Description
Machine learning is reshaping how we simulate the Earth system, and data-driven emulators of the ocean and sea ice are becoming stable and skillful. Yet purely data-driven models lack physical interpretability with respect to individual dynamical and thermodynamic processes, and their generalizability to a changing climate remains uncertain. At GEOMAR we therefore also follow a second, complementary route: keep the physics and learn only what we do not know. This talk presents that route for sea ice.
Sea ice plays a central role in the climate system by regulating exchanges of heat and momentum between ocean and atmosphere. Representing its evolution numerically is challenging, as growing model complexity increases computational cost while key processes remain unresolved and their parametrizations poorly constrained. We investigate a hybrid framework that bridges numerical and data-driven modeling: a machine-learning-enabled numerical sea-ice thermodynamic model. The end-to-end differentiable implementation of zero-layer column thermodynamics in Python allows sensitivities with respect to model parameters to be computed directly. This enables gradient-based parameter optimization and, more importantly, the description of individual parametrizations by process-emulating neural-network components, jointly trained and evaluated with the numerical model against snow and ice thickness observations from ice mass-balance buoys.
These low-complexity components remain physically interpretable owing to their explicit input–output relationships and local pointwise operation, in contrast to high-dimensional ML models. The hybrid model improves the representation of snow and melt processes while retaining the stability and interpretability of the numerical backbone, a compromise between physical constraint and data-driven flexibility. We discuss what governs whether such hybrid models remain stable and physically plausible—in particular training strategy and imbalanced observational data—and close with an outlook on sea-ice physics learned directly from observations.