OceanRep: A Foundation Model for Ocean Dynamics

OceanRep proposes a novel AI foundation model for ocean dynamics, a cornerstone for understanding and predicting climate change. Inspired by the success of AtmoRep, a deep learning model for atmospheric dynamics, OceanRep seeks to extend this framework to the ocean. In order to leverage transformer models and large-scale, multi-resolution oceanographic data (e.g., from ocean model FESOM2), the design is based on vision transformers, modified to handle four-dimensional data represented by space-time tokens, and with a U-net-type backbone to capture intricate interactions within the ocean system. For pre-training, BERT-style masking is used.

Preliminary results demonstrate OceanRep's ability to generate skillful week scale forecasts using data from a 1-degree resolution FESOM2 simulation. Ultimately, the project aims to create a robust model capable of simulating ocean and sea ice dynamics over decades. This will allow for extensive numerical experimentation and rapid generation of accurate "what-if'' scenarios. These capabilities hold immense value for climate adaptation strategies, policy development, and scientific exploration of the intricate dynamics governing the Earth system.