GLONET: Mercator's end-to-end neural Global Ocean forecasting system

TL;DR

GLONET introduces a physics-informed, operator-learning global ocean forecasting system that fuses large-scale Fourier-based dynamics with local CNN refinements within an encoder–decoder latent framework. Trained on GLORYS12 data and evaluated against GLO12 and Xihe using a comprehensive, process-aware validation suite (including CLASS4-based observation benchmarks, derived quantities, and Lagrangian analyses), GLONET demonstrates competitive performance for currents, SSH, and SLA while preserving dynamical coherence across scales. The study highlights the importance of multi-scale fusion, temporal encoding, and rigorous, process-oriented evaluation to assess data-driven ocean forecasts, revealing both strengths and remaining challenges in SST prediction and deep-ocean consistency. Operationally, GLONET offers rapid 10-day forecasts (less than 10 seconds) on a multi-GPU pipeline and provides daily products via the EDITO platform, underscoring its potential as a scalable, near-real-time forecasting tool for oceanography and marine operations.

Abstract

Accurate ocean forecasting is crucial in different areas ranging from science to decision making. Recent advancements in data-driven models have shown significant promise, particularly in weather forecasting community, but yet no data-driven approaches have matched the accuracy and the scalability of traditional global ocean forecasting systems that rely on physics-driven numerical models and can be very computationally expensive, depending on their spatial resolution or complexity. Here, we introduce GLONET, a global ocean neural network-based forecasting system, developed by Mercator Ocean International. GLONET is trained on the global Mercator Ocean physical reanalysis GLORYS12 to integrate physics-based principles through neural operators and networks, which dynamically capture local-global interactions within a unified, scalable framework, ensuring high small-scale accuracy and efficient dynamics. GLONET's performance is assessed and benchmarked against two other forecasting systems: the global Mercator Ocean analysis and forecasting 1/12 high-resolution physical system GLO12 and a recent neural-based system also trained from GLORYS12. A series of comprehensive validation metrics is proposed, specifically tailored for neural network-based ocean forecasting systems, which extend beyond traditional point-wise error assessments that can introduce bias towards neural networks optimized primarily to minimize such metrics. The preliminary evaluation of GLONET shows promising results, for temperature, sea surface height, salinity and ocean currents. GLONET's experimental daily forecast are accessible through the European Digital Twin Ocean platform EDITO.

Figures (19)

• Figure 1: Overview of GLONET's architecture containing different modules, particularly time-block designed to learn feature maps encapsulating initial conditions along with forcings. A spatial module architectured to learn multi-scale dynamics, and finally and encoder-decoder to fuse multi-scale circulations into a unified latent space.
• Figure 2: Dispersion and evolution of RMSD as a function of forecast lead time for salinity (top left) and temperature (top middle) in the 5–100 m layer, SLA (top right), and SST/(Temperature of the first level for models) (bottom left) from drifting buoys and zonal and meridional ocean currents (middle of bottom right). The thick lines represent the $75\%$ distribution, while the thin lines correspond to the $95\%$ distribution, the dot represent the median of the distribution
• Figure 3: Normalized MAE difference for SST/(Temperature of the first level for models) and zonal current between GLO12 and GLONET (left panels) and XIHE (righ panels). Positive values (red) denote improvement relative to GLO12, while negative values (blue) indicate degradation.
• Figure 4: RMSE computed at each depth level and averaged over all lead times for 3D variables (U and V currents, temperature, and salinity) for the GLONET (orange line) and Xihe (blue line) models, covering January to July 2024. GLORYS12 serves as the reference, with 10-day forecasts initialized weekly on Wednesdays from a nowcast analysis performed with GLO12 seven days behind real-time.
• Figure 5: RMSE averaged across common depth levels of 3D variables (U and V currents, temperature, salinity), along with RMSE of sea surface height (SSH) for the GLONET and Xihe models. Calculations span from January to July 2024, using GLORYS12 as the reference. Forecasts are initialized weekly on Wednesdays, following the operational protocol.