[Paper Review] Data-driven ensemble prediction of the global ocean

Qiusheng Huang, Xiaohui Zhong+4 more|arXiv (Cornell University)|Mar 20, 2026

Oceanographic and Atmospheric Processes0 citations

TL;DR

FuXi-ONS is the first machine-learning ensemble forecasting system for the global ocean, producing probabilistic 5-day to 365-day forecasts on a 1° grid for multiple variables, with a learned, physically structured perturbation mechanism and atmospheric conditioning for stability and efficiency.

ABSTRACT

Data-driven models have advanced deterministic ocean forecasting, but extending machine learning to probabilistic global ocean prediction remains an open challenge. Here we introduce FuXi-ONS, the first machine-learning ensemble forecasting system for the global ocean, providing 5-day forecasts on a global 1° grid up to 365 days for sea-surface temperature, sea-surface height, subsurface temperature, salinity and ocean currents. Rather than relying on repeated integration of computationally expensive numerical models, FuXi-ONS learns physically structured perturbations and incorporates an atmospheric encoding module to stabilize long-range forecasts. Evaluated against GLORYS12 reanalysis, FuXi-ONS improves both ensemble-mean skill and probabilistic forecast quality relative to deterministic and noise-perturbed baselines, and shows competitive performance against established seasonal forecast references for SST and Niño3.4 variability, while running orders of magnitude faster than conventional ensemble systems. These results provide a strong example of machine learning advancing a core problem in ocean science, and establish a practical path toward efficient probabilistic ocean forecasting and climate risk assessment.

Motivation & Objective

• Extend machine learning from deterministic ocean forecasting to probabilistic ensemble prediction for the global ocean.
• Demonstrate calibrated and physically meaningful forecast spread across variables and depths for subseasonal to annual lead times.
• Achieve substantial computational efficiency relative to conventional numerical ensemble systems, enabling practical probabilistic forecasting and risk assessment.
• Incorporate atmospheric conditioning to stabilize long-range forecasts while preserving large-scale predictability.

Proposed method

• Three-part architecture: a structured-noise generation module, an atmospheric conditioning module, and an ocean forecasting module.
• Structured perturbations are generated via state-dependent amplitudes and an SPDE-based Matérn spectral sampler to produce spatially coherent, flow-dependent noise.
• Atmospheric conditioning encodes initial atmospheric state into latent features to stabilize forecasts over long lead times.
• Forecasts are produced autoregressively at 5-day intervals up to 365 days, with ensemble members formed by resampling structured perturbations at each step.
• Training uses GLORYS12 reanalysis as both input and target, with ERA5 atmospheric conditioning, and a 1° × 1° global grid.
• Evaluation compares FuXi-ONS to deterministic and noise-perturbed baselines as well as established seasonal references (e.g., NMME, IRI-D, IRI-ALL).

Figure 1: Depth-averaged forecast skill Comparison of FuXi-ONS with baselines over all lead times. Forecast performance over the 2021–2023 test period for salinity (S), temperature (ST), zonal current (SU), meridional current (SV), and sea surface height (SSH) as a function of lead time up to 360 da

Experimental results

Research questions

• RQ1 Can a data-driven approach generate a calibrated, probabilistic ensemble forecast for the global ocean at subseasonal to annual horizons?
• RQ2 How does a learned ensemble compare to deterministic forecasts and externally imposed perturbations in terms of skill, spread, and physical realism across variables and depths?
• RQ3 Is the data-driven ensemble both competitive with traditional numerical ensembles and computationally efficient for operational-like use?
• RQ4 What is the vertical (depth-dependent) structure of forecast improvements and uncertainty representation in a global ocean setting?

Key findings

• FuXi-ONS provides the best overall probabilistic performance (CRPS, SSR) and strongest deterministic performance among learned baselines across salinity, temperature, zonal and meridional currents, and sea-surface height.
• FuXi-ONS outperforms a deterministic backbone (FuXi-Aim) and a Perlin-perturbed ensemble (FuXi-Aim-Perlin), with the largest gains in long-range forecasts and a coherent vertical structure of improvement.
• The ensemble spread is more calibrated and flow-dependent than Perlin baselines, with substantial advantages in the middle/deeper ocean layers, especially for current fields.
• Compared with established seasonal references, FuXi-ONS shows competitive or superior performance for Niño3.4 and SST fields, including better RMSE, CRPS, ACC, and SSR in SST forecasts and ENSO evolution.
• FuXi-ONS delivers these probabilistic forecasts orders of magnitude faster than conventional ensemble systems, illustrating a practical path toward operational AI-driven ocean ensemble prediction.

Figure 2: Depth-dependent normalized improvement of FuXi-ONS as a function of forecast lead time. Columns correspond to salinity (S), temperature (ST), zonal current (SU), and meridional current (SV). Rows show the relative changes in CRPS, SSR, RMSE, and ACC, respectively, across forecast lead time