Accurate reconstruction of upper-ocean temperature profiles is vital for climate studies, marine engineering, and operational forecasting. However, traditional approaches based on numerical ocean models and data assimilation systems, such as reanalysis products, require substantial computational resources and direct access to subsurface observations, limiting their applicability in resource-constrained or real-time operational contexts. This study introduces an AI-based approach that reconstructs global temperature profiles down to 200 m depth using only ERA5-derived surface and atmospheric forcing features, including meteorological variables and energy balance components with temporal sequences spanning the preceding 6 days. The AI model has been trained with Continuous Ranked Probability Score (CRPS) loss function, allowing to output members of a non-parametric ensemble, enabling reliable uncertainty estimates alongside point reconstructions. Performance was evaluated basin-wise across the Mediterranean Sea, North Atlantic, South Atlantic, North Pacific, South Pacific, and Indian Ocean, and compared against daily-mean profiles from Copernicus CMEMS GLORYS12V1 reanalysis. The errors and metrics statistics of the AI model are comparable to those of GLORYS12V1 in the whole world. Although the reanalysis product generally exhibits slightly lower errors as a result of its assimilation of subsurface observations, the AI approach delivers strong and reliable performance and, in certain basins, achieves equivalent or better accuracy. Furthermore, feature sensitivity analysis confirms thermodynamic boundary conditions and large-scale geographical context as the primary drivers in shaping the upper-ocean thermal structure. The aim of this work is to explore the capability of AI to learn complex meteocean interactions, by illustrating a method which can offer significant advantages in computational cost, inference speed, and scalability, making it well-suited for integration into near-real-time forecasting workflows and resource-constrained applications, thereby supporting more accessible and frequent ocean state estimation.
Mattia Cavaiola (Thu,) studied this question.