[Paper Review] AI-GOMS: Large AI-Driven Global Ocean Modeling System

Wei Xiong, Yanfei Xiang+5 more|arXiv (Cornell University)|Aug 6, 2023

Oceanographic and Atmospheric Processes16 citations

TL;DR

AI-GOMS presents a large AI-driven global ocean modeling framework with a backbone Fourier-based Masked Autoencoder and lightweight downstream modules for regional downscaling, wave decoding, and biochemistry coupling, enabling 30-day global ocean prediction at 1/4° with 15 depth layers and transferability to downstream tasks.

ABSTRACT

Ocean modeling is a powerful tool for simulating the physical, chemical, and biological processes of the ocean, which is the foundation for marine science research and operational oceanography. Modern numerical ocean modeling mainly consists of governing equations and numerical algorithms. Nonlinear instability, computational expense, low reusability efficiency and high coupling costs have gradually become the main bottlenecks for the further development of numerical ocean modeling. Recently, artificial intelligence-based modeling in scientific computing has shown revolutionary potential for digital twins and scientific simulations, but the bottlenecks of numerical ocean modeling have not been further solved. Here, we present AI-GOMS, a large AI-driven global ocean modeling system, for accurate and efficient global ocean daily prediction. AI-GOMS consists of a backbone model with the Fourier-based Masked Autoencoder structure for basic ocean variable prediction and lightweight fine-tuning models incorporating regional downscaling, wave decoding, and biochemistry coupling modules. AI-GOMS has achieved the best performance in 30 days of prediction for the global ocean basic variables with 15 depth layers at 1/4° spatial resolution. Beyond the good performance in statistical metrics, AI-GOMS realizes the simulation of mesoscale eddies in the Kuroshio region at 1/12° spatial resolution and ocean stratification in the tropical Pacific Ocean. AI-GOMS provides a new backbone-downstream paradigm for Earth system modeling, which makes the system transferable, scalable and reusable.

Motivation & Objective

• Motivate the need for data-driven approaches to overcome bottlenecks in traditional numerical ocean modeling, such as nonlinear instability and high coupling costs.
• Propose a large AI-driven framework (AI-GOMS) that can learn global ocean dynamics and provide accurate daily predictions.
• Demonstrate transferability of a trained backbone model to downstream ocean-related tasks with low fine-tuning cost.

Proposed method

• Introduce a backbone model with an asymmetric encoder–decoder using Fourier-based attention blocks and a random patch-masking strategy.
• Enable multi-source data input by using a patch-embedding and patch-recovery scheme that supports 2D, 3D, and sparse inputs and allows data assimilation.
• Train the backbone on HYCOM global reanalysis data to predict five basic ocean variables across 15 depth layers at 1/4° resolution.
• Develop lightweight downstream modules for regional downscaling, wave decoding, and biochemistry coupling that can be fine-tuned with low cost.
• Use a backbone–downstream architecture where downstream modules ingest backbone features plus scenario-specific inputs for task-specific predictions.

Experimental results

Research questions

• RQ1 Can a large AI model accurately simulate global ocean variables and their vertical profiles over 30 days at 1/4° resolution?
• RQ2 Does a lightweight downstream fine-tuning approach enable accurate regional downscaling, wave height decoding, and biochemical variable prediction with low training cost?
• RQ3 Is the backbone model transferable to downstream ocean-related tasks while maintaining physical consistency (e.g., eddy structures and stratification)?

Key findings

• The backbone model predicts five ocean variables daily across 15 depth layers at 1/4° resolution with strong long-term prediction capability.
• AI-GOMS achieves better latitude-weighted ACC and RMSE than FourCastNet for all variables over 30 days of prediction.
• The regional downscaling module resolves mesoscale eddies in the Kuroshio region at 1/12° with ACC > 0.6 over 7 days for velocity and SSH.
• The wave decoding module enables 30-day significant wave height prediction using lightweight fine-tuning.
• The biochemistry coupling module predicts eight biochemical variables by fusing backbone features with biochemical conditions via a lightweight fine-tuning model.
• The design supports transferability of the backbone to downstream scenarios with low fine-tuning cost.