This session aims to explore the application of ML methods in ocean science, with a focus on advancing our understanding and addressing key challenges in the field. Our objective is to foster discussions, share recent advancements, and explore future directions in the field of ML methods for ocean science.
A wide range of machine learning techniques can be considered including supervised learning, unsupervised learning, interpretable techniques, and physics-informed and generative models. The applications to be addressed span both observational and modeling approaches.
Observational approaches include for example:
- Identifying patterns and features in oceanic fields
- Filling observational gaps of in-situ or satellite observations
- Inferring unobserved variables or unobserved scales
- Automating quality control of data
- Modeling approaches can address (but are not restricted to):
- Designing new parameterization schemes in ocean models
- Emulating partially or completely ocean models
- Parameter tuning and model uncertainty
The session also welcomes submissions at the interface between modeling and observations, such as data assimilation, data-model fusion, or bias correction.
Researchers and practitioners working in the domain of ocean science, as well as those interested in the application of ML methods, are encouraged to attend and participate in this session.
Hybrid modeling integrates data-driven Machine Learning (ML) components, such as Neural Networks (NN), into physics-based numerical models to improve the accuracy, stability, and adaptability of dynamical simulations. Rather than replacing established physical laws, hybrid models augment them by learning corrections that compensate for unresolved processes, reduce systematic biases, or dynamically calibrate uncertain parameters.
In oceanic and atmospheric numerical models, unresolved dynamics are represented through sub-grid-scale (SGS) parameterizations coupled to the Navier–Stokes equations. As these parameterizations constitute a major source of uncertainty, recent work has increasingly explored Artificial Intelligence (AI) to improve their modeling and constraint. A particularly promising strategy is online learning, in which the AI model is embedded within the numerical solver and trained while interacting with the evolving system dynamics. This setup allows the model to learn temporal dependencies across multiple solver steps and to optimize long-term behavior. Although online learning has demonstrated improved forecast skill and stability over long horizons compared to the more widely used offline learning strategy, its application to high-dimensional ocean models is limited by two key challenges: the requirement for fully differentiable solvers and the high computational and memory costs associated with backpropagation through long trajectories.
To overcome these limitations, we introduce a new family of gradient-approximation methods that selectively simplify intermediate Jacobians in the backpropagation chain. The resulting gradients closely approximate the exact full gradients over long trajectories, preserving the dominant sensitivities required for effective online learning and substantially reducing computational and memory overhead.
We evaluate the proposed methods using two case studies of increasing complexity. We first consider a hybrid neural–Lorenz-63 model in which an AI component compensates for missing dynamics. The framework is then extended to a semi-realistic hybrid quasi-geostrophic model of the Northwestern Mediterranean Sea, demonstrating two complementary enhancement strategies: the calibration of a biased physical parameter (bottom drag) and a NN-based correction of bottom-layer momentum tendencies. Together, these experiments show that our Jacobian-approximation strategies enable stable and efficient online learning across both low-dimensional chaotic systems and high-dimensional ocean models. Although our configurations remain simpler than fully operational ocean models, our results provide a foundation for scaling online learning to realistic ocean applications and, ultimately, for integrating AI-based corrections into next-generation forecasting systems.
How to cite: González Zamora, E., Ouala, S., and Tandeo, P.: Efficient Gradient-Approximation Methods for Online Learning in Hybrid Neural–Physical Ocean Models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2346, https://doi.org/10.5194/egusphere-egu26-2346, 2026.
Incorporating physical laws into neural networks has long been a central topic in geophysical machine learning. While purely data-driven approaches can achieve strong prediction skill, they often lack physical consistency and degrade under sparse observations or long lead times. In this study, we impose a simple yet fundamental constraint, global volume conservation, by introducing a dedicated volume-conserving layer into neural networks. We apply this volume-conserved network in both an idealized shallow-water model and a realistic global sea level anomaly prediction task, and show systematic improvements in prediction skill, reaching up to 25%. The improvement increases as observation points decreasing and leading time increasing, and the predictions follow physical laws strictly. In addition, although post-processing also enforce physical consistency, the constrained model achieves substantially lower prediction errors, with reductions of up to 15%. These results demonstrate the effectiveness of embedding hard physical constraints as network layers for improving both accuracy and physical fidelity.
How to cite: Li, Y. and Tang, Y.: Improving Global Sea Level Prediction with Hard Physical Constraints in Neural Networks, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6777, https://doi.org/10.5194/egusphere-egu26-6777, 2026.
Substantial natural variability can obscure the detection of anthropogenic long-term trends in the marine carbonate system (e.g., ocean acidification). Yet the magnitude of the trends and variability remains poorly constrained due to limited marine carbonate system observations. Here, we use a Bayesian machine-learning approach based on Gaussian Process Regression (GPR) to decompose total variability of ocean acidification-related variables into seasonal, interannual and long-term components. The method is first applied to three decades of observations from the Line P carbon program, the longest marine carbonate system timeseries in the Northeast Pacific (1990-2019), typically taking samples three times per year. We found that over the period from 1990 to 2019, the local oceanic uptake of anthropogenic carbon dioxide from the atmosphere was the main driver of long-term changes in the marine carbonate system, including acidification. The seasonal cycle of dissolved inorganic carbon and the aragonite saturation state (both indicators of ocean acidification) was the dominant contributor to total variability in the top 60-70 m of the water column, with a mean surface seasonal amplitude of 35 ± 3 µmol kg−1 and 0.31 ± 0.04, respectively. In this depth range, the magnitude of the interannual variability was at least half of the seasonal variability for most variables. We then apply GPR to output from a global ocean biogeochemical model subsampled as per availability of observations, to assess the observational effort required to detect future ocean carbon trends, with a particular focus on detecting signals related to potential marine carbon dioxide removal interventions.
How to cite: Franco, A. C., Monahan, A. H., Ianson, D., and Bernardello, R.: Using Gaussian Process Regression to disentangle marine carbonate system trends and variability, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-12925, https://doi.org/10.5194/egusphere-egu26-12925, 2026.
Marine biogeochemical models typically contain tens to hundreds of parameters and are notoriously challenging to tune to sparse and noisy observations, in particular for specific regional conditions. While ensemble-based methods can automate this process and are also used for data assimilation, they do not scale well to large numbers of unknown parameters. Gradient-based methods, on the other hand, scale well with high dimensionalities but require adjoint models. However, state-of-the-art differentiable programming frameworks such as PyTorch eliminate the need for manual adjoint implementations through automatic differentiation, that is, by using the chain rule to automatically compute analytic derivatives.
We introduce a fully differentiable framework for tracer transport and marine biogeochemical (BGC) processes in PyTorch. We implement advection and diffusion operators based on popular models written in Fortran, e.g. the General Ocean Turbulence Model (GOTM) for water columns. As GOTM's vertical mixing formulation requires implicit time stepping, we provide efficient differentiable solvers for batched tridiagonal systems with custom backward methods derived by implicit differentiation. Furthermore, our framework includes a PyTorch base class for differentiable BGC models with an interface similar to the Framework for Aquatic Biogeochemical Models (FABM). We provide several examples, including a re-implementation of the popular ecosystem model ECOSMO. As our operators are implemented in PyTorch, they can easily be combined with established neural network layers and optimizers.
We demonstrate our framework by performing model tuning and data assimilation in BGC models using 4DVar on sparse and noisy observations. We investigate the scaling behaviour of our tridiagonal solver for various batch and system sizes with both GPU and CPU computation. Our contribution has the potential to enhance data assimilation, speed up parameter tuning workflows and improve the accuracy of biogeochemical modelling.
How to cite: Nimtz, P. R., Demir, K. T., Zinchenko, V., Frion, A., and Greenberg, D. S.: Fully differentiable transport operators enable gradient-based parameter tuning and data assimilation of marine biogeochemical models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19761, https://doi.org/10.5194/egusphere-egu26-19761, 2026.
Data-driven approaches, particularly deep learning, are rapidly transforming earth system modeling. OceanBench has established a standardized benchmark for global short-range data-driven ocean forecasting, providing operationally consistent datasets and evaluation protocols that support reproducible development and assessment of ML-based ocean forecasting systems.
Building on this foundation, we introduce new extensions to OceanBench that broaden its accessibility and applicability under realistic computational constraints. These include the integration of coarser-resolution (~1°) global models, enabling computationally efficient experimentation, regional evaluation capabilities, and the inclusion of new candidate models spanning both physics-based and machine-learning approaches. By supporting multiple resolutions and modeling paradigms, the extended OceanBench framework enables more flexible and application-relevant assessment of ocean forecasts, accelerating research and operational adoption of data-driven and hybrid ocean modeling systems.
How to cite: El Aouni, A., Gaudel, Q., Aissa-Abdi, Z., Bricaud, C., and Ruggiero, G.: OceanBench: A Benchmark for Data-Driven Global Ocean Forecasting systems, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20905, https://doi.org/10.5194/egusphere-egu26-20905, 2026.
Data-driven models promise higher-fidelity Earth system forecasts at a fraction of the computational cost of numerical models, enabling the use of large ensembles for more robust statistics. Consequently, the number of purely data-driven atmospheric models has grown explosively in recent years. However, the sheer diversity of architectures and the absence of a clear "winner" pose a significant design challenge for those seeking to replicate these successes in oceanography.
GraphCast was one of the first models in this arena and remains state-of-the-art. Based on graph neural networks, it lacks specific atmospheric inductive biases, such as fixed physical dimensions, conservation laws, or explicit evolution equations. Its only relational inductive bias is the physical proximity between interacting elements. When provided with an appropriate graph, this principle should hold equally well for the ocean, making GraphCast an ideal candidate for cross-domain application.
To test this hypothesis, we introduce ARCO-OCEAN: a new Analysis-Ready, Cloud-Optimized curated dataset designed for training such models. ARCO-OCEAN contains global reanalyses and hindcasts of multiple Earth system components, including ocean physical state, waves, sea ice, and atmospheric/hydrological forcing. Widely available through the AWS Open Data program, this dataset decouples AI/ML-related methodological development from domain-specific scientific knowledge (e.g., variable selection, spatial and temporal resolution) and data engineering (e.g., choice of format, chunking), relieving data scientists of the heavy burden of data preparation.
We detail the specific design choices of ARCO-OCEAN intended for coupled atmosphere-ocean modeling at subseasonal-to-seasonal timescales. Finally, by equipping GraphCast with land-masking capabilities and a global ocean mesh graph, we present preliminary results on its training performance within the ocean domain.
How to cite: Campanella, S., Salon, S., Querin, S., and Bortolussi, L.: Can GraphCast learn skillful subseasonal-to-seasonal global ocean forecasting using the ARCO-OCEAN testbed?, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20245, https://doi.org/10.5194/egusphere-egu26-20245, 2026.
Sea surface height (SSH) derived from satellite altimetry is essential for oceanographic research and marine monitoring. To improve SSH prediction accuracy, we propose a set of physics-informed methods based on neural networks (NNs). The main strategies include: (1) integrating a geostrophic constraint (GC) into the loss function; (2) incorporating land mask information (MI) to mitigate artifacts introduced by the land points in ocean data.
Utilizing altimeter satellite gridded absolute dynamic topography data, we evaluate three mainstream spatiotemporal predictive NNs—SimVPv2 (SV), PredRNNv2 (PR), and PredFormer (PF)—each exhibiting distinct inductive biases inherent to their architectures, to assess their performance under the proposed strategies. The results indicate that both strategies can significantly improve SSH prediction, though their effects vary across architectures. While SV shows limited improvement from MI, PR benefits the most, which can likely be attributed to its gating mechanism and recurrent architecture. In contrast, GC enhances the performance of SV more effectively than that of PR. However, both strategies degrade the performance of PF, a Vision Transformer (ViT)-based model that differs fundamentally from SV and PR. To our knowledge, this study is the first to identify land-induced artifacts in spatiotemporal predictive NNs and to implement a land mask input strategy to mitigate their impact on ocean forecasting.
Building upon these findings, we further explored the potential of multivariable inputs. Contrary to expectations, our experiments of concatenating wind speed with SSH as inputs reveal that directly combining heterogeneous oceanic variables is suboptimal. This finding highlights a broader multimodal integration problem in applying NNs to oceanography, which remains an open challenge.
How to cite: Huang, L., Shu, Y., Yao, J., and Liu, D.: Investigation of Physics-Informed Methods for Improving Sea Surface Height Prediction Based on Neural Networks , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9123, https://doi.org/10.5194/egusphere-egu26-9123, 2026.
End-to-end neural schemes have become state-of-the-art approaches for the reconstruction of ocean variables from irregularly-sampled observations, especially for sea surface dynamics (e.g., SST, SLA, ocean colour…).While most studies rely on the direct application of state-of-the-art architectures developed in imaging science, especially Unets, a class of approaches explicitly leverage state-space formulation and generalize in a neural fashion established data assimilation schemes such as 4DVar algorithms and EnKF schemes. Most of these approaches have been demonstrated for toy examples or intermediate-complexity case-studies. Here, we focus on 4DVarNet architectures which generalizes weak-constraint 4DVar solvers. Drawing inspirations from unrolled neural architectures used in computational imaging, especially in diffusion and flow matching models, we extend the original 4DVarNet architectures to a broader class of unrolled architectures which differ according to the specific parameterization of the considered iterative residual update. Leveraging diffusion-based Unet schemes with time embedding blocks, the resulting 4DVarNet schemes range from 1-million-parameter configurations to 50-million-parameter ones. Through an application to satellite altimetry and Sea Level Anomaly mapping, we assess the performance of the proposed architectures. Our contributions are three-fold: (i) we report state-of-the-art performance of considered neural global SLA mapping schemes compared to the state-of-the-art (eg, MIOST, NeuROST); (ii) unrolled architectures with just very few iterations, typically 5 to 10, reach the best mapping performance, (iii) the best unrolled architecture explicitly benefits from the knowledge conveyed by the underlying variational representation of the mapping problem. We discuss how these results could pave the way towards at-scale demonstrations of end-to-end neural DA schemes for the reconstruction of global ocean states from partial observations, including uncertainty quantification issues.
How to cite: Fablet, R., Zhu, D., de Nauily, P., Botvynko, D., and le Sommer, J.: Scaling End-to-end neural DA up to real-world problems: a case study for global-scale SLAmapping and 4DVarNets, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-21696, https://doi.org/10.5194/egusphere-egu26-21696, 2026.
The SWOT (Surface Water Ocean Topography) mission is currently providing unpreceded high-resolution measurements of Sea Surface Height (SSH), revealing ocean features at finer scales. Nevertheless, the two-dimensional observations of KaRIn altimeter of SWOT suffer from instrumental and geophysical correction errors. This noise degradation is polluting the high frequencies of SWOT signal, thus hiding the submesoscale dynamics from oceanographers. For this reason, Tréboutte et al. (2023) has developed a convolutional neural network (CNN) based on UNet architecture to separate the noise from the physical signals contained in the SSH. This work has already demonstrated great results on SWOT measurements. However, last version of the algorithm delivers poor performance in certain oceanic conditions. Therefore, we modify the training procedure to obtain a more robust version of the algorithm. We show that we manage to mitigate these issues significantly, avoiding biases and artefacts in the denoised observations.
This data is also incomplete. SWOT measurements are sometimes distorted by various factors, such as rain cells, boats, icebergs, etc. To address these errors, editing is applied to remove erroneous pixels from the data. However, this lost data is valuable to many users. That is why we have also developed a deep learning inpainting methodology using a CNN to retrieve the missing physical information. We demonstrate that it is possible to accurately restore measurements lost after the editing step, better than classical interpolation approaches.
How to cite: Meis, G., Tréboutte, A., Pujol, M.-I., Ballarotta, M., and Dibarboure, G.: Enhancing two-dimensional SWOT oceanic measurements using deep learning approaches for denoising and inpainting, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11229, https://doi.org/10.5194/egusphere-egu26-11229, 2026.
Sea surface temperature (SST) is a fundamental variable influencing the variability of the ocean and atmosphere on synoptic, decadal and climate timescales. Satellites play a major role in its estimation and particularly measurements from infrared (IR) radiometers, which provide high-resolution observations of SST. However, IR retrievals are contaminated by the presence of clouds that are therefore removed resulting in gaps in the retrieved fields. Because many applications rely on a gap-free SST field, including marine heatwaves studies and ocean reanalysis, a high-quality reconstruction of missing SST is required.
Traditional techniques to address this issue include Empirical orthogonal functions (EOFs) and Optimal interpolation (OI). However, those techniques often result in over-smoothing, even where observations are present. Recently, deep learning (DL) techniques have been employed, leveraging their capacity of capturing non-linearities to better reconstruct data with gaps.
Recently Asperti et al. (2025) developed DL models based on U-Net and transformer architectures with several configurations implemented in the Italian Seas to reconstruct SST using Level 3 products. The results show that DL based models are promising to reconstruct SST fields even close to complex coastlines. In this work, we extend the methodology introduced in their study to the entire Mediterranean Sea, starting from the best performing configuration, based on U-Net architecture. Here the method used to train the neural network is to add an additional cloud mask from a randomly picked day, to the input SST, in order to have a ground truth to use for the loss computation. The extended Mediterranean Sea model skill is comparable to the model in Asperti et al. (2025) on the overlapping regions. Since the modulation of observed fields is negligible by U-Net, our model shows better skill compared to the Level 4 products based on OI. Finally, we will also present results from an independent validation against in-situ drifter SST observations that are mainly located in the western Mediterranean basin. Level 3 SST products show discrepancies relative to drifters in terms of both overall error and mean bias, which are preserved by the U-Net in cloud-free regions. In reconstructed regions, only a modest degradation in skill relative to drifter observations is observed, indicating that the reconstruction introduces limited additional error.
Deep Learning for Sea Surface Temperature Reconstruction under Cloud Occlusion by Asperti et al. 2025. Applied Ocean Research. In review. https://arxiv.org/abs/2412.03413
How to cite: Tartufoli, B., Aydogdu, A., Pinardi, N., Asperti, A., and Oddo, P.: Sea surface temperature reconstruction in the Mediterranean Sea using deep learning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-21240, https://doi.org/10.5194/egusphere-egu26-21240, 2026.
Accurately forecasting Sea Surface Temperature (SST) is critical for understanding ocean dynamics, climate change impacts, and marine ecosystem management (Brito-Morales et al., 2020; Gattuso et al., 2018). In recent years, Graph Neural Networks (GNNs) have emerged as a powerful tool for spatiotemporal oceanographic forecasting, offering advantages over traditional Euclidean deep learning models by operating on unstructured grids (Liang et al., 2023; Zhang et al., 2025). However, the transition from structured satellite-derived data to mesh-based representations often introduces numerical artifacts, particularly due to the grid-to-mesh coupling mechanisms (Cuervo-Londoño et al., 2026; Cuervo-Londoño, Sánchez, et al., 2025).
This study investigates the origin of "Voronoi-induced artifacts" in GNN architectures applied to SST forecasting in the Northwest African region and the Canary Islands. We demonstrate that the grid-to-mesh association is algebraically equivalent to an order-k Voronoi partition (Cuervo-Londoño, Reyes, et al., 2025; Okabe et al., 2000), implying that the way nodes are distributed and how they associate with the underlying data grid significantly influences the quality of the predictions. To address these issues, we propose and evaluate four different mesh configurations: structured quadrangular meshes (Holmberg et al., 2024; Lam et al., 2023) and three unstructured approaches, including novel bathymetry-aware meshes.
Our findings reveal that connectivity plays a decisive role in mitigating artifact formation. Specifically, using approximately four connections per node under optimized grid-to-mesh association rules significantly reduces errors. Furthermore, the results show that densifying the node distribution according to the seabed’s topography (bathymetry) not only reduce spatial artifacts but also increases forecast accuracy. The bathymetry-based meshes with optimized connectivity (3-4 connections) achieved a 30% improvement in performance compared to traditional structured mesh baselines. These insights suggest that incorporating geographical and topological priors into GNN design is essential for developing robust and reliable machine-learning surrogates for physical oceanography (Reichstein et al., 2019).
Acknowledgments: This work was supported by the projects SIRENA and SIRENA 2, funded by the collaboration of the Biodiversity Foundation of the Ministry for the Ecological Transition and the Demographic Challenge, through the Pleamar Program, and are co-financed by the European Union through the EMFAF (European Maritime, Fisheries and Aquaculture Fund).
References
Cuervo-Londoño, G. A., Reyes, J. G., Rodríguez-Santana, Á., & Sánchez, J. (2025). Voronoi-Induced Artifacts from Grid-to-Mesh Coupling and Bathymetry-Aware Meshes in Graph Neural Networks for Sea Surface Temperature Forecasting. Electronics, 14(24), 4841. https://doi.org/10.3390/electronics14244841
Cuervo-Londoño, G. A., Sánchez, J., & Rodríguez-Santana, Á. (2025). Deep Learning Weather Models for Subregional Ocean Forecasting: A Case Study on the Canary Current Upwelling System (No. arXiv:2505.24429). arXiv.https://doi.org/10.48550/arXiv.2505.24429
Cuervo-Londoño, G. A., Sánchez, J., & Rodríguez-Santana, Á. (2026). Forecasting Sea Surface Temperature from Satellite Images with Graph Neural Networks. In M. Castrillón-Santana, C. M. Travieso-González, O. Deniz Suarez, D. Freire-Obregón, D. Hernández-Sosa, J. Lorenzo-Navarro, & O. J. Santana (Eds.), Computer Analysis of Images and Patterns (pp. 329–339). Springer Nature Switzerland. https://doi.org/10.1007/978-3-032-05060-1_28
How to cite: Cuervo Londoño, G. A., Rodríguez Santana, Á., and Sánchez, J.: Mitigating Voronoi-induced artifacts in GNN-based sea surface temperature forecasting using bathymetry-aware adaptive meshes, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-15146, https://doi.org/10.5194/egusphere-egu26-15146, 2026.
The Argo Program is a global network of 4,000 autonomous drifting floats that provide essential, real-time data on the upper 2,000 meters of the ocean. By measuring temperature and salinity, Argo has become the primary source of information for monitoring ocean warming, sea-level rise, and climate variability. However, the massive volume of data generated—totaling millions of profiles—presents a significant challenge for Quality Control (QC).
Traditionally, delayed-mode quality control has relied heavily on human expertise and the "trained eye" of scientists to identify instrumental drifts and sensor malfunctions. To address the cost and limitations of manual inspection, we introduce Argo-YOLO, an innovative approach that transposes computer vision techniques into the field of physical oceanography.
By converting oceanographic profiles into graphical representations, our system utilizes the YOLO (You Only Look Once) deep learning architecture to "scan" the data, mimicking the visual diagnostic capabilities of expert oceanographers. This method enables high-speed, systematic detection of instrumental drifts, sensor malfunctions, and profile anomalies across the entire Argo dataset while maintaining the nuanced precision of human analysis.
Initial results demonstrate that Argo-YOLO faithfully reproduces expert visual diagnostics with high performance: 97% accuracy in identifying valid profiles with only 3% false alarms, and 96% success in detecting anomalous profiles with 4% missed detections.
These results confirm the viability of computer vision for operational oceanographic quality control.
Argo-YOLO demonstrates how computer vision can be successfully adapted to oceanographic challenges, representing a major step toward automated, scalable quality control in global ocean observing systems and ensuring the integrity of long-term climate records in an era of "Big Data" oceanography.
How to cite: Carval, T., Tosello, V., Dobler, D., and Lebeaud, A.: Argo-YOLO: Leveraging Computer Vision for Automated Quality Control of Argo Ocean Profiles, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11527, https://doi.org/10.5194/egusphere-egu26-11527, 2026.
Climate change is reshaping ocean ecosystems faster than we can monitor them. Predicting shifts in productivity, carbon uptake, and oxygen levels requires forecasting interacting biogeochemical variables, a task where traditional process-based models struggle with computational cost and parameter uncertainty.
BG4Sea is a machine-learned seasonal forecast that was trained on Mercator Océan's operational biogeochemical analysis. The model can generate skillful seasonal predictions of the carbon cycle, nutrients, oxygen, pH, chlorophyll, and plankton dynamics at a fraction of the computational cost, all while remaining competitive even at longer forecasting horizons. However, while the model demonstrates skill when evaluated against reanalysis data, this is likely to share the parametrization assumptions and constraints that are characteristic of process-based models.
This contribution explores strategies for evaluating against real-world measurements and for using observations to guide and constrain the model. We investigate “global-first” approaches, which prioritize remote-sensing data, as well as “regional-first” approaches, which use the model’s grid-independent structure to produce region-specific updates from in-situ stations.
How to cite: Martinez Balbontin, G., Charantonis, A., Bereziat, D., and Ciavatta, S.: Guiding Machine-Learned Biogeochemical Forecasts with Observations, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13007, https://doi.org/10.5194/egusphere-egu26-13007, 2026.
Harmful algal blooms (HABs) are massive proliferations of microalgae in aquatic ecosystems that may be harmful to the ecosystems or to society. Predicting HABs in spatially complex coastal environments requires understanding the potential environmental drivers that may determine microalgal population dynamics. When considering the study of HABs we may evaluate if these processes are spatially invariant or if they demonstrate site-specific dynamics. Machine learning models often achieve high training performance but fail when extrapolating to unseen locations due to site-specific overfitting. We developed a methodological framework integrating hierarchical modelling, spatially explicit machine learning, and interpretable AI techniques to quantify spatial heterogeneity in HAB environmental drivers.
Gambierdiscus spp is a genus of benthic marine microalgae (dinoflagellate) that are found in coastal areas and that produce potent marine toxins which are transferred mainly to fish. We analysed 348 observations of Gambierdiscus spp. abundances across 32 sites in the Balearic Islands (2021-2024), integrating field abundance data with satellite-derived oceanographic variables (temperature, nutrients, hydrodynamics) from Copernicus Marine Service. Seven modelling approaches were compared: Generalized Additive Mixed Models (GAMM), Generalized Additive Models (GAM), Geographically Weighted Regression (GWR), Random Forest (RF), Geographic Random Forest (GRF), XGBoost, and Geographic XGBoost. A three-phase feature selection procedure (temporal lag optimization, collinearity removal via VIF, LASSO regularization) reduced 61 candidate predictors to 12 ecologically interpretable variables optimized for spatial modelling.
Model validation employed Leave-One-Out Cross-Validation (LOO-CV) to assess true spatial generalization rather than interpolation. Machine learning models achieved high training performance (R²=0.75-0.85) but collapsed under spatial extrapolation (R²_LOO=0.30-0.40). In contrast, GAMM demonstrated superior spatial transferability (R²_LOO=0.47), attributable to its explicit separation of fixed environmental relationships from hierarchical site-specific random effects. SHAP (SHapley Additive exPlanations) analysis on island-stratified Random Forest models quantified spatial non-stationarity: temperature importance varied 13-fold across islands (SHAP: 0.05-0.64), while phosphate exhibited 2.6-fold consistency (SHAP: 0.10-0.26). Partial dependence plots verify that drivers operate through fundamentally different mechanisms across the archipelago.
Significant spatial clustering (Moran's I=0.346, p<0.001) with persistent hotspots and coldspots validated non-stationarity. Phosphate emerged as the only universal driver, while temperature, substrate, and hydrodynamics exhibited location-dependent effects. Our findings demonstrate that interpretable ML combined with spatial cross-validation effectively diagnoses when environmental relationships transfer versus when they require location-specific calibration, providing a generalizable framework for spatial prediction in heterogeneous ocean systems.
How to cite: Dorado Guerra, D. Y., Gimeno Monforte, S., Alcaraz Cazorla, C., and Diogène Fadini, J.: Spatial Non-Stationarity in Harmful Algal Bloom Drivers for the benthic dinoflagellate Gambieridscus spp in the Balearic Islands, Revealed Through Interpretable Machine Learning and Hierarchical Modelling, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-17203, https://doi.org/10.5194/egusphere-egu26-17203, 2026.
Short-term forecasting of Arctic essential climate variables (ECVs) requires methods that can exploit the growing diversity of forthcoming satellite observations while remaining robust to sparse and heterogeneous sampling. This study targets sea-ice concentration (SIC) and sea-ice thickness (SIT) forecasting using observations from the new Copernicus Sentinel Expansion missions: ROSE-L providing high-resolution (≈500 m) SIC, CIMR delivering intermediate-resolution (≈5 km) SIC and thin sea ice thickness, and CRISTAL altimeter supplying SIT and sea surface height at similar scales. We propose an online multiresolution neural forecasting framework designed to ingest irregular satellite swaths across resolutions and sensor types, and to produce observation-conditioned nowcasts compatible with operational constraints. The model combines multiscale forecast architectures to explicitly handle intermittency, scale disparities, and sensor-dependent information content. Beyond its operational relevance, the framework is used as a research tool to investigate predictability across scales, enabling a systematic analysis of how submesoscale ice processes impact short-term forecast skill at coarser resolutions. Forecast performance is assessed using resolution-aware metrics, revealing scale-dependent gains in ice-edge sharpness, thin-ice variability, and short-lead SIT evolution compared to baseline methods. By explicitly combining ROSE-L, CIMR, and CRISTAL observations within a unified multiresolution framework, this work enables a direct assessment of how high-resolution sea-ice variability propagates across scales and impacts short-term predictability in operational Arctic ECV forecasts.
How to cite: Beauchamp, M., de Nailly, P., le Guillouzic, M., Singha, S., Rasmussen, T., Sievers, I., and Fablet, R.: Multiscale data-driven forecasting of Sea Ice Essential Climate Variables, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-21323, https://doi.org/10.5194/egusphere-egu26-21323, 2026.
Abstract: Argo floats provide a global dataset of subsurface temperature and salinity profiles but lack direct velocity observations. This limits the reconstruction of Lagrangian ocean transport using the Argo data. We propose a physics-informed machine learning emulator that infers latent horizontal velocity fields from Argo hydrographic observations. The model learns a neural velocity representation using 3D temperature–salinity gradients, which is constrained by advection–diffusion equations. This approach implicitly recovers flow patterns that are consistent with the observed changes in properties and enables the simulation of synthetic trajectory without the input of explicit velocity data. Sparse years are handled via physics-based self-supervision and spatio-temporal regularization. Preliminary experiments in the Mediterranean Sea demonstrate that the learned velocities reproduce qualitatively the known major gyres and boundary currents, achieving realistic float displacements and energy spectra that are comparable to those in reanalysis fields. This framework offers a new way to reconstruct Lagrangian dynamics directly from hydrography, providing an efficient, observation-driven alternative to numerical trajectory modeling.
How to cite: Gasana Elysee, M., Pirro, A., Poulain, P.-M., Mauri, E., Manzoni, L., and Menna, M.: Learning Implicit Subsurface Velocity Fields from Argo Hydrography Using Physics-Informed Neural Emulation, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13260, https://doi.org/10.5194/egusphere-egu26-13260, 2026.
Providing accurate and timely warnings on storm surges is essential to limit the impact of flooding in coastal areas. These warnings are based on hydrodynamic models of the area which traditionally rely on numerical solvers to predict water levels. These models are preferably run in an ensemble to also provide uncertainty information about the forecast. In addition to forecasts, these models are also used as part of climate scenarios to provide statistics on storm surges under future climate. A major bottleneck in especially the latter two applications is the computational cost of the model.
In recent years, machine learning models have been developed that can partly or fully emulate numerical models at reduced computational cost once trained, enabling faster forecasts, larger ensembles, or longer climate runs. These emulators come in various forms, from predicting the hydrodynamics of the entire region of interest (more closely mimicking existing numerical models) to predicting water levels at selected points of interest (more closely aligning with available observational data). In this presentation we will discuss our work towards emulating the hydrodynamics of the North Sea for storm surge prediction using either type of emulator. We will demonstrate the performance of the emulators on multiple cases ranging from test problems to more realistic settings. Additionally, we will discuss how known physics of the system or observational data can be incorporated into the surrogate models, with the goal of making the model more generalizable and reducing the data requirements for training.
How to cite: Tromp, W., Zhao, J., and Verlaan, M.: Machine learning emulators for predicting storm surges in the North Sea , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16754, https://doi.org/10.5194/egusphere-egu26-16754, 2026.
In this study, a hybrid architecture combining convolutional neural networks for spatial reconstruction and long short-term memory networks for temporal forecasting is used to predict sea-level variations in the German Bight. This new framework is applied to a series of sea level data ranging from academic to realistic data. Experiments with monochromatic waves demonstrate the model’s ability to deliver accurate short-term forecasts with minimal errors. Forecasts of real tidal constituents, including M2 and the sum of M2 and M4 tides, confirm robust model performance over lead times up to 48 h. A key result is that deep learning can reconstruct basin-wide sea level from a limited number of coastal gauge stations. Therefore, in the forecast experiments, adding data from coastal observations (mimicking data assimilation) significantly improves prediction accuracy. The study highlights the potential of deep learning to supplement traditional numerical models, particularly in regions with dense observational coverage. Key factors influencing model performance are identified, among them spatial signal complexity and steepness of gradients. An overall result is that deep learning can complement numerical models in operational ocean forecasting and provide a valuable tool for evidence-based coastal management in data-rich regions.
How to cite: Yildiz, I., Stanev, E. V., and Staneva, J.: From monochromatic waves to realistic tides: deep learning for short-term forecasting of coastal ocean, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5209, https://doi.org/10.5194/egusphere-egu26-5209, 2026.
Marine research and operational services require accurate sea-surface current (SSC) data. Because direct observations are sparse and spatially incomplete, spatially consistent SSC fields are most commonly obtained from numerical ocean models. These models are physically comprehensive but computationally expensive, as they integrate the full three-dimensional ocean state even when only surface currents are required. This makes their routine use inefficient for applications that primarily need surface information.
Here we develop a convolutional U-shaped neural network to partially emulate daily-mean SSC variability in the Baltic Sea. The emulator is formulated as a one-day state-update operator that predicts next-day zonal and meridional SSC components from the previous-day SSC field and prescribed near-surface atmospheric forcing. The network is trained on nine years (2015–2023) of SSC fields from the Copernicus Marine Service Baltic Sea Physical Reanalysis, together with near-surface atmospheric forcing from the fifth-generation European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis (ERA5), interpolated to the SSC grid. Both datasets are used at 1-nautical-mile spatial resolution and daily temporal resolution. Predictive performance is evaluated on an independent test year (2024).
Occlusion sensitivity-based input selection indicates that SSC persistence (SSC on day t) and near-surface wind forcing (wind on day t+1) capture the dominant controls on day-to-day SSC evolution (SSC on day t+1), allowing the input space to be reduced to four channels by excluding additional atmospheric variables. One-day emulation achieves high skill across most of the basin, with spatially averaged vector errors of 2.4–2.6 cm s⁻¹ and correlations exceeding 0.9. When deployed in an autoregressive mode, errors increase smoothly with lead time and correlations decrease to approximately 0.65 by day 21. However, large parts of the coastal and interior Baltic Sea retain correlations above 0.9 and vector errors below 10 cm s⁻¹ even at multi-week lead times, indicating stable and spatially localized error growth.
To interpret the learned dynamics, we apply two explainability analyses: layer-wise relevance propagation (LRP) and diagonal Jacobian elasticity (DJE). LRP identifies which input information supports the formation of the forecast by propagating the predicted output backward through the network and assigning each input grid point a relevance score that reflects its contribution to the forward computation, independent of local sensitivity or numerical scaling. DJE, which we term here, characterizes how the forecast responds to small input perturbations by using the model’s Jacobian—the set of partial derivatives linking outputs to inputs—to quantify local, co-located sensitivities. The results show that SSC persistence provides the primary structural support for predictions in energetic boundary and strait regions, while wind forcing dominates the local sensitivity of predicted SSC over the interior basin and offshore waters. These diagnostics indicate that the network learns physically plausible state-memory and wind-driven adjustment patterns rather than relying on diffuse, non-local correlations.
How to cite: Barzandeh, A., Manß, C., Stahl, F., Maljutenko, I., Rikka, S., and Raudsepp, U.: Partial Emulation of Simulated Sea-Surface Currents in the Baltic Sea: An Assessment of Explainability and Potential Forecast Skill, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-17936, https://doi.org/10.5194/egusphere-egu26-17936, 2026.
Sea surface temperature (SST) is critically important for understanding ocean dynamics and supporting various marine activities, making accurate short-term SST forecasting highly significant. However, accurately modeling the multi-scale variability of SST remains challenging for existing deep learning (DL) models. This study introduces the coupled Transformer–CNN network (CoTCN), a hybrid architecture designed to leverage the multiscale variability of SST. The CoTCN combines the strengths of Transformers and convolutional neural networks (CNNs), significantly enhancing SST forecasts’ spatial continuity and predictive accuracy. Compared to five state-of-the-art DL models based on Transformers or CNNs that include convolutional long short-term memory (ConvLSTM), ConvGRU, adaptive Fourier neural operator (AFNO), PredRNN, and SwinLSTM, the CoTCN demonstrates superior performance in global and local areas of SST forecasting. At 1-day lead time, the CoTCN reduces the global average root-mean-square error (RMSE) by over 15%, with forecast errors ranging from 0.20 °C to 0.53 °C across 1–10-day lead times. Moreover, the CoTCN effectively mitigates the checkerboard artifacts inherent to the Vision Transformer (ViT) architecture. These findings highlight the effectiveness of the CoTCN in capturing SST’s multiscale features and underscore the promising potential of hybrid architectures for future DL models.
How to cite: Zhang, T., Lin, P., Liu, H., Wang, P., Wang, Y., Xu, K., Zheng, W., Li, Y., Jiang, J., Zhao, L., and Chen, J.: A Coupled Transformer-CNN Network: Advancing Sea Surface Temperature Forecast Accuracy, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-8845, https://doi.org/10.5194/egusphere-egu26-8845, 2026.
The Surface Water and Ocean Topography (SWOT) satellite mission, launched in December 2022, provides revolutionary measurements of the sea surface height (SSH) variations with unprecedented spatial resolution down to Ο(1 km). As a result, SWOT products have significant potential in monitoring ocean dynamics down to the submesoscale. However, the repeat cycle of 21 days introduces a barrier to fully capture these dynamics as they vary on the order of days. To fully exploit the potential of the satellite mission and simplify processing requirements for potential users, we propose a physics-informed neural network (PINN) to generate gridded SSH products from SWOT L3 along-track snapshots. The neural network has a U-Net-like architecture combined with residual learning to consider the spatial variations of the SSH field, and takes time, geolocations, and gridded SSH from conventional altimetry missions as input features, while the SWOT observations serve as ground truth. In addition to the classical data loss, the PINN model applies direct constraints on the model's trainable parameters by forcing them to fulfill the next-order correction of the quasi-geostrophic theory (SQG+1), which has been demonstrated to be able to capture cyclogeostrophic balance and frontogenesis attributed to submesoscale dynamics. To this end, the high resolution of SWOT observations is kept, while the velocities and pressure fields associated with the SQG+1 theory are predicted. We conducted experiments using both simulated data and real-world data. Both experiments demonstrate the benefits of incorporating physical loss to achieve higher generalizability, thereby filling the gaps between SWOT tracks reasonably. Based on the real-world data, 2-km gridded SSH products with a temporal resolution of five days are achieved. The proposed method shows promising potential for generating high-resolution gridded products while considering physical constraints. The product will be beneficial for the community to analyze mesoscale to submesoscale ocean dynamics, and compare with other sources of surface and in-situ data in the upper ocean.
How to cite: Gou, J., Dù, R. S., Smith, K. S., Soja, B., and Bodner, A.: Physics-informed neural network for gridded SSH from SWOT observations considering the next-order balanced model, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-15749, https://doi.org/10.5194/egusphere-egu26-15749, 2026.
Harmful algal blooms (HABs) pose a persistent challenge to coastal ecosystems, fisheries, and public health, particularly in urbanized coastal regions subject to strong hydrodynamic forcing and meteorological variability. HAB dynamics emerge from the interaction of biologically driven growth processes and physically governed transport and dispersion, operating across disparate spatial and temporal scales. However, most existing data-driven forecasting approaches treat these processes implicitly and holistically, limiting physical interpretability, robustness under nonstationary forcing, and the ability to represent forecast uncertainty.
This study proposes a physics-informed diffusion-based framework for HAB forecasting in coastal environments, with the objective of explicitly separating biological and physical drivers within a generative probabilistic model. The central hypothesis is that decoupling meteorological and hydrodynamic forces can improve the physical consistency and generalizability of HAB forecasts while enabling uncertainty-aware prediction. To this end, future HAB states are formulated as conditional samples generated through a reverse diffusion process guided by physically meaningful environmental inputs.
The proposed framework adopts a dual-forcing architecture. A meteorological branch encodes atmospheric variables—including air temperature, precipitation, wind speed, and radiative forcing—that primarily regulate phytoplankton growth potential and bloom initiation. In parallel, a hydrodynamic branch incorporates tidal dynamics and wave-related information to represent advection, mixing, and dispersion processes governing the spatial evolution of algal biomass in coastal waters. Physical consistency is promoted by embedding the advection–diffusion equation as a soft constraint within the hydrodynamic latent space, encouraging mass-conserving and physically plausible transport behavior without imposing a fully deterministic dynamical model.
By leveraging diffusion probabilistic modeling, the framework generates ensemble-based forecasts that characterize the conditional probability distribution of future HAB states rather than single deterministic trajectories. Forecast outputs are formulated in terms of a probabilistic HAB severity index, facilitating interpretable, risk-informed early warning analogous to probabilistic weather forecasting systems. Model development is designed to integrate multi-source environmental datasets, including high-frequency meteorological observations, wave and tidal records, and routine coastal water-quality monitoring.
The framework is developed with a focus on tidally energetic coastal systems, with the Hong Kong coastal region serving as a representative application domain. Overall, this study outlines a physically interpretable and uncertainty-aware modeling paradigm for HAB forecasting and provides a conceptual foundation for next-generation early-warning systems in coastal environments.
How to cite: Liu, Z.: Physics-Informed Diffusion Model for HAB Forecasting in Hong Kong Coastal Waters, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9163, https://doi.org/10.5194/egusphere-egu26-9163, 2026.
The ocean floor is littered with seamounts, most of which are volcanic in origin. Seamounts are important in the marine geosciences because they are oceanographic ‘dip-sticks’, biodiversity hotspots, scatterers of tsunami waves, and hazards for navigation. Research ships with single beam echo-sounders have discovered many small seamounts and some large ones while satellite altimetry has led to discovery of many large seamounts and some small ones. The exact number of seamounts in the world’s ocean basins and their margins remains, however, unknown. Here we use machine learning in an attempt to locate all seamounts, to estimate their height and volume and to speculate on their origin. We use the seamounts found by Hillier & Watts (2007) along ship track from single beam echo-sounder data acquired on 5585 individual research cruises during 1950 to 2002 as a ‘training’ data set and the SRTM15+V2.7 (GEBCO 2025) topographic grid that combines shipboard single beam and multibeam (swath) bathymetry data acquired on 2154 individual research cruises during 1980 to 2024 with predicted bathymetry from satellite altimeter data in regions of sparse ship tracks to determine the 6 main attributes (channels) of seamounts, 4 of which refer to their slopes. We then use the SRTM15+V2.7 (GEBCO 2025) topographic grid together with machine learning to update the global seamount census of Hillier & Watts (2007). Preliminary results in two pilot study areas on old and young oceanic crust in the Pacific Ocean indicate that machine learning yields up to a factor of 2 more seamounts than were identified in the training data set. The implications of these results are examined for volcanism on Earth and on other terrestrial planets.
References:
Hillier, J.K., Watts, A.B., 2007. Global distribution of seamounts from ship-track bathymetry data. Geophys. Res. Letts. 34, 1-5, doi:10.1029/2007GL029874.
Tozer, B., Sandwell, D.T., Smith, W.H.F., Olson, C., Beale, J.R., and Wessel, P., 2019 Global Bathymetry and Topography at 15 Arc Sec: SRTM15+. Earth and Space Science 6, doi:10.1029/2019EA000658
How to cite: Wang, Z. and Watts, A. B.: Global distribution of seamounts from machine learning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2457, https://doi.org/10.5194/egusphere-egu26-2457, 2026.
Global ocean circulation has a significant impact on climate variability, where ~80% of the ocean energy transfer occurs in small-scale processes. While the existing record of altimetry goes back thirty years and has enabled the assimilation of gridded sea surface height maps, their operational resolution of 0.25° is not high enough to study these mesoscale eddies, and we are therefore in need of methods that can improve their resolution globally.
The recently launched SWOT satellite with ~2km resolution now offers the first data record with sufficient resolution to reveal these processes in observations, and offers the possibility of drastically improving sea surface state maps. However, its sparse temporal and spatial record brings challenges for global assimilation.
We propose a generative machine learning approach to downscale existing gridded Level 4 sea surface height to the fine resolution of SWOT. Our methodology involves two steps: first, training a conditional diffusion downscaling model on high resolution simulated data as a prior joint distribution over sea state observations, including height, temperature and salinity. Secondly, a data assimilation scheme via a Bayesian posterior formulation that generates high resolution sea surface state maps assimilated with a set of observations. We evaluate our methodology both in simulated and observing system experiments that demonstrate the efficacy of our approach as well as their scalability to global context in evaluations of major currents. Under the Bayesian formulation we also find that the diffusion model produces well calibrated predictive uncertainty estimates, which further underlines the applicability of diffusion models as a computationally efficient method in this domain. Our high resolution sea surface height maps open up new insights into mesoscale eddies.
How to cite: Lehmann, N., Shah, A., Bamber, J., and Zhu, X.: OceanBottle: Sea Surface State Data Assimilation and Downscaling, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11308, https://doi.org/10.5194/egusphere-egu26-11308, 2026.
High-dimensional ocean datasets, e.g. of global sea surface temperature, provide crucial insight to the dynamic of physical ocean characteristics such as seasonal cycle, ENSO, and global trend, but the dimensionality often results in computational complexity. Deep learning methods, such as variational autoencoders (VAEs), offer dimension reduction techniques that retain nonlinearities while expressing the system state in a meaningful lower-dimensional latent space. We explore whether encoded spatially limited observations, such as from satellites, buoys, or ship tracks, could be assimilated in the latent space. First, we developed a VAE to create a low-dimensional representation of a global sea surface temperature anomalies dataset. Next, we built a sample environment to demonstrate data assimilation within the latent space by creating spatially incomplete observations from the global dataset by selecting specific regions and adding noise. Accordingly, we developed an observational encoder to map these observations into the latent space of the VAE. For the latent data assimilation, we created a Bayesian update (e.g. Kalman filter) and decoded assimilated observations to evaluate results. We report on the assimilation of encoded limited observations within the latent space and discuss possible applications and future development of this approach.
How to cite: Carsey, S., Hornschild, A., and Saynisch-Wagner, J.: Development of an Observational Encoder for Data Assimilation in the Latent Space of a Variational Autoencoder (with Sea Surface Temperature) , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6860, https://doi.org/10.5194/egusphere-egu26-6860, 2026.
We present the development of a next-generation family of Lat–Lon–Cap (LLC) global ocean simulations, culminating in LLC8640, a 1/96◦ (≈ 1 km) realistic global 'nature run’ that, once complete, will represent the highest-resolution global ocean model produced under realistic conditions. This effort advances well beyond the widely used LLC4320 configuration by addressing long-standing dynamical biases through coordinated improvements in resolution, physical formulation, and forcing.
Key advances include increased vertical and horizontal resolution, updated global bathymetry, non-linear free surface, explicit ice-shelf cavities around Greenland and Antarctica, hourly atmospheric forcing, realistic river discharge, and improved astronomical tidal forcing. Together, these developments directly target deficiencies in earlier LLC models, including a misplaced Gulf Stream, a crude representation of Antarctic shelf circulation, and weak tropical instability waves. Particular emphasis is placed on the equatorial ocean, where Green’s-function-based approaches are used to optimize turbulence parameterizations and reduce persistent discrepancies between global models and observations. Early results from the ongoing lower-resolution spin-up already demonstrate markedly improved realism, including a more accurate Gulf Stream path and a strengthened, more realistic equatorial undercurrent.
The modeling strategy employs a staged spin-up across resolutions: a multi-year 1/12◦ (LLC1080 ) integration to equilibrate large-scale circulation and kinetic energy; a subsequent 1/48◦ (LLC4320 ) phase to sharpen mesoscale and submesoscale dynamics; and a final month-long 1/96◦ (LLC8640 ) integration producing several petabytes of hourly three-dimensional velocity, temperature, and salinity fields. The resulting dataset will provide an unprecedented global benchmark for studies of internal tides and waves, submesoscale turbulence and mixing parameterizations, and SWOT-era sea-surface height variability.
How to cite: Momeni, K., Menemenlis, D., Zhang, K. Q., and Peltier, W. R.: A Trailblazing Global Ocean Simulation in the Time of Wide Swath Altimetry, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2382, https://doi.org/10.5194/egusphere-egu26-2382, 2026.
The global ocean carbon sink is a critical component of the Earth’s climate but current models are limited in their predictive capability because of the high computational cost that a biogeochemical model, required to simulate air-sea CO₂ fluxes, entails. In this study, we investigate the ability of deep learning (DL) methods to reconstruct monthly air-sea CO₂ fluxes using only the physical output of an ocean circulation model, thereby exploring a data-driven alternative to a costly biogeochemical model. We used a collection of global simulations from the ocean biogeochemistry model NEMO-MOPS at a horizontal resolution of 0.25°, which differ in their atmospheric forcing components. The simulations span 61 years (1958-2018), providing a long, high-resolution dataset that captures substantial interannual to decadal variability. Our objectives are threefold: (1) to assess how accurately DL models can reconstruct CO₂ fluxes from physical variables alone, (2) to evaluate the generalization of these models across unseen years and forcing regimes, and (3) to identify the relative importance of physical drivers and their temporal lags in predicting air-sea CO₂ exchange. To this end, we train a point-wise Long Short-Term Memory (LSTM) network augmented with a temporal attention mechanism, which enables dynamically weight information from different time steps, to predict present-month CO₂ fluxes. To this end, we use eight physical predictors from the current month and the preceding five months. Standard regression metrics indicate an overall accurate reconstruction even though extreme CO₂ outgassing events are often underestimated. Seasonal and interannual variations are mostly well reconstructed across different ocean regimes. Spatial patterns are also well reconstructed, even though the DL model is trained only with local features (not including latitude and longitude information). This is a promising result in terms of generalizing to other physical settings, which we aim to test in future experiments. We finally interpret the learned relationships, by computing the Shapley values to quantify the contribution of each physical driver across time lags. Overall, our work highlights the potential of combining DL based techniques and explainable AI as a scalable and transparent complement to traditional Earth system modeling for studying ocean carbon cycle dynamics.
How to cite: Mohanty, S., Patara, L., Rath, W., Kazempour, D., and Kröger, P.: Reconstructing surface ocean carbon flux from physical parameters using deep learning methods, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-22647, https://doi.org/10.5194/egusphere-egu26-22647, 2026.
The Marmara Sea, covering approximately 11,350 km² in northwestern Turkey, links the Black Sea and the Aegean Sea via the Bosporus and Dardanelles straits. It is bordered by densely populated and industrialized cities such as Istanbul. The Marmara Sea is facing eutrophication and mucilage outbreaks, necessitating the monitoring of key indicators, including chlorophyll-a, which serves as an indicator of phytoplankton abundance. Atmospheric dust deposition can play a significant role in providing nutrients such as nitrogen, phosphorus, silica, and iron to the surface ocean, thereby affecting phytoplankton growth. Excessive phytoplankton growth and the accumulation of organic matter trigger mucilage formation under suitable conditions. The region is influenced by dust transported from regional and distant sources, such as the Sahara Desert.
In this study, spatio-temporal dynamics of chlorophyll-a (Chl-a), Aerosol Optical Depth (AOD), Sea Surface Temperature (SST), Particulate Organic Carbon (POC), Photosynthetically Active Radiation (PAR), and precipitation were investigated on a monthly scale using MODIS-derived products from 2005 to 2020. Time series analysis and machine learning models such as HGB (Histogram Gradient Boosting), Random Forest, and Multiple Linear Regression were performed for exploring temporal patterns, relationships, and modeling Chl-a, respectively. Chl-a showed a moderate negative correlation with SST (r = –0.52) and a strong positive correlation with POC (r = 0.80), while its relationship with AOD was negligible. It should be noted that during desert dust episodes, a significant lagged correlation was observed between Chl-a and AOD. The observed Chl-a values ranged between 0.6 and 19.50 mg/m³ over the study period, with the highest values observed in April and the lowest values occurring between June and November. Modeling Chl-a based on satellite-derived environmental variables showed that the Histogram Gradient Boosting algorithm achieved the highest performance, yielding r = 0.807, R² = 0.645, RMSE = 1.870, MAE = 1.218, and MBE = 0.062. These results highlighted the strong influence of SST and POC on Chl-a variability, while AOD appears to have minimal direct impact. Further investigation of the impact of the high dust deposition periods during dust storm events is suggested for the Marmara Sea.
How to cite: Demir, B., Aydin, Y., and Olgun, N.: Machine-Learning Assessment of Chlorophyll-a Responses to Atmospheric Dust and Environmental Factors Using Remote Sensing Data in the Marmara Sea, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1199, https://doi.org/10.5194/egusphere-egu26-1199, 2026.
Artificial Neural Networks (ANN) are applied to estimate the interannual variability of monthly-mean satellite-derived chlorophyll-a (CHL) at a global scale in the 1997-2025 period, as function of different physical variables and climate teleconnection indices. Among other variables, satellite-derived sea-surface height (SSH) proved to be a good single predictor for the CHL, showing significant CHL-SSH correlation in most of the world ocean between 60°S and 60°N (where the most continuous data series are available). This correlation, generally low for a linear estimation, opens the possibility to CHL reconstruction using higher-performance non-linear techniques like ANN. The ANN-model successfully reproduces the CHL interannual variability: 59% of the modeled CHL present correlations > 0.90. Then, the ANN-model can be used to predict CHL beyond the training period, showing a good predictability at least one season ahead. On the other hand, a similar exercise for the reconstruction/predictability of CHL is subsequently carried out using selected teleconnection indices as predictors, presenting an alternative simpler method to estimate the CHL variability in key regions along the world ocean. Thus, the proposed methods open the possibility to predict not only CHL but other related biogeochemical variables.
How to cite: Rivas, D.: Estimation of global satellite-derived chlorophyll-a as function of physical drivers using shallow neural networks, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-15274, https://doi.org/10.5194/egusphere-egu26-15274, 2026.
Exchange of water masses across the Greenland-Scotland ridge is an important part of the AMOC. The complicated bathymetry of the ridge is not properly resolved in standard CMIP models with around 1 degree resolution and this reduces confidence in simulated exchanges and their variability. Previously, a nested-domain approach with finer resolution in selected areas has been applied. Here, we perform a pilot study of the alternative approach of a fine-resolution ocean emulator. We use daily 3D salinity and temperature fields of the GLORYS reanalysis in original (target) and 4x reduced (input) resolution and demonstrate that a fine-resolution emulator consisting of a simple U-net architecture trained on 10 years of input/target can be used to reconstruct the target field from the input fields outside the training period with a significant skill compared to simple interpolation. We apply temporal and spatial scrambling to assess input feature importance. Our study suggests that the fine resolution model in a nested setup can be replaced with an ocean emulator leading to substantial gains in overall execution speed.
How to cite: Schmith, T., Beauchamp, M., Devilliers, M., Gierisch, A., and Olsen, S.: Fine-resolution ocean emulator for the Greenland-Scotland Ridge, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11454, https://doi.org/10.5194/egusphere-egu26-11454, 2026.
Growing evidence suggests that the present-day Atlantic Meridional Overturning Circulation (AMOC) operates in a bistable regime and may transition to a weakened or collapsed (“OFF”) state under climate change forcing, with severe global climate impacts. In addition to deterministic forcing, stochastic variability can induce noise-induced transitions between stable AMOC states. Quantifying the probability and pathways of such transitions is therefore critical.
Previous work (Soons et al., 2024) applied Large Deviation Theory (LDT) to a stochastic box ocean model (Wood et al., 2019) to estimate the most probable pathways for noise-induced AMOC collapse and recovery. While effective, this approach requires explicit knowledge of system properties, such as the Jacobian, limiting its applicability to higher-dimensional, more complex climate models.
Here, we adapt a recently proposed deep reinforcement learning framework (Lin et al., 2025) to compute most probable transition pathways in stochastic dynamical systems without prior knowledge of the governing equations. Applied to the stochastic box ocean model, the method robustly identifies physically consistent collapse and recovery pathways, comparable to those obtained using LDT. Finally, we demonstrate the feasibility of this framework in a more complex ocean model.
How to cite: Guardamagna, F. and Dijkstra, H.: Estimating Most Probable AMOC Collapse and Recovery Pathways Using Deep Reinforcement Learning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-7643, https://doi.org/10.5194/egusphere-egu26-7643, 2026.
Anthropogenic climate change induces sea level changes (SLC) that must be accurately estimated to improve understanding of both past and future changes, facilitate timely adaptation and mitigate coastal risk. The rate and acceleration of global mean sea level and the associated uncertainty has been thoroughly assessed for the period since 1900. For the period since 1993, regional assessments have been produced, leveraging tide gauge records and satellite altimetry, allowing nations to understand and adapt more appropriately to local sea-level changes. However, improved regional timeseries are needed to robustly detect potential accelerations in local SLC.
This study proposes a novel data-driven approach for reconstructing regional SLC from tide gauges. We use a Reduced-Space Ensemble Kalman Smoother associated with the statistical Analog Prediction. This method, named RedAnDA, has been previously applied to reconstruct past temperature and salinity fields in the tropical Pacific, with good results. In this work, RedAnDA derives empirical orthogonal functions from satellite altimetry to extrapolate spatial features of the variability, as well as Analogs to predict monthly SLC associated with interannual-to-decadal variability. The uncertainty is quantified from the spread within the ensemble and takes various components into consideration, such as non-linearity in the dynamics or sampling issues. Tide gauge and altimetry input datasets are pre-processed (for instance for vertical land motion) using state-of-the-art methods.
The RedAnDA performance is assessed by comparing the reconstruction to altimetry and existing tide gauge reconstructions, to evaluate our results over the recent period. In comparison to other reconstruction methods, RedAnDA can assess monthly changes associated with interannual variability over the 20th century, relying only on observational-based information. We further test the method by doing reconstructions which only assimilate 50-75% of the tide gauges, using the remaining ones for validation. These different tests show that RedAnDA can provide important additional regional information on SLC in the 20th century, including new estimates of the acceleration in regional SLC.
How to cite: Oulhen, E., Slangen, A. B. A., and Palmer, M. D.: Reconstructing regional 20th century sea level changes from tide gauges using the Analog Method, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20090, https://doi.org/10.5194/egusphere-egu26-20090, 2026.
Decadal climate predictions are essential for climate adaptation, yet remain challenging due to the complex interplay of initial conditions and external forcings. A key factor in achieving skillful forecasts is the upper ocean, which plays a central role in modulating decadal-scale climate variability, including phenomena such as ENSO, the Atlantic Multidecadal Variability, and the Indian Ocean Dipole. Accurately capturing the ocean’s memory is therefore critical, but traditional numerical models are computationally demanding and often exhibit systematic biases. While machine learning has shown promise in improving medium-range weather forecasts, its application to decadal climate prediction remains limited.
This work explores the feasibility of using machine learning to predict sea surface temperature (SST) and ocean heat content (OHC) on decadal timescales. We develop an autoregressive model based on a UNet-like convolutional neural network, trained on 1,000 years of data from a pre-industrial control run from the fully coupled MPI-ESM. This simulation provides a controlled environment to study predictability arising from internal ocean dynamics. Inputs include SST, OHC, a land-sea mask, and top-of-atmosphere solar radiation to encode the seasonal cycle. We conduct a systematic study of input design, to assess how the representation of past states influences model stability and predictive skill. Our results suggest that machine learning can be a viable and flexible approach for decadal ocean prediction. Additionally, we find that longer input windows and coarser resolution may improve long-term stability, potentially offering new insights into how climate memory is encoded.
How to cite: Meyer, F., Kadow, C., and Baehr, J.: Machine Learning for Decadal Ocean Prediction - Exploring the Feasibility of Capturing Climate Memory in the Upper Ocean, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-12894, https://doi.org/10.5194/egusphere-egu26-12894, 2026.
The Kuroshio Current is a western boundary current in the northwestern Pacific, and its transport and path variability significantly affect air-sea interactions, thus modulating North Pacific climate, as well as ecosystems. The Japan Meteorological Agency (JMA) 137°E repeat hydrographic section, occupied every winter from 1967 to 2023 (57 years), provides a long and consistent benchmark for diagnosing the variability of the Kuroshio Current system. Here, we analyze these repeated occupations and derive the vertical structure of zonal geostrophic velocity and associated transport. Our analysis reveals that the Kuroshio Current system exhibits substantial variability and intrinsic asymmetry in its transport, axis position, and vertical hydrographic structure. To capture the asymmetric hydrographic patterns associated with these transport fluctuations, we extract leading variability in the vertical structure using empirical orthogonal functions and apply a 1×5 self-organizing map (SOM) to classify distinct circulation patterns. The SOM yields five physically interpretable nodes: two large-meander (LM) nodes (moderate and extreme) and three distinct non-LM nodes. The extreme LM node features a southward shift of the Kuroshio axis to around 30°N accompanied by a significant weakening of the recirculation gyre. Moderate LM events exhibit a less pronounced southward shift near 31°N. The non-LM nodes can be characterized by (i) strengthened recirculation with near-normal net transport, (ii) enhanced net eastward transport, and (iii) reduced net transport. The heaving of isopycnal lines mostly accounts for thermohaline anomalies throughout the nodes, whereas spicing plays a partial role only in the extreme LM node. This study argues that variation in the thickness of the Subtropical Mode Water (STMW) accounts for upper ocean heat content and consequently for volume transport, underpinning STMW thickness as a metric integrating variability across the Kuroshio Current system along the 137°E section.
How to cite: Park, H.-J., Kim, Y. S., and Na, H.: Nodes of the Kuroshio Current system from multidecadal repeat observations along 137°E, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16459, https://doi.org/10.5194/egusphere-egu26-16459, 2026.
Real-time, high-fidelity ocean state estimation is a prerequisite for Earth system digital twins, yet faces a dilemma between the computational bottlenecks of traditional assimilation and the grid-based fidelity losses of deep learning. Here we present ADAF-Ocean, a geometry-agnostic framework that resolves this by assimilating multi-source observations directly at their original resolutions. Leveraging a neural process-based architecture, our approach learns a continuous mapping from heterogeneous inputs, such as sparse profiles and satellite imagery, thereby maximizing information extraction while enforcing multivariate physical consistency. Although purely data-driven, ADAF-Ocean is capable of implicitly learning the coupling patterns between thermodynamic and kinematic variables directly from high-fidelity datasets. Evaluations show that superior analysis accuracy gives rise to emergent physical coherence. Serving as superior initial conditions for a DL forecast model, these coherent fields sustain a significant forecast skill advantage for up to 20 days. Furthermore, by quantifying the contribution of individual observational sources, this framework establishes a trustworthy pathway for AI-driven oceanography, bridging data-driven efficiency with the rigorous standards of Earth system monitoring.
How to cite: Xiang, Y.: Advancing Ocean State Estimation with efficient and scalable AI, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-15437, https://doi.org/10.5194/egusphere-egu26-15437, 2026.
Posters virtual: Mon, 4 May, 14:00–18:00 | vPoster spot A
EGU26-707 | ECS | Posters virtual | VPS31
Data-Driven Modelling and Assimilation of the Sub-Seasonal Evolution of Sea Surface TemperatureMon, 04 May, 14:06–14:09 (CEST) vPoster spot A
Sea Surface Temperature (SST), due to its influence on air-sea interactions, is a critical input into weather models. While physics-based ocean models are continually improving to better represent SST in weather models, data-driven methods offer a promising alternative. In this work, we present an implementation of nonlinear operator inference on a satellite-based SST field (10 km spatial resolution, 1 day temporal resolution) in the northern Indian Ocean, which is known to significantly impact the Indian monsoon. For the prediction of SST, a reduced-order model with a polynomial structure is built non-intrusively from satellite data over a 30-day training period, showing the first five proper orthogonal decomposition modes to capture the SST evolution. A moving-window assimilation scheme utilises the reduced-order model adjoint to correct the prior state, enforcing the model equations over the assimilation window with state observations. Results show that this framework corrects drift, extending the prediction horizon from one week to twenty days. We demonstrate the efficacy of the discovered models using error metrics and their ability to accurately capture lateral SST gradients. Importantly, the inferred operators from the reduced-order model enable the derivation of an explicit adjoint directly from the data, overcoming the computational constraints of General Circulation Models that prohibit rapid adjoint-based assimilation. The performance of the reduced-order model over multiple seasons will also be presented, including the effects of training with data from several years.
How to cite: Madduri, S. H. Y. K., Mathur, M., and Kalur, A.: Data-Driven Modelling and Assimilation of the Sub-Seasonal Evolution of Sea Surface Temperature, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-707, https://doi.org/10.5194/egusphere-egu26-707, 2026.
