FuXi-Ocean: A Global Ocean Forecasting System with Sub-Daily Resolution

TL;DR

FuXi-Ocean addresses the demand for high-resolution sub-daily global ocean forecasts by introducing a data-driven autoregressive model with a Mixture-of-Time module that adaptively fuses multiple temporal contexts. The architecture captures variable-specific multiscale temporal dynamics and mitigates error accumulation, achieving six-hour forecasts at eddy-resolving $1/12^\circ$ resolution down to $0$–$1500$ m with data-efficient training (~$9$ years). Empirical results on HYCOM-RD and IV-TT demonstrate superior SST and broader variable performance compared with traditional numerical models, while maintaining reasonable computational efficiency. This work paves the way for practical, high-frequency ocean forecasting at global scales, with potential benefits for maritime safety, environmental monitoring, and resource management.

Abstract

Accurate, high-resolution ocean forecasting is crucial for maritime operations and environmental monitoring. While traditional numerical models are capable of producing sub-daily, eddy-resolving forecasts, they are computationally intensive and face challenges in maintaining accuracy at fine spatial and temporal scales. In contrast, recent data-driven approaches offer improved computational efficiency and emerging potential, yet typically operate at daily resolution and struggle with sub-daily predictions due to error accumulation over time. We introduce FuXi-Ocean, the first data-driven global ocean forecasting model achieving six-hourly predictions at eddy-resolving 1/12° spatial resolution, reaching depths of up to 1500 meters. The model architecture integrates a context-aware feature extraction module with a predictive network employing stacked attention blocks. The core innovation is the Mixture-of-Time (MoT) module, which adaptively integrates predictions from multiple temporal contexts by learning variable-specific reliability , mitigating cumulative errors in sequential forecasting. Through comprehensive experimental evaluation, FuXi-Ocean demonstrates superior skill in predicting key variables, including temperature, salinity, and currents, across multiple depths.

Figures (10)

• Figure 1: Architecture of our ocean forecasting framework. Our model processes sequential ocean states $\text{X}^{t-3}$ through $\text{X}^{t}$ to predict $\widehat{\text{X}}^{t+1}$. (a) The main pipeline consists of a shared encoder $\mathbf{E}$ that transforms input states into latent representations, modulated by spatiotemporal features $\text{F}_S$ from network $\mathbf{M}$. The fused representations feed into the prediction module, whose outputs are processed by decoders $\mathbf{D}$ with skip connections. (b) The prediction module employs stacked attention blocks with adaptive layer normalization (AdaLN) and feed-forward networks (FFN), capturing complex temporal dynamics. (c) The Mixture-of-Time (MoT) module performs channel-wise selection across the four decoder outputs from different temporal skip connections. For each channel, MoT identifies the top-K temporal dependencies using matrix $\text{V}$ (derived from spatially-averaged MAE metrics and softmax) and computes the optimal weighted average to synthesize the final prediction $\widehat{\text{X}}^{t+1}$.
• Figure 2: Global RMSE distribution of sea surface. From top to bottom, the results represent salinity (psu), temperature (°C), and ocean current U/V components (m/s), and from left to right correspond to different forecast lead times. Each subplot represents the average RMSE (lower is better) for the test set.
• Figure 3: Depth-dependent RMSE Distributions of salinity and temperature varying with lead time. Each subplot represents the RMSE (lower is better) varying with depth at the current lead time. Blue represents salinity results, while red represents temperature results.
• Figure 4: The SST comparison of different methods based on the IV-TT evaluation framework. The x-axis represents the forecast time, and the y-axis represents the RMSE (lower is better).
• Figure 5: The ablation study of different methods on the HYCOM-RD. The five subplots respectively show salinity, temperature, ocean current UV components, and sea surface height. The x-axis represents the forecast lead time, and the y-axis represents the RMSE (lower is better). Note that for each variable, we average the RMSE over depth.