al. The era5 global reanalysis. Quarterly Journal of the Royal Meteorological Society146, 1999–2049 (2020). URL https://rmets.onlinelibrary. wiley.com/doi/abs/10.1002/qj.3803
doi:10.1002/qj.3803 1999
Show all 13 references
-
[9]
Clementi, E. et al. Mediterranean sea physical analysis and forecast (cmems med-currents, eas6 system) (version 1) [data set] (2021). URL https://doi.org/ 10.25423/CMCC/MEDSEA ANALYSISFORECAST PHY 006 013 EAS6
2021 doi
-
[10]
& Masina, S
Storto, A. & Masina, S. C-glorsv5: an improved multipurpose global ocean eddy- permitting physical reanalysis. Earth System Science Data 8, 679–696 (2016). URL https://essd.copernicus.org/articles/8/679/2016/
2016
-
[11]
& Brox, T
Ronneberger, O., Fischer, P. & Brox, T. Navab, N., Hornegger, J., Wells, W. M. & Frangi, A. F. (eds) U-net: Convolutional networks for biomedical image seg- mentation. (eds Navab, N., Hornegger, J., Wells, W. M. & Frangi, A. F.) Medical Image Computing and Computer-Assisted In...
2015
-
[12]
Liu, Z. et al. . (ed.) Swin transformer: Hierarchical vision transformer using shifted windows. (ed..) 2021 IEEE/CVF International Conference on Computer 28 Vision (ICCV), 9992–10002 (IEEE, 2021)
2021
-
[13]
Liu, Z. et al. . (ed.) Swin transformer v2: Scaling up capacity and resolution. (ed..) 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 11999–12009 (IEEE, 2022). 29
2022
Reviewed August 5, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.