Learning sea-ice physics from data: a hybrid ML-numerical modelling framework

Sea ice remains challenging to predict across spatial and temporal scales, from hourly floe-scale motion to seasonal regional forecasts and long-term climate projections. Increasingly complex numerical models have been developed to represent the strongly nonlinear dynamics and thermodynamics of sea ice, but their growing computational cost together with uncertainties in initial conditions and unresolved physical processes continues to limit their applicability. In recent years, machine-learning approaches trained on satellite observations or numerical model output have shown promise in predicting sea-ice evolution. However, while often accurate, such purely data-driven models provide limited physical interpretability, hindering the analysis of individual dynamic and thermodynamic processes and their response to climate change. Here we present a hybrid modelling framework that bridges these two approaches: a machine learning-enabled numerical sea-ice model. At its core is a differentiable implementation of a dynamic-thermodynamic sea-ice model in Python, which allows the computation of sensitivities with respect to model parameters as well as initial and boundary conditions. This enables systematic parameter optimization against observations and, more importantly, facilitates the replacement of individual parameterizations with lightweight machine-learning components. Once trained, these lightweight components remain physically interpretable due to their low complexity, explicit input-output relationships, and strictly local (pointwise) operation, in contrast to black-box, high-dimensional ML models. The hybrid model is embedded in an efficient data-loading infrastructure that provides access to diverse observational data sources, including satellite, buoy, and in-situ measurements, for training and evaluation. We demonstrate the capabilities of this framework by comparing alternative data-driven thermodynamic parameterizations for sea-ice and snow growth and melting rates trained on buoy and satellite data, and by assessing their impact on large-scale sea-ice evolution.