South China Sea "神针(Trident)": Solution for ocean forecast foundation models

Artificial intelligence (AI) has rapidly advanced short-term forecasting in the Earth system, driving progress in the development of both ocean and meteorological foundation models. Meteorological foundation models lead the way and have already been incorporated into global operational weather forecasting systems. The Pangu global model, trained on the ECMWF Reanalysis v.5 (ERA5) dataset at a 25-kilometer (km) resolution, which can resolve major large- to meso- and small-scale weather systems, has demonstrated competitive forecast skill relative to traditional numerical weather prediction systems in selected benchmarks.1 Moreover, the ECMWF integrates its Artificial Intelligence Forecasting System (AIFS) into operational forecasting.2 The development of AI-driven meteorological foundation models provides rapid and accurate weather forecasting, supporting many socio-economic aspects such as energy, insurance, and shipping. Motivated by the success of meteorological foundation models, substantial efforts are now underway to develop global ocean foundation models. Using global ocean reanalysis datasets, nascent ocean foundation models such as Xihe,3 Wenhai,4 and Langya have been developed. These global ocean foundation models, trained at a higher computational resolution (such as GLORYS12 at 1/12° or approximately 10-km resolution) than those for the atmosphere, are reported to well forecast large- to mesoscale ocean dynamic processes. Despite the use of higher-resolution training datasets, it is still challenging to represent and forecast sub-mesoscale to small-scale ocean processes in ocean foundation models. Here, we examine the causes of this limitation and propose a practical pathway for developing AI-driven regional ocean foundation models, using the newly developed South China Sea (SCS) “神针(Trident)” foundation model as an example.

The motivation for developing Trident is seeking a pathway to develop an ocean foundation model distinct from meteorological models, driven by two fundamental considerations: (1) a much higher-resolution training dataset is required to resolve sub-mesoscale to small-scale ocean dynamics and (2) the scarcity of available ocean observations compared with those of the atmosphere. In terms of dynamical scales, the atmospheric Rossby deformation radius spans hundreds to over 1,000 km, whereas the ocean Rossby deformation radius is much smaller, typically only 10–50 km. A 25-km-resolution ERA5 reanalysis dataset is therefore sufficient to resolve large- to small-scale major weather systems reasonably well for the training of meteorology foundational models. However, even at a 10-km resolution, ocean numerical models can resolve large- and mesoscale processes well but struggle with sub-mesoscale and small-scale processes such as sub-mesoscale eddies, internal waves, fronts, and ocean responses to transient weather systems. Resolving these processes requires model resolutions at a km or even sub-km scale, far beyond the ability of available global ocean reanalysis datasets. It should also be noted that such high spatial resolution requires correspondingly high temporal resolution to adequately capture the rapid evolution of mesoscale, sub-mesoscale, and small-scale ocean processes. Many of these processes evolve on timescales of hours to days, and insufficient temporal sampling can lead to phase errors or omission of transient features. In terms of observations, atmospheric observations benefit from a mature, dense, and global three-dimensional network comprising satellites, radar, and ground stations, enabling continuous monitoring of multi-scale weather systems. In contrast, ocean observations are primarily limited to the sea surface from satellites. Subsurface layers depend heavily on limited data collected from moorings and profiling floats, leaving large areas of the ocean interior unsampled. Therefore, the relatively coarse reanalysis data and sparse observations hinder the development, improvement, and operational application of ocean foundation models to a much greater extent than those in meteorology.