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

Yuan Gao, Hao Wu, Fan Xu, Yanfei Xiang, Ruijian Gou, Ruiqi Shu, Qingsong Wen, Xian Wu, Kun Wang, Xiaomeng Huang · 2025

Read NeuralOM : Neural Ocean Model for Subseasonal-to-Seasonal Simulation by Yuan Gao, Hao Wu, Fan Xu, Yanfei Xiang, Ruijian Gou, Ruiqi Shu, Qingsong Wen, Xian Wu, Kun Wang, Xiaomeng Huang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-21020.txt.

Introduction

Yuan Gao and co-authors present NeuralOM, a neural model that simulates the ocean weeks to months ahead. Forecasting slow systems such as oceans and climate is hard for learning models, because small errors pile up until the forecast drifts. Their 2025 framework attacks that drift with two ideas built into one neural operator design.

First, a staged correction scheme breaks each forecast step into fine refinement passes, which keeps errors from compounding over long runs. Second, a graph network carries built-in message rules that mirror physics, such as flows down a gradient, which keeps the model honest and cheap to run. On global ocean tests at seasonal range, the model beat leading rivals in accuracy and steadiness. At a sixty-day lead it cut the main error score by about thirteen percent against the best rival.

Catalog record

CREATORYuan Gao, Hao Wu, Fan Xu, Yanfei Xiang, Ruijian Gou, Ruiqi Shu, Qingsong Wen, Xian Wu, Kun Wang, Xiaomeng Huang
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE49 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-21020

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

NeuralOM : Neural Ocean Model for Subseasonal-to-Seasonal Simulation Yuan Gao, Hao Wu, Fan Xu, Yanfei Xiang, Ruijian Gou, Ruiqi Shu, Qingsong Wen, Xian Wu, Kun Wang, Xiaomeng Huang 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. 1 Tsinghua University, 2 Shenzhen Loop Area Institute, 3 Ocean University of China, 4 Squirrel Ai Learning, 5 Tencent, 6 Nanyang Technological University {yuangao24, wu-h25, xiangyf22, srq24}@mails.tsinghua.edu.cn; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected] Code — https://github.com/YuanGao-YG/NeuralOM Introduction The long-term simulation of complex physical systems ( Wang et al. 2024b ; Wu et al. 2025c ) , such as the global ocean and climate, is a cornerstone of modern science yet remains a great challenge. While machine learning (ML) models offer a computationally efficient path to complement traditional numerical solvers, they face a fundamental bottleneck in such tasks: long-term stability . This issue is especially severe in slow-changing systems , where the state evolves in a subtle manner. In standard autoregressive frameworks, minor prediction errors at each step inevitably accumulate, causing the simulation to diverge catastrophically from the true physical dynamics, a well-known problem of compounding error . This error cascade often leads data-driven models to fail in long-term rollouts. As shown in Figure 1 (bottom left), a state-of-the-art baseline model generates substantial, physically unrealistic artifacts in its long-range forecast. 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). To address this challenge, existing machine learning approaches have been explored, but they are limited in two key aspects. ❶. First , they often struggle to capture the fine-grained dynamics inherent to slow-changing systems. Many models are designed to predict sharp, significant shifts, making them ill-suited for learning the small, incremental updates that define ocean dynamics. This typically results in either over-smoothing the predictions or failing to capture critical physical anomalies. ❷. Second , the internal mechanisms of these models are often physics-agnostic. For instance, standard message passing in Graph Neural Networks (GNNs) ( Keisler 2022 ) treats all node interactions uniformly, unable to distinguish between fundamentally different physical processes like gradient-driven flows and multiplicative couplings between variables. Furthermore, they lack the ability to adaptively model phenomena across different spatio-temporal scales, from large-scale, slow-moving currents to small-scale, energetic eddies. To systematically address these issues, we introduce NeuralOM , a neural operator framework designed for stable, high-fidelity simulation of slow-changing dynamics. NeuralOM ’s core is built on two key innovations. First, we propose a Progressive Residual Correction Framework . This framework decomposes the complex forecasting task into a cascade of refinement steps, where each step focuses on learning and correcting the residual error from the previous one. This strategy of “small, iterative corrections” enables the model to capture extremely fine-grained dynamic changes, thereby effectively suppressing long-term error accumulation. Second, we design a novel Physics-Guided Graph Network with a built-in adaptive messaging mechanism . It incorporates physical priors into the network by explicitly modeling gradient-driven flows and multiplicative couplings between variables. Crucially, its aggregation scheme is not fixed but dynamically adjusts, allowing it to inherently handle multi-scale dynamics. We validate NeuralOM on the challenging benchmark of global Subseasonal-to-Seasonal (S2S) ocean simulation and forecasting. Extensive experiments provide strong evidence of its superior performance: Experimental Observations . ❶ State-of-the-art accuracy in long-term simulation : As shown in Table 1 , in 60 -day simulation, NeuralOM achieves a normalized RMSE of just 0.7014 , a 13.3 % reduction compared to the best-performing baseline, WenHai ( 0.8091 ). It also improves the Anomaly Correlation Coefficient (ACC) by 12.0 % , indicating more accurate tracking of the physical field’s evolution. ❷ Superior visual fidelity and physical consistency : As illustrated in Figure 1 and Figure 3 , unlike baseline models like CirT that collapse due to error explosion, NeuralOM produces physically consistent forecasts that remain highly aligned with the ground truth even at 60 days, successfully preserving critical fine-scale structures. ❸ Validated effectiveness of our designs through ablation studies : As detailed in Table 2 and Table 3 , removing any of our core components, either the Progressive Residual Correction Framework or the Physics-Guided Adaptive Graph Messaging, leads to a significant drop in performance. In particular, incorporating climatology priors improves the ACC metric by more than 5 -fold , highlighting that every design choice is crucial to achieve accurate and stable simulations. In summary, our main contributions are threefold: (1) a progressive residual correction framework designed for slow-changing systems; (2) a physics-guided graph network with an adaptive, multi-scale messaging mechanism; and (3) state-of-the-art performance on the S2S ocean simulation benchmark, establishing a new paradigm for data-driven scientific computing. Related Work In recent years, previous works in spatio-temporal mining have achieved stupendous achievements ( Pfaff et al. 2020 ; Li et al. 2020 ; Wu et al. 2023 ; Wu et al. 2024d ; Wang et al. 2024a ; Wu et al. 2024b ; Wu et al. 2024a ; Wu et al. 2025b ; Wu et al. 2025a ; Wu et al. 2025d ; Jia et al. 2025 ) . However, many of these methods are only applicable in idealized or relative simple scenario, whereas ocean system is a complex dynamical system in the real world, requiring the design of specialized methods for accurate prediction ( Wu et al. 2024c ) . Research in data-driven ocean simulation is primarily divided into two streams: task-specific predictive models and general-purpose global foundation models. Task-Specific ML Models. These models aim to simulate specific oceanic or atmospheric phenomena and have achieved impressive results in their respective domains. Examples include work on El Nio-Southern Oscillation (ENSO) prediction ( Ham et al. 2019 ; Hu et al. 2021 ; Chang et al. 2023 ; Lyu et al. 2024 ; Chen et al. 2025 ) , as well as models for the

Read the complete text