Ice flow numerical models are essential for predicting the evolution of ice sheets in a warming climate. Recent research emphasizes the need for higher-order and even full-Stokes flow models instead of the traditional Shallow-Ice Approximation whose assumptions are not valid in certain critical but spatially limited areas. These higher-order models are however computationally intensive and difficult to use at the continental scale. The purpose of this work, therefore, is to develop a new technique that reduces the computational cost of ice flow models while maximizing their accuracy. To this end, several ice flow models of varying order of complexity have been implemented in the Ice Sheet System Model, a massively parallelized finite element software developed at the Jet Propulsion Laboratory. Analysis and comparison of model results on both synthetic and real geometries shows that sophisticated models are only needed in the grounding line area, transition between grounded and floating ice, whereas simpler models yield accurate results in most of the model domain. There is therefore a strong need for coupling such models in order to balance computational cost and physical accuracy. Several techniques and frameworks dedicated to model coupling already exist and are investigated. A new technique adapted to the specificities of ice flow models is developed: the Tiling method, a multi-model computation strategy based on the superposition and linking of different numerical models. A mathematical analysis of a mixed Tiling formulation is first performed to define the conditions of application. The treatment of the junction between full-Stokes and simpler models that decouple horizontal and vertical equation is then elaborated in order to rigorously combine all velocity components. This method is finally implemented in the Ice Sheet System Model to design hybrid models that combine several ice flow approximations of varying order of complexity. Following a validation on synthetic geometries, this method is applied to real cases, such as Pine Island Glacier, in West Antarctica, to illustrate its relevance. Hybrid models have the potential to significantly improve physical accuracy by combining models in their domain of validity, while preserving the computational cost and being compatible with the actual computational resources.
