Leveraging an Atmospheric Foundational Model for Subregional Sea Surface Temperature Forecasting
Víctor Medina Centro de Tecnologías de la Imagen (CTIM) Instituto Universitario de Cibernética, Empresas y Sociedad (IUCES) University of Las Palmas de Gran Canaria, Spain [email protected] Giovanny A. Cuervo-Londoño Oceanografía Física y Geofísica Aplicada (OFYGA) Instituto Universitari · 2025
Read Leveraging an Atmospheric Foundational Model for Subregional Sea Surface Temperature Forecasting by Víctor Medina Centro de Tecnologías de la Imagen (CTIM) Instituto Universitario de Cibernética, Empresas y Sociedad (IUCES) University of Las Palmas de Gran Canaria, Spain [email protected] Giovanny A. Cuervo-Londoño Oceanografía Física y Geofísica Aplicada (OFYGA) Instituto Universitari (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-25563.txt.
Introduction
Victor Medina, Giovanny A Cuervo-Londono, and Javier Sanchez forecast sea surface temperature. Their 2025 paper adapts an atmospheric foundation model. Classic numeric forecasting faces computational cost limits. Scalability is another limitation of numeric methods.
Fine-tuning uses high-resolution oceanographic reanalysis data. The target region is the Canary Upwelling System. Staged fine-tuning incorporates latitude-weighted error metrics. Hyperparameters were optimized for quick learning during tuning. The model reached low root-mean-square error overall. Anomaly correlation stayed very high in tests. Finer coastal details proved harder to capture accurately. Data-driven ocean forecasting feasibility is showed here. Models pretrained in other domains can transfer. Future work may add variables and physics-informed networks.
Catalog record
| CREATOR | Víctor Medina Centro de Tecnologías de la Imagen (CTIM) Instituto Universitario de Cibernética, Empresas y Sociedad (IUCES) University of Las Palmas de Gran Canaria, Spain [email protected] Giovanny A. Cuervo-Londoño Oceanografía Física y Geofísica Aplicada (OFYGA) Instituto Universitari |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 49 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2510-25563
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Leveraging an Atmospheric Foundational Model for Subregional Sea Surface Temperature Forecasting Víctor Medina Centro de Tecnologías de la Imagen (CTIM) Instituto Universitario de Cibernética, Empresas y Sociedad (IUCES) University of Las Palmas de Gran Canaria, Spain [email protected] Giovanny A. Cuervo-Londoño Oceanografía Física y Geofísica Aplicada (OFYGA) Instituto Universitario de Investigación en Acuicultura Sostenible y Ecosistemas Marinos (ECOAQUA) University of Las Palmas de Gran Canaria, Spain [email protected] Javier Sánchez Centro de Tecnologías de la Imagen (CTIM) Instituto Universitario de Cibernética, Empresas y Sociedad (IUCES) University of Las Palmas de Gran Canaria, Spain [email protected] Abstract The accurate prediction of oceanographic variables is crucial for understanding climate change, managing marine resources, and optimizing maritime activities. Traditional ocean forecasting relies on numerical models; however, these approaches face limitations in terms of computational cost and scalability. In this study, we adapt Aurora, a foundational deep learning model originally designed for atmospheric forecasting, to predict sea surface temperature (SST) in the Canary Upwelling System. By fine-tuning this model with high-resolution oceanographic reanalysis data, we demonstrate its ability to capture complex spatiotemporal patterns while reducing computational demands. Our methodology involves a staged fine-tuning process, incorporating latitude-weighted error metrics and optimizing hyperparameters for efficient learning. The experimental results show that the model achieves a low RMSE of 0.119K, maintaining high anomaly correlation coefficients (ACC ≈ 0.997 \approx 0.997 ). The model successfully reproduces large-scale SST structures but faces challenges in capturing finer details in coastal regions. This work contributes to the field of data-driven ocean forecasting by demonstrating the feasibility of using deep learning models pre-trained in different domains for oceanic applications. Future improvements include integrating additional oceanographic variables, increasing spatial resolution, and exploring physics-informed neural networks to enhance interpretability and understanding. These advancements can improve climate modeling and ocean prediction accuracy, supporting decision-making in environmental and economic sectors. K eywords Forecasting \cdot Sea Surface Temperature \cdot Upwelling System \cdot Oceanography \cdot Foundational Model \cdot Deep Learning 1 Introduction The ocean plays a crucial role in multiple aspects of the planet, from regulating the global climate [ 1 ] to the availability of strategic marine resources [ 2 ] . Understanding its processes and predicting changes is essential for decision-making in the fishing industry and maritime transport sectors [ 3 ] . Anticipating the behavior of oceanographic variables is important in the current scenario [ 4 ] , marked by climate change and the urgent need for sustainable policies [ 5 ] , including the development of new tools to assess the present and future states of the Earth, as current approaches are still limited [ 6 ] . Ocean analysis is critical since it influences global climate, storm formation, and heat exchange with the atmosphere through sea surface temperature [ 7 ] . Additionally, marine biodiversity depends on the dynamic balance of the ocean’s physical and chemical properties. Human activities also require information about ocean conditions to ensure safety, optimize maritime routes, manage energy resources, and implement ecological policies [ 8 ] . Traditional ocean prediction methodologies rely on numerical models that solve complex fluid dynamics equations. However, these models struggle to capture data across multiple spatial and temporal scales and have high computational costs. These limitations, combined with the increasing volume of data, create new research opportunities. In this context, deep learning has emerged as a promising alternative to overcome these challenges, enabling the processing of large volumes of data across different scales. Furthermore, with the growing quantity and diversity of data from in situ sources, satellites, and reanalysis, we now have an opportunity to address these challenges [ 9 ] . These new models learn complex relationships from data without explicitly relying on physical equations. By removing the strict dependence on theoretical formulations, it becomes feasible to capture patterns that emerge from the interaction of multiple variables. This approach can provide more detailed solutions, reduce computational costs, and integrate heterogeneous data. Models based on deep learning could enable the development of more efficient methods in various scenarios [ 10 ] . The field of atmospheric prediction has witnessed a transformative shift with the advent of deep learning models. Recent breakthroughs include models like FourCastNet [ 11 ] , known for its speed and accuracy in predicting extreme weather events, and GraphCast [ 12 ] , which demonstrates remarkable skill in medium-range global weather forecasting based on Graph Neural Networks. Furthermore, models such as NeuralGCM [ 13 ] explore the integration of neural networks within general circulation models, aiming to enhance the representation of sub-grid scale processes. More recently, the Aurora [ 14 ] model has emerged, showcasing the potential of deep learning to not only predict atmospheric conditions but also to extend these capabilities to coupled systems by learning complex interactions across different Earth system components. These advancements signify a new era in weather and climate modeling, promising more efficient and accurate predictions for a range of applications. This work aims to adapt the Aurora foundational model, initially developed for global atmospheric applications, to subregional ocean prediction. The model is fine-tuned for predicting the evolution of the SST in the Canary Upwelling System. We are interested in adapting the model for predicting the ocean’s potential temperature ( \theta_{0} ) in a specific geographic area and period. Its three-dimensional architecture enables the efficient integration of complex and heterogeneous data. Additionally, its reduced training time, ability to leverage previously learned representations, and flexibility to incorporate data from various sources make it an ideal candidate for tackling current challenges in ocean prediction. We use data from the GLORYS12V1 [ 15 ] ocean prediction system of the Copernicus Marine Service [ 16 ] . This corresponds to the global ocean physical reanalysis, which provides three-dimensional daily fields with a horizontal resolution of 1/12° and 50 vertical levels. This dataset, based on the assimilation of satellite and in situ observations, provides information on potential temperature, salinity, currents, and two-dimensional variables, such as sea level and mixed-layer thickness. This model offers advantages in terms of computational cost reduction and prediction quality, although limitations exist due to the complexity of the domain and physical constraints. This should be seen as complementary to other methodologies rather than a complete replacement. Section 2 summarizes state-of-the-art works. Section 3 details the dataset used in this work and the area of study. Section 4 explains the Aurora model and how we adapted it to our problem. The experimental configuration in Section 5 addresses various details about the training process, the loss function, and the metrics used to evaluate the model. Section 6 tackles the fine-tuning process and shows some preliminary results. Finally, the conclusions in Section 8 summarize the main ideas and contributions of the work and propose some ideas for future research. 2 Related Work Prediction in oceanography is essential for anticipating and understanding natural phenomena that affect global climate and various human activities [ 17 ] . Sea temperature plays a fundamental role in the distribution of water masses, global circulation, and the ocean-atmosphere energy balance [ 18 , 19 ] . Thus, estimating changes in the physical and chemical properties of the ocean, such as temperature, is crucial for multiple activities [ 20 ] , in addition to being key to anticipating natural