Phytoplankton play a key role in maintaining marine ecosystems and regulating global carbon dioxide concentrations through photosynthesis. Thus, it is crucial to assess and understand their temporal variations. However, fluctuations of phytoplankton biomass on multi-decadal and longer timescales remain uncertain, in contrast to seasonal and interannual ones, due to the lack of long-term observations on a global scale and the uncertainties related to the complex balance of processes that control their fate. As phytoplankton growth depends on the availability of nutrients in the sunlit upper ocean, which is closely linked to the stratification of the ocean, one can assume that at first order changes in phytoplankton is related to changes in ocean and atmosphere dynamics.
Over the last few years, several conventional data-driven deterministic approaches have been trained from physical observations (used as predictors) to reconstruct satellite ocean color time series (i.e., Chlorophyll-a concentration, Chl, which is used as a proxy of the phytoplankton biomass) and investigate their multi-decadal variability. Deterministic methods, such as encoder-decoder architecture U-Net, LSTM, FourCastNet, are robust but tend to fail in capturing probabilistic uncertainty because they produce deterministic outcomes. Additionally, these methods struggle with handling extreme and highly complex real-world scenarios. This study proposes a novel application of score-based generative diffusion models to address these challenges and present a comparative analysis against U-Net and FourCastNet. Probabilistic conditional diffusion model has been pretrained on simulation data and subsequently fine-tuned to learn the parameters using satellite observation data. This generative model learns the inherited uncertainty by generating ensembles of possible Chl mapping and analyzing the variability within the ensemble. The model can then be sampled efficiently to produce realistic Chl ensembles, conditioned on physical predictors and the baseline model U-Net. The ensembles from the diffusion model show greater reliability and accuracy, particularly in extreme event classification.
Our results demonstrate that when conditioned with a U-Net (meaning this input together with eight physical predictors), diffusion behaves better than the baseline method, especially when the number of samples is increased. It is visible from the spatial maps of standard deviation that as the sample size increases, the model's predictions stabilize and become more concentrated around the mean which leads to a reduction in the spread of outcomes.
How to cite: Lakra, M., Fablet, R., Drumetz, L., Martinez, E., Pauthenet, E., and Nga Nguyen, T. T.: Probabilistic Diffusion Models for Ocean Chlorophyll-a Prediction, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-18806, https://doi.org/10.5194/egusphere-egu25-18806, 2025.
Understanding the seasonal dynamics of plankton in the Atlantic Ocean is the first step towards the proper assessment of marine ecosystem health and productivity. Ocean colour and surface chlorophyll (chl-a) distribution serve as proxies for phytoplankton biomass, providing insights into marine food web dynamics and biogeochemical cycles. This study examines the response of the total chlorophyll concentration to physical drivers observable by remote sensing in the Atlantic Ocean using a combination of multivariate Principal Component Analysis (PCA) and deep learning models. The results show that the Sea Surface Salinity (SSS), Absolute Dynamic Topography (ADT), and Sea Surface Temperature (SST) are found to be the predominant drivers of physical variability across the ocean, with distinct spatial patterns. The clustering of the principal components identifies regions characterised by distinct physical processes. Based on these clusters, we devised a Transformer Encoder model to predict chl-a concentrations in three distinct regions. The model outperformed climatological baselines, especially in the temperate and tropical regions, though accuracy varied seasonally, with higher accuracy in winter months and increased complexity in summer due to more dynamic oceanographic conditions. A SHAP-based sensitivity analysis showed that ADT and SSS dominate chl-a variability, particularly during summer months, while SST and wind stress also contribute significantly during transitional periods. The study highlights the necessity to account for both seasonal and regional differences in predictive modelling, and it underscores the importance of continuing to develop spatio-temporal models to improve forecasting accuracy for marine ecosystem management and conservation.
How to cite: Chauhan, A., Smith, P., Rodrigues, F., Christensen, A., Nardelli, B. B., John, M. St., and Mariani, P.: Deep Learning Models to Identify Seasonal Drivers of Chlorophyll Changes in the Atlantic Ocean, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-19668, https://doi.org/10.5194/egusphere-egu25-19668, 2025.
Satellite observations provide a global or near-global coverage of the World Ocean. They are however affected by clouds (among others), which severely reduce their spatial coverage. Different methods have been proposed in the literature to reconstruct missing data in satellite observations. For many applications of satellite observations, it has been increasingly important to accurately reflect the underlying uncertainty of the reconstructed observations. In this study, we investigate the use of a denoising diffusion model to reconstruct missing observations. Such methods can naturally provide an ensemble of reconstructions where each member is spatially coherent with the scales of variability and with the available data. Rather than providing a single reconstruction, an ensemble of possible reconstructions can be computed, and the ensemble spread reflects the underlying uncertainty. We show how this method can be trained from a collection of satellite data without requiring a prior interpolation of missing data and without resorting to data from a numerical model. The reconstruction method is tested with chlorophyll a concentration from the Ocean and Land Colour Instrument (OLCI) sensor (aboard the satellites Sentinel-3A and Sentinel-3B) on a small area of the Black Sea and compared with the neural network DINCAE (Data-INterpolating Convolutional Auto-Encoder). The quality of the reconstruction is assessed using independent test data.
The spatial scales of the reconstructed data are assessed via a variogram, and the accuracy and statistical validity of the reconstructed ensemble are quantified using the continuous ranked probability score and its decomposition into reliability, resolution, and uncertainty.
The diffusion method compared favorably against the U-Net DINCAE. The RMSE of the reconstructed data using the denoising diffusion model was smaller than the corresponding reconstruction of DINCAE. The main advantage of the diffusion model is, however, the ability to reproduce an ensemble of possible reconstructed conditions on the available data. Each of these reconstructions contains small-scale information comparable to the scales of variability in the original data, avoiding a common problem where the results of U-Net and autoencoders produce images that are too smooth, as the information on small scales can typically not be recovered under clouds with a certain extent. The overall conclusion is robust when applying this technique to other areas of the Black Sea.
The ensembles of reconstructed data generated by the diffusion model can be used, for example, in the detection of gradients and fronts in the satellite images or in the estimation of the error in derived quantities, where information on how the error is correlated in space is also needed.
How to cite: Barth, A., Brajard, J., Alvera-Azcárate, A., Mohamed, B., Troupin, C., and Beckers, J.-M.: Reconstruction of missing satellite data using a Probabilistic Denoising Diffusion Model applied to chlorophyll a concentration, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-11139, https://doi.org/10.5194/egusphere-egu25-11139, 2025.
Biogeochemical models are computational approximations of systems of differential equations used to represent and predict biogeochemical constituents of the ocean. These might include carbon and nutrients cycles, and its interactions with biological components, such as different types of plankton. Unfortunately, these models tend to be constrained by their complex parametrization and computational cost, limiting their practical application and scalability.
Autoencoders are neural networks that are trained to learn a compressed representation of a dataset, typically with the goal of reconstructing the input to its original or a specified target dimension. But the bottleneck of this compression, or the latent space of the autoencoder, can offer interesting insights into the dominant features of the system.
Here we train different types of autoencoders to capture the main spatiotemporal dynamics from data modeled by the biogeochemical analysis BIO4 (based on NEMO-PISCES). This not only provides a basis for the development of computationally efficient emulators, but it can help us detect patterns and relationships that might not be immediately apparent in the high-dimensional output of the model. This offers interesting insights into how the model actually captures its constituting components.
Such compressed representations can also be used for parameter sensitivity analysis, to develop data assimilation frameworks, and as tools for uncertainty quantification and outlier detection.
How to cite: Martinez Balbontin, G. and Ciavatta, S.: Peeking Into a Marine Biogeochemical Model with an Autoencoder , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-11622, https://doi.org/10.5194/egusphere-egu25-11622, 2025.
Over the past decades, global ocean oxygen inventories have declined by 0.5–3.3% relative to historical averages, with significant uncertainties in data-sparse regions such as the South Pacific and Indian Oceans. These gaps hinder accurate estimates of deoxygenation rates, potentially leading to underestimation of its magnitude. In this context, gridded oxygen products are essential for assessing global and regional trends and projecting the impacts of deoxygenation on marine ecosystems. However, traditional Optimal Interpolation (OI) methods are known to underestimate ocean oxygen loss, particularly in poorly observed areas.
To address these limitations, we propose a novel approach to build a gridded oxygen concentration product. Specifically, we develop a neural network emulator of oxygen concentration based on temperature and salinity measurements. This neural network is then used to generate emulated oxygen concentration data, which are combined with dissolved oxygen measurements to produce a new global gridded oxygen concentration product spanning 1965 to 2022. We evaluate our product against climatological estimates from the World Ocean Atlas and other gridded oxygen products. Future work will leverage this gridded product to study the regional evolution of ocean deoxygenation, particularly in Oxygen Minimum Zone (OMZ) regions.
How to cite: Lachkar, Z. and Ouala, S.: A Novel Global Gridded Ocean Oxygen Product Derived from Neural Network Emulators (1965–2022), EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-15255, https://doi.org/10.5194/egusphere-egu25-15255, 2025.
Climate change risks are often assessed using climate impact indicators (CIIs) determined for various socio-economic scenarios. Ideally, for every scenario an impact model, e.g. an ecological model or a hydrological model, processes outputs of a climate model to produce CIIs. However sometimes, even if outputs of a climate model are available for all scenarios, computation costs of the impact model can limit the number of scenarios with available CIIs.
To fill this gap, we propose to infer CIIs for unseen scenarios, i.e. scenarios not processed by the impact model, with an interpretable equation. This equation is discovered using symbolic regression on a scenario processed by the impact model. Specifically, we discover an equation that predicts CIIs based on climate impact drivers (CIDs), where CIDs are variables of the climate model averaged monthly and spatially.
In our application, the impact model is a biogeochemical model of the Mediterranean Sea driven by the same regional climate model for two scenarios: RCP4.5 and RCP8.5. Our CII is the annual mean Net Primary Production (NPP) summed over an offshore area in the Gulf of Lion (located in the North-western Mediterranean basin), where NPP is the total rate of organic carbon production by photosynthesis of marine phytoplankton minus their respiration.
Preliminary results show that the discovered equation reproduces well the trend and the interannual variability of NPP for the testing scenario RCP4.5, unseen during the training. Indeed, the scenario RCP8.5 is preferred for training as it spans a wider range of climatological contexts. If our preliminary results are confirmed, we could extend our approach to a large ensemble of climate models, in order to characterize the uncertainty of CIIs.
How to cite: Le Roux, E., Tandeo, P., Granero Belinchon, C., Baklouti, M., Le Sommer, J., Sevault, F., Somot, S., Doury, A., and Al Najar, M.: Equation discovery for climate impact: symbolic regression to emulate climate impact indicators for unseen scenarios, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-3836, https://doi.org/10.5194/egusphere-egu25-3836, 2025.
Phytoplankton is a central component of marine ecosystems. It contributes to biogeochemical cycles by absorbing carbon through photosynthesis at the ocean surface and transporting it deeper through sinking and subduction—hence contributing to the biological carbon pump. Plankton also represents the first link in marine food webs, supporting a wide range of marine life, from other plankters to the most productive fisheries on earth.
Satellites can help monitor phytoplankton over large-scales thanks to ocean color sensors. Current products provide daily, 4 km-resolution fields of chlorophyll-a concentration (Chla, the most widely used estimator of phytoplankton biomass) as well as its distribution in a few groups, hence estimating broad community composition. To produce these operational maps, the concentration of pigments measured by HPLC (High-Performance Liquid Chromatography) from in situ samples is regressed on reflectances at a few wavelengths matched to those samples in space and time. While incredibly useful, these models still display 30% error for Chla and at least as much when predicting community composition.
Numerous studies have shown the importance of considering mesoscale ocean structures, such as fronts and eddies, as they have a significant influence on the production and distribution of phytoplankton. These structures span tens to hundreds of kilometers and can be observed through ocean color but also infrared and radio wave satellite data.
In this work, we develop a deep learning model to predict the concentration of three phytoplankton size classes: pico-, nano-, and micro-phytoplankton. The in situ values are derived from over 7000 HPLC measurements spanning the globe, from 1997 to 2021. We use a Multi-Layer Perceptron to naturally combine reflectances with other satellite-derived variables that describe ocean physics (sea surface temperature, sea level anomalies, etc.) as input. The MLP is preceded by convolutional layers to summarise arrays of the input variables covering dozens of kilometers around the in situ observations. These two approaches are meant to capture the effect of mesoscale oceanic structures on the abundance and composition of phytoplankton.
This approach improves the estimation of phytoplankton communities on a global scale. It paves the way for in-depth studies on the influence of mesoscale structure in specific oceanic regions. Furthermore, it lays the groundwork for the future integration of the temporal dimension into the model, enabling a more comprehensive representation of ecological dynamics.
How to cite: Labourdette, E., Sauzède, R., Abbas-Turki, L., and Irisson, J. O.: Extending the inputs of deep learning models to capture the mesoscale context and better predict phytoplankton community composition, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-17211, https://doi.org/10.5194/egusphere-egu25-17211, 2025.
Mesoscale ocean eddies are dynamic structures controlling a significant proportion of water exchanges between the surface and the deep ocean, and therefore of heat, carbon and nutrient transfers. The eddy dynamics, i.e. changes in height, velocity and energy, are classically computed through complex ocean equations such as the quasi-geostrophic balance. However, those computations are time-consuming and slow down decision-making in operational situations. Some recent studies have managed to define eddy dynamics with simple properties - centroid position, amplitude, radius, current velocity, and horizontal displacement - and to predict their future evolution with machine learning models (Wang et al., 2020). We aim to implement a simple machine learning model to predict eddy properties that can reconstruct eddy dynamics and to include it in operational tools.
In this study, we simplified eddy structures, converting their 2D/3D gridded physical space into a parametric space, characterized by the eddy properties obtained with the AMEDA algorithm (Le Vu et al., 2017). Thus we considered eddies as 2D ellipse structures with additional properties - centroid position, amplitude, semi-axis of ellipse, rotation angle, maximal current velocity, and horizontal displacements. Explainable simple ML models were trained to learn the evolution of those parameters between two consecutive time steps. Here we selected two approaches of the least square regression model: the global linear regression on the whole training dataset and the local linear regression based on the nearest neighbors observations. Performances of each model are evaluated with the RMSE metric and compared to identify which model gives the most satisfactory results for eddy prediction.
Our analysis shows better performances with the local linear regression. However, the choice of more adapted models or a better selection of eddy properties would enhance the prediction of eddies. The next steps to the inclusion of the model in operational tools will be the consideration of eddy interactions - splitting and merging -, the uncertainty quantification and the data assimilation of eddy dynamics with an object-oriented approach.
References
Wang, X., Wang, H., Liu, D., Wang, W., 2020. The Prediction of Oceanic Mesoscale Eddy Properties and Propagation Trajectories Based on Machine Learning. Water 12, 2521. https://doi.org/10.3390/w12092521
Le Vu, B., Stegner, A., Arsouze, T., 2018. Angular Momentum Eddy Detection and Tracking Algorithm (AMEDA) and Its Application to Coastal Eddy Formation. Journal of Atmospheric and Oceanic Technology 35, 739–762. https://doi.org/10.1175/JTECH-D-17-0010.1
How to cite: Dealbera, S., Tandeo, P., Granero-Belinchon, C., Raynaud, S., and Boussidi, B.: Object-oriented mesoscale eddy prediction, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4043, https://doi.org/10.5194/egusphere-egu25-4043, 2025.
Submesoscale eddies are oceanic structures that occur on horizontal scales from 0.1-10 km, vertical scales from 0.01-1 km, and last from hours to several days. They are characterised by a Rossby number of Ro = ζ/f ~ O(1), where surface vertical vorticity ζ is similar to Coriolis frequency f. Submesoscale eddies are important in setting the stratification in the ocean surface mixed layer, mediating air-sea exchanges, and transporting energy between large and small scale motions. However, the study of submesoscale eddies on a global scale has been hindered by a shortage of global, long-term datasets. To fill this gap, we train and apply an unsupervised machine learning method adapted from the Profile Classification Model (PCM) to density profiles collected by Argo floats over global ocean from 2000-2021, producing the first global observational climatology of submesoscale activity. The adapted PCM identifies regions with high submesoscale activity using solely the density profiles and without any additional information on the velocity, location, or horizontal density gradients. The climatology shows that submesoscale activity peaks in spring in both hemispheres and lags behind the maxima of mixed layer depth by one month, suggesting that submesoscale eddies play important role in re-stratifying the mixed layer. Hotspots of submesoscale activity can be found in the Norwegian Sea and the Drake Passage in spring. This observational reconstruction of submesoscale activity enables the study of submesoscale distribution, seasonality, and inter-annual variation on a global scale.
How to cite: Yao, L. and Taylor, J. R.: Global Climatology of Submesoscale Activity Using Machine Learning, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-7671, https://doi.org/10.5194/egusphere-egu25-7671, 2025.
As an ocean and climate modeller, I propose to expose a few venues of ocean modelling where Machine Learning (ML) is expected to break through persistent challenges. My prime target is the numerical representation of the global ocean, with distinguishable coarse spatial scale (25 to 100 km) and long duration (at least 100 years). Observations are not sufficient (too sparse in space, particularly at depth, and too short in time, spanning only the last few decades) to be used directly as the sole ground truth. Hence it is compulsory to consider perfect model set-ups, besides training on observed database. Current challenges in ocean modelling that ML could contribute to solving, are the following : equilibration of simulations, quantification of sensitivity to parameters, parameterizations of unresolved processes (due to reduced spatial resolution and/or complexity) and quantification of structural uncertainties. I will introduce a few ML-based solutions to these challenges based on recent bibliography and my own activities. Overall, we need to build capacity in bridging the gaps between these centennial global ocean simulations, useful for climate applications, process models at regional scale, global ocean hindcasts (simulations with data assimilation), large eddy simulations and models of the past, present and future climate. To reach this goal, I advocate combining various ML architectures, factoring in uncertainties of every pieces of this hierarchy.
How to cite: Deshayes, J.: Ocean models for climate applications : progress expected from Machine Learning, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-13540, https://doi.org/10.5194/egusphere-egu25-13540, 2025.
The Atlantic Meridional Overturning Circulation (AMOC) is a critical component of the Earth system and one of the most prominent tipping element. In a warming climate, the AMOC is at risk of collapse due to increased freshwater input in the North Atlantic. Such an extreme event could lead to severe consequences for the global climate, with strong socio-economics impacts. Such a tipping event has been demonstrated to occur in conceptual, intermediate complexity and, recently, in the Community Earth System Model (CESM). Therefore, Reliable early warning signals are required for detecting whether the AMOC is approaching a tipping point. To estimate the distance of the AMOC to tipping, we propose a novel methodology, based on a Convolutional Neural Network (CNN) which uses sea surface salinity and temperature across the Atlantic as input. First, we validate our approach within the model of intermediate complexity Climber-X, demonstrating its ability to generalize to different forcing rates and in the presence of noise. We also explore the use of alternative climate variables such as the full-depth salinity profile at 35°S. Second, we assess the generalization capability of our methodology to a model of higher complexity. To this end, we use the CNN trained on Climber-X and successfully apply it to the AMOC collapse recently simulated in the CESM model. To demonstrate the physical consistency of the CNN model and increase its interpretability, we identify the most relevant regions to estimate the distance of the AMOC to tipping via the Layer-wise Relevance propagation technique.
How to cite: Guardamagna, F., Sinet, S., and Dijkstra, H.: Estimating the distance to the AMOC tipping point using convolutional neural networks, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4190, https://doi.org/10.5194/egusphere-egu25-4190, 2025.
The effective spatial and temporal scales resolved by Earth System Models (ESMs) remain a key limitation in reducing uncertainties in climate projections. While increasing model resolution is computationally prohibitive, machine learning (ML)-based parameterizations offer a promising alternative. However, these approaches often face generalization challenges in ‘out-of-sample’ scenarios, leading to numerical instabilities when integrated into ESMs. In this study, we aim to tackle these challenges by developing a data-driven discretization neural network for multidirectional advection in ocean models. The canonical 1D advection problem is revisited by using neural networks to predict the coefficients of a three-node stencil trained on high-resolution solutions projected onto coarser spatial and temporal grids. Conventional discretizations generalize to all scalar fields, while the data-driven approach is, by construction, tied to the training data. First, it is shown that we can normalize inputs with min-max scaling to achieve generalization, while training on coarsened high-resolution data across multiple grid configurations reduces sensitivity to time steps and mesh resolution. We find that coarsening based on triangular test functions, instead of averaging, enables unique mapping of the fine-scale variations of high-resolution solutions, leading to monotonicity of the neural network. Hybrid ML discretizations that predict advective fluxes are investigated, with a focus on enforcing desirable numerical properties—such as monotonicity, accuracy, and stability. Finally, we aim to test the numerical and generalization properties of the new data-driven discretization on 2D geostrophic flows. These results provide guidance for the development of better end-to-end data-driven parameterizations and discretizations in ESMs.
How to cite: Nasser, A.-A. and Adcroft, A.: Generalizing machine-learned discretization for climate simulations: addressing ‘out-of-sample’ challenges for 2D data-driven advection discretization, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-12958, https://doi.org/10.5194/egusphere-egu25-12958, 2025.
The spin-up of the ocean component is a critical step in coupled global climate simulations, allowing the model to achieve a physically consistent equilibrium by stabilising key variables such as temperature, salinity, and ocean currents. Without an adequate spin-up, residual drifts can undermine the accuracy and reliability of long-term climate projections. This study explores data-driven strategies to accelerate the spin-up, reducing computational costs while preserving the fidelity of simulated climate states. Using a low-resolution configuration of the EC-Earth4 Earth System Model (ESM), we tested few deterministic approaches to optimise the spin-up phase. A key method relies on iterative adjustments of the oceanic state by projecting multi-decadal trends in temperature and salinity. Empirical Orthogonal Function (EOF) analysis was employed to filter internal variability and generate new initial conditions that minimise numerical instabilities. Additionally, vertical stability was ensured to reduce energy imbalances and maintain physical consistency. Overall, our approach can significantly enhance the efficiency of spin-up processes in coupled climate models by at least a factor of two. These findings pave the way for the development of more sustainable and sophisticated strategies (e.g. exploiting machine learning and AI techniques) in climate modelling. Such advancements will be particularly helpful for high-resolution simulations, where achieving computational efficiency is critical.
How to cite: Sozza, A., Davini, P., and Corti, S.: Data-driven approaches for accelerating ocean spin-up in coupled climate simulations, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4573, https://doi.org/10.5194/egusphere-egu25-4573, 2025.
The combination of Machine Learning (ML) with geoscientific models has become an active area of research, but many technical challenges still remain because of the heterogeneous nature of programming languages, library environments and hardwares. Much efforts have been made over the recent years to propose different frameworks to perform online deployment of ML components within geoscientific models. One common drawback to all these solutions is the complexity of the required software environment. The latter often relies on versioned libraries and codes, both for the geoscientific and the ML models. Thus, ensuring the reproducibility of hybrid geoscientific model experiments is challenging, as it requires describing several tools and how to deploy them. This becomes even more problematic as the number of coupling solutions for hybrid modeling increases and may be unfamiliar to the members of the different modeling communities.
Here, we introduce Morays as an example of a community-based workflow for sharing reproducible hybrid ocean model experiments. Morays uses a GitHub organization to host hybrid experiments material that leverage the OASIS coupler (https://oasis.cerfacs.fr/en), which is widely used in European climate models. Our framework is based on a Python library (https://github.com/meom-group/eophis) that facilitates the use of OASIS for deploying hybrid modeling pipelines bridging FORTRAN solvers and ML models implemented in Python. The geoscientific model and ML scripts are executed separately and exchange data through the coupling API.
In this presentation, we will showcase several successful deployments of hybrid ocean model experiments with the NEMO ocean/sea-ice modeling framework. These experiments implement ML-based parameterizations and model correction schemes for improving different aspects of model solution (vertical physics, eddy parameterization, surface fluxes). All the experiments are shared openly in a dedicated GitHub organization (https://github.com/morays-community), as individual repositories following a standard template. We will present the material available to the community (tutorials, test cases), explain how to contribute, and discuss the broader perspective of reproducible workflow for future hybrid geoscientific models.
How to cite: Barge, A., Meunier, E., Contreras, M., Kamm, D., and Le Sommer, J.: Morays-community: a framework to share reproducible hybrid Machine Learning and Ocean modeling experiments., EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-18745, https://doi.org/10.5194/egusphere-egu25-18745, 2025.
Accurate ocean forecasting is essential for a range of critical applications, from maritime safety to climate adaptation strategies. Given the inherent uncertainties in ocean dynamics, the ability to predict a range of probable ocean states is key to informed decision-making. Here, we present MerCast, a probabilistic ocean forecasting model designed to redefine global-scale prediction by quantifying uncertainty in ocean state estimates. Trained on decades of high-resolution reanalysis products, MerCast integrates diffusion models to generate ensembles of daily forecasts at 1/4-degree resolution, dynamically capturing local-global interactions while preserving fine-scale ocean features essential for accurate predictions.
MerCast's performance is rigorously evaluated using an array of metrics tailored for stochastic forecasting systems, including ensemble spread, probabilistic error assessments, and metrics designed for process-oriented evaluations. Initial results highlight MerCast's skill in forecasting critical variables such as sea surface height, temperature, salinity, and ocean currents, with superior resilience to error accumulation over extended forecast horizons. This work establishes a foundational step toward integrating probabilistic methods in operational ocean forecasting, bridging the gap between efficiency, accuracy, and uncertainty quantification.
How to cite: El Aouni, A., Ruggiero, G., Gaudel, Q., Van Gennip, S., Drillet, Y., and Drevillon, M.: Probabilistic Global Ocean Forecasting Through Diffusion-Based Ensembles, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-12581, https://doi.org/10.5194/egusphere-egu25-12581, 2025.
The El Niño-Southern Oscillation (ENSO) is a global significant signal in marine science and exerts substantial climatic and socioeconomic impacts worldwide. However, the long-term prediction of ENSO remains a challenge because of its diversity, irregularity and asymmetry. Here, we develop a spatiotemporal fusion transformer network (STFTN), which designed a parallel encoder structure to effectively extract spatiotemporal information from sea surface temperature anomaly and Niño3.4 index simultaneously, thereby enhancing the precision of Niño3.4 index forecasts. STFTN leverages the attention mechanism within its parallel encoder structure to extract global characteristics and establish remote dependencies on targets. With this structure, STFTN displays better prediction accuracy in different lead months. Furthermore, the activation map used in STFTN visualizes the contribution of the predictors to the output which helps to comprehend the factors contributing to ENSO events. The results highlight the potential of our model of ENSO forecasts and comprehension.
How to cite: Zhao, A. and Du, Z.: ENSO Forecasts with Spatiotemporal Fusion Transformer Network, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-2550, https://doi.org/10.5194/egusphere-egu25-2550, 2025.
Accurate forecasting of ocean dynamics is essential for understanding the distribution of heat, salinity, and nutrients in the ocean. While data-driven machine learning models offer promising solutions for ocean forecasting and emulating ocean models, they often lack physical consistency (i.e., adherence to the physical laws of fluid dynamics) and explainability. In this study, we introduce a deep neural network architecture leveraging Fourier Neural Operators (FNO) for efficient forecasting of ocean surface dynamics: sea level, temperature, and salinity. FNOs excel in learning resolution-invariant solutions of partial differential equations (PDEs), offering a scalable alternative to traditional physics-based models. Operating in Fourier space enables differentiation to be treated as multiplication, which is the basis of spectral methods used for solving PDEs, including the Navier-Stokes equations that govern hydrodynamic models. Therefore, it is intuitive that by directly parameterizing the integral kernel in Fourier space, the model can learn PDE solutions more efficiently. FNOs also enable training on low-resolution data and evaluation on high-resolution data, which helps minimize the growth of autoregressive errors.
Our model is trained on the Baltic Sea Physics Analysis and Forecast dataset to predict sea surface parameters, including sea level, temperature, and salinity. The Baltic Sea is a non-tidal, semi-enclosed sea with a complex coastline, shallow sea, significant salinity gradients, and permanent stratification, which makes it a unique and challenging testbed for ocean modelling. Input variables include the initial state, atmospheric forcing, and bathymetry, and the model is trained to predict ocean surface dynamics (sea level, temperature, and salinity) and learn the mapping from time t to t+1. In the inference step, the model is initialized with the initial sea surface inputs from an out-of-sample testing dataset and iteratively generates forecasts for τ time steps. Evaluation of the model demonstrates competitive forecasting skill compared to physical models, while significantly reducing computational costs. This study highlights the potential of FNOs to advance knowledge-driven machine learning models for ocean forecasting. These models, as cost-effective alternatives to high-resolution physical ocean models, can pave the way for more efficient, scalable approaches to understanding and predicting ocean dynamics.
How to cite: Jahanmard, V., Ellmann, A., and Delpeche-Ellmann, N.: Fourier Neural Operators for Emulating Ocean Models: Towards a Knowledge-Driven Machine Learning, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4354, https://doi.org/10.5194/egusphere-egu25-4354, 2025.
Mixed layer eddies (MLE) are submesoscale structures, characterized by spatial and temporal scales of O(10 km) and O(1 day), generated by mixed layer instability under conditions of strong horizontal buoyancy gradient and weak stratification. MLE produces mixed layer restratification, which has important implications for global ocean and climate dynamics. Existing parameterizations represent MLE effects with a streamfunction that depends on the horizontal buoyancy gradient, mixed layer depth, and the Coriolis parameter. Machine learning techniques have recently been proposed for improving existing MLE parameterizations. Bodner et al., (2024) proposed an approach for predicting submesoscale vertical buoyancy fluxes using a convolutional neural network (CNN), showing an improvement compared to previous parameterizations.
In this study, we analyze the impact of a new MLE parameterization - based on Bodner et al. (2024) - in a global ocean model simulation performed with NEMO (eORCA25). The implementation of the CNN parameterization in NEMO is performed through EOPHIS (https://github.com/meom-group/eophis/). The CNN simulation (MLE-CNN) is compared with a simulation with a standard parametrization and a simulation without MLE parameterization. With the CNN parameterization, maximum winter mixed layer depths are reduced by 10% with respect to the simulation without parameterization, which is comparable to the reduction obtained with the standard parameterization. The CNN parameterization differs from the standard parameterization in terms of spatial variability. For example, in the tropical region, the CNN produces a vertical heat flux across the mixed layer that can reach twice the magnitude of the standard parameterization. Mixed layer depth from simulations will be compared with observations.
How to cite: Contreras, M., Barge, A., Le Sommer, J., Bodner, A., and Balwada, D.: Assessing the impact of the new mixed layer eddy parameterization based on machine learning in NEMO, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-11727, https://doi.org/10.5194/egusphere-egu25-11727, 2025.
The dynamics of the ocean are dictated by processes that occur over a wide spectrum of scales. Of particular importance are the motions that occur at and around the Rossby radius of deformation, between approximately 10 and 1000 km, so called mesoscale eddies. Mesoscales exchange energy with large-scale ocean currents, thus influencing global ocean circulation. Mesoscale eddies extract potential energy (PE) from the large scale via baroclinic instability, and transfer kinetic energy (KE) upscale via the backscatter effect (inverse cascade). Accurately capturing the global ocean circulation, and the role of mesoscale eddies, is imperative in the development of reliable climate models. However, the resolution required to resolve mesoscales and their contribution to the global ocean energy cycle is far too computationally expensive, particularly for long climate integrations or large ensembles. Thus, the contributions of mesoscale eddies to the energy cycle must be parameterised in terms of the coarse-resolution flow variables of climate models.
Most parameterisations of mesoscale eddies have focussed on resolving individual aspects of the energy cycle. Our approach aims to simultaneously address the downscale transfer of PE and the upscale transfer of KE by leveraging high-resolution simulations and machine learning. This endeavour relies on a theoretical framework that projects the buoyancy flux onto the momentum equations, resulting in an eddy forcing captured by the divergence of the Eliassen-Palm (EP) flux tensor. We develop our parameterisation using the idealised two-layer double-gyre (DG) configuration of MOM6 (ocean component of GFDL + NCAR model). High-resolution DG data is used to train an artificial neural network offline on the correlation between spatially-filtered (large scale) flow features with EP fluxes (subgrid-scale forcing). This parameterisation is shown to improve the representation of the eddy energy cycle in a DG configuration of MOM6. Our results are part of an ongoing effort towards a comprehensive parameterisation capable of capturing the entirety of the mesoscale eddy energy cycle.
How to cite: Everard, K., Perezhogin, P., Balwada, D., and Zanna, L.: Leveraging machine learning to parameterise ocean mesoscale eddies, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-13051, https://doi.org/10.5194/egusphere-egu25-13051, 2025.
We discuss the use of systematic ‘a posteriori’ calibration in the development of complicated (but theory-based) parameterizations. With ‘a posteriori’ calibration, model error is assessed using the results of forward simulations, thereby incorporating numerical error, numerical stability, model-specific implementation details, and alleviating the need for explicit data for all parameterized model components. We show how calibration illuminates the parameterization development trade-off between reductions in model bias, producing better predictions, and increased parametric complexity, the latter which can decrease a model’s ability to extrapolate, increase both the data requirements and computational expense of the calibration. We illustrate the importance of a posteriori calibration by describing the iterative development of CATKE, a new parameterization we develop within CliMA for the fluxes associated with small- or "micro-scale" ocean turbulent mixing on scales between 1 and 100 meters. For calibration we use Ensemble Kalman Inversion to minimize the error between a set of large eddy simulations (="the truth") and predictions of the parameterization and this way find optimal values for the free parameters. Without systematic calibration we cannot make informed choices about parameterization development because we cannot distinguish between structural error and error due to non-optimal parameter values.
How to cite: Constantinou, N., Wagner, G., Hillier, A., Silvestri, S., Souza, A., Burns, K., Hill, C., Campin, J.-M., Marshall, J., and Ferrari, R.: Calibration-driven parameterization development, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-7180, https://doi.org/10.5194/egusphere-egu25-7180, 2025.
The study of mesoscale oceanic eddy dynamics requires regular, high-resolution space-time grids of topography observations. However, most observations come from the constellation of altimetry satellites, which measure the topography along very fine and still very sparse tracks, and surface currents must therefore be calculated using level 4 topography maps. These level 4 maps used operationally are produced by methods based on objective analysis (OA, Le Traon et al., 1998), such as historically used in the DUACS production until end 2024, or variational resolution, such as MIOST (Ubelmann et al., 2022), but their spatial resolution limits the scales of dynamics that can be resolved. While OA and MIOST can capture mesoscale dynamics down to approximately 150–200 km, sub-mesoscale features remain inaccessible with these methods.
Recent advancements in neural network-based mapping models have the potential to refine the resolution of mesoscale topography reconstruction. The NeurOST model developed by S. A. Martin (2024), for instance, improves the spatial resolution by 30% compared with existing conventional methods like OA, establishing itself as a state-of-the-art technique in level-4 topography mapping. While the 4DVarNet model developped by Febvre et al. (2024) has proven effective in Observing System Simulation Experiments (OSSE) over the Gulf Stream, it has not yet been applied on real altimetric observations or on a global scale.
In this study, we leverage the 4DVarNet model to estimate global surface current maps from both conventional nadir altimetry and SWOT KaRIn swath data. The model was trained on GLORYS12V1 reanalysis data over the Gulf Stream and the Agulhas Current, and subsequently applied to global altimetric observations, including SWOT KaRIn.
Our results show that 4DVarNet-derived topography maps from nadir altimetry improve the effective resolution OA and over NeurOST in regions of high variability and strong currents, such as the Gulf Stream, Kuroshio, Agulhas and Brazil currents. The inclusion of SWOT KaRIn data further enhances the effective resolution and significantly reduces mapping errors. 4DVarNet's reconstructions also reveal more small-scale vortex structures and deformations compared to NeurOST. The resulting maps seem to improve our ability to observe eddy dynamics and their impact on energy transfer between different scales.
Nevertheless, the model still needs many improvements to provide satisfactory topography on a global scale. Ongoing and future work includes further investigation into the contribution of additional geophysical variables to the topography reconstruction performance of 4DVarNet, such as bathymetry, sea surface temperature, salinity and ocean color, and the exploration of an unsupervised learning scheme for better generalization to real altimetric data. These developments aim to improve the model's applicability to diverse oceanic regions and enhance its ability in capturing sub-mesoscale eddy dynamics.
How to cite: Laloue, A., Anadon, C., Treboutte, A., Ballarotta, M., Pujol, M.-I., and Fablet, R.: Level 4 global topography mapping with 4DVarNet, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-9778, https://doi.org/10.5194/egusphere-egu25-9778, 2025.
The 4D variational assimilation (4DVar) framework is widely used in classical numerical weather prediction and geophysical data assimilation. However, a crucial assumption in 4DVar is that the model state that is close to the true state corresponds to the minimizer of the 4DVar cost function. Using a single-layer quasi-geostrophic (QG) model, we study scenarios where this assumption breaks down, particularly in the presence of model errors and suboptimal initialization. By introducing controlled perturbations in the initial conditions—we design experiments to investigate the sensitivity of 4DVar solutions. We find that minimizing the 4DVar score does not always correlate with achieving lower accuracy, suggesting the presence of local minima in the optimization process.
4DVarNet, an end-to-end neural network based on variational data assimilation formulation, is trained in a supervised manner to solve the data assimilation task. This study aims to understand the advantage of trainable solvers that solve the same optimization problem using supervised learning, generating more accurate solutions efficiently. Through this case study based on observing system simulation experiments for sea surface geophysical fields, we show that supervised learning can overcome the minimization challenges of 4DVar when faced with observations that are irregular and highly sparse which are critical to address problems in ocean reconstruction. The advantage of learning allows 4DVarNet to discover hidden representations that are suitable for solving specific data assimilation tasks with better accuracy.
How to cite: Roy, S. K. and Fablet, R.: Performance Gains and Advantages of 4DVarNet in End-to-End Learning for Data Assimilation, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4300, https://doi.org/10.5194/egusphere-egu25-4300, 2025.
Satellite observations provide indispensable data that is assimilated into numerical ocean models to correct errors and biases. Traditionally, sea surface height (SSH) from satellite altimeter tracks, sea surface temperature (SST), and more recently, sea surface salinity (SSS), have been assimilated into these models. Temperature and salinity are part of the governing equations of ocean dynamics, and SSH is directly derived from the state of the resolved ocean, making these variables a first choice for data assimilation. However, satellite-derived Chlorophyll-a (Chl-a) data, which offer high-resolution information, is not typically assimilated. This is primarily because this variable is not solved by the physical models, and the biochemical models that simulate broader marine ecosystems, including phytoplankton dynamics and nutrient cycles which do estimate Chl-a, are computationally expensive and not used in operational models.
In this study, we utilize a ten-year free run of a biochemical ocean model of the Gulf of Mexico to simulate satellite observations, including altimeter tracks, SST, SSS, and Chl-a. We trained and tested various machine learning architectures, including Convolutional Neural Networks (CNNs), Autoregressive Convolutional Neural Networks (AR-CNNs), and Vision Transformers, to learn the relationship between these variables and the SSH. The trained models were then used to estimate sea surface height from the simulated observations to estimate the current and future state of the sea surface height, leveraging the autoregressive properties of one of the tested architectures. Our results demonstrate that this approach outperforms the traditional interpolations in metrics like the RMSE. Finally, we applied the best-performing models to real satellite observations, highlighting the potential of improving SSH estimation quality.
How to cite: Velasco-Zavala, J., Zavala-Romero, O., Sheinbaum, J., Miranda, J., Hiron, L., Bozec, A., Bulusu, S., and Chassignet, E.: Enhancing sea surface height estimation using satellite-derived chlorophyll-a and temperature data via machine learning: a case study in the Gulf of Mexico, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-13573, https://doi.org/10.5194/egusphere-egu25-13573, 2025.
Understanding the dynamics of phytoplankton communities in response to physical
environmental changes is essential for evaluating the impact of climate change on marine
ecosystems. Satellite observations provide a rich dataset spanning over two decades,
capturing physical sea surface parameters such as temperature, salinity, and sea surface
height, alongside biological insights such as ocean color. Ocean color data, in particular, is
processed to estimate sea surface chlorophyll-a concentrations — a widely recognized proxy
for phytoplankton biomass. Recent advancements in ocean color observation have further
enabled the characterization of phytoplankton community structure in terms of functional
groups or size classes.
However, linking satellite-derived physical parameters to biological indicators remains
challenging due to spatial and temporal variability.
Can physical data reliably predict patterns in ocean color, such as chlorophyll-a
concentrations and phytoplankton community structures, and potentially assess their
variations? This study addresses this question through a deep-learning approach, utilizing
an attention-based autoencoder model to learn relationships between physical variables and
ocean color data, including chlorophyll-a concentrations and phytoplankton size classes at
weekly and 1° spatial resolution.
Our trained deep-learning model effectively captures patterns and correlations between
physical parameters, chlorophyll concentrations, and phytoplankton size classes. It enables
detailed exploration of how physical factors influence biological variability across different
temporal scales. Utilizing a phytoplankton database spanning 1997–2023, this approach
demonstrates promising results in replicating chlorophyll concentrations, inferring
phytoplankton size classes, and shedding light on the potential links between physical and
biological data.
This study highlights the potential of machine learning for ecological research, contributing to
more accurate trend analyses. Understanding phytoplankton variability is critical for marine
ecosystem management, given their role in global carbon cycling. This methodology
underscores the value of deep-learning to anticipate phytoplankton dynamics under
changing environmental conditions.
How to cite: Ollier, L., ElHourany, R., and Levy, M.: Deep learning algorithm to uncover links between satellite-derived physical drivers and biological fields., EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-6755, https://doi.org/10.5194/egusphere-egu25-6755, 2025.
Ocean salinity is a fundamental variable that determines seawater density and, therefore, stratification and oceanic dynamics. To understand how salinity is affected and how it contributes to ocean processes, its variability must be studied, particularly through in situ observations. Unfortunately, while temperature observations were limited during the 20th century, salinity observations were even sparser, as some instruments were designed to measure temperature only. The development of the Argo observing system since 2002 has improved sampling and reduced the disparity between both variables, enabling better assessment of salinity variability over the past 20 years at interannual to decadal scales. In this study, we estimate salinity covariability with temperature from the Argo period to reconstruct monthly subsurface salinity fields, in the tropical Pacific between 1930 and 2001, leveraging temperature observations. The analysis is performed using the data-driven RedAnDA method, which combines Data Assimilation, Analog Prediction, and Reduced-space Interpolation, first validated using synthetic data. We reconstruct the 20th century interannual variability of salinity associated with ENSO events both at the surface and in the subsurface. Notably, thanks to the coupling with temperature, the representation of stratification and its modulation by vertical salinity gradients is enhanced. This new method and product provide for the first time the possibility to extend the hydrological time series consistently in the past, offering potential new insights into mechanisms generating decadal variability in the Pacific.
How to cite: Oulhen, E., Kolodziejczyk, N., Tandeo, P., Blanke, B., and Sévellec, F.: Reconstructing historical salinity fields over the 20th century using a data-driven method and Argo data, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-10637, https://doi.org/10.5194/egusphere-egu25-10637, 2025.
Understanding ocean salinity is crucial for tracking changes in the Earth's water cycle and climate. However, collecting accurate salinity data has been challenging due to limited observations, especially in certain regions. This study focuses on the development of a method to create 2-dimensional maps of ocean salinity and its trends on pressure surfaces from sparse observations. An unsupervised classification technique called Gaussian Mixture Modeling (GMM) is used to identify coherent regions where temperature and salinity are tightly related at constant pressure. By grouping similar ocean regions using GMM, we are able to predict missing salinity data and fill gaps in historical salinity records from 1970 to 2014. The results show that this approach effectively estimates past salinity data. In the South Atlantic, at a pressure of 539 dbar, the root mean square error of salinity and of the linear trend of salinity are 0.040 g kg⁻¹ and 2.1 10⁻³g kg⁻¹ yr⁻¹. The method could help fill in missing salinity observations and thus improve our understanding of the intensification of the global water cycle in response to climate change.
How to cite: Piedagnel, E., Sohail, T., and Zika, J.: Mapping ocean salinity data using Gaussian Mixture Modeling., EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-1385, https://doi.org/10.5194/egusphere-egu25-1385, 2025.
Posters on site: Thu, 1 May, 16:15–18:00 | Hall X4
ENSO exerts profound impacts on global climate change through ocean-atmosphere interactions and serves as a critical factor in global climate prediction. However, its prediction remains challenging due to the complex spatiotemporal interactions and evolution processes, as well as the varying degrees of correlation and teleconnection across different geographical regions. To address this issue, this study proposes an advanced ENSO forecasting framework based on regional predictions and model ensemble. The framework leverages a graph self-attention mechanism (GAT) to learn and capture the spatiotemporal dependency signals of ENSO, which are then incorporated as physical constraints into a spatiotemporal graph convolutional neural network (STGCN) for regional predictions. Furthermore, machine learning algorithms, including XGBoost and SVR are employed to integrate the predictions from different regions. Experimental results based on reanalysis data demonstrate the effectiveness and robustness of the proposed framework, achieving a correlation skill exceeding 0.8 within a 12-month lead prediction period, and significantly improving the computational efficiency by filtering key signals.
How to cite: Xiao, H., Song, Z., and Wang, L.: Regional Ensemble ENSO Prediction Based on Graph Neural Networks with Self-Attention, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-2635, https://doi.org/10.5194/egusphere-egu25-2635, 2025.
The airborne radar altimeter can be extrapolated to a variety of parameters, including sea surface height, sea surface wind speed, significant wave height, and the topography of land, sea ice and ice cap. However, the airborne radar altimeter observation data contains signal error terms such as airborne platform jitter and ocean waves, which will lead to a large bias in the observation data. Here, we propose a method based on the combination of bandpass filtering and adaptive feature AI analysis to achieve the inversion of high-resolution sea level anomaly (SLA) data from airborne radar altimeter aliased signals.
For the airborne altimeter along-track data, statistical analyses were first performed. After that, the along-track data are filtered to remove the influence of ocean waves signals and flight platform oscillations, and the secondary interpolation is fitted based on the interval of the airborne altimeter data. According to the sampling interval of the altimeter data, the mean sea surface (MSS) and tide data under the along-track are processed to obtain the corresponding SLA data. The same interpolation method is used to process AVISO and SWOT L3 data. Finally, through the deep learning framework, the adaptive feature AI analysis is constructed to invert the SLA data, optimise the model and achieve accurate SLA prediction. The experimental results show that the RMSE of the SLA of the airborne altimeter inversion data with the along-track SWOT L3 and AVISO data are 1.12cm and 0.44cm, respectively, and the airborne altimeter data can acquire more small-scale change signals. This study verifies the working mechanism of the new system payload and lays a solid data and algorithm foundation for the development of subsequent satellite payloads.
How to cite: Ren, M., Yu, F., Zhang, X., Tang, J., and Chen, G.: Research on sea level inversion method from airborne radar altimeter , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-2984, https://doi.org/10.5194/egusphere-egu25-2984, 2025.
The irregular and incomplete coverage of in-situ ocean temperature profile observations is a major problem for various scientific applications in ocean and climate research and operational fields. However, high-resolution gridded datasets are needed to support applications. Here, we explore a physics-informed machine learning approach based on partial convolutions with multi-branch U-Net neural network structure to reconstruct the subsurface temperature profile fields with 0.1°×0.1° weekly resolution in Western Pacific Ocean. The input data include in-situ temperature profile observations, high-resolution satellite remote-sensing products (including sea surface height, sea surface temperature, sea surface salinity, etc.), and a coarse-resolution (1°× 1°) gridded subsurface temperature product (IAPv4). We show that the new reconstruction retained the large-scale features represented by the 1°× 1° temperature gridded data but added mesoscale features (because of the inputs of high-resolution satellite data). The application of physical constraints for subsurface vertical structure improves the reconstruction near thermocline. The root mean square error (RMSE) can be reduced by ~12% in the target region in average with greater improvements in the upper layer (0-700m). Further analysis shows the small-scale information is performed well also in the sparse observation coverage area with some typical mesoscale vortex features can be identified, and the features in the strait and offshore regions can be effectively improved compared with coarse resolution 1°× 1° temperature gridded data. The successful application of machine learning in this study provides confidence for the accurate reconstruction of high-resolution ocean and climate data, which can improve and complement the existing data assimilation and objective analysis methods for reconstructing multi-scale ocean information in complex regions.
How to cite: Wei, W., Cheng, L., and Tian, T.: Physics-Informed Machine Learning Reconstruction of High Resolution Ocean Subsurface Temperature Profiles From In-Situ and Satellite Observations, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-3468, https://doi.org/10.5194/egusphere-egu25-3468, 2025.
Ocean General Circulation Models (hereafter OGCM) are critical to the study of past and present climate, and the production of future projections. Unfortunately, they require large amounts of computations at simulation time. On the other hand, deep learning emulators trained on reanalyses are starting to deliver accurate short-term predictions, using comparatively small computational resources, yet they struggle to deliver long term predictions, are not interpretable and do no't factor in the uncertainty in physical parameters. As a result, they cannot be used by climate scientists to understand mechanisms of climate variability, such as tipping points, nor adjustment processes to greenhouse gas emissions.
Aiming to take the best from each world and establish a close interaction between emulators and OGCM, we investigate whether an emulator can be used to provide state variables to an ocean model, which would then handle the temporal integration using physics equations. Namely, we trained a diffusion model on a large dataset of ocean variables produced by NEMO, analysed the newly generated states, propose metrics to assess their physical consistency, and use them as initial conditions of simulations to assess their compatibility (physical and numerical) with NEMO.
Overall, we want to determine whether unconstrained generative models are able to produce realistic solutions, and to assess the tolerance of OGCM to externally generated ocean states, what we consider as a first step towards building an hybrid OGCM.
How to cite: Meunier, E., Lguensat, R., Gachon, G., Kamm, D., and Deshayes, J.: Can diffusion model generate ocean states compatible with OGCM ? , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4189, https://doi.org/10.5194/egusphere-egu25-4189, 2025.
The ocean's biogeochemistry is crucial for understanding the global ocean carbon cycle. Within the climate ocean model Nemo, the PISCES module (Pelagic Interactions Scheme for Carbon and Ecosystem Studies), is based on the numerical calculation of 24 different biological, physical and chemical variables which contribute to a complex bio-geo-chemical relationship to be able to estimate the net source and sinks of primary carbon production. In this work, we want to present the first steps toward using the Deep Neural Networks as a multi-variate problem trained on the model output to predict the next sequences and replace the module with an emulator solely based on machine learning (ML).
How to cite: Rahpoe, N. and Bernardello, R.: A Deep Learn Emulator for Ocean Biogeochemical Modelling, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4191, https://doi.org/10.5194/egusphere-egu25-4191, 2025.
The eddy-permitting NEMO model (ORCA025) is known to exhibit sub-par Southern Ocean circulation features, such as a too weak Antarctic Circumpolar Current transport and cool and warm biases on the Antarctic shelf. The ORCA025 model sits in the numerical grey zone, which is where the horizontal grid resolution can only resolve mesoscale processes over part of the domain. In other parts of the domain, the eddies need to be parameterised, such as high-latitude regions. This difficulty in representing eddies has in-part contributed to the poor Southern Ocean circulation, leading to great uncertainty in key climate metrics such as carbon and heat transport, and the Antarctic ice mass balance. The key question is, how do we parameterise mesoscale eddies where they are most needed, without being detrimental to the resolved flow. Scale- and flow-aware parameterisations have been implemented in NEMO and have led to improvements in some flow characteristics. However, an alternative approach is to leverage data, physics, and machine learning to develop an improved eddy parameterisation.
As part of the project, AI4PEX, we aim to develop a data- and physics-driven mesoscale eddy parameterisation that better captures the dynamical feedback between mesoscale eddies and the large-scale ocean circulation, reducing model uncertainty. In our work, we will attempt to improve an eddy parameterisation that is available in NEMO, GEOMETRIC. To do this we will use a Neural Network trained on high resolution data from realistic global models ORCA12/ORCA36. To reduce the black-box nature of the Neural Network, we will design a loss function that is informed by the physics of mesoscale eddies. Initial investigation of the eddy parameterisation will take place offline in an idealised configuration.
How to cite: Wilder, T., Kuhlbrodt, T., and Swaminathan, R.: Developing a data and physics driven machine learning mesoscale eddy parameterisation for NEMO, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4485, https://doi.org/10.5194/egusphere-egu25-4485, 2025.
In the ocean, submesoscale physical phenomena O(100m) to O(1km) have been reported to play a key role in ocean oxygen ventilation, nutrient supply to the surface ocean, and carbon export, as well as the transfer of energy to larger scales [1]. However, due to limitations in computational resources, current ocean general circulation models are frequently run at resolutions on the order of O(10km) to O(100km) and cannot directly resolve submesoscale turbulence (i.e., subgrid-scale phenomena). Therefore, parameterization schemes are required to simulate these subgrid-scale phenomena.
Recent advances in machine learning have triggered the active exploration of data-driven approaches to parameterization for subgrid-scale phenomena that utilize data from observations and simulations. In previous studies [2, 3], the neural network is trained directly using the same variables as the neural network's output, such as viscosity and diffusivity coefficients. However, this approach does not guarantee that the inferred model parameters accurately represent the state of subgrid-scale phenomena they aim to reproduce. We propose a novel parameterization method for estimating diffusivity and viscosity parameters to parameterize subgrid-scale phenomena and have implemented this method in MITgcm, an ocean simulator [4, 5]. This method trains a neural network using the state variables (i.e., velocity fields, potential temperature, and salinity) derived from the simulation results at a resolution that can directly resolve subgrid-scale phenomena. Therefore, unlike previous studies, the diffusivity and viscosity parameters inferred by the trained network can reproduce the global state of subgrid-scale phenomena.
The ocean simulator MITgcm is implemented in Fortran, which does not have a built-in package to compute gradients within the neural network, in contrast to deep learning libraries (e.g., PyTorch) like Python. In our previous work [4, 5], we used a quasi-newton optimization method, which does not require computation of these gradients. However, the optimization performance of this method was limited. In this study, we use adjoint code within MITgcm to compute gradients for optimizing neural networks and examine the effect of different optimizers on training performance.
Acknowledgments
This work used computational resources of supercomputer Fugaku provided by the RIKEN Center for Computational Science through the HPCI System Research Project (Project ID: hp240394).
References
[1] M. Lévy et. al (2024), The impact of fine-scale currents on biogeochemical cycles in a changing ocean, Annual Review of Marine Science, 16(1), 191–215.
[2] Y. Han et. al (2020), A moist physics parameterization based on deep learning, Journal of Advances in Modeling Earth Systems, 12(9), e2020MS002076.
[3] Y. Zhu et. al (2022), Physics-informed deep-learning parameterization of ocean vertical mixing improves climate simulations, National Science Review, 9(8), nwac044.
[4] R. Irie et. al (2024), Parameterizing ocean vertical mixing using deep learning trained from high-resolution simulations, EGU General Assembly 2024, EGU24-2297.
[5] R. Irie et. al (2024), Optimizing a deep-learning model for parameterizing submesoscale phenomena in an ocean simulator, Workshop on Scientific Machine Learning and Its Industrial Applications.
How to cite: Irie, R., Stewart, H., Hisada, M., and Yaguchi, T.: Performance evaluation and optimization of a deep learning parameterization method trained from submesoscale-permitting ocean simulations, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4606, https://doi.org/10.5194/egusphere-egu25-4606, 2025.
Nighttime ship detection plays a vital role in understanding oceanic patterns and human activities in marine environments. As an observational approach in ocean science, it enables researchers to monitor vessel distribution patterns, analyze maritime traffic flows, and collect valuable data about human interactions with marine ecosystems. While the VIIRS Day-Night Band (DNB) sensor enables nighttime vessel detection from space, conventional detection methods primarily rely on threshold-based techniques, which show limitations in handling complex environmental factors such as cloud coverage and varying atmospheric conditions. To overcome these challenges, this study presents an automated ship detection approach that combines VIIRS DNB imagery with AutoML techniques. Our AutoML framework automatically optimizes model parameters and features to adapt to various environmental conditions, providing more robust detection capabilities compared to traditional threshold-based methods. The methodology incorporates AIS data for model training and validation to enhance detection accuracy. Our experimental results demonstrate improved detection performance across diverse maritime environments and weather conditions, effectively addressing the limitations of conventional threshold-based approaches. This research contributes to advancing pattern recognition in oceanic observations by providing an automated approach for identifying vessel activities in nighttime satellite imagery.
How to cite: Seong, N., Jung, O., Jung, Y., and Song, S.-H.: Nighttime Ship Detection Using VIIRS DNB Data: An AutoML Approach, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-5626, https://doi.org/10.5194/egusphere-egu25-5626, 2025.
Accurate ocean forecasting models are crucial for both scientific research and practical application, such as understanding ocean dynamics and efficient ship route planning. While traditional numerical ocean models have proven effective, they require substantial computational resources due to the complexity of solving partial differential equations. In recent years, data-driven weather forecasting models have demonstrated their ability to provide accurate predictions at lower computational costs compared to conventional numerical weather prediction models while their application to ocean forecasting remains limited.
This study explores a data-driven ocean forecasting model for 10-day global forecasting, employing a multi-scale graph neural network (GNN) to capture the multi-scale features of ocean variables while incorporating graph structures that account for land masks. To reflect the effects of atmospheric forcing, surface atmospheric variables are combined with ocean variables and used as GNN’s node input features. The model was initially trained on paired reanalysis data samples with a 1-day interval to minimize the mean squared error. Subsequently, it was fine-tuned using auto-regressive rollouts across multiple time steps. The forecasting process involves autoregressive steps, where the predicted ocean variables from the previous step and weather forecasting variables provided by an operational center are used as inputs for the next step.
Preliminary experiments comparing the proposed model with persistent forecasts showed the skillfulness of the proposed model. Sensitivity experiments were conducted to evaluate the impact of atmospheric forcing by replacing weather forecasting data with climatological data. The evaluation was conducted over a one-year period across the global ocean employing reanalysis data as references. The results showed that using weather forecasting data improved the accuracy of surface ocean variable predictions compared to using climatology. Specifically, the RMSE was reduced by 6.6%, 6.2%, and 1.0% for 3-day-ahead, 5-day-ahead, and 10-day-ahead forecasts, respectively, representing the median improvement across the period and variables. The improvements varied across variables; for instance, salinity showed a consistent improvement of almost 1% across all lead times, whereas northward velocity showed greater improvements at shorter lead times, such as an improvement of 22% at 3-day-ahead forecasts.
The results indicate that it is crucial for data-driven ocean models to incorporate atmospheric forcing, similar to numerical ocean models. These findings suggest that the multi-scale GNN-based ocean forecasting model that integrates atmospheric forcing offers a potential approach for 10-day global ocean forecasting.
How to cite: Hirabayashi, Y., Matsuoka, D., and Kimura, K.: Data-driven Ocean Forecasting Models with Multi-Scale Graph Neural Networks for 10-day Global Forecasting, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-8030, https://doi.org/10.5194/egusphere-egu25-8030, 2025.
Motivation The increasing amount of plastic debris in the oceans calls for quick action to prevent irreversibly damaging our world’s largest ecosystem. To this end, tracking plastic debris and understanding its dynamics could facilitate collection campaigns and help monitor the evolution of the threat. To achieve this goal, accurate models are necessary to predict the dynamics of floating objects at the ocean surface, which are subject to currents and winds. Physical models and remote sensing data estimate these influencing forces. However, using them directly in process-based models still leads to a significant gap between the true dynamics and the predicted trajectory. Hence, we aim to minimize this gap by resorting to data-driven machine-learning methods.
Data We can identify two different scenarios where the dynamics of floating objects differ: trajectories close to coastal regions and trajectories in the open ocean. As a consequence, we focus on two different datasets: the first aims to predict dynamics in coastal regions for 24 hours. The second focuses on open-ocean dynamics, where we try to predict trajectories for multiple days. As target variables, we use data from the Global Drifter program, which contains several thousand GPS-tracked free-floating buoys. The contextual information about the ocean surface current is extracted from Copernicus Marine and HYCOM. Wind data is taken from ERA5.
Approach We develop a denoising diffusion model that generates multiple trajectories based on surface current and wind, as provided by physical models. In contrast to the unstructured i.i.d. Gaussian noise in standard denoising diffusion, we use a more suitable process: Brownian motion noise, which has a small variance close to the start of the trajectories and increases with time. The denoiser model is autoregressive and based on a multilayer recurrent neural network that iteratively learns to remove the noise from random realizations of this Brownian motion.
Results We found that the model not only outperforms physical models on the coastal dataset but also provides a posterior distribution of the predicted trajectories, thus offering a measure of uncertainty without additional overhead.
How to cite: Donner, C., Goshtasbpour, S., Dalsasso, E., Volpi, M., Russwurm, M., and Tuia, D.: Autoregressive denoising diffusion for predicting trajectories of floating objects in oceans, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-11223, https://doi.org/10.5194/egusphere-egu25-11223, 2025.
Big data has become increasingly important in marine geoscience, where in situ measurements are often limited, leaving large portions of the seafloor unsampled. To address this gap, we present a data-driven approach that leverages non-parametric machine learning algorithms—specifically, an ensemble of k-Nearest Neighbors (kNN) and Random Forest regressors—to predict a global geospatial prediction of median grain size (D50) at a 2-arc minute resolution. Our methodology incorporates parametric uncertainty quantification in the form of distance-to-nearest-neighbor metrics in feature space, thereby creating spatially explicit uncertainty maps that highlight regions where additional data collection would most effectively improve model predictions. This emphasis on parametric uncertainty serves as a roadmap for data-driven exploration, reducing the time, energy, and cost associated with collecting or curating a comprehensive dataset.
We train the model on ~40,000 publicly available, seafloor grain size measurements and iteratively optimize hyperparameters based on prediction error and out-of-sample validation. The final model is a global prediction of seafloor grain size with a correlation of ~0.65 between observed and predicted grain size values. We also apply a ranked noise grid analysis to select predictor variables that minimize the overall predictive error, ensuring the feature set is robust and agnostic to human bias.
Regions with sparse data coverage or atypical geological conditions manifest as areas of high uncertainty, underscoring the need for targeted sampling. By mapping this uncertainty, our framework facilitates strategic data acquisition efforts and reduces curation time and cost. We demonstrate the impact of sampling high uncertainty regions on not only improving predictions in the newly sampled geographical location but are also geologically similar (close in parameter space) around the globe. In doing so, it demonstrates how the synergy between machine learning approaches and systematic data-driven exploration can enhance the dependency of global seafloor property models. Our predicted grain size map provides a proxy for further regional and global studies that rely on grain size measurements, while more broadly highlighting the transformative potential of machine learning methods to refine our approach to data exploration and curation.
How to cite: Renzaglia, J., Lee, T., and Le, A.: Global Seafloor Grain-Size Prediction: A Data-Driven Approach, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-12879, https://doi.org/10.5194/egusphere-egu25-12879, 2025.
OceanRep proposes a novel AI foundation model for ocean dynamics, a cornerstone for understanding and predicting climate change. Inspired by the success of AtmoRep, a deep learning model for atmospheric dynamics, OceanRep seeks to extend this framework to the ocean. In order to leverage transformer models and large-scale, multi-resolution oceanographic data (e.g., from ocean model FESOM2), the design is based on vision transformers, modified to handle four-dimensional data represented by space-time tokens, and with a U-net-type backbone to capture intricate interactions within the ocean system. For pre-training, BERT-style masking is used.
Preliminary results demonstrate OceanRep's ability to generate skillful week scale forecasts using data from a 1-degree resolution FESOM2 simulation. Ultimately, the project aims to create a robust model capable of simulating ocean and sea ice dynamics over decades. This will allow for extensive numerical experimentation and rapid generation of accurate "what-if'' scenarios. These capabilities hold immense value for climate adaptation strategies, policy development, and scientific exploration of the intricate dynamics governing the Earth system.
How to cite: Nowak, K., Koldunov, N., Jung, T., Danilov, S., Lessing, C., and Luise, I.: OceanRep: A Foundation Model for Ocean Dynamics, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-15061, https://doi.org/10.5194/egusphere-egu25-15061, 2025.
Accurate high-resolution forecasting of oceanic and atmospheric states remains a critical challenge. This study introduces an AI-based regional downscaling framework employing a U-Net deep learning architecture, trained on coarse-resolution simulations. By embedding physical constraints, the model effectively bridges scales, capturing fine-grained dynamics unresolved in traditional approaches.
The framework significantly enhances computational efficiency, reducing forecast times from hours to seconds per region while maintaining high accuracy. Its integration with data-parallel computing units enables scalable multi-region applications. Applied within a coupled ocean-atmosphere-wave-tide system, the model excels in reproducing extreme events and mesoscale dynamics.
This work highlights the potential of AI in offering scalable, precise solutions for forecasting, climate science, and disaster management.
How to cite: Huang, Y., Chen, R., Ji, L., and Zhang, S.: AI-Driven Regional Downscaling for High-Resolution Oceanic and Atmospheric Forecasting, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-16391, https://doi.org/10.5194/egusphere-egu25-16391, 2025.
Differentiable programming has emerged as a powerful tool in geoscientific modelling, offering new possibilities for optimization and parameter calibration. However, this approach requires the underlying physical models to be differentiable in order to compute gradients and apply optimization algorithms. In practice, current-generation geoscientific models are generally not differentiable, which limits the use of variational approaches to calibrate their parameters. In the past few years, several strategies have been proposed to overcome this limitation.
Here, we explore the use of deep learning techniques for the calibration of vertical physics schemes of current-generation ocean models. We propose to build conditional emulators of single column ocean models to approximate the gradient of their solution with respect to their physical parameters. Our baseline is a single column ocean model, implemented in Jax, which provides a differentiable framework for the calibration of ocean vertical physics schemes. We leverage this framework to generate sets of simulations for the design of deep conditional emulators of the model, and assess their ability to approximate the gradient of the model in an inverse problem setting.
We focus on several idealized cases corresponding to different forcing conditions, starting from the Kato-Philips case. It describes the evolution of a water column with no heat flux and uniform wind friction velocity. We obtain various trajectories for uniformly sampled n-uplets defining the initial conditions, friction velocity, and physical parameters. With this dataset, we train and test different kinds of neural networks, exploring architectures and losses, to make the most of temporal and spatial dependencies.
Comparison with the fully differentiable baseline solution shows that deep conditional emulators are able to predict the system states both forward and backward, with different initial and forcing conditions, and can be used to calibrate ocean model parameters. Our results therefore illustrate how deep emulators are a potential solution to take over the non-differentiability of existing geoscientific models, and solve inverse problems for their calibration.
How to cite: Durif, A., Mouttapa, G., Le Sommer, J., and Fablet, R.: Deep Conditional Emulators for calibrating ocean vertical physics schemes, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-17009, https://doi.org/10.5194/egusphere-egu25-17009, 2025.
Sea surface height (SSH) data derived from satellite altimetry are widely used in data assimilation to enhance the representations of ocean currents and subsurface temperature and salinity structure. However, accurately projecting SSH onto subsurface temperature and salinity structures presents significant challenges. Consistent adjustment to temperature and salinity profiles are required to conserve the potential vorticity, which depends on the vertical density gradient. Otherwise, SSH assimilation can produce adverse effects (Fu and Zhu, 2011). Several methods have been proposed to address this issue, including the CH96(Cooper and Haines, 1996) method used by Chang et al (2023), which constructs pseudo profile derived from altimetry data by preserving density structures. However, when tidal forcing is applied to an ocean model, the CH96 method becomes challenging to use due to the significant difficulty in removing tidal signals. To overcome these limitations, this study proposes a Transformer-based machine learning approach to reconstruct T/S (Temperature and Salinity) profiles from SSH. Transformers are well-suited for capturing complex correlations through attention mechanisms (Vaswani et al., 2017), making them ideal for learning T/S profiles influenced by diverse and intricate variables. Monthly GLORYS data from 2010 to 2020 was utilized to train a model for reconstructing T/S profiles. The data was structured into 1/2° grids, where learning was conducted grid-by-grid to capture spatiotemporal variability. For improved accuracy and better incorporation of surrounding grid influences, a combination of 4D-Var techniques and CNNs was employed. This approach learns patterns by grouping four neighboring grids into a quadrilateral for joint training, ensuring that the final profiles account for interactions across grids. During prediction, the surface information of a target point is distributed to its four neighboring low-resolution grids to generate profiles, which are then interpolated into a high-resolution 1/12° grid. The final profile is computed using inverse distance weighting (IDW) interpolation, prioritizing the influence of closer profiles for spatial consistency. Model performance was validated by comparing predicted profiles with low-resolution maps for 2021–2022 over the northwest Pacific region (10°S–45°N, 120°–170°E), achieving an RMSE of 0.55 for temperature and 0.12 for salinity. The model will be further validated against in-situ observational data. We plan to conduct experiments to investigate the impact of assimilation of the reconstructed profiles and compared against CH96-derived profiles to evaluate their accuracy and advantages.
How to cite: Lee, G.-M. and Kim, Y.-H.: Machine Learning-Based Reconstruction of T/S Profiles from Satellite-Derived SSH Using Transformer Models, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-18226, https://doi.org/10.5194/egusphere-egu25-18226, 2025.
In the last decades, mesoscale air-sea interactions have received increasing interest from the scientific community. Mesoscale thermal (sea surface temperature influence, TFB) and mechanical (oceanic surface current influence, CFB) air-sea interactions have been shown to have a strong influence on the wind up to the troposphere and on ocean dynamics. However, from an oceanic perspective, running an atmospheric model is very expensive. To overcome this issue, we have developed a convolutional neural network (CNN) that aims to reproduce the mesoscale ocean-atmosphere interactions. Training was performed with simulated data from a realistic coupled ocean-atmosphere tropical channel simulation (45°S- 45°N) using NEMO for the ocean model, WRF for the atmosphere model, and the OASIS3-MCT coupler. As a first step, the CNN was trained over two energetic regions (the Agulhas Current and the Kuroshio) to predict mesoscale surface stress anomalies from large-scale atmospheric and mesoscale oceanic inputs. Validation over the Gulf Stream and other regions shows that the CNN successfully reproduces the surface stress anomalies associated with both TFB and CFB. In a second step, to parameterize the mesoscale ocean-atmosphere interactions, we coupled the CNN to NEMO via an Eophis library (pyOASIS) and ran a simulation over the tropical channel configuration. In this talk, we will present our main results in terms of oceanic energetics and ocean-atmosphere energy transfer.
How to cite: Ernout, N., Renault, L., Simon, E., Benshila, R., Zhang, S., and Le Sommer, J.: Toward a New Parameterization of Fine-Scale Ocean-Atmosphere Interactions Based on a Machine Learning Approach, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-18713, https://doi.org/10.5194/egusphere-egu25-18713, 2025.
Systematic biases pose a significant challenge in ocean general circulation models, where numerical approximations, unresolved physical processes, and parameterization choices can lead to state-dependent errors. Addressing these biases is crucial for improving forecasts of the Earth’s climate system, yet remains nontrivial—particularly given the sparse nature of ocean observations, which complicates bias detection and correction.
One promising route is to harness analysis increments within a Machine Learning (ML) framework to learn state-dependent systematic errors from archived data assimilation corrections. For instance, neural networks can be used to train a model with the ocean state as input and the Data Assimilation corrections as output. By training on these increments, the ML model learns how errors systematically depend on the local physical state.
In our work, we use outputs from ocean reanalysis data using variational data assimilation and the NEMO ocean model. The ML-based correction is embedded in NEMO’s tendency equations as an additional forcing term, allowing the model to evolve more realistically by accounting for state-dependent systematic errors in temperature and salinity.
However, the sparsity of ocean observations can lead to “punctual” analysis increments that contain not only model biases but also noise from intermittent measurement coverage, errors, and initial-condition uncertainties. To mitigate this issue, we apply a two dimensional low-pass filter to remove high-frequency fluctuations in both the ocean fields and the analysis increments, preserving larger-scale patterns.
We adopt a feed-forward neural network (NN) that processes vertical profiles. By focusing on the ocean’s vertical stratification and processes, the network is trained on these filtered analysis increments and learns the non-linear relationships linking NEMO’s state variables (temperature, salinity) to the corrections identified by the variational scheme. Through this level-specific, column-oriented NN, the model more effectively adjusts for systematic errors.
In this poster, we present preliminary results on offline validation of the trained NN—predicting analysis increments on independent test data beyond the training period—without yet applying these corrections in a fully integrated forecast. Our preliminary findings show how well the NN reproduces systematic biases at various depths and in different oceanic regions, even under sparse data conditions and complex multi-scale dynamics. This demonstration highlights the potential of combining analysis increments with ML to systematically reduce model errors in next-generation ocean prediction systems, setting the stage for future work that integrates these learned corrections into an online, real-time workflow.
How to cite: Nunziante, G., Storto, A., and Yang, C.: Systematic error correction in numerical ocean models with artificial neural networks, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-19733, https://doi.org/10.5194/egusphere-egu25-19733, 2025.
Understanding internal wave is essential, as they exert a profound influence on a multitude of oceanic processes, including mixing and the transfer of energy across a vast range of spatial scales. The phase of internal waves can undergo a rapid alteration during propagation, resulting in the formation of broad spectral peaks. In this study, we introduce a stochastic model designed to parametrise the spectral properties of coastal internal waves. This model employs a Lorentzian function to characterise the broad internal tide peaks and a Matern function for the energy continuum. The efficacy of our model is validated using long-term in-situ mooring temperature data from the Australian Northwest Shelf (NWS) and Timor Sea. By optimising the model parameters using debiased Whittle likelihood in the frequency domain, our approach is able to reproduce the spectrum of internal wave incoherent peaks and the continuum of energy down to the buoyancy frequency. The fitted parameters allow for a comparison of internal wave properties between sites, depths, and seasons. The decorrelation timescale, indicative of the extent of the phase shift, exhibited a median value between 3 and 5 days and demonstrated minimal variation across sites and depths. The depth variation for the energy continuum amplitude and the amplitude of the semidiurnal peak exhibited an internal wave mode-1-like structure, particularly at the deeper mooring sites. The greatest amplitudes were observed within the surface mixed layer and thermocline. The slope parameter of the continuum exhibited a median value slightly less than the content slope in Garret-Munk spectral model and demonstrated seasonal variation, with a more rapid decay of energy in the summer compared to winter. The parameters obtained through our method can be further utilised to construct more realistic internal tide boundary conditions using Gaussian processes, thereby enabling more sophisticated modelling of internal waves in coastal regions.
How to cite: Zheng, Y., Rayson, M., Jones, N., and Astfalck, L.: A Machine Learning Parametrisation for the Internal Gravity Wave Spectrum , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-637, https://doi.org/10.5194/egusphere-egu25-637, 2025.
Herein we apply Nonlinear Autoregressive models with exogenous Inputs (NARX) to estimate the interannual variability of satellite-derived chlorophyll-a (CHL) at a global scale, as function of sea-surface height (SSH) from a satellite product provided by Copernicus. A previous analysis shows that SSH is one of the top drivers of CHL in key regions of the tropical and south Atlantic, which is herein corroborated at a global scale, showing a 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 NARX. Herein the NARX model was generated with 10 neurons in the hidden layer, trained with a Levenberg-Marquardt algorithm, and applied to the CHL and SSH monthly composites from Oct 1997 to Sep 2024. A noise level of 0.57 for the model correlations was defined as the 95th percentile of 10,000 NARX-modeled random series. This noise level is exceeded by 97% of the CHL-anomaly series modeled for the 1997-2024 period. The NARX-model successfully reproduces the CHL interannual variability: 59% of the modeled CHL present correlations > 0.90. Then, the NARX-model can be potentially used to predict CHL beyond the training period. In this study’s next stage, the predictability of CHL will be evaluated using SSH for a post-training period, and an ultimate goal for the NARX-model will be a predictability assessment using numerical-model predictions. Thus, the proposed method opens the possibility for reconstruction and prediction not only for CHL but also for other related biogeochemical variables.
How to cite: Rivas, D., Fransner, F., and Keenlyside, N.: Estimation of global satellite-derived chlorophyll-a as function of sea-surface height using shallow neural networks, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-766, https://doi.org/10.5194/egusphere-egu25-766, 2025.
Marine ecosystems are a vital component of the global carbon cycle. Our understanding of the cycle within the ocean relies on a combination of numerical models and satellite observations, which are combined through data assimilation (DA) methods. Here we developed a global ensemble DA system for marine ecosystem prediction using the NEMO-MEDUSA coupled ocean-biogeochemistry model and the Parallel Data Assimilation Framework. Unlike deterministic DA systems, the ensemble approach provides flow-dependent uncertainty estimates, improving the reliability of global marine ecosystem forecasts.
How to cite: Chen, Y. and Partridge, D.: Phytoplankton carbon assimilation in a global ensemble marine ecosystem data assimilation system, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-3630, https://doi.org/10.5194/egusphere-egu25-3630, 2025.
Gas hydrate is an important future alternative marine energy resource to fossil fuels, with the advantages of high energy, large reserves, wide distribution, and shallow burial. Accurate identification of gas hydrate reservoirs and estimation of hydrate saturation are the prerequisites for the development and utilization of gas hydrate resources. This research focuses on the difficult issues of hydrate identification, combined with the multidisciplinary technology of ocean-geology-artificial intelligence (AI). The effective hydrate formation identification technology method is studied and put forward based on the geophysical attributes. The method has been verified in the Dongsha area of the northern South China Sea. This study uses machine learning algorithms to analyze whether the sediment contains gas hydrates. Several commonly used machine learning algorithms are selected, such as random forest, Bagging, AdaBoost, and K-Nearest Neighbor (KNN). These algorithms are used to analyze the data of the P-wave velocity and density with high sensitivity to the change of hydrate. The parameters of different algorithm models are optimized through training, and the identification and classification effects of different algorithm models are compared. Finally, the results show that these algorithms could well distinguish whether there is hydrate in the sediment, among those, the KNN algorithm has a good application. The results show method based on machine learning can improve the identification accuracy of gas hydrate. The identification method of this research provides strong technical support for the subsequent exploration and development of hydrates.
How to cite: Tian, D. and Yang, S.: Identification of gas hydrate based on machine learning in the northern South China Sea , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-5478, https://doi.org/10.5194/egusphere-egu25-5478, 2025.
Since the late 1800s, and especially in the last century, the natural biogeochemical cycle of lead (Pb) in the ocean has been severely perturbed by anthropogenic emissions generated by the use of leaded gasoline, waste incineration, coal combustion and non-ferrous metal smelting. Lead and its isotopes are powerful tools to study the pathways of Pb pollution from land to sea and, simultaneously, investigate biogeochemical processes in the ocean. For these reasons, the study of Pb concentrations and isotope compositions of seawater is a core part of the international marine geochemistry programme GEOTRACES. However, the scarcity and sparsity of in situ measurements of Pb concentrations and isotope compositions do not allow for a comprehensive understanding of Pb pollution pathways and marine biogeochemical cycling on a global scale.
We present here three machine learning models developed to map seawater Pb concentrations and isotope compositions leveraging the global GEOTRACES dataset together with historical data. The models are based on the non-linear regression algorithm XGBoost and use climatologies of oceanographic and atmospheric variables as features from which to predict Pb concentrations, 206Pb/207Pb, and 208Pb/207Pb. Using Shapley Additive Values (SHAP), we found that seawater temperature, atmospheric dust and black carbon, and salinity are the most important features for mapping Pb concentrations. Dissolved oxygen concentration, salinity, temperature, and atmospheric dust are the most important features for mapping 206Pb/207Pb, while atmospheric black carbon and dust, seawater temperature, and surface chlorophyll-a for 208Pb/207Pb. The output of our models shows that (i) the highest levels of pollution are found in the surface Indian Ocean, (ii) pollution from previous decades is sinking in the North Atlantic and Pacific Ocean, and (iii) waters characterised by a highly anthropogenic Pb isotope fingerprint are spreading from the Southern Ocean throughout the Southern Hemisphere at intermediate depths. The analysis of the uncertainty associated with the mapped distribution of Pb concentrations, 206Pb/207Pb, and 208Pb/207Pb suggests that the Southern Ocean is the key area to prioritise in future sampling campaigns.
How to cite: Olivelli, A., Arcucci, R., Rehkämper, M., and van de Flierdt, T.: Mapping the global distribution of lead and its isotopes in seawater with explainable machine learning, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-6069, https://doi.org/10.5194/egusphere-egu25-6069, 2025.
The improvement of remote sensing systems together with the emergence of new model fitting algorithms based on sophisticated methods, such as machine-learning techniques, have allowed the determination of the partial pressure of carbon dioxide (pCO2,sw) in the Canary Islands waters based on mathematical modeling. Among all the fitted models, the most powerful one seems to be the bootstrap aggregation (bagging), giving an RMSE < 6 µatm (R2 > 0.95), although the multilinear regression (MLR), neural network (NN) and categorical boosting (CatBoost) also have a good predictive performance, with RMSE ranging from 9 to 13 µatm for 360 < pCO2,sw < 481 µatm. Using the most reliable model that uses sea surface temperature (SST), Chlorophyll a (Chla), and mixed layer depth (MLD), it was determined that during the period comprised between 2019 and 2024, the Canary basin behaved as a slight net sink of atmospheric CO2, with an average daily flux of -1.45 ± 0.08 mmol m-2 d-1, resulting in the sequestration of -2.59 ± 0.15 Tg CO2 yr-1.
How to cite: González-Dávila, M., Sánchez-Mendoza, I., González-Santana, D., Curbelo-Hernández, D., Estupiñan, D., Suarez de Tangil, M., González, A. G., and Santana-Casiano, J. M.: Modeling pCO2,sw in the Canary Islands region based on satellite measurements, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-8349, https://doi.org/10.5194/egusphere-egu25-8349, 2025.
Shipwrecks have long fascinated people with their stories of mysteries and hidden treasures. UNESCO estimates that more than three million shipwrecks lie undiscovered in the world’s oceans and lakes, yet less than 10% of these have been precisely located. Beyond their historical and archaeological significance, shipwrecks can pose significant environmental threats. Instead of treasures, they often conceal harmful substances like fuels and corroded heavy metals, which, if released, can harm surrounding ecosystems and nearby communities.
This study introduces an innovative artificial intelligence (AI) approach, leveraging convolutional neural networks (CNNs) and open-access remote sensing data, to detect and map shipwrecks in remote coral reefs. The method is designed to identify wrecks based on the environmental footprint they leave, referred to as "Black Reefs", even in cases where the shipwreck itself has completely degraded.
One of the primary challenges was the limited availability of known black reef locations, which restricted the training dataset. To address this, a supervised fully convolutional neural network architecture, called SimpleNet, was employed. This architecture is specifically suited for scenarios with small labelled datasets. From a shortlist of eight suitable reefs (e.g., Kenn, Nikumaroro, Kingman, Kanton, and Rose), five were used for generating training and evaluation data, while the remaining were excluded due to low-resolution imagery or cloud interference.
Image tiles of 256 x 256 x 3 bands were extracted from the training reefs, resulting in approximately 1,600 labelled images. For evaluation, small sections of Kenn and Rose reefs were used to train the model, while other portions served as test datasets. Training was conducted using the IRIDIS supercomputer at the University of Southampton, utilizing 12 CPUs, one node with 264 GB of memory, and MATLAB 9.6 (2019b). The training process took approximately two hours.
The results demonstrate that even with limited training data, the SimpleNet architecture, featuring just eight fully convolutional layers, can efficiently identify and classify black reefs, indicating the presence of shipwrecks. Moreover, the algorithm provides a tool for monitoring reef discoloration and assessing ecological impacts over time through time-series imagery.
This study underscores the potential of AI-driven methods to enhance shipwreck detection and environmental monitoring, offering an efficient, cost-effective solution for tackling the challenges posed by limited ground data and inaccessible regions.
How to cite: Karamitrou, A., Sturt, F., and Bogiatzis, P.: Hidden Wrecks and Black Reefs: Harnessing AI to Unveil Maritime Mysteries and Environmental Risks, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-9298, https://doi.org/10.5194/egusphere-egu25-9298, 2025.
Understanding the response to climate change of the Venice Lagoon is fundamental for the conservation and sustainable management of a vulnerable environment, with important ecological and socio-economic consequences. Deterministic dynamic models that can reproduce the behavior of the lagoon have a very high computational cost, that limits substantially their applicability, particularly considering the multiple and multidecadal simulations required to analyses climate change. This study explores the use of artificial neural networks (ANNs) to model the relationships between climate drivers and key parameters (temperature and salinity) of the Venice lagoon to understand their different dynamics within the lagoon environment. We carry on a sensitivity study on the various drivers utilized and examine the simultaneous presence of different response patterns within the lagoon. The analysis is based on the combination in situ observations of the lagoon water temperature and salinity with large-scale data from the Copernicus Marine Services’ reanalysis to estimate how the main physical parameters of the lagoons are driven by key climatic drivers. The sensitivity analysis was conducted by excluding from the ANN or randomizing single drivers to assess their importance for describing the variability of the lagoon environment. This analysis allow to identify three clusters, defining three areas of the lagoon, whose differences that can be physically interpreted. The riverine cluster (central/northern lagoon) is influenced by the presence of small tributaries and, consequently, by local precipitation; The marine cluster is located in the part of the lagoon near the sea outlets, where salinity and temperature values are strongly influenced by marine salinity and temperature; The mixed cluster (in the south lagoon) where both the marine and riverine regimes overlap with comparable effects on salinity and temperature.
Financial support from ICSC – Centro Nazionale di Ricerca in High Performance Computing, Big Data and Quantum Computing, funded by European Union – NextGenerationEU. Project code CN_00000033, CUP C83C22000560007 and from NBFC – National Biodiversity Future Center, funded by European Union – NextGenerationEU. Project code CN_00000033, CUP F87G22000290001
How to cite: Bozzeda, F., Sigovini, M., and Lionello, P.: Using artificial intelligence for exploring the climatic drivers of the Venice Lagoon environmental variability and response to climate change., EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-9911, https://doi.org/10.5194/egusphere-egu25-9911, 2025.
Measuring sea surface currents (SSC) directly is challenging. Instead, SSC are often inferred from indirect measurements like altimetry. However, altimetry-based methods only provide large-scale (>100 km) geostrophically-balanced velocity estimates of SSC. Here, we present a statistical inversion model to predict fine-scale SSC using remotely sensed sea surface temperature (SST) data. Our approach employs Gaussian Process (GP) regression, where the GP is informed by a two-dimensional tracer transport equation. This method yields a predictive distribution of SSC, from which we can generate an ensemble of surface currents to derive both predictions and prediction uncertainties. Our approach incorporates prior knowledge of the SSC length scales and variances that appear in the covariance function of the GP, which are then estimated from the SST data. The framework naturally handles noisy and incomplete SST data (e.g., due to cloud cover), without the need for pre-filtering. We validate the inversion model through an observing system simulation experiment (OSSE), which demonstrates that GP-based statistical inversion outperforms existing methods, especially when the measurement signal-to-noise ratio is low. When applied to Himawari-9 satellite SST data over the Australian North-West Shelf, our method successfully resolves SSC down to the sub-mesoscale. We anticipate our framework being used to improve understanding of fine-scale ocean dynamics, and to facilitate the coherent propagation of uncertainty into downstream applications such as ocean particle tracking.
How to cite: de Kreij, R., Zammit Mangion, A., Rayson, M., Jones, N., and Zulberti, A.: Statistical inversion of surface tracers to infer fine-scale near-surface ocean currents, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-10550, https://doi.org/10.5194/egusphere-egu25-10550, 2025.
Precise forecasting of sea surface currents is crucial for diverse applications, including navigation, pollution control, and ecosystem monitoring. Traditional high-resolution hydrodynamic models like NEMO generate detailed short-term forecasts but are computationally expensive and resource-intensive. To overcome these limitations, we present sciCUN: a deep learning framework designed for surface current inference using CNN-U-Net architecture.
In summary, sciCUN utilizes the zonal and meridional wind components, mean sea level pressure, air temperature, and dew point temperature from ECMWF Reanalysis v5 (ERA5) for the current day, along with the high-resolution zonal and meridional sea surface current velocity fields from the Copernicus Marine Service Baltic Sea Physics Reanalysis for the previous day, as input features. It then generates the high-resolution zonal and meridional sea surface current velocity fields for the current day.
As a case study, sciCUN was implemented in the Gulf of Riga domain. The model was trained to capture the influence of atmospheric forcing on preceding sea surface currents over a training period spanning 1993 to 2019. Its predictive performance was subsequently validated through a 4-year testing phase (2020–2023). Results showed that while prediction accuracy was slightly lower in coastal regions near river mouths and the Irbe Strait—areas where hydrodynamic models typically employ boundary conditions—sciCUN exhibited strong overall performance. The model achieved an average Euclidean distance of 2.30 cm/s between its predictions and reference data, with an average component-wise mean absolute error of 1.45 cm/s and correlation coefficient of 92.
How to cite: Barzandeh, A., Maljutenko, I., Rikka, S., and Raudsepp, U.: A Surrogate Model for Daily Sea Surface Current Fields Prediction Using CNN-UNET , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-16890, https://doi.org/10.5194/egusphere-egu25-16890, 2025.
Through global warming, ocean deoxygenation is considered as a major concern since it consequently reduces the quality and the quantity of suitable habitats for marine life. Eutrophication plays a major role in its depletion which enhances respiration at different depths. Many species such as fishes, benthic worms or even plankton are affected by this phenomenon.
This study aims to get a better understanding of benthic worm species on the continental shelf of the Black Sea which is well known for high frequency oxic stresses. Our main objective is to map species through their biological traits (i.e. body length, burial depth, reproductive frequency…) in order to assess their vulnerability towards environmental variations that occur at this location.
Unfortunately, in the oceanographic field, one of the major issues is the sparsity of in-situ observations, especially when it comes to benthic biology. Therefore, we have decided to use a multivariate approach allowing us to use related datasets with significantly better spatial and temporal coverage. This multivariate approach is implemented using deep learning in order to get complete maps of traits on our domain. An adapted convolutional neural network allowing to capture non-linearities is used to reconstruct the traits repartitions.
Thus, as an input for the neural network, we consider our traits dataset and environmental variables which are likely to enhance their reconstruction; Surface currents, particulate organic carbon, oxygen concentration and bathymetry are considered. A chosen period from 2008 to 2017 is selected. Traits datasets are located by stations (238) and were constructed through fuzzy coding and rescaled by their biomass.
The neural network architecture is composed of an encoder and a decoder where the encoder considers a gappy and non-gridded dataset. The encoder uses a series of convolutional layers followed by max pooling layers which reduce the size of the dataset. The decoder does essentially the reverse operation by considering convolutional and interpolation layers.
In order to avoid overfitting, the model has skip connections which ensure to keep information from the input dataset. For additional information please refer to Barth et al 2022. The model gives the reconstructed trait repartition and the standard error of the reconstruction.
This study will be helpful in the understanding of benthic traits repartition and will aim to link their patterns to environmental factors. This will help to get a deeper understanding of the ecological role and functions of this poorly known ecosystem. This work is carried in the frame of NECCTON European project.
How to cite: Dechenne, A., Chevalier, S., Gregoire, M., Alvera-Azcarate, A., and Barth, A.: Deriving benthic traits through deep learning methods, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-10653, https://doi.org/10.5194/egusphere-egu25-10653, 2025.
Marine Primary Production (MPP) is a key component in understanding ocean ecosystems and their atmospheric carbon sequestration capacity. However, numerous challenges exist for obtaining MPP estimates. Algorithm variability is a significant issue, since various MPP models (chlorophyll-based or carbon-based algorithms) yield divergent results. Furthermore, the lack of observational data and periodic vertical profiles of the surface ocean hinder the ability to validate and refine such models.
This work focuses on improving MPP estimations by extending the state-of-the-art Carbon, Absorption, and Fluorescence Euphotic-resolving (CAFE) net primary production model with machine learning techniques to overcome current limitations. To improve the model's accessibility and versatility to be extended with data-driven methods, the original C code was rewritten in Python, resulting in a more user-friendly version named PyCAFE [https://github.com/jcrespinesteve/PYCAFE.git]. Using PyCAFE, simulations of MPP from 2003 to 2023 were conducted, producing a comprehensive dataset for training, validation, and testing. First, we train a random forest (RF) model using 500 random locations to emulate PyCAFE and to test global upscaling of MPP estimates. Our results show that the RF model has a strong capability for extrapolating MPP predictions with high accuracy [R2=0.96]. Second, we develop a hybrid model approach to simulate MPP: the HYPE-CAFE model (HYbrid marine Primary production Estimates based on the Carbon, Absorption, and Fluorescence Euphotic-resolving model). HYPE-CAFE combines the physical processes of the PyCAFE model with a neural network predicting the light-use efficiency (LUE), i.e., MPP is calculated as the product of absorbed photons and the predicted LUE. Preliminary results indicate that HYPE-CAFE provides an improvement over the predictions made with the CAFE model alone, especially in regions with variable environmental conditions. However, the lack of observational data limits the learning process. Therefore, in a next step we test a transfer learning approach to improve MPP predictions by HYPE-CAFE.
In conclusion, this project paves the way for the development of advanced hybrid modeling approaches, such as HYPE-CAFE, for global MPP estimation, and offers a transformative avenue for deepening our understanding of global ocean productivity, particularly in the context of climate change.
How to cite: Crespin, J., Benson, V., and Winkler, A. J.: HYPE-CAFE: Towards a Hybrid Model for Improved Marine Primary Production Estimates, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-19488, https://doi.org/10.5194/egusphere-egu25-19488, 2025.
Extreme marine biological events, such as harmful algal blooms and mass mortalities, are increasingly driven by climate variability and anthropogenic pressures, profoundly impacting marine ecosystems. The Black Sea, with its distinct stratification, salinity gradients, and diverse phytoplankton functional groups, is particularly vulnerable to these changes. Understanding and forecasting the interactions between physical, chemical, and biological variables in this region is crucial for effective ecosystem management.
We present a neural network-based surrogate modeling framework to analyze and predict the dynamics of the Black Sea ecosystem. A 3D convolutional encoder-decoder network is trained on simulation data (1950–2014) produced be the University of Liège, including daily basin-scale values of temperature, salinity, nutrients, chlorophyll, and phytoplankton biomass. The model processes time series of spatial maps as input and predicts chlorophyll concentrations and the distributions of phytoplankton functional groups for the subsequent two weeks.
This approach efficiently captures complex interdependencies between variables, offering a computationally efficient alternative to traditional process-based models. By perturbing input variables, the model identifies key drivers of chlorophyll variability, enabling rapid scenario testing to explore the impacts of environmental changes on the ecosystem.
Our findings demonstrate the potential of neural network-based surrogate models to advance understanding of phytoplankton dynamics and support decision-making in marine ecosystem management.
How to cite: Smith, P. A. H., Chauhan, A., Christensen, A., St. John, M., Rodrigues, F., and Mariani, P.: Understanding Drivers of Phytoplankton Variability in the Black Sea Using Convolutional Neural Networks, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-19563, https://doi.org/10.5194/egusphere-egu25-19563, 2025.
This work is dedicated to analysing the simulations of the quasi-geostrophic double-gyre model from dynamical systems point of view to discover nonlinear low-order structures in this turbulent regime. The double-gyre is simulated by a stratified quasi-geostrophic model which is solved using high-resolution CABARET scheme [1]. The statistically stationary simulations of the double-gyre model are considered for 400 years after a 100-year spin-up period. Double-gyre simulations are coarse-grained (symbolized) based on the Taken’s embedding theorem [2] which is proved promising for identifying nonlinear patterns from the stochastic background in the turbulent flow signals. To analyse the coarse-grained time series, Permutation Entropy [3-6] is deployed to quantify repetitive mutual ordering between subsequent time series values using the deviations from uniformity in the distribution of occurrences for symbolic ordinal patterns. Based on permutation entropy analysis, the large-scale double-gyre circulation and its eastward jet demonstrate highly nonlinear behaviour while smaller-scale eddies spread throughout the domain behave linearly. The results of this dynamical system analysis are also compared with data-driven and multi-scale reduced-order models previously developed for this ocean circulation [7,8].
References:
[1] Karabasov, S.A., Berloff, P. S. & Goloviznin, V. M. (2009). CABARET in the ocean gyres, Ocean Modelling, 30(2-3), 155–168.
[2] Takens, F. (2006, October). Detecting strange attractors in turbulence. In Dynamical Systems and Turbulence, Warwick 1980: proceedings of a symposium held at the University of Warwick 1979/80 (pp. 366-381). Berlin, Heidelberg: Springer Berlin Heidelberg.
[3] Bandt, C., & Pompe, B. (2002). Permutation entropy: A natural complexity measure for time series, Physical Review Letters, 88, 174102.
[4] Rosso, O. A., Larrondo, H. A., Martin, M. T., Plastino, A., & Fuentes, M. A. (2007), Distinguishing noise from chaos, Physical Review Letters, 99 (15), 1–5.
[5] Kobayashi, W., Gotoda, H., Kandani, S., Ohmichi, Y., & Matsuyama, S. (2019). Spatiotemporal dynamics of turbulent coaxial jet analyzed by symbolic information-theory quantifiers and complex-network approach, Chaos: An Interdisciplinary Journal of Nonlinear Science, 29 (12), 123110.
[6] Gryazev, V., Riabov, V., Markesteijn, A., Armani, U., Toropov, V., & Karabasov, S. A. (2024). A Dynamical System Method for Finding Flow Structures from Jet LES Data. In 30th AIAA/CEAS Aeroacoustics Conference (2024), 3087.
[7] Naghibi, E., Armani, U., Gryazev, V., Toropov, V., & Karabasov, S., (2024). Reconstruction of the North Atlantic Double-gyre Circulation with Genetic Programming, Springer Proceedings in Mathematics and Statistics, Proceeding of ATSF Conference 2024.
[8] Naghibi, S. E., Karabasov, S. A., Jalali, M. A., & Sadati, S. H. (2019). Fast spectral solutions of the double-gyre problem in a turbulent flow regime. Applied Mathematical Modelling, 66, 745-767.
How to cite: Naghibi, E., Gryazev, V., and Karabasov, S.: A Dynamical System Approach for Finding Nonlinear Flow Structures in Double-gyre Circulation, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-20725, https://doi.org/10.5194/egusphere-egu25-20725, 2025.
Posters virtual: Fri, 2 May, 14:00–15:45 | vPoster spot 2
EGU25-7577 | Posters virtual | VPS30
Understanding Marine Heat Waves in the Chesapeake Bay: Drivers, Variability, and Predictive Insights Using Machine LearningFri, 02 May, 14:00–15:45 (CEST) | vP2.2
Marine heat waves (MHWs) pose significant threats to coastal ecosystems, with particularly severe impacts in shallow waters where their magnitude is often amplified. The Chesapeake Bay, the largest estuary in the United States, is highly vulnerable to these events, which have increased in frequency and duration in recent decades. MHWs in the Chesapeake Bay have critical implications for its ecological balance, including effects on fish populations, habitat degradation, and water quality. Despite their growing prevalence, the underlying causes of these events and the factors regulating their variability remain poorly understood. Our study employs machine learning approaches to elucidate the drivers of marine heat waves in the Chesapeake Bay and to quantify their contributions to these extreme temperature events. By incorporating a comprehensive set of potential predictors, including local air temperature, wind forcing, river discharge, and Atlantic Ocean temperature, the model reveals the key mechanisms driving the onset, intensity, and persistence of MHWs in the Chesapeake Bay. Advanced feature selection techniques isolate the most relevant variables, while model outputs are validated against observed data to ensure accuracy and robustness. Our results suggest that local air temperature and ocean temperature anomalies from the Atlantic Ocean are dominant in triggering MHWs. These findings shed light on the complex interactions between atmospheric, hydrological, and oceanographic processes in shaping extreme thermal events in estuarine systems.
How to cite: Li, C., Xiong, N., and Zhao, J.: Understanding Marine Heat Waves in the Chesapeake Bay: Drivers, Variability, and Predictive Insights Using Machine Learning, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-7577, https://doi.org/10.5194/egusphere-egu25-7577, 2025.
EGU25-9680 | Posters virtual | VPS30
Temporal Fusion Transformers for Improved Coastal Dynamics Forecasting in the Western Black SeaFri, 02 May, 14:00–15:45 (CEST) | vP2.3
The paper investigates the potential of Artificial Intelligence (AI) and Machine Learning (ML) techniques, specifically Temporal Fusion Transformers (TFTs), to enhance the prediction of coastal dynamics along the Western Black Sea coast. We aim to bridge the gap between in-situ observations from five meteo-oceanographic stations and modelled geospatial marine data from the Copernicus Marine Service. TFTs are employed to refine predictions of shallow water dynamics by considering atmospheric influences, focusing on wave-wind correlations. Atmospheric pressure and temperature are treated as latitude-dependent constants, with specific investigations into extreme events like freezing and solar radiation-induced turbulence.
The analysis utilizes a dataset of meteorological information collected by the Maritime Hydrographic Directorate (MHD) since 2015. The study relies on data gathered from seven automated weather stations at lighthouses along the Romanian coastline. The stations, part of the Romanian Navy - Marine Meteorological Surveillance Network, continuously gather meteorological parameters at specific ground-level heights, including wind speed and direction. The Copernicus Marine Service (CMEMS) wave reanalysis dataset for the Black Sea provides a comprehensive record of wave conditions with a spatial resolution of approximately 2.5 km and hourly temporal resolution.
Explainable AI (XAI) is exploited to ensure transparent model interpretations and identify key influential input variables, including static, encoder, and decoder variables. Data attribution strategies address missing data concerns, while ensemble modelling enhances overall prediction robustness. The models demonstrate a significant improvement in prediction accuracy compared to traditional methods. This research provides a deeper understanding of atmosphere-marine interactions and demonstrates the efficacy of AI/ML in bridging observational and modelled data gaps for maritime safety and coastal management along the Western Black Sea coast.
Acknowledgements: The research of the M.E.M., P.P., M.G.I., and L.D. was conducted as part of the "Forecasting and observing the open-to-coastal ocean for Copernicus users" FOCCUS Project (https://foccus-project.eu/), funded by the European Union (Grant Agreement No. 101133911). Views and opinions expressed are, however, those of the author(s) only and do not necessarily reflect those of the European Union or the European Health and Digital Executive Agency (HaDEA). Neither the European Union nor the granting authority can be held responsible for them.
The presented results of the M.E.M., P.P., M.G.I., and L.D. have been carried out with financial support from the Sectorial Research-Development Plan of the Romanian Ministry of National Defence, PSCD 2021–2024 Project (097/2021, 092/2022, 097/2023, 097/2024): „Development of an integrated monitoring system to increase the quality of hydro-oceanographic data in the area of responsibility of the Romanian Naval Forces".
Thanks are extended to the relevant departments of INOE-2000 for their help through the "Core Program with the National Research Development and Innovation Plan 2022-2027" with the support of MCID, project no. PN 23 05/2023, contract 11N/2023.
How to cite: Mihailov, M. E., Ichim, M. G., Chirosca, A. V., Chirosca, G., Dutu, L., and Popov, P.: Temporal Fusion Transformers for Improved Coastal Dynamics Forecasting in the Western Black Sea , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-9680, https://doi.org/10.5194/egusphere-egu25-9680, 2025.
EGU25-14016 | ECS | Posters virtual | VPS30
Predicting GHG Emissions in Shipping: A Case Study Of CanadaFri, 02 May, 14:00–15:45 (CEST) | vP2.4
Shipping remains a crucial element of global trade and commerce, facilitating over 90% of international trade by volume. The maritime industry’s advanced logistics chains are vital for the timely delivery of goods, supporting both economic growth and employment. However, it is also a significant source of pollution, accounting for approximately 3% of global greenhouse gas (GHG) emissions, and contributing 13% of nitrogen oxides (NOx) and 12% of sulfur oxides (SOx). Additionally, shipping emits harmful pollutants, including particulate matter (PM), black carbon (BC), and methane (CH4). These emissions not only impact the global climate but also pose severe health risks to communities near shorelines, contributing to asthma, respiratory and cardiovascular diseases, lung cancer, and premature death.
The International Maritime Organization (IMO) is actively engaged in mitigating these environmental impacts as part of its support for the UN Sustainable Development Goal 13, which addresses climate change in alignment with the 2015 Paris Agreement. The IMO has implemented several regulations to curb GHG emissions from shipping, beginning with mandatory energy efficiency measures introduced on July 15, 2011. Subsequent regulations include the Initial IMO GHG Strategy (2018) and the updated Strategy on Reduction of GHG Emissions from Ships (2023). The 2023 strategy sets ambitious targets to achieve near-zero GHG emissions from international shipping by around 2050, with interim goals of reducing emissions by at least 20% by 2030 and 70-80% by 2040. It also aims to cut the carbon intensity of international shipping by at least 40% by 2030, measured as CO2 emissions per unit of transport work. As of January 1, 2023, ships are required to calculate their Energy Efficiency Existing Ship Index (EEXI) and establish an annual operational Carbon Intensity Indicator (CII), with ratings from A to E indicating energy efficiency (International Maritime Organization).
In response to evolving regulations aimed at reducing GHG emissions, we propose a machine learning framework to improve emission predictions, with a particular focus on the Saint Lawrence River. Currently, emissions in the Canadian shipping sector are calculated a posteriori, with Environment and Climate Change Canada (ECCC) providing a national marine emissions inventory and a comprehensive visualization tool. This tool enables users to analyze shipping activities and emissions across Canada by filtering data through various parameters.
Our proposed work is designed to predict GHG emissions for vessels navigating the Saint Lawrence River, with plans for broader application across Canada. By employing a bottom-up methodology, we create a detailed emissions inventory based on individual vessel activities, leveraging Automatic Identification System (AIS) data to capture the spatiotemporal dynamics of shipping (Spire). To enhance accuracy, we incorporate vessel-specific information from CLARKSONS, including engine type, fuel type, and power, along with meteorological data such as current speed to account for external factors affecting emissions. Machine learning models, particularly deep learning techniques, are employed in the prediction phase, enabling the model to continually improve with new data. This scalable approach not only enhances environmental monitoring but also supports national efforts to reduce GHG emissions from marine transportation across Canada.
How to cite: El aissi, A. and Benabbou, L.: Predicting GHG Emissions in Shipping: A Case Study Of Canada, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-14016, https://doi.org/10.5194/egusphere-egu25-14016, 2025.
EGU25-14039 | ECS | Posters virtual | VPS30
A Hybrid Machine Learning Model For Ship Speed Through Water: Solve And PredictFri, 02 May, 14:00–15:45 (CEST) | vP2.5
The maritime transport industry faces a significant challenge: reducing its greenhouse gas (GHG) emissions by 50% compared to 2008 levels. A crucial factor in calculating and optimizing these emissions is accurately predicting ship speed through water. While various models exist, few effectively combine both physical principles and machine learning approaches, leading to limitations in prediction accuracy.
The paper proposes a hybrid model with two main components: ''Solve'' Component: A physics-based approach that uses a Physics-Informed Neural Network (PINN) to determine the theoretical speed a ship would achieve in calm water conditions, based on fundamental physical principles and equations. ''Predict'' Component: A data-driven approach that takes the theoretical calm water speed and adjusts it based on real-world conditions using machine learning algorithms, producing actual speed predictions.
The Solve Phase centers around a differential equation relating three key parameters: propulsion power (P), draft (T), and speed through calm water (Vw), the equation takes the form:
The model uses a PINN to solve a differential equation that links propulsion power (P), draft (T), and calm water speed (Vw) to generate initial speed estimates. The PINN uses a loss function that incorporates both initial conditions and differential equation residuals. A major challenge arises because Vw is theoretical and cannot be directly measured. This issue is addressed using historical data by identifying periods when sea conditions were calm to use as training data.
The model creates a bridge between its solve and predict phases. In the first approach, focused on training data generation, the system utilizes the trained PINN to generate collocation points. From these points, it creates training triplets consisting of propulsion power (Pi), draft (Ti), and calm water speed (Vwi). This approach uses a straightforward mean squared error loss function to train the neural network. The second approach takes a different path by using propulsion power (P) and draft (T) as direct inputs to the neural network. What makes this approach unique is that it incorporates the PINN directly into the loss function. This integration allows physical principles from the differential equation to directly influence the predictions, creating a stronger connection between the physical model and the machine learning component.
The predict phase begins by taking the calm water speed predictions generated from the solve phase and enhances them by incorporating various real-world factors that affect ship movement. These factors include maritime conditions, meteorological data, and current conditions, providing a comprehensive view of the actual sailing environment. To process this combined data, we use machine learning algorithms such as Xgboost. The final output of this phase is the real speed through water (Vwr), which represents a more realistic prediction that accounts for all environmental factors affecting the ship's speed.
The model offers a groundbreaking approach to maritime speed prediction by generalizing across vessel types and integrating physical principles with machine learning. By incorporating operational and meteorological data, it provides more accurate speed predictions that optimize fuel consumption and support the maritime industry's greenhouse gas emission reduction goals, bridging environmental protection with operational efficiency.
How to cite: Atanane, A., Elmimouni, Z., and Benabbou, L.: A Hybrid Machine Learning Model For Ship Speed Through Water: Solve And Predict, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-14039, https://doi.org/10.5194/egusphere-egu25-14039, 2025.
