Work

Artificial intelligence for ocean phenomena forecasting

Year: 2026

Type: article

Abstract: Overview of the AI-based ocean phenomena forecasting system. The workflow integrates multi-source data from distinct platforms: (i) reanalysis datasets providing 3D physical fields (U, V, T, S); (ii) in situ observations from sparse platforms like Argo floats and buoys; and (iii) satellite remote sensing, which provides high-coverage surface variables (SST, SSH, and SSS). OP are categorized by dominant physical drivers and timescales: Atmosphere-forced high-frequency hazards (hourly-daily), Ocean internal mesoscale & wave dynamics (daily-weekly), and Coupled climate-cryosphere modes (monthly-yearly). The general framework for the Large Ocean Model (LOM), illustrating how AI architectures (e.g., Adaptive Fourier Neural Operators, Transformers, and Diffusion models) process these multi-source inputs to generate forecasts. Advancements in artificial intelligence (AI) are ushering in a new era of ocean forecasting. AI-based ocean forecasting falls into two categories: ocean phenomena (OP) forecasting and ocean state variable (OSV) forecasting. Compared to OSV forecasting, OP forecasting reaches a broader audience of end-users, including policymakers and the general public. Despite rapid progress in AI-driven ocean forecasting, existing studies remain fragmented across phenomena and timescales, and a comprehensive synthesis remains lacking. Hence, this paper presents in-depth reviews of different OP forecasting types, including hourly-daily: atmosphere-forced high-frequency hazards, daily-weekly/monthly: ocean-internal mesoscale & wave dynamics, and monthly-yearly: coupled climate-cryosphere modes with distinct dominant physical drivers and timescales. General AI frameworks, combined with physical knowledge and a tailored training strategy, can be effectively applied to a range of OP forecasting tasks. While OP forecasting models operate independently at the phenomenon level, OSV large models remain disconnected from OP forecasting, limiting their role in disaster mitigation and human impact assessment. To bridge the gap between OSV and OP forecasting, the Large Ocean Model (LOM) of OP was proposed as a possible means of connecting them. Finally, five essential insights into the design and practical implementation of AI models in ocean forecasting are presented.

Source: 中国科学通报：英文版

DOI: 10.1016/j.scib.2026.04.002