Using machine learning, data assimilation and their combination to improve a new generation of sea ice models
Alberto Carrassi
Novel sea-ice models treat the ice more like a brittle solid than a fluid, thus making possible to simulate the sea-ice multifractal and anisotropic character. Nevertheless, they still are very sensitive to their uncertain parameters and rely on parametrizations for key processes, such as, but not only the melt ponds. We will show how state-of-the-art ensemble variational data assimilation methods can be successful in estimating key parameters using observation of the physical quantities. Furthermore most data assimilation methods rely on a quasi-linear and Gaussian assumption and fail to preserve fundamental physical and dynamical balances in the sea ice. We will present a novel combined machine learning and data assimilation method whereby an ensemble Kalman filter is hybridised with a variational auto-encoder. Finally, we investigate the combination of geophysical sea-ice models together with neural networks in a hybrid modelling setup. On the one hand, deep learning can surrogate computationally expensive sea-ice models, on the other hand, deep learning is used to learn parametrization of the sea-ice melt ponds that have a major role on the albedo and thus on the general energy balance.