KIST-Ocean: Data-Driven Ocean Modeling

• KIST-Ocean is a deep-learning 3D ocean general circulation model that predicts global ocean states by learning from past conditions and atmospheric forcing.
• It employs a U-shaped visual attention network with partial convolutions, ensuring high-resolution forecasts and realistic coastal variability.
• The model achieves efficient 5-day interval predictions validated by physical tests on Kelvin and Rossby waves and other ocean dynamic processes.

Searching arXiv for the KIST-Ocean paper and closely related ocean-modeling work. KIST-Ocean is a data-driven global three-dimensional ocean general circulation model built with a U-shaped visual attention adversarial network and designed to predict the global ocean state in an autoregressive fashion at 5-day intervals out to 200 days (Kim et al., 31 Jul 2025). It is positioned as a deep-learning ocean component for weather-to-climate prediction, with an emphasis on ocean memory, subsurface dynamics, and ocean response to atmospheric forcing. Unlike standard numerical ocean models, it does not solve the primitive equations explicitly; instead, it learns a mapping from past ocean states and surface boundary forcing to future ocean states from data (Kim et al., 31 Jul 2025). The model predicts only oceanic variables, not atmospheric forcing variables, and is evaluated both by conventional forecast skill metrics and by process-oriented tests involving Kelvin waves, Rossby waves, and wind-driven vertical motions (Kim et al., 31 Jul 2025).

1. Definition and scientific scope

KIST-Ocean is presented as the Korea Institute of Science and Technology’s Ocean model: a deep-learning-based global 3D ocean general circulation model intended to emulate the predictive role of an OGCM within a future coupled ocean–atmosphere system (Kim et al., 31 Jul 2025). Its scientific motivation is the extension of deep learning from medium-range atmospheric forecasting toward subseasonal-to-seasonal and climate-relevant prediction, where the ocean’s large heat capacity, slow evolution, and long memory become central (Kim et al., 31 Jul 2025).

The model targets a specific gap in prior data-driven ocean prediction. Earlier deep-learning ocean models are described as often using non-autoregressive forecasting, separate models for each lead time, monthly timestep resolution, or incomplete prognostic-variable sets, and as lacking rigorous tests of physical response to boundary forcing (Kim et al., 31 Jul 2025). In that framing, KIST-Ocean is not merely a statistical forecaster of sea-surface variables. It is designed to represent three-dimensional ocean evolution, including potential temperature, currents, salinity, SST, and sea ice concentration, under prescribed surface forcing (Kim et al., 31 Jul 2025).

A central claim of the work is that KIST-Ocean is the first study to verify oceanic responses driven by the atmosphere, such as tropical waves and Ekman divergence, against traditional theory (Kim et al., 31 Jul 2025). This places the model at the intersection of machine-learning-based surrogate modeling and process-based ocean dynamics. A plausible implication is that the authors regard physical interpretability of the learned dynamics, rather than forecast skill alone, as essential for the model’s relevance to Earth system prediction.

2. State representation, inputs, outputs, and forecasting setup

The model operates on a $1^\circ \times 1^\circ$ global grid from $0^\circ$ to $360^\circ$ longitude and $79^\circ$S to $80^\circ$N latitude, with 15 vertical levels (Kim et al., 31 Jul 2025). Its input tensor comprises 68 channels: 62 oceanic variables and 6 surface boundary forcing variables (Kim et al., 31 Jul 2025). The output is the 62 oceanic variables at $t+5$ days (Kim et al., 31 Jul 2025).

The ocean variables include 3D potential temperature $\theta$, zonal current UO, meridional current VO, salinity S, and 2D SST and sea ice concentration (SIC), while the forcing variables are zonal and meridional wind stress $(\tau_x,\tau_y)$, downward longwave radiation, downward shortwave radiation, latent heat flux, and sensible heat flux (Kim et al., 31 Jul 2025). The forecasting procedure is autoregressive: the model output at one step is fed back as part of the next input, repeated 40 times to reach lead times from 5 to 200 days (Kim et al., 31 Jul 2025).

Two inference modes define the forcing assumptions. In KIST-O$_\text{GT}$, ground-truth forcing is prescribed during inference; in KIST-O$_\text{Clim}$, climatological forcing is used instead (Kim et al., 31 Jul 2025). These are described as upper and lower bounds on what a future coupled deep-learning ocean–atmosphere system could achieve (Kim et al., 31 Jul 2025). This distinction is methodologically important because the model does not predict the atmospheric state. It therefore isolates the oceanic forecasting problem while still exposing the dependence of forecast skill on external forcing realism.

The paper emphasizes computational efficiency. KIST-Ocean has 6.6 million parameters; pretraining takes about 33.3 hours on one NVIDIA A100, fine-tuning about 2.4 hours, and a 200-day forecast requires about 6–7 seconds on one NVIDIA A100 (Kim et al., 31 Jul 2025). This is presented as substantially cheaper than traditional numerical OGCM integration. The claim is operational rather than theoretical: the model’s significance lies partly in inference cost at global scale.

3. Network architecture and core algorithmic components

KIST-Ocean uses a generator–discriminator GAN framework in which the generator is a U-shaped visual attention network and the discriminator is a PatchGAN (Kim et al., 31 Jul 2025). The generator follows a multiscale encoder–decoder design: a down-sampling path extracts multiscale and global features, an up-sampling path reconstructs high-resolution outputs, and skip connections preserve local information and aid training with limited data (Kim et al., 31 Jul 2025). The visual attention network is chosen for efficiency through large kernel attention, which expands receptive fields without large parameter overhead (Kim et al., 31 Jul 2025).

At each time step, the generator takes the 68-channel input tensor, passes it through a stem layer, applies alternating VAN stages and down-sampling blocks twice, then applies VAN stages and up-sampling blocks to restore resolution, and outputs 62 channels for the future ocean variables (Kim et al., 31 Jul 2025). Each VAN stage includes batch normalization, point-wise convolution, GELU, large kernel attention, point-wise convolution, batch normalization, point-wise convolution, $0^\circ$0 depth-wise convolution, GELU, point-wise convolution, and layer normalization, with two residual connections (Kim et al., 31 Jul 2025). Circular padding is applied to preserve longitudinal continuity (Kim et al., 31 Jul 2025).

The discriminator receives predicted and ground-truth fields at $0^\circ$1 days and processes them with six $0^\circ$2 depth-wise convolution layers, four $0^\circ$3 max-pooling layers, GELU activations, and a final sigmoid (Kim et al., 31 Jul 2025). It outputs a tensor of size $0^\circ$4, where each element corresponds to the probability that a $0^\circ$5 patch is real (Kim et al., 31 Jul 2025). The use of a PatchGAN is specifically motivated as a way to preserve high-frequency details and local realism under autoregressive rollout (Kim et al., 31 Jul 2025).

The generator and discriminator are trained with a combined adversarial and reconstruction objective. The generator loss is

$0^\circ$6

and the discriminator loss is

$0^\circ$7

with $0^\circ$8 and $0^\circ$9 (Kim et al., 31 Jul 2025). The reconstruction term is an area-weighted $360^\circ$0 loss,

$360^\circ$1

with latitude weight

$360^\circ$2

These equations make explicit that the training objective balances distributional realism against pointwise reconstruction fidelity (Kim et al., 31 Jul 2025).

4. Partial convolution, adversarial training, and transfer learning

A distinctive feature of KIST-Ocean is the use of partial convolution to handle land–sea masking (Kim et al., 31 Jul 2025). Ordinary convolution on geophysical grids can mix valid ocean values with masked land values, distorting coastal variability; partial convolution instead excludes masked values and renormalizes by the number of valid points (Kim et al., 31 Jul 2025). The paper defines

$360^\circ$3

where $360^\circ$4 is the feature map, $360^\circ$5 the binary mask, $360^\circ$6 the kernel region, $360^\circ$7 the filter weights, and $360^\circ$8 prevents division by zero (Kim et al., 31 Jul 2025). All convolution operations in the generator, except point-wise convolutions, are replaced with partial convolutions (Kim et al., 31 Jul 2025).

The role of adversarial training is to reduce predictive distribution drift in autoregressive forecasts (Kim et al., 31 Jul 2025). Because model outputs are recursively fed back as inputs, long-lead predictions can drift away from the data manifold. The PatchGAN discriminator penalizes such drift by enforcing local realism in predicted fields (Kim et al., 31 Jul 2025). According to the ablation summary, removing adversarial training degrades skill, produces blurrier outputs, underrepresents realistic current structures, and increases distribution drift and smoothing (Kim et al., 31 Jul 2025).

Transfer learning is used to address the limited duration of reanalysis-quality global ocean observations. KIST-Ocean is pretrained on CESM2 Large Ensemble historical simulations using member 1301.012 for 1850–2014 and member 1301.013 for 1850–2004 for training, with 2005–2014 from member 1301.013 for validation (Kim et al., 31 Jul 2025). It is then fine-tuned on 1982–2013 reanalysis data: GODAS for 3D ocean variables and wind stress, OISST for SST and SIC, and ERA5 for surface fluxes (Kim et al., 31 Jul 2025). The paper reports that transfer learning has a modest overall effect but becomes more beneficial at deeper layers, especially for potential temperature and salinity (Kim et al., 31 Jul 2025).

The ablation results are methodologically central. Removing partial convolution causes faster skill decay, stronger distribution drift, unrealistic strong currents, and poorer coastal and ocean variability representation (Kim et al., 31 Jul 2025). Removing adversarial training degrades skill and realism; removing transfer learning has a smaller but still noticeable effect, especially in the deep ocean (Kim et al., 31 Jul 2025). A FourCastNet-based baseline with 25 million parameters performs worse than KIST-Ocean despite being about four times larger, which the authors use to argue that architecture and training design matter more than raw parameter count (Kim et al., 31 Jul 2025).

5. Training data, evaluation metrics, and forecast skill

The training pipeline is divided into pretraining, fine-tuning, and hindcast evaluation (Kim et al., 31 Jul 2025). Pretraining on CESM2-LE uses 23,360 training samples and 730 validation samples (Kim et al., 31 Jul 2025). Fine-tuning on reanalysis data from 1982–2013 uses 2,336 samples, and testing or hindcast evaluation covers 2014–2023 with 730 samples (Kim et al., 31 Jul 2025).

Evaluation uses RMSE and ACC. The global mean RMSE is defined as

$360^\circ$9

while gridpoint ACC and global ACC are defined with a Fisher $79^\circ$0-transform to avoid bias from directly averaging correlations (Kim et al., 31 Jul 2025). The evaluation framework is thus tailored to anomaly skill rather than raw-state matching.

The headline forecast result is that KIST-Ocean substantially outperformed persistence forecasts: among 2,480 targets, 81.3% were better in ACC and 86.5% were better in RMSE (Kim et al., 31 Jul 2025). For KIST-O$79^\circ$1, the maximum lead times with significant ACC skill are 200 days for potential temperature, 85 days for zonal current UO, 40 days for meridional current VO, and 100 days for salinity (Kim et al., 31 Jul 2025). KIST-O$79^\circ$2 performs worse, especially for SST, but retains meaningful skill, which is interpreted as evidence that the model has learned both ocean memory from the initial state and dynamical influence from surface forcing (Kim et al., 31 Jul 2025).

The paper also compares monthly SST prediction against NMME systems, specifically COLA-RSMAS-CCSM4, GFDL-SPEAR, NASA-GEOSS2S, NCEP-CFSv2, and CanSIPS-IC3 (Kim et al., 31 Jul 2025). KIST-O$79^\circ$3 outperformed the NMME models for global, Pacific, Atlantic, and Southern Ocean SST up to 6 months lead, while KIST-O$79^\circ$4 had one-month skill comparable to the best NMME models and remained within their range at two months, retaining Atlantic skill up to 6 months (Kim et al., 31 Jul 2025). Forecast gaps between the two forcing modes were larger in the 20–50° latitude bands, and forecasts initialized in late boreal winter were generally better (Kim et al., 31 Jul 2025).

These results are presented as evidence of subseasonal-to-seasonal utility. This suggests that the model’s learned dynamics are not confined to short-lag interpolation. Rather, they persist over lead times associated with ocean memory and coupled climate variability.

6. Representation of ocean–atmosphere coupling dynamics

A major emphasis of KIST-Ocean is process fidelity under atmospheric forcing (Kim et al., 31 Jul 2025). The paper argues that the model reproduces eastward-propagating downwelling Kelvin waves, westward-propagating upwelling Rossby waves, Rossby-wave reflection off the Maritime Continent as Kelvin waves, and opposite-phase responses under easterly forcing in the tropical Pacific (Kim et al., 31 Jul 2025). These responses are tied explicitly to ENSO-relevant physics, including thermocline deepening, suppression of upwelling of cold subsurface water, and El Niño development under westerly wind bursts (Kim et al., 31 Jul 2025).

The paper further states that KIST-Ocean reproduced westerly-wind-burst-driven warming in the 2013–2014 case and the 2015 strong El Niño evolution when realistic forcing was used (Kim et al., 31 Jul 2025). It also reports that Rossby wave speed varies with latitude in the physically correct way: slower at higher latitudes and faster in the Southern Hemisphere than the Northern Hemisphere in the tested setup (Kim et al., 31 Jul 2025). The comparison is made with the first-baroclinic long-wave theoretical approximation

$79^\circ$5

where the negative sign indicates westward propagation, $79^\circ$6, $79^\circ$7, and $79^\circ$8 (Kim et al., 31 Jul 2025). The reported agreement is used as an explicit physical validation rather than a generic visualization of propagating anomalies.

The model also captures vertical motions induced by rotating wind stress. Cyclonic forcing leads to Ekman divergence, upwelling, and surface and subsurface cooling, whereas anticyclonic forcing leads to convergence, downwelling, and warming (Kim et al., 31 Jul 2025). Since KIST-Ocean does not explicitly output vertical velocity $79^\circ$9, the paper diagnoses it from continuity:

$80^\circ$0

and

$80^\circ$1

These experiments are interpreted as demonstrating vertical heat redistribution and Ekman pumping processes central to coupled ocean–atmosphere dynamics (Kim et al., 31 Jul 2025).

The authors connect these results to the delayed oscillator, recharge–discharge process, and Bjerknes feedback in ENSO theory (Kim et al., 31 Jul 2025). Because the model is not yet fully coupled, these mechanisms are represented under prescribed forcing rather than interactive atmosphere–ocean feedback. A plausible implication is that the paper treats KIST-Ocean as a proof of dynamical competence of a learned ocean component, rather than a completed coupled climate model.

7. Position within ocean modeling and Earth system AI

KIST-Ocean is framed as both a surrogate OGCM and a step toward AI-based Earth system modeling (Kim et al., 31 Jul 2025). Relative to traditional numerical ocean models, it is described as a neural surrogate learned from data, far faster at inference, and capable of reproducing the mapping from forcing and previous ocean states to future ocean states without explicit PDE integration (Kim et al., 31 Jul 2025). Relative to atmospheric deep-learning systems such as Pangu-Weather, GraphCast, FourCastNet, FengWu, and FuXi, it targets the 3D global ocean and is explicitly designed to model ocean response to atmospheric forcing (Kim et al., 31 Jul 2025).

The broader significance claimed in the paper is that learned ocean dynamics can reproduce key atmosphere-driven responses, preserve essential climate dynamics, and potentially extend to a fully coupled ocean–atmosphere system (Kim et al., 31 Jul 2025). The authors are explicit that the present model is not yet a coupled system because surface forcing is prescribed during inference (Kim et al., 31 Jul 2025). This caveat is important because it limits the scope of interpretation: KIST-Ocean demonstrates learned ocean general circulation under external forcing, not autonomous climate-system evolution.

Within that scope, the model is presented as a proof of concept that global ocean circulation can be learned from data in a physically meaningful way (Kim et al., 31 Jul 2025). Its claimed significance therefore has three layers: forecast efficiency, subsurface multivariate prediction skill, and physically interpretable response to wind forcing. For researchers in ocean modeling and machine learning, the most consequential aspect may be the conjunction of those layers rather than any one of them in isolation.

A possible misconception is to read KIST-Ocean as a replacement for numerical OGCMs in all settings. The paper does not make that claim. Instead, it presents the model as a component of a future coupled deep-learning ocean–atmosphere framework, with prescribed forcing used as a controlled testbed for learned ocean dynamics (Kim et al., 31 Jul 2025). In that narrower but technically significant sense, KIST-Ocean represents a data-driven surrogate OGCM whose central achievement is not only predictive skill, but physically consistent simulation of ocean–atmosphere coupling phenomena relevant to ENSO and seasonal climate variability (Kim et al., 31 Jul 2025).