Large data pre-trained neural networks have shown outstanding performance on downstream tasks in computer vision and natural language understanding. Such foundation models are rapidly transforming the way we address challenges in the design, development and deployment of neural network systems. Pre-training of neural networks on large geophysical datasets offers new ways to integrate observations and simulations, and opens new pathways for rapid model development. Here, I will present recent advances in the development of a joint ocean and atmospheric foundation model and will demonstrate how these models can be adapted and applied to support a range of downstream tasks.

Our pre-trained model functions as a single adaptable system, removing the need to develop bespoke models for every application. This reduces both data requirements and development time while improving consistency across diverse information sources. Once trained, the model can be applied or fine-tuned for tasks such as characterisation through embeddings, global and regional forecasting, and traditional computer vision applications including classification, segmentation, and tracking.

Beyond the technical benefits, foundation models simplify the development of machine learning tools, shorten the path from research to application, and allow the same model to be applied across different datasets and applications.

Biography:

Flowershift is an Australian company developing machine learning and artificial intelligence applications for environmental intelligence. Justin Freeman will present flowershift’s development of ocean–atmosphere foundation models and their application to operational oceanography.