- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
EddyCast forecasts daily upper-ocean temperature (10 levels) and salinity (15 levels) in the Northwest Pacific for seven days, trained on GLORYS12V1 reanalysis. A compact 3D convolutional encoder–decoder is extended with a "Physics-as-Bypass" branch. Physical features, either EAG (SST Laplacians at t and t−1 and their difference) or VMoD (learnable vertical modes), pass through a small network and are added to the output through a zero-initialized 1×1×1 projection. An EDM head gives a probabilistic variant. EddyCast beats seven baselines (RMSE 0.0106 vs 0.0118 for WenHai) with about 3M parameters, and with the same EAG features and backbone it beats feeding the features as inputs or as an auxiliary loss.
Critical strengths
- Table 2 compares the same EAG features used as input, as loss and as bypass, with the backbone and training fixed. This is cleaner evidence for the integration mechanism than the cross-architecture comparison.
- The accuracy–cost trade-off is good: about 3M parameters, 25.7 ms for a seven-step rollout, and the lowest PhysRMSE at every lead.
Critical weaknesses
- The task is not forecasting as presented. Per App. A.1 and Table 3, every rollout step receives the target-day SSH, SST, SSS, surface currents and mixed-layer depth, taken from the reanalysis. These are not available when a forecast is issued, and they carry much of the skill: removing them (w/o Surf) gives RMSE 0.0126, worse than the no-physics model (0.0124). The results therefore measure subsurface estimation given future surface truth, not seven-day forecast skill. Persistence does not receive this information.
- A physics-specific benefit is not shown. Zero-initialized residual branches are a standard device (e.g. ControlNet, ReZero), and EAG is three simple SST-derived maps, so the gain over input concatenation (0.0114 → 0.0106) may come from the extra trainable branch. There is no parameter-matched generic bypass (the same branch fed learned or random features, or without zero initialization). The vertical analysis is confounded by the channel layout: the "vertical" axis concatenates 10 temperature and 15 salinity levels with different depths (Table 4), and the depth-resolved trust curves in Fig. 3(c, e) jump exactly at layer 10, the temperature/salinity boundary. In Fig. 6 the best model (K = 2) also has the lowest EOF alignment (about 0.45 vs 0.64 for K = 4 and 8), which does not support the claim that forecasting utility follows prior–data alignment.
Questions for authors
- How does performance change when the target-day surface fields are replaced by persistence or by a forecast available at initialization? Did all baselines receive the same surface inputs?
- Does the advantage hold against a parameter-matched generic bypass, and against the same bypass without zero initialization?
- What is the spatial extent of each sample? Fig. 3(d) shows a grid of about 36 × 36 points (about 3° × 3°), and training uses ±18-pixel crop jitter. Are training and evaluation done on small windows or on the full 120–155°E, 10–40°N domain, and how many test cases are there? WenHai and Swin3D show strong block artifacts in Fig. 9; how were these global models adapted and tuned for this setting?
- The manuscript does not contain the AI use statement required by the ICLR 2027 Author Guidelines. Please add it.