文章背景与核心概要

全球海洋预报对于气候监测、海上航行以及极端事件预警至关重要。传统的基于物理的海洋模型计算成本极高，而现有的深度学习方法主要依赖于规则网格（structured grids），导致计算资源被掩码陆地单元白白浪费，且无法根据局部流场的复杂性在动态异构的海洋区域施加自适应分辨率。

为了克服这些局限性，本文推出了 OceanLight——一个创新性地将几何自适应非结构网格标记化（tokenization）与图神经网络（GNN）骨干网络相结合的高效全球海洋预报框架。OceanLight 在逐点预报准确率和动能谱保真度上均超越了业务化数值分析以及现有的 AI 模型，同时在地转平衡一致性方面超越了所有现有的 AI 海洋模型。此外，它还能可靠地表征中尺度涡旋，捕捉超越逐点统计优化的相干海洋结构。在实现这些卓越性能的同时，相较于传统的规则网格基线，OceanLight 的 GPU 内存消耗降低了 62%，浮点运算量（FLOPs）降低了 70%，为可扩展的数据驱动海洋学建立了一个全新的通用范式。

OceanLight: Efficient Global Ocean Forecasting via Geometry-Adaptive Unstructured Mesh Representation

arXiv ID: 2608.16070
Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
Submission Date: August 17, 2026
Authors: Wei Wu, Xiang Wang, Hongze Leng, Qingye Min, Junxing Zhu, Junqiang Song

📌 Summary

OceanLight is a state-of-the-art deep learning framework designed for efficient and reliable global ocean forecasting. Traditional physics-based models are computationally expensive, while existing deep learning approaches primarily rely on structured grids, wasting compute resources on masked land cells and enforcing uniform resolutions across dynamic ocean regions.

To overcome these limitations, OceanLight introduces a geometry-adaptive unstructured mesh tokenization coupled with a Graph Neural Network (GNN) backbone. This design yields superior pointwise forecast accuracy and kinetic energy spectral fidelity compared to both operational numerical analyses and existing AI models. Furthermore, it excels in geostrophic balance consistency and accurately represents mesoscale eddies beyond simple pointwise statistical optimization—all while reducing GPU memory consumption by 62% and FLOPs by 70% relative to traditional structured-grid baselines.

🧭 Abstract

可靠的全球海洋预报对于气候监测、海上导航和极端事件早期预警至关重要。基于物理的海洋预报模型带来了沉重的计算成本，而现有的深度学习方法主要依赖于规则网格架构，不仅在被掩码的陆地单元上产生了不必要的计算，还在动态异构的海洋区域中强制使用统一分辨率，而忽略了局部的流场复杂度。在这里，我们推出了 OceanLight，这是一个高效的全球海洋预报框架，它创新性地将几何自适应非结构网格标记化与图神经网络（GNN）骨干相结合。OceanLight 的逐点预报准确率和动能谱保真度超越了业务数值分析和最先进的基于 AI 的模型，同时在地转平衡一致性上超越了所有基于 AI 的海洋模型。此外，OceanLight 表现出可靠的中尺度涡旋表征能力，能够捕捉超越逐点统计优化的相干海洋结构。与规则网格基线相比，这些性能的实现伴随着 62% 的 GPU 内存消耗降低和 70% 的 FLOPs 降低。我们的非结构网格表示为可扩展的数据驱动海洋学建立了一个可泛化的范式。

Reliable global ocean forecasting is critical for climate monitoring, marine navigation, and extreme event early warning. Physics-based ocean forecasting models impose prohibitive computational costs, while existing deep learning approaches predominantly rely on structured-grid architectures, incurring unnecessary computation on masked land cells and enforcing uniform resolution across dynamically heterogeneous ocean regions regardless of local flow complexity. Here we present OceanLight, an efficient global ocean forecasting framework innovatively combining geometry-adaptive unstructured mesh tokenization with a graph neural network (GNN) backbone. OceanLight achieves pointwise forecast accuracy and kinetic energy spectral fidelity exceeding both operational numerical analyses and state-of-the-art AI-based models, while surpassing all AI-based ocean models in geostrophic balance consistency. Furthermore, OceanLight demonstrates reliable mesoscale eddy representation, capturing coherent ocean structures beyond pointwise statistical optimization. These capabilities are delivered with a 62% reduction in GPU memory consumption and 70% reduction in FLOPs relative to structured-grid baselines. Our unstructured mesh representation establishes a generalizable paradigm for scalable data-driven oceanography.

📋 Document Metadata

🔗 Full-Text & External Resources

• Access Paper: View PDF | HTML (Experimental) | TeX Source
• Citations & References: Google Scholar | Semantic Scholar | NASA ADS