Ocean-Centric Models

• Ocean-Centric Models are frameworks that center on the ocean's dynamics by using multi-depth state representations and dedicated variables for stratification and circulation.
• They integrate diverse methodologies—including deep learning backbones, observation-driven presets, and asymmetric ocean–atmosphere coupling—to capture key patterns like ENSO behavior.
• The approach provides practical insights for designing models that respect physical consistency, enabling enhanced prediction of phenomena such as sea surface temperature and wave energetics.

An ocean-centric model is a model whose state representation, inductive biases, training data, and evaluation criteria are organized around the structure, dynamics, and observation modalities of the ocean rather than around a generic spatiotemporal or multimodal template. In recent literature, the term has been used for global ocean forecasting systems, ocean-atmosphere coupled predictors in which the ocean provides the dominant source of memory, ocean remote-sensing foundation models, SAR-specific marine feature extractors, and ocean-domain language or multimodal models (Xiong et al., 2023). In a broader physical sense, the same orientation appears in earlier work that treats ocean circulation as a primary organizer of climate or regards currents, winds, and waves as the dominant environmental system to which secondary processes respond (Ji et al., 2018).

1. Scope and defining characteristics

Across the cited literature, an ocean-centric model is distinguished less by a single architecture than by a recurring set of design choices. The ocean is treated as the primary prognostic substrate; the atmosphere typically enters as boundary forcing or coupled driver; variables are selected to preserve upper-ocean stratification, circulation, or wave energetics; and benchmarks emphasize ocean-relevant quantities such as SST, SSH, thermocline structure, EKE, significant wave height, or marine biogeochemical fields (Wang et al., 2024).

A concise taxonomy of representative systems is given below.

Domain	Representative systems
Global ocean forecasting and simulation	AI-GOMS, Ola, NeuralOM, ORCA-DL, HybridOM, AxiomOcean, KIST-Ocean, OceanCastNet (Xiong et al., 2023)
Ocean remote sensing and observation FMs	OceanMAE, OceanSAR-2, Sentinel-3 ocean-colour FM (Stamer et al., 9 Apr 2026)
Ocean-domain language and multimodal knowledge models	OceanGPT, OceanPile (Bi et al., 2023)

In the physical-science antecedents of the term, the same logic appears without modern deep learning. The exoplanet circulation study frames an ocean-centric climate perspective as one in which large-scale ocean circulation redistributes heat, shapes sea-ice distribution, and interacts with the atmosphere in ways that can enhance or suppress habitability (Ji et al., 2018). The combined ocean-and-oil framework for adaptive spill monitoring is likewise ocean-centric because currents, winds, and waves are solved first, while oil is treated as a transported contaminant that responds to them without feeding back on the environment (Hodgson et al., 2019).

Taken together, these uses suggest that “ocean-centric” names a modeling paradigm in which the ocean is not an auxiliary field or boundary condition, but the central system around which the model is organized.

2. Ocean state representation as the organizing principle

A defining property of ocean-centric forecasting models is that they predict multi-depth ocean state variables rather than a thin set of surface proxies. AI-GOMS predicts five basic ocean variables—temperature, salinity, zonal velocity, meridional velocity, and sea surface height—with temperature, salinity, and currents defined on 15 depth levels from 0 to 500 m, at global 1/4° resolution and daily lead times out to 30 days (Xiong et al., 2023). AxiomOcean pushes this further by jointly predicting upper-ocean temperature, salinity, and three-dimensional currents at global 1/12° resolution on 23 levels down to 643 m, explicitly targeting preservation of vertical hierarchy and cross-layer dependence (Wu et al., 11 May 2026).

Other systems adopt related but distinct upper-ocean state vectors. NeuralOM uses sea salinity, sea temperature, zonal velocity, and meridional velocity over 23 depth levels from 0.49–643.57 m, plus SSH at the surface, on a 0.5° grid for subseasonal-to-seasonal simulation (Gao et al., 27 May 2025). ORCA-DL predicts SST and SSH at the surface and potential temperature, salinity, zonal velocity, and meridional velocity on 16 levels from 10 to 1000 m, using zonal and meridional wind stress as atmospheric input (Guo et al., 2024). KIST-Ocean uses 15 levels from 5 to 600 m for potential temperature, zonal and meridional currents, and salinity, plus SST and sea-ice concentration at the surface, with 5-day time stepping (Kim et al., 31 Jul 2025).

This representation choice is consequential. In these models, upper-ocean stratification, mixed-layer adjustment, subsurface heat storage, and thermocline variability are not side effects; they are the objects of prediction. AxiomOcean formalizes this using an increment formulation,

$$X_{t+\Delta t} = X_t + M(X_{t-1}, X_t, F_{t-1}, F_t),$$

where $X_t$ is the multivariate 3D ocean state and $F_t$ is atmospheric forcing (Wu et al., 11 May 2026). ORCA-DL uses a related lead-time-aware formulation,

$$\hat{O}^{t+\Delta t} = \mathrm{ORCA\text{-}DL}(O^{t}, A^{t}, \Delta t, M),$$

with $O^t$ the ocean state, $A^t$ atmospheric wind stress, $\Delta t$ lead time, and $M$ the ocean–land mask (Guo et al., 2024).

This suggests that ocean-centricity at the state level is inseparable from vertical structure. Models that preserve the layered water column tend to evaluate not only surface skill but also thermocline geometry, upper-ocean heat content, and depth-dependent errors, because those quantities are central to short-range and seasonal predictability.

3. Architectural patterns and learned ocean dynamics

Ocean-centric architectures are typically built around operators or backbones chosen to respect basin geometry, multiscale circulation, wave-like structure, or vertical coupling. AI-GOMS uses a Fourier-based Masked Autoencoder backbone with Adaptive Fourier Neural Operator blocks, taking topography, atmospheric forcings, and 3D ocean states as a 67-channel tensor and learning daily transitions in an autoregressive fashion (Xiong et al., 2023). The model is explicitly framed as replacing the dynamic core of a numerical general circulation model with a learned backbone, then attaching downstream ocean modules for regional downscaling, wave decoding, and biogeochemical coupling.

NeuralOM adopts a different route. It maps the ocean onto a multi-scale graph and introduces a Multi-scale Interactive Messaging block with gradient-like differences, multiplicative couplings, and cosine similarity on edges, together with gated sum/mean aggregation on nodes, to capture fronts, shear, nonlinear couplings, and slow-varying basin-scale structure (Gao et al., 27 May 2025). The same paper explicitly links climatology subtraction and multi-stage residual refinement to the slowly changing character of the ocean.

AxiomOcean makes vertical structure itself a primary architectural object. Its fully three-dimensional encoder–backbone–decoder fuses a 3D ocean encoder, a 2D atmospheric encoder, and a 3D U-Swin Transformer backbone so that attention operates across horizontal and vertical dimensions rather than only across flattened depth channels (Wu et al., 11 May 2026). KIST-Ocean, by contrast, remains horizontally two-dimensional but builds a deep-learning OGCM around a U-shaped visual attention adversarial network with partial convolution, allowing it to represent coastal complexity and reduce predictive distribution drift in autoregressive rollouts (Kim et al., 31 Jul 2025).

HybridOM occupies a distinct gray-box position. It writes the state evolution as

$$\frac{dX}{dt} = M_{\text{phy}}(X, A) + N_\theta(X, A),$$

with $M_{\text{phy}}$ a differentiable physical skeleton and $X_t$0 a neural corrector, and then advances the hybrid system through a solver (Shu et al., 31 Jan 2026). For tracers $X_t$1, the skeleton uses flux-form advection–diffusion,

$X_t$2

so conservation structure is retained in the numerical core while unresolved processes are learned (Shu et al., 31 Jan 2026).

These designs differ substantially, but their commonality is clear: they are not generic sequence predictors with ocean data attached. They encode ocean geometry, ocean scales, or ocean balance relations directly into representation and training.

4. Coupling, energetics, and physical consistency

A second hallmark of the ocean-centric model is that atmosphere–ocean interaction is usually asymmetric. The ocean is treated as the slow, memory-bearing subsystem; the atmosphere enters through forcing variables or surface boundary information. Ola makes this asymmetry explicit by using separate neural operators for atmosphere and ocean,

$X_t$3

with atmosphere conditioned on SST and ocean conditioned on 10 m winds, 2 m temperature, and MSLP, plus the cosine of the solar zenith angle (Wang et al., 2024). The model is evaluated primarily through ENSO behavior, upper-ocean thermal structure, and equatorial wave propagation, all of which emphasize the ocean as the dominant source of seasonal memory.

KIST-Ocean adopts a similar forcing–response logic in a global three-dimensional ocean general circulation model. It accurately captures Kelvin and Rossby wave propagation in the tropical Pacific and vertical motions induced by cyclonic and anticyclonic wind stress, thereby representing key ocean–atmosphere coupling mechanisms associated with ENSO (Kim et al., 31 Jul 2025). ORCA-DL is likewise forced by zonal and meridional wind stress rather than fully coupled atmospheric evolution, yet it achieves strong skill for ENSO, PDO, IPO, AMO, and upper-ocean marine heatwaves on seasonal-to-decadal horizons (Guo et al., 2024).

In wave forecasting, OceanCastNet organizes the problem around energy balance rather than generic autoregression. Its mapping

$X_t$4

makes wave height depend explicitly on prior wave state and a short window of wind forcing (Zhang et al., 2024). The paper argues that this preserves an “energy-rich” long-term state and avoids the dissipation common in long autoregressive rollouts. HybridOM pursues physical consistency through differentiable flux-form physics and flux-gated regional downscaling, so that regional high-resolution states are constrained by coarse-scale fluxes rather than by unconstrained super-resolution of state variables (Shu et al., 31 Jan 2026).

This body of work suggests two recurring principles. First, ocean-centric models typically allocate asymmetry to coupling: the atmosphere is fast forcing, the ocean is slow memory. Second, physical consistency is pursued not only through explicit conservation laws but also through architectural bias, forcing design, and diagnostics tied to mesoscale variance, EKE, OHC, or wave energetics.

5. Observation-driven ocean-centric foundation models

Ocean-centricity is not confined to dynamical forecasting. In remote sensing, it denotes pretraining on marine observations and encoding ocean-relevant descriptors or physics in the latent representation. OceanMAE is an ocean-specific masked autoencoder for Sentinel-2 imagery that augments multispectral input with bathymetry, chlorophyll level, and Secchi depth during self-supervised pretraining (Stamer et al., 9 Apr 2026). It is evaluated on MADOS and MARIDA for marine pollutant and debris segmentation and on MagicBathyNet for bathymetry regression, with the strongest gains on marine segmentation and competitive, task-dependent gains on bathymetry (Stamer et al., 9 Apr 2026).

OceanSAR-2 does the same for ocean SAR. It is trained on physically calibrated Sentinel-1 Wave Mode $X_t$5 scenes with improved SSL training and dynamic data curation, and transfers to geophysical pattern classification, significant wave height estimation, ocean surface wind estimation, and iceberg detection (Tuel et al., 12 Jan 2026). The paper argues that this is ocean-centric because the entire design—data, architecture, SSL objective, and benchmarks—is optimized around the physics and phenomenology of the ocean as seen by SAR rather than around land-dominated EO statistics (Tuel et al., 12 Jan 2026).

A related Sentinel-3 ocean-colour foundation model adapts the Prithvi-EO masked autoencoder to OLCI and optionally SLSTR SST, pretraining on global ocean tiles stratified by Longhurst provinces and then fine-tuning for chlorophyll-a and primary production (Dawson et al., 25 Sep 2025). The model is designed around ocean colour as a proxy for phytoplankton biomass, pigments, and bio-optical properties, and shows strong label efficiency when fine-tuned on small, high-quality in-situ datasets (Dawson et al., 25 Sep 2025).

These observation FMs enlarge the meaning of “ocean-centric model.” In this setting, the ocean is not primarily a prognostic dynamical field but a domain whose sensor physics, spatial statistics, and semantic tasks require specialized pretraining rather than transfer from land-heavy EO corpora.

6. Ocean-domain language and multimodal knowledge models

The same paradigm now appears in language and multimodal foundation models. OceanGPT is described as the first-ever LLM in the ocean domain, pre-trained on 67,633 cleaned ocean science documents and instruction-tuned on an ocean-specific instruction set exceeding 150,000 examples generated through a multi-agent framework (Bi et al., 2023). Its benchmark, OCEANBENCH, covers 15 task types and is intended to test ocean knowledge and reasoning rather than generic dialogue (Bi et al., 2023).

OceanPile generalizes this further by providing OceanCorpus, OceanInstruction, and OceanBenchmark as a large-scale multimodal corpus for ocean foundation models (Xue et al., 25 Apr 2026). OceanCorpus integrates sonar, underwater imagery, marine science visuals, and scientific text; OceanInstruction is synthesized through a pipeline guided by a hierarchical Ocean Concept Knowledge Graph; and OceanBenchmark evaluates textual and multimodal marine competence (Xue et al., 25 Apr 2026). Fine-tuning Qwen-based models on OceanPile yields substantial gains on Ocean Science VQA, Sonar VQA, and Marine Organisms VQA, indicating that ocean-centricity in LLMs can be induced primarily through domain-aligned data and task design rather than through new neural architectures (Xue et al., 25 Apr 2026).

A plausible implication is that ocean-centricity is becoming modality-agnostic. In forecasting models it centers state evolution and forcing; in remote sensing it centers spectral or SAR phenomenology; in language and multimodal systems it centers marine knowledge graphs, ocean-specific instructions, and marine evaluation suites.

7. Limitations, controversies, and future directions

A consistent limitation across ocean-centric AI forecasting models is the incomplete treatment of physical constraints. AI-GOMS explicitly notes that exact mass or energy conservation is not embedded in the loss and that fidelity is inherited primarily from reanalysis and architecture (Xiong et al., 2023). NeuralOM likewise relies on architectural priors and prescribed true atmospheric forcing, while modeling only the upper $X_t$6 m and not the deeper ocean (Gao et al., 27 May 2025). AxiomOcean preserves upper-ocean vertical structure effectively but is limited to 643 m and to physical variables, with no biogeochemical tracers or explicit coupled sea-ice dynamics (Wu et al., 11 May 2026).

Observation-driven foundation models face a different set of issues. OceanMAE notes that benefits are strongest for segmentation and are task-dependent for bathymetry, while descriptor quality and alignment can be limiting (Stamer et al., 9 Apr 2026). OceanSAR-2 is trained only on Sentinel-1 Wave Mode and presently uses single-channel $X_t$7, so cross-mode, cross-mission, and multi-polarization generalization remain open problems (Tuel et al., 12 Jan 2026). The Sentinel-3 ocean-colour FM is constrained by sparse labeled data and by the range of chlorophyll and primary-production observations used for fine-tuning (Dawson et al., 25 Sep 2025).

Language and multimodal ocean models inherit the familiar difficulties of domain LLMs. OceanGPT explicitly notes data bias and hallucination as persistent problems (Bi et al., 2023). OceanPile, despite multi-stage quality control, still depends on heterogeneous open-access sources and automatically enriched multimodal annotations, which implies residual noise and uneven regional or sensor coverage (Xue et al., 25 Apr 2026).

Future work in this area is converging on several themes. One is tighter physics integration: HybridOM moves toward this with a differentiable numerical skeleton and flux-form constraints (Shu et al., 31 Jan 2026). Another is fuller coupling: Ola, KIST-Ocean, and ORCA-DL all indicate that ocean-centric components can serve as the memory-bearing core of larger AI Earth-system models (Wang et al., 2024). A third is depth and modality expansion: extending upper-ocean models toward full-depth circulation, adding biogeochemistry or sea ice, and linking forecasting backbones to ocean observation and ocean knowledge FMs (Wu et al., 11 May 2026). This suggests that the mature form of the ocean-centric model may be a family of interoperable systems—forecasting, observational, and multimodal—organized around the ocean as a central dynamical, observational, and semantic domain.