Trustworthy AI for Ocean Dynamics: A Semantics-Driven Approach to Southern Ocean Circulation (Invited Presentation)

Artificial Intelligence offers transformative potential in climate science, particularly in understanding complex and observationally sparse regions like the Southern Ocean. However, this potential comes with the dual challenge of making AI algorithms that are fit-for-purpose and managing limitations, such as its ‘black box’ aspect. In this talk, we present a methodology grounded in AI-semantics for new knowledge. First, an unsupervised AI framework that centers validation and uncertainty quantification using manifold mapping and clustering, identifies distinct dynamical regimes within the Southern Ocean: areas of sparsity in the equations governing the circulation. These regimes provide new, physically meaningful insights into ocean circulation. Second, emphasizing the need to critically validate neural network predictions, we use a semantics-driven neural network approach that ensures predictions are grounded in geophysical fluid dynamics, and subsequently demonstrate how this can be used to gain new knowledge. By using neural networks enhanced with explainable AI, we track these dynamical regimes using sparse available data, and use the neural network’s changes in input relevance to guide hypothesis generation and testing. We reveal how fundamental shifts in the drivers of Southern Ocean circulation are governed under climate change. This work serves both as an exploration of AI's promise and a blueprint for responsibly navigating the challenges it presents in climate science, discussing standards for reliable, transparent, and hypothesis-driven AI applications.