AI models are increasingly trained on geoscientific observations and simulations. A growing body of theory investigates their representations and generalisation, but many existing results rely on assumptions that are difficult to reconcile with geoscientific data. These data reflect nonlinear, multiscale and non-stationary systems, are often incomplete, and may change across spatial scales, temporal regimes and physical conditions. Understanding how learning behaves under these conditions remains an open theoretical challenge.

This session focuses on the theory of learned representations in geoscientific AI. We seek contributions that investigate how embeddings and latent spaces encode spatial, temporal and physical structure, and how these representations depend on the data, model architecture, training dynamics and physical constraints. Relevant questions include the geometry and identifiability of latent representations, inductive biases, invariance and equivariance, stability and generalisation across scales and regimes.

The session brings together geoscience, theoretical machine learning, applied mathematics, statistics, statistical physics, dynamical systems and fluid dynamics. Contributions may develop mathematical or statistical theory, introduce diagnostics derived from theoretical principles, or test theoretical predictions using observations, simulations and idealised systems. The emphasis is on explaining model behaviour rather than reporting predictive performance alone, and on determining when learned representations are stable, transferable and physically meaningful.