Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Data-driven ensemble prediction of the global ocean

Qiusheng Huang, Xiaohui Zhong, Anboyu Guo, Ziyi Peng, Lei Chen, Hao Li

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.

Data-driven ensemble prediction of the global ocean

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.
Paper Structure (14 sections, 10 equations, 5 figures)

This paper contains 14 sections, 10 equations, 5 figures.

Table of Contents

  1. Introduction
  2. Results
  3. Overall ocean forecast skill
  4. Vertical structure of forecast improvement
  5. Comparison with numerical ensemble models
  6. Discussion
  7. Methods
  8. Data
  9. Problem formulation and model overview
  10. Structured noise generation
  11. Atmospheric conditioning
  12. Forecasting backbone and Auxiliary conditioning
  13. Prediction loss and training objective
  14. Evaluation metrics

Figures (5)

  • 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 days. Rows show the continuous ranked probability score (CRPS), spread-skill ratio (SSR), root mean square error (RMSE), and anomaly correlation coefficient (ACC). CRPS and SSR are reported only for the two ensemble methods, FuXi-ONS and FuXi-Aim-Perlin. RMSE and ACC further include the deterministic baseline FuXi-Aim and persistence. Lower values indicate better performance for CRPS and RMSE, whereas higher values indicate better performance for SSR and ACC. FuXi-ONS consistently provides the best overall probabilistic performance and the strongest deterministic performance of the ensemble mean across most variables and lead times.
  • 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 times and depth levels. For CRPS and SSR, the normalized improvement is computed relative to FuXi-Aim-Perlin; for RMSE and ACC, it is computed relative to FuXi-Aim.
  • Figure 3: Niño3.4 forecasts for the 2021-02 initialization. Niño3.4 index forecasts from GLORY, NMME, IRI-D, IRI-ALL, and FuXi-ONS as a function of lead season. IRI-D uses only dynamical models, whereas IRI-ALL combines dynamical and statistical models. FuXi-ONS w1--w6 denote forecasts initialized every 5 days within the month, specifically 1 February, 6 February, 11 February, 16 February, 21 February, and 26 February 2021, and the FuXi-ONS mean denotes the average over these six forecasts. The horizontal axis is expressed in overlapping 3-month seasons, where each value is the mean over the labeled season, for example FMA denotes the average of February, March, and April.
  • Figure 4: Comparison of FuXi-ONS with NMME for sea-surface temperature forecasting. RMSE, ACC, CRPS, and SSR of monthly SST forecasts as a function of lead time. The blue solid line denotes FuXi-ONS, and the red line with circle markers denotes NMME. All metrics are averaged over forecasts initialized in 2021.
  • Figure 5: Architecture of the FuXi-ONS ensemble forecasting framework. (a) Structured-noise generation module. Given the current ocean state $X_t$, the model predicts state-dependent perturbation statistics and samples spatially correlated noise through an SPDE-based Matérn formulation, yielding the structured perturbation field $N_t^{(m)}$. (b) Forecast module. The initial atmospheric state $A_{\mathrm{init}}$ and auxiliary input $C$ are encoded into latent conditioning features $F_{\mathrm{air}}$ and $F_c$. The perturbed ocean state $X_t + N_t^{(m)}$ is then mapped to the next-step forecast $\hat{X}_{t+1}^{(m)}$. Multiple ensemble members are generated by resampling $N_t^{(m)}$ at each forecast step.