NeuralOM: Neural Ocean Model for Subseasonal-to-Seasonal Simulation

TL;DR

NeuralOM tackles the challenge of accurate, long-horizon ocean simulations by addressing error accumulation through a Progressive Residual Correction framework and enforcing physical consistency with a Physics-Guided Graph Network. The method combines a multi-stage residual refinement process with physics-informed edge interactions and a dynamic, multi-scale aggregation scheme to capture both fine-scale and large-scale dynamics. Empirical results on global S2S ocean forecasting show state-of-the-art long-term fidelity, robustness under realistic forcings, and improved performance for extreme events, with ablations confirming the necessity of climatology priors and each architectural component. The work offers a practical, scalable approach for data-driven scientific computing in oceanography and climate, with potential extensions to enforce conservation laws and reduce computational cost.

Abstract

Long-term, high-fidelity simulation of slow-changing physical systems, such as the ocean and climate, presents a fundamental challenge in scientific computing. Traditional autoregressive machine learning models often fail in these tasks as minor errors accumulate and lead to rapid forecast degradation. To address this problem, we propose NeuralOM, a general neural operator framework designed for simulating complex, slow-changing dynamics. NeuralOM's core consists of two key innovations: (1) a Progressive Residual Correction Framework that decomposes the forecasting task into a series of fine-grained refinement steps, effectively suppressing long-term error accumulation; and (2) a Physics-Guided Graph Network whose built-in adaptive messaging mechanism explicitly models multi-scale physical interactions, such as gradient-driven flows and multiplicative couplings, thereby enhancing physical consistency while maintaining computational efficiency. We validate NeuralOM on the challenging task of global Subseasonal-to-Seasonal (S2S) ocean simulation. Extensive experiments demonstrate that NeuralOM not only surpasses state-of-the-art models in forecast accuracy and long-term stability, but also excels in simulating extreme events. For instance, at a 60-day lead time, NeuralOM achieves a 13.3% lower RMSE compared to the best-performing baseline, offering a stable, efficient, and physically-aware paradigm for data-driven scientific computing. Code link: https://github.com/YuanGao-YG/NeuralOM.

Figures (12)

• Figure 1: NeuralOM overcomes the long-term instability challenge in simulating slow-changing physical systems. Shown is a 60-day Subseasonal-to-Seasonal (S2S) simulation for sea surface temperature anomaly. (Bottom Left) A state-of-the-art baseline (CirT) collapses due to compounding errors, producing physically unrealistic artifacts and failing to capture the ocean state. (Bottom Right) In contrast, NeuralOM maintains high fidelity and long-term stability, accurately reproducing the complex patterns of the ground-truth, as highlighted in the detailed view (red dashed lines).
• Figure 2: Overview of Our NeuralOM. (a) The overall architecture of the progressive residual correction framework, input variables (subtracting climatology for periodic variables), progressive residual correction stage, and visualization of simulation ocean variables; (b) The proposed physics-guided graph network; (c) The global ocean simulation module uses a rollout approach to generate future results.
• Figure 3:
• Figure 4: The latitude-weighted RMSE and ACC results of several important ocean surface variables.
• Figure 5: Performance on forecasting and extreme event assessment.(a) Visualizations of 30-day forecasts. Our model's outputs align closely with the ground truth, while baseline models exhibit significant artifacts. (b) CSI and SEDI scores for extreme surface current events. NeuralOM achieves higher scores than baselines at both 30 and 60-day lead times.