OceanNet: a principled neural operator-based digital twin for regional oceans.

Area of Science:

• Oceanography
• Artificial Intelligence
• Computational Fluid Dynamics

Background:

• Ocean modeling faces challenges from complex bathymetry, land interactions, vertical structure, and flow non-linearity.
• Data-driven methods show promise but require adaptation for oceanic complexities.
• Existing numerical ocean models are computationally intensive.

Purpose of the Study:

• To introduce OceanNet, a neural operator-based digital twin for regional sea surface height emulation.
• To address challenges in ocean modeling using a physics-inspired deep learning approach.
• To enable cost-effective and efficient seasonal prediction of ocean currents.

Main Methods:

• Utilized a Fourier neural operator architecture for sea surface height emulation.
• Implemented a predictor-evaluate-corrector integration scheme to enhance stability and mitigate error growth.
• Incorporated a spectral regularizer to counteract spectral bias at smaller scales.
• Trained OceanNet on historical sea surface height data for the northwest Atlantic Ocean.

Main Results:

• OceanNet achieved competitive forecast skill compared to a state-of-the-art dynamical ocean model.
• Demonstrated a significant reduction in computational cost, by 500,000 times.
• Successfully emulated sea surface height for the Gulf Stream and Loop Current eddies.

Conclusions:

• Physics-inspired deep neural operators offer a viable and efficient alternative to high-resolution numerical ocean models.
• OceanNet represents a significant advancement in data-driven ocean modeling and forecasting.
• The approach shows potential for broader applications in climate and oceanographic research.